Launch multiple wal-redo postgres instances

Too bad we don't have a --dry-run in PRs :(

.circleci/config.yml

@@ -1,20 +1,19 @@
 version: 2.1
 
-orbs:
-  python: circleci/python@1.4.0
-
 executors:
   zenith-build-executor:
     resource_class: xlarge
     docker:
-      - image: cimg/rust:1.52.1
+      - image: cimg/rust:1.56.1
+  zenith-python-executor:
+    docker:
+      - image: cimg/python:3.7.10  # Oldest available 3.7 with Ubuntu 20.04 (for GLIBC and Rust) at CirlceCI
 
 jobs:
-  check-codestyle:
+  check-codestyle-rust:
     executor: zenith-build-executor
     steps:
       - checkout
-
       - run:
           name: rustfmt
           when: always
@@ -81,6 +80,8 @@ jobs:
       build_type:
         type: enum
         enum: ["debug", "release"]
+    environment:
+      BUILD_TYPE: << parameters.build_type >>
     steps:
       - run:
           name: apt install dependencies
@@ -110,25 +111,26 @@ jobs:
             # Require an exact match. While an out of date cache might speed up the build,
             # there's no way to clean out old packages, so the cache grows every time something
             # changes.
-            - v03-rust-cache-deps-<< parameters.build_type >>-{{ checksum "Cargo.lock" }}
+            - v04-rust-cache-deps-<< parameters.build_type >>-{{ checksum "Cargo.lock" }}
 
         # Build the rust code, including test binaries
       - run:
           name: Rust build << parameters.build_type >>
           command: |
-            export CARGO_INCREMENTAL=0
-            BUILD_TYPE="<< parameters.build_type >>"
             if [[ $BUILD_TYPE == "debug" ]]; then
-              echo "Build in debug mode"
-              cargo build --bins --tests
+              cov_prefix=(scripts/coverage "--profraw-prefix=$CIRCLE_JOB" --dir=/tmp/zenith/coverage run)
+              CARGO_FLAGS=
             elif [[ $BUILD_TYPE == "release" ]]; then
-              echo "Build in release mode"
-              cargo build --release --bins --tests
+              cov_prefix=()
+              CARGO_FLAGS=--release
             fi
 
+            export CARGO_INCREMENTAL=0
+            "${cov_prefix[@]}" cargo build $CARGO_FLAGS --bins --tests
+
       - save_cache:
           name: Save rust cache
-          key: v03-rust-cache-deps-<< parameters.build_type >>-{{ checksum "Cargo.lock" }}
+          key: v04-rust-cache-deps-<< parameters.build_type >>-{{ checksum "Cargo.lock" }}
           paths:
             - ~/.cargo/registry
             - ~/.cargo/git
@@ -138,53 +140,100 @@ jobs:
         # has to run separately from cargo fmt section
         # since needs to run with dependencies
       - run:
-          name: clippy
+          name: cargo clippy
           command: |
-            ./run_clippy.sh
+            if [[ $BUILD_TYPE == "debug" ]]; then
+              cov_prefix=(scripts/coverage "--profraw-prefix=$CIRCLE_JOB" --dir=/tmp/zenith/coverage run)
+            elif [[ $BUILD_TYPE == "release" ]]; then
+              cov_prefix=()
+            fi
+
+            "${cov_prefix[@]}" ./run_clippy.sh
 
         # Run rust unit tests
-      - run: cargo test
+      - run:
+          name: cargo test
+          command: |
+            if [[ $BUILD_TYPE == "debug" ]]; then
+              cov_prefix=(scripts/coverage "--profraw-prefix=$CIRCLE_JOB" --dir=/tmp/zenith/coverage run)
+            elif [[ $BUILD_TYPE == "release" ]]; then
+              cov_prefix=()
+            fi
+
+            "${cov_prefix[@]}" cargo test
 
         # Install the rust binaries, for use by test jobs
-        # `--locked` is required; otherwise, `cargo install` will ignore Cargo.lock.
-        # FIXME: this is a really silly way to install; maybe we should just output
-        # a tarball as an artifact? Or a .deb package?
       - run:
-          name: cargo install
+          name: Install rust binaries
           command: |
-            export CARGO_INCREMENTAL=0
-            BUILD_TYPE="<< parameters.build_type >>"
             if [[ $BUILD_TYPE == "debug" ]]; then
-              echo "Install debug mode"
-              CARGO_FLAGS="--debug"
+              cov_prefix=(scripts/coverage "--profraw-prefix=$CIRCLE_JOB" --dir=/tmp/zenith/coverage run)
             elif [[ $BUILD_TYPE == "release" ]]; then
-              echo "Install release mode"
-              # The default is release mode; there is no --release flag.
-              CARGO_FLAGS=""
+              cov_prefix=()
+            fi
+
+            binaries=$(
+              "${cov_prefix[@]}" cargo metadata --format-version=1 --no-deps |
+              jq -r '.packages[].targets[] | select(.kind | index("bin")) | .name'
+            )
+
+            test_exe_paths=$(
+              "${cov_prefix[@]}" cargo test --message-format=json --no-run |
+              jq -r '.executable | select(. != null)'
+            )
+
+            mkdir -p /tmp/zenith/bin
+            mkdir -p /tmp/zenith/test_bin
+            mkdir -p /tmp/zenith/etc
+
+            # Install target binaries
+            for bin in $binaries; do
+              SRC=target/$BUILD_TYPE/$bin
+              DST=/tmp/zenith/bin/$bin
+              cp $SRC $DST
+              echo $DST >> /tmp/zenith/etc/binaries.list
+            done
+
+            # Install test executables (for code coverage)
+            if [[ $BUILD_TYPE == "debug" ]]; then
+              for bin in $test_exe_paths; do
+                SRC=$bin
+                DST=/tmp/zenith/test_bin/$(basename $bin)
+                cp $SRC $DST
+                echo $DST >> /tmp/zenith/etc/binaries.list
+              done
             fi
-            cargo install $CARGO_FLAGS --locked --root /tmp/zenith --path pageserver
-            cargo install $CARGO_FLAGS --locked --root /tmp/zenith --path walkeeper
-            cargo install $CARGO_FLAGS --locked --root /tmp/zenith --path zenith
 
         # Install the postgres binaries, for use by test jobs
-        # FIXME: this is a silly way to do "install"; maybe just output a standard
-        # postgres package, whatever the favored form is (tarball? .deb package?)
-        # Note that pg_regress needs some build artifacts that probably aren't
-        # in the usual package...?
       - run:
-          name: postgres install
+          name: Install postgres binaries
           command: |
             cp -a tmp_install /tmp/zenith/pg_install
 
-        # Save the rust output binaries for other jobs in this workflow.
+        # Save the rust binaries and coverage data for other jobs in this workflow.
       - persist_to_workspace:
           root: /tmp/zenith
           paths:
             - "*"
 
+  check-codestyle-python:
+    executor: zenith-python-executor
+    steps:
+      - checkout
+      - run:
+          name: Install deps
+          command: pipenv --python 3.7 install --dev
+      - run:
+          name: Run yapf to ensure code format
+          when: always
+          command: pipenv run yapf --recursive --diff .
+      - run:
+          name: Run mypy to check types
+          when: always
+          command: pipenv run mypy .
+
   run-pytest:
-    #description: "Run pytest"
-    executor: python/default
+    executor: zenith-python-executor
     parameters:
       # pytest args to specify the tests to run.
       #
@@ -210,6 +259,11 @@ jobs:
       run_in_parallel:
         type: boolean
         default: true
+      save_perf_report:
+        type: boolean
+        default: false
+    environment:
+      BUILD_TYPE: << parameters.build_type >>
     steps:
       - attach_workspace:
           at: /tmp/zenith
@@ -219,20 +273,26 @@ jobs:
           steps:
             - run: git submodule update --init --depth 1
       - run:
-          name: Install pipenv & deps
-          working_directory: test_runner
-          command: |
-            pip install pipenv
-            pipenv install
+          name: Install deps
+          command: pipenv --python 3.7 install
       - run:
           name: Run pytest
-          working_directory: test_runner
+          # pytest doesn't output test logs in real time, so CI job may fail with
+          # `Too long with no output` error, if a test is running for a long time.
+          # In that case, tests should have internal timeouts that are less than
+          # no_output_timeout, specified here.
+          no_output_timeout: 10m
           environment:
             - ZENITH_BIN: /tmp/zenith/bin
             - POSTGRES_DISTRIB_DIR: /tmp/zenith/pg_install
             - TEST_OUTPUT: /tmp/test_output
+            # this variable will be embedded in perf test report
+            # and is needed to distinguish different environments
+            - PLATFORM: zenith-local-ci
           command: |
-            TEST_SELECTION="<< parameters.test_selection >>"
+            PERF_REPORT_DIR="$(realpath test_runner/perf-report-local)"
+
+            TEST_SELECTION="test_runner/<< parameters.test_selection >>"
             EXTRA_PARAMS="<< parameters.extra_params >>"
             if [ -z "$TEST_SELECTION" ]; then
               echo "test_selection must be set"
@@ -240,18 +300,47 @@ jobs:
             fi
             if << parameters.run_in_parallel >>; then
               EXTRA_PARAMS="-n4 $EXTRA_PARAMS"
-            fi;
+            fi
+            if << parameters.save_perf_report >>; then
+              if [[ $CIRCLE_BRANCH == "main" ]]; then
+                mkdir -p "$PERF_REPORT_DIR"
+                EXTRA_PARAMS="--out-dir $PERF_REPORT_DIR $EXTRA_PARAMS"
+              fi
+            fi
+
+            export GITHUB_SHA=$CIRCLE_SHA1
+
+            if [[ $BUILD_TYPE == "debug" ]]; then
+              cov_prefix=(scripts/coverage "--profraw-prefix=$CIRCLE_JOB" --dir=/tmp/zenith/coverage run)
+            elif [[ $BUILD_TYPE == "release" ]]; then
+              cov_prefix=()
+            fi
+
             # Run the tests.
             #
             # The junit.xml file allows CircleCI to display more fine-grained test information
             # in its "Tests" tab in the results page.
-            # -s prevents pytest from capturing output, which helps to see
-            # what's going on if the test hangs
             # --verbose prints name of each test (helpful when there are
             # multiple tests in one file)
             # -rA prints summary in the end
             # -n4 uses four processes to run tests via pytest-xdist
-            pipenv run pytest --junitxml=$TEST_OUTPUT/junit.xml --tb=short -s --verbose -rA $TEST_SELECTION $EXTRA_PARAMS
+            # -s is not used to prevent pytest from capturing output, because tests are running
+            # in parallel and logs are mixed between different tests
+            "${cov_prefix[@]}" pipenv run pytest \
+              --junitxml=$TEST_OUTPUT/junit.xml \
+              --tb=short \
+              --verbose \
+              -m "not remote_cluster" \
+              -rA $TEST_SELECTION $EXTRA_PARAMS
+
+            if << parameters.save_perf_report >>; then
+              if [[ $CIRCLE_BRANCH == "main" ]]; then
+                # TODO: reuse scripts/git-upload
+                export REPORT_FROM="$PERF_REPORT_DIR"
+                export REPORT_TO=local
+                scripts/generate_and_push_perf_report.sh
+              fi
+            fi
       - run:
           # CircleCI artifacts are preserved one file at a time, so skipping
           # this step isn't a good idea. If you want to extract the
@@ -260,13 +349,72 @@ jobs:
           when: always
           command: |
             du -sh /tmp/test_output/*
-            find /tmp/test_output -type f ! -name "pg.log" ! -name "pageserver.log" ! -name "wal_acceptor.log" ! -name "regression.diffs" ! -name "junit.xml" ! -name "*.filediff" ! -name "*.stdout" ! -name "*.stderr" -delete
+            find /tmp/test_output -type f ! -name "pg.log" ! -name "pageserver.log" ! -name "safekeeper.log" ! -name "regression.diffs" ! -name "junit.xml" ! -name "*.filediff" ! -name "*.stdout" ! -name "*.stderr" -delete
             du -sh /tmp/test_output/*
       - store_artifacts:
           path: /tmp/test_output
       # The store_test_results step tells CircleCI where to find the junit.xml file.
       - store_test_results:
           path: /tmp/test_output
+      # Save coverage data (if any)
+      - persist_to_workspace:
+          root: /tmp/zenith
+          paths:
+            - "*"
+
+  coverage-report:
+    executor: zenith-build-executor
+    steps:
+      - attach_workspace:
+          at: /tmp/zenith
+      - checkout
+      - restore_cache:
+          name: Restore rust cache
+          keys:
+            # Require an exact match. While an out of date cache might speed up the build,
+            # there's no way to clean out old packages, so the cache grows every time something
+            # changes.
+            - v04-rust-cache-deps-debug-{{ checksum "Cargo.lock" }}
+      - run:
+          name: Install llvm-tools
+          command: |
+            # TODO: install a proper symbol demangler, e.g. rustfilt
+            # TODO: we should embed this into a docker image
+            rustup component add llvm-tools-preview
+      - run:
+          name: Build coverage report
+          command: |
+            COMMIT_URL=https://github.com/zenithdb/zenith/commit/$CIRCLE_SHA1
+
+            scripts/coverage \
+              --dir=/tmp/zenith/coverage report \
+              --input-objects=/tmp/zenith/etc/binaries.list \
+              --commit-url=$COMMIT_URL \
+              --format=github
+      - run:
+          name: Upload coverage report
+          command: |
+            LOCAL_REPO=$CIRCLE_PROJECT_USERNAME/$CIRCLE_PROJECT_REPONAME
+            REPORT_URL=https://zenithdb.github.io/zenith-coverage-data/$CIRCLE_SHA1
+            COMMIT_URL=https://github.com/zenithdb/zenith/commit/$CIRCLE_SHA1
+
+            scripts/git-upload \
+              --repo=https://$VIP_VAP_ACCESS_TOKEN@github.com/zenithdb/zenith-coverage-data.git \
+              --message="Add code coverage for $COMMIT_URL" \
+              copy /tmp/zenith/coverage/report $CIRCLE_SHA1 # COPY FROM TO_RELATIVE
+
+            # Add link to the coverage report to the commit
+            curl -f -X POST \
+            https://api.github.com/repos/$LOCAL_REPO/statuses/$CIRCLE_SHA1 \
+            -H "Accept: application/vnd.github.v3+json" \
+            --user "$CI_ACCESS_TOKEN" \
+            --data \
+              "{
+                \"state\": \"success\",
+                \"context\": \"zenith-coverage\",
+                \"description\": \"Coverage report is ready\",
+                \"target_url\": \"$REPORT_URL\"
+              }"
 
   # Build zenithdb/zenith:latest image and push it to Docker hub
   docker-image:
@@ -283,7 +431,7 @@ jobs:
           name: Build and push Docker image
           command: |
             echo $DOCKER_PWD | docker login -u $DOCKER_LOGIN --password-stdin
-            docker build -t zenithdb/zenith:latest . && docker push zenithdb/zenith:latest
+            docker build --build-arg GIT_VERSION=$CIRCLE_SHA1 -t zenithdb/zenith:latest . && docker push zenithdb/zenith:latest
 
   # Trigger a new remote CI job
   remote-ci-trigger:
@@ -325,15 +473,15 @@ jobs:
                 \"inputs\": {
                   \"ci_job_name\": \"zenith-remote-ci\",
                   \"commit_hash\": \"$CIRCLE_SHA1\",
-                  \"remote_repo\": \"$LOCAL_REPO\",
-                  \"zenith_image_branch\": \"$CIRCLE_BRANCH\"
+                  \"remote_repo\": \"$LOCAL_REPO\"
                 }
               }"
 
 workflows:
   build_and_test:
     jobs:
-      - check-codestyle
+      - check-codestyle-rust
+      - check-codestyle-python
       - build-postgres:
           name: build-postgres-<< matrix.build_type >>
           matrix:
@@ -368,8 +516,15 @@ workflows:
           build_type: release
           test_selection: performance
           run_in_parallel: false
+          save_perf_report: true
           requires:
             - build-zenith-release
+      - coverage-report:
+          # Context passes credentials for gh api
+          context: CI_ACCESS_TOKEN
+          requires:
+            # TODO: consider adding more
+            - other-tests-debug
       - docker-image:
           # Context gives an ability to login
           context: Docker Hub

.github/workflows/benchmarking.yml

@@ -0,0 +1,114 @@
+name: benchmarking
+
+on:
+  # uncomment to run on push for debugging your PR
+  # push:
+  #   branches: [ mybranch ]
+  schedule:
+    # * is a special character in YAML so you have to quote this string
+    #          ┌───────────── minute (0 - 59)
+    #          │ ┌───────────── hour (0 - 23)
+    #          │ │ ┌───────────── day of the month (1 - 31)
+    #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    - cron:  '36 7 * * *' # run once a day, timezone is utc
+
+  workflow_dispatch: # adds ability to run this manually
+
+env:
+  BASE_URL: "https://console.zenith.tech"
+
+jobs:
+  bench:
+    # this workflow runs on self hosteed runner
+    # it's environment is quite different from usual guthub runner
+    # probably the most important difference is that it doesnt start from clean workspace each time
+    # e g if you install system packages they are not cleaned up since you install them directly in host machine
+    # not a container or something
+    # See documentation for more info: https://docs.github.com/en/actions/hosting-your-own-runners/about-self-hosted-runners
+    runs-on: [self-hosted, zenith-benchmarker]
+
+    env:
+      PG_BIN: "/usr/pgsql-13/bin"
+
+    steps:
+    - name: Checkout zenith repo
+      uses: actions/checkout@v2
+
+    # actions/setup-python@v2 is not working correctly on self-hosted runners
+    # see https://github.com/actions/setup-python/issues/162
+    # and probably https://github.com/actions/setup-python/issues/162#issuecomment-865387976 in particular
+    # so the simplest solution to me is to use already installed system python and spin virtualenvs for job runs.
+    # there is Python 3.7.10 already installed on the machine so use it to install pipenv and then use pipenv's virtuealenvs
+    - name: Install pipenv & deps
+      run: |
+        python3 -m pip install --upgrade pipenv wheel
+        # since pip/pipenv caches are reused there shouldn't be any troubles with install every time
+        pipenv install
+
+    - name: Show versions
+      run: |
+        echo Python
+        python3 --version
+        pipenv run python3 --version
+        echo Pipenv
+        pipenv --version
+        echo Pgbench
+        $PG_BIN/pgbench --version
+
+    # FIXME cluster setup is skipped due to various changes in console API
+    # for now pre created cluster is used. When API gain some stability
+    # after massive changes dynamic cluster setup will be revived.
+    # So use pre created cluster. It needs to be started manually, but stop is automatic after 5 minutes of inactivity
+    - name: Setup cluster
+      env:
+        BENCHMARK_CONSOLE_USER_PASSWORD: "${{ secrets.BENCHMARK_CONSOLE_USER_PASSWORD }}"
+        BENCHMARK_CONSOLE_ACCESS_TOKEN: "${{ secrets.BENCHMARK_CONSOLE_ACCESS_TOKEN }}"
+        BENCHMARK_CLUSTER_ID: "${{ secrets.BENCHMARK_CLUSTER_ID }}"
+      shell: bash
+      run: |
+        set -e
+
+        echo "Starting cluster"
+        CLUSTER=$(curl -s --fail --show-error -X POST $BASE_URL/api/v1/clusters/$BENCHMARK_CLUSTER_ID/start \
+            -H "Authorization: Bearer $BENCHMARK_CONSOLE_ACCESS_TOKEN")
+        echo $CLUSTER | python -m json.tool
+
+        echo "Waiting for cluster to become ready"
+        sleep 10
+
+        echo "CLUSTER_ID=$BENCHMARK_CLUSTER_ID" >> $GITHUB_ENV
+        CLUSTER=$(curl -s --fail --show-error -X GET $BASE_URL/api/v1/clusters/$BENCHMARK_CLUSTER_ID.json \
+            -H "Authorization: Bearer $BENCHMARK_CONSOLE_ACCESS_TOKEN")
+        echo $CLUSTER | python -m json.tool
+
+    - name: Run benchmark
+      # pgbench is installed system wide from official repo
+      # https://download.postgresql.org/pub/repos/yum/13/redhat/rhel-7-x86_64/
+      # via
+      # sudo tee /etc/yum.repos.d/pgdg.repo<<EOF
+      # [pgdg13]
+      # name=PostgreSQL 13 for RHEL/CentOS 7 - x86_64
+      # baseurl=https://download.postgresql.org/pub/repos/yum/13/redhat/rhel-7-x86_64/
+      # enabled=1
+      # gpgcheck=0
+      # EOF
+      # sudo yum makecache
+      # sudo yum install postgresql13-contrib
+      # actual binaries are located in /usr/pgsql-13/bin/
+      env:
+        TEST_PG_BENCH_TRANSACTIONS_MATRIX: "5000,10000,20000"
+        TEST_PG_BENCH_SCALES_MATRIX: "10,15"
+        PLATFORM: "zenith-staging"
+        BENCHMARK_CONSOLE_ACCESS_TOKEN: "${{ secrets.BENCHMARK_CONSOLE_ACCESS_TOKEN }}"
+        BENCHMARK_CONNSTR: "${{ secrets.BENCHMARK_STAGING_CONNSTR }}"
+        REMOTE_ENV: "1" # indicate to test harness that we do not have zenith binaries locally
+      run: |
+        mkdir -p perf-report-staging
+        pipenv run pytest test_runner/performance/ -v -m "remote_cluster" --skip-interfering-proc-check --out-dir perf-report-staging
+
+    - name: Submit result
+      env:
+        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
+      run: |
+        REPORT_FROM=$(realpath perf-report-staging) REPORT_TO=staging scripts/generate_and_push_perf_report.sh

.gitignore

@@ -7,3 +7,7 @@ test_output/
 .vscode
 /.zenith
 /integration_tests/.zenith
+
+# Coverage
+*.profraw
+*.profdata

.yapfignore

@@ -0,0 +1,10 @@
+# This file is only read when `yapf` is run from this directory.
+# Hence we only top-level directories here to avoid confusion.
+# See source code for the exact file format: https://github.com/google/yapf/blob/c6077954245bc3add82dafd853a1c7305a6ebd20/yapf/yapflib/file_resources.py#L40-L43
+vendor/
+target/
+tmp_install/
+__pycache__/
+test_output/
+.zenith/
+.git/

Dockerfile

@@ -21,11 +21,15 @@ RUN rm -rf postgres_install/build
 # net time waste in a lot of cases. Copying Cargo.lock with empty lib.rs should do the work.
 #
 FROM zenithdb/build:buster AS build
+
+ARG GIT_VERSION
+RUN if [ -z "$GIT_VERSION" ]; then echo "GIT_VERSION is reqired, use build_arg to pass it"; exit 1; fi
+
 WORKDIR /zenith
 COPY --from=pg-build /zenith/tmp_install/include/postgresql/server tmp_install/include/postgresql/server
 
 COPY . .
-RUN cargo build --release
+RUN GIT_VERSION=$GIT_VERSION cargo build --release
 
 #
 # Copy binaries to resulting image.
@@ -37,7 +41,7 @@ RUN apt-get update && apt-get -yq install libreadline-dev libseccomp-dev openssl
     mkdir zenith_install
 
 COPY --from=build /zenith/target/release/pageserver /usr/local/bin
-COPY --from=build /zenith/target/release/wal_acceptor /usr/local/bin
+COPY --from=build /zenith/target/release/safekeeper /usr/local/bin
 COPY --from=build /zenith/target/release/proxy /usr/local/bin
 COPY --from=pg-build /zenith/tmp_install postgres_install
 COPY docker-entrypoint.sh /docker-entrypoint.sh

Dockerfile.alpine

@@ -81,7 +81,7 @@ FROM alpine:3.13
 RUN apk add --update openssl build-base libseccomp-dev
 RUN apk --no-cache --update --repository https://dl-cdn.alpinelinux.org/alpine/edge/testing add rocksdb
 COPY --from=build /zenith/target/release/pageserver /usr/local/bin
-COPY --from=build /zenith/target/release/wal_acceptor /usr/local/bin
+COPY --from=build /zenith/target/release/safekeeper /usr/local/bin
 COPY --from=build /zenith/target/release/proxy /usr/local/bin
 COPY --from=pg-build /zenith/tmp_install /usr/local
 COPY docker-entrypoint.sh /docker-entrypoint.sh

Makefile

@@ -10,32 +10,43 @@ endif
 # We differentiate between release / debug build types using the BUILD_TYPE
 # environment variable.
 #
+BUILD_TYPE ?= debug
 ifeq ($(BUILD_TYPE),release)
 	PG_CONFIGURE_OPTS = --enable-debug
-	PG_CFLAGS = -O2 -g3 ${CFLAGS}
-else
+	PG_CFLAGS = -O2 -g3 $(CFLAGS)
+	# Unfortunately, `--profile=...` is a nightly feature
+	CARGO_BUILD_FLAGS += --release
+else ifeq ($(BUILD_TYPE),debug)
 	PG_CONFIGURE_OPTS = --enable-debug --enable-cassert --enable-depend
-	PG_CFLAGS = -O0 -g3 ${CFLAGS}
+	PG_CFLAGS = -O0 -g3 $(CFLAGS)
+else
+$(error Bad build type `$(BUILD_TYPE)', see Makefile for options)
 endif
 
+# Choose whether we should be silent or verbose
+CARGO_BUILD_FLAGS += --$(if $(filter s,$(MAKEFLAGS)),quiet,verbose)
+# Fix for a corner case when make doesn't pass a jobserver
+CARGO_BUILD_FLAGS += $(filter -j1,$(MAKEFLAGS))
+
+# This option has a side effect of passing make jobserver to cargo.
+# However, we shouldn't do this if `make -n` (--dry-run) has been asked.
+CARGO_CMD_PREFIX += $(if $(filter n,$(MAKEFLAGS)),,+)
+# Force cargo not to print progress bar
+CARGO_CMD_PREFIX += CARGO_TERM_PROGRESS_WHEN=never CI=1
+
 #
 # Top level Makefile to build Zenith and PostgreSQL
 #
+.PHONY: all
 all: zenith postgres
 
-# We don't want to run 'cargo build' in parallel with the postgres build,
-# because interleaving cargo build output with postgres build output looks
-# confusing. Also, 'cargo build' is parallel on its own, so it would be too
-# much parallelism. (Recursive invocation of postgres target still gets any
-# '-j' flag from the command line, so 'make -j' is still useful.)
-.NOTPARALLEL:
-
 ### Zenith Rust bits
 #
 # The 'postgres_ffi' depends on the Postgres headers.
 .PHONY: zenith
 zenith: postgres-headers
-	cargo build
+	+@echo "Compiling Zenith"
+	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS)
 
 ### PostgreSQL parts
 tmp_install/build/config.status:
@@ -57,10 +68,10 @@ postgres-headers: postgres-configure
 	+@echo "Installing PostgreSQL headers"
 	$(MAKE) -C tmp_install/build/src/include MAKELEVEL=0 install
 
-
 # Compile and install PostgreSQL and contrib/zenith
 .PHONY: postgres
-postgres: postgres-configure
+postgres: postgres-configure \
+		  postgres-headers # to prevent `make install` conflicts with zenith's `postgres-headers`
 	+@echo "Compiling PostgreSQL"
 	$(MAKE) -C tmp_install/build MAKELEVEL=0 install
 	+@echo "Compiling contrib/zenith"
@@ -68,18 +79,21 @@ postgres: postgres-configure
 	+@echo "Compiling contrib/zenith_test_utils"
 	$(MAKE) -C tmp_install/build/contrib/zenith_test_utils install
 
+.PHONY: postgres-clean
 postgres-clean:
 	$(MAKE) -C tmp_install/build MAKELEVEL=0 clean
 
 # This doesn't remove the effects of 'configure'.
+.PHONY: clean
 clean:
-	cd tmp_install/build && ${MAKE} clean
-	cargo clean
+	cd tmp_install/build && $(MAKE) clean
+	$(CARGO_CMD_PREFIX) cargo clean
 
 # This removes everything
+.PHONY: distclean
 distclean:
 	rm -rf tmp_install
-	cargo clean
+	$(CARGO_CMD_PREFIX) cargo clean
 
 .PHONY: fmt
 fmt:

Pipfile

@@ -1 +0,0 @@
-./test_runner/Pipfile

Pipfile

@@ -0,0 +1,30 @@
+[[source]]
+url = "https://pypi.python.org/simple"
+verify_ssl = true
+name = "pypi"
+
+[packages]
+pytest = ">=6.0.0"
+typing-extensions = "*"
+pyjwt = {extras = ["crypto"], version = "*"}
+requests = "*"
+pytest-xdist = "*"
+asyncpg = "*"
+cached-property = "*"
+psycopg2-binary = "*"
+jinja2 = "*"
+
+[dev-packages]
+# Behavior may change slightly between versions. These are run continuously,
+# so we pin exact versions to avoid suprising breaks. Update if comfortable.
+yapf = "==0.31.0"
+mypy = "==0.910"
+# Non-pinned packages follow.
+pipenv = "*"
+flake8 = "*"
+types-requests = "*"
+types-psycopg2 = "*"
+
+[requires]
+# we need at least 3.7, but pipenv doesn't allow to say this directly
+python_version = "3"

Pipfile.lock

@@ -1 +0,0 @@
-./test_runner/Pipfile.lock

Pipfile.lock

@@ -0,0 +1,652 @@
+{
+    "_meta": {
+        "hash": {
+            "sha256": "c309cb963a7b07ae3d30e9cbf08b495f77bdecc0e5356fc89d133c4fbcb65b2b"
+        },
+        "pipfile-spec": 6,
+        "requires": {
+            "python_version": "3"
+        },
+        "sources": [
+            {
+                "name": "pypi",
+                "url": "https://pypi.python.org/simple",
+                "verify_ssl": true
+            }
+        ]
+    },
+    "default": {
+        "asyncpg": {
+            "hashes": [
+                "sha256:129d501f3d30616afd51eb8d3142ef51ba05374256bd5834cec3ef4956a9b317",
+                "sha256:29ef6ae0a617fc13cc2ac5dc8e9b367bb83cba220614b437af9b67766f4b6b20",
+                "sha256:41704c561d354bef01353835a7846e5606faabbeb846214dfcf666cf53319f18",
+                "sha256:556b0e92e2b75dc028b3c4bc9bd5162ddf0053b856437cf1f04c97f9c6837d03",
+                "sha256:8ff5073d4b654e34bd5eaadc01dc4d68b8a9609084d835acd364cd934190a08d",
+                "sha256:a458fc69051fbb67d995fdda46d75a012b5d6200f91e17d23d4751482640ed4c",
+                "sha256:a7095890c96ba36f9f668eb552bb020dddb44f8e73e932f8573efc613ee83843",
+                "sha256:a738f4807c853623d3f93f0fea11f61be6b0e5ca16ea8aeb42c2c7ee742aa853",
+                "sha256:c4fc0205fe4ddd5aeb3dfdc0f7bafd43411181e1f5650189608e5971cceacff1",
+                "sha256:dd2fa063c3344823487d9ddccb40802f02622ddf8bf8a6cc53885ee7a2c1c0c6",
+                "sha256:ddffcb85227bf39cd1bedd4603e0082b243cf3b14ced64dce506a15b05232b83",
+                "sha256:e36c6806883786b19551bb70a4882561f31135dc8105a59662e0376cf5b2cbc5",
+                "sha256:eed43abc6ccf1dc02e0d0efc06ce46a411362f3358847c6b0ec9a43426f91ece"
+            ],
+            "index": "pypi",
+            "version": "==0.24.0"
+        },
+        "attrs": {
+            "hashes": [
+                "sha256:149e90d6d8ac20db7a955ad60cf0e6881a3f20d37096140088356da6c716b0b1",
+                "sha256:ef6aaac3ca6cd92904cdd0d83f629a15f18053ec84e6432106f7a4d04ae4f5fb"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
+            "version": "==21.2.0"
+        },
+        "cached-property": {
+            "hashes": [
+                "sha256:9fa5755838eecbb2d234c3aa390bd80fbd3ac6b6869109bfc1b499f7bd89a130",
+                "sha256:df4f613cf7ad9a588cc381aaf4a512d26265ecebd5eb9e1ba12f1319eb85a6a0"
+            ],
+            "index": "pypi",
+            "version": "==1.5.2"
+        },
+        "certifi": {
+            "hashes": [
+                "sha256:78884e7c1d4b00ce3cea67b44566851c4343c120abd683433ce934a68ea58872",
+                "sha256:d62a0163eb4c2344ac042ab2bdf75399a71a2d8c7d47eac2e2ee91b9d6339569"
+            ],
+            "version": "==2021.10.8"
+        },
+        "cffi": {
+            "hashes": [
+                "sha256:00c878c90cb53ccfaae6b8bc18ad05d2036553e6d9d1d9dbcf323bbe83854ca3",
+                "sha256:0104fb5ae2391d46a4cb082abdd5c69ea4eab79d8d44eaaf79f1b1fd806ee4c2",
+                "sha256:06c48159c1abed75c2e721b1715c379fa3200c7784271b3c46df01383b593636",
+                "sha256:0808014eb713677ec1292301ea4c81ad277b6cdf2fdd90fd540af98c0b101d20",
+                "sha256:10dffb601ccfb65262a27233ac273d552ddc4d8ae1bf93b21c94b8511bffe728",
+                "sha256:14cd121ea63ecdae71efa69c15c5543a4b5fbcd0bbe2aad864baca0063cecf27",
+                "sha256:17771976e82e9f94976180f76468546834d22a7cc404b17c22df2a2c81db0c66",
+                "sha256:181dee03b1170ff1969489acf1c26533710231c58f95534e3edac87fff06c443",
+                "sha256:23cfe892bd5dd8941608f93348c0737e369e51c100d03718f108bf1add7bd6d0",
+                "sha256:263cc3d821c4ab2213cbe8cd8b355a7f72a8324577dc865ef98487c1aeee2bc7",
+                "sha256:2756c88cbb94231c7a147402476be2c4df2f6078099a6f4a480d239a8817ae39",
+                "sha256:27c219baf94952ae9d50ec19651a687b826792055353d07648a5695413e0c605",
+                "sha256:2a23af14f408d53d5e6cd4e3d9a24ff9e05906ad574822a10563efcef137979a",
+                "sha256:31fb708d9d7c3f49a60f04cf5b119aeefe5644daba1cd2a0fe389b674fd1de37",
+                "sha256:3415c89f9204ee60cd09b235810be700e993e343a408693e80ce7f6a40108029",
+                "sha256:3773c4d81e6e818df2efbc7dd77325ca0dcb688116050fb2b3011218eda36139",
+                "sha256:3b96a311ac60a3f6be21d2572e46ce67f09abcf4d09344c49274eb9e0bf345fc",
+                "sha256:3f7d084648d77af029acb79a0ff49a0ad7e9d09057a9bf46596dac9514dc07df",
+                "sha256:41d45de54cd277a7878919867c0f08b0cf817605e4eb94093e7516505d3c8d14",
+                "sha256:4238e6dab5d6a8ba812de994bbb0a79bddbdf80994e4ce802b6f6f3142fcc880",
+                "sha256:45db3a33139e9c8f7c09234b5784a5e33d31fd6907800b316decad50af323ff2",
+                "sha256:45e8636704eacc432a206ac7345a5d3d2c62d95a507ec70d62f23cd91770482a",
+                "sha256:4958391dbd6249d7ad855b9ca88fae690783a6be9e86df65865058ed81fc860e",
+                "sha256:4a306fa632e8f0928956a41fa8e1d6243c71e7eb59ffbd165fc0b41e316b2474",
+                "sha256:57e9ac9ccc3101fac9d6014fba037473e4358ef4e89f8e181f8951a2c0162024",
+                "sha256:59888172256cac5629e60e72e86598027aca6bf01fa2465bdb676d37636573e8",
+                "sha256:5e069f72d497312b24fcc02073d70cb989045d1c91cbd53979366077959933e0",
+                "sha256:64d4ec9f448dfe041705426000cc13e34e6e5bb13736e9fd62e34a0b0c41566e",
+                "sha256:6dc2737a3674b3e344847c8686cf29e500584ccad76204efea14f451d4cc669a",
+                "sha256:74fdfdbfdc48d3f47148976f49fab3251e550a8720bebc99bf1483f5bfb5db3e",
+                "sha256:75e4024375654472cc27e91cbe9eaa08567f7fbdf822638be2814ce059f58032",
+                "sha256:786902fb9ba7433aae840e0ed609f45c7bcd4e225ebb9c753aa39725bb3e6ad6",
+                "sha256:8b6c2ea03845c9f501ed1313e78de148cd3f6cad741a75d43a29b43da27f2e1e",
+                "sha256:91d77d2a782be4274da750752bb1650a97bfd8f291022b379bb8e01c66b4e96b",
+                "sha256:91ec59c33514b7c7559a6acda53bbfe1b283949c34fe7440bcf917f96ac0723e",
+                "sha256:920f0d66a896c2d99f0adbb391f990a84091179542c205fa53ce5787aff87954",
+                "sha256:a5263e363c27b653a90078143adb3d076c1a748ec9ecc78ea2fb916f9b861962",
+                "sha256:abb9a20a72ac4e0fdb50dae135ba5e77880518e742077ced47eb1499e29a443c",
+                "sha256:c2051981a968d7de9dd2d7b87bcb9c939c74a34626a6e2f8181455dd49ed69e4",
+                "sha256:c21c9e3896c23007803a875460fb786118f0cdd4434359577ea25eb556e34c55",
+                "sha256:c2502a1a03b6312837279c8c1bd3ebedf6c12c4228ddbad40912d671ccc8a962",
+                "sha256:d4d692a89c5cf08a8557fdeb329b82e7bf609aadfaed6c0d79f5a449a3c7c023",
+                "sha256:da5db4e883f1ce37f55c667e5c0de439df76ac4cb55964655906306918e7363c",
+                "sha256:e7022a66d9b55e93e1a845d8c9eba2a1bebd4966cd8bfc25d9cd07d515b33fa6",
+                "sha256:ef1f279350da2c586a69d32fc8733092fd32cc8ac95139a00377841f59a3f8d8",
+                "sha256:f54a64f8b0c8ff0b64d18aa76675262e1700f3995182267998c31ae974fbc382",
+                "sha256:f5c7150ad32ba43a07c4479f40241756145a1f03b43480e058cfd862bf5041c7",
+                "sha256:f6f824dc3bce0edab5f427efcfb1d63ee75b6fcb7282900ccaf925be84efb0fc",
+                "sha256:fd8a250edc26254fe5b33be00402e6d287f562b6a5b2152dec302fa15bb3e997",
+                "sha256:ffaa5c925128e29efbde7301d8ecaf35c8c60ffbcd6a1ffd3a552177c8e5e796"
+            ],
+            "version": "==1.15.0"
+        },
+        "charset-normalizer": {
+            "hashes": [
+                "sha256:e019de665e2bcf9c2b64e2e5aa025fa991da8720daa3c1138cadd2fd1856aed0",
+                "sha256:f7af805c321bfa1ce6714c51f254e0d5bb5e5834039bc17db7ebe3a4cec9492b"
+            ],
+            "markers": "python_version >= '3'",
+            "version": "==2.0.7"
+        },
+        "cryptography": {
+            "hashes": [
+                "sha256:07bb7fbfb5de0980590ddfc7f13081520def06dc9ed214000ad4372fb4e3c7f6",
+                "sha256:18d90f4711bf63e2fb21e8c8e51ed8189438e6b35a6d996201ebd98a26abbbe6",
+                "sha256:1ed82abf16df40a60942a8c211251ae72858b25b7421ce2497c2eb7a1cee817c",
+                "sha256:22a38e96118a4ce3b97509443feace1d1011d0571fae81fc3ad35f25ba3ea999",
+                "sha256:2d69645f535f4b2c722cfb07a8eab916265545b3475fdb34e0be2f4ee8b0b15e",
+                "sha256:4a2d0e0acc20ede0f06ef7aa58546eee96d2592c00f450c9acb89c5879b61992",
+                "sha256:54b2605e5475944e2213258e0ab8696f4f357a31371e538ef21e8d61c843c28d",
+                "sha256:7075b304cd567694dc692ffc9747f3e9cb393cc4aa4fb7b9f3abd6f5c4e43588",
+                "sha256:7b7ceeff114c31f285528ba8b390d3e9cfa2da17b56f11d366769a807f17cbaa",
+                "sha256:7eba2cebca600a7806b893cb1d541a6e910afa87e97acf2021a22b32da1df52d",
+                "sha256:928185a6d1ccdb816e883f56ebe92e975a262d31cc536429041921f8cb5a62fd",
+                "sha256:9933f28f70d0517686bd7de36166dda42094eac49415459d9bdf5e7df3e0086d",
+                "sha256:a688ebcd08250eab5bb5bca318cc05a8c66de5e4171a65ca51db6bd753ff8953",
+                "sha256:abb5a361d2585bb95012a19ed9b2c8f412c5d723a9836418fab7aaa0243e67d2",
+                "sha256:c10c797ac89c746e488d2ee92bd4abd593615694ee17b2500578b63cad6b93a8",
+                "sha256:ced40344e811d6abba00295ced98c01aecf0c2de39481792d87af4fa58b7b4d6",
+                "sha256:d57e0cdc1b44b6cdf8af1d01807db06886f10177469312fbde8f44ccbb284bc9",
+                "sha256:d99915d6ab265c22873f1b4d6ea5ef462ef797b4140be4c9d8b179915e0985c6",
+                "sha256:eb80e8a1f91e4b7ef8b33041591e6d89b2b8e122d787e87eeb2b08da71bb16ad",
+                "sha256:ebeddd119f526bcf323a89f853afb12e225902a24d29b55fe18dd6fcb2838a76"
+            ],
+            "version": "==35.0.0"
+        },
+        "execnet": {
+            "hashes": [
+                "sha256:8f694f3ba9cc92cab508b152dcfe322153975c29bda272e2fd7f3f00f36e47c5",
+                "sha256:a295f7cc774947aac58dde7fdc85f4aa00c42adf5d8f5468fc630c1acf30a142"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
+            "version": "==1.9.0"
+        },
+        "idna": {
+            "hashes": [
+                "sha256:84d9dd047ffa80596e0f246e2eab0b391788b0503584e8945f2368256d2735ff",
+                "sha256:9d643ff0a55b762d5cdb124b8eaa99c66322e2157b69160bc32796e824360e6d"
+            ],
+            "markers": "python_version >= '3'",
+            "version": "==3.3"
+        },
+        "importlib-metadata": {
+            "hashes": [
+                "sha256:b618b6d2d5ffa2f16add5697cf57a46c76a56229b0ed1c438322e4e95645bd15",
+                "sha256:f284b3e11256ad1e5d03ab86bb2ccd6f5339688ff17a4d797a0fe7df326f23b1"
+            ],
+            "markers": "python_version < '3.8'",
+            "version": "==4.8.1"
+        },
+        "iniconfig": {
+            "hashes": [
+                "sha256:011e24c64b7f47f6ebd835bb12a743f2fbe9a26d4cecaa7f53bc4f35ee9da8b3",
+                "sha256:bc3af051d7d14b2ee5ef9969666def0cd1a000e121eaea580d4a313df4b37f32"
+            ],
+            "version": "==1.1.1"
+        },
+        "jinja2": {
+            "hashes": [
+                "sha256:827a0e32839ab1600d4eb1c4c33ec5a8edfbc5cb42dafa13b81f182f97784b45",
+                "sha256:8569982d3f0889eed11dd620c706d39b60c36d6d25843961f33f77fb6bc6b20c"
+            ],
+            "index": "pypi",
+            "version": "==3.0.2"
+        },
+        "markupsafe": {
+            "hashes": [
+                "sha256:01a9b8ea66f1658938f65b93a85ebe8bc016e6769611be228d797c9d998dd298",
+                "sha256:023cb26ec21ece8dc3907c0e8320058b2e0cb3c55cf9564da612bc325bed5e64",
+                "sha256:0446679737af14f45767963a1a9ef7620189912317d095f2d9ffa183a4d25d2b",
+                "sha256:04635854b943835a6ea959e948d19dcd311762c5c0c6e1f0e16ee57022669194",
+                "sha256:0717a7390a68be14b8c793ba258e075c6f4ca819f15edfc2a3a027c823718567",
+                "sha256:0955295dd5eec6cb6cc2fe1698f4c6d84af2e92de33fbcac4111913cd100a6ff",
+                "sha256:0d4b31cc67ab36e3392bbf3862cfbadac3db12bdd8b02a2731f509ed5b829724",
+                "sha256:10f82115e21dc0dfec9ab5c0223652f7197feb168c940f3ef61563fc2d6beb74",
+                "sha256:168cd0a3642de83558a5153c8bd34f175a9a6e7f6dc6384b9655d2697312a646",
+                "sha256:1d609f577dc6e1aa17d746f8bd3c31aa4d258f4070d61b2aa5c4166c1539de35",
+                "sha256:1f2ade76b9903f39aa442b4aadd2177decb66525062db244b35d71d0ee8599b6",
+                "sha256:20dca64a3ef2d6e4d5d615a3fd418ad3bde77a47ec8a23d984a12b5b4c74491a",
+                "sha256:2a7d351cbd8cfeb19ca00de495e224dea7e7d919659c2841bbb7f420ad03e2d6",
+                "sha256:2d7d807855b419fc2ed3e631034685db6079889a1f01d5d9dac950f764da3dad",
+                "sha256:2ef54abee730b502252bcdf31b10dacb0a416229b72c18b19e24a4509f273d26",
+                "sha256:36bc903cbb393720fad60fc28c10de6acf10dc6cc883f3e24ee4012371399a38",
+                "sha256:37205cac2a79194e3750b0af2a5720d95f786a55ce7df90c3af697bfa100eaac",
+                "sha256:3c112550557578c26af18a1ccc9e090bfe03832ae994343cfdacd287db6a6ae7",
+                "sha256:3dd007d54ee88b46be476e293f48c85048603f5f516008bee124ddd891398ed6",
+                "sha256:4296f2b1ce8c86a6aea78613c34bb1a672ea0e3de9c6ba08a960efe0b0a09047",
+                "sha256:47ab1e7b91c098ab893b828deafa1203de86d0bc6ab587b160f78fe6c4011f75",
+                "sha256:49e3ceeabbfb9d66c3aef5af3a60cc43b85c33df25ce03d0031a608b0a8b2e3f",
+                "sha256:4dc8f9fb58f7364b63fd9f85013b780ef83c11857ae79f2feda41e270468dd9b",
+                "sha256:4efca8f86c54b22348a5467704e3fec767b2db12fc39c6d963168ab1d3fc9135",
+                "sha256:53edb4da6925ad13c07b6d26c2a852bd81e364f95301c66e930ab2aef5b5ddd8",
+                "sha256:5855f8438a7d1d458206a2466bf82b0f104a3724bf96a1c781ab731e4201731a",
+                "sha256:594c67807fb16238b30c44bdf74f36c02cdf22d1c8cda91ef8a0ed8dabf5620a",
+                "sha256:5b6d930f030f8ed98e3e6c98ffa0652bdb82601e7a016ec2ab5d7ff23baa78d1",
+                "sha256:5bb28c636d87e840583ee3adeb78172efc47c8b26127267f54a9c0ec251d41a9",
+                "sha256:60bf42e36abfaf9aff1f50f52644b336d4f0a3fd6d8a60ca0d054ac9f713a864",
+                "sha256:611d1ad9a4288cf3e3c16014564df047fe08410e628f89805e475368bd304914",
+                "sha256:6300b8454aa6930a24b9618fbb54b5a68135092bc666f7b06901f897fa5c2fee",
+                "sha256:63f3268ba69ace99cab4e3e3b5840b03340efed0948ab8f78d2fd87ee5442a4f",
+                "sha256:6557b31b5e2c9ddf0de32a691f2312a32f77cd7681d8af66c2692efdbef84c18",
+                "sha256:693ce3f9e70a6cf7d2fb9e6c9d8b204b6b39897a2c4a1aa65728d5ac97dcc1d8",
+                "sha256:6a7fae0dd14cf60ad5ff42baa2e95727c3d81ded453457771d02b7d2b3f9c0c2",
+                "sha256:6c4ca60fa24e85fe25b912b01e62cb969d69a23a5d5867682dd3e80b5b02581d",
+                "sha256:6fcf051089389abe060c9cd7caa212c707e58153afa2c649f00346ce6d260f1b",
+                "sha256:7d91275b0245b1da4d4cfa07e0faedd5b0812efc15b702576d103293e252af1b",
+                "sha256:89c687013cb1cd489a0f0ac24febe8c7a666e6e221b783e53ac50ebf68e45d86",
+                "sha256:8d206346619592c6200148b01a2142798c989edcb9c896f9ac9722a99d4e77e6",
+                "sha256:905fec760bd2fa1388bb5b489ee8ee5f7291d692638ea5f67982d968366bef9f",
+                "sha256:97383d78eb34da7e1fa37dd273c20ad4320929af65d156e35a5e2d89566d9dfb",
+                "sha256:984d76483eb32f1bcb536dc27e4ad56bba4baa70be32fa87152832cdd9db0833",
+                "sha256:99df47edb6bda1249d3e80fdabb1dab8c08ef3975f69aed437cb69d0a5de1e28",
+                "sha256:9f02365d4e99430a12647f09b6cc8bab61a6564363f313126f775eb4f6ef798e",
+                "sha256:a30e67a65b53ea0a5e62fe23682cfe22712e01f453b95233b25502f7c61cb415",
+                "sha256:ab3ef638ace319fa26553db0624c4699e31a28bb2a835c5faca8f8acf6a5a902",
+                "sha256:aca6377c0cb8a8253e493c6b451565ac77e98c2951c45f913e0b52facdcff83f",
+                "sha256:add36cb2dbb8b736611303cd3bfcee00afd96471b09cda130da3581cbdc56a6d",
+                "sha256:b2f4bf27480f5e5e8ce285a8c8fd176c0b03e93dcc6646477d4630e83440c6a9",
+                "sha256:b7f2d075102dc8c794cbde1947378051c4e5180d52d276987b8d28a3bd58c17d",
+                "sha256:baa1a4e8f868845af802979fcdbf0bb11f94f1cb7ced4c4b8a351bb60d108145",
+                "sha256:be98f628055368795d818ebf93da628541e10b75b41c559fdf36d104c5787066",
+                "sha256:bf5d821ffabf0ef3533c39c518f3357b171a1651c1ff6827325e4489b0e46c3c",
+                "sha256:c47adbc92fc1bb2b3274c4b3a43ae0e4573d9fbff4f54cd484555edbf030baf1",
+                "sha256:cdfba22ea2f0029c9261a4bd07e830a8da012291fbe44dc794e488b6c9bb353a",
+                "sha256:d6c7ebd4e944c85e2c3421e612a7057a2f48d478d79e61800d81468a8d842207",
+                "sha256:d7f9850398e85aba693bb640262d3611788b1f29a79f0c93c565694658f4071f",
+                "sha256:d8446c54dc28c01e5a2dbac5a25f071f6653e6e40f3a8818e8b45d790fe6ef53",
+                "sha256:deb993cacb280823246a026e3b2d81c493c53de6acfd5e6bfe31ab3402bb37dd",
+                "sha256:e0f138900af21926a02425cf736db95be9f4af72ba1bb21453432a07f6082134",
+                "sha256:e9936f0b261d4df76ad22f8fee3ae83b60d7c3e871292cd42f40b81b70afae85",
+                "sha256:f0567c4dc99f264f49fe27da5f735f414c4e7e7dd850cfd8e69f0862d7c74ea9",
+                "sha256:f5653a225f31e113b152e56f154ccbe59eeb1c7487b39b9d9f9cdb58e6c79dc5",
+                "sha256:f826e31d18b516f653fe296d967d700fddad5901ae07c622bb3705955e1faa94",
+                "sha256:f8ba0e8349a38d3001fae7eadded3f6606f0da5d748ee53cc1dab1d6527b9509",
+                "sha256:f9081981fe268bd86831e5c75f7de206ef275defcb82bc70740ae6dc507aee51",
+                "sha256:fa130dd50c57d53368c9d59395cb5526eda596d3ffe36666cd81a44d56e48872"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==2.0.1"
+        },
+        "packaging": {
+            "hashes": [
+                "sha256:096d689d78ca690e4cd8a89568ba06d07ca097e3306a4381635073ca91479966",
+                "sha256:14317396d1e8cdb122989b916fa2c7e9ca8e2be9e8060a6eff75b6b7b4d8a7e0"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==21.2"
+        },
+        "pluggy": {
+            "hashes": [
+                "sha256:4224373bacce55f955a878bf9cfa763c1e360858e330072059e10bad68531159",
+                "sha256:74134bbf457f031a36d68416e1509f34bd5ccc019f0bcc952c7b909d06b37bd3"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==1.0.0"
+        },
+        "psycopg2-binary": {
+            "hashes": [
+                "sha256:0b7dae87f0b729922e06f85f667de7bf16455d411971b2043bbd9577af9d1975",
+                "sha256:0f2e04bd2a2ab54fa44ee67fe2d002bb90cee1c0f1cc0ebc3148af7b02034cbd",
+                "sha256:123c3fb684e9abfc47218d3784c7b4c47c8587951ea4dd5bc38b6636ac57f616",
+                "sha256:1473c0215b0613dd938db54a653f68251a45a78b05f6fc21af4326f40e8360a2",
+                "sha256:14db1752acdd2187d99cb2ca0a1a6dfe57fc65c3281e0f20e597aac8d2a5bd90",
+                "sha256:1e3a362790edc0a365385b1ac4cc0acc429a0c0d662d829a50b6ce743ae61b5a",
+                "sha256:1e85b74cbbb3056e3656f1cc4781294df03383127a8114cbc6531e8b8367bf1e",
+                "sha256:20f1ab44d8c352074e2d7ca67dc00843067788791be373e67a0911998787ce7d",
+                "sha256:24b0b6688b9f31a911f2361fe818492650795c9e5d3a1bc647acbd7440142a4f",
+                "sha256:2f62c207d1740b0bde5c4e949f857b044818f734a3d57f1d0d0edc65050532ed",
+                "sha256:3242b9619de955ab44581a03a64bdd7d5e470cc4183e8fcadd85ab9d3756ce7a",
+                "sha256:35c4310f8febe41f442d3c65066ca93cccefd75013df3d8c736c5b93ec288140",
+                "sha256:4235f9d5ddcab0b8dbd723dca56ea2922b485ea00e1dafacf33b0c7e840b3d32",
+                "sha256:542875f62bc56e91c6eac05a0deadeae20e1730be4c6334d8f04c944fcd99759",
+                "sha256:5ced67f1e34e1a450cdb48eb53ca73b60aa0af21c46b9b35ac3e581cf9f00e31",
+                "sha256:661509f51531ec125e52357a489ea3806640d0ca37d9dada461ffc69ee1e7b6e",
+                "sha256:7360647ea04db2e7dff1648d1da825c8cf68dc5fbd80b8fb5b3ee9f068dcd21a",
+                "sha256:736b8797b58febabb85494142c627bd182b50d2a7ec65322983e71065ad3034c",
+                "sha256:8c13d72ed6af7fd2c8acbd95661cf9477f94e381fce0792c04981a8283b52917",
+                "sha256:988b47ac70d204aed01589ed342303da7c4d84b56c2f4c4b8b00deda123372bf",
+                "sha256:995fc41ebda5a7a663a254a1dcac52638c3e847f48307b5416ee373da15075d7",
+                "sha256:a36c7eb6152ba5467fb264d73844877be8b0847874d4822b7cf2d3c0cb8cdcb0",
+                "sha256:aed4a9a7e3221b3e252c39d0bf794c438dc5453bc2963e8befe9d4cd324dff72",
+                "sha256:aef9aee84ec78af51107181d02fe8773b100b01c5dfde351184ad9223eab3698",
+                "sha256:b0221ca5a9837e040ebf61f48899926b5783668b7807419e4adae8175a31f773",
+                "sha256:b4d7679a08fea64573c969f6994a2631908bb2c0e69a7235648642f3d2e39a68",
+                "sha256:c250a7ec489b652c892e4f0a5d122cc14c3780f9f643e1a326754aedf82d9a76",
+                "sha256:ca86db5b561b894f9e5f115d6a159fff2a2570a652e07889d8a383b5fae66eb4",
+                "sha256:cfc523edecddaef56f6740d7de1ce24a2fdf94fd5e704091856a201872e37f9f",
+                "sha256:d92272c7c16e105788efe2cfa5d680f07e34e0c29b03c1908f8636f55d5f915a",
+                "sha256:da113b70f6ec40e7d81b43d1b139b9db6a05727ab8be1ee559f3a69854a69d34",
+                "sha256:f6fac64a38f6768e7bc7b035b9e10d8a538a9fadce06b983fb3e6fa55ac5f5ce",
+                "sha256:f8559617b1fcf59a9aedba2c9838b5b6aa211ffedecabca412b92a1ff75aac1a",
+                "sha256:fbb42a541b1093385a2d8c7eec94d26d30437d0e77c1d25dae1dcc46741a385e"
+            ],
+            "index": "pypi",
+            "version": "==2.9.1"
+        },
+        "py": {
+            "hashes": [
+                "sha256:21b81bda15b66ef5e1a777a21c4dcd9c20ad3efd0b3f817e7a809035269e1bd3",
+                "sha256:3b80836aa6d1feeaa108e046da6423ab8f6ceda6468545ae8d02d9d58d18818a"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==1.10.0"
+        },
+        "pycparser": {
+            "hashes": [
+                "sha256:2d475327684562c3a96cc71adf7dc8c4f0565175cf86b6d7a404ff4c771f15f0",
+                "sha256:7582ad22678f0fcd81102833f60ef8d0e57288b6b5fb00323d101be910e35705"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==2.20"
+        },
+        "pyjwt": {
+            "extras": [
+                "crypto"
+            ],
+            "hashes": [
+                "sha256:b888b4d56f06f6dcd777210c334e69c737be74755d3e5e9ee3fe67dc18a0ee41",
+                "sha256:e0c4bb8d9f0af0c7f5b1ec4c5036309617d03d56932877f2f7a0beeb5318322f"
+            ],
+            "index": "pypi",
+            "version": "==2.3.0"
+        },
+        "pyparsing": {
+            "hashes": [
+                "sha256:c203ec8783bf771a155b207279b9bccb8dea02d8f0c9e5f8ead507bc3246ecc1",
+                "sha256:ef9d7589ef3c200abe66653d3f1ab1033c3c419ae9b9bdb1240a85b024efc88b"
+            ],
+            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==2.4.7"
+        },
+        "pytest": {
+            "hashes": [
+                "sha256:131b36680866a76e6781d13f101efb86cf674ebb9762eb70d3082b6f29889e89",
+                "sha256:7310f8d27bc79ced999e760ca304d69f6ba6c6649c0b60fb0e04a4a77cacc134"
+            ],
+            "index": "pypi",
+            "version": "==6.2.5"
+        },
+        "pytest-forked": {
+            "hashes": [
+                "sha256:6aa9ac7e00ad1a539c41bec6d21011332de671e938c7637378ec9710204e37ca",
+                "sha256:dc4147784048e70ef5d437951728825a131b81714b398d5d52f17c7c144d8815"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
+            "version": "==1.3.0"
+        },
+        "pytest-xdist": {
+            "hashes": [
+                "sha256:7b61ebb46997a0820a263553179d6d1e25a8c50d8a8620cd1aa1e20e3be99168",
+                "sha256:89b330316f7fc475f999c81b577c2b926c9569f3d397ae432c0c2e2496d61ff9"
+            ],
+            "index": "pypi",
+            "version": "==2.4.0"
+        },
+        "requests": {
+            "hashes": [
+                "sha256:6c1246513ecd5ecd4528a0906f910e8f0f9c6b8ec72030dc9fd154dc1a6efd24",
+                "sha256:b8aa58f8cf793ffd8782d3d8cb19e66ef36f7aba4353eec859e74678b01b07a7"
+            ],
+            "index": "pypi",
+            "version": "==2.26.0"
+        },
+        "toml": {
+            "hashes": [
+                "sha256:806143ae5bfb6a3c6e736a764057db0e6a0e05e338b5630894a5f779cabb4f9b",
+                "sha256:b3bda1d108d5dd99f4a20d24d9c348e91c4db7ab1b749200bded2f839ccbe68f"
+            ],
+            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==0.10.2"
+        },
+        "typing-extensions": {
+            "hashes": [
+                "sha256:49f75d16ff11f1cd258e1b988ccff82a3ca5570217d7ad8c5f48205dd99a677e",
+                "sha256:d8226d10bc02a29bcc81df19a26e56a9647f8b0a6d4a83924139f4a8b01f17b7",
+                "sha256:f1d25edafde516b146ecd0613dabcc61409817af4766fbbcfb8d1ad4ec441a34"
+            ],
+            "index": "pypi",
+            "version": "==3.10.0.2"
+        },
+        "urllib3": {
+            "hashes": [
+                "sha256:4987c65554f7a2dbf30c18fd48778ef124af6fab771a377103da0585e2336ece",
+                "sha256:c4fdf4019605b6e5423637e01bc9fe4daef873709a7973e195ceba0a62bbc844"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4' and python_version < '4'",
+            "version": "==1.26.7"
+        },
+        "zipp": {
+            "hashes": [
+                "sha256:71c644c5369f4a6e07636f0aa966270449561fcea2e3d6747b8d23efaa9d7832",
+                "sha256:9fe5ea21568a0a70e50f273397638d39b03353731e6cbbb3fd8502a33fec40bc"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==3.6.0"
+        }
+    },
+    "develop": {
+        "backports.entry-points-selectable": {
+            "hashes": [
+                "sha256:988468260ec1c196dab6ae1149260e2f5472c9110334e5d51adcb77867361f6a",
+                "sha256:a6d9a871cde5e15b4c4a53e3d43ba890cc6861ec1332c9c2428c92f977192acc"
+            ],
+            "markers": "python_version >= '2.7'",
+            "version": "==1.1.0"
+        },
+        "certifi": {
+            "hashes": [
+                "sha256:78884e7c1d4b00ce3cea67b44566851c4343c120abd683433ce934a68ea58872",
+                "sha256:d62a0163eb4c2344ac042ab2bdf75399a71a2d8c7d47eac2e2ee91b9d6339569"
+            ],
+            "version": "==2021.10.8"
+        },
+        "distlib": {
+            "hashes": [
+                "sha256:c8b54e8454e5bf6237cc84c20e8264c3e991e824ef27e8f1e81049867d861e31",
+                "sha256:d982d0751ff6eaaab5e2ec8e691d949ee80eddf01a62eaa96ddb11531fe16b05"
+            ],
+            "version": "==0.3.3"
+        },
+        "filelock": {
+            "hashes": [
+                "sha256:7afc856f74fa7006a289fd10fa840e1eebd8bbff6bffb69c26c54a0512ea8cf8",
+                "sha256:bb2a1c717df74c48a2d00ed625e5a66f8572a3a30baacb7657add1d7bac4097b"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==3.3.2"
+        },
+        "flake8": {
+            "hashes": [
+                "sha256:479b1304f72536a55948cb40a32dce8bb0ffe3501e26eaf292c7e60eb5e0428d",
+                "sha256:806e034dda44114815e23c16ef92f95c91e4c71100ff52813adf7132a6ad870d"
+            ],
+            "index": "pypi",
+            "version": "==4.0.1"
+        },
+        "importlib-metadata": {
+            "hashes": [
+                "sha256:b618b6d2d5ffa2f16add5697cf57a46c76a56229b0ed1c438322e4e95645bd15",
+                "sha256:f284b3e11256ad1e5d03ab86bb2ccd6f5339688ff17a4d797a0fe7df326f23b1"
+            ],
+            "markers": "python_version < '3.8'",
+            "version": "==4.8.1"
+        },
+        "mccabe": {
+            "hashes": [
+                "sha256:ab8a6258860da4b6677da4bd2fe5dc2c659cff31b3ee4f7f5d64e79735b80d42",
+                "sha256:dd8d182285a0fe56bace7f45b5e7d1a6ebcbf524e8f3bd87eb0f125271b8831f"
+            ],
+            "version": "==0.6.1"
+        },
+        "mypy": {
+            "hashes": [
+                "sha256:088cd9c7904b4ad80bec811053272986611b84221835e079be5bcad029e79dd9",
+                "sha256:0aadfb2d3935988ec3815952e44058a3100499f5be5b28c34ac9d79f002a4a9a",
+                "sha256:119bed3832d961f3a880787bf621634ba042cb8dc850a7429f643508eeac97b9",
+                "sha256:1a85e280d4d217150ce8cb1a6dddffd14e753a4e0c3cf90baabb32cefa41b59e",
+                "sha256:3c4b8ca36877fc75339253721f69603a9c7fdb5d4d5a95a1a1b899d8b86a4de2",
+                "sha256:3e382b29f8e0ccf19a2df2b29a167591245df90c0b5a2542249873b5c1d78212",
+                "sha256:42c266ced41b65ed40a282c575705325fa7991af370036d3f134518336636f5b",
+                "sha256:53fd2eb27a8ee2892614370896956af2ff61254c275aaee4c230ae771cadd885",
+                "sha256:704098302473cb31a218f1775a873b376b30b4c18229421e9e9dc8916fd16150",
+                "sha256:7df1ead20c81371ccd6091fa3e2878559b5c4d4caadaf1a484cf88d93ca06703",
+                "sha256:866c41f28cee548475f146aa4d39a51cf3b6a84246969f3759cb3e9c742fc072",
+                "sha256:a155d80ea6cee511a3694b108c4494a39f42de11ee4e61e72bc424c490e46457",
+                "sha256:adaeee09bfde366d2c13fe6093a7df5df83c9a2ba98638c7d76b010694db760e",
+                "sha256:b6fb13123aeef4a3abbcfd7e71773ff3ff1526a7d3dc538f3929a49b42be03f0",
+                "sha256:b94e4b785e304a04ea0828759172a15add27088520dc7e49ceade7834275bedb",
+                "sha256:c0df2d30ed496a08de5daed2a9ea807d07c21ae0ab23acf541ab88c24b26ab97",
+                "sha256:c6c2602dffb74867498f86e6129fd52a2770c48b7cd3ece77ada4fa38f94eba8",
+                "sha256:ceb6e0a6e27fb364fb3853389607cf7eb3a126ad335790fa1e14ed02fba50811",
+                "sha256:d9dd839eb0dc1bbe866a288ba3c1afc33a202015d2ad83b31e875b5905a079b6",
+                "sha256:e4dab234478e3bd3ce83bac4193b2ecd9cf94e720ddd95ce69840273bf44f6de",
+                "sha256:ec4e0cd079db280b6bdabdc807047ff3e199f334050db5cbb91ba3e959a67504",
+                "sha256:ecd2c3fe726758037234c93df7e98deb257fd15c24c9180dacf1ef829da5f921",
+                "sha256:ef565033fa5a958e62796867b1df10c40263ea9ded87164d67572834e57a174d"
+            ],
+            "index": "pypi",
+            "version": "==0.910"
+        },
+        "mypy-extensions": {
+            "hashes": [
+                "sha256:090fedd75945a69ae91ce1303b5824f428daf5a028d2f6ab8a299250a846f15d",
+                "sha256:2d82818f5bb3e369420cb3c4060a7970edba416647068eb4c5343488a6c604a8"
+            ],
+            "version": "==0.4.3"
+        },
+        "pipenv": {
+            "hashes": [
+                "sha256:05958fadcd70b2de6a27542fcd2bd72dd5c59c6d35307fdac3e06361fb06e30e",
+                "sha256:d180f5be4775c552fd5e69ae18a9d6099d9dafb462efe54f11c72cb5f4d5e977"
+            ],
+            "index": "pypi",
+            "version": "==2021.5.29"
+        },
+        "platformdirs": {
+            "hashes": [
+                "sha256:367a5e80b3d04d2428ffa76d33f124cf11e8fff2acdaa9b43d545f5c7d661ef2",
+                "sha256:8868bbe3c3c80d42f20156f22e7131d2fb321f5bc86a2a345375c6481a67021d"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==2.4.0"
+        },
+        "pycodestyle": {
+            "hashes": [
+                "sha256:720f8b39dde8b293825e7ff02c475f3077124006db4f440dcbc9a20b76548a20",
+                "sha256:eddd5847ef438ea1c7870ca7eb78a9d47ce0cdb4851a5523949f2601d0cbbe7f"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
+            "version": "==2.8.0"
+        },
+        "pyflakes": {
+            "hashes": [
+                "sha256:05a85c2872edf37a4ed30b0cce2f6093e1d0581f8c19d7393122da7e25b2b24c",
+                "sha256:3bb3a3f256f4b7968c9c788781e4ff07dce46bdf12339dcda61053375426ee2e"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==2.4.0"
+        },
+        "six": {
+            "hashes": [
+                "sha256:1e61c37477a1626458e36f7b1d82aa5c9b094fa4802892072e49de9c60c4c926",
+                "sha256:8abb2f1d86890a2dfb989f9a77cfcfd3e47c2a354b01111771326f8aa26e0254"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==1.16.0"
+        },
+        "toml": {
+            "hashes": [
+                "sha256:806143ae5bfb6a3c6e736a764057db0e6a0e05e338b5630894a5f779cabb4f9b",
+                "sha256:b3bda1d108d5dd99f4a20d24d9c348e91c4db7ab1b749200bded2f839ccbe68f"
+            ],
+            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==0.10.2"
+        },
+        "typed-ast": {
+            "hashes": [
+                "sha256:01ae5f73431d21eead5015997ab41afa53aa1fbe252f9da060be5dad2c730ace",
+                "sha256:067a74454df670dcaa4e59349a2e5c81e567d8d65458d480a5b3dfecec08c5ff",
+                "sha256:0fb71b8c643187d7492c1f8352f2c15b4c4af3f6338f21681d3681b3dc31a266",
+                "sha256:1b3ead4a96c9101bef08f9f7d1217c096f31667617b58de957f690c92378b528",
+                "sha256:2068531575a125b87a41802130fa7e29f26c09a2833fea68d9a40cf33902eba6",
+                "sha256:209596a4ec71d990d71d5e0d312ac935d86930e6eecff6ccc7007fe54d703808",
+                "sha256:2c726c276d09fc5c414693a2de063f521052d9ea7c240ce553316f70656c84d4",
+                "sha256:398e44cd480f4d2b7ee8d98385ca104e35c81525dd98c519acff1b79bdaac363",
+                "sha256:52b1eb8c83f178ab787f3a4283f68258525f8d70f778a2f6dd54d3b5e5fb4341",
+                "sha256:5feca99c17af94057417d744607b82dd0a664fd5e4ca98061480fd8b14b18d04",
+                "sha256:7538e495704e2ccda9b234b82423a4038f324f3a10c43bc088a1636180f11a41",
+                "sha256:760ad187b1041a154f0e4d0f6aae3e40fdb51d6de16e5c99aedadd9246450e9e",
+                "sha256:777a26c84bea6cd934422ac2e3b78863a37017618b6e5c08f92ef69853e765d3",
+                "sha256:95431a26309a21874005845c21118c83991c63ea800dd44843e42a916aec5899",
+                "sha256:9ad2c92ec681e02baf81fdfa056fe0d818645efa9af1f1cd5fd6f1bd2bdfd805",
+                "sha256:9c6d1a54552b5330bc657b7ef0eae25d00ba7ffe85d9ea8ae6540d2197a3788c",
+                "sha256:aee0c1256be6c07bd3e1263ff920c325b59849dc95392a05f258bb9b259cf39c",
+                "sha256:af3d4a73793725138d6b334d9d247ce7e5f084d96284ed23f22ee626a7b88e39",
+                "sha256:b36b4f3920103a25e1d5d024d155c504080959582b928e91cb608a65c3a49e1a",
+                "sha256:b9574c6f03f685070d859e75c7f9eeca02d6933273b5e69572e5ff9d5e3931c3",
+                "sha256:bff6ad71c81b3bba8fa35f0f1921fb24ff4476235a6e94a26ada2e54370e6da7",
+                "sha256:c190f0899e9f9f8b6b7863debfb739abcb21a5c054f911ca3596d12b8a4c4c7f",
+                "sha256:c907f561b1e83e93fad565bac5ba9c22d96a54e7ea0267c708bffe863cbe4075",
+                "sha256:cae53c389825d3b46fb37538441f75d6aecc4174f615d048321b716df2757fb0",
+                "sha256:dd4a21253f42b8d2b48410cb31fe501d32f8b9fbeb1f55063ad102fe9c425e40",
+                "sha256:dde816ca9dac1d9c01dd504ea5967821606f02e510438120091b84e852367428",
+                "sha256:f2362f3cb0f3172c42938946dbc5b7843c2a28aec307c49100c8b38764eb6927",
+                "sha256:f328adcfebed9f11301eaedfa48e15bdece9b519fb27e6a8c01aa52a17ec31b3",
+                "sha256:f8afcf15cc511ada719a88e013cec87c11aff7b91f019295eb4530f96fe5ef2f",
+                "sha256:fb1bbeac803adea29cedd70781399c99138358c26d05fcbd23c13016b7f5ec65"
+            ],
+            "markers": "python_version < '3.8'",
+            "version": "==1.4.3"
+        },
+        "types-psycopg2": {
+            "hashes": [
+                "sha256:77ed80f2668582654623e04fb3d741ecce93effcc39c929d7e02f4a917a538ce",
+                "sha256:98a6e0e9580cd7eb4bd4d20f7c7063d154b2589a2b90c0ce4e3ca6085cde77c6"
+            ],
+            "index": "pypi",
+            "version": "==2.9.1"
+        },
+        "types-requests": {
+            "hashes": [
+                "sha256:b279284e51f668e38ee12d9665e4d789089f532dc2a0be4a1508ca0efd98ba9e",
+                "sha256:ba1d108d512e294b6080c37f6ae7cb2a2abf527560e2b671d1786c1fc46b541a"
+            ],
+            "index": "pypi",
+            "version": "==2.25.11"
+        },
+        "typing-extensions": {
+            "hashes": [
+                "sha256:49f75d16ff11f1cd258e1b988ccff82a3ca5570217d7ad8c5f48205dd99a677e",
+                "sha256:d8226d10bc02a29bcc81df19a26e56a9647f8b0a6d4a83924139f4a8b01f17b7",
+                "sha256:f1d25edafde516b146ecd0613dabcc61409817af4766fbbcfb8d1ad4ec441a34"
+            ],
+            "index": "pypi",
+            "version": "==3.10.0.2"
+        },
+        "virtualenv": {
+            "hashes": [
+                "sha256:4b02e52a624336eece99c96e3ab7111f469c24ba226a53ec474e8e787b365814",
+                "sha256:576d05b46eace16a9c348085f7d0dc8ef28713a2cabaa1cf0aea41e8f12c9218"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
+            "version": "==20.10.0"
+        },
+        "virtualenv-clone": {
+            "hashes": [
+                "sha256:418ee935c36152f8f153c79824bb93eaf6f0f7984bae31d3f48f350b9183501a",
+                "sha256:44d5263bceed0bac3e1424d64f798095233b64def1c5689afa43dc3223caf5b0"
+            ],
+            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
+            "version": "==0.5.7"
+        },
+        "yapf": {
+            "hashes": [
+                "sha256:408fb9a2b254c302f49db83c59f9aa0b4b0fd0ec25be3a5c51181327922ff63d",
+                "sha256:e3a234ba8455fe201eaa649cdac872d590089a18b661e39bbac7020978dd9c2e"
+            ],
+            "index": "pypi",
+            "version": "==0.31.0"
+        },
+        "zipp": {
+            "hashes": [
+                "sha256:71c644c5369f4a6e07636f0aa966270449561fcea2e3d6747b8d23efaa9d7832",
+                "sha256:9fe5ea21568a0a70e50f273397638d39b03353731e6cbbb3fd8502a33fec40bc"
+            ],
+            "markers": "python_version >= '3.6'",
+            "version": "==3.6.0"
+        }
+    }
+}

README.md

@@ -28,12 +28,12 @@ apt install build-essential libtool libreadline-dev zlib1g-dev flex bison libsec
 libssl-dev clang pkg-config libpq-dev
 ```
 
-[Rust] 1.52 or later is also required.
+[Rust] 1.55 or later is also required.
 
 To run the `psql` client, install the `postgresql-client` package or modify `PATH` and `LD_LIBRARY_PATH` to include `tmp_install/bin` and `tmp_install/lib`, respectively.
 
-To run the integration tests (not required to use the code), install
-Python (3.6 or higher), and install python3 packages with `pipenv` using `pipenv install` in the project directory.
+To run the integration tests or Python scripts (not required to use the code), install
+Python (3.7 or higher), and install python3 packages using `pipenv install` in the project directory.
 
 2. Build zenith and patched postgres
 ```sh
@@ -47,17 +47,26 @@ make -j5
 # Create repository in .zenith with proper paths to binaries and data
 # Later that would be responsibility of a package install script
 > ./target/debug/zenith init
+initializing tenantid c03ba6b7ad4c5e9cf556f059ade44229
+created initial timeline 5b014a9e41b4b63ce1a1febc04503636 timeline.lsn 0/169C3C8
+created main branch
 pageserver init succeeded
 
-# start pageserver
+# start pageserver and safekeeper
 > ./target/debug/zenith start
-Starting pageserver at '127.0.0.1:64000' in .zenith
+Starting pageserver at 'localhost:64000' in '.zenith'
 Pageserver started
+initializing for single for 7676
+Starting safekeeper at 'localhost:5454' in '.zenith/safekeepers/single'
+Safekeeper started
 
-# start postgres on top on the pageserver
+# start postgres compute node
 > ./target/debug/zenith pg start main
-Starting postgres node at 'host=127.0.0.1 port=55432 user=stas'
+Starting new postgres main on main...
+Extracting base backup to create postgres instance: path=.zenith/pgdatadirs/tenants/c03ba6b7ad4c5e9cf556f059ade44229/main port=55432
+Starting postgres node at 'host=127.0.0.1 port=55432 user=zenith_admin dbname=postgres'
 waiting for server to start.... done
+server started
 
 # check list of running postgres instances
 > ./target/debug/zenith pg list
@@ -108,10 +117,9 @@ postgres=# insert into t values(2,2);
 INSERT 0 1
 ```
 
-6. If you want to run tests afterwards (see below), you have to stop pageserver and all postgres instances you have just started:
+6. If you want to run tests afterwards (see below), you have to stop all the running the pageserver, safekeeper and postgres instances
+   you have just started. You can stop them all with one command:
 ```sh
-> ./target/debug/zenith pg stop migration_check
-> ./target/debug/zenith pg stop main
 > ./target/debug/zenith stop
 ```
 
@@ -121,7 +129,7 @@ INSERT 0 1
 git clone --recursive https://github.com/zenithdb/zenith.git
 make # builds also postgres and installs it to ./tmp_install
 cd test_runner
-pytest
+pipenv run pytest
 ```
 
 ## Documentation

control_plane/Cargo.toml

@@ -18,7 +18,7 @@ regex = "1"
 anyhow = "1.0"
 thiserror = "1"
 bytes = "1.0.1"
-nix = "0.20"
+nix = "0.23"
 url = "2.2.2"
 hex = { version = "0.4.3", features = ["serde"] }
 reqwest = { version = "0.11", features = ["blocking", "json"] }

control_plane/safekeepers.conf

@@ -0,0 +1,20 @@
+# Page server and three safekeepers.
+[pageserver]
+pg_port = 64000
+http_port = 9898
+auth_type = 'Trust'
+
+[[safekeepers]]
+name = 'sk1'
+pg_port = 5454
+http_port = 7676
+
+[[safekeepers]]
+name = 'sk2'
+pg_port = 5455
+http_port = 7677
+
+[[safekeepers]]
+name = 'sk3'
+pg_port = 5456
+http_port = 7678

control_plane/simple.conf

@@ -0,0 +1,11 @@
+# Minimal zenith environment with one safekeeper. This is equivalent to the built-in
+# defaults that you get with no --config
+[pageserver]
+pg_port = 64000
+http_port = 9898
+auth_type = 'Trust'
+
+[[safekeepers]]
+name = 'single'
+pg_port = 5454
+http_port = 7676

control_plane/src/compute.rs

@@ -39,8 +39,6 @@ impl ComputeControlPlane {
     // |  |- <tenant_id>
     // |  |   |- <branch name>
     pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
-        // TODO: since pageserver do not have config file yet we believe here that
-        // it is running on default port. Change that when pageserver will have config.
         let pageserver = Arc::new(PageServerNode::from_env(&env));
 
         let mut nodes = BTreeMap::default();
@@ -75,40 +73,59 @@ impl ComputeControlPlane {
             .unwrap_or(self.base_port)
     }
 
-    pub fn local(local_env: &LocalEnv, pageserver: &Arc<PageServerNode>) -> ComputeControlPlane {
-        ComputeControlPlane {
-            base_port: 65431,
-            pageserver: Arc::clone(pageserver),
-            nodes: BTreeMap::new(),
-            env: local_env.clone(),
+    // FIXME: see also parse_point_in_time in branches.rs.
+    fn parse_point_in_time(
+        &self,
+        tenantid: ZTenantId,
+        s: &str,
+    ) -> Result<(ZTimelineId, Option<Lsn>)> {
+        let mut strings = s.split('@');
+        let name = strings.next().unwrap();
+
+        let lsn: Option<Lsn>;
+        if let Some(lsnstr) = strings.next() {
+            lsn = Some(
+                Lsn::from_str(lsnstr)
+                    .with_context(|| "invalid LSN in point-in-time specification")?,
+            );
+        } else {
+            lsn = None
         }
+
+        // Resolve the timeline ID, given the human-readable branch name
+        let timeline_id = self
+            .pageserver
+            .branch_get_by_name(&tenantid, name)?
+            .timeline_id;
+
+        Ok((timeline_id, lsn))
     }
 
     pub fn new_node(
         &mut self,
         tenantid: ZTenantId,
-        branch_name: &str,
+        name: &str,
+        timeline_spec: &str,
         port: Option<u16>,
     ) -> Result<Arc<PostgresNode>> {
-        let timeline_id = self
-            .pageserver
-            .branch_get_by_name(&tenantid, branch_name)?
-            .timeline_id;
+        // Resolve the human-readable timeline spec into timeline ID and LSN
+        let (timelineid, lsn) = self.parse_point_in_time(tenantid, timeline_spec)?;
 
         let port = port.unwrap_or_else(|| self.get_port());
         let node = Arc::new(PostgresNode {
-            name: branch_name.to_owned(),
+            name: name.to_owned(),
             address: SocketAddr::new("127.0.0.1".parse().unwrap(), port),
             env: self.env.clone(),
             pageserver: Arc::clone(&self.pageserver),
             is_test: false,
-            timelineid: timeline_id,
+            timelineid,
+            lsn,
             tenantid,
             uses_wal_proposer: false,
         });
 
         node.create_pgdata()?;
-        node.setup_pg_conf(self.env.auth_type)?;
+        node.setup_pg_conf(self.env.pageserver.auth_type)?;
 
         self.nodes
             .insert((tenantid, node.name.clone()), Arc::clone(&node));
@@ -127,6 +144,7 @@ pub struct PostgresNode {
     pageserver: Arc<PageServerNode>,
     is_test: bool,
     pub timelineid: ZTimelineId,
+    pub lsn: Option<Lsn>, // if it's a read-only node. None for primary
     pub tenantid: ZTenantId,
     uses_wal_proposer: bool,
 }
@@ -161,9 +179,12 @@ impl PostgresNode {
         let port: u16 = conf.parse_field("port", &context)?;
         let timelineid: ZTimelineId = conf.parse_field("zenith.zenith_timeline", &context)?;
         let tenantid: ZTenantId = conf.parse_field("zenith.zenith_tenant", &context)?;
-
         let uses_wal_proposer = conf.get("wal_acceptors").is_some();
 
+        // parse recovery_target_lsn, if any
+        let recovery_target_lsn: Option<Lsn> =
+            conf.parse_field_optional("recovery_target_lsn", &context)?;
+
         // ok now
         Ok(PostgresNode {
             address: SocketAddr::new("127.0.0.1".parse().unwrap(), port),
@@ -172,12 +193,13 @@ impl PostgresNode {
             pageserver: Arc::clone(pageserver),
             is_test: false,
             timelineid,
+            lsn: recovery_target_lsn,
             tenantid,
             uses_wal_proposer,
         })
     }
 
-    fn sync_walkeepers(&self) -> Result<Lsn> {
+    fn sync_safekeepers(&self) -> Result<Lsn> {
         let pg_path = self.env.pg_bin_dir().join("postgres");
         let sync_handle = Command::new(pg_path)
             .arg("--sync-safekeepers")
@@ -202,7 +224,7 @@ impl PostgresNode {
         }
 
         let lsn = Lsn::from_str(std::str::from_utf8(&sync_output.stdout)?.trim())?;
-        println!("Walkeepers synced on {}", lsn);
+        println!("Safekeepers synced on {}", lsn);
         Ok(lsn)
     }
 
@@ -233,7 +255,7 @@ impl PostgresNode {
         // Read the archive directly from the `CopyOutReader`
         tar::Archive::new(copyreader)
             .unpack(&self.pgdata())
-            .with_context(|| "extracting page backup failed")?;
+            .with_context(|| "extracting base backup failed")?;
 
         Ok(())
     }
@@ -265,10 +287,15 @@ impl PostgresNode {
         conf.append("max_replication_slots", "10");
         conf.append("hot_standby", "on");
         conf.append("shared_buffers", "1MB");
+        conf.append("max_wal_size", "100GB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
-        conf.append("wal_sender_timeout", "0");
         conf.append("wal_level", "replica");
+        // wal_sender_timeout is the maximum time to wait for WAL replication.
+        // It also defines how often the walreciever will send a feedback message to the wal sender.
+        //conf.append("wal_sender_timeout", "5s");
+        //conf.append("max_replication_flush_lag", "160MB");
+        //conf.append("max_replication_apply_lag", "1500MB");
         conf.append("listen_addresses", &self.address.ip().to_string());
         conf.append("port", &self.address.port().to_string());
 
@@ -301,11 +328,30 @@ impl PostgresNode {
         conf.append("zenith.page_server_connstring", &pageserver_connstr);
         conf.append("zenith.zenith_tenant", &self.tenantid.to_string());
         conf.append("zenith.zenith_timeline", &self.timelineid.to_string());
+        if let Some(lsn) = self.lsn {
+            conf.append("recovery_target_lsn", &lsn.to_string());
+        }
         conf.append_line("");
 
-        // Configure the node to stream WAL directly to the pageserver
-        conf.append("synchronous_standby_names", "pageserver"); // TODO: add a new function arg?
-        conf.append("zenith.callmemaybe_connstring", &self.connstr());
+        if !self.env.safekeepers.is_empty() {
+            // Configure the node to connect to the safekeepers
+            conf.append("synchronous_standby_names", "walproposer");
+
+            let wal_acceptors = self
+                .env
+                .safekeepers
+                .iter()
+                .map(|sk| format!("localhost:{}", sk.pg_port))
+                .collect::<Vec<String>>()
+                .join(",");
+            conf.append("wal_acceptors", &wal_acceptors);
+        } else {
+            // Configure the node to stream WAL directly to the pageserver
+            // This isn't really a supported configuration, but can be useful for
+            // testing.
+            conf.append("synchronous_standby_names", "pageserver");
+            conf.append("zenith.callmemaybe_connstring", &self.connstr());
+        }
 
         let mut file = File::create(self.pgdata().join("postgresql.conf"))?;
         file.write_all(conf.to_string().as_bytes())?;
@@ -314,12 +360,14 @@ impl PostgresNode {
     }
 
     fn load_basebackup(&self) -> Result<()> {
-        let lsn = if self.uses_wal_proposer {
+        let backup_lsn = if let Some(lsn) = self.lsn {
+            Some(lsn)
+        } else if self.uses_wal_proposer {
             // LSN 0 means that it is bootstrap and we need to download just
             // latest data from the pageserver. That is a bit clumsy but whole bootstrap
             // procedure evolves quite actively right now, so let's think about it again
             // when things would be more stable (TODO).
-            let lsn = self.sync_walkeepers()?;
+            let lsn = self.sync_safekeepers()?;
             if lsn == Lsn(0) {
                 None
             } else {
@@ -329,7 +377,7 @@ impl PostgresNode {
             None
         };
 
-        self.do_basebackup(lsn)?;
+        self.do_basebackup(backup_lsn)?;
 
         Ok(())
     }
@@ -406,6 +454,10 @@ impl PostgresNode {
         // 3. Load basebackup
         self.load_basebackup()?;
 
+        if self.lsn.is_some() {
+            File::create(self.pgdata().join("standby.signal"))?;
+        }
+
         // 4. Finally start the compute node postgres
         println!("Starting postgres node at '{}'", self.connstr());
         self.pg_ctl(&["start"], auth_token)

control_plane/src/lib.rs

@@ -13,6 +13,7 @@ use std::path::Path;
 pub mod compute;
 pub mod local_env;
 pub mod postgresql_conf;
+pub mod safekeeper;
 pub mod storage;
 
 /// Read a PID file

control_plane/src/local_env.rs

@@ -7,46 +7,102 @@
 use anyhow::{Context, Result};
 use serde::{Deserialize, Serialize};
 use std::env;
+use std::fmt::Write;
 use std::fs;
-use std::path::PathBuf;
+use std::path::{Path, PathBuf};
 use std::process::{Command, Stdio};
-use zenith_utils::auth::{encode_from_key_path, Claims, Scope};
+use zenith_utils::auth::{encode_from_key_file, Claims, Scope};
 use zenith_utils::postgres_backend::AuthType;
 use zenith_utils::zid::ZTenantId;
 
 //
-// This data structures represent deserialized zenith CLI config
+// This data structures represents zenith CLI config
+//
+// It is deserialized from the .zenith/config file, or the config file passed
+// to 'zenith init --config=<path>' option. See control_plane/simple.conf for
+// an example.
 //
 #[derive(Serialize, Deserialize, Clone, Debug)]
 pub struct LocalEnv {
-    // Pageserver connection settings
-    pub pageserver_pg_port: u16,
-    pub pageserver_http_port: u16,
-
-    // Base directory for both pageserver and compute nodes
+    // Base directory for all the nodes (the pageserver, safekeepers and
+    // compute nodes).
+    //
+    // This is not stored in the config file. Rather, this is the path where the
+    // config file itself is. It is read from the ZENITH_REPO_DIR env variable or
+    // '.zenith' if not given.
+    #[serde(skip)]
     pub base_data_dir: PathBuf,
 
     // Path to postgres distribution. It's expected that "bin", "include",
     // "lib", "share" from postgres distribution are there. If at some point
     // in time we will be able to run against vanilla postgres we may split that
     // to four separate paths and match OS-specific installation layout.
+    #[serde(default)]
     pub pg_distrib_dir: PathBuf,
 
     // Path to pageserver binary.
+    #[serde(default)]
     pub zenith_distrib_dir: PathBuf,
 
-    // keeping tenant id in config to reduce copy paste when running zenith locally with single tenant
-    #[serde(with = "hex")]
-    pub tenantid: ZTenantId,
+    // Default tenant ID to use with the 'zenith' command line utility, when
+    // --tenantid is not explicitly specified.
+    #[serde(with = "opt_tenantid_serde")]
+    #[serde(default)]
+    pub default_tenantid: Option<ZTenantId>,
 
-    // jwt auth token used for communication with pageserver
-    pub auth_token: String,
+    // used to issue tokens during e.g pg start
+    #[serde(default)]
+    pub private_key_path: PathBuf,
+
+    pub pageserver: PageServerConf,
+
+    #[serde(default)]
+    pub safekeepers: Vec<SafekeeperConf>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+#[serde(default)]
+pub struct PageServerConf {
+    // Pageserver connection settings
+    pub pg_port: u16,
+    pub http_port: u16,
 
     // used to determine which auth type is used
     pub auth_type: AuthType,
 
-    // used to issue tokens during e.g pg start
-    pub private_key_path: PathBuf,
+    // jwt auth token used for communication with pageserver
+    pub auth_token: String,
+}
+
+impl Default for PageServerConf {
+    fn default() -> Self {
+        Self {
+            pg_port: 0,
+            http_port: 0,
+            auth_type: AuthType::Trust,
+            auth_token: "".to_string(),
+        }
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+#[serde(default)]
+pub struct SafekeeperConf {
+    pub name: String,
+    pub pg_port: u16,
+    pub http_port: u16,
+    pub sync: bool,
+}
+
+impl Default for SafekeeperConf {
+    fn default() -> Self {
+        Self {
+            name: "".to_string(),
+            pg_port: 0,
+            http_port: 0,
+            sync: true,
+        }
+    }
 }
 
 impl LocalEnv {
@@ -62,6 +118,10 @@ impl LocalEnv {
         Ok(self.zenith_distrib_dir.join("pageserver"))
     }
 
+    pub fn safekeeper_bin(&self) -> Result<PathBuf> {
+        Ok(self.zenith_distrib_dir.join("safekeeper"))
+    }
+
     pub fn pg_data_dirs_path(&self) -> PathBuf {
         self.base_data_dir.join("pgdatadirs").join("tenants")
     }
@@ -76,6 +136,187 @@ impl LocalEnv {
     pub fn pageserver_data_dir(&self) -> PathBuf {
         self.base_data_dir.clone()
     }
+
+    pub fn safekeeper_data_dir(&self, node_name: &str) -> PathBuf {
+        self.base_data_dir.join("safekeepers").join(node_name)
+    }
+
+    /// Create a LocalEnv from a config file.
+    ///
+    /// Unlike 'load_config', this function fills in any defaults that are missing
+    /// from the config file.
+    pub fn create_config(toml: &str) -> Result<LocalEnv> {
+        let mut env: LocalEnv = toml::from_str(toml)?;
+
+        // Find postgres binaries.
+        // Follow POSTGRES_DISTRIB_DIR if set, otherwise look in "tmp_install".
+        if env.pg_distrib_dir == Path::new("") {
+            if let Some(postgres_bin) = env::var_os("POSTGRES_DISTRIB_DIR") {
+                env.pg_distrib_dir = postgres_bin.into();
+            } else {
+                let cwd = env::current_dir()?;
+                env.pg_distrib_dir = cwd.join("tmp_install")
+            }
+        }
+        if !env.pg_distrib_dir.join("bin/postgres").exists() {
+            anyhow::bail!(
+                "Can't find postgres binary at {}",
+                env.pg_distrib_dir.display()
+            );
+        }
+
+        // Find zenith binaries.
+        if env.zenith_distrib_dir == Path::new("") {
+            env.zenith_distrib_dir = env::current_exe()?.parent().unwrap().to_owned();
+        }
+        if !env.zenith_distrib_dir.join("pageserver").exists() {
+            anyhow::bail!("Can't find pageserver binary.");
+        }
+        if !env.zenith_distrib_dir.join("safekeeper").exists() {
+            anyhow::bail!("Can't find safekeeper binary.");
+        }
+
+        // If no initial tenant ID was given, generate it.
+        if env.default_tenantid.is_none() {
+            env.default_tenantid = Some(ZTenantId::generate());
+        }
+
+        env.base_data_dir = base_path();
+
+        Ok(env)
+    }
+
+    /// Locate and load config
+    pub fn load_config() -> Result<LocalEnv> {
+        let repopath = base_path();
+
+        if !repopath.exists() {
+            anyhow::bail!(
+                "Zenith config is not found in {}. You need to run 'zenith init' first",
+                repopath.to_str().unwrap()
+            );
+        }
+
+        // TODO: check that it looks like a zenith repository
+
+        // load and parse file
+        let config = fs::read_to_string(repopath.join("config"))?;
+        let mut env: LocalEnv = toml::from_str(config.as_str())?;
+
+        env.base_data_dir = repopath;
+
+        Ok(env)
+    }
+
+    // this function is used only for testing purposes in CLI e g generate tokens during init
+    pub fn generate_auth_token(&self, claims: &Claims) -> Result<String> {
+        let private_key_path = if self.private_key_path.is_absolute() {
+            self.private_key_path.to_path_buf()
+        } else {
+            self.base_data_dir.join(&self.private_key_path)
+        };
+
+        let key_data = fs::read(private_key_path)?;
+        encode_from_key_file(claims, &key_data)
+    }
+
+    //
+    // Initialize a new Zenith repository
+    //
+    pub fn init(&mut self) -> Result<()> {
+        // check if config already exists
+        let base_path = &self.base_data_dir;
+        if base_path == Path::new("") {
+            anyhow::bail!("repository base path is missing");
+        }
+        if base_path.exists() {
+            anyhow::bail!(
+                "directory '{}' already exists. Perhaps already initialized?",
+                base_path.to_str().unwrap()
+            );
+        }
+
+        fs::create_dir(&base_path)?;
+
+        // generate keys for jwt
+        // openssl genrsa -out private_key.pem 2048
+        let private_key_path;
+        if self.private_key_path == PathBuf::new() {
+            private_key_path = base_path.join("auth_private_key.pem");
+            let keygen_output = Command::new("openssl")
+                .arg("genrsa")
+                .args(&["-out", private_key_path.to_str().unwrap()])
+                .arg("2048")
+                .stdout(Stdio::null())
+                .output()
+                .with_context(|| "failed to generate auth private key")?;
+            if !keygen_output.status.success() {
+                anyhow::bail!(
+                    "openssl failed: '{}'",
+                    String::from_utf8_lossy(&keygen_output.stderr)
+                );
+            }
+            self.private_key_path = Path::new("auth_private_key.pem").to_path_buf();
+
+            let public_key_path = base_path.join("auth_public_key.pem");
+            // openssl rsa -in private_key.pem -pubout -outform PEM -out public_key.pem
+            let keygen_output = Command::new("openssl")
+                .arg("rsa")
+                .args(&["-in", private_key_path.to_str().unwrap()])
+                .arg("-pubout")
+                .args(&["-outform", "PEM"])
+                .args(&["-out", public_key_path.to_str().unwrap()])
+                .stdout(Stdio::null())
+                .output()
+                .with_context(|| "failed to generate auth private key")?;
+            if !keygen_output.status.success() {
+                anyhow::bail!(
+                    "openssl failed: '{}'",
+                    String::from_utf8_lossy(&keygen_output.stderr)
+                );
+            }
+        }
+
+        self.pageserver.auth_token =
+            self.generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
+
+        fs::create_dir_all(self.pg_data_dirs_path())?;
+
+        for safekeeper in self.safekeepers.iter() {
+            fs::create_dir_all(self.safekeeper_data_dir(&safekeeper.name))?;
+        }
+
+        let mut conf_content = String::new();
+
+        // Currently, the user first passes a config file with 'zenith init --config=<path>'
+        // We read that in, in `create_config`, and fill any missing defaults. Then it's saved
+        // to .zenith/config. TODO: We lose any formatting and comments along the way, which is
+        // a bit sad.
+        write!(
+            &mut conf_content,
+            r#"# This file describes a locale deployment of the page server
+# and safekeeeper node. It is read by the 'zenith' command-line
+# utility.
+"#
+        )?;
+
+        // Convert the LocalEnv to a toml file.
+        //
+        // This could be as simple as this:
+        //
+        // conf_content += &toml::to_string_pretty(env)?;
+        //
+        // But it results in a "values must be emitted before tables". I'm not sure
+        // why, AFAICS the table, i.e. 'safekeepers: Vec<SafekeeperConf>' is last.
+        // Maybe rust reorders the fields to squeeze avoid padding or something?
+        // In any case, converting to toml::Value first, and serializing that, works.
+        // See https://github.com/alexcrichton/toml-rs/issues/142
+        conf_content += &toml::to_string_pretty(&toml::Value::try_from(&self)?)?;
+
+        fs::write(base_path.join("config"), conf_content)?;
+
+        Ok(())
+    }
 }
 
 fn base_path() -> PathBuf {
@@ -85,118 +326,29 @@ fn base_path() -> PathBuf {
     }
 }
 
-//
-// Initialize a new Zenith repository
-//
-pub fn init(
-    pageserver_pg_port: u16,
-    pageserver_http_port: u16,
-    tenantid: ZTenantId,
-    auth_type: AuthType,
-) -> Result<()> {
-    // check if config already exists
-    let base_path = base_path();
-    if base_path.exists() {
-        anyhow::bail!(
-            "{} already exists. Perhaps already initialized?",
-            base_path.to_str().unwrap()
-        );
+/// Serde routines for Option<ZTenantId>. The serialized form is a hex string.
+mod opt_tenantid_serde {
+    use serde::{Deserialize, Deserializer, Serialize, Serializer};
+    use std::str::FromStr;
+    use zenith_utils::zid::ZTenantId;
+
+    pub fn serialize<S>(tenantid: &Option<ZTenantId>, ser: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        tenantid.map(|t| t.to_string()).serialize(ser)
     }
-    fs::create_dir(&base_path)?;
 
-    // ok, now check that expected binaries are present
-
-    // Find postgres binaries. Follow POSTGRES_DISTRIB_DIR if set, otherwise look in "tmp_install".
-    let pg_distrib_dir: PathBuf = {
-        if let Some(postgres_bin) = env::var_os("POSTGRES_DISTRIB_DIR") {
-            postgres_bin.into()
-        } else {
-            let cwd = env::current_dir()?;
-            cwd.join("tmp_install")
+    pub fn deserialize<'de, D>(des: D) -> Result<Option<ZTenantId>, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        let s: Option<String> = Option::deserialize(des)?;
+        if let Some(s) = s {
+            return Ok(Some(
+                ZTenantId::from_str(&s).map_err(serde::de::Error::custom)?,
+            ));
         }
-    };
-    if !pg_distrib_dir.join("bin/postgres").exists() {
-        anyhow::bail!("Can't find postgres binary at {:?}", pg_distrib_dir);
+        Ok(None)
     }
-
-    // generate keys for jwt
-    // openssl genrsa -out private_key.pem 2048
-    let private_key_path = base_path.join("auth_private_key.pem");
-    let keygen_output = Command::new("openssl")
-        .arg("genrsa")
-        .args(&["-out", private_key_path.to_str().unwrap()])
-        .arg("2048")
-        .stdout(Stdio::null())
-        .output()
-        .with_context(|| "failed to generate auth private key")?;
-    if !keygen_output.status.success() {
-        anyhow::bail!(
-            "openssl failed: '{}'",
-            String::from_utf8_lossy(&keygen_output.stderr)
-        );
-    }
-
-    let public_key_path = base_path.join("auth_public_key.pem");
-    // openssl rsa -in private_key.pem -pubout -outform PEM -out public_key.pem
-    let keygen_output = Command::new("openssl")
-        .arg("rsa")
-        .args(&["-in", private_key_path.to_str().unwrap()])
-        .arg("-pubout")
-        .args(&["-outform", "PEM"])
-        .args(&["-out", public_key_path.to_str().unwrap()])
-        .stdout(Stdio::null())
-        .output()
-        .with_context(|| "failed to generate auth private key")?;
-    if !keygen_output.status.success() {
-        anyhow::bail!(
-            "openssl failed: '{}'",
-            String::from_utf8_lossy(&keygen_output.stderr)
-        );
-    }
-
-    let auth_token =
-        encode_from_key_path(&Claims::new(None, Scope::PageServerApi), &private_key_path)?;
-
-    // Find zenith binaries.
-    let zenith_distrib_dir = env::current_exe()?.parent().unwrap().to_owned();
-    if !zenith_distrib_dir.join("pageserver").exists() {
-        anyhow::bail!("Can't find pageserver binary.",);
-    }
-
-    let conf = LocalEnv {
-        pageserver_pg_port,
-        pageserver_http_port,
-        pg_distrib_dir,
-        zenith_distrib_dir,
-        base_data_dir: base_path,
-        tenantid,
-        auth_token,
-        auth_type,
-        private_key_path,
-    };
-
-    fs::create_dir_all(conf.pg_data_dirs_path())?;
-
-    let toml = toml::to_string_pretty(&conf)?;
-    fs::write(conf.base_data_dir.join("config"), toml)?;
-
-    Ok(())
-}
-
-// Locate and load config
-pub fn load_config() -> Result<LocalEnv> {
-    let repopath = base_path();
-
-    if !repopath.exists() {
-        anyhow::bail!(
-            "Zenith config is not found in {}. You need to run 'zenith init' first",
-            repopath.to_str().unwrap()
-        );
-    }
-
-    // TODO: check that it looks like a zenith repository
-
-    // load and parse file
-    let config = fs::read_to_string(repopath.join("config"))?;
-    toml::from_str(config.as_str()).map_err(|e| e.into())
 }

control_plane/src/postgresql_conf.rs

@@ -83,6 +83,22 @@ impl PostgresConf {
             .with_context(|| format!("could not parse '{}' option {}", field_name, context))
     }
 
+    pub fn parse_field_optional<T>(&self, field_name: &str, context: &str) -> Result<Option<T>>
+    where
+        T: FromStr,
+        <T as FromStr>::Err: std::error::Error + Send + Sync + 'static,
+    {
+        if let Some(val) = self.get(field_name) {
+            let result = val
+                .parse::<T>()
+                .with_context(|| format!("could not parse '{}' option {}", field_name, context))?;
+
+            Ok(Some(result))
+        } else {
+            Ok(None)
+        }
+    }
+
     ///
     /// Note: if you call this multiple times for the same option, the config
     /// file will a line for each call. It would be nice to have a function

control_plane/src/safekeeper.rs

@@ -0,0 +1,281 @@
+use std::io::Write;
+use std::net::TcpStream;
+use std::path::PathBuf;
+use std::process::Command;
+use std::sync::Arc;
+use std::time::Duration;
+use std::{io, result, thread};
+
+use anyhow::bail;
+use nix::errno::Errno;
+use nix::sys::signal::{kill, Signal};
+use nix::unistd::Pid;
+use postgres::Config;
+use reqwest::blocking::{Client, RequestBuilder, Response};
+use reqwest::{IntoUrl, Method};
+use thiserror::Error;
+use zenith_utils::http::error::HttpErrorBody;
+use zenith_utils::postgres_backend::AuthType;
+
+use crate::local_env::{LocalEnv, SafekeeperConf};
+use crate::read_pidfile;
+use crate::storage::PageServerNode;
+use zenith_utils::connstring::connection_address;
+use zenith_utils::connstring::connection_host_port;
+
+#[derive(Error, Debug)]
+pub enum SafekeeperHttpError {
+    #[error("Reqwest error: {0}")]
+    Transport(#[from] reqwest::Error),
+
+    #[error("Error: {0}")]
+    Response(String),
+}
+
+type Result<T> = result::Result<T, SafekeeperHttpError>;
+
+pub trait ResponseErrorMessageExt: Sized {
+    fn error_from_body(self) -> Result<Self>;
+}
+
+impl ResponseErrorMessageExt for Response {
+    fn error_from_body(self) -> Result<Self> {
+        let status = self.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            return Ok(self);
+        }
+
+        // reqwest do not export it's error construction utility functions, so lets craft the message ourselves
+        let url = self.url().to_owned();
+        Err(SafekeeperHttpError::Response(
+            match self.json::<HttpErrorBody>() {
+                Ok(err_body) => format!("Error: {}", err_body.msg),
+                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
+            },
+        ))
+    }
+}
+
+//
+// Control routines for safekeeper.
+//
+// Used in CLI and tests.
+//
+#[derive(Debug)]
+pub struct SafekeeperNode {
+    pub name: String,
+
+    pub conf: SafekeeperConf,
+
+    pub pg_connection_config: Config,
+    pub env: LocalEnv,
+    pub http_client: Client,
+    pub http_base_url: String,
+
+    pub pageserver: Arc<PageServerNode>,
+}
+
+impl SafekeeperNode {
+    pub fn from_env(env: &LocalEnv, conf: &SafekeeperConf) -> SafekeeperNode {
+        let pageserver = Arc::new(PageServerNode::from_env(env));
+
+        println!("initializing for {} for {}", conf.name, conf.http_port);
+
+        SafekeeperNode {
+            name: conf.name.clone(),
+            conf: conf.clone(),
+            pg_connection_config: Self::safekeeper_connection_config(conf.pg_port),
+            env: env.clone(),
+            http_client: Client::new(),
+            http_base_url: format!("http://localhost:{}/v1", conf.http_port),
+            pageserver,
+        }
+    }
+
+    /// Construct libpq connection string for connecting to this safekeeper.
+    fn safekeeper_connection_config(port: u16) -> Config {
+        // TODO safekeeper authentication not implemented yet
+        format!("postgresql://no_user@localhost:{}/no_db", port)
+            .parse()
+            .unwrap()
+    }
+
+    pub fn datadir_path(&self) -> PathBuf {
+        self.env.safekeeper_data_dir(&self.name)
+    }
+
+    pub fn pid_file(&self) -> PathBuf {
+        self.datadir_path().join("safekeeper.pid")
+    }
+
+    pub fn start(&self) -> anyhow::Result<()> {
+        print!(
+            "Starting safekeeper at '{}' in '{}'",
+            connection_address(&self.pg_connection_config),
+            self.datadir_path().display()
+        );
+        io::stdout().flush().unwrap();
+
+        // Configure connection to page server
+        //
+        // FIXME: We extract the host and port from the connection string instead of using
+        // the connection string directly, because the 'safekeeper' binary expects
+        // host:port format. That's a bit silly when we already have a full libpq connection
+        // string at hand.
+        let pageserver_conn = {
+            let (host, port) = connection_host_port(&self.pageserver.pg_connection_config);
+            format!("{}:{}", host, port)
+        };
+
+        let listen_pg = format!("localhost:{}", self.conf.pg_port);
+        let listen_http = format!("localhost:{}", self.conf.http_port);
+
+        let mut cmd = Command::new(self.env.safekeeper_bin()?);
+        cmd.args(&["-D", self.datadir_path().to_str().unwrap()])
+            .args(&["--listen-pg", &listen_pg])
+            .args(&["--listen-http", &listen_http])
+            .args(&["--pageserver", &pageserver_conn])
+            .args(&["--recall", "1 second"])
+            .arg("--daemonize")
+            .env_clear()
+            .env("RUST_BACKTRACE", "1");
+        if !self.conf.sync {
+            cmd.arg("--no-sync");
+        }
+
+        if self.env.pageserver.auth_type == AuthType::ZenithJWT {
+            cmd.env("PAGESERVER_AUTH_TOKEN", &self.env.pageserver.auth_token);
+        }
+
+        let var = "LLVM_PROFILE_FILE";
+        if let Some(val) = std::env::var_os(var) {
+            cmd.env(var, val);
+        }
+
+        if !cmd.status()?.success() {
+            bail!(
+                "Safekeeper failed to start. See '{}' for details.",
+                self.datadir_path().join("safekeeper.log").display()
+            );
+        }
+
+        // It takes a while for the safekeeper to start up. Wait until it is
+        // open for business.
+        const RETRIES: i8 = 15;
+        for retries in 1..RETRIES {
+            match self.check_status() {
+                Ok(_) => {
+                    println!("\nSafekeeper started");
+                    return Ok(());
+                }
+                Err(err) => {
+                    match err {
+                        SafekeeperHttpError::Transport(err) => {
+                            if err.is_connect() && retries < 5 {
+                                print!(".");
+                                io::stdout().flush().unwrap();
+                            } else {
+                                if retries == 5 {
+                                    println!() // put a line break after dots for second message
+                                }
+                                println!(
+                                    "Safekeeper not responding yet, err {} retrying ({})...",
+                                    err, retries
+                                );
+                            }
+                        }
+                        SafekeeperHttpError::Response(msg) => {
+                            bail!("safekeeper failed to start: {} ", msg)
+                        }
+                    }
+                    thread::sleep(Duration::from_secs(1));
+                }
+            }
+        }
+        bail!("safekeeper failed to start in {} seconds", RETRIES);
+    }
+
+    ///
+    /// Stop the server.
+    ///
+    /// If 'immediate' is true, we use SIGQUIT, killing the process immediately.
+    /// Otherwise we use SIGTERM, triggering a clean shutdown
+    ///
+    /// If the server is not running, returns success
+    ///
+    pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
+        let pid_file = self.pid_file();
+        if !pid_file.exists() {
+            println!("Safekeeper {} is already stopped", self.name);
+            return Ok(());
+        }
+        let pid = read_pidfile(&pid_file)?;
+        let pid = Pid::from_raw(pid);
+
+        let sig = if immediate {
+            println!("Stop safekeeper immediately");
+            Signal::SIGQUIT
+        } else {
+            println!("Stop safekeeper gracefully");
+            Signal::SIGTERM
+        };
+        match kill(pid, sig) {
+            Ok(_) => (),
+            Err(Errno::ESRCH) => {
+                println!(
+                    "Safekeeper with pid {} does not exist, but a PID file was found",
+                    pid
+                );
+                return Ok(());
+            }
+            Err(err) => bail!(
+                "Failed to send signal to safekeeper with pid {}: {}",
+                pid,
+                err.desc()
+            ),
+        }
+
+        let address = connection_address(&self.pg_connection_config);
+
+        // TODO Remove this "timeout" and handle it on caller side instead.
+        // Shutting down may take a long time,
+        // if safekeeper flushes a lot of data
+        for _ in 0..100 {
+            if let Err(_e) = TcpStream::connect(&address) {
+                println!("Safekeeper stopped receiving connections");
+
+                //Now check status
+                match self.check_status() {
+                    Ok(_) => {
+                        println!("Safekeeper status is OK. Wait a bit.");
+                        thread::sleep(Duration::from_secs(1));
+                    }
+                    Err(err) => {
+                        println!("Safekeeper status is: {}", err);
+                        return Ok(());
+                    }
+                }
+            } else {
+                println!("Safekeeper still receives connections");
+                thread::sleep(Duration::from_secs(1));
+            }
+        }
+
+        bail!("Failed to stop safekeeper with pid {}", pid);
+    }
+
+    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> RequestBuilder {
+        // TODO: authentication
+        //if self.env.auth_type == AuthType::ZenithJWT {
+        //    builder = builder.bearer_auth(&self.env.safekeeper_auth_token)
+        //}
+        self.http_client.request(method, url)
+    }
+
+    pub fn check_status(&self) -> Result<()> {
+        self.http_request(Method::GET, format!("{}/{}", self.http_base_url, "status"))
+            .send()?
+            .error_from_body()?;
+        Ok(())
+    }
+}

control_plane/src/storage.rs

@@ -5,7 +5,8 @@ use std::process::Command;
 use std::time::Duration;
 use std::{io, result, thread};
 
-use anyhow::{anyhow, bail};
+use anyhow::bail;
+use nix::errno::Errno;
 use nix::sys::signal::{kill, Signal};
 use nix::unistd::Pid;
 use pageserver::http::models::{BranchCreateRequest, TenantCreateRequest};
@@ -20,6 +21,7 @@ use zenith_utils::zid::ZTenantId;
 use crate::local_env::LocalEnv;
 use crate::read_pidfile;
 use pageserver::branches::BranchInfo;
+use pageserver::tenant_mgr::TenantInfo;
 use zenith_utils::connstring::connection_address;
 
 #[derive(Error, Debug)]
@@ -62,7 +64,6 @@ impl ResponseErrorMessageExt for Response {
 //
 #[derive(Debug)]
 pub struct PageServerNode {
-    pub kill_on_exit: bool,
     pub pg_connection_config: Config,
     pub env: LocalEnv,
     pub http_client: Client,
@@ -71,34 +72,33 @@ pub struct PageServerNode {
 
 impl PageServerNode {
     pub fn from_env(env: &LocalEnv) -> PageServerNode {
-        let password = if env.auth_type == AuthType::ZenithJWT {
-            &env.auth_token
+        let password = if env.pageserver.auth_type == AuthType::ZenithJWT {
+            &env.pageserver.auth_token
         } else {
             ""
         };
 
         PageServerNode {
-            kill_on_exit: false,
             pg_connection_config: Self::pageserver_connection_config(
                 password,
-                env.pageserver_pg_port,
+                env.pageserver.pg_port,
             ),
             env: env.clone(),
             http_client: Client::new(),
-            http_base_url: format!("http://localhost:{}/v1", env.pageserver_http_port),
+            http_base_url: format!("http://localhost:{}/v1", env.pageserver.http_port),
         }
     }
 
+    /// Construct libpq connection string for connecting to the pageserver.
     fn pageserver_connection_config(password: &str, port: u16) -> Config {
         format!("postgresql://no_user:{}@localhost:{}/no_db", password, port)
             .parse()
             .unwrap()
     }
 
-    pub fn init(&self, create_tenant: Option<&str>, enable_auth: bool) -> anyhow::Result<()> {
-        let mut cmd = Command::new(self.env.pageserver_bin()?);
-        let listen_pg = format!("localhost:{}", self.env.pageserver_pg_port);
-        let listen_http = format!("localhost:{}", self.env.pageserver_http_port);
+    pub fn init(&self, create_tenant: Option<&str>) -> anyhow::Result<()> {
+        let listen_pg = format!("localhost:{}", self.env.pageserver.pg_port);
+        let listen_http = format!("localhost:{}", self.env.pageserver.http_port);
         let mut args = vec![
             "--init",
             "-D",
@@ -111,27 +111,29 @@ impl PageServerNode {
             &listen_http,
         ];
 
-        if enable_auth {
+        let auth_type_str = &self.env.pageserver.auth_type.to_string();
+        if self.env.pageserver.auth_type != AuthType::Trust {
             args.extend(&["--auth-validation-public-key-path", "auth_public_key.pem"]);
-            args.extend(&["--auth-type", "ZenithJWT"]);
         }
+        args.extend(&["--auth-type", auth_type_str]);
 
         if let Some(tenantid) = create_tenant {
             args.extend(&["--create-tenant", tenantid])
         }
 
-        let status = cmd
-            .args(args)
-            .env_clear()
-            .env("RUST_BACKTRACE", "1")
-            .status()
-            .expect("pageserver init failed");
+        let mut cmd = Command::new(self.env.pageserver_bin()?);
+        cmd.args(args).env_clear().env("RUST_BACKTRACE", "1");
 
-        if status.success() {
-            Ok(())
-        } else {
-            Err(anyhow!("pageserver init failed"))
+        let var = "LLVM_PROFILE_FILE";
+        if let Some(val) = std::env::var_os(var) {
+            cmd.env(var, val);
         }
+
+        if !cmd.status()?.success() {
+            bail!("pageserver init failed");
+        }
+
+        Ok(())
     }
 
     pub fn repo_path(&self) -> PathBuf {
@@ -152,10 +154,15 @@ impl PageServerNode {
 
         let mut cmd = Command::new(self.env.pageserver_bin()?);
         cmd.args(&["-D", self.repo_path().to_str().unwrap()])
-            .arg("-d")
+            .arg("--daemonize")
             .env_clear()
             .env("RUST_BACKTRACE", "1");
 
+        let var = "LLVM_PROFILE_FILE";
+        if let Some(val) = std::env::var_os(var) {
+            cmd.env(var, val);
+        }
+
         if !cmd.status()?.success() {
             bail!(
                 "Pageserver failed to start. See '{}' for details.",
@@ -199,23 +206,69 @@ impl PageServerNode {
         bail!("pageserver failed to start in {} seconds", RETRIES);
     }
 
-    pub fn stop(&self) -> anyhow::Result<()> {
-        let pid = read_pidfile(&self.pid_file())?;
-        let pid = Pid::from_raw(pid);
-        if kill(pid, Signal::SIGTERM).is_err() {
-            bail!("Failed to kill pageserver with pid {}", pid);
+    ///
+    /// Stop the server.
+    ///
+    /// If 'immediate' is true, we use SIGQUIT, killing the process immediately.
+    /// Otherwise we use SIGTERM, triggering a clean shutdown
+    ///
+    /// If the server is not running, returns success
+    ///
+    pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
+        let pid_file = self.pid_file();
+        if !pid_file.exists() {
+            println!("Pageserver is already stopped");
+            return Ok(());
         }
+        let pid = Pid::from_raw(read_pidfile(&pid_file)?);
 
-        // wait for pageserver stop
-        let address = connection_address(&self.pg_connection_config);
-        for _ in 0..5 {
-            let stream = TcpStream::connect(&address);
-            thread::sleep(Duration::from_secs(1));
-            if let Err(_e) = stream {
-                println!("Pageserver stopped");
+        let sig = if immediate {
+            println!("Stop pageserver immediately");
+            Signal::SIGQUIT
+        } else {
+            println!("Stop pageserver gracefully");
+            Signal::SIGTERM
+        };
+        match kill(pid, sig) {
+            Ok(_) => (),
+            Err(Errno::ESRCH) => {
+                println!(
+                    "Pageserver with pid {} does not exist, but a PID file was found",
+                    pid
+                );
                 return Ok(());
             }
-            println!("Stopping pageserver on {}", address);
+            Err(err) => bail!(
+                "Failed to send signal to pageserver with pid {}: {}",
+                pid,
+                err.desc()
+            ),
+        }
+
+        let address = connection_address(&self.pg_connection_config);
+
+        // TODO Remove this "timeout" and handle it on caller side instead.
+        // Shutting down may take a long time,
+        // if pageserver checkpoints a lot of data
+        for _ in 0..100 {
+            if let Err(_e) = TcpStream::connect(&address) {
+                println!("Pageserver stopped receiving connections");
+
+                //Now check status
+                match self.check_status() {
+                    Ok(_) => {
+                        println!("Pageserver status is OK. Wait a bit.");
+                        thread::sleep(Duration::from_secs(1));
+                    }
+                    Err(err) => {
+                        println!("Pageserver status is: {}", err);
+                        return Ok(());
+                    }
+                }
+            } else {
+                println!("Pageserver still receives connections");
+                thread::sleep(Duration::from_secs(1));
+            }
         }
 
         bail!("Failed to stop pageserver with pid {}", pid);
@@ -234,8 +287,8 @@ impl PageServerNode {
 
     fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> RequestBuilder {
         let mut builder = self.http_client.request(method, url);
-        if self.env.auth_type == AuthType::ZenithJWT {
-            builder = builder.bearer_auth(&self.env.auth_token)
+        if self.env.pageserver.auth_type == AuthType::ZenithJWT {
+            builder = builder.bearer_auth(&self.env.pageserver.auth_token)
         }
         builder
     }
@@ -247,7 +300,7 @@ impl PageServerNode {
         Ok(())
     }
 
-    pub fn tenant_list(&self) -> Result<Vec<String>> {
+    pub fn tenant_list(&self) -> Result<Vec<TenantInfo>> {
         Ok(self
             .http_request(Method::GET, format!("{}/{}", self.http_base_url, "tenant"))
             .send()?
@@ -310,11 +363,3 @@ impl PageServerNode {
             .json()?)
     }
 }
-
-impl Drop for PageServerNode {
-    fn drop(&mut self) {
-        if self.kill_on_exit {
-            let _ = self.stop();
-        }
-    }
-}

docker-entrypoint.sh

@@ -7,7 +7,7 @@ if [ "$1" = 'pageserver' ]; then
         pageserver --init -D /data --postgres-distrib /usr/local
     fi
     echo "Staring pageserver at 0.0.0.0:6400"
-    pageserver -l 0.0.0.0:6400 -D /data
+    pageserver -l 0.0.0.0:6400 --listen-http 0.0.0.0:9898 -D /data
 else
     "$@"
 fi

docs/docker.md

@@ -4,7 +4,7 @@
 
 Currently we build two main images:
 
-- [zenithdb/zenith](https://hub.docker.com/repository/docker/zenithdb/zenith) — image with pre-built `pageserver`, `wal_acceptor` and `proxy` binaries and all the required runtime dependencies. Built from [/Dockerfile](/Dockerfile).
+- [zenithdb/zenith](https://hub.docker.com/repository/docker/zenithdb/zenith) — image with pre-built `pageserver`, `safekeeper` and `proxy` binaries and all the required runtime dependencies. Built from [/Dockerfile](/Dockerfile).
 - [zenithdb/compute-node](https://hub.docker.com/repository/docker/zenithdb/compute-node) — compute node image with pre-built Postgres binaries from [zenithdb/postgres](https://github.com/zenithdb/postgres).
 
 And two intermediate images used either to reduce build time or to deliver some additional binary tools from other repos:

docs/glossary.md

@@ -51,11 +51,14 @@ Each PostgreSQL fork is considered a separate relish.
 
 ### Layer
 
-Each layer corresponds to the specific version of a relish Segment in a range of LSNs.
+A layer contains data needed to reconstruct any page versions within the
+layer's Segment and range of LSNs.
+
 There are two kinds of layers, in-memory and on-disk layers. In-memory
 layers are used to ingest incoming WAL, and provide fast access
 to the recent page versions. On-disk layers are stored as files on disk, and
-are immutable.
+are immutable. See pageserver/src/layered_repository/README.md for more.
+
 ### Layer file (on-disk layer)
 
 Layered repository on-disk format is based on immutable files.  The

docs/multitenancy.md

@@ -56,4 +56,4 @@ Tenant id is passed to postgres via GUC the same way as the timeline. Tenant id
 
 ### Safety
 
-For now particular tenant can only appear on a particular pageserver. Set of WAL acceptors are also pinned to particular (tenantid, timeline) pair so there can only be one writer for particular (tenantid, timeline).
+For now particular tenant can only appear on a particular pageserver. Set of safekeepers are also pinned to particular (tenantid, timeline) pair so there can only be one writer for particular (tenantid, timeline).

docs/settings.md

@@ -0,0 +1,128 @@
+## Pageserver
+
+### listen_pg_addr
+
+Network interface and port number to listen at for connections from
+the compute nodes and safekeepers. The default is `127.0.0.1:64000`.
+
+### listen_http_addr
+
+Network interface and port number to listen at for admin connections.
+The default is `127.0.0.1:9898`.
+
+### checkpoint_distance
+
+`checkpoint_distance` is the amount of incoming WAL that is held in
+the open layer, before it's flushed to local disk. It puts an upper
+bound on how much WAL needs to be re-processed after a pageserver
+crash. It is a soft limit, the pageserver can momentarily go above it,
+but it will trigger a checkpoint operation to get it back below the
+limit.
+
+`checkpoint_distance` also determines how much WAL needs to be kept
+durable in the safekeeper.  The safekeeper must have capacity to hold
+this much WAL, with some headroom, otherwise you can get stuck in a
+situation where the safekeeper is full and stops accepting new WAL,
+but the pageserver is not flushing out and releasing the space in the
+safekeeper because it hasn't reached checkpoint_distance yet.
+
+`checkpoint_distance` also controls how often the WAL is uploaded to
+S3.
+
+The unit is # of bytes.
+
+### checkpoint_period
+
+The pageserver checks whether `checkpoint_distance` has been reached
+every `checkpoint_period` seconds. Default is 1 s, which should be
+fine.
+
+### gc_horizon
+
+`gz_horizon` determines how much history is retained, to allow
+branching and read replicas at an older point in time. The unit is #
+of bytes of WAL. Page versions older than this are garbage collected
+away.
+
+### gc_period
+
+Interval at which garbage collection is triggered. Default is 100 s.
+
+### superuser
+
+Name of the initial superuser role, passed to initdb when a new tenant
+is initialized. It doesn't affect anything after initialization. The
+default is Note: The default is 'zenith_admin', and the console
+depends on that, so if you change it, bad things will happen.
+
+### page_cache_size
+
+Size of the page cache, to hold materialized page versions. Unit is
+number of 8 kB blocks. The default is 8192, which means 64 MB.
+
+### max_file_descriptors
+
+Max number of file descriptors to hold open concurrently for accessing
+layer files. This should be kept well below the process/container/OS
+limit (see `ulimit -n`), as the pageserver also needs file descriptors
+for other files and for sockets for incoming connections.
+
+### postgres-distrib
+
+A directory with Postgres installation to use during pageserver activities.
+Inside that dir, a `bin/postgres` binary should be present.
+
+The default distrib dir is `./tmp_install/`.
+
+### workdir (-D)
+
+A directory in the file system, where pageserver will store its files.
+The default is `./.zenith/`.
+
+### Remote storage
+
+There's a way to automatically backup and restore some of the pageserver's data from working dir to the remote storage.
+The backup system is disabled by default and can be enabled for either of the currently available storages:
+
+#### Local FS storage
+
+##### remote-storage-local-path
+
+Pageserver can back up and restore some of its workdir contents to another directory.
+For that, only a path to that directory needs to be specified as a parameter.
+
+#### S3 storage
+
+Pageserver can back up and restore some of its workdir contents to S3.
+Full set of S3 credentials is needed for that as parameters:
+
+##### remote-storage-s3-bucket
+
+Name of the bucket to connect to, example: "some-sample-bucket".
+
+##### remote-storage-region
+
+Name of the region where the bucket is located at, example: "eu-north-1"
+
+##### remote-storage-access-key
+
+Access key to connect to the bucket ("login" part of the credentials), example: "AKIAIOSFODNN7EXAMPLE"
+
+##### remote-storage-secret-access-key
+
+Secret access key to connect to the bucket ("password" part of the credentials), example: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY"
+
+#### General remote storage configuration
+
+Pagesever allows only one remote storage configured concurrently and errors if parameters from multiple different remote configurations are used.
+No default values are used for the remote storage configuration parameters.
+
+##### remote-storage-max-concurrent-sync
+
+Max number of concurrent connections to open for uploading to or
+downloading from S3.
+The default value is 100.
+
+## safekeeper
+
+TODO

docs/sourcetree.md

@@ -79,3 +79,61 @@ Helpers for exposing Prometheus metrics from the server.
 `/zenith_utils`:
 
 Helpers that are shared between other crates in this repository.
+
+## Using Python
+Note that Debian/Ubuntu Python packages are stale, as it commonly happens,
+so manual installation of dependencies is not recommended.
+
+A single virtual environment with all dependencies is described in the single `Pipfile`.
+
+### Prerequisites
+- Install Python 3.7 (the minimal supported version)
+    - Later version (e.g. 3.8) is ok if you don't write Python code
+    - You can install Python 3.7 separately, e.g.:
+      ```bash
+      # In Ubuntu
+      sudo add-apt-repository ppa:deadsnakes/ppa
+      sudo apt update
+      sudo apt install python3.7
+      ```
+- Install `pipenv`
+    - Exact version of `pipenv` is not important, you can use Debian/Ubuntu package `pipenv`.
+- Install dependencies via either
+  * `pipenv --python 3.7 install --dev` if you will write Python code, or
+  * `pipenv install` if you only want to run Python scripts and don't have Python 3.7.
+
+Run `pipenv shell` to activate the virtual environment.
+Alternatively, use `pipenv run` to run a single command in the venv, e.g. `pipenv run pytest`.
+
+### Obligatory checks
+We force code formatting via `yapf` and type hints via `mypy`.
+Run the following commands in the repository's root (next to `setup.cfg`):
+
+```bash
+pipenv run yapf -ri .  # All code is reformatted
+pipenv run mypy .  # Ensure there are no typing errors
+```
+
+**WARNING**: do not run `mypy` from a directory other than the root of the repository.
+Otherwise it will not find its configuration.
+
+Also consider:
+
+* Running `flake8` (or a linter of your choice, e.g. `pycodestyle`) and fixing possible defects, if any.
+* Adding more type hints to your code to avoid `Any`.
+
+### Changing dependencies
+You have to update `Pipfile.lock` if you have changed `Pipfile`:
+
+```bash
+pipenv --python 3.7 install --dev  # Re-create venv for Python 3.7 and install recent pipenv inside
+pipenv run pipenv --version  # Should be at least 2021.5.29
+pipenv run pipenv lock  # Regenerate Pipfile.lock
+```
+
+As the minimal supported version is Python 3.7 and we use it in CI,
+you have to use a Python 3.7 environment when updating `Pipfile.lock`.
+Otherwise some back-compatibility packages will be missing.
+
+It is also important to run recent `pipenv`.
+Older versions remove markers from `Pipfile.lock`.

pageserver/Cargo.toml

@@ -5,7 +5,7 @@ authors = ["Stas Kelvich <stas@zenith.tech>"]
 edition = "2018"
 
 [dependencies]
-bookfile = "^0.3"
+bookfile = { git = "https://github.com/zenithdb/bookfile.git", branch="generic-readext" }
 chrono = "0.4.19"
 rand = "0.8.3"
 regex = "1.4.5"
@@ -17,7 +17,7 @@ lazy_static = "1.4.0"
 log = "0.4.14"
 clap = "2.33.0"
 daemonize = "0.4.1"
-tokio = { version = "1.11", features = ["process", "macros", "fs"] }
+tokio = { version = "1.11", features = ["process", "sync", "macros", "fs", "rt", "io-util", "time"] }
 postgres-types = { git = "https://github.com/zenithdb/rust-postgres.git", rev="9eb0dbfbeb6a6c1b79099b9f7ae4a8c021877858" }
 postgres-protocol = { git = "https://github.com/zenithdb/rust-postgres.git", rev="9eb0dbfbeb6a6c1b79099b9f7ae4a8c021877858" }
 postgres = { git = "https://github.com/zenithdb/rust-postgres.git", rev="9eb0dbfbeb6a6c1b79099b9f7ae4a8c021877858" }
@@ -32,9 +32,15 @@ serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 toml = "0.5"
 scopeguard = "1.1.0"
-rust-s3 = { version = "0.27.0-rc4", features = ["no-verify-ssl"] }
 async-trait = "0.1"
 const_format = "0.2.21"
+tracing = "0.1.27"
+signal-hook = "0.3.10"
+url = "2"
+nix = "0.23"
+once_cell = "1.8.0"
+
+rust-s3 = { version = "0.27.0-rc4", features = ["no-verify-ssl"] }
 
 postgres_ffi = { path = "../postgres_ffi" }
 zenith_metrics = { path = "../zenith_metrics" }
@@ -43,3 +49,4 @@ workspace_hack = { path = "../workspace_hack" }
 
 [dev-dependencies]
 hex-literal = "0.3"
+tempfile = "3.2"

pageserver/README.md

@@ -41,7 +41,7 @@ Legend:
 +--+
 
 ....
-.  .   Component that we will need, but doesn't exist at the moment. A TODO.
+.  .   Component at its early development phase.
 ....
 
 --->   Data flow
@@ -116,13 +116,49 @@ Remove old on-disk layer files that are no longer needed according to the
 PITR retention policy
 
 
-TODO: Backup service
---------------------
+### Backup service
 
-The backup service is responsible for periodically pushing the chunks to S3.
+The backup service, responsible for storing pageserver recovery data externally.
 
-TODO: How/when do restore from S3? Whenever we get a GetPage@LSN request for
-a chunk we don't currently have? Or when an external Control Plane tells us?
+Currently, pageserver stores its files in a filesystem directory it's pointed to.
+That working directory could be rather ephemeral for such cases as "a pageserver pod running in k8s with no persistent volumes attached".
+Therefore, the server interacts with external, more reliable storage to back up and restore its state.
+
+The code for storage support is extensible and can support arbitrary ones as long as they implement a certain Rust trait.
+There are the following implementations present:
+* local filesystem — to use in tests mainly
+* AWS S3           - to use in production
+
+Implementation details are covered in the [backup readme](./src/remote_storage/README.md) and corresponding Rust file docs.
+
+The backup service is disabled by default and can be enabled to interact with a single remote storage.
+
+CLI examples:
+* Local FS: `${PAGESERVER_BIN} --remote-storage-local-path="/some/local/path/"`
+* AWS S3  : `${PAGESERVER_BIN} --remote-storage-s3-bucket="some-sample-bucket" --remote-storage-region="eu-north-1" --remote-storage-access-key="SOMEKEYAAAAASADSAH*#" --remote-storage-secret-access-key="SOMEsEcReTsd292v"`
+
+For Amazon AWS S3, a key id and secret access key could be located in `~/.aws/credentials` if awscli was ever configured to work with the desired bucket, on the AWS Settings page for a certain user. Also note, that the bucket names does not contain any protocols when used on AWS.
+For local S3 installations, refer to the their documentation for name format and credentials.
+
+Similar to other pageserver settings, toml config file can be used to configure either of the storages as backup targets.
+Required sections are:
+
+```toml
+[remote_storage]
+local_path = '/Users/someonetoignore/Downloads/tmp_dir/'
+```
+
+or
+
+```toml
+[remote_storage]
+bucket_name = 'some-sample-bucket'
+bucket_region = 'eu-north-1'
+access_key_id = 'SOMEKEYAAAAASADSAH*#'
+secret_access_key = 'SOMEsEcReTsd292v'
+```
+
+Also, `AWS_SECRET_ACCESS_KEY` and `AWS_ACCESS_KEY_ID` variables can be used to specify the credentials instead of any of the ways above.
 
 TODO: Sharding
 --------------------

pageserver/src/basebackup.rs

@@ -10,9 +10,10 @@
 //! This module is responsible for creation of such tarball
 //! from data stored in object storage.
 //!
-use anyhow::Result;
+use anyhow::{Context, Result};
 use bytes::{BufMut, BytesMut};
 use log::*;
+use std::fmt::Write as FmtWrite;
 use std::io;
 use std::io::Write;
 use std::sync::Arc;
@@ -31,7 +32,7 @@ use zenith_utils::lsn::Lsn;
 pub struct Basebackup<'a> {
     ar: Builder<&'a mut dyn Write>,
     timeline: &'a Arc<dyn Timeline>,
-    lsn: Lsn,
+    pub lsn: Lsn,
     prev_record_lsn: Lsn,
 }
 
@@ -83,7 +84,7 @@ impl<'a> Basebackup<'a> {
 
         info!(
             "taking basebackup lsn={}, prev_lsn={}",
-            backup_prev, backup_lsn
+            backup_lsn, backup_prev
         );
 
         Ok(Basebackup {
@@ -97,7 +98,6 @@ impl<'a> Basebackup<'a> {
     pub fn send_tarball(&mut self) -> anyhow::Result<()> {
         // Create pgdata subdirs structure
         for dir in pg_constants::PGDATA_SUBDIRS.iter() {
-            info!("send subdir {:?}", *dir);
             let header = new_tar_header_dir(*dir)?;
             self.ar.append(&header, &mut io::empty())?;
         }
@@ -242,20 +242,16 @@ impl<'a> Basebackup<'a> {
     fn add_pgcontrol_file(&mut self) -> anyhow::Result<()> {
         let checkpoint_bytes = self
             .timeline
-            .get_page_at_lsn(RelishTag::Checkpoint, 0, self.lsn)?;
-        let pg_control_bytes =
-            self.timeline
-                .get_page_at_lsn(RelishTag::ControlFile, 0, self.lsn)?;
+            .get_page_at_lsn(RelishTag::Checkpoint, 0, self.lsn)
+            .context("failed to get checkpoint bytes")?;
+        let pg_control_bytes = self
+            .timeline
+            .get_page_at_lsn(RelishTag::ControlFile, 0, self.lsn)
+            .context("failed get control bytes")?;
         let mut pg_control = ControlFileData::decode(&pg_control_bytes)?;
         let mut checkpoint = CheckPoint::decode(&checkpoint_bytes)?;
 
-        // Generate new pg_control and WAL needed for bootstrap
-        let checkpoint_segno = self.lsn.segment_number(pg_constants::WAL_SEGMENT_SIZE);
-        let checkpoint_lsn = XLogSegNoOffsetToRecPtr(
-            checkpoint_segno,
-            XLOG_SIZE_OF_XLOG_LONG_PHD as u32,
-            pg_constants::WAL_SEGMENT_SIZE,
-        );
+        // Generate new pg_control needed for bootstrap
         checkpoint.redo = normalize_lsn(self.lsn, pg_constants::WAL_SEGMENT_SIZE).0;
 
         //reset some fields we don't want to preserve
@@ -264,19 +260,24 @@ impl<'a> Basebackup<'a> {
         checkpoint.oldestActiveXid = 0;
 
         //save new values in pg_control
-        pg_control.checkPoint = checkpoint_lsn;
+        pg_control.checkPoint = 0;
         pg_control.checkPointCopy = checkpoint;
         pg_control.state = pg_constants::DB_SHUTDOWNED;
 
         // add zenith.signal file
-        let xl_prev = if self.prev_record_lsn == Lsn(0) {
-            0xBAD0 // magic value to indicate that we don't know prev_lsn
+        let mut zenith_signal = String::new();
+        if self.prev_record_lsn == Lsn(0) {
+            if self.lsn == self.timeline.get_ancestor_lsn() {
+                write!(zenith_signal, "PREV LSN: none")?;
+            } else {
+                write!(zenith_signal, "PREV LSN: invalid")?;
+            }
         } else {
-            self.prev_record_lsn.0
-        };
+            write!(zenith_signal, "PREV LSN: {}", self.prev_record_lsn)?;
+        }
         self.ar.append(
-            &new_tar_header("zenith.signal", 8)?,
-            &xl_prev.to_le_bytes()[..],
+            &new_tar_header("zenith.signal", zenith_signal.len() as u64)?,
+            zenith_signal.as_bytes(),
         )?;
 
         //send pg_control
@@ -285,14 +286,11 @@ impl<'a> Basebackup<'a> {
         self.ar.append(&header, &pg_control_bytes[..])?;
 
         //send wal segment
-        let wal_file_name = XLogFileName(
-            1, // FIXME: always use Postgres timeline 1
-            checkpoint_segno,
-            pg_constants::WAL_SEGMENT_SIZE,
-        );
+        let segno = self.lsn.segment_number(pg_constants::WAL_SEGMENT_SIZE);
+        let wal_file_name = XLogFileName(PG_TLI, segno, pg_constants::WAL_SEGMENT_SIZE);
         let wal_file_path = format!("pg_wal/{}", wal_file_name);
         let header = new_tar_header(&wal_file_path, pg_constants::WAL_SEGMENT_SIZE as u64)?;
-        let wal_seg = generate_wal_segment(&pg_control);
+        let wal_seg = generate_wal_segment(segno, pg_control.system_identifier);
         assert!(wal_seg.len() == pg_constants::WAL_SEGMENT_SIZE);
         self.ar.append(&header, &wal_seg[..])?;
         Ok(())

pageserver/src/bin/dump_layerfile.rs

@@ -4,11 +4,14 @@
 use anyhow::Result;
 use clap::{App, Arg};
 use pageserver::layered_repository::dump_layerfile_from_path;
+use pageserver::virtual_file;
 use std::path::PathBuf;
+use zenith_utils::GIT_VERSION;
 
 fn main() -> Result<()> {
     let arg_matches = App::new("Zenith dump_layerfile utility")
         .about("Dump contents of one layer file, for debugging")
+        .version(GIT_VERSION)
         .arg(
             Arg::with_name("path")
                 .help("Path to file to dump")
@@ -19,6 +22,9 @@ fn main() -> Result<()> {
 
     let path = PathBuf::from(arg_matches.value_of("path").unwrap());
 
+    // Basic initialization of things that don't change after startup
+    virtual_file::init(10);
+
     dump_layerfile_from_path(&path)?;
 
     Ok(())

pageserver/src/bin/pageserver.rs

@@ -2,37 +2,35 @@
 // Main entry point for the Page Server executable
 //
 
-use log::*;
-use pageserver::defaults::*;
 use serde::{Deserialize, Serialize};
 use std::{
     env,
-    net::TcpListener,
+    num::{NonZeroU32, NonZeroUsize},
     path::{Path, PathBuf},
     str::FromStr,
     thread,
 };
-use zenith_utils::{auth::JwtAuth, logging, postgres_backend::AuthType};
+use tracing::*;
+use zenith_utils::{auth::JwtAuth, logging, postgres_backend::AuthType, tcp_listener, GIT_VERSION};
 
 use anyhow::{bail, ensure, Context, Result};
+
 use clap::{App, Arg, ArgMatches};
 use daemonize::Daemonize;
 
 use pageserver::{
-    branches,
-    defaults::{
-        DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_PG_LISTEN_ADDR,
-        DEFAULT_RELISH_STORAGE_MAX_CONCURRENT_SYNC_LIMITS,
-    },
-    http, page_service, relish_storage, tenant_mgr, PageServerConf, RelishStorageConfig,
-    RelishStorageKind, S3Config, LOG_FILE_NAME,
+    branches, defaults::*, http, page_cache, page_service, remote_storage, tenant_mgr,
+    virtual_file, PageServerConf, RemoteStorageConfig, RemoteStorageKind, S3Config, LOG_FILE_NAME,
 };
 use zenith_utils::http::endpoint;
+use zenith_utils::postgres_backend;
+use zenith_utils::shutdown::exit_now;
+use zenith_utils::signals::{self, Signal};
 
 use const_format::formatcp;
 
 /// String arguments that can be declared via CLI or config file
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, PartialEq, Eq, Clone)]
 struct CfgFileParams {
     listen_pg_addr: Option<String>,
     listen_http_addr: Option<String>,
@@ -40,16 +38,29 @@ struct CfgFileParams {
     checkpoint_period: Option<String>,
     gc_horizon: Option<String>,
     gc_period: Option<String>,
+    open_mem_limit: Option<String>,
+    page_cache_size: Option<String>,
+    max_file_descriptors: Option<String>,
     pg_distrib_dir: Option<String>,
     auth_validation_public_key_path: Option<String>,
     auth_type: Option<String>,
-    // see https://github.com/alexcrichton/toml-rs/blob/6c162e6562c3e432bf04c82a3d1d789d80761a86/examples/enum_external.rs for enum deserialisation examples
-    relish_storage: Option<RelishStorage>,
-    relish_storage_max_concurrent_sync: Option<String>,
+    remote_storage_max_concurrent_sync: Option<String>,
+    remote_storage_max_sync_errors: Option<String>,
+    /////////////////////////////////
+    //// Don't put `Option<String>` and other "simple" values below.
+    ////
+    /// `Option<RemoteStorage>` is a <a href='https://toml.io/en/v1.0.0#table'>table</a> in TOML.
+    /// Values in TOML cannot be defined after tables (other tables can),
+    /// and [`toml`] crate serializes all fields in the order of their appearance.
+    ////////////////////////////////
+    remote_storage: Option<RemoteStorage>,
 }
 
-#[derive(Serialize, Deserialize, Clone)]
-enum RelishStorage {
+#[derive(Serialize, Deserialize, PartialEq, Eq, Clone)]
+// Without this attribute, enums with values won't be serialized by the `toml` library (but can be deserialized nonetheless!).
+// See https://github.com/alexcrichton/toml-rs/blob/6c162e6562c3e432bf04c82a3d1d789d80761a86/examples/enum_external.rs for the examples
+#[serde(untagged)]
+enum RemoteStorage {
     Local {
         local_path: String,
     },
@@ -70,33 +81,37 @@ impl CfgFileParams {
             arg_matches.value_of(arg_name).map(str::to_owned)
         };
 
-        let relish_storage = if let Some(local_path) = get_arg("relish-storage-local-path") {
-            Some(RelishStorage::Local { local_path })
+        let remote_storage = if let Some(local_path) = get_arg("remote-storage-local-path") {
+            Some(RemoteStorage::Local { local_path })
         } else if let Some((bucket_name, bucket_region)) =
-            get_arg("relish-storage-s3-bucket").zip(get_arg("relish-storage-region"))
+            get_arg("remote-storage-s3-bucket").zip(get_arg("remote-storage-region"))
         {
-            Some(RelishStorage::AwsS3 {
+            Some(RemoteStorage::AwsS3 {
                 bucket_name,
                 bucket_region,
-                access_key_id: get_arg("relish-storage-access-key"),
-                secret_access_key: get_arg("relish-storage-secret-access-key"),
+                access_key_id: get_arg("remote-storage-access-key"),
+                secret_access_key: get_arg("remote-storage-secret-access-key"),
             })
         } else {
             None
         };
 
         Self {
-            listen_pg_addr: get_arg("listen-pg"),
-            listen_http_addr: get_arg("listen-http"),
+            listen_pg_addr: get_arg("listen_pg_addr"),
+            listen_http_addr: get_arg("listen_http_addr"),
             checkpoint_distance: get_arg("checkpoint_distance"),
             checkpoint_period: get_arg("checkpoint_period"),
             gc_horizon: get_arg("gc_horizon"),
             gc_period: get_arg("gc_period"),
+            open_mem_limit: get_arg("open_mem_limit"),
+            page_cache_size: get_arg("page_cache_size"),
+            max_file_descriptors: get_arg("max_file_descriptors"),
             pg_distrib_dir: get_arg("postgres-distrib"),
             auth_validation_public_key_path: get_arg("auth-validation-public-key-path"),
             auth_type: get_arg("auth-type"),
-            relish_storage,
-            relish_storage_max_concurrent_sync: get_arg("relish-storage-max-concurrent-sync"),
+            remote_storage,
+            remote_storage_max_concurrent_sync: get_arg("remote-storage-max-concurrent-sync"),
+            remote_storage_max_sync_errors: get_arg("remote-storage-max-sync-errors"),
         }
     }
 
@@ -110,15 +125,21 @@ impl CfgFileParams {
             checkpoint_period: self.checkpoint_period.or(other.checkpoint_period),
             gc_horizon: self.gc_horizon.or(other.gc_horizon),
             gc_period: self.gc_period.or(other.gc_period),
+            open_mem_limit: self.open_mem_limit.or(other.open_mem_limit),
+            page_cache_size: self.page_cache_size.or(other.page_cache_size),
+            max_file_descriptors: self.max_file_descriptors.or(other.max_file_descriptors),
             pg_distrib_dir: self.pg_distrib_dir.or(other.pg_distrib_dir),
             auth_validation_public_key_path: self
                 .auth_validation_public_key_path
                 .or(other.auth_validation_public_key_path),
             auth_type: self.auth_type.or(other.auth_type),
-            relish_storage: self.relish_storage.or(other.relish_storage),
-            relish_storage_max_concurrent_sync: self
-                .relish_storage_max_concurrent_sync
-                .or(other.relish_storage_max_concurrent_sync),
+            remote_storage: self.remote_storage.or(other.remote_storage),
+            remote_storage_max_concurrent_sync: self
+                .remote_storage_max_concurrent_sync
+                .or(other.remote_storage_max_concurrent_sync),
+            remote_storage_max_sync_errors: self
+                .remote_storage_max_sync_errors
+                .or(other.remote_storage_max_sync_errors),
         }
     }
 
@@ -154,6 +175,21 @@ impl CfgFileParams {
             None => DEFAULT_GC_PERIOD,
         };
 
+        let open_mem_limit: usize = match self.open_mem_limit.as_ref() {
+            Some(open_mem_limit_str) => open_mem_limit_str.parse()?,
+            None => DEFAULT_OPEN_MEM_LIMIT,
+        };
+
+        let page_cache_size: usize = match self.page_cache_size.as_ref() {
+            Some(page_cache_size_str) => page_cache_size_str.parse()?,
+            None => DEFAULT_PAGE_CACHE_SIZE,
+        };
+
+        let max_file_descriptors: usize = match self.max_file_descriptors.as_ref() {
+            Some(max_file_descriptors_str) => max_file_descriptors_str.parse()?,
+            None => DEFAULT_MAX_FILE_DESCRIPTORS,
+        };
+
         let pg_distrib_dir = match self.pg_distrib_dir.as_ref() {
             Some(pg_distrib_dir_str) => PathBuf::from(pg_distrib_dir_str),
             None => env::current_dir()?.join("tmp_install"),
@@ -187,31 +223,34 @@ impl CfgFileParams {
             );
         }
 
-        let max_concurrent_sync = match self.relish_storage_max_concurrent_sync.as_deref() {
-            Some(relish_storage_max_concurrent_sync) => {
-                relish_storage_max_concurrent_sync.parse()?
-            }
-            None => DEFAULT_RELISH_STORAGE_MAX_CONCURRENT_SYNC_LIMITS,
+        let max_concurrent_sync = match self.remote_storage_max_concurrent_sync.as_deref() {
+            Some(number_str) => number_str.parse()?,
+            None => NonZeroUsize::new(DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNC).unwrap(),
         };
-        let relish_storage_config = self.relish_storage.as_ref().map(|storage_params| {
+        let max_sync_errors = match self.remote_storage_max_sync_errors.as_deref() {
+            Some(number_str) => number_str.parse()?,
+            None => NonZeroU32::new(DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS).unwrap(),
+        };
+        let remote_storage_config = self.remote_storage.as_ref().map(|storage_params| {
             let storage = match storage_params.clone() {
-                RelishStorage::Local { local_path } => {
-                    RelishStorageKind::LocalFs(PathBuf::from(local_path))
+                RemoteStorage::Local { local_path } => {
+                    RemoteStorageKind::LocalFs(PathBuf::from(local_path))
                 }
-                RelishStorage::AwsS3 {
+                RemoteStorage::AwsS3 {
                     bucket_name,
                     bucket_region,
                     access_key_id,
                     secret_access_key,
-                } => RelishStorageKind::AwsS3(S3Config {
+                } => RemoteStorageKind::AwsS3(S3Config {
                     bucket_name,
                     bucket_region,
                     access_key_id,
                     secret_access_key,
                 }),
             };
-            RelishStorageConfig {
+            RemoteStorageConfig {
                 max_concurrent_sync,
+                max_sync_errors,
                 storage,
             }
         });
@@ -225,6 +264,9 @@ impl CfgFileParams {
             checkpoint_period,
             gc_horizon,
             gc_period,
+            open_mem_limit,
+            page_cache_size,
+            max_file_descriptors,
 
             superuser: String::from(DEFAULT_SUPERUSER),
 
@@ -234,7 +276,7 @@ impl CfgFileParams {
 
             auth_validation_public_key_path,
             auth_type,
-            relish_storage_config,
+            remote_storage_config,
         })
     }
 }
@@ -243,18 +285,19 @@ fn main() -> Result<()> {
     zenith_metrics::set_common_metrics_prefix("pageserver");
     let arg_matches = App::new("Zenith page server")
         .about("Materializes WAL stream to pages and serves them to the postgres")
+        .version(GIT_VERSION)
         .arg(
-            Arg::with_name("listen-pg")
+            Arg::with_name("listen_pg_addr")
                 .short("l")
-                .long("listen-pg")
-                .alias("listen") // keep some compatibility
+                .long("listen_pg_addr")
+                .aliases(&["listen", "listen-pg"]) // keep some compatibility
                 .takes_value(true)
                 .help(formatcp!("listen for incoming page requests on ip:port (default: {DEFAULT_PG_LISTEN_ADDR})")),
         )
         .arg(
-            Arg::with_name("listen-http")
-                .long("listen-http")
-                .alias("http_endpoint") // keep some compatibility
+            Arg::with_name("listen_http_addr")
+                .long("listen_http_addr")
+                .aliases(&["http_endpoint", "listen-http"]) // keep some compatibility
                 .takes_value(true)
                 .help(formatcp!("http endpoint address for metrics and management API calls on ip:port (default: {DEFAULT_HTTP_LISTEN_ADDR})")),
         )
@@ -295,6 +338,25 @@ fn main() -> Result<()> {
                 .takes_value(true)
                 .help("Interval between garbage collector iterations"),
         )
+        .arg(
+            Arg::with_name("open_mem_limit")
+                .long("open_mem_limit")
+                .takes_value(true)
+                .help("Amount of memory reserved for buffering incoming WAL"),
+        )
+        .arg(
+
+            Arg::with_name("page_cache_size")
+                .long("page_cache_size")
+                .takes_value(true)
+                .help("Number of pages in the page cache"),
+        )
+        .arg(
+            Arg::with_name("max_file_descriptors")
+                .long("max_file_descriptors")
+                .takes_value(true)
+                .help("Max number of file descriptors to keep open for files"),
+        )
         .arg(
             Arg::with_name("workdir")
                 .short("D")
@@ -328,45 +390,45 @@ fn main() -> Result<()> {
                 .help("Authentication scheme type. One of: Trust, MD5, ZenithJWT"),
         )
         .arg(
-            Arg::with_name("relish-storage-local-path")
-                .long("relish-storage-local-path")
+            Arg::with_name("remote-storage-local-path")
+                .long("remote-storage-local-path")
                 .takes_value(true)
-                .help("Path to the local directory, to be used as an external relish storage")
+                .help("Path to the local directory, to be used as an external remote storage")
                 .conflicts_with_all(&[
-                    "relish-storage-s3-bucket",
-                    "relish-storage-region",
-                    "relish-storage-access-key",
-                    "relish-storage-secret-access-key",
+                    "remote-storage-s3-bucket",
+                    "remote-storage-region",
+                    "remote-storage-access-key",
+                    "remote-storage-secret-access-key",
                 ]),
         )
         .arg(
-            Arg::with_name("relish-storage-s3-bucket")
-                .long("relish-storage-s3-bucket")
+            Arg::with_name("remote-storage-s3-bucket")
+                .long("remote-storage-s3-bucket")
                 .takes_value(true)
-                .help("Name of the AWS S3 bucket to use an external relish storage")
-                .requires("relish-storage-region"),
+                .help("Name of the AWS S3 bucket to use an external remote storage")
+                .requires("remote-storage-region"),
         )
         .arg(
-            Arg::with_name("relish-storage-region")
-                .long("relish-storage-region")
+            Arg::with_name("remote-storage-region")
+                .long("remote-storage-region")
                 .takes_value(true)
                 .help("Region of the AWS S3 bucket"),
         )
         .arg(
-            Arg::with_name("relish-storage-access-key")
-                .long("relish-storage-access-key")
+            Arg::with_name("remote-storage-access-key")
+                .long("remote-storage-access-key")
                 .takes_value(true)
                 .help("Credentials to access the AWS S3 bucket"),
         )
         .arg(
-            Arg::with_name("relish-storage-secret-access-key")
-                .long("relish-storage-secret-access-key")
+            Arg::with_name("remote-storage-secret-access-key")
+                .long("remote-storage-secret-access-key")
                 .takes_value(true)
                 .help("Credentials to access the AWS S3 bucket"),
         )
         .arg(
-            Arg::with_name("relish-storage-max-concurrent-sync")
-                .long("relish-storage-max-concurrent-sync")
+            Arg::with_name("remote-storage-max-concurrent-sync")
+                .long("remote-storage-max-concurrent-sync")
                 .takes_value(true)
                 .help("Maximum allowed concurrent synchronisations with storage"),
         )
@@ -426,6 +488,11 @@ fn main() -> Result<()> {
     // as a ref.
     let conf: &'static PageServerConf = Box::leak(Box::new(conf));
 
+    // Basic initialization of things that don't change after startup
+    virtual_file::init(conf.max_file_descriptors);
+
+    page_cache::init(conf);
+
     // Create repo and exit if init was requested
     if init {
         branches::init_pageserver(conf, create_tenant).context("Failed to init pageserver")?;
@@ -447,7 +514,9 @@ fn main() -> Result<()> {
 
 fn start_pageserver(conf: &'static PageServerConf) -> Result<()> {
     // Initialize logger
-    let (_scope_guard, log_file) = logging::init(LOG_FILE_NAME, conf.daemonize)?;
+    let log_file = logging::init(LOG_FILE_NAME, conf.daemonize)?;
+
+    info!("version: {}", GIT_VERSION);
 
     // TODO: Check that it looks like a valid repository before going further
 
@@ -456,14 +525,15 @@ fn start_pageserver(conf: &'static PageServerConf) -> Result<()> {
         "Starting pageserver http handler on {}",
         conf.listen_http_addr
     );
-    let http_listener = TcpListener::bind(conf.listen_http_addr.clone())?;
+    let http_listener = tcp_listener::bind(conf.listen_http_addr.clone())?;
 
     info!(
         "Starting pageserver pg protocol handler on {}",
         conf.listen_pg_addr
     );
-    let pageserver_listener = TcpListener::bind(conf.listen_pg_addr.clone())?;
+    let pageserver_listener = tcp_listener::bind(conf.listen_pg_addr.clone())?;
 
+    // XXX: Don't spawn any threads before daemonizing!
     if conf.daemonize {
         info!("daemonizing...");
 
@@ -478,18 +548,21 @@ fn start_pageserver(conf: &'static PageServerConf) -> Result<()> {
             .stdout(stdout)
             .stderr(stderr);
 
-        match daemonize.start() {
+        // XXX: The parent process should exit abruptly right after
+        // it has spawned a child to prevent coverage machinery from
+        // dumping stats into a `profraw` file now owned by the child.
+        // Otherwise, the coverage data will be damaged.
+        match daemonize.exit_action(|| exit_now(0)).start() {
             Ok(_) => info!("Success, daemonized"),
-            Err(e) => error!("could not daemonize: {:#}", e),
+            Err(err) => error!(%err, "could not daemonize"),
         }
     }
 
-    // keep join handles for spawned threads
-    // don't spawn threads before daemonizing
-    let mut join_handles = Vec::new();
+    let signals = signals::install_shutdown_handlers()?;
+    let mut threads = vec![];
 
-    if let Some(handle) = relish_storage::run_storage_sync_thread(conf)? {
-        join_handles.push(handle);
+    if let Some(handle) = remote_storage::run_storage_sync_thread(conf)? {
+        threads.push(handle);
     }
     // Initialize tenant manager.
     tenant_mgr::init(conf);
@@ -508,30 +581,204 @@ fn start_pageserver(conf: &'static PageServerConf) -> Result<()> {
     // Spawn a new thread for the http endpoint
     // bind before launching separate thread so the error reported before startup exits
     let cloned = auth.clone();
-    let http_endpoint_thread = thread::Builder::new()
-        .name("http_endpoint_thread".into())
-        .spawn(move || {
-            let router = http::make_router(conf, cloned);
-            endpoint::serve_thread_main(router, http_listener)
-        })?;
-
-    join_handles.push(http_endpoint_thread);
+    threads.push(
+        thread::Builder::new()
+            .name("http_endpoint_thread".into())
+            .spawn(move || {
+                let router = http::make_router(conf, cloned);
+                endpoint::serve_thread_main(router, http_listener)
+            })?,
+    );
 
     // Spawn a thread to listen for connections. It will spawn further threads
     // for each connection.
-    let page_service_thread = thread::Builder::new()
-        .name("Page Service thread".into())
-        .spawn(move || {
-            page_service::thread_main(conf, auth, pageserver_listener, conf.auth_type)
-        })?;
+    threads.push(
+        thread::Builder::new()
+            .name("Page Service thread".into())
+            .spawn(move || {
+                page_service::thread_main(conf, auth, pageserver_listener, conf.auth_type)
+            })?,
+    );
 
-    join_handles.push(page_service_thread);
+    signals.handle(|signal| match signal {
+        Signal::Quit => {
+            info!(
+                "Got {}. Terminating in immediate shutdown mode",
+                signal.name()
+            );
+            std::process::exit(111);
+        }
 
-    for handle in join_handles.into_iter() {
-        handle
-            .join()
-            .expect("thread panicked")
-            .expect("thread exited with an error")
-    }
-    Ok(())
+        Signal::Interrupt | Signal::Terminate => {
+            info!(
+                "Got {}. Terminating gracefully in fast shutdown mode",
+                signal.name()
+            );
+
+            postgres_backend::set_pgbackend_shutdown_requested();
+            tenant_mgr::shutdown_all_tenants()?;
+            endpoint::shutdown();
+
+            for handle in std::mem::take(&mut threads) {
+                handle
+                    .join()
+                    .expect("thread panicked")
+                    .expect("thread exited with an error");
+            }
+
+            info!("Shut down successfully completed");
+            std::process::exit(0);
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn page_server_conf_toml_serde() {
+        let params = CfgFileParams {
+            listen_pg_addr: Some("listen_pg_addr_VALUE".to_string()),
+            listen_http_addr: Some("listen_http_addr_VALUE".to_string()),
+            checkpoint_distance: Some("checkpoint_distance_VALUE".to_string()),
+            checkpoint_period: Some("checkpoint_period_VALUE".to_string()),
+            gc_horizon: Some("gc_horizon_VALUE".to_string()),
+            gc_period: Some("gc_period_VALUE".to_string()),
+            open_mem_limit: Some("open_mem_limit_VALUE".to_string()),
+            page_cache_size: Some("page_cache_size_VALUE".to_string()),
+            max_file_descriptors: Some("max_file_descriptors_VALUE".to_string()),
+            pg_distrib_dir: Some("pg_distrib_dir_VALUE".to_string()),
+            auth_validation_public_key_path: Some(
+                "auth_validation_public_key_path_VALUE".to_string(),
+            ),
+            auth_type: Some("auth_type_VALUE".to_string()),
+            remote_storage: Some(RemoteStorage::Local {
+                local_path: "remote_storage_local_VALUE".to_string(),
+            }),
+            remote_storage_max_concurrent_sync: Some(
+                "remote_storage_max_concurrent_sync_VALUE".to_string(),
+            ),
+            remote_storage_max_sync_errors: Some(
+                "remote_storage_max_sync_errors_VALUE".to_string(),
+            ),
+        };
+
+        let toml_string = toml::to_string(&params).expect("Failed to serialize correct config");
+        let toml_pretty_string =
+            toml::to_string_pretty(&params).expect("Failed to serialize correct config");
+        assert_eq!(
+            r#"listen_pg_addr = 'listen_pg_addr_VALUE'
+listen_http_addr = 'listen_http_addr_VALUE'
+checkpoint_distance = 'checkpoint_distance_VALUE'
+checkpoint_period = 'checkpoint_period_VALUE'
+gc_horizon = 'gc_horizon_VALUE'
+gc_period = 'gc_period_VALUE'
+open_mem_limit = 'open_mem_limit_VALUE'
+page_cache_size = 'page_cache_size_VALUE'
+max_file_descriptors = 'max_file_descriptors_VALUE'
+pg_distrib_dir = 'pg_distrib_dir_VALUE'
+auth_validation_public_key_path = 'auth_validation_public_key_path_VALUE'
+auth_type = 'auth_type_VALUE'
+remote_storage_max_concurrent_sync = 'remote_storage_max_concurrent_sync_VALUE'
+remote_storage_max_sync_errors = 'remote_storage_max_sync_errors_VALUE'
+
+[remote_storage]
+local_path = 'remote_storage_local_VALUE'
+"#,
+            toml_pretty_string
+        );
+
+        let params_from_serialized: CfgFileParams = toml::from_str(&toml_string)
+            .expect("Failed to deserialize the serialization result of the config");
+        let params_from_serialized_pretty: CfgFileParams = toml::from_str(&toml_pretty_string)
+            .expect("Failed to deserialize the prettified serialization result of the config");
+        assert!(
+            params_from_serialized == params,
+            "Expected the same config in the end of config -> serialize -> deserialize chain"
+        );
+        assert!(
+            params_from_serialized_pretty == params,
+            "Expected the same config in the end of config -> serialize pretty -> deserialize chain"
+        );
+    }
+
+    #[test]
+    fn credentials_omitted_during_serialization() {
+        let params = CfgFileParams {
+            listen_pg_addr: Some("listen_pg_addr_VALUE".to_string()),
+            listen_http_addr: Some("listen_http_addr_VALUE".to_string()),
+            checkpoint_distance: Some("checkpoint_distance_VALUE".to_string()),
+            checkpoint_period: Some("checkpoint_period_VALUE".to_string()),
+            gc_horizon: Some("gc_horizon_VALUE".to_string()),
+            gc_period: Some("gc_period_VALUE".to_string()),
+            open_mem_limit: Some("open_mem_limit_VALUE".to_string()),
+            page_cache_size: Some("page_cache_size_VALUE".to_string()),
+            max_file_descriptors: Some("max_file_descriptors_VALUE".to_string()),
+            pg_distrib_dir: Some("pg_distrib_dir_VALUE".to_string()),
+            auth_validation_public_key_path: Some(
+                "auth_validation_public_key_path_VALUE".to_string(),
+            ),
+            auth_type: Some("auth_type_VALUE".to_string()),
+            remote_storage: Some(RemoteStorage::AwsS3 {
+                bucket_name: "bucket_name_VALUE".to_string(),
+                bucket_region: "bucket_region_VALUE".to_string(),
+                access_key_id: Some("access_key_id_VALUE".to_string()),
+                secret_access_key: Some("secret_access_key_VALUE".to_string()),
+            }),
+            remote_storage_max_concurrent_sync: Some(
+                "remote_storage_max_concurrent_sync_VALUE".to_string(),
+            ),
+            remote_storage_max_sync_errors: Some(
+                "remote_storage_max_sync_errors_VALUE".to_string(),
+            ),
+        };
+
+        let toml_string = toml::to_string(&params).expect("Failed to serialize correct config");
+        let toml_pretty_string =
+            toml::to_string_pretty(&params).expect("Failed to serialize correct config");
+        assert_eq!(
+            r#"listen_pg_addr = 'listen_pg_addr_VALUE'
+listen_http_addr = 'listen_http_addr_VALUE'
+checkpoint_distance = 'checkpoint_distance_VALUE'
+checkpoint_period = 'checkpoint_period_VALUE'
+gc_horizon = 'gc_horizon_VALUE'
+gc_period = 'gc_period_VALUE'
+open_mem_limit = 'open_mem_limit_VALUE'
+page_cache_size = 'page_cache_size_VALUE'
+max_file_descriptors = 'max_file_descriptors_VALUE'
+pg_distrib_dir = 'pg_distrib_dir_VALUE'
+auth_validation_public_key_path = 'auth_validation_public_key_path_VALUE'
+auth_type = 'auth_type_VALUE'
+remote_storage_max_concurrent_sync = 'remote_storage_max_concurrent_sync_VALUE'
+remote_storage_max_sync_errors = 'remote_storage_max_sync_errors_VALUE'
+
+[remote_storage]
+bucket_name = 'bucket_name_VALUE'
+bucket_region = 'bucket_region_VALUE'
+"#,
+            toml_pretty_string
+        );
+
+        let params_from_serialized: CfgFileParams = toml::from_str(&toml_string)
+            .expect("Failed to deserialize the serialization result of the config");
+        let params_from_serialized_pretty: CfgFileParams = toml::from_str(&toml_pretty_string)
+            .expect("Failed to deserialize the prettified serialization result of the config");
+
+        let mut expected_params = params;
+        expected_params.remote_storage = Some(RemoteStorage::AwsS3 {
+            bucket_name: "bucket_name_VALUE".to_string(),
+            bucket_region: "bucket_region_VALUE".to_string(),
+            access_key_id: None,
+            secret_access_key: None,
+        });
+        assert!(
+            params_from_serialized == expected_params,
+            "Expected the config without credentials in the end of a 'config -> serialize -> deserialize' chain"
+        );
+        assert!(
+            params_from_serialized_pretty == expected_params,
+            "Expected the config without credentials in the end of a 'config -> serialize pretty -> deserialize' chain"
+        );
+    }
 }

pageserver/src/branches.rs

@@ -4,7 +4,7 @@
 // TODO: move all paths construction to conf impl
 //
 
-use anyhow::{bail, ensure, Context, Result};
+use anyhow::{bail, Context, Result};
 use postgres_ffi::ControlFileData;
 use serde::{Deserialize, Serialize};
 use std::{
@@ -14,15 +14,16 @@ use std::{
     str::FromStr,
     sync::Arc,
 };
-use zenith_utils::zid::{ZTenantId, ZTimelineId};
+use tracing::*;
 
-use log::*;
 use zenith_utils::crashsafe_dir;
 use zenith_utils::logging;
 use zenith_utils::lsn::Lsn;
+use zenith_utils::zid::{ZTenantId, ZTimelineId};
 
 use crate::tenant_mgr;
 use crate::walredo::WalRedoManager;
+use crate::CheckpointConfig;
 use crate::{repository::Repository, PageServerConf};
 use crate::{restore_local_repo, LOG_FILE_NAME};
 
@@ -35,15 +36,14 @@ pub struct BranchInfo {
     pub ancestor_id: Option<String>,
     pub ancestor_lsn: Option<String>,
     pub current_logical_size: usize,
-    pub current_logical_size_non_incremental: usize,
+    pub current_logical_size_non_incremental: Option<usize>,
 }
 
 impl BranchInfo {
     pub fn from_path<T: AsRef<Path>>(
         path: T,
-        conf: &PageServerConf,
-        tenantid: &ZTenantId,
         repo: &Arc<dyn Repository>,
+        include_non_incremental_logical_size: bool,
     ) -> Result<Self> {
         let name = path
             .as_ref()
@@ -56,27 +56,22 @@ impl BranchInfo {
 
         let timeline = repo.get_timeline(timeline_id)?;
 
-        let ancestor_path = conf.ancestor_path(&timeline_id, tenantid);
-        let mut ancestor_id: Option<String> = None;
-        let mut ancestor_lsn: Option<String> = None;
+        // we use ancestor lsn zero if we don't have an ancestor, so turn this into an option based on timeline id
+        let (ancestor_id, ancestor_lsn) = match timeline.get_ancestor_timeline_id() {
+            Some(ancestor_id) => (
+                Some(ancestor_id.to_string()),
+                Some(timeline.get_ancestor_lsn().to_string()),
+            ),
+            None => (None, None),
+        };
 
-        if ancestor_path.exists() {
-            let ancestor = std::fs::read_to_string(ancestor_path)?;
-            let mut strings = ancestor.split('@');
-
-            ancestor_id = Some(
-                strings
-                    .next()
-                    .with_context(|| "wrong branch ancestor point in time format")?
-                    .to_owned(),
-            );
-            ancestor_lsn = Some(
-                strings
-                    .next()
-                    .with_context(|| "wrong branch ancestor point in time format")?
-                    .to_owned(),
-            );
-        }
+        // non incremental size calculation can be heavy, so let it be optional
+        // needed for tests to check size calculation
+        let current_logical_size_non_incremental = include_non_incremental_logical_size
+            .then(|| {
+                timeline.get_current_logical_size_non_incremental(timeline.get_last_record_lsn())
+            })
+            .transpose()?;
 
         Ok(BranchInfo {
             name,
@@ -85,8 +80,7 @@ impl BranchInfo {
             ancestor_id,
             ancestor_lsn,
             current_logical_size: timeline.get_current_logical_size(),
-            current_logical_size_non_incremental: timeline
-                .get_current_logical_size_non_incremental(timeline.get_last_record_lsn())?,
+            current_logical_size_non_incremental,
         })
     }
 }
@@ -100,7 +94,7 @@ pub struct PointInTime {
 pub fn init_pageserver(conf: &'static PageServerConf, create_tenant: Option<&str>) -> Result<()> {
     // Initialize logger
     // use true as daemonize parameter because otherwise we pollute zenith cli output with a few pages long output of info messages
-    let (_scope_guard, _log_file) = logging::init(LOG_FILE_NAME, true)?;
+    let _log_file = logging::init(LOG_FILE_NAME, true)?;
 
     // We don't use the real WAL redo manager, because we don't want to spawn the WAL redo
     // process during repository initialization.
@@ -145,7 +139,11 @@ pub fn create_repo(
 
     info!("created directory structure in {}", repo_dir.display());
 
-    let tli = create_timeline(conf, None, &tenantid)?;
+    // create a new timeline directory
+    let timeline_id = ZTimelineId::generate();
+    let timelinedir = conf.timeline_path(&timeline_id, &tenantid);
+
+    crashsafe_dir::create_dir(&timelinedir)?;
 
     let repo = Arc::new(crate::layered_repository::LayeredRepository::new(
         conf,
@@ -157,7 +155,7 @@ pub fn create_repo(
     // Load data into pageserver
     // TODO To implement zenith import we need to
     //      move data loading out of create_repo()
-    bootstrap_timeline(conf, tenantid, tli, repo.as_ref())?;
+    bootstrap_timeline(conf, tenantid, timeline_id, repo.as_ref())?;
 
     Ok(repo)
 }
@@ -176,13 +174,16 @@ fn get_lsn_from_controlfile(path: &Path) -> Result<Lsn> {
 // to get bootstrap data for timeline initialization.
 //
 fn run_initdb(conf: &'static PageServerConf, initdbpath: &Path) -> Result<()> {
-    info!("running initdb... ");
+    info!("running initdb in {}... ", initdbpath.display());
 
     let initdb_path = conf.pg_bin_dir().join("initdb");
     let initdb_output = Command::new(initdb_path)
         .args(&["-D", initdbpath.to_str().unwrap()])
         .args(&["-U", &conf.superuser])
         .arg("--no-instructions")
+        // This is only used for a temporary installation that is deleted shortly after,
+        // so no need to fsync it
+        .arg("--no-sync")
         .env_clear()
         .env("LD_LIBRARY_PATH", conf.pg_lib_dir().to_str().unwrap())
         .env("DYLD_LIBRARY_PATH", conf.pg_lib_dir().to_str().unwrap())
@@ -195,7 +196,6 @@ fn run_initdb(conf: &'static PageServerConf, initdbpath: &Path) -> Result<()> {
             String::from_utf8_lossy(&initdb_output.stderr)
         );
     }
-    info!("initdb succeeded");
 
     Ok(())
 }
@@ -210,6 +210,8 @@ fn bootstrap_timeline(
     tli: ZTimelineId,
     repo: &dyn Repository,
 ) -> Result<()> {
+    let _enter = info_span!("bootstrapping", timeline = %tli, tenant = %tenantid).entered();
+
     let initdb_path = conf.tenant_path(&tenantid).join("tmp");
 
     // Init temporarily repo to get bootstrap data
@@ -218,17 +220,17 @@ fn bootstrap_timeline(
 
     let lsn = get_lsn_from_controlfile(&pgdata_path)?.align();
 
-    info!("bootstrap_timeline {:?} at lsn {}", pgdata_path, lsn);
-
     // Import the contents of the data directory at the initial checkpoint
     // LSN, and any WAL after that.
-    let timeline = repo.create_empty_timeline(tli)?;
+    // Initdb lsn will be equal to last_record_lsn which will be set after import.
+    // Because we know it upfront avoid having an option or dummy zero value by passing it to create_empty_timeline.
+    let timeline = repo.create_empty_timeline(tli, lsn)?;
     restore_local_repo::import_timeline_from_postgres_datadir(
         &pgdata_path,
-        timeline.as_ref(),
+        timeline.writer().as_ref(),
         lsn,
     )?;
-    timeline.checkpoint()?;
+    timeline.checkpoint(CheckpointConfig::Forced)?;
 
     println!(
         "created initial timeline {} timeline.lsn {}",
@@ -246,29 +248,38 @@ fn bootstrap_timeline(
     Ok(())
 }
 
-pub(crate) fn get_tenants(conf: &PageServerConf) -> Result<Vec<String>> {
-    let tenants_dir = conf.tenants_path();
-
-    std::fs::read_dir(&tenants_dir)?
-        .map(|dir_entry_res| {
-            let dir_entry = dir_entry_res?;
-            ensure!(dir_entry.file_type()?.is_dir());
-            Ok(dir_entry.file_name().to_str().unwrap().to_owned())
-        })
-        .collect()
-}
-
-pub(crate) fn get_branches(conf: &PageServerConf, tenantid: &ZTenantId) -> Result<Vec<BranchInfo>> {
+pub(crate) fn get_branches(
+    conf: &PageServerConf,
+    tenantid: &ZTenantId,
+    include_non_incremental_logical_size: bool,
+) -> Result<Vec<BranchInfo>> {
     let repo = tenant_mgr::get_repository_for_tenant(*tenantid)?;
 
     // Each branch has a corresponding record (text file) in the refs/branches
     // with timeline_id.
     let branches_dir = conf.branches_path(tenantid);
 
-    std::fs::read_dir(&branches_dir)?
+    std::fs::read_dir(&branches_dir)
+        .with_context(|| {
+            format!(
+                "Found no branches directory '{}' for tenant {}",
+                branches_dir.display(),
+                tenantid
+            )
+        })?
         .map(|dir_entry_res| {
-            let dir_entry = dir_entry_res?;
-            BranchInfo::from_path(dir_entry.path(), conf, tenantid, &repo)
+            let dir_entry = dir_entry_res.with_context(|| {
+                format!(
+                    "Failed to list branches directory '{}' content for tenant {}",
+                    branches_dir.display(),
+                    tenantid
+                )
+            })?;
+            BranchInfo::from_path(
+                dir_entry.path(),
+                &repo,
+                include_non_incremental_logical_size,
+            )
         })
         .collect()
 }
@@ -311,26 +322,26 @@ pub(crate) fn create_branch(
         );
     }
 
-    // create a new timeline directory for it
-    let newtli = create_timeline(conf, Some(startpoint), tenantid)?;
+    let new_timeline_id = ZTimelineId::generate();
 
-    // Let the Repository backend do its initialization
-    repo.branch_timeline(startpoint.timelineid, newtli, startpoint.lsn)?;
+    // Forward entire timeline creation routine to repository
+    // backend, so it can do all needed initialization
+    repo.branch_timeline(startpoint.timelineid, new_timeline_id, startpoint.lsn)?;
 
     // Remember the human-readable branch name for the new timeline.
     // FIXME: there's a race condition, if you create a branch with the same
     // name concurrently.
-    let data = newtli.to_string();
+    let data = new_timeline_id.to_string();
     fs::write(conf.branch_path(branchname, tenantid), data)?;
 
     Ok(BranchInfo {
         name: branchname.to_string(),
-        timeline_id: newtli,
+        timeline_id: new_timeline_id,
         latest_valid_lsn: startpoint.lsn,
-        ancestor_id: None,
-        ancestor_lsn: None,
+        ancestor_id: Some(startpoint.timelineid.to_string()),
+        ancestor_lsn: Some(startpoint.lsn.to_string()),
         current_logical_size: 0,
-        current_logical_size_non_incremental: 0,
+        current_logical_size_non_incremental: Some(0),
     })
 }
 
@@ -406,24 +417,3 @@ fn parse_point_in_time(
 
     bail!("could not parse point-in-time {}", s);
 }
-
-fn create_timeline(
-    conf: &PageServerConf,
-    ancestor: Option<PointInTime>,
-    tenantid: &ZTenantId,
-) -> Result<ZTimelineId> {
-    // Create initial timeline
-
-    let timelineid = ZTimelineId::generate();
-
-    let timelinedir = conf.timeline_path(&timelineid, tenantid);
-
-    fs::create_dir(&timelinedir)?;
-
-    if let Some(ancestor) = ancestor {
-        let data = format!("{}@{}", ancestor.timelineid, ancestor.lsn);
-        fs::write(timelinedir.join("ancestor"), data)?;
-    }
-
-    Ok(timelineid)
-}

pageserver/src/http/openapi_spec.yml

@@ -25,6 +25,11 @@ paths:
         schema:
           type: string
           format: hex
+      - name: include-non-incremental-logical-size
+        in: query
+        schema:
+          type: string
+          description: Controls calculation of current_logical_size_non_incremental
     get:
       description: Get branches for tenant
       responses:
@@ -73,6 +78,11 @@ paths:
         required: true
         schema:
           type: string
+      - name: include-non-incremental-logical-size
+        in: query
+        schema:
+          type: string
+          description: Controls calculation of current_logical_size_non_incremental
     get:
       description: Get branches for tenant
       responses:
@@ -164,13 +174,13 @@ paths:
       description: Get tenants list
       responses:
         "200":
-          description: OK
+          description: TenantInfo
           content:
             application/json:
               schema:
                 type: array
                 items:
-                  type: string
+                  $ref: "#/components/schemas/TenantInfo"
         "401":
           description: Unauthorized Error
           content:
@@ -243,6 +253,16 @@ components:
       scheme: bearer
       bearerFormat: JWT
   schemas:
+    TenantInfo:
+      type: object
+      required:
+        - id
+        - state
+      properties:
+        id:
+          type: string
+        state:
+          type: string
     BranchInfo:
       type: object
       required:
@@ -250,7 +270,6 @@ components:
         - timeline_id
         - latest_valid_lsn
         - current_logical_size
-        - current_logical_size_non_incremental
       properties:
         name:
           type: string

pageserver/src/http/routes.rs

@@ -1,4 +1,3 @@
-use std::str::FromStr;
 use std::sync::Arc;
 
 use anyhow::Result;
@@ -6,6 +5,7 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use routerify::{ext::RequestExt, RouterBuilder};
+use tracing::*;
 use zenith_utils::auth::JwtAuth;
 use zenith_utils::http::endpoint::attach_openapi_ui;
 use zenith_utils::http::endpoint::auth_middleware;
@@ -15,6 +15,8 @@ use zenith_utils::http::{
     endpoint,
     error::HttpErrorBody,
     json::{json_request, json_response},
+    request::get_request_param,
+    request::parse_request_param,
 };
 
 use super::models::BranchCreateRequest;
@@ -56,33 +58,6 @@ fn get_config(request: &Request<Body>) -> &'static PageServerConf {
     get_state(request).conf
 }
 
-fn get_request_param<'a>(
-    request: &'a Request<Body>,
-    param_name: &str,
-) -> Result<&'a str, ApiError> {
-    match request.param(param_name) {
-        Some(arg) => Ok(arg),
-        None => {
-            return Err(ApiError::BadRequest(format!(
-                "no {} specified in path param",
-                param_name
-            )))
-        }
-    }
-}
-
-fn parse_request_param<T: FromStr>(
-    request: &Request<Body>,
-    param_name: &str,
-) -> Result<T, ApiError> {
-    match get_request_param(request, param_name)?.parse() {
-        Ok(v) => Ok(v),
-        Err(_) => Err(ApiError::BadRequest(
-            "failed to parse tenant id".to_string(),
-        )),
-    }
-}
-
 // healthcheck handler
 async fn status_handler(_: Request<Body>) -> Result<Response<Body>, ApiError> {
     Ok(Response::builder()
@@ -98,6 +73,7 @@ async fn branch_create_handler(mut request: Request<Body>) -> Result<Response<Bo
     check_permission(&request, Some(request_data.tenant_id))?;
 
     let response_data = tokio::task::spawn_blocking(move || {
+        let _enter = info_span!("/branch_create", name = %request_data.name, tenant = %request_data.tenant_id, startpoint=%request_data.start_point).entered();
         branches::create_branch(
             get_config(&request),
             &request_data.name,
@@ -110,29 +86,53 @@ async fn branch_create_handler(mut request: Request<Body>) -> Result<Response<Bo
     Ok(json_response(StatusCode::CREATED, response_data)?)
 }
 
+// Gate non incremental logical size calculation behind a flag
+// after pgbench -i -s100 calculation took 28ms so if multiplied by the number of timelines
+// and tenants it can take noticeable amount of time. Also the value currently used only in tests
+fn get_include_non_incremental_logical_size(request: &Request<Body>) -> bool {
+    request
+        .uri()
+        .query()
+        .map(|v| {
+            url::form_urlencoded::parse(v.as_bytes())
+                .into_owned()
+                .any(|(param, _)| param == "include-non-incremental-logical-size")
+        })
+        .unwrap_or(false)
+}
+
 async fn branch_list_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
     let tenantid: ZTenantId = parse_request_param(&request, "tenant_id")?;
 
+    let include_non_incremental_logical_size = get_include_non_incremental_logical_size(&request);
+
     check_permission(&request, Some(tenantid))?;
 
     let response_data = tokio::task::spawn_blocking(move || {
-        crate::branches::get_branches(get_config(&request), &tenantid)
+        let _enter = info_span!("branch_list", tenant = %tenantid).entered();
+        crate::branches::get_branches(
+            get_config(&request),
+            &tenantid,
+            include_non_incremental_logical_size,
+        )
     })
     .await
     .map_err(ApiError::from_err)??;
     Ok(json_response(StatusCode::OK, response_data)?)
 }
 
-// TODO add to swagger
 async fn branch_detail_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
     let tenantid: ZTenantId = parse_request_param(&request, "tenant_id")?;
-    let branch_name: &str = get_request_param(&request, "branch_name")?;
+    let branch_name: String = get_request_param(&request, "branch_name")?.to_string();
     let conf = get_state(&request).conf;
-    let path = conf.branch_path(branch_name, &tenantid);
+    let path = conf.branch_path(&branch_name, &tenantid);
+
+    let include_non_incremental_logical_size = get_include_non_incremental_logical_size(&request);
 
     let response_data = tokio::task::spawn_blocking(move || {
+        let _enter = info_span!("branch_detail", tenant = %tenantid, branch=%branch_name).entered();
         let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
-        BranchInfo::from_path(path, conf, &tenantid, &repo)
+        BranchInfo::from_path(path, &repo, include_non_incremental_logical_size)
     })
     .await
     .map_err(ApiError::from_err)??;
@@ -144,10 +144,13 @@ async fn tenant_list_handler(request: Request<Body>) -> Result<Response<Body>, A
     // check for management permission
     check_permission(&request, None)?;
 
-    let response_data =
-        tokio::task::spawn_blocking(move || crate::branches::get_tenants(get_config(&request)))
-            .await
-            .map_err(ApiError::from_err)??;
+    let response_data = tokio::task::spawn_blocking(move || {
+        let _enter = info_span!("tenant_list").entered();
+        crate::tenant_mgr::list_tenants()
+    })
+    .await
+    .map_err(ApiError::from_err)??;
+
     Ok(json_response(StatusCode::OK, response_data)?)
 }
 
@@ -158,6 +161,7 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
     let request_data: TenantCreateRequest = json_request(&mut request).await?;
 
     let response_data = tokio::task::spawn_blocking(move || {
+        let _enter = info_span!("tenant_create", tenant = %request_data.tenant_id).entered();
         tenant_mgr::create_repository_for_tenant(get_config(&request), request_data.tenant_id)
     })
     .await

pageserver/src/layered_repository/README.md

@@ -1,12 +1,56 @@
 # Overview
 
-The on-disk format is based on immutable files. The page server
-receives a stream of incoming WAL, parses the WAL records to determine
-which pages they apply to, and accumulates the incoming changes in
-memory. Every now and then, the accumulated changes are written out to
-new immutable files. This process is called checkpointing. Old versions
-of on-disk files that are not needed by any timeline are removed by GC
-process.
+The on-disk format is based on immutable files. The page server receives a
+stream of incoming WAL, parses the WAL records to determine which pages they
+apply to, and accumulates the incoming changes in memory. Every now and then,
+the accumulated changes are written out to new immutable files. This process is
+called checkpointing. Old versions of on-disk files that are not needed by any
+timeline are removed by GC process.
+
+The main responsibility of the Page Server is to process the incoming WAL, and
+reprocess it into a format that allows reasonably quick access to any page
+version.
+
+The incoming WAL contains updates to arbitrary pages in the system. The
+distribution depends on the workload: the updates could be totally random, or
+there could be a long stream of updates to a single relation when data is bulk
+loaded, for example, or something in between. The page server slices the
+incoming WAL per relation and page, and packages the sliced WAL into
+suitably-sized "layer files". The layer files contain all the history of the
+database, back to some reasonable retention period. This system replaces the
+base backups and the WAL archive used in a traditional PostgreSQL
+installation. The layer files are immutable, they are not modified in-place
+after creation. New layer files are created for new incoming WAL, and old layer
+files are removed when they are no longer needed. We could also replace layer
+files with new files that contain the same information, merging small files for
+example, but that hasn't been implemented yet.
+
+
+Cloud Storage                   Page Server                   Safekeeper
+                     Local disk                Memory            WAL
+
+|AAAA|               |AAAA|AAAA|               |AA
+|BBBB|               |BBBB|BBBB|               |
+|CCCC|CCCC|  <----   |CCCC|CCCC|CCCC|   <---   |CC     <----   ADEBAABED
+|DDDD|DDDD|          |DDDD|DDDD|               |DDD
+|EEEE|               |EEEE|EEEE|EEEE|          |E
+
+
+In this illustration, WAL is received as a stream from the Safekeeper, from the
+right.  It is immediately captured by the page server and stored quickly in
+memory. The page server memory can be thought of as a quick "reorder buffer",
+used to hold the incoming WAL and reorder it so that we keep the WAL records for
+the same page and relation close to each other.
+
+From the page server memory, whenever enough WAL has been accumulated for one
+relation segment, it is moved to local disk, as a new layer file, and the memory
+is released.
+
+From the local disk, the layers are further copied to Cloud Storage, for
+long-term archival. After a layer has been copied to Cloud Storage, it can be
+removed from local disk, although we currently keep everything locally for fast
+access. If a layer is needed that isn't found locally, it is fetched from Cloud
+Storage and stored in local disk.
 
 # Terms used in layered repository
 
@@ -14,32 +58,9 @@ process.
 - Segment - one slice of a Relish that is stored in a LayeredTimeline.
 - Layer -  specific version of a relish Segment in a range of LSNs.
 
-Layers can be InMemory or OnDisk:
-- InMemory layer is not durably stored and needs to rebuild from WAL on pageserver start.
-- OnDisk layer is durably stored.
+# Layer map
 
-OnDisk layers can be Image or Delta:
-- ImageLayer represents an image or a snapshot of a segment at one particular LSN.
-- DeltaLayer represents a collection of WAL records or page images in a range of LSNs.
-
-Dropped segments are always represented on disk by DeltaLayer.
-
-LSN range defined by start_lsn and end_lsn:
-- start_lsn is inclusive.
-- end_lsn is exclusive.
-
-For an open in-memory layer, the end_lsn is MAX_LSN. For a frozen
-in-memory layer or a delta layer, it is a valid end bound. An image
-layer represents snapshot at one LSN, so end_lsn is always the
-snapshot LSN + 1
-
-Layers can be open or historical:
-- Open layer is a writeable one. Only InMemory layer can be open.
-FIXME: If open layer is dropped, it is not writeable, so it should be turned into historical, 
-but now it is not implemented - see bug #569.
-- Historical layer is the one that cannot be modified anymore. Now only OnDisk layers can be historical.
-
-- LayerMap - a map that tracks what layers exist for all the relishes in a timeline.
+The LayerMap tracks what layers exist for all the relishes in a timeline.
 
 LayerMap consists of two data structures:
 - segs - All the layers keyed by segment tag
@@ -54,8 +75,55 @@ TODO: Are there any exceptions to this?
 For example, timeline.list_rels(lsn) will return all segments that are visible in this timeline at the LSN,
 including ones that were not modified in this timeline and thus don't have a layer in the timeline's LayerMap.
 
-TODO:
-Describe GC and checkpoint interval settings.
+
+# Different kinds of layers
+
+A layer can be in different states:
+
+- Open - a layer where new WAL records can be appended to.
+- Closed - a layer that is read-only, no new WAL records can be appended to it
+- Historic: synonym for closed
+- InMemory: A layer that needs to be rebuilt from WAL on pageserver start.
+To avoid OOM errors, InMemory layers can be spilled to disk into ephemeral file.
+- OnDisk: A layer that is stored on disk. If its end-LSN is older than
+  disk_consistent_lsn, it is known to be fully flushed and fsync'd to local disk.
+- Frozen layer: an in-memory layer that is Closed.
+
+TODO: Clarify the difference between Closed, Historic and Frozen.
+
+There are two kinds of OnDisk layers:
+- ImageLayer represents an image or a snapshot of a 10 MB relish segment, at one particular LSN.
+- DeltaLayer represents a collection of WAL records or page images in a range of LSNs, for one
+  relish segment.
+
+Dropped segments are always represented on disk by DeltaLayer.
+
+# Layer life cycle
+
+LSN range defined by start_lsn and end_lsn:
+- start_lsn is inclusive.
+- end_lsn is exclusive.
+
+For an open in-memory layer, the end_lsn is MAX_LSN. For a frozen in-memory
+layer or a delta layer, it is a valid end bound. An image layer represents
+snapshot at one LSN, so end_lsn is always the snapshot LSN + 1
+
+Every layer starts its life as an Open In-Memory layer. When the page server
+receives the first WAL record for a segment, it creates a new In-Memory layer
+for it, and puts it to the layer map. Later, the layer is old enough, its
+contents are written to disk, as On-Disk layers. This process is called
+"evicting" a layer.
+
+Layer eviction is a two-step process: First, the layer is marked as closed, so
+that it no longer accepts new WAL records, and the layer map is updated
+accordingly. If a new WAL record for that segment arrives after this step, a new
+Open layer is created to hold it. After this first step, the layer is a Closed
+InMemory state. This first step is called "freezing" the layer.
+
+In the second step, new Delta and Image layers are created, containing all the
+data in the Frozen InMemory layer. When the new layers are ready, the original
+frozen layer is replaced with the new layers in the layer map, and the original
+frozen layer is dropped, releasing the memory.
 
 # Layer files (On-disk layers)
 
@@ -366,6 +434,8 @@ is a newer layer file there. TODO: This optimization hasn't been
 implemented! The GC algorithm will currently keep the file on the
 'main' branch anyway, for as long as the child branch exists.
 
+TODO:
+Describe GC and checkpoint interval settings.
 
 # TODO: On LSN ranges
 

pageserver/src/layered_repository/blob.rs

@@ -1,4 +1,5 @@
-use std::{fs::File, io::Write};
+use std::io::{Read, Write};
+use std::os::unix::prelude::FileExt;
 
 use anyhow::Result;
 use bookfile::{BookWriter, BoundedReader, ChapterId, ChapterWriter};
@@ -10,7 +11,7 @@ pub struct BlobRange {
     size: usize,
 }
 
-pub fn read_blob(reader: &BoundedReader<&'_ File>, range: &BlobRange) -> Result<Vec<u8>> {
+pub fn read_blob<F: FileExt>(reader: &BoundedReader<&'_ F>, range: &BlobRange) -> Result<Vec<u8>> {
     let mut buf = vec![0u8; range.size];
     reader.read_exact_at(&mut buf, range.offset)?;
     Ok(buf)
@@ -28,14 +29,14 @@ impl<W: Write> BlobWriter<W> {
         Self { writer, offset: 0 }
     }
 
-    pub fn write_blob(&mut self, blob: &[u8]) -> Result<BlobRange> {
-        self.writer.write_all(blob)?;
+    pub fn write_blob_from_reader(&mut self, r: &mut impl Read) -> Result<BlobRange> {
+        let len = std::io::copy(r, &mut self.writer)?;
 
         let range = BlobRange {
             offset: self.offset,
-            size: blob.len(),
+            size: len as usize,
         };
-        self.offset += blob.len() as u64;
+        self.offset += len as u64;
         Ok(range)
     }
 

pageserver/src/layered_repository/delta_layer.rs

@@ -39,21 +39,22 @@
 //!
 use crate::layered_repository::blob::BlobWriter;
 use crate::layered_repository::filename::{DeltaFileName, PathOrConf};
+use crate::layered_repository::page_versions::PageVersions;
 use crate::layered_repository::storage_layer::{
     Layer, PageReconstructData, PageReconstructResult, PageVersion, SegmentTag,
 };
+use crate::virtual_file::VirtualFile;
 use crate::waldecoder;
 use crate::PageServerConf;
 use crate::{ZTenantId, ZTimelineId};
 use anyhow::{bail, ensure, Result};
 use log::*;
 use serde::{Deserialize, Serialize};
-use std::collections::BTreeMap;
+use zenith_utils::vec_map::VecMap;
 // avoid binding to Write (conflicts with std::io::Write)
 // while being able to use std::fmt::Write's methods
 use std::fmt::Write as _;
 use std::fs;
-use std::fs::File;
 use std::io::{BufWriter, Write};
 use std::ops::Bound::Included;
 use std::path::{Path, PathBuf};
@@ -139,15 +140,21 @@ pub struct DeltaLayerInner {
     /// loaded into memory yet.
     loaded: bool,
 
+    book: Option<Book<VirtualFile>>,
+
     /// All versions of all pages in the file are are kept here.
     /// Indexed by block number and LSN.
-    page_version_metas: BTreeMap<(u32, Lsn), BlobRange>,
+    page_version_metas: VecMap<(u32, Lsn), BlobRange>,
 
     /// `relsizes` tracks the size of the relation at different points in time.
-    relsizes: BTreeMap<Lsn, u32>,
+    relsizes: VecMap<Lsn, u32>,
 }
 
 impl Layer for DeltaLayer {
+    fn get_tenant_id(&self) -> ZTenantId {
+        self.tenantid
+    }
+
     fn get_timeline_id(&self) -> ZTimelineId {
         self.timelineid
     }
@@ -169,29 +176,7 @@ impl Layer for DeltaLayer {
     }
 
     fn filename(&self) -> PathBuf {
-        PathBuf::from(
-            DeltaFileName {
-                seg: self.seg,
-                start_lsn: self.start_lsn,
-                end_lsn: self.end_lsn,
-                dropped: self.dropped,
-            }
-            .to_string(),
-        )
-    }
-
-    fn path(&self) -> Option<PathBuf> {
-        Some(Self::path_for(
-            &self.path_or_conf,
-            self.timelineid,
-            self.tenantid,
-            &DeltaFileName {
-                seg: self.seg,
-                start_lsn: self.start_lsn,
-                end_lsn: self.end_lsn,
-                dropped: self.dropped,
-            },
-        ))
+        PathBuf::from(self.layer_name().to_string())
     }
 
     /// Look up given page in the cache.
@@ -199,44 +184,63 @@ impl Layer for DeltaLayer {
         &self,
         blknum: u32,
         lsn: Lsn,
+        cached_img_lsn: Option<Lsn>,
         reconstruct_data: &mut PageReconstructData,
     ) -> Result<PageReconstructResult> {
         let mut need_image = true;
 
         assert!(self.seg.blknum_in_seg(blknum));
 
+        match &cached_img_lsn {
+            Some(cached_lsn) if &self.end_lsn <= cached_lsn => {
+                return Ok(PageReconstructResult::Cached)
+            }
+            _ => {}
+        }
+
         {
             // Open the file and lock the metadata in memory
-            // TODO: avoid opening the file for each read
-            let (_path, book) = self.open_book()?;
-            let page_version_reader = book.chapter_reader(PAGE_VERSIONS_CHAPTER)?;
             let inner = self.load()?;
+            let page_version_reader = inner
+                .book
+                .as_ref()
+                .expect("should be loaded in load call above")
+                .chapter_reader(PAGE_VERSIONS_CHAPTER)?;
 
             // Scan the metadata BTreeMap backwards, starting from the given entry.
             let minkey = (blknum, Lsn(0));
             let maxkey = (blknum, lsn);
-            let mut iter = inner
+            let iter = inner
                 .page_version_metas
-                .range((Included(&minkey), Included(&maxkey)));
-            while let Some(((_blknum, _entry_lsn), blob_range)) = iter.next_back() {
+                .slice_range((Included(&minkey), Included(&maxkey)))
+                .iter()
+                .rev();
+            for ((_blknum, pv_lsn), blob_range) in iter {
+                match &cached_img_lsn {
+                    Some(cached_lsn) if pv_lsn <= cached_lsn => {
+                        return Ok(PageReconstructResult::Cached)
+                    }
+                    _ => {}
+                }
+
                 let pv = PageVersion::des(&read_blob(&page_version_reader, blob_range)?)?;
 
-                if let Some(img) = pv.page_image {
-                    // Found a page image, return it
-                    reconstruct_data.page_img = Some(img);
-                    need_image = false;
-                    break;
-                } else if let Some(rec) = pv.record {
-                    let will_init = rec.will_init;
-                    reconstruct_data.records.push(rec);
-                    if will_init {
-                        // This WAL record initializes the page, so no need to go further back
+                match pv {
+                    PageVersion::Page(img) => {
+                        // Found a page image, return it
+                        reconstruct_data.page_img = Some(img);
                         need_image = false;
                         break;
                     }
-                } else {
-                    // No base image, and no WAL record. Huh?
-                    bail!("no page image or WAL record for requested page");
+                    PageVersion::Wal(rec) => {
+                        let will_init = rec.will_init;
+                        reconstruct_data.records.push((*pv_lsn, rec));
+                        if will_init {
+                            // This WAL record initializes the page, so no need to go further back
+                            need_image = false;
+                            break;
+                        }
+                    }
                 }
             }
 
@@ -246,7 +250,7 @@ impl Layer for DeltaLayer {
         // If an older page image is needed to reconstruct the page, let the
         // caller know.
         if need_image {
-            Ok(PageReconstructResult::Continue(self.start_lsn))
+            Ok(PageReconstructResult::Continue(Lsn(self.start_lsn.0 - 1)))
         } else {
             Ok(PageReconstructResult::Complete)
         }
@@ -262,15 +266,15 @@ impl Layer for DeltaLayer {
 
         // Scan the BTreeMap backwards, starting from the given entry.
         let inner = self.load()?;
-        let mut iter = inner.relsizes.range((Included(&Lsn(0)), Included(&lsn)));
+        let slice = inner
+            .relsizes
+            .slice_range((Included(&Lsn(0)), Included(&lsn)));
 
-        let result;
-        if let Some((_entry_lsn, entry)) = iter.next_back() {
-            result = *entry;
+        if let Some((_entry_lsn, entry)) = slice.last() {
+            Ok(*entry)
         } else {
-            bail!("could not find seg size in delta layer");
+            Err(anyhow::anyhow!("could not find seg size in delta layer"))
         }
-        Ok(result)
     }
 
     /// Does this segment exist at given LSN?
@@ -290,17 +294,20 @@ impl Layer for DeltaLayer {
     ///
     fn unload(&self) -> Result<()> {
         let mut inner = self.inner.lock().unwrap();
-        inner.page_version_metas = BTreeMap::new();
-        inner.relsizes = BTreeMap::new();
+        inner.page_version_metas = VecMap::default();
+        inner.relsizes = VecMap::default();
         inner.loaded = false;
+
+        // Note: we keep the Book open. Is that a good idea? The virtual file
+        // machinery has its own rules for closing the file descriptor if it's not
+        // needed, but the Book struct uses up some memory, too.
+
         Ok(())
     }
 
     fn delete(&self) -> Result<()> {
         // delete underlying file
-        if let Some(path) = self.path() {
-            fs::remove_file(path)?;
-        }
+        fs::remove_file(self.path())?;
         Ok(())
     }
 
@@ -308,6 +315,10 @@ impl Layer for DeltaLayer {
         true
     }
 
+    fn is_in_memory(&self) -> bool {
+        false
+    }
+
     /// debugging function to print out the contents of the layer
     fn dump(&self) -> Result<()> {
         println!(
@@ -317,31 +328,38 @@ impl Layer for DeltaLayer {
 
         println!("--- relsizes ---");
         let inner = self.load()?;
-        for (k, v) in inner.relsizes.iter() {
+        for (k, v) in inner.relsizes.as_slice() {
             println!("  {}: {}", k, v);
         }
         println!("--- page versions ---");
-        let (_path, book) = self.open_book()?;
+
+        let path = self.path();
+        let file = std::fs::File::open(&path)?;
+        let book = Book::new(file)?;
+
         let chapter = book.chapter_reader(PAGE_VERSIONS_CHAPTER)?;
-        for ((blk, lsn), blob_range) in inner.page_version_metas.iter() {
+        for ((blk, lsn), blob_range) in inner.page_version_metas.as_slice() {
             let mut desc = String::new();
 
             let buf = read_blob(&chapter, blob_range)?;
             let pv = PageVersion::des(&buf)?;
 
-            if let Some(img) = pv.page_image.as_ref() {
-                write!(&mut desc, " img {} bytes", img.len())?;
-            }
-            if let Some(rec) = pv.record.as_ref() {
-                let wal_desc = waldecoder::describe_wal_record(&rec.rec);
-                write!(
-                    &mut desc,
-                    " rec {} bytes will_init: {} {}",
-                    rec.rec.len(),
-                    rec.will_init,
-                    wal_desc
-                )?;
+            match pv {
+                PageVersion::Page(img) => {
+                    write!(&mut desc, " img {} bytes", img.len())?;
+                }
+                PageVersion::Wal(rec) => {
+                    let wal_desc = waldecoder::describe_wal_record(&rec.rec);
+                    write!(
+                        &mut desc,
+                        " rec {} bytes will_init: {} {}",
+                        rec.rec.len(),
+                        rec.will_init,
+                        wal_desc
+                    )?;
+                }
             }
+
             println!("  blk {} at {}: {}", blk, lsn, desc);
         }
 
@@ -365,14 +383,14 @@ impl DeltaLayer {
     }
 
     /// Create a new delta file, using the given page versions and relsizes.
-    /// The page versions are passed by an iterator; the iterator must return
-    /// page versions in blknum+lsn order.
+    /// The page versions are passed in a PageVersions struct. If 'cutoff' is
+    /// given, only page versions with LSN < cutoff are included.
     ///
-    /// This is used to write the in-memory layer to disk. The in-memory layer uses the same
-    /// data structure with two btreemaps as we do, so passing the btreemaps is currently
-    /// expedient.
+    /// This is used to write the in-memory layer to disk. The page_versions and
+    /// relsizes are thus passed in the same format as they are in the in-memory
+    /// layer, as that's expedient.
     #[allow(clippy::too_many_arguments)]
-    pub fn create<'a>(
+    pub fn create(
         conf: &'static PageServerConf,
         timelineid: ZTimelineId,
         tenantid: ZTenantId,
@@ -380,8 +398,9 @@ impl DeltaLayer {
         start_lsn: Lsn,
         end_lsn: Lsn,
         dropped: bool,
-        page_versions: impl Iterator<Item = (&'a (u32, Lsn), &'a PageVersion)>,
-        relsizes: BTreeMap<Lsn, u32>,
+        page_versions: &PageVersions,
+        cutoff: Option<Lsn>,
+        relsizes: VecMap<Lsn, u32>,
     ) -> Result<DeltaLayer> {
         if seg.rel.is_blocky() {
             assert!(!relsizes.is_empty());
@@ -396,46 +415,51 @@ impl DeltaLayer {
             end_lsn,
             dropped,
             inner: Mutex::new(DeltaLayerInner {
-                loaded: true,
-                page_version_metas: BTreeMap::new(),
+                loaded: false,
+                book: None,
+                page_version_metas: VecMap::default(),
                 relsizes,
             }),
         };
         let mut inner = delta_layer.inner.lock().unwrap();
 
-        // Write the in-memory btreemaps into a file
-        let path = delta_layer
-            .path()
-            .expect("DeltaLayer is supposed to have a layer path on disk");
-
+        // Write the data into a file
+        //
+        // Note: Because we open the file in write-only mode, we cannot
+        // reuse the same VirtualFile for reading later. That's why we don't
+        // set inner.book here. The first read will have to re-open it.
+        //
         // Note: This overwrites any existing file. There shouldn't be any.
         // FIXME: throw an error instead?
-        let file = File::create(&path)?;
+        let path = delta_layer.path();
+        let file = VirtualFile::create(&path)?;
         let buf_writer = BufWriter::new(file);
         let book = BookWriter::new(buf_writer, DELTA_FILE_MAGIC)?;
 
         let mut page_version_writer = BlobWriter::new(book, PAGE_VERSIONS_CHAPTER);
 
-        for (key, page_version) in page_versions {
-            let buf = PageVersion::ser(page_version)?;
-            let blob_range = page_version_writer.write_blob(&buf)?;
+        let page_versions_iter = page_versions.ordered_page_version_iter(cutoff);
+        for (blknum, lsn, pos) in page_versions_iter {
+            let blob_range =
+                page_version_writer.write_blob_from_reader(&mut page_versions.reader(pos)?)?;
 
-            let old = inner.page_version_metas.insert(*key, blob_range);
-
-            assert!(old.is_none());
+            inner
+                .page_version_metas
+                .append((blknum, lsn), blob_range)
+                .unwrap();
         }
 
         let book = page_version_writer.close()?;
 
         // Write out page versions
         let mut chapter = book.new_chapter(PAGE_VERSION_METAS_CHAPTER);
-        let buf = BTreeMap::ser(&inner.page_version_metas)?;
+        let buf = VecMap::ser(&inner.page_version_metas)?;
         chapter.write_all(&buf)?;
         let book = chapter.close()?;
 
         // and relsizes to separate chapter
         let mut chapter = book.new_chapter(REL_SIZES_CHAPTER);
-        let buf = BTreeMap::ser(&inner.relsizes)?;
+        let buf = VecMap::ser(&inner.relsizes)?;
         chapter.write_all(&buf)?;
         let book = chapter.close()?;
 
@@ -464,25 +488,6 @@ impl DeltaLayer {
         Ok(delta_layer)
     }
 
-    fn open_book(&self) -> Result<(PathBuf, Book<File>)> {
-        let path = Self::path_for(
-            &self.path_or_conf,
-            self.timelineid,
-            self.tenantid,
-            &DeltaFileName {
-                seg: self.seg,
-                start_lsn: self.start_lsn,
-                end_lsn: self.end_lsn,
-                dropped: self.dropped,
-            },
-        );
-
-        let file = File::open(&path)?;
-        let book = Book::new(file)?;
-
-        Ok((path, book))
-    }
-
     ///
     /// Load the contents of the file into memory
     ///
@@ -494,7 +499,14 @@ impl DeltaLayer {
             return Ok(inner);
         }
 
-        let (path, book) = self.open_book()?;
+        let path = self.path();
+
+        // Open the file if it's not open already.
+        if inner.book.is_none() {
+            let file = VirtualFile::open(&path)?;
+            inner.book = Some(Book::new(file)?);
+        }
+        let book = inner.book.as_ref().unwrap();
 
         match &self.path_or_conf {
             PathOrConf::Conf(_) => {
@@ -522,18 +534,16 @@ impl DeltaLayer {
         }
 
         let chapter = book.read_chapter(PAGE_VERSION_METAS_CHAPTER)?;
-        let page_version_metas = BTreeMap::des(&chapter)?;
+        let page_version_metas = VecMap::des(&chapter)?;
 
         let chapter = book.read_chapter(REL_SIZES_CHAPTER)?;
-        let relsizes = BTreeMap::des(&chapter)?;
+        let relsizes = VecMap::des(&chapter)?;
 
         debug!("loaded from {}", &path.display());
 
-        *inner = DeltaLayerInner {
-            loaded: true,
-            page_version_metas,
-            relsizes,
-        };
+        inner.page_version_metas = page_version_metas;
+        inner.relsizes = relsizes;
+        inner.loaded = true;
 
         Ok(inner)
     }
@@ -555,8 +565,9 @@ impl DeltaLayer {
             dropped: filename.dropped,
             inner: Mutex::new(DeltaLayerInner {
                 loaded: false,
-                page_version_metas: BTreeMap::new(),
-                relsizes: BTreeMap::new(),
+                book: None,
+                page_version_metas: VecMap::default(),
+                relsizes: VecMap::default(),
             }),
         }
     }
@@ -564,7 +575,10 @@ impl DeltaLayer {
     /// Create a DeltaLayer struct representing an existing file on disk.
     ///
     /// This variant is only used for debugging purposes, by the 'dump_layerfile' binary.
-    pub fn new_for_path(path: &Path, book: &Book<File>) -> Result<Self> {
+    pub fn new_for_path<F>(path: &Path, book: &Book<F>) -> Result<Self>
+    where
+        F: std::os::unix::prelude::FileExt,
+    {
         let chapter = book.read_chapter(SUMMARY_CHAPTER)?;
         let summary = Summary::des(&chapter)?;
 
@@ -578,9 +592,29 @@ impl DeltaLayer {
             dropped: summary.dropped,
             inner: Mutex::new(DeltaLayerInner {
                 loaded: false,
-                page_version_metas: BTreeMap::new(),
-                relsizes: BTreeMap::new(),
+                book: None,
+                page_version_metas: VecMap::default(),
+                relsizes: VecMap::default(),
             }),
         })
     }
+
+    fn layer_name(&self) -> DeltaFileName {
+        DeltaFileName {
+            seg: self.seg,
+            start_lsn: self.start_lsn,
+            end_lsn: self.end_lsn,
+            dropped: self.dropped,
+        }
+    }
+
+    /// Path to the layer file in pageserver workdir.
+    pub fn path(&self) -> PathBuf {
+        Self::path_for(
+            &self.path_or_conf,
+            self.timelineid,
+            self.tenantid,
+            &self.layer_name(),
+        )
+    }
 }

pageserver/src/layered_repository/ephemeral_file.rs

@@ -0,0 +1,298 @@
+//! Implementation of append-only file data structure
+//! used to keep in-memory layers spilled on disk.
+
+use crate::page_cache;
+use crate::page_cache::PAGE_SZ;
+use crate::page_cache::{ReadBufResult, WriteBufResult};
+use crate::virtual_file::VirtualFile;
+use crate::PageServerConf;
+use lazy_static::lazy_static;
+use std::cmp::min;
+use std::collections::HashMap;
+use std::fs::OpenOptions;
+use std::io::{Error, ErrorKind, Seek, SeekFrom, Write};
+use std::ops::DerefMut;
+use std::path::PathBuf;
+use std::sync::{Arc, RwLock};
+use zenith_utils::zid::ZTenantId;
+use zenith_utils::zid::ZTimelineId;
+
+use std::os::unix::fs::FileExt;
+
+lazy_static! {
+    ///
+    /// This is the global cache of file descriptors (File objects).
+    ///
+    static ref EPHEMERAL_FILES: RwLock<EphemeralFiles> = RwLock::new(EphemeralFiles {
+        next_file_id: 1,
+        files: HashMap::new(),
+    });
+}
+
+pub struct EphemeralFiles {
+    next_file_id: u64,
+
+    files: HashMap<u64, Arc<VirtualFile>>,
+}
+
+pub struct EphemeralFile {
+    file_id: u64,
+    _tenantid: ZTenantId,
+    _timelineid: ZTimelineId,
+    file: Arc<VirtualFile>,
+
+    pos: u64,
+}
+
+impl EphemeralFile {
+    pub fn create(
+        conf: &PageServerConf,
+        tenantid: ZTenantId,
+        timelineid: ZTimelineId,
+    ) -> Result<EphemeralFile, std::io::Error> {
+        let mut l = EPHEMERAL_FILES.write().unwrap();
+        let file_id = l.next_file_id;
+        l.next_file_id += 1;
+
+        let filename = conf
+            .timeline_path(&timelineid, &tenantid)
+            .join(PathBuf::from(format!("ephemeral-{}", file_id)));
+
+        let file = VirtualFile::open_with_options(
+            &filename,
+            OpenOptions::new().read(true).write(true).create(true),
+        )?;
+        let file_rc = Arc::new(file);
+        l.files.insert(file_id, file_rc.clone());
+
+        Ok(EphemeralFile {
+            file_id,
+            _tenantid: tenantid,
+            _timelineid: timelineid,
+            file: file_rc,
+            pos: 0,
+        })
+    }
+
+    pub fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), Error> {
+        let mut off = 0;
+        while off < PAGE_SZ {
+            let n = self
+                .file
+                .read_at(&mut buf[off..], blkno as u64 * PAGE_SZ as u64 + off as u64)?;
+
+            if n == 0 {
+                // Reached EOF. Fill the rest of the buffer with zeros.
+                const ZERO_BUF: [u8; PAGE_SZ] = [0u8; PAGE_SZ];
+
+                buf[off..].copy_from_slice(&ZERO_BUF[off..]);
+                break;
+            }
+
+            off += n as usize;
+        }
+        Ok(())
+    }
+}
+
+impl FileExt for EphemeralFile {
+    fn read_at(&self, dstbuf: &mut [u8], offset: u64) -> Result<usize, Error> {
+        // Look up the right page
+        let blkno = (offset / PAGE_SZ as u64) as u32;
+        let off = offset as usize % PAGE_SZ;
+        let len = min(PAGE_SZ - off, dstbuf.len());
+
+        let read_guard;
+        let mut write_guard;
+
+        let cache = page_cache::get();
+        let buf = match cache.read_ephemeral_buf(self.file_id, blkno) {
+            ReadBufResult::Found(guard) => {
+                read_guard = guard;
+                read_guard.as_ref()
+            }
+            ReadBufResult::NotFound(guard) => {
+                // Read the page from disk into the buffer
+                write_guard = guard;
+                self.fill_buffer(write_guard.deref_mut(), blkno)?;
+                write_guard.mark_valid();
+
+                // And then fall through to read the requested slice from the
+                // buffer.
+                write_guard.as_ref()
+            }
+        };
+
+        dstbuf[0..len].copy_from_slice(&buf[off..(off + len)]);
+        Ok(len)
+    }
+
+    fn write_at(&self, srcbuf: &[u8], offset: u64) -> Result<usize, Error> {
+        // Look up the right page
+        let blkno = (offset / PAGE_SZ as u64) as u32;
+        let off = offset as usize % PAGE_SZ;
+        let len = min(PAGE_SZ - off, srcbuf.len());
+
+        let mut write_guard;
+        let cache = page_cache::get();
+        let buf = match cache.write_ephemeral_buf(self.file_id, blkno) {
+            WriteBufResult::Found(guard) => {
+                write_guard = guard;
+                write_guard.deref_mut()
+            }
+            WriteBufResult::NotFound(guard) => {
+                // Read the page from disk into the buffer
+                // TODO: if we're overwriting the whole page, no need to read it in first
+                write_guard = guard;
+                self.fill_buffer(write_guard.deref_mut(), blkno)?;
+                write_guard.mark_valid();
+
+                // And then fall through to modify it.
+                write_guard.deref_mut()
+            }
+        };
+
+        buf[off..(off + len)].copy_from_slice(&srcbuf[0..len]);
+        write_guard.mark_dirty();
+        Ok(len)
+    }
+}
+
+impl Write for EphemeralFile {
+    fn write(&mut self, buf: &[u8]) -> Result<usize, Error> {
+        let n = self.write_at(buf, self.pos)?;
+        self.pos += n as u64;
+        Ok(n)
+    }
+
+    fn flush(&mut self) -> Result<(), std::io::Error> {
+        todo!()
+    }
+}
+
+impl Seek for EphemeralFile {
+    fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
+        match pos {
+            SeekFrom::Start(offset) => {
+                self.pos = offset;
+            }
+            SeekFrom::End(_offset) => {
+                return Err(Error::new(
+                    ErrorKind::Other,
+                    "SeekFrom::End not supported by EphemeralFile",
+                ));
+            }
+            SeekFrom::Current(offset) => {
+                let pos = self.pos as i128 + offset as i128;
+                if pos < 0 {
+                    return Err(Error::new(
+                        ErrorKind::InvalidInput,
+                        "offset would be negative",
+                    ));
+                }
+                if pos > u64::MAX as i128 {
+                    return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
+                }
+                self.pos = pos as u64;
+            }
+        }
+        Ok(self.pos)
+    }
+}
+
+impl Drop for EphemeralFile {
+    fn drop(&mut self) {
+        // drop all pages from page cache
+        let cache = page_cache::get();
+        cache.drop_buffers_for_ephemeral(self.file_id);
+
+        // remove entry from the hash map
+        EPHEMERAL_FILES.write().unwrap().files.remove(&self.file_id);
+
+        // unlink file
+        // FIXME: print error
+        let _ = std::fs::remove_file(&self.file.path);
+    }
+}
+
+pub fn writeback(file_id: u64, blkno: u32, buf: &[u8]) -> Result<(), std::io::Error> {
+    if let Some(file) = EPHEMERAL_FILES.read().unwrap().files.get(&file_id) {
+        file.write_all_at(buf, blkno as u64 * PAGE_SZ as u64)?;
+        Ok(())
+    } else {
+        Err(std::io::Error::new(
+            ErrorKind::Other,
+            "could not write back page, not found in ephemeral files hash",
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use rand::seq::SliceRandom;
+    use rand::thread_rng;
+    use std::fs;
+    use std::str::FromStr;
+
+    fn repo_harness(
+        test_name: &str,
+    ) -> Result<(&'static PageServerConf, ZTenantId, ZTimelineId), Error> {
+        let repo_dir = PageServerConf::test_repo_dir(test_name);
+        let _ = fs::remove_dir_all(&repo_dir);
+        let conf = PageServerConf::dummy_conf(repo_dir);
+        // Make a static copy of the config. This can never be free'd, but that's
+        // OK in a test.
+        let conf: &'static PageServerConf = Box::leak(Box::new(conf));
+
+        let tenantid = ZTenantId::from_str("11000000000000000000000000000000").unwrap();
+        let timelineid = ZTimelineId::from_str("22000000000000000000000000000000").unwrap();
+        fs::create_dir_all(conf.timeline_path(&timelineid, &tenantid))?;
+
+        Ok((conf, tenantid, timelineid))
+    }
+
+    // Helper function to slurp contents of a file, starting at the current position,
+    // into a string
+    fn read_string(efile: &EphemeralFile, offset: u64, len: usize) -> Result<String, Error> {
+        let mut buf = Vec::new();
+        buf.resize(len, 0u8);
+
+        efile.read_exact_at(&mut buf, offset)?;
+
+        Ok(String::from_utf8_lossy(&buf)
+            .trim_end_matches('\0')
+            .to_string())
+    }
+
+    #[test]
+    fn test_ephemeral_files() -> Result<(), Error> {
+        let (conf, tenantid, timelineid) = repo_harness("ephemeral_files")?;
+
+        let mut file_a = EphemeralFile::create(conf, tenantid, timelineid)?;
+
+        file_a.write_all(b"foo")?;
+        assert_eq!("foo", read_string(&file_a, 0, 20)?);
+
+        file_a.write_all(b"bar")?;
+        assert_eq!("foobar", read_string(&file_a, 0, 20)?);
+
+        // Open a lot of files, enough to cause some page evictions.
+        let mut efiles = Vec::new();
+        for fileno in 0..100 {
+            let mut efile = EphemeralFile::create(conf, tenantid, timelineid)?;
+            efile.write_all(format!("file {}", fileno).as_bytes())?;
+            assert_eq!(format!("file {}", fileno), read_string(&efile, 0, 10)?);
+            efiles.push((fileno, efile));
+        }
+
+        // Check that all the files can still be read from. Use them in random order for
+        // good measure.
+        efiles.as_mut_slice().shuffle(&mut thread_rng());
+        for (fileno, efile) in efiles.iter_mut() {
+            assert_eq!(format!("file {}", fileno), read_string(efile, 0, 10)?);
+        }
+
+        Ok(())
+    }
+}

pageserver/src/layered_repository/filename.rs

@@ -13,6 +13,8 @@ use anyhow::Result;
 use log::*;
 use zenith_utils::lsn::Lsn;
 
+use super::metadata::METADATA_FILE_NAME;
+
 // Note: LayeredTimeline::load_layer_map() relies on this sort order
 #[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Clone)]
 pub struct DeltaFileName {
@@ -35,7 +37,7 @@ impl DeltaFileName {
     /// Parse a string as a delta file name. Returns None if the filename does not
     /// match the expected pattern.
     ///
-    pub fn from_str(fname: &str) -> Option<Self> {
+    pub fn parse_str(fname: &str) -> Option<Self> {
         let rel;
         let mut parts;
         if let Some(rest) = fname.strip_prefix("rel_") {
@@ -168,7 +170,7 @@ impl ImageFileName {
     /// Parse a string as an image file name. Returns None if the filename does not
     /// match the expected pattern.
     ///
-    pub fn from_str(fname: &str) -> Option<Self> {
+    pub fn parse_str(fname: &str) -> Option<Self> {
         let rel;
         let mut parts;
         if let Some(rest) = fname.strip_prefix("rel_") {
@@ -286,11 +288,11 @@ pub fn list_files(
         let fname = direntry?.file_name();
         let fname = fname.to_str().unwrap();
 
-        if let Some(deltafilename) = DeltaFileName::from_str(fname) {
+        if let Some(deltafilename) = DeltaFileName::parse_str(fname) {
             deltafiles.push(deltafilename);
-        } else if let Some(imgfilename) = ImageFileName::from_str(fname) {
+        } else if let Some(imgfilename) = ImageFileName::parse_str(fname) {
             imgfiles.push(imgfilename);
-        } else if fname == "metadata" || fname == "ancestor" || fname.ends_with(".old") {
+        } else if fname == METADATA_FILE_NAME || fname.ends_with(".old") {
             // ignore these
         } else {
             warn!("unrecognized filename in timeline dir: {}", fname);

pageserver/src/layered_repository/global_layer_map.rs

@@ -0,0 +1,142 @@
+//!
+//! Global registry of open layers.
+//!
+//! Whenever a new in-memory layer is created to hold incoming WAL, it is registered
+//! in [`GLOBAL_LAYER_MAP`], so that we can keep track of the total number of
+//! in-memory layers in the system, and know when we need to evict some to release
+//! memory.
+//!
+//! Each layer is assigned a unique ID when it's registered in the global registry.
+//! The ID can be used to relocate the layer later, without having to hold locks.
+//!
+
+use std::sync::atomic::{AtomicU8, Ordering};
+use std::sync::{Arc, RwLock};
+
+use super::inmemory_layer::InMemoryLayer;
+
+use lazy_static::lazy_static;
+
+const MAX_USAGE_COUNT: u8 = 5;
+
+lazy_static! {
+    pub static ref GLOBAL_LAYER_MAP: RwLock<InMemoryLayers> =
+        RwLock::new(InMemoryLayers::default());
+}
+
+// TODO these types can probably be smaller
+#[derive(PartialEq, Eq, Clone, Copy)]
+pub struct LayerId {
+    index: usize,
+    tag: u64, // to avoid ABA problem
+}
+
+enum SlotData {
+    Occupied(Arc<InMemoryLayer>),
+    /// Vacant slots form a linked list, the value is the index
+    /// of the next vacant slot in the list.
+    Vacant(Option<usize>),
+}
+
+struct Slot {
+    tag: u64,
+    data: SlotData,
+    usage_count: AtomicU8, // for clock algorithm
+}
+
+#[derive(Default)]
+pub struct InMemoryLayers {
+    slots: Vec<Slot>,
+    num_occupied: usize,
+
+    // Head of free-slot list.
+    next_empty_slot_idx: Option<usize>,
+}
+
+impl InMemoryLayers {
+    pub fn insert(&mut self, layer: Arc<InMemoryLayer>) -> LayerId {
+        let slot_idx = match self.next_empty_slot_idx {
+            Some(slot_idx) => slot_idx,
+            None => {
+                let idx = self.slots.len();
+                self.slots.push(Slot {
+                    tag: 0,
+                    data: SlotData::Vacant(None),
+                    usage_count: AtomicU8::new(0),
+                });
+                idx
+            }
+        };
+        let slots_len = self.slots.len();
+
+        let slot = &mut self.slots[slot_idx];
+
+        match slot.data {
+            SlotData::Occupied(_) => {
+                panic!("an occupied slot was in the free list");
+            }
+            SlotData::Vacant(next_empty_slot_idx) => {
+                self.next_empty_slot_idx = next_empty_slot_idx;
+            }
+        }
+
+        slot.data = SlotData::Occupied(layer);
+        slot.usage_count.store(1, Ordering::Relaxed);
+
+        self.num_occupied += 1;
+        assert!(self.num_occupied <= slots_len);
+
+        LayerId {
+            index: slot_idx,
+            tag: slot.tag,
+        }
+    }
+
+    pub fn get(&self, layer_id: &LayerId) -> Option<Arc<InMemoryLayer>> {
+        let slot = self.slots.get(layer_id.index)?; // TODO should out of bounds indexes just panic?
+        if slot.tag != layer_id.tag {
+            return None;
+        }
+
+        if let SlotData::Occupied(layer) = &slot.data {
+            let _ = slot.usage_count.fetch_update(
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+                |old_usage_count| {
+                    if old_usage_count < MAX_USAGE_COUNT {
+                        Some(old_usage_count + 1)
+                    } else {
+                        None
+                    }
+                },
+            );
+            Some(Arc::clone(layer))
+        } else {
+            None
+        }
+    }
+
+    // TODO this won't be a public API in the future
+    pub fn remove(&mut self, layer_id: &LayerId) {
+        let slot = &mut self.slots[layer_id.index];
+
+        if slot.tag != layer_id.tag {
+            return;
+        }
+
+        match &slot.data {
+            SlotData::Occupied(_layer) => {
+                // TODO evict the layer
+            }
+            SlotData::Vacant(_) => unimplemented!(),
+        }
+
+        slot.data = SlotData::Vacant(self.next_empty_slot_idx);
+        self.next_empty_slot_idx = Some(layer_id.index);
+
+        assert!(self.num_occupied > 0);
+        self.num_occupied -= 1;
+
+        slot.tag = slot.tag.wrapping_add(1);
+    }
+}

pageserver/src/layered_repository/image_layer.rs

@@ -27,15 +27,15 @@ use crate::layered_repository::storage_layer::{
 };
 use crate::layered_repository::LayeredTimeline;
 use crate::layered_repository::RELISH_SEG_SIZE;
+use crate::virtual_file::VirtualFile;
 use crate::PageServerConf;
 use crate::{ZTenantId, ZTimelineId};
-use anyhow::{anyhow, bail, ensure, Result};
+use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::Bytes;
 use log::*;
 use serde::{Deserialize, Serialize};
 use std::convert::TryInto;
 use std::fs;
-use std::fs::File;
 use std::io::{BufWriter, Write};
 use std::path::{Path, PathBuf};
 use std::sync::{Mutex, MutexGuard};
@@ -104,9 +104,8 @@ enum ImageType {
 }
 
 pub struct ImageLayerInner {
-    /// If false, the 'image_type' has not been
-    /// loaded into memory yet.
-    loaded: bool,
+    /// If None, the 'image_type' has not been loaded into memory yet.
+    book: Option<Book<VirtualFile>>,
 
     /// Derived from filename and bookfile chapter metadata
     image_type: ImageType,
@@ -114,25 +113,11 @@ pub struct ImageLayerInner {
 
 impl Layer for ImageLayer {
     fn filename(&self) -> PathBuf {
-        PathBuf::from(
-            ImageFileName {
-                seg: self.seg,
-                lsn: self.lsn,
-            }
-            .to_string(),
-        )
+        PathBuf::from(self.layer_name().to_string())
     }
 
-    fn path(&self) -> Option<PathBuf> {
-        Some(Self::path_for(
-            &self.path_or_conf,
-            self.timelineid,
-            self.tenantid,
-            &ImageFileName {
-                seg: self.seg,
-                lsn: self.lsn,
-            },
-        ))
+    fn get_tenant_id(&self) -> ZTenantId {
+        self.tenantid
     }
 
     fn get_timeline_id(&self) -> ZTimelineId {
@@ -161,16 +146,20 @@ impl Layer for ImageLayer {
         &self,
         blknum: u32,
         lsn: Lsn,
+        cached_img_lsn: Option<Lsn>,
         reconstruct_data: &mut PageReconstructData,
     ) -> Result<PageReconstructResult> {
         assert!(lsn >= self.lsn);
 
+        match cached_img_lsn {
+            Some(cached_lsn) if self.lsn <= cached_lsn => return Ok(PageReconstructResult::Cached),
+            _ => {}
+        }
+
         let inner = self.load()?;
 
         let base_blknum = blknum % RELISH_SEG_SIZE;
 
-        let (_path, book) = self.open_book()?;
-
         let buf = match &inner.image_type {
             ImageType::Blocky { num_blocks } => {
                 if base_blknum >= *num_blocks {
@@ -180,14 +169,23 @@ impl Layer for ImageLayer {
                 let mut buf = vec![0u8; BLOCK_SIZE];
                 let offset = BLOCK_SIZE as u64 * base_blknum as u64;
 
-                let chapter = book.chapter_reader(BLOCKY_IMAGES_CHAPTER)?;
+                let chapter = inner
+                    .book
+                    .as_ref()
+                    .unwrap()
+                    .chapter_reader(BLOCKY_IMAGES_CHAPTER)?;
                 chapter.read_exact_at(&mut buf, offset)?;
 
                 buf
             }
             ImageType::NonBlocky => {
                 ensure!(base_blknum == 0);
-                book.read_chapter(NONBLOCKY_IMAGE_CHAPTER)?.into_vec()
+                inner
+                    .book
+                    .as_ref()
+                    .unwrap()
+                    .read_chapter(NONBLOCKY_IMAGE_CHAPTER)?
+                    .into_vec()
             }
         };
 
@@ -209,22 +207,13 @@ impl Layer for ImageLayer {
         Ok(true)
     }
 
-    ///
-    /// Release most of the memory used by this layer. If it's accessed again later,
-    /// it will need to be loaded back.
-    ///
     fn unload(&self) -> Result<()> {
-        let mut inner = self.inner.lock().unwrap();
-        inner.image_type = ImageType::Blocky { num_blocks: 0 };
-        inner.loaded = false;
         Ok(())
     }
 
     fn delete(&self) -> Result<()> {
         // delete underlying file
-        if let Some(path) = self.path() {
-            fs::remove_file(path)?;
-        }
+        fs::remove_file(self.path())?;
         Ok(())
     }
 
@@ -232,6 +221,10 @@ impl Layer for ImageLayer {
         false
     }
 
+    fn is_in_memory(&self) -> bool {
+        false
+    }
+
     /// debugging function to print out the contents of the layer
     fn dump(&self) -> Result<()> {
         println!(
@@ -244,8 +237,11 @@ impl Layer for ImageLayer {
         match inner.image_type {
             ImageType::Blocky { num_blocks } => println!("({}) blocks ", num_blocks),
             ImageType::NonBlocky => {
-                let (_path, book) = self.open_book()?;
-                let chapter = book.read_chapter(NONBLOCKY_IMAGE_CHAPTER)?;
+                let chapter = inner
+                    .book
+                    .as_ref()
+                    .unwrap()
+                    .read_chapter(NONBLOCKY_IMAGE_CHAPTER)?;
                 println!("non-blocky ({} bytes)", chapter.len());
             }
         }
@@ -293,19 +289,22 @@ impl ImageLayer {
             seg,
             lsn,
             inner: Mutex::new(ImageLayerInner {
-                loaded: true,
+                book: None,
                 image_type: image_type.clone(),
             }),
         };
         let inner = layer.inner.lock().unwrap();
 
         // Write the images into a file
-        let path = layer
-            .path()
-            .expect("ImageLayer is supposed to have a layer path on disk");
+        //
+        // Note: Because we open the file in write-only mode, we cannot
+        // reuse the same VirtualFile for reading later. That's why we don't
+        // set inner.book here. The first read will have to re-open it.
+        //
         // Note: This overwrites any existing file. There shouldn't be any.
         // FIXME: throw an error instead?
-        let file = File::create(&path)?;
+        let path = layer.path();
+        let file = VirtualFile::create(&path)?;
         let buf_writer = BufWriter::new(file);
         let book = BookWriter::new(buf_writer, IMAGE_FILE_MAGIC)?;
 
@@ -340,7 +339,7 @@ impl ImageLayer {
         let writer = book.close()?;
         writer.get_ref().sync_all()?;
 
-        trace!("saved {}", &path.display());
+        trace!("saved {}", path.display());
 
         drop(inner);
 
@@ -392,11 +391,19 @@ impl ImageLayer {
         // quick exit if already loaded
         let mut inner = self.inner.lock().unwrap();
 
-        if inner.loaded {
+        if inner.book.is_some() {
             return Ok(inner);
         }
 
-        let (path, book) = self.open_book()?;
+        let path = self.path();
+        let file = VirtualFile::open(&path)
+            .with_context(|| format!("Failed to open virtual file '{}'", path.display()))?;
+        let book = Book::new(file).with_context(|| {
+            format!(
+                "Failed to open virtual file '{}' as a bookfile",
+                path.display()
+            )
+        })?;
 
         match &self.path_or_conf {
             PathOrConf::Conf(_) => {
@@ -437,30 +444,13 @@ impl ImageLayer {
         debug!("loaded from {}", &path.display());
 
         *inner = ImageLayerInner {
-            loaded: true,
+            book: Some(book),
             image_type,
         };
 
         Ok(inner)
     }
 
-    fn open_book(&self) -> Result<(PathBuf, Book<File>)> {
-        let path = Self::path_for(
-            &self.path_or_conf,
-            self.timelineid,
-            self.tenantid,
-            &ImageFileName {
-                seg: self.seg,
-                lsn: self.lsn,
-            },
-        );
-
-        let file = File::open(&path)?;
-        let book = Book::new(file)?;
-
-        Ok((path, book))
-    }
-
     /// Create an ImageLayer struct representing an existing file on disk
     pub fn new(
         conf: &'static PageServerConf,
@@ -475,7 +465,7 @@ impl ImageLayer {
             seg: filename.seg,
             lsn: filename.lsn,
             inner: Mutex::new(ImageLayerInner {
-                loaded: false,
+                book: None,
                 image_type: ImageType::Blocky { num_blocks: 0 },
             }),
         }
@@ -484,7 +474,10 @@ impl ImageLayer {
     /// Create an ImageLayer struct representing an existing file on disk.
     ///
     /// This variant is only used for debugging purposes, by the 'dump_layerfile' binary.
-    pub fn new_for_path(path: &Path, book: &Book<File>) -> Result<ImageLayer> {
+    pub fn new_for_path<F>(path: &Path, book: &Book<F>) -> Result<ImageLayer>
+    where
+        F: std::os::unix::prelude::FileExt,
+    {
         let chapter = book.read_chapter(SUMMARY_CHAPTER)?;
         let summary = Summary::des(&chapter)?;
 
@@ -495,9 +488,26 @@ impl ImageLayer {
             seg: summary.seg,
             lsn: summary.lsn,
             inner: Mutex::new(ImageLayerInner {
-                loaded: false,
+                book: None,
                 image_type: ImageType::Blocky { num_blocks: 0 },
             }),
         })
     }
+
+    fn layer_name(&self) -> ImageFileName {
+        ImageFileName {
+            seg: self.seg,
+            lsn: self.lsn,
+        }
+    }
+
+    /// Path to the layer file in pageserver workdir.
+    pub fn path(&self) -> PathBuf {
+        Self::path_for(
+            &self.path_or_conf,
+            self.timelineid,
+            self.tenantid,
+            &self.layer_name(),
+        )
+    }
 }

pageserver/src/layered_repository/inmemory_layer.rs

@@ -1,7 +1,10 @@
+//! An in-memory layer stores recently received PageVersions.
+//! The page versions are held in a BTreeMap. To avoid OOM errors, the map size is limited
+//! and layers can be spilled to disk into ephemeral files.
 //!
-//! An in-memory layer stores recently received page versions in memory. The page versions
-//! are held in a BTreeMap, and there's another BTreeMap to track the size of the relation.
+//! And there's another BTreeMap to track the size of the relation.
 //!
+use crate::layered_repository::ephemeral_file::EphemeralFile;
 use crate::layered_repository::filename::DeltaFileName;
 use crate::layered_repository::storage_layer::{
     Layer, PageReconstructData, PageReconstructResult, PageVersion, SegmentTag, RELISH_SEG_SIZE,
@@ -12,17 +15,15 @@ use crate::layered_repository::{DeltaLayer, ImageLayer};
 use crate::repository::WALRecord;
 use crate::PageServerConf;
 use crate::{ZTenantId, ZTimelineId};
-use anyhow::{bail, ensure, Result};
+use anyhow::{ensure, Result};
 use bytes::Bytes;
 use log::*;
-use std::cmp::min;
-use std::collections::BTreeMap;
-use std::ops::Bound::Included;
 use std::path::PathBuf;
 use std::sync::{Arc, RwLock};
-
-use zenith_utils::accum::Accum;
 use zenith_utils::lsn::Lsn;
+use zenith_utils::vec_map::VecMap;
+
+use super::page_versions::PageVersions;
 
 pub struct InMemoryLayer {
     conf: &'static PageServerConf,
@@ -36,9 +37,6 @@ pub struct InMemoryLayer {
     ///
     start_lsn: Lsn,
 
-    /// Frozen in-memory layers have an inclusive end LSN.
-    end_lsn: Option<Lsn>,
-
     /// LSN of the oldest page version stored in this layer
     oldest_pending_lsn: Lsn,
 
@@ -51,14 +49,19 @@ pub struct InMemoryLayer {
 }
 
 pub struct InMemoryLayerInner {
+    /// Frozen layers have an exclusive end LSN.
+    /// Writes are only allowed when this is None
+    end_lsn: Option<Lsn>,
+
     /// If this relation was dropped, remember when that happened.
-    drop_lsn: Option<Lsn>,
+    /// The drop LSN is recorded in [`end_lsn`].
+    dropped: bool,
 
     ///
     /// All versions of all pages in the layer are are kept here.
     /// Indexed by block number and LSN.
     ///
-    page_versions: BTreeMap<(u32, Lsn), PageVersion>,
+    page_versions: PageVersions,
 
     ///
     /// `segsizes` tracks the size of the segment at different points in time.
@@ -67,28 +70,20 @@ pub struct InMemoryLayerInner {
     /// so that determining the size never depends on the predecessor layer. For
     /// a non-blocky rel, 'segsizes' is not used and is always empty.
     ///
-    segsizes: BTreeMap<Lsn, u32>,
-
-    /// Writes are only allowed when true.
-    /// Set to false when this layer is in the process of being replaced.
-    writeable: bool,
+    segsizes: VecMap<Lsn, u32>,
 }
 
 impl InMemoryLayerInner {
-    fn check_writeable(&self) -> WriteResult<()> {
-        if self.writeable {
-            Ok(())
-        } else {
-            Err(NonWriteableError)
-        }
+    fn assert_writeable(&self) {
+        assert!(self.end_lsn.is_none());
     }
 
     fn get_seg_size(&self, lsn: Lsn) -> u32 {
         // Scan the BTreeMap backwards, starting from the given entry.
-        let mut iter = self.segsizes.range((Included(&Lsn(0)), Included(&lsn)));
+        let slice = self.segsizes.slice_range(..=lsn);
 
         // We make sure there is always at least one entry
-        if let Some((_entry_lsn, entry)) = iter.next_back() {
+        if let Some((_entry_lsn, entry)) = slice.last() {
             *entry
         } else {
             panic!("could not find seg size in in-memory layer");
@@ -97,34 +92,32 @@ impl InMemoryLayerInner {
 }
 
 impl Layer for InMemoryLayer {
-    // An in-memory layer doesn't really have a filename as it's not stored on disk,
-    // but we construct a filename as if it was a delta layer
+    // An in-memory layer can be spilled to disk into ephemeral file,
+    // This function is used only for debugging, so we don't need to be very precise.
+    // Construct a filename as if it was a delta layer.
     fn filename(&self) -> PathBuf {
         let inner = self.inner.read().unwrap();
 
         let end_lsn;
-        let dropped;
-        if let Some(drop_lsn) = inner.drop_lsn {
+        if let Some(drop_lsn) = inner.end_lsn {
             end_lsn = drop_lsn;
-            dropped = true;
         } else {
             end_lsn = Lsn(u64::MAX);
-            dropped = false;
         }
 
         let delta_filename = DeltaFileName {
             seg: self.seg,
             start_lsn: self.start_lsn,
             end_lsn,
-            dropped,
+            dropped: inner.dropped,
         }
         .to_string();
 
         PathBuf::from(format!("inmem-{}", delta_filename))
     }
 
-    fn path(&self) -> Option<PathBuf> {
-        None
+    fn get_tenant_id(&self) -> ZTenantId {
+        self.tenantid
     }
 
     fn get_timeline_id(&self) -> ZTimelineId {
@@ -140,14 +133,10 @@ impl Layer for InMemoryLayer {
     }
 
     fn get_end_lsn(&self) -> Lsn {
-        if let Some(end_lsn) = self.end_lsn {
-            return Lsn(end_lsn.0 + 1);
-        }
-
         let inner = self.inner.read().unwrap();
 
-        if let Some(drop_lsn) = inner.drop_lsn {
-            drop_lsn
+        if let Some(end_lsn) = inner.end_lsn {
+            end_lsn
         } else {
             Lsn(u64::MAX)
         }
@@ -155,7 +144,7 @@ impl Layer for InMemoryLayer {
 
     fn is_dropped(&self) -> bool {
         let inner = self.inner.read().unwrap();
-        inner.drop_lsn.is_some()
+        inner.dropped
     }
 
     /// Look up given page in the cache.
@@ -163,6 +152,7 @@ impl Layer for InMemoryLayer {
         &self,
         blknum: u32,
         lsn: Lsn,
+        cached_img_lsn: Option<Lsn>,
         reconstruct_data: &mut PageReconstructData,
     ) -> Result<PageReconstructResult> {
         let mut need_image = true;
@@ -172,27 +162,35 @@ impl Layer for InMemoryLayer {
         {
             let inner = self.inner.read().unwrap();
 
-            // Scan the BTreeMap backwards, starting from reconstruct_data.lsn.
-            let minkey = (blknum, Lsn(0));
-            let maxkey = (blknum, lsn);
-            let mut iter = inner
+            // Scan the page versions backwards, starting from `lsn`.
+            let iter = inner
                 .page_versions
-                .range((Included(&minkey), Included(&maxkey)));
-            while let Some(((_blknum, _entry_lsn), entry)) = iter.next_back() {
-                if let Some(img) = &entry.page_image {
-                    reconstruct_data.page_img = Some(img.clone());
-                    need_image = false;
-                    break;
-                } else if let Some(rec) = &entry.record {
-                    reconstruct_data.records.push(rec.clone());
-                    if rec.will_init {
-                        // This WAL record initializes the page, so no need to go further back
+                .get_block_lsn_range(blknum, ..=lsn)
+                .iter()
+                .rev();
+            for (entry_lsn, pos) in iter {
+                match &cached_img_lsn {
+                    Some(cached_lsn) if entry_lsn <= cached_lsn => {
+                        return Ok(PageReconstructResult::Cached)
+                    }
+                    _ => {}
+                }
+
+                let pv = inner.page_versions.get_page_version(*pos)?;
+                match pv {
+                    PageVersion::Page(img) => {
+                        reconstruct_data.page_img = Some(img);
                         need_image = false;
                         break;
                     }
-                } else {
-                    // No base image, and no WAL record. Huh?
-                    bail!("no page image or WAL record for requested page");
+                    PageVersion::Wal(rec) => {
+                        reconstruct_data.records.push((*entry_lsn, rec.clone()));
+                        if rec.will_init {
+                            // This WAL record initializes the page, so no need to go further back
+                            need_image = false;
+                            break;
+                        }
+                    }
                 }
             }
             // release lock on 'inner'
@@ -233,9 +231,13 @@ impl Layer for InMemoryLayer {
         assert!(lsn >= self.start_lsn);
 
         // Is the requested LSN after the segment was dropped?
-        if let Some(drop_lsn) = inner.drop_lsn {
-            if lsn >= drop_lsn {
-                return Ok(false);
+        if inner.dropped {
+            if let Some(end_lsn) = inner.end_lsn {
+                if lsn >= end_lsn {
+                    return Ok(false);
+                }
+            } else {
+                panic!("dropped in-memory layer with no end LSN");
             }
         }
 
@@ -253,66 +255,57 @@ impl Layer for InMemoryLayer {
     /// Nothing to do here. When you drop the last reference to the layer, it will
     /// be deallocated.
     fn delete(&self) -> Result<()> {
-        Ok(())
+        panic!("can't delete an InMemoryLayer")
     }
 
     fn is_incremental(&self) -> bool {
         self.incremental
     }
 
+    fn is_in_memory(&self) -> bool {
+        true
+    }
+
     /// debugging function to print out the contents of the layer
     fn dump(&self) -> Result<()> {
         let inner = self.inner.read().unwrap();
 
         let end_str = inner
-            .drop_lsn
+            .end_lsn
             .as_ref()
-            .map(|drop_lsn| drop_lsn.to_string())
+            .map(Lsn::to_string)
             .unwrap_or_default();
 
         println!(
-            "----- in-memory layer for tli {} seg {} {}-{} ----",
-            self.timelineid, self.seg, self.start_lsn, end_str
+            "----- in-memory layer for tli {} seg {} {}-{} {} ----",
+            self.timelineid, self.seg, self.start_lsn, end_str, inner.dropped,
         );
 
-        for (k, v) in inner.segsizes.iter() {
+        for (k, v) in inner.segsizes.as_slice() {
             println!("segsizes {}: {}", k, v);
         }
 
-        for (k, v) in inner.page_versions.iter() {
-            println!(
-                "blk {} at {}: {}/{}\n",
-                k.0,
-                k.1,
-                v.page_image.is_some(),
-                v.record.is_some()
-            );
+        for (blknum, lsn, pos) in inner.page_versions.ordered_page_version_iter(None) {
+            let pv = inner.page_versions.get_page_version(pos)?;
+            let pv_description = match pv {
+                PageVersion::Page(_img) => "page",
+                PageVersion::Wal(_rec) => "wal",
+            };
+
+            println!("blk {} at {}: {}\n", blknum, lsn, pv_description);
         }
 
         Ok(())
     }
 }
 
-/// Write failed because the layer is in process of being replaced.
-/// See [`LayeredTimeline::perform_write_op`] for how to handle this error.
-#[derive(Debug)]
-pub struct NonWriteableError;
-
-pub type WriteResult<T> = std::result::Result<T, NonWriteableError>;
-
-/// Helper struct to cleanup `InMemoryLayer::freeze` return signature.
-pub struct FreezeLayers {
-    /// Replacement layer for the layer which freeze was called on.
-    pub frozen: Arc<InMemoryLayer>,
-    /// New open layer containing leftover data.
-    pub open: Option<Arc<InMemoryLayer>>,
+/// A result of an inmemory layer data being written to disk.
+pub struct LayersOnDisk {
+    pub delta_layers: Vec<DeltaLayer>,
+    pub image_layers: Vec<ImageLayer>,
 }
 
 impl InMemoryLayer {
-    fn assert_not_frozen(&self) {
-        assert!(self.end_lsn.is_none());
-    }
-
     /// Return the oldest page version that's stored in this layer
     pub fn get_oldest_pending_lsn(&self) -> Lsn {
         self.oldest_pending_lsn
@@ -337,25 +330,26 @@ impl InMemoryLayer {
         );
 
         // The segment is initially empty, so initialize 'segsizes' with 0.
-        let mut segsizes = BTreeMap::new();
+        let mut segsizes = VecMap::default();
         if seg.rel.is_blocky() {
-            segsizes.insert(start_lsn, 0);
+            segsizes.append(start_lsn, 0).unwrap();
         }
 
+        let file = EphemeralFile::create(conf, tenantid, timelineid)?;
+
         Ok(InMemoryLayer {
             conf,
             timelineid,
             tenantid,
             seg,
             start_lsn,
-            end_lsn: None,
             oldest_pending_lsn,
             incremental: false,
             inner: RwLock::new(InMemoryLayerInner {
-                drop_lsn: None,
-                page_versions: BTreeMap::new(),
+                end_lsn: None,
+                dropped: false,
+                page_versions: PageVersions::new(file),
                 segsizes,
-                writeable: true,
             }),
         })
     }
@@ -363,33 +357,18 @@ impl InMemoryLayer {
     // Write operations
 
     /// Remember new page version, as a WAL record over previous version
-    pub fn put_wal_record(&self, blknum: u32, rec: WALRecord) -> WriteResult<u32> {
-        self.put_page_version(
-            blknum,
-            rec.lsn,
-            PageVersion {
-                page_image: None,
-                record: Some(rec),
-            },
-        )
+    pub fn put_wal_record(&self, lsn: Lsn, blknum: u32, rec: WALRecord) -> Result<u32> {
+        self.put_page_version(blknum, lsn, PageVersion::Wal(rec))
     }
 
     /// Remember new page version, as a full page image
-    pub fn put_page_image(&self, blknum: u32, lsn: Lsn, img: Bytes) -> WriteResult<u32> {
-        self.put_page_version(
-            blknum,
-            lsn,
-            PageVersion {
-                page_image: Some(img),
-                record: None,
-            },
-        )
+    pub fn put_page_image(&self, blknum: u32, lsn: Lsn, img: Bytes) -> Result<u32> {
+        self.put_page_version(blknum, lsn, PageVersion::Page(img))
     }
 
     /// Common subroutine of the public put_wal_record() and put_page_image() functions.
     /// Adds the page version to the in-memory tree
-    pub fn put_page_version(&self, blknum: u32, lsn: Lsn, pv: PageVersion) -> WriteResult<u32> {
-        self.assert_not_frozen();
+    pub fn put_page_version(&self, blknum: u32, lsn: Lsn, pv: PageVersion) -> Result<u32> {
         assert!(self.seg.blknum_in_seg(blknum));
 
         trace!(
@@ -401,9 +380,9 @@ impl InMemoryLayer {
         );
         let mut inner = self.inner.write().unwrap();
 
-        inner.check_writeable()?;
+        inner.assert_writeable();
 
-        let old = inner.page_versions.insert((blknum, lsn), pv);
+        let old = inner.page_versions.append_or_update_last(blknum, lsn, pv)?;
 
         if old.is_some() {
             // We already had an entry for this LSN. That's odd..
@@ -439,16 +418,15 @@ impl InMemoryLayer {
                 // subsequent call to initialize the gap page.
                 let gapstart = self.seg.segno * RELISH_SEG_SIZE + oldsize;
                 for gapblknum in gapstart..blknum {
-                    let zeropv = PageVersion {
-                        page_image: Some(ZERO_PAGE.clone()),
-                        record: None,
-                    };
+                    let zeropv = PageVersion::Page(ZERO_PAGE.clone());
                     trace!(
                         "filling gap blk {} with zeros for write of {}",
                         gapblknum,
                         blknum
                     );
-                    let old = inner.page_versions.insert((gapblknum, lsn), zeropv);
+                    let old = inner
+                        .page_versions
+                        .append_or_update_last(gapblknum, lsn, zeropv)?;
                     // We already had an entry for this LSN. That's odd..
 
                     if old.is_some() {
@@ -459,53 +437,47 @@ impl InMemoryLayer {
                     }
                 }
 
-                inner.segsizes.insert(lsn, newsize);
+                inner.segsizes.append_or_update_last(lsn, newsize).unwrap();
                 return Ok(newsize - oldsize);
             }
         }
+
         Ok(0)
     }
 
     /// Remember that the relation was truncated at given LSN
-    pub fn put_truncation(&self, lsn: Lsn, segsize: u32) -> WriteResult<()> {
+    pub fn put_truncation(&self, lsn: Lsn, segsize: u32) {
         assert!(
             self.seg.rel.is_blocky(),
             "put_truncation() called on a non-blocky rel"
         );
-        self.assert_not_frozen();
 
         let mut inner = self.inner.write().unwrap();
-        inner.check_writeable()?;
+        inner.assert_writeable();
 
         // check that this we truncate to a smaller size than segment was before the truncation
         let oldsize = inner.get_seg_size(lsn);
         assert!(segsize < oldsize);
 
-        let old = inner.segsizes.insert(lsn, segsize);
+        let (old, _delta_size) = inner.segsizes.append_or_update_last(lsn, segsize).unwrap();
 
         if old.is_some() {
             // We already had an entry for this LSN. That's odd..
             warn!("Inserting truncation, but had an entry for the LSN already");
         }
-
-        Ok(())
     }
 
     /// Remember that the segment was dropped at given LSN
-    pub fn drop_segment(&self, lsn: Lsn) -> WriteResult<()> {
-        self.assert_not_frozen();
-
+    pub fn drop_segment(&self, lsn: Lsn) {
         let mut inner = self.inner.write().unwrap();
 
-        inner.check_writeable()?;
-
-        assert!(inner.drop_lsn.is_none());
-        inner.drop_lsn = Some(lsn);
-        inner.writeable = false;
+        assert!(inner.end_lsn.is_none());
+        assert!(!inner.dropped);
+        inner.dropped = true;
+        assert!(self.start_lsn < lsn);
+        inner.end_lsn = Some(lsn);
 
         trace!("dropped segment {} at {}", self.seg, lsn);
-
-        Ok(())
     }
 
     ///
@@ -533,136 +505,57 @@ impl InMemoryLayer {
         );
 
         // Copy the segment size at the start LSN from the predecessor layer.
-        let mut segsizes = BTreeMap::new();
+        let mut segsizes = VecMap::default();
         if seg.rel.is_blocky() {
             let size = src.get_seg_size(start_lsn)?;
-            segsizes.insert(start_lsn, size);
+            segsizes.append(start_lsn, size).unwrap();
         }
 
+        let file = EphemeralFile::create(conf, tenantid, timelineid)?;
+
         Ok(InMemoryLayer {
             conf,
             timelineid,
             tenantid,
             seg,
             start_lsn,
-            end_lsn: None,
             oldest_pending_lsn,
             incremental: true,
             inner: RwLock::new(InMemoryLayerInner {
-                drop_lsn: None,
-                page_versions: BTreeMap::new(),
+                end_lsn: None,
+                dropped: false,
+                page_versions: PageVersions::new(file),
                 segsizes,
-                writeable: true,
             }),
         })
     }
 
     pub fn is_writeable(&self) -> bool {
         let inner = self.inner.read().unwrap();
-        inner.writeable
+        inner.end_lsn.is_none()
     }
 
-    /// Splits `self` into two InMemoryLayers: `frozen` and `open`.
-    /// All data up to and including `cutoff_lsn`
-    /// is copied to `frozen`, while the remaining data is copied to `open`.
-    /// After completion, self is non-writeable, but not frozen.
-    pub fn freeze(self: Arc<Self>, cutoff_lsn: Lsn) -> Result<FreezeLayers> {
-        info!(
-            "freezing in memory layer {} on timeline {} at {} (oldest {})",
-            self.filename().display(),
-            self.timelineid,
-            cutoff_lsn,
-            self.oldest_pending_lsn
-        );
+    /// Make the layer non-writeable. Only call once.
+    /// Records the end_lsn for non-dropped layers.
+    /// `end_lsn` is inclusive
+    pub fn freeze(&self, end_lsn: Lsn) {
+        let mut inner = self.inner.write().unwrap();
 
-        self.assert_not_frozen();
-
-        let self_ref = self.clone();
-        let mut inner = self_ref.inner.write().unwrap();
-        // Dropped layers don't need any special freeze actions,
-        // they are marked as non-writeable at drop and just
-        // written out to disk by checkpointer.
-        if inner.drop_lsn.is_some() {
-            assert!(!inner.writeable);
-            info!(
-                "freezing in memory layer for {} on timeline {} is dropped at {}",
-                self.seg,
-                self.timelineid,
-                inner.drop_lsn.unwrap()
-            );
-
-            // There should be no newer layer that refers this non-writeable layer,
-            // because layer that is created after dropped one represents a new rel.
-            return Ok(FreezeLayers {
-                frozen: self,
-                open: None,
-            });
-        }
-        assert!(inner.writeable);
-        inner.writeable = false;
-
-        // Divide all the page versions into old and new
-        // at the 'cutoff_lsn' point.
-        let mut before_segsizes = BTreeMap::new();
-        let mut after_segsizes = BTreeMap::new();
-        let mut after_oldest_lsn: Accum<Lsn> = Accum(None);
-        for (lsn, size) in inner.segsizes.iter() {
-            if *lsn > cutoff_lsn {
-                after_segsizes.insert(*lsn, *size);
-                after_oldest_lsn.accum(min, *lsn);
-            } else {
-                before_segsizes.insert(*lsn, *size);
-            }
-        }
-
-        let mut before_page_versions = BTreeMap::new();
-        let mut after_page_versions = BTreeMap::new();
-        for ((blknum, lsn), pv) in inner.page_versions.iter() {
-            if *lsn > cutoff_lsn {
-                after_page_versions.insert((*blknum, *lsn), pv.clone());
-                after_oldest_lsn.accum(min, *lsn);
-            } else {
-                before_page_versions.insert((*blknum, *lsn), pv.clone());
-            }
-        }
-
-        let frozen = Arc::new(InMemoryLayer {
-            conf: self.conf,
-            tenantid: self.tenantid,
-            timelineid: self.timelineid,
-            seg: self.seg,
-            start_lsn: self.start_lsn,
-            end_lsn: Some(cutoff_lsn),
-            oldest_pending_lsn: self.start_lsn,
-            incremental: self.incremental,
-            inner: RwLock::new(InMemoryLayerInner {
-                drop_lsn: inner.drop_lsn,
-                page_versions: before_page_versions,
-                segsizes: before_segsizes,
-                writeable: false,
-            }),
-        });
-
-        let open = if !after_segsizes.is_empty() || !after_page_versions.is_empty() {
-            let mut new_open = Self::create_successor_layer(
-                self.conf,
-                frozen.clone(),
-                self.timelineid,
-                self.tenantid,
-                cutoff_lsn + 1,
-                after_oldest_lsn.0.unwrap(),
-            )?;
-
-            let new_inner = new_open.inner.get_mut().unwrap();
-            new_inner.page_versions.append(&mut after_page_versions);
-            new_inner.segsizes.append(&mut after_segsizes);
-
-            Some(Arc::new(new_open))
+        if inner.end_lsn.is_some() {
+            assert!(inner.dropped);
         } else {
-            None
-        };
+            assert!(!inner.dropped);
+            assert!(self.start_lsn < end_lsn + 1);
+            inner.end_lsn = Some(Lsn(end_lsn.0 + 1));
 
-        Ok(FreezeLayers { frozen, open })
+            if let Some((lsn, _)) = inner.segsizes.as_slice().last() {
+                assert!(lsn <= &end_lsn, "{:?} {:?}", lsn, end_lsn);
+            }
+
+            for (_blk, lsn, _pv) in inner.page_versions.ordered_page_version_iter(None) {
+                assert!(lsn <= end_lsn);
+            }
+        }
     }
 
     /// Write the this frozen in-memory layer to disk.
@@ -673,16 +566,15 @@ impl InMemoryLayer {
     /// WAL records between start and end LSN. (The delta layer is not needed
     /// when a new relish is created with a single LSN, so that the start and
     /// end LSN are the same.)
-    pub fn write_to_disk(&self, timeline: &LayeredTimeline) -> Result<Vec<Arc<dyn Layer>>> {
+    pub fn write_to_disk(&self, timeline: &LayeredTimeline) -> Result<LayersOnDisk> {
         trace!(
-            "write_to_disk {} end_lsn is {} get_end_lsn is {}",
+            "write_to_disk {} get_end_lsn is {}",
             self.filename().display(),
-            self.end_lsn.unwrap_or(Lsn(0)),
             self.get_end_lsn()
         );
 
         // Grab the lock in read-mode. We hold it over the I/O, but because this
-        // layer is not writeable anymore, no one should be trying to aquire the
+        // layer is not writeable anymore, no one should be trying to acquire the
         // write lock on it, so we shouldn't block anyone. There's one exception
         // though: another thread might have grabbed a reference to this layer
         // in `get_layer_for_write' just before the checkpointer called
@@ -691,46 +583,42 @@ impl InMemoryLayer {
         // would have to wait until we release it. That race condition is very
         // rare though, so we just accept the potential latency hit for now.
         let inner = self.inner.read().unwrap();
-        assert!(!inner.writeable);
+        let end_lsn_exclusive = inner.end_lsn.unwrap();
 
-        if let Some(drop_lsn) = inner.drop_lsn {
+        if inner.dropped {
             let delta_layer = DeltaLayer::create(
                 self.conf,
                 self.timelineid,
                 self.tenantid,
                 self.seg,
                 self.start_lsn,
-                drop_lsn,
+                end_lsn_exclusive,
                 true,
-                inner.page_versions.iter(),
+                &inner.page_versions,
+                None,
                 inner.segsizes.clone(),
             )?;
             trace!(
                 "freeze: created delta layer for dropped segment {} {}-{}",
                 self.seg,
                 self.start_lsn,
-                drop_lsn
+                end_lsn_exclusive
             );
-            return Ok(vec![Arc::new(delta_layer)]);
+            return Ok(LayersOnDisk {
+                delta_layers: vec![delta_layer],
+                image_layers: Vec::new(),
+            });
         }
 
-        let end_lsn = self.end_lsn.unwrap();
+        // Since `end_lsn` is inclusive, subtract 1.
+        // We want to make an ImageLayer for the last included LSN,
+        // so the DeltaLayer should exclude that LSN.
+        let end_lsn_inclusive = Lsn(end_lsn_exclusive.0 - 1);
 
-        let mut before_segsizes = BTreeMap::new();
-        for (lsn, size) in inner.segsizes.iter() {
-            if *lsn <= end_lsn {
-                before_segsizes.insert(*lsn, *size);
-            }
-        }
-        let mut before_page_versions = inner.page_versions.iter().filter(|tup| {
-            let ((_blknum, lsn), _pv) = tup;
+        let mut delta_layers = Vec::new();
 
-            *lsn < end_lsn
-        });
-
-        let mut frozen_layers: Vec<Arc<dyn Layer>> = Vec::new();
-
-        if self.start_lsn != end_lsn {
+        if self.start_lsn != end_lsn_inclusive {
+            let (segsizes, _) = inner.segsizes.split_at(&end_lsn_exclusive);
             // Write the page versions before the cutoff to disk.
             let delta_layer = DeltaLayer::create(
                 self.conf,
@@ -738,29 +626,41 @@ impl InMemoryLayer {
                 self.tenantid,
                 self.seg,
                 self.start_lsn,
-                end_lsn,
+                end_lsn_inclusive,
                 false,
-                before_page_versions,
-                before_segsizes,
+                &inner.page_versions,
+                Some(end_lsn_inclusive),
+                segsizes,
             )?;
-            frozen_layers.push(Arc::new(delta_layer));
+            delta_layers.push(delta_layer);
             trace!(
                 "freeze: created delta layer {} {}-{}",
                 self.seg,
                 self.start_lsn,
-                end_lsn
+                end_lsn_inclusive
             );
         } else {
-            assert!(before_page_versions.next().is_none());
+            assert!(inner
+                .page_versions
+                .ordered_page_version_iter(None)
+                .next()
+                .is_none());
         }
 
         drop(inner);
 
         // Write a new base image layer at the cutoff point
-        let image_layer = ImageLayer::create_from_src(self.conf, timeline, self, end_lsn)?;
-        frozen_layers.push(Arc::new(image_layer));
-        trace!("freeze: created image layer {} at {}", self.seg, end_lsn);
+        let image_layer =
+            ImageLayer::create_from_src(self.conf, timeline, self, end_lsn_inclusive)?;
+        trace!(
+            "freeze: created image layer {} at {}",
+            self.seg,
+            end_lsn_inclusive
+        );
 
-        Ok(frozen_layers)
+        Ok(LayersOnDisk {
+            delta_layers,
+            image_layers: vec![image_layer],
+        })
     }
 }

pageserver/src/layered_repository/layer_map.rs

@@ -21,6 +21,8 @@ use std::sync::Arc;
 use zenith_metrics::{register_int_gauge, IntGauge};
 use zenith_utils::lsn::Lsn;
 
+use super::global_layer_map::{LayerId, GLOBAL_LAYER_MAP};
+
 lazy_static! {
     static ref NUM_INMEMORY_LAYERS: IntGauge =
         register_int_gauge!("pageserver_inmemory_layers", "Number of layers in memory")
@@ -68,7 +70,9 @@ impl LayerMap {
     pub fn get_open(&self, tag: &SegmentTag) -> Option<Arc<InMemoryLayer>> {
         let segentry = self.segs.get(tag)?;
 
-        segentry.open.as_ref().map(Arc::clone)
+        segentry
+            .open_layer_id
+            .and_then(|layer_id| GLOBAL_LAYER_MAP.read().unwrap().get(&layer_id))
     }
 
     ///
@@ -77,7 +81,7 @@ impl LayerMap {
     pub fn insert_open(&mut self, layer: Arc<InMemoryLayer>) {
         let segentry = self.segs.entry(layer.get_seg_tag()).or_default();
 
-        segentry.update_open(Arc::clone(&layer));
+        let layer_id = segentry.update_open(Arc::clone(&layer));
 
         let oldest_pending_lsn = layer.get_oldest_pending_lsn();
 
@@ -89,7 +93,7 @@ impl LayerMap {
         // Also add it to the binary heap
         let open_layer_entry = OpenLayerEntry {
             oldest_pending_lsn: layer.get_oldest_pending_lsn(),
-            layer,
+            layer_id,
             generation: self.current_generation,
         };
         self.open_layers.push(open_layer_entry);
@@ -97,24 +101,35 @@ impl LayerMap {
         NUM_INMEMORY_LAYERS.inc();
     }
 
-    /// Remove the oldest in-memory layer
-    pub fn pop_oldest_open(&mut self) {
-        // Pop it from the binary heap
-        let oldest_entry = self.open_layers.pop().unwrap();
-        let segtag = oldest_entry.layer.get_seg_tag();
+    /// Remove an open in-memory layer
+    pub fn remove_open(&mut self, layer_id: LayerId) {
+        // Note: we don't try to remove the entry from the binary heap.
+        // It will be removed lazily by peek_oldest_open() when it's made it to
+        // the top of the heap.
 
-        // Also remove it from the SegEntry of this segment
-        let mut segentry = self.segs.get_mut(&segtag).unwrap();
-        if Arc::ptr_eq(segentry.open.as_ref().unwrap(), &oldest_entry.layer) {
-            segentry.open = None;
-        } else {
-            // We could have already updated segentry.open for
-            // dropped (non-writeable) layer. This is fine.
-            assert!(!oldest_entry.layer.is_writeable());
-            assert!(oldest_entry.layer.is_dropped());
+        let layer_opt = {
+            let mut global_map = GLOBAL_LAYER_MAP.write().unwrap();
+            let layer_opt = global_map.get(&layer_id);
+            global_map.remove(&layer_id);
+            // TODO it's bad that a ref can still exist after being evicted from cache
+            layer_opt
+        };
+
+        if let Some(layer) = layer_opt {
+            let mut segentry = self.segs.get_mut(&layer.get_seg_tag()).unwrap();
+
+            if segentry.open_layer_id == Some(layer_id) {
+                // Also remove it from the SegEntry of this segment
+                segentry.open_layer_id = None;
+            } else {
+                // We could have already updated segentry.open for
+                // dropped (non-writeable) layer. This is fine.
+                assert!(!layer.is_writeable());
+                assert!(layer.is_dropped());
+            }
+
+            NUM_INMEMORY_LAYERS.dec();
         }
-
-        NUM_INMEMORY_LAYERS.dec();
     }
 
     ///
@@ -199,10 +214,17 @@ impl LayerMap {
     }
 
     /// Return the oldest in-memory layer, along with its generation number.
-    pub fn peek_oldest_open(&self) -> Option<(Arc<InMemoryLayer>, u64)> {
-        self.open_layers
-            .peek()
-            .map(|oldest_entry| (Arc::clone(&oldest_entry.layer), oldest_entry.generation))
+    pub fn peek_oldest_open(&mut self) -> Option<(LayerId, Arc<InMemoryLayer>, u64)> {
+        let global_map = GLOBAL_LAYER_MAP.read().unwrap();
+
+        while let Some(oldest_entry) = self.open_layers.peek() {
+            if let Some(layer) = global_map.get(&oldest_entry.layer_id) {
+                return Some((oldest_entry.layer_id, layer, oldest_entry.generation));
+            } else {
+                self.open_layers.pop();
+            }
+        }
+        None
     }
 
     /// Increment the generation number used to stamp open in-memory layers. Layers
@@ -225,8 +247,12 @@ impl LayerMap {
     pub fn dump(&self) -> Result<()> {
         println!("Begin dump LayerMap");
         for (seg, segentry) in self.segs.iter() {
-            if let Some(open) = &segentry.open {
-                open.dump()?;
+            if let Some(open) = &segentry.open_layer_id {
+                if let Some(layer) = GLOBAL_LAYER_MAP.read().unwrap().get(open) {
+                    layer.dump()?;
+                } else {
+                    println!("layer not found in global map");
+                }
             }
 
             for layer in segentry.historic.iter() {
@@ -259,7 +285,7 @@ impl IntervalItem for dyn Layer {
 /// IntervalTree.
 #[derive(Default)]
 struct SegEntry {
-    open: Option<Arc<InMemoryLayer>>,
+    open_layer_id: Option<LayerId>,
     historic: IntervalTree<dyn Layer>,
 }
 
@@ -275,10 +301,10 @@ impl SegEntry {
     }
 
     pub fn get(&self, lsn: Lsn) -> Option<Arc<dyn Layer>> {
-        if let Some(open) = &self.open {
-            if open.get_start_lsn() <= lsn {
-                let x: Arc<dyn Layer> = Arc::clone(open) as _;
-                return Some(x);
+        if let Some(open_layer_id) = &self.open_layer_id {
+            let open_layer = GLOBAL_LAYER_MAP.read().unwrap().get(open_layer_id)?;
+            if open_layer.get_start_lsn() <= lsn {
+                return Some(open_layer);
             }
         }
 
@@ -297,11 +323,16 @@ impl SegEntry {
     // Set new open layer for a SegEntry.
     // It's ok to rewrite previous open layer,
     // but only if it is not writeable anymore.
-    pub fn update_open(&mut self, layer: Arc<InMemoryLayer>) {
-        if let Some(prev_open) = &self.open {
-            assert!(!prev_open.is_writeable());
+    pub fn update_open(&mut self, layer: Arc<InMemoryLayer>) -> LayerId {
+        if let Some(prev_open_layer_id) = &self.open_layer_id {
+            if let Some(prev_open_layer) = GLOBAL_LAYER_MAP.read().unwrap().get(prev_open_layer_id)
+            {
+                assert!(!prev_open_layer.is_writeable());
+            }
         }
-        self.open = Some(layer);
+        let open_layer_id = GLOBAL_LAYER_MAP.write().unwrap().insert(layer);
+        self.open_layer_id = Some(open_layer_id);
+        open_layer_id
     }
 
     pub fn insert_historic(&mut self, layer: Arc<dyn Layer>) {
@@ -316,9 +347,9 @@ impl SegEntry {
 /// recently-added entries (i.e after last call to increment_generation()) from older
 /// entries with the same 'oldest_pending_lsn'.
 struct OpenLayerEntry {
-    pub oldest_pending_lsn: Lsn, // copy of layer.get_oldest_pending_lsn()
-    pub generation: u64,
-    pub layer: Arc<InMemoryLayer>,
+    oldest_pending_lsn: Lsn, // copy of layer.get_oldest_pending_lsn()
+    generation: u64,
+    layer_id: LayerId,
 }
 impl Ord for OpenLayerEntry {
     fn cmp(&self, other: &Self) -> Ordering {
@@ -383,6 +414,13 @@ mod tests {
         forknum: 0,
     });
 
+    lazy_static! {
+        static ref DUMMY_TIMELINEID: ZTimelineId =
+            ZTimelineId::from_str("00000000000000000000000000000000").unwrap();
+        static ref DUMMY_TENANTID: ZTenantId =
+            ZTenantId::from_str("00000000000000000000000000000000").unwrap();
+    }
+
     /// Construct a dummy InMemoryLayer for testing
     fn dummy_inmem_layer(
         conf: &'static PageServerConf,
@@ -393,8 +431,8 @@ mod tests {
         Arc::new(
             InMemoryLayer::create(
                 conf,
-                ZTimelineId::from_str("00000000000000000000000000000000").unwrap(),
-                ZTenantId::from_str("00000000000000000000000000000000").unwrap(),
+                *DUMMY_TIMELINEID,
+                *DUMMY_TENANTID,
                 SegmentTag {
                     rel: TESTREL_A,
                     segno,
@@ -410,6 +448,7 @@ mod tests {
     fn test_open_layers() -> Result<()> {
         let conf = PageServerConf::dummy_conf(PageServerConf::test_repo_dir("dummy_inmem_layer"));
         let conf = Box::leak(Box::new(conf));
+        std::fs::create_dir_all(conf.timeline_path(&DUMMY_TIMELINEID, &DUMMY_TENANTID))?;
 
         let mut layers = LayerMap::default();
 
@@ -426,10 +465,10 @@ mod tests {
         // A helper function (closure) to pop the next oldest open entry from the layer map,
         // and assert that it is what we'd expect
         let mut assert_pop_layer = |expected_segno: u32, expected_generation: u64| {
-            let (l, generation) = layers.peek_oldest_open().unwrap();
+            let (layer_id, l, generation) = layers.peek_oldest_open().unwrap();
             assert!(l.get_seg_tag().segno == expected_segno);
             assert!(generation == expected_generation);
-            layers.pop_oldest_open();
+            layers.remove_open(layer_id);
         };
 
         assert_pop_layer(0, gen1); // 0x100

pageserver/src/layered_repository/metadata.rs

@@ -0,0 +1,226 @@
+//! Every image of a certain timeline from [`crate::layered_repository::LayeredRepository`]
+//! has a metadata that needs to be stored persistently.
+//!
+//! Later, the file gets is used in [`crate::remote_storage::storage_sync`] as a part of
+//! external storage import and export operations.
+//!
+//! The module contains all structs and related helper methods related to timeline metadata.
+
+use std::{convert::TryInto, path::PathBuf};
+
+use anyhow::ensure;
+use zenith_utils::{
+    bin_ser::BeSer,
+    lsn::Lsn,
+    zid::{ZTenantId, ZTimelineId},
+};
+
+use crate::{
+    layered_repository::{METADATA_CHECKSUM_SIZE, METADATA_MAX_DATA_SIZE, METADATA_MAX_SAFE_SIZE},
+    PageServerConf,
+};
+
+/// The name of the metadata file pageserver creates per timeline.
+pub const METADATA_FILE_NAME: &str = "metadata";
+
+/// Metadata stored on disk for each timeline
+///
+/// The fields correspond to the values we hold in memory, in LayeredTimeline.
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
+pub struct TimelineMetadata {
+    disk_consistent_lsn: Lsn,
+    // This is only set if we know it. We track it in memory when the page
+    // server is running, but we only track the value corresponding to
+    // 'last_record_lsn', not 'disk_consistent_lsn' which can lag behind by a
+    // lot. We only store it in the metadata file when we flush *all* the
+    // in-memory data so that 'last_record_lsn' is the same as
+    // 'disk_consistent_lsn'.  That's OK, because after page server restart, as
+    // soon as we reprocess at least one record, we will have a valid
+    // 'prev_record_lsn' value in memory again. This is only really needed when
+    // doing a clean shutdown, so that there is no more WAL beyond
+    // 'disk_consistent_lsn'
+    prev_record_lsn: Option<Lsn>,
+    ancestor_timeline: Option<ZTimelineId>,
+    ancestor_lsn: Lsn,
+    latest_gc_cutoff_lsn: Lsn,
+    initdb_lsn: Lsn,
+}
+
+/// Points to a place in pageserver's local directory,
+/// where certain timeline's metadata file should be located.
+pub fn metadata_path(
+    conf: &'static PageServerConf,
+    timelineid: ZTimelineId,
+    tenantid: ZTenantId,
+) -> PathBuf {
+    conf.timeline_path(&timelineid, &tenantid)
+        .join(METADATA_FILE_NAME)
+}
+
+impl TimelineMetadata {
+    pub fn new(
+        disk_consistent_lsn: Lsn,
+        prev_record_lsn: Option<Lsn>,
+        ancestor_timeline: Option<ZTimelineId>,
+        ancestor_lsn: Lsn,
+        latest_gc_cutoff_lsn: Lsn,
+        initdb_lsn: Lsn,
+    ) -> Self {
+        Self {
+            disk_consistent_lsn,
+            prev_record_lsn,
+            ancestor_timeline,
+            ancestor_lsn,
+            latest_gc_cutoff_lsn,
+            initdb_lsn,
+        }
+    }
+
+    pub fn from_bytes(metadata_bytes: &[u8]) -> anyhow::Result<Self> {
+        ensure!(
+            metadata_bytes.len() == METADATA_MAX_SAFE_SIZE,
+            "metadata bytes size is wrong"
+        );
+
+        let data = &metadata_bytes[..METADATA_MAX_DATA_SIZE];
+        let calculated_checksum = crc32c::crc32c(data);
+
+        let checksum_bytes: &[u8; METADATA_CHECKSUM_SIZE] =
+            metadata_bytes[METADATA_MAX_DATA_SIZE..].try_into()?;
+        let expected_checksum = u32::from_le_bytes(*checksum_bytes);
+        ensure!(
+            calculated_checksum == expected_checksum,
+            "metadata checksum mismatch"
+        );
+
+        let data = TimelineMetadata::from(serialize::DeTimelineMetadata::des_prefix(data)?);
+        assert!(data.disk_consistent_lsn.is_aligned());
+
+        Ok(data)
+    }
+
+    pub fn to_bytes(&self) -> anyhow::Result<Vec<u8>> {
+        let serializeable_metadata = serialize::SeTimelineMetadata::from(self);
+        let mut metadata_bytes = serialize::SeTimelineMetadata::ser(&serializeable_metadata)?;
+        assert!(metadata_bytes.len() <= METADATA_MAX_DATA_SIZE);
+        metadata_bytes.resize(METADATA_MAX_SAFE_SIZE, 0u8);
+
+        let checksum = crc32c::crc32c(&metadata_bytes[..METADATA_MAX_DATA_SIZE]);
+        metadata_bytes[METADATA_MAX_DATA_SIZE..].copy_from_slice(&u32::to_le_bytes(checksum));
+        Ok(metadata_bytes)
+    }
+
+    /// [`Lsn`] that corresponds to the corresponding timeline directory
+    /// contents, stored locally in the pageserver workdir.
+    pub fn disk_consistent_lsn(&self) -> Lsn {
+        self.disk_consistent_lsn
+    }
+
+    pub fn prev_record_lsn(&self) -> Option<Lsn> {
+        self.prev_record_lsn
+    }
+
+    pub fn ancestor_timeline(&self) -> Option<ZTimelineId> {
+        self.ancestor_timeline
+    }
+
+    pub fn ancestor_lsn(&self) -> Lsn {
+        self.ancestor_lsn
+    }
+
+    pub fn latest_gc_cutoff_lsn(&self) -> Lsn {
+        self.latest_gc_cutoff_lsn
+    }
+
+    pub fn initdb_lsn(&self) -> Lsn {
+        self.initdb_lsn
+    }
+}
+
+/// This module is for direct conversion of metadata to bytes and back.
+/// For a certain metadata, besides the conversion a few verification steps has to
+/// be done, so all serde derives are hidden from the user, to avoid accidental
+/// verification-less metadata creation.
+mod serialize {
+    use serde::{Deserialize, Serialize};
+    use zenith_utils::{lsn::Lsn, zid::ZTimelineId};
+
+    use super::TimelineMetadata;
+
+    #[derive(Serialize)]
+    pub(super) struct SeTimelineMetadata<'a> {
+        disk_consistent_lsn: &'a Lsn,
+        prev_record_lsn: &'a Option<Lsn>,
+        ancestor_timeline: &'a Option<ZTimelineId>,
+        ancestor_lsn: &'a Lsn,
+        latest_gc_cutoff_lsn: &'a Lsn,
+        initdb_lsn: &'a Lsn,
+    }
+
+    impl<'a> From<&'a TimelineMetadata> for SeTimelineMetadata<'a> {
+        fn from(other: &'a TimelineMetadata) -> Self {
+            Self {
+                disk_consistent_lsn: &other.disk_consistent_lsn,
+                prev_record_lsn: &other.prev_record_lsn,
+                ancestor_timeline: &other.ancestor_timeline,
+                ancestor_lsn: &other.ancestor_lsn,
+                latest_gc_cutoff_lsn: &other.latest_gc_cutoff_lsn,
+                initdb_lsn: &other.initdb_lsn,
+            }
+        }
+    }
+
+    #[derive(Deserialize)]
+    pub(super) struct DeTimelineMetadata {
+        disk_consistent_lsn: Lsn,
+        prev_record_lsn: Option<Lsn>,
+        ancestor_timeline: Option<ZTimelineId>,
+        ancestor_lsn: Lsn,
+        latest_gc_cutoff_lsn: Lsn,
+        initdb_lsn: Lsn,
+    }
+
+    impl From<DeTimelineMetadata> for TimelineMetadata {
+        fn from(other: DeTimelineMetadata) -> Self {
+            Self {
+                disk_consistent_lsn: other.disk_consistent_lsn,
+                prev_record_lsn: other.prev_record_lsn,
+                ancestor_timeline: other.ancestor_timeline,
+                ancestor_lsn: other.ancestor_lsn,
+                latest_gc_cutoff_lsn: other.latest_gc_cutoff_lsn,
+                initdb_lsn: other.initdb_lsn,
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::repository::repo_harness::TIMELINE_ID;
+
+    use super::*;
+
+    #[test]
+    fn metadata_serializes_correctly() {
+        let original_metadata = TimelineMetadata {
+            disk_consistent_lsn: Lsn(0x200),
+            prev_record_lsn: Some(Lsn(0x100)),
+            ancestor_timeline: Some(TIMELINE_ID),
+            ancestor_lsn: Lsn(0),
+            latest_gc_cutoff_lsn: Lsn(0),
+            initdb_lsn: Lsn(0),
+        };
+
+        let metadata_bytes = original_metadata
+            .to_bytes()
+            .expect("Should serialize correct metadata to bytes");
+
+        let deserialized_metadata = TimelineMetadata::from_bytes(&metadata_bytes)
+            .expect("Should deserialize its own bytes");
+
+        assert_eq!(
+            deserialized_metadata, original_metadata,
+            "Metadata that was serialized to bytes and deserialized back should not change"
+        );
+    }
+}

pageserver/src/layered_repository/page_versions.rs

@@ -0,0 +1,252 @@
+//!
+//! Data structure to ingest incoming WAL into an append-only file.
+//!
+//! - The file is considered temporary, and will be discarded on crash
+//! - based on a B-tree
+//!
+
+use std::os::unix::fs::FileExt;
+use std::{collections::HashMap, ops::RangeBounds, slice};
+
+use anyhow::Result;
+
+use std::cmp::min;
+use std::io::Seek;
+
+use zenith_utils::{lsn::Lsn, vec_map::VecMap};
+
+use super::storage_layer::PageVersion;
+use crate::layered_repository::ephemeral_file::EphemeralFile;
+
+use zenith_utils::bin_ser::BeSer;
+
+const EMPTY_SLICE: &[(Lsn, u64)] = &[];
+
+pub struct PageVersions {
+    map: HashMap<u32, VecMap<Lsn, u64>>,
+
+    /// The PageVersion structs are stored in a serialized format in this file.
+    /// Each serialized PageVersion is preceded by a 'u32' length field.
+    /// The 'map' stores offsets into this file.
+    file: EphemeralFile,
+}
+
+impl PageVersions {
+    pub fn new(file: EphemeralFile) -> PageVersions {
+        PageVersions {
+            map: HashMap::new(),
+            file,
+        }
+    }
+
+    pub fn append_or_update_last(
+        &mut self,
+        blknum: u32,
+        lsn: Lsn,
+        page_version: PageVersion,
+    ) -> Result<Option<u64>> {
+        // remember starting position
+        let pos = self.file.stream_position()?;
+
+        // make room for the 'length' field by writing zeros as a placeholder.
+        self.file.seek(std::io::SeekFrom::Start(pos + 4)).unwrap();
+
+        page_version.ser_into(&mut self.file).unwrap();
+
+        // write the 'length' field.
+        let len = self.file.stream_position()? - pos - 4;
+        let lenbuf = u32::to_ne_bytes(len as u32);
+        self.file.write_all_at(&lenbuf, pos)?;
+
+        let map = self.map.entry(blknum).or_insert_with(VecMap::default);
+        Ok(map.append_or_update_last(lsn, pos as u64).unwrap().0)
+    }
+
+    /// Get all [`PageVersion`]s in a block
+    fn get_block_slice(&self, blknum: u32) -> &[(Lsn, u64)] {
+        self.map
+            .get(&blknum)
+            .map(VecMap::as_slice)
+            .unwrap_or(EMPTY_SLICE)
+    }
+
+    /// Get a range of [`PageVersions`] in a block
+    pub fn get_block_lsn_range<R: RangeBounds<Lsn>>(&self, blknum: u32, range: R) -> &[(Lsn, u64)] {
+        self.map
+            .get(&blknum)
+            .map(|vec_map| vec_map.slice_range(range))
+            .unwrap_or(EMPTY_SLICE)
+    }
+
+    /// Iterate through [`PageVersion`]s in (block, lsn) order.
+    /// If a [`cutoff_lsn`] is set, only show versions with `lsn < cutoff_lsn`
+    pub fn ordered_page_version_iter(&self, cutoff_lsn: Option<Lsn>) -> OrderedPageVersionIter<'_> {
+        let mut ordered_blocks: Vec<u32> = self.map.keys().cloned().collect();
+        ordered_blocks.sort_unstable();
+
+        let slice = ordered_blocks
+            .first()
+            .map(|&blknum| self.get_block_slice(blknum))
+            .unwrap_or(EMPTY_SLICE);
+
+        OrderedPageVersionIter {
+            page_versions: self,
+            ordered_blocks,
+            cur_block_idx: 0,
+            cutoff_lsn,
+            cur_slice_iter: slice.iter(),
+        }
+    }
+
+    /// Returns a 'Read' that reads the page version at given offset.
+    pub fn reader(&self, pos: u64) -> Result<PageVersionReader, std::io::Error> {
+        // read length
+        let mut lenbuf = [0u8; 4];
+        self.file.read_exact_at(&mut lenbuf, pos)?;
+        let len = u32::from_ne_bytes(lenbuf);
+
+        Ok(PageVersionReader {
+            file: &self.file,
+            pos: pos + 4,
+            end_pos: pos + 4 + len as u64,
+        })
+    }
+
+    pub fn get_page_version(&self, pos: u64) -> Result<PageVersion> {
+        let mut reader = self.reader(pos)?;
+        Ok(PageVersion::des_from(&mut reader)?)
+    }
+}
+
+pub struct PageVersionReader<'a> {
+    file: &'a EphemeralFile,
+    pos: u64,
+    end_pos: u64,
+}
+
+impl<'a> std::io::Read for PageVersionReader<'a> {
+    fn read(&mut self, buf: &mut [u8]) -> Result<usize, std::io::Error> {
+        let len = min(buf.len(), (self.end_pos - self.pos) as usize);
+        let n = self.file.read_at(&mut buf[..len], self.pos)?;
+        self.pos += n as u64;
+        Ok(n)
+    }
+}
+
+pub struct OrderedPageVersionIter<'a> {
+    page_versions: &'a PageVersions,
+
+    ordered_blocks: Vec<u32>,
+    cur_block_idx: usize,
+
+    cutoff_lsn: Option<Lsn>,
+
+    cur_slice_iter: slice::Iter<'a, (Lsn, u64)>,
+}
+
+impl OrderedPageVersionIter<'_> {
+    fn is_lsn_before_cutoff(&self, lsn: &Lsn) -> bool {
+        if let Some(cutoff_lsn) = self.cutoff_lsn.as_ref() {
+            lsn < cutoff_lsn
+        } else {
+            true
+        }
+    }
+}
+
+impl<'a> Iterator for OrderedPageVersionIter<'a> {
+    type Item = (u32, Lsn, u64);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        loop {
+            if let Some((lsn, pos)) = self.cur_slice_iter.next() {
+                if self.is_lsn_before_cutoff(lsn) {
+                    let blknum = self.ordered_blocks[self.cur_block_idx];
+                    return Some((blknum, *lsn, *pos));
+                }
+            }
+
+            let next_block_idx = self.cur_block_idx + 1;
+            let blknum: u32 = *self.ordered_blocks.get(next_block_idx)?;
+            self.cur_block_idx = next_block_idx;
+            self.cur_slice_iter = self.page_versions.get_block_slice(blknum).iter();
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use bytes::Bytes;
+
+    use super::*;
+    use crate::PageServerConf;
+    use std::fs;
+    use std::str::FromStr;
+    use zenith_utils::zid::{ZTenantId, ZTimelineId};
+
+    fn repo_harness(test_name: &str) -> Result<(&'static PageServerConf, ZTenantId, ZTimelineId)> {
+        let repo_dir = PageServerConf::test_repo_dir(test_name);
+        let _ = fs::remove_dir_all(&repo_dir);
+        let conf = PageServerConf::dummy_conf(repo_dir);
+        // Make a static copy of the config. This can never be free'd, but that's
+        // OK in a test.
+        let conf: &'static PageServerConf = Box::leak(Box::new(conf));
+
+        let tenantid = ZTenantId::from_str("11000000000000000000000000000000").unwrap();
+        let timelineid = ZTimelineId::from_str("22000000000000000000000000000000").unwrap();
+        fs::create_dir_all(conf.timeline_path(&timelineid, &tenantid))?;
+
+        Ok((conf, tenantid, timelineid))
+    }
+
+    #[test]
+    fn test_ordered_iter() -> Result<()> {
+        let (conf, tenantid, timelineid) = repo_harness("test_ordered_iter")?;
+
+        let file = EphemeralFile::create(conf, tenantid, timelineid)?;
+
+        let mut page_versions = PageVersions::new(file);
+
+        const BLOCKS: u32 = 1000;
+        const LSNS: u64 = 50;
+
+        let empty_page = Bytes::from_static(&[0u8; 8192]);
+        let empty_page_version = PageVersion::Page(empty_page);
+
+        for blknum in 0..BLOCKS {
+            for lsn in 0..LSNS {
+                let old = page_versions.append_or_update_last(
+                    blknum,
+                    Lsn(lsn),
+                    empty_page_version.clone(),
+                )?;
+                assert!(old.is_none());
+            }
+        }
+
+        let mut iter = page_versions.ordered_page_version_iter(None);
+        for blknum in 0..BLOCKS {
+            for lsn in 0..LSNS {
+                let (actual_blknum, actual_lsn, _pv) = iter.next().unwrap();
+                assert_eq!(actual_blknum, blknum);
+                assert_eq!(Lsn(lsn), actual_lsn);
+            }
+        }
+        assert!(iter.next().is_none());
+        assert!(iter.next().is_none()); // should be robust against excessive next() calls
+
+        const CUTOFF_LSN: Lsn = Lsn(30);
+        let mut iter = page_versions.ordered_page_version_iter(Some(CUTOFF_LSN));
+        for blknum in 0..BLOCKS {
+            for lsn in 0..CUTOFF_LSN.0 {
+                let (actual_blknum, actual_lsn, _pv) = iter.next().unwrap();
+                assert_eq!(actual_blknum, blknum);
+                assert_eq!(Lsn(lsn), actual_lsn);
+            }
+        }
+        assert!(iter.next().is_none());
+        assert!(iter.next().is_none()); // should be robust against excessive next() calls
+
+        Ok(())
+    }
+}

pageserver/src/layered_repository/storage_layer.rs

@@ -4,7 +4,7 @@
 
 use crate::relish::RelishTag;
 use crate::repository::WALRecord;
-use crate::ZTimelineId;
+use crate::{ZTenantId, ZTimelineId};
 use anyhow::Result;
 use bytes::Bytes;
 use serde::{Deserialize, Serialize};
@@ -51,23 +51,10 @@ impl SegmentTag {
 ///
 /// A page version can be stored as a full page image, or as WAL record that needs
 /// to be applied over the previous page version to reconstruct this version.
-///
-/// It's also possible to have both a WAL record and a page image in the same
-/// PageVersion. That happens if page version is originally stored as a WAL record
-/// but it is later reconstructed by a GetPage@LSN request by performing WAL
-/// redo. The get_page_at_lsn() code will store the reconstructed pag image next to
-/// the WAL record in that case. TODO: That's pretty accidental, not the result
-/// of any grand design. If we want to keep reconstructed page versions around, we
-/// probably should have a separate buffer cache so that we could control the
-/// replacement policy globally. Or if we keep a reconstructed page image, we
-/// could throw away the WAL record.
-///
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct PageVersion {
-    /// an 8kb page image
-    pub page_image: Option<Bytes>,
-    /// WAL record to get from previous page version to this one.
-    pub record: Option<WALRecord>,
+pub enum PageVersion {
+    Page(Bytes),
+    Wal(WALRecord),
 }
 
 ///
@@ -78,7 +65,7 @@ pub struct PageVersion {
 /// 'records' contains the records to apply over the base image.
 ///
 pub struct PageReconstructData {
-    pub records: Vec<WALRecord>,
+    pub records: Vec<(Lsn, WALRecord)>,
     pub page_img: Option<Bytes>,
 }
 
@@ -93,6 +80,8 @@ pub enum PageReconstructResult {
     /// the returned LSN. This is usually considered an error, but might be OK
     /// in some circumstances.
     Missing(Lsn),
+    /// Use the cached image at `cached_img_lsn` as the base image
+    Cached,
 }
 
 ///
@@ -104,6 +93,8 @@ pub enum PageReconstructResult {
 /// in-memory and on-disk layers.
 ///
 pub trait Layer: Send + Sync {
+    fn get_tenant_id(&self) -> ZTenantId;
+
     /// Identify the timeline this relish belongs to
     fn get_timeline_id(&self) -> ZTimelineId;
 
@@ -123,10 +114,6 @@ pub trait Layer: Send + Sync {
     /// Is the segment represented by this layer dropped by PostgreSQL?
     fn is_dropped(&self) -> bool;
 
-    /// Gets the physical location of the layer on disk.
-    /// Some layers, such as in-memory, might not have the location.
-    fn path(&self) -> Option<PathBuf>;
-
     /// Filename used to store this layer on disk. (Even in-memory layers
     /// implement this, to print a handy unique identifier for the layer for
     /// log messages, even though they're never not on disk.)
@@ -142,6 +129,9 @@ pub trait Layer: Send + Sync {
     /// of the *relish*, not the beginning of the segment. The requested
     /// 'blknum' must be covered by this segment.
     ///
+    /// `cached_img_lsn` should be set to a cached page image's lsn < `lsn`.
+    /// This function will only return data after `cached_img_lsn`.
+    ///
     /// See PageReconstructResult for possible return values. The collected data
     /// is appended to reconstruct_data; the caller should pass an empty struct
     /// on first call. If this returns PageReconstructResult::Continue, look up
@@ -151,6 +141,7 @@ pub trait Layer: Send + Sync {
         &self,
         blknum: u32,
         lsn: Lsn,
+        cached_img_lsn: Option<Lsn>,
         reconstruct_data: &mut PageReconstructData,
     ) -> Result<PageReconstructResult>;
 
@@ -166,6 +157,9 @@ pub trait Layer: Send + Sync {
     /// the previous non-incremental layer.
     fn is_incremental(&self) -> bool;
 
+    /// Returns true for layers that are represented in memory.
+    fn is_in_memory(&self) -> bool;
+
     /// Release memory used by this layer. There is no corresponding 'load'
     /// function, that's done implicitly when you call one of the get-functions.
     fn unload(&self) -> Result<()>;

pageserver/src/lib.rs

@@ -1,6 +1,8 @@
+use layered_repository::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use zenith_utils::postgres_backend::AuthType;
 use zenith_utils::zid::{ZTenantId, ZTimelineId};
 
+use std::num::{NonZeroU32, NonZeroUsize};
 use std::path::PathBuf;
 use std::time::Duration;
 
@@ -11,12 +13,15 @@ pub mod basebackup;
 pub mod branches;
 pub mod http;
 pub mod layered_repository;
+pub mod page_cache;
 pub mod page_service;
 pub mod relish;
-pub mod relish_storage;
+pub mod remote_storage;
 pub mod repository;
 pub mod restore_local_repo;
 pub mod tenant_mgr;
+pub mod tenant_threads;
+pub mod virtual_file;
 pub mod waldecoder;
 pub mod walreceiver;
 pub mod walredo;
@@ -34,13 +39,18 @@ pub mod defaults {
     // would be more appropriate. But a low value forces the code to be exercised more,
     // which is good for now to trigger bugs.
     pub const DEFAULT_CHECKPOINT_DISTANCE: u64 = 256 * 1024 * 1024;
-    pub const DEFAULT_CHECKPOINT_PERIOD: Duration = Duration::from_secs(1);
+    pub const DEFAULT_CHECKPOINT_PERIOD: Duration = Duration::from_secs(10);
 
     pub const DEFAULT_GC_HORIZON: u64 = 64 * 1024 * 1024;
-    pub const DEFAULT_GC_PERIOD: Duration = Duration::from_secs(100);
+    pub const DEFAULT_GC_PERIOD: Duration = Duration::from_secs(10);
 
     pub const DEFAULT_SUPERUSER: &str = "zenith_admin";
-    pub const DEFAULT_RELISH_STORAGE_MAX_CONCURRENT_SYNC_LIMITS: usize = 100;
+    pub const DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNC: usize = 100;
+    pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
+
+    pub const DEFAULT_OPEN_MEM_LIMIT: usize = 128 * 1024 * 1024;
+    pub const DEFAULT_PAGE_CACHE_SIZE: usize = 8192;
+    pub const DEFAULT_MAX_FILE_DESCRIPTORS: usize = 100;
 }
 
 lazy_static! {
@@ -69,6 +79,10 @@ pub struct PageServerConf {
     pub gc_period: Duration,
     pub superuser: String,
 
+    pub open_mem_limit: usize,
+    pub page_cache_size: usize,
+    pub max_file_descriptors: usize,
+
     // Repository directory, relative to current working directory.
     // Normally, the page server changes the current working directory
     // to the repository, and 'workdir' is always '.'. But we don't do
@@ -82,7 +96,7 @@ pub struct PageServerConf {
     pub auth_type: AuthType,
 
     pub auth_validation_public_key_path: Option<PathBuf>,
-    pub relish_storage_config: Option<RelishStorageConfig>,
+    pub remote_storage_config: Option<RemoteStorageConfig>,
 }
 
 impl PageServerConf {
@@ -91,7 +105,7 @@ impl PageServerConf {
     //
 
     fn tenants_path(&self) -> PathBuf {
-        self.workdir.join("tenants")
+        self.workdir.join(TENANTS_SEGMENT_NAME)
     }
 
     fn tenant_path(&self, tenantid: &ZTenantId) -> PathBuf {
@@ -115,17 +129,13 @@ impl PageServerConf {
     }
 
     fn timelines_path(&self, tenantid: &ZTenantId) -> PathBuf {
-        self.tenant_path(tenantid).join("timelines")
+        self.tenant_path(tenantid).join(TIMELINES_SEGMENT_NAME)
     }
 
     fn timeline_path(&self, timelineid: &ZTimelineId, tenantid: &ZTenantId) -> PathBuf {
         self.timelines_path(tenantid).join(timelineid.to_string())
     }
 
-    fn ancestor_path(&self, timelineid: &ZTimelineId, tenantid: &ZTenantId) -> PathBuf {
-        self.timeline_path(timelineid, tenantid).join("ancestor")
-    }
-
     //
     // Postgres distribution paths
     //
@@ -151,6 +161,9 @@ impl PageServerConf {
             checkpoint_period: Duration::from_secs(10),
             gc_horizon: defaults::DEFAULT_GC_HORIZON,
             gc_period: Duration::from_secs(10),
+            open_mem_limit: defaults::DEFAULT_OPEN_MEM_LIMIT,
+            page_cache_size: defaults::DEFAULT_PAGE_CACHE_SIZE,
+            max_file_descriptors: defaults::DEFAULT_MAX_FILE_DESCRIPTORS,
             listen_pg_addr: defaults::DEFAULT_PG_LISTEN_ADDR.to_string(),
             listen_http_addr: defaults::DEFAULT_HTTP_LISTEN_ADDR.to_string(),
             superuser: "zenith_admin".to_string(),
@@ -158,23 +171,34 @@ impl PageServerConf {
             pg_distrib_dir: "".into(),
             auth_type: AuthType::Trust,
             auth_validation_public_key_path: None,
-            relish_storage_config: None,
+            remote_storage_config: None,
         }
     }
 }
 
-/// External relish storage configuration, enough for creating a client for that storage.
-#[derive(Debug, Clone)]
-pub struct RelishStorageConfig {
-    /// Limits the number of concurrent sync operations between pageserver and relish storage.
-    pub max_concurrent_sync: usize,
-    /// The storage connection configuration.
-    pub storage: RelishStorageKind,
+/// Config for the Repository checkpointer
+#[derive(Debug, Clone, Copy)]
+pub enum CheckpointConfig {
+    // Flush in-memory data that is older than this
+    Distance(u64),
+    // Flush all in-memory data
+    Forced,
 }
 
-/// A kind of a relish storage to connect to, with its connection configuration.
+/// External backup storage configuration, enough for creating a client for that storage.
 #[derive(Debug, Clone)]
-pub enum RelishStorageKind {
+pub struct RemoteStorageConfig {
+    /// Max allowed number of concurrent sync operations between pageserver and the remote storage.
+    pub max_concurrent_sync: NonZeroUsize,
+    /// Max allowed errors before the sync task is considered failed and evicted.
+    pub max_sync_errors: NonZeroU32,
+    /// The storage connection configuration.
+    pub storage: RemoteStorageKind,
+}
+
+/// A kind of a remote storage to connect to, with its connection configuration.
+#[derive(Debug, Clone)]
+pub enum RemoteStorageKind {
     /// Storage based on local file system.
     /// Specify a root folder to place all stored relish data into.
     LocalFs(PathBuf),

pageserver/src/page_cache.rs

@@ -0,0 +1,778 @@
+//!
+//! Global page cache
+//!
+//! The page cache uses up most of the memory in the page server. It is shared
+//! by all tenants, and it is used to store different kinds of pages. Sharing
+//! the cache allows memory to be dynamically allocated where it's needed the
+//! most.
+//!
+//! The page cache consists of fixed-size buffers, 8 kB each to match the
+//! PostgreSQL buffer size, and a Slot struct for each buffer to contain
+//! information about what's stored in the buffer.
+//!
+//! # Locking
+//!
+//! There are two levels of locking involved: There's one lock for the "mapping"
+//! from page identifier (tenant ID, timeline ID, rel, block, LSN) to the buffer
+//! slot, and a separate lock on each slot. To read or write the contents of a
+//! slot, you must hold the lock on the slot in read or write mode,
+//! respectively. To change the mapping of a slot, i.e. to evict a page or to
+//! assign a buffer for a page, you must hold the mapping lock and the lock on
+//! the slot at the same time.
+//!
+//! Whenever you need to hold both locks simultenously, the slot lock must be
+//! acquired first. This consistent ordering avoids deadlocks. To look up a page
+//! in the cache, you would first look up the mapping, while holding the mapping
+//! lock, and then lock the slot. You must release the mapping lock in between,
+//! to obey the lock ordering and avoid deadlock.
+//!
+//! A slot can momentarily have invalid contents, even if it's already been
+//! inserted to the mapping, but you must hold the write-lock on the slot until
+//! the contents are valid. If you need to release the lock without initializing
+//! the contents, you must remove the mapping first. We make that easy for the
+//! callers with PageWriteGuard: when lock_for_write() returns an uninitialized
+//! page, the caller must explicitly call guard.mark_valid() after it has
+//! initialized it. If the guard is dropped without calling mark_valid(), the
+//! mapping is automatically removed and the slot is marked free.
+//!
+
+use std::{
+    collections::{hash_map::Entry, HashMap},
+    convert::TryInto,
+    sync::{
+        atomic::{AtomicU8, AtomicUsize, Ordering},
+        RwLock, RwLockReadGuard, RwLockWriteGuard,
+    },
+};
+
+use once_cell::sync::OnceCell;
+use tracing::error;
+use zenith_utils::{
+    lsn::Lsn,
+    zid::{ZTenantId, ZTimelineId},
+};
+
+use crate::layered_repository::writeback_ephemeral_file;
+use crate::{relish::RelTag, PageServerConf};
+
+static PAGE_CACHE: OnceCell<PageCache> = OnceCell::new();
+const TEST_PAGE_CACHE_SIZE: usize = 10;
+
+///
+/// Initialize the page cache. This must be called once at page server startup.
+///
+pub fn init(conf: &'static PageServerConf) {
+    if PAGE_CACHE
+        .set(PageCache::new(conf.page_cache_size))
+        .is_err()
+    {
+        panic!("page cache already initialized");
+    }
+}
+
+///
+/// Get a handle to the page cache.
+///
+pub fn get() -> &'static PageCache {
+    //
+    // In unit tests, page server startup doesn't happen and no one calls
+    // page_cache::init(). Initialize it here with a tiny cache, so that the
+    // page cache is usable in unit tests.
+    //
+    if cfg!(test) {
+        PAGE_CACHE.get_or_init(|| PageCache::new(TEST_PAGE_CACHE_SIZE))
+    } else {
+        PAGE_CACHE.get().expect("page cache not initialized")
+    }
+}
+
+pub const PAGE_SZ: usize = postgres_ffi::pg_constants::BLCKSZ as usize;
+const MAX_USAGE_COUNT: u8 = 5;
+
+///
+/// CacheKey uniquely identifies a "thing" to cache in the page cache.
+///
+#[derive(Debug, PartialEq, Eq, Clone)]
+enum CacheKey {
+    MaterializedPage {
+        hash_key: MaterializedPageHashKey,
+        lsn: Lsn,
+    },
+    EphemeralPage {
+        file_id: u64,
+        blkno: u32,
+    },
+}
+
+#[derive(Debug, PartialEq, Eq, Hash, Clone)]
+struct MaterializedPageHashKey {
+    tenant_id: ZTenantId,
+    timeline_id: ZTimelineId,
+    rel_tag: RelTag,
+    blknum: u32,
+}
+
+#[derive(Clone)]
+struct Version {
+    lsn: Lsn,
+    slot_idx: usize,
+}
+
+struct Slot {
+    inner: RwLock<SlotInner>,
+    usage_count: AtomicU8,
+}
+
+struct SlotInner {
+    key: Option<CacheKey>,
+    buf: &'static mut [u8; PAGE_SZ],
+    dirty: bool,
+}
+
+impl Slot {
+    /// Increment usage count on the buffer, with ceiling at MAX_USAGE_COUNT.
+    fn inc_usage_count(&self) {
+        let _ = self
+            .usage_count
+            .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |val| {
+                if val == MAX_USAGE_COUNT {
+                    None
+                } else {
+                    Some(val + 1)
+                }
+            });
+    }
+
+    /// Decrement usage count on the buffer, unless it's already zero.  Returns
+    /// the old usage count.
+    fn dec_usage_count(&self) -> u8 {
+        let count_res =
+            self.usage_count
+                .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |val| {
+                    if val == 0 {
+                        None
+                    } else {
+                        Some(val - 1)
+                    }
+                });
+
+        match count_res {
+            Ok(usage_count) => usage_count,
+            Err(usage_count) => usage_count,
+        }
+    }
+}
+
+pub struct PageCache {
+    /// This contains the mapping from the cache key to buffer slot that currently
+    /// contains the page, if any.
+    ///
+    /// TODO: This is protected by a single lock. If that becomes a bottleneck,
+    /// this HashMap can be replaced with a more concurrent version, there are
+    /// plenty of such crates around.
+    ///
+    /// If you add support for caching different kinds of objects, each object kind
+    /// can have a separate mapping map, next to this field.
+    materialized_page_map: RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,
+
+    ephemeral_page_map: RwLock<HashMap<(u64, u32), usize>>,
+
+    /// The actual buffers with their metadata.
+    slots: Box<[Slot]>,
+
+    /// Index of the next candidate to evict, for the Clock replacement algorithm.
+    /// This is interpreted modulo the page cache size.
+    next_evict_slot: AtomicUsize,
+}
+
+///
+/// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
+/// until the guard is dropped.
+///
+pub struct PageReadGuard<'i>(RwLockReadGuard<'i, SlotInner>);
+
+impl std::ops::Deref for PageReadGuard<'_> {
+    type Target = [u8; PAGE_SZ];
+
+    fn deref(&self) -> &Self::Target {
+        self.0.buf
+    }
+}
+
+///
+/// PageWriteGuard is a lease on a buffer for modifying it. The page is kept locked
+/// until the guard is dropped.
+///
+/// Counterintuitively, this is used even for a read, if the requested page is not
+/// currently found in the page cache. In that case, the caller of lock_for_read()
+/// is expected to fill in the page contents and call mark_valid(). Similarly
+/// lock_for_write() can return an invalid buffer that the caller is expected to
+/// to initialize.
+///
+pub struct PageWriteGuard<'i> {
+    inner: RwLockWriteGuard<'i, SlotInner>,
+
+    // Are the page contents currently valid?
+    valid: bool,
+}
+
+impl std::ops::DerefMut for PageWriteGuard<'_> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.inner.buf
+    }
+}
+
+impl std::ops::Deref for PageWriteGuard<'_> {
+    type Target = [u8; PAGE_SZ];
+
+    fn deref(&self) -> &Self::Target {
+        self.inner.buf
+    }
+}
+
+impl PageWriteGuard<'_> {
+    /// Mark that the buffer contents are now valid.
+    pub fn mark_valid(&mut self) {
+        assert!(self.inner.key.is_some());
+        assert!(
+            !self.valid,
+            "mark_valid called on a buffer that was already valid"
+        );
+        self.valid = true;
+    }
+    pub fn mark_dirty(&mut self) {
+        // only ephemeral pages can be dirty ATM.
+        assert!(matches!(
+            self.inner.key,
+            Some(CacheKey::EphemeralPage { .. })
+        ));
+        self.inner.dirty = true;
+    }
+}
+
+impl Drop for PageWriteGuard<'_> {
+    ///
+    /// If the buffer was allocated for a page that was not already in the
+    /// cache, but the lock_for_read/write() caller dropped the buffer without
+    /// initializing it, remove the mapping from the page cache.
+    ///
+    fn drop(&mut self) {
+        assert!(self.inner.key.is_some());
+        if !self.valid {
+            let self_key = self.inner.key.as_ref().unwrap();
+            PAGE_CACHE.get().unwrap().remove_mapping(self_key);
+            self.inner.key = None;
+            self.inner.dirty = false;
+        }
+    }
+}
+
+/// lock_for_read() return value
+pub enum ReadBufResult<'a> {
+    Found(PageReadGuard<'a>),
+    NotFound(PageWriteGuard<'a>),
+}
+
+/// lock_for_write() return value
+pub enum WriteBufResult<'a> {
+    Found(PageWriteGuard<'a>),
+    NotFound(PageWriteGuard<'a>),
+}
+
+impl PageCache {
+    //
+    // Section 1.1: Public interface functions for looking up and memorizing materialized page
+    // versions in the page cache
+    //
+
+    /// Look up a materialized page version.
+    ///
+    /// The 'lsn' is an upper bound, this will return the latest version of
+    /// the given block, but not newer than 'lsn'. Returns the actual LSN of the
+    /// returned page.
+    pub fn lookup_materialized_page(
+        &self,
+        tenant_id: ZTenantId,
+        timeline_id: ZTimelineId,
+        rel_tag: RelTag,
+        blknum: u32,
+        lsn: Lsn,
+    ) -> Option<(Lsn, PageReadGuard)> {
+        let mut cache_key = CacheKey::MaterializedPage {
+            hash_key: MaterializedPageHashKey {
+                tenant_id,
+                timeline_id,
+                rel_tag,
+                blknum,
+            },
+            lsn,
+        };
+
+        if let Some(guard) = self.try_lock_for_read(&mut cache_key) {
+            if let CacheKey::MaterializedPage { hash_key: _, lsn } = cache_key {
+                Some((lsn, guard))
+            } else {
+                panic!("unexpected key type in slot");
+            }
+        } else {
+            None
+        }
+    }
+
+    ///
+    /// Store an image of the given page in the cache.
+    ///
+    pub fn memorize_materialized_page(
+        &self,
+        tenant_id: ZTenantId,
+        timeline_id: ZTimelineId,
+        rel_tag: RelTag,
+        blknum: u32,
+        lsn: Lsn,
+        img: &[u8],
+    ) {
+        let cache_key = CacheKey::MaterializedPage {
+            hash_key: MaterializedPageHashKey {
+                tenant_id,
+                timeline_id,
+                rel_tag,
+                blknum,
+            },
+            lsn,
+        };
+
+        match self.lock_for_write(&cache_key) {
+            WriteBufResult::Found(write_guard) => {
+                // We already had it in cache. Another thread must've put it there
+                // concurrently. Check that it had the same contents that we
+                // replayed.
+                assert!(*write_guard == img);
+            }
+            WriteBufResult::NotFound(mut write_guard) => {
+                write_guard.copy_from_slice(img);
+                write_guard.mark_valid();
+            }
+        }
+    }
+
+    // Section 1.2: Public interface functions for working with Ephemeral pages.
+
+    pub fn read_ephemeral_buf(&self, file_id: u64, blkno: u32) -> ReadBufResult {
+        let mut cache_key = CacheKey::EphemeralPage { file_id, blkno };
+
+        self.lock_for_read(&mut cache_key)
+    }
+
+    pub fn write_ephemeral_buf(&self, file_id: u64, blkno: u32) -> WriteBufResult {
+        let cache_key = CacheKey::EphemeralPage { file_id, blkno };
+
+        self.lock_for_write(&cache_key)
+    }
+
+    /// Immediately drop all buffers belonging to given file, without writeback
+    pub fn drop_buffers_for_ephemeral(&self, drop_file_id: u64) {
+        for slot_idx in 0..self.slots.len() {
+            let slot = &self.slots[slot_idx];
+
+            let mut inner = slot.inner.write().unwrap();
+            if let Some(key) = &inner.key {
+                match key {
+                    CacheKey::EphemeralPage { file_id, blkno: _ } if *file_id == drop_file_id => {
+                        // remove mapping for old buffer
+                        self.remove_mapping(key);
+                        inner.key = None;
+                        inner.dirty = false;
+                    }
+                    _ => {}
+                }
+            }
+        }
+    }
+
+    //
+    // Section 2: Internal interface functions for lookup/update.
+    //
+    // To add support for a new kind of "thing" to cache, you will need
+    // to add public interface routines above, and code to deal with the
+    // "mappings" after this section. But the routines in this section should
+    // not require changes.
+
+    /// Look up a page in the cache.
+    ///
+    /// If the search criteria is not exact, *cache_key is updated with the key
+    /// for exact key of the returned page. (For materialized pages, that means
+    /// that the LSN in 'cache_key' is updated with the LSN of the returned page
+    /// version.)
+    ///
+    /// If no page is found, returns None and *cache_key is left unmodified.
+    ///
+    fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
+        let cache_key_orig = cache_key.clone();
+        if let Some(slot_idx) = self.search_mapping(cache_key) {
+            // The page was found in the mapping. Lock the slot, and re-check
+            // that it's still what we expected (because we released the mapping
+            // lock already, another thread could have evicted the page)
+            let slot = &self.slots[slot_idx];
+            let inner = slot.inner.read().unwrap();
+            if inner.key.as_ref() == Some(cache_key) {
+                slot.inc_usage_count();
+                return Some(PageReadGuard(inner));
+            } else {
+                // search_mapping might have modified the search key; restore it.
+                *cache_key = cache_key_orig;
+            }
+        }
+        None
+    }
+
+    /// Return a locked buffer for given block.
+    ///
+    /// Like try_lock_for_read(), if the search criteria is not exact and the
+    /// page is already found in the cache, *cache_key is updated.
+    ///
+    /// If the page is not found in the cache, this allocates a new buffer for
+    /// it. The caller may then initialize the buffer with the contents, and
+    /// call mark_valid().
+    ///
+    /// Example usage:
+    ///
+    /// ```ignore
+    /// let cache = page_cache::get();
+    ///
+    /// match cache.lock_for_read(&key) {
+    ///     ReadBufResult::Found(read_guard) => {
+    ///         // The page was found in cache. Use it
+    ///     },
+    ///     ReadBufResult::NotFound(write_guard) => {
+    ///         // The page was not found in cache. Read it from disk into the
+    ///         // buffer.
+    ///         //read_my_page_from_disk(write_guard);
+    ///
+    ///         // The buffer contents are now valid. Tell the page cache.
+    ///         write_guard.mark_valid();
+    ///     },
+    /// }
+    /// ```
+    ///
+    fn lock_for_read(&self, cache_key: &mut CacheKey) -> ReadBufResult {
+        loop {
+            // First check if the key already exists in the cache.
+            if let Some(read_guard) = self.try_lock_for_read(cache_key) {
+                return ReadBufResult::Found(read_guard);
+            }
+
+            // Not found. Find a victim buffer
+            let (slot_idx, mut inner) = self.find_victim();
+
+            // Insert mapping for this. At this point, we may find that another
+            // thread did the same thing concurrently. In that case, we evicted
+            // our victim buffer unnecessarily. Put it into the free list and
+            // continue with the slot that the other thread chose.
+            if let Some(_existing_slot_idx) = self.try_insert_mapping(cache_key, slot_idx) {
+                // TODO: put to free list
+
+                // We now just loop back to start from beginning. This is not
+                // optimal, we'll perform the lookup in the mapping again, which
+                // is not really necessary because we already got
+                // 'existing_slot_idx'.  But this shouldn't happen often enough
+                // to matter much.
+                continue;
+            }
+
+            // Make the slot ready
+            let slot = &self.slots[slot_idx];
+            inner.key = Some(cache_key.clone());
+            inner.dirty = false;
+            slot.usage_count.store(1, Ordering::Relaxed);
+
+            return ReadBufResult::NotFound(PageWriteGuard {
+                inner,
+                valid: false,
+            });
+        }
+    }
+
+    /// Look up a page in the cache and lock it in write mode. If it's not
+    /// found, returns None.
+    ///
+    /// When locking a page for writing, the search criteria is always "exact".
+    fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
+        if let Some(slot_idx) = self.search_mapping_for_write(cache_key) {
+            // The page was found in the mapping. Lock the slot, and re-check
+            // that it's still what we expected (because we don't released the mapping
+            // lock already, another thread could have evicted the page)
+            let slot = &self.slots[slot_idx];
+            let inner = slot.inner.write().unwrap();
+            if inner.key.as_ref() == Some(cache_key) {
+                slot.inc_usage_count();
+                return Some(PageWriteGuard { inner, valid: true });
+            }
+        }
+        None
+    }
+
+    /// Return a write-locked buffer for given block.
+    ///
+    /// Similar to lock_for_read(), but the returned buffer is write-locked and
+    /// may be modified by the caller even if it's already found in the cache.
+    fn lock_for_write(&self, cache_key: &CacheKey) -> WriteBufResult {
+        loop {
+            // First check if the key already exists in the cache.
+            if let Some(write_guard) = self.try_lock_for_write(cache_key) {
+                return WriteBufResult::Found(write_guard);
+            }
+
+            // Not found. Find a victim buffer
+            let (slot_idx, mut inner) = self.find_victim();
+
+            // Insert mapping for this. At this point, we may find that another
+            // thread did the same thing concurrently. In that case, we evicted
+            // our victim buffer unnecessarily. Put it into the free list and
+            // continue with the slot that the other thread chose.
+            if let Some(_existing_slot_idx) = self.try_insert_mapping(cache_key, slot_idx) {
+                // TODO: put to free list
+
+                // We now just loop back to start from beginning. This is not
+                // optimal, we'll perform the lookup in the mapping again, which
+                // is not really necessary because we already got
+                // 'existing_slot_idx'.  But this shouldn't happen often enough
+                // to matter much.
+                continue;
+            }
+
+            // Make the slot ready
+            let slot = &self.slots[slot_idx];
+            inner.key = Some(cache_key.clone());
+            inner.dirty = false;
+            slot.usage_count.store(1, Ordering::Relaxed);
+
+            return WriteBufResult::NotFound(PageWriteGuard {
+                inner,
+                valid: false,
+            });
+        }
+    }
+
+    //
+    // Section 3: Mapping functions
+    //
+
+    /// Search for a page in the cache using the given search key.
+    ///
+    /// Returns the slot index, if any. If the search criteria is not exact,
+    /// *cache_key is updated with the actual key of the found page.
+    ///
+    /// NOTE: We don't hold any lock on the mapping on return, so the slot might
+    /// get recycled for an unrelated page immediately after this function
+    /// returns.  The caller is responsible for re-checking that the slot still
+    /// contains the page with the same key before using it.
+    ///
+    fn search_mapping(&self, cache_key: &mut CacheKey) -> Option<usize> {
+        match cache_key {
+            CacheKey::MaterializedPage { hash_key, lsn } => {
+                let map = self.materialized_page_map.read().unwrap();
+                let versions = map.get(hash_key)?;
+
+                let version_idx = match versions.binary_search_by_key(lsn, |v| v.lsn) {
+                    Ok(version_idx) => version_idx,
+                    Err(0) => return None,
+                    Err(version_idx) => version_idx - 1,
+                };
+                let version = &versions[version_idx];
+                *lsn = version.lsn;
+                Some(version.slot_idx)
+            }
+            CacheKey::EphemeralPage { file_id, blkno } => {
+                let map = self.ephemeral_page_map.read().unwrap();
+                Some(*map.get(&(*file_id, *blkno))?)
+            }
+        }
+    }
+
+    /// Search for a page in the cache using the given search key.
+    ///
+    /// Like 'search_mapping, but performs an "exact" search. Used for
+    /// allocating a new buffer.
+    fn search_mapping_for_write(&self, key: &CacheKey) -> Option<usize> {
+        match key {
+            CacheKey::MaterializedPage { hash_key, lsn } => {
+                let map = self.materialized_page_map.read().unwrap();
+                let versions = map.get(hash_key)?;
+
+                if let Ok(version_idx) = versions.binary_search_by_key(lsn, |v| v.lsn) {
+                    Some(versions[version_idx].slot_idx)
+                } else {
+                    None
+                }
+            }
+            CacheKey::EphemeralPage { file_id, blkno } => {
+                let map = self.ephemeral_page_map.read().unwrap();
+                Some(*map.get(&(*file_id, *blkno))?)
+            }
+        }
+    }
+
+    ///
+    /// Remove mapping for given key.
+    ///
+    fn remove_mapping(&self, old_key: &CacheKey) {
+        match old_key {
+            CacheKey::MaterializedPage {
+                hash_key: old_hash_key,
+                lsn: old_lsn,
+            } => {
+                let mut map = self.materialized_page_map.write().unwrap();
+                if let Entry::Occupied(mut old_entry) = map.entry(old_hash_key.clone()) {
+                    let versions = old_entry.get_mut();
+
+                    if let Ok(version_idx) = versions.binary_search_by_key(old_lsn, |v| v.lsn) {
+                        versions.remove(version_idx);
+                        if versions.is_empty() {
+                            old_entry.remove_entry();
+                        }
+                    }
+                } else {
+                    panic!("could not find old key in mapping")
+                }
+            }
+            CacheKey::EphemeralPage { file_id, blkno } => {
+                let mut map = self.ephemeral_page_map.write().unwrap();
+                map.remove(&(*file_id, *blkno))
+                    .expect("could not find old key in mapping");
+            }
+        }
+    }
+
+    ///
+    /// Insert mapping for given key.
+    ///
+    /// If a mapping already existed for the given key, returns the slot index
+    /// of the existing mapping and leaves it untouched.
+    fn try_insert_mapping(&self, new_key: &CacheKey, slot_idx: usize) -> Option<usize> {
+        match new_key {
+            CacheKey::MaterializedPage {
+                hash_key: new_key,
+                lsn: new_lsn,
+            } => {
+                let mut map = self.materialized_page_map.write().unwrap();
+                let versions = map.entry(new_key.clone()).or_default();
+                match versions.binary_search_by_key(new_lsn, |v| v.lsn) {
+                    Ok(version_idx) => Some(versions[version_idx].slot_idx),
+                    Err(version_idx) => {
+                        versions.insert(
+                            version_idx,
+                            Version {
+                                lsn: *new_lsn,
+                                slot_idx,
+                            },
+                        );
+                        None
+                    }
+                }
+            }
+            CacheKey::EphemeralPage { file_id, blkno } => {
+                let mut map = self.ephemeral_page_map.write().unwrap();
+                match map.entry((*file_id, *blkno)) {
+                    Entry::Occupied(entry) => Some(*entry.get()),
+                    Entry::Vacant(entry) => {
+                        entry.insert(slot_idx);
+                        None
+                    }
+                }
+            }
+        }
+    }
+
+    //
+    // Section 4: Misc internal helpers
+    //
+
+    /// Find a slot to evict.
+    ///
+    /// On return, the slot is empty and write-locked.
+    fn find_victim(&self) -> (usize, RwLockWriteGuard<SlotInner>) {
+        let iter_limit = self.slots.len() * 2;
+        let mut iters = 0;
+        loop {
+            let slot_idx = self.next_evict_slot.fetch_add(1, Ordering::Relaxed) % self.slots.len();
+
+            let slot = &self.slots[slot_idx];
+
+            if slot.dec_usage_count() == 0 || iters >= iter_limit {
+                let mut inner = slot.inner.write().unwrap();
+
+                if let Some(old_key) = &inner.key {
+                    if inner.dirty {
+                        if let Err(err) = Self::writeback(old_key, inner.buf) {
+                            // Writing the page to disk failed.
+                            //
+                            // FIXME: What to do here, when? We could propagate the error to the
+                            // caller, but victim buffer is generally unrelated to the original
+                            // call. It can even belong to a different tenant. Currently, we
+                            // report the error to the log and continue the clock sweep to find
+                            // a different victim. But if the problem persists, the page cache
+                            // could fill up with dirty pages that we cannot evict, and we will
+                            // loop retrying the writebacks indefinitely.
+                            error!("writeback of buffer {:?} failed: {}", old_key, err);
+                            continue;
+                        }
+                    }
+
+                    // remove mapping for old buffer
+                    self.remove_mapping(old_key);
+                    inner.dirty = false;
+                    inner.key = None;
+                }
+                return (slot_idx, inner);
+            }
+
+            iters += 1;
+        }
+    }
+
+    fn writeback(cache_key: &CacheKey, buf: &[u8]) -> Result<(), std::io::Error> {
+        match cache_key {
+            CacheKey::MaterializedPage {
+                hash_key: _,
+                lsn: _,
+            } => {
+                panic!("unexpected dirty materialized page");
+            }
+            CacheKey::EphemeralPage { file_id, blkno } => {
+                writeback_ephemeral_file(*file_id, *blkno, buf)
+            }
+        }
+    }
+
+    /// Initialize a new page cache
+    ///
+    /// This should be called only once at page server startup.
+    fn new(num_pages: usize) -> Self {
+        assert!(num_pages > 0, "page cache size must be > 0");
+
+        let page_buffer = Box::leak(vec![0u8; num_pages * PAGE_SZ].into_boxed_slice());
+
+        let slots = page_buffer
+            .chunks_exact_mut(PAGE_SZ)
+            .map(|chunk| {
+                let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();
+
+                Slot {
+                    inner: RwLock::new(SlotInner {
+                        key: None,
+                        buf,
+                        dirty: false,
+                    }),
+                    usage_count: AtomicU8::new(0),
+                }
+            })
+            .collect();
+
+        Self {
+            materialized_page_map: Default::default(),
+            ephemeral_page_map: Default::default(),
+            slots,
+            next_evict_slot: AtomicUsize::new(0),
+        }
+    }
+}

pageserver/src/page_service.rs

@@ -10,10 +10,9 @@
 //     *callmemaybe <zenith timelineid> $url* -- ask pageserver to start walreceiver on $url
 //
 
-use anyhow::{anyhow, bail, ensure, Result};
+use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::{Buf, BufMut, Bytes, BytesMut};
 use lazy_static::lazy_static;
-use log::*;
 use regex::Regex;
 use std::net::TcpListener;
 use std::str;
@@ -21,10 +20,12 @@ use std::str::FromStr;
 use std::sync::Arc;
 use std::thread;
 use std::{io, net::TcpStream};
+use tracing::*;
 use zenith_metrics::{register_histogram_vec, HistogramVec};
 use zenith_utils::auth::{self, JwtAuth};
 use zenith_utils::auth::{Claims, Scope};
 use zenith_utils::lsn::Lsn;
+use zenith_utils::postgres_backend::is_socket_read_timed_out;
 use zenith_utils::postgres_backend::PostgresBackend;
 use zenith_utils::postgres_backend::{self, AuthType};
 use zenith_utils::pq_proto::{
@@ -187,17 +188,32 @@ pub fn thread_main(
     listener: TcpListener,
     auth_type: AuthType,
 ) -> anyhow::Result<()> {
-    loop {
+    let mut join_handles = Vec::new();
+
+    while !tenant_mgr::shutdown_requested() {
         let (socket, peer_addr) = listener.accept()?;
         debug!("accepted connection from {}", peer_addr);
         socket.set_nodelay(true).unwrap();
         let local_auth = auth.clone();
-        thread::spawn(move || {
-            if let Err(err) = page_service_conn_main(conf, local_auth, socket, auth_type) {
-                error!("page server thread exiting with error: {:#}", err);
-            }
-        });
+
+        let handle = thread::Builder::new()
+            .name("serving Page Service thread".into())
+            .spawn(move || {
+                if let Err(err) = page_service_conn_main(conf, local_auth, socket, auth_type) {
+                    error!(%err, "page server thread exited with error");
+                }
+            })
+            .unwrap();
+
+        join_handles.push(handle);
     }
+
+    debug!("page_service loop terminated. wait for connections to cancel");
+    for handle in join_handles.into_iter() {
+        handle.join().unwrap();
+    }
+
+    Ok(())
 }
 
 fn page_service_conn_main(
@@ -216,7 +232,7 @@ fn page_service_conn_main(
     }
 
     let mut conn_handler = PageServerHandler::new(conf, auth);
-    let pgbackend = PostgresBackend::new(socket, auth_type, None)?;
+    let pgbackend = PostgresBackend::new(socket, auth_type, None, true)?;
     pgbackend.run(&mut conn_handler)
 }
 
@@ -260,50 +276,66 @@ impl PageServerHandler {
         timelineid: ZTimelineId,
         tenantid: ZTenantId,
     ) -> anyhow::Result<()> {
+        let _enter = info_span!("pagestream", timeline = %timelineid, tenant = %tenantid).entered();
+
         // Check that the timeline exists
         let timeline = tenant_mgr::get_timeline_for_tenant(tenantid, timelineid)?;
 
         /* switch client to COPYBOTH */
         pgb.write_message(&BeMessage::CopyBothResponse)?;
 
-        while let Some(message) = pgb.read_message()? {
-            trace!("query({:?}): {:?}", timelineid, message);
+        while !tenant_mgr::shutdown_requested() {
+            match pgb.read_message() {
+                Ok(message) => {
+                    if let Some(message) = message {
+                        trace!("query: {:?}", message);
 
-            let copy_data_bytes = match message {
-                FeMessage::CopyData(bytes) => bytes,
-                _ => continue,
-            };
+                        let copy_data_bytes = match message {
+                            FeMessage::CopyData(bytes) => bytes,
+                            _ => continue,
+                        };
 
-            let zenith_fe_msg = PagestreamFeMessage::parse(copy_data_bytes)?;
+                        let zenith_fe_msg = PagestreamFeMessage::parse(copy_data_bytes)?;
 
-            let response = match zenith_fe_msg {
-                PagestreamFeMessage::Exists(req) => SMGR_QUERY_TIME
-                    .with_label_values(&["get_rel_exists"])
-                    .observe_closure_duration(|| {
-                        self.handle_get_rel_exists_request(&*timeline, &req)
-                    }),
-                PagestreamFeMessage::Nblocks(req) => SMGR_QUERY_TIME
-                    .with_label_values(&["get_rel_size"])
-                    .observe_closure_duration(|| self.handle_get_nblocks_request(&*timeline, &req)),
-                PagestreamFeMessage::GetPage(req) => SMGR_QUERY_TIME
-                    .with_label_values(&["get_page_at_lsn"])
-                    .observe_closure_duration(|| {
-                        self.handle_get_page_at_lsn_request(&*timeline, &req)
-                    }),
-            };
+                        let response = match zenith_fe_msg {
+                            PagestreamFeMessage::Exists(req) => SMGR_QUERY_TIME
+                                .with_label_values(&["get_rel_exists"])
+                                .observe_closure_duration(|| {
+                                    self.handle_get_rel_exists_request(&*timeline, &req)
+                                }),
+                            PagestreamFeMessage::Nblocks(req) => SMGR_QUERY_TIME
+                                .with_label_values(&["get_rel_size"])
+                                .observe_closure_duration(|| {
+                                    self.handle_get_nblocks_request(&*timeline, &req)
+                                }),
+                            PagestreamFeMessage::GetPage(req) => SMGR_QUERY_TIME
+                                .with_label_values(&["get_page_at_lsn"])
+                                .observe_closure_duration(|| {
+                                    self.handle_get_page_at_lsn_request(&*timeline, &req)
+                                }),
+                        };
 
-            let response = response.unwrap_or_else(|e| {
-                // print the all details to the log with {:#}, but for the client the
-                // error message is enough
-                error!("error reading relation or page version: {:#}", e);
-                PagestreamBeMessage::Error(PagestreamErrorResponse {
-                    message: e.to_string(),
-                })
-            });
+                        let response = response.unwrap_or_else(|e| {
+                            // print the all details to the log with {:#}, but for the client the
+                            // error message is enough
+                            error!("error reading relation or page version: {:#}", e);
+                            PagestreamBeMessage::Error(PagestreamErrorResponse {
+                                message: e.to_string(),
+                            })
+                        });
 
-            pgb.write_message(&BeMessage::CopyData(&response.serialize()))?;
+                        pgb.write_message(&BeMessage::CopyData(&response.serialize()))?;
+                    } else {
+                        break;
+                    }
+                }
+                Err(e) => {
+                    if !is_socket_read_timed_out(&e) {
+                        return Err(e);
+                    }
+                }
+            }
         }
-
         Ok(())
     }
 
@@ -363,6 +395,8 @@ impl PageServerHandler {
         timeline: &dyn Timeline,
         req: &PagestreamExistsRequest,
     ) -> Result<PagestreamBeMessage> {
+        let _enter = info_span!("get_rel_exists", rel = %req.rel, req_lsn = %req.lsn).entered();
+
         let tag = RelishTag::Relation(req.rel);
         let lsn = Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest)?;
 
@@ -378,6 +412,7 @@ impl PageServerHandler {
         timeline: &dyn Timeline,
         req: &PagestreamNblocksRequest,
     ) -> Result<PagestreamBeMessage> {
+        let _enter = info_span!("get_nblocks", rel = %req.rel, req_lsn = %req.lsn).entered();
         let tag = RelishTag::Relation(req.rel);
         let lsn = Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest)?;
 
@@ -397,6 +432,8 @@ impl PageServerHandler {
         timeline: &dyn Timeline,
         req: &PagestreamGetPageRequest,
     ) -> Result<PagestreamBeMessage> {
+        let _enter = info_span!("get_page", rel = %req.rel, blkno = &req.blkno, req_lsn = %req.lsn)
+            .entered();
         let tag = RelishTag::Relation(req.rel);
         let lsn = Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest)?;
 
@@ -414,17 +451,25 @@ impl PageServerHandler {
         lsn: Option<Lsn>,
         tenantid: ZTenantId,
     ) -> anyhow::Result<()> {
+        let span = info_span!("basebackup", timeline = %timelineid, tenant = %tenantid, lsn = field::Empty);
+        let _enter = span.enter();
+
         // check that the timeline exists
         let timeline = tenant_mgr::get_timeline_for_tenant(tenantid, timelineid)?;
+        if let Some(lsn) = lsn {
+            timeline
+                .check_lsn_is_in_scope(lsn)
+                .context("invalid basebackup lsn")?;
+        }
 
-        /* switch client to COPYOUT */
+        // switch client to COPYOUT
         pgb.write_message(&BeMessage::CopyOutResponse)?;
-        info!("sent CopyOut");
 
         /* Send a tarball of the latest layer on the timeline */
         {
             let mut writer = CopyDataSink { pgb };
             let mut basebackup = basebackup::Basebackup::new(&mut writer, &timeline, lsn)?;
+            span.record("lsn", &basebackup.lsn.to_string().as_str());
             basebackup.send_tarball()?;
         }
         pgb.write_message(&BeMessage::CopyDone)?;
@@ -529,11 +574,6 @@ impl postgres_backend::Handler for PageServerHandler {
                 None
             };
 
-            info!(
-                "got basebackup command. tenantid=\"{}\" timelineid=\"{}\" lsn=\"{:#?}\"",
-                tenantid, timelineid, lsn
-            );
-
             // Check that the timeline exists
             self.handle_basebackup_request(pgb, timelineid, lsn, tenantid)?;
             pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
@@ -551,6 +591,9 @@ impl postgres_backend::Handler for PageServerHandler {
 
             self.check_permission(Some(tenantid))?;
 
+            let _enter =
+                info_span!("callmemaybe", timeline = %timelineid, tenant = %tenantid).entered();
+
             // Check that the timeline exists
             tenant_mgr::get_timeline_for_tenant(tenantid, timelineid)?;
 
@@ -573,6 +616,9 @@ impl postgres_backend::Handler for PageServerHandler {
 
             self.check_permission(Some(tenantid))?;
 
+            let _enter =
+                info_span!("branch_create", name = %branchname, tenant = %tenantid).entered();
+
             let branch =
                 branches::create_branch(self.conf, &branchname, &startpoint_str, &tenantid)?;
             let branch = serde_json::to_vec(&branch)?;
@@ -589,14 +635,16 @@ impl postgres_backend::Handler for PageServerHandler {
 
             let tenantid = ZTenantId::from_str(caps.get(1).unwrap().as_str())?;
 
-            let branches = crate::branches::get_branches(self.conf, &tenantid)?;
+            // since these handlers for tenant/branch commands are deprecated (in favor of http based ones)
+            // just use false in place of include non incremental logical size
+            let branches = crate::branches::get_branches(self.conf, &tenantid, false)?;
             let branches_buf = serde_json::to_vec(&branches)?;
 
             pgb.write_message_noflush(&SINGLE_COL_ROWDESC)?
                 .write_message_noflush(&BeMessage::DataRow(&[Some(&branches_buf)]))?
                 .write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
         } else if query_string.starts_with("tenant_list") {
-            let tenants = crate::branches::get_tenants(self.conf)?;
+            let tenants = crate::tenant_mgr::list_tenants()?;
             let tenants_buf = serde_json::to_vec(&tenants)?;
 
             pgb.write_message_noflush(&SINGLE_COL_ROWDESC)?
@@ -648,9 +696,7 @@ impl postgres_backend::Handler for PageServerHandler {
                 .unwrap_or(Ok(self.conf.gc_horizon))?;
 
             let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
-
             let result = repo.gc_iteration(Some(timelineid), gc_horizon, true)?;
-
             pgb.write_message_noflush(&BeMessage::RowDescription(&[
                 RowDescriptor::int8_col(b"layer_relfiles_total"),
                 RowDescriptor::int8_col(b"layer_relfiles_needed_by_cutoff"),

pageserver/src/relish_storage.rs

@@ -1,83 +0,0 @@
-//! Abstractions for the page server to store its relish layer data in the external storage.
-//!
-//! Main purpose of this module subtree is to provide a set of abstractions to manage the storage state
-//! in a way, optimal for page server.
-//!
-//! The abstractions hide multiple custom external storage API implementations,
-//! such as AWS S3, local filesystem, etc., located in the submodules.
-
-mod local_fs;
-mod rust_s3;
-/// A queue-based storage with the background machinery behind it to synchronize
-/// local page server layer files with external storage.
-mod synced_storage;
-
-use std::path::Path;
-use std::thread;
-
-use anyhow::Context;
-
-use self::local_fs::LocalFs;
-pub use self::synced_storage::schedule_timeline_upload;
-use crate::relish_storage::rust_s3::RustS3;
-use crate::{PageServerConf, RelishStorageKind};
-
-pub fn run_storage_sync_thread(
-    config: &'static PageServerConf,
-) -> anyhow::Result<Option<thread::JoinHandle<anyhow::Result<()>>>> {
-    match &config.relish_storage_config {
-        Some(relish_storage_config) => {
-            let max_concurrent_sync = relish_storage_config.max_concurrent_sync;
-            match &relish_storage_config.storage {
-                RelishStorageKind::LocalFs(root) => synced_storage::run_storage_sync_thread(
-                    config,
-                    LocalFs::new(root.clone())?,
-                    max_concurrent_sync,
-                ),
-                RelishStorageKind::AwsS3(s3_config) => synced_storage::run_storage_sync_thread(
-                    config,
-                    RustS3::new(s3_config)?,
-                    max_concurrent_sync,
-                ),
-            }
-        }
-        None => Ok(None),
-    }
-}
-
-/// Storage (potentially remote) API to manage its state.
-#[async_trait::async_trait]
-pub trait RelishStorage: Send + Sync {
-    type RelishStoragePath;
-
-    fn derive_destination(
-        page_server_workdir: &Path,
-        relish_local_path: &Path,
-    ) -> anyhow::Result<Self::RelishStoragePath>;
-
-    async fn list_relishes(&self) -> anyhow::Result<Vec<Self::RelishStoragePath>>;
-
-    async fn download_relish(
-        &self,
-        from: &Self::RelishStoragePath,
-        to: &Path,
-    ) -> anyhow::Result<()>;
-
-    async fn delete_relish(&self, path: &Self::RelishStoragePath) -> anyhow::Result<()>;
-
-    async fn upload_relish(&self, from: &Path, to: &Self::RelishStoragePath) -> anyhow::Result<()>;
-}
-
-fn strip_workspace_prefix<'a>(
-    page_server_workdir: &'a Path,
-    relish_local_path: &'a Path,
-) -> anyhow::Result<&'a Path> {
-    relish_local_path
-        .strip_prefix(page_server_workdir)
-        .with_context(|| {
-            format!(
-                "Unexpected: relish local path '{}' is not relevant to server workdir",
-                relish_local_path.display(),
-            )
-        })
-}

pageserver/src/relish_storage/local_fs.rs

@@ -1,158 +0,0 @@
-//! Local filesystem relish storage.
-//!
-//! Page server already stores layer data on the server, when freezing it.
-//! This storage serves a way to
-//!
-//! * test things locally simply
-//! * allow to compabre both binary sets
-//! * help validating the relish storage API
-
-use std::{
-    future::Future,
-    path::{Path, PathBuf},
-    pin::Pin,
-};
-
-use anyhow::{bail, Context};
-
-use super::{strip_workspace_prefix, RelishStorage};
-
-pub struct LocalFs {
-    root: PathBuf,
-}
-
-impl LocalFs {
-    /// Atetmpts to create local FS relish storage, also creates the directory provided, if not exists.
-    pub fn new(root: PathBuf) -> anyhow::Result<Self> {
-        if !root.exists() {
-            std::fs::create_dir_all(&root).with_context(|| {
-                format!(
-                    "Failed to create all directories in the given root path {}",
-                    root.display(),
-                )
-            })?;
-        }
-        Ok(Self { root })
-    }
-
-    fn resolve_in_storage(&self, path: &Path) -> anyhow::Result<PathBuf> {
-        if path.is_relative() {
-            Ok(self.root.join(path))
-        } else if path.starts_with(&self.root) {
-            Ok(path.to_path_buf())
-        } else {
-            bail!(
-                "Path '{}' does not belong to the current storage",
-                path.display()
-            )
-        }
-    }
-}
-
-#[async_trait::async_trait]
-impl RelishStorage for LocalFs {
-    type RelishStoragePath = PathBuf;
-
-    fn derive_destination(
-        page_server_workdir: &Path,
-        relish_local_path: &Path,
-    ) -> anyhow::Result<Self::RelishStoragePath> {
-        Ok(strip_workspace_prefix(page_server_workdir, relish_local_path)?.to_path_buf())
-    }
-
-    async fn list_relishes(&self) -> anyhow::Result<Vec<Self::RelishStoragePath>> {
-        Ok(get_all_files(&self.root).await?.into_iter().collect())
-    }
-
-    async fn download_relish(
-        &self,
-        from: &Self::RelishStoragePath,
-        to: &Path,
-    ) -> anyhow::Result<()> {
-        let file_path = self.resolve_in_storage(from)?;
-        if file_path.exists() && file_path.is_file() {
-            create_target_directory(to).await?;
-            tokio::fs::copy(file_path, to).await?;
-            Ok(())
-        } else {
-            bail!(
-                "File '{}' either does not exist or is not a file",
-                file_path.display()
-            )
-        }
-    }
-
-    async fn delete_relish(&self, path: &Self::RelishStoragePath) -> anyhow::Result<()> {
-        let file_path = self.resolve_in_storage(path)?;
-        if file_path.exists() && file_path.is_file() {
-            Ok(tokio::fs::remove_file(file_path).await?)
-        } else {
-            bail!(
-                "File '{}' either does not exist or is not a file",
-                file_path.display()
-            )
-        }
-    }
-
-    async fn upload_relish(&self, from: &Path, to: &Self::RelishStoragePath) -> anyhow::Result<()> {
-        let target_file_path = self.resolve_in_storage(to)?;
-        create_target_directory(&target_file_path).await?;
-
-        tokio::fs::copy(&from, &target_file_path)
-            .await
-            .with_context(|| {
-                format!(
-                    "Failed to upload relish '{}' to local storage",
-                    from.display(),
-                )
-            })?;
-        Ok(())
-    }
-}
-
-fn get_all_files<'a, P>(
-    directory_path: P,
-) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<PathBuf>>> + Send + Sync + 'a>>
-where
-    P: AsRef<Path> + Send + Sync + 'a,
-{
-    Box::pin(async move {
-        let directory_path = directory_path.as_ref();
-        if directory_path.exists() {
-            if directory_path.is_dir() {
-                let mut paths = Vec::new();
-                let mut dir_contents = tokio::fs::read_dir(directory_path).await?;
-                while let Some(dir_entry) = dir_contents.next_entry().await? {
-                    let file_type = dir_entry.file_type().await?;
-                    let entry_path = dir_entry.path();
-                    if file_type.is_symlink() {
-                        log::debug!("{:?} us a symlink, skipping", entry_path)
-                    } else if file_type.is_dir() {
-                        paths.extend(get_all_files(entry_path).await?.into_iter())
-                    } else {
-                        paths.push(dir_entry.path());
-                    }
-                }
-                Ok(paths)
-            } else {
-                bail!("Path '{}' is not a directory", directory_path.display())
-            }
-        } else {
-            Ok(Vec::new())
-        }
-    })
-}
-
-async fn create_target_directory(target_file_path: &Path) -> anyhow::Result<()> {
-    let target_dir = match target_file_path.parent() {
-        Some(parent_dir) => parent_dir,
-        None => bail!(
-            "Relish path '{}' has no parent directory",
-            target_file_path.display()
-        ),
-    };
-    if !target_dir.exists() {
-        tokio::fs::create_dir_all(target_dir).await?;
-    }
-    Ok(())
-}

pageserver/src/relish_storage/rust_s3.rs

@@ -1,144 +0,0 @@
-//! A wrapper around AWS S3 client library `rust_s3` to be used a relish storage.
-
-use std::path::Path;
-
-use anyhow::Context;
-use s3::{bucket::Bucket, creds::Credentials, region::Region};
-
-use crate::{relish_storage::strip_workspace_prefix, S3Config};
-
-use super::RelishStorage;
-
-const S3_FILE_SEPARATOR: char = '/';
-
-#[derive(Debug)]
-pub struct S3ObjectKey(String);
-
-impl S3ObjectKey {
-    fn key(&self) -> &str {
-        &self.0
-    }
-}
-
-/// AWS S3 relish storage.
-pub struct RustS3 {
-    bucket: Bucket,
-}
-
-impl RustS3 {
-    /// Creates the relish storage, errors if incorrect AWS S3 configuration provided.
-    pub fn new(aws_config: &S3Config) -> anyhow::Result<Self> {
-        let region = aws_config
-            .bucket_region
-            .parse::<Region>()
-            .context("Failed to parse the s3 region from config")?;
-        let credentials = Credentials::new(
-            aws_config.access_key_id.as_deref(),
-            aws_config.secret_access_key.as_deref(),
-            None,
-            None,
-            None,
-        )
-        .context("Failed to create the s3 credentials")?;
-        Ok(Self {
-            bucket: Bucket::new_with_path_style(
-                aws_config.bucket_name.as_str(),
-                region,
-                credentials,
-            )
-            .context("Failed to create the s3 bucket")?,
-        })
-    }
-}
-
-#[async_trait::async_trait]
-impl RelishStorage for RustS3 {
-    type RelishStoragePath = S3ObjectKey;
-
-    fn derive_destination(
-        page_server_workdir: &Path,
-        relish_local_path: &Path,
-    ) -> anyhow::Result<Self::RelishStoragePath> {
-        let relative_path = strip_workspace_prefix(page_server_workdir, relish_local_path)?;
-        let mut key = String::new();
-        for segment in relative_path {
-            key.push(S3_FILE_SEPARATOR);
-            key.push_str(&segment.to_string_lossy());
-        }
-        Ok(S3ObjectKey(key))
-    }
-
-    async fn list_relishes(&self) -> anyhow::Result<Vec<Self::RelishStoragePath>> {
-        let list_response = self
-            .bucket
-            .list(String::new(), None)
-            .await
-            .context("Failed to list s3 objects")?;
-
-        Ok(list_response
-            .into_iter()
-            .flat_map(|response| response.contents)
-            .map(|s3_object| S3ObjectKey(s3_object.key))
-            .collect())
-    }
-
-    async fn download_relish(
-        &self,
-        from: &Self::RelishStoragePath,
-        to: &Path,
-    ) -> anyhow::Result<()> {
-        let mut target_file = std::fs::OpenOptions::new()
-            .write(true)
-            .open(to)
-            .with_context(|| format!("Failed to open target s3 destination at {}", to.display()))?;
-        let code = self
-            .bucket
-            .get_object_stream(from.key(), &mut target_file)
-            .await
-            .with_context(|| format!("Failed to download s3 object with key {}", from.key()))?;
-        if code != 200 {
-            Err(anyhow::format_err!(
-                "Received non-200 exit code during downloading object from directory, code: {}",
-                code
-            ))
-        } else {
-            Ok(())
-        }
-    }
-
-    async fn delete_relish(&self, path: &Self::RelishStoragePath) -> anyhow::Result<()> {
-        let (_, code) = self
-            .bucket
-            .delete_object(path.key())
-            .await
-            .with_context(|| format!("Failed to delete s3 object with key {}", path.key()))?;
-        if code != 200 {
-            Err(anyhow::format_err!(
-                "Received non-200 exit code during deleting object with key '{}', code: {}",
-                path.key(),
-                code
-            ))
-        } else {
-            Ok(())
-        }
-    }
-
-    async fn upload_relish(&self, from: &Path, to: &Self::RelishStoragePath) -> anyhow::Result<()> {
-        let mut local_file = tokio::fs::OpenOptions::new().read(true).open(from).await?;
-
-        let code = self
-            .bucket
-            .put_object_stream(&mut local_file, to.key())
-            .await
-            .with_context(|| format!("Failed to create s3 object with key {}", to.key()))?;
-        if code != 200 {
-            Err(anyhow::format_err!(
-                "Received non-200 exit code during creating object with key '{}', code: {}",
-                to.key(),
-                code
-            ))
-        } else {
-            Ok(())
-        }
-    }
-}

pageserver/src/relish_storage/synced_storage.rs

@@ -1,52 +0,0 @@
-use std::time::Duration;
-use std::{collections::BinaryHeap, sync::Mutex, thread};
-
-use crate::{relish_storage::RelishStorage, PageServerConf};
-
-lazy_static::lazy_static! {
-    static ref UPLOAD_QUEUE: Mutex<BinaryHeap<SyncTask>> = Mutex::new(BinaryHeap::new());
-}
-
-pub fn schedule_timeline_upload(_local_timeline: ()) {
-    // UPLOAD_QUEUE
-    //     .lock()
-    //     .unwrap()
-    //     .push(SyncTask::Upload(local_timeline))
-}
-
-#[derive(Debug, PartialEq, Eq, PartialOrd, Ord)]
-enum SyncTask {}
-
-pub fn run_storage_sync_thread<
-    P: std::fmt::Debug,
-    S: 'static + RelishStorage<RelishStoragePath = P>,
->(
-    config: &'static PageServerConf,
-    relish_storage: S,
-    max_concurrent_sync: usize,
-) -> anyhow::Result<Option<thread::JoinHandle<anyhow::Result<()>>>> {
-    let runtime = tokio::runtime::Builder::new_current_thread()
-        .enable_all()
-        .build()?;
-
-    let handle = thread::Builder::new()
-        .name("Queue based relish storage sync".to_string())
-        .spawn(move || loop {
-            let mut queue_accessor = UPLOAD_QUEUE.lock().unwrap();
-            log::debug!("Upload queue length: {}", queue_accessor.len());
-            let next_task = queue_accessor.pop();
-            drop(queue_accessor);
-            match next_task {
-                Some(task) => runtime.block_on(async {
-                    // suppress warnings
-                    let _ = (config, task, &relish_storage, max_concurrent_sync);
-                    todo!("omitted for brevity")
-                }),
-                None => {
-                    thread::sleep(Duration::from_secs(1));
-                    continue;
-                }
-            }
-        })?;
-    Ok(Some(handle))
-}

pageserver/src/remote_storage.rs

@@ -0,0 +1,182 @@
+//! A set of generic storage abstractions for the page server to use when backing up and restoring its state from the external storage.
+//! This particular module serves as a public API border between pageserver and the internal storage machinery.
+//! No other modules from this tree are supposed to be used directly by the external code.
+//!
+//! There are a few components the storage machinery consists of:
+//! * [`RemoteStorage`] trait a CRUD-like generic abstraction to use for adapting external storages with a few implementations:
+//!     * [`local_fs`] allows to use local file system as an external storage
+//!     * [`rust_s3`] uses AWS S3 bucket entirely as an external storage
+//!
+//! * synchronization logic at [`storage_sync`] module that keeps pageserver state (both runtime one and the workdir files) and storage state in sync.
+//!
+//! * public API via to interact with the external world: [`run_storage_sync_thread`] and [`schedule_timeline_checkpoint_upload`]
+//!
+//! Here's a schematic overview of all interactions backup and the rest of the pageserver perform:
+//!
+//! +------------------------+                                    +--------->-------+
+//! |                        |  - - - (init async loop) - - - ->  |                 |
+//! |                        |                                    |                 |
+//! |                        |  ------------------------------->  |      async      |
+//! |       pageserver       |    (schedule checkpoint upload)    | upload/download |
+//! |                        |                                    |      loop       |
+//! |                        |  <-------------------------------  |                 |
+//! |                        |    (register downloaded layers)    |                 |
+//! +------------------------+                                    +---------<-------+
+//!                                                                         |
+//!                                                                         |
+//!                                          CRUD layer file operations     |
+//!                                     (upload/download/delete/list, etc.) |
+//!                                                                         V
+//!                                                            +------------------------+
+//!                                                            |                        |
+//!                                                            | [`RemoteStorage`] impl |
+//!                                                            |                        |
+//!                                                            | pageserver assumes it  |
+//!                                                            | owns exclusive write   |
+//!                                                            | access to this storage |
+//!                                                            +------------------------+
+//!
+//! First, during startup, the pageserver inits the storage sync thread with the async loop, or leaves the loop unitialised, if configured so.
+//! Some time later, during pageserver checkpoints, in-memory data is flushed onto disk along with its metadata.
+//! If the storage sync loop was successfully started before, pageserver schedules the new image uploads after every checkpoint.
+//! See [`crate::layered_repository`] for the upload calls and the adjacent logic.
+//!
+//! The storage logic considers `image` as a set of local files, fully representing a certain timeline at given moment (identified with `disk_consistent_lsn`).
+//! Timeline can change its state, by adding more files on disk and advancing its `disk_consistent_lsn`: this happens after pageserver checkpointing and is followed
+//! by the storage upload, if enabled.
+//! When a certain image gets uploaded, the sync loop remembers the fact, preventing further reuploads of the same image state.
+//! No files are deleted from either local or remote storage, only the missing ones locally/remotely get downloaded/uploaded, local metadata file will be overwritten
+//! when the newer timeline is downloaded.
+//!
+//! Meanwhile, the loop inits the storage connection and checks the remote files stored.
+//! This is done once at startup only, relying on the fact that pageserver uses the storage alone (ergo, nobody else uploads the files to the storage but this server).
+//! Based on the remote image data, the storage sync logic queues image downloads, while accepting any potential upload tasks from pageserver and managing the tasks by their priority.
+//! On the image download, a [`crate::tenant_mgr::register_relish_download`] function is called to register the new image in pageserver, initializing all related threads and internal state.
+//!
+//! When the pageserver terminates, the upload loop finishes a current image sync task (if any) and exits.
+//!
+//! NOTES:
+//! * pageserver assumes it has exclusive write access to the remote storage. If supported, the way multiple pageservers can be separated in the same storage
+//! (i.e. using different directories in the local filesystem external storage), but totally up to the storage implementation and not covered with the trait API.
+//!
+//! * the uploads do not happen right after pageserver startup, they are registered when
+//!     1. pageserver does the checkpoint, which happens further in the future after the server start
+//!     2. pageserver loads the timeline from disk for the first time
+//!
+//! * the uploads do not happen right after the upload registration: the sync loop might be occupied with other tasks, or tasks with bigger priority could be waiting already
+//!
+//! * all synchronization tasks (including the public API to register uploads and downloads and the sync queue management) happens on an image scale: a big set of remote files,
+//! enough to represent (and recover, if needed) a certain timeline state. On the contrary, all internal storage CRUD calls are made per reilsh file from those images.
+//! This way, the synchronization is able to download the image partially, if some state was synced before, but exposes correctly synced images only.
+
+mod local_fs;
+mod rust_s3;
+mod storage_sync;
+
+use std::{
+    path::{Path, PathBuf},
+    thread,
+};
+
+use anyhow::Context;
+use tokio::io;
+use zenith_utils::zid::{ZTenantId, ZTimelineId};
+
+pub use self::storage_sync::schedule_timeline_checkpoint_upload;
+use self::{local_fs::LocalFs, rust_s3::S3};
+use crate::{PageServerConf, RemoteStorageKind};
+
+/// Any timeline has its own id and its own tenant it belongs to,
+/// the sync processes group timelines by both for simplicity.
+#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Clone, Copy, Hash)]
+pub struct TimelineSyncId(ZTenantId, ZTimelineId);
+
+/// Based on the config, initiates the remote storage connection and starts a separate thread
+/// that ensures that pageserver and the remote storage are in sync with each other.
+/// If no external configuraion connection given, no thread or storage initialization is done.
+pub fn run_storage_sync_thread(
+    config: &'static PageServerConf,
+) -> anyhow::Result<Option<thread::JoinHandle<anyhow::Result<()>>>> {
+    match &config.remote_storage_config {
+        Some(storage_config) => {
+            let max_concurrent_sync = storage_config.max_concurrent_sync;
+            let max_sync_errors = storage_config.max_sync_errors;
+            let handle = match &storage_config.storage {
+                RemoteStorageKind::LocalFs(root) => storage_sync::spawn_storage_sync_thread(
+                    config,
+                    LocalFs::new(root.clone(), &config.workdir)?,
+                    max_concurrent_sync,
+                    max_sync_errors,
+                ),
+                RemoteStorageKind::AwsS3(s3_config) => storage_sync::spawn_storage_sync_thread(
+                    config,
+                    S3::new(s3_config, &config.workdir)?,
+                    max_concurrent_sync,
+                    max_sync_errors,
+                ),
+            };
+            handle.map(Some)
+        }
+        None => Ok(None),
+    }
+}
+
+/// Storage (potentially remote) API to manage its state.
+/// This storage tries to be unaware of any layered repository context,
+/// providing basic CRUD operations with storage files.
+#[async_trait::async_trait]
+trait RemoteStorage: Send + Sync {
+    /// A way to uniquely reference a file in the remote storage.
+    type StoragePath;
+
+    /// Attempts to derive the storage path out of the local path, if the latter is correct.
+    fn storage_path(&self, local_path: &Path) -> anyhow::Result<Self::StoragePath>;
+
+    /// Gets the download path of the given storage file.
+    fn local_path(&self, storage_path: &Self::StoragePath) -> anyhow::Result<PathBuf>;
+
+    /// Lists all items the storage has right now.
+    async fn list(&self) -> anyhow::Result<Vec<Self::StoragePath>>;
+
+    /// Streams the local file contents into remote into the remote storage entry.
+    async fn upload(
+        &self,
+        from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        to: &Self::StoragePath,
+    ) -> anyhow::Result<()>;
+
+    /// Streams the remote storage entry contents into the buffered writer given, returns the filled writer.
+    async fn download(
+        &self,
+        from: &Self::StoragePath,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()>;
+
+    /// Streams a given byte range of the remote storage entry contents into the buffered writer given, returns the filled writer.
+    async fn download_range(
+        &self,
+        from: &Self::StoragePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()>;
+
+    async fn delete(&self, path: &Self::StoragePath) -> anyhow::Result<()>;
+}
+
+fn strip_path_prefix<'a>(prefix: &'a Path, path: &'a Path) -> anyhow::Result<&'a Path> {
+    if prefix == path {
+        anyhow::bail!(
+            "Prefix and the path are equal, cannot strip: '{}'",
+            prefix.display()
+        )
+    } else {
+        path.strip_prefix(prefix).with_context(|| {
+            format!(
+                "Path '{}' is not prefixed with '{}'",
+                path.display(),
+                prefix.display(),
+            )
+        })
+    }
+}

pageserver/src/remote_storage/README.md

@@ -0,0 +1,75 @@
+# Non-implementation details
+
+This document describes the current state of the backup system in pageserver, existing limitations and concerns, why some things are done the way they are the future development plans.
+Detailed description on how the synchronization works and how it fits into the rest of the pageserver can be found in the [storage module](./../remote_storage.rs) and its submodules.
+Ideally, this document should disappear after current implementation concerns are mitigated, with the remaining useful knowledge bits moved into rustdocs.
+
+## Approach
+
+Backup functionality is a new component, appeared way after the core DB functionality was implemented.
+Pageserver layer functionality is also quite volatile at the moment, there's a risk its local file management changes over time.
+
+To avoid adding more chaos into that, backup functionality is currently designed as a relatively standalone component, with the majority of its logic placed in a standalone async loop.
+This way, the backups are managed in background, not affecting directly other pageserver parts: this way the backup and restoration process may lag behind, but eventually keep up with the reality. To track that, a set of prometheus metrics is exposed from pageserver.
+
+## What's done
+
+Current implementation
+* provides remote storage wrappers for AWS S3 and local FS
+* uploads layers, frozen by pageserver checkpoint thread
+* downloads and registers layers, found on the remote storage, but missing locally
+
+No good optimisations or performance testing is done, the feature is disabled by default and gets polished over time.
+It's planned to deal with all questions that are currently on and prepare the feature to be enabled by default in cloud environments.
+
+### Peculiarities
+
+As mentioned, the backup component is rather new and under development currently, so not all things are done properly from the start.
+Here's the list of known compromises with comments:
+
+* Remote storage model is the same as the `tenants/` directory contents of the pageserver's local workdir storage.
+This is relatively simple to implement, but may be costly to use in AWS S3: an initial data image contains ~782 relish files and a metadata file, ~31 MB combined.
+AWS charges per API call and for traffic either, layers are expected to be updated frequently, so this model most probably is ineffective.
+Additionally, pageservers might need to migrate images between tenants, which does not improve the situation.
+
+Storage sync API operates images when backing up or restoring a backup, so we're fluent to repack the layer contents the way we want to, which most probably will be done later.
+
+* no proper file comparison
+
+Currently, every layer contains `Lsn` in their name, to map the data it holds against a certain DB state.
+Then the images with same ids and different `Lsn`'s are compared, files are considered equal if their local file paths are equal (for remote files, "local file path" is their download destination).
+No file contents assertion is done currently, but should be.
+AWS S3 returns file checksums during the `list` operation, so that can be used to ensure the backup consistency, but that needs further research and, since current pageserver impl also needs to deal with layer file checksums.
+
+For now, due to this, we consider local workdir files as source of truth, not removing them ever and adjusting remote files instead, if image files mismatch.
+
+* sad rust-s3 api
+
+rust-s3 is not very pleasant to use:
+1. it returns `anyhow::Result` and it's hard to distinguish "missing file" cases from "no connection" one, for instance
+2. at least one function it its API that we need (`get_object_stream`) has `async` keyword and blocks (!), see details [here](https://github.com/zenithdb/zenith/pull/752#discussion_r728373091)
+3. it's a prerelease library with unclear maintenance status
+4. noisy on debug level
+
+But it's already used in the project, so for now it's reused to avoid bloating the dependency tree.
+Based on previous evaluation, even `rusoto-s3` could be a better choice over this library, but needs further benchmarking.
+
+
+* gc and branches are ignored
+
+So far, we don't consider non-main images and don't adjust the remote storage based on GC thread loop results.
+Only checkpointer loop affects the remote storage.
+
+* more layers should be downloaded on demand
+
+Since we download and load remote layers into pageserver, there's a possibility a need for those layers' ancestors arise.
+Most probably, every downloaded image's ancestor is not present in locally too, but currently there's no logic for downloading such ancestors and their metadata,
+so the pageserver is unable to respond property on requests to such ancestors.
+
+To implement the downloading, more `tenant_mgr` refactoring is needed to properly handle web requests for layers and handle the state changes.
+[Here](https://github.com/zenithdb/zenith/pull/689#issuecomment-931216193) are the details about initial state management updates needed.
+
+* no IT tests
+
+Automated S3 testing is lacking currently, due to no convenient way to enable backups during the tests.
+After it's fixed, benchmark runs should also be carried out to find bottlenecks.

pageserver/src/remote_storage/local_fs.rs

@@ -0,0 +1,689 @@
+//! Local filesystem acting as a remote storage.
+//! Multiple pageservers can use the same "storage" of this kind by using different storage roots.
+//!
+//! This storage used in pageserver tests, but can also be used in cases when a certain persistent
+//! volume is mounted to the local FS.
+
+use std::{
+    future::Future,
+    path::{Path, PathBuf},
+    pin::Pin,
+};
+
+use anyhow::{bail, ensure, Context};
+use tokio::{
+    fs,
+    io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
+};
+use tracing::*;
+
+use super::{strip_path_prefix, RemoteStorage};
+
+pub struct LocalFs {
+    pageserver_workdir: &'static Path,
+    root: PathBuf,
+}
+
+impl LocalFs {
+    /// Attempts to create local FS storage, along with its root directory.
+    pub fn new(root: PathBuf, pageserver_workdir: &'static Path) -> anyhow::Result<Self> {
+        if !root.exists() {
+            std::fs::create_dir_all(&root).with_context(|| {
+                format!(
+                    "Failed to create all directories in the given root path '{}'",
+                    root.display(),
+                )
+            })?;
+        }
+        Ok(Self {
+            pageserver_workdir,
+            root,
+        })
+    }
+
+    fn resolve_in_storage(&self, path: &Path) -> anyhow::Result<PathBuf> {
+        if path.is_relative() {
+            Ok(self.root.join(path))
+        } else if path.starts_with(&self.root) {
+            Ok(path.to_path_buf())
+        } else {
+            bail!(
+                "Path '{}' does not belong to the current storage",
+                path.display()
+            )
+        }
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for LocalFs {
+    type StoragePath = PathBuf;
+
+    fn storage_path(&self, local_path: &Path) -> anyhow::Result<Self::StoragePath> {
+        Ok(self.root.join(
+            strip_path_prefix(self.pageserver_workdir, local_path)
+                .context("local path does not belong to this storage")?,
+        ))
+    }
+
+    fn local_path(&self, storage_path: &Self::StoragePath) -> anyhow::Result<PathBuf> {
+        let relative_path = strip_path_prefix(&self.root, storage_path)
+            .context("local path does not belong to this storage")?;
+        Ok(self.pageserver_workdir.join(relative_path))
+    }
+
+    async fn list(&self) -> anyhow::Result<Vec<Self::StoragePath>> {
+        Ok(get_all_files(&self.root).await?.into_iter().collect())
+    }
+
+    async fn upload(
+        &self,
+        mut from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        to: &Self::StoragePath,
+    ) -> anyhow::Result<()> {
+        let target_file_path = self.resolve_in_storage(to)?;
+        create_target_directory(&target_file_path).await?;
+        let mut destination = io::BufWriter::new(
+            fs::OpenOptions::new()
+                .write(true)
+                .create(true)
+                .open(&target_file_path)
+                .await
+                .with_context(|| {
+                    format!(
+                        "Failed to open target fs destination at '{}'",
+                        target_file_path.display()
+                    )
+                })?,
+        );
+
+        io::copy(&mut from, &mut destination)
+            .await
+            .with_context(|| {
+                format!(
+                    "Failed to upload file to the local storage at '{}'",
+                    target_file_path.display()
+                )
+            })?;
+        destination.flush().await.with_context(|| {
+            format!(
+                "Failed to upload file to the local storage at '{}'",
+                target_file_path.display()
+            )
+        })?;
+        Ok(())
+    }
+
+    async fn download(
+        &self,
+        from: &Self::StoragePath,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()> {
+        let file_path = self.resolve_in_storage(from)?;
+
+        if file_path.exists() && file_path.is_file() {
+            let mut source = io::BufReader::new(
+                fs::OpenOptions::new()
+                    .read(true)
+                    .open(&file_path)
+                    .await
+                    .with_context(|| {
+                        format!(
+                            "Failed to open source file '{}' to use in the download",
+                            file_path.display()
+                        )
+                    })?,
+            );
+            io::copy(&mut source, to).await.with_context(|| {
+                format!(
+                    "Failed to download file '{}' from the local storage",
+                    file_path.display()
+                )
+            })?;
+            source.flush().await?;
+            Ok(())
+        } else {
+            bail!(
+                "File '{}' either does not exist or is not a file",
+                file_path.display()
+            )
+        }
+    }
+
+    async fn download_range(
+        &self,
+        from: &Self::StoragePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()> {
+        if let Some(end_exclusive) = end_exclusive {
+            ensure!(
+                end_exclusive > start_inclusive,
+                "Invalid range, start ({}) is bigger then end ({:?})",
+                start_inclusive,
+                end_exclusive
+            );
+            if start_inclusive == end_exclusive.saturating_sub(1) {
+                return Ok(());
+            }
+        }
+        let file_path = self.resolve_in_storage(from)?;
+
+        if file_path.exists() && file_path.is_file() {
+            let mut source = io::BufReader::new(
+                fs::OpenOptions::new()
+                    .read(true)
+                    .open(&file_path)
+                    .await
+                    .with_context(|| {
+                        format!(
+                            "Failed to open source file '{}' to use in the download",
+                            file_path.display()
+                        )
+                    })?,
+            );
+            source
+                .seek(io::SeekFrom::Start(start_inclusive))
+                .await
+                .context("Failed to seek to the range start in a local storage file")?;
+            match end_exclusive {
+                Some(end_exclusive) => {
+                    io::copy(&mut source.take(end_exclusive - start_inclusive), to).await
+                }
+                None => io::copy(&mut source, to).await,
+            }
+            .with_context(|| {
+                format!(
+                    "Failed to download file '{}' range from the local storage",
+                    file_path.display()
+                )
+            })?;
+            Ok(())
+        } else {
+            bail!(
+                "File '{}' either does not exist or is not a file",
+                file_path.display()
+            )
+        }
+    }
+
+    async fn delete(&self, path: &Self::StoragePath) -> anyhow::Result<()> {
+        let file_path = self.resolve_in_storage(path)?;
+        if file_path.exists() && file_path.is_file() {
+            Ok(fs::remove_file(file_path).await?)
+        } else {
+            bail!(
+                "File '{}' either does not exist or is not a file",
+                file_path.display()
+            )
+        }
+    }
+}
+
+fn get_all_files<'a, P>(
+    directory_path: P,
+) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<PathBuf>>> + Send + Sync + 'a>>
+where
+    P: AsRef<Path> + Send + Sync + 'a,
+{
+    Box::pin(async move {
+        let directory_path = directory_path.as_ref();
+        if directory_path.exists() {
+            if directory_path.is_dir() {
+                let mut paths = Vec::new();
+                let mut dir_contents = fs::read_dir(directory_path).await?;
+                while let Some(dir_entry) = dir_contents.next_entry().await? {
+                    let file_type = dir_entry.file_type().await?;
+                    let entry_path = dir_entry.path();
+                    if file_type.is_symlink() {
+                        debug!("{:?} us a symlink, skipping", entry_path)
+                    } else if file_type.is_dir() {
+                        paths.extend(get_all_files(entry_path).await?.into_iter())
+                    } else {
+                        paths.push(dir_entry.path());
+                    }
+                }
+                Ok(paths)
+            } else {
+                bail!("Path '{}' is not a directory", directory_path.display())
+            }
+        } else {
+            Ok(Vec::new())
+        }
+    })
+}
+
+async fn create_target_directory(target_file_path: &Path) -> anyhow::Result<()> {
+    let target_dir = match target_file_path.parent() {
+        Some(parent_dir) => parent_dir,
+        None => bail!(
+            "File path '{}' has no parent directory",
+            target_file_path.display()
+        ),
+    };
+    if !target_dir.exists() {
+        fs::create_dir_all(target_dir).await?;
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod pure_tests {
+    use crate::{
+        layered_repository::metadata::METADATA_FILE_NAME,
+        repository::repo_harness::{RepoHarness, TIMELINE_ID},
+    };
+
+    use super::*;
+
+    #[test]
+    fn storage_path_positive() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("storage_path_positive")?;
+        let storage_root = PathBuf::from("somewhere").join("else");
+        let storage = LocalFs {
+            pageserver_workdir: &repo_harness.conf.workdir,
+            root: storage_root.clone(),
+        };
+
+        let local_path = repo_harness.timeline_path(&TIMELINE_ID).join("file_name");
+        let expected_path = storage_root.join(local_path.strip_prefix(&repo_harness.conf.workdir)?);
+
+        assert_eq!(
+            expected_path,
+            storage.storage_path(&local_path).expect("Matching path should map to storage path normally"),
+            "File paths from pageserver workdir should be stored in local fs storage with the same path they have relative to the workdir"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn storage_path_negatives() -> anyhow::Result<()> {
+        #[track_caller]
+        fn storage_path_error(storage: &LocalFs, mismatching_path: &Path) -> String {
+            match storage.storage_path(mismatching_path) {
+                Ok(wrong_path) => panic!(
+                    "Expected path '{}' to error, but got storage path: {:?}",
+                    mismatching_path.display(),
+                    wrong_path,
+                ),
+                Err(e) => format!("{:?}", e),
+            }
+        }
+
+        let repo_harness = RepoHarness::create("storage_path_negatives")?;
+        let storage_root = PathBuf::from("somewhere").join("else");
+        let storage = LocalFs {
+            pageserver_workdir: &repo_harness.conf.workdir,
+            root: storage_root,
+        };
+
+        let error_string = storage_path_error(&storage, &repo_harness.conf.workdir);
+        assert!(error_string.contains("does not belong to this storage"));
+        assert!(error_string.contains(repo_harness.conf.workdir.to_str().unwrap()));
+
+        let mismatching_path_str = "/something/else";
+        let error_message = storage_path_error(&storage, Path::new(mismatching_path_str));
+        assert!(
+            error_message.contains(mismatching_path_str),
+            "Error should mention wrong path"
+        );
+        assert!(
+            error_message.contains(repo_harness.conf.workdir.to_str().unwrap()),
+            "Error should mention server workdir"
+        );
+        assert!(error_message.contains("does not belong to this storage"));
+
+        Ok(())
+    }
+
+    #[test]
+    fn local_path_positive() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("local_path_positive")?;
+        let storage_root = PathBuf::from("somewhere").join("else");
+        let storage = LocalFs {
+            pageserver_workdir: &repo_harness.conf.workdir,
+            root: storage_root.clone(),
+        };
+
+        let name = "not a metadata";
+        let local_path = repo_harness.timeline_path(&TIMELINE_ID).join(name);
+        assert_eq!(
+            local_path,
+            storage
+                .local_path(
+                    &storage_root.join(local_path.strip_prefix(&repo_harness.conf.workdir)?)
+                )
+                .expect("For a valid input, valid S3 info should be parsed"),
+            "Should be able to parse metadata out of the correctly named remote delta file"
+        );
+
+        let local_metadata_path = repo_harness
+            .timeline_path(&TIMELINE_ID)
+            .join(METADATA_FILE_NAME);
+        let remote_metadata_path = storage.storage_path(&local_metadata_path)?;
+        assert_eq!(
+            local_metadata_path,
+            storage
+                .local_path(&remote_metadata_path)
+                .expect("For a valid input, valid local path should be parsed"),
+            "Should be able to parse metadata out of the correctly named remote metadata file"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn local_path_negatives() -> anyhow::Result<()> {
+        #[track_caller]
+        #[allow(clippy::ptr_arg)] // have to use &PathBuf due to `storage.local_path` parameter requirements
+        fn local_path_error(storage: &LocalFs, storage_path: &PathBuf) -> String {
+            match storage.local_path(storage_path) {
+                Ok(wrong_path) => panic!(
+                    "Expected local path input {:?} to cause an error, but got file path: {:?}",
+                    storage_path, wrong_path,
+                ),
+                Err(e) => format!("{:?}", e),
+            }
+        }
+
+        let repo_harness = RepoHarness::create("local_path_negatives")?;
+        let storage_root = PathBuf::from("somewhere").join("else");
+        let storage = LocalFs {
+            pageserver_workdir: &repo_harness.conf.workdir,
+            root: storage_root,
+        };
+
+        let totally_wrong_path = "wrong_wrong_wrong";
+        let error_message = local_path_error(&storage, &PathBuf::from(totally_wrong_path));
+        assert!(error_message.contains(totally_wrong_path));
+
+        Ok(())
+    }
+
+    #[test]
+    fn download_destination_matches_original_path() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_destination_matches_original_path")?;
+        let original_path = repo_harness.timeline_path(&TIMELINE_ID).join("some name");
+
+        let storage_root = PathBuf::from("somewhere").join("else");
+        let dummy_storage = LocalFs {
+            pageserver_workdir: &repo_harness.conf.workdir,
+            root: storage_root,
+        };
+
+        let storage_path = dummy_storage.storage_path(&original_path)?;
+        let download_destination = dummy_storage.local_path(&storage_path)?;
+
+        assert_eq!(
+            original_path, download_destination,
+            "'original path -> storage path -> matching fs path' transformation should produce the same path as the input one for the correct path"
+        );
+
+        Ok(())
+    }
+}
+
+#[cfg(test)]
+mod fs_tests {
+    use super::*;
+    use crate::repository::repo_harness::{RepoHarness, TIMELINE_ID};
+
+    use std::io::Write;
+    use tempfile::tempdir;
+
+    #[tokio::test]
+    async fn upload_file() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("upload_file")?;
+        let storage = create_storage()?;
+
+        let source = create_file_for_upload(
+            &storage.pageserver_workdir.join("whatever"),
+            "whatever_contents",
+        )
+        .await?;
+        let target_path = PathBuf::from("/").join("somewhere").join("else");
+        match storage.upload(source, &target_path).await {
+            Ok(()) => panic!("Should not allow storing files with wrong target path"),
+            Err(e) => {
+                let message = format!("{:?}", e);
+                assert!(message.contains(&target_path.display().to_string()));
+                assert!(message.contains("does not belong to the current storage"));
+            }
+        }
+        assert!(storage.list().await?.is_empty());
+
+        let target_path_1 = upload_dummy_file(&repo_harness, &storage, "upload_1").await?;
+        assert_eq!(
+            storage.list().await?,
+            vec![target_path_1.clone()],
+            "Should list a single file after first upload"
+        );
+
+        let target_path_2 = upload_dummy_file(&repo_harness, &storage, "upload_2").await?;
+        assert_eq!(
+            list_files_sorted(&storage).await?,
+            vec![target_path_1.clone(), target_path_2.clone()],
+            "Should list a two different files after second upload"
+        );
+
+        Ok(())
+    }
+
+    fn create_storage() -> anyhow::Result<LocalFs> {
+        let pageserver_workdir = Box::leak(Box::new(tempdir()?.path().to_owned()));
+        let storage = LocalFs::new(tempdir()?.path().to_owned(), pageserver_workdir)?;
+        Ok(storage)
+    }
+
+    #[tokio::test]
+    async fn download_file() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_file")?;
+        let storage = create_storage()?;
+        let upload_name = "upload_1";
+        let upload_target = upload_dummy_file(&repo_harness, &storage, upload_name).await?;
+
+        let mut content_bytes = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        storage.download(&upload_target, &mut content_bytes).await?;
+        content_bytes.flush().await?;
+
+        let contents = String::from_utf8(content_bytes.into_inner().into_inner())?;
+        assert_eq!(
+            dummy_contents(upload_name),
+            contents,
+            "We should upload and download the same contents"
+        );
+
+        let non_existing_path = PathBuf::from("somewhere").join("else");
+        match storage.download(&non_existing_path, &mut io::sink()).await {
+            Ok(_) => panic!("Should not allow downloading non-existing storage files"),
+            Err(e) => {
+                let error_string = e.to_string();
+                assert!(error_string.contains("does not exist"));
+                assert!(error_string.contains(&non_existing_path.display().to_string()));
+            }
+        }
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn download_file_range_positive() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_file_range_positive")?;
+        let storage = create_storage()?;
+        let upload_name = "upload_1";
+        let upload_target = upload_dummy_file(&repo_harness, &storage, upload_name).await?;
+
+        let mut full_range_bytes = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        storage
+            .download_range(&upload_target, 0, None, &mut full_range_bytes)
+            .await?;
+        full_range_bytes.flush().await?;
+        assert_eq!(
+            dummy_contents(upload_name),
+            String::from_utf8(full_range_bytes.into_inner().into_inner())?,
+            "Download full range should return the whole upload"
+        );
+
+        let mut zero_range_bytes = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        let same_byte = 1_000_000_000;
+        storage
+            .download_range(
+                &upload_target,
+                same_byte,
+                Some(same_byte + 1), // exclusive end
+                &mut zero_range_bytes,
+            )
+            .await?;
+        zero_range_bytes.flush().await?;
+        assert!(
+            zero_range_bytes.into_inner().into_inner().is_empty(),
+            "Zero byte range should not download any part of the file"
+        );
+
+        let uploaded_bytes = dummy_contents(upload_name).into_bytes();
+        let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
+
+        let mut first_part_remote = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        storage
+            .download_range(
+                &upload_target,
+                0,
+                Some(first_part_local.len() as u64),
+                &mut first_part_remote,
+            )
+            .await?;
+        first_part_remote.flush().await?;
+        let first_part_remote = first_part_remote.into_inner().into_inner();
+        assert_eq!(
+            first_part_local,
+            first_part_remote.as_slice(),
+            "First part bytes should be returned when requrested"
+        );
+
+        let mut second_part_remote = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        storage
+            .download_range(
+                &upload_target,
+                first_part_local.len() as u64,
+                Some((first_part_local.len() + second_part_local.len()) as u64),
+                &mut second_part_remote,
+            )
+            .await?;
+        second_part_remote.flush().await?;
+        let second_part_remote = second_part_remote.into_inner().into_inner();
+        assert_eq!(
+            second_part_local,
+            second_part_remote.as_slice(),
+            "Second part bytes should be returned when requrested"
+        );
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn download_file_range_negative() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_file_range_negative")?;
+        let storage = create_storage()?;
+        let upload_name = "upload_1";
+        let upload_target = upload_dummy_file(&repo_harness, &storage, upload_name).await?;
+
+        let start = 10000;
+        let end = 234;
+        assert!(start > end, "Should test an incorrect range");
+        match storage
+            .download_range(&upload_target, start, Some(end), &mut io::sink())
+            .await
+        {
+            Ok(_) => panic!("Should not allow downloading wrong ranges"),
+            Err(e) => {
+                let error_string = e.to_string();
+                assert!(error_string.contains("Invalid range"));
+                assert!(error_string.contains(&start.to_string()));
+                assert!(error_string.contains(&end.to_string()));
+            }
+        }
+
+        let non_existing_path = PathBuf::from("somewhere").join("else");
+        match storage
+            .download_range(&non_existing_path, 1, Some(3), &mut io::sink())
+            .await
+        {
+            Ok(_) => panic!("Should not allow downloading non-existing storage file ranges"),
+            Err(e) => {
+                let error_string = e.to_string();
+                assert!(error_string.contains("does not exist"));
+                assert!(error_string.contains(&non_existing_path.display().to_string()));
+            }
+        }
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn delete_file() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("delete_file")?;
+        let storage = create_storage()?;
+        let upload_name = "upload_1";
+        let upload_target = upload_dummy_file(&repo_harness, &storage, upload_name).await?;
+
+        storage.delete(&upload_target).await?;
+        assert!(storage.list().await?.is_empty());
+
+        match storage.delete(&upload_target).await {
+            Ok(()) => panic!("Should not allow deleting non-existing storage files"),
+            Err(e) => {
+                let error_string = e.to_string();
+                assert!(error_string.contains("does not exist"));
+                assert!(error_string.contains(&upload_target.display().to_string()));
+            }
+        }
+        Ok(())
+    }
+
+    async fn upload_dummy_file(
+        harness: &RepoHarness,
+        storage: &LocalFs,
+        name: &str,
+    ) -> anyhow::Result<PathBuf> {
+        let timeline_path = harness.timeline_path(&TIMELINE_ID);
+        let relative_timeline_path = timeline_path.strip_prefix(&harness.conf.workdir)?;
+        let storage_path = storage.root.join(relative_timeline_path).join(name);
+        storage
+            .upload(
+                create_file_for_upload(
+                    &storage.pageserver_workdir.join(name),
+                    &dummy_contents(name),
+                )
+                .await?,
+                &storage_path,
+            )
+            .await?;
+        Ok(storage_path)
+    }
+
+    async fn create_file_for_upload(
+        path: &Path,
+        contents: &str,
+    ) -> anyhow::Result<io::BufReader<fs::File>> {
+        std::fs::create_dir_all(path.parent().unwrap())?;
+        let mut file_for_writing = std::fs::OpenOptions::new()
+            .write(true)
+            .create_new(true)
+            .open(path)?;
+        write!(file_for_writing, "{}", contents)?;
+        drop(file_for_writing);
+        Ok(io::BufReader::new(
+            fs::OpenOptions::new().read(true).open(&path).await?,
+        ))
+    }
+
+    fn dummy_contents(name: &str) -> String {
+        format!("contents for {}", name)
+    }
+
+    async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<PathBuf>> {
+        let mut files = storage.list().await?;
+        files.sort();
+        Ok(files)
+    }
+}

pageserver/src/remote_storage/rust_s3.rs

@@ -0,0 +1,373 @@
+//! AWS S3 storage wrapper around `rust_s3` library.
+//! Currently does not allow multiple pageservers to use the same bucket concurrently: objects are
+//! placed in the root of the bucket.
+
+use std::path::{Path, PathBuf};
+
+use anyhow::Context;
+use s3::{bucket::Bucket, creds::Credentials, region::Region};
+use tokio::io::{self, AsyncWriteExt};
+
+use crate::{
+    remote_storage::{strip_path_prefix, RemoteStorage},
+    S3Config,
+};
+
+const S3_FILE_SEPARATOR: char = '/';
+
+#[derive(Debug, Eq, PartialEq)]
+pub struct S3ObjectKey(String);
+
+impl S3ObjectKey {
+    fn key(&self) -> &str {
+        &self.0
+    }
+
+    fn download_destination(&self, pageserver_workdir: &Path) -> PathBuf {
+        pageserver_workdir.join(self.0.split(S3_FILE_SEPARATOR).collect::<PathBuf>())
+    }
+}
+
+/// AWS S3 storage.
+pub struct S3 {
+    pageserver_workdir: &'static Path,
+    bucket: Bucket,
+}
+
+impl S3 {
+    /// Creates the storage, errors if incorrect AWS S3 configuration provided.
+    pub fn new(aws_config: &S3Config, pageserver_workdir: &'static Path) -> anyhow::Result<Self> {
+        let region = aws_config
+            .bucket_region
+            .parse::<Region>()
+            .context("Failed to parse the s3 region from config")?;
+        let credentials = Credentials::new(
+            aws_config.access_key_id.as_deref(),
+            aws_config.secret_access_key.as_deref(),
+            None,
+            None,
+            None,
+        )
+        .context("Failed to create the s3 credentials")?;
+        Ok(Self {
+            bucket: Bucket::new_with_path_style(
+                aws_config.bucket_name.as_str(),
+                region,
+                credentials,
+            )
+            .context("Failed to create the s3 bucket")?,
+            pageserver_workdir,
+        })
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for S3 {
+    type StoragePath = S3ObjectKey;
+
+    fn storage_path(&self, local_path: &Path) -> anyhow::Result<Self::StoragePath> {
+        let relative_path = strip_path_prefix(self.pageserver_workdir, local_path)?;
+        let mut key = String::new();
+        for segment in relative_path {
+            key.push(S3_FILE_SEPARATOR);
+            key.push_str(&segment.to_string_lossy());
+        }
+        Ok(S3ObjectKey(key))
+    }
+
+    fn local_path(&self, storage_path: &Self::StoragePath) -> anyhow::Result<PathBuf> {
+        Ok(storage_path.download_destination(self.pageserver_workdir))
+    }
+
+    async fn list(&self) -> anyhow::Result<Vec<Self::StoragePath>> {
+        let list_response = self
+            .bucket
+            .list(String::new(), None)
+            .await
+            .context("Failed to list s3 objects")?;
+
+        Ok(list_response
+            .into_iter()
+            .flat_map(|response| response.contents)
+            .map(|s3_object| S3ObjectKey(s3_object.key))
+            .collect())
+    }
+
+    async fn upload(
+        &self,
+        mut from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        to: &Self::StoragePath,
+    ) -> anyhow::Result<()> {
+        let mut upload_contents = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
+        io::copy(&mut from, &mut upload_contents)
+            .await
+            .context("Failed to read the upload contents")?;
+        upload_contents
+            .flush()
+            .await
+            .context("Failed to read the upload contents")?;
+        let upload_contents = upload_contents.into_inner().into_inner();
+
+        let (_, code) = self
+            .bucket
+            .put_object(to.key(), &upload_contents)
+            .await
+            .with_context(|| format!("Failed to create s3 object with key {}", to.key()))?;
+        if code != 200 {
+            Err(anyhow::format_err!(
+                "Received non-200 exit code during creating object with key '{}', code: {}",
+                to.key(),
+                code
+            ))
+        } else {
+            Ok(())
+        }
+    }
+
+    async fn download(
+        &self,
+        from: &Self::StoragePath,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()> {
+        let (data, code) = self
+            .bucket
+            .get_object(from.key())
+            .await
+            .with_context(|| format!("Failed to download s3 object with key {}", from.key()))?;
+        if code != 200 {
+            Err(anyhow::format_err!(
+                "Received non-200 exit code during downloading object, code: {}",
+                code
+            ))
+        } else {
+            // we don't have to write vector into the destination this way, `to_write_all` would be enough.
+            // but we want to prepare for migration on `rusoto`, that has a streaming HTTP body instead here, with
+            // which it makes more sense to use `io::copy`.
+            io::copy(&mut data.as_slice(), to)
+                .await
+                .context("Failed to write downloaded data into the destination buffer")?;
+            Ok(())
+        }
+    }
+
+    async fn download_range(
+        &self,
+        from: &Self::StoragePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+        to: &mut (impl io::AsyncWrite + Unpin + Send + Sync),
+    ) -> anyhow::Result<()> {
+        // S3 accepts ranges as https://www.w3.org/Protocols/rfc2616/rfc2616-sec14.html#sec14.35
+        // and needs both ends to be exclusive
+        let end_inclusive = end_exclusive.map(|end| end.saturating_sub(1));
+        let (data, code) = self
+            .bucket
+            .get_object_range(from.key(), start_inclusive, end_inclusive)
+            .await
+            .with_context(|| format!("Failed to download s3 object with key {}", from.key()))?;
+        if code != 206 {
+            Err(anyhow::format_err!(
+                "Received non-206 exit code during downloading object range, code: {}",
+                code
+            ))
+        } else {
+            // see `download` function above for the comment on why `Vec<u8>` buffer is copied this way
+            io::copy(&mut data.as_slice(), to)
+                .await
+                .context("Failed to write downloaded range into the destination buffer")?;
+            Ok(())
+        }
+    }
+
+    async fn delete(&self, path: &Self::StoragePath) -> anyhow::Result<()> {
+        let (_, code) = self
+            .bucket
+            .delete_object(path.key())
+            .await
+            .with_context(|| format!("Failed to delete s3 object with key {}", path.key()))?;
+        if code != 204 {
+            Err(anyhow::format_err!(
+                "Received non-204 exit code during deleting object with key '{}', code: {}",
+                path.key(),
+                code
+            ))
+        } else {
+            Ok(())
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::{
+        layered_repository::metadata::METADATA_FILE_NAME,
+        repository::repo_harness::{RepoHarness, TIMELINE_ID},
+    };
+
+    use super::*;
+
+    #[test]
+    fn download_destination() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_destination")?;
+
+        let local_path = repo_harness.timeline_path(&TIMELINE_ID).join("test_name");
+        let relative_path = local_path.strip_prefix(&repo_harness.conf.workdir)?;
+
+        let key = S3ObjectKey(format!(
+            "{}{}",
+            S3_FILE_SEPARATOR,
+            relative_path
+                .iter()
+                .map(|segment| segment.to_str().unwrap())
+                .collect::<Vec<_>>()
+                .join(&S3_FILE_SEPARATOR.to_string()),
+        ));
+
+        assert_eq!(
+            local_path,
+            key.download_destination(&repo_harness.conf.workdir),
+            "Download destination should consist of s3 path joined with the pageserver workdir prefix"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn storage_path_positive() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("storage_path_positive")?;
+
+        let segment_1 = "matching";
+        let segment_2 = "file";
+        let local_path = &repo_harness.conf.workdir.join(segment_1).join(segment_2);
+        let expected_key = S3ObjectKey(format!(
+            "{SEPARATOR}{}{SEPARATOR}{}",
+            segment_1,
+            segment_2,
+            SEPARATOR = S3_FILE_SEPARATOR,
+        ));
+
+        let actual_key = dummy_storage(&repo_harness.conf.workdir)
+            .storage_path(local_path)
+            .expect("Matching path should map to S3 path normally");
+        assert_eq!(
+            expected_key,
+            actual_key,
+            "S3 key from the matching path should contain all segments after the workspace prefix, separated with S3 separator"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn storage_path_negatives() -> anyhow::Result<()> {
+        #[track_caller]
+        fn storage_path_error(storage: &S3, mismatching_path: &Path) -> String {
+            match storage.storage_path(mismatching_path) {
+                Ok(wrong_key) => panic!(
+                    "Expected path '{}' to error, but got S3 key: {:?}",
+                    mismatching_path.display(),
+                    wrong_key,
+                ),
+                Err(e) => e.to_string(),
+            }
+        }
+
+        let repo_harness = RepoHarness::create("storage_path_negatives")?;
+        let storage = dummy_storage(&repo_harness.conf.workdir);
+
+        let error_message = storage_path_error(&storage, &repo_harness.conf.workdir);
+        assert!(
+            error_message.contains("Prefix and the path are equal"),
+            "Message '{}' does not contain the required string",
+            error_message
+        );
+
+        let mismatching_path = PathBuf::from("somewhere").join("else");
+        let error_message = storage_path_error(&storage, &mismatching_path);
+        assert!(
+            error_message.contains(mismatching_path.to_str().unwrap()),
+            "Error should mention wrong path"
+        );
+        assert!(
+            error_message.contains(repo_harness.conf.workdir.to_str().unwrap()),
+            "Error should mention server workdir"
+        );
+        assert!(
+            error_message.contains("is not prefixed with"),
+            "Message '{}' does not contain a required string",
+            error_message
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn local_path_positive() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("local_path_positive")?;
+        let storage = dummy_storage(&repo_harness.conf.workdir);
+        let timeline_dir = repo_harness.timeline_path(&TIMELINE_ID);
+        let relative_timeline_path = timeline_dir.strip_prefix(&repo_harness.conf.workdir)?;
+
+        let s3_key = create_s3_key(&relative_timeline_path.join("not a metadata"));
+        assert_eq!(
+            s3_key.download_destination(&repo_harness.conf.workdir),
+            storage
+                .local_path(&s3_key)
+                .expect("For a valid input, valid S3 info should be parsed"),
+            "Should be able to parse metadata out of the correctly named remote delta file"
+        );
+
+        let s3_key = create_s3_key(&relative_timeline_path.join(METADATA_FILE_NAME));
+        assert_eq!(
+            s3_key.download_destination(&repo_harness.conf.workdir),
+            storage
+                .local_path(&s3_key)
+                .expect("For a valid input, valid S3 info should be parsed"),
+            "Should be able to parse metadata out of the correctly named remote metadata file"
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn download_destination_matches_original_path() -> anyhow::Result<()> {
+        let repo_harness = RepoHarness::create("download_destination_matches_original_path")?;
+        let original_path = repo_harness.timeline_path(&TIMELINE_ID).join("some name");
+
+        let dummy_storage = dummy_storage(&repo_harness.conf.workdir);
+
+        let key = dummy_storage.storage_path(&original_path)?;
+        let download_destination = dummy_storage.local_path(&key)?;
+
+        assert_eq!(
+            original_path, download_destination,
+            "'original path -> storage key -> matching fs path' transformation should produce the same path as the input one for the correct path"
+        );
+
+        Ok(())
+    }
+
+    fn dummy_storage(pageserver_workdir: &'static Path) -> S3 {
+        S3 {
+            pageserver_workdir,
+            bucket: Bucket::new(
+                "dummy-bucket",
+                "us-east-1".parse().unwrap(),
+                Credentials::anonymous().unwrap(),
+            )
+            .unwrap(),
+        }
+    }
+
+    fn create_s3_key(relative_file_path: &Path) -> S3ObjectKey {
+        S3ObjectKey(
+            relative_file_path
+                .iter()
+                .fold(String::new(), |mut path_string, segment| {
+                    path_string.push(S3_FILE_SEPARATOR);
+                    path_string.push_str(segment.to_str().unwrap());
+                    path_string
+                }),
+        )
+    }
+}

pageserver/src/repository.rs

@@ -1,9 +1,10 @@
 use crate::relish::*;
+use crate::CheckpointConfig;
 use anyhow::Result;
-use bytes::{Buf, BufMut, Bytes, BytesMut};
+use bytes::Bytes;
 use serde::{Deserialize, Serialize};
 use std::collections::HashSet;
-use std::ops::AddAssign;
+use std::ops::{AddAssign, Deref};
 use std::sync::Arc;
 use std::time::Duration;
 use zenith_utils::lsn::{Lsn, RecordLsn};
@@ -13,18 +14,28 @@ use zenith_utils::zid::ZTimelineId;
 /// A repository corresponds to one .zenith directory. One repository holds multiple
 /// timelines, forked off from the same initial call to 'initdb'.
 pub trait Repository: Send + Sync {
+    fn shutdown(&self) -> Result<()>;
+
+    /// Stops all timeline-related process in the repository and removes the timeline data from memory.
+    fn unload_timeline(&self, timeline_id: ZTimelineId) -> Result<()>;
+
     /// Get Timeline handle for given zenith timeline ID.
     fn get_timeline(&self, timelineid: ZTimelineId) -> Result<Arc<dyn Timeline>>;
 
     /// Create a new, empty timeline. The caller is responsible for loading data into it
-    fn create_empty_timeline(&self, timelineid: ZTimelineId) -> Result<Arc<dyn Timeline>>;
+    /// Initdb lsn is provided for timeline impl to be able to perform checks for some operations against it.
+    fn create_empty_timeline(
+        &self,
+        timelineid: ZTimelineId,
+        initdb_lsn: Lsn,
+    ) -> Result<Arc<dyn Timeline>>;
 
     /// Branch a timeline
     fn branch_timeline(&self, src: ZTimelineId, dst: ZTimelineId, start_lsn: Lsn) -> Result<()>;
 
-    /// perform one garbage collection iteration.
-    /// garbage collection is periodically performed by gc thread,
-    /// but it can be explicitly requested through page server api.
+    /// perform one garbage collection iteration, removing old data files from disk.
+    /// this funtion is periodically called by gc thread.
+    /// also it can be explicitly requested through page server api 'do_gc' command.
     ///
     /// 'timelineid' specifies the timeline to GC, or None for all.
     /// `horizon` specifies delta from last lsn to preserve all object versions (pitr interval).
@@ -37,6 +48,10 @@ pub trait Repository: Send + Sync {
         horizon: u64,
         checkpoint_before_gc: bool,
     ) -> Result<GcResult>;
+
+    /// perform one checkpoint iteration, flushing in-memory data on disk.
+    /// this function is periodically called by checkponter thread.
+    fn checkpoint_iteration(&self, cconf: CheckpointConfig) -> Result<()>;
 }
 
 ///
@@ -117,17 +132,64 @@ pub trait Timeline: Send + Sync {
     /// Get a list of all existing non-relational objects
     fn list_nonrels(&self, lsn: Lsn) -> Result<HashSet<RelishTag>>;
 
+    /// Get the ancestor's timeline id
+    fn get_ancestor_timeline_id(&self) -> Option<ZTimelineId>;
+
+    /// Get the LSN where this branch was created
+    fn get_ancestor_lsn(&self) -> Lsn;
+
     //------------------------------------------------------------------------------
     // Public PUT functions, to update the repository with new page versions.
     //
     // These are called by the WAL receiver to digest WAL records.
     //------------------------------------------------------------------------------
 
+    /// Atomically get both last and prev.
+    fn get_last_record_rlsn(&self) -> RecordLsn;
+    /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
+    fn get_last_record_lsn(&self) -> Lsn;
+    fn get_prev_record_lsn(&self) -> Lsn;
+    fn get_start_lsn(&self) -> Lsn;
+    fn get_disk_consistent_lsn(&self) -> Lsn;
+
+    /// Mutate the timeline with a [`TimelineWriter`].
+    fn writer<'a>(&'a self) -> Box<dyn TimelineWriter + 'a>;
+
+    ///
+    /// Flush to disk all data that was written with the put_* functions
+    ///
+    /// NOTE: This has nothing to do with checkpoint in PostgreSQL. We don't
+    /// know anything about them here in the repository.
+    fn checkpoint(&self, cconf: CheckpointConfig) -> Result<()>;
+
+    ///
+    /// Check that it is valid to request operations with that lsn.
+    fn check_lsn_is_in_scope(&self, lsn: Lsn) -> Result<()>;
+
+    /// Retrieve current logical size of the timeline
+    ///
+    /// NOTE: counted incrementally, includes ancestors,
+    /// doesnt support TwoPhase relishes yet
+    fn get_current_logical_size(&self) -> usize;
+
+    /// Does the same as get_current_logical_size but counted on demand.
+    /// Used in tests to ensure thet incremental and non incremental variants match.
+    fn get_current_logical_size_non_incremental(&self, lsn: Lsn) -> Result<usize>;
+
+    /// An escape hatch to allow "casting" a generic Timeline to LayeredTimeline.
+    fn upgrade_to_layered_timeline(&self) -> &crate::layered_repository::LayeredTimeline;
+}
+
+/// Various functions to mutate the timeline.
+// TODO Currently, Deref is used to allow easy access to read methods from this trait.
+// This is probably considered a bad practice in Rust and should be fixed eventually,
+// but will cause large code changes.
+pub trait TimelineWriter: Deref<Target = dyn Timeline> {
     /// Put a new page version that can be constructed from a WAL record
     ///
     /// This will implicitly extend the relation, if the page is beyond the
     /// current end-of-file.
-    fn put_wal_record(&self, tag: RelishTag, blknum: u32, rec: WALRecord) -> Result<()>;
+    fn put_wal_record(&self, lsn: Lsn, tag: RelishTag, blknum: u32, rec: WALRecord) -> Result<()>;
 
     /// Like put_wal_record, but with ready-made image of the page.
     fn put_page_image(&self, tag: RelishTag, blknum: u32, lsn: Lsn, img: Bytes) -> Result<()>;
@@ -143,34 +205,10 @@ pub trait Timeline: Send + Sync {
     /// Advance requires aligned LSN as an argument and would wake wait_lsn() callers.
     /// Previous last record LSN is stored alongside the latest and can be read.
     fn advance_last_record_lsn(&self, lsn: Lsn);
-    /// Atomically get both last and prev.
-    fn get_last_record_rlsn(&self) -> RecordLsn;
-    /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
-    fn get_last_record_lsn(&self) -> Lsn;
-    fn get_prev_record_lsn(&self) -> Lsn;
-    fn get_start_lsn(&self) -> Lsn;
-
-    ///
-    /// Flush to disk all data that was written with the put_* functions
-    ///
-    /// NOTE: This has nothing to do with checkpoint in PostgreSQL. We don't
-    /// know anything about them here in the repository.
-    fn checkpoint(&self) -> Result<()>;
-
-    /// Retrieve current logical size of the timeline
-    ///
-    /// NOTE: counted incrementally, includes ancestors,
-    /// doesnt support TwoPhase relishes yet
-    fn get_current_logical_size(&self) -> usize;
-
-    /// Does the same as get_current_logical_size but counted on demand.
-    /// Used in tests to ensure thet incremental and non incremental variants match.
-    fn get_current_logical_size_non_incremental(&self, lsn: Lsn) -> Result<usize>;
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
 pub struct WALRecord {
-    pub lsn: Lsn, // LSN at the *end* of the record
     pub will_init: bool,
     pub rec: Bytes,
     // Remember the offset of main_data in rec,
@@ -179,68 +217,29 @@ pub struct WALRecord {
     pub main_data_offset: u32,
 }
 
-impl WALRecord {
-    pub fn pack(&self, buf: &mut BytesMut) {
-        buf.put_u64(self.lsn.0);
-        buf.put_u8(self.will_init as u8);
-        buf.put_u32(self.main_data_offset);
-        buf.put_u32(self.rec.len() as u32);
-        buf.put_slice(&self.rec[..]);
-    }
-    pub fn unpack(buf: &mut Bytes) -> WALRecord {
-        let lsn = Lsn::from(buf.get_u64());
-        let will_init = buf.get_u8() != 0;
-        let main_data_offset = buf.get_u32();
-        let mut dst = vec![0u8; buf.get_u32() as usize];
-        buf.copy_to_slice(&mut dst);
-        WALRecord {
-            lsn,
-            will_init,
-            rec: Bytes::from(dst),
-            main_data_offset,
-        }
-    }
-}
-
-///
-/// Tests that should work the same with any Repository/Timeline implementation.
-///
-#[allow(clippy::bool_assert_comparison)]
 #[cfg(test)]
-mod tests {
+pub mod repo_harness {
+    use bytes::BytesMut;
+    use std::{fs, path::PathBuf};
+
+    use crate::{
+        layered_repository::{LayeredRepository, TIMELINES_SEGMENT_NAME},
+        walredo::{WalRedoError, WalRedoManager},
+        PageServerConf,
+    };
+
     use super::*;
-    use crate::layered_repository::LayeredRepository;
-    use crate::walredo::{WalRedoError, WalRedoManager};
-    use crate::PageServerConf;
     use hex_literal::hex;
-    use postgres_ffi::pg_constants;
-    use postgres_ffi::xlog_utils::SIZEOF_CHECKPOINT;
-    use std::fs;
-    use std::path::PathBuf;
     use zenith_utils::zid::ZTenantId;
 
-    const TIMELINE_ID: ZTimelineId =
+    pub const TIMELINE_ID: ZTimelineId =
         ZTimelineId::from_array(hex!("11223344556677881122334455667788"));
-    const NEW_TIMELINE_ID: ZTimelineId =
+    pub const NEW_TIMELINE_ID: ZTimelineId =
         ZTimelineId::from_array(hex!("AA223344556677881122334455667788"));
 
-    /// Arbitrary relation tag, for testing.
-    const TESTREL_A: RelishTag = RelishTag::Relation(RelTag {
-        spcnode: 0,
-        dbnode: 111,
-        relnode: 1000,
-        forknum: 0,
-    });
-    const TESTREL_B: RelishTag = RelishTag::Relation(RelTag {
-        spcnode: 0,
-        dbnode: 111,
-        relnode: 1001,
-        forknum: 0,
-    });
-
     /// Convenience function to create a page image with given string as the only content
     #[allow(non_snake_case)]
-    fn TEST_IMG(s: &str) -> Bytes {
+    pub fn TEST_IMG(s: &str) -> Bytes {
         let mut buf = BytesMut::new();
         buf.extend_from_slice(s.as_bytes());
         buf.resize(8192, 0);
@@ -248,28 +247,17 @@ mod tests {
         buf.freeze()
     }
 
-    fn assert_current_logical_size(timeline: &Arc<dyn Timeline>, lsn: Lsn) {
-        let incremental = timeline.get_current_logical_size();
-        let non_incremental = timeline
-            .get_current_logical_size_non_incremental(lsn)
-            .unwrap();
-        assert_eq!(incremental, non_incremental);
-    }
-
-    static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; 8192]);
-    static ZERO_CHECKPOINT: Bytes = Bytes::from_static(&[0u8; SIZEOF_CHECKPOINT]);
-
-    struct RepoHarness {
-        conf: &'static PageServerConf,
-        tenant_id: ZTenantId,
+    pub struct RepoHarness {
+        pub conf: &'static PageServerConf,
+        pub tenant_id: ZTenantId,
     }
 
     impl RepoHarness {
-        fn create(test_name: &'static str) -> Result<Self> {
+        pub fn create(test_name: &'static str) -> Result<Self> {
             let repo_dir = PageServerConf::test_repo_dir(test_name);
             let _ = fs::remove_dir_all(&repo_dir);
             fs::create_dir_all(&repo_dir)?;
-            fs::create_dir_all(&repo_dir.join("timelines"))?;
+            fs::create_dir_all(&repo_dir.join(TIMELINES_SEGMENT_NAME))?;
 
             let conf = PageServerConf::dummy_conf(repo_dir);
             // Make a static copy of the config. This can never be free'd, but that's
@@ -282,7 +270,7 @@ mod tests {
             Ok(Self { conf, tenant_id })
         }
 
-        fn load(&self) -> Box<dyn Repository> {
+        pub fn load(&self) -> Box<dyn Repository> {
             let walredo_mgr = Arc::new(TestRedoManager);
 
             Box::new(LayeredRepository::new(
@@ -293,11 +281,80 @@ mod tests {
             ))
         }
 
-        fn timeline_path(&self, timeline_id: &ZTimelineId) -> PathBuf {
+        pub fn timeline_path(&self, timeline_id: &ZTimelineId) -> PathBuf {
             self.conf.timeline_path(timeline_id, &self.tenant_id)
         }
     }
 
+    // Mock WAL redo manager that doesn't do much
+    struct TestRedoManager;
+
+    impl WalRedoManager for TestRedoManager {
+        fn request_redo(
+            &self,
+            rel: RelishTag,
+            blknum: u32,
+            lsn: Lsn,
+            base_img: Option<Bytes>,
+            records: Vec<(Lsn, WALRecord)>,
+        ) -> Result<Bytes, WalRedoError> {
+            let s = format!(
+                "redo for {} blk {} to get to {}, with {} and {} records",
+                rel,
+                blknum,
+                lsn,
+                if base_img.is_some() {
+                    "base image"
+                } else {
+                    "no base image"
+                },
+                records.len()
+            );
+            println!("{}", s);
+            Ok(TEST_IMG(&s))
+        }
+    }
+}
+
+///
+/// Tests that should work the same with any Repository/Timeline implementation.
+///
+#[allow(clippy::bool_assert_comparison)]
+#[cfg(test)]
+mod tests {
+    use crate::layered_repository::metadata::METADATA_FILE_NAME;
+
+    use super::repo_harness::*;
+    use super::*;
+    use postgres_ffi::{pg_constants, xlog_utils::SIZEOF_CHECKPOINT};
+    use std::fs;
+
+    /// Arbitrary relation tag, for testing.
+    const TESTREL_A_REL_TAG: RelTag = RelTag {
+        spcnode: 0,
+        dbnode: 111,
+        relnode: 1000,
+        forknum: 0,
+    };
+    const TESTREL_A: RelishTag = RelishTag::Relation(TESTREL_A_REL_TAG);
+    const TESTREL_B: RelishTag = RelishTag::Relation(RelTag {
+        spcnode: 0,
+        dbnode: 111,
+        relnode: 1001,
+        forknum: 0,
+    });
+
+    fn assert_current_logical_size(timeline: &Arc<dyn Timeline>, lsn: Lsn) {
+        let incremental = timeline.get_current_logical_size();
+        let non_incremental = timeline
+            .get_current_logical_size_non_incremental(lsn)
+            .unwrap();
+        assert_eq!(incremental, non_incremental);
+    }
+
+    static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; 8192]);
+    static ZERO_CHECKPOINT: Bytes = Bytes::from_static(&[0u8; SIZEOF_CHECKPOINT]);
+
     #[test]
     fn test_relsize() -> Result<()> {
         let repo = RepoHarness::create("test_relsize")?.load();
@@ -305,15 +362,16 @@ mod tests {
         //repo.get_timeline("11223344556677881122334455667788");
 
         // Create timeline to work on
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
 
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x30), TEST_IMG("foo blk 0 at 3"))?;
-        tline.put_page_image(TESTREL_A, 1, Lsn(0x40), TEST_IMG("foo blk 1 at 4"))?;
-        tline.put_page_image(TESTREL_A, 2, Lsn(0x50), TEST_IMG("foo blk 2 at 5"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x30), TEST_IMG("foo blk 0 at 3"))?;
+        writer.put_page_image(TESTREL_A, 1, Lsn(0x40), TEST_IMG("foo blk 1 at 4"))?;
+        writer.put_page_image(TESTREL_A, 2, Lsn(0x50), TEST_IMG("foo blk 2 at 5"))?;
 
-        tline.advance_last_record_lsn(Lsn(0x50));
+        writer.advance_last_record_lsn(Lsn(0x50));
 
         assert_current_logical_size(&tline, Lsn(0x50));
 
@@ -359,8 +417,8 @@ mod tests {
         );
 
         // Truncate last block
-        tline.put_truncation(TESTREL_A, Lsn(0x60), 2)?;
-        tline.advance_last_record_lsn(Lsn(0x60));
+        writer.put_truncation(TESTREL_A, Lsn(0x60), 2)?;
+        writer.advance_last_record_lsn(Lsn(0x60));
         assert_current_logical_size(&tline, Lsn(0x60));
 
         // Check reported size and contents after truncation
@@ -382,13 +440,13 @@ mod tests {
         );
 
         // Truncate to zero length
-        tline.put_truncation(TESTREL_A, Lsn(0x68), 0)?;
-        tline.advance_last_record_lsn(Lsn(0x68));
+        writer.put_truncation(TESTREL_A, Lsn(0x68), 0)?;
+        writer.advance_last_record_lsn(Lsn(0x68));
         assert_eq!(tline.get_relish_size(TESTREL_A, Lsn(0x68))?.unwrap(), 0);
 
         // Extend from 0 to 2 blocks, leaving a gap
-        tline.put_page_image(TESTREL_A, 1, Lsn(0x70), TEST_IMG("foo blk 1"))?;
-        tline.advance_last_record_lsn(Lsn(0x70));
+        writer.put_page_image(TESTREL_A, 1, Lsn(0x70), TEST_IMG("foo blk 1"))?;
+        writer.advance_last_record_lsn(Lsn(0x70));
         assert_eq!(tline.get_relish_size(TESTREL_A, Lsn(0x70))?.unwrap(), 2);
         assert_eq!(tline.get_page_at_lsn(TESTREL_A, 0, Lsn(0x70))?, ZERO_PAGE);
         assert_eq!(
@@ -422,26 +480,27 @@ mod tests {
         let repo = RepoHarness::create("test_drop_extend")?.load();
 
         // Create timeline to work on
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
 
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
-        tline.advance_last_record_lsn(Lsn(0x20));
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
+        writer.advance_last_record_lsn(Lsn(0x20));
 
         // Check that rel exists and size is correct
         assert_eq!(tline.get_rel_exists(TESTREL_A, Lsn(0x20))?, true);
         assert_eq!(tline.get_relish_size(TESTREL_A, Lsn(0x20))?.unwrap(), 1);
 
         // Drop relish
-        tline.drop_relish(TESTREL_A, Lsn(0x30))?;
-        tline.advance_last_record_lsn(Lsn(0x30));
+        writer.drop_relish(TESTREL_A, Lsn(0x30))?;
+        writer.advance_last_record_lsn(Lsn(0x30));
 
         // Check that rel is not visible anymore
         assert_eq!(tline.get_rel_exists(TESTREL_A, Lsn(0x30))?, false);
         assert!(tline.get_relish_size(TESTREL_A, Lsn(0x30))?.is_none());
 
         // Extend it again
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("foo blk 0 at 4"))?;
-        tline.advance_last_record_lsn(Lsn(0x40));
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("foo blk 0 at 4"))?;
+        writer.advance_last_record_lsn(Lsn(0x40));
 
         // Check that rel exists and size is correct
         assert_eq!(tline.get_rel_exists(TESTREL_A, Lsn(0x40))?, true);
@@ -458,7 +517,8 @@ mod tests {
         let repo = RepoHarness::create("test_truncate_extend")?.load();
 
         // Create timeline to work on
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
 
         //from storage_layer.rs
         const RELISH_SEG_SIZE: u32 = 10 * 1024 * 1024 / 8192;
@@ -468,10 +528,10 @@ mod tests {
         for blkno in 0..relsize {
             let lsn = Lsn(0x20);
             let data = format!("foo blk {} at {}", blkno, lsn);
-            tline.put_page_image(TESTREL_A, blkno, lsn, TEST_IMG(&data))?;
+            writer.put_page_image(TESTREL_A, blkno, lsn, TEST_IMG(&data))?;
         }
 
-        tline.advance_last_record_lsn(Lsn(0x20));
+        writer.advance_last_record_lsn(Lsn(0x20));
 
         // The relation was created at LSN 2, not visible at LSN 1 yet.
         assert_eq!(tline.get_rel_exists(TESTREL_A, Lsn(0x10))?, false);
@@ -495,8 +555,8 @@ mod tests {
 
         // Truncate relation so that second segment was dropped
         // - only leave one page
-        tline.put_truncation(TESTREL_A, Lsn(0x60), 1)?;
-        tline.advance_last_record_lsn(Lsn(0x60));
+        writer.put_truncation(TESTREL_A, Lsn(0x60), 1)?;
+        writer.advance_last_record_lsn(Lsn(0x60));
 
         // Check reported size and contents after truncation
         assert_eq!(tline.get_relish_size(TESTREL_A, Lsn(0x60))?.unwrap(), 1);
@@ -529,9 +589,9 @@ mod tests {
         for blkno in 0..relsize {
             let lsn = Lsn(0x80);
             let data = format!("foo blk {} at {}", blkno, lsn);
-            tline.put_page_image(TESTREL_A, blkno, lsn, TEST_IMG(&data))?;
+            writer.put_page_image(TESTREL_A, blkno, lsn, TEST_IMG(&data))?;
         }
-        tline.advance_last_record_lsn(Lsn(0x80));
+        writer.advance_last_record_lsn(Lsn(0x80));
 
         assert_eq!(tline.get_rel_exists(TESTREL_A, Lsn(0x80))?, true);
         assert_eq!(
@@ -556,15 +616,16 @@ mod tests {
     #[test]
     fn test_large_rel() -> Result<()> {
         let repo = RepoHarness::create("test_large_rel")?.load();
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
 
         let mut lsn = 0x10;
         for blknum in 0..pg_constants::RELSEG_SIZE + 1 {
             let img = TEST_IMG(&format!("foo blk {} at {}", blknum, Lsn(lsn)));
             lsn += 0x10;
-            tline.put_page_image(TESTREL_A, blknum as u32, Lsn(lsn), img)?;
+            writer.put_page_image(TESTREL_A, blknum as u32, Lsn(lsn), img)?;
         }
-        tline.advance_last_record_lsn(Lsn(lsn));
+        writer.advance_last_record_lsn(Lsn(lsn));
 
         assert_current_logical_size(&tline, Lsn(lsn));
 
@@ -575,8 +636,8 @@ mod tests {
 
         // Truncate one block
         lsn += 0x10;
-        tline.put_truncation(TESTREL_A, Lsn(lsn), pg_constants::RELSEG_SIZE)?;
-        tline.advance_last_record_lsn(Lsn(lsn));
+        writer.put_truncation(TESTREL_A, Lsn(lsn), pg_constants::RELSEG_SIZE)?;
+        writer.advance_last_record_lsn(Lsn(lsn));
         assert_eq!(
             tline.get_relish_size(TESTREL_A, Lsn(lsn))?.unwrap(),
             pg_constants::RELSEG_SIZE
@@ -585,8 +646,8 @@ mod tests {
 
         // Truncate another block
         lsn += 0x10;
-        tline.put_truncation(TESTREL_A, Lsn(lsn), pg_constants::RELSEG_SIZE - 1)?;
-        tline.advance_last_record_lsn(Lsn(lsn));
+        writer.put_truncation(TESTREL_A, Lsn(lsn), pg_constants::RELSEG_SIZE - 1)?;
+        writer.advance_last_record_lsn(Lsn(lsn));
         assert_eq!(
             tline.get_relish_size(TESTREL_A, Lsn(lsn))?.unwrap(),
             pg_constants::RELSEG_SIZE - 1
@@ -598,8 +659,8 @@ mod tests {
         let mut size: i32 = 3000;
         while size >= 0 {
             lsn += 0x10;
-            tline.put_truncation(TESTREL_A, Lsn(lsn), size as u32)?;
-            tline.advance_last_record_lsn(Lsn(lsn));
+            writer.put_truncation(TESTREL_A, Lsn(lsn), size as u32)?;
+            writer.advance_last_record_lsn(Lsn(lsn));
             assert_eq!(
                 tline.get_relish_size(TESTREL_A, Lsn(lsn))?.unwrap(),
                 size as u32
@@ -618,17 +679,18 @@ mod tests {
     #[test]
     fn test_list_rels_drop() -> Result<()> {
         let repo = RepoHarness::create("test_list_rels_drop")?.load();
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
         const TESTDB: u32 = 111;
 
         // Import initial dummy checkpoint record, otherwise the get_timeline() call
         // after branching fails below
-        tline.put_page_image(RelishTag::Checkpoint, 0, Lsn(0x10), ZERO_CHECKPOINT.clone())?;
+        writer.put_page_image(RelishTag::Checkpoint, 0, Lsn(0x10), ZERO_CHECKPOINT.clone())?;
 
         // Create a relation on the timeline
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
 
-        tline.advance_last_record_lsn(Lsn(0x30));
+        writer.advance_last_record_lsn(Lsn(0x30));
 
         // Check that list_rels() lists it after LSN 2, but no before it
         assert!(!tline.list_rels(0, TESTDB, Lsn(0x10))?.contains(&TESTREL_A));
@@ -638,14 +700,17 @@ mod tests {
         // Create a branch, check that the relation is visible there
         repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x30))?;
         let newtline = repo.get_timeline(NEW_TIMELINE_ID)?;
+        let new_writer = newtline.writer();
 
         assert!(newtline
             .list_rels(0, TESTDB, Lsn(0x30))?
             .contains(&TESTREL_A));
 
         // Drop it on the branch
-        newtline.drop_relish(TESTREL_A, Lsn(0x40))?;
-        newtline.advance_last_record_lsn(Lsn(0x40));
+        new_writer.drop_relish(TESTREL_A, Lsn(0x40))?;
+        new_writer.advance_last_record_lsn(Lsn(0x40));
+
+        drop(new_writer);
 
         // Check that it's no longer listed on the branch after the point where it was dropped
         assert!(newtline
@@ -656,7 +721,7 @@ mod tests {
             .contains(&TESTREL_A));
 
         // Run checkpoint and garbage collection and check that it's still not visible
-        newtline.checkpoint()?;
+        newtline.checkpoint(CheckpointConfig::Forced)?;
         repo.gc_iteration(Some(NEW_TIMELINE_ID), 0, true)?;
 
         assert!(!newtline
@@ -672,29 +737,31 @@ mod tests {
     #[test]
     fn test_branch() -> Result<()> {
         let repo = RepoHarness::create("test_branch")?.load();
-        let tline = repo.create_empty_timeline(TIMELINE_ID)?;
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        let writer = tline.writer();
 
         // Import initial dummy checkpoint record, otherwise the get_timeline() call
         // after branching fails below
-        tline.put_page_image(RelishTag::Checkpoint, 0, Lsn(0x10), ZERO_CHECKPOINT.clone())?;
+        writer.put_page_image(RelishTag::Checkpoint, 0, Lsn(0x10), ZERO_CHECKPOINT.clone())?;
 
         // Create a relation on the timeline
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x30), TEST_IMG("foo blk 0 at 3"))?;
-        tline.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("foo blk 0 at 4"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x20), TEST_IMG("foo blk 0 at 2"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x30), TEST_IMG("foo blk 0 at 3"))?;
+        writer.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("foo blk 0 at 4"))?;
 
         // Create another relation
-        tline.put_page_image(TESTREL_B, 0, Lsn(0x20), TEST_IMG("foobar blk 0 at 2"))?;
+        writer.put_page_image(TESTREL_B, 0, Lsn(0x20), TEST_IMG("foobar blk 0 at 2"))?;
 
-        tline.advance_last_record_lsn(Lsn(0x40));
+        writer.advance_last_record_lsn(Lsn(0x40));
         assert_current_logical_size(&tline, Lsn(0x40));
 
         // Branch the history, modify relation differently on the new timeline
         repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x30))?;
         let newtline = repo.get_timeline(NEW_TIMELINE_ID)?;
+        let new_writer = newtline.writer();
 
-        newtline.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("bar blk 0 at 4"))?;
-        newtline.advance_last_record_lsn(Lsn(0x40));
+        new_writer.put_page_image(TESTREL_A, 0, Lsn(0x40), TEST_IMG("bar blk 0 at 4"))?;
+        new_writer.advance_last_record_lsn(Lsn(0x40));
 
         // Check page contents on both branches
         assert_eq!(
@@ -719,16 +786,178 @@ mod tests {
         Ok(())
     }
 
+    fn make_some_layers(tline: &Arc<dyn Timeline>, start_lsn: Lsn) -> Result<()> {
+        let mut lsn = start_lsn;
+        {
+            let writer = tline.writer();
+            // Create a relation on the timeline
+            writer.put_page_image(
+                TESTREL_A,
+                0,
+                lsn,
+                TEST_IMG(&format!("foo blk 0 at {}", lsn)),
+            )?;
+            lsn += 0x10;
+            writer.put_page_image(
+                TESTREL_A,
+                0,
+                lsn,
+                TEST_IMG(&format!("foo blk 0 at {}", lsn)),
+            )?;
+            writer.advance_last_record_lsn(lsn);
+        }
+        tline.checkpoint(CheckpointConfig::Forced)?;
+        {
+            let writer = tline.writer();
+            lsn += 0x10;
+            writer.put_page_image(
+                TESTREL_A,
+                0,
+                lsn,
+                TEST_IMG(&format!("foo blk 0 at {}", lsn)),
+            )?;
+            lsn += 0x10;
+            writer.put_page_image(
+                TESTREL_A,
+                0,
+                lsn,
+                TEST_IMG(&format!("foo blk 0 at {}", lsn)),
+            )?;
+            writer.advance_last_record_lsn(lsn);
+        }
+        tline.checkpoint(CheckpointConfig::Forced)
+    }
+
+    #[test]
+    fn test_prohibit_branch_creation_on_garbage_collected_data() -> Result<()> {
+        let repo =
+            RepoHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")?.load();
+
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        make_some_layers(&tline, Lsn(0x20))?;
+
+        // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
+        repo.gc_iteration(Some(TIMELINE_ID), 0x10, false)?;
+
+        // try to branch at lsn 25, should fail because we already garbage collected the data
+        match repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x25)) {
+            Ok(_) => panic!("branching should have failed"),
+            Err(err) => {
+                assert!(err.to_string().contains("invalid branch start lsn"));
+                assert!(err
+                    .source()
+                    .unwrap()
+                    .to_string()
+                    .contains("we might've already garbage collected needed data"))
+            }
+        }
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> Result<()> {
+        let repo = RepoHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")?.load();
+
+        repo.create_empty_timeline(TIMELINE_ID, Lsn(0x50))?;
+        // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
+        match repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x25)) {
+            Ok(_) => panic!("branching should have failed"),
+            Err(err) => {
+                assert!(&err.to_string().contains("invalid branch start lsn"));
+                assert!(&err
+                    .source()
+                    .unwrap()
+                    .to_string()
+                    .contains("is earlier than initdb lsn"));
+            }
+        }
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_prohibit_get_page_at_lsn_for_garbage_collected_pages() -> Result<()> {
+        let repo =
+            RepoHarness::create("test_prohibit_get_page_at_lsn_for_garbage_collected_pages")?
+                .load();
+
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+        make_some_layers(&tline, Lsn(0x20))?;
+
+        repo.gc_iteration(Some(TIMELINE_ID), 0x10, false)?;
+
+        match tline.get_page_at_lsn(TESTREL_A, 0, Lsn(0x25)) {
+            Ok(_) => panic!("request for page should have failed"),
+            Err(err) => assert!(err
+                .to_string()
+                .contains("tried to request a page version that was garbage collected")),
+        }
+        Ok(())
+    }
+
+    #[test]
+    fn test_retain_data_in_parent_which_is_needed_for_child() -> Result<()> {
+        let repo =
+            RepoHarness::create("test_retain_data_in_parent_which_is_needed_for_child")?.load();
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+
+        make_some_layers(&tline, Lsn(0x20))?;
+
+        repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x40))?;
+        let newtline = repo.get_timeline(NEW_TIMELINE_ID)?;
+
+        // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
+        repo.gc_iteration(Some(TIMELINE_ID), 0x10, false)?;
+        assert!(newtline.get_page_at_lsn(TESTREL_A, 0, Lsn(0x25)).is_ok());
+
+        Ok(())
+    }
+
+    #[test]
+    fn test_parent_keeps_data_forever_after_branching() -> Result<()> {
+        let harness = RepoHarness::create("test_parent_keeps_data_forever_after_branching")?;
+        let repo = harness.load();
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
+
+        make_some_layers(&tline, Lsn(0x20))?;
+
+        repo.branch_timeline(TIMELINE_ID, NEW_TIMELINE_ID, Lsn(0x40))?;
+        let newtline = repo.get_timeline(NEW_TIMELINE_ID)?;
+
+        make_some_layers(&newtline, Lsn(0x60))?;
+
+        // run gc on parent
+        repo.gc_iteration(Some(TIMELINE_ID), 0x10, false)?;
+
+        // check that the layer in parent before the branching point is still there
+        let tline_dir = harness.conf.timeline_path(&TIMELINE_ID, &harness.tenant_id);
+
+        let expected_image_layer_path = tline_dir.join(format!(
+            "rel_{}_{}_{}_{}_{}_{:016X}_{:016X}",
+            TESTREL_A_REL_TAG.spcnode,
+            TESTREL_A_REL_TAG.dbnode,
+            TESTREL_A_REL_TAG.relnode,
+            TESTREL_A_REL_TAG.forknum,
+            0, // seg is 0
+            0x20,
+            0x30,
+        ));
+        assert!(fs::metadata(&expected_image_layer_path).is_ok());
+
+        Ok(())
+    }
+
     #[test]
     fn corrupt_metadata() -> Result<()> {
         const TEST_NAME: &str = "corrupt_metadata";
         let harness = RepoHarness::create(TEST_NAME)?;
         let repo = harness.load();
 
-        repo.create_empty_timeline(TIMELINE_ID)?;
+        repo.create_empty_timeline(TIMELINE_ID, Lsn(0))?;
         drop(repo);
 
-        let metadata_path = harness.timeline_path(&TIMELINE_ID).join("metadata");
+        let metadata_path = harness.timeline_path(&TIMELINE_ID).join(METADATA_FILE_NAME);
 
         assert!(metadata_path.is_file());
 
@@ -739,7 +968,11 @@ mod tests {
 
         let new_repo = harness.load();
         let err = new_repo.get_timeline(TIMELINE_ID).err().unwrap();
-        assert!(err.to_string().contains("checksum"));
+        assert_eq!(err.to_string(), "failed to load metadata");
+        assert_eq!(
+            err.source().unwrap().to_string(),
+            "metadata checksum mismatch"
+        );
 
         Ok(())
     }
@@ -750,7 +983,12 @@ mod tests {
         let harness = RepoHarness::create(TEST_NAME)?;
         let repo = harness.load();
 
-        repo.create_empty_timeline(TIMELINE_ID)?;
+        // Create a timeline with disk_consistent_lsn = 8000
+        let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0x8000))?;
+        let writer = tline.writer();
+        writer.advance_last_record_lsn(Lsn(0x8000));
+        drop(writer);
+        repo.checkpoint_iteration(CheckpointConfig::Forced)?;
         drop(repo);
 
         let timeline_path = harness.timeline_path(&TIMELINE_ID);
@@ -764,68 +1002,86 @@ mod tests {
             Ok(())
         };
 
-        let image_filename = format!("pg_control_0_{:016X}", 8000);
-        let delta_filename = format!("pg_control_0_{:016X}_{:016X}", 8000, 8008);
-
-        make_empty_file(&image_filename)?;
-        make_empty_file(&delta_filename)?;
-
-        let new_repo = harness.load();
-        new_repo.get_timeline(TIMELINE_ID).unwrap();
-        drop(new_repo);
-
-        let check_old = |filename: &str, num: u32| {
+        // Helper function to check that a relation file exists, and a corresponding
+        // <filename>.0.old file does not.
+        let assert_exists = |filename: &str| {
             let path = timeline_path.join(filename);
-            assert!(!path.exists());
+            assert!(path.exists(), "file {} was removed", filename);
 
-            let backup_path = timeline_path.join(format!("{}.{}.old", filename, num));
-            assert!(backup_path.exists());
+            // Check that there is no .old file
+            let backup_path = timeline_path.join(format!("{}.0.old", filename));
+            assert!(
+                !backup_path.exists(),
+                "unexpected backup file {}",
+                backup_path.display()
+            );
         };
 
-        check_old(&image_filename, 0);
-        check_old(&delta_filename, 0);
+        // Helper function to check that a relation file does *not* exists, and a corresponding
+        // <filename>.<num>.old file does.
+        let assert_is_renamed = |filename: &str, num: u32| {
+            let path = timeline_path.join(filename);
+            assert!(
+                !path.exists(),
+                "file {} was not removed as expected",
+                filename
+            );
 
-        make_empty_file(&image_filename)?;
-        make_empty_file(&delta_filename)?;
+            let backup_path = timeline_path.join(format!("{}.{}.old", filename, num));
+            assert!(
+                backup_path.exists(),
+                "backup file {} was not created",
+                backup_path.display()
+            );
+        };
+
+        // These files are considered to be in the future and will be renamed out
+        // of the way
+        let future_filenames = vec![
+            format!("pg_control_0_{:016X}", 0x8001),
+            format!("pg_control_0_{:016X}_{:016X}", 0x8001, 0x8008),
+        ];
+        // But these are not:
+        let past_filenames = vec![
+            format!("pg_control_0_{:016X}", 0x8000),
+            format!("pg_control_0_{:016X}_{:016X}", 0x7000, 0x8001),
+        ];
+
+        for filename in future_filenames.iter().chain(past_filenames.iter()) {
+            make_empty_file(filename)?;
+        }
+
+        // Load the timeline. This will cause the files in the "future" to be renamed
+        // away.
+        let new_repo = harness.load();
+        new_repo.get_timeline(TIMELINE_ID).unwrap();
+        drop(new_repo);
+
+        for filename in future_filenames.iter() {
+            assert_is_renamed(filename, 0);
+        }
+        for filename in past_filenames.iter() {
+            assert_exists(filename);
+        }
+
+        // Create the future files again, and load again. They should be renamed to
+        // *.1.old this time.
+        for filename in future_filenames.iter() {
+            make_empty_file(filename)?;
+        }
 
         let new_repo = harness.load();
         new_repo.get_timeline(TIMELINE_ID).unwrap();
         drop(new_repo);
 
-        check_old(&image_filename, 0);
-        check_old(&delta_filename, 0);
-        check_old(&image_filename, 1);
-        check_old(&delta_filename, 1);
+        for filename in future_filenames.iter() {
+            assert_is_renamed(filename, 0);
+            assert_is_renamed(filename, 1);
+        }
+        for filename in past_filenames.iter() {
+            assert_exists(filename);
+        }
 
         Ok(())
     }
-
-    // Mock WAL redo manager that doesn't do much
-    struct TestRedoManager;
-
-    impl WalRedoManager for TestRedoManager {
-        fn request_redo(
-            &self,
-            rel: RelishTag,
-            blknum: u32,
-            lsn: Lsn,
-            base_img: Option<Bytes>,
-            records: Vec<WALRecord>,
-        ) -> Result<Bytes, WalRedoError> {
-            let s = format!(
-                "redo for {} blk {} to get to {}, with {} and {} records",
-                rel,
-                blknum,
-                lsn,
-                if base_img.is_some() {
-                    "base image"
-                } else {
-                    "no base image"
-                },
-                records.len()
-            );
-            println!("{}", s);
-            Ok(TEST_IMG(&s))
-        }
-    }
 }

pageserver/src/restore_local_repo.rs

@@ -2,17 +2,17 @@
 //! Import data and WAL from a PostgreSQL data directory and WAL segments into
 //! zenith Timeline.
 //!
-use log::*;
 use postgres_ffi::nonrelfile_utils::clogpage_precedes;
 use postgres_ffi::nonrelfile_utils::slru_may_delete_clogsegment;
 use std::cmp::min;
 use std::fs;
 use std::fs::File;
-use std::io::Read;
-use std::path::Path;
+use std::io::{Read, Seek, SeekFrom};
+use std::path::{Path, PathBuf};
 
-use anyhow::{bail, Result};
-use bytes::{Buf, Bytes};
+use anyhow::{anyhow, bail, Result};
+use bytes::{Buf, Bytes, BytesMut};
+use tracing::*;
 
 use crate::relish::*;
 use crate::repository::*;
@@ -34,9 +34,11 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; 8192]);
 ///
 pub fn import_timeline_from_postgres_datadir(
     path: &Path,
-    timeline: &dyn Timeline,
+    writer: &dyn TimelineWriter,
     lsn: Lsn,
 ) -> Result<()> {
+    let mut pg_control: Option<ControlFileData> = None;
+
     // Scan 'global'
     for direntry in fs::read_dir(path.join("global"))? {
         let direntry = direntry?;
@@ -44,10 +46,10 @@ pub fn import_timeline_from_postgres_datadir(
             None => continue,
 
             Some("pg_control") => {
-                import_control_file(timeline, lsn, &direntry.path())?;
+                pg_control = Some(import_control_file(writer, lsn, &direntry.path())?);
             }
             Some("pg_filenode.map") => import_nonrel_file(
-                timeline,
+                writer,
                 lsn,
                 RelishTag::FileNodeMap {
                     spcnode: pg_constants::GLOBALTABLESPACE_OID,
@@ -59,7 +61,7 @@ pub fn import_timeline_from_postgres_datadir(
             // Load any relation files into the page server
             _ => import_relfile(
                 &direntry.path(),
-                timeline,
+                writer,
                 lsn,
                 pg_constants::GLOBALTABLESPACE_OID,
                 0,
@@ -86,7 +88,7 @@ pub fn import_timeline_from_postgres_datadir(
 
                 Some("PG_VERSION") => continue,
                 Some("pg_filenode.map") => import_nonrel_file(
-                    timeline,
+                    writer,
                     lsn,
                     RelishTag::FileNodeMap {
                         spcnode: pg_constants::DEFAULTTABLESPACE_OID,
@@ -98,7 +100,7 @@ pub fn import_timeline_from_postgres_datadir(
                 // Load any relation files into the page server
                 _ => import_relfile(
                     &direntry.path(),
-                    timeline,
+                    writer,
                     lsn,
                     pg_constants::DEFAULTTABLESPACE_OID,
                     dboid,
@@ -108,24 +110,36 @@ pub fn import_timeline_from_postgres_datadir(
     }
     for entry in fs::read_dir(path.join("pg_xact"))? {
         let entry = entry?;
-        import_slru_file(timeline, lsn, SlruKind::Clog, &entry.path())?;
+        import_slru_file(writer, lsn, SlruKind::Clog, &entry.path())?;
     }
     for entry in fs::read_dir(path.join("pg_multixact").join("members"))? {
         let entry = entry?;
-        import_slru_file(timeline, lsn, SlruKind::MultiXactMembers, &entry.path())?;
+        import_slru_file(writer, lsn, SlruKind::MultiXactMembers, &entry.path())?;
     }
     for entry in fs::read_dir(path.join("pg_multixact").join("offsets"))? {
         let entry = entry?;
-        import_slru_file(timeline, lsn, SlruKind::MultiXactOffsets, &entry.path())?;
+        import_slru_file(writer, lsn, SlruKind::MultiXactOffsets, &entry.path())?;
     }
     for entry in fs::read_dir(path.join("pg_twophase"))? {
         let entry = entry?;
         let xid = u32::from_str_radix(entry.path().to_str().unwrap(), 16)?;
-        import_nonrel_file(timeline, lsn, RelishTag::TwoPhase { xid }, &entry.path())?;
+        import_nonrel_file(writer, lsn, RelishTag::TwoPhase { xid }, &entry.path())?;
     }
     // TODO: Scan pg_tblspc
 
-    timeline.advance_last_record_lsn(lsn);
+    writer.advance_last_record_lsn(lsn);
+
+    // Import WAL. This is needed even when starting from a shutdown checkpoint, because
+    // this reads the checkpoint record itself, advancing the tip of the timeline to
+    // *after* the checkpoint record. And crucially, it initializes the 'prev_lsn'
+    let pg_control = pg_control.ok_or_else(|| anyhow!("pg_control file not found"))?;
+    import_wal(
+        &path.join("pg_wal"),
+        writer,
+        Lsn(pg_control.checkPointCopy.redo),
+        lsn,
+        &mut pg_control.checkPointCopy.clone(),
+    )?;
 
     Ok(())
 }
@@ -133,12 +147,13 @@ pub fn import_timeline_from_postgres_datadir(
 // subroutine of import_timeline_from_postgres_datadir(), to load one relation file.
 fn import_relfile(
     path: &Path,
-    timeline: &dyn Timeline,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     spcoid: Oid,
     dboid: Oid,
 ) -> Result<()> {
     // Does it look like a relation file?
+    trace!("importing rel file {}", path.display());
 
     let p = parse_relfilename(path.file_name().unwrap().to_str().unwrap());
     if let Err(e) = p {
@@ -166,14 +181,14 @@ fn import_relfile(
             }
 
             // TODO: UnexpectedEof is expected
-            Err(e) => match e.kind() {
+            Err(err) => match err.kind() {
                 std::io::ErrorKind::UnexpectedEof => {
                     // reached EOF. That's expected.
                     // FIXME: maybe check that we read the full length of the file?
                     break;
                 }
                 _ => {
-                    bail!("error reading file {}: {:#}", path.display(), e);
+                    bail!("error reading file {}: {:#}", path.display(), err);
                 }
             },
         };
@@ -190,7 +205,7 @@ fn import_relfile(
 /// are just slurped into the repository as one blob.
 ///
 fn import_nonrel_file(
-    timeline: &dyn Timeline,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     tag: RelishTag,
     path: &Path,
@@ -200,7 +215,7 @@ fn import_nonrel_file(
     // read the whole file
     file.read_to_end(&mut buffer)?;
 
-    info!("importing non-rel file {}", path.display());
+    trace!("importing non-rel file {}", path.display());
 
     timeline.put_page_image(tag, 0, lsn, Bytes::copy_from_slice(&buffer[..]))?;
     Ok(())
@@ -211,13 +226,17 @@ fn import_nonrel_file(
 ///
 /// The control file is imported as is, but we also extract the checkpoint record
 /// from it and store it separated.
-fn import_control_file(timeline: &dyn Timeline, lsn: Lsn, path: &Path) -> Result<()> {
+fn import_control_file(
+    timeline: &dyn TimelineWriter,
+    lsn: Lsn,
+    path: &Path,
+) -> Result<ControlFileData> {
     let mut file = File::open(path)?;
     let mut buffer = Vec::new();
     // read the whole file
     file.read_to_end(&mut buffer)?;
 
-    info!("importing control file {}", path.display());
+    trace!("importing control file {}", path.display());
 
     // Import it as ControlFile
     timeline.put_page_image(
@@ -232,19 +251,24 @@ fn import_control_file(timeline: &dyn Timeline, lsn: Lsn, path: &Path) -> Result
     let checkpoint_bytes = pg_control.checkPointCopy.encode();
     timeline.put_page_image(RelishTag::Checkpoint, 0, lsn, checkpoint_bytes)?;
 
-    Ok(())
+    Ok(pg_control)
 }
 
 ///
 /// Import an SLRU segment file
 ///
-fn import_slru_file(timeline: &dyn Timeline, lsn: Lsn, slru: SlruKind, path: &Path) -> Result<()> {
+fn import_slru_file(
+    timeline: &dyn TimelineWriter,
+    lsn: Lsn,
+    slru: SlruKind,
+    path: &Path,
+) -> Result<()> {
     // Does it look like an SLRU file?
     let mut file = File::open(path)?;
     let mut buf: [u8; 8192] = [0u8; 8192];
     let segno = u32::from_str_radix(path.file_name().unwrap().to_str().unwrap(), 16)?;
 
-    info!("importing slru file {}", path.display());
+    trace!("importing slru file {}", path.display());
 
     let mut rpageno = 0;
     loop {
@@ -260,14 +284,14 @@ fn import_slru_file(timeline: &dyn Timeline, lsn: Lsn, slru: SlruKind, path: &Pa
             }
 
             // TODO: UnexpectedEof is expected
-            Err(e) => match e.kind() {
+            Err(err) => match err.kind() {
                 std::io::ErrorKind::UnexpectedEof => {
                     // reached EOF. That's expected.
                     // FIXME: maybe check that we read the full length of the file?
                     break;
                 }
                 _ => {
-                    bail!("error reading file {}: {:#}", path.display(), e);
+                    bail!("error reading file {}: {:#}", path.display(), err);
                 }
             },
         };
@@ -279,19 +303,119 @@ fn import_slru_file(timeline: &dyn Timeline, lsn: Lsn, slru: SlruKind, path: &Pa
     Ok(())
 }
 
+/// Scan PostgreSQL WAL files in given directory and load all records between
+/// 'startpoint' and 'endpoint' into the repository.
+fn import_wal(
+    walpath: &Path,
+    timeline: &dyn TimelineWriter,
+    startpoint: Lsn,
+    endpoint: Lsn,
+    checkpoint: &mut CheckPoint,
+) -> Result<()> {
+    let mut waldecoder = WalStreamDecoder::new(startpoint);
+
+    let mut segno = startpoint.segment_number(pg_constants::WAL_SEGMENT_SIZE);
+    let mut offset = startpoint.segment_offset(pg_constants::WAL_SEGMENT_SIZE);
+    let mut last_lsn = startpoint;
+
+    while last_lsn <= endpoint {
+        // FIXME: assume postgresql tli 1 for now
+        let filename = XLogFileName(1, segno, pg_constants::WAL_SEGMENT_SIZE);
+        let mut buf = Vec::new();
+
+        // Read local file
+        let mut path = walpath.join(&filename);
+
+        // It could be as .partial
+        if !PathBuf::from(&path).exists() {
+            path = walpath.join(filename + ".partial");
+        }
+
+        // Slurp the WAL file
+        let mut file = File::open(&path)?;
+
+        if offset > 0 {
+            file.seek(SeekFrom::Start(offset as u64))?;
+        }
+
+        let nread = file.read_to_end(&mut buf)?;
+        if nread != pg_constants::WAL_SEGMENT_SIZE - offset as usize {
+            // Maybe allow this for .partial files?
+            error!("read only {} bytes from WAL file", nread);
+        }
+
+        waldecoder.feed_bytes(&buf);
+
+        let mut nrecords = 0;
+        while last_lsn <= endpoint {
+            if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
+                let mut checkpoint_modified = false;
+
+                let decoded = decode_wal_record(recdata.clone());
+                save_decoded_record(
+                    checkpoint,
+                    &mut checkpoint_modified,
+                    timeline,
+                    &decoded,
+                    recdata,
+                    lsn,
+                )?;
+                last_lsn = lsn;
+
+                if checkpoint_modified {
+                    let checkpoint_bytes = checkpoint.encode();
+                    timeline.put_page_image(
+                        RelishTag::Checkpoint,
+                        0,
+                        last_lsn,
+                        checkpoint_bytes,
+                    )?;
+                }
+
+                // Now that this record has been fully handled, including updating the
+                // checkpoint data, let the repository know that it is up-to-date to this LSN
+                timeline.advance_last_record_lsn(last_lsn);
+                nrecords += 1;
+
+                trace!("imported record at {} (end {})", lsn, endpoint);
+            }
+        }
+
+        debug!("imported {} records up to {}", nrecords, last_lsn);
+
+        segno += 1;
+        offset = 0;
+    }
+
+    if last_lsn != startpoint {
+        debug!(
+            "reached end of WAL at {}, updating checkpoint info",
+            last_lsn
+        );
+
+        timeline.advance_last_record_lsn(last_lsn);
+    } else {
+        info!("no WAL to import at {}", last_lsn);
+    }
+
+    Ok(())
+}
+
 ///
 /// Helper function to parse a WAL record and call the Timeline's PUT functions for all the
 /// relations/pages that the record affects.
 ///
 pub fn save_decoded_record(
     checkpoint: &mut CheckPoint,
-    timeline: &dyn Timeline,
+    checkpoint_modified: &mut bool,
+    timeline: &dyn TimelineWriter,
     decoded: &DecodedWALRecord,
     recdata: Bytes,
     lsn: Lsn,
 ) -> Result<()> {
-    checkpoint.update_next_xid(decoded.xl_xid);
-
+    if checkpoint.update_next_xid(decoded.xl_xid) {
+        *checkpoint_modified = true;
+    }
     // Iterate through all the blocks that the record modifies, and
     // "put" a separate copy of the record for each block.
     for blk in decoded.blocks.iter() {
@@ -302,14 +426,43 @@ pub fn save_decoded_record(
             forknum: blk.forknum as u8,
         });
 
-        let rec = WALRecord {
-            lsn,
-            will_init: blk.will_init || blk.apply_image,
-            rec: recdata.clone(),
-            main_data_offset: decoded.main_data_offset as u32,
-        };
+        //
+        // Instead of storing full-page-image WAL record,
+        // it is better to store extracted image: we can skip wal-redo
+        // in this case. Also some FPI records may contain multiple (up to 32) pages,
+        // so them have to be copied multiple times.
+        //
+        if blk.apply_image
+            && blk.has_image
+            && decoded.xl_rmid == pg_constants::RM_XLOG_ID
+            && (decoded.xl_info == pg_constants::XLOG_FPI
+                || decoded.xl_info == pg_constants::XLOG_FPI_FOR_HINT)
+            // compression of WAL is not yet supported: fall back to storing the original WAL record
+            && (blk.bimg_info & pg_constants::BKPIMAGE_IS_COMPRESSED) == 0
+        {
+            // Extract page image from FPI record
+            let img_len = blk.bimg_len as usize;
+            let img_offs = blk.bimg_offset as usize;
+            let mut image = BytesMut::with_capacity(pg_constants::BLCKSZ as usize);
+            image.extend_from_slice(&recdata[img_offs..img_offs + img_len]);
 
-        timeline.put_wal_record(tag, blk.blkno, rec)?;
+            if blk.hole_length != 0 {
+                let tail = image.split_off(blk.hole_offset as usize);
+                image.resize(image.len() + blk.hole_length as usize, 0u8);
+                image.unsplit(tail);
+            }
+            image[0..4].copy_from_slice(&((lsn.0 >> 32) as u32).to_le_bytes());
+            image[4..8].copy_from_slice(&(lsn.0 as u32).to_le_bytes());
+            assert_eq!(image.len(), pg_constants::BLCKSZ as usize);
+            timeline.put_page_image(tag, blk.blkno, lsn, image.freeze())?;
+        } else {
+            let rec = WALRecord {
+                will_init: blk.will_init || blk.apply_image,
+                rec: recdata.clone(),
+                main_data_offset: decoded.main_data_offset as u32,
+            };
+            timeline.put_wal_record(lsn, tag, blk.blkno, rec)?;
+        }
     }
 
     let mut buf = decoded.record.clone();
@@ -374,7 +527,7 @@ pub fn save_decoded_record(
         } else {
             assert!(info == pg_constants::CLOG_TRUNCATE);
             let xlrec = XlClogTruncate::decode(&mut buf);
-            save_clog_truncate_record(checkpoint, timeline, lsn, &xlrec)?;
+            save_clog_truncate_record(checkpoint, checkpoint_modified, timeline, lsn, &xlrec)?;
         }
     } else if decoded.xl_rmid == pg_constants::RM_XACT_ID {
         let info = decoded.xl_info & pg_constants::XLOG_XACT_OPMASK;
@@ -443,10 +596,17 @@ pub fn save_decoded_record(
             )?;
         } else if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
             let xlrec = XlMultiXactCreate::decode(&mut buf);
-            save_multixact_create_record(checkpoint, timeline, lsn, &xlrec, decoded)?;
+            save_multixact_create_record(
+                checkpoint,
+                checkpoint_modified,
+                timeline,
+                lsn,
+                &xlrec,
+                decoded,
+            )?;
         } else if info == pg_constants::XLOG_MULTIXACT_TRUNCATE_ID {
             let xlrec = XlMultiXactTruncate::decode(&mut buf);
-            save_multixact_truncate_record(checkpoint, timeline, lsn, &xlrec)?;
+            save_multixact_truncate_record(checkpoint, checkpoint_modified, timeline, lsn, &xlrec)?;
         }
     } else if decoded.xl_rmid == pg_constants::RM_RELMAP_ID {
         let xlrec = XlRelmapUpdate::decode(&mut buf);
@@ -455,7 +615,10 @@ pub fn save_decoded_record(
         let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
         if info == pg_constants::XLOG_NEXTOID {
             let next_oid = buf.get_u32_le();
-            checkpoint.nextOid = next_oid;
+            if checkpoint.nextOid != next_oid {
+                checkpoint.nextOid = next_oid;
+                *checkpoint_modified = true;
+            }
         } else if info == pg_constants::XLOG_CHECKPOINT_ONLINE
             || info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
         {
@@ -471,6 +634,7 @@ pub fn save_decoded_record(
             );
             if (checkpoint.oldestXid.wrapping_sub(xlog_checkpoint.oldestXid) as i32) < 0 {
                 checkpoint.oldestXid = xlog_checkpoint.oldestXid;
+                *checkpoint_modified = true;
             }
         }
     }
@@ -478,7 +642,11 @@ pub fn save_decoded_record(
 }
 
 /// Subroutine of save_decoded_record(), to handle an XLOG_DBASE_CREATE record.
-fn save_xlog_dbase_create(timeline: &dyn Timeline, lsn: Lsn, rec: &XlCreateDatabase) -> Result<()> {
+fn save_xlog_dbase_create(
+    timeline: &dyn TimelineWriter,
+    lsn: Lsn,
+    rec: &XlCreateDatabase,
+) -> Result<()> {
     let db_id = rec.db_id;
     let tablespace_id = rec.tablespace_id;
     let src_db_id = rec.src_db_id;
@@ -555,7 +723,11 @@ fn save_xlog_dbase_create(timeline: &dyn Timeline, lsn: Lsn, rec: &XlCreateDatab
 /// Subroutine of save_decoded_record(), to handle an XLOG_SMGR_TRUNCATE record.
 ///
 /// This is the same logic as in PostgreSQL's smgr_redo() function.
-fn save_xlog_smgr_truncate(timeline: &dyn Timeline, lsn: Lsn, rec: &XlSmgrTruncate) -> Result<()> {
+fn save_xlog_smgr_truncate(
+    timeline: &dyn TimelineWriter,
+    lsn: Lsn,
+    rec: &XlSmgrTruncate,
+) -> Result<()> {
     let spcnode = rec.rnode.spcnode;
     let dbnode = rec.rnode.dbnode;
     let relnode = rec.rnode.relnode;
@@ -617,7 +789,7 @@ fn save_xlog_smgr_truncate(timeline: &dyn Timeline, lsn: Lsn, rec: &XlSmgrTrunca
 /// Subroutine of save_decoded_record(), to handle an XLOG_XACT_* records.
 ///
 fn save_xact_record(
-    timeline: &dyn Timeline,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     parsed: &XlXactParsedRecord,
     decoded: &DecodedWALRecord,
@@ -628,12 +800,12 @@ fn save_xact_record(
     let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
     let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
     let rec = WALRecord {
-        lsn,
         will_init: false,
         rec: decoded.record.clone(),
         main_data_offset: decoded.main_data_offset as u32,
     };
     timeline.put_wal_record(
+        lsn,
         RelishTag::Slru {
             slru: SlruKind::Clog,
             segno,
@@ -649,6 +821,7 @@ fn save_xact_record(
             let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
             let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
             timeline.put_wal_record(
+                lsn,
                 RelishTag::Slru {
                     slru: SlruKind::Clog,
                     segno,
@@ -674,7 +847,8 @@ fn save_xact_record(
 
 fn save_clog_truncate_record(
     checkpoint: &mut CheckPoint,
-    timeline: &dyn Timeline,
+    checkpoint_modified: &mut bool,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     xlrec: &XlClogTruncate,
 ) -> Result<()> {
@@ -692,6 +866,7 @@ fn save_clog_truncate_record(
     // TODO Figure out if there will be any issues with replica.
     checkpoint.oldestXid = xlrec.oldest_xid;
     checkpoint.oldestXidDB = xlrec.oldest_xid_db;
+    *checkpoint_modified = true;
 
     // TODO Treat AdvanceOldestClogXid() or write a comment why we don't need it
 
@@ -734,13 +909,13 @@ fn save_clog_truncate_record(
 
 fn save_multixact_create_record(
     checkpoint: &mut CheckPoint,
-    timeline: &dyn Timeline,
+    checkpoint_modified: &mut bool,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     xlrec: &XlMultiXactCreate,
     decoded: &DecodedWALRecord,
 ) -> Result<()> {
     let rec = WALRecord {
-        lsn,
         will_init: false,
         rec: decoded.record.clone(),
         main_data_offset: decoded.main_data_offset as u32,
@@ -749,6 +924,7 @@ fn save_multixact_create_record(
     let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
     let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
     timeline.put_wal_record(
+        lsn,
         RelishTag::Slru {
             slru: SlruKind::MultiXactOffsets,
             segno,
@@ -768,6 +944,7 @@ fn save_multixact_create_record(
         let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
         let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
         timeline.put_wal_record(
+            lsn,
             RelishTag::Slru {
                 slru: SlruKind::MultiXactMembers,
                 segno,
@@ -790,9 +967,11 @@ fn save_multixact_create_record(
     }
     if xlrec.mid >= checkpoint.nextMulti {
         checkpoint.nextMulti = xlrec.mid + 1;
+        *checkpoint_modified = true;
     }
     if xlrec.moff + xlrec.nmembers > checkpoint.nextMultiOffset {
         checkpoint.nextMultiOffset = xlrec.moff + xlrec.nmembers;
+        *checkpoint_modified = true;
     }
     let max_mbr_xid = xlrec.members.iter().fold(0u32, |acc, mbr| {
         if mbr.xid.wrapping_sub(acc) as i32 > 0 {
@@ -802,18 +981,22 @@ fn save_multixact_create_record(
         }
     });
 
-    checkpoint.update_next_xid(max_mbr_xid);
+    if checkpoint.update_next_xid(max_mbr_xid) {
+        *checkpoint_modified = true;
+    }
     Ok(())
 }
 
 fn save_multixact_truncate_record(
     checkpoint: &mut CheckPoint,
-    timeline: &dyn Timeline,
+    checkpoint_modified: &mut bool,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     xlrec: &XlMultiXactTruncate,
 ) -> Result<()> {
     checkpoint.oldestMulti = xlrec.end_trunc_off;
     checkpoint.oldestMultiDB = xlrec.oldest_multi_db;
+    *checkpoint_modified = true;
 
     // PerformMembersTruncation
     let maxsegment: i32 = mx_offset_to_member_segment(pg_constants::MAX_MULTIXACT_OFFSET);
@@ -847,7 +1030,7 @@ fn save_multixact_truncate_record(
 }
 
 fn save_relmap_page(
-    timeline: &dyn Timeline,
+    timeline: &dyn TimelineWriter,
     lsn: Lsn,
     xlrec: &XlRelmapUpdate,
     decoded: &DecodedWALRecord,

pageserver/src/tenant_mgr.rs

@@ -4,39 +4,100 @@
 use crate::branches;
 use crate::layered_repository::LayeredRepository;
 use crate::repository::{Repository, Timeline};
+use crate::tenant_threads;
 use crate::walredo::PostgresRedoManager;
 use crate::PageServerConf;
 use anyhow::{anyhow, bail, Context, Result};
 use lazy_static::lazy_static;
-use log::info;
-use std::collections::hash_map::Entry;
+use log::*;
+use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
+use std::fmt;
 use std::fs;
 use std::str::FromStr;
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::{Arc, Mutex, MutexGuard};
 use zenith_utils::zid::{ZTenantId, ZTimelineId};
 
 lazy_static! {
-    static ref REPOSITORY: Mutex<HashMap<ZTenantId, Arc<dyn Repository>>> =
-        Mutex::new(HashMap::new());
+    static ref TENANTS: Mutex<HashMap<ZTenantId, Tenant>> = Mutex::new(HashMap::new());
 }
 
-fn access_repository() -> MutexGuard<'static, HashMap<ZTenantId, Arc<dyn Repository>>> {
-    REPOSITORY.lock().unwrap()
+struct Tenant {
+    state: TenantState,
+    repo: Option<Arc<dyn Repository>>,
 }
 
-pub fn init(conf: &'static PageServerConf) {
-    let mut m = access_repository();
-    for dir_entry in fs::read_dir(conf.tenants_path()).unwrap() {
-        let tenantid =
-            ZTenantId::from_str(dir_entry.unwrap().file_name().to_str().unwrap()).unwrap();
-        let repo = init_repo(conf, tenantid);
-        info!("initialized storage for tenant: {}", &tenantid);
-        m.insert(tenantid, repo);
+#[derive(Debug, Serialize, Deserialize, Clone, Copy, PartialEq, Eq)]
+pub enum TenantState {
+    // This tenant only exists in cloud storage. It cannot be accessed.
+    CloudOnly,
+    // This tenant exists in cloud storage, and we are currently downloading it to local disk.
+    // It cannot be accessed yet, not until it's been fully downloaded to local disk.
+    Downloading,
+    // All data for this tenant is complete on local disk, but we haven't loaded the Repository,
+    // Timeline and Layer structs into memory yet, so it cannot be accessed yet.
+    //Ready,
+    // This tenant exists on local disk, and the layer map has been loaded into memory.
+    // The local disk might have some newer files that don't exist in cloud storage yet.
+    Active,
+    // Tenant is active, but there is no walreceiver connection.
+    Idle,
+    // This tenant exists on local disk, and the layer map has been loaded into memory.
+    // The local disk might have some newer files that don't exist in cloud storage yet.
+    // The tenant cannot be accessed anymore for any reason, but graceful shutdown.
+    Stopping,
+}
+
+/// A remote storage timeline synchronization event, that needs another step
+/// to be fully completed.
+#[derive(Debug)]
+pub enum PostTimelineSyncStep {
+    /// The timeline cannot be synchronized anymore due to some sync issues.
+    /// Needs to be removed from pageserver, to avoid further data diverging.
+    Evict,
+    /// A new timeline got downloaded and needs to be loaded into pageserver.
+    RegisterDownload,
+}
+
+impl fmt::Display for TenantState {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            TenantState::CloudOnly => f.write_str("CloudOnly"),
+            TenantState::Downloading => f.write_str("Downloading"),
+            TenantState::Active => f.write_str("Active"),
+            TenantState::Idle => f.write_str("Idle"),
+            TenantState::Stopping => f.write_str("Stopping"),
+        }
     }
 }
 
-fn init_repo(conf: &'static PageServerConf, tenant_id: ZTenantId) -> Arc<LayeredRepository> {
+fn access_tenants() -> MutexGuard<'static, HashMap<ZTenantId, Tenant>> {
+    TENANTS.lock().unwrap()
+}
+
+static SHUTDOWN_REQUESTED: AtomicBool = AtomicBool::new(false);
+
+pub fn init(conf: &'static PageServerConf) {
+    for dir_entry in fs::read_dir(conf.tenants_path()).unwrap() {
+        let tenantid =
+            ZTenantId::from_str(dir_entry.unwrap().file_name().to_str().unwrap()).unwrap();
+
+        {
+            let mut m = access_tenants();
+            let tenant = Tenant {
+                state: TenantState::CloudOnly,
+                repo: None,
+            };
+            m.insert(tenantid, tenant);
+        }
+
+        init_repo(conf, tenantid);
+        info!("initialized storage for tenant: {}", &tenantid);
+    }
+}
+
+fn init_repo(conf: &'static PageServerConf, tenant_id: ZTenantId) {
     // Set up a WAL redo manager, for applying WAL records.
     let walredo_mgr = PostgresRedoManager::new(conf, tenant_id);
 
@@ -45,32 +106,64 @@ fn init_repo(conf: &'static PageServerConf, tenant_id: ZTenantId) -> Arc<Layered
         conf,
         Arc::new(walredo_mgr),
         tenant_id,
-        true,
+        false,
     ));
-    LayeredRepository::launch_checkpointer_thread(conf, repo.clone());
-    LayeredRepository::launch_gc_thread(conf, repo.clone());
-    repo
+
+    let mut m = access_tenants();
+    let tenant = m.get_mut(&tenant_id).unwrap();
+    tenant.repo = Some(repo);
+    tenant.state = TenantState::Idle;
 }
 
-// TODO kb Currently unused function, will later be used when the relish storage downloads a new layer.
-// Relevant PR: https://github.com/zenithdb/zenith/pull/686
-pub fn register_relish_download(
+pub fn perform_post_timeline_sync_steps(
     conf: &'static PageServerConf,
-    tenant_id: ZTenantId,
-    timeline_id: ZTimelineId,
+    post_sync_steps: HashMap<(ZTenantId, ZTimelineId), PostTimelineSyncStep>,
 ) {
-    log::info!(
-        "Registering new download, tenant id {}, timeline id: {}",
-        tenant_id,
-        timeline_id
-    );
-    match access_repository().entry(tenant_id) {
-        Entry::Occupied(o) => init_timeline(o.get().as_ref(), timeline_id),
-        Entry::Vacant(v) => {
-            log::info!("New repo initialized");
-            let new_repo = init_repo(conf, tenant_id);
-            init_timeline(new_repo.as_ref(), timeline_id);
-            v.insert(new_repo);
+    if post_sync_steps.is_empty() {
+        return;
+    }
+
+    info!("Performing {} post-sync steps", post_sync_steps.len());
+    trace!("Steps: {:?}", post_sync_steps);
+
+    {
+        let mut m = access_tenants();
+        for &(tenant_id, timeline_id) in post_sync_steps.keys() {
+            let tenant = m.entry(tenant_id).or_insert_with(|| Tenant {
+                state: TenantState::Downloading,
+                repo: None,
+            });
+            tenant.state = TenantState::Downloading;
+            match &tenant.repo {
+                Some(repo) => {
+                    init_timeline(repo.as_ref(), timeline_id);
+                    tenant.state = TenantState::Idle;
+                    return;
+                }
+                None => log::warn!("Initialize new repo"),
+            }
+            tenant.state = TenantState::Idle;
+        }
+    }
+
+    for ((tenant_id, timeline_id), post_sync_step) in post_sync_steps {
+        match post_sync_step {
+            PostTimelineSyncStep::Evict => {
+                if let Err(e) = get_repository_for_tenant(tenant_id)
+                    .and_then(|repo| repo.unload_timeline(timeline_id))
+                {
+                    error!(
+                        "Failed to remove repository for tenant {}, timeline {}: {:#}",
+                        tenant_id, timeline_id, e
+                    )
+                }
+            }
+            PostTimelineSyncStep::RegisterDownload => {
+                // init repo updates Tenant state
+                init_repo(conf, tenant_id);
+                let new_repo = get_repository_for_tenant(tenant_id).unwrap();
+                init_timeline(new_repo.as_ref(), timeline_id);
+            }
         }
     }
 }
@@ -82,29 +175,96 @@ fn init_timeline(repo: &dyn Repository, timeline_id: ZTimelineId) {
     }
 }
 
+// Check this flag in the thread loops to know when to exit
+pub fn shutdown_requested() -> bool {
+    SHUTDOWN_REQUESTED.load(Ordering::Relaxed)
+}
+
+pub fn shutdown_all_tenants() -> Result<()> {
+    SHUTDOWN_REQUESTED.swap(true, Ordering::Relaxed);
+
+    let tenantids = list_tenantids()?;
+
+    for tenantid in &tenantids {
+        set_tenant_state(*tenantid, TenantState::Stopping)?;
+    }
+
+    for tenantid in tenantids {
+        // Wait for checkpointer and GC to finish their job
+        tenant_threads::wait_for_tenant_threads_to_stop(tenantid);
+
+        let repo = get_repository_for_tenant(tenantid)?;
+        debug!("shutdown tenant {}", tenantid);
+        repo.shutdown()?;
+    }
+    Ok(())
+}
+
 pub fn create_repository_for_tenant(
     conf: &'static PageServerConf,
     tenantid: ZTenantId,
 ) -> Result<()> {
-    let mut m = access_repository();
-
-    // First check that the tenant doesn't exist already
-    if m.get(&tenantid).is_some() {
-        bail!("tenant {} already exists", tenantid);
+    {
+        let mut m = access_tenants();
+        // First check that the tenant doesn't exist already
+        if m.get(&tenantid).is_some() {
+            bail!("tenant {} already exists", tenantid);
+        }
+        let tenant = Tenant {
+            state: TenantState::CloudOnly,
+            repo: None,
+        };
+        m.insert(tenantid, tenant);
     }
+
     let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenantid));
     let repo = branches::create_repo(conf, tenantid, wal_redo_manager)?;
 
-    m.insert(tenantid, repo);
+    let mut m = access_tenants();
+    let tenant = m.get_mut(&tenantid).unwrap();
+    tenant.repo = Some(repo);
+    tenant.state = TenantState::Idle;
 
     Ok(())
 }
 
+// If tenant is not found in the repository, return CloudOnly state
+pub fn get_tenant_state(tenantid: ZTenantId) -> TenantState {
+    let m = access_tenants();
+    match m.get(&tenantid) {
+        Some(tenant) => tenant.state,
+        None => TenantState::CloudOnly,
+    }
+}
+
+pub fn set_tenant_state(tenantid: ZTenantId, newstate: TenantState) -> Result<TenantState> {
+    let mut m = access_tenants();
+    let tenant = m.get_mut(&tenantid);
+
+    match tenant {
+        Some(tenant) => {
+            if newstate == TenantState::Idle && tenant.state != TenantState::Active {
+                // Only Active tenant can become Idle
+                return Ok(tenant.state);
+            }
+            info!("set_tenant_state: {} -> {}", tenant.state, newstate);
+            tenant.state = newstate;
+            Ok(tenant.state)
+        }
+        None => bail!("Tenant not found for tenant {}", tenantid),
+    }
+}
+
 pub fn get_repository_for_tenant(tenantid: ZTenantId) -> Result<Arc<dyn Repository>> {
-    access_repository()
+    let m = access_tenants();
+    let tenant = m
         .get(&tenantid)
-        .map(Arc::clone)
-        .ok_or_else(|| anyhow!("repository not found for tenant name {}", tenantid))
+        .ok_or_else(|| anyhow!("Tenant not found for tenant {}", tenantid))?;
+
+    match &tenant.repo {
+        Some(repo) => Ok(Arc::clone(repo)),
+        None => anyhow::bail!("Repository for tenant {} is not yet valid", tenantid),
+    }
 }
 
 pub fn get_timeline_for_tenant(
@@ -115,3 +275,33 @@ pub fn get_timeline_for_tenant(
         .get_timeline(timelineid)
         .with_context(|| format!("cannot fetch timeline {}", timelineid))
 }
+
+fn list_tenantids() -> Result<Vec<ZTenantId>> {
+    let m = access_tenants();
+    m.iter()
+        .map(|v| {
+            let (tenantid, _) = v;
+            Ok(*tenantid)
+        })
+        .collect()
+}
+
+#[derive(Serialize, Deserialize, Clone)]
+pub struct TenantInfo {
+    #[serde(with = "hex")]
+    pub id: ZTenantId,
+    pub state: TenantState,
+}
+
+pub fn list_tenants() -> Result<Vec<TenantInfo>> {
+    let m = access_tenants();
+    m.iter()
+        .map(|v| {
+            let (id, tenant) = v;
+            Ok(TenantInfo {
+                id: *id,
+                state: tenant.state,
+            })
+        })
+        .collect()
+}

pageserver/src/tenant_threads.rs

@@ -0,0 +1,149 @@
+//! This module contains functions to serve per-tenant background processes,
+//! such as checkpointer and GC
+use crate::tenant_mgr;
+use crate::tenant_mgr::TenantState;
+use crate::CheckpointConfig;
+use crate::PageServerConf;
+use anyhow::Result;
+use lazy_static::lazy_static;
+use std::collections::HashMap;
+use std::sync::Mutex;
+use std::thread::JoinHandle;
+use std::time::Duration;
+use tracing::*;
+use zenith_metrics::{register_int_gauge_vec, IntGaugeVec};
+use zenith_utils::zid::ZTenantId;
+
+struct TenantHandleEntry {
+    checkpointer_handle: Option<JoinHandle<()>>,
+    gc_handle: Option<JoinHandle<()>>,
+}
+
+// Preserve handles to wait for thread completion
+// at shutdown
+lazy_static! {
+    static ref TENANT_HANDLES: Mutex<HashMap<ZTenantId, TenantHandleEntry>> =
+        Mutex::new(HashMap::new());
+}
+
+lazy_static! {
+    static ref TENANT_THREADS_COUNT: IntGaugeVec = register_int_gauge_vec!(
+        "tenant_threads_count",
+        "Number of live tenant threads",
+        &["tenant_thread_type"]
+    )
+    .expect("failed to define a metric");
+}
+
+// Launch checkpointer and GC for the tenant.
+// It's possible that the threads are running already,
+// if so, just don't spawn new ones.
+pub fn start_tenant_threads(conf: &'static PageServerConf, tenantid: ZTenantId) {
+    let mut handles = TENANT_HANDLES.lock().unwrap();
+    let h = handles
+        .entry(tenantid)
+        .or_insert_with(|| TenantHandleEntry {
+            checkpointer_handle: None,
+            gc_handle: None,
+        });
+
+    if h.checkpointer_handle.is_none() {
+        h.checkpointer_handle = std::thread::Builder::new()
+            .name("Checkpointer thread".into())
+            .spawn(move || {
+                checkpoint_loop(tenantid, conf).expect("Checkpointer thread died");
+            })
+            .ok();
+    }
+
+    if h.gc_handle.is_none() {
+        h.gc_handle = std::thread::Builder::new()
+            .name("GC thread".into())
+            .spawn(move || {
+                gc_loop(tenantid, conf).expect("GC thread died");
+            })
+            .ok();
+    }
+}
+
+pub fn wait_for_tenant_threads_to_stop(tenantid: ZTenantId) {
+    let mut handles = TENANT_HANDLES.lock().unwrap();
+    if let Some(h) = handles.get_mut(&tenantid) {
+        h.checkpointer_handle.take().map(JoinHandle::join);
+        trace!("checkpointer for tenant {} has stopped", tenantid);
+        h.gc_handle.take().map(JoinHandle::join);
+        trace!("gc for tenant {} has stopped", tenantid);
+    }
+    handles.remove(&tenantid);
+}
+
+///
+/// Checkpointer thread's main loop
+///
+fn checkpoint_loop(tenantid: ZTenantId, conf: &'static PageServerConf) -> Result<()> {
+    let gauge = TENANT_THREADS_COUNT.with_label_values(&["checkpointer"]);
+    gauge.inc();
+    scopeguard::defer! {
+        gauge.dec();
+    }
+
+    loop {
+        if tenant_mgr::get_tenant_state(tenantid) != TenantState::Active {
+            break;
+        }
+
+        std::thread::sleep(conf.checkpoint_period);
+        trace!("checkpointer thread for tenant {} waking up", tenantid);
+
+        // checkpoint timelines that have accumulated more than CHECKPOINT_DISTANCE
+        // bytes of WAL since last checkpoint.
+        let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
+        repo.checkpoint_iteration(CheckpointConfig::Distance(conf.checkpoint_distance))?;
+    }
+
+    trace!(
+        "checkpointer thread stopped for tenant {} state is {}",
+        tenantid,
+        tenant_mgr::get_tenant_state(tenantid)
+    );
+    Ok(())
+}
+
+///
+/// GC thread's main loop
+///
+fn gc_loop(tenantid: ZTenantId, conf: &'static PageServerConf) -> Result<()> {
+    let gauge = TENANT_THREADS_COUNT.with_label_values(&["gc"]);
+    gauge.inc();
+    scopeguard::defer! {
+        gauge.dec();
+    }
+
+    loop {
+        if tenant_mgr::get_tenant_state(tenantid) != TenantState::Active {
+            break;
+        }
+
+        trace!("gc thread for tenant {} waking up", tenantid);
+
+        // Garbage collect old files that are not needed for PITR anymore
+        if conf.gc_horizon > 0 {
+            let repo = tenant_mgr::get_repository_for_tenant(tenantid)?;
+            repo.gc_iteration(None, conf.gc_horizon, false).unwrap();
+        }
+
+        // TODO Write it in more adequate way using
+        // condvar.wait_timeout() or something
+        let mut sleep_time = conf.gc_period.as_secs();
+        while sleep_time > 0 && tenant_mgr::get_tenant_state(tenantid) == TenantState::Active {
+            sleep_time -= 1;
+            std::thread::sleep(Duration::from_secs(1));
+        }
+    }
+    trace!(
+        "GC thread stopped for tenant {} state is {}",
+        tenantid,
+        tenant_mgr::get_tenant_state(tenantid)
+    );
+    Ok(())
+}

pageserver/src/virtual_file.rs

@@ -0,0 +1,619 @@
+//!
+//! VirtualFile is like a normal File, but it's not bound directly to
+//! a file descriptor. Instead, the file is opened when it's read from,
+//! and if too many files are open globally in the system, least-recently
+//! used ones are closed.
+//!
+//! To track which files have been recently used, we use the clock algorithm
+//! with a 'recently_used' flag on each slot.
+//!
+//! This is similar to PostgreSQL's virtual file descriptor facility in
+//! src/backend/storage/file/fd.c
+//!
+use std::fs::{File, OpenOptions};
+use std::io::{Error, ErrorKind, Read, Seek, SeekFrom, Write};
+use std::os::unix::fs::FileExt;
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
+use std::sync::{RwLock, RwLockWriteGuard};
+
+use once_cell::sync::OnceCell;
+
+///
+/// A virtual file descriptor. You can use this just like std::fs::File, but internally
+/// the underlying file is closed if the system is low on file descriptors,
+/// and re-opened when it's accessed again.
+///
+/// Like with std::fs::File, multiple threads can read/write the file concurrently,
+/// holding just a shared reference the same VirtualFile, using the read_at() / write_at()
+/// functions from the FileExt trait. But the functions from the Read/Write/Seek traits
+/// require a mutable reference, because they modify the "current position".
+///
+/// Each VirtualFile has a physical file descriptor in the global OPEN_FILES array, at the
+/// slot that 'handle points to, if the underlying file is currently open. If it's not
+/// currently open, the 'handle' can still point to the slot where it was last kept. The
+/// 'tag' field is used to detect whether the handle still is valid or not.
+///
+pub struct VirtualFile {
+    /// Lazy handle to the global file descriptor cache. The slot that this points to
+    /// might contain our File, or it may be empty, or it may contain a File that
+    /// belongs to a different VirtualFile.
+    handle: RwLock<SlotHandle>,
+
+    /// Current file position
+    pos: u64,
+
+    /// File path and options to use to open it.
+    ///
+    /// Note: this only contains the options needed to re-open it. For example,
+    /// if a new file is created, we only pass the create flag when it's initially
+    /// opened, in the VirtualFile::create() function, and strip the flag before
+    /// storing it here.
+    pub path: PathBuf,
+    open_options: OpenOptions,
+}
+
+#[derive(PartialEq, Clone, Copy)]
+struct SlotHandle {
+    /// Index into OPEN_FILES.slots
+    index: usize,
+
+    /// Value of 'tag' in the slot. If slot's tag doesn't match, then the slot has
+    /// been recycled and no longer contains the FD for this virtual file.
+    tag: u64,
+}
+
+/// OPEN_FILES is the global array that holds the physical file descriptors that
+/// are currently open. Each slot in the array is protected by a separate lock,
+/// so that different files can be accessed independently. The lock must be held
+/// in write mode to replace the slot with a different file, but a read mode
+/// is enough to operate on the file, whether you're reading or writing to it.
+///
+/// OPEN_FILES starts in uninitialized state, and it's initialized by
+/// the virtual_file::init() function. It must be called exactly once at page
+/// server startup.
+static OPEN_FILES: OnceCell<OpenFiles> = OnceCell::new();
+
+struct OpenFiles {
+    slots: &'static [Slot],
+
+    /// clock arm for the clock algorithm
+    next: AtomicUsize,
+}
+
+struct Slot {
+    inner: RwLock<SlotInner>,
+
+    /// has this file been used since last clock sweep?
+    recently_used: AtomicBool,
+}
+
+struct SlotInner {
+    /// Counter that's incremented every time a different file is stored here.
+    /// To avoid the ABA problem.
+    tag: u64,
+
+    /// the underlying file
+    file: Option<File>,
+}
+
+impl OpenFiles {
+    /// Find a slot to use, evicting an existing file descriptor if needed.
+    ///
+    /// On return, we hold a lock on the slot, and its 'tag' has been updated
+    /// recently_used has been set. It's all ready for reuse.
+    fn find_victim_slot(&self) -> (SlotHandle, RwLockWriteGuard<SlotInner>) {
+        //
+        // Run the clock algorithm to find a slot to replace.
+        //
+        let num_slots = self.slots.len();
+        let mut retries = 0;
+        let mut slot;
+        let mut slot_guard;
+        let index;
+        loop {
+            let next = self.next.fetch_add(1, Ordering::AcqRel) % num_slots;
+            slot = &self.slots[next];
+
+            // If the recently_used flag on this slot is set, continue the clock
+            // sweep. Otherwise try to use this slot. If we cannot acquire the
+            // lock, also continue the clock sweep.
+            //
+            // We only continue in this manner for a while, though. If we loop
+            // through the array twice without finding a victim, just pick the
+            // next slot and wait until we can reuse it. This way, we avoid
+            // spinning in the extreme case that all the slots are busy with an
+            // I/O operation.
+            if retries < num_slots * 2 {
+                if !slot.recently_used.swap(false, Ordering::Release) {
+                    if let Ok(guard) = slot.inner.try_write() {
+                        slot_guard = guard;
+                        index = next;
+                        break;
+                    }
+                }
+                retries += 1;
+            } else {
+                slot_guard = slot.inner.write().unwrap();
+                index = next;
+                break;
+            }
+        }
+
+        //
+        // We now have the victim slot locked. If it was in use previously, close the
+        // old file.
+        //
+        if let Some(old_file) = slot_guard.file.take() {
+            drop(old_file);
+        }
+
+        // Prepare the slot for reuse and return it
+        slot_guard.tag += 1;
+        slot.recently_used.store(true, Ordering::Relaxed);
+        (
+            SlotHandle {
+                index,
+                tag: slot_guard.tag,
+            },
+            slot_guard,
+        )
+    }
+}
+
+impl VirtualFile {
+    /// Open a file in read-only mode. Like File::open.
+    pub fn open(path: &Path) -> Result<VirtualFile, std::io::Error> {
+        Self::open_with_options(path, OpenOptions::new().read(true))
+    }
+
+    /// Create a new file for writing. If the file exists, it will be truncated.
+    /// Like File::create.
+    pub fn create(path: &Path) -> Result<VirtualFile, std::io::Error> {
+        Self::open_with_options(
+            path,
+            OpenOptions::new().write(true).create(true).truncate(true),
+        )
+    }
+
+    /// Open a file with given options.
+    ///
+    /// Note: If any custom flags were set in 'open_options' through OpenOptionsExt,
+    /// they will be applied also when the file is subsequently re-opened, not only
+    /// on the first time. Make sure that's sane!
+    pub fn open_with_options(
+        path: &Path,
+        open_options: &OpenOptions,
+    ) -> Result<VirtualFile, std::io::Error> {
+        let (handle, mut slot_guard) = get_open_files().find_victim_slot();
+
+        let file = open_options.open(path)?;
+
+        // Strip all options other than read and write.
+        //
+        // It would perhaps be nicer to check just for the read and write flags
+        // explicitly, but OpenOptions doesn't contain any functions to read flags,
+        // only to set them.
+        let mut reopen_options = open_options.clone();
+        reopen_options.create(false);
+        reopen_options.create_new(false);
+        reopen_options.truncate(false);
+
+        let vfile = VirtualFile {
+            handle: RwLock::new(handle),
+            pos: 0,
+            path: path.to_path_buf(),
+            open_options: reopen_options,
+        };
+
+        slot_guard.file.replace(file);
+
+        Ok(vfile)
+    }
+
+    /// Call File::sync_all() on the underlying File.
+    pub fn sync_all(&self) -> Result<(), Error> {
+        self.with_file(|file| file.sync_all())?
+    }
+
+    /// Helper function that looks up the underlying File for this VirtualFile,
+    /// opening it and evicting some other File if necessary. It calls 'func'
+    /// with the physical File.
+    fn with_file<F, R>(&self, mut func: F) -> Result<R, Error>
+    where
+        F: FnMut(&File) -> R,
+    {
+        let open_files = get_open_files();
+
+        let mut handle_guard = {
+            // Read the cached slot handle, and see if the slot that it points to still
+            // contains our File.
+            //
+            // We only need to hold the handle lock while we read the current handle. If
+            // another thread closes the file and recycles the slot for a different file,
+            // we will notice that the handle we read is no longer valid and retry.
+            let mut handle = *self.handle.read().unwrap();
+            loop {
+                // Check if the slot contains our File
+                {
+                    let slot = &open_files.slots[handle.index];
+                    let slot_guard = slot.inner.read().unwrap();
+                    if slot_guard.tag == handle.tag {
+                        if let Some(file) = &slot_guard.file {
+                            // Found a cached file descriptor.
+                            slot.recently_used.store(true, Ordering::Relaxed);
+                            return Ok(func(file));
+                        }
+                    }
+                }
+
+                // The slot didn't contain our File. We will have to open it ourselves,
+                // but before that, grab a write lock on handle in the VirtualFile, so
+                // that no other thread will try to concurrently open the same file.
+                let handle_guard = self.handle.write().unwrap();
+
+                // If another thread changed the handle while we were not holding the lock,
+                // then the handle might now be valid again. Loop back to retry.
+                if *handle_guard != handle {
+                    handle = *handle_guard;
+                    continue;
+                }
+                break handle_guard;
+            }
+        };
+
+        // We need to open the file ourselves. The handle in the VirtualFile is
+        // now locked in write-mode. Find a free slot to put it in.
+        let (handle, mut slot_guard) = open_files.find_victim_slot();
+
+        // Open the physical file
+        let file = self.open_options.open(&self.path)?;
+
+        // Perform the requested operation on it
+        //
+        // TODO: We could downgrade the locks to read mode before calling
+        // 'func', to allow a little bit more concurrency, but the standard
+        // library RwLock doesn't allow downgrading without releasing the lock,
+        // and that doesn't seem worth the trouble. (parking_lot RwLock would
+        // allow it)
+        let result = func(&file);
+
+        // Store the File in the slot and update the handle in the VirtualFile
+        // to point to it.
+        slot_guard.file.replace(file);
+
+        *handle_guard = handle;
+
+        Ok(result)
+    }
+}
+
+impl Drop for VirtualFile {
+    /// If a VirtualFile is dropped, close the underlying file if it was open.
+    fn drop(&mut self) {
+        let handle = self.handle.get_mut().unwrap();
+
+        // We could check with a read-lock first, to avoid waiting on an
+        // unrelated I/O.
+        let slot = &get_open_files().slots[handle.index];
+        let mut slot_guard = slot.inner.write().unwrap();
+        if slot_guard.tag == handle.tag {
+            slot.recently_used.store(false, Ordering::Relaxed);
+            slot_guard.file.take();
+        }
+    }
+}
+
+impl Read for VirtualFile {
+    fn read(&mut self, buf: &mut [u8]) -> Result<usize, Error> {
+        let pos = self.pos;
+        let n = self.read_at(buf, pos)?;
+        self.pos += n as u64;
+        Ok(n)
+    }
+}
+
+impl Write for VirtualFile {
+    fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
+        let pos = self.pos;
+        let n = self.write_at(buf, pos)?;
+        self.pos += n as u64;
+        Ok(n)
+    }
+
+    fn flush(&mut self) -> Result<(), std::io::Error> {
+        // flush is no-op for File (at least on unix), so we don't need to do
+        // anything here either.
+        Ok(())
+    }
+}
+
+impl Seek for VirtualFile {
+    fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
+        match pos {
+            SeekFrom::Start(offset) => {
+                self.pos = offset;
+            }
+            SeekFrom::End(offset) => {
+                self.pos = self.with_file(|mut file| file.seek(SeekFrom::End(offset)))??
+            }
+            SeekFrom::Current(offset) => {
+                let pos = self.pos as i128 + offset as i128;
+                if pos < 0 {
+                    return Err(Error::new(
+                        ErrorKind::InvalidInput,
+                        "offset would be negative",
+                    ));
+                }
+                if pos > u64::MAX as i128 {
+                    return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
+                }
+                self.pos = pos as u64;
+            }
+        }
+        Ok(self.pos)
+    }
+}
+
+impl FileExt for VirtualFile {
+    fn read_at(&self, buf: &mut [u8], offset: u64) -> Result<usize, Error> {
+        self.with_file(|file| file.read_at(buf, offset))?
+    }
+
+    fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
+        self.with_file(|file| file.write_at(buf, offset))?
+    }
+}
+
+impl OpenFiles {
+    fn new(num_slots: usize) -> OpenFiles {
+        let mut slots = Box::new(Vec::with_capacity(num_slots));
+        for _ in 0..num_slots {
+            let slot = Slot {
+                recently_used: AtomicBool::new(false),
+                inner: RwLock::new(SlotInner { tag: 0, file: None }),
+            };
+            slots.push(slot);
+        }
+
+        OpenFiles {
+            next: AtomicUsize::new(0),
+            slots: Box::leak(slots),
+        }
+    }
+}
+
+///
+/// Initialize the virtual file module. This must be called once at page
+/// server startup.
+///
+pub fn init(num_slots: usize) {
+    if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
+        panic!("virtual_file::init called twice");
+    }
+}
+
+const TEST_MAX_FILE_DESCRIPTORS: usize = 10;
+
+// Get a handle to the global slots array.
+fn get_open_files() -> &'static OpenFiles {
+    //
+    // In unit tests, page server startup doesn't happen and no one calls
+    // virtual_file::init(). Initialize it here, with a small array.
+    //
+    // This applies to the virtual file tests below, but all other unit
+    // tests too, so the virtual file facility is always usable in
+    // unit tests.
+    //
+    if cfg!(test) {
+        OPEN_FILES.get_or_init(|| OpenFiles::new(TEST_MAX_FILE_DESCRIPTORS))
+    } else {
+        OPEN_FILES.get().expect("virtual_file::init not called yet")
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use rand::seq::SliceRandom;
+    use rand::thread_rng;
+    use rand::Rng;
+    use std::sync::Arc;
+    use std::thread;
+
+    // Helper function to slurp contents of a file, starting at the current position,
+    // into a string
+    fn read_string<FD>(vfile: &mut FD) -> Result<String, Error>
+    where
+        FD: Read,
+    {
+        let mut buf = String::new();
+        vfile.read_to_string(&mut buf)?;
+        Ok(buf)
+    }
+
+    // Helper function to slurp a portion of a file into a string
+    fn read_string_at<FD>(vfile: &mut FD, pos: u64, len: usize) -> Result<String, Error>
+    where
+        FD: FileExt,
+    {
+        let mut buf = Vec::new();
+        buf.resize(len, 0);
+        vfile.read_exact_at(&mut buf, pos)?;
+        Ok(String::from_utf8(buf).unwrap())
+    }
+
+    #[test]
+    fn test_virtual_files() -> Result<(), Error> {
+        // The real work is done in the test_files() helper function. This
+        // allows us to run the same set of tests against a native File, and
+        // VirtualFile. We trust the native Files and wouldn't need to test them,
+        // but this allows us to verify that the operations return the same
+        // results with VirtualFiles as with native Files. (Except that with
+        // native files, you will run out of file descriptors if the ulimit
+        // is low enough.)
+        test_files("virtual_files", |path, open_options| {
+            VirtualFile::open_with_options(path, open_options)
+        })
+    }
+
+    #[test]
+    fn test_physical_files() -> Result<(), Error> {
+        test_files("physical_files", |path, open_options| {
+            open_options.open(path)
+        })
+    }
+
+    fn test_files<OF, FD>(testname: &str, openfunc: OF) -> Result<(), Error>
+    where
+        FD: Read + Write + Seek + FileExt,
+        OF: Fn(&Path, &OpenOptions) -> Result<FD, std::io::Error>,
+    {
+        let testdir = crate::PageServerConf::test_repo_dir(testname);
+        std::fs::create_dir_all(&testdir)?;
+
+        let path_a = testdir.join("file_a");
+        let mut file_a = openfunc(
+            &path_a,
+            OpenOptions::new().write(true).create(true).truncate(true),
+        )?;
+        file_a.write_all(b"foobar")?;
+
+        // cannot read from a file opened in write-only mode
+        assert!(read_string(&mut file_a).is_err());
+
+        // Close the file and re-open for reading
+        let mut file_a = openfunc(&path_a, OpenOptions::new().read(true))?;
+
+        // cannot write to a file opened in read-only mode
+        assert!(file_a.write(b"bar").is_err());
+
+        // Try simple read
+        assert_eq!("foobar", read_string(&mut file_a)?);
+
+        // It's positioned at the EOF now.
+        assert_eq!("", read_string(&mut file_a)?);
+
+        // Test seeks.
+        assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
+        assert_eq!("oobar", read_string(&mut file_a)?);
+
+        assert_eq!(file_a.seek(SeekFrom::End(-2))?, 4);
+        assert_eq!("ar", read_string(&mut file_a)?);
+
+        assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
+        assert_eq!(file_a.seek(SeekFrom::Current(2))?, 3);
+        assert_eq!("bar", read_string(&mut file_a)?);
+
+        assert_eq!(file_a.seek(SeekFrom::Current(-5))?, 1);
+        assert_eq!("oobar", read_string(&mut file_a)?);
+
+        // Test erroneous seeks to before byte 0
+        assert!(file_a.seek(SeekFrom::End(-7)).is_err());
+        assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
+        assert!(file_a.seek(SeekFrom::Current(-2)).is_err());
+
+        // the erroneous seek should have left the position unchanged
+        assert_eq!("oobar", read_string(&mut file_a)?);
+
+        // Create another test file, and try FileExt functions on it.
+        let path_b = testdir.join("file_b");
+        let mut file_b = openfunc(
+            &path_b,
+            OpenOptions::new()
+                .read(true)
+                .write(true)
+                .create(true)
+                .truncate(true),
+        )?;
+        file_b.write_all_at(b"BAR", 3)?;
+        file_b.write_all_at(b"FOO", 0)?;
+
+        assert_eq!(read_string_at(&mut file_b, 2, 3)?, "OBA");
+
+        // Open a lot of files, enough to cause some evictions. (Or to be precise,
+        // open the same file many times. The effect is the same.)
+        //
+        // leave file_a positioned at offset 1 before we start
+        assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
+
+        let mut vfiles = Vec::new();
+        for _ in 0..100 {
+            let mut vfile = openfunc(&path_b, OpenOptions::new().read(true))?;
+            assert_eq!("FOOBAR", read_string(&mut vfile)?);
+            vfiles.push(vfile);
+        }
+
+        // make sure we opened enough files to definitely cause evictions.
+        assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
+
+        // The underlying file descriptor for 'file_a' should be closed now. Try to read
+        // from it again. We left the file positioned at offset 1 above.
+        assert_eq!("oobar", read_string(&mut file_a)?);
+
+        // Check that all the other FDs still work too. Use them in random order for
+        // good measure.
+        vfiles.as_mut_slice().shuffle(&mut thread_rng());
+        for vfile in vfiles.iter_mut() {
+            assert_eq!("OOBAR", read_string_at(vfile, 1, 5)?);
+        }
+
+        Ok(())
+    }
+
+    /// Test using VirtualFiles from many threads concurrently. This tests both using
+    /// a lot of VirtualFiles concurrently, causing evictions, and also using the same
+    /// VirtualFile from multiple threads concurrently.
+    #[test]
+    fn test_vfile_concurrency() -> Result<(), Error> {
+        const SIZE: usize = 8 * 1024;
+        const VIRTUAL_FILES: usize = 100;
+        const THREADS: usize = 100;
+        const SAMPLE: [u8; SIZE] = [0xADu8; SIZE];
+
+        let testdir = crate::PageServerConf::test_repo_dir("vfile_concurrency");
+        std::fs::create_dir_all(&testdir)?;
+
+        // Create a test file.
+        let test_file_path = testdir.join("concurrency_test_file");
+        {
+            let file = File::create(&test_file_path)?;
+            file.write_all_at(&SAMPLE, 0)?;
+        }
+
+        // Open the file many times.
+        let mut files = Vec::new();
+        for _ in 0..VIRTUAL_FILES {
+            let f = VirtualFile::open_with_options(&test_file_path, OpenOptions::new().read(true))?;
+            files.push(f);
+        }
+        let files = Arc::new(files);
+
+        // Launch many threads, and use the virtual files concurrently in random order.
+        let mut threads = Vec::new();
+        for threadno in 0..THREADS {
+            let builder =
+                thread::Builder::new().name(format!("test_vfile_concurrency thread {}", threadno));
+
+            let files = files.clone();
+            let thread = builder
+                .spawn(move || {
+                    let mut buf = [0u8; SIZE];
+                    let mut rng = rand::thread_rng();
+                    for _ in 1..1000 {
+                        let f = &files[rng.gen_range(0..files.len())];
+                        f.read_exact_at(&mut buf, 0).unwrap();
+                        assert!(buf == SAMPLE);
+                    }
+                })
+                .unwrap();
+            threads.push(thread);
+        }
+
+        for thread in threads {
+            thread.join().unwrap();
+        }
+
+        Ok(())
+    }
+}

pageserver/src/waldecoder.rs

@@ -72,6 +72,10 @@ impl WalStreamDecoder {
     ///     Err(WalDecodeError): an error occured while decoding, meaning the input was invalid.
     ///
     pub fn poll_decode(&mut self) -> Result<Option<(Lsn, Bytes)>, WalDecodeError> {
+        let recordbuf;
+
+        // Run state machine that validates page headers, and reassembles records
+        // that cross page boundaries.
         loop {
             // parse and verify page boundaries as we go
             if self.lsn.segment_offset(pg_constants::WAL_SEGMENT_SIZE) == 0 {
@@ -120,29 +124,41 @@ impl WalStreamDecoder {
                 self.lsn += self.padlen as u64;
                 self.padlen = 0;
             } else if self.contlen == 0 {
-                // need to have at least the xl_tot_len field
+                assert!(self.recordbuf.is_empty());
 
+                // need to have at least the xl_tot_len field
                 if self.inputbuf.remaining() < 4 {
                     return Ok(None);
                 }
 
-                // read xl_tot_len FIXME: assumes little-endian
+                // peek xl_tot_len at the beginning of the record.
+                // FIXME: assumes little-endian
                 self.startlsn = self.lsn;
-                let xl_tot_len = self.inputbuf.get_u32_le();
+                let xl_tot_len = (&self.inputbuf[0..4]).get_u32_le();
                 if (xl_tot_len as usize) < XLOG_SIZE_OF_XLOG_RECORD {
                     return Err(WalDecodeError {
                         msg: format!("invalid xl_tot_len {}", xl_tot_len),
                         lsn: self.lsn,
                     });
                 }
-                self.lsn += 4;
 
-                self.recordbuf.clear();
-                self.recordbuf.reserve(xl_tot_len as usize);
-                self.recordbuf.put_u32_le(xl_tot_len);
-
-                self.contlen = xl_tot_len - 4;
-                continue;
+                // Fast path for the common case that the whole record fits on the page.
+                let pageleft = self.lsn.remaining_in_block() as u32;
+                if self.inputbuf.remaining() >= xl_tot_len as usize && xl_tot_len <= pageleft {
+                    // Take the record from the 'inputbuf', and validate it.
+                    recordbuf = self.inputbuf.copy_to_bytes(xl_tot_len as usize);
+                    self.lsn += xl_tot_len as u64;
+                    break;
+                } else {
+                    // Need to assemble the record from pieces. Remember the size of the
+                    // record, and loop back. On next iteration, we will reach the 'else'
+                    // branch below, and copy the part of the record that was on this page
+                    // to 'recordbuf'.  Subsequent iterations will skip page headers, and
+                    // append the continuations from the next pages to 'recordbuf'.
+                    self.recordbuf.reserve(xl_tot_len as usize);
+                    self.contlen = xl_tot_len;
+                    continue;
+                }
             } else {
                 // we're continuing a record, possibly from previous page.
                 let pageleft = self.lsn.remaining_in_block() as u32;
@@ -159,47 +175,42 @@ impl WalStreamDecoder {
                 self.contlen -= n as u32;
 
                 if self.contlen == 0 {
-                    let recordbuf = std::mem::replace(&mut self.recordbuf, BytesMut::new());
-
-                    let recordbuf = recordbuf.freeze();
-                    let mut buf = recordbuf.clone();
-
-                    let xlogrec = XLogRecord::from_bytes(&mut buf);
-
-                    // XLOG_SWITCH records are special. If we see one, we need to skip
-                    // to the next WAL segment.
-                    if xlogrec.is_xlog_switch_record() {
-                        trace!("saw xlog switch record at {}", self.lsn);
-                        self.padlen =
-                            self.lsn.calc_padding(pg_constants::WAL_SEGMENT_SIZE as u64) as u32;
-                    } else {
-                        // Pad to an 8-byte boundary
-                        self.padlen = self.lsn.calc_padding(8u32) as u32;
-                    }
-
-                    let mut crc = crc32c_append(0, &recordbuf[XLOG_RECORD_CRC_OFFS + 4..]);
-                    crc = crc32c_append(crc, &recordbuf[0..XLOG_RECORD_CRC_OFFS]);
-                    if crc != xlogrec.xl_crc {
-                        return Err(WalDecodeError {
-                            msg: "WAL record crc mismatch".into(),
-                            lsn: self.lsn,
-                        });
-                    }
-
-                    // Always align resulting LSN on 0x8 boundary -- that is important for getPage()
-                    // and WalReceiver integration. Since this code is used both for WalReceiver and
-                    // initial WAL import let's force alignment right here.
-                    let result = (self.lsn.align(), recordbuf);
-                    return Ok(Some(result));
+                    // The record is now complete.
+                    recordbuf = std::mem::replace(&mut self.recordbuf, BytesMut::new()).freeze();
+                    break;
                 }
                 continue;
             }
         }
-        // check record boundaries
 
-        // deal with continuation records
+        // We now have a record in the 'recordbuf' local variable.
+        let xlogrec = XLogRecord::from_slice(&recordbuf[0..XLOG_SIZE_OF_XLOG_RECORD]);
 
-        // deal with xlog_switch records
+        let mut crc = 0;
+        crc = crc32c_append(crc, &recordbuf[XLOG_RECORD_CRC_OFFS + 4..]);
+        crc = crc32c_append(crc, &recordbuf[0..XLOG_RECORD_CRC_OFFS]);
+        if crc != xlogrec.xl_crc {
+            return Err(WalDecodeError {
+                msg: "WAL record crc mismatch".into(),
+                lsn: self.lsn,
+            });
+        }
+
+        // XLOG_SWITCH records are special. If we see one, we need to skip
+        // to the next WAL segment.
+        if xlogrec.is_xlog_switch_record() {
+            trace!("saw xlog switch record at {}", self.lsn);
+            self.padlen = self.lsn.calc_padding(pg_constants::WAL_SEGMENT_SIZE as u64) as u32;
+        } else {
+            // Pad to an 8-byte boundary
+            self.padlen = self.lsn.calc_padding(8u32) as u32;
+        }
+
+        // Always align resulting LSN on 0x8 boundary -- that is important for getPage()
+        // and WalReceiver integration. Since this code is used both for WalReceiver and
+        // initial WAL import let's force alignment right here.
+        let result = (self.lsn.align(), recordbuf);
+        Ok(Some(result))
     }
 }
 
@@ -218,17 +229,18 @@ pub struct DecodedBkpBlock {
     pub blkno: u32,
 
     /* copy of the fork_flags field from the XLogRecordBlockHeader */
-    flags: u8,
+    pub flags: u8,
 
     /* Information on full-page image, if any */
-    has_image: bool,       /* has image, even for consistency checking */
+    pub has_image: bool,   /* has image, even for consistency checking */
     pub apply_image: bool, /* has image that should be restored */
     pub will_init: bool,   /* record doesn't need previous page version to apply */
     //char	   *bkp_image;
-    hole_offset: u16,
-    hole_length: u16,
-    bimg_len: u16,
-    bimg_info: u8,
+    pub hole_offset: u16,
+    pub hole_length: u16,
+    pub bimg_offset: u32,
+    pub bimg_len: u16,
+    pub bimg_info: u8,
 
     /* Buffer holding the rmgr-specific data associated with this block */
     has_data: bool,
@@ -848,8 +860,19 @@ pub fn decode_wal_record(record: Bytes) -> DecodedWALRecord {
     }
 
     // 3. Decode blocks.
+    let mut ptr = record.len() - buf.remaining();
+    for blk in blocks.iter_mut() {
+        if blk.has_image {
+            blk.bimg_offset = ptr as u32;
+            ptr += blk.bimg_len as usize;
+        }
+        if blk.has_data {
+            ptr += blk.data_len as usize;
+        }
+    }
     // We don't need them, so just skip blocks_total_len bytes
     buf.advance(blocks_total_len as usize);
+    assert_eq!(ptr, record.len() - buf.remaining());
 
     let main_data_offset = (xlogrec.xl_tot_len - main_data_len) as usize;
 

pageserver/src/walreceiver.rs

@@ -8,11 +8,12 @@
 use crate::relish::*;
 use crate::restore_local_repo;
 use crate::tenant_mgr;
+use crate::tenant_mgr::TenantState;
+use crate::tenant_threads;
 use crate::waldecoder::*;
 use crate::PageServerConf;
 use anyhow::{bail, Error, Result};
 use lazy_static::lazy_static;
-use log::*;
 use postgres::fallible_iterator::FallibleIterator;
 use postgres::replication::ReplicationIter;
 use postgres::{Client, NoTls, SimpleQueryMessage, SimpleQueryRow};
@@ -25,8 +26,10 @@ use std::str::FromStr;
 use std::sync::Mutex;
 use std::thread;
 use std::thread::sleep;
+use std::thread::JoinHandle;
 use std::thread_local;
 use std::time::{Duration, SystemTime};
+use tracing::*;
 use zenith_utils::lsn::Lsn;
 use zenith_utils::zid::ZTenantId;
 use zenith_utils::zid::ZTimelineId;
@@ -36,6 +39,8 @@ use zenith_utils::zid::ZTimelineId;
 //
 struct WalReceiverEntry {
     wal_producer_connstr: String,
+    wal_receiver_handle: Option<JoinHandle<()>>,
+    tenantid: ZTenantId,
 }
 
 lazy_static! {
@@ -50,6 +55,36 @@ thread_local! {
     pub(crate) static IS_WAL_RECEIVER: Cell<bool> = Cell::new(false);
 }
 
+// Wait for walreceiver to stop
+// Now it stops when pageserver shutdown is requested.
+// In future we can make this more granular and send shutdown signals
+// per tenant/timeline to cancel inactive walreceivers.
+// TODO deal with blocking pg connections
+pub fn stop_wal_receiver(timelineid: ZTimelineId) {
+    let mut receivers = WAL_RECEIVERS.lock().unwrap();
+    if let Some(r) = receivers.get_mut(&timelineid) {
+        r.wal_receiver_handle.take();
+        // r.wal_receiver_handle.take().map(JoinHandle::join);
+    }
+}
+
+pub fn drop_wal_receiver(timelineid: ZTimelineId, tenantid: ZTenantId) {
+    let mut receivers = WAL_RECEIVERS.lock().unwrap();
+    receivers.remove(&timelineid);
+
+    // Check if it was the last walreceiver of the tenant.
+    // TODO now we store one WalReceiverEntry per timeline,
+    // so this iterator looks a bit strange.
+    for (_timelineid, entry) in receivers.iter() {
+        if entry.tenantid == tenantid {
+            return;
+        }
+    }
+
+    // When last walreceiver of the tenant is gone, change state to Idle
+    tenant_mgr::set_tenant_state(tenantid, TenantState::Idle).unwrap();
+}
+
 // Launch a new WAL receiver, or tell one that's running about change in connection string
 pub fn launch_wal_receiver(
     conf: &'static PageServerConf,
@@ -64,19 +99,24 @@ pub fn launch_wal_receiver(
             receiver.wal_producer_connstr = wal_producer_connstr.into();
         }
         None => {
-            let receiver = WalReceiverEntry {
-                wal_producer_connstr: wal_producer_connstr.into(),
-            };
-            receivers.insert(timelineid, receiver);
-
-            // Also launch a new thread to handle this connection
-            let _walreceiver_thread = thread::Builder::new()
+            let wal_receiver_handle = thread::Builder::new()
                 .name("WAL receiver thread".into())
                 .spawn(move || {
                     IS_WAL_RECEIVER.with(|c| c.set(true));
                     thread_main(conf, timelineid, tenantid);
                 })
                 .unwrap();
+
+            let receiver = WalReceiverEntry {
+                wal_producer_connstr: wal_producer_connstr.into(),
+                wal_receiver_handle: Some(wal_receiver_handle),
+                tenantid,
+            };
+            receivers.insert(timelineid, receiver);
+
+            // Update tenant state and start tenant threads, if they are not running yet.
+            tenant_mgr::set_tenant_state(tenantid, TenantState::Active).unwrap();
+            tenant_threads::start_tenant_threads(conf, tenantid);
         }
     };
 }
@@ -96,16 +136,18 @@ fn get_wal_producer_connstr(timelineid: ZTimelineId) -> String {
 // This is the entry point for the WAL receiver thread.
 //
 fn thread_main(conf: &'static PageServerConf, timelineid: ZTimelineId, tenantid: ZTenantId) {
-    info!(
-        "WAL receiver thread started for timeline : '{}'",
-        timelineid
-    );
+    let _enter = info_span!("WAL receiver", timeline = %timelineid, tenant = %tenantid).entered();
+    info!("WAL receiver thread started");
+
+    let mut retry_count = 10;
 
     //
     // Make a connection to the WAL safekeeper, or directly to the primary PostgreSQL server,
     // and start streaming WAL from it. If the connection is lost, keep retrying.
+    // TODO How long should we retry in case of losing connection?
+    // Should we retry at all or we can wait for the next callmemaybe request?
     //
-    loop {
+    while !tenant_mgr::shutdown_requested() && retry_count > 0 {
         // Look up the current WAL producer address
         let wal_producer_connstr = get_wal_producer_connstr(timelineid);
 
@@ -116,9 +158,20 @@ fn thread_main(conf: &'static PageServerConf, timelineid: ZTimelineId, tenantid:
                 "WAL streaming connection failed ({}), retrying in 1 second",
                 e
             );
+            retry_count -= 1;
             sleep(Duration::from_secs(1));
+        } else {
+            info!(
+                "walreceiver disconnected tenant {}, timelineid {}",
+                tenantid, timelineid
+            );
+            break;
         }
     }
+    info!("WAL streaming shut down");
+    // Drop it from list of active WAL_RECEIVERS
+    // so that next callmemaybe request launched a new thread
+    drop_wal_receiver(timelineid, tenantid);
 }
 
 fn walreceiver_main(
@@ -169,8 +222,8 @@ fn walreceiver_main(
     startpoint += startpoint.calc_padding(8u32);
 
     info!(
-        "last_record_lsn {} starting replication from {} for timeline {}, server is at {}...",
-        last_rec_lsn, startpoint, timelineid, end_of_wal
+        "last_record_lsn {} starting replication from {}, server is at {}...",
+        last_rec_lsn, startpoint, end_of_wal
     );
 
     let query = format!("START_REPLICATION PHYSICAL {}", startpoint);
@@ -198,27 +251,32 @@ fn walreceiver_main(
                 waldecoder.feed_bytes(data);
 
                 while let Some((lsn, recdata)) = waldecoder.poll_decode()? {
-                    // Save old checkpoint value to compare with it after decoding WAL record
-                    let old_checkpoint_bytes = checkpoint.encode();
-                    let decoded = decode_wal_record(recdata.clone());
+                    let _enter = info_span!("processing record", lsn = %lsn).entered();
 
                     // It is important to deal with the aligned records as lsn in getPage@LSN is
                     // aligned and can be several bytes bigger. Without this alignment we are
                     // at risk of hittind a deadlock.
                     assert!(lsn.is_aligned());
 
+                    let writer = timeline.writer();
+
+                    let mut checkpoint_modified = false;
+
+                    let decoded = decode_wal_record(recdata.clone());
                     restore_local_repo::save_decoded_record(
                         &mut checkpoint,
-                        &*timeline,
+                        &mut checkpoint_modified,
+                        writer.as_ref(),
                         &decoded,
                         recdata,
                         lsn,
                     )?;
 
-                    let new_checkpoint_bytes = checkpoint.encode();
                     // Check if checkpoint data was updated by save_decoded_record
-                    if new_checkpoint_bytes != old_checkpoint_bytes {
-                        timeline.put_page_image(
+                    if checkpoint_modified {
+                        let new_checkpoint_bytes = checkpoint.encode();
+
+                        writer.put_page_image(
                             RelishTag::Checkpoint,
                             0,
                             lsn,
@@ -228,7 +286,7 @@ fn walreceiver_main(
 
                     // Now that this record has been fully handled, including updating the
                     // checkpoint data, let the repository know that it is up-to-date to this LSN
-                    timeline.advance_last_record_lsn(lsn);
+                    writer.advance_last_record_lsn(lsn);
                     last_rec_lsn = lsn;
                 }
 
@@ -263,17 +321,35 @@ fn walreceiver_main(
         };
 
         if let Some(last_lsn) = status_update {
-            // TODO: More thought should go into what values are sent here.
             let last_lsn = PgLsn::from(u64::from(last_lsn));
+
+            // The last LSN we processed. It is not guaranteed to survive pageserver crash.
             let write_lsn = last_lsn;
-            let flush_lsn = last_lsn;
-            let apply_lsn = PgLsn::from(0);
+            // This value doesn't guarantee data durability, but it's ok.
+            // In setup with WAL service, pageserver durability is guaranteed by safekeepers.
+            // In setup without WAL service, we just don't care.
+            let flush_lsn = write_lsn;
+            // `disk_consistent_lsn` is the LSN at which page server guarantees persistence of all received data
+            // Depending on the setup we recieve WAL directly from Compute Node or
+            // from a WAL service.
+            //
+            // Senders use the feedback to determine if we are caught up:
+            // - Safekeepers are free to remove WAL preceding `apply_lsn`,
+            // as it will never be requested by this page server.
+            // - Compute Node uses 'apply_lsn' to calculate a lag for back pressure mechanism
+            // (delay WAL inserts to avoid lagging pageserver responses and WAL overflow).
+            let apply_lsn = PgLsn::from(u64::from(timeline.get_disk_consistent_lsn()));
             let ts = SystemTime::now();
             const NO_REPLY: u8 = 0;
-
             physical_stream.standby_status_update(write_lsn, flush_lsn, apply_lsn, ts, NO_REPLY)?;
         }
+
+        if tenant_mgr::shutdown_requested() {
+            debug!("stop walreceiver because pageserver shutdown is requested");
+            break;
+        }
     }
+
     Ok(())
 }
 

pageserver/src/walredo.rs

@@ -22,23 +22,24 @@ use byteorder::{ByteOrder, LittleEndian};
 use bytes::{Buf, BufMut, Bytes, BytesMut};
 use lazy_static::lazy_static;
 use log::*;
+use nix::poll::*;
 use serde::Serialize;
 use std::fs;
 use std::fs::OpenOptions;
 use std::io::prelude::*;
-use std::io::Error;
+use std::io::{Error, ErrorKind};
+use std::os::unix::io::AsRawFd;
 use std::path::PathBuf;
 use std::process::Stdio;
+use std::process::{Child, ChildStderr, ChildStdin, ChildStdout, Command};
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::Mutex;
 use std::time::Duration;
 use std::time::Instant;
-use tokio::io::AsyncBufReadExt;
-use tokio::io::{AsyncReadExt, AsyncWriteExt};
-use tokio::process::{ChildStdin, ChildStdout, Command};
-use tokio::time::timeout;
 use zenith_metrics::{register_histogram, register_int_counter, Histogram, IntCounter};
 use zenith_utils::bin_ser::BeSer;
 use zenith_utils::lsn::Lsn;
+use zenith_utils::nonblock::set_nonblock;
 use zenith_utils::zid::ZTenantId;
 
 use crate::relish::*;
@@ -53,6 +54,8 @@ use postgres_ffi::nonrelfile_utils::transaction_id_set_status;
 use postgres_ffi::pg_constants;
 use postgres_ffi::XLogRecord;
 
+const N_WAL_REDO_PROCS: usize = 1;
+
 ///
 /// `RelTag` + block number (`blknum`) gives us a unique id of the page in the cluster.
 ///
@@ -82,7 +85,7 @@ pub trait WalRedoManager: Send + Sync {
         blknum: u32,
         lsn: Lsn,
         base_img: Option<Bytes>,
-        records: Vec<WALRecord>,
+        records: Vec<(Lsn, WALRecord)>,
     ) -> Result<Bytes, WalRedoError>;
 }
 
@@ -99,7 +102,7 @@ impl crate::walredo::WalRedoManager for DummyRedoManager {
         _blknum: u32,
         _lsn: Lsn,
         _base_img: Option<Bytes>,
-        _records: Vec<WALRecord>,
+        _records: Vec<(Lsn, WALRecord)>,
     ) -> Result<Bytes, WalRedoError> {
         Err(WalRedoError::InvalidState)
     }
@@ -139,8 +142,8 @@ pub struct PostgresRedoManager {
     tenantid: ZTenantId,
     conf: &'static PageServerConf,
 
-    runtime: tokio::runtime::Runtime,
-    process: Mutex<Option<PostgresRedoProcess>>,
+    round_robin: AtomicUsize,
+    processes: [Mutex<Option<PostgresRedoProcess>>; N_WAL_REDO_PROCS],
 }
 
 #[derive(Debug)]
@@ -150,9 +153,16 @@ struct WalRedoRequest {
     lsn: Lsn,
 
     base_img: Option<Bytes>,
-    records: Vec<WALRecord>,
+    records: Vec<(Lsn, WALRecord)>,
 }
 
+impl WalRedoRequest {
+    // Can this request be served by zenith redo funcitons
+    // or we need to pass it to wal-redo postgres process?
+    fn can_apply_in_zenith(&self) -> bool {
+        !matches!(self.rel, RelishTag::Relation(_))
+    }
+}
 /// An error happened in WAL redo
 #[derive(Debug, thiserror::Error)]
 pub enum WalRedoError {
@@ -161,6 +171,8 @@ pub enum WalRedoError {
 
     #[error("cannot perform WAL redo now")]
     InvalidState,
+    #[error("cannot perform WAL redo for this request")]
+    InvalidRequest,
 }
 
 ///
@@ -179,10 +191,9 @@ impl WalRedoManager for PostgresRedoManager {
         blknum: u32,
         lsn: Lsn,
         base_img: Option<Bytes>,
-        records: Vec<WALRecord>,
+        records: Vec<(Lsn, WALRecord)>,
     ) -> Result<Bytes, WalRedoError> {
         let start_time;
-        let lock_time;
         let end_time;
 
         let request = WalRedoRequest {
@@ -194,26 +205,38 @@ impl WalRedoManager for PostgresRedoManager {
         };
 
         start_time = Instant::now();
-        let result = {
-            let mut process_guard = self.process.lock().unwrap();
-            lock_time = Instant::now();
+        let result;
+
+        if request.can_apply_in_zenith() {
+            result = self.handle_apply_request_zenith(&request);
+
+            end_time = Instant::now();
+            WAL_REDO_TIME.observe(end_time.duration_since(start_time).as_secs_f64());
+        } else {
+            let rr = self.round_robin.fetch_add(1, Ordering::Relaxed) % N_WAL_REDO_PROCS;
+            let mut process_guard = self.processes[rr].lock().unwrap();
+            let lock_time = Instant::now();
 
             // launch the WAL redo process on first use
             if process_guard.is_none() {
-                let p = self
-                    .runtime
-                    .block_on(PostgresRedoProcess::launch(self.conf, &self.tenantid))?;
+                let p = PostgresRedoProcess::launch(self.conf, &self.tenantid, rr)?;
                 *process_guard = Some(p);
             }
             let process = process_guard.as_mut().unwrap();
 
-            self.runtime
-                .block_on(self.handle_apply_request(process, &request))
-        };
-        end_time = Instant::now();
+            result = self.handle_apply_request_postgres(process, &request);
 
-        WAL_REDO_WAIT_TIME.observe(lock_time.duration_since(start_time).as_secs_f64());
-        WAL_REDO_TIME.observe(end_time.duration_since(lock_time).as_secs_f64());
+            WAL_REDO_WAIT_TIME.observe(lock_time.duration_since(start_time).as_secs_f64());
+            end_time = Instant::now();
+            WAL_REDO_TIME.observe(end_time.duration_since(lock_time).as_secs_f64());
+
+            // If something went wrong, don't try to reuse the process. Kill it, and
+            // next request will launch a new one.
+            if result.is_err() {
+                let process = process_guard.take().unwrap();
+                process.kill();
+            }
+        }
 
         result
     }
@@ -224,31 +247,57 @@ impl PostgresRedoManager {
     /// Create a new PostgresRedoManager.
     ///
     pub fn new(conf: &'static PageServerConf, tenantid: ZTenantId) -> PostgresRedoManager {
-        // We block on waiting for requests on the walredo request channel, but
-        // use async I/O to communicate with the child process. Initialize the
-        // runtime for the async part.
-        let runtime = tokio::runtime::Builder::new_current_thread()
-            .enable_all()
-            .build()
-            .unwrap();
-
         // The actual process is launched lazily, on first request.
         PostgresRedoManager {
-            runtime,
             tenantid,
             conf,
-            process: Mutex::new(None),
+            round_robin: AtomicUsize::new(0),
+            processes: [(); N_WAL_REDO_PROCS].map(|_| Mutex::new(None)),
         }
     }
 
     ///
-    /// Process one request for WAL redo.
+    /// Process one request for WAL redo using wal-redo postgres
     ///
-    async fn handle_apply_request(
+    fn handle_apply_request_postgres(
         &self,
         process: &mut PostgresRedoProcess,
         request: &WalRedoRequest,
     ) -> Result<Bytes, WalRedoError> {
+        let blknum = request.blknum;
+        let lsn = request.lsn;
+        let base_img = request.base_img.clone();
+        let records = &request.records;
+        let nrecords = records.len();
+
+        let start = Instant::now();
+
+        let apply_result: Result<Bytes, Error>;
+
+        if let RelishTag::Relation(rel) = request.rel {
+            // Relational WAL records are applied using wal-redo-postgres
+            let buf_tag = BufferTag { rel, blknum };
+            apply_result = process.apply_wal_records(buf_tag, base_img, records);
+
+            let duration = start.elapsed();
+
+            debug!(
+                "postgres applied {} WAL records in {} us to reconstruct page image at LSN {}",
+                nrecords,
+                duration.as_micros(),
+                lsn
+            );
+
+            apply_result.map_err(WalRedoError::IoError)
+        } else {
+            Err(WalRedoError::InvalidRequest)
+        }
+    }
+
+    ///
+    /// Process one request for WAL redo using custom zenith code
+    ///
+    fn handle_apply_request_zenith(&self, request: &WalRedoRequest) -> Result<Bytes, WalRedoError> {
         let rel = request.rel;
         let blknum = request.blknum;
         let lsn = request.lsn;
@@ -260,178 +309,158 @@ impl PostgresRedoManager {
         let start = Instant::now();
 
         let apply_result: Result<Bytes, Error>;
-        if let RelishTag::Relation(rel) = rel {
-            // Relational WAL records are applied using wal-redo-postgres
-            let buf_tag = BufferTag { rel, blknum };
-            apply_result = process.apply_wal_records(buf_tag, base_img, records).await;
+
+        // Non-relational WAL records are handled here, with custom code that has the
+        // same effects as the corresponding Postgres WAL redo function.
+        const ZERO_PAGE: [u8; 8192] = [0u8; 8192];
+        let mut page = BytesMut::new();
+        if let Some(fpi) = base_img {
+            // If full-page image is provided, then use it...
+            page.extend_from_slice(&fpi[..]);
         } else {
-            // Non-relational WAL records are handled here, with custom code that has the
-            // same effects as the corresponding Postgres WAL redo function.
-            const ZERO_PAGE: [u8; 8192] = [0u8; 8192];
-            let mut page = BytesMut::new();
-            if let Some(fpi) = base_img {
-                // If full-page image is provided, then use it...
-                page.extend_from_slice(&fpi[..]);
-            } else {
-                // otherwise initialize page with zeros
-                page.extend_from_slice(&ZERO_PAGE);
+            // otherwise initialize page with zeros
+            page.extend_from_slice(&ZERO_PAGE);
+        }
+        // Apply all collected WAL records
+        for (_lsn, record) in records {
+            let mut buf = record.rec.clone();
+
+            WAL_REDO_RECORD_COUNTER.inc();
+
+            // 1. Parse XLogRecord struct
+            // FIXME: refactor to avoid code duplication.
+            let xlogrec = XLogRecord::from_bytes(&mut buf);
+
+            //move to main data
+            // TODO probably, we should store some records in our special format
+            // to avoid this weird parsing on replay
+            let skip = (record.main_data_offset - pg_constants::SIZEOF_XLOGRECORD) as usize;
+            if buf.remaining() > skip {
+                buf.advance(skip);
             }
-            // Apply all collected WAL records
-            for record in records {
-                let mut buf = record.rec.clone();
 
-                WAL_REDO_RECORD_COUNTER.inc();
-
-                // 1. Parse XLogRecord struct
-                // FIXME: refactor to avoid code duplication.
-                let xlogrec = XLogRecord::from_bytes(&mut buf);
-
-                //move to main data
-                // TODO probably, we should store some records in our special format
-                // to avoid this weird parsing on replay
-                let skip = (record.main_data_offset - pg_constants::SIZEOF_XLOGRECORD) as usize;
-                if buf.remaining() > skip {
-                    buf.advance(skip);
-                }
-
-                if xlogrec.xl_rmid == pg_constants::RM_XACT_ID {
-                    // Transaction manager stuff
-                    let rec_segno = match rel {
-                        RelishTag::Slru { slru, segno } => {
-                            assert!(
-                                slru == SlruKind::Clog,
-                                "Not valid XACT relish tag {:?}",
-                                rel
+            if xlogrec.xl_rmid == pg_constants::RM_XACT_ID {
+                // Transaction manager stuff
+                let rec_segno = match rel {
+                    RelishTag::Slru { slru, segno } => {
+                        assert!(
+                            slru == SlruKind::Clog,
+                            "Not valid XACT relish tag {:?}",
+                            rel
+                        );
+                        segno
+                    }
+                    _ => panic!("Not valid XACT relish tag {:?}", rel),
+                };
+                let parsed_xact =
+                    XlXactParsedRecord::decode(&mut buf, xlogrec.xl_xid, xlogrec.xl_info);
+                if parsed_xact.info == pg_constants::XLOG_XACT_COMMIT
+                    || parsed_xact.info == pg_constants::XLOG_XACT_COMMIT_PREPARED
+                {
+                    transaction_id_set_status(
+                        parsed_xact.xid,
+                        pg_constants::TRANSACTION_STATUS_COMMITTED,
+                        &mut page,
+                    );
+                    for subxact in &parsed_xact.subxacts {
+                        let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
+                        let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                        let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                        // only update xids on the requested page
+                        if rec_segno == segno && blknum == rpageno {
+                            transaction_id_set_status(
+                                *subxact,
+                                pg_constants::TRANSACTION_STATUS_COMMITTED,
+                                &mut page,
                             );
-                            segno
-                        }
-                        _ => panic!("Not valid XACT relish tag {:?}", rel),
-                    };
-                    let parsed_xact =
-                        XlXactParsedRecord::decode(&mut buf, xlogrec.xl_xid, xlogrec.xl_info);
-                    if parsed_xact.info == pg_constants::XLOG_XACT_COMMIT
-                        || parsed_xact.info == pg_constants::XLOG_XACT_COMMIT_PREPARED
-                    {
-                        transaction_id_set_status(
-                            parsed_xact.xid,
-                            pg_constants::TRANSACTION_STATUS_COMMITTED,
-                            &mut page,
-                        );
-                        for subxact in &parsed_xact.subxacts {
-                            let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
-                            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            // only update xids on the requested page
-                            if rec_segno == segno && blknum == rpageno {
-                                transaction_id_set_status(
-                                    *subxact,
-                                    pg_constants::TRANSACTION_STATUS_COMMITTED,
-                                    &mut page,
-                                );
-                            }
-                        }
-                    } else if parsed_xact.info == pg_constants::XLOG_XACT_ABORT
-                        || parsed_xact.info == pg_constants::XLOG_XACT_ABORT_PREPARED
-                    {
-                        transaction_id_set_status(
-                            parsed_xact.xid,
-                            pg_constants::TRANSACTION_STATUS_ABORTED,
-                            &mut page,
-                        );
-                        for subxact in &parsed_xact.subxacts {
-                            let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
-                            let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                            // only update xids on the requested page
-                            if rec_segno == segno && blknum == rpageno {
-                                transaction_id_set_status(
-                                    *subxact,
-                                    pg_constants::TRANSACTION_STATUS_ABORTED,
-                                    &mut page,
-                                );
-                            }
                         }
                     }
-                } else if xlogrec.xl_rmid == pg_constants::RM_MULTIXACT_ID {
-                    // Multixact operations
-                    let info = xlogrec.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
-                    if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
-                        let xlrec = XlMultiXactCreate::decode(&mut buf);
-                        if let RelishTag::Slru {
-                            slru,
-                            segno: rec_segno,
-                        } = rel
-                        {
-                            if slru == SlruKind::MultiXactMembers {
-                                for i in 0..xlrec.nmembers {
-                                    let pageno =
-                                        i / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
-                                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                                    if segno == rec_segno && rpageno == blknum {
-                                        // update only target block
-                                        let offset = xlrec.moff + i;
-                                        let memberoff = mx_offset_to_member_offset(offset);
-                                        let flagsoff = mx_offset_to_flags_offset(offset);
-                                        let bshift = mx_offset_to_flags_bitshift(offset);
-                                        let mut flagsval =
-                                            LittleEndian::read_u32(&page[flagsoff..flagsoff + 4]);
-                                        flagsval &= !(((1
-                                            << pg_constants::MXACT_MEMBER_BITS_PER_XACT)
-                                            - 1)
+                } else if parsed_xact.info == pg_constants::XLOG_XACT_ABORT
+                    || parsed_xact.info == pg_constants::XLOG_XACT_ABORT_PREPARED
+                {
+                    transaction_id_set_status(
+                        parsed_xact.xid,
+                        pg_constants::TRANSACTION_STATUS_ABORTED,
+                        &mut page,
+                    );
+                    for subxact in &parsed_xact.subxacts {
+                        let pageno = *subxact as u32 / pg_constants::CLOG_XACTS_PER_PAGE;
+                        let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                        let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                        // only update xids on the requested page
+                        if rec_segno == segno && blknum == rpageno {
+                            transaction_id_set_status(
+                                *subxact,
+                                pg_constants::TRANSACTION_STATUS_ABORTED,
+                                &mut page,
+                            );
+                        }
+                    }
+                }
+            } else if xlogrec.xl_rmid == pg_constants::RM_MULTIXACT_ID {
+                // Multixact operations
+                let info = xlogrec.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+                if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
+                    let xlrec = XlMultiXactCreate::decode(&mut buf);
+                    if let RelishTag::Slru {
+                        slru,
+                        segno: rec_segno,
+                    } = rel
+                    {
+                        if slru == SlruKind::MultiXactMembers {
+                            for i in 0..xlrec.nmembers {
+                                let pageno = i / pg_constants::MULTIXACT_MEMBERS_PER_PAGE as u32;
+                                let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                                let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                                if segno == rec_segno && rpageno == blknum {
+                                    // update only target block
+                                    let offset = xlrec.moff + i;
+                                    let memberoff = mx_offset_to_member_offset(offset);
+                                    let flagsoff = mx_offset_to_flags_offset(offset);
+                                    let bshift = mx_offset_to_flags_bitshift(offset);
+                                    let mut flagsval =
+                                        LittleEndian::read_u32(&page[flagsoff..flagsoff + 4]);
+                                    flagsval &=
+                                        !(((1 << pg_constants::MXACT_MEMBER_BITS_PER_XACT) - 1)
                                             << bshift);
-                                        flagsval |= xlrec.members[i as usize].status << bshift;
-                                        LittleEndian::write_u32(
-                                            &mut page[flagsoff..flagsoff + 4],
-                                            flagsval,
-                                        );
-                                        LittleEndian::write_u32(
-                                            &mut page[memberoff..memberoff + 4],
-                                            xlrec.members[i as usize].xid,
-                                        );
-                                    }
+                                    flagsval |= xlrec.members[i as usize].status << bshift;
+                                    LittleEndian::write_u32(
+                                        &mut page[flagsoff..flagsoff + 4],
+                                        flagsval,
+                                    );
+                                    LittleEndian::write_u32(
+                                        &mut page[memberoff..memberoff + 4],
+                                        xlrec.members[i as usize].xid,
+                                    );
                                 }
-                            } else {
-                                // Multixact offsets SLRU
-                                let offs = (xlrec.mid
-                                    % pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32
-                                    * 4) as usize;
-                                LittleEndian::write_u32(&mut page[offs..offs + 4], xlrec.moff);
                             }
                         } else {
-                            panic!();
+                            // Multixact offsets SLRU
+                            let offs = (xlrec.mid % pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32
+                                * 4) as usize;
+                            LittleEndian::write_u32(&mut page[offs..offs + 4], xlrec.moff);
                         }
                     } else {
                         panic!();
                     }
+                } else {
+                    panic!();
                 }
             }
-
-            apply_result = Ok::<Bytes, Error>(page.freeze());
         }
 
+        apply_result = Ok::<Bytes, Error>(page.freeze());
+
         let duration = start.elapsed();
 
-        let result: Result<Bytes, WalRedoError>;
-
         debug!(
-            "applied {} WAL records in {} ms to reconstruct page image at LSN {}",
+            "zenith applied {} WAL records in {} ms to reconstruct page image at LSN {}",
             nrecords,
             duration.as_millis(),
             lsn
         );
 
-        if let Err(e) = apply_result {
-            error!("could not apply WAL records: {:#}", e);
-            result = Err(WalRedoError::IoError(e));
-        } else {
-            let img = apply_result.unwrap();
-
-            result = Ok(img);
-        }
-
-        // The caller is responsible for sending the response
-        result
+        apply_result.map_err(WalRedoError::IoError)
     }
 }
 
@@ -439,22 +468,27 @@ impl PostgresRedoManager {
 /// Handle to the Postgres WAL redo process
 ///
 struct PostgresRedoProcess {
+    child: Child,
     stdin: ChildStdin,
     stdout: ChildStdout,
+    stderr: ChildStderr,
 }
 
 impl PostgresRedoProcess {
     //
     // Start postgres binary in special WAL redo mode.
     //
-    async fn launch(
+    fn launch(
         conf: &PageServerConf,
         tenantid: &ZTenantId,
+        id: usize,
     ) -> Result<PostgresRedoProcess, Error> {
         // FIXME: We need a dummy Postgres cluster to run the process in. Currently, we
         // just create one with constant name. That fails if you try to launch more than
         // one WAL redo manager concurrently.
-        let datadir = conf.tenant_path(tenantid).join("wal-redo-datadir");
+        let datadir = conf
+            .tenant_path(tenantid)
+            .join(format! {"wal-redo-datadir-{}", id});
 
         // Create empty data directory for wal-redo postgres, deleting old one first.
         if datadir.exists() {
@@ -471,7 +505,6 @@ impl PostgresRedoProcess {
             .env("LD_LIBRARY_PATH", conf.pg_lib_dir().to_str().unwrap())
             .env("DYLD_LIBRARY_PATH", conf.pg_lib_dir().to_str().unwrap())
             .output()
-            .await
             .expect("failed to execute initdb");
 
         if !initdb.status.success() {
@@ -508,108 +541,139 @@ impl PostgresRedoProcess {
             datadir.display()
         );
 
-        let stdin = child.stdin.take().expect("failed to open child's stdin");
-        let stderr = child.stderr.take().expect("failed to open child's stderr");
-        let stdout = child.stdout.take().expect("failed to open child's stdout");
+        let stdin = child.stdin.take().unwrap();
+        let stdout = child.stdout.take().unwrap();
+        let stderr = child.stderr.take().unwrap();
 
-        // This async block reads the child's stderr, and forwards it to the logger
-        let f_stderr = async {
-            let mut stderr_buffered = tokio::io::BufReader::new(stderr);
+        set_nonblock(stdin.as_raw_fd())?;
+        set_nonblock(stdout.as_raw_fd())?;
+        set_nonblock(stderr.as_raw_fd())?;
 
-            let mut line = String::new();
-            loop {
-                let res = stderr_buffered.read_line(&mut line).await;
-                if res.is_err() {
-                    debug!("could not convert line to utf-8");
-                    continue;
-                }
-                if res.unwrap() == 0 {
-                    break;
-                }
-                error!("wal-redo-postgres: {}", line.trim());
-                line.clear();
-            }
-            Ok::<(), Error>(())
-        };
-        tokio::spawn(f_stderr);
+        Ok(PostgresRedoProcess {
+            child,
+            stdin,
+            stdout,
+            stderr,
+        })
+    }
 
-        Ok(PostgresRedoProcess { stdin, stdout })
+    fn kill(mut self) {
+        let _ = self.child.kill();
+        if let Ok(exit_status) = self.child.wait() {
+            error!("wal-redo-postgres exited with code {}", exit_status);
+        }
+        drop(self);
     }
 
     //
     // Apply given WAL records ('records') over an old page image. Returns
     // new page image.
     //
-    async fn apply_wal_records(
+    fn apply_wal_records(
         &mut self,
         tag: BufferTag,
         base_img: Option<Bytes>,
-        records: &[WALRecord],
+        records: &[(Lsn, WALRecord)],
     ) -> Result<Bytes, std::io::Error> {
-        let stdout = &mut self.stdout;
-        // Buffer the writes to avoid a lot of small syscalls.
-        let mut stdin = tokio::io::BufWriter::new(&mut self.stdin);
+        // Serialize all the messages to send the WAL redo process first.
+        //
+        // This could be problematic if there are millions of records to replay,
+        // but in practice the number of records is usually so small that it doesn't
+        // matter, and it's better to keep this code simple.
+        let mut writebuf: Vec<u8> = Vec::new();
+        build_begin_redo_for_block_msg(tag, &mut writebuf);
+        if let Some(img) = base_img {
+            build_push_page_msg(tag, &img, &mut writebuf);
+        }
+        for (lsn, rec) in records.iter() {
+            build_apply_record_msg(*lsn, &rec.rec, &mut writebuf);
+        }
+        build_get_page_msg(tag, &mut writebuf);
+        WAL_REDO_RECORD_COUNTER.inc_by(records.len() as u64);
+
+        // The input is now in 'writebuf'. Do a blind write first, writing as much as
+        // we can, before calling poll(). That skips one call to poll() if the stdin is
+        // already available for writing, which it almost certainly is because the
+        // process is idle.
+        let mut nwrite = self.stdin.write(&writebuf)?;
+
+        // We expect the WAL redo process to respond with an 8k page image. We read it
+        // into this buffer.
+        let mut resultbuf = vec![0; pg_constants::BLCKSZ.into()];
+        let mut nresult: usize = 0; // # of bytes read into 'resultbuf' so far
+
+        // Prepare for calling poll()
+        let mut pollfds = [
+            PollFd::new(self.stdout.as_raw_fd(), PollFlags::POLLIN),
+            PollFd::new(self.stderr.as_raw_fd(), PollFlags::POLLIN),
+            PollFd::new(self.stdin.as_raw_fd(), PollFlags::POLLOUT),
+        ];
 
         // We do three things simultaneously: send the old base image and WAL records to
         // the child process's stdin, read the result from child's stdout, and forward any logging
         // information that the child writes to its stderr to the page server's log.
-        //
-        // 'f_stdin' handles writing the base image and WAL records to the child process.
-        // 'f_stdout' below reads the result back. And 'f_stderr', which was spawned into the
-        // tokio runtime in the 'launch' function already, forwards the logging.
-        let f_stdin = async {
-            // Send base image, if any. (If the record initializes the page, previous page
-            // version is not needed.)
-            timeout(
-                TIMEOUT,
-                stdin.write_all(&build_begin_redo_for_block_msg(tag)),
-            )
-            .await??;
-            if base_img.is_some() {
-                timeout(
-                    TIMEOUT,
-                    stdin.write_all(&build_push_page_msg(tag, base_img.unwrap())),
-                )
-                .await??;
+        while nresult < pg_constants::BLCKSZ.into() {
+            // If we have more data to write, wake up if 'stdin' becomes writeable or
+            // we have data to read. Otherwise only wake up if there's data to read.
+            let nfds = if nwrite < writebuf.len() { 3 } else { 2 };
+            let n = nix::poll::poll(&mut pollfds[0..nfds], TIMEOUT.as_millis() as i32)?;
+
+            if n == 0 {
+                return Err(Error::new(ErrorKind::Other, "WAL redo timed out"));
             }
 
-            // Send WAL records.
-            for rec in records.iter() {
-                let r = rec.clone();
+            // If we have some messages in stderr, forward them to the log.
+            let err_revents = pollfds[1].revents().unwrap();
+            if err_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
+                let mut errbuf: [u8; 16384] = [0; 16384];
+                let n = self.stderr.read(&mut errbuf)?;
 
-                WAL_REDO_RECORD_COUNTER.inc();
+                // The message might not be split correctly into lines here. But this is
+                // good enough, the important thing is to get the message to the log.
+                if n > 0 {
+                    error!(
+                        "wal-redo-postgres: {}",
+                        String::from_utf8_lossy(&errbuf[0..n])
+                    );
 
-                stdin
-                    .write_all(&build_apply_record_msg(r.lsn, r.rec))
-                    .await?;
-
-                //debug!("sent WAL record to wal redo postgres process ({:X}/{:X}",
-                //       r.lsn >> 32, r.lsn & 0xffff_ffff);
+                    // To make sure we capture all log from the process if it fails, keep
+                    // reading from the stderr, before checking the stdout.
+                    continue;
+                }
+            } else if err_revents.contains(PollFlags::POLLHUP) {
+                return Err(Error::new(
+                    ErrorKind::BrokenPipe,
+                    "WAL redo process closed its stderr unexpectedly",
+                ));
             }
-            //debug!("sent {} WAL records to wal redo postgres process ({:X}/{:X}",
-            //       records.len(), lsn >> 32, lsn & 0xffff_ffff);
 
-            // Send GetPage command to get the result back
-            timeout(TIMEOUT, stdin.write_all(&build_get_page_msg(tag))).await??;
-            timeout(TIMEOUT, stdin.flush()).await??;
-            //debug!("sent GetPage for {}", tag.blknum);
-            Ok::<(), Error>(())
-        };
+            // If we have more data to write and 'stdin' is writeable, do write.
+            if nwrite < writebuf.len() {
+                let in_revents = pollfds[2].revents().unwrap();
+                if in_revents & (PollFlags::POLLERR | PollFlags::POLLOUT) != PollFlags::empty() {
+                    nwrite += self.stdin.write(&writebuf[nwrite..])?;
+                } else if in_revents.contains(PollFlags::POLLHUP) {
+                    // We still have more data to write, but the process closed the pipe.
+                    return Err(Error::new(
+                        ErrorKind::BrokenPipe,
+                        "WAL redo process closed its stdin unexpectedly",
+                    ));
+                }
+            }
 
-        // Read back new page image
-        let f_stdout = async {
-            let mut buf = [0u8; 8192];
+            // If we have some data in stdout, read it to the result buffer.
+            let out_revents = pollfds[0].revents().unwrap();
+            if out_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
+                nresult += self.stdout.read(&mut resultbuf[nresult..])?;
+            } else if out_revents.contains(PollFlags::POLLHUP) {
+                return Err(Error::new(
+                    ErrorKind::BrokenPipe,
+                    "WAL redo process closed its stdout unexpectedly",
+                ));
+            }
+        }
 
-            timeout(TIMEOUT, stdout.read_exact(&mut buf)).await??;
-            //debug!("got response for {}", tag.blknum);
-            Ok::<[u8; 8192], Error>(buf)
-        };
-
-        let res = tokio::try_join!(f_stdout, f_stdin)?;
-
-        let buf = res.0;
-
-        Ok::<Bytes, Error>(Bytes::from(std::vec::Vec::from(buf)))
+        Ok(Bytes::from(resultbuf))
     }
 }
 
@@ -617,81 +681,42 @@ impl PostgresRedoProcess {
 // process. See vendor/postgres/src/backend/tcop/zenith_wal_redo.c for
 // explanation of the protocol.
 
-fn build_begin_redo_for_block_msg(tag: BufferTag) -> Bytes {
+fn build_begin_redo_for_block_msg(tag: BufferTag, buf: &mut Vec<u8>) {
     let len = 4 + 1 + 4 * 4;
-    let mut buf = BytesMut::with_capacity(1 + len);
 
     buf.put_u8(b'B');
     buf.put_u32(len as u32);
 
-    // FIXME: this is a temporary hack that should go away when we refactor
-    // the postgres protocol serialization + handlers.
-    //
-    // BytesMut is a dynamic growable buffer, used a lot in tokio code but
-    // not in the std library. To write to a BytesMut from a serde serializer,
-    // we need to either:
-    // - pre-allocate the required buffer space. This is annoying because we
-    //   shouldn't care what the exact serialized size is-- that's the
-    //   serializer's job.
-    // - Or, we need to create a temporary "writer" (which implements the
-    //   `Write` trait). It's a bit awkward, because the writer consumes the
-    //   underlying BytesMut, and we need to extract it later with
-    //   `into_inner`.
-    let mut writer = buf.writer();
-    tag.ser_into(&mut writer)
+    tag.ser_into(buf)
         .expect("serialize BufferTag should always succeed");
-    let buf = writer.into_inner();
-
-    debug_assert!(buf.len() == 1 + len);
-
-    buf.freeze()
 }
 
-fn build_push_page_msg(tag: BufferTag, base_img: Bytes) -> Bytes {
+fn build_push_page_msg(tag: BufferTag, base_img: &[u8], buf: &mut Vec<u8>) {
     assert!(base_img.len() == 8192);
 
     let len = 4 + 1 + 4 * 4 + base_img.len();
-    let mut buf = BytesMut::with_capacity(1 + len);
 
     buf.put_u8(b'P');
     buf.put_u32(len as u32);
-    let mut writer = buf.writer();
-    tag.ser_into(&mut writer)
+    tag.ser_into(buf)
         .expect("serialize BufferTag should always succeed");
-    let mut buf = writer.into_inner();
     buf.put(base_img);
-
-    debug_assert!(buf.len() == 1 + len);
-
-    buf.freeze()
 }
 
-fn build_apply_record_msg(endlsn: Lsn, rec: Bytes) -> Bytes {
+fn build_apply_record_msg(endlsn: Lsn, rec: &[u8], buf: &mut Vec<u8>) {
     let len = 4 + 8 + rec.len();
-    let mut buf = BytesMut::with_capacity(1 + len);
 
     buf.put_u8(b'A');
     buf.put_u32(len as u32);
     buf.put_u64(endlsn.0);
     buf.put(rec);
-
-    debug_assert!(buf.len() == 1 + len);
-
-    buf.freeze()
 }
 
-fn build_get_page_msg(tag: BufferTag) -> Bytes {
+fn build_get_page_msg(tag: BufferTag, buf: &mut Vec<u8>) {
     let len = 4 + 1 + 4 * 4;
-    let mut buf = BytesMut::with_capacity(1 + len);
 
     buf.put_u8(b'G');
     buf.put_u32(len as u32);
-    let mut writer = buf.writer();
-    tag.ser_into(&mut writer)
+    tag.ser_into(buf)
         .expect("serialize BufferTag should always succeed");
-    let buf = writer.into_inner();
-
-    debug_assert!(buf.len() == 1 + len);
-
-    buf.freeze()
 }

postgres_ffi/src/xlog_utils.rs

@@ -9,7 +9,6 @@
 
 use crate::pg_constants;
 use crate::CheckPoint;
-use crate::ControlFileData;
 use crate::FullTransactionId;
 use crate::XLogLongPageHeaderData;
 use crate::XLogPageHeaderData;
@@ -18,8 +17,8 @@ use crate::XLOG_PAGE_MAGIC;
 
 use anyhow::{bail, Result};
 use byteorder::{ByteOrder, LittleEndian};
+use bytes::BytesMut;
 use bytes::{Buf, Bytes};
-use bytes::{BufMut, BytesMut};
 use crc32c::*;
 use log::*;
 use std::cmp::max;
@@ -44,6 +43,9 @@ pub const XLOG_SIZE_OF_XLOG_RECORD: usize = std::mem::size_of::<XLogRecord>();
 #[allow(clippy::identity_op)]
 pub const SIZE_OF_XLOG_RECORD_DATA_HEADER_SHORT: usize = 1 * 2;
 
+// PG timeline is always 1, changing it doesn't have useful meaning in Zenith.
+pub const PG_TLI: u32 = 1;
+
 pub type XLogRecPtr = u64;
 pub type TimeLineID = u32;
 pub type TimestampTz = i64;
@@ -185,8 +187,13 @@ fn find_end_of_wal_segment(
             let xl_tot_len = LittleEndian::read_u32(&buf[page_offs..page_offs + 4]) as usize;
             if xl_tot_len == 0 {
                 info!(
-                    "find_end_of_wal_segment reached zeros at {:?}",
-                    Lsn(XLogSegNoOffsetToRecPtr(segno, offs as u32, wal_seg_size))
+                    "find_end_of_wal_segment reached zeros at {:?}, last records ends at {:?}",
+                    Lsn(XLogSegNoOffsetToRecPtr(segno, offs as u32, wal_seg_size)),
+                    Lsn(XLogSegNoOffsetToRecPtr(
+                        segno,
+                        last_valid_rec_pos as u32,
+                        wal_seg_size
+                    ))
                 );
                 break; // zeros, reached the end
             }
@@ -301,12 +308,17 @@ pub fn find_end_of_wal(
                     high_segno,
                 );
             }
+            let start_offset = if start_lsn.segment_number(wal_seg_size) == high_segno {
+                start_lsn.segment_offset(wal_seg_size)
+            } else {
+                0
+            };
             high_offs = find_end_of_wal_segment(
                 data_dir,
                 high_segno,
                 high_tli,
                 wal_seg_size,
-                start_lsn.segment_offset(wal_seg_size),
+                start_offset,
             )?;
         }
         let high_ptr = XLogSegNoOffsetToRecPtr(high_segno, high_offs, wal_seg_size);
@@ -329,7 +341,12 @@ pub fn main() {
 }
 
 impl XLogRecord {
-    pub fn from_bytes(buf: &mut Bytes) -> XLogRecord {
+    pub fn from_slice(buf: &[u8]) -> XLogRecord {
+        use zenith_utils::bin_ser::LeSer;
+        XLogRecord::des(buf).unwrap()
+    }
+
+    pub fn from_bytes<B: Buf>(buf: &mut B) -> XLogRecord {
         use zenith_utils::bin_ser::LeSer;
         XLogRecord::des_from(&mut buf.reader()).unwrap()
     }
@@ -377,10 +394,12 @@ impl CheckPoint {
         Ok(CheckPoint::des(buf)?)
     }
 
-    // Update next XID based on provided new_xid and stored epoch.
-    // Next XID should be greater than new_xid.
-    // Also take in account 32-bit wrap-around.
-    pub fn update_next_xid(&mut self, xid: u32) {
+    /// Update next XID based on provided new_xid and stored epoch.
+    /// Next XID should be greater than new_xid. This handles 32-bit
+    /// XID wraparound correctly.
+    ///
+    /// Returns 'true' if the XID was updated.
+    pub fn update_next_xid(&mut self, xid: u32) -> bool {
         let xid = xid.wrapping_add(XID_CHECKPOINT_INTERVAL - 1) & !(XID_CHECKPOINT_INTERVAL - 1);
         let full_xid = self.nextXid.value;
         let new_xid = std::cmp::max(xid + 1, pg_constants::FIRST_NORMAL_TRANSACTION_ID);
@@ -391,35 +410,37 @@ impl CheckPoint {
                 // wrap-around
                 epoch += 1;
             }
-            self.nextXid = FullTransactionId {
-                value: (epoch << 32) | new_xid as u64,
-            };
+            let nextXid = (epoch << 32) | new_xid as u64;
+
+            if nextXid != self.nextXid.value {
+                self.nextXid = FullTransactionId { value: nextXid };
+                return true;
+            }
         }
+        false
     }
 }
 
 //
-// Generate new WAL segment with single XLOG_CHECKPOINT_SHUTDOWN record.
+// Generate new, empty WAL segment.
 // We need this segment to start compute node.
-// In order to minimize changes in Postgres core, we prefer to
-// provide WAL segment from which is can extract checkpoint record in standard way,
-// rather then implement some alternative mechanism.
 //
-pub fn generate_wal_segment(pg_control: &ControlFileData) -> Bytes {
+pub fn generate_wal_segment(segno: u64, system_id: u64) -> Bytes {
     let mut seg_buf = BytesMut::with_capacity(pg_constants::WAL_SEGMENT_SIZE as usize);
 
+    let pageaddr = XLogSegNoOffsetToRecPtr(segno, 0, pg_constants::WAL_SEGMENT_SIZE);
     let hdr = XLogLongPageHeaderData {
         std: {
             XLogPageHeaderData {
                 xlp_magic: XLOG_PAGE_MAGIC as u16,
                 xlp_info: pg_constants::XLP_LONG_HEADER,
-                xlp_tli: 1, // FIXME: always use Postgres timeline 1
-                xlp_pageaddr: pg_control.checkPoint - XLOG_SIZE_OF_XLOG_LONG_PHD as u64,
+                xlp_tli: PG_TLI,
+                xlp_pageaddr: pageaddr,
                 xlp_rem_len: 0,
                 ..Default::default() // Put 0 in padding fields.
             }
         },
-        xlp_sysid: pg_control.system_identifier,
+        xlp_sysid: system_id,
         xlp_seg_size: pg_constants::WAL_SEGMENT_SIZE as u32,
         xlp_xlog_blcksz: XLOG_BLCKSZ as u32,
     };
@@ -427,36 +448,6 @@ pub fn generate_wal_segment(pg_control: &ControlFileData) -> Bytes {
     let hdr_bytes = hdr.encode();
     seg_buf.extend_from_slice(&hdr_bytes);
 
-    let rec_hdr = XLogRecord {
-        xl_tot_len: (XLOG_SIZE_OF_XLOG_RECORD
-            + SIZE_OF_XLOG_RECORD_DATA_HEADER_SHORT
-            + SIZEOF_CHECKPOINT) as u32,
-        xl_xid: 0, //0 is for InvalidTransactionId
-        xl_prev: 0,
-        xl_info: pg_constants::XLOG_CHECKPOINT_SHUTDOWN,
-        xl_rmid: pg_constants::RM_XLOG_ID,
-        xl_crc: 0,
-        ..Default::default() // Put 0 in padding fields.
-    };
-
-    let mut rec_shord_hdr_bytes = BytesMut::new();
-    rec_shord_hdr_bytes.put_u8(pg_constants::XLR_BLOCK_ID_DATA_SHORT);
-    rec_shord_hdr_bytes.put_u8(SIZEOF_CHECKPOINT as u8);
-
-    let rec_bytes = rec_hdr.encode();
-    let checkpoint_bytes = pg_control.checkPointCopy.encode();
-
-    //calculate record checksum
-    let mut crc = 0;
-    crc = crc32c_append(crc, &rec_shord_hdr_bytes[..]);
-    crc = crc32c_append(crc, &checkpoint_bytes[..]);
-    crc = crc32c_append(crc, &rec_bytes[0..XLOG_RECORD_CRC_OFFS]);
-
-    seg_buf.extend_from_slice(&rec_bytes[0..XLOG_RECORD_CRC_OFFS]);
-    seg_buf.put_u32_le(crc);
-    seg_buf.extend_from_slice(&rec_shord_hdr_bytes);
-    seg_buf.extend_from_slice(&checkpoint_bytes);
-
     //zero out the rest of the file
     seg_buf.resize(pg_constants::WAL_SEGMENT_SIZE, 0);
     seg_buf.freeze()

pre-commit.py

@@ -37,20 +37,27 @@ def rustfmt(fix_inplace: bool = False, no_color: bool = False) -> str:
     return cmd
 
 
+def yapf(fix_inplace: bool) -> str:
+    cmd = "pipenv run yapf --recursive"
+    if fix_inplace:
+        cmd += " --in-place"
+    else:
+        cmd += " --diff"
+    return cmd
+
+
+def mypy() -> str:
+    return "pipenv run mypy"
+
+
 def get_commit_files() -> List[str]:
-    files = subprocess.check_output(
-        "git diff --cached --name-only --diff-filter=ACM".split()
-    )
+    files = subprocess.check_output("git diff --cached --name-only --diff-filter=ACM".split())
     return files.decode().splitlines()
 
 
-def check(
-    name: str, suffix: str, cmd: str, changed_files: List[str], no_color: bool = False
-):
+def check(name: str, suffix: str, cmd: str, changed_files: List[str], no_color: bool = False):
     print(f"Checking: {name} ", end="")
-    applicable_files = list(
-        filter(lambda fname: fname.strip().endswith(suffix), changed_files)
-    )
+    applicable_files = list(filter(lambda fname: fname.strip().endswith(suffix), changed_files))
     if not applicable_files:
         print(colorify("[NOT APPLICABLE]", Color.CYAN, no_color))
         return
@@ -59,7 +66,14 @@ def check(
     res = subprocess.run(cmd.split(), capture_output=True)
     if res.returncode != 0:
         print(colorify("[FAILED]", Color.RED, no_color))
-        print("Please inspect the output below and run make fmt to fix automatically\n")
+        if name == "mypy":
+            print("Please inspect the output below and fix type mismatches.")
+        else:
+            print("Please inspect the output below and run make fmt to fix automatically.")
+        if suffix == ".py":
+            print("If the output is empty, ensure that you've installed Python tooling by\n"
+                  "running 'pipenv install --dev' in the current directory (no root needed)")
+        print()
         print(res.stdout.decode())
         exit(1)
 
@@ -68,12 +82,11 @@ def check(
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--fix-inplace", action="store_true", help="apply fixes inplace"
-    )
-    parser.add_argument(
-        "--no-color", action="store_true", help="disable colored output", default=not sys.stdout.isatty()
-    )
+    parser.add_argument("--fix-inplace", action="store_true", help="apply fixes inplace")
+    parser.add_argument("--no-color",
+                        action="store_true",
+                        help="disable colored output",
+                        default=not sys.stdout.isatty())
     args = parser.parse_args()
 
     files = get_commit_files()
@@ -87,3 +100,17 @@ if __name__ == "__main__":
         changed_files=files,
         no_color=args.no_color,
     )
+    check(
+        name="yapf",
+        suffix=".py",
+        cmd=yapf(fix_inplace=args.fix_inplace),
+        changed_files=files,
+        no_color=args.no_color,
+    )
+    check(
+        name="mypy",
+        suffix=".py",
+        cmd=mypy(),
+        changed_files=files,
+        no_color=args.no_color,
+    )

proxy/src/cplane_api.rs

@@ -1,67 +1,139 @@
-use anyhow::{bail, Context, Result};
+use anyhow::{anyhow, bail, Context};
 use serde::{Deserialize, Serialize};
 use std::net::{SocketAddr, ToSocketAddrs};
 
-pub struct CPlaneApi {
-    auth_endpoint: &'static str,
-}
+use crate::state::ProxyWaiters;
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Default)]
 pub struct DatabaseInfo {
     pub host: String,
     pub port: u16,
     pub dbname: String,
     pub user: String,
-    pub password: String,
+    pub password: Option<String>,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+#[serde(untagged)]
+enum ProxyAuthResponse {
+    Ready { conn_info: DatabaseInfo },
+    Error { error: String },
+    NotReady { ready: bool }, // TODO: get rid of `ready`
 }
 
 impl DatabaseInfo {
-    pub fn socket_addr(&self) -> Result<SocketAddr> {
+    pub fn socket_addr(&self) -> anyhow::Result<SocketAddr> {
         let host_port = format!("{}:{}", self.host, self.port);
         host_port
             .to_socket_addrs()
             .with_context(|| format!("cannot resolve {} to SocketAddr", host_port))?
             .next()
-            .ok_or_else(|| anyhow::Error::msg("cannot resolve at least one SocketAddr"))
-    }
-
-    pub fn conn_string(&self) -> String {
-        format!(
-            "dbname={} user={} password={}",
-            self.dbname, self.user, self.password
-        )
+            .ok_or_else(|| anyhow!("cannot resolve at least one SocketAddr"))
     }
 }
 
-impl CPlaneApi {
-    pub fn new(auth_endpoint: &'static str) -> CPlaneApi {
-        CPlaneApi { auth_endpoint }
-    }
+impl From<DatabaseInfo> for tokio_postgres::Config {
+    fn from(db_info: DatabaseInfo) -> Self {
+        let mut config = tokio_postgres::Config::new();
 
+        config
+            .host(&db_info.host)
+            .port(db_info.port)
+            .dbname(&db_info.dbname)
+            .user(&db_info.user);
+
+        if let Some(password) = db_info.password {
+            config.password(password);
+        }
+
+        config
+    }
+}
+
+pub struct CPlaneApi<'a> {
+    auth_endpoint: &'a str,
+    waiters: &'a ProxyWaiters,
+}
+
+impl<'a> CPlaneApi<'a> {
+    pub fn new(auth_endpoint: &'a str, waiters: &'a ProxyWaiters) -> Self {
+        Self {
+            auth_endpoint,
+            waiters,
+        }
+    }
+}
+
+impl CPlaneApi<'_> {
     pub fn authenticate_proxy_request(
         &self,
         user: &str,
         database: &str,
         md5_response: &[u8],
         salt: &[u8; 4],
-    ) -> Result<DatabaseInfo> {
+        psql_session_id: &str,
+    ) -> anyhow::Result<DatabaseInfo> {
         let mut url = reqwest::Url::parse(self.auth_endpoint)?;
         url.query_pairs_mut()
             .append_pair("login", user)
             .append_pair("database", database)
             .append_pair("md5response", std::str::from_utf8(md5_response)?)
-            .append_pair("salt", &hex::encode(salt));
+            .append_pair("salt", &hex::encode(salt))
+            .append_pair("psql_session_id", psql_session_id);
 
-        println!("cplane request: {}", url.as_str());
+        let waiter = self.waiters.register(psql_session_id.to_owned());
 
+        println!("cplane request: {}", url);
         let resp = reqwest::blocking::get(url)?;
+        if !resp.status().is_success() {
+            bail!("Auth failed: {}", resp.status())
+        }
 
-        if resp.status().is_success() {
-            let conn_info: DatabaseInfo = serde_json::from_str(resp.text()?.as_str())?;
-            println!("got conn info: #{:?}", conn_info);
-            Ok(conn_info)
-        } else {
-            bail!("Auth failed")
+        let auth_info: ProxyAuthResponse = serde_json::from_str(resp.text()?.as_str())?;
+        println!("got auth info: #{:?}", auth_info);
+
+        use ProxyAuthResponse::*;
+        match auth_info {
+            Ready { conn_info } => Ok(conn_info),
+            Error { error } => bail!(error),
+            NotReady { .. } => waiter.wait()?.map_err(|e| anyhow!(e)),
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn test_proxy_auth_response() {
+        // Ready
+        let auth: ProxyAuthResponse = serde_json::from_value(json!({
+            "ready": true,
+            "conn_info": DatabaseInfo::default(),
+        }))
+        .unwrap();
+        assert!(matches!(
+            auth,
+            ProxyAuthResponse::Ready {
+                conn_info: DatabaseInfo { .. }
+            }
+        ));
+
+        // Error
+        let auth: ProxyAuthResponse = serde_json::from_value(json!({
+            "ready": false,
+            "error": "too bad, so sad",
+        }))
+        .unwrap();
+        assert!(matches!(auth, ProxyAuthResponse::Error { .. }));
+
+        // NotReady
+        let auth: ProxyAuthResponse = serde_json::from_value(json!({
+            "ready": false,
+        }))
+        .unwrap();
+        assert!(matches!(auth, ProxyAuthResponse::NotReady { .. }));
+    }
+}

proxy/src/main.rs

@@ -5,78 +5,21 @@
 /// (control plane API in our case) and can create new databases and accounts
 /// in somewhat transparent manner (again via communication with control plane API).
 ///
-use std::{
-    collections::HashMap,
-    net::{SocketAddr, TcpListener},
-    sync::{mpsc, Arc, Mutex},
-    thread,
-};
-
-use anyhow::{anyhow, bail, ensure, Context};
-use clap::{App, Arg, ArgMatches};
-
-use cplane_api::DatabaseInfo;
-use rustls::{internal::pemfile, NoClientAuth, ProtocolVersion, ServerConfig};
+use anyhow::bail;
+use clap::{App, Arg};
+use state::{ProxyConfig, ProxyState};
+use std::thread;
+use zenith_utils::{tcp_listener, GIT_VERSION};
 
 mod cplane_api;
 mod mgmt;
 mod proxy;
-
-pub struct ProxyConf {
-    /// main entrypoint for users to connect to
-    pub proxy_address: SocketAddr,
-
-    /// http management endpoint. Upon user account creation control plane
-    /// will notify us here, so that we can 'unfreeze' user session.
-    pub mgmt_address: SocketAddr,
-
-    /// send unauthenticated users to this URI
-    pub redirect_uri: String,
-
-    /// control plane address where we would check auth.
-    pub auth_endpoint: String,
-
-    pub ssl_config: Option<Arc<ServerConfig>>,
-}
-
-pub struct ProxyState {
-    pub conf: ProxyConf,
-    pub waiters: Mutex<HashMap<String, mpsc::Sender<anyhow::Result<DatabaseInfo>>>>,
-}
-
-fn configure_ssl(arg_matches: &ArgMatches) -> anyhow::Result<Option<Arc<ServerConfig>>> {
-    let (key_path, cert_path) = match (
-        arg_matches.value_of("ssl-key"),
-        arg_matches.value_of("ssl-cert"),
-    ) {
-        (Some(key_path), Some(cert_path)) => (key_path, cert_path),
-        (None, None) => return Ok(None),
-        _ => bail!("either both or neither ssl-key and ssl-cert must be specified"),
-    };
-
-    let key = {
-        let key_bytes = std::fs::read(key_path).context("SSL key file")?;
-        let mut keys = pemfile::pkcs8_private_keys(&mut &key_bytes[..])
-            .map_err(|_| anyhow!("couldn't read TLS keys"))?;
-        ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
-        keys.pop().unwrap()
-    };
-
-    let cert_chain = {
-        let cert_chain_bytes = std::fs::read(cert_path).context("SSL cert file")?;
-        pemfile::certs(&mut &cert_chain_bytes[..])
-            .map_err(|_| anyhow!("couldn't read TLS certificates"))?
-    };
-
-    let mut config = ServerConfig::new(NoClientAuth::new());
-    config.set_single_cert(cert_chain, key)?;
-    config.versions = vec![ProtocolVersion::TLSv1_3];
-
-    Ok(Some(Arc::new(config)))
-}
+mod state;
+mod waiters;
 
 fn main() -> anyhow::Result<()> {
     let arg_matches = App::new("Zenith proxy/router")
+        .version(GIT_VERSION)
         .arg(
             Arg::with_name("proxy")
                 .short("p")
@@ -125,38 +68,47 @@ fn main() -> anyhow::Result<()> {
         )
         .get_matches();
 
-    let conf = ProxyConf {
+    let ssl_config = match (
+        arg_matches.value_of("ssl-key"),
+        arg_matches.value_of("ssl-cert"),
+    ) {
+        (Some(key_path), Some(cert_path)) => {
+            Some(crate::state::configure_ssl(key_path, cert_path)?)
+        }
+        (None, None) => None,
+        _ => bail!("either both or neither ssl-key and ssl-cert must be specified"),
+    };
+
+    let config = ProxyConfig {
         proxy_address: arg_matches.value_of("proxy").unwrap().parse()?,
         mgmt_address: arg_matches.value_of("mgmt").unwrap().parse()?,
         redirect_uri: arg_matches.value_of("uri").unwrap().parse()?,
         auth_endpoint: arg_matches.value_of("auth-endpoint").unwrap().parse()?,
-        ssl_config: configure_ssl(&arg_matches)?,
+        ssl_config,
     };
-    let state = ProxyState {
-        conf,
-        waiters: Mutex::new(HashMap::new()),
-    };
-    let state: &'static ProxyState = Box::leak(Box::new(state));
+    let state: &ProxyState = Box::leak(Box::new(ProxyState::new(config)));
+
+    println!("Version: {}", GIT_VERSION);
 
     // Check that we can bind to address before further initialization
     println!("Starting proxy on {}", state.conf.proxy_address);
-    let pageserver_listener = TcpListener::bind(state.conf.proxy_address)?;
+    let pageserver_listener = tcp_listener::bind(state.conf.proxy_address)?;
 
     println!("Starting mgmt on {}", state.conf.mgmt_address);
-    let mgmt_listener = TcpListener::bind(state.conf.mgmt_address)?;
+    let mgmt_listener = tcp_listener::bind(state.conf.mgmt_address)?;
 
-    let threads = vec![
+    let threads = [
         // Spawn a thread to listen for connections. It will spawn further threads
         // for each connection.
         thread::Builder::new()
-            .name("Proxy thread".into())
+            .name("Listener thread".into())
             .spawn(move || proxy::thread_main(state, pageserver_listener))?,
         thread::Builder::new()
             .name("Mgmt thread".into())
             .spawn(move || mgmt::thread_main(state, mgmt_listener))?,
     ];
 
-    for t in threads.into_iter() {
+    for t in threads {
         t.join().unwrap()?;
     }
 

proxy/src/mgmt.rs

@@ -3,7 +3,6 @@ use std::{
     thread,
 };
 
-use anyhow::bail;
 use bytes::Bytes;
 use serde::Deserialize;
 use zenith_utils::{
@@ -25,22 +24,23 @@ pub fn thread_main(state: &'static ProxyState, listener: TcpListener) -> anyhow:
         socket.set_nodelay(true).unwrap();
 
         thread::spawn(move || {
-            if let Err(err) = mgmt_conn_main(state, socket) {
+            if let Err(err) = handle_connection(state, socket) {
                 println!("error: {}", err);
             }
         });
     }
 }
 
-pub fn mgmt_conn_main(state: &'static ProxyState, socket: TcpStream) -> anyhow::Result<()> {
+fn handle_connection(state: &ProxyState, socket: TcpStream) -> anyhow::Result<()> {
     let mut conn_handler = MgmtHandler { state };
-    let pgbackend = PostgresBackend::new(socket, AuthType::Trust, None)?;
+    let pgbackend = PostgresBackend::new(socket, AuthType::Trust, None, true)?;
     pgbackend.run(&mut conn_handler)
 }
 
-struct MgmtHandler {
-    state: &'static ProxyState,
+struct MgmtHandler<'a> {
+    state: &'a ProxyState,
 }
+
 /// Serialized examples:
 // {
 //     "session_id": "71d6d03e6d93d99a",
@@ -64,18 +64,18 @@ struct MgmtHandler {
 // // to test manually by sending a query to mgmt interface:
 // psql -h 127.0.0.1 -p 9999 -c '{"session_id":"4f10dde522e14739","result":{"Success":{"host":"127.0.0.1","port":5432,"dbname":"stas","user":"stas","password":"stas"}}}'
 #[derive(Deserialize)]
-pub struct PsqlSessionResponse {
+struct PsqlSessionResponse {
     session_id: String,
     result: PsqlSessionResult,
 }
 
 #[derive(Deserialize)]
-pub enum PsqlSessionResult {
+enum PsqlSessionResult {
     Success(DatabaseInfo),
     Failure(String),
 }
 
-impl postgres_backend::Handler for MgmtHandler {
+impl postgres_backend::Handler for MgmtHandler<'_> {
     fn process_query(
         &mut self,
         pgb: &mut PostgresBackend,
@@ -96,32 +96,26 @@ fn try_process_query(
     query_string: Bytes,
 ) -> anyhow::Result<()> {
     let query_string = query_from_cstring(query_string);
-
     println!("Got mgmt query: '{}'", std::str::from_utf8(&query_string)?);
 
     let resp: PsqlSessionResponse = serde_json::from_slice(&query_string)?;
 
-    let waiters = mgmt.state.waiters.lock().unwrap();
-
-    let sender = waiters
-        .get(&resp.session_id)
-        .ok_or_else(|| anyhow::Error::msg("psql_session_id is not found"))?;
-
-    match resp.result {
-        PsqlSessionResult::Success(db_info) => {
-            sender.send(Ok(db_info))?;
+    use PsqlSessionResult::*;
+    let msg = match resp.result {
+        Success(db_info) => Ok(db_info),
+        Failure(message) => Err(message),
+    };
 
+    match mgmt.state.waiters.notify(&resp.session_id, msg) {
+        Ok(()) => {
             pgb.write_message_noflush(&SINGLE_COL_ROWDESC)?
                 .write_message_noflush(&BeMessage::DataRow(&[Some(b"ok")]))?
-                .write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
-            pgb.flush()?;
-            Ok(())
+                .write_message(&BeMessage::CommandComplete(b"SELECT 1"))?;
         }
-
-        PsqlSessionResult::Failure(message) => {
-            sender.send(Err(anyhow::Error::msg(message.clone())))?;
-
-            bail!("psql session request failed: {}", message)
+        Err(e) => {
+            pgb.write_message(&BeMessage::ErrorResponse(e.to_string()))?;
         }
     }
+
+    Ok(())
 }

proxy/src/proxy.rs

@@ -1,18 +1,12 @@
-use crate::cplane_api::CPlaneApi;
-use crate::cplane_api::DatabaseInfo;
+use crate::cplane_api::{CPlaneApi, DatabaseInfo};
 use crate::ProxyState;
-
-use anyhow::bail;
+use anyhow::{anyhow, bail};
+use std::net::TcpStream;
+use std::{io, thread};
 use tokio_postgres::NoTls;
-
-use rand::Rng;
-use std::io::Write;
-use std::{io, sync::mpsc::channel, thread};
-use zenith_utils::postgres_backend::Stream;
-use zenith_utils::postgres_backend::{PostgresBackend, ProtoState};
-use zenith_utils::pq_proto::*;
+use zenith_utils::postgres_backend::{self, PostgresBackend, ProtoState, Stream};
+use zenith_utils::pq_proto::{BeMessage as Be, FeMessage as Fe, *};
 use zenith_utils::sock_split::{ReadStream, WriteStream};
-use zenith_utils::{postgres_backend, pq_proto::BeMessage};
 
 ///
 /// Main proxy listener loop.
@@ -28,272 +22,259 @@ pub fn thread_main(
         println!("accepted connection from {}", peer_addr);
         socket.set_nodelay(true).unwrap();
 
-        thread::spawn(move || {
-            if let Err(err) = proxy_conn_main(state, socket) {
-                println!("error: {}", err);
-            }
-        });
+        thread::Builder::new()
+            .name("Proxy thread".into())
+            .spawn(move || {
+                if let Err(err) = proxy_conn_main(state, socket) {
+                    println!("error: {}", err);
+                }
+            })?;
     }
 }
 
-// XXX: clean up fields
+// TODO: clean up fields
 struct ProxyConnection {
     state: &'static ProxyState,
-
-    cplane: CPlaneApi,
-
-    user: String,
-    database: String,
-
-    pgb: PostgresBackend,
-    md5_salt: [u8; 4],
-
     psql_session_id: String,
+    pgb: PostgresBackend,
 }
 
-pub fn proxy_conn_main(
-    state: &'static ProxyState,
-    socket: std::net::TcpStream,
-) -> anyhow::Result<()> {
-    let mut conn = ProxyConnection {
+pub fn proxy_conn_main(state: &'static ProxyState, socket: TcpStream) -> anyhow::Result<()> {
+    let conn = ProxyConnection {
         state,
-        cplane: CPlaneApi::new(&state.conf.auth_endpoint),
-        user: "".into(),
-        database: "".into(),
+        psql_session_id: hex::encode(rand::random::<[u8; 8]>()),
         pgb: PostgresBackend::new(
             socket,
             postgres_backend::AuthType::MD5,
             state.conf.ssl_config.clone(),
+            false,
         )?,
-        md5_salt: [0u8; 4],
-        psql_session_id: "".into(),
     };
 
-    // Check StartupMessage
-    // This will set conn.existing_user and we can decide on next actions
-    conn.handle_startup()?;
+    let (client, server) = conn.handle_client()?;
 
-    // both scenarious here should end up producing database connection string
-    let db_info = if conn.is_existing_user() {
-        conn.handle_existing_user()?
-    } else {
-        conn.handle_new_user()?
+    let server = zenith_utils::sock_split::BidiStream::from_tcp(server);
+
+    let client = match client {
+        Stream::Bidirectional(bidi_stream) => bidi_stream,
+        _ => panic!("invalid stream type"),
     };
 
-    // XXX: move that inside handle_new_user/handle_existing_user to be able to
-    // report wrong connection error.
-    proxy_pass(conn.pgb, db_info)
+    proxy(client.split(), server.split())
 }
 
 impl ProxyConnection {
-    fn is_existing_user(&self) -> bool {
-        self.user.ends_with("@zenith")
+    fn handle_client(mut self) -> anyhow::Result<(Stream, TcpStream)> {
+        let mut authenticate = || {
+            let (username, dbname) = self.handle_startup()?;
+
+            // Both scenarios here should end up producing database credentials
+            if username.ends_with("@zenith") {
+                self.handle_existing_user(&username, &dbname)
+            } else {
+                self.handle_new_user()
+            }
+        };
+
+        let conn = match authenticate() {
+            Ok(db_info) => connect_to_db(db_info),
+            Err(e) => {
+                // Report the error to the client
+                self.pgb.write_message(&Be::ErrorResponse(e.to_string()))?;
+                bail!("failed to handle client: {:?}", e);
+            }
+        };
+
+        // We'll get rid of this once migration to async is complete
+        let (pg_version, db_stream) = {
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()?;
+
+            let (pg_version, stream) = runtime.block_on(conn)?;
+            let stream = stream.into_std()?;
+            stream.set_nonblocking(false)?;
+
+            (pg_version, stream)
+        };
+
+        // Let the client send new requests
+        self.pgb
+            .write_message_noflush(&BeMessage::ParameterStatus(
+                BeParameterStatusMessage::ServerVersion(&pg_version),
+            ))?
+            .write_message(&Be::ReadyForQuery)?;
+
+        Ok((self.pgb.into_stream(), db_stream))
     }
 
-    fn handle_startup(&mut self) -> anyhow::Result<()> {
+    fn handle_startup(&mut self) -> anyhow::Result<(String, String)> {
+        let have_tls = self.pgb.tls_config.is_some();
         let mut encrypted = false;
+
         loop {
-            let msg = self.pgb.read_message()?;
-            println!("got message {:?}", msg);
-            match msg {
-                Some(FeMessage::StartupMessage(m)) => {
-                    println!("got startup message {:?}", m);
+            let mut msg = match self.pgb.read_message()? {
+                Some(Fe::StartupMessage(msg)) => msg,
+                None => bail!("connection is lost"),
+                bad => bail!("unexpected message type: {:?}", bad),
+            };
+            println!("got message: {:?}", msg);
 
-                    match m.kind {
-                        StartupRequestCode::NegotiateGss => {
-                            self.pgb
-                                .write_message(&BeMessage::EncryptionResponse(false))?;
-                        }
-                        StartupRequestCode::NegotiateSsl => {
-                            println!("SSL requested");
-                            if self.pgb.tls_config.is_some() {
-                                self.pgb
-                                    .write_message(&BeMessage::EncryptionResponse(true))?;
-                                self.pgb.start_tls()?;
-                                encrypted = true;
-                            } else {
-                                self.pgb
-                                    .write_message(&BeMessage::EncryptionResponse(false))?;
-                            }
-                        }
-                        StartupRequestCode::Normal => {
-                            if self.state.conf.ssl_config.is_some() && !encrypted {
-                                self.pgb.write_message(&BeMessage::ErrorResponse(
-                                    "must connect with TLS".to_string(),
-                                ))?;
-                                bail!("client did not connect with TLS");
-                            }
-                            self.user = m
-                                .params
-                                .get("user")
-                                .ok_or_else(|| {
-                                    anyhow::Error::msg("user is required in startup packet")
-                                })?
-                                .into();
-                            self.database = m
-                                .params
-                                .get("database")
-                                .ok_or_else(|| {
-                                    anyhow::Error::msg("database is required in startup packet")
-                                })?
-                                .into();
-
-                            break;
-                        }
-                        StartupRequestCode::Cancel => break,
+            match msg.kind {
+                StartupRequestCode::NegotiateGss => {
+                    self.pgb.write_message(&Be::EncryptionResponse(false))?;
+                }
+                StartupRequestCode::NegotiateSsl => {
+                    self.pgb.write_message(&Be::EncryptionResponse(have_tls))?;
+                    if have_tls {
+                        self.pgb.start_tls()?;
+                        encrypted = true;
                     }
                 }
-                None => {
-                    bail!("connection closed")
-                }
-                unexpected => {
-                    bail!("unexpected message type : {:?}", unexpected)
+                StartupRequestCode::Normal => {
+                    if have_tls && !encrypted {
+                        bail!("must connect with TLS");
+                    }
+
+                    let mut get_param = |key| {
+                        msg.params
+                            .remove(key)
+                            .ok_or_else(|| anyhow!("{} is missing in startup packet", key))
+                    };
+
+                    return Ok((get_param("user")?, get_param("database")?));
                 }
+                // TODO: implement proper stmt cancellation
+                StartupRequestCode::Cancel => bail!("query cancellation is not supported"),
             }
         }
-        Ok(())
     }
 
-    fn handle_existing_user(&mut self) -> anyhow::Result<DatabaseInfo> {
-        // ask password
-        rand::thread_rng().fill(&mut self.md5_salt);
+    fn handle_existing_user(&mut self, user: &str, db: &str) -> anyhow::Result<DatabaseInfo> {
+        let md5_salt = rand::random::<[u8; 4]>();
+
+        // Ask password
         self.pgb
-            .write_message(&BeMessage::AuthenticationMD5Password(&self.md5_salt))?;
+            .write_message(&Be::AuthenticationMD5Password(&md5_salt))?;
         self.pgb.state = ProtoState::Authentication; // XXX
 
-        // check password
-        println!("handle_existing_user");
-        let msg = self.pgb.read_message()?;
-        println!("got message {:?}", msg);
-        if let Some(FeMessage::PasswordMessage(m)) = msg {
-            println!("got password message '{:?}'", m);
+        // Check password
+        let msg = match self.pgb.read_message()? {
+            Some(Fe::PasswordMessage(msg)) => msg,
+            None => bail!("connection is lost"),
+            bad => bail!("unexpected message type: {:?}", bad),
+        };
+        println!("got message: {:?}", msg);
 
-            assert!(self.is_existing_user());
+        let (_trailing_null, md5_response) = msg
+            .split_last()
+            .ok_or_else(|| anyhow!("unexpected password message"))?;
 
-            let (_trailing_null, md5_response) = m
-                .split_last()
-                .ok_or_else(|| anyhow::Error::msg("unexpected password message"))?;
+        let cplane = CPlaneApi::new(&self.state.conf.auth_endpoint, &self.state.waiters);
+        let db_info = cplane.authenticate_proxy_request(
+            user,
+            db,
+            md5_response,
+            &md5_salt,
+            &self.psql_session_id,
+        )?;
 
-            match self.cplane.authenticate_proxy_request(
-                self.user.as_str(),
-                self.database.as_str(),
-                md5_response,
-                &self.md5_salt,
-            ) {
-                Err(e) => {
-                    self.pgb
-                        .write_message(&BeMessage::ErrorResponse(format!("{}", e)))?;
+        self.pgb
+            .write_message_noflush(&Be::AuthenticationOk)?
+            .write_message_noflush(&BeParameterStatusMessage::encoding())?;
 
-                    bail!("auth failed: {}", e);
-                }
-                Ok(conn_info) => {
-                    self.pgb
-                        .write_message_noflush(&BeMessage::AuthenticationOk)?;
-                    self.pgb
-                        .write_message_noflush(&BeMessage::ParameterStatus)?;
-                    self.pgb.write_message(&BeMessage::ReadyForQuery)?;
-
-                    Ok(conn_info)
-                }
-            }
-        } else {
-            bail!("protocol violation");
-        }
+        Ok(db_info)
     }
 
     fn handle_new_user(&mut self) -> anyhow::Result<DatabaseInfo> {
-        let mut psql_session_id_buf = [0u8; 8];
-        rand::thread_rng().fill(&mut psql_session_id_buf);
-        self.psql_session_id = hex::encode(psql_session_id_buf);
+        let greeting = hello_message(&self.state.conf.redirect_uri, &self.psql_session_id);
 
-        let hello_message = format!("☀️  Welcome to Zenith!
-
-To proceed with database creation, open the following link:
-
-    {redirect_uri}{sess_id}
-
-It needs to be done once and we will send you '.pgpass' file, which will allow you to access or create
-databases without opening the browser.
-
-", redirect_uri = self.state.conf.redirect_uri, sess_id = self.psql_session_id);
+        // First, register this session
+        let waiter = self.state.waiters.register(self.psql_session_id.clone());
 
+        // Give user a URL to spawn a new database
         self.pgb
-            .write_message_noflush(&BeMessage::AuthenticationOk)?;
-        self.pgb
-            .write_message_noflush(&BeMessage::ParameterStatus)?;
-        self.pgb
-            .write_message(&BeMessage::NoticeResponse(hello_message))?;
-
-        // await for database creation
-        let (tx, rx) = channel::<anyhow::Result<DatabaseInfo>>();
-        let _ = self
-            .state
-            .waiters
-            .lock()
-            .unwrap()
-            .insert(self.psql_session_id.clone(), tx);
+            .write_message_noflush(&Be::AuthenticationOk)?
+            .write_message_noflush(&BeParameterStatusMessage::encoding())?
+            .write_message(&Be::NoticeResponse(greeting))?;
 
         // Wait for web console response
-        // XXX: respond with error to client
-        let dbinfo = rx.recv()??;
+        let db_info = waiter.wait()?.map_err(|e| anyhow!(e))?;
 
-        self.pgb.write_message_noflush(&BeMessage::NoticeResponse(
-            "Connecting to database.".to_string(),
-        ))?;
-        self.pgb.write_message(&BeMessage::ReadyForQuery)?;
+        self.pgb
+            .write_message_noflush(&Be::NoticeResponse("Connecting to database.".into()))?;
 
-        Ok(dbinfo)
+        Ok(db_info)
     }
 }
 
+fn hello_message(redirect_uri: &str, session_id: &str) -> String {
+    format!(
+        concat![
+            "☀️  Welcome to Zenith!\n",
+            "To proceed with database creation, open the following link:\n\n",
+            "    {redirect_uri}{session_id}\n\n",
+            "It needs to be done once and we will send you '.pgpass' file,\n",
+            "which will allow you to access or create ",
+            "databases without opening your web browser."
+        ],
+        redirect_uri = redirect_uri,
+        session_id = session_id,
+    )
+}
+
 /// Create a TCP connection to a postgres database, authenticate with it, and receive the ReadyForQuery message
-async fn connect_to_db(db_info: DatabaseInfo) -> anyhow::Result<tokio::net::TcpStream> {
+async fn connect_to_db(db_info: DatabaseInfo) -> anyhow::Result<(String, tokio::net::TcpStream)> {
     let mut socket = tokio::net::TcpStream::connect(db_info.socket_addr()?).await?;
-    let config = db_info.conn_string().parse::<tokio_postgres::Config>()?;
-    let _ = config.connect_raw(&mut socket, NoTls).await?;
-    Ok(socket)
+    let config = tokio_postgres::Config::from(db_info);
+    let (client, conn) = config.connect_raw(&mut socket, NoTls).await?;
+
+    let query = client.query_one("select current_setting('server_version')", &[]);
+
+    tokio::pin!(query, conn);
+
+    let version = tokio::select!(
+        x = query => x?.try_get(0)?,
+        _ = conn => bail!("connection closed too early"),
+    );
+
+    Ok((version, socket))
 }
 
 /// Concurrently proxy both directions of the client and server connections
 fn proxy(
-    client_read: ReadStream,
-    client_write: WriteStream,
-    server_read: ReadStream,
-    server_write: WriteStream,
+    (client_read, client_write): (ReadStream, WriteStream),
+    (server_read, server_write): (ReadStream, WriteStream),
 ) -> anyhow::Result<()> {
-    fn do_proxy(mut reader: ReadStream, mut writer: WriteStream) -> io::Result<()> {
-        std::io::copy(&mut reader, &mut writer)?;
-        writer.flush()?;
-        writer.shutdown(std::net::Shutdown::Both)
+    fn do_proxy(mut reader: impl io::Read, mut writer: WriteStream) -> io::Result<u64> {
+        /// FlushWriter will make sure that every message is sent as soon as possible
+        struct FlushWriter<W>(W);
+
+        impl<W: io::Write> io::Write for FlushWriter<W> {
+            fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+                // `std::io::copy` is guaranteed to exit if we return an error,
+                // so we can afford to lose `res` in case `flush` fails
+                let res = self.0.write(buf);
+                if res.is_ok() {
+                    self.flush()?;
+                }
+                res
+            }
+
+            fn flush(&mut self) -> io::Result<()> {
+                self.0.flush()
+            }
+        }
+
+        let res = std::io::copy(&mut reader, &mut FlushWriter(&mut writer));
+        writer.shutdown(std::net::Shutdown::Both)?;
+        res
     }
 
     let client_to_server_jh = thread::spawn(move || do_proxy(client_read, server_write));
 
-    let res1 = do_proxy(server_read, client_write);
-    let res2 = client_to_server_jh.join().unwrap();
-    res1?;
-    res2?;
+    do_proxy(server_read, client_write)?;
+    client_to_server_jh.join().unwrap()?;
 
     Ok(())
 }
-
-/// Proxy a client connection to a postgres database
-fn proxy_pass(pgb: PostgresBackend, db_info: DatabaseInfo) -> anyhow::Result<()> {
-    let runtime = tokio::runtime::Builder::new_current_thread()
-        .enable_all()
-        .build()?;
-    let db_stream = runtime.block_on(connect_to_db(db_info))?;
-    let db_stream = db_stream.into_std()?;
-    db_stream.set_nonblocking(false)?;
-
-    let db_stream = zenith_utils::sock_split::BidiStream::from_tcp(db_stream);
-    let (db_read, db_write) = db_stream.split();
-
-    let stream = match pgb.into_stream() {
-        Stream::Bidirectional(bidi_stream) => bidi_stream,
-        _ => bail!("invalid stream"),
-    };
-
-    let (client_read, client_write) = stream.split();
-    proxy(client_read, client_write, db_read, db_write)
-}

proxy/src/state.rs

@@ -0,0 +1,62 @@
+use crate::cplane_api::DatabaseInfo;
+use anyhow::{anyhow, ensure, Context};
+use rustls::{internal::pemfile, NoClientAuth, ProtocolVersion, ServerConfig};
+use std::net::SocketAddr;
+use std::sync::Arc;
+
+pub type SslConfig = Arc<ServerConfig>;
+
+pub struct ProxyConfig {
+    /// main entrypoint for users to connect to
+    pub proxy_address: SocketAddr,
+
+    /// http management endpoint. Upon user account creation control plane
+    /// will notify us here, so that we can 'unfreeze' user session.
+    pub mgmt_address: SocketAddr,
+
+    /// send unauthenticated users to this URI
+    pub redirect_uri: String,
+
+    /// control plane address where we would check auth.
+    pub auth_endpoint: String,
+
+    pub ssl_config: Option<SslConfig>,
+}
+
+pub type ProxyWaiters = crate::waiters::Waiters<Result<DatabaseInfo, String>>;
+
+pub struct ProxyState {
+    pub conf: ProxyConfig,
+    pub waiters: ProxyWaiters,
+}
+
+impl ProxyState {
+    pub fn new(conf: ProxyConfig) -> Self {
+        Self {
+            conf,
+            waiters: ProxyWaiters::default(),
+        }
+    }
+}
+
+pub fn configure_ssl(key_path: &str, cert_path: &str) -> anyhow::Result<SslConfig> {
+    let key = {
+        let key_bytes = std::fs::read(key_path).context("SSL key file")?;
+        let mut keys = pemfile::pkcs8_private_keys(&mut &key_bytes[..])
+            .map_err(|_| anyhow!("couldn't read TLS keys"))?;
+        ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
+        keys.pop().unwrap()
+    };
+
+    let cert_chain = {
+        let cert_chain_bytes = std::fs::read(cert_path).context("SSL cert file")?;
+        pemfile::certs(&mut &cert_chain_bytes[..])
+            .map_err(|_| anyhow!("couldn't read TLS certificates"))?
+    };
+
+    let mut config = ServerConfig::new(NoClientAuth::new());
+    config.set_single_cert(cert_chain, key)?;
+    config.versions = vec![ProtocolVersion::TLSv1_3];
+
+    Ok(config.into())
+}

proxy/src/waiters.rs

@@ -0,0 +1,58 @@
+use anyhow::{anyhow, Context};
+use std::collections::HashMap;
+use std::sync::{mpsc, Mutex};
+
+pub struct Waiters<T>(pub(self) Mutex<HashMap<String, mpsc::Sender<T>>>);
+
+impl<T> Default for Waiters<T> {
+    fn default() -> Self {
+        Waiters(Default::default())
+    }
+}
+
+impl<T> Waiters<T> {
+    pub fn register(&self, key: String) -> Waiter<T> {
+        let (tx, rx) = mpsc::channel();
+
+        // TODO: use `try_insert` (unstable)
+        let prev = self.0.lock().unwrap().insert(key.clone(), tx);
+        assert!(matches!(prev, None)); // assert_matches! is nightly-only
+
+        Waiter {
+            receiver: rx,
+            registry: self,
+            key,
+        }
+    }
+
+    pub fn notify(&self, key: &str, value: T) -> anyhow::Result<()>
+    where
+        T: Send + Sync + 'static,
+    {
+        let tx = self
+            .0
+            .lock()
+            .unwrap()
+            .remove(key)
+            .ok_or_else(|| anyhow!("key {} not found", key))?;
+        tx.send(value).context("channel hangup")
+    }
+}
+
+pub struct Waiter<'a, T> {
+    receiver: mpsc::Receiver<T>,
+    registry: &'a Waiters<T>,
+    key: String,
+}
+
+impl<T> Waiter<'_, T> {
+    pub fn wait(self) -> anyhow::Result<T> {
+        self.receiver.recv().context("channel hangup")
+    }
+}
+
+impl<T> Drop for Waiter<'_, T> {
+    fn drop(&mut self) {
+        self.registry.0.lock().unwrap().remove(&self.key);
+    }
+}

pytest.ini

@@ -0,0 +1,9 @@
+[pytest]
+addopts =
+    -m 'not remote_cluster'
+markers =
+    remote_cluster
+minversion = 6.0
+log_format = %(asctime)s.%(msecs)-3d %(levelname)s [%(filename)s:%(lineno)d] %(message)s
+log_date_format = %Y-%m-%d %H:%M:%S
+log_cli = true

scripts/coverage

@@ -0,0 +1,510 @@
+#!/usr/bin/env python3
+
+# Here'a good link in case you're interested in learning more
+# about current deficiencies of rust code coverage story:
+# https://github.com/rust-lang/rust/issues?q=is%3Aissue+is%3Aopen+instrument-coverage+label%3AA-code-coverage
+#
+# Also a couple of inspirational tools which I deliberately ended up not using:
+#  * https://github.com/mozilla/grcov
+#  * https://github.com/taiki-e/cargo-llvm-cov
+#  * https://github.com/llvm/llvm-project/tree/main/llvm/test/tools/llvm-cov
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from textwrap import dedent
+from typing import Any, Iterable, List, Optional
+
+import argparse
+import json
+import os
+import shutil
+import subprocess
+import sys
+
+
+def intersperse(sep: Any, iterable: Iterable[Any]):
+    fst = True
+    for item in iterable:
+        if not fst:
+            yield sep
+        fst = False
+        yield item
+
+
+def find_demangler(demangler=None):
+    known_tools = ['c++filt', 'rustfilt', 'llvm-cxxfilt']
+
+    if demangler:
+        # Explicit argument has precedence over `known_tools`
+        demanglers = [demangler]
+    else:
+        demanglers = known_tools
+
+    for demangler in demanglers:
+        if shutil.which(demangler):
+            return demangler
+
+    raise Exception(' '.join([
+        'Failed to find symbol demangler.',
+        'Please install it or provide another tool',
+        f"(e.g. {', '.join(known_tools)})",
+    ]))
+
+
+class Cargo:
+    def __init__(self, cwd: Path):
+        self.cwd = cwd
+        self.target_dir = Path(os.environ.get('CARGO_TARGET_DIR', cwd / 'target')).resolve()
+        self._rustlib_dir = None
+
+    @property
+    def rustlib_dir(self):
+        if not self._rustlib_dir:
+            cmd = [
+                'cargo',
+                '-Zunstable-options',
+                'rustc',
+                '--print=target-libdir',
+            ]
+            self._rustlib_dir = Path(subprocess.check_output(cmd, cwd=self.cwd, text=True)).parent
+
+        return self._rustlib_dir
+
+    def binaries(self, profile: str) -> List[str]:
+        executables = []
+
+        # This will emit json messages containing test binaries names
+        cmd = [
+            'cargo',
+            'test',
+            '--no-run',
+            '--message-format=json',
+        ]
+        env = dict(os.environ, PROFILE=profile)
+        output = subprocess.check_output(cmd, cwd=self.cwd, env=env, text=True)
+
+        for line in output.splitlines(keepends=False):
+            meta = json.loads(line)
+            exe = meta.get('executable')
+            if exe:
+                executables.append(exe)
+
+        # Metadata contains crate names, which can be used
+        # to recover names of executables, e.g. `pageserver`
+        cmd = [
+            'cargo',
+            'metadata',
+            '--format-version=1',
+            '--no-deps',
+        ]
+        meta = json.loads(subprocess.check_output(cmd, cwd=self.cwd))
+
+        for pkg in meta.get('packages', []):
+            for target in pkg.get('targets', []):
+                if 'bin' in target['kind']:
+                    exe = self.target_dir / profile / target['name']
+                    if exe.exists():
+                        executables.append(str(exe))
+
+        return executables
+
+
+@dataclass
+class LLVM:
+    cargo: Cargo
+
+    def resolve_tool(self, name: str) -> str:
+        exe = self.cargo.rustlib_dir / 'bin' / name
+        if exe.exists():
+            return str(exe)
+
+        if not shutil.which(name):
+            # Show a user-friendly warning
+            raise Exception(' '.join([
+                f"It appears that you don't have `{name}` installed.",
+                "Please execute `rustup component add llvm-tools-preview`,",
+                "or install it via your package manager of choice.",
+                "LLVM tools should be the same version as LLVM in `rustc --version --verbose`.",
+            ]))
+
+        return name
+
+    def profdata(self, input_dir: Path, output_profdata: Path):
+        profraws = [f for f in input_dir.iterdir() if f.suffix == '.profraw']
+        if not profraws:
+            raise Exception(f'No profraw files found at {input_dir}')
+
+        with open(input_dir / 'profraw.list', 'w') as input_files:
+            profraw_mtime = 0
+            for profraw in profraws:
+                profraw_mtime = max(profraw_mtime, profraw.stat().st_mtime_ns)
+                print(profraw, file=input_files)
+            input_files.flush()
+
+            try:
+                profdata_mtime = output_profdata.stat().st_mtime_ns
+            except FileNotFoundError:
+                profdata_mtime = 0
+
+            # An obvious make-ish optimization
+            if profraw_mtime >= profdata_mtime:
+                subprocess.check_call([
+                    self.resolve_tool('llvm-profdata'),
+                    'merge',
+                    '-sparse',
+                    f'-input-files={input_files.name}',
+                    f'-output={output_profdata}',
+                ])
+
+    def _cov(self,
+             *extras,
+             subcommand: str,
+             profdata: Path,
+             objects: List[str],
+             sources: List[str],
+             demangler: Optional[str] = None) -> None:
+
+        cwd = self.cargo.cwd
+        objects = list(intersperse('-object', objects))
+        extras = list(extras)
+
+        # For some reason `rustc` produces relative paths to src files,
+        # so we force it to cut the $PWD prefix.
+        # see: https://github.com/rust-lang/rust/issues/34701#issuecomment-739809584
+        if sources:
+            extras.append(f'-path-equivalence=.,{cwd.resolve()}')
+
+        if demangler:
+            extras.append(f'-Xdemangler={demangler}')
+
+        cmd = [
+            self.resolve_tool('llvm-cov'),
+            subcommand,  # '-dump-collected-paths',  # classified debug flag
+            '-instr-profile',
+            str(profdata),
+            *extras,
+            *objects,
+            *sources,
+        ]
+        subprocess.check_call(cmd, cwd=cwd)
+
+    def cov_report(self, **kwargs) -> None:
+        self._cov(subcommand='report', **kwargs)
+
+    def cov_export(self, *, kind: str, **kwargs) -> None:
+        extras = [f'-format={kind}']
+        self._cov(subcommand='export', *extras, **kwargs)
+
+    def cov_show(self, *, kind: str, output_dir: Optional[Path] = None, **kwargs) -> None:
+        extras = [f'-format={kind}']
+        if output_dir:
+            extras.append(f'-output-dir={output_dir}')
+
+        self._cov(subcommand='show', *extras, **kwargs)
+
+
+@dataclass
+class Report(ABC):
+    """ Common properties of a coverage report """
+
+    llvm: LLVM
+    demangler: str
+    profdata: Path
+    objects: List[str]
+    sources: List[str]
+
+    def _common_kwargs(self):
+        return dict(profdata=self.profdata,
+                    objects=self.objects,
+                    sources=self.sources,
+                    demangler=self.demangler)
+
+    @abstractmethod
+    def generate(self):
+        pass
+
+    def open(self):
+        # Do nothing by default
+        pass
+
+
+class SummaryReport(Report):
+    def generate(self):
+        self.llvm.cov_report(**self._common_kwargs())
+
+
+class TextReport(Report):
+    def generate(self):
+        self.llvm.cov_show(kind='text', **self._common_kwargs())
+
+
+class LcovReport(Report):
+    def generate(self):
+        self.llvm.cov_export(kind='lcov', **self._common_kwargs())
+
+
+@dataclass
+class HtmlReport(Report):
+    output_dir: Path
+
+    def generate(self):
+        self.llvm.cov_show(kind='html', output_dir=self.output_dir, **self._common_kwargs())
+        print(f'HTML report is located at `{self.output_dir}`')
+
+    def open(self):
+        tool = dict(linux='xdg-open', darwin='open').get(sys.platform)
+        if not tool:
+            raise Exception(f'Unknown platform {sys.platform}')
+
+        subprocess.check_call([tool, self.output_dir / 'index.html'],
+                              stdout=subprocess.DEVNULL,
+                              stderr=subprocess.DEVNULL)
+
+
+@dataclass
+class GithubPagesReport(HtmlReport):
+    output_dir: Path
+    commit_url: str
+
+    def generate(self):
+        def index_path(path):
+            return path / 'index.html'
+
+        common = self._common_kwargs()
+        # Provide default sources if there's none
+        common.setdefault('sources', ['.'])
+
+        self.llvm.cov_show(kind='html', output_dir=self.output_dir, **common)
+        shutil.copy(index_path(self.output_dir), self.output_dir / 'local.html')
+
+        with TemporaryDirectory() as tmp:
+            output_dir = Path(tmp)
+            args = dict(common, sources=[])
+            self.llvm.cov_show(kind='html', output_dir=output_dir, **args)
+            shutil.copy(index_path(output_dir), self.output_dir / 'all.html')
+
+        with open(index_path(self.output_dir), 'w') as index:
+            commit_sha = self.commit_url.rsplit('/', maxsplit=1)[-1][:10]
+
+            html = f"""
+                <!DOCTYPE html>
+                <html>
+                    <head>
+                        <title>Coverage ({commit_sha})</title>
+                    </head>
+                    <body>
+                        <h1>
+                            Coverage report for commit
+                                <a href="{self.commit_url}">
+                                    {commit_sha}
+                                </a>
+                        </h1>
+
+                        <p>
+                            <a href="./local.html">
+                                <b>Show only local sources</b>
+                            </a>
+                        </p>
+
+                        <p>
+                            <a href="./all.html">
+                                Show all sources (including dependencies)
+                            </a>
+                        </p>
+                    </body>
+                </html>
+            """
+            index.write(dedent(html))
+
+        print(f'HTML report is located at `{self.output_dir}`')
+
+
+class State:
+    def __init__(self, cwd: Path, top_dir: Optional[Path], profraw_prefix: Optional[str]):
+        # Use hostname by default
+        profraw_prefix = profraw_prefix or '%h'
+
+        self.cwd = cwd
+        self.cargo = Cargo(self.cwd)
+        self.llvm = LLVM(self.cargo)
+
+        self.top_dir = top_dir or self.cargo.target_dir / 'coverage'
+        self.report_dir = self.top_dir / 'report'
+
+        # Directory for raw coverage data emitted by executables
+        self.profraw_dir = self.top_dir / 'profraw'
+        self.profraw_dir.mkdir(parents=True, exist_ok=True)
+
+        # Aggregated coverage data
+        self.profdata_file = self.top_dir / 'coverage.profdata'
+
+        # Dump all coverage data files into a dedicated directory.
+        # Each filename is parameterized by PID & executable's signature.
+        os.environ['LLVM_PROFILE_FILE'] = str(self.profraw_dir /
+                                              f'cov-{profraw_prefix}-%p-%m.profraw')
+
+        os.environ['RUSTFLAGS'] = ' '.join([
+            os.environ.get('RUSTFLAGS', ''),
+            # Enable LLVM's source-based coverage
+            # see: https://clang.llvm.org/docs/SourceBasedCodeCoverage.html
+            # see: https://blog.rust-lang.org/inside-rust/2020/11/12/source-based-code-coverage.html
+            '-Zinstrument-coverage',
+            # Link every bit of code to prevent "holes" in coverage report
+            # see: https://doc.rust-lang.org/rustc/codegen-options/index.html#link-dead-code
+            '-Clink-dead-code',
+            # Some of the paths that `rustc` embeds into binaries are absolute, others are relative.
+            # The point is, we can't have both, because depending on `-path-equivalence`, `llvm-cov`
+            # either will cripple absolute paths or won't be able to show relative paths at all.
+            # There's no way to turn relative paths into absolute, so we strip $PWD prefix.
+            # Only source files of deps (e.g. `$HOME/.cargo`) will keep their absolute paths,
+            # but we won't include them in report by default (but see `--all`).
+            f'--remap-path-prefix {self.cwd}=',
+        ])
+
+        # XXX: God, have mercy on our souls...
+        # see: https://github.com/rust-lang/rust/pull/90132
+        os.environ['RUSTC_BOOTSTRAP'] = '1'
+
+    def do_run(self, args):
+        subprocess.check_call([*args.command, *args.args])
+
+    def do_report(self, args):
+        if args.all and args.sources:
+            raise Exception('--all should not be used with sources')
+
+        # see man for `llvm-cov show [sources]`
+        if args.all:
+            sources = []
+        elif not args.sources:
+            sources = ['.']
+        else:
+            sources = args.sources
+
+        print('* Merging profraw files')
+        self.llvm.profdata(self.profraw_dir, self.profdata_file)
+
+        objects = []
+        if args.input_objects:
+            print('* Collecting object files using --input-objects')
+            with open(args.input_objects) as f:
+                objects.extend(f.read().splitlines(keepends=False))
+        if args.cargo_objects == 'true' or (args.cargo_objects == 'auto'
+                                            and not args.input_objects):
+            print('* Collecting object files using cargo')
+            objects.extend(self.cargo.binaries(args.profile))
+
+        params = dict(llvm=self.llvm,
+                      demangler=find_demangler(args.demangler),
+                      profdata=self.profdata_file,
+                      objects=objects,
+                      sources=sources)
+
+        formats = {
+            'html':
+            lambda: HtmlReport(**params, output_dir=self.report_dir),
+            'text':
+            lambda: TextReport(**params),
+            'lcov':
+            lambda: LcovReport(**params),
+            'summary':
+            lambda: SummaryReport(**params),
+            'github':
+            lambda: GithubPagesReport(
+                **params, output_dir=self.report_dir, commit_url=args.commit_url),
+        }
+
+        report = formats.get(args.format)()
+        if not report:
+            raise Exception('Format `{args.format}` is not supported')
+
+        print(f'* Rendering coverage report ({args.format})')
+        report.generate()
+
+        if args.open:
+            print('* Opening the report')
+            report.open()
+
+    def do_clean(self, args):
+        # Wipe everything if no filters have been provided
+        if not (args.report or args.prof):
+            shutil.rmtree(self.top_dir, ignore_errors=True)
+        else:
+            if args.report:
+                shutil.rmtree(self.report_dir, ignore_errors=True)
+            if args.prof:
+                self.profdata_file.unlink(missing_ok=True)
+
+
+def main():
+    app = sys.argv[0]
+    example = f"""
+prerequisites:
+    # alternatively, install a system package for `llvm-tools`
+    rustup component add llvm-tools-preview
+
+self-contained example:
+    {app} run make
+    {app} run pipenv run pytest test_runner
+    {app} run cargo test
+    {app} report --open
+    """
+
+    parser = argparse.ArgumentParser(description='Coverage report builder',
+                                     formatter_class=argparse.RawDescriptionHelpFormatter,
+                                     epilog=example)
+    parser.add_argument('--dir', type=Path, help='output directory')
+    parser.add_argument('--profraw-prefix', metavar='STRING', type=str)
+
+    commands = parser.add_subparsers(title='commands', dest='subparser_name')
+
+    p_run = commands.add_parser('run', help='run a command with magic env')
+    p_run.add_argument('command', nargs=1)
+    p_run.add_argument('args', nargs=argparse.REMAINDER)
+
+    p_report = commands.add_parser('report', help='generate a coverage report')
+    p_report.add_argument('--profile',
+                          default='debug',
+                          choices=('debug', 'release'),
+                          help='cargo build profile')
+    p_report.add_argument('--format',
+                          default='html',
+                          choices=('html', 'text', 'summary', 'lcov', 'github'),
+                          help='report format')
+    p_report.add_argument('--input-objects',
+                          metavar='FILE',
+                          type=Path,
+                          help='file containing list of binaries')
+    p_report.add_argument('--cargo-objects',
+                          default='auto',
+                          choices=('auto', 'true', 'false'),
+                          help='use cargo for auto discovery of binaries')
+    p_report.add_argument('--commit-url', type=str, help='required for --format=github')
+    p_report.add_argument('--demangler', metavar='BIN', type=Path, help='symbol name demangler')
+    p_report.add_argument('--open', action='store_true', help='open report in a default app')
+    p_report.add_argument('--all', action='store_true', help='show everything, e.g. deps')
+    p_report.add_argument('sources', nargs='*', type=Path, help='source file or directory')
+
+    p_clean = commands.add_parser('clean', help='wipe coverage artifacts')
+    p_clean.add_argument('--report', action='store_true', help='pick generated report')
+    p_clean.add_argument('--prof', action='store_true', help='pick *.profdata & *.profraw')
+
+    args = parser.parse_args()
+    state = State(cwd=Path.cwd(), top_dir=args.dir, profraw_prefix=args.profraw_prefix)
+
+    commands = {
+        'run': state.do_run,
+        'report': state.do_report,
+        'clean': state.do_clean,
+    }
+
+    action = commands.get(args.subparser_name)
+    if action:
+        action(args)
+    else:
+        parser.print_help()
+
+
+if __name__ == '__main__':
+    main()

scripts/generate_and_push_perf_report.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+# this is a shortcut script to avoid duplication in CI
+
+set -eux -o pipefail
+
+SCRIPT_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
+
+git clone https://$VIP_VAP_ACCESS_TOKEN@github.com/zenithdb/zenith-perf-data.git
+cd zenith-perf-data
+mkdir -p reports/
+mkdir -p data/$REPORT_TO
+
+cp $REPORT_FROM/* data/$REPORT_TO
+
+echo "Generating report"
+pipenv run python $SCRIPT_DIR/generate_perf_report_page.py --input-dir data/$REPORT_TO --out reports/$REPORT_TO.html 
+echo "Uploading perf result"
+git add data reports
+git \
+    -c "user.name=vipvap" \
+    -c "user.email=vipvap@zenith.tech" \
+    commit \
+    --author="vipvap <vipvap@zenith.tech>" \
+    -m "add performance test result for $GITHUB_SHA zenith revision"
+
+git push https://$VIP_VAP_ACCESS_TOKEN@github.com/zenithdb/zenith-perf-data.git master

scripts/generate_perf_report_page.py

@@ -0,0 +1,207 @@
+#!/usr/bin/env python3
+import argparse
+from dataclasses import dataclass
+from pathlib import Path
+import json
+from typing import Any, Dict, List, Optional, Tuple, cast
+from jinja2 import Template
+
+# skip 'input' columns. They are included in the header and just blow the table
+EXCLUDE_COLUMNS = frozenset({
+    'scale',
+    'duration',
+    'number_of_clients',
+    'number_of_threads',
+    'init_start_timestamp',
+    'init_end_timestamp',
+    'run_start_timestamp',
+    'run_end_timestamp',
+})
+
+KEY_EXCLUDE_FIELDS = frozenset({
+    'init_start_timestamp',
+    'init_end_timestamp',
+    'run_start_timestamp',
+    'run_end_timestamp',
+})
+NEGATIVE_COLOR = 'negative'
+POSITIVE_COLOR = 'positive'
+
+
+@dataclass
+class SuitRun:
+    revision: str
+    values: Dict[str, Any]
+
+
+@dataclass
+class SuitRuns:
+    platform: str
+    suit: str
+    common_columns: List[Tuple[str, str]]
+    value_columns: List[str]
+    runs: List[SuitRun]
+
+
+@dataclass
+class RowValue:
+    value: str
+    color: str
+    ratio: str
+
+
+def get_columns(values: List[Dict[Any, Any]]) -> Tuple[List[Tuple[str, str]], List[str]]:
+    value_columns = []
+    common_columns = []
+    for item in values:
+        if item['name'] in KEY_EXCLUDE_FIELDS:
+            continue
+        if item['report'] != 'test_param':
+            value_columns.append(cast(str, item['name']))
+        else:
+            common_columns.append((cast(str, item['name']), cast(str, item['value'])))
+    value_columns.sort()
+    common_columns.sort(key=lambda x: x[0])  # sort by name
+    return common_columns, value_columns
+
+
+def format_ratio(ratio: float, report: str) -> Tuple[str, str]:
+    color = ''
+    sign = '+' if ratio > 0 else ''
+    if abs(ratio) < 0.05:
+        return f'&nbsp({sign}{ratio:.2f})', color
+
+    if report not in {'test_param', 'higher_is_better', 'lower_is_better'}:
+        raise ValueError(f'Unknown report type: {report}')
+
+    if report == 'test_param':
+        return f'{ratio:.2f}', color
+
+    if ratio > 0:
+        if report == 'higher_is_better':
+            color = POSITIVE_COLOR
+        elif report == 'lower_is_better':
+            color = NEGATIVE_COLOR
+    elif ratio < 0:
+        if report == 'higher_is_better':
+            color = NEGATIVE_COLOR
+        elif report == 'lower_is_better':
+            color = POSITIVE_COLOR
+
+    return f'&nbsp({sign}{ratio:.2f})', color
+
+
+def extract_value(name: str, suit_run: SuitRun) -> Optional[Dict[str, Any]]:
+    for item in suit_run.values['data']:
+        if item['name'] == name:
+            return cast(Dict[str, Any], item)
+    return None
+
+
+def get_row_values(columns: List[str], run_result: SuitRun,
+                   prev_result: Optional[SuitRun]) -> List[RowValue]:
+    row_values = []
+    for column in columns:
+        current_value = extract_value(column, run_result)
+        if current_value is None:
+            # should never happen
+            raise ValueError(f'{column} not found in {run_result.values}')
+
+        value = current_value["value"]
+        if isinstance(value, float):
+            value = f'{value:.2f}'
+
+        if prev_result is None:
+            row_values.append(RowValue(value, '', ''))
+            continue
+
+        prev_value = extract_value(column, prev_result)
+        if prev_value is None:
+            # this might happen when new metric is added and there is no value for it in previous run
+            # let this be here, TODO add proper handling when this actually happens
+            raise ValueError(f'{column} not found in previous result')
+        ratio = float(value) / float(prev_value['value']) - 1
+        ratio_display, color = format_ratio(ratio, current_value['report'])
+        row_values.append(RowValue(value, color, ratio_display))
+    return row_values
+
+
+@dataclass
+class SuiteRunTableRow:
+    revision: str
+    values: List[RowValue]
+
+
+def prepare_rows_from_runs(value_columns: List[str], runs: List[SuitRun]) -> List[SuiteRunTableRow]:
+    rows = []
+    prev_run = None
+    for run in runs:
+        rows.append(
+            SuiteRunTableRow(revision=run.revision,
+                             values=get_row_values(value_columns, run, prev_run)))
+        prev_run = run
+
+    return rows
+
+
+def main(args: argparse.Namespace) -> None:
+    input_dir = Path(args.input_dir)
+    grouped_runs: Dict[str, SuitRuns] = {}
+    # we have files in form: <ctr>_<rev>.json
+    # fill them in the hashmap so we have grouped items for the
+    # same run configuration (scale, duration etc.) ordered by counter.
+    for item in sorted(input_dir.iterdir(), key=lambda x: int(x.name.split('_')[0])):
+        run_data = json.loads(item.read_text())
+        revision = run_data['revision']
+
+        for suit_result in run_data['result']:
+            key = "{}{}".format(run_data['platform'], suit_result['suit'])
+            # pack total duration as a synthetic value
+            total_duration = suit_result['total_duration']
+            suit_result['data'].append({
+                'name': 'total_duration',
+                'value': total_duration,
+                'unit': 's',
+                'report': 'lower_is_better',
+            })
+            common_columns, value_columns = get_columns(suit_result['data'])
+
+            grouped_runs.setdefault(
+                key,
+                SuitRuns(
+                    platform=run_data['platform'],
+                    suit=suit_result['suit'],
+                    common_columns=common_columns,
+                    value_columns=value_columns,
+                    runs=[],
+                ),
+            )
+
+            grouped_runs[key].runs.append(SuitRun(revision=revision, values=suit_result))
+    context = {}
+    for result in grouped_runs.values():
+        suit = result.suit
+        context[suit] = {
+            'common_columns': result.common_columns,
+            'value_columns': result.value_columns,
+            'platform': result.platform,
+            # reverse the order so newest results are on top of the table
+            'rows': reversed(prepare_rows_from_runs(result.value_columns, result.runs)),
+        }
+
+    template = Template((Path(__file__).parent / 'perf_report_template.html').read_text())
+
+    Path(args.out).write_text(template.render(context=context))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--input-dir',
+        dest='input_dir',
+        required=True,
+        help='Directory with jsons generated by the test suite',
+    )
+    parser.add_argument('--out', required=True, help='Output html file path')
+    args = parser.parse_args()
+    main(args)

scripts/git-upload

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+
+from contextlib import contextmanager
+from tempfile import TemporaryDirectory
+from pathlib import Path
+
+import argparse
+import os
+import shutil
+import subprocess
+import sys
+
+
+def absolute_path(path):
+    return Path(path).resolve()
+
+
+def relative_path(path):
+    path = Path(path)
+    if path.is_absolute():
+        raise Exception(f'path `{path}` must be relative!')
+    return path
+
+
+@contextmanager
+def chdir(cwd: Path):
+    old = os.getcwd()
+    os.chdir(cwd)
+    try:
+        yield cwd
+    finally:
+        os.chdir(old)
+
+
+def run(cmd, *args, **kwargs):
+    print('$', ' '.join(cmd))
+    subprocess.check_call(cmd, *args, **kwargs)
+
+
+class GitRepo:
+    def __init__(self, url):
+        self.url = url
+        self.cwd = TemporaryDirectory()
+
+        subprocess.check_call([
+            'git',
+            'clone',
+            str(url),
+            self.cwd.name,
+        ])
+
+    def is_dirty(self):
+        res = subprocess.check_output(['git', 'status', '--porcelain'], text=True).strip()
+        return bool(res)
+
+    def update(self, message, action, branch=None):
+        with chdir(self.cwd.name):
+            if not branch:
+                cmd = ['git', 'branch', '--show-current']
+                branch = subprocess.check_output(cmd, text=True).strip()
+
+            # Run action in repo's directory
+            action()
+
+            run(['git', 'add', '.'])
+
+            if not self.is_dirty():
+                print('No changes detected, quitting')
+                return
+
+            run([
+                'git',
+                '-c',
+                'user.name=vipvap',
+                '-c',
+                'user.email=vipvap@zenith.tech',
+                'commit',
+                '--author="vipvap <vipvap@zenith.tech>"',
+                f'--message={message}',
+            ])
+
+            for _ in range(5):
+                try:
+                    run(['git', 'fetch', 'origin', branch])
+                    run(['git', 'rebase', f'origin/{branch}'])
+                    run(['git', 'push', 'origin', branch])
+                    return
+
+                except subprocess.CalledProcessError as e:
+                    print(f'failed to update branch `{branch}`: {e}', file=sys.stderr)
+
+            raise Exception(f'failed to update branch `{branch}`')
+
+
+def do_copy(args):
+    src = args.src
+    dst = args.dst
+
+    try:
+        if src.is_dir():
+            shutil.copytree(src, dst)
+        else:
+            shutil.copy(src, dst)
+    except FileExistsError:
+        if args.forbid_overwrite:
+            raise
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Git upload tool')
+    parser.add_argument('--repo', type=str, metavar='URL', required=True, help='git repo url')
+    parser.add_argument('--message', type=str, metavar='TEXT', help='commit message')
+
+    commands = parser.add_subparsers(title='commands', dest='subparser_name')
+
+    p_copy = commands.add_parser('copy', help='copy file into the repo')
+    p_copy.add_argument('src', type=absolute_path, help='source path')
+    p_copy.add_argument('dst', type=relative_path, help='relative dest path')
+    p_copy.add_argument('--forbid-overwrite', action='store_true', help='do not allow overwrites')
+
+    args = parser.parse_args()
+
+    commands = {
+        'copy': do_copy,
+    }
+
+    action = commands.get(args.subparser_name)
+    if action:
+        message = args.message or 'update'
+        GitRepo(args.repo).update(message, lambda: action(args))
+    else:
+        parser.print_usage()
+
+
+if __name__ == '__main__':
+    main()

scripts/perf_report_template.html

@@ -0,0 +1,52 @@
+<!DOCTYPE html>
+<html>
+
+<body>
+    <style>
+        table,
+        th,
+        td {
+            border: 1px solid black;
+            border-collapse: collapse;
+        }
+
+        .positive {
+            background-color: rgba(0, 255, 0, 0.8)
+        }
+
+        .negative {
+            background-color: rgba(255, 0, 0, 0.65)
+        }
+    </style>
+
+    <h2>Zenith Performance Tests</h2>
+
+    {% for suit_name, suit_data in context.items() %}
+    <h3>Runs for {{ suit_name }} </h3>
+    <b>platform:</b> {{ suit_data.platform }}<br>
+    {% for common_column_name, common_column_value in suit_data.common_columns %}
+    <b>{{ common_column_name }}</b>: {{ common_column_value }}<br>
+    {% endfor %}
+    <br>
+
+    <table>
+        <tr>
+            <th>revision</th>
+            {% for column_name in suit_data.value_columns %}
+            <th>{{ column_name }}</th>
+            {% endfor %}
+        </tr>
+        {% for row in suit_data.rows %}
+        <tr>
+            <td><a href=https://github.com/zenithdb/zenith/commit/{{ row.revision }}>{{ row.revision[:6] }}</a></td>
+            {% for column_value in row.values %}
+            <td class="{{ column_value.color }}">{{ column_value.value }}{{column_value.ratio}}</td>
+            {% endfor %}
+        </tr>
+        {% endfor %}
+    </table>
+    {% endfor %}
+
+</body>
+
+</html>

setup.cfg

@@ -10,8 +10,11 @@ max-line-length = 100
 [yapf]
 based_on_style = pep8
 column_limit = 100
+split_all_top_level_comma_separated_values = true
 
 [mypy]
+# mypy uses regex
+exclude = ^vendor/
 # some tests don't typecheck when this flag is set
 check_untyped_defs = false
 
@@ -21,7 +24,11 @@ disallow_untyped_decorators = false
 disallow_untyped_defs = false
 strict = true
 
-[mypy-psycopg2.*]
+[mypy-asyncpg.*]
+# There is some work in progress, though: https://github.com/MagicStack/asyncpg/pull/577
+ignore_missing_imports = true
+
+[mypy-cached_property.*]
 ignore_missing_imports = true
 
 [mypy-pytest.*]

test_runner/Pipfile

@@ -1,22 +0,0 @@
-[[source]]
-url = "https://pypi.python.org/simple"
-verify_ssl = true
-name = "pypi"
-
-[packages]
-pytest = ">=6.0.0"
-psycopg2 = "*"
-typing-extensions = "*"
-pyjwt = {extras = ["crypto"], version = "*"}
-requests = "*"
-pytest-xdist = "*"
-asyncpg = "*"
-
-[dev-packages]
-yapf = "*"
-flake8 = "*"
-mypy = "*"
-
-[requires]
-# we need at least 3.6, but pipenv doesn't allow to say this directly
-python_version = "3"

test_runner/Pipfile.lock

@@ -1,376 +0,0 @@
-{
-    "_meta": {
-        "hash": {
-            "sha256": "3cdc048691824d0b93912b6b78a0aa01dc98f278212c1badb0cc2edbd2103c3a"
-        },
-        "pipfile-spec": 6,
-        "requires": {
-            "python_version": "3"
-        },
-        "sources": [
-            {
-                "name": "pypi",
-                "url": "https://pypi.python.org/simple",
-                "verify_ssl": true
-            }
-        ]
-    },
-    "default": {
-        "asyncpg": {
-            "hashes": [
-                "sha256:129d501f3d30616afd51eb8d3142ef51ba05374256bd5834cec3ef4956a9b317",
-                "sha256:29ef6ae0a617fc13cc2ac5dc8e9b367bb83cba220614b437af9b67766f4b6b20",
-                "sha256:41704c561d354bef01353835a7846e5606faabbeb846214dfcf666cf53319f18",
-                "sha256:556b0e92e2b75dc028b3c4bc9bd5162ddf0053b856437cf1f04c97f9c6837d03",
-                "sha256:8ff5073d4b654e34bd5eaadc01dc4d68b8a9609084d835acd364cd934190a08d",
-                "sha256:a458fc69051fbb67d995fdda46d75a012b5d6200f91e17d23d4751482640ed4c",
-                "sha256:a7095890c96ba36f9f668eb552bb020dddb44f8e73e932f8573efc613ee83843",
-                "sha256:a738f4807c853623d3f93f0fea11f61be6b0e5ca16ea8aeb42c2c7ee742aa853",
-                "sha256:c4fc0205fe4ddd5aeb3dfdc0f7bafd43411181e1f5650189608e5971cceacff1",
-                "sha256:dd2fa063c3344823487d9ddccb40802f02622ddf8bf8a6cc53885ee7a2c1c0c6",
-                "sha256:ddffcb85227bf39cd1bedd4603e0082b243cf3b14ced64dce506a15b05232b83",
-                "sha256:e36c6806883786b19551bb70a4882561f31135dc8105a59662e0376cf5b2cbc5",
-                "sha256:eed43abc6ccf1dc02e0d0efc06ce46a411362f3358847c6b0ec9a43426f91ece"
-            ],
-            "index": "pypi",
-            "version": "==0.24.0"
-        },
-        "attrs": {
-            "hashes": [
-                "sha256:149e90d6d8ac20db7a955ad60cf0e6881a3f20d37096140088356da6c716b0b1",
-                "sha256:ef6aaac3ca6cd92904cdd0d83f629a15f18053ec84e6432106f7a4d04ae4f5fb"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
-            "version": "==21.2.0"
-        },
-        "certifi": {
-            "hashes": [
-                "sha256:2bbf76fd432960138b3ef6dda3dde0544f27cbf8546c458e60baf371917ba9ee",
-                "sha256:50b1e4f8446b06f41be7dd6338db18e0990601dce795c2b1686458aa7e8fa7d8"
-            ],
-            "version": "==2021.5.30"
-        },
-        "cffi": {
-            "hashes": [
-                "sha256:06c54a68935738d206570b20da5ef2b6b6d92b38ef3ec45c5422c0ebaf338d4d",
-                "sha256:0c0591bee64e438883b0c92a7bed78f6290d40bf02e54c5bf0978eaf36061771",
-                "sha256:19ca0dbdeda3b2615421d54bef8985f72af6e0c47082a8d26122adac81a95872",
-                "sha256:22b9c3c320171c108e903d61a3723b51e37aaa8c81255b5e7ce102775bd01e2c",
-                "sha256:26bb2549b72708c833f5abe62b756176022a7b9a7f689b571e74c8478ead51dc",
-                "sha256:33791e8a2dc2953f28b8d8d300dde42dd929ac28f974c4b4c6272cb2955cb762",
-                "sha256:3c8d896becff2fa653dc4438b54a5a25a971d1f4110b32bd3068db3722c80202",
-                "sha256:4373612d59c404baeb7cbd788a18b2b2a8331abcc84c3ba40051fcd18b17a4d5",
-                "sha256:487d63e1454627c8e47dd230025780e91869cfba4c753a74fda196a1f6ad6548",
-                "sha256:48916e459c54c4a70e52745639f1db524542140433599e13911b2f329834276a",
-                "sha256:4922cd707b25e623b902c86188aca466d3620892db76c0bdd7b99a3d5e61d35f",
-                "sha256:55af55e32ae468e9946f741a5d51f9896da6b9bf0bbdd326843fec05c730eb20",
-                "sha256:57e555a9feb4a8460415f1aac331a2dc833b1115284f7ded7278b54afc5bd218",
-                "sha256:5d4b68e216fc65e9fe4f524c177b54964af043dde734807586cf5435af84045c",
-                "sha256:64fda793737bc4037521d4899be780534b9aea552eb673b9833b01f945904c2e",
-                "sha256:6d6169cb3c6c2ad50db5b868db6491a790300ade1ed5d1da29289d73bbe40b56",
-                "sha256:7bcac9a2b4fdbed2c16fa5681356d7121ecabf041f18d97ed5b8e0dd38a80224",
-                "sha256:80b06212075346b5546b0417b9f2bf467fea3bfe7352f781ffc05a8ab24ba14a",
-                "sha256:818014c754cd3dba7229c0f5884396264d51ffb87ec86e927ef0be140bfdb0d2",
-                "sha256:8eb687582ed7cd8c4bdbff3df6c0da443eb89c3c72e6e5dcdd9c81729712791a",
-                "sha256:99f27fefe34c37ba9875f224a8f36e31d744d8083e00f520f133cab79ad5e819",
-                "sha256:9f3e33c28cd39d1b655ed1ba7247133b6f7fc16fa16887b120c0c670e35ce346",
-                "sha256:a8661b2ce9694ca01c529bfa204dbb144b275a31685a075ce123f12331be790b",
-                "sha256:a9da7010cec5a12193d1af9872a00888f396aba3dc79186604a09ea3ee7c029e",
-                "sha256:aedb15f0a5a5949ecb129a82b72b19df97bbbca024081ed2ef88bd5c0a610534",
-                "sha256:b315d709717a99f4b27b59b021e6207c64620790ca3e0bde636a6c7f14618abb",
-                "sha256:ba6f2b3f452e150945d58f4badd92310449876c4c954836cfb1803bdd7b422f0",
-                "sha256:c33d18eb6e6bc36f09d793c0dc58b0211fccc6ae5149b808da4a62660678b156",
-                "sha256:c9a875ce9d7fe32887784274dd533c57909b7b1dcadcc128a2ac21331a9765dd",
-                "sha256:c9e005e9bd57bc987764c32a1bee4364c44fdc11a3cc20a40b93b444984f2b87",
-                "sha256:d2ad4d668a5c0645d281dcd17aff2be3212bc109b33814bbb15c4939f44181cc",
-                "sha256:d950695ae4381ecd856bcaf2b1e866720e4ab9a1498cba61c602e56630ca7195",
-                "sha256:e22dcb48709fc51a7b58a927391b23ab37eb3737a98ac4338e2448bef8559b33",
-                "sha256:e8c6a99be100371dbb046880e7a282152aa5d6127ae01783e37662ef73850d8f",
-                "sha256:e9dc245e3ac69c92ee4c167fbdd7428ec1956d4e754223124991ef29eb57a09d",
-                "sha256:eb687a11f0a7a1839719edd80f41e459cc5366857ecbed383ff376c4e3cc6afd",
-                "sha256:eb9e2a346c5238a30a746893f23a9535e700f8192a68c07c0258e7ece6ff3728",
-                "sha256:ed38b924ce794e505647f7c331b22a693bee1538fdf46b0222c4717b42f744e7",
-                "sha256:f0010c6f9d1a4011e429109fda55a225921e3206e7f62a0c22a35344bfd13cca",
-                "sha256:f0c5d1acbfca6ebdd6b1e3eded8d261affb6ddcf2186205518f1428b8569bb99",
-                "sha256:f10afb1004f102c7868ebfe91c28f4a712227fe4cb24974350ace1f90e1febbf",
-                "sha256:f174135f5609428cc6e1b9090f9268f5c8935fddb1b25ccb8255a2d50de6789e",
-                "sha256:f3ebe6e73c319340830a9b2825d32eb6d8475c1dac020b4f0aa774ee3b898d1c",
-                "sha256:f627688813d0a4140153ff532537fbe4afea5a3dffce1f9deb7f91f848a832b5",
-                "sha256:fd4305f86f53dfd8cd3522269ed7fc34856a8ee3709a5e28b2836b2db9d4cd69"
-            ],
-            "version": "==1.14.6"
-        },
-        "charset-normalizer": {
-            "hashes": [
-                "sha256:5d209c0a931f215cee683b6445e2d77677e7e75e159f78def0db09d68fafcaa6",
-                "sha256:5ec46d183433dcbd0ab716f2d7f29d8dee50505b3fdb40c6b985c7c4f5a3591f"
-            ],
-            "markers": "python_version >= '3'",
-            "version": "==2.0.6"
-        },
-        "cryptography": {
-            "hashes": [
-                "sha256:0a7dcbcd3f1913f664aca35d47c1331fce738d44ec34b7be8b9d332151b0b01e",
-                "sha256:1eb7bb0df6f6f583dd8e054689def236255161ebbcf62b226454ab9ec663746b",
-                "sha256:21ca464b3a4b8d8e86ba0ee5045e103a1fcfac3b39319727bc0fc58c09c6aff7",
-                "sha256:34dae04a0dce5730d8eb7894eab617d8a70d0c97da76b905de9efb7128ad7085",
-                "sha256:3520667fda779eb788ea00080124875be18f2d8f0848ec00733c0ec3bb8219fc",
-                "sha256:3c4129fc3fdc0fa8e40861b5ac0c673315b3c902bbdc05fc176764815b43dd1d",
-                "sha256:3fa3a7ccf96e826affdf1a0a9432be74dc73423125c8f96a909e3835a5ef194a",
-                "sha256:5b0fbfae7ff7febdb74b574055c7466da334a5371f253732d7e2e7525d570498",
-                "sha256:695104a9223a7239d155d7627ad912953b540929ef97ae0c34c7b8bf30857e89",
-                "sha256:8695456444f277af73a4877db9fc979849cd3ee74c198d04fc0776ebc3db52b9",
-                "sha256:94cc5ed4ceaefcbe5bf38c8fba6a21fc1d365bb8fb826ea1688e3370b2e24a1c",
-                "sha256:94fff993ee9bc1b2440d3b7243d488c6a3d9724cc2b09cdb297f6a886d040ef7",
-                "sha256:9965c46c674ba8cc572bc09a03f4c649292ee73e1b683adb1ce81e82e9a6a0fb",
-                "sha256:a00cf305f07b26c351d8d4e1af84ad7501eca8a342dedf24a7acb0e7b7406e14",
-                "sha256:a305600e7a6b7b855cd798e00278161b681ad6e9b7eca94c721d5f588ab212af",
-                "sha256:cd65b60cfe004790c795cc35f272e41a3df4631e2fb6b35aa7ac6ef2859d554e",
-                "sha256:d2a6e5ef66503da51d2110edf6c403dc6b494cc0082f85db12f54e9c5d4c3ec5",
-                "sha256:d9ec0e67a14f9d1d48dd87a2531009a9b251c02ea42851c060b25c782516ff06",
-                "sha256:f44d141b8c4ea5eb4dbc9b3ad992d45580c1d22bf5e24363f2fbf50c2d7ae8a7"
-            ],
-            "version": "==3.4.8"
-        },
-        "execnet": {
-            "hashes": [
-                "sha256:8f694f3ba9cc92cab508b152dcfe322153975c29bda272e2fd7f3f00f36e47c5",
-                "sha256:a295f7cc774947aac58dde7fdc85f4aa00c42adf5d8f5468fc630c1acf30a142"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
-            "version": "==1.9.0"
-        },
-        "idna": {
-            "hashes": [
-                "sha256:14475042e284991034cb48e06f6851428fb14c4dc953acd9be9a5e95c7b6dd7a",
-                "sha256:467fbad99067910785144ce333826c71fb0e63a425657295239737f7ecd125f3"
-            ],
-            "markers": "python_version >= '3'",
-            "version": "==3.2"
-        },
-        "iniconfig": {
-            "hashes": [
-                "sha256:011e24c64b7f47f6ebd835bb12a743f2fbe9a26d4cecaa7f53bc4f35ee9da8b3",
-                "sha256:bc3af051d7d14b2ee5ef9969666def0cd1a000e121eaea580d4a313df4b37f32"
-            ],
-            "version": "==1.1.1"
-        },
-        "packaging": {
-            "hashes": [
-                "sha256:7dc96269f53a4ccec5c0670940a4281106dd0bb343f47b7471f779df49c2fbe7",
-                "sha256:c86254f9220d55e31cc94d69bade760f0847da8000def4dfe1c6b872fd14ff14"
-            ],
-            "markers": "python_version >= '3.6'",
-            "version": "==21.0"
-        },
-        "pluggy": {
-            "hashes": [
-                "sha256:4224373bacce55f955a878bf9cfa763c1e360858e330072059e10bad68531159",
-                "sha256:74134bbf457f031a36d68416e1509f34bd5ccc019f0bcc952c7b909d06b37bd3"
-            ],
-            "markers": "python_version >= '3.6'",
-            "version": "==1.0.0"
-        },
-        "psycopg2": {
-            "hashes": [
-                "sha256:079d97fc22de90da1d370c90583659a9f9a6ee4007355f5825e5f1c70dffc1fa",
-                "sha256:2087013c159a73e09713294a44d0c8008204d06326006b7f652bef5ace66eebb",
-                "sha256:2c992196719fadda59f72d44603ee1a2fdcc67de097eea38d41c7ad9ad246e62",
-                "sha256:7640e1e4d72444ef012e275e7b53204d7fab341fb22bc76057ede22fe6860b25",
-                "sha256:7f91312f065df517187134cce8e395ab37f5b601a42446bdc0f0d51773621854",
-                "sha256:830c8e8dddab6b6716a4bf73a09910c7954a92f40cf1d1e702fb93c8a919cc56",
-                "sha256:89409d369f4882c47f7ea20c42c5046879ce22c1e4ea20ef3b00a4dfc0a7f188",
-                "sha256:bf35a25f1aaa8a3781195595577fcbb59934856ee46b4f252f56ad12b8043bcf",
-                "sha256:de5303a6f1d0a7a34b9d40e4d3bef684ccc44a49bbe3eb85e3c0bffb4a131b7c"
-            ],
-            "index": "pypi",
-            "version": "==2.9.1"
-        },
-        "py": {
-            "hashes": [
-                "sha256:21b81bda15b66ef5e1a777a21c4dcd9c20ad3efd0b3f817e7a809035269e1bd3",
-                "sha256:3b80836aa6d1feeaa108e046da6423ab8f6ceda6468545ae8d02d9d58d18818a"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==1.10.0"
-        },
-        "pycparser": {
-            "hashes": [
-                "sha256:2d475327684562c3a96cc71adf7dc8c4f0565175cf86b6d7a404ff4c771f15f0",
-                "sha256:7582ad22678f0fcd81102833f60ef8d0e57288b6b5fb00323d101be910e35705"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==2.20"
-        },
-        "pyjwt": {
-            "extras": [
-                "crypto"
-            ],
-            "hashes": [
-                "sha256:934d73fbba91b0483d3857d1aff50e96b2a892384ee2c17417ed3203f173fca1",
-                "sha256:fba44e7898bbca160a2b2b501f492824fc8382485d3a6f11ba5d0c1937ce6130"
-            ],
-            "index": "pypi",
-            "version": "==2.1.0"
-        },
-        "pyparsing": {
-            "hashes": [
-                "sha256:c203ec8783bf771a155b207279b9bccb8dea02d8f0c9e5f8ead507bc3246ecc1",
-                "sha256:ef9d7589ef3c200abe66653d3f1ab1033c3c419ae9b9bdb1240a85b024efc88b"
-            ],
-            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==2.4.7"
-        },
-        "pytest": {
-            "hashes": [
-                "sha256:131b36680866a76e6781d13f101efb86cf674ebb9762eb70d3082b6f29889e89",
-                "sha256:7310f8d27bc79ced999e760ca304d69f6ba6c6649c0b60fb0e04a4a77cacc134"
-            ],
-            "index": "pypi",
-            "version": "==6.2.5"
-        },
-        "pytest-forked": {
-            "hashes": [
-                "sha256:6aa9ac7e00ad1a539c41bec6d21011332de671e938c7637378ec9710204e37ca",
-                "sha256:dc4147784048e70ef5d437951728825a131b81714b398d5d52f17c7c144d8815"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4'",
-            "version": "==1.3.0"
-        },
-        "pytest-xdist": {
-            "hashes": [
-                "sha256:7b61ebb46997a0820a263553179d6d1e25a8c50d8a8620cd1aa1e20e3be99168",
-                "sha256:89b330316f7fc475f999c81b577c2b926c9569f3d397ae432c0c2e2496d61ff9"
-            ],
-            "index": "pypi",
-            "version": "==2.4.0"
-        },
-        "requests": {
-            "hashes": [
-                "sha256:6c1246513ecd5ecd4528a0906f910e8f0f9c6b8ec72030dc9fd154dc1a6efd24",
-                "sha256:b8aa58f8cf793ffd8782d3d8cb19e66ef36f7aba4353eec859e74678b01b07a7"
-            ],
-            "index": "pypi",
-            "version": "==2.26.0"
-        },
-        "toml": {
-            "hashes": [
-                "sha256:806143ae5bfb6a3c6e736a764057db0e6a0e05e338b5630894a5f779cabb4f9b",
-                "sha256:b3bda1d108d5dd99f4a20d24d9c348e91c4db7ab1b749200bded2f839ccbe68f"
-            ],
-            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==0.10.2"
-        },
-        "typing-extensions": {
-            "hashes": [
-                "sha256:49f75d16ff11f1cd258e1b988ccff82a3ca5570217d7ad8c5f48205dd99a677e",
-                "sha256:d8226d10bc02a29bcc81df19a26e56a9647f8b0a6d4a83924139f4a8b01f17b7",
-                "sha256:f1d25edafde516b146ecd0613dabcc61409817af4766fbbcfb8d1ad4ec441a34"
-            ],
-            "index": "pypi",
-            "version": "==3.10.0.2"
-        },
-        "urllib3": {
-            "hashes": [
-                "sha256:39fb8672126159acb139a7718dd10806104dec1e2f0f6c88aab05d17df10c8d4",
-                "sha256:f57b4c16c62fa2760b7e3d97c35b255512fb6b59a259730f36ba32ce9f8e342f"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3, 3.4' and python_version < '4'",
-            "version": "==1.26.6"
-        }
-    },
-    "develop": {
-        "flake8": {
-            "hashes": [
-                "sha256:07528381786f2a6237b061f6e96610a4167b226cb926e2aa2b6b1d78057c576b",
-                "sha256:bf8fd333346d844f616e8d47905ef3a3384edae6b4e9beb0c5101e25e3110907"
-            ],
-            "index": "pypi",
-            "version": "==3.9.2"
-        },
-        "mccabe": {
-            "hashes": [
-                "sha256:ab8a6258860da4b6677da4bd2fe5dc2c659cff31b3ee4f7f5d64e79735b80d42",
-                "sha256:dd8d182285a0fe56bace7f45b5e7d1a6ebcbf524e8f3bd87eb0f125271b8831f"
-            ],
-            "version": "==0.6.1"
-        },
-        "mypy": {
-            "hashes": [
-                "sha256:088cd9c7904b4ad80bec811053272986611b84221835e079be5bcad029e79dd9",
-                "sha256:0aadfb2d3935988ec3815952e44058a3100499f5be5b28c34ac9d79f002a4a9a",
-                "sha256:119bed3832d961f3a880787bf621634ba042cb8dc850a7429f643508eeac97b9",
-                "sha256:1a85e280d4d217150ce8cb1a6dddffd14e753a4e0c3cf90baabb32cefa41b59e",
-                "sha256:3c4b8ca36877fc75339253721f69603a9c7fdb5d4d5a95a1a1b899d8b86a4de2",
-                "sha256:3e382b29f8e0ccf19a2df2b29a167591245df90c0b5a2542249873b5c1d78212",
-                "sha256:42c266ced41b65ed40a282c575705325fa7991af370036d3f134518336636f5b",
-                "sha256:53fd2eb27a8ee2892614370896956af2ff61254c275aaee4c230ae771cadd885",
-                "sha256:704098302473cb31a218f1775a873b376b30b4c18229421e9e9dc8916fd16150",
-                "sha256:7df1ead20c81371ccd6091fa3e2878559b5c4d4caadaf1a484cf88d93ca06703",
-                "sha256:866c41f28cee548475f146aa4d39a51cf3b6a84246969f3759cb3e9c742fc072",
-                "sha256:a155d80ea6cee511a3694b108c4494a39f42de11ee4e61e72bc424c490e46457",
-                "sha256:adaeee09bfde366d2c13fe6093a7df5df83c9a2ba98638c7d76b010694db760e",
-                "sha256:b6fb13123aeef4a3abbcfd7e71773ff3ff1526a7d3dc538f3929a49b42be03f0",
-                "sha256:b94e4b785e304a04ea0828759172a15add27088520dc7e49ceade7834275bedb",
-                "sha256:c0df2d30ed496a08de5daed2a9ea807d07c21ae0ab23acf541ab88c24b26ab97",
-                "sha256:c6c2602dffb74867498f86e6129fd52a2770c48b7cd3ece77ada4fa38f94eba8",
-                "sha256:ceb6e0a6e27fb364fb3853389607cf7eb3a126ad335790fa1e14ed02fba50811",
-                "sha256:d9dd839eb0dc1bbe866a288ba3c1afc33a202015d2ad83b31e875b5905a079b6",
-                "sha256:e4dab234478e3bd3ce83bac4193b2ecd9cf94e720ddd95ce69840273bf44f6de",
-                "sha256:ec4e0cd079db280b6bdabdc807047ff3e199f334050db5cbb91ba3e959a67504",
-                "sha256:ecd2c3fe726758037234c93df7e98deb257fd15c24c9180dacf1ef829da5f921",
-                "sha256:ef565033fa5a958e62796867b1df10c40263ea9ded87164d67572834e57a174d"
-            ],
-            "index": "pypi",
-            "version": "==0.910"
-        },
-        "mypy-extensions": {
-            "hashes": [
-                "sha256:090fedd75945a69ae91ce1303b5824f428daf5a028d2f6ab8a299250a846f15d",
-                "sha256:2d82818f5bb3e369420cb3c4060a7970edba416647068eb4c5343488a6c604a8"
-            ],
-            "version": "==0.4.3"
-        },
-        "pycodestyle": {
-            "hashes": [
-                "sha256:514f76d918fcc0b55c6680472f0a37970994e07bbb80725808c17089be302068",
-                "sha256:c389c1d06bf7904078ca03399a4816f974a1d590090fecea0c63ec26ebaf1cef"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==2.7.0"
-        },
-        "pyflakes": {
-            "hashes": [
-                "sha256:7893783d01b8a89811dd72d7dfd4d84ff098e5eed95cfa8905b22bbffe52efc3",
-                "sha256:f5bc8ecabc05bb9d291eb5203d6810b49040f6ff446a756326104746cc00c1db"
-            ],
-            "markers": "python_version >= '2.7' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==2.3.1"
-        },
-        "toml": {
-            "hashes": [
-                "sha256:806143ae5bfb6a3c6e736a764057db0e6a0e05e338b5630894a5f779cabb4f9b",
-                "sha256:b3bda1d108d5dd99f4a20d24d9c348e91c4db7ab1b749200bded2f839ccbe68f"
-            ],
-            "markers": "python_version >= '2.6' and python_version not in '3.0, 3.1, 3.2, 3.3'",
-            "version": "==0.10.2"
-        },
-        "typing-extensions": {
-            "hashes": [
-                "sha256:49f75d16ff11f1cd258e1b988ccff82a3ca5570217d7ad8c5f48205dd99a677e",
-                "sha256:d8226d10bc02a29bcc81df19a26e56a9647f8b0a6d4a83924139f4a8b01f17b7",
-                "sha256:f1d25edafde516b146ecd0613dabcc61409817af4766fbbcfb8d1ad4ec441a34"
-            ],
-            "index": "pypi",
-            "version": "==3.10.0.2"
-        },
-        "yapf": {
-            "hashes": [
-                "sha256:408fb9a2b254c302f49db83c59f9aa0b4b0fd0ec25be3a5c51181327922ff63d",
-                "sha256:e3a234ba8455fe201eaa649cdac872d590089a18b661e39bbac7020978dd9c2e"
-            ],
-            "index": "pypi",
-            "version": "==0.31.0"
-        }
-    }
-}

test_runner/README.md

@@ -3,18 +3,13 @@
 This directory contains integration tests.
 
 Prerequisites:
-- Python 3.6 or later
-- Dependencies: install them via `pipenv install`. Note that Debian/Ubuntu
-  packages are stale, as it commonly happens, so manual installation is not
-  recommended.
-  Run `pipenv shell` to activate the venv or use `pipenv run` to run a single
-  command in the venv, e.g. `pipenv run pytest`.
+- Correctly configured Python, see [`/docs/sourcetree.md`](/docs/sourcetree.md#using-python)
 - Zenith and Postgres binaries
-    - See the root README.md for build directions
+    - See the root [README.md](/README.md) for build directions
     - Tests can be run from the git tree; or see the environment variables
       below to run from other directories.
 - The zenith git repo, including the postgres submodule
-  (for some tests, e.g. pg_regress)
+  (for some tests, e.g. `pg_regress`)
 
 ### Test Organization
 
@@ -35,15 +30,15 @@ be stored under a directory `test_output`.
 
 You can run all the tests with:
 
-`pytest`
+`pipenv run pytest`
 
 If you want to run all the tests in a particular file:
 
-`pytest test_pgbench.py`
+`pipenv run pytest test_pgbench.py`
 
 If you want to run all tests that have the string "bench" in their names:
 
-`pytest -k bench`
+`pipenv run pytest -k bench`
 
 Useful environment variables:
 
@@ -53,8 +48,8 @@ Useful environment variables:
 should go.
 `TEST_SHARED_FIXTURES`: Try to re-use a single pageserver for all the tests.
 
-Let stdout and stderr go to the terminal instead of capturing them:
-`pytest -s ...`
+Let stdout, stderr and `INFO` log messages go to the terminal instead of capturing them:
+`pytest -s --log-cli-level=INFO ...`
 (Note many tests capture subprocess outputs separately, so this may not
 show much.)
 
@@ -62,44 +57,51 @@ Exit after the first test failure:
 `pytest -x ...`
 (there are many more pytest options; run `pytest -h` to see them.)
 
+### Writing a test
 
-### Building new tests
+Every test needs a Zenith Environment, or ZenithEnv to operate in. A Zenith Environment
+is like a little cloud-in-a-box, and consists of a Pageserver, 0-N Safekeepers, and
+compute Postgres nodes. The connections between them can be configured to use JWT
+authentication tokens, and some other configuration options can be tweaked too.
 
-The tests make heavy use of pytest fixtures. You can read about how they work here: https://docs.pytest.org/en/stable/fixture.html
+The easiest way to get access to a Zenith Environment is by using the `zenith_simple_env`
+fixture. The 'simple' env may be shared across multiple tests, so don't shut down the nodes
+or make other destructive changes in that environment. Also don't assume that
+there are no tenants or branches or data in the cluster. For convenience, there is a
+branch called `empty`, though. The convention is to create a test-specific branch of
+that and load any test data there, instead of the 'main' branch.
 
-Essentially, this means that each time you see a fixture named as an input parameter, the function with that name will be run and passed as a parameter to the function.
-
-So this code:
+For more complicated cases, you can build a custom Zenith Environment, with the `zenith_env`
+fixture:
 
 ```python
-def test_something(zenith_cli, pg_bin):
-    pass
+def test_foobar(zenith_env_builder: ZenithEnvBuilder):
+    # Prescribe the environment.
+    # We want to have 3 safekeeper nodes, and use JWT authentication in the
+    # connections to the page server
+    zenith_env_builder.num_safekeepers = 3
+    zenith_env_builder.set_pageserver_auth(True)
+
+    # Now create the environment. This initializes the repository, and starts
+    # up the page server and the safekeepers
+    env = zenith_env_builder.init()
+
+    # Run the test
+    ...
 ```
 
-... will run the fixtures called `zenith_cli` and `pg_bin` and deliver those results to the test function.
+For more information about pytest fixtures, see https://docs.pytest.org/en/stable/fixture.html
 
-Fixtures can't be imported using the normal python syntax. Instead, use this:
+At the end of a test, all the nodes in the environment are automatically stopped, so you
+don't need to worry about cleaning up. Logs and test data are preserved for the analysis,
+in a directory under `../test_output/<testname>`
 
-```python
-pytest_plugins = ("fixtures.something")
-```
+### Before submitting a patch
+Ensure that you pass all [obligatory checks](/docs/sourcetree.md#obligatory-checks).
 
-That will make all the fixtures in the `fixtures/something.py` file available.
-
-Anything that's likely to be used in multiple tests should be built into a fixture.
-
-Note that fixtures can clean up after themselves if they use the `yield` syntax.
-Cleanup will happen even if the test fails (raises an unhandled exception).
-Python destructors, e.g. `__del__()` aren't recommended for cleanup.
-
-
-### Code quality
-
-Before submitting a patch, please consider:
+Also consider:
 
 * Writing a couple of docstrings to clarify the reasoning behind a new test.
-* Running `flake8` (or a linter of your choice, e.g. `pycodestyle`) and fixing possible defects, if any.
-* Formatting the code with `yapf -r -i .` (TODO: implement an opt-in pre-commit hook for that).
-* (Optional) Typechecking the code with `mypy .`. Currently this mostly affects `fixtures/zenith_fixtures.py`.
-
-The tools can be installed with `pipenv install --dev`.
+* Adding more type hints to your code to avoid `Any`, especially:
+  * For fixture parameters, they are not automatically deduced.
+  * For function arguments and return values.

test_runner/batch_others/test_auth.py

@@ -1,21 +1,22 @@
-
 from contextlib import closing
 from typing import Iterator
 from uuid import uuid4
 import psycopg2
-from fixtures.zenith_fixtures import PortDistributor, Postgres, ZenithCli, ZenithPageserver, PgBin
+from fixtures.zenith_fixtures import ZenithEnvBuilder
 import pytest
 
-
 pytest_plugins = ("fixtures.zenith_fixtures")
 
 
-def test_pageserver_auth(pageserver_auth_enabled: ZenithPageserver):
-    ps = pageserver_auth_enabled
+def test_pageserver_auth(zenith_env_builder: ZenithEnvBuilder):
+    zenith_env_builder.pageserver_auth_enabled = True
+    env = zenith_env_builder.init()
 
-    tenant_token = ps.auth_keys.generate_tenant_token(ps.initial_tenant)
-    invalid_tenant_token = ps.auth_keys.generate_tenant_token(uuid4().hex)
-    management_token = ps.auth_keys.generate_management_token()
+    ps = env.pageserver
+
+    tenant_token = env.auth_keys.generate_tenant_token(env.initial_tenant)
+    invalid_tenant_token = env.auth_keys.generate_tenant_token(uuid4().hex)
+    management_token = env.auth_keys.generate_management_token()
 
     # this does not invoke auth check and only decodes jwt and checks it for validity
     # check both tokens
@@ -23,57 +24,41 @@ def test_pageserver_auth(pageserver_auth_enabled: ZenithPageserver):
     ps.safe_psql("status", password=management_token)
 
     # tenant can create branches
-    ps.safe_psql(f"branch_create {ps.initial_tenant} new1 main", password=tenant_token)
+    ps.safe_psql(f"branch_create {env.initial_tenant} new1 main", password=tenant_token)
     # console can create branches for tenant
-    ps.safe_psql(f"branch_create {ps.initial_tenant} new2 main", password=management_token)
+    ps.safe_psql(f"branch_create {env.initial_tenant} new2 main", password=management_token)
 
     # fail to create branch using token with different tenantid
     with pytest.raises(psycopg2.DatabaseError, match='Tenant id mismatch. Permission denied'):
-        ps.safe_psql(f"branch_create {ps.initial_tenant} new2 main", password=invalid_tenant_token)
+        ps.safe_psql(f"branch_create {env.initial_tenant} new2 main", password=invalid_tenant_token)
 
     # create tenant using management token
     ps.safe_psql(f"tenant_create {uuid4().hex}", password=management_token)
 
     # fail to create tenant using tenant token
-    with pytest.raises(psycopg2.DatabaseError, match='Attempt to access management api with tenant scope. Permission denied'):
+    with pytest.raises(
+            psycopg2.DatabaseError,
+            match='Attempt to access management api with tenant scope. Permission denied'):
         ps.safe_psql(f"tenant_create {uuid4().hex}", password=tenant_token)
 
 
 @pytest.mark.parametrize('with_wal_acceptors', [False, True])
-def test_compute_auth_to_pageserver(
-    zenith_cli: ZenithCli,
-    wa_factory,
-    pageserver_auth_enabled: ZenithPageserver,
-    repo_dir: str,
-    with_wal_acceptors: bool,
-    pg_bin: PgBin,
-    port_distributor: PortDistributor,
-):
-    ps = pageserver_auth_enabled
-    # since we are in progress of refactoring protocols between compute safekeeper and page server
-    # use hardcoded management token in safekeeper
-    management_token = ps.auth_keys.generate_management_token()
+def test_compute_auth_to_pageserver(zenith_env_builder: ZenithEnvBuilder, with_wal_acceptors: bool):
+    zenith_env_builder.pageserver_auth_enabled = True
+    if with_wal_acceptors:
+        zenith_env_builder.num_safekeepers = 3
+    env = zenith_env_builder.init()
 
     branch = f"test_compute_auth_to_pageserver{with_wal_acceptors}"
-    zenith_cli.run(["branch", branch, "empty"])
-    if with_wal_acceptors:
-        wa_factory.start_n_new(3, management_token)
+    env.zenith_cli(["branch", branch, "main"])
 
-    with Postgres(
-        zenith_cli=zenith_cli,
-        repo_dir=repo_dir,
-        pg_bin=pg_bin,
-        tenant_id=ps.initial_tenant,
-        port=port_distributor.get_port(),
-    ).create_start(
-        branch,
-        wal_acceptors=wa_factory.get_connstrs() if with_wal_acceptors else None,
-    ) as pg:
-        with closing(pg.connect()) as conn:
-            with conn.cursor() as cur:
-                # we rely upon autocommit after each statement
-                # as waiting for acceptors happens there
-                cur.execute('CREATE TABLE t(key int primary key, value text)')
-                cur.execute("INSERT INTO t SELECT generate_series(1,100000), 'payload'")
-                cur.execute('SELECT sum(key) FROM t')
-                assert cur.fetchone() == (5000050000, )
+    pg = env.postgres.create_start(branch)
+
+    with closing(pg.connect()) as conn:
+        with conn.cursor() as cur:
+            # we rely upon autocommit after each statement
+            # as waiting for acceptors happens there
+            cur.execute('CREATE TABLE t(key int primary key, value text)')
+            cur.execute("INSERT INTO t SELECT generate_series(1,100000), 'payload'")
+            cur.execute('SELECT sum(key) FROM t')
+            assert cur.fetchone() == (5000050000, )

test_runner/batch_others/test_branch_behind.py

@@ -1,6 +1,11 @@
 import subprocess
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+from contextlib import closing
 
+import psycopg2.extras
+import pytest
+from fixtures.log_helper import log
+from fixtures.utils import print_gc_result
+from fixtures.zenith_fixtures import ZenithEnv
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -8,18 +13,27 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 #
 # Create a couple of branches off the main branch, at a historical point in time.
 #
-def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, pg_bin):
+def test_branch_behind(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
     # Branch at the point where only 100 rows were inserted
-    zenith_cli.run(["branch", "test_branch_behind", "empty"])
+    env.zenith_cli(["branch", "test_branch_behind", "empty"])
 
-    pgmain = postgres.create_start('test_branch_behind')
-    print("postgres is running on 'test_branch_behind' branch")
+    pgmain = env.postgres.create_start('test_branch_behind')
+    log.info("postgres is running on 'test_branch_behind' branch")
 
     main_pg_conn = pgmain.connect()
     main_cur = main_pg_conn.cursor()
 
+    main_cur.execute("SHOW zenith.zenith_timeline")
+    timeline = main_cur.fetchone()[0]
+
     # Create table, and insert the first 100 rows
     main_cur.execute('CREATE TABLE foo (t text)')
+
+    # keep some early lsn to test branch creation on out of date lsn
+    main_cur.execute('SELECT pg_current_wal_insert_lsn()')
+    gced_lsn = main_cur.fetchone()[0]
+
     main_cur.execute('''
         INSERT INTO foo
             SELECT 'long string to consume some space' || g
@@ -27,7 +41,7 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     ''')
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
     lsn_a = main_cur.fetchone()[0]
-    print('LSN after 100 rows: ' + lsn_a)
+    log.info(f'LSN after 100 rows: {lsn_a}')
 
     # Insert some more rows. (This generates enough WAL to fill a few segments.)
     main_cur.execute('''
@@ -37,10 +51,10 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     ''')
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
     lsn_b = main_cur.fetchone()[0]
-    print('LSN after 200100 rows: ' + lsn_b)
+    log.info(f'LSN after 200100 rows: {lsn_b}')
 
     # Branch at the point where only 100 rows were inserted
-    zenith_cli.run(["branch", "test_branch_behind_hundred", "test_branch_behind@" + lsn_a])
+    env.zenith_cli(["branch", "test_branch_behind_hundred", "test_branch_behind@" + lsn_a])
 
     # Insert many more rows. This generates enough WAL to fill a few segments.
     main_cur.execute('''
@@ -52,13 +66,13 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
 
     main_cur.execute('SELECT pg_current_wal_insert_lsn()')
     lsn_c = main_cur.fetchone()[0]
-    print('LSN after 400100 rows: ' + lsn_c)
+    log.info(f'LSN after 400100 rows: {lsn_c}')
 
     # Branch at the point where only 200100 rows were inserted
-    zenith_cli.run(["branch", "test_branch_behind_more", "test_branch_behind@" + lsn_b])
+    env.zenith_cli(["branch", "test_branch_behind_more", "test_branch_behind@" + lsn_b])
 
-    pg_hundred = postgres.create_start("test_branch_behind_hundred")
-    pg_more = postgres.create_start("test_branch_behind_more")
+    pg_hundred = env.postgres.create_start("test_branch_behind_hundred")
+    pg_more = env.postgres.create_start("test_branch_behind_more")
 
     # On the 'hundred' branch, we should see only 100 rows
     hundred_pg_conn = pg_hundred.connect()
@@ -79,14 +93,34 @@ def test_branch_behind(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     # Check bad lsn's for branching
 
     # branch at segment boundary
-    zenith_cli.run(["branch", "test_branch_segment_boundary", "test_branch_behind@0/3000000"])
-    pg = postgres.create_start("test_branch_segment_boundary")
+    env.zenith_cli(["branch", "test_branch_segment_boundary", "test_branch_behind@0/3000000"])
+    pg = env.postgres.create_start("test_branch_segment_boundary")
     cur = pg.connect().cursor()
     cur.execute('SELECT 1')
     assert cur.fetchone() == (1, )
 
     # branch at pre-initdb lsn
-    try:
-        zenith_cli.run(["branch", "test_branch_preinitdb", "test_branch_behind@0/42"])
-    except subprocess.CalledProcessError:
-        print("Branch creation with pre-initdb LSN failed (as expected)")
+    with pytest.raises(Exception, match="invalid branch start lsn"):
+        env.zenith_cli(["branch", "test_branch_preinitdb", "test_branch_behind@0/42"])
+
+    # check that we cannot create branch based on garbage collected data
+    with closing(env.pageserver.connect()) as psconn:
+        with psconn.cursor(cursor_factory=psycopg2.extras.DictCursor) as pscur:
+            # call gc to advace latest_gc_cutoff_lsn
+            pscur.execute(f"do_gc {env.initial_tenant} {timeline} 0")
+            row = pscur.fetchone()
+            print_gc_result(row)
+
+    with pytest.raises(Exception, match="invalid branch start lsn"):
+        # this gced_lsn is pretty random, so if gc is disabled this woudln't fail
+        env.zenith_cli(["branch", "test_branch_create_fail", f"test_branch_behind@{gced_lsn}"])
+
+    # check that after gc everything is still there
+    hundred_cur.execute('SELECT count(*) FROM foo')
+    assert hundred_cur.fetchone() == (100, )
+
+    more_cur.execute('SELECT count(*) FROM foo')
+    assert more_cur.fetchone() == (200100, )
+
+    main_cur.execute('SELECT count(*) FROM foo')
+    assert main_cur.fetchone() == (400100, )

test_runner/batch_others/test_clog_truncate.py

@@ -3,7 +3,8 @@ import os
 
 from contextlib import closing
 
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -11,20 +12,24 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 #
 # Test compute node start after clog truncation
 #
-def test_clog_truncate(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, pg_bin):
+def test_clog_truncate(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_clog_truncate", "empty"])
+    env.zenith_cli(["branch", "test_clog_truncate", "empty"])
 
     # set agressive autovacuum to make sure that truncation will happen
     config = [
-        'autovacuum_max_workers=10', 'autovacuum_vacuum_threshold=0',
-        'autovacuum_vacuum_insert_threshold=0', 'autovacuum_vacuum_cost_delay=0',
-        'autovacuum_vacuum_cost_limit=10000', 'autovacuum_naptime =1s',
+        'autovacuum_max_workers=10',
+        'autovacuum_vacuum_threshold=0',
+        'autovacuum_vacuum_insert_threshold=0',
+        'autovacuum_vacuum_cost_delay=0',
+        'autovacuum_vacuum_cost_limit=10000',
+        'autovacuum_naptime =1s',
         'autovacuum_freeze_max_age=100000'
     ]
 
-    pg = postgres.create_start('test_clog_truncate', config_lines=config)
-    print('postgres is running on test_clog_truncate branch')
+    pg = env.postgres.create_start('test_clog_truncate', config_lines=config)
+    log.info('postgres is running on test_clog_truncate branch')
 
     # Install extension containing function needed for test
     pg.safe_psql('CREATE EXTENSION zenith_test_utils')
@@ -33,22 +38,22 @@ def test_clog_truncate(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:
             cur.execute('select test_consume_xids(1000*1000*10);')
-            print('xids consumed')
+            log.info('xids consumed')
 
             # call a checkpoint to trigger TruncateSubtrans
             cur.execute('CHECKPOINT;')
 
             # ensure WAL flush
             cur.execute('select txid_current()')
-            print(cur.fetchone())
+            log.info(cur.fetchone())
 
     # wait for autovacuum to truncate the pg_xact
     # XXX Is it worth to add a timeout here?
     pg_xact_0000_path = os.path.join(pg.pg_xact_dir_path(), '0000')
-    print("pg_xact_0000_path = " + pg_xact_0000_path)
+    log.info(f"pg_xact_0000_path = {pg_xact_0000_path}")
 
     while os.path.isfile(pg_xact_0000_path):
-        print("file exists. wait for truncation. " "pg_xact_0000_path = " + pg_xact_0000_path)
+        log.info(f"file exists. wait for truncation. " "pg_xact_0000_path = {pg_xact_0000_path}")
         time.sleep(5)
 
     # checkpoint to advance latest lsn
@@ -59,14 +64,14 @@ def test_clog_truncate(zenith_cli, pageserver: ZenithPageserver, postgres: Postg
             lsn_after_truncation = cur.fetchone()[0]
 
     # create new branch after clog truncation and start a compute node on it
-    print('create branch at lsn_after_truncation ' + lsn_after_truncation)
-    zenith_cli.run(
+    log.info(f'create branch at lsn_after_truncation {lsn_after_truncation}')
+    env.zenith_cli(
         ["branch", "test_clog_truncate_new", "test_clog_truncate@" + lsn_after_truncation])
 
-    pg2 = postgres.create_start('test_clog_truncate_new')
-    print('postgres is running on test_clog_truncate_new branch')
+    pg2 = env.postgres.create_start('test_clog_truncate_new')
+    log.info('postgres is running on test_clog_truncate_new branch')
 
     # check that new node doesn't contain truncated segment
     pg_xact_0000_path_new = os.path.join(pg2.pg_xact_dir_path(), '0000')
-    print("pg_xact_0000_path_new = " + pg_xact_0000_path_new)
+    log.info(f"pg_xact_0000_path_new = {pg_xact_0000_path_new}")
     assert os.path.isfile(pg_xact_0000_path_new) is False

test_runner/batch_others/test_config.py

@@ -1,6 +1,7 @@
 from contextlib import closing
 
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -8,13 +9,14 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 #
 # Test starting Postgres with custom options
 #
-def test_config(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, pg_bin):
+def test_config(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_config", "empty"])
+    env.zenith_cli(["branch", "test_config", "empty"])
 
     # change config
-    pg = postgres.create_start('test_config', config_lines=['log_min_messages=debug1'])
-    print('postgres is running on test_config branch')
+    pg = env.postgres.create_start('test_config', config_lines=['log_min_messages=debug1'])
+    log.info('postgres is running on test_config branch')
 
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:

test_runner/batch_others/test_createdropdb.py

@@ -2,7 +2,8 @@ import os
 import pathlib
 
 from contextlib import closing
-from fixtures.zenith_fixtures import ZenithPageserver, PostgresFactory, ZenithCli, check_restored_datadir_content
+from fixtures.zenith_fixtures import ZenithEnv, check_restored_datadir_content
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -10,16 +11,12 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 #
 # Test CREATE DATABASE when there have been relmapper changes
 #
-def test_createdb(
-    zenith_cli: ZenithCli,
-    pageserver: ZenithPageserver,
-    postgres: PostgresFactory,
-    pg_bin,
-):
-    zenith_cli.run(["branch", "test_createdb", "empty"])
+def test_createdb(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    env.zenith_cli(["branch", "test_createdb", "empty"])
 
-    pg = postgres.create_start('test_createdb')
-    print("postgres is running on 'test_createdb' branch")
+    pg = env.postgres.create_start('test_createdb')
+    log.info("postgres is running on 'test_createdb' branch")
 
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:
@@ -32,28 +29,24 @@ def test_createdb(
             lsn = cur.fetchone()[0]
 
     # Create a branch
-    zenith_cli.run(["branch", "test_createdb2", "test_createdb@" + lsn])
+    env.zenith_cli(["branch", "test_createdb2", "test_createdb@" + lsn])
 
-    pg2 = postgres.create_start('test_createdb2')
+    pg2 = env.postgres.create_start('test_createdb2')
 
     # Test that you can connect to the new database on both branches
     for db in (pg, pg2):
         db.connect(dbname='foodb').close()
 
+
 #
 # Test DROP DATABASE
 #
-def test_dropdb(
-    zenith_cli: ZenithCli,
-    pageserver: ZenithPageserver,
-    postgres: PostgresFactory,
-    pg_bin,
-    test_output_dir
-):
-    zenith_cli.run(["branch", "test_dropdb", "empty"])
+def test_dropdb(zenith_simple_env: ZenithEnv, test_output_dir):
+    env = zenith_simple_env
+    env.zenith_cli(["branch", "test_dropdb", "empty"])
 
-    pg = postgres.create_start('test_dropdb')
-    print("postgres is running on 'test_dropdb' branch")
+    pg = env.postgres.create_start('test_dropdb')
+    log.info("postgres is running on 'test_dropdb' branch")
 
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:
@@ -65,7 +58,6 @@ def test_dropdb(
             cur.execute("SELECT oid FROM pg_database WHERE datname='foodb';")
             dboid = cur.fetchone()[0]
 
-
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:
             cur.execute('DROP DATABASE foodb')
@@ -75,28 +67,29 @@ def test_dropdb(
             cur.execute('SELECT pg_current_wal_insert_lsn()')
             lsn_after_drop = cur.fetchone()[0]
 
-
     # Create two branches before and after database drop.
-    zenith_cli.run(["branch", "test_before_dropdb", "test_dropdb@" + lsn_before_drop])
-    pg_before = postgres.create_start('test_before_dropdb')
+    env.zenith_cli(["branch", "test_before_dropdb", "test_dropdb@" + lsn_before_drop])
+    pg_before = env.postgres.create_start('test_before_dropdb')
 
-    zenith_cli.run(["branch", "test_after_dropdb", "test_dropdb@" + lsn_after_drop])
-    pg_after = postgres.create_start('test_after_dropdb')
+    env.zenith_cli(["branch", "test_after_dropdb", "test_dropdb@" + lsn_after_drop])
+    pg_after = env.postgres.create_start('test_after_dropdb')
 
     # Test that database exists on the branch before drop
     pg_before.connect(dbname='foodb').close()
 
     # Test that database subdir exists on the branch before drop
+    assert pg_before.pgdata_dir
     dbpath = pathlib.Path(pg_before.pgdata_dir) / 'base' / str(dboid)
-    print(dbpath)
+    log.info(dbpath)
 
     assert os.path.isdir(dbpath) == True
 
     # Test that database subdir doesn't exist on the branch after drop
+    assert pg_after.pgdata_dir
     dbpath = pathlib.Path(pg_after.pgdata_dir) / 'base' / str(dboid)
-    print(dbpath)
+    log.info(dbpath)
 
     assert os.path.isdir(dbpath) == False
 
     # Check that we restore the content of the datadir correctly
-    check_restored_datadir_content(zenith_cli, test_output_dir, pg, pageserver.service_port.pg)
+    check_restored_datadir_content(test_output_dir, env, pg)

test_runner/batch_others/test_createuser.py

@@ -1,6 +1,7 @@
 from contextlib import closing
 
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -8,11 +9,12 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 #
 # Test CREATE USER to check shared catalog restore
 #
-def test_createuser(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, pg_bin):
-    zenith_cli.run(["branch", "test_createuser", "empty"])
+def test_createuser(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    env.zenith_cli(["branch", "test_createuser", "empty"])
 
-    pg = postgres.create_start('test_createuser')
-    print("postgres is running on 'test_createuser' branch")
+    pg = env.postgres.create_start('test_createuser')
+    log.info("postgres is running on 'test_createuser' branch")
 
     with closing(pg.connect()) as conn:
         with conn.cursor() as cur:
@@ -25,9 +27,9 @@ def test_createuser(zenith_cli, pageserver: ZenithPageserver, postgres: Postgres
             lsn = cur.fetchone()[0]
 
     # Create a branch
-    zenith_cli.run(["branch", "test_createuser2", "test_createuser@" + lsn])
+    env.zenith_cli(["branch", "test_createuser2", "test_createuser@" + lsn])
 
-    pg2 = postgres.create_start('test_createuser2')
+    pg2 = env.postgres.create_start('test_createuser2')
 
     # Test that you can connect to new branch as a new user
     assert pg2.safe_psql('select current_user', username='testuser') == [('testuser', )]

test_runner/batch_others/test_multixact.py

@@ -1,4 +1,5 @@
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver, check_restored_datadir_content
+from fixtures.zenith_fixtures import ZenithEnv, check_restored_datadir_content
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
@@ -9,13 +10,13 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 # it only checks next_multixact_id field in restored pg_control,
 # since we don't have functions to check multixact internals.
 #
-def test_multixact(pageserver: ZenithPageserver, postgres: PostgresFactory,
-                    pg_bin, zenith_cli, base_dir, test_output_dir):
+def test_multixact(zenith_simple_env: ZenithEnv, test_output_dir):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_multixact", "empty"])
-    pg = postgres.create_start('test_multixact')
+    env.zenith_cli(["branch", "test_multixact", "empty"])
+    pg = env.postgres.create_start('test_multixact')
 
-    print("postgres is running on 'test_multixact' branch")
+    log.info("postgres is running on 'test_multixact' branch")
     pg_conn = pg.connect()
     cur = pg_conn.cursor()
 
@@ -52,10 +53,10 @@ def test_multixact(pageserver: ZenithPageserver, postgres: PostgresFactory,
     assert int(next_multixact_id) > int(next_multixact_id_old)
 
     # Branch at this point
-    zenith_cli.run(["branch", "test_multixact_new", "test_multixact@" + lsn])
-    pg_new = postgres.create_start('test_multixact_new')
+    env.zenith_cli(["branch", "test_multixact_new", "test_multixact@" + lsn])
+    pg_new = env.postgres.create_start('test_multixact_new')
 
-    print("postgres is running on 'test_multixact_new' branch")
+    log.info("postgres is running on 'test_multixact_new' branch")
     pg_new_conn = pg_new.connect()
     cur_new = pg_new_conn.cursor()
 
@@ -66,4 +67,4 @@ def test_multixact(pageserver: ZenithPageserver, postgres: PostgresFactory,
     assert next_multixact_id_new == next_multixact_id
 
     # Check that we restore the content of the datadir correctly
-    check_restored_datadir_content(zenith_cli, test_output_dir, pg_new, pageserver.service_port.pg)
+    check_restored_datadir_content(test_output_dir, env, pg_new)

test_runner/batch_others/test_old_request_lsn.py

@@ -1,9 +1,11 @@
 from contextlib import closing
 
-from fixtures.zenith_fixtures import PostgresFactory, ZenithPageserver
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
+
 #
 # Test where Postgres generates a lot of WAL, and it's garbage collected away, but
 # no pages are evicted so that Postgres uses an old LSN in a GetPage request.
@@ -14,11 +16,12 @@ pytest_plugins = ("fixtures.zenith_fixtures")
 # just a hint that the page hasn't been modified since that LSN, and the page
 # server should return the latest page version regardless of the LSN.
 #
-def test_old_request_lsn(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, pg_bin):
+def test_old_request_lsn(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_old_request_lsn", "empty"])
-    pg = postgres.create_start('test_old_request_lsn')
-    print('postgres is running on test_old_request_lsn branch')
+    env.zenith_cli(["branch", "test_old_request_lsn", "empty"])
+    pg = env.postgres.create_start('test_old_request_lsn')
+    log.info('postgres is running on test_old_request_lsn branch')
 
     pg_conn = pg.connect()
     cur = pg_conn.cursor()
@@ -27,7 +30,7 @@ def test_old_request_lsn(zenith_cli, pageserver: ZenithPageserver, postgres: Pos
     cur.execute("SHOW zenith.zenith_timeline")
     timeline = cur.fetchone()[0]
 
-    psconn = pageserver.connect()
+    psconn = env.pageserver.connect()
     pscur = psconn.cursor()
 
     # Create table, and insert some rows. Make it big enough that it doesn't fit in
@@ -46,20 +49,20 @@ def test_old_request_lsn(zenith_cli, pageserver: ZenithPageserver, postgres: Pos
         from pg_settings where name = 'shared_buffers'
     ''')
     row = cur.fetchone()
-    print(f'shared_buffers is {row[0]}, table size {row[1]}');
+    log.info(f'shared_buffers is {row[0]}, table size {row[1]}')
     assert int(row[0]) < int(row[1])
 
-    cur.execute('VACUUM foo');
+    cur.execute('VACUUM foo')
 
     # Make a lot of updates on a single row, generating a lot of WAL. Trigger
     # garbage collections so that the page server will remove old page versions.
     for i in range(10):
-        pscur.execute(f"do_gc {pageserver.initial_tenant} {timeline} 0")
+        pscur.execute(f"do_gc {env.initial_tenant} {timeline} 0")
         for j in range(100):
-            cur.execute('UPDATE foo SET val = val + 1 WHERE id = 1;');
+            cur.execute('UPDATE foo SET val = val + 1 WHERE id = 1;')
 
     # All (or at least most of) the updates should've been on the same page, so
     # that we haven't had to evict any dirty pages for a long time. Now run
     # a query that sends GetPage@LSN requests with the old LSN.
-    cur.execute("SELECT COUNT(*), SUM(val) FROM foo");
+    cur.execute("SELECT COUNT(*), SUM(val) FROM foo")
     assert cur.fetchone() == (100000, 101000)

test_runner/batch_others/test_pageserver_api.py

@@ -3,25 +3,28 @@ from uuid import uuid4
 import pytest
 import psycopg2
 import requests
-from fixtures.zenith_fixtures import ZenithPageserver, ZenithPageserverHttpClient
+from fixtures.zenith_fixtures import ZenithEnv, ZenithEnvBuilder, ZenithPageserverHttpClient
+from typing import cast
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
 
-def test_status_psql(pageserver):
-    assert pageserver.safe_psql('status') == [
+def test_status_psql(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    assert env.pageserver.safe_psql('status') == [
         ('hello world', ),
     ]
 
 
-def test_branch_list_psql(pageserver: ZenithPageserver, zenith_cli):
+def test_branch_list_psql(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_branch_list_main", "empty"])
+    env.zenith_cli(["branch", "test_branch_list_main", "empty"])
 
-    conn = pageserver.connect()
+    conn = env.pageserver.connect()
     cur = conn.cursor()
 
-    cur.execute(f'branch_list {pageserver.initial_tenant}')
+    cur.execute(f'branch_list {env.initial_tenant}')
     branches = json.loads(cur.fetchone()[0])
     # Filter out branches created by other tests
     branches = [x for x in branches if x['name'].startswith('test_branch_list')]
@@ -34,10 +37,10 @@ def test_branch_list_psql(pageserver: ZenithPageserver, zenith_cli):
     assert 'ancestor_lsn' in branches[0]
 
     # Create another branch, and start Postgres on it
-    zenith_cli.run(['branch', 'test_branch_list_experimental', 'test_branch_list_main'])
-    zenith_cli.run(['pg', 'create', 'test_branch_list_experimental'])
+    env.zenith_cli(['branch', 'test_branch_list_experimental', 'test_branch_list_main'])
+    env.zenith_cli(['pg', 'create', 'test_branch_list_experimental'])
 
-    cur.execute(f'branch_list {pageserver.initial_tenant}')
+    cur.execute(f'branch_list {env.initial_tenant}')
     new_branches = json.loads(cur.fetchone()[0])
     # Filter out branches created by other tests
     new_branches = [x for x in new_branches if x['name'].startswith('test_branch_list')]
@@ -53,18 +56,22 @@ def test_branch_list_psql(pageserver: ZenithPageserver, zenith_cli):
     conn.close()
 
 
-def test_tenant_list_psql(pageserver: ZenithPageserver, zenith_cli):
-    res = zenith_cli.run(["tenant", "list"])
-    res.check_returncode()
-    tenants = res.stdout.splitlines()
-    assert tenants == [pageserver.initial_tenant]
+def test_tenant_list_psql(zenith_env_builder: ZenithEnvBuilder):
+    # don't use zenith_simple_env, because there might be other tenants there,
+    # left over from other tests.
+    env = zenith_env_builder.init()
 
-    conn = pageserver.connect()
+    res = env.zenith_cli(["tenant", "list"])
+    res.check_returncode()
+    tenants = sorted(map(lambda t: t.split()[0], res.stdout.splitlines()))
+    assert tenants == [env.initial_tenant]
+
+    conn = env.pageserver.connect()
     cur = conn.cursor()
 
     # check same tenant cannot be created twice
-    with pytest.raises(psycopg2.DatabaseError, match=f'tenant {pageserver.initial_tenant} already exists'):
-        cur.execute(f'tenant_create {pageserver.initial_tenant}')
+    with pytest.raises(psycopg2.DatabaseError, match=f'tenant {env.initial_tenant} already exists'):
+        cur.execute(f'tenant_create {env.initial_tenant}')
 
     # create one more tenant
     tenant1 = uuid4().hex
@@ -73,20 +80,20 @@ def test_tenant_list_psql(pageserver: ZenithPageserver, zenith_cli):
     cur.execute('tenant_list')
 
     # compare tenants list
-    new_tenants = sorted(json.loads(cur.fetchone()[0]))
-    assert sorted([pageserver.initial_tenant, tenant1]) == new_tenants
+    new_tenants = sorted(map(lambda t: cast(str, t['id']), json.loads(cur.fetchone()[0])))
+    assert sorted([env.initial_tenant, tenant1]) == new_tenants
 
 
 def check_client(client: ZenithPageserverHttpClient, initial_tenant: str):
     client.check_status()
 
     # check initial tenant is there
-    assert initial_tenant in set(client.tenant_list())
+    assert initial_tenant in {t['id'] for t in client.tenant_list()}
 
     # create new tenant and check it is also there
     tenant_id = uuid4()
     client.tenant_create(tenant_id)
-    assert tenant_id.hex in set(client.tenant_list())
+    assert tenant_id.hex in {t['id'] for t in client.tenant_list()}
 
     # create branch
     branch_name = uuid4().hex
@@ -96,11 +103,17 @@ def check_client(client: ZenithPageserverHttpClient, initial_tenant: str):
     assert branch_name in {b['name'] for b in client.branch_list(tenant_id)}
 
 
-def test_pageserver_http_api_client(pageserver: ZenithPageserver):
-    client = pageserver.http_client()
-    check_client(client, pageserver.initial_tenant)
+def test_pageserver_http_api_client(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    client = env.pageserver.http_client()
+    check_client(client, env.initial_tenant)
 
 
-def test_pageserver_http_api_client_auth_enabled(pageserver_auth_enabled: ZenithPageserver):
-    client = pageserver_auth_enabled.http_client(auth_token=pageserver_auth_enabled.auth_keys.generate_management_token())
-    check_client(client, pageserver_auth_enabled.initial_tenant)
+def test_pageserver_http_api_client_auth_enabled(zenith_env_builder: ZenithEnvBuilder):
+    zenith_env_builder.pageserver_auth_enabled = True
+    env = zenith_env_builder.init()
+
+    management_token = env.auth_keys.generate_management_token()
+
+    client = env.pageserver.http_client(auth_token=management_token)
+    check_client(client, env.initial_tenant)

test_runner/batch_others/test_pageserver_restart.py

@@ -4,21 +4,22 @@ import time
 
 from contextlib import closing
 from multiprocessing import Process, Value
-from fixtures.zenith_fixtures import WalAcceptorFactory, ZenithPageserver, PostgresFactory
+from fixtures.zenith_fixtures import ZenithEnvBuilder
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
+
 # Check that dead minority doesn't prevent the commits: execute insert n_inserts
 # times, with fault_probability chance of getting a wal acceptor down or up
 # along the way. 2 of 3 are always alive, so the work keeps going.
-def test_pageserver_restart(zenith_cli, pageserver: ZenithPageserver, postgres: PostgresFactory, wa_factory: WalAcceptorFactory):
-
+def test_pageserver_restart(zenith_env_builder: ZenithEnvBuilder):
     # One safekeeper is enough for this test.
-    wa_factory.start_n_new(1)
+    zenith_env_builder.num_safekeepers = 1
+    env = zenith_env_builder.init()
 
-    zenith_cli.run(["branch", "test_pageserver_restart", "empty"])
-    pg = postgres.create_start('test_pageserver_restart',
-                               wal_acceptors=wa_factory.get_connstrs())
+    env.zenith_cli(["branch", "test_pageserver_restart", "main"])
+    pg = env.postgres.create_start('test_pageserver_restart')
 
     pg_conn = pg.connect()
     cur = pg_conn.cursor()
@@ -40,14 +41,14 @@ def test_pageserver_restart(zenith_cli, pageserver: ZenithPageserver, postgres:
         from pg_settings where name = 'shared_buffers'
     ''')
     row = cur.fetchone()
-    print("shared_buffers is {}, table size {}", row[0], row[1]);
+    log.info(f"shared_buffers is {row[0]}, table size {row[1]}")
     assert int(row[0]) < int(row[1])
 
     # Stop and restart pageserver. This is a more or less graceful shutdown, although
     # the page server doesn't currently have a shutdown routine so there's no difference
     # between stopping and crashing.
-    pageserver.stop();
-    pageserver.start();
+    env.pageserver.stop()
+    env.pageserver.start()
 
     # Stopping the pageserver breaks the connection from the postgres backend to
     # the page server, and causes the next query on the connection to fail. Start a new
@@ -61,6 +62,5 @@ def test_pageserver_restart(zenith_cli, pageserver: ZenithPageserver, postgres:
     assert cur.fetchone() == (100000, )
 
     # Stop the page server by force, and restart it
-    pageserver.stop();
-    pageserver.start();
-
+    env.pageserver.stop()
+    env.pageserver.start()

test_runner/batch_others/test_parallel_copy.py

@@ -0,0 +1,54 @@
+from io import BytesIO
+import asyncio
+import asyncpg
+import subprocess
+from fixtures.zenith_fixtures import ZenithEnv, Postgres
+from fixtures.log_helper import log
+
+pytest_plugins = ("fixtures.zenith_fixtures")
+
+
+async def repeat_bytes(buf, repetitions: int):
+    for i in range(repetitions):
+        yield buf
+
+
+async def copy_test_data_to_table(pg: Postgres, worker_id: int, table_name: str):
+    buf = BytesIO()
+    for i in range(1000):
+        buf.write(
+            f"{i}\tLoaded by worker {worker_id}. Long string to consume some space.\n".encode())
+    buf.seek(0)
+
+    copy_input = repeat_bytes(buf.read(), 5000)
+
+    pg_conn = await pg.connect_async()
+    await pg_conn.copy_to_table(table_name, source=copy_input)
+
+
+async def parallel_load_same_table(pg: Postgres, n_parallel: int):
+    workers = []
+    for worker_id in range(n_parallel):
+        worker = copy_test_data_to_table(pg, worker_id, f'copytest')
+        workers.append(asyncio.create_task(worker))
+
+    # await all workers
+    await asyncio.gather(*workers)
+
+
+# Load data into one table with COPY TO from 5 parallel connections
+def test_parallel_copy(zenith_simple_env: ZenithEnv, n_parallel=5):
+    env = zenith_simple_env
+    # Create a branch for us
+    env.zenith_cli(["branch", "test_parallel_copy", "empty"])
+
+    pg = env.postgres.create_start('test_parallel_copy')
+    log.info("postgres is running on 'test_parallel_copy' branch")
+
+    # Create test table
+    conn = pg.connect()
+    cur = conn.cursor()
+    cur.execute(f'CREATE TABLE copytest (i int, t text)')
+
+    # Run COPY TO to load the table with parallel connections.
+    asyncio.run(parallel_load_same_table(pg, n_parallel))

test_runner/batch_others/test_pgbench.py

@@ -1,14 +1,16 @@
-from fixtures.zenith_fixtures import PostgresFactory
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
 
 pytest_plugins = ("fixtures.zenith_fixtures")
 
 
-def test_pgbench(postgres: PostgresFactory, pg_bin, zenith_cli):
+def test_pgbench(zenith_simple_env: ZenithEnv, pg_bin):
+    env = zenith_simple_env
     # Create a branch for us
-    zenith_cli.run(["branch", "test_pgbench", "empty"])
+    env.zenith_cli(["branch", "test_pgbench", "empty"])
 
-    pg = postgres.create_start('test_pgbench')
-    print("postgres is running on 'test_pgbench' branch")
+    pg = env.postgres.create_start('test_pgbench')
+    log.info("postgres is running on 'test_pgbench' branch")
 
     connstr = pg.connstr()
 

test_runner/batch_others/test_readonly_node.py

@@ -0,0 +1,91 @@
+import pytest
+from fixtures.log_helper import log
+from fixtures.zenith_fixtures import ZenithEnv
+
+pytest_plugins = ("fixtures.zenith_fixtures")
+
+
+#
+# Create read-only compute nodes, anchored at historical points in time.
+#
+# This is very similar to the 'test_branch_behind' test, but instead of
+# creating branches, creates read-only nodes.
+#
+def test_readonly_node(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    env.zenith_cli(["branch", "test_readonly_node", "empty"])
+
+    pgmain = env.postgres.create_start('test_readonly_node')
+    log.info("postgres is running on 'test_readonly_node' branch")
+
+    main_pg_conn = pgmain.connect()
+    main_cur = main_pg_conn.cursor()
+
+    # Create table, and insert the first 100 rows
+    main_cur.execute('CREATE TABLE foo (t text)')
+
+    main_cur.execute('''
+        INSERT INTO foo
+            SELECT 'long string to consume some space' || g
+            FROM generate_series(1, 100) g
+    ''')
+    main_cur.execute('SELECT pg_current_wal_insert_lsn()')
+    lsn_a = main_cur.fetchone()[0]
+    log.info('LSN after 100 rows: ' + lsn_a)
+
+    # Insert some more rows. (This generates enough WAL to fill a few segments.)
+    main_cur.execute('''
+        INSERT INTO foo
+            SELECT 'long string to consume some space' || g
+            FROM generate_series(1, 200000) g
+    ''')
+    main_cur.execute('SELECT pg_current_wal_insert_lsn()')
+    lsn_b = main_cur.fetchone()[0]
+    log.info('LSN after 200100 rows: ' + lsn_b)
+
+    # Insert many more rows. This generates enough WAL to fill a few segments.
+    main_cur.execute('''
+        INSERT INTO foo
+            SELECT 'long string to consume some space' || g
+            FROM generate_series(1, 200000) g
+    ''')
+
+    main_cur.execute('SELECT pg_current_wal_insert_lsn()')
+    lsn_c = main_cur.fetchone()[0]
+    log.info('LSN after 400100 rows: ' + lsn_c)
+
+    # Create first read-only node at the point where only 100 rows were inserted
+    pg_hundred = env.postgres.create_start("test_readonly_node_hundred",
+                                           branch=f'test_readonly_node@{lsn_a}')
+
+    # And another at the point where 200100 rows were inserted
+    pg_more = env.postgres.create_start("test_readonly_node_more",
+                                        branch=f'test_readonly_node@{lsn_b}')
+
+    # On the 'hundred' node, we should see only 100 rows
+    hundred_pg_conn = pg_hundred.connect()
+    hundred_cur = hundred_pg_conn.cursor()
+    hundred_cur.execute('SELECT count(*) FROM foo')
+    assert hundred_cur.fetchone() == (100, )
+
+    # On the 'more' node, we should see 100200 rows
+    more_pg_conn = pg_more.connect()
+    more_cur = more_pg_conn.cursor()
+    more_cur.execute('SELECT count(*) FROM foo')
+    assert more_cur.fetchone() == (200100, )
+
+    # All the rows are visible on the main branch
+    main_cur.execute('SELECT count(*) FROM foo')
+    assert main_cur.fetchone() == (400100, )
+
+    # Check creating a node at segment boundary
+    pg = env.postgres.create_start("test_branch_segment_boundary",
+                                   branch="test_readonly_node@0/3000000")
+    cur = pg.connect().cursor()
+    cur.execute('SELECT 1')
+    assert cur.fetchone() == (1, )
+
+    # Create node at pre-initdb lsn
+    with pytest.raises(Exception, match="invalid basebackup lsn"):
+        # compute node startup with invalid LSN should fail
+        env.zenith_cli(["pg", "start", "test_readonly_node_preinitdb", "test_readonly_node@0/42"])