pageserver: enable setting a target disk range

## Problem

There was an edge case where initial logical size calculation can be
downloading a layer that wasn't hit by the test's `SELECT`, and it's
on-disk but still marked as remote in the pageserver's internal state,
so evicting it fails.


https://neon-github-public-dev.s3.amazonaws.com/reports/pr-5648/6630099807/index.html#categories/dee044ec96f666edb90a77c01099a941/e38e97a2735ffa8c/

## Summary of changes

Use pageserver API to learn about layers, instead of inspecting local
disk, so that we will always agree with the pageserver about which layer
are local.

.github/actions/allure-report-generate/action.yml

@@ -76,8 +76,8 @@ runs:
           rm -f ${ALLURE_ZIP}
         fi
       env:
-        ALLURE_VERSION: 2.23.1
-        ALLURE_ZIP_SHA256: 11141bfe727504b3fd80c0f9801eb317407fd0ac983ebb57e671f14bac4bcd86
+        ALLURE_VERSION: 2.24.0
+        ALLURE_ZIP_SHA256: 60b1d6ce65d9ef24b23cf9c2c19fd736a123487c38e54759f1ed1a7a77353c90
 
     # Potentially we could have several running build for the same key (for example, for the main branch), so we use improvised lock for this
     - name: Acquire lock

.github/workflows/approved-for-ci-run.yml

@@ -16,21 +16,29 @@ on:
       # Actual magic happens here:
       - labeled
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
+
 env:
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PR_NUMBER: ${{ github.event.pull_request.number }}
   BRANCH: "ci-run/pr-${{ github.event.pull_request.number }}"
 
-permissions: write-all
+# No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
+permissions: {}
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
 
 jobs:
   remove-label:
     # Remove `approved-for-ci-run` label if the workflow is triggered by changes in a PR.
     # The PR should be reviewed and labelled manually again.
 
+    permissions:
+      pull-requests: write # For `gh pr edit`
+
     if: |
       contains(fromJSON('["opened", "synchronize", "reopened", "closed"]'), github.event.action) &&
       contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
@@ -43,6 +51,10 @@ jobs:
   create-or-update-pr-for-ci-run:
     # Create local PR for an `approved-for-ci-run` labelled PR to run CI pipeline in it.
 
+    permissions:
+      pull-requests: write # for `gh pr edit`
+      # For `git push` and `gh pr create` we use CI_ACCESS_TOKEN
+
     if: |
       github.event.action == 'labeled' &&
       contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
@@ -75,7 +87,7 @@ jobs:
             Feel free to review/comment/discuss the original PR #${PR_NUMBER}.
           EOF
 
-          ALREADY_CREATED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${HEAD} --base main --json number --jq '.[].number')"
+          ALREADY_CREATED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${BRANCH} --base main --json number --jq '.[].number')"
           if [ -z "${ALREADY_CREATED}" ]; then
             gh pr --repo "${GITHUB_REPOSITORY}" create --title "CI run for PR #${PR_NUMBER}" \
                                                        --body-file "body.md" \
@@ -87,6 +99,10 @@ jobs:
   cleanup:
     # Close PRs and delete branchs if the original PR is closed.
 
+    permissions:
+      contents: write # for `--delete-branch` flag in `gh pr close`
+      pull-requests: write # for `gh pr close`
+
     if: |
       github.event.action == 'closed' &&
       github.event.pull_request.head.repo.full_name != github.repository
@@ -94,8 +110,9 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-      - run: |
-          CLOSED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${HEAD} --json 'closed' --jq '.[].closed')"
+      - name: Close PR and delete `ci-run/pr-${{ env.PR_NUMBER }}` branch
+        run: |
+          CLOSED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${BRANCH} --json 'closed' --jq '.[].closed')"
           if [ "${CLOSED}" == "false" ]; then
             gh pr --repo "${GITHUB_REPOSITORY}" close "${BRANCH}" --delete-branch
           fi

.github/workflows/build_and_test.yml

@@ -320,6 +320,9 @@ jobs:
       - name: Build neon extensions
         run: mold -run make neon-pg-ext -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: mold -run make walproposer-lib -j$(nproc)
+
       - name: Run cargo build
         run: |
           ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
@@ -335,6 +338,16 @@ jobs:
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
 
+          # Run separate tests for real Azure Blob Storage
+          # XXX: replace region with `eu-central-1`-like region
+          export AZURE_STORAGE_ACCOUNT="${{ secrets.AZURE_STORAGE_ACCOUNT_DEV }}"
+          export AZURE_STORAGE_ACCESS_KEY="${{ secrets.AZURE_STORAGE_ACCESS_KEY_DEV }}"
+          export ENABLE_REAL_AZURE_REMOTE_STORAGE=y
+          export REMOTE_STORAGE_AZURE_CONTAINER=neon-github-sandbox
+          export REMOTE_STORAGE_AZURE_REGION=eastus2
+          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_azure
+
       - name: Install rust binaries
         run: |
           # Install target binaries
@@ -834,7 +847,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.17.5
+      VM_BUILDER_VERSION: v0.18.2
 
     steps:
       - name: Checkout
@@ -1091,7 +1104,10 @@ jobs:
           GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
         run: |
           if [[ "$GITHUB_REF_NAME" == "main" ]]; then
-            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}}
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=false
+
+            # TODO: move deployPreprodRegion to release (`"$GITHUB_REF_NAME" == "release"` block), once Staging support different compute tag prefixes for different regions
+            gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=true
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
             gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f disclamerAcknowledged=true
           else

.github/workflows/neon_extra_builds.yml

@@ -32,7 +32,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -90,18 +90,21 @@ jobs:
 
       - name: Build postgres v14
         if: steps.cache_pg_14.outputs.cache-hit != 'true'
-        run: make postgres-v14 -j$(nproc)
+        run: make postgres-v14 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v15
         if: steps.cache_pg_15.outputs.cache-hit != 'true'
-        run: make postgres-v15 -j$(nproc)
+        run: make postgres-v15 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v16
         if: steps.cache_pg_16.outputs.cache-hit != 'true'
-        run: make postgres-v16 -j$(nproc)
+        run: make postgres-v16 -j$(sysctl -n hw.ncpu)
 
       - name: Build neon extensions
-        run: make neon-pg-ext -j$(nproc)
+        run: make neon-pg-ext -j$(sysctl -n hw.ncpu)
+
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(sysctl -n hw.ncpu)
 
       - name: Run cargo build
         run: cargo build --all --release
@@ -126,7 +129,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -135,6 +138,9 @@ jobs:
       - name: Get postgres headers
         run: make postgres-headers -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(nproc)
+
       - name: Produce the build stats
         run: cargo build --all --release --timings
 

CODEOWNERS

@@ -5,7 +5,7 @@
 /libs/remote_storage/ @neondatabase/storage
 /libs/safekeeper_api/ @neondatabase/safekeepers
 /libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
-/pageserver/ @neondatabase/compute @neondatabase/storage
+/pageserver/ @neondatabase/storage
 /pgxn/ @neondatabase/compute
 /proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers

Cargo.toml

@@ -26,6 +26,7 @@ members = [
     "libs/tracing-utils",
     "libs/postgres_ffi/wal_craft",
     "libs/vm_monitor",
+    "libs/walproposer",
 ]
 
 [workspace.package]
@@ -36,6 +37,10 @@ license = "Apache-2.0"
 [workspace.dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
 async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
+azure_core = "0.16"
+azure_identity = "0.16"
+azure_storage = "0.16"
+azure_storage_blobs = "0.16"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
@@ -51,6 +56,7 @@ bindgen = "0.65"
 bstr = "1.0"
 byteorder = "1.4"
 bytes = "1.0"
+camino = "1.1.6"
 cfg-if = "1.0.0"
 chrono = { version = "0.4", default-features = false, features = ["clock"] }
 clap = { version = "4.0", features = ["derive"] }
@@ -75,10 +81,11 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http-types = "2"
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
-hyper-tungstenite = "0.9"
+hyper-tungstenite = "0.11"
 inotify = "0.10.2"
 itertools = "0.10"
 jsonwebtoken = "8"
@@ -107,6 +114,7 @@ reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
+rustc-hash = "1.1.0"
 rustls = "0.21"
 rustls-pemfile = "1"
 rustls-split = "0.3"
@@ -153,11 +161,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -178,6 +186,7 @@ tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
+walproposer = { version = "0.1", path = "./libs/walproposer/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
@@ -186,14 +195,14 @@ workspace_hack = { version = "0.1", path = "./workspace_hack/" }
 criterion = "0.5.1"
 rcgen = "0.11"
 rstest = "0.18"
-tempfile = "3.4"
+camino-tempfile = "1.0.2"
 tonic-build = "0.9"
 
 [patch.crates-io]
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
 
 ################# Binary contents sections
 

Dockerfile.compute-node

@@ -224,8 +224,8 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.7.tar.gz -
 FROM build-deps AS vector-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.0.tar.gz -O pgvector.tar.gz && \
-    echo "d8aa3504b215467ca528525a6de12c3f85f9891b091ce0e5864dd8a9b757f77b pgvector.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.1.tar.gz -O pgvector.tar.gz && \
+    echo "cc7a8e034a96e30a819911ac79d32f6bc47bdd1aa2de4d7d4904e26b83209dc8 pgvector.tar.gz" | sha256sum --check && \
     mkdir pgvector-src && cd pgvector-src && tar xvzf ../pgvector.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -368,8 +368,8 @@ RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.18.tar
 FROM build-deps AS plpgsql-check-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.4.0.tar.gz -O plpgsql_check.tar.gz && \
-    echo "9ba58387a279b35a3bfa39ee611e5684e6cddb2ba046ddb2c5190b3bd2ca254a plpgsql_check.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.5.3.tar.gz -O plpgsql_check.tar.gz && \
+    echo "6631ec3e7fb3769eaaf56e3dfedb829aa761abf163d13dba354b4c218508e1c0 plpgsql_check.tar.gz" | sha256sum --check && \
     mkdir plpgsql_check-src && cd plpgsql_check-src && tar xvzf ../plpgsql_check.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
@@ -589,8 +589,7 @@ RUN case "${PG_VERSION}" in \
     echo "${PG_EMBEDDING_CHECKSUM} pg_embedding.tar.gz" | sha256sum --check && \
     mkdir pg_embedding-src && cd pg_embedding-src && tar xvzf ../pg_embedding.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
-    make -j $(getconf _NPROCESSORS_ONLN) install && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/embedding.control
+    make -j $(getconf _NPROCESSORS_ONLN) install
 
 #########################################################################################
 #
@@ -616,11 +615,7 @@ RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgre
 #########################################################################################
 #
 # Layer "rust extensions"
-# This layer is used to build `pgx` deps
-#
-# FIXME: This needs to be updated to latest version of 'pgrx' (it was renamed from
-# 'pgx' to 'pgrx') for PostgreSQL 16. And that in turn requires bumping the pgx
-# dependency on all the rust extension that depend on it, too.
+# This layer is used to build `pgrx` deps
 #
 #########################################################################################
 FROM build-deps AS rust-extensions-build
@@ -636,22 +631,12 @@ USER nonroot
 WORKDIR /home/nonroot
 ARG PG_VERSION
 
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version ${PG_VERSION}" && exit 1 \
-        ;; \
-    esac && \
-    curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
     chmod +x rustup-init && \
     ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
     rm rustup-init && \
-    cargo install --locked --version 0.7.3 cargo-pgx && \
-    /bin/bash -c 'cargo pgx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
+    cargo install --locked --version 0.10.2 cargo-pgrx && \
+    /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
 
 USER root
 
@@ -665,23 +650,11 @@ USER root
 FROM rust-extensions-build AS pg-jsonschema-pg-build
 ARG PG_VERSION
 
-# caeab60d70b2fd3ae421ec66466a3abbb37b7ee6 made on 06/03/2023
-# there is no release tag yet, but we need it due to the superuser fix in the control file, switch to git tag after release >= 0.1.5
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version \"${PG_VERSION}\"" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/supabase/pg_jsonschema/archive/caeab60d70b2fd3ae421ec66466a3abbb37b7ee6.tar.gz -O pg_jsonschema.tar.gz && \
-    echo "54129ce2e7ee7a585648dbb4cef6d73f795d94fe72f248ac01119992518469a4 pg_jsonschema.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v0.2.0.tar.gz -O pg_jsonschema.tar.gz && \
+    echo "9118fc508a6e231e7a39acaa6f066fcd79af17a5db757b47d2eefbe14f7794f0 pg_jsonschema.tar.gz" | sha256sum --check && \
     mkdir pg_jsonschema-src && cd pg_jsonschema-src && tar xvzf ../pg_jsonschema.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx = "0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    cargo pgx install --release && \
+    sed -i 's/pgrx = "0.10.2"/pgrx = { version = "0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_jsonschema.control
 
 #########################################################################################
@@ -694,26 +667,11 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-graphql-pg-build
 ARG PG_VERSION
 
-# b4988843647450a153439be367168ed09971af85 made on 22/02/2023 (from remove-pgx-contrib-spiext branch)
-# Currently pgx version bump to >= 0.7.2  causes "call to unsafe function" compliation errors in
-# pgx-contrib-spiext. There is a branch that removes that dependency, so use it. It is on the
-# same 1.1 version we've used before.
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/yrashk/pg_graphql/archive/b4988843647450a153439be367168ed09971af85.tar.gz -O pg_graphql.tar.gz && \
-    echo "0c7b0e746441b2ec24187d0e03555faf935c2159e2839bddd14df6dafbc8c9bd pg_graphql.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/supabase/pg_graphql/archive/refs/tags/v1.4.0.tar.gz -O pg_graphql.tar.gz && \
+    echo "bd8dc7230282b3efa9ae5baf053a54151ed0e66881c7c53750e2d0c765776edc pg_graphql.tar.gz" | sha256sum --check && \
     mkdir pg_graphql-src && cd pg_graphql-src && tar xvzf ../pg_graphql.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx = "~0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    sed -i 's/pgx-tests = "~0.7.1"/pgx-tests = "0.7.3"/g' Cargo.toml && \
-    cargo pgx install --release && \
+    sed -i 's/pgrx = "=0.10.2"/pgrx = { version = "0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
     # it's needed to enable extension because it uses untrusted C language
     sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_graphql.control && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_graphql.control
@@ -728,21 +686,11 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-tiktoken-pg-build
 ARG PG_VERSION
 
-# 801f84f08c6881c8aa30f405fafbf00eec386a72 made on 10/03/2023
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/kelvich/pg_tiktoken/archive/801f84f08c6881c8aa30f405fafbf00eec386a72.tar.gz -O pg_tiktoken.tar.gz && \
-    echo "52f60ac800993a49aa8c609961842b611b6b1949717b69ce2ec9117117e16e4a pg_tiktoken.tar.gz" | sha256sum --check && \
+# 26806147b17b60763039c6a6878884c41a262318 made on 26/09/2023
+RUN wget https://github.com/kelvich/pg_tiktoken/archive/26806147b17b60763039c6a6878884c41a262318.tar.gz -O pg_tiktoken.tar.gz && \
+    echo "e64e55aaa38c259512d3e27c572da22c4637418cf124caba904cd50944e5004e pg_tiktoken.tar.gz" | sha256sum --check && \
     mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xvzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \
-    cargo pgx install --release && \
+    cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_tiktoken.control
 
 #########################################################################################
@@ -755,21 +703,15 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-pgx-ulid-build
 ARG PG_VERSION
 
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.0.tar.gz -O pgx_ulid.tar.gz && \
-    echo "908b7358e6f846e87db508ae5349fb56a88ee6305519074b12f3d5b0ff09f791 pgx_ulid.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.3.tar.gz -O pgx_ulid.tar.gz && \
+    echo "ee5db82945d2d9f2d15597a80cf32de9dca67b897f605beb830561705f12683c pgx_ulid.tar.gz" | sha256sum --check && \
     mkdir pgx_ulid-src && cd pgx_ulid-src && tar xvzf ../pgx_ulid.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx        = "=0.7.3"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    cargo pgx install --release && \
+    echo "******************* Apply a patch for Postgres 16 support; delete in the next release ******************" && \
+    wget https://github.com/pksunkara/pgx_ulid/commit/f84954cf63fc8c80d964ac970d9eceed3c791196.patch && \
+    patch -p1 < f84954cf63fc8c80d964ac970d9eceed3c791196.patch && \
+    echo "********************************************************************************************************" && \
+    sed -i 's/pgrx       = "=0.10.2"/pgrx = { version = "=0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/ulid.control
 
 #########################################################################################

Makefile

@@ -62,7 +62,7 @@ all: neon postgres neon-pg-ext
 #
 # The 'postgres_ffi' depends on the Postgres headers.
 .PHONY: neon
-neon: postgres-headers
+neon: postgres-headers walproposer-lib
 	+@echo "Compiling Neon"
 	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS)
 
@@ -153,18 +153,6 @@ neon-pg-ext-%: postgres-%
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile install
 
-# pg_embedding was temporarily released as hnsw from this repo, when we only
-# supported PostgreSQL 14 and 15
-neon-pg-ext-v14: neon-pg-ext-hnsw-v14
-neon-pg-ext-v15: neon-pg-ext-hnsw-v15
-
-neon-pg-ext-hnsw-%: postgres-headers-% postgres-%
-	+@echo "Compiling hnsw $*"
-	mkdir -p $(POSTGRES_INSTALL_DIR)/build/hnsw-$*
-	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
-		-C $(POSTGRES_INSTALL_DIR)/build/hnsw-$* \
-		-f $(ROOT_PROJECT_DIR)/pgxn/hnsw/Makefile install
-
 .PHONY: neon-pg-ext-clean-%
 neon-pg-ext-clean-%:
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
@@ -179,9 +167,42 @@ neon-pg-ext-clean-%:
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
 	-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 	-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile clean
-	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
-	-C $(POSTGRES_INSTALL_DIR)/build/hnsw-$* \
-	-f $(ROOT_PROJECT_DIR)/pgxn/hnsw/Makefile clean
+
+# Build walproposer as a static library. walproposer source code is located
+# in the pgxn/neon directory.
+# 
+# We also need to include libpgport.a and libpgcommon.a, because walproposer
+# uses some functions from those libraries.
+# 
+# Some object files are removed from libpgport.a and libpgcommon.a because
+# they depend on openssl and other libraries that are not included in our
+# Rust build.
+.PHONY: walproposer-lib
+walproposer-lib: neon-pg-ext-v16
+	+@echo "Compiling walproposer-lib"
+	mkdir -p $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgport.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgcommon.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+ifeq ($(UNAME_S),Linux)
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgport.a \
+		pg_strong_random.o
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgcommon.a \
+		pg_crc32c.o \
+		hmac_openssl.o \
+		cryptohash_openssl.o \
+		scram-common.o \
+		md5_common.o \
+		checksum_helper.o
+endif
+
+.PHONY: walproposer-lib-clean
+walproposer-lib-clean:
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile clean
 
 .PHONY: neon-pg-ext
 neon-pg-ext: \

NOTICE

@@ -1,5 +1,5 @@
 Neon
 Copyright 2022 Neon Inc.
 
-The PostgreSQL submodules in vendor/postgres-v14 and vendor/postgres-v15 are licensed under the
-PostgreSQL license. See vendor/postgres-v14/COPYRIGHT and vendor/postgres-v15/COPYRIGHT.
+The PostgreSQL submodules in vendor/ are licensed under the PostgreSQL license.
+See vendor/postgres-vX/COPYRIGHT for details.

README.md

@@ -29,13 +29,13 @@ See developer documentation in [SUMMARY.md](/docs/SUMMARY.md) for more informati
 ```bash
 apt install build-essential libtool libreadline-dev zlib1g-dev flex bison libseccomp-dev \
 libssl-dev clang pkg-config libpq-dev cmake postgresql-client protobuf-compiler \
-libcurl4-openssl-dev openssl python-poetry lsof
+libcurl4-openssl-dev openssl python-poetry lsof libicu-dev
 ```
 * On Fedora, these packages are needed:
 ```bash
 dnf install flex bison readline-devel zlib-devel openssl-devel \
   libseccomp-devel perl clang cmake postgresql postgresql-contrib protobuf-compiler \
-  protobuf-devel libcurl-devel openssl poetry lsof
+  protobuf-devel libcurl-devel openssl poetry lsof libicu-devel
 ```
 * On Arch based systems, these packages are needed:
 ```bash

compute_tools/src/compute.rs

@@ -252,7 +252,7 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
                     IF NOT EXISTS (
                         SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
                     THEN
-                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN IN ROLE pg_read_all_data, pg_write_all_data;
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION IN ROLE pg_read_all_data, pg_write_all_data;
                         IF array_length(roles, 1) IS NOT NULL THEN
                             EXECUTE format('GRANT neon_superuser TO %s',
                                            array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
@@ -692,10 +692,11 @@ impl ComputeNode {
         // Proceed with post-startup configuration. Note, that order of operations is important.
         let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
         create_neon_superuser(spec, &mut client)?;
+        cleanup_instance(&mut client)?;
         handle_roles(spec, &mut client)?;
         handle_databases(spec, &mut client)?;
         handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
-        handle_grants(spec, self.connstr.as_str())?;
+        handle_grants(spec, &mut client, self.connstr.as_str())?;
         handle_extensions(spec, &mut client)?;
         create_availability_check_data(&mut client)?;
 
@@ -731,10 +732,11 @@ impl ComputeNode {
         // Disable DDL forwarding because control plane already knows about these roles/databases.
         if spec.mode == ComputeMode::Primary {
             client.simple_query("SET neon.forward_ddl = false")?;
+            cleanup_instance(&mut client)?;
             handle_roles(&spec, &mut client)?;
             handle_databases(&spec, &mut client)?;
             handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, self.connstr.as_str())?;
+            handle_grants(&spec, &mut client, self.connstr.as_str())?;
             handle_extensions(&spec, &mut client)?;
         }
 
@@ -1039,7 +1041,7 @@ LIMIT 100",
         let remote_extensions = spec
             .remote_extensions
             .as_ref()
-            .ok_or(anyhow::anyhow!("Remote extensions are not configured",))?;
+            .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
 
         info!("parse shared_preload_libraries from spec.cluster.settings");
         let mut libs_vec = Vec::new();

compute_tools/src/monitor.rs

@@ -1,5 +1,5 @@
 use std::sync::Arc;
-use std::{thread, time};
+use std::{thread, time::Duration};
 
 use chrono::{DateTime, Utc};
 use postgres::{Client, NoTls};
@@ -7,7 +7,7 @@ use tracing::{debug, info};
 
 use crate::compute::ComputeNode;
 
-const MONITOR_CHECK_INTERVAL: u64 = 500; // milliseconds
+const MONITOR_CHECK_INTERVAL: Duration = Duration::from_millis(500);
 
 // Spin in a loop and figure out the last activity time in the Postgres.
 // Then update it in the shared state. This function never errors out.
@@ -17,13 +17,12 @@ fn watch_compute_activity(compute: &ComputeNode) {
     let connstr = compute.connstr.as_str();
     // Define `client` outside of the loop to reuse existing connection if it's active.
     let mut client = Client::connect(connstr, NoTls);
-    let timeout = time::Duration::from_millis(MONITOR_CHECK_INTERVAL);
 
     info!("watching Postgres activity at {}", connstr);
 
     loop {
         // Should be outside of the write lock to allow others to read while we sleep.
-        thread::sleep(timeout);
+        thread::sleep(MONITOR_CHECK_INTERVAL);
 
         match &mut client {
             Ok(cli) => {

compute_tools/src/pg_helpers.rs

@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::fmt::Write;
 use std::fs;
 use std::fs::File;
@@ -205,22 +206,37 @@ pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
 }
 
 /// Build a list of existing Postgres databases
-pub fn get_existing_dbs(client: &mut Client) -> Result<Vec<Database>> {
-    let postgres_dbs = client
+pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>> {
+    // `pg_database.datconnlimit = -2` means that the database is in the
+    // invalid state. See:
+    //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+    let postgres_dbs: Vec<Database> = client
         .query(
-            "SELECT datname, datdba::regrole::text as owner
-               FROM pg_catalog.pg_database;",
+            "SELECT
+                datname AS name,
+                datdba::regrole::text AS owner,
+                NOT datallowconn AS restrict_conn,
+                datconnlimit = - 2 AS invalid
+            FROM
+                pg_catalog.pg_database;",
             &[],
         )?
         .iter()
         .map(|row| Database {
-            name: row.get("datname"),
+            name: row.get("name"),
             owner: row.get("owner"),
+            restrict_conn: row.get("restrict_conn"),
+            invalid: row.get("invalid"),
             options: None,
         })
         .collect();
 
-    Ok(postgres_dbs)
+    let dbs_map = postgres_dbs
+        .iter()
+        .map(|db| (db.name.clone(), db.clone()))
+        .collect::<HashMap<_, _>>();
+
+    Ok(dbs_map)
 }
 
 /// Wait for Postgres to become ready to accept connections. It's ready to

compute_tools/src/spec.rs

@@ -13,7 +13,7 @@ use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
-use compute_api::spec::{ComputeSpec, Database, PgIdent, Role};
+use compute_api::spec::{ComputeSpec, PgIdent, Role};
 
 // Do control plane request and return response if any. In case of error it
 // returns a bool flag indicating whether it makes sense to retry the request
@@ -161,6 +161,38 @@ pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Compute could be unexpectedly shut down, for example, during the
+/// database dropping. This leaves the database in the invalid state,
+/// which prevents new db creation with the same name. This function
+/// will clean it up before proceeding with catalog updates. All
+/// possible future cleanup operations may go here too.
+#[instrument(skip_all)]
+pub fn cleanup_instance(client: &mut Client) -> Result<()> {
+    let existing_dbs = get_existing_dbs(client)?;
+
+    for (_, db) in existing_dbs {
+        if db.invalid {
+            // After recent commit in Postgres, interrupted DROP DATABASE
+            // leaves the database in the invalid state. According to the
+            // commit message, the only option for user is to drop it again.
+            // See:
+            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+            //
+            // Postgres Neon extension is done the way, that db is de-registered
+            // in the control plane metadata only after it is dropped. So there is
+            // a chance that it still thinks that db should exist. This means
+            // that it will be re-created by `handle_databases()`. Yet, it's fine
+            // as user can just repeat drop (in vanilla Postgres they would need
+            // to do the same, btw).
+            let query = format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote());
+            info!("dropping invalid database {}", db.name);
+            client.execute(query.as_str(), &[])?;
+        }
+    }
+
+    Ok(())
+}
+
 /// Given a cluster spec json and open transaction it handles roles creation,
 /// deletion and update.
 #[instrument(skip_all)]
@@ -270,7 +302,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
             }
             RoleAction::Create => {
                 let mut query: String = format!(
-                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
+                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -379,13 +411,13 @@ fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent
 /// which together provide us idempotency.
 #[instrument(skip_all)]
 pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Print a list of existing Postgres databases (only in debug mode)
     if span_enabled!(Level::INFO) {
         info!("postgres databases:");
-        for r in &existing_dbs {
-            info!("    {}:{}", r.name, r.owner);
+        for (dbname, db) in &existing_dbs {
+            info!("    {}:{}", dbname, db.owner);
         }
     }
 
@@ -439,8 +471,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 "rename_db" => {
                     let new_name = op.new_name.as_ref().unwrap();
 
-                    // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-                    if existing_dbs.iter().any(|r| r.name == op.name) {
+                    if existing_dbs.get(&op.name).is_some() {
                         let query: String = format!(
                             "ALTER DATABASE {} RENAME TO {}",
                             op.name.pg_quote(),
@@ -457,14 +488,12 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
     }
 
     // Refresh Postgres databases info to handle possible renames
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     info!("cluster spec databases:");
     for db in &spec.cluster.databases {
         let name = &db.name;
-
-        // XXX: with a limited number of databases it is fine, but consider making it a HashMap
-        let pg_db = existing_dbs.iter().find(|r| r.name == *name);
+        let pg_db = existing_dbs.get(name);
 
         enum DatabaseAction {
             None,
@@ -530,13 +559,32 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
-    info!("cluster spec grants:");
+pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) -> Result<()> {
+    info!("modifying database permissions");
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Do some per-database access adjustments. We'd better do this at db creation time,
     // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
     // atomically.
     for db in &spec.cluster.databases {
+        match existing_dbs.get(&db.name) {
+            Some(pg_db) => {
+                if pg_db.restrict_conn || pg_db.invalid {
+                    info!(
+                        "skipping grants for db {} (invalid: {}, connections not allowed: {})",
+                        db.name, pg_db.invalid, pg_db.restrict_conn
+                    );
+                    continue;
+                }
+            }
+            None => {
+                bail!(
+                    "database {} doesn't exist in Postgres after handle_databases()",
+                    db.name
+                );
+            }
+        }
+
         let mut conf = Config::from_str(connstr)?;
         conf.dbname(&db.name);
 
@@ -575,6 +623,11 @@ pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
 
         // Explicitly grant CREATE ON SCHEMA PUBLIC to the web_access user.
         // This is needed because since postgres 15 this privilege is removed by default.
+        // TODO: web_access isn't created for almost 1 year. It could be that we have
+        // active users of 1 year old projects, but hopefully not, so check it and
+        // remove this code if possible. The worst thing that could happen is that
+        // user won't be able to use public schema in NEW databases created in the
+        // very OLD project.
         let grant_query = "DO $$\n\
                 BEGIN\n\
                     IF EXISTS(\n\

compute_tools/tests/pg_helpers_tests.rs

@@ -28,7 +28,7 @@ mod pg_helpers_tests {
         assert_eq!(
             spec.cluster.settings.as_pg_settings(),
             r#"fsync = off
-wal_level = replica
+wal_level = logical
 hot_standby = on
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on

control_plane/Cargo.toml

@@ -6,6 +6,7 @@ license.workspace = true
 
 [dependencies]
 anyhow.workspace = true
+camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
 git-version.workspace = true

control_plane/src/attachment_service.rs

@@ -1,5 +1,6 @@
 use crate::{background_process, local_env::LocalEnv};
 use anyhow::anyhow;
+use camino::Utf8PathBuf;
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 use std::{path::PathBuf, process::Child};
@@ -47,8 +48,9 @@ impl AttachmentService {
         }
     }
 
-    fn pid_file(&self) -> PathBuf {
-        self.env.base_data_dir.join("attachment_service.pid")
+    fn pid_file(&self) -> Utf8PathBuf {
+        Utf8PathBuf::from_path_buf(self.env.base_data_dir.join("attachment_service.pid"))
+            .expect("non-Unicode path")
     }
 
     pub fn start(&self) -> anyhow::Result<Child> {

control_plane/src/background_process.rs

@@ -16,12 +16,13 @@ use std::ffi::OsStr;
 use std::io::Write;
 use std::os::unix::prelude::AsRawFd;
 use std::os::unix::process::CommandExt;
-use std::path::{Path, PathBuf};
+use std::path::Path;
 use std::process::{Child, Command};
 use std::time::Duration;
 use std::{fs, io, thread};
 
 use anyhow::Context;
+use camino::{Utf8Path, Utf8PathBuf};
 use nix::errno::Errno;
 use nix::fcntl::{FcntlArg, FdFlag};
 use nix::sys::signal::{kill, Signal};
@@ -45,9 +46,9 @@ const NOTICE_AFTER_RETRIES: u64 = 50;
 /// it itself.
 pub enum InitialPidFile<'t> {
     /// Create a pidfile, to allow future CLI invocations to manipulate the process.
-    Create(&'t Path),
+    Create(&'t Utf8Path),
     /// The process will create the pidfile itself, need to wait for that event.
-    Expect(&'t Path),
+    Expect(&'t Utf8Path),
 }
 
 /// Start a background child process using the parameters given.
@@ -85,7 +86,7 @@ where
         .stdout(process_log_file)
         .stderr(same_file_for_stderr)
         .args(args);
-    let filled_cmd = fill_aws_secrets_vars(fill_rust_env_vars(background_command));
+    let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
     filled_cmd.envs(envs);
 
     let pid_file_to_check = match initial_pid_file {
@@ -137,7 +138,11 @@ where
 }
 
 /// Stops the process, using the pid file given. Returns Ok also if the process is already not running.
-pub fn stop_process(immediate: bool, process_name: &str, pid_file: &Path) -> anyhow::Result<()> {
+pub fn stop_process(
+    immediate: bool,
+    process_name: &str,
+    pid_file: &Utf8Path,
+) -> anyhow::Result<()> {
     let pid = match pid_file::read(pid_file)
         .with_context(|| format!("read pid_file {pid_file:?}"))?
     {
@@ -233,11 +238,13 @@ fn fill_rust_env_vars(cmd: &mut Command) -> &mut Command {
     filled_cmd
 }
 
-fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
+fn fill_remote_storage_secrets_vars(mut cmd: &mut Command) -> &mut Command {
     for env_key in [
         "AWS_ACCESS_KEY_ID",
         "AWS_SECRET_ACCESS_KEY",
         "AWS_SESSION_TOKEN",
+        "AZURE_STORAGE_ACCOUNT",
+        "AZURE_STORAGE_ACCESS_KEY",
     ] {
         if let Ok(value) = std::env::var(env_key) {
             cmd = cmd.env(env_key, value);
@@ -252,9 +259,9 @@ fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
 ///    will remain held until the cmd exits.
 fn pre_exec_create_pidfile<P>(cmd: &mut Command, path: P) -> &mut Command
 where
-    P: Into<PathBuf>,
+    P: Into<Utf8PathBuf>,
 {
-    let path: PathBuf = path.into();
+    let path: Utf8PathBuf = path.into();
     // SAFETY
     // pre_exec is marked unsafe because it runs between fork and exec.
     // Why is that dangerous in various ways?
@@ -311,7 +318,7 @@ where
 
 fn process_started<F>(
     pid: Pid,
-    pid_file_to_check: Option<&Path>,
+    pid_file_to_check: Option<&Utf8Path>,
     status_check: &F,
 ) -> anyhow::Result<bool>
 where

control_plane/src/bin/attachment_service.rs

@@ -13,6 +13,7 @@ use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
 use utils::logging::{self, LogFormat};
+use utils::signals::{ShutdownSignals, Signal};
 
 use utils::{
     http::{
@@ -223,6 +224,7 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
     if attach_req.pageserver_id.is_some() {
         tenant_state.generation += 1;
     }
+    tenant_state.pageserver = attach_req.pageserver_id;
     let generation = tenant_state.generation;
 
     locked.save().await.map_err(ApiError::InternalServerError)?;
@@ -267,7 +269,16 @@ async fn main() -> anyhow::Result<()> {
     let server = hyper::Server::from_tcp(http_listener)?.serve(service);
 
     tracing::info!("Serving on {0}", args.listen);
-    server.await?;
+
+    tokio::task::spawn(server);
+
+    ShutdownSignals::handle(|signal| match signal {
+        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
+            tracing::info!("Got {}. Terminating", signal.name());
+            // We're just a test helper: no graceful shutdown.
+            std::process::exit(0);
+        }
+    })?;
 
     Ok(())
 }

control_plane/src/bin/neon_local.rs

@@ -116,6 +116,7 @@ fn main() -> Result<()> {
             "attachment_service" => handle_attachment_service(sub_args, &env),
             "safekeeper" => handle_safekeeper(sub_args, &env),
             "endpoint" => handle_endpoint(sub_args, &env),
+            "mappings" => handle_mappings(sub_args, &mut env),
             "pg" => bail!("'pg' subcommand has been renamed to 'endpoint'"),
             _ => bail!("unexpected subcommand {sub_name}"),
         };
@@ -816,6 +817,38 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
     Ok(())
 }
 
+fn handle_mappings(sub_match: &ArgMatches, env: &mut local_env::LocalEnv) -> Result<()> {
+    let (sub_name, sub_args) = match sub_match.subcommand() {
+        Some(ep_subcommand_data) => ep_subcommand_data,
+        None => bail!("no mappings subcommand provided"),
+    };
+
+    match sub_name {
+        "map" => {
+            let branch_name = sub_args
+                .get_one::<String>("branch-name")
+                .expect("branch-name argument missing");
+
+            let tenant_id = sub_args
+                .get_one::<String>("tenant-id")
+                .map(|x| TenantId::from_str(x))
+                .expect("tenant-id argument missing")
+                .expect("malformed tenant-id arg");
+
+            let timeline_id = sub_args
+                .get_one::<String>("timeline-id")
+                .map(|x| TimelineId::from_str(x))
+                .expect("timeline-id argument missing")
+                .expect("malformed timeline-id arg");
+
+            env.register_branch_mapping(branch_name.to_owned(), tenant_id, timeline_id)?;
+
+            Ok(())
+        }
+        other => unimplemented!("mappings subcommand {other}"),
+    }
+}
+
 fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
     fn get_pageserver(env: &local_env::LocalEnv, args: &ArgMatches) -> Result<PageServerNode> {
         let node_id = if let Some(id_str) = args.get_one::<String>("pageserver-id") {
@@ -1084,6 +1117,7 @@ fn cli() -> Command {
     // --id, when using a pageserver command
     let pageserver_id_arg = Arg::new("pageserver-id")
         .long("id")
+        .global(true)
         .help("pageserver id")
         .required(false);
     // --pageserver-id when using a non-pageserver command
@@ -1254,17 +1288,20 @@ fn cli() -> Command {
             Command::new("pageserver")
                 .arg_required_else_help(true)
                 .about("Manage pageserver")
+                .arg(pageserver_id_arg)
                 .subcommand(Command::new("status"))
-                .arg(pageserver_id_arg.clone())
-                .subcommand(Command::new("start").about("Start local pageserver")
-                .arg(pageserver_id_arg.clone())
-                .arg(pageserver_config_args.clone()))
-                .subcommand(Command::new("stop").about("Stop local pageserver")
-                .arg(pageserver_id_arg.clone())
-                            .arg(stop_mode_arg.clone()))
-                .subcommand(Command::new("restart").about("Restart local pageserver")
-                .arg(pageserver_id_arg.clone())
-                .arg(pageserver_config_args.clone()))
+                .subcommand(Command::new("start")
+                    .about("Start local pageserver")
+                    .arg(pageserver_config_args.clone())
+                )
+                .subcommand(Command::new("stop")
+                    .about("Stop local pageserver")
+                    .arg(stop_mode_arg.clone())
+                )
+                .subcommand(Command::new("restart")
+                    .about("Restart local pageserver")
+                    .arg(pageserver_config_args.clone())
+                )
         )
         .subcommand(
             Command::new("attachment_service")
@@ -1321,8 +1358,8 @@ fn cli() -> Command {
                     .about("Start postgres.\n If the endpoint doesn't exist yet, it is created.")
                     .arg(endpoint_id_arg.clone())
                     .arg(tenant_id_arg.clone())
-                    .arg(branch_name_arg)
-                    .arg(timeline_id_arg)
+                    .arg(branch_name_arg.clone())
+                    .arg(timeline_id_arg.clone())
                     .arg(lsn_arg)
                     .arg(pg_port_arg)
                     .arg(http_port_arg)
@@ -1335,7 +1372,7 @@ fn cli() -> Command {
                 .subcommand(
                     Command::new("stop")
                     .arg(endpoint_id_arg)
-                    .arg(tenant_id_arg)
+                    .arg(tenant_id_arg.clone())
                     .arg(
                         Arg::new("destroy")
                             .help("Also delete data directory (now optional, should be default in future)")
@@ -1346,6 +1383,18 @@ fn cli() -> Command {
                 )
 
         )
+        .subcommand(
+            Command::new("mappings")
+                .arg_required_else_help(true)
+                .about("Manage neon_local branch name mappings")
+                .subcommand(
+                    Command::new("map")
+                        .about("Create new mapping which cannot exist already")
+                        .arg(branch_name_arg.clone())
+                        .arg(tenant_id_arg.clone())
+                        .arg(timeline_id_arg.clone())
+                )
+        )
         // Obsolete old name for 'endpoint'. We now just print an error if it's used.
         .subcommand(
             Command::new("pg")

control_plane/src/broker.rs

@@ -7,7 +7,7 @@
 //! ```
 use anyhow::Context;
 
-use std::path::PathBuf;
+use camino::Utf8PathBuf;
 
 use crate::{background_process, local_env};
 
@@ -30,7 +30,7 @@ pub fn start_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
         || {
             let url = broker.client_url();
             let status_url = url.join("status").with_context(|| {
-                format!("Failed to append /status path to broker endpoint {url}",)
+                format!("Failed to append /status path to broker endpoint {url}")
             })?;
             let request = client
                 .get(status_url)
@@ -50,6 +50,7 @@ pub fn stop_broker_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
     background_process::stop_process(true, "storage_broker", &storage_broker_pid_file_path(env))
 }
 
-fn storage_broker_pid_file_path(env: &local_env::LocalEnv) -> PathBuf {
-    env.base_data_dir.join("storage_broker.pid")
+fn storage_broker_pid_file_path(env: &local_env::LocalEnv) -> Utf8PathBuf {
+    Utf8PathBuf::from_path_buf(env.base_data_dir.join("storage_broker.pid"))
+        .expect("non-Unicode path")
 }

control_plane/src/endpoint.rs

@@ -253,7 +253,7 @@ impl Endpoint {
         conf.append("shared_buffers", "1MB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
-        conf.append("wal_level", "replica");
+        conf.append("wal_level", "logical");
         // 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");

control_plane/src/pageserver.rs

@@ -14,6 +14,7 @@ use std::process::{Child, Command};
 use std::{io, result};
 
 use anyhow::{bail, Context};
+use camino::Utf8PathBuf;
 use pageserver_api::models::{self, TenantInfo, TimelineInfo};
 use postgres_backend::AuthType;
 use postgres_connection::{parse_host_port, PgConnectionConfig};
@@ -144,7 +145,7 @@ impl PageServerNode {
     pub fn initialize(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
         // First, run `pageserver --init` and wait for it to write a config into FS and exit.
         self.pageserver_init(config_overrides)
-            .with_context(|| format!("Failed to run init for pageserver node {}", self.conf.id,))
+            .with_context(|| format!("Failed to run init for pageserver node {}", self.conf.id))
     }
 
     pub fn repo_path(&self) -> PathBuf {
@@ -154,8 +155,9 @@ impl PageServerNode {
     /// The pid file is created by the pageserver process, with its pid stored inside.
     /// Other pageservers cannot lock the same file and overwrite it for as long as the current
     /// pageserver runs. (Unless someone removes the file manually; never do that!)
-    fn pid_file(&self) -> PathBuf {
-        self.repo_path().join("pageserver.pid")
+    fn pid_file(&self) -> Utf8PathBuf {
+        Utf8PathBuf::from_path_buf(self.repo_path().join("pageserver.pid"))
+            .expect("non-Unicode path")
     }
 
     pub fn start(&self, config_overrides: &[&str]) -> anyhow::Result<Child> {

control_plane/src/safekeeper.rs

@@ -11,6 +11,7 @@ use std::process::Child;
 use std::{io, result};
 
 use anyhow::Context;
+use camino::Utf8PathBuf;
 use postgres_connection::PgConnectionConfig;
 use reqwest::blocking::{Client, RequestBuilder, Response};
 use reqwest::{IntoUrl, Method};
@@ -97,8 +98,9 @@ impl SafekeeperNode {
         SafekeeperNode::datadir_path_by_id(&self.env, self.id)
     }
 
-    pub fn pid_file(&self) -> PathBuf {
-        self.datadir_path().join("safekeeper.pid")
+    pub fn pid_file(&self) -> Utf8PathBuf {
+        Utf8PathBuf::from_path_buf(self.datadir_path().join("safekeeper.pid"))
+            .expect("non-Unicode path")
     }
 
     pub fn start(&self, extra_opts: Vec<String>) -> anyhow::Result<Child> {

deny.toml

@@ -23,7 +23,7 @@ vulnerability = "deny"
 unmaintained = "warn"
 yanked = "warn"
 notice = "warn"
-ignore = ["RUSTSEC-2023-0052"]
+ignore = []
 
 # This section is considered when running `cargo deny check licenses`
 # More documentation for the licenses section can be found here:

docker-compose/compute_wrapper/var/db/postgres/specs/spec.json

@@ -25,7 +25,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

docs/error-handling.md

@@ -188,11 +188,60 @@ that.
 
 ## Error message style
 
+### PostgreSQL extensions
+
 PostgreSQL has a style guide for writing error messages:
 
 https://www.postgresql.org/docs/current/error-style-guide.html
 
 Follow that guide when writing error messages in the PostgreSQL
-extension. We don't follow it strictly in the pageserver and
-safekeeper, but the advice in the PostgreSQL style guide is generally
-good, and you can't go wrong by following it.
+extensions.
+
+### Neon Rust code
+
+#### Anyhow Context
+
+When adding anyhow `context()`, use form `present-tense-verb+action`.
+
+Example:
+- Bad: `file.metadata().context("could not get file metadata")?;`
+- Good: `file.metadata().context("get file metadata")?;`
+
+#### Logging Errors
+
+When logging any error `e`, use `could not {e:#}` or `failed to {e:#}`.
+
+If `e` is an `anyhow` error and you want to log the backtrace that it contains,
+use `{e:?}` instead of `{e:#}`.
+
+#### Rationale
+
+The `{:#}` ("alternate Display") of an `anyhow` error chain is concatenation fo the contexts, using `: `.
+
+For example, the following Rust code will result in output
+```
+ERROR  failed to list users: load users from server: parse response: invalid json
+```
+
+This is more concise / less noisy than what happens if you do `.context("could not ...")?` at each level, i.e.:
+
+```
+ERROR  could not list users: could not load users from server: could not parse response: invalid json
+```
+
+
+```rust
+fn main() {
+  match list_users().context("list users") else {
+    Ok(_) => ...,
+    Err(e) => tracing::error!("failed to {e:#}"),
+  }
+}
+fn list_users() {
+  http_get_users().context("load users from server")?;
+}
+fn http_get_users() {
+  let response = client....?;
+  response.parse().context("parse response")?; // fails with serde error "invalid json"
+}
+```

docs/pageserver-services.md

@@ -96,6 +96,16 @@ prefix_in_bucket = '/test_prefix/'
 
 `AWS_SECRET_ACCESS_KEY` and `AWS_ACCESS_KEY_ID` env variables can be used to specify the S3 credentials if needed.
 
+or
+
+```toml
+[remote_storage]
+container_name = 'some-container-name'
+container_region = 'us-east'
+prefix_in_container = '/test-prefix/'
+```
+
+`AZURE_STORAGE_ACCOUNT` and `AZURE_STORAGE_ACCESS_KEY` env variables can be used to specify the azure credentials if needed.
 
 ## Repository background tasks
 

docs/rfcs/028-pageserver-migration.md

@@ -0,0 +1,599 @@
+# Seamless tenant migration
+
+- Author: john@neon.tech
+- Created on 2023-08-11
+- Implemented on ..
+
+## Summary
+
+The preceding [generation numbers RFC](025-generation-numbers.md) may be thought of as "making tenant
+migration safe". Following that,
+this RFC is about how those migrations are to be done:
+
+1. Seamlessly (without interruption to client availability)
+2. Quickly (enabling faster operations)
+3. Efficiently (minimizing I/O and $ cost)
+
+These points are in priority order: if we have to sacrifice
+efficiency to make a migration seamless for clients, we will
+do so, etc.
+
+This is accomplished by introducing two high level changes:
+
+- A dual-attached state for tenants, used in a control-plane-orchestrated
+  migration procedure that preserves availability during a migration.
+- Warm secondary locations for tenants, where on-disk content is primed
+  for a fast migration of the tenant from its current attachment to this
+  secondary location.
+
+## Motivation
+
+Migrating tenants between pageservers is essential to operating a service
+at scale, in several contexts:
+
+1. Responding to a pageserver node failure by migrating tenants to other pageservers
+2. Balancing load and capacity across pageservers, for example when a user expands their
+   database and they need to migrate to a pageserver with more capacity.
+3. Restarting pageservers for upgrades and maintenance
+
+The current situation steps for migration are:
+
+- detach from old node; skip if old node is dead; (the [skip part is still WIP](https://github.com/neondatabase/cloud/issues/5426)).
+- attach to new node
+- re-configure endpoints to use the new node
+
+Once [generation numbers](025-generation-numbers.md) are implemented,
+the detach step is no longer critical for correctness. So, we can
+
+- attach to a new node,
+- re-configure endpoints to use the new node, and then
+- detach from the old node.
+
+However, this still does not meet our seamless/fast/efficient goals:
+
+- Not fast: The new node will have to download potentially large amounts
+  of data from S3, which may take many minutes.
+- Not seamless: If we attach to a new pageserver before detaching an old one,
+  the new one might delete some objects that interrupt availability of reads on the old one.
+- Not efficient: the old pageserver will continue uploading
+  S3 content during the migration that will never be read.
+
+The user expectations for availability are:
+
+- For planned maintenance, there should be zero availability
+  gap. This expectation is fulfilled by this RFC.
+- For unplanned changes (e.g. node failures), there should be
+  minimal availability gap. This RFC provides the _mechanism_
+  to fail over quickly, but does not provide the failure _detection_
+  nor failover _policy_.
+
+## Non Goals
+
+- Defining service tiers with different storage strategies: the same
+  level of HA & overhead will apply to all tenants. This doesn't rule out
+  adding such tiers in future.
+- Enabling pageserver failover in the absence of a control plane: the control
+  plane will remain the source of truth for what should be attached where.
+- Totally avoiding availability gaps on unplanned migrations during
+  a failure (we expect a small, bounded window of
+  read unavailability of very recent LSNs)
+- Workload balancing: this RFC defines the mechanism for moving tenants
+  around, not the higher level logic for deciding who goes where.
+- Defining all possible configuration flows for tenants: the migration process
+  defined in this RFC demonstrates the sufficiency of the pageserver API, but
+  is not the only kind of configuration change the control plane will ever do.
+  The APIs defined here should let the control plane move tenants around in
+  whatever way is needed while preserving data safety and read availability.
+
+## Impacted components
+
+Pageserver, control plane
+
+## Terminology
+
+- **Attachment**: a tenant is _attached_ to a pageserver if it has
+  been issued a generation number, and is running an instance of
+  the `Tenant` type, ingesting the WAL, and available to serve
+  page reads.
+- **Location**: locations are a superset of attachments. A location
+  is a combination of a tenant and a pageserver. We may _attach_ at a _location_.
+
+- **Secondary location**: a location which is not currently attached.
+- **Warm secondary location**: a location which is not currently attached, but is endeavoring to maintain a warm local cache of layers. We avoid calling this a _warm standby_ to avoid confusion with similar postgres features.
+
+## Implementation (high level)
+
+### Warm secondary locations
+
+To enable faster migrations, we will identify at least one _secondary location_
+for each tenant. This secondary location will keep a warm cache of layers
+for the tenant, so that if it is later attached, it can catch up with the
+latest LSN quickly: rather than downloading everything, it only has to replay
+the recent part of the WAL to advance from the remote_consistent_offset to the
+most recent LSN in the WAL.
+
+The control plane is responsible for selecting secondary locations, and
+calling into pageservers to configure tenants into a secondary mode at this
+new location, as well as attaching the tenant in its existing primary location.
+
+The attached pageserver for a tenant will publish a [layer heatmap](#layer-heatmap)
+to advise secondaries of which layers should be downloaded.
+
+### Location modes
+
+Currently, we consider a tenant to be in one of two states on a pageserver:
+
+- Attached: active `Tenant` object, and layers on local disk
+- Detached: no layers on local disk, no runtime state.
+
+We will extend this with finer-grained modes, whose purpose will become
+clear in later sections:
+
+- **AttachedSingle**: equivalent the existing attached state.
+- **AttachedMulti**: like AttachedSingle, holds an up to date generation, but
+  does not do deletions.
+- **AttachedStale**: like AttachedSingle, holds a stale generation,
+  do not do any remote storage operations.
+- **Secondary**: keep local state on disk, periodically update from S3.
+- **Detached**: equivalent to existing detached state.
+
+To control these finer grained states, a new pageserver API endpoint will be added.
+
+### Cutover procedure
+
+Define old location and new location as "Node A" and "Node B". Consider
+the case where both nodes are available, and Node B was previously configured
+as a secondary location for the tenant we are migrating.
+
+The cutover procedure is orchestrated by the control plane, calling into
+the pageservers' APIs:
+
+1. Call to Node A requesting it to flush to S3 and enter AttachedStale state
+2. Increment generation, and call to Node B requesting it to enter AttachedMulti
+   state with the new generation.
+3. Call to Node B, requesting it to download the latest hot layers from remote storage,
+   according to the latest heatmap flushed by Node A.
+4. Wait for Node B's WAL ingestion to catch up with node A's
+5. Update endpoints to use node B instead of node A
+6. Call to node B requesting it to enter state AttachedSingle.
+7. Call to node A requesting it to enter state Secondary
+
+The following table summarizes how the state of the system advances:
+
+|     Step      |     Node A     |     Node B     | Node used by endpoints |
+| :-----------: | :------------: | :------------: | :--------------------: |
+| 1 (_initial_) | AttachedSingle |   Secondary    |           A            |
+|       2       | AttachedStale  | AttachedMulti  |           A            |
+|       3       | AttachedStale  | AttachedMulti  |           A            |
+|       4       | AttachedStale  | AttachedMulti  |           A            |
+| 5 (_cutover_) | AttachedStale  | AttachedMulti  |           B            |
+|       6       | AttachedStale  | AttachedSingle |           B            |
+|  7 (_final_)  |   Secondary    | AttachedSingle |           B            |
+
+The procedure described for a clean handover from a live node to a secondary
+is also used for failure cases and for migrations to a location that is not
+configured as a secondary, by simply skipping irrelevant steps, as described in
+the following sections.
+
+#### Migration from an unresponsive node
+
+If node A is unavailable, then all calls into
+node A are skipped and we don't wait for B to catch up before
+switching updating the endpoints to use B.
+
+#### Migration to a location that is not a secondary
+
+If node B is initially in Detached state, the procedure is identical. Since Node B
+is coming from a Detached state rather than Secondary, the download of layers and
+catch up with WAL will take much longer.
+
+We might do this if:
+
+- Attached and secondary locations are both critically low on disk, and we need
+  to migrate to a third node with more resources available.
+- We are migrating a tenant which does not use secondary locations to save on cost.
+
+#### Permanent migration away from a node
+
+In the final step of the migration, we generally request the original node to enter a Secondary
+state. This is typical if we are doing a planned migration during maintenance, or to
+balance CPU/network load away from a node.
+
+One might also want to permanently migrate away: this can be done by simply removing the secondary
+location after the migration is complete, or as an optimization by substituting the Detached state
+for the Secondary state in the final step.
+
+#### Cutover diagram
+
+```mermaid
+sequenceDiagram
+participant CP as Control plane
+participant A as Node A
+participant B as Node B
+participant E as Endpoint
+
+CP->>A: PUT Flush & go to AttachedStale
+note right of A: A continues to ingest WAL
+CP->>B: PUT AttachedMulti
+CP->>B: PUT Download layers from latest heatmap
+note right of B: B downloads from S3
+loop Poll until download complete
+CP->>B: GET download status
+end
+activate B
+note right of B: B ingests WAL
+loop Poll until catch up
+CP->>B: GET visible WAL
+CP->>A: GET visible WAL
+end
+deactivate B
+CP->>E: Configure to use Node B
+E->>B: Connect for reads
+CP->>B: PUT AttachedSingle
+CP->>A: PUT Secondary
+```
+
+#### Cutover from an unavailable pageserver
+
+This case is far simpler: we may skip straight to our intended
+end state.
+
+```mermaid
+sequenceDiagram
+participant A as Node A
+participant CP as Control plane
+participant B as Node B
+participant E as Endpoint
+
+note right of A: Node A offline
+activate A
+CP->>B: PUT AttachedSingle
+CP->>E: Configure to use Node B
+E->>B: Connect for reads
+deactivate A
+```
+
+## Implementation (detail)
+
+### Purpose of AttachedMulti, AttachedStale
+
+#### AttachedMulti
+
+Ordinarily, an attached pageserver whose generation is the latest may delete
+layers at will (e.g. during compaction). If a previous generation pageserver
+is also still attached, and in use by endpoints, then this layer deletion could
+lead to a loss of availability for the endpoint when reading from the previous
+generation pageserver.
+
+The _AttachedMulti_ state simply disables deletions. These will be enqueued
+in `RemoteTimelineClient` until the control plane transitions the
+node into AttachedSingle, which unblocks deletions.  Other remote storage operations
+such as uploads are not blocked.
+
+AttachedMulti is not required for data safety, only to preserve availability
+on pageservers running with stale generations.
+
+A node enters AttachedMulti only when explicitly asked to by the control plane. It should
+only remain in this state for the duration of a migration.
+
+If a control plane bug leaves
+the node in AttachedMulti for a long time, then we must avoid unbounded memory use from enqueued
+deletions. This may be accomplished simply, by dropping enqueued deletions when some modest
+threshold of delayed deletions (e.g. 10k layers per tenant) is reached. As with all deletions,
+it is safe to skip them, and the leaked objects will be eventually cleaned up by scrub or
+by timeline deletion.
+
+During AttachedMulti, the Tenant is free to drop layers from local disk in response to
+disk pressure: only the deletion of remote layers is blocked.
+
+#### AttachedStale
+
+Currently, a pageserver with a stale generation number will continue to
+upload layers, but be prevented from completing deletions. This is safe, but inefficient: layers uploaded by this stale generation
+will not be read back by future generations of pageservers.
+
+The _AttachedStale_ state disables S3 uploads. The stale pageserver
+will continue to ingest the WAL and write layers to local disk, but not to
+do any uploads to S3.
+
+A node may enter AttachedStale in two ways:
+
+- Explicitly, when control plane calls into the node at the start of a migration.
+- Implicitly, when the node tries to validate some deletions and discovers
+  that its generation is stale.
+
+The AttachedStale state also disables sending consumption metrics from
+that location: it is interpreted as an indication that some other pageserver
+is already attached or is about to be attached, and that new pageserver will
+be responsible for sending consumption metrics.
+
+#### Disk Pressure & AttachedStale
+
+Over long periods of time, a tenant location in AttachedStale will accumulate data
+on local disk, as it cannot evict any layers written since it entered the
+AttachStale state. We rely on the control plane to revert the location to
+Secondary or Detached at the end of a migration.
+
+This scenario is particularly noteworthy when evacuating all tenants on a pageserver:
+since _all_ the attached tenants will go into AttachedStale, we will be doing no
+uploads at all, therefore ingested data will cause disk usage to increase continuously.
+Under nominal conditions, the available disk space on pageservers should be sufficient
+to complete the evacuation before this becomes a problem, but we must also handle
+the case where we hit a low disk situation while in this state.
+
+The concept of disk pressure already exists in the pageserver: the `disk_usage_eviction_task`
+touches each Tenant when it determines that a low-disk condition requires
+some layer eviction. Having selected layers for eviction, the eviction
+task calls `Timeline::evict_layers`.
+
+**Safety**: If evict_layers is called while in AttachedStale state, and some of the to-be-evicted
+layers are not yet uploaded to S3, then the block on uploads will be lifted. This
+will result in leaking some objects once a migration is complete, but will enable
+the node to manage its disk space properly: if a node is left with some tenants
+in AttachedStale indefinitely due to a network partition or control plane bug,
+these tenants will not cause a full disk condition.
+
+### Warm secondary updates
+
+#### Layer heatmap
+
+The secondary location's job is to serve reads **with the same quality of service as the original location
+was serving them around the time of a migration**. This does not mean the secondary
+location needs the whole set of layers: inactive layers that might soon
+be evicted on the attached pageserver need not be downloaded by the
+secondary. A totally idle tenant only needs to maintain enough on-disk
+state to enable a fast cold start (i.e. the most recent image layers are
+typically sufficient).
+
+To enable this, we introduce the concept of a _layer heatmap_, which
+acts as an advisory input to secondary locations to decide which
+layers to download from S3.
+
+#### Attached pageserver
+
+The attached pageserver, if in state AttachedSingle, periodically
+uploads a serialized heat map to S3. It may skip this if there
+is no change since the last time it uploaded (e.g. if the tenant
+is totally idle).
+
+Additionally, when the tenant is flushed to remote storage prior to a migration
+(the first step in [cutover procedure](#cutover-procedure)), 
+the heatmap is written out. This enables a future attached pageserver
+to get an up to date view when deciding which layers to download.
+
+#### Secondary location behavior
+
+Secondary warm locations run a simple loop, implemented separately from
+the main `Tenant` type, which represents attached tenants:
+
+- Download the layer heatmap
+- Select any "hot enough" layers to download, if there is sufficient
+  free disk space.
+- Download layers, if they were not previously evicted (see below)
+- Download the latest index_part.json
+- Check if any layers currently on disk are no longer referenced by
+  IndexPart & delete them
+
+Note that the heatmap is only advisory: if a secondary location has plenty
+of disk space, it may choose to retain layers that aren't referenced
+by the heatmap, as long as they are still referenced by the IndexPart. Conversely,
+if a node is very low on disk space, it might opt to raise the heat threshold required
+to both downloading a layer, until more disk space is available.
+
+#### Secondary locations & disk pressure
+
+Secondary locations are subject to eviction on disk pressure, just as
+attached locations are.  For eviction purposes, the access time of a
+layer in a secondary location will be the access time given in the heatmap,
+rather than the literal time at which the local layer file was accessed.
+
+The heatmap will indicate which layers are in local storage on the attached
+location.  The secondary will always attempt to get back to having that
+set of layers on disk, but to avoid flapping, it will remember the access
+time of the layer it was most recently asked to evict, and layers whose
+access time is below that will not be re-downloaded.
+
+The resulting behavior is that after a layer is evicted from a secondary
+location, it is only re-downloaded once the attached pageserver accesses
+the layer and uploads a heatmap reflecting that access time.  On a pageserver
+restart, the secondary location will attempt to download all layers in
+the heatmap again, if they are not on local disk.
+
+This behavior will be slightly different when secondary locations are
+used for "low energy tenants", but that is beyond the scope of this RFC.
+
+### Location configuration API
+
+Currently, the `/tenant/<tenant_id>/config` API defines various
+tunables like compaction settings, which apply to the tenant irrespective
+of which pageserver it is running on.
+
+A new "location config" structure will be introduced, which defines
+configuration which is per-tenant, but local to a particular pageserver,
+such as the attachment mode and whether it is a secondary.
+
+The pageserver will expose a new per-tenant API for setting
+the state: `/tenant/<tenant_id>/location/config`.
+
+Body content:
+
+```
+{
+  state: 'enum{Detached, Secondary, AttachedSingle, AttachedMulti, AttachedStale}',
+  generation: Option<u32>,
+  configuration: `Option<TenantConfig>`
+  flush: bool
+}
+```
+
+Existing `/attach` and `/detach` endpoint will have the same
+behavior as calling `/location/config` with `AttachedSingle` and `Detached`
+states respectively. These endpoints will be deprecated and later
+removed.
+
+The generation attribute is mandatory for entering `AttachedSingle` or
+`AttachedMulti`.
+
+The configuration attribute is mandatory when entering any state other
+than `Detached`. This configuration is the same as the body for
+the existing `/tenant/<tenant_id>/config` endpoint.
+
+The `flush` argument indicates whether the pageservers should flush
+to S3 before proceeding: this only has any effect if the node is
+currently in AttachedSingle or AttachedMulti. This is used
+during the first phase of migration, when transitioning the
+old pageserver to AttachedSingle.
+
+The `/re-attach` API response will be extended to include a `state` as
+well as a `generation`, enabling the pageserver to enter the
+correct state for each tenant on startup.
+
+### Database schema for locations
+
+A new table `ProjectLocation`:
+
+- pageserver_id: int
+- tenant_id: TenantId
+- generation: Option<int>
+- state: `enum(Secondary, AttachedSingle, AttachedMulti)`
+
+Notes:
+
+- It is legacy for a Project to have zero `ProjectLocation`s
+- The `pageserver` column in `Project` now means "to which pageserver should
+  endpoints connect", rather than simply which pageserver is attached.
+- The `generation` column in `Project` remains, and is incremented and used
+  to set the generation of `ProjectLocation` rows when they are set into
+  an attached state.
+- The `Detached` state is implicitly represented as the absence of
+  a `ProjectLocation`.
+
+### Executing migrations
+
+Migrations will be implemented as Go functions, within the
+existing `Operation` framework in the control plane. These
+operations are persistent, such that they will always keep
+trying until completion: this property is important to avoid
+leaving garbage behind on pageservers, such as AttachedStale
+locations.
+
+### Recovery from failures during migration
+
+During migration, the control plane may encounter failures of either
+the original or new pageserver, or both:
+
+- If the original fails, skip past waiting for the new pageserver
+  to catch up, and put it into AttachedSingle immediately.
+- If the new node fails, put the old pageserver into Secondary
+  and then back into AttachedSingle (this has the effect of
+  retaining on-disk state and granting it a fresh generation number).
+- If both nodes fail, keep trying until one of them is available
+  again.
+
+### Control plane -> Pageserver reconciliation
+
+A migration may be done while the old node is unavailable,
+in which case the old node may still be running in an AttachedStale
+state.
+
+In this case, it is undesirable to have the migration `Operation`
+stay alive until the old node eventually comes back online
+and can be cleaned up. To handle this, the control plane
+should run a background reconciliation process to compare
+a pageserver's attachments with the database, and clean up
+any that shouldn't be there any more.
+
+Note that there will be no work to do if the old node was really
+offline, as during startup it will call into `/re-attach` and
+be updated that way. The reconciliation will only be needed
+if the node was unavailable but still running.
+
+## Alternatives considered
+
+### Only enabling secondary locations for tenants on a higher service tier
+
+This will make sense in future, especially for tiny databases that may be
+downloaded from S3 in milliseconds when needed.
+
+However, it is not wise to do it immediately, because pageservers contain
+a mixture of higher and lower tier workloads. If we had 1 tenant with
+a secondary location and 9 without, then those other 9 tenants will do
+a lot of I/O as they try to recover from S3, which may degrade the
+service of the tenant which had a secondary location.
+
+Until we segregate tenant on different service tiers on different pageserver
+nodes, or implement & test QoS to ensure that tenants with secondaries are
+not harmed by tenants without, we should use the same failover approach
+for all the tenants.
+
+### Hot secondary locations (continuous WAL replay)
+
+Instead of secondary locations populating their caches from S3, we could
+have them consume the WAL from safekeepers. The downsides of this would be:
+
+- Double load on safekeepers, which are a less scalable service than S3
+- Secondary locations' on-disk state would end up subtly different to
+  the remote state, which would make synchronizing with S3 more complex/expensive
+  when going into attached state.
+
+The downside of only updating secondary locations from S3 is that we will
+have a delay during migration from replaying the LSN range between what's
+in S3 and what's in the pageserver. This range will be very small on
+planned migrations, as we have the old pageserver flush to S3 immediately
+before attaching the new pageserver. On unplanned migrations (old pageserver
+is unavailable), the range of LSNs to replay is bounded by the flush frequency
+on the old pageserver. However, the migration doesn't have to wait for the
+replay: it's just that not-yet-replayed LSNs will be unavailable for read
+until the new pageserver catches up.
+
+We expect that pageserver reads of the most recent LSNs will be relatively
+rare, as for an active endpoint those pages will usually still be in the postgres
+page cache: this leads us to prefer synchronizing from S3 on secondary
+locations, rather than consuming the WAL from safekeepers.
+
+### Cold secondary locations
+
+It is not functionally necessary to keep warm caches on secondary locations at all. However, if we do not, then
+we would experience a de-facto availability loss in unplanned migrations, as reads to the new node would take an extremely long time (many seconds, perhaps minutes).
+
+Warm caches on secondary locations are necessary to meet
+our availability goals.
+
+### Pageserver-granularity failover
+
+Instead of migrating tenants individually, we could have entire spare nodes,
+and on a node death, move all its work to one of these spares.
+
+This approach is avoided for several reasons:
+
+- we would still need fine-grained tenant migration for other
+  purposes such as balancing load
+- by sharing the spare capacity over many peers rather than one spare node,
+  these peers may use the capacity for other purposes, until it is needed
+  to handle migrated tenants. e.g. for keeping a deeper cache of their
+  attached tenants.
+
+### Readonly during migration
+
+We could simplify migrations by making both previous and new nodes go into a
+readonly state, then flush remote content from the previous node, then activate
+attachment on the secondary node.
+
+The downside to this approach is a potentially large gap in readability of
+recent LSNs while loading data onto the new node. To avoid this, it is worthwhile
+to incur the extra cost of double-replaying the WAL onto old and new nodes' local
+storage during a migration.
+
+### Peer-to-peer pageserver communication
+
+Rather than uploading the heatmap to S3, attached pageservers could make it
+available to peers.
+
+Currently, pageservers have no peer to peer communication, so adding this
+for heatmaps would incur significant overhead in deployment and configuration
+of the service, and ensuring that when a new pageserver is deployed, other
+pageservers are updated to be aware of it.
+
+As well as simplifying implementation, putting heatmaps in S3 will be useful
+for future analytics purposes -- gathering aggregated statistics on activity
+pattersn across many tenants may be done directly from data in S3.

libs/compute_api/src/spec.rs

@@ -200,6 +200,12 @@ pub struct Database {
     pub name: PgIdent,
     pub owner: PgIdent,
     pub options: GenericOptions,
+    // These are derived flags, not present in the spec file.
+    // They are never set by the control plane.
+    #[serde(skip_deserializing, default)]
+    pub restrict_conn: bool,
+    #[serde(skip_deserializing, default)]
+    pub invalid: bool,
 }
 
 /// Common type representing both SQL statement params with or without value,

libs/compute_api/tests/cluster_spec.json

@@ -76,7 +76,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

libs/consumption_metrics/src/lib.rs

@@ -3,9 +3,9 @@
 //!
 use chrono::{DateTime, Utc};
 use rand::Rng;
-use serde::Serialize;
+use serde::{Deserialize, Serialize};
 
-#[derive(Serialize, Debug, Clone, Copy, Eq, PartialEq, Ord, PartialOrd)]
+#[derive(Serialize, serde::Deserialize, Debug, Clone, Copy, Eq, PartialEq, Ord, PartialOrd)]
 #[serde(tag = "type")]
 pub enum EventType {
     #[serde(rename = "absolute")]
@@ -27,7 +27,8 @@ impl EventType {
     }
 
     pub fn incremental_timerange(&self) -> Option<std::ops::Range<&DateTime<Utc>>> {
-        // these can most likely be thought of as Range or RangeFull
+        // these can most likely be thought of as Range or RangeFull, at least pageserver creates
+        // incremental ranges where the stop and next start are equal.
         use EventType::*;
         match self {
             Incremental {
@@ -41,15 +42,25 @@ impl EventType {
     pub fn is_incremental(&self) -> bool {
         matches!(self, EventType::Incremental { .. })
     }
+
+    /// Returns the absolute time, or for incremental ranges, the stop time.
+    pub fn recorded_at(&self) -> &DateTime<Utc> {
+        use EventType::*;
+
+        match self {
+            Absolute { time } => time,
+            Incremental { stop_time, .. } => stop_time,
+        }
+    }
 }
 
-#[derive(Serialize, Debug, Clone, Eq, PartialEq, Ord, PartialOrd)]
-pub struct Event<Extra> {
+#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq, Ord, PartialOrd)]
+pub struct Event<Extra, Metric> {
     #[serde(flatten)]
     #[serde(rename = "type")]
     pub kind: EventType,
 
-    pub metric: &'static str,
+    pub metric: Metric,
     pub idempotency_key: String,
     pub value: u64,
 
@@ -58,19 +69,45 @@ pub struct Event<Extra> {
 }
 
 pub fn idempotency_key(node_id: &str) -> String {
-    format!(
-        "{}-{}-{:04}",
-        Utc::now(),
-        node_id,
-        rand::thread_rng().gen_range(0..=9999)
-    )
+    IdempotencyKey::generate(node_id).to_string()
+}
+
+/// Downstream users will use these to detect upload retries.
+pub struct IdempotencyKey<'a> {
+    now: chrono::DateTime<Utc>,
+    node_id: &'a str,
+    nonce: u16,
+}
+
+impl std::fmt::Display for IdempotencyKey<'_> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}-{}-{:04}", self.now, self.node_id, self.nonce)
+    }
+}
+
+impl<'a> IdempotencyKey<'a> {
+    pub fn generate(node_id: &'a str) -> Self {
+        IdempotencyKey {
+            now: Utc::now(),
+            node_id,
+            nonce: rand::thread_rng().gen_range(0..=9999),
+        }
+    }
+
+    pub fn for_tests(now: DateTime<Utc>, node_id: &'a str, nonce: u16) -> Self {
+        IdempotencyKey {
+            now,
+            node_id,
+            nonce,
+        }
+    }
 }
 
 pub const CHUNK_SIZE: usize = 1000;
 
 // Just a wrapper around a slice of events
 // to serialize it as `{"events" : [ ] }
-#[derive(serde::Serialize)]
+#[derive(serde::Serialize, serde::Deserialize)]
 pub struct EventChunk<'a, T: Clone> {
     pub events: std::borrow::Cow<'a, [T]>,
 }

libs/metrics/src/wrappers.rs

@@ -1,6 +1,6 @@
 use std::io::{Read, Result, Write};
 
-/// A wrapper for an object implementing [Read](std::io::Read)
+/// A wrapper for an object implementing [Read]
 /// which allows a closure to observe the amount of bytes read.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -51,17 +51,17 @@ impl<'a, T> CountedReader<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get an immutable reference to the underlying [Read] implementor
     pub fn inner(&self) -> &T {
         &self.reader
     }
 
-    /// Get a mutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get a mutable reference to the underlying [Read] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.reader
     }
 
-    /// Consume the wrapper and return the underlying [Read](std::io::Read) implementor
+    /// Consume the wrapper and return the underlying [Read] implementor
     pub fn into_inner(self) -> T {
         self.reader
     }
@@ -75,7 +75,7 @@ impl<T: Read> Read for CountedReader<'_, T> {
     }
 }
 
-/// A wrapper for an object implementing [Write](std::io::Write)
+/// A wrapper for an object implementing [Write]
 /// which allows a closure to observe the amount of bytes written.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -122,17 +122,17 @@ impl<'a, T> CountedWriter<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get an immutable reference to the underlying [Write] implementor
     pub fn inner(&self) -> &T {
         &self.writer
     }
 
-    /// Get a mutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get a mutable reference to the underlying [Write] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.writer
     }
 
-    /// Consume the wrapper and return the underlying [Write](std::io::Write) implementor
+    /// Consume the wrapper and return the underlying [Write] implementor
     pub fn into_inner(self) -> T {
         self.writer
     }

libs/pageserver_api/src/models.rs

@@ -10,6 +10,7 @@ use serde_with::{serde_as, DisplayFromStr};
 use strum_macros;
 use utils::{
     completion,
+    generation::Generation,
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
@@ -218,6 +219,8 @@ impl std::ops::Deref for TenantCreateRequest {
     }
 }
 
+/// An alternative representation of `pageserver::tenant::TenantConf` with
+/// simpler types.
 #[derive(Serialize, Deserialize, Debug, Default)]
 pub struct TenantConfig {
     pub checkpoint_distance: Option<u64>,
@@ -243,6 +246,39 @@ pub struct TenantConfig {
     pub gc_feedback: Option<bool>,
 }
 
+/// A flattened analog of a `pagesever::tenant::LocationMode`, which
+/// lists out all possible states (and the virtual "Detached" state)
+/// in a flat form rather than using rust-style enums.
+#[derive(Serialize, Deserialize, Debug)]
+pub enum LocationConfigMode {
+    AttachedSingle,
+    AttachedMulti,
+    AttachedStale,
+    Secondary,
+    Detached,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+pub struct LocationConfigSecondary {
+    pub warm: bool,
+}
+
+/// An alternative representation of `pageserver::tenant::LocationConf`,
+/// for use in external-facing APIs.
+#[derive(Serialize, Deserialize, Debug)]
+pub struct LocationConfig {
+    pub mode: LocationConfigMode,
+    /// If attaching, in what generation?
+    #[serde(default)]
+    pub generation: Option<Generation>,
+    #[serde(default)]
+    pub secondary_conf: Option<LocationConfigSecondary>,
+
+    // If requesting mode `Secondary`, configuration for that.
+    // Custom storage configuration for the tenant, if any
+    pub tenant_conf: TenantConfig,
+}
+
 #[serde_as]
 #[derive(Serialize, Deserialize)]
 #[serde(transparent)]
@@ -253,6 +289,16 @@ pub struct StatusResponse {
     pub id: NodeId,
 }
 
+#[serde_as]
+#[derive(Serialize, Deserialize, Debug)]
+#[serde(deny_unknown_fields)]
+pub struct TenantLocationConfigRequest {
+    #[serde_as(as = "DisplayFromStr")]
+    pub tenant_id: TenantId,
+    #[serde(flatten)]
+    pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
+}
+
 #[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
@@ -363,8 +409,15 @@ pub struct TimelineInfo {
     pub latest_gc_cutoff_lsn: Lsn,
     #[serde_as(as = "DisplayFromStr")]
     pub disk_consistent_lsn: Lsn,
+
+    /// The LSN that we have succesfully uploaded to remote storage
     #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn: Lsn,
+
+    /// The LSN that we are advertizing to safekeepers
+    #[serde_as(as = "DisplayFromStr")]
+    pub remote_consistent_lsn_visible: Lsn,
+
     pub current_logical_size: Option<u64>, // is None when timeline is Unloaded
     /// Sum of the size of all layer files.
     /// If a layer is present in both local FS and S3, it counts only once.

libs/pageserver_api/src/reltag.rs

@@ -22,9 +22,9 @@ use postgres_ffi::Oid;
 /// [See more related comments here](https:///github.com/postgres/postgres/blob/99c5852e20a0987eca1c38ba0c09329d4076b6a0/src/include/storage/relfilenode.h#L57).
 ///
 // FIXME: should move 'forknum' as last field to keep this consistent with Postgres.
-// Then we could replace the custo Ord and PartialOrd implementations below with
-// deriving them.
-#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize, Deserialize)]
+// Then we could replace the custom Ord and PartialOrd implementations below with
+// deriving them. This will require changes in walredoproc.c.
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize)]
 pub struct RelTag {
     pub forknum: u8,
     pub spcnode: Oid,
@@ -40,21 +40,9 @@ impl PartialOrd for RelTag {
 
 impl Ord for RelTag {
     fn cmp(&self, other: &Self) -> Ordering {
-        let mut cmp = self.spcnode.cmp(&other.spcnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.dbnode.cmp(&other.dbnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.relnode.cmp(&other.relnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.forknum.cmp(&other.forknum);
-
-        cmp
+        // Custom ordering where we put forknum to the end of the list
+        let other_tup = (other.spcnode, other.dbnode, other.relnode, other.forknum);
+        (self.spcnode, self.dbnode, self.relnode, self.forknum).cmp(&other_tup)
     }
 }
 

libs/postgres_backend/src/lib.rs

@@ -19,8 +19,8 @@ use tracing::{debug, error, info, trace};
 
 use pq_proto::framed::{ConnectionError, Framed, FramedReader, FramedWriter};
 use pq_proto::{
-    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_INTERNAL_ERROR,
-    SQLSTATE_SUCCESSFUL_COMPLETION,
+    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_ADMIN_SHUTDOWN,
+    SQLSTATE_INTERNAL_ERROR, SQLSTATE_SUCCESSFUL_COMPLETION,
 };
 
 /// An error, occurred during query processing:
@@ -30,6 +30,9 @@ pub enum QueryError {
     /// The connection was lost while processing the query.
     #[error(transparent)]
     Disconnected(#[from] ConnectionError),
+    /// We were instructed to shutdown while processing the query
+    #[error("Shutting down")]
+    Shutdown,
     /// Some other error
     #[error(transparent)]
     Other(#[from] anyhow::Error),
@@ -44,7 +47,8 @@ impl From<io::Error> for QueryError {
 impl QueryError {
     pub fn pg_error_code(&self) -> &'static [u8; 5] {
         match self {
-            Self::Disconnected(_) => b"08006",         // connection failure
+            Self::Disconnected(_) => b"08006", // connection failure
+            Self::Shutdown => SQLSTATE_ADMIN_SHUTDOWN,
             Self::Other(_) => SQLSTATE_INTERNAL_ERROR, // internal error
         }
     }
@@ -396,7 +400,20 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         // socket might be already closed, e.g. if previously received error,
         // so ignore result.
         self.framed.shutdown().await.ok();
-        ret
+        match ret {
+            Ok(()) => Ok(()),
+            Err(QueryError::Shutdown) => {
+                info!("Stopped due to shutdown");
+                Ok(())
+            }
+            Err(QueryError::Disconnected(e)) => {
+                info!("Disconnected ({e:#})");
+                // Disconnection is not an error: we just use it that way internally to drop
+                // out of loops.
+                Ok(())
+            }
+            e => e,
+        }
     }
 
     async fn run_message_loop<F, S>(
@@ -416,15 +433,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received during handshake");
-                return Ok(())
+                return Err(QueryError::Shutdown)
             },
 
-            result = self.handshake(handler) => {
-                // Handshake complete.
-                result?;
-                if self.state == ProtoState::Closed {
-                    return Ok(()); // EOF during handshake
-                }
+            handshake_r = self.handshake(handler) => {
+                handshake_r?;
             }
         );
 
@@ -435,17 +448,34 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received in run_message_loop");
-                Ok(None)
+                return Err(QueryError::Shutdown)
             },
             msg = self.read_message() => { msg },
         )? {
             trace!("got message {:?}", msg);
 
             let result = self.process_message(handler, msg, &mut query_string).await;
-            self.flush().await?;
+            tokio::select!(
+                biased;
+                _ = shutdown_watcher() => {
+                    // We were requested to shut down.
+                    tracing::info!("shutdown request received during response flush");
+
+                    // If we exited process_message with a shutdown error, there may be
+                    // some valid response content on in our transmit buffer: permit sending
+                    // this within a short timeout.  This is a best effort thing so we don't
+                    // care about the result.
+                    tokio::time::timeout(std::time::Duration::from_millis(500), self.flush()).await.ok();
+
+                    return Err(QueryError::Shutdown)
+                },
+                flush_r = self.flush() => {
+                    flush_r?;
+                }
+            );
+
             match result? {
                 ProcessMsgResult::Continue => {
-                    self.flush().await?;
                     continue;
                 }
                 ProcessMsgResult::Break => break,
@@ -550,7 +580,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during handshake".to_string()),
+                    )));
                 }
             }
         }
@@ -589,7 +621,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during auth".to_string()),
+                    )));
                 }
             }
         }
@@ -913,6 +947,7 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> AsyncWrite for CopyDataWriter<'a, I
 pub fn short_error(e: &QueryError) -> String {
     match e {
         QueryError::Disconnected(connection_error) => connection_error.to_string(),
+        QueryError::Shutdown => "shutdown".to_string(),
         QueryError::Other(e) => format!("{e:#}"),
     }
 }
@@ -929,6 +964,9 @@ fn log_query_error(query: &str, e: &QueryError) {
         QueryError::Disconnected(other_connection_error) => {
             error!("query handler for '{query}' failed with connection error: {other_connection_error:?}")
         }
+        QueryError::Shutdown => {
+            info!("query handler for '{query}' cancelled during tenant shutdown")
+        }
         QueryError::Other(e) => {
             error!("query handler for '{query}' failed: {e:?}");
         }

libs/postgres_ffi/src/lib.rs

@@ -131,6 +131,7 @@ pub const MAX_SEND_SIZE: usize = XLOG_BLCKSZ * 16;
 
 // Export some version independent functions that are used outside of this mod
 pub use v14::xlog_utils::encode_logical_message;
+pub use v14::xlog_utils::from_pg_timestamp;
 pub use v14::xlog_utils::get_current_timestamp;
 pub use v14::xlog_utils::to_pg_timestamp;
 pub use v14::xlog_utils::XLogFileName;

libs/postgres_ffi/src/pg_constants.rs

@@ -137,9 +137,12 @@ pub const XLOG_HEAP_INSERT: u8 = 0x00;
 pub const XLOG_HEAP_DELETE: u8 = 0x10;
 pub const XLOG_HEAP_UPDATE: u8 = 0x20;
 pub const XLOG_HEAP_HOT_UPDATE: u8 = 0x40;
+pub const XLOG_HEAP_LOCK: u8 = 0x60;
 pub const XLOG_HEAP_INIT_PAGE: u8 = 0x80;
 pub const XLOG_HEAP2_VISIBLE: u8 = 0x40;
 pub const XLOG_HEAP2_MULTI_INSERT: u8 = 0x50;
+pub const XLOG_HEAP2_LOCK_UPDATED: u8 = 0x60;
+pub const XLH_LOCK_ALL_FROZEN_CLEARED: u8 = 0x01;
 pub const XLH_INSERT_ALL_FROZEN_SET: u8 = (1 << 5) as u8;
 pub const XLH_INSERT_ALL_VISIBLE_CLEARED: u8 = (1 << 0) as u8;
 pub const XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED: u8 = (1 << 0) as u8;
@@ -217,6 +220,10 @@ pub const XLOG_CHECKPOINT_ONLINE: u8 = 0x10;
 pub const XLP_FIRST_IS_CONTRECORD: u16 = 0x0001;
 pub const XLP_LONG_HEADER: u16 = 0x0002;
 
+/* From replication/slot.h */
+pub const REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN: usize = 4*4  /* offset of `slotdata` in ReplicationSlotOnDisk  */
+   + 64 /* NameData */  + 4*4;
+
 /* From fsm_internals.h */
 const FSM_NODES_PER_PAGE: usize = BLCKSZ as usize - SIZEOF_PAGE_HEADER_DATA - 4;
 const FSM_NON_LEAF_NODES_PER_PAGE: usize = BLCKSZ as usize / 2 - 1;

libs/postgres_ffi/src/xlog_utils.rs

@@ -136,21 +136,42 @@ pub fn get_current_timestamp() -> TimestampTz {
     to_pg_timestamp(SystemTime::now())
 }
 
-pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
-    const UNIX_EPOCH_JDATE: u64 = 2440588; /* == date2j(1970, 1, 1) */
-    const POSTGRES_EPOCH_JDATE: u64 = 2451545; /* == date2j(2000, 1, 1) */
+// Module to reduce the scope of the constants
+mod timestamp_conversions {
+    use std::time::Duration;
+
+    use super::*;
+
+    const UNIX_EPOCH_JDATE: u64 = 2440588; // == date2j(1970, 1, 1)
+    const POSTGRES_EPOCH_JDATE: u64 = 2451545; // == date2j(2000, 1, 1)
     const SECS_PER_DAY: u64 = 86400;
     const USECS_PER_SEC: u64 = 1000000;
-    match time.duration_since(SystemTime::UNIX_EPOCH) {
-        Ok(n) => {
-            ((n.as_secs() - ((POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY))
-                * USECS_PER_SEC
-                + n.subsec_micros() as u64) as i64
+    const SECS_DIFF_UNIX_TO_POSTGRES_EPOCH: u64 =
+        (POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY;
+
+    pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
+        match time.duration_since(SystemTime::UNIX_EPOCH) {
+            Ok(n) => {
+                ((n.as_secs() - SECS_DIFF_UNIX_TO_POSTGRES_EPOCH) * USECS_PER_SEC
+                    + n.subsec_micros() as u64) as i64
+            }
+            Err(_) => panic!("SystemTime before UNIX EPOCH!"),
         }
-        Err(_) => panic!("SystemTime before UNIX EPOCH!"),
+    }
+
+    pub fn from_pg_timestamp(time: TimestampTz) -> SystemTime {
+        let time: u64 = time
+            .try_into()
+            .expect("timestamp before millenium (postgres epoch)");
+        let since_unix_epoch = time + SECS_DIFF_UNIX_TO_POSTGRES_EPOCH * USECS_PER_SEC;
+        SystemTime::UNIX_EPOCH
+            .checked_add(Duration::from_micros(since_unix_epoch))
+            .expect("SystemTime overflow")
     }
 }
 
+pub use timestamp_conversions::{from_pg_timestamp, to_pg_timestamp};
+
 // Returns (aligned) end_lsn of the last record in data_dir with WAL segments.
 // start_lsn must point to some previously known record boundary (beginning of
 // the next record). If no valid record after is found, start_lsn is returned
@@ -481,4 +502,24 @@ pub fn encode_logical_message(prefix: &str, message: &str) -> Vec<u8> {
     wal
 }
 
-// If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_ts_conversion() {
+        let now = SystemTime::now();
+        let round_trip = from_pg_timestamp(to_pg_timestamp(now));
+
+        let now_since = now.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        let round_trip_since = round_trip.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        assert_eq!(now_since.as_micros(), round_trip_since.as_micros());
+
+        let now_pg = get_current_timestamp();
+        let round_trip_pg = to_pg_timestamp(from_pg_timestamp(now_pg));
+
+        assert_eq!(now_pg, round_trip_pg);
+    }
+
+    // If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+}

libs/postgres_ffi/wal_craft/Cargo.toml

@@ -12,7 +12,7 @@ log.workspace = true
 once_cell.workspace = true
 postgres.workspace = true
 postgres_ffi.workspace = true
-tempfile.workspace = true
+camino-tempfile.workspace = true
 
 workspace_hack.workspace = true
 

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -1,4 +1,5 @@
 use anyhow::{bail, ensure};
+use camino_tempfile::{tempdir, Utf8TempDir};
 use log::*;
 use postgres::types::PgLsn;
 use postgres::Client;
@@ -8,7 +9,6 @@ use std::cmp::Ordering;
 use std::path::{Path, PathBuf};
 use std::process::Command;
 use std::time::{Duration, Instant};
-use tempfile::{tempdir, TempDir};
 
 macro_rules! xlog_utils_test {
     ($version:ident) => {
@@ -33,7 +33,7 @@ pub struct Conf {
 
 pub struct PostgresServer {
     process: std::process::Child,
-    _unix_socket_dir: TempDir,
+    _unix_socket_dir: Utf8TempDir,
     client_config: postgres::Config,
 }
 

libs/pq_proto/src/lib.rs

@@ -670,6 +670,7 @@ pub fn read_cstr(buf: &mut Bytes) -> Result<Bytes, ProtocolError> {
 }
 
 pub const SQLSTATE_INTERNAL_ERROR: &[u8; 5] = b"XX000";
+pub const SQLSTATE_ADMIN_SHUTDOWN: &[u8; 5] = b"57P01";
 pub const SQLSTATE_SUCCESSFUL_COMPLETION: &[u8; 5] = b"00000";
 
 impl<'a> BeMessage<'a> {

libs/remote_storage/Cargo.toml

@@ -13,6 +13,8 @@ aws-types.workspace = true
 aws-config.workspace = true
 aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
+bytes.workspace = true
+camino.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 serde.workspace = true
 serde_json.workspace = true
@@ -25,7 +27,15 @@ metrics.workspace = true
 utils.workspace = true
 pin-project-lite.workspace = true
 workspace_hack.workspace = true
+azure_core.workspace = true
+azure_identity.workspace = true
+azure_storage.workspace = true
+azure_storage_blobs.workspace = true
+futures-util.workspace = true
+http-types.workspace = true
+itertools.workspace = true
 
 [dev-dependencies]
-tempfile.workspace = true
+camino-tempfile.workspace = true
 test-context.workspace = true
+rand.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -0,0 +1,356 @@
+//! Azure Blob Storage wrapper
+
+use std::env;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+use std::{borrow::Cow, collections::HashMap, io::Cursor};
+
+use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
+use anyhow::Result;
+use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::Header;
+use azure_identity::DefaultAzureCredential;
+use azure_storage::StorageCredentials;
+use azure_storage_blobs::prelude::ClientBuilder;
+use azure_storage_blobs::{
+    blob::operations::GetBlobBuilder,
+    prelude::{BlobClient, ContainerClient},
+};
+use futures_util::StreamExt;
+use http_types::StatusCode;
+use tokio::io::AsyncRead;
+use tracing::debug;
+
+use crate::s3_bucket::RequestKind;
+use crate::{
+    AzureConfig, ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage,
+    StorageMetadata,
+};
+
+pub struct AzureBlobStorage {
+    client: ContainerClient,
+    prefix_in_container: Option<String>,
+    max_keys_per_list_response: Option<NonZeroU32>,
+    concurrency_limiter: ConcurrencyLimiter,
+}
+
+impl AzureBlobStorage {
+    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+        debug!(
+            "Creating azure remote storage for azure container {}",
+            azure_config.container_name
+        );
+
+        let account = env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT");
+
+        // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
+        // otherwise try the token based credentials.
+        let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
+            StorageCredentials::access_key(account.clone(), access_key)
+        } else {
+            let token_credential = DefaultAzureCredential::default();
+            StorageCredentials::token_credential(Arc::new(token_credential))
+        };
+
+        let builder = ClientBuilder::new(account, credentials);
+
+        let client = builder.container_client(azure_config.container_name.to_owned());
+
+        let max_keys_per_list_response =
+            if let Some(limit) = azure_config.max_keys_per_list_response {
+                Some(
+                    NonZeroU32::new(limit as u32)
+                        .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
+                )
+            } else {
+                None
+            };
+
+        Ok(AzureBlobStorage {
+            client,
+            prefix_in_container: azure_config.prefix_in_container.to_owned(),
+            max_keys_per_list_response,
+            concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+        })
+    }
+
+    pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
+        assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path
+            .get_path()
+            .as_str()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        match &self.prefix_in_container {
+            Some(prefix) => {
+                if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix.clone() + path_string
+                } else {
+                    format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
+                }
+            }
+            None => path_string.to_string(),
+        }
+    }
+
+    fn name_to_relative_path(&self, key: &str) -> RemotePath {
+        let relative_path =
+            match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
+                Some(stripped) => stripped,
+                // we rely on Azure to return properly prefixed paths
+                // for requests with a certain prefix
+                None => panic!(
+                    "Key {key} does not start with container prefix {:?}",
+                    self.prefix_in_container
+                ),
+            };
+        RemotePath(
+            relative_path
+                .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                .collect(),
+        )
+    }
+
+    async fn download_for_builder(
+        &self,
+        metadata: StorageMetadata,
+        builder: GetBlobBuilder,
+    ) -> Result<Download, DownloadError> {
+        let mut response = builder.into_stream();
+
+        // TODO give proper streaming response instead of buffering into RAM
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::new();
+        while let Some(part) = response.next().await {
+            let part = part.map_err(to_download_error)?;
+            let data = part
+                .data
+                .collect()
+                .await
+                .map_err(|e| DownloadError::Other(e.into()))?;
+            buf.extend_from_slice(&data.slice(..));
+        }
+        Ok(Download {
+            download_stream: Box::pin(Cursor::new(buf)),
+            metadata: Some(metadata),
+        })
+    }
+    // TODO get rid of this function once we have metadata included in the response
+    // https://github.com/Azure/azure-sdk-for-rust/issues/1439
+    async fn get_metadata(
+        &self,
+        blob_client: &BlobClient,
+    ) -> Result<StorageMetadata, DownloadError> {
+        let builder = blob_client.get_metadata();
+
+        let response = builder.into_future().await.map_err(to_download_error)?;
+        let mut map = HashMap::new();
+
+        for md in response.metadata.iter() {
+            map.insert(
+                md.name().as_str().to_string(),
+                md.value().as_str().to_string(),
+            );
+        }
+        Ok(StorageMetadata(map))
+    }
+
+    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+        self.concurrency_limiter
+            .acquire(kind)
+            .await
+            .expect("semaphore is never closed")
+    }
+}
+
+fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
+    let mut res = Metadata::new();
+    for (k, v) in metadata.0.into_iter() {
+        res.insert(k, v);
+    }
+    res
+}
+
+fn to_download_error(error: azure_core::Error) -> DownloadError {
+    if let Some(http_err) = error.as_http_error() {
+        match http_err.status() {
+            StatusCode::NotFound => DownloadError::NotFound,
+            StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
+            _ => DownloadError::Other(anyhow::Error::new(error)),
+        }
+    } else {
+        DownloadError::Other(error.into())
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for AzureBlobStorage {
+    async fn list_prefixes(
+        &self,
+        prefix: Option<&RemotePath>,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        // get the passed prefix or if it is not set use prefix_in_bucket value
+        let list_prefix = prefix
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone())
+            .map(|mut p| {
+                // required to end with a separator
+                // otherwise request will return only the entry of a prefix
+                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                }
+                p
+            });
+
+        let mut builder = self
+            .client
+            .list_blobs()
+            .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+
+        if let Some(prefix) = list_prefix {
+            builder = builder.prefix(Cow::from(prefix.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(entry) = response.next().await {
+            let entry = entry.map_err(to_download_error)?;
+            let name_iter = entry
+                .blobs
+                .prefixes()
+                .map(|prefix| self.name_to_relative_path(&prefix.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let folder_name = folder
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone());
+
+        let mut builder = self.client.list_blobs();
+
+        if let Some(folder_name) = folder_name {
+            builder = builder.prefix(Cow::from(folder_name.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(l) = response.next().await {
+            let entry = l.map_err(anyhow::Error::new)?;
+            let name_iter = entry
+                .blobs
+                .blobs()
+                .map(|bl| self.name_to_relative_path(&bl.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn upload(
+        &self,
+        mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Put).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+
+        // TODO FIX THIS UGLY HACK and don't buffer the entire object
+        // into RAM here, but use the streaming interface. For that,
+        // we'd have to change the interface though...
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::with_capacity(data_size_bytes);
+        tokio::io::copy(&mut from, &mut buf).await?;
+        let body = azure_core::Body::Bytes(buf.into());
+
+        let mut builder = blob_client.put_block_blob(body);
+
+        if let Some(metadata) = metadata {
+            builder = builder.metadata(to_azure_metadata(metadata));
+        }
+
+        let _response = builder.into_future().await?;
+
+        Ok(())
+    }
+
+    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let builder = blob_client.get();
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn download_byte_range(
+        &self,
+        from: &RemotePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+    ) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let mut builder = blob_client.get();
+
+        if let Some(end_exclusive) = end_exclusive {
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        } else {
+            // Open ranges are not supported by the SDK so we work around
+            // by setting the upper limit extremely high (but high enough
+            // to still be representable by signed 64 bit integers).
+            // TODO remove workaround once the SDK adds open range support
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1438
+            let end_exclusive = u64::MAX / 4;
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        }
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+        let builder = blob_client.delete();
+
+        match builder.into_future().await {
+            Ok(_response) => Ok(()),
+            Err(e) => {
+                if let Some(http_err) = e.as_http_error() {
+                    if http_err.status() == StatusCode::NotFound {
+                        return Ok(());
+                    }
+                }
+                Err(anyhow::Error::new(e))
+            }
+        }
+    }
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        // Permit is already obtained by inner delete function
+
+        // TODO batch requests are also not supported by the SDK
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+        for path in paths {
+            self.delete(path).await?;
+        }
+        Ok(())
+    }
+}

libs/remote_storage/src/lib.rs

@@ -4,7 +4,10 @@
 //! [`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
 //!   * [`s3_bucket`] uses AWS S3 bucket as an external storage
+//!   * [`azure_blob`] allows to use Azure Blob storage as an external storage
 //!
+
+mod azure_blob;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
@@ -13,18 +16,23 @@ use std::{
     collections::HashMap,
     fmt::Debug,
     num::{NonZeroU32, NonZeroUsize},
-    path::{Path, PathBuf},
     pin::Pin,
     sync::Arc,
 };
 
 use anyhow::{bail, Context};
+use camino::{Utf8Path, Utf8PathBuf};
 
-use tokio::io;
+use serde::{Deserialize, Serialize};
+use tokio::{io, sync::Semaphore};
 use toml_edit::Item;
 use tracing::info;
 
-pub use self::{local_fs::LocalFs, s3_bucket::S3Bucket, simulate_failures::UnreliableWrapper};
+pub use self::{
+    azure_blob::AzureBlobStorage, local_fs::LocalFs, s3_bucket::S3Bucket,
+    simulate_failures::UnreliableWrapper,
+};
+use s3_bucket::RequestKind;
 
 /// How many different timelines can be processed simultaneously when synchronizing layers with the remote storage.
 /// During regular work, pageserver produces one layer file per timeline checkpoint, with bursts of concurrency
@@ -38,26 +46,53 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
+/// We set this a little bit low as we currently buffer the entire file into RAM
+///
+/// Here, a limit of max 20k concurrent connections was noted.
+/// <https://learn.microsoft.com/en-us/answers/questions/1301863/is-there-any-limitation-to-concurrent-connections>
+pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 30;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
 /// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
+/// As defined in S3 docs
+pub const MAX_KEYS_PER_DELETE: usize = 1000;
+
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 
 /// Path on the remote storage, relative to some inner prefix.
 /// The prefix is an implementation detail, that allows representing local paths
 /// as the remote ones, stripping the local storage prefix away.
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
-pub struct RemotePath(PathBuf);
+pub struct RemotePath(Utf8PathBuf);
+
+impl Serialize for RemotePath {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        serializer.collect_str(self)
+    }
+}
+
+impl<'de> Deserialize<'de> for RemotePath {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let str = String::deserialize(deserializer)?;
+        Ok(Self(Utf8PathBuf::from(&str)))
+    }
+}
 
 impl std::fmt::Display for RemotePath {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.0.display())
+        std::fmt::Display::fmt(&self.0, f)
     }
 }
 
 impl RemotePath {
-    pub fn new(relative_path: &Path) -> anyhow::Result<Self> {
+    pub fn new(relative_path: &Utf8Path) -> anyhow::Result<Self> {
         anyhow::ensure!(
             relative_path.is_relative(),
             "Path {relative_path:?} is not relative"
@@ -66,27 +101,31 @@ impl RemotePath {
     }
 
     pub fn from_string(relative_path: &str) -> anyhow::Result<Self> {
-        Self::new(Path::new(relative_path))
+        Self::new(Utf8Path::new(relative_path))
     }
 
-    pub fn with_base(&self, base_path: &Path) -> PathBuf {
+    pub fn with_base(&self, base_path: &Utf8Path) -> Utf8PathBuf {
         base_path.join(&self.0)
     }
 
     pub fn object_name(&self) -> Option<&str> {
-        self.0.file_name().and_then(|os_str| os_str.to_str())
+        self.0.file_name()
     }
 
-    pub fn join(&self, segment: &Path) -> Self {
+    pub fn join(&self, segment: &Utf8Path) -> Self {
         Self(self.0.join(segment))
     }
 
-    pub fn get_path(&self) -> &PathBuf {
+    pub fn get_path(&self) -> &Utf8PathBuf {
         &self.0
     }
 
     pub fn extension(&self) -> Option<&str> {
-        self.0.extension()?.to_str()
+        self.0.extension()
+    }
+
+    pub fn strip_prefix(&self, p: &RemotePath) -> Result<&Utf8Path, std::path::StripPrefixError> {
+        self.0.strip_prefix(&p.0)
     }
 }
 
@@ -190,6 +229,7 @@ impl std::error::Error for DownloadError {}
 pub enum GenericRemoteStorage {
     LocalFs(LocalFs),
     AwsS3(Arc<S3Bucket>),
+    AzureBlob(Arc<AzureBlobStorage>),
     Unreliable(Arc<UnreliableWrapper>),
 }
 
@@ -201,6 +241,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_files(folder).await,
             Self::AwsS3(s) => s.list_files(folder).await,
+            Self::AzureBlob(s) => s.list_files(folder).await,
             Self::Unreliable(s) => s.list_files(folder).await,
         }
     }
@@ -215,6 +256,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_prefixes(prefix).await,
             Self::AwsS3(s) => s.list_prefixes(prefix).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
             Self::Unreliable(s) => s.list_prefixes(prefix).await,
         }
     }
@@ -229,6 +271,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
         }
     }
@@ -237,6 +280,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.download(from).await,
             Self::AwsS3(s) => s.download(from).await,
+            Self::AzureBlob(s) => s.download(from).await,
             Self::Unreliable(s) => s.download(from).await,
         }
     }
@@ -256,6 +300,10 @@ impl GenericRemoteStorage {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
             }
+            Self::AzureBlob(s) => {
+                s.download_byte_range(from, start_inclusive, end_exclusive)
+                    .await
+            }
             Self::Unreliable(s) => {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
@@ -267,6 +315,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete(path).await,
             Self::AwsS3(s) => s.delete(path).await,
+            Self::AzureBlob(s) => s.delete(path).await,
             Self::Unreliable(s) => s.delete(path).await,
         }
     }
@@ -275,6 +324,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete_objects(paths).await,
             Self::AwsS3(s) => s.delete_objects(paths).await,
+            Self::AzureBlob(s) => s.delete_objects(paths).await,
             Self::Unreliable(s) => s.delete_objects(paths).await,
         }
     }
@@ -284,7 +334,7 @@ impl GenericRemoteStorage {
     pub fn from_config(storage_config: &RemoteStorageConfig) -> anyhow::Result<Self> {
         Ok(match &storage_config.storage {
             RemoteStorageKind::LocalFs(root) => {
-                info!("Using fs root '{}' as a remote storage", root.display());
+                info!("Using fs root '{root}' as a remote storage");
                 Self::LocalFs(LocalFs::new(root.clone())?)
             }
             RemoteStorageKind::AwsS3(s3_config) => {
@@ -292,6 +342,11 @@ impl GenericRemoteStorage {
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
                 Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
             }
+            RemoteStorageKind::AzureContainer(azure_config) => {
+                info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
+                      azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+            }
         })
     }
 
@@ -352,10 +407,13 @@ pub struct RemoteStorageConfig {
 pub enum RemoteStorageKind {
     /// Storage based on local file system.
     /// Specify a root folder to place all stored files into.
-    LocalFs(PathBuf),
+    LocalFs(Utf8PathBuf),
     /// AWS S3 based storage, storing all files in the S3 bucket
     /// specified by the config
     AwsS3(S3Config),
+    /// Azure Blob based storage, storing all files in the container
+    /// specified by the config
+    AzureContainer(AzureConfig),
 }
 
 /// AWS S3 bucket coordinates and access credentials to manage the bucket contents (read and write).
@@ -395,11 +453,45 @@ impl Debug for S3Config {
     }
 }
 
+/// Azure  bucket coordinates and access credentials to manage the bucket contents (read and write).
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureConfig {
+    /// Name of the container to connect to.
+    pub container_name: String,
+    /// The region where the bucket is located at.
+    pub container_region: String,
+    /// A "subfolder" in the container, to use the same container separately by multiple remote storage users at once.
+    pub prefix_in_container: Option<String>,
+    /// Azure has various limits on its API calls, we need not to exceed those.
+    /// See [`DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT`] for more details.
+    pub concurrency_limit: NonZeroUsize,
+    pub max_keys_per_list_response: Option<i32>,
+}
+
+impl Debug for AzureConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AzureConfig")
+            .field("bucket_name", &self.container_name)
+            .field("bucket_region", &self.container_region)
+            .field("prefix_in_bucket", &self.prefix_in_container)
+            .field("concurrency_limit", &self.concurrency_limit)
+            .field(
+                "max_keys_per_list_response",
+                &self.max_keys_per_list_response,
+            )
+            .finish()
+    }
+}
+
 impl RemoteStorageConfig {
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
         let bucket_region = toml.get("bucket_region");
+        let container_name = toml.get("container_name");
+        let container_region = toml.get("container_region");
+
+        let use_azure = container_name.is_some() && container_region.is_some();
 
         let max_concurrent_syncs = NonZeroUsize::new(
             parse_optional_integer("max_concurrent_syncs", toml)?
@@ -413,9 +505,13 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'max_sync_errors' as a positive integer")?;
 
+        let default_concurrency_limit = if use_azure {
+            DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT
+        } else {
+            DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
+        };
         let concurrency_limit = NonZeroUsize::new(
-            parse_optional_integer("concurrency_limit", toml)?
-                .unwrap_or(DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
+            parse_optional_integer("concurrency_limit", toml)?.unwrap_or(default_concurrency_limit),
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
@@ -424,33 +520,70 @@ impl RemoteStorageConfig {
                 .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
                 .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
 
-        let storage = match (local_path, bucket_name, bucket_region) {
+        let endpoint = toml
+            .get("endpoint")
+            .map(|endpoint| parse_toml_string("endpoint", endpoint))
+            .transpose()?;
+
+        let storage = match (
+            local_path,
+            bucket_name,
+            bucket_region,
+            container_name,
+            container_region,
+        ) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
-            (None, None, None) => return Ok(None),
-            (_, Some(_), None) => {
+            (None, None, None, None, None) => return Ok(None),
+            (_, Some(_), None, ..) => {
                 bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
             }
-            (_, None, Some(_)) => {
+            (_, None, Some(_), ..) => {
                 bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
             }
-            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
-                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
-                prefix_in_bucket: toml
-                    .get("prefix_in_bucket")
-                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
-                    .transpose()?,
-                endpoint: toml
-                    .get("endpoint")
-                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
-                    .transpose()?,
-                concurrency_limit,
-                max_keys_per_list_response,
-            }),
-            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(PathBuf::from(
-                parse_toml_string("local_path", local_path)?,
-            )),
-            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
+            (None, Some(bucket_name), Some(bucket_region), ..) => {
+                RemoteStorageKind::AwsS3(S3Config {
+                    bucket_name: parse_toml_string("bucket_name", bucket_name)?,
+                    bucket_region: parse_toml_string("bucket_region", bucket_region)?,
+                    prefix_in_bucket: toml
+                        .get("prefix_in_bucket")
+                        .map(|prefix_in_bucket| {
+                            parse_toml_string("prefix_in_bucket", prefix_in_bucket)
+                        })
+                        .transpose()?,
+                    endpoint,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (_, _, _, Some(_), None) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (_, _, _, None, Some(_)) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (None, None, None, Some(container_name), Some(container_region)) => {
+                RemoteStorageKind::AzureContainer(AzureConfig {
+                    container_name: parse_toml_string("container_name", container_name)?,
+                    container_region: parse_toml_string("container_region", container_region)?,
+                    prefix_in_container: toml
+                        .get("prefix_in_container")
+                        .map(|prefix_in_container| {
+                            parse_toml_string("prefix_in_container", prefix_in_container)
+                        })
+                        .transpose()?,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (Some(local_path), None, None, None, None) => RemoteStorageKind::LocalFs(
+                Utf8PathBuf::from(parse_toml_string("local_path", local_path)?),
+            ),
+            (Some(_), Some(_), ..) => {
+                bail!("'local_path' and 'bucket_name' are mutually exclusive")
+            }
+            (Some(_), _, _, Some(_), Some(_)) => {
+                bail!("local_path and 'container_name' are mutually exclusive")
+            }
         };
 
         Ok(Some(RemoteStorageConfig {
@@ -486,29 +619,69 @@ fn parse_toml_string(name: &str, item: &Item) -> anyhow::Result<String> {
     Ok(s.to_string())
 }
 
+struct ConcurrencyLimiter {
+    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
+    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
+    // The helps to ensure we don't exceed the thresholds.
+    write: Arc<Semaphore>,
+    read: Arc<Semaphore>,
+}
+
+impl ConcurrencyLimiter {
+    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
+        match kind {
+            RequestKind::Get => &self.read,
+            RequestKind::Put => &self.write,
+            RequestKind::List => &self.read,
+            RequestKind::Delete => &self.write,
+        }
+    }
+
+    async fn acquire(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
+        self.for_kind(kind).acquire().await
+    }
+
+    async fn acquire_owned(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
+        Arc::clone(self.for_kind(kind)).acquire_owned().await
+    }
+
+    fn new(limit: usize) -> ConcurrencyLimiter {
+        Self {
+            read: Arc::new(Semaphore::new(limit)),
+            write: Arc::new(Semaphore::new(limit)),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
 
     #[test]
     fn test_object_name() {
-        let k = RemotePath::new(Path::new("a/b/c")).unwrap();
+        let k = RemotePath::new(Utf8Path::new("a/b/c")).unwrap();
         assert_eq!(k.object_name(), Some("c"));
 
-        let k = RemotePath::new(Path::new("a/b/c/")).unwrap();
+        let k = RemotePath::new(Utf8Path::new("a/b/c/")).unwrap();
         assert_eq!(k.object_name(), Some("c"));
 
-        let k = RemotePath::new(Path::new("a/")).unwrap();
+        let k = RemotePath::new(Utf8Path::new("a/")).unwrap();
         assert_eq!(k.object_name(), Some("a"));
 
         // XXX is it impossible to have an empty key?
-        let k = RemotePath::new(Path::new("")).unwrap();
+        let k = RemotePath::new(Utf8Path::new("")).unwrap();
         assert_eq!(k.object_name(), None);
     }
 
     #[test]
     fn rempte_path_cannot_be_created_from_absolute_ones() {
-        let err = RemotePath::new(Path::new("/")).expect_err("Should fail on absolute paths");
+        let err = RemotePath::new(Utf8Path::new("/")).expect_err("Should fail on absolute paths");
         assert_eq!(err.to_string(), "Path \"/\" is not relative");
     }
 }

libs/remote_storage/src/local_fs.rs

@@ -4,15 +4,10 @@
 //! This storage used in tests, but can also be used in cases when a certain persistent
 //! volume is mounted to the local FS.
 
-use std::{
-    borrow::Cow,
-    future::Future,
-    io::ErrorKind,
-    path::{Path, PathBuf},
-    pin::Pin,
-};
+use std::{borrow::Cow, future::Future, io::ErrorKind, pin::Pin};
 
 use anyhow::{bail, ensure, Context};
+use camino::{Utf8Path, Utf8PathBuf};
 use tokio::{
     fs,
     io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
@@ -28,20 +23,20 @@ const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
 
 #[derive(Debug, Clone)]
 pub struct LocalFs {
-    storage_root: PathBuf,
+    storage_root: Utf8PathBuf,
 }
 
 impl LocalFs {
     /// Attempts to create local FS storage, along with its root directory.
     /// Storage root will be created (if does not exist) and transformed into an absolute path (if passed as relative).
-    pub fn new(mut storage_root: PathBuf) -> anyhow::Result<Self> {
+    pub fn new(mut storage_root: Utf8PathBuf) -> anyhow::Result<Self> {
         if !storage_root.exists() {
             std::fs::create_dir_all(&storage_root).with_context(|| {
                 format!("Failed to create all directories in the given root path {storage_root:?}")
             })?;
         }
         if !storage_root.is_absolute() {
-            storage_root = storage_root.canonicalize().with_context(|| {
+            storage_root = storage_root.canonicalize_utf8().with_context(|| {
                 format!("Failed to represent path {storage_root:?} as an absolute path")
             })?;
         }
@@ -50,7 +45,7 @@ impl LocalFs {
     }
 
     // mirrors S3Bucket::s3_object_to_relative_path
-    fn local_file_to_relative_path(&self, key: PathBuf) -> RemotePath {
+    fn local_file_to_relative_path(&self, key: Utf8PathBuf) -> RemotePath {
         let relative_path = key
             .strip_prefix(&self.storage_root)
             .expect("relative path must contain storage_root as prefix");
@@ -59,22 +54,18 @@ impl LocalFs {
 
     async fn read_storage_metadata(
         &self,
-        file_path: &Path,
+        file_path: &Utf8Path,
     ) -> anyhow::Result<Option<StorageMetadata>> {
         let metadata_path = storage_metadata_path(file_path);
         if metadata_path.exists() && metadata_path.is_file() {
             let metadata_string = fs::read_to_string(&metadata_path).await.with_context(|| {
-                format!(
-                    "Failed to read metadata from the local storage at '{}'",
-                    metadata_path.display()
-                )
+                format!("Failed to read metadata from the local storage at '{metadata_path}'")
             })?;
 
             serde_json::from_str(&metadata_string)
                 .with_context(|| {
                     format!(
-                        "Failed to deserialize metadata from the local storage at '{}'",
-                        metadata_path.display()
+                        "Failed to deserialize metadata from the local storage at '{metadata_path}'",
                     )
                 })
                 .map(|metadata| Some(StorageMetadata(metadata)))
@@ -171,25 +162,21 @@ impl RemoteStorage for LocalFs {
             }
         }
 
-        // Note that PathBuf starts_with only considers full path segments, but
+        // Note that Utf8PathBuf starts_with only considers full path segments, but
         // object prefixes are arbitrary strings, so we need the strings for doing
         // starts_with later.
-        let prefix = full_path.to_string_lossy();
+        let prefix = full_path.as_str();
 
         let mut files = vec![];
-        let mut directory_queue = vec![initial_dir.clone()];
+        let mut directory_queue = vec![initial_dir];
         while let Some(cur_folder) = directory_queue.pop() {
-            let mut entries = fs::read_dir(cur_folder.clone()).await?;
-            while let Some(entry) = entries.next_entry().await? {
-                let file_name: PathBuf = entry.file_name().into();
-                let full_file_name = cur_folder.clone().join(&file_name);
-                if full_file_name
-                    .to_str()
-                    .map(|s| s.starts_with(prefix.as_ref()))
-                    .unwrap_or(false)
-                {
+            let mut entries = cur_folder.read_dir_utf8()?;
+            while let Some(Ok(entry)) = entries.next() {
+                let file_name = entry.file_name();
+                let full_file_name = cur_folder.join(file_name);
+                if full_file_name.as_str().starts_with(prefix) {
                     let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
-                    files.push(file_remote_path.clone());
+                    files.push(file_remote_path);
                     if full_file_name.is_dir() {
                         directory_queue.push(full_file_name);
                     }
@@ -230,10 +217,7 @@ impl RemoteStorage for LocalFs {
                 .open(&temp_file_path)
                 .await
                 .with_context(|| {
-                    format!(
-                        "Failed to open target fs destination at '{}'",
-                        target_file_path.display()
-                    )
+                    format!("Failed to open target fs destination at '{target_file_path}'")
                 })?,
         );
 
@@ -244,8 +228,7 @@ impl RemoteStorage for LocalFs {
             .await
             .with_context(|| {
                 format!(
-                    "Failed to upload file (write temp) to the local storage at '{}'",
-                    temp_file_path.display()
+                    "Failed to upload file (write temp) to the local storage at '{temp_file_path}'",
                 )
             })?;
 
@@ -262,8 +245,7 @@ impl RemoteStorage for LocalFs {
 
         destination.flush().await.with_context(|| {
             format!(
-                "Failed to upload (flush temp) file to the local storage at '{}'",
-                temp_file_path.display()
+                "Failed to upload (flush temp) file to the local storage at '{temp_file_path}'",
             )
         })?;
 
@@ -271,8 +253,7 @@ impl RemoteStorage for LocalFs {
             .await
             .with_context(|| {
                 format!(
-                    "Failed to upload (rename) file to the local storage at '{}'",
-                    target_file_path.display()
+                    "Failed to upload (rename) file to the local storage at '{target_file_path}'",
                 )
             })?;
 
@@ -286,8 +267,7 @@ impl RemoteStorage for LocalFs {
             .await
             .with_context(|| {
                 format!(
-                    "Failed to write metadata to the local storage at '{}'",
-                    storage_metadata_path.display()
+                    "Failed to write metadata to the local storage at '{storage_metadata_path}'",
                 )
             })?;
         }
@@ -393,16 +373,16 @@ impl RemoteStorage for LocalFs {
     }
 }
 
-fn storage_metadata_path(original_path: &Path) -> PathBuf {
+fn storage_metadata_path(original_path: &Utf8Path) -> Utf8PathBuf {
     path_with_suffix_extension(original_path, "metadata")
 }
 
 fn get_all_files<'a, P>(
     directory_path: P,
     recursive: bool,
-) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<PathBuf>>> + Send + Sync + 'a>>
+) -> Pin<Box<dyn Future<Output = anyhow::Result<Vec<Utf8PathBuf>>> + Send + Sync + 'a>>
 where
-    P: AsRef<Path> + Send + Sync + 'a,
+    P: AsRef<Utf8Path> + Send + Sync + 'a,
 {
     Box::pin(async move {
         let directory_path = directory_path.as_ref();
@@ -412,7 +392,13 @@ where
                 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();
+                    let entry_path =
+                        Utf8PathBuf::from_path_buf(dir_entry.path()).map_err(|pb| {
+                            anyhow::Error::msg(format!(
+                                "non-Unicode path: {}",
+                                pb.to_string_lossy()
+                            ))
+                        })?;
                     if file_type.is_symlink() {
                         debug!("{entry_path:?} is a symlink, skipping")
                     } else if file_type.is_dir() {
@@ -435,13 +421,10 @@ where
     })
 }
 
-async fn create_target_directory(target_file_path: &Path) -> anyhow::Result<()> {
+async fn create_target_directory(target_file_path: &Utf8Path) -> 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()
-        ),
+        None => bail!("File path '{target_file_path}' has no parent directory"),
     };
     if !target_dir.exists() {
         fs::create_dir_all(target_dir).await?;
@@ -449,13 +432,9 @@ async fn create_target_directory(target_file_path: &Path) -> anyhow::Result<()>
     Ok(())
 }
 
-fn file_exists(file_path: &Path) -> anyhow::Result<bool> {
+fn file_exists(file_path: &Utf8Path) -> anyhow::Result<bool> {
     if file_path.exists() {
-        ensure!(
-            file_path.is_file(),
-            "file path '{}' is not a file",
-            file_path.display()
-        );
+        ensure!(file_path.is_file(), "file path '{file_path}' is not a file");
         Ok(true)
     } else {
         Ok(false)
@@ -466,13 +445,13 @@ fn file_exists(file_path: &Path) -> anyhow::Result<bool> {
 mod fs_tests {
     use super::*;
 
+    use camino_tempfile::tempdir;
     use std::{collections::HashMap, io::Write};
-    use tempfile::tempdir;
 
     async fn read_and_assert_remote_file_contents(
         storage: &LocalFs,
         #[allow(clippy::ptr_arg)]
-        // have to use &PathBuf due to `storage.local_path` parameter requirements
+        // have to use &Utf8PathBuf due to `storage.local_path` parameter requirements
         remote_storage_path: &RemotePath,
         expected_metadata: Option<&StorageMetadata>,
     ) -> anyhow::Result<String> {
@@ -519,7 +498,7 @@ mod fs_tests {
     async fn upload_file_negatives() -> anyhow::Result<()> {
         let storage = create_storage()?;
 
-        let id = RemotePath::new(Path::new("dummy"))?;
+        let id = RemotePath::new(Utf8Path::new("dummy"))?;
         let content = std::io::Cursor::new(b"12345");
 
         // Check that you get an error if the size parameter doesn't match the actual
@@ -544,7 +523,8 @@ mod fs_tests {
     }
 
     fn create_storage() -> anyhow::Result<LocalFs> {
-        LocalFs::new(tempdir()?.path().to_owned())
+        let storage_root = tempdir()?.path().to_path_buf();
+        LocalFs::new(storage_root)
     }
 
     #[tokio::test]
@@ -561,7 +541,7 @@ mod fs_tests {
         );
 
         let non_existing_path = "somewhere/else";
-        match storage.download(&RemotePath::new(Path::new(non_existing_path))?).await {
+        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?).await {
             Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
             other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
         }
@@ -775,7 +755,7 @@ mod fs_tests {
     }
 
     async fn create_file_for_upload(
-        path: &Path,
+        path: &Utf8Path,
         contents: &str,
     ) -> anyhow::Result<(io::BufReader<fs::File>, usize)> {
         std::fs::create_dir_all(path.parent().unwrap())?;

libs/remote_storage/src/s3_bucket.rs

@@ -4,7 +4,7 @@
 //! allowing multiple api users to independently work with the same S3 bucket, if
 //! their bucket prefixes are both specified and different.
 
-use std::sync::Arc;
+use std::borrow::Cow;
 
 use anyhow::Context;
 use aws_config::{
@@ -24,23 +24,20 @@ use aws_sdk_s3::{
 use aws_smithy_http::body::SdkBody;
 use hyper::Body;
 use scopeguard::ScopeGuard;
-use tokio::{
-    io::{self, AsyncRead},
-    sync::Semaphore,
-};
+use tokio::io::{self, AsyncRead};
 use tokio_util::io::ReaderStream;
 use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
-    Download, DownloadError, RemotePath, RemoteStorage, S3Config, REMOTE_STORAGE_PREFIX_SEPARATOR,
+    ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage, S3Config,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
-const MAX_DELETE_OBJECTS_REQUEST_SIZE: usize = 1000;
-
 pub(super) mod metrics;
 
-use self::metrics::{AttemptOutcome, RequestKind};
+use self::metrics::AttemptOutcome;
+pub(super) use self::metrics::RequestKind;
 
 /// AWS S3 storage.
 pub struct S3Bucket {
@@ -48,10 +45,7 @@ pub struct S3Bucket {
     bucket_name: String,
     prefix_in_bucket: Option<String>,
     max_keys_per_list_response: Option<i32>,
-    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
-    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
-    // The helps to ensure we don't exceed the thresholds.
-    concurrency_limiter: Arc<Semaphore>,
+    concurrency_limiter: ConcurrencyLimiter,
 }
 
 #[derive(Default)]
@@ -118,7 +112,7 @@ impl S3Bucket {
             bucket_name: aws_config.bucket_name.clone(),
             max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
-            concurrency_limiter: Arc::new(Semaphore::new(aws_config.concurrency_limit.get())),
+            concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
         })
     }
 
@@ -144,12 +138,11 @@ impl S3Bucket {
         assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
         let path_string = path
             .get_path()
-            .to_string_lossy()
-            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR)
-            .to_string();
+            .as_str()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
         match &self.prefix_in_bucket {
-            Some(prefix) => prefix.clone() + "/" + &path_string,
-            None => path_string,
+            Some(prefix) => prefix.clone() + "/" + path_string,
+            None => path_string.to_string(),
         }
     }
 
@@ -157,7 +150,7 @@ impl S3Bucket {
         let started_at = start_counting_cancelled_wait(kind);
         let permit = self
             .concurrency_limiter
-            .acquire()
+            .acquire(kind)
             .await
             .expect("semaphore is never closed");
 
@@ -173,8 +166,7 @@ impl S3Bucket {
         let started_at = start_counting_cancelled_wait(kind);
         let permit = self
             .concurrency_limiter
-            .clone()
-            .acquire_owned()
+            .acquire_owned(kind)
             .await
             .expect("semaphore is never closed");
 
@@ -500,7 +492,7 @@ impl RemoteStorage for S3Bucket {
             delete_objects.push(obj_id);
         }
 
-        for chunk in delete_objects.chunks(MAX_DELETE_OBJECTS_REQUEST_SIZE) {
+        for chunk in delete_objects.chunks(MAX_KEYS_PER_DELETE) {
             let started_at = start_measuring_requests(kind);
 
             let resp = self
@@ -522,6 +514,20 @@ impl RemoteStorage for S3Bucket {
                         .deleted_objects_total
                         .inc_by(chunk.len() as u64);
                     if let Some(errors) = resp.errors {
+                        // Log a bounded number of the errors within the response:
+                        // these requests can carry 1000 keys so logging each one
+                        // would be too verbose, especially as errors may lead us
+                        // to retry repeatedly.
+                        const LOG_UP_TO_N_ERRORS: usize = 10;
+                        for e in errors.iter().take(LOG_UP_TO_N_ERRORS) {
+                            tracing::warn!(
+                                "DeleteObjects key {} failed: {}: {}",
+                                e.key.as_ref().map(Cow::from).unwrap_or("".into()),
+                                e.code.as_ref().map(Cow::from).unwrap_or("".into()),
+                                e.message.as_ref().map(Cow::from).unwrap_or("".into())
+                            );
+                        }
+
                         return Err(anyhow::format_err!(
                             "Failed to delete {} objects",
                             errors.len()
@@ -566,8 +572,8 @@ fn start_measuring_requests(
 
 #[cfg(test)]
 mod tests {
+    use camino::Utf8Path;
     use std::num::NonZeroUsize;
-    use std::path::Path;
 
     use crate::{RemotePath, S3Bucket, S3Config};
 
@@ -576,7 +582,7 @@ mod tests {
         let all_paths = ["", "some/path", "some/path/"];
         let all_paths: Vec<RemotePath> = all_paths
             .iter()
-            .map(|x| RemotePath::new(Path::new(x)).expect("bad path"))
+            .map(|x| RemotePath::new(Utf8Path::new(x)).expect("bad path"))
             .collect();
         let prefixes = [
             None,

libs/remote_storage/src/s3_bucket/metrics.rs

@@ -6,7 +6,7 @@ use once_cell::sync::Lazy;
 pub(super) static BUCKET_METRICS: Lazy<BucketMetrics> = Lazy::new(Default::default);
 
 #[derive(Clone, Copy, Debug)]
-pub(super) enum RequestKind {
+pub(crate) enum RequestKind {
     Get = 0,
     Put = 1,
     Delete = 2,

libs/remote_storage/tests/test_real_azure.rs

@@ -0,0 +1,625 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use camino::Utf8Path;
+use once_cell::sync::OnceCell;
+use remote_storage::{
+    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
+
+const BASE_PREFIX: &str = "test";
+
+/// Tests that the Azure client can list all prefixes, even if the response comes paginated and requires multiple HTTP queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified.
+/// See the client creation in [`create_azure_client`] for details on the required env vars.
+/// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// where
+/// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real Azure enabled and `#[cfg(test)]` Rust configuration used, the Azure client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to Azure.
+///
+/// Lastly, the test attempts to clean up and remove all uploaded Azure files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledAzureWithTestBlobs)]
+#[tokio::test]
+async fn azure_pagination_should_work(
+    ctx: &mut MaybeEnabledAzureWithTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
+        .context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+/// Tests that Azure client can list all files in a folder, even if the response comes paginated and requirees multiple Azure queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
+/// See `Azure_pagination_should_work` for more information.
+///
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// Then performs the following queries:
+///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
+#[test_context(MaybeEnabledAzureWithSimpleTestBlobs)]
+#[tokio::test]
+async fn azure_list_files_works(
+    ctx: &mut MaybeEnabledAzureWithSimpleTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithSimpleTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithSimpleTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithSimpleTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let base_prefix =
+        RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
+    let root_files = test_client
+        .list_files(None)
+        .await
+        .context("client list root files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_files,
+        ctx.remote_blobs.clone(),
+        "remote storage list_files on root mismatches with the uploads."
+    );
+    let nested_remote_files = test_client
+        .list_files(Some(&base_prefix))
+        .await
+        .context("client list nested files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let trim_remote_blobs: HashSet<_> = ctx
+        .remote_blobs
+        .iter()
+        .map(|x| x.get_path())
+        .filter(|x| x.starts_with("folder1"))
+        .map(|x| RemotePath::new(x).expect("must be valid path"))
+        .collect();
+    assert_eq!(
+        nested_remote_files, trim_remote_blobs,
+        "remote storage list_files on subdirrectory mismatches with the uploads."
+    );
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_non_exising_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
+    ))
+    .with_context(|| "RemotePath conversion")?;
+
+    ctx.client.delete(&path).await.expect("should succeed");
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_objects_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path2 = RemotePath::new(Utf8Path::new(format!("{}/path2", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data1 = "remote blob data1".as_bytes();
+    let data1_len = data1.len();
+    let data2 = "remote blob data2".as_bytes();
+    let data2_len = data2.len();
+    let data3 = "remote blob data3".as_bytes();
+    let data3_len = data3.len();
+    ctx.client
+        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
+        .await?;
+
+    ctx.client.delete_objects(&[path1, path2]).await?;
+
+    let prefixes = ctx.client.list_prefixes(None).await?;
+
+    assert_eq!(prefixes.len(), 1);
+
+    ctx.client.delete_objects(&[path3]).await?;
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let MaybeEnabledAzure::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data = "remote blob data here".as_bytes();
+    let data_len = data.len() as u64;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data), data.len(), &path, None)
+        .await?;
+
+    async fn download_and_compare(mut dl: Download) -> anyhow::Result<Vec<u8>> {
+        let mut buf = Vec::new();
+        tokio::io::copy(&mut dl.download_stream, &mut buf).await?;
+        Ok(buf)
+    }
+    // Normal download request
+    let dl = ctx.client.download(&path).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(data_len))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // partial range (end specified)
+    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..10]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(data_len * 100))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[8..]);
+
+    // Partial range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..]);
+
+    // Full range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
+}
+
+fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+        )
+        .expect("logging init failed");
+    });
+}
+
+struct EnabledAzure {
+    client: Arc<GenericRemoteStorage>,
+    base_prefix: &'static str,
+}
+
+impl EnabledAzure {
+    async fn setup(max_keys_in_list_response: Option<i32>) -> Self {
+        let client = create_azure_client(max_keys_in_list_response)
+            .context("Azure client creation")
+            .expect("Azure client creation failed");
+
+        EnabledAzure {
+            client,
+            base_prefix: BASE_PREFIX,
+        }
+    }
+}
+
+enum MaybeEnabledAzure {
+    Enabled(EnabledAzure),
+    Disabled,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzure {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        Self::Enabled(EnabledAzure::setup(None).await)
+    }
+}
+
+enum MaybeEnabledAzureWithTestBlobs {
+    Enabled(AzureWithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithTestBlobs),
+}
+
+struct AzureWithTestBlobs {
+    enabled: EnabledAzure,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+// NOTE: the setups for the list_prefixes test and the list_files test are very similar
+// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
+// whereas the list_files function is concerned with listing files.
+// See `RemoteStorage::list_files` documentation for more details
+enum MaybeEnabledAzureWithSimpleTestBlobs {
+    Enabled(AzureWithSimpleTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithSimpleTestBlobs),
+}
+struct AzureWithSimpleTestBlobs {
+    enabled: EnabledAzure,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_azure_client(
+    max_keys_per_list_response: Option<i32>,
+) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    use rand::Rng;
+
+    let remote_storage_azure_container = env::var("REMOTE_STORAGE_AZURE_CONTAINER").context(
+        "`REMOTE_STORAGE_AZURE_CONTAINER` env var is not set, but real Azure tests are enabled",
+    )?;
+    let remote_storage_azure_region = env::var("REMOTE_STORAGE_AZURE_REGION").context(
+        "`REMOTE_STORAGE_AZURE_REGION` env var is not set, but real Azure tests are enabled",
+    )?;
+
+    // due to how time works, we've had test runners use the same nanos as bucket prefixes.
+    // millis is just a debugging aid for easier finding the prefix later.
+    let millis = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random Azure test prefix part calculation")?
+        .as_millis();
+
+    // because nanos can be the same for two threads so can millis, add randomness
+    let random = rand::thread_rng().gen::<u32>();
+
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AzureContainer(AzureConfig {
+            container_name: remote_storage_azure_container,
+            container_region: remote_storage_azure_region,
+            prefix_in_container: Some(format!("test_{millis}_{random:08x}/")),
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response,
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+async fn upload_simple_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}

libs/remote_storage/tests/test_real_s3.rs

@@ -2,11 +2,12 @@ use std::collections::HashSet;
 use std::env;
 use std::num::{NonZeroU32, NonZeroUsize};
 use std::ops::ControlFlow;
-use std::path::{Path, PathBuf};
+use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
+use camino::Utf8Path;
 use once_cell::sync::OnceCell;
 use remote_storage::{
     GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
@@ -55,7 +56,7 @@ async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3WithTestBlobs) -> any
     let test_client = Arc::clone(&ctx.enabled.client);
     let expected_remote_prefixes = ctx.remote_prefixes.clone();
 
-    let base_prefix = RemotePath::new(Path::new(ctx.enabled.base_prefix))
+    let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
         .context("common_prefix construction")?;
     let root_remote_prefixes = test_client
         .list_prefixes(None)
@@ -108,7 +109,7 @@ async fn s3_list_files_works(ctx: &mut MaybeEnabledS3WithSimpleTestBlobs) -> any
     };
     let test_client = Arc::clone(&ctx.enabled.client);
     let base_prefix =
-        RemotePath::new(Path::new("folder1")).context("common_prefix construction")?;
+        RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
     let root_files = test_client
         .list_files(None)
         .await
@@ -129,9 +130,9 @@ async fn s3_list_files_works(ctx: &mut MaybeEnabledS3WithSimpleTestBlobs) -> any
     let trim_remote_blobs: HashSet<_> = ctx
         .remote_blobs
         .iter()
-        .map(|x| x.get_path().to_str().expect("must be valid name"))
+        .map(|x| x.get_path())
         .filter(|x| x.starts_with("folder1"))
-        .map(|x| RemotePath::new(Path::new(x)).expect("must be valid name"))
+        .map(|x| RemotePath::new(x).expect("must be valid path"))
         .collect();
     assert_eq!(
         nested_remote_files, trim_remote_blobs,
@@ -148,10 +149,9 @@ async fn s3_delete_non_exising_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result
         MaybeEnabledS3::Disabled => return Ok(()),
     };
 
-    let path = RemotePath::new(&PathBuf::from(format!(
-        "{}/for_sure_there_is_nothing_there_really",
-        ctx.base_prefix,
-    )))
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
+    ))
     .with_context(|| "RemotePath conversion")?;
 
     ctx.client.delete(&path).await.expect("should succeed");
@@ -167,13 +167,13 @@ async fn s3_delete_objects_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()>
         MaybeEnabledS3::Disabled => return Ok(()),
     };
 
-    let path1 = RemotePath::new(&PathBuf::from(format!("{}/path1", ctx.base_prefix,)))
+    let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
-    let path2 = RemotePath::new(&PathBuf::from(format!("{}/path2", ctx.base_prefix,)))
+    let path2 = RemotePath::new(Utf8Path::new(format!("{}/path2", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
-    let path3 = RemotePath::new(&PathBuf::from(format!("{}/path3", ctx.base_prefix,)))
+    let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
     let data1 = "remote blob data1".as_bytes();
@@ -378,21 +378,30 @@ impl AsyncTestContext for MaybeEnabledS3WithSimpleTestBlobs {
 fn create_s3_client(
     max_keys_per_list_response: Option<i32>,
 ) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    use rand::Rng;
+
     let remote_storage_s3_bucket = env::var("REMOTE_STORAGE_S3_BUCKET")
         .context("`REMOTE_STORAGE_S3_BUCKET` env var is not set, but real S3 tests are enabled")?;
     let remote_storage_s3_region = env::var("REMOTE_STORAGE_S3_REGION")
         .context("`REMOTE_STORAGE_S3_REGION` env var is not set, but real S3 tests are enabled")?;
-    let random_prefix_part = std::time::SystemTime::now()
+
+    // due to how time works, we've had test runners use the same nanos as bucket prefixes.
+    // millis is just a debugging aid for easier finding the prefix later.
+    let millis = std::time::SystemTime::now()
         .duration_since(UNIX_EPOCH)
         .context("random s3 test prefix part calculation")?
-        .as_nanos();
+        .as_millis();
+
+    // because nanos can be the same for two threads so can millis, add randomness
+    let random = rand::thread_rng().gen::<u32>();
+
     let remote_storage_config = RemoteStorageConfig {
         max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
         max_sync_errors: NonZeroU32::new(5).unwrap(),
         storage: RemoteStorageKind::AwsS3(S3Config {
             bucket_name: remote_storage_s3_bucket,
             bucket_region: remote_storage_s3_region,
-            prefix_in_bucket: Some(format!("pagination_should_work_test_{random_prefix_part}/")),
+            prefix_in_bucket: Some(format!("test_{millis}_{random:08x}/")),
             endpoint: None,
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
@@ -418,10 +427,10 @@ async fn upload_s3_data(
     for i in 1..upload_tasks_count + 1 {
         let task_client = Arc::clone(client);
         upload_tasks.spawn(async move {
-            let prefix = PathBuf::from(format!("{base_prefix_str}/sub_prefix_{i}/"));
-            let blob_prefix = RemotePath::new(&prefix)
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
                 .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Path::new(&format!("blob_{i}")));
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
             debug!("Creating remote item {i} at path {blob_path:?}");
 
             let data = format!("remote blob data {i}").into_bytes();
@@ -503,8 +512,10 @@ async fn upload_simple_s3_data(
         let task_client = Arc::clone(client);
         upload_tasks.spawn(async move {
             let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
-            let blob_path = RemotePath::new(&blob_path)
-                .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
             debug!("Creating remote item {i} at path {blob_path:?}");
 
             let data = format!("remote blob data {i}").into_bytes();

libs/utils/Cargo.toml

@@ -10,6 +10,7 @@ async-trait.workspace = true
 anyhow.workspace = true
 bincode.workspace = true
 bytes.workspace = true
+camino.workspace = true
 chrono.workspace = true
 heapless.workspace = true
 hex = { workspace = true, features = ["serde"] }
@@ -53,7 +54,7 @@ byteorder.workspace = true
 bytes.workspace = true
 criterion.workspace = true
 hex-literal.workspace = true
-tempfile.workspace = true
+camino-tempfile.workspace = true
 
 [[bench]]
 name = "benchmarks"

libs/utils/src/auth.rs

@@ -2,9 +2,9 @@
 
 use serde;
 use std::fs;
-use std::path::Path;
 
 use anyhow::Result;
+use camino::Utf8Path;
 use jsonwebtoken::{
     decode, encode, Algorithm, DecodingKey, EncodingKey, Header, TokenData, Validation,
 };
@@ -65,7 +65,7 @@ impl JwtAuth {
         }
     }
 
-    pub fn from_key_path(key_path: &Path) -> Result<Self> {
+    pub fn from_key_path(key_path: &Utf8Path) -> Result<Self> {
         let public_key = fs::read(key_path)?;
         Ok(Self::new(DecodingKey::from_ed_pem(&public_key)?))
     }

libs/utils/src/crashsafe.rs

@@ -1,14 +1,14 @@
 use std::{
     borrow::Cow,
-    ffi::OsStr,
     fs::{self, File},
     io,
-    path::{Path, PathBuf},
 };
 
+use camino::{Utf8Path, Utf8PathBuf};
+
 /// Similar to [`std::fs::create_dir`], except we fsync the
 /// created directory and its parent.
-pub fn create_dir(path: impl AsRef<Path>) -> io::Result<()> {
+pub fn create_dir(path: impl AsRef<Utf8Path>) -> io::Result<()> {
     let path = path.as_ref();
 
     fs::create_dir(path)?;
@@ -18,7 +18,7 @@ pub fn create_dir(path: impl AsRef<Path>) -> io::Result<()> {
 
 /// Similar to [`std::fs::create_dir_all`], except we fsync all
 /// newly created directories and the pre-existing parent.
-pub fn create_dir_all(path: impl AsRef<Path>) -> io::Result<()> {
+pub fn create_dir_all(path: impl AsRef<Utf8Path>) -> io::Result<()> {
     let mut path = path.as_ref();
 
     let mut dirs_to_create = Vec::new();
@@ -30,7 +30,7 @@ pub fn create_dir_all(path: impl AsRef<Path>) -> io::Result<()> {
             Ok(_) => {
                 return Err(io::Error::new(
                     io::ErrorKind::AlreadyExists,
-                    format!("non-directory found in path: {}", path.display()),
+                    format!("non-directory found in path: {path}"),
                 ));
             }
             Err(ref e) if e.kind() == io::ErrorKind::NotFound => {}
@@ -44,7 +44,7 @@ pub fn create_dir_all(path: impl AsRef<Path>) -> io::Result<()> {
             None => {
                 return Err(io::Error::new(
                     io::ErrorKind::InvalidInput,
-                    format!("can't find parent of path '{}'", path.display()).as_str(),
+                    format!("can't find parent of path '{path}'"),
                 ));
             }
         }
@@ -70,21 +70,18 @@ pub fn create_dir_all(path: impl AsRef<Path>) -> io::Result<()> {
 
 /// Adds a suffix to the file(directory) name, either appending the suffix to the end of its extension,
 /// or if there's no extension, creates one and puts a suffix there.
-pub fn path_with_suffix_extension(original_path: impl AsRef<Path>, suffix: &str) -> PathBuf {
-    let new_extension = match original_path
-        .as_ref()
-        .extension()
-        .map(OsStr::to_string_lossy)
-    {
+pub fn path_with_suffix_extension(
+    original_path: impl AsRef<Utf8Path>,
+    suffix: &str,
+) -> Utf8PathBuf {
+    let new_extension = match original_path.as_ref().extension() {
         Some(extension) => Cow::Owned(format!("{extension}.{suffix}")),
         None => Cow::Borrowed(suffix),
     };
-    original_path
-        .as_ref()
-        .with_extension(new_extension.as_ref())
+    original_path.as_ref().with_extension(new_extension)
 }
 
-pub fn fsync_file_and_parent(file_path: &Path) -> io::Result<()> {
+pub fn fsync_file_and_parent(file_path: &Utf8Path) -> io::Result<()> {
     let parent = file_path.parent().ok_or_else(|| {
         io::Error::new(
             io::ErrorKind::Other,
@@ -97,7 +94,7 @@ pub fn fsync_file_and_parent(file_path: &Path) -> io::Result<()> {
     Ok(())
 }
 
-pub fn fsync(path: &Path) -> io::Result<()> {
+pub fn fsync(path: &Utf8Path) -> io::Result<()> {
     File::open(path)
         .map_err(|e| io::Error::new(e.kind(), format!("Failed to open the file {path:?}: {e}")))
         .and_then(|file| {
@@ -111,19 +108,18 @@ pub fn fsync(path: &Path) -> io::Result<()> {
         .map_err(|e| io::Error::new(e.kind(), format!("Failed to fsync file {path:?}: {e}")))
 }
 
-pub async fn fsync_async(path: impl AsRef<std::path::Path>) -> Result<(), std::io::Error> {
-    tokio::fs::File::open(path).await?.sync_all().await
+pub async fn fsync_async(path: impl AsRef<Utf8Path>) -> Result<(), std::io::Error> {
+    tokio::fs::File::open(path.as_ref()).await?.sync_all().await
 }
 
 #[cfg(test)]
 mod tests {
-    use tempfile::tempdir;
 
     use super::*;
 
     #[test]
     fn test_create_dir_fsyncd() {
-        let dir = tempdir().unwrap();
+        let dir = camino_tempfile::tempdir().unwrap();
 
         let existing_dir_path = dir.path();
         let err = create_dir(existing_dir_path).unwrap_err();
@@ -139,7 +135,7 @@ mod tests {
 
     #[test]
     fn test_create_dir_all_fsyncd() {
-        let dir = tempdir().unwrap();
+        let dir = camino_tempfile::tempdir().unwrap();
 
         let existing_dir_path = dir.path();
         create_dir_all(existing_dir_path).unwrap();
@@ -166,29 +162,29 @@ mod tests {
 
     #[test]
     fn test_path_with_suffix_extension() {
-        let p = PathBuf::from("/foo/bar");
+        let p = Utf8PathBuf::from("/foo/bar");
         assert_eq!(
-            &path_with_suffix_extension(p, "temp").to_string_lossy(),
+            &path_with_suffix_extension(p, "temp").to_string(),
             "/foo/bar.temp"
         );
-        let p = PathBuf::from("/foo/bar");
+        let p = Utf8PathBuf::from("/foo/bar");
         assert_eq!(
-            &path_with_suffix_extension(p, "temp.temp").to_string_lossy(),
+            &path_with_suffix_extension(p, "temp.temp").to_string(),
             "/foo/bar.temp.temp"
         );
-        let p = PathBuf::from("/foo/bar.baz");
+        let p = Utf8PathBuf::from("/foo/bar.baz");
         assert_eq!(
-            &path_with_suffix_extension(p, "temp.temp").to_string_lossy(),
+            &path_with_suffix_extension(p, "temp.temp").to_string(),
             "/foo/bar.baz.temp.temp"
         );
-        let p = PathBuf::from("/foo/bar.baz");
+        let p = Utf8PathBuf::from("/foo/bar.baz");
         assert_eq!(
-            &path_with_suffix_extension(p, ".temp").to_string_lossy(),
+            &path_with_suffix_extension(p, ".temp").to_string(),
             "/foo/bar.baz..temp"
         );
-        let p = PathBuf::from("/foo/bar/dir/");
+        let p = Utf8PathBuf::from("/foo/bar/dir/");
         assert_eq!(
-            &path_with_suffix_extension(p, ".temp").to_string_lossy(),
+            &path_with_suffix_extension(p, ".temp").to_string(),
             "/foo/bar/dir..temp"
         );
     }

libs/utils/src/fs_ext.rs

@@ -55,8 +55,6 @@ where
 
 #[cfg(test)]
 mod test {
-    use std::path::PathBuf;
-
     use crate::fs_ext::{is_directory_empty, list_dir};
 
     use super::ignore_absent_files;
@@ -65,7 +63,7 @@ mod test {
     fn is_empty_dir() {
         use super::PathExt;
 
-        let dir = tempfile::tempdir().unwrap();
+        let dir = camino_tempfile::tempdir().unwrap();
         let dir_path = dir.path();
 
         // test positive case
@@ -75,7 +73,7 @@ mod test {
         );
 
         // invoke on a file to ensure it returns an error
-        let file_path: PathBuf = dir_path.join("testfile");
+        let file_path = dir_path.join("testfile");
         let f = std::fs::File::create(&file_path).unwrap();
         drop(f);
         assert!(file_path.is_empty_dir().is_err());
@@ -87,7 +85,7 @@ mod test {
 
     #[tokio::test]
     async fn is_empty_dir_async() {
-        let dir = tempfile::tempdir().unwrap();
+        let dir = camino_tempfile::tempdir().unwrap();
         let dir_path = dir.path();
 
         // test positive case
@@ -97,7 +95,7 @@ mod test {
         );
 
         // invoke on a file to ensure it returns an error
-        let file_path: PathBuf = dir_path.join("testfile");
+        let file_path = dir_path.join("testfile");
         let f = std::fs::File::create(&file_path).unwrap();
         drop(f);
         assert!(is_directory_empty(&file_path).await.is_err());
@@ -109,10 +107,9 @@ mod test {
 
     #[test]
     fn ignore_absent_files_works() {
-        let dir = tempfile::tempdir().unwrap();
-        let dir_path = dir.path();
+        let dir = camino_tempfile::tempdir().unwrap();
 
-        let file_path: PathBuf = dir_path.join("testfile");
+        let file_path = dir.path().join("testfile");
 
         ignore_absent_files(|| std::fs::remove_file(&file_path)).expect("should execute normally");
 
@@ -126,17 +123,17 @@ mod test {
 
     #[tokio::test]
     async fn list_dir_works() {
-        let dir = tempfile::tempdir().unwrap();
+        let dir = camino_tempfile::tempdir().unwrap();
         let dir_path = dir.path();
 
         assert!(list_dir(dir_path).await.unwrap().is_empty());
 
-        let file_path: PathBuf = dir_path.join("testfile");
+        let file_path = dir_path.join("testfile");
         let _ = std::fs::File::create(&file_path).unwrap();
 
         assert_eq!(&list_dir(dir_path).await.unwrap(), &["testfile"]);
 
-        let another_dir_path: PathBuf = dir_path.join("testdir");
+        let another_dir_path = dir_path.join("testdir");
         std::fs::create_dir(another_dir_path).unwrap();
 
         let expected = &["testdir", "testfile"];

libs/utils/src/generation.rs

@@ -89,6 +89,22 @@ impl Generation {
             Self::Broken => panic!("Attempted to use a broken generation"),
         }
     }
+
+    pub fn next(&self) -> Generation {
+        match self {
+            Self::Valid(n) => Self::Valid(*n + 1),
+            Self::None => Self::Valid(1),
+            Self::Broken => panic!("Attempted to use a broken generation"),
+        }
+    }
+
+    pub fn into(self) -> Option<u32> {
+        if let Self::Valid(v) = self {
+            Some(v)
+        } else {
+            None
+        }
+    }
 }
 
 impl Serialize for Generation {

libs/utils/src/http/error.rs

@@ -1,8 +1,9 @@
 use hyper::{header, Body, Response, StatusCode};
 use serde::{Deserialize, Serialize};
+use std::borrow::Cow;
 use std::error::Error as StdError;
 use thiserror::Error;
-use tracing::error;
+use tracing::{error, info};
 
 #[derive(Debug, Error)]
 pub enum ApiError {
@@ -24,6 +25,12 @@ pub enum ApiError {
     #[error("Precondition failed: {0}")]
     PreconditionFailed(Box<str>),
 
+    #[error("Resource temporarily unavailable: {0}")]
+    ResourceUnavailable(Cow<'static, str>),
+
+    #[error("Shutting down")]
+    ShuttingDown,
+
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -52,6 +59,14 @@ impl ApiError {
                 self.to_string(),
                 StatusCode::PRECONDITION_FAILED,
             ),
+            ApiError::ShuttingDown => HttpErrorBody::response_from_msg_and_status(
+                "Shutting down".to_string(),
+                StatusCode::SERVICE_UNAVAILABLE,
+            ),
+            ApiError::ResourceUnavailable(err) => HttpErrorBody::response_from_msg_and_status(
+                err.to_string(),
+                StatusCode::SERVICE_UNAVAILABLE,
+            ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                 err.to_string(),
                 StatusCode::INTERNAL_SERVER_ERROR,
@@ -101,10 +116,12 @@ pub async fn route_error_handler(err: routerify::RouteError) -> Response<Body> {
 
 pub fn api_error_handler(api_error: ApiError) -> Response<Body> {
     // Print a stack trace for Internal Server errors
-    if let ApiError::InternalServerError(_) = api_error {
-        error!("Error processing HTTP request: {api_error:?}");
-    } else {
-        error!("Error processing HTTP request: {api_error:#}");
+
+    match api_error {
+        ApiError::ResourceUnavailable(_) => info!("Error processing HTTP request: {api_error:#}"),
+        ApiError::NotFound(_) => info!("Error processing HTTP request: {api_error:#}"),
+        ApiError::InternalServerError(_) => error!("Error processing HTTP request: {api_error:?}"),
+        _ => error!("Error processing HTTP request: {api_error:#}"),
     }
 
     api_error.into_response()

libs/utils/src/id.rs

@@ -1,4 +1,3 @@
-use std::ffi::OsStr;
 use std::{fmt, str::FromStr};
 
 use anyhow::Context;
@@ -215,12 +214,11 @@ pub struct TimelineId(Id);
 
 id_newtype!(TimelineId);
 
-impl TryFrom<Option<&OsStr>> for TimelineId {
+impl TryFrom<Option<&str>> for TimelineId {
     type Error = anyhow::Error;
 
-    fn try_from(value: Option<&OsStr>) -> Result<Self, Self::Error> {
+    fn try_from(value: Option<&str>) -> Result<Self, Self::Error> {
         value
-            .and_then(OsStr::to_str)
             .unwrap_or_default()
             .parse::<TimelineId>()
             .with_context(|| format!("Could not parse timeline id from {:?}", value))

libs/utils/src/lock_file.rs

@@ -11,10 +11,10 @@ use std::{
     io::{Read, Write},
     ops::Deref,
     os::unix::prelude::AsRawFd,
-    path::{Path, PathBuf},
 };
 
 use anyhow::Context;
+use camino::{Utf8Path, Utf8PathBuf};
 use nix::{errno::Errno::EAGAIN, fcntl};
 
 use crate::crashsafe;
@@ -23,7 +23,7 @@ use crate::crashsafe;
 /// Returned by [`create_exclusive`].
 #[must_use]
 pub struct UnwrittenLockFile {
-    path: PathBuf,
+    path: Utf8PathBuf,
     file: fs::File,
 }
 
@@ -60,7 +60,7 @@ impl UnwrittenLockFile {
 ///
 /// It is not an error if the file already exists.
 /// It is an error if the file is already locked.
-pub fn create_exclusive(lock_file_path: &Path) -> anyhow::Result<UnwrittenLockFile> {
+pub fn create_exclusive(lock_file_path: &Utf8Path) -> anyhow::Result<UnwrittenLockFile> {
     let lock_file = fs::OpenOptions::new()
         .create(true) // O_CREAT
         .write(true)
@@ -101,7 +101,7 @@ pub enum LockFileRead {
 /// Open & try to lock the lock file at the given `path`, returning a [handle][`LockFileRead`] to
 /// inspect its content. It is not an `Err(...)` if the file does not exist or is already locked.
 /// Check the [`LockFileRead`] variants for details.
-pub fn read_and_hold_lock_file(path: &Path) -> anyhow::Result<LockFileRead> {
+pub fn read_and_hold_lock_file(path: &Utf8Path) -> anyhow::Result<LockFileRead> {
     let res = fs::OpenOptions::new().read(true).open(path);
     let mut lock_file = match res {
         Ok(f) => f,

libs/utils/src/logging.rs

@@ -216,6 +216,30 @@ impl std::fmt::Debug for PrettyLocation<'_, '_> {
     }
 }
 
+/// When you will store a secret but want to make sure it won't
+/// be accidentally logged, wrap it in a SecretString, whose Debug
+/// implementation does not expose the contents.
+#[derive(Clone, Eq, PartialEq)]
+pub struct SecretString(String);
+
+impl SecretString {
+    pub fn get_contents(&self) -> &str {
+        self.0.as_str()
+    }
+}
+
+impl From<String> for SecretString {
+    fn from(s: String) -> Self {
+        Self(s)
+    }
+}
+
+impl std::fmt::Debug for SecretString {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "[SECRET]")
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use metrics::{core::Opts, IntCounterVec};

libs/utils/src/lsn.rs

@@ -1,9 +1,9 @@
 #![warn(missing_docs)]
 
+use camino::Utf8Path;
 use serde::{Deserialize, Serialize};
 use std::fmt;
 use std::ops::{Add, AddAssign};
-use std::path::Path;
 use std::str::FromStr;
 use std::sync::atomic::{AtomicU64, Ordering};
 
@@ -44,11 +44,9 @@ impl Lsn {
     /// Parse an LSN from a filename in the form `0000000000000000`
     pub fn from_filename<F>(filename: F) -> Result<Self, LsnParseError>
     where
-        F: AsRef<Path>,
+        F: AsRef<Utf8Path>,
     {
-        let filename: &Path = filename.as_ref();
-        let filename = filename.to_str().ok_or(LsnParseError)?;
-        Lsn::from_hex(filename)
+        Lsn::from_hex(filename.as_ref().as_str())
     }
 
     /// Parse an LSN from a string in the form `0000000000000000`

libs/utils/src/pid_file.rs

@@ -49,9 +49,10 @@
 //! At this point, `B` and `C` are running, which is hazardous.
 //! Morale of the story: don't unlink pidfiles, ever.
 
-use std::{ops::Deref, path::Path};
+use std::ops::Deref;
 
 use anyhow::Context;
+use camino::Utf8Path;
 use nix::unistd::Pid;
 
 use crate::lock_file::{self, LockFileRead};
@@ -84,7 +85,7 @@ impl Deref for PidFileGuard {
 /// The claim ends as soon as the returned guard object is dropped.
 /// To maintain the claim for the remaining lifetime of the current process,
 /// use [`std::mem::forget`] or similar.
-pub fn claim_for_current_process(path: &Path) -> anyhow::Result<PidFileGuard> {
+pub fn claim_for_current_process(path: &Utf8Path) -> anyhow::Result<PidFileGuard> {
     let unwritten_lock_file = lock_file::create_exclusive(path).context("lock file")?;
     // if any of the next steps fail, we drop the file descriptor and thereby release the lock
     let guard = unwritten_lock_file
@@ -132,7 +133,7 @@ pub enum PidFileRead {
 ///
 /// On success, this function returns a [`PidFileRead`].
 /// Check its docs for a description of the meaning of its different variants.
-pub fn read(pidfile: &Path) -> anyhow::Result<PidFileRead> {
+pub fn read(pidfile: &Utf8Path) -> anyhow::Result<PidFileRead> {
     let res = lock_file::read_and_hold_lock_file(pidfile).context("read and hold pid file")?;
     let ret = match res {
         LockFileRead::NotExist => PidFileRead::NotExist,

libs/utils/src/seqwait.rs

@@ -58,7 +58,7 @@ where
 // to get that.
 impl<T: Ord> PartialOrd for Waiter<T> {
     fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        other.wake_num.partial_cmp(&self.wake_num)
+        Some(self.cmp(other))
     }
 }
 

libs/vm_monitor/README.md

@@ -27,8 +27,8 @@ and old one if it exists.
 * the filecache: a struct that allows communication with the Postgres file cache.
 On startup, we connect to the filecache and hold on to the connection for the
 entire monitor lifetime.
-* the cgroup watcher: the `CgroupWatcher` manages the `neon-postgres` cgroup by
-listening for `memory.high` events and setting its `memory.{high,max}` values.
+* the cgroup watcher: the `CgroupWatcher` polls the `neon-postgres` cgroup's memory
+usage and sends rolling aggregates to the runner.
 * the runner: the runner marries the filecache and cgroup watcher together,
 communicating with the agent throught the `Dispatcher`, and then calling filecache
 and cgroup watcher functions as needed to upscale and downscale

libs/vm_monitor/src/cgroup.rs

@@ -1,161 +1,38 @@
-use std::{
-    fmt::{Debug, Display},
-    fs,
-    pin::pin,
-    sync::atomic::{AtomicU64, Ordering},
-};
+use std::fmt::{self, Debug, Formatter};
+use std::time::{Duration, Instant};
 
-use anyhow::{anyhow, bail, Context};
+use anyhow::{anyhow, Context};
 use cgroups_rs::{
-    freezer::FreezerController,
-    hierarchies::{self, is_cgroup2_unified_mode, UNIFIED_MOUNTPOINT},
+    hierarchies::{self, is_cgroup2_unified_mode},
     memory::MemController,
-    MaxValue,
-    Subsystem::{Freezer, Mem},
+    Subsystem,
 };
-use inotify::{EventStream, Inotify, WatchMask};
-use tokio::sync::mpsc::{self, error::TryRecvError};
-use tokio::time::{Duration, Instant};
-use tokio_stream::{Stream, StreamExt};
+use tokio::sync::watch;
 use tracing::{info, warn};
 
-use crate::protocol::Resources;
-use crate::MiB;
-
-/// Monotonically increasing counter of the number of memory.high events
-/// the cgroup has experienced.
-///
-/// We use this to determine if a modification to the `memory.events` file actually
-/// changed the `high` field. If not, we don't care about the change. When we
-/// read the file, we check the `high` field in the file against `MEMORY_EVENT_COUNT`
-/// to see if it changed since last time.
-pub static MEMORY_EVENT_COUNT: AtomicU64 = AtomicU64::new(0);
-
-/// Monotonically increasing counter that gives each cgroup event a unique id.
-///
-/// This allows us to answer questions like "did this upscale arrive before this
-/// memory.high?". This static is also used by the `Sequenced` type to "tag" values
-/// with a sequence number. As such, prefer to used the `Sequenced` type rather
-/// than this static directly.
-static EVENT_SEQUENCE_NUMBER: AtomicU64 = AtomicU64::new(0);
-
-/// A memory event type reported in memory.events.
-#[derive(Debug, Eq, PartialEq, Copy, Clone)]
-pub enum MemoryEvent {
-    Low,
-    High,
-    Max,
-    Oom,
-    OomKill,
-    OomGroupKill,
-}
-
-impl MemoryEvent {
-    fn as_str(&self) -> &str {
-        match self {
-            MemoryEvent::Low => "low",
-            MemoryEvent::High => "high",
-            MemoryEvent::Max => "max",
-            MemoryEvent::Oom => "oom",
-            MemoryEvent::OomKill => "oom_kill",
-            MemoryEvent::OomGroupKill => "oom_group_kill",
-        }
-    }
-}
-
-impl Display for MemoryEvent {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.as_str())
-    }
-}
-
 /// Configuration for a `CgroupWatcher`
 #[derive(Debug, Clone)]
 pub struct Config {
-    // The target difference between the total memory reserved for the cgroup
-    // and the value of the cgroup's memory.high.
-    //
-    // In other words, memory.high + oom_buffer_bytes will equal the total memory that the cgroup may
-    // use (equal to system memory, minus whatever's taken out for the file cache).
-    oom_buffer_bytes: u64,
+    /// Interval at which we should be fetching memory statistics
+    memory_poll_interval: Duration,
 
-    // The amount of memory, in bytes, below a proposed new value for
-    // memory.high that the cgroup's memory usage must be for us to downscale
-    //
-    // In other words, we can downscale only when:
-    //
-    //   memory.current + memory_high_buffer_bytes < (proposed) memory.high
-    //
-    // TODO: there's some minor issues with this approach -- in particular, that we might have
-    // memory in use by the kernel's page cache that we're actually ok with getting rid of.
-    pub(crate) memory_high_buffer_bytes: u64,
-
-    // The maximum duration, in milliseconds, that we're allowed to pause
-    // the cgroup for while waiting for the autoscaler-agent to upscale us
-    max_upscale_wait: Duration,
-
-    // The required minimum time, in milliseconds, that we must wait before re-freezing
-    // the cgroup while waiting for the autoscaler-agent to upscale us.
-    do_not_freeze_more_often_than: Duration,
-
-    // The amount of memory, in bytes, that we should periodically increase memory.high
-    // by while waiting for the autoscaler-agent to upscale us.
-    //
-    // This exists to avoid the excessive throttling that happens when a cgroup is above its
-    // memory.high for too long. See more here:
-    // https://github.com/neondatabase/autoscaling/issues/44#issuecomment-1522487217
-    memory_high_increase_by_bytes: u64,
-
-    // The period, in milliseconds, at which we should repeatedly increase the value
-    // of the cgroup's memory.high while we're waiting on upscaling and memory.high
-    // is still being hit.
-    //
-    // Technically speaking, this actually serves as a rate limit to moderate responding to
-    // memory.high events, but these are roughly equivalent if the process is still allocating
-    // memory.
-    memory_high_increase_every: Duration,
-}
-
-impl Config {
-    /// Calculate the new value for the cgroups memory.high based on system memory
-    pub fn calculate_memory_high_value(&self, total_system_mem: u64) -> u64 {
-        total_system_mem.saturating_sub(self.oom_buffer_bytes)
-    }
+    /// The number of samples used in constructing aggregated memory statistics
+    memory_history_len: usize,
+    /// The number of most recent samples that will be periodically logged.
+    ///
+    /// Each sample is logged exactly once. Increasing this value means that recent samples will be
+    /// logged less frequently, and vice versa.
+    ///
+    /// For simplicity, this value must be greater than or equal to `memory_history_len`.
+    memory_history_log_interval: usize,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
-            oom_buffer_bytes: 100 * MiB,
-            memory_high_buffer_bytes: 100 * MiB,
-            // while waiting for upscale, don't freeze for more than 20ms every 1s
-            max_upscale_wait: Duration::from_millis(20),
-            do_not_freeze_more_often_than: Duration::from_millis(1000),
-            // while waiting for upscale, increase memory.high by 10MiB every 25ms
-            memory_high_increase_by_bytes: 10 * MiB,
-            memory_high_increase_every: Duration::from_millis(25),
-        }
-    }
-}
-
-/// Used to represent data that is associated with a certain point in time, such
-/// as an upscale request or memory.high event.
-///
-/// Internally, creating a `Sequenced` uses a static atomic counter to obtain
-/// a unique sequence number. Sequence numbers are monotonically increasing,
-/// allowing us to answer questions like "did this upscale happen after this
-/// memory.high event?" by comparing the sequence numbers of the two events.
-#[derive(Debug, Clone)]
-pub struct Sequenced<T> {
-    seqnum: u64,
-    data: T,
-}
-
-impl<T> Sequenced<T> {
-    pub fn new(data: T) -> Self {
-        Self {
-            seqnum: EVENT_SEQUENCE_NUMBER.fetch_add(1, Ordering::AcqRel),
-            data,
+            memory_poll_interval: Duration::from_millis(100),
+            memory_history_len: 5, // use 500ms of history for decision-making
+            memory_history_log_interval: 20, // but only log every ~2s (otherwise it's spammy)
         }
     }
 }
@@ -170,74 +47,14 @@ impl<T> Sequenced<T> {
 pub struct CgroupWatcher {
     pub config: Config,
 
-    /// The sequence number of the last upscale.
-    ///
-    /// If we receive a memory.high event that has a _lower_ sequence number than
-    /// `last_upscale_seqnum`, then we know it occured before the upscale, and we
-    /// can safely ignore it.
-    ///
-    /// Note: Like the `events` field, this doesn't _need_ interior mutability but we
-    /// use it anyways so that methods take `&self`, not `&mut self`.
-    last_upscale_seqnum: AtomicU64,
-
-    /// A channel on which we send messages to request upscale from the dispatcher.
-    upscale_requester: mpsc::Sender<()>,
-
     /// The actual cgroup we are watching and managing.
     cgroup: cgroups_rs::Cgroup,
 }
 
-/// Read memory.events for the desired event type.
-///
-/// `path` specifies the path to the desired `memory.events` file.
-/// For more info, see the `memory.events` section of the [kernel docs]
-/// <https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>
-fn get_event_count(path: &str, event: MemoryEvent) -> anyhow::Result<u64> {
-    let contents = fs::read_to_string(path)
-        .with_context(|| format!("failed to read memory.events from {path}"))?;
-
-    // Then contents of the file look like:
-    // low 42
-    // high 101
-    // ...
-    contents
-        .lines()
-        .filter_map(|s| s.split_once(' '))
-        .find(|(e, _)| *e == event.as_str())
-        .ok_or_else(|| anyhow!("failed to find entry for memory.{event} events in {path}"))
-        .and_then(|(_, count)| {
-            count
-                .parse::<u64>()
-                .with_context(|| format!("failed to parse memory.{event} as u64"))
-        })
-}
-
-/// Create an event stream that produces events whenever the file at the provided
-/// path is modified.
-fn create_file_watcher(path: &str) -> anyhow::Result<EventStream<[u8; 1024]>> {
-    info!("creating file watcher for {path}");
-    let inotify = Inotify::init().context("failed to initialize file watcher")?;
-    inotify
-        .watches()
-        .add(path, WatchMask::MODIFY)
-        .with_context(|| format!("failed to start watching {path}"))?;
-    inotify
-        // The inotify docs use [0u8; 1024] so we'll just copy them. We only need
-        // to store one event at a time - if the event gets written over, that's
-        // ok. We still see that there is an event. For more information, see:
-        // https://man7.org/linux/man-pages/man7/inotify.7.html
-        .into_event_stream([0u8; 1024])
-        .context("failed to start inotify event stream")
-}
-
 impl CgroupWatcher {
     /// Create a new `CgroupWatcher`.
     #[tracing::instrument(skip_all, fields(%name))]
-    pub fn new(
-        name: String,
-        // A channel on which to send upscale requests
-        upscale_requester: mpsc::Sender<()>,
-    ) -> anyhow::Result<(Self, impl Stream<Item = Sequenced<u64>>)> {
+    pub fn new(name: String) -> anyhow::Result<Self> {
         // TODO: clarify exactly why we need v2
         // Make sure cgroups v2 (aka unified) are supported
         if !is_cgroup2_unified_mode() {
@@ -245,451 +62,203 @@ impl CgroupWatcher {
         }
         let cgroup = cgroups_rs::Cgroup::load(hierarchies::auto(), &name);
 
-        // Start monitoring the cgroup for memory events. In general, for
-        // cgroups v2 (aka unified), metrics are reported in files like
-        // > `/sys/fs/cgroup/{name}/{metric}`
-        // We are looking for `memory.high` events, which are stored in the
-        // file `memory.events`. For more info, see the `memory.events` section
-        // of https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
-        let path = format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name);
-        let memory_events = create_file_watcher(&path)
-            .with_context(|| format!("failed to create event watcher for {path}"))?
-            // This would be nice with with .inspect_err followed by .ok
-            .filter_map(move |_| match get_event_count(&path, MemoryEvent::High) {
-                Ok(high) => Some(high),
-                Err(error) => {
-                    // TODO: Might want to just panic here
-                    warn!(?error, "failed to read high events count from {}", &path);
-                    None
-                }
-            })
-            // Only report the event if the memory.high count increased
-            .filter_map(|high| {
-                if MEMORY_EVENT_COUNT.fetch_max(high, Ordering::AcqRel) < high {
-                    Some(high)
-                } else {
-                    None
-                }
-            })
-            .map(Sequenced::new);
-
-        let initial_count = get_event_count(
-            &format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name),
-            MemoryEvent::High,
-        )?;
-
-        info!(initial_count, "initial memory.high event count");
-
-        // Hard update `MEMORY_EVENT_COUNT` since there could have been processes
-        // running in the cgroup before that caused it to be non-zero.
-        MEMORY_EVENT_COUNT.fetch_max(initial_count, Ordering::AcqRel);
-
-        Ok((
-            Self {
-                cgroup,
-                upscale_requester,
-                last_upscale_seqnum: AtomicU64::new(0),
-                config: Default::default(),
-            },
-            memory_events,
-        ))
+        Ok(Self {
+            cgroup,
+            config: Default::default(),
+        })
     }
 
     /// The entrypoint for the `CgroupWatcher`.
     #[tracing::instrument(skip_all)]
-    pub async fn watch<E>(
+    pub async fn watch(
         &self,
-        // These are ~dependency injected~ (fancy, I know) because this function
-        // should never return.
-        // -> therefore: when we tokio::spawn it, we don't await the JoinHandle.
-        // -> therefore: if we want to stick it in an Arc so many threads can access
-        //    it, methods can never take mutable access.
-        //     - note: we use the Arc strategy so that a) we can call this function
-        //             right here and b) the runner can call the set/get_memory methods
-        // -> since calling recv() on a tokio::sync::mpsc::Receiver takes &mut self,
-        //    we just pass them in here instead of holding them in fields, as that
-        //    would require this method to take &mut self.
-        mut upscales: mpsc::Receiver<Sequenced<Resources>>,
-        events: E,
-    ) -> anyhow::Result<()>
-    where
-        E: Stream<Item = Sequenced<u64>>,
-    {
-        let mut wait_to_freeze = pin!(tokio::time::sleep(Duration::ZERO));
-        let mut last_memory_high_increase_at: Option<Instant> = None;
-        let mut events = pin!(events);
-
-        // Are we waiting to be upscaled? Could be true if we request upscale due
-        // to a memory.high event and it does not arrive in time.
-        let mut waiting_on_upscale = false;
-
-        loop {
-            tokio::select! {
-                upscale = upscales.recv() => {
-                    let Sequenced { seqnum, data } = upscale
-                        .context("failed to listen on upscale notification channel")?;
-                    waiting_on_upscale = false;
-                    last_memory_high_increase_at = None;
-                    self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-                    info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-                }
-                event = events.next() => {
-                    let Some(Sequenced { seqnum, .. }) = event else {
-                        bail!("failed to listen for memory.high events")
-                    };
-                    // The memory.high came before our last upscale, so we consider
-                    // it resolved
-                    if self.last_upscale_seqnum.fetch_max(seqnum, Ordering::AcqRel) > seqnum {
-                        info!(
-                            "received memory.high event, but it came before our last upscale -> ignoring it"
-                        );
-                        continue;
-                    }
-
-                    // The memory.high came after our latest upscale. We don't
-                    // want to do anything yet, so peek the next event in hopes
-                    // that it's an upscale.
-                    if let Some(upscale_num) = self
-                        .upscaled(&mut upscales)
-                        .context("failed to check if we were upscaled")?
-                    {
-                        if upscale_num > seqnum {
-                            info!(
-                                "received memory.high event, but it came before our last upscale -> ignoring it"
-                            );
-                            continue;
-                        }
-                    }
-
-                    // If it's been long enough since we last froze, freeze the
-                    // cgroup and request upscale
-                    if wait_to_freeze.is_elapsed() {
-                        info!("received memory.high event -> requesting upscale");
-                        waiting_on_upscale = self
-                            .handle_memory_high_event(&mut upscales)
-                            .await
-                            .context("failed to handle upscale")?;
-                        wait_to_freeze
-                            .as_mut()
-                            .reset(Instant::now() + self.config.do_not_freeze_more_often_than);
-                        continue;
-                    }
-
-                    // Ok, we can't freeze, just request upscale
-                    if !waiting_on_upscale {
-                        info!("received memory.high event, but too soon to refreeze -> requesting upscale");
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to request upscaling because we got upscaled");
-                            continue;
-                        }
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-                        waiting_on_upscale = true;
-                        continue;
-                    }
-
-                    // Shoot, we can't freeze or and we're still waiting on upscale,
-                    // increase memory.high to reduce throttling
-                    let can_increase_memory_high = match last_memory_high_increase_at {
-                        None => true,
-                        Some(t) => t.elapsed() > self.config.memory_high_increase_every,
-                    };
-                    if can_increase_memory_high {
-                        info!(
-                            "received memory.high event, \
-                            but too soon to refreeze and already requested upscale \
-                            -> increasing memory.high"
-                        );
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to increase memory.high because got upscaled");
-                            continue;
-                        }
-
-                        // Request upscale anyways (the agent will handle deduplicating
-                        // requests)
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-
-                        let memory_high =
-                            self.get_high_bytes().context("failed to get memory.high")?;
-                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
-                        info!(
-                            current_high_bytes = memory_high,
-                            new_high_bytes = new_high,
-                            "updating memory.high"
-                        );
-                        self.set_high_bytes(new_high)
-                            .context("failed to set memory.high")?;
-                        last_memory_high_increase_at = Some(Instant::now());
-                        continue;
-                    }
-
-                    info!("received memory.high event, but can't do anything");
-                }
-            };
-        }
-    }
-
-    /// Handle a `memory.high`, returning whether we are still waiting on upscale
-    /// by the time the function returns.
-    ///
-    /// The general plan for handling a `memory.high` event is as follows:
-    /// 1. Freeze the cgroup
-    /// 2. Start a timer for `self.config.max_upscale_wait`
-    /// 3. Request upscale
-    /// 4. After the timer elapses or we receive upscale, thaw the cgroup.
-    /// 5. Return whether or not we are still waiting for upscale. If we are,
-    ///    we'll increase the cgroups memory.high to avoid getting oom killed
-    #[tracing::instrument(skip_all)]
-    async fn handle_memory_high_event(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<bool> {
-        // Immediately freeze the cgroup before doing anything else.
-        info!("received memory.high event -> freezing cgroup");
-        self.freeze().context("failed to freeze cgroup")?;
-
-        // We'll use this for logging durations
-        let start_time = Instant::now();
-
-        // Await the upscale until we have to unfreeze
-        let timed =
-            tokio::time::timeout(self.config.max_upscale_wait, self.await_upscale(upscales));
-
-        // Request the upscale
-        info!(
-            wait = ?self.config.max_upscale_wait,
-            "sending request for immediate upscaling",
-        );
-        self.upscale_requester
-            .send(())
-            .await
-            .context("failed to request upscale")?;
-
-        let waiting_on_upscale = match timed.await {
-            Ok(Ok(())) => {
-                info!(elapsed = ?start_time.elapsed(), "received upscale in time");
-                false
-            }
-            // **important**: unfreeze the cgroup before ?-reporting the error
-            Ok(Err(e)) => {
-                info!("error waiting for upscale -> thawing cgroup");
-                self.thaw()
-                    .context("failed to thaw cgroup after errored waiting for upscale")?;
-                Err(e.context("failed to await upscale"))?
-            }
-            Err(_) => {
-                info!(elapsed = ?self.config.max_upscale_wait, "timed out waiting for upscale");
-                true
-            }
-        };
-
-        info!("thawing cgroup");
-        self.thaw().context("failed to thaw cgroup")?;
-
-        Ok(waiting_on_upscale)
-    }
-
-    /// Checks whether we were just upscaled, returning the upscale's sequence
-    /// number if so.
-    #[tracing::instrument(skip_all)]
-    fn upscaled(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<Option<u64>> {
-        let Sequenced { seqnum, data } = match upscales.try_recv() {
-            Ok(upscale) => upscale,
-            Err(TryRecvError::Empty) => return Ok(None),
-            Err(TryRecvError::Disconnected) => {
-                bail!("upscale notification channel was disconnected")
-            }
-        };
-
-        // Make sure to update the last upscale sequence number
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-        Ok(Some(seqnum))
-    }
-
-    /// Await an upscale event, discarding any `memory.high` events received in
-    /// the process.
-    ///
-    /// This is used in `handle_memory_high_event`, where we need to listen
-    /// for upscales in particular so we know if we can thaw the cgroup early.
-    #[tracing::instrument(skip_all)]
-    async fn await_upscale(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+        updates: watch::Sender<(Instant, MemoryHistory)>,
     ) -> anyhow::Result<()> {
-        let Sequenced { seqnum, .. } = upscales
-            .recv()
-            .await
-            .context("error listening for upscales")?;
+        // this requirement makes the code a bit easier to work with; see the config for more.
+        assert!(self.config.memory_history_len <= self.config.memory_history_log_interval);
 
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        Ok(())
-    }
+        let mut ticker = tokio::time::interval(self.config.memory_poll_interval);
+        ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+        // ticker.reset_immediately(); // FIXME: enable this once updating to tokio >= 1.30.0
 
-    /// Get the cgroup's name.
-    pub fn path(&self) -> &str {
-        self.cgroup.path()
-    }
-}
+        let mem_controller = self.memory()?;
 
-/// Represents a set of limits we apply to a cgroup to control memory usage.
-///
-/// Setting these values also affects the thresholds for receiving usage alerts.
-#[derive(Debug)]
-pub struct MemoryLimits {
-    high: u64,
-    max: u64,
-}
+        // buffer for samples that will be logged. once full, it remains so.
+        let history_log_len = self.config.memory_history_log_interval;
+        let mut history_log_buf = vec![MemoryStatus::zeroed(); history_log_len];
 
-impl MemoryLimits {
-    pub fn new(high: u64, max: u64) -> Self {
-        Self { max, high }
-    }
-}
+        for t in 0_u64.. {
+            ticker.tick().await;
 
-// Methods for manipulating the actual cgroup
-impl CgroupWatcher {
-    /// Get a handle on the freezer subsystem.
-    fn freezer(&self) -> anyhow::Result<&FreezerController> {
-        if let Some(Freezer(freezer)) = self
-            .cgroup
-            .subsystems()
-            .iter()
-            .find(|sub| matches!(sub, Freezer(_)))
-        {
-            Ok(freezer)
-        } else {
-            anyhow::bail!("could not find freezer subsystem")
+            let now = Instant::now();
+            let mem = Self::memory_usage(mem_controller);
+
+            let i = t as usize % history_log_len;
+            history_log_buf[i] = mem;
+
+            // We're taking *at most* memory_history_len values; we may be bounded by the total
+            // number of samples that have come in so far.
+            let samples_count = (t + 1).min(self.config.memory_history_len as u64) as usize;
+            // NB: in `ring_buf_recent_values_iter`, `i` is *inclusive*, which matches the fact
+            // that we just inserted a value there, so the end of the iterator will *include* the
+            // value at i, rather than stopping just short of it.
+            let samples = ring_buf_recent_values_iter(&history_log_buf, i, samples_count);
+
+            let summary = MemoryHistory {
+                avg_non_reclaimable: samples.map(|h| h.non_reclaimable).sum::<u64>()
+                    / samples_count as u64,
+                samples_count,
+                samples_span: self.config.memory_poll_interval * (samples_count - 1) as u32,
+            };
+
+            // Log the current history if it's time to do so. Because `history_log_buf` has length
+            // equal to the logging interval, we can just log the entire buffer every time we set
+            // the last entry, which also means that for this log line, we can ignore that it's a
+            // ring buffer (because all the entries are in order of increasing time).
+            if i == history_log_len - 1 {
+                info!(
+                    history = ?MemoryStatus::debug_slice(&history_log_buf),
+                    summary = ?summary,
+                    "Recent cgroup memory statistics history"
+                );
+            }
+
+            updates
+                .send((now, summary))
+                .context("failed to send MemoryHistory")?;
         }
-    }
 
-    /// Attempt to freeze the cgroup.
-    pub fn freeze(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .freeze()
-            .context("failed to freeze")
-    }
-
-    /// Attempt to thaw the cgroup.
-    pub fn thaw(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .thaw()
-            .context("failed to thaw")
+        unreachable!()
     }
 
     /// Get a handle on the memory subsystem.
-    ///
-    /// Note: this method does not require `self.memory_update_lock` because
-    /// getting a handle to the subsystem does not access any of the files we
-    /// care about, such as memory.high and memory.events
     fn memory(&self) -> anyhow::Result<&MemController> {
-        if let Some(Mem(memory)) = self
-            .cgroup
+        self.cgroup
             .subsystems()
             .iter()
-            .find(|sub| matches!(sub, Mem(_)))
-        {
-            Ok(memory)
-        } else {
-            anyhow::bail!("could not find memory subsystem")
-        }
-    }
-
-    /// Get cgroup current memory usage.
-    pub fn current_memory_usage(&self) -> anyhow::Result<u64> {
-        Ok(self
-            .memory()
-            .context("failed to get memory subsystem")?
-            .memory_stat()
-            .usage_in_bytes)
-    }
-
-    /// Set cgroup memory.high threshold.
-    pub fn set_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
-        self.memory()
-            .context("failed to get memory subsystem")?
-            .set_mem(cgroups_rs::memory::SetMemory {
-                low: None,
-                high: Some(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64)),
-                min: None,
-                max: None,
+            .find_map(|sub| match sub {
+                Subsystem::Mem(c) => Some(c),
+                _ => None,
             })
-            .context("failed to set memory.high")
+            .ok_or_else(|| anyhow!("could not find memory subsystem"))
     }
 
-    /// Set cgroup memory.high and memory.max.
-    pub fn set_limits(&self, limits: &MemoryLimits) -> anyhow::Result<()> {
-        info!(
-            limits.high,
-            limits.max,
-            path = self.path(),
-            "writing new memory limits",
-        );
-        self.memory()
-            .context("failed to get memory subsystem while setting memory limits")?
-            .set_mem(cgroups_rs::memory::SetMemory {
-                min: None,
-                low: None,
-                high: Some(MaxValue::Value(
-                    u64::min(limits.high, i64::MAX as u64) as i64
-                )),
-                max: Some(MaxValue::Value(u64::min(limits.max, i64::MAX as u64) as i64)),
-            })
-            .context("failed to set memory limits")
-    }
-
-    /// Given some amount of available memory, set the desired cgroup memory limits
-    pub fn set_memory_limits(&mut self, available_memory: u64) -> anyhow::Result<()> {
-        let new_high = self.config.calculate_memory_high_value(available_memory);
-        let limits = MemoryLimits::new(new_high, available_memory);
-        info!(
-            path = self.path(),
-            memory = ?limits,
-            "setting cgroup memory",
-        );
-        self.set_limits(&limits)
-            .context("failed to set cgroup memory limits")?;
-        Ok(())
-    }
-
-    /// Get memory.high threshold.
-    pub fn get_high_bytes(&self) -> anyhow::Result<u64> {
-        let high = self
-            .memory()
-            .context("failed to get memory subsystem while getting memory statistics")?
-            .get_mem()
-            .map(|mem| mem.high)
-            .context("failed to get memory statistics from subsystem")?;
-        match high {
-            Some(MaxValue::Max) => Ok(i64::MAX as u64),
-            Some(MaxValue::Value(high)) => Ok(high as u64),
-            None => anyhow::bail!("failed to read memory.high from memory subsystem"),
+    /// Given a handle on the memory subsystem, returns the current memory information
+    fn memory_usage(mem_controller: &MemController) -> MemoryStatus {
+        let stat = mem_controller.memory_stat().stat;
+        MemoryStatus {
+            non_reclaimable: stat.active_anon + stat.inactive_anon,
         }
     }
 }
+
+// Helper function for `CgroupWatcher::watch`
+fn ring_buf_recent_values_iter<T>(
+    buf: &[T],
+    last_value_idx: usize,
+    count: usize,
+) -> impl '_ + Iterator<Item = &T> {
+    // Assertion carried over from `CgroupWatcher::watch`, to make the logic in this function
+    // easier (we only have to add `buf.len()` once, rather than a dynamic number of times).
+    assert!(count <= buf.len());
+
+    buf.iter()
+        // 'cycle' because the values could wrap around
+        .cycle()
+        // with 'cycle', this skip is more like 'offset', and functionally this is
+        // offsettting by 'last_value_idx - count (mod buf.len())', but we have to be
+        // careful to avoid underflow, so we pre-add buf.len().
+        // The '+ 1' is because `last_value_idx` is inclusive, rather than exclusive.
+        .skip((buf.len() + last_value_idx + 1 - count) % buf.len())
+        .take(count)
+}
+
+/// Summary of recent memory usage
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryHistory {
+    /// Rolling average of non-reclaimable memory usage samples over the last `history_period`
+    pub avg_non_reclaimable: u64,
+
+    /// The number of samples used to construct this summary
+    pub samples_count: usize,
+    /// Total timespan between the first and last sample used for this summary
+    pub samples_span: Duration,
+}
+
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryStatus {
+    non_reclaimable: u64,
+}
+
+impl MemoryStatus {
+    fn zeroed() -> Self {
+        MemoryStatus { non_reclaimable: 0 }
+    }
+
+    fn debug_slice(slice: &[Self]) -> impl '_ + Debug {
+        struct DS<'a>(&'a [MemoryStatus]);
+
+        impl<'a> Debug for DS<'a> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_struct("[MemoryStatus]")
+                    .field(
+                        "non_reclaimable[..]",
+                        &Fields(self.0, |stat: &MemoryStatus| {
+                            BytesToGB(stat.non_reclaimable)
+                        }),
+                    )
+                    .finish()
+            }
+        }
+
+        struct Fields<'a, F>(&'a [MemoryStatus], F);
+
+        impl<'a, F: Fn(&MemoryStatus) -> T, T: Debug> Debug for Fields<'a, F> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_list().entries(self.0.iter().map(&self.1)).finish()
+            }
+        }
+
+        struct BytesToGB(u64);
+
+        impl Debug for BytesToGB {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.write_fmt(format_args!(
+                    "{:.3}Gi",
+                    self.0 as f64 / (1_u64 << 30) as f64
+                ))
+            }
+        }
+
+        DS(slice)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn ring_buf_iter() {
+        let buf = vec![0_i32, 1, 2, 3, 4, 5, 6, 7, 8, 9];
+
+        let values = |offset, count| {
+            super::ring_buf_recent_values_iter(&buf, offset, count)
+                .copied()
+                .collect::<Vec<i32>>()
+        };
+
+        // Boundary conditions: start, end, and entire thing:
+        assert_eq!(values(0, 1), [0]);
+        assert_eq!(values(3, 4), [0, 1, 2, 3]);
+        assert_eq!(values(9, 4), [6, 7, 8, 9]);
+        assert_eq!(values(9, 10), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]);
+
+        // "normal" operation: no wraparound
+        assert_eq!(values(7, 4), [4, 5, 6, 7]);
+
+        // wraparound:
+        assert_eq!(values(0, 4), [7, 8, 9, 0]);
+        assert_eq!(values(1, 4), [8, 9, 0, 1]);
+        assert_eq!(values(2, 4), [9, 0, 1, 2]);
+        assert_eq!(values(2, 10), [3, 4, 5, 6, 7, 8, 9, 0, 1, 2]);
+    }
+}

libs/vm_monitor/src/dispatcher.rs

@@ -12,12 +12,10 @@ use futures::{
     stream::{SplitSink, SplitStream},
     SinkExt, StreamExt,
 };
-use tokio::sync::mpsc;
 use tracing::info;
 
-use crate::cgroup::Sequenced;
 use crate::protocol::{
-    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, Resources, PROTOCOL_MAX_VERSION,
+    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, PROTOCOL_MAX_VERSION,
     PROTOCOL_MIN_VERSION,
 };
 
@@ -36,13 +34,6 @@ pub struct Dispatcher {
     /// We send messages to the agent through `sink`
     sink: SplitSink<WebSocket, Message>,
 
-    /// Used to notify the cgroup when we are upscaled.
-    pub(crate) notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-
-    /// When the cgroup requests upscale it will send on this channel. In response
-    /// we send an `UpscaleRequst` to the agent.
-    pub(crate) request_upscale_events: mpsc::Receiver<()>,
-
     /// The protocol version we have agreed to use with the agent. This is negotiated
     /// during the creation of the dispatcher, and should be the highest shared protocol
     /// version.
@@ -61,11 +52,7 @@ impl Dispatcher {
     /// 1. Wait for the agent to sent the range of protocols it supports.
     /// 2. Send a protocol version that works for us as well, or an error if there
     ///    is no compatible version.
-    pub async fn new(
-        stream: WebSocket,
-        notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-        request_upscale_events: mpsc::Receiver<()>,
-    ) -> anyhow::Result<Self> {
+    pub async fn new(stream: WebSocket) -> anyhow::Result<Self> {
         let (mut sink, mut source) = stream.split();
 
         // Figure out the highest protocol version we both support
@@ -119,22 +106,10 @@ impl Dispatcher {
         Ok(Self {
             sink,
             source,
-            notify_upscale_events,
-            request_upscale_events,
             proto_version: highest_shared_version,
         })
     }
 
-    /// Notify the cgroup manager that we have received upscale and wait for
-    /// the acknowledgement.
-    #[tracing::instrument(skip_all, fields(?resources))]
-    pub async fn notify_upscale(&self, resources: Sequenced<Resources>) -> anyhow::Result<()> {
-        self.notify_upscale_events
-            .send(resources)
-            .await
-            .context("failed to send resources and oneshot sender across channel")
-    }
-
     /// Send a message to the agent.
     ///
     /// Although this function is small, it has one major benefit: it is the only

libs/vm_monitor/src/runner.rs

@@ -4,19 +4,17 @@
 //! This is the "Monitor" part of the monitor binary and is the main entrypoint for
 //! all functionality.
 
-use std::sync::Arc;
+use std::fmt::Debug;
 use std::time::{Duration, Instant};
-use std::{fmt::Debug, mem};
 
 use anyhow::{bail, Context};
 use axum::extract::ws::{Message, WebSocket};
 use futures::StreamExt;
-use tokio::sync::broadcast;
-use tokio::sync::mpsc;
+use tokio::sync::{broadcast, watch};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};
 
-use crate::cgroup::{CgroupWatcher, MemoryLimits, Sequenced};
+use crate::cgroup::{self, CgroupWatcher};
 use crate::dispatcher::Dispatcher;
 use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
@@ -28,7 +26,7 @@ use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args
 pub struct Runner {
     config: Config,
     filecache: Option<FileCacheState>,
-    cgroup: Option<Arc<CgroupWatcher>>,
+    cgroup: Option<CgroupState>,
     dispatcher: Dispatcher,
 
     /// We "mint" new message ids by incrementing this counter and taking the value.
@@ -45,6 +43,14 @@ pub struct Runner {
     kill: broadcast::Receiver<()>,
 }
 
+#[derive(Debug)]
+struct CgroupState {
+    watcher: watch::Receiver<(Instant, cgroup::MemoryHistory)>,
+    /// If [`cgroup::MemoryHistory::avg_non_reclaimable`] exceeds `threshold`, we send upscale
+    /// requests.
+    threshold: u64,
+}
+
 /// Configuration for a `Runner`
 #[derive(Debug)]
 pub struct Config {
@@ -62,16 +68,56 @@ pub struct Config {
     /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
+
+    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// other words, providing a ceiling for the highest value of the threshold by enforcing that
+    /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
+    /// threshold.
+    ///
+    /// For example, a value of `0.1` means that 10% of total memory must remain after exceeding
+    /// the threshold, so the value of the cgroup threshold would always be capped at 90% of total
+    /// memory.
+    ///
+    /// The default value of `0.15` means that we *guarantee* sending upscale requests if the
+    /// cgroup is using more than 85% of total memory (even if we're *not* separately reserving
+    /// memory for the file cache).
+    cgroup_min_overhead_fraction: f64,
+
+    cgroup_downscale_threshold_buffer_bytes: u64,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
             sys_buffer_bytes: 100 * MiB,
+            cgroup_min_overhead_fraction: 0.15,
+            cgroup_downscale_threshold_buffer_bytes: 100 * MiB,
         }
     }
 }
 
+impl Config {
+    fn cgroup_threshold(&self, total_mem: u64, file_cache_disk_size: u64) -> u64 {
+        // If the file cache is in tmpfs, then it will count towards shmem usage of the cgroup,
+        // and thus be non-reclaimable, so we should allow for additional memory usage.
+        //
+        // If the file cache sits on disk, our desired stable system state is for it to be fully
+        // page cached (its contents should only be paged to/from disk in situations where we can't
+        // upscale fast enough). Page-cached memory is reclaimable, so we need to lower the
+        // threshold for non-reclaimable memory so we scale up *before* the kernel starts paging
+        // out the file cache.
+        let memory_remaining_for_cgroup = total_mem.saturating_sub(file_cache_disk_size);
+
+        // Even if we're not separately making room for the file cache (if it's in tmpfs), we still
+        // want our threshold to be met gracefully instead of letting postgres get OOM-killed.
+        // So we guarantee that there's at least `cgroup_min_overhead_fraction` of total memory
+        // remaining above the threshold.
+        let max_threshold = (total_mem as f64 * (1.0 - self.cgroup_min_overhead_fraction)) as u64;
+
+        memory_remaining_for_cgroup.min(max_threshold)
+    }
+}
+
 impl Runner {
     /// Create a new monitor.
     #[tracing::instrument(skip_all, fields(?config, ?args))]
@@ -87,12 +133,7 @@ impl Runner {
             "invalid monitor Config: sys_buffer_bytes cannot be 0"
         );
 
-        // *NOTE*: the dispatcher and cgroup manager talk through these channels
-        // so make sure they each get the correct half, nothing is droppped, etc.
-        let (notified_send, notified_recv) = mpsc::channel(1);
-        let (requesting_send, requesting_recv) = mpsc::channel(1);
-
-        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+        let dispatcher = Dispatcher::new(ws)
             .await
             .context("error creating new dispatcher")?;
 
@@ -106,9 +147,10 @@ impl Runner {
             kill,
         };
 
-        let mut file_cache_reserved_bytes = 0;
         let mem = get_total_system_memory();
 
+        let mut file_cache_disk_size = 0;
+
         // We need to process file cache initialization before cgroup initialization, so that the memory
         // allocated to the file cache is appropriately taken into account when we decide the cgroup's
         // memory limits.
@@ -144,43 +186,40 @@ impl Runner {
             if actual_size != new_size {
                 info!("file cache size actually got set to {actual_size}")
             }
-            // Mark the resources given to the file cache as reserved, but only if it's in memory.
-            if !args.file_cache_on_disk {
-                file_cache_reserved_bytes = actual_size;
+
+            if args.file_cache_on_disk {
+                file_cache_disk_size = actual_size;
             }
 
             state.filecache = Some(file_cache);
         }
 
         if let Some(name) = &args.cgroup {
-            let (mut cgroup, cgroup_event_stream) =
-                CgroupWatcher::new(name.clone(), requesting_send)
-                    .context("failed to create cgroup manager")?;
+            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
+            // now, and then set limits later.
+            info!("initializing cgroup");
 
-            let available = mem - file_cache_reserved_bytes;
+            let cgroup =
+                CgroupWatcher::new(name.clone()).context("failed to create cgroup manager")?;
 
-            cgroup
-                .set_memory_limits(available)
-                .context("failed to set cgroup memory limits")?;
-
-            let cgroup = Arc::new(cgroup);
-
-            // Some might call this . . . cgroup v2
-            let cgroup_clone = Arc::clone(&cgroup);
+            let init_value = cgroup::MemoryHistory {
+                avg_non_reclaimable: 0,
+                samples_count: 0,
+                samples_span: Duration::ZERO,
+            };
+            let (hist_tx, hist_rx) = watch::channel((Instant::now(), init_value));
 
             spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
-                cgroup_clone.watch(notified_recv, cgroup_event_stream).await
+                cgroup.watch(hist_tx).await
             });
 
-            state.cgroup = Some(cgroup);
-        } else {
-            // *NOTE*: We need to forget the sender so that its drop impl does not get ran.
-            // This allows us to poll it in `Monitor::run` regardless of whether we
-            // are managing a cgroup or not. If we don't forget it, all receives will
-            // immediately return an error because the sender is droped and it will
-            // claim all select! statements, effectively turning `Monitor::run` into
-            // `loop { fail to receive }`.
-            mem::forget(requesting_send);
+            let threshold = state.config.cgroup_threshold(mem, file_cache_disk_size);
+            info!(threshold, "set initial cgroup threshold",);
+
+            state.cgroup = Some(CgroupState {
+                watcher: hist_rx,
+                threshold,
+            });
         }
 
         Ok(state)
@@ -200,28 +239,51 @@ impl Runner {
 
         let requested_mem = target.mem;
         let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
-        let expected_file_cache_mem_usage = self
+        let (expected_file_cache_size, expected_file_cache_disk_size) = self
             .filecache
             .as_ref()
-            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
-            .unwrap_or(0);
-        let mut new_cgroup_mem_high = 0;
+            .map(|file_cache| {
+                let size = file_cache.config.calculate_cache_size(usable_system_memory);
+                match file_cache.config.in_memory {
+                    true => (size, 0),
+                    false => (size, size),
+                }
+            })
+            .unwrap_or((0, 0));
         if let Some(cgroup) = &self.cgroup {
-            new_cgroup_mem_high = cgroup
+            let (last_time, last_history) = *cgroup.watcher.borrow();
+
+            // NB: The ordering of these conditions is intentional. During startup, we should deny
+            // downscaling until we have enough information to determine that it's safe to do so
+            // (i.e. enough samples have come in). But if it's been a while and we *still* haven't
+            // received any information, we should *fail* instead of just denying downscaling.
+            //
+            // `last_time` is set to `Instant::now()` on startup, so checking `last_time.elapsed()`
+            // serves double-duty: it trips if we haven't received *any* metrics for long enough,
+            // OR if we haven't received metrics *recently enough*.
+            //
+            // TODO: make the duration here configurable.
+            if last_time.elapsed() > Duration::from_secs(5) {
+                bail!("haven't gotten cgroup memory stats recently enough to determine downscaling information");
+            } else if last_history.samples_count <= 1 {
+                let status = "haven't received enough cgroup memory stats yet";
+                info!(status, "discontinuing downscale");
+                return Ok((false, status.to_owned()));
+            }
+
+            let new_threshold = self
                 .config
-                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+                .cgroup_threshold(usable_system_memory, expected_file_cache_disk_size);
 
-            let current = cgroup
-                .current_memory_usage()
-                .context("failed to fetch cgroup memory")?;
+            let current = last_history.avg_non_reclaimable;
 
-            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+            if new_threshold < current + self.config.cgroup_downscale_threshold_buffer_bytes {
                 let status = format!(
-                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
-                    "calculated memory.high too low",
-                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    "{}: {} MiB (new threshold) < {} (current usage) + {} (downscale buffer)",
+                    "calculated memory threshold too low",
+                    bytes_to_mebibytes(new_threshold),
                     bytes_to_mebibytes(current),
-                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                    bytes_to_mebibytes(self.config.cgroup_downscale_threshold_buffer_bytes)
                 );
 
                 info!(status, "discontinuing downscale");
@@ -232,14 +294,14 @@ impl Runner {
 
         // The downscaling has been approved. Downscale the file cache, then the cgroup.
         let mut status = vec![];
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let actual_usage = file_cache
-                .set_file_cache_size(expected_file_cache_mem_usage)
+                .set_file_cache_size(expected_file_cache_size)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
             let message = format!(
                 "set file cache size to {} MiB (in memory = {})",
@@ -250,28 +312,18 @@ impl Runner {
             status.push(message);
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-
-            if file_cache_mem_usage != expected_file_cache_mem_usage {
-                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
-            }
-
-            let limits = MemoryLimits::new(
-                // new_cgroup_mem_high is initialized to 0 but it is guarancontextd to not be here
-                // since it is properly initialized in the previous cgroup if let block
-                new_cgroup_mem_high,
-                available_memory,
-            );
-            cgroup
-                .set_limits(&limits)
-                .context("failed to set cgroup memory limits")?;
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
 
             let message = format!(
-                "set cgroup memory.high to {} MiB, of new max {} MiB",
-                bytes_to_mebibytes(new_cgroup_mem_high),
-                bytes_to_mebibytes(available_memory)
+                "set cgroup memory threshold from {} MiB to {} MiB, of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
+            cgroup.threshold = new_threshold;
             info!("downscale: {message}");
             status.push(message);
         }
@@ -292,8 +344,7 @@ impl Runner {
         let new_mem = resources.mem;
         let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
 
-        // Get the file cache's expected contribution to the memory usage
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
             info!(
@@ -306,8 +357,8 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
 
             if actual_usage != expected_usage {
@@ -319,19 +370,18 @@ impl Runner {
             }
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
+
             info!(
-                target = bytes_to_mebibytes(new_cgroup_mem_high),
-                total = bytes_to_mebibytes(new_mem),
-                name = cgroup.path(),
-                "updating cgroup memory.high",
+                "set cgroup memory threshold from {} MiB to {} MiB of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
-            let limits = MemoryLimits::new(new_cgroup_mem_high, available_memory);
-            cgroup
-                .set_limits(&limits)
-                .context("failed to set file cache size")?;
+            cgroup.threshold = new_threshold;
         }
 
         Ok(())
@@ -349,10 +399,6 @@ impl Runner {
                 self.handle_upscale(granted)
                     .await
                     .context("failed to handle upscale")?;
-                self.dispatcher
-                    .notify_upscale(Sequenced::new(granted))
-                    .await
-                    .context("failed to notify notify cgroup of upscale")?;
                 Ok(Some(OutboundMsg::new(
                     OutboundMsgKind::UpscaleConfirmation {},
                     id,
@@ -396,33 +442,53 @@ impl Runner {
                         Err(e) => bail!("failed to receive kill signal: {e}")
                     }
                 }
-                // we need to propagate an upscale request
-                request = self.dispatcher.request_upscale_events.recv() => {
-                    if request.is_none() {
-                        bail!("failed to listen for upscale event from cgroup")
+
+                // New memory stats from the cgroup, *may* need to request upscaling, if we've
+                // exceeded the threshold
+                result = self.cgroup.as_mut().unwrap().watcher.changed(), if self.cgroup.is_some() => {
+                    result.context("failed to receive from cgroup memory stats watcher")?;
+
+                    let cgroup = self.cgroup.as_ref().unwrap();
+
+                    let (_time, cgroup_mem_stat) = *cgroup.watcher.borrow();
+
+                    // If we haven't exceeded the threshold, then we're all ok
+                    if cgroup_mem_stat.avg_non_reclaimable < cgroup.threshold {
+                        continue;
                     }
 
-                    // If it's been less than 1 second since the last time we requested upscaling,
-                    // ignore the event, to avoid spamming the agent (otherwise, this can happen
-                    // ~1k times per second).
+                    // Otherwise, we generally want upscaling. But, if it's been less than 1 second
+                    // since the last time we requested upscaling, ignore the event, to avoid
+                    // spamming the agent.
                     if let Some(t) = self.last_upscale_request_at {
                         let elapsed = t.elapsed();
                         if elapsed < Duration::from_secs(1) {
-                            info!(elapsed_millis = elapsed.as_millis(), "cgroup asked for upscale but too soon to forward the request, ignoring");
+                            info!(
+                                elapsed_millis = elapsed.as_millis(),
+                                avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                                threshold = bytes_to_mebibytes(cgroup.threshold),
+                                "cgroup memory stats are high enough to upscale but too soon to forward the request, ignoring",
+                            );
                             continue;
                         }
                     }
 
                     self.last_upscale_request_at = Some(Instant::now());
 
-                    info!("cgroup asking for upscale; forwarding request");
+                    info!(
+                        avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                        threshold = bytes_to_mebibytes(cgroup.threshold),
+                        "cgroup memory stats are high enough to upscale, requesting upscale",
+                    );
+
                     self.counter += 2; // Increment, preserving parity (i.e. keep the
                                        // counter odd). See the field comment for more.
                     self.dispatcher
                         .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
                         .await
                         .context("failed to send message")?;
-                }
+                },
+
                 // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
@@ -450,11 +516,14 @@ impl Runner {
                                     Ok(Some(out)) => out,
                                     Ok(None) => continue,
                                     Err(e) => {
-                                        let error = e.to_string();
-                                        warn!(?error, "error handling message");
+                                        // use {:#} for our logging because the display impl only
+                                        // gives the outermost cause, and the debug impl
+                                        // pretty-prints the error, whereas {:#} contains all the
+                                        // causes, but is compact (no newlines).
+                                        warn!(error = format!("{e:#}"), "error handling message");
                                         OutboundMsg::new(
                                             OutboundMsgKind::InternalError {
-                                                error
+                                                error: e.to_string(),
                                             },
                                             message.id
                                         )

libs/walproposer/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "walproposer"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+utils.workspace = true
+postgres_ffi.workspace = true
+
+workspace_hack.workspace = true
+
+[build-dependencies]
+anyhow.workspace = true
+bindgen.workspace = true

libs/walproposer/bindgen_deps.h

@@ -0,0 +1 @@
+#include "walproposer.h"

libs/walproposer/build.rs

@@ -0,0 +1,113 @@
+use std::{env, path::PathBuf, process::Command};
+
+use anyhow::{anyhow, Context};
+use bindgen::CargoCallbacks;
+
+fn main() -> anyhow::Result<()> {
+    // Tell cargo to invalidate the built crate whenever the wrapper changes
+    println!("cargo:rerun-if-changed=bindgen_deps.h");
+
+    // Finding the location of built libraries and Postgres C headers:
+    // - if POSTGRES_INSTALL_DIR is set look into it, otherwise look into `<project_root>/pg_install`
+    // - if there's a `bin/pg_config` file use it for getting include server, otherwise use `<project_root>/pg_install/{PG_MAJORVERSION}/include/postgresql/server`
+    let pg_install_dir = if let Some(postgres_install_dir) = env::var_os("POSTGRES_INSTALL_DIR") {
+        postgres_install_dir.into()
+    } else {
+        PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pg_install")
+    };
+
+    let pg_install_abs = std::fs::canonicalize(pg_install_dir)?;
+    let walproposer_lib_dir = pg_install_abs.join("build/walproposer-lib");
+    let walproposer_lib_search_str = walproposer_lib_dir
+        .to_str()
+        .ok_or(anyhow!("Bad non-UTF path"))?;
+
+    let pgxn_neon = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pgxn/neon");
+    let pgxn_neon = std::fs::canonicalize(pgxn_neon)?;
+    let pgxn_neon = pgxn_neon.to_str().ok_or(anyhow!("Bad non-UTF path"))?;
+
+    println!("cargo:rustc-link-lib=static=pgport");
+    println!("cargo:rustc-link-lib=static=pgcommon");
+    println!("cargo:rustc-link-lib=static=walproposer");
+    println!("cargo:rustc-link-search={walproposer_lib_search_str}");
+
+    let pg_config_bin = pg_install_abs.join("v16").join("bin").join("pg_config");
+    let inc_server_path: String = if pg_config_bin.exists() {
+        let output = Command::new(pg_config_bin)
+            .arg("--includedir-server")
+            .output()
+            .context("failed to execute `pg_config --includedir-server`")?;
+
+        if !output.status.success() {
+            panic!("`pg_config --includedir-server` failed")
+        }
+
+        String::from_utf8(output.stdout)
+            .context("pg_config output is not UTF-8")?
+            .trim_end()
+            .into()
+    } else {
+        let server_path = pg_install_abs
+            .join("v16")
+            .join("include")
+            .join("postgresql")
+            .join("server")
+            .into_os_string();
+        server_path
+            .into_string()
+            .map_err(|s| anyhow!("Bad postgres server path {s:?}"))?
+    };
+
+    // The bindgen::Builder is the main entry point
+    // to bindgen, and lets you build up options for
+    // the resulting bindings.
+    let bindings = bindgen::Builder::default()
+        // The input header we would like to generate
+        // bindings for.
+        .header("bindgen_deps.h")
+        // Tell cargo to invalidate the built crate whenever any of the
+        // included header files changed.
+        .parse_callbacks(Box::new(CargoCallbacks))
+        .allowlist_type("WalProposer")
+        .allowlist_type("WalProposerConfig")
+        .allowlist_type("walproposer_api")
+        .allowlist_function("WalProposerCreate")
+        .allowlist_function("WalProposerStart")
+        .allowlist_function("WalProposerBroadcast")
+        .allowlist_function("WalProposerPoll")
+        .allowlist_function("WalProposerFree")
+        .allowlist_var("DEBUG5")
+        .allowlist_var("DEBUG4")
+        .allowlist_var("DEBUG3")
+        .allowlist_var("DEBUG2")
+        .allowlist_var("DEBUG1")
+        .allowlist_var("LOG")
+        .allowlist_var("INFO")
+        .allowlist_var("NOTICE")
+        .allowlist_var("WARNING")
+        .allowlist_var("ERROR")
+        .allowlist_var("FATAL")
+        .allowlist_var("PANIC")
+        .allowlist_var("WPEVENT")
+        .allowlist_var("WL_LATCH_SET")
+        .allowlist_var("WL_SOCKET_READABLE")
+        .allowlist_var("WL_SOCKET_WRITEABLE")
+        .allowlist_var("WL_TIMEOUT")
+        .allowlist_var("WL_SOCKET_CLOSED")
+        .allowlist_var("WL_SOCKET_MASK")
+        .clang_arg("-DWALPROPOSER_LIB")
+        .clang_arg(format!("-I{pgxn_neon}"))
+        .clang_arg(format!("-I{inc_server_path}"))
+        // Finish the builder and generate the bindings.
+        .generate()
+        // Unwrap the Result and panic on failure.
+        .expect("Unable to generate bindings");
+
+    // Write the bindings to the $OUT_DIR/bindings.rs file.
+    let out_path = PathBuf::from(env::var("OUT_DIR").unwrap()).join("bindings.rs");
+    bindings
+        .write_to_file(out_path)
+        .expect("Couldn't write bindings!");
+
+    Ok(())
+}

libs/walproposer/src/api_bindings.rs

@@ -0,0 +1,455 @@
+#![allow(dead_code)]
+
+use std::ffi::CStr;
+use std::ffi::CString;
+
+use crate::bindings::uint32;
+use crate::bindings::walproposer_api;
+use crate::bindings::PGAsyncReadResult;
+use crate::bindings::PGAsyncWriteResult;
+use crate::bindings::Safekeeper;
+use crate::bindings::Size;
+use crate::bindings::StringInfoData;
+use crate::bindings::TimeLineID;
+use crate::bindings::TimestampTz;
+use crate::bindings::WalProposer;
+use crate::bindings::WalProposerConnStatusType;
+use crate::bindings::WalProposerConnectPollStatusType;
+use crate::bindings::WalProposerExecStatusType;
+use crate::bindings::WalproposerShmemState;
+use crate::bindings::XLogRecPtr;
+use crate::walproposer::ApiImpl;
+use crate::walproposer::WaitResult;
+
+extern "C" fn get_shmem_state(wp: *mut WalProposer) -> *mut WalproposerShmemState {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_shmem_state()
+    }
+}
+
+extern "C" fn start_streaming(wp: *mut WalProposer, startpos: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).start_streaming(startpos)
+    }
+}
+
+extern "C" fn get_flush_rec_ptr(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_flush_rec_ptr()
+    }
+}
+
+extern "C" fn get_current_timestamp(wp: *mut WalProposer) -> TimestampTz {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_current_timestamp()
+    }
+}
+
+extern "C" fn conn_error_message(sk: *mut Safekeeper) -> *mut ::std::os::raw::c_char {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let msg = (*api).conn_error_message(&mut (*sk));
+        let msg = CString::new(msg).unwrap();
+        // TODO: fix leaking error message
+        msg.into_raw()
+    }
+}
+
+extern "C" fn conn_status(sk: *mut Safekeeper) -> WalProposerConnStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_status(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_start(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_start(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_poll(sk: *mut Safekeeper) -> WalProposerConnectPollStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_poll(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_send_query(sk: *mut Safekeeper, query: *mut ::std::os::raw::c_char) -> bool {
+    let query = unsafe { CStr::from_ptr(query) };
+    let query = query.to_str().unwrap();
+
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_send_query(&mut (*sk), query)
+    }
+}
+
+extern "C" fn conn_get_query_result(sk: *mut Safekeeper) -> WalProposerExecStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_get_query_result(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_flush(sk: *mut Safekeeper) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_flush(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_finish(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_finish(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_async_read(
+    sk: *mut Safekeeper,
+    buf: *mut *mut ::std::os::raw::c_char,
+    amount: *mut ::std::os::raw::c_int,
+) -> PGAsyncReadResult {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let (res, result) = (*api).conn_async_read(&mut (*sk));
+
+        // This function has guarantee that returned buf will be valid until
+        // the next call. So we can store a Vec in each Safekeeper and reuse
+        // it on the next call.
+        let mut inbuf = take_vec_u8(&mut (*sk).inbuf).unwrap_or_default();
+
+        inbuf.clear();
+        inbuf.extend_from_slice(res);
+
+        // Put a Vec back to sk->inbuf and return data ptr.
+        *buf = store_vec_u8(&mut (*sk).inbuf, inbuf);
+        *amount = res.len() as i32;
+
+        result
+    }
+}
+
+extern "C" fn conn_async_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> PGAsyncWriteResult {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_async_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn conn_blocking_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_blocking_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn recovery_download(
+    sk: *mut Safekeeper,
+    _timeline: TimeLineID,
+    startpos: XLogRecPtr,
+    endpos: XLogRecPtr,
+) -> bool {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).recovery_download(&mut (*sk), startpos, endpos)
+    }
+}
+
+extern "C" fn wal_read(
+    sk: *mut Safekeeper,
+    buf: *mut ::std::os::raw::c_char,
+    startptr: XLogRecPtr,
+    count: Size,
+) {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_read(&mut (*sk), buf, startptr)
+    }
+}
+
+extern "C" fn wal_reader_allocate(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_reader_allocate(&mut (*sk));
+    }
+}
+
+extern "C" fn free_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).free_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn init_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).init_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn update_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).update_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).add_safekeeper_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn wait_event_set(
+    wp: *mut WalProposer,
+    timeout: ::std::os::raw::c_long,
+    event_sk: *mut *mut Safekeeper,
+    events: *mut uint32,
+) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let result = (*api).wait_event_set(&mut (*wp), timeout);
+        match result {
+            WaitResult::Latch => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_LATCH_SET;
+                1
+            }
+            WaitResult::Timeout => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_TIMEOUT;
+                0
+            }
+            WaitResult::Network(sk, event_mask) => {
+                *event_sk = sk;
+                *events = event_mask;
+                1
+            }
+        }
+    }
+}
+
+extern "C" fn strong_random(
+    wp: *mut WalProposer,
+    buf: *mut ::std::os::raw::c_void,
+    len: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, len);
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).strong_random(buf)
+    }
+}
+
+extern "C" fn get_redo_start_lsn(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_redo_start_lsn()
+    }
+}
+
+extern "C" fn finish_sync_safekeepers(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).finish_sync_safekeepers(lsn)
+    }
+}
+
+extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).process_safekeeper_feedback(&mut (*wp), commit_lsn)
+    }
+}
+
+extern "C" fn confirm_wal_streamed(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).confirm_wal_streamed(&mut (*wp), lsn)
+    }
+}
+
+extern "C" fn log_internal(
+    wp: *mut WalProposer,
+    level: ::std::os::raw::c_int,
+    line: *const ::std::os::raw::c_char,
+) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let line = CStr::from_ptr(line);
+        let line = line.to_str().unwrap();
+        (*api).log_internal(&mut (*wp), Level::from(level as u32), line)
+    }
+}
+
+extern "C" fn after_election(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).after_election(&mut (*wp))
+    }
+}
+
+#[derive(Debug)]
+pub enum Level {
+    Debug5,
+    Debug4,
+    Debug3,
+    Debug2,
+    Debug1,
+    Log,
+    Info,
+    Notice,
+    Warning,
+    Error,
+    Fatal,
+    Panic,
+    WPEvent,
+}
+
+impl Level {
+    pub fn from(elevel: u32) -> Level {
+        use crate::bindings::*;
+
+        match elevel {
+            DEBUG5 => Level::Debug5,
+            DEBUG4 => Level::Debug4,
+            DEBUG3 => Level::Debug3,
+            DEBUG2 => Level::Debug2,
+            DEBUG1 => Level::Debug1,
+            LOG => Level::Log,
+            INFO => Level::Info,
+            NOTICE => Level::Notice,
+            WARNING => Level::Warning,
+            ERROR => Level::Error,
+            FATAL => Level::Fatal,
+            PANIC => Level::Panic,
+            WPEVENT => Level::WPEvent,
+            _ => panic!("unknown log level {}", elevel),
+        }
+    }
+}
+
+pub(crate) fn create_api() -> walproposer_api {
+    walproposer_api {
+        get_shmem_state: Some(get_shmem_state),
+        start_streaming: Some(start_streaming),
+        get_flush_rec_ptr: Some(get_flush_rec_ptr),
+        get_current_timestamp: Some(get_current_timestamp),
+        conn_error_message: Some(conn_error_message),
+        conn_status: Some(conn_status),
+        conn_connect_start: Some(conn_connect_start),
+        conn_connect_poll: Some(conn_connect_poll),
+        conn_send_query: Some(conn_send_query),
+        conn_get_query_result: Some(conn_get_query_result),
+        conn_flush: Some(conn_flush),
+        conn_finish: Some(conn_finish),
+        conn_async_read: Some(conn_async_read),
+        conn_async_write: Some(conn_async_write),
+        conn_blocking_write: Some(conn_blocking_write),
+        recovery_download: Some(recovery_download),
+        wal_read: Some(wal_read),
+        wal_reader_allocate: Some(wal_reader_allocate),
+        free_event_set: Some(free_event_set),
+        init_event_set: Some(init_event_set),
+        update_event_set: Some(update_event_set),
+        add_safekeeper_event_set: Some(add_safekeeper_event_set),
+        wait_event_set: Some(wait_event_set),
+        strong_random: Some(strong_random),
+        get_redo_start_lsn: Some(get_redo_start_lsn),
+        finish_sync_safekeepers: Some(finish_sync_safekeepers),
+        process_safekeeper_feedback: Some(process_safekeeper_feedback),
+        confirm_wal_streamed: Some(confirm_wal_streamed),
+        log_internal: Some(log_internal),
+        after_election: Some(after_election),
+    }
+}
+
+impl std::fmt::Display for Level {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+/// Take ownership of `Vec<u8>` from StringInfoData.
+pub(crate) fn take_vec_u8(pg: &mut StringInfoData) -> Option<Vec<u8>> {
+    if pg.data.is_null() {
+        return None;
+    }
+
+    let ptr = pg.data as *mut u8;
+    let length = pg.len as usize;
+    let capacity = pg.maxlen as usize;
+
+    pg.data = std::ptr::null_mut();
+    pg.len = 0;
+    pg.maxlen = 0;
+
+    unsafe { Some(Vec::from_raw_parts(ptr, length, capacity)) }
+}
+
+/// Store `Vec<u8>` in StringInfoData.
+fn store_vec_u8(pg: &mut StringInfoData, vec: Vec<u8>) -> *mut ::std::os::raw::c_char {
+    let ptr = vec.as_ptr() as *mut ::std::os::raw::c_char;
+    let length = vec.len();
+    let capacity = vec.capacity();
+
+    assert!(pg.data.is_null());
+
+    pg.data = ptr;
+    pg.len = length as i32;
+    pg.maxlen = capacity as i32;
+
+    std::mem::forget(vec);
+
+    ptr
+}

libs/walproposer/src/lib.rs

@@ -0,0 +1,14 @@
+pub mod bindings {
+    #![allow(non_upper_case_globals)]
+    #![allow(non_camel_case_types)]
+    #![allow(non_snake_case)]
+    // bindgen creates some unsafe code with no doc comments.
+    #![allow(clippy::missing_safety_doc)]
+    // noted at 1.63 that in many cases there's a u32 -> u32 transmutes in bindgen code.
+    #![allow(clippy::useless_transmute)]
+
+    include!(concat!(env!("OUT_DIR"), "/bindings.rs"));
+}
+
+pub mod api_bindings;
+pub mod walproposer;

libs/walproposer/src/walproposer.rs

@@ -0,0 +1,485 @@
+use std::ffi::CString;
+
+use postgres_ffi::WAL_SEGMENT_SIZE;
+use utils::id::TenantTimelineId;
+
+use crate::{
+    api_bindings::{create_api, take_vec_u8, Level},
+    bindings::{
+        Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate, WalProposerFree,
+        WalProposerStart,
+    },
+};
+
+/// Rust high-level wrapper for C walproposer API. Many methods are not required
+/// for simple cases, hence todo!() in default implementations.
+///
+/// Refer to `pgxn/neon/walproposer.h` for documentation.
+pub trait ApiImpl {
+    fn get_shmem_state(&self) -> &mut crate::bindings::WalproposerShmemState {
+        todo!()
+    }
+
+    fn start_streaming(&self, _startpos: u64) {
+        todo!()
+    }
+
+    fn get_flush_rec_ptr(&self) -> u64 {
+        todo!()
+    }
+
+    fn get_current_timestamp(&self) -> i64 {
+        todo!()
+    }
+
+    fn conn_error_message(&self, _sk: &mut Safekeeper) -> String {
+        todo!()
+    }
+
+    fn conn_status(&self, _sk: &mut Safekeeper) -> crate::bindings::WalProposerConnStatusType {
+        todo!()
+    }
+
+    fn conn_connect_start(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_connect_poll(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerConnectPollStatusType {
+        todo!()
+    }
+
+    fn conn_send_query(&self, _sk: &mut Safekeeper, _query: &str) -> bool {
+        todo!()
+    }
+
+    fn conn_get_query_result(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerExecStatusType {
+        todo!()
+    }
+
+    fn conn_flush(&self, _sk: &mut Safekeeper) -> i32 {
+        todo!()
+    }
+
+    fn conn_finish(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_async_read(&self, _sk: &mut Safekeeper) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+        todo!()
+    }
+
+    fn conn_async_write(
+        &self,
+        _sk: &mut Safekeeper,
+        _buf: &[u8],
+    ) -> crate::bindings::PGAsyncWriteResult {
+        todo!()
+    }
+
+    fn conn_blocking_write(&self, _sk: &mut Safekeeper, _buf: &[u8]) -> bool {
+        todo!()
+    }
+
+    fn recovery_download(&self, _sk: &mut Safekeeper, _startpos: u64, _endpos: u64) -> bool {
+        todo!()
+    }
+
+    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) {
+        todo!()
+    }
+
+    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn free_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn init_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn update_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn add_safekeeper_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn wait_event_set(&self, _wp: &mut WalProposer, _timeout_millis: i64) -> WaitResult {
+        todo!()
+    }
+
+    fn strong_random(&self, _buf: &mut [u8]) -> bool {
+        todo!()
+    }
+
+    fn get_redo_start_lsn(&self) -> u64 {
+        todo!()
+    }
+
+    fn finish_sync_safekeepers(&self, _lsn: u64) {
+        todo!()
+    }
+
+    fn process_safekeeper_feedback(&self, _wp: &mut WalProposer, _commit_lsn: u64) {
+        todo!()
+    }
+
+    fn confirm_wal_streamed(&self, _wp: &mut WalProposer, _lsn: u64) {
+        todo!()
+    }
+
+    fn log_internal(&self, _wp: &mut WalProposer, _level: Level, _msg: &str) {
+        todo!()
+    }
+
+    fn after_election(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+}
+
+pub enum WaitResult {
+    Latch,
+    Timeout,
+    Network(*mut Safekeeper, u32),
+}
+
+pub struct Config {
+    /// Tenant and timeline id
+    pub ttid: TenantTimelineId,
+    /// List of safekeepers in format `host:port`
+    pub safekeepers_list: Vec<String>,
+    /// Safekeeper reconnect timeout in milliseconds
+    pub safekeeper_reconnect_timeout: i32,
+    /// Safekeeper connection timeout in milliseconds
+    pub safekeeper_connection_timeout: i32,
+    /// walproposer mode, finish when all safekeepers are synced or subscribe
+    /// to WAL streaming
+    pub sync_safekeepers: bool,
+}
+
+/// WalProposer main struct. C methods are reexported as Rust functions.
+pub struct Wrapper {
+    wp: *mut WalProposer,
+    _safekeepers_list_vec: Vec<u8>,
+}
+
+impl Wrapper {
+    pub fn new(api: Box<dyn ApiImpl>, config: Config) -> Wrapper {
+        let neon_tenant = CString::new(config.ttid.tenant_id.to_string())
+            .unwrap()
+            .into_raw();
+        let neon_timeline = CString::new(config.ttid.timeline_id.to_string())
+            .unwrap()
+            .into_raw();
+
+        let mut safekeepers_list_vec = CString::new(config.safekeepers_list.join(","))
+            .unwrap()
+            .into_bytes_with_nul();
+        assert!(safekeepers_list_vec.len() == safekeepers_list_vec.capacity());
+        let safekeepers_list = safekeepers_list_vec.as_mut_ptr() as *mut i8;
+
+        let callback_data = Box::into_raw(Box::new(api)) as *mut ::std::os::raw::c_void;
+
+        let c_config = WalProposerConfig {
+            neon_tenant,
+            neon_timeline,
+            safekeepers_list,
+            safekeeper_reconnect_timeout: config.safekeeper_reconnect_timeout,
+            safekeeper_connection_timeout: config.safekeeper_connection_timeout,
+            wal_segment_size: WAL_SEGMENT_SIZE as i32, // default 16MB
+            syncSafekeepers: config.sync_safekeepers,
+            systemId: 0,
+            pgTimeline: 1,
+            callback_data,
+        };
+        let c_config = Box::into_raw(Box::new(c_config));
+
+        let api = create_api();
+        let wp = unsafe { WalProposerCreate(c_config, api) };
+        Wrapper {
+            wp,
+            _safekeepers_list_vec: safekeepers_list_vec,
+        }
+    }
+
+    pub fn start(&self) {
+        unsafe { WalProposerStart(self.wp) }
+    }
+}
+
+impl Drop for Wrapper {
+    fn drop(&mut self) {
+        unsafe {
+            let config = (*self.wp).config;
+            drop(Box::from_raw(
+                (*config).callback_data as *mut Box<dyn ApiImpl>,
+            ));
+            drop(CString::from_raw((*config).neon_tenant));
+            drop(CString::from_raw((*config).neon_timeline));
+            drop(Box::from_raw(config));
+
+            for i in 0..(*self.wp).n_safekeepers {
+                let sk = &mut (*self.wp).safekeeper[i as usize];
+                take_vec_u8(&mut sk.inbuf);
+            }
+
+            WalProposerFree(self.wp);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::{
+        cell::Cell,
+        sync::{atomic::AtomicUsize, mpsc::sync_channel},
+    };
+
+    use utils::id::TenantTimelineId;
+
+    use crate::{api_bindings::Level, walproposer::Wrapper};
+
+    use super::ApiImpl;
+
+    #[derive(Clone, Copy, Debug)]
+    struct WaitEventsData {
+        sk: *mut crate::bindings::Safekeeper,
+        event_mask: u32,
+    }
+
+    struct MockImpl {
+        // data to return from wait_event_set
+        wait_events: Cell<WaitEventsData>,
+        // walproposer->safekeeper messages
+        expected_messages: Vec<Vec<u8>>,
+        expected_ptr: AtomicUsize,
+        // safekeeper->walproposer messages
+        safekeeper_replies: Vec<Vec<u8>>,
+        replies_ptr: AtomicUsize,
+        // channel to send LSN to the main thread
+        sync_channel: std::sync::mpsc::SyncSender<u64>,
+    }
+
+    impl MockImpl {
+        fn check_walproposer_msg(&self, msg: &[u8]) {
+            let ptr = self
+                .expected_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.expected_messages.len() {
+                panic!("unexpected message from walproposer");
+            }
+
+            let expected_msg = &self.expected_messages[ptr];
+            assert_eq!(msg, expected_msg.as_slice());
+        }
+
+        fn next_safekeeper_reply(&self) -> &[u8] {
+            let ptr = self
+                .replies_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.safekeeper_replies.len() {
+                panic!("no more safekeeper replies");
+            }
+
+            &self.safekeeper_replies[ptr]
+        }
+    }
+
+    impl ApiImpl for MockImpl {
+        fn get_current_timestamp(&self) -> i64 {
+            println!("get_current_timestamp");
+            0
+        }
+
+        fn conn_status(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnStatusType {
+            println!("conn_status");
+            crate::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
+        }
+
+        fn conn_connect_start(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("conn_connect_start");
+        }
+
+        fn conn_connect_poll(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnectPollStatusType {
+            println!("conn_connect_poll");
+            crate::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
+        }
+
+        fn conn_send_query(&self, _: &mut crate::bindings::Safekeeper, query: &str) -> bool {
+            println!("conn_send_query: {}", query);
+            true
+        }
+
+        fn conn_get_query_result(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerExecStatusType {
+            println!("conn_get_query_result");
+            crate::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
+        }
+
+        fn conn_async_read(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+            println!("conn_async_read");
+            let reply = self.next_safekeeper_reply();
+            println!("conn_async_read result: {:?}", reply);
+            (
+                reply,
+                crate::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS,
+            )
+        }
+
+        fn conn_blocking_write(&self, _: &mut crate::bindings::Safekeeper, buf: &[u8]) -> bool {
+            println!("conn_blocking_write: {:?}", buf);
+            self.check_walproposer_msg(buf);
+            true
+        }
+
+        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("wal_reader_allocate")
+        }
+
+        fn free_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("free_event_set")
+        }
+
+        fn init_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("init_event_set")
+        }
+
+        fn update_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "update_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn add_safekeeper_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn wait_event_set(
+            &self,
+            _: &mut crate::bindings::WalProposer,
+            timeout_millis: i64,
+        ) -> super::WaitResult {
+            let data = self.wait_events.get();
+            println!(
+                "wait_event_set, timeout_millis={}, res={:?}",
+                timeout_millis, data
+            );
+            super::WaitResult::Network(data.sk, data.event_mask)
+        }
+
+        fn strong_random(&self, buf: &mut [u8]) -> bool {
+            println!("strong_random");
+            buf.fill(0);
+            true
+        }
+
+        fn finish_sync_safekeepers(&self, lsn: u64) {
+            self.sync_channel.send(lsn).unwrap();
+            panic!("sync safekeepers finished at lsn={}", lsn);
+        }
+
+        fn log_internal(&self, _wp: &mut crate::bindings::WalProposer, level: Level, msg: &str) {
+            println!("walprop_log[{}] {}", level, msg);
+        }
+
+        fn after_election(&self, _wp: &mut crate::bindings::WalProposer) {
+            println!("after_election");
+        }
+    }
+
+    /// Test that walproposer can successfully connect to safekeeper and finish
+    /// sync_safekeepers. API is mocked in MockImpl.
+    ///
+    /// Run this test with valgrind to detect leaks:
+    /// `valgrind --leak-check=full target/debug/deps/walproposer-<build>`
+    #[test]
+    fn test_simple_sync_safekeepers() -> anyhow::Result<()> {
+        let ttid = TenantTimelineId::new(
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+        );
+
+        let (sender, receiver) = sync_channel(1);
+
+        let my_impl: Box<dyn ApiImpl> = Box::new(MockImpl {
+            wait_events: Cell::new(WaitEventsData {
+                sk: std::ptr::null_mut(),
+                event_mask: 0,
+            }),
+            expected_messages: vec![
+                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160000, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 158, 76, 143, 54, 6, 60, 108, 110,
+                    147, 188, 32, 214, 90, 130, 15, 61, 158, 76, 143, 54, 6, 60, 108, 110, 147,
+                    188, 32, 214, 90, 130, 15, 61, 1, 0, 0, 0, 0, 0, 0, 1,
+                ],
+                // VoteRequest(VoteRequest { term: 3 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            expected_ptr: AtomicUsize::new(0),
+            safekeeper_replies: vec![
+                // Greeting(AcceptorGreeting { term: 2, node_id: NodeId(1) })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
+                ],
+                // VoteResponse(VoteResponse { term: 3, vote_given: 1, flush_lsn: 0/539, truncate_lsn: 0/539, term_history: [(2, 0/539)], timeline_start_lsn: 0/539 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 57,
+                    5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0,
+                    0, 57, 5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            replies_ptr: AtomicUsize::new(0),
+            sync_channel: sender,
+        });
+        let config = crate::walproposer::Config {
+            ttid,
+            safekeepers_list: vec!["localhost:5000".to_string()],
+            safekeeper_reconnect_timeout: 1000,
+            safekeeper_connection_timeout: 10000,
+            sync_safekeepers: true,
+        };
+
+        let wp = Wrapper::new(my_impl, config);
+
+        // walproposer will panic when it finishes sync_safekeepers
+        std::panic::catch_unwind(|| wp.start()).unwrap_err();
+        // validate the resulting LSN
+        assert_eq!(receiver.recv()?, 1337);
+        Ok(())
+        // drop() will free up resources here
+    }
+}

pageserver/Cargo.toml

@@ -17,6 +17,8 @@ async-stream.workspace = true
 async-trait.workspace = true
 byteorder.workspace = true
 bytes.workspace = true
+camino.workspace = true
+camino-tempfile.workspace = true
 chrono = { workspace = true, features = ["serde"] }
 clap = { workspace = true, features = ["string"] }
 close_fds.workspace = true
@@ -84,7 +86,6 @@ strum_macros.workspace = true
 [dev-dependencies]
 criterion.workspace = true
 hex-literal.workspace = true
-tempfile.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time", "test-util"] }
 
 [[bench]]

pageserver/benches/bench_walredo.rs

@@ -11,10 +11,7 @@ use std::sync::{Arc, Barrier};
 
 use bytes::{Buf, Bytes};
 use pageserver::{
-    config::PageServerConf,
-    repository::Key,
-    walrecord::NeonWalRecord,
-    walredo::{PostgresRedoManager, WalRedoError},
+    config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
 };
 use utils::{id::TenantId, lsn::Lsn};
 
@@ -25,7 +22,7 @@ fn redo_scenarios(c: &mut Criterion) {
     // input to the stderr.
     // utils::logging::init(utils::logging::LogFormat::Plain).unwrap();
 
-    let repo_dir = tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
+    let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
     let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
     let conf = Box::leak(Box::new(conf));
@@ -35,9 +32,15 @@ fn redo_scenarios(c: &mut Criterion) {
 
     let manager = Arc::new(manager);
 
-    tracing::info!("executing first");
-    short().execute(&manager).unwrap();
-    tracing::info!("first executed");
+    {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        tracing::info!("executing first");
+        short().execute(rt.handle(), &manager).unwrap();
+        tracing::info!("first executed");
+    }
 
     let thread_counts = [1, 2, 4, 8, 16];
 
@@ -80,9 +83,14 @@ fn add_multithreaded_walredo_requesters(
     assert_ne!(threads, 0);
 
     if threads == 1 {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        let handle = rt.handle();
         b.iter_batched_ref(
             || Some(input_factory()),
-            |input| execute_all(input.take(), manager),
+            |input| execute_all(input.take(), handle, manager),
             criterion::BatchSize::PerIteration,
         );
     } else {
@@ -98,19 +106,26 @@ fn add_multithreaded_walredo_requesters(
                     let manager = manager.clone();
                     let barrier = barrier.clone();
                     let work_rx = work_rx.clone();
-                    move || loop {
-                        // queue up and wait if we want to go another round
-                        if work_rx.lock().unwrap().recv().is_err() {
-                            break;
+                    move || {
+                        let rt = tokio::runtime::Builder::new_current_thread()
+                            .enable_all()
+                            .build()
+                            .unwrap();
+                        let handle = rt.handle();
+                        loop {
+                            // queue up and wait if we want to go another round
+                            if work_rx.lock().unwrap().recv().is_err() {
+                                break;
+                            }
+
+                            let input = Some(input_factory());
+
+                            barrier.wait();
+
+                            execute_all(input, handle, &manager).unwrap();
+
+                            barrier.wait();
                         }
-
-                        let input = Some(input_factory());
-
-                        barrier.wait();
-
-                        execute_all(input, &manager).unwrap();
-
-                        barrier.wait();
                     }
                 })
             })
@@ -152,15 +167,19 @@ impl Drop for JoinOnDrop {
     }
 }
 
-fn execute_all<I>(input: I, manager: &PostgresRedoManager) -> Result<(), WalRedoError>
+fn execute_all<I>(
+    input: I,
+    handle: &tokio::runtime::Handle,
+    manager: &PostgresRedoManager,
+) -> anyhow::Result<()>
 where
     I: IntoIterator<Item = Request>,
 {
     // just fire all requests as fast as possible
     input.into_iter().try_for_each(|req| {
-        let page = req.execute(manager)?;
+        let page = req.execute(handle, manager)?;
         assert_eq!(page.remaining(), 8192);
-        Ok::<_, WalRedoError>(())
+        anyhow::Ok(())
     })
 }
 
@@ -473,9 +492,11 @@ struct Request {
 }
 
 impl Request {
-    fn execute(self, manager: &PostgresRedoManager) -> Result<Bytes, WalRedoError> {
-        use pageserver::walredo::WalRedoManager;
-
+    fn execute(
+        self,
+        rt: &tokio::runtime::Handle,
+        manager: &PostgresRedoManager,
+    ) -> anyhow::Result<Bytes> {
         let Request {
             key,
             lsn,
@@ -484,6 +505,6 @@ impl Request {
             pg_version,
         } = self;
 
-        manager.request_redo(key, lsn, base_img, records, pg_version)
+        rt.block_on(manager.request_redo(key, lsn, base_img, records, pg_version))
     }
 }

pageserver/ctl/Cargo.toml

@@ -9,6 +9,7 @@ license.workspace = true
 [dependencies]
 anyhow.workspace = true
 bytes.workspace = true
+camino.workspace = true
 clap = { workspace = true, features = ["string"] }
 git-version.workspace = true
 pageserver = { path = ".." }

pageserver/ctl/src/layer_map_analyzer.rs

@@ -3,11 +3,14 @@
 //! Currently it only analyzes holes, which are regions within the layer range that the layer contains no updates for. In the future it might do more analysis (maybe key quantiles?) but it should never return sensitive data.
 
 use anyhow::Result;
+use camino::{Utf8Path, Utf8PathBuf};
+use pageserver::context::{DownloadBehavior, RequestContext};
+use pageserver::task_mgr::TaskKind;
 use pageserver::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use std::cmp::Ordering;
 use std::collections::BinaryHeap;
 use std::ops::Range;
-use std::{fs, path::Path, str};
+use std::{fs, str};
 
 use pageserver::page_cache::PAGE_SZ;
 use pageserver::repository::{Key, KEY_SIZE};
@@ -96,9 +99,9 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {
 }
 
 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
-async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
+async fn get_holes(path: &Utf8Path, max_holes: usize, ctx: &RequestContext) -> Result<Vec<Hole>> {
     let file = FileBlockReader::new(VirtualFile::open(path).await?);
-    let summary_blk = file.read_blk(0).await?;
+    let summary_blk = file.read_blk(0, ctx).await?;
     let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
     let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
         actual_summary.index_start_blk,
@@ -125,6 +128,7 @@ async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
                 prev_key = Some(curr.next());
                 true
             },
+            ctx,
         )
         .await?;
     let mut holes = heap.into_vec();
@@ -135,6 +139,7 @@ async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
 pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
     let storage_path = &cmd.path;
     let max_holes = cmd.max_holes.unwrap_or(DEFAULT_MAX_HOLES);
+    let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
 
     // Initialize virtual_file (file desriptor cache) and page cache which are needed to access layer persistent B-Tree.
     pageserver::virtual_file::init(10);
@@ -163,7 +168,9 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
                     parse_filename(&layer.file_name().into_string().unwrap())
                 {
                     if layer_file.is_delta {
-                        layer_file.holes = get_holes(&layer.path(), max_holes).await?;
+                        let layer_path =
+                            Utf8PathBuf::from_path_buf(layer.path()).expect("non-Unicode path");
+                        layer_file.holes = get_holes(&layer_path, max_holes, &ctx).await?;
                         n_deltas += 1;
                     }
                     layers.push(layer_file);

pageserver/ctl/src/layers.rs

@@ -1,7 +1,10 @@
 use std::path::{Path, PathBuf};
 
 use anyhow::Result;
+use camino::Utf8Path;
 use clap::Subcommand;
+use pageserver::context::{DownloadBehavior, RequestContext};
+use pageserver::task_mgr::TaskKind;
 use pageserver::tenant::block_io::BlockCursor;
 use pageserver::tenant::disk_btree::DiskBtreeReader;
 use pageserver::tenant::storage_layer::delta_layer::{BlobRef, Summary};
@@ -44,12 +47,12 @@ pub(crate) enum LayerCmd {
     },
 }
 
-async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
-    let path = path.as_ref();
+async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
+    let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
     virtual_file::init(10);
     page_cache::init(100);
     let file = FileBlockReader::new(VirtualFile::open(path).await?);
-    let summary_blk = file.read_blk(0).await?;
+    let summary_blk = file.read_blk(0, ctx).await?;
     let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
     let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
         actual_summary.index_start_blk,
@@ -67,11 +70,12 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
                 all.push((curr, BlobRef(value_offset)));
                 true
             },
+            ctx,
         )
         .await?;
     let cursor = BlockCursor::new_fileblockreader(&file);
     for (k, v) in all {
-        let value = cursor.read_blob(v.pos()).await?;
+        let value = cursor.read_blob(v.pos(), ctx).await?;
         println!("key:{} value_len:{}", k, value.len());
     }
     // TODO(chi): special handling for last key?
@@ -79,6 +83,7 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
 }
 
 pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
+    let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
     match cmd {
         LayerCmd::List { path } => {
             for tenant in fs::read_dir(path.join(TENANTS_SEGMENT_NAME))? {
@@ -153,7 +158,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                         );
 
                         if layer_file.is_delta {
-                            read_delta_file(layer.path()).await?;
+                            read_delta_file(layer.path(), &ctx).await?;
                         } else {
                             anyhow::bail!("not supported yet :(");
                         }

pageserver/ctl/src/main.rs

@@ -8,6 +8,7 @@ mod draw_timeline_dir;
 mod layer_map_analyzer;
 mod layers;
 
+use camino::{Utf8Path, Utf8PathBuf};
 use clap::{Parser, Subcommand};
 use layers::LayerCmd;
 use pageserver::{
@@ -18,7 +19,6 @@ use pageserver::{
     virtual_file,
 };
 use postgres_ffi::ControlFileData;
-use std::path::{Path, PathBuf};
 use utils::{lsn::Lsn, project_git_version};
 
 project_git_version!(GIT_VERSION);
@@ -49,7 +49,7 @@ enum Commands {
 #[derive(Parser)]
 struct MetadataCmd {
     /// Input metadata file path
-    metadata_path: PathBuf,
+    metadata_path: Utf8PathBuf,
     /// Replace disk consistent Lsn
     disk_consistent_lsn: Option<Lsn>,
     /// Replace previous record Lsn
@@ -61,13 +61,13 @@ struct MetadataCmd {
 #[derive(Parser)]
 struct PrintLayerFileCmd {
     /// Pageserver data path
-    path: PathBuf,
+    path: Utf8PathBuf,
 }
 
 #[derive(Parser)]
 struct AnalyzeLayerMapCmd {
     /// Pageserver data path
-    path: PathBuf,
+    path: Utf8PathBuf,
     /// Max holes
     max_holes: Option<usize>,
 }
@@ -102,7 +102,7 @@ async fn main() -> anyhow::Result<()> {
     Ok(())
 }
 
-fn read_pg_control_file(control_file_path: &Path) -> anyhow::Result<()> {
+fn read_pg_control_file(control_file_path: &Utf8Path) -> anyhow::Result<()> {
     let control_file = ControlFileData::decode(&std::fs::read(control_file_path)?)?;
     println!("{control_file:?}");
     let control_file_initdb = Lsn(control_file.checkPoint);
@@ -114,7 +114,7 @@ fn read_pg_control_file(control_file_path: &Path) -> anyhow::Result<()> {
     Ok(())
 }
 
-async fn print_layerfile(path: &Path) -> anyhow::Result<()> {
+async fn print_layerfile(path: &Utf8Path) -> anyhow::Result<()> {
     // Basic initialization of things that don't change after startup
     virtual_file::init(10);
     page_cache::init(100);

pageserver/src/basebackup.rs

@@ -13,6 +13,7 @@
 use anyhow::{anyhow, bail, ensure, Context};
 use bytes::{BufMut, BytesMut};
 use fail::fail_point;
+use postgres_ffi::pg_constants;
 use std::fmt::Write as FmtWrite;
 use std::time::SystemTime;
 use tokio::io;
@@ -180,6 +181,7 @@ where
             }
         }
 
+        let mut min_restart_lsn: Lsn = Lsn::MAX;
         // Create tablespace directories
         for ((spcnode, dbnode), has_relmap_file) in
             self.timeline.list_dbdirs(self.lsn, self.ctx).await?
@@ -213,6 +215,34 @@ where
                     self.add_rel(rel, rel).await?;
                 }
             }
+
+            for (path, content) in self.timeline.list_aux_files(self.lsn, self.ctx).await? {
+                if path.starts_with("pg_replslot") {
+                    let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
+                    let restart_lsn = Lsn(u64::from_le_bytes(
+                        content[offs..offs + 8].try_into().unwrap(),
+                    ));
+                    info!("Replication slot {} restart LSN={}", path, restart_lsn);
+                    min_restart_lsn = Lsn::min(min_restart_lsn, restart_lsn);
+                }
+                let header = new_tar_header(&path, content.len() as u64)?;
+                self.ar
+                    .append(&header, &*content)
+                    .await
+                    .context("could not add aux file to basebackup tarball")?;
+            }
+        }
+        if min_restart_lsn != Lsn::MAX {
+            info!(
+                "Min restart LSN for logical replication is {}",
+                min_restart_lsn
+            );
+            let data = min_restart_lsn.0.to_le_bytes();
+            let header = new_tar_header("restart.lsn", data.len() as u64)?;
+            self.ar
+                .append(&header, &data[..])
+                .await
+                .context("could not add restart.lsn file to basebackup tarball")?;
         }
         for xid in self
             .timeline

pageserver/src/bin/pageserver.rs

@@ -2,12 +2,15 @@
 
 use std::env::{var, VarError};
 use std::sync::Arc;
-use std::{env, ops::ControlFlow, path::Path, str::FromStr};
+use std::time::Duration;
+use std::{env, ops::ControlFlow, str::FromStr};
 
 use anyhow::{anyhow, Context};
+use camino::Utf8Path;
 use clap::{Arg, ArgAction, Command};
 
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
+use pageserver::control_plane_client::ControlPlaneClient;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::WALRECEIVER_RUNTIME;
@@ -20,6 +23,7 @@ use metrics::set_build_info_metric;
 use pageserver::{
     config::{defaults::*, PageServerConf},
     context::{DownloadBehavior, RequestContext},
+    deletion_queue::DeletionQueue,
     http, page_cache, page_service, task_mgr,
     task_mgr::TaskKind,
     task_mgr::{BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME},
@@ -63,21 +67,17 @@ fn main() -> anyhow::Result<()> {
 
     let workdir = arg_matches
         .get_one::<String>("workdir")
-        .map(Path::new)
-        .unwrap_or_else(|| Path::new(".neon"));
+        .map(Utf8Path::new)
+        .unwrap_or_else(|| Utf8Path::new(".neon"));
     let workdir = workdir
-        .canonicalize()
-        .with_context(|| format!("Error opening workdir '{}'", workdir.display()))?;
+        .canonicalize_utf8()
+        .with_context(|| format!("Error opening workdir '{workdir}'"))?;
 
     let cfg_file_path = workdir.join("pageserver.toml");
 
     // Set CWD to workdir for non-daemon modes
-    env::set_current_dir(&workdir).with_context(|| {
-        format!(
-            "Failed to set application's current dir to '{}'",
-            workdir.display()
-        )
-    })?;
+    env::set_current_dir(&workdir)
+        .with_context(|| format!("Failed to set application's current dir to '{workdir}'"))?;
 
     let conf = match initialize_config(&cfg_file_path, arg_matches, &workdir)? {
         ControlFlow::Continue(conf) => conf,
@@ -113,12 +113,8 @@ fn main() -> anyhow::Result<()> {
 
     let tenants_path = conf.tenants_path();
     if !tenants_path.exists() {
-        utils::crashsafe::create_dir_all(conf.tenants_path()).with_context(|| {
-            format!(
-                "Failed to create tenants root dir at '{}'",
-                tenants_path.display()
-            )
-        })?;
+        utils::crashsafe::create_dir_all(conf.tenants_path())
+            .with_context(|| format!("Failed to create tenants root dir at '{tenants_path}'"))?;
     }
 
     // Initialize up failpoints support
@@ -135,9 +131,9 @@ fn main() -> anyhow::Result<()> {
 }
 
 fn initialize_config(
-    cfg_file_path: &Path,
+    cfg_file_path: &Utf8Path,
     arg_matches: clap::ArgMatches,
-    workdir: &Path,
+    workdir: &Utf8Path,
 ) -> anyhow::Result<ControlFlow<(), &'static PageServerConf>> {
     let init = arg_matches.get_flag("init");
     let update_config = init || arg_matches.get_flag("update-config");
@@ -145,33 +141,22 @@ fn initialize_config(
     let (mut toml, config_file_exists) = if cfg_file_path.is_file() {
         if init {
             anyhow::bail!(
-                "Config file '{}' already exists, cannot init it, use --update-config to update it",
-                cfg_file_path.display()
+                "Config file '{cfg_file_path}' already exists, cannot init it, use --update-config to update it",
             );
         }
         // Supplement the CLI arguments with the config file
-        let cfg_file_contents = std::fs::read_to_string(cfg_file_path).with_context(|| {
-            format!(
-                "Failed to read pageserver config at '{}'",
-                cfg_file_path.display()
-            )
-        })?;
+        let cfg_file_contents = std::fs::read_to_string(cfg_file_path)
+            .with_context(|| format!("Failed to read pageserver config at '{cfg_file_path}'"))?;
         (
             cfg_file_contents
                 .parse::<toml_edit::Document>()
                 .with_context(|| {
-                    format!(
-                        "Failed to parse '{}' as pageserver config",
-                        cfg_file_path.display()
-                    )
+                    format!("Failed to parse '{cfg_file_path}' as pageserver config")
                 })?,
             true,
         )
     } else if cfg_file_path.exists() {
-        anyhow::bail!(
-            "Config file '{}' exists but is not a regular file",
-            cfg_file_path.display()
-        );
+        anyhow::bail!("Config file '{cfg_file_path}' exists but is not a regular file");
     } else {
         // We're initializing the tenant, so there's no config file yet
         (
@@ -190,7 +175,7 @@ fn initialize_config(
 
             for (key, item) in doc.iter() {
                 if config_file_exists && update_config && key == "id" && toml.contains_key(key) {
-                    anyhow::bail!("Pageserver config file exists at '{}' and has node id already, it cannot be overridden", cfg_file_path.display());
+                    anyhow::bail!("Pageserver config file exists at '{cfg_file_path}' and has node id already, it cannot be overridden");
                 }
                 toml.insert(key, item.clone());
             }
@@ -202,18 +187,11 @@ fn initialize_config(
         .context("Failed to parse pageserver configuration")?;
 
     if update_config {
-        info!("Writing pageserver config to '{}'", cfg_file_path.display());
+        info!("Writing pageserver config to '{cfg_file_path}'");
 
-        std::fs::write(cfg_file_path, toml.to_string()).with_context(|| {
-            format!(
-                "Failed to write pageserver config to '{}'",
-                cfg_file_path.display()
-            )
-        })?;
-        info!(
-            "Config successfully written to '{}'",
-            cfg_file_path.display()
-        )
+        std::fs::write(cfg_file_path, toml.to_string())
+            .with_context(|| format!("Failed to write pageserver config to '{cfg_file_path}'"))?;
+        info!("Config successfully written to '{cfg_file_path}'")
     }
 
     Ok(if init {
@@ -223,6 +201,51 @@ fn initialize_config(
     })
 }
 
+struct WaitForPhaseResult<F: std::future::Future + Unpin> {
+    timeout_remaining: Duration,
+    skipped: Option<F>,
+}
+
+/// During startup, we apply a timeout to our waits for readiness, to avoid
+/// stalling the whole service if one Tenant experiences some problem.  Each
+/// phase may consume some of the timeout: this function returns the updated
+/// timeout for use in the next call.
+async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
+where
+    F: std::future::Future + Unpin,
+{
+    let initial_t = Instant::now();
+    let skipped = match tokio::time::timeout(timeout, &mut fut).await {
+        Ok(_) => None,
+        Err(_) => {
+            tracing::info!(
+                timeout_millis = timeout.as_millis(),
+                %phase,
+                "Startup phase timed out, proceeding anyway"
+            );
+            Some(fut)
+        }
+    };
+
+    WaitForPhaseResult {
+        timeout_remaining: timeout
+            .checked_sub(Instant::now().duration_since(initial_t))
+            .unwrap_or(Duration::ZERO),
+        skipped,
+    }
+}
+
+fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
+    let elapsed = started_at.elapsed();
+    let secs = elapsed.as_secs_f64();
+    STARTUP_DURATION.with_label_values(&[phase]).set(secs);
+
+    info!(
+        elapsed_ms = elapsed.as_millis(),
+        "{human_phase} ({secs:.3}s since start)"
+    )
+}
+
 fn start_pageserver(
     launch_ts: &'static LaunchTimestamp,
     conf: &'static PageServerConf,
@@ -230,16 +253,6 @@ fn start_pageserver(
     // Monotonic time for later calculating startup duration
     let started_startup_at = Instant::now();
 
-    let startup_checkpoint = move |phase: &str, human_phase: &str| {
-        let elapsed = started_startup_at.elapsed();
-        let secs = elapsed.as_secs_f64();
-        STARTUP_DURATION.with_label_values(&[phase]).set(secs);
-        info!(
-            elapsed_ms = elapsed.as_millis(),
-            "{human_phase} ({secs:.3}s since start)"
-        )
-    };
-
     // Print version and launch timestamp to the log,
     // and expose them as prometheus metrics.
     // A changed version string indicates changed software.
@@ -346,12 +359,25 @@ fn start_pageserver(
         }
     };
 
+    // Top-level cancellation token for the process
+    let shutdown_pageserver = tokio_util::sync::CancellationToken::new();
+
     // Set up remote storage client
     let remote_storage = create_remote_storage_client(conf)?;
 
+    // Set up deletion queue
+    let (deletion_queue, deletion_workers) = DeletionQueue::new(
+        remote_storage.clone(),
+        ControlPlaneClient::new(conf, &shutdown_pageserver),
+        conf,
+    );
+    if let Some(deletion_workers) = deletion_workers {
+        deletion_workers.spawn_with(BACKGROUND_RUNTIME.handle());
+    }
+
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.
-    startup_checkpoint("initial", "Starting loading tenants");
+    startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
     STARTUP_IS_LOADING.set(1);
 
     // Startup staging or optimizing:
@@ -365,6 +391,7 @@ fn start_pageserver(
     // consumer side) will be dropped once we can start the background jobs. Currently it is behind
     // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
     // (background_task_maximum_delay).
+    let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
     let (init_done_tx, init_done_rx) = utils::completion::channel();
 
     let (init_logical_size_done_tx, init_logical_size_done_rx) = utils::completion::channel();
@@ -372,20 +399,21 @@ fn start_pageserver(
     let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
 
     let order = pageserver::InitializationOrder {
-        initial_tenant_load: Some(init_done_tx),
+        initial_tenant_load_remote: Some(init_done_tx),
+        initial_tenant_load: Some(init_remote_done_tx),
         initial_logical_size_can_start: init_done_rx.clone(),
         initial_logical_size_attempt: Some(init_logical_size_done_tx),
         background_jobs_can_start: background_jobs_barrier.clone(),
     };
 
     // Scan the local 'tenants/' directory and start loading the tenants
-    let shutdown_pageserver = tokio_util::sync::CancellationToken::new();
-
+    let deletion_queue_client = deletion_queue.new_client();
     BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
         conf,
         TenantSharedResources {
             broker_client: broker_client.clone(),
             remote_storage: remote_storage.clone(),
+            deletion_queue_client,
         },
         order,
         shutdown_pageserver.clone(),
@@ -396,55 +424,93 @@ fn start_pageserver(
         let shutdown_pageserver = shutdown_pageserver.clone();
         let drive_init = async move {
             // NOTE: unlike many futures in pageserver, this one is cancellation-safe
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial load completed"));
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial load completed")
+            });
 
-            init_done_rx.wait().await;
-            startup_checkpoint("initial_tenant_load", "Initial load completed");
-            STARTUP_IS_LOADING.set(0);
+            let timeout = conf.background_task_maximum_delay;
+
+            let init_remote_done = std::pin::pin!(async {
+                init_remote_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load_remote",
+                    "Remote part of initial load completed",
+                );
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_remote_skipped,
+            } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
+
+            let init_load_done = std::pin::pin!(async {
+                init_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load",
+                    "Initial load completed",
+                );
+                STARTUP_IS_LOADING.set(0);
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_load_skipped,
+            } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
 
             // initial logical sizes can now start, as they were waiting on init_done_rx.
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            let mut init_sizes_done = std::pin::pin!(init_logical_size_done_rx.wait());
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial logical sizes completed")
+            });
 
-            let timeout = conf.background_task_maximum_delay;
+            let logical_sizes_done = std::pin::pin!(async {
+                init_logical_size_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_logical_sizes",
+                    "Initial logical sizes completed",
+                );
+            });
 
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-
-            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
-                Ok(_) => {
-                    startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed");
-                    None
-                }
-                Err(_) => {
-                    tracing::info!(
-                        timeout_millis = timeout.as_millis(),
-                        "Initial logical size timeout elapsed; starting background jobs"
-                    );
-                    Some(init_sizes_done)
-                }
-            };
+            let WaitForPhaseResult {
+                timeout_remaining: _,
+                skipped: logical_sizes_skipped,
+            } = wait_for_phase("initial_logical_sizes", logical_sizes_done, timeout).await;
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            // allow background jobs to start
+            // allow background jobs to start: we either completed prior stages, or they reached timeout
+            // and were skipped.  It is important that we do not let them block background jobs indefinitely,
+            // because things like consumption metrics for billing are blocked by this barrier.
             drop(background_jobs_can_start);
-            startup_checkpoint("background_jobs_can_start", "Starting background jobs");
-
-            if let Some(init_sizes_done) = init_sizes_done {
-                // ending up here is not a bug; at the latest logical sizes will be queried by
-                // consumption metrics.
-                let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-                init_sizes_done.await;
-
-                scopeguard::ScopeGuard::into_inner(guard);
-
-                startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed after timeout (background jobs already started)");
+            startup_checkpoint(
+                started_startup_at,
+                "background_jobs_can_start",
+                "Starting background jobs",
+            );
 
+            // We are done. If we skipped any phases due to timeout, run them to completion here so that
+            // they will eventually update their startup_checkpoint, and so that we do not declare the
+            // 'complete' stage until all the other stages are really done.
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before waiting for skipped phases done")
+            });
+            if let Some(f) = init_remote_skipped {
+                f.await;
             }
+            if let Some(f) = init_load_skipped {
+                f.await;
+            }
+            if let Some(f) = logical_sizes_skipped {
+                f.await;
+            }
+            scopeguard::ScopeGuard::into_inner(guard);
 
-            startup_checkpoint("complete", "Startup complete");
+            startup_checkpoint(started_startup_at, "complete", "Startup complete");
         };
 
         async move {
@@ -481,9 +547,10 @@ fn start_pageserver(
             http::routes::State::new(
                 conf,
                 http_auth.clone(),
-                remote_storage,
+                remote_storage.clone(),
                 broker_client.clone(),
                 disk_usage_eviction_state,
+                deletion_queue.new_client(),
             )
             .context("Failed to initialize router state")?,
         );
@@ -518,6 +585,9 @@ fn start_pageserver(
             // creates a child context with the right DownloadBehavior.
             DownloadBehavior::Error,
         );
+
+        let local_disk_storage = conf.workdir.join("last_consumption_metrics.json");
+
         task_mgr::spawn(
             crate::BACKGROUND_RUNTIME.handle(),
             TaskKind::MetricsCollection,
@@ -544,6 +614,7 @@ fn start_pageserver(
                     conf.cached_metric_collection_interval,
                     conf.synthetic_size_calculation_interval,
                     conf.id,
+                    local_disk_storage,
                     metrics_ctx,
                 )
                 .instrument(info_span!("metrics_collection"))
@@ -579,6 +650,7 @@ fn start_pageserver(
                     pageserver_listener,
                     conf.pg_auth_type,
                     libpq_ctx,
+                    task_mgr::shutdown_token(),
                 )
                 .await
             },
@@ -607,7 +679,12 @@ fn start_pageserver(
             // Right now that tree doesn't reach very far, and `task_mgr` is used instead.
             // The plan is to change that over time.
             shutdown_pageserver.take();
-            BACKGROUND_RUNTIME.block_on(pageserver::shutdown_pageserver(0));
+            let bg_remote_storage = remote_storage.clone();
+            let bg_deletion_queue = deletion_queue.clone();
+            BACKGROUND_RUNTIME.block_on(pageserver::shutdown_pageserver(
+                bg_remote_storage.map(|_| bg_deletion_queue),
+                0,
+            ));
             unreachable!()
         }
     })
@@ -619,7 +696,7 @@ fn create_remote_storage_client(
     let config = if let Some(config) = &conf.remote_storage_config {
         config
     } else {
-        // No remote storage configured.
+        tracing::warn!("no remote storage configured, this is a deprecated configuration");
         return Ok(None);
     };
 

pageserver/src/config.rs

@@ -11,17 +11,18 @@ use std::env;
 use storage_broker::Uri;
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::ConnectionId;
+use utils::logging::SecretString;
 
 use once_cell::sync::OnceCell;
 use reqwest::Url;
 use std::num::NonZeroUsize;
-use std::path::{Path, PathBuf};
 use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
 use toml_edit;
 use toml_edit::{Document, Item};
 
+use camino::{Utf8Path, Utf8PathBuf};
 use postgres_backend::AuthType;
 use utils::{
     id::{NodeId, TenantId, TimelineId},
@@ -36,8 +37,8 @@ use crate::tenant::{
     TIMELINES_SEGMENT_NAME,
 };
 use crate::{
-    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TIMELINE_DELETE_MARK_SUFFIX,
-    TIMELINE_UNINIT_MARK_SUFFIX,
+    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TENANT_LOCATION_CONFIG_NAME,
+    TIMELINE_DELETE_MARK_SUFFIX, TIMELINE_UNINIT_MARK_SUFFIX,
 };
 
 pub mod defaults {
@@ -64,7 +65,7 @@ pub mod defaults {
         super::ConfigurableSemaphore::DEFAULT_INITIAL.get();
 
     pub const DEFAULT_METRIC_COLLECTION_INTERVAL: &str = "10 min";
-    pub const DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL: &str = "1 hour";
+    pub const DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL: &str = "0s";
     pub const DEFAULT_METRIC_COLLECTION_ENDPOINT: Option<reqwest::Url> = None;
     pub const DEFAULT_SYNTHETIC_SIZE_CALCULATION_INTERVAL: &str = "10 min";
     pub const DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY: &str = "10s";
@@ -152,9 +153,9 @@ pub struct PageServerConf {
     // that during unit testing, because the current directory is global
     // to the process but different unit tests work on different
     // repositories.
-    pub workdir: PathBuf,
+    pub workdir: Utf8PathBuf,
 
-    pub pg_distrib_dir: PathBuf,
+    pub pg_distrib_dir: Utf8PathBuf,
 
     // Authentication
     /// authentication method for the HTTP mgmt API
@@ -163,7 +164,7 @@ pub struct PageServerConf {
     pub pg_auth_type: AuthType,
     /// Path to a file containing public key for verifying JWT tokens.
     /// Used for both mgmt and compute auth, if enabled.
-    pub auth_validation_public_key_path: Option<PathBuf>,
+    pub auth_validation_public_key_path: Option<Utf8PathBuf>,
 
     pub remote_storage_config: Option<RemoteStorageConfig>,
 
@@ -207,6 +208,13 @@ pub struct PageServerConf {
     pub background_task_maximum_delay: Duration,
 
     pub control_plane_api: Option<Url>,
+
+    /// JWT token for use with the control plane API.
+    pub control_plane_api_token: Option<SecretString>,
+
+    /// If true, pageserver will make best-effort to operate without a control plane: only
+    /// for use in major incidents.
+    pub control_plane_emergency_mode: bool,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -249,15 +257,15 @@ struct PageServerConfigBuilder {
     page_cache_size: BuilderValue<usize>,
     max_file_descriptors: BuilderValue<usize>,
 
-    workdir: BuilderValue<PathBuf>,
+    workdir: BuilderValue<Utf8PathBuf>,
 
-    pg_distrib_dir: BuilderValue<PathBuf>,
+    pg_distrib_dir: BuilderValue<Utf8PathBuf>,
 
     http_auth_type: BuilderValue<AuthType>,
     pg_auth_type: BuilderValue<AuthType>,
 
     //
-    auth_validation_public_key_path: BuilderValue<Option<PathBuf>>,
+    auth_validation_public_key_path: BuilderValue<Option<Utf8PathBuf>>,
     remote_storage_config: BuilderValue<Option<RemoteStorageConfig>>,
 
     id: BuilderValue<NodeId>,
@@ -283,6 +291,8 @@ struct PageServerConfigBuilder {
     background_task_maximum_delay: BuilderValue<Duration>,
 
     control_plane_api: BuilderValue<Option<Url>>,
+    control_plane_api_token: BuilderValue<Option<SecretString>>,
+    control_plane_emergency_mode: BuilderValue<bool>,
 }
 
 impl Default for PageServerConfigBuilder {
@@ -300,10 +310,12 @@ impl Default for PageServerConfigBuilder {
             superuser: Set(DEFAULT_SUPERUSER.to_string()),
             page_cache_size: Set(DEFAULT_PAGE_CACHE_SIZE),
             max_file_descriptors: Set(DEFAULT_MAX_FILE_DESCRIPTORS),
-            workdir: Set(PathBuf::new()),
-            pg_distrib_dir: Set(env::current_dir()
-                .expect("cannot access current directory")
-                .join("pg_install")),
+            workdir: Set(Utf8PathBuf::new()),
+            pg_distrib_dir: Set(Utf8PathBuf::from_path_buf(
+                env::current_dir().expect("cannot access current directory"),
+            )
+            .expect("non-Unicode path")
+            .join("pg_install")),
             http_auth_type: Set(AuthType::Trust),
             pg_auth_type: Set(AuthType::Trust),
             auth_validation_public_key_path: Set(None),
@@ -347,6 +359,8 @@ impl Default for PageServerConfigBuilder {
             .unwrap()),
 
             control_plane_api: Set(None),
+            control_plane_api_token: Set(None),
+            control_plane_emergency_mode: Set(false),
         }
     }
 }
@@ -384,11 +398,11 @@ impl PageServerConfigBuilder {
         self.max_file_descriptors = BuilderValue::Set(max_file_descriptors)
     }
 
-    pub fn workdir(&mut self, workdir: PathBuf) {
+    pub fn workdir(&mut self, workdir: Utf8PathBuf) {
         self.workdir = BuilderValue::Set(workdir)
     }
 
-    pub fn pg_distrib_dir(&mut self, pg_distrib_dir: PathBuf) {
+    pub fn pg_distrib_dir(&mut self, pg_distrib_dir: Utf8PathBuf) {
         self.pg_distrib_dir = BuilderValue::Set(pg_distrib_dir)
     }
 
@@ -402,7 +416,7 @@ impl PageServerConfigBuilder {
 
     pub fn auth_validation_public_key_path(
         &mut self,
-        auth_validation_public_key_path: Option<PathBuf>,
+        auth_validation_public_key_path: Option<Utf8PathBuf>,
     ) {
         self.auth_validation_public_key_path = BuilderValue::Set(auth_validation_public_key_path)
     }
@@ -475,8 +489,16 @@ impl PageServerConfigBuilder {
         self.background_task_maximum_delay = BuilderValue::Set(delay);
     }
 
-    pub fn control_plane_api(&mut self, api: Url) {
-        self.control_plane_api = BuilderValue::Set(Some(api))
+    pub fn control_plane_api(&mut self, api: Option<Url>) {
+        self.control_plane_api = BuilderValue::Set(api)
+    }
+
+    pub fn control_plane_api_token(&mut self, token: Option<SecretString>) {
+        self.control_plane_api_token = BuilderValue::Set(token)
+    }
+
+    pub fn control_plane_emergency_mode(&mut self, enabled: bool) {
+        self.control_plane_emergency_mode = BuilderValue::Set(enabled)
     }
 
     pub fn build(self) -> anyhow::Result<PageServerConf> {
@@ -567,6 +589,12 @@ impl PageServerConfigBuilder {
             control_plane_api: self
                 .control_plane_api
                 .ok_or(anyhow!("missing control_plane_api"))?,
+            control_plane_api_token: self
+                .control_plane_api_token
+                .ok_or(anyhow!("missing control_plane_api_token"))?,
+            control_plane_emergency_mode: self
+                .control_plane_emergency_mode
+                .ok_or(anyhow!("missing control_plane_emergency_mode"))?,
         })
     }
 }
@@ -576,34 +604,63 @@ impl PageServerConf {
     // Repository paths, relative to workdir.
     //
 
-    pub fn tenants_path(&self) -> PathBuf {
+    pub fn tenants_path(&self) -> Utf8PathBuf {
         self.workdir.join(TENANTS_SEGMENT_NAME)
     }
 
-    pub fn tenant_path(&self, tenant_id: &TenantId) -> PathBuf {
+    pub fn deletion_prefix(&self) -> Utf8PathBuf {
+        self.workdir.join("deletion")
+    }
+
+    pub fn deletion_list_path(&self, sequence: u64) -> Utf8PathBuf {
+        // Encode a version in the filename, so that if we ever switch away from JSON we can
+        // increment this.
+        const VERSION: u8 = 1;
+
+        self.deletion_prefix()
+            .join(format!("{sequence:016x}-{VERSION:02x}.list"))
+    }
+
+    pub fn deletion_header_path(&self) -> Utf8PathBuf {
+        // Encode a version in the filename, so that if we ever switch away from JSON we can
+        // increment this.
+        const VERSION: u8 = 1;
+
+        self.deletion_prefix().join(format!("header-{VERSION:02x}"))
+    }
+
+    pub fn tenant_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenants_path().join(tenant_id.to_string())
     }
 
-    pub fn tenant_attaching_mark_file_path(&self, tenant_id: &TenantId) -> PathBuf {
+    pub fn tenant_attaching_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id)
             .join(TENANT_ATTACHING_MARKER_FILENAME)
     }
 
-    pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> PathBuf {
+    pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(IGNORED_TENANT_FILE_NAME)
     }
 
     /// Points to a place in pageserver's local directory,
     /// where certain tenant's tenantconf file should be located.
-    pub fn tenant_config_path(&self, tenant_id: &TenantId) -> PathBuf {
+    ///
+    /// Legacy: superseded by tenant_location_config_path.  Eventually
+    /// remove this function.
+    pub fn tenant_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(TENANT_CONFIG_NAME)
     }
 
-    pub fn timelines_path(&self, tenant_id: &TenantId) -> PathBuf {
+    pub fn tenant_location_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
+        self.tenant_path(tenant_id)
+            .join(TENANT_LOCATION_CONFIG_NAME)
+    }
+
+    pub fn timelines_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(TIMELINES_SEGMENT_NAME)
     }
 
-    pub fn timeline_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> PathBuf {
+    pub fn timeline_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> Utf8PathBuf {
         self.timelines_path(tenant_id).join(timeline_id.to_string())
     }
 
@@ -611,7 +668,7 @@ impl PageServerConf {
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
-    ) -> PathBuf {
+    ) -> Utf8PathBuf {
         path_with_suffix_extension(
             self.timeline_path(&tenant_id, &timeline_id),
             TIMELINE_UNINIT_MARK_SUFFIX,
@@ -622,19 +679,19 @@ impl PageServerConf {
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
-    ) -> PathBuf {
+    ) -> Utf8PathBuf {
         path_with_suffix_extension(
             self.timeline_path(&tenant_id, &timeline_id),
             TIMELINE_DELETE_MARK_SUFFIX,
         )
     }
 
-    pub fn tenant_deleted_mark_file_path(&self, tenant_id: &TenantId) -> PathBuf {
+    pub fn tenant_deleted_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id)
             .join(TENANT_DELETED_MARKER_FILE_NAME)
     }
 
-    pub fn traces_path(&self) -> PathBuf {
+    pub fn traces_path(&self) -> Utf8PathBuf {
         self.workdir.join("traces")
     }
 
@@ -643,7 +700,7 @@ impl PageServerConf {
         tenant_id: &TenantId,
         timeline_id: &TimelineId,
         connection_id: &ConnectionId,
-    ) -> PathBuf {
+    ) -> Utf8PathBuf {
         self.traces_path()
             .join(tenant_id.to_string())
             .join(timeline_id.to_string())
@@ -652,20 +709,20 @@ impl PageServerConf {
 
     /// Points to a place in pageserver's local directory,
     /// where certain timeline's metadata file should be located.
-    pub fn metadata_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> PathBuf {
+    pub fn metadata_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> Utf8PathBuf {
         self.timeline_path(tenant_id, timeline_id)
             .join(METADATA_FILE_NAME)
     }
 
     /// Turns storage remote path of a file into its local path.
-    pub fn local_path(&self, remote_path: &RemotePath) -> PathBuf {
+    pub fn local_path(&self, remote_path: &RemotePath) -> Utf8PathBuf {
         remote_path.with_base(&self.workdir)
     }
 
     //
     // Postgres distribution paths
     //
-    pub fn pg_distrib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_distrib_dir(&self, pg_version: u32) -> anyhow::Result<Utf8PathBuf> {
         let path = self.pg_distrib_dir.clone();
 
         #[allow(clippy::manual_range_patterns)]
@@ -675,10 +732,10 @@ impl PageServerConf {
         }
     }
 
-    pub fn pg_bin_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_bin_dir(&self, pg_version: u32) -> anyhow::Result<Utf8PathBuf> {
         Ok(self.pg_distrib_dir(pg_version)?.join("bin"))
     }
-    pub fn pg_lib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_lib_dir(&self, pg_version: u32) -> anyhow::Result<Utf8PathBuf> {
         Ok(self.pg_distrib_dir(pg_version)?.join("lib"))
     }
 
@@ -686,7 +743,7 @@ impl PageServerConf {
     /// validating the input and failing on errors.
     ///
     /// This leaves any options not present in the file in the built-in defaults.
-    pub fn parse_and_validate(toml: &Document, workdir: &Path) -> anyhow::Result<Self> {
+    pub fn parse_and_validate(toml: &Document, workdir: &Utf8Path) -> anyhow::Result<Self> {
         let mut builder = PageServerConfigBuilder::default();
         builder.workdir(workdir.to_owned());
 
@@ -705,10 +762,10 @@ impl PageServerConf {
                     builder.max_file_descriptors(parse_toml_u64(key, item)? as usize)
                 }
                 "pg_distrib_dir" => {
-                    builder.pg_distrib_dir(PathBuf::from(parse_toml_string(key, item)?))
+                    builder.pg_distrib_dir(Utf8PathBuf::from(parse_toml_string(key, item)?))
                 }
                 "auth_validation_public_key_path" => builder.auth_validation_public_key_path(Some(
-                    PathBuf::from(parse_toml_string(key, item)?),
+                    Utf8PathBuf::from(parse_toml_string(key, item)?),
                 )),
                 "http_auth_type" => builder.http_auth_type(parse_toml_from_str(key, item)?),
                 "pg_auth_type" => builder.pg_auth_type(parse_toml_from_str(key, item)?),
@@ -747,7 +804,26 @@ impl PageServerConf {
                 },
                 "ondemand_download_behavior_treat_error_as_warn" => builder.ondemand_download_behavior_treat_error_as_warn(parse_toml_bool(key, item)?),
                 "background_task_maximum_delay" => builder.background_task_maximum_delay(parse_toml_duration(key, item)?),
-                "control_plane_api" => builder.control_plane_api(parse_toml_string(key, item)?.parse().context("failed to parse control plane URL")?),
+                "control_plane_api" => {
+                    let parsed = parse_toml_string(key, item)?;
+                    if parsed.is_empty() {
+                        builder.control_plane_api(None)
+                    } else {
+                        builder.control_plane_api(Some(parsed.parse().context("failed to parse control plane URL")?))
+                    }
+                },
+                "control_plane_api_token" => {
+                    let parsed = parse_toml_string(key, item)?;
+                    if parsed.is_empty() {
+                        builder.control_plane_api_token(None)
+                    } else {
+                        builder.control_plane_api_token(Some(parsed.into()))
+                    }
+                },
+                "control_plane_emergency_mode" => {
+                    builder.control_plane_emergency_mode(parse_toml_bool(key, item)?)
+
+                },
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -761,8 +837,7 @@ impl PageServerConf {
             ensure!(
                 auth_validation_public_key_path.exists(),
                 format!(
-                    "Can't find auth_validation_public_key at '{}'",
-                    auth_validation_public_key_path.display()
+                    "Can't find auth_validation_public_key at '{auth_validation_public_key_path}'",
                 )
             );
         }
@@ -878,12 +953,12 @@ impl PageServerConf {
     }
 
     #[cfg(test)]
-    pub fn test_repo_dir(test_name: &str) -> PathBuf {
-        PathBuf::from(format!("../tmp_check/test_{test_name}"))
+    pub fn test_repo_dir(test_name: &str) -> Utf8PathBuf {
+        Utf8PathBuf::from(format!("../tmp_check/test_{test_name}"))
     }
 
-    pub fn dummy_conf(repo_dir: PathBuf) -> Self {
-        let pg_distrib_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../pg_install");
+    pub fn dummy_conf(repo_dir: Utf8PathBuf) -> Self {
+        let pg_distrib_dir = Utf8PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../pg_install");
 
         PageServerConf {
             id: NodeId(0),
@@ -917,6 +992,8 @@ impl PageServerConf {
             ondemand_download_behavior_treat_error_as_warn: false,
             background_task_maximum_delay: Duration::ZERO,
             control_plane_api: None,
+            control_plane_api_token: None,
+            control_plane_emergency_mode: false,
         }
     }
 }
@@ -1049,8 +1126,8 @@ mod tests {
         num::{NonZeroU32, NonZeroUsize},
     };
 
+    use camino_tempfile::{tempdir, Utf8TempDir};
     use remote_storage::{RemoteStorageKind, S3Config};
-    use tempfile::{tempdir, TempDir};
     use utils::serde_percent::Percent;
 
     use super::*;
@@ -1089,8 +1166,7 @@ background_task_maximum_delay = '334 s'
         let broker_endpoint = storage_broker::DEFAULT_ENDPOINT;
         // we have to create dummy values to overcome the validation errors
         let config_string = format!(
-            "pg_distrib_dir='{}'\nid=10\nbroker_endpoint = '{broker_endpoint}'",
-            pg_distrib_dir.display()
+            "pg_distrib_dir='{pg_distrib_dir}'\nid=10\nbroker_endpoint = '{broker_endpoint}'",
         );
         let toml = config_string.parse()?;
 
@@ -1140,7 +1216,9 @@ background_task_maximum_delay = '334 s'
                 background_task_maximum_delay: humantime::parse_duration(
                     defaults::DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY
                 )?,
-                control_plane_api: None
+                control_plane_api: None,
+                control_plane_api_token: None,
+                control_plane_emergency_mode: false
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1155,8 +1233,7 @@ background_task_maximum_delay = '334 s'
         let broker_endpoint = storage_broker::DEFAULT_ENDPOINT;
 
         let config_string = format!(
-            "{ALL_BASE_VALUES_TOML}pg_distrib_dir='{}'\nbroker_endpoint = '{broker_endpoint}'",
-            pg_distrib_dir.display()
+            "{ALL_BASE_VALUES_TOML}pg_distrib_dir='{pg_distrib_dir}'\nbroker_endpoint = '{broker_endpoint}'",
         );
         let toml = config_string.parse()?;
 
@@ -1196,7 +1273,9 @@ background_task_maximum_delay = '334 s'
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
                 background_task_maximum_delay: Duration::from_secs(334),
-                control_plane_api: None
+                control_plane_api: None,
+                control_plane_api_token: None,
+                control_plane_emergency_mode: false
             },
             "Should be able to parse all basic config values correctly"
         );
@@ -1215,23 +1294,18 @@ background_task_maximum_delay = '334 s'
         let identical_toml_declarations = &[
             format!(
                 r#"[remote_storage]
-local_path = '{}'"#,
-                local_storage_path.display()
-            ),
-            format!(
-                "remote_storage={{local_path='{}'}}",
-                local_storage_path.display()
+local_path = '{local_storage_path}'"#,
             ),
+            format!("remote_storage={{local_path='{local_storage_path}'}}"),
         ];
 
         for remote_storage_config_str in identical_toml_declarations {
             let config_string = format!(
                 r#"{ALL_BASE_VALUES_TOML}
-pg_distrib_dir='{}'
+pg_distrib_dir='{pg_distrib_dir}'
 broker_endpoint = '{broker_endpoint}'
 
 {remote_storage_config_str}"#,
-                pg_distrib_dir.display(),
             );
 
             let toml = config_string.parse()?;
@@ -1294,11 +1368,10 @@ concurrency_limit = {s3_concurrency_limit}"#
         for remote_storage_config_str in identical_toml_declarations {
             let config_string = format!(
                 r#"{ALL_BASE_VALUES_TOML}
-pg_distrib_dir='{}'
+pg_distrib_dir='{pg_distrib_dir}'
 broker_endpoint = '{broker_endpoint}'
 
 {remote_storage_config_str}"#,
-                pg_distrib_dir.display(),
             );
 
             let toml = config_string.parse()?;
@@ -1340,12 +1413,11 @@ broker_endpoint = '{broker_endpoint}'
 
         let config_string = format!(
             r#"{ALL_BASE_VALUES_TOML}
-pg_distrib_dir='{}'
+pg_distrib_dir='{pg_distrib_dir}'
 broker_endpoint = '{broker_endpoint}'
 
 [tenant_config]
 trace_read_requests = {trace_read_requests}"#,
-            pg_distrib_dir.display(),
         );
 
         let toml = config_string.parse()?;
@@ -1365,7 +1437,7 @@ trace_read_requests = {trace_read_requests}"#,
         let (workdir, pg_distrib_dir) = prepare_fs(&tempdir)?;
 
         let pageserver_conf_toml = format!(
-            r#"pg_distrib_dir = "{}"
+            r#"pg_distrib_dir = "{pg_distrib_dir}"
 metric_collection_endpoint = "http://sample.url"
 metric_collection_interval = "10min"
 id = 222
@@ -1383,7 +1455,6 @@ kind = "LayerAccessThreshold"
 period = "20m"
 threshold = "20m"
 "#,
-            pg_distrib_dir.display(),
         );
         let toml: Document = pageserver_conf_toml.parse()?;
         let conf = PageServerConf::parse_and_validate(&toml, &workdir)?;
@@ -1408,6 +1479,8 @@ threshold = "20m"
             Some(DiskUsageEvictionTaskConfig {
                 max_usage_pct: Percent::new(80).unwrap(),
                 min_avail_bytes: 0,
+                target_avail_bytes: None,
+                target_usage_pct: None,
                 period: Duration::from_secs(10),
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
@@ -1424,7 +1497,7 @@ threshold = "20m"
         Ok(())
     }
 
-    fn prepare_fs(tempdir: &TempDir) -> anyhow::Result<(PathBuf, PathBuf)> {
+    fn prepare_fs(tempdir: &Utf8TempDir) -> anyhow::Result<(Utf8PathBuf, Utf8PathBuf)> {
         let tempdir_path = tempdir.path();
 
         let workdir = tempdir_path.join("workdir");

pageserver/src/consumption_metrics.rs

@@ -1,188 +1,56 @@
-//!
 //! Periodically collect consumption metrics for all active tenants
 //! and push them to a HTTP endpoint.
-//! Cache metrics to send only the updated ones.
-//!
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
+use crate::tenant::tasks::BackgroundLoopKind;
 use crate::tenant::{mgr, LogicalSizeCalculationCause};
-use anyhow;
-use chrono::{DateTime, Utc};
-use consumption_metrics::{idempotency_key, Event, EventChunk, EventType, CHUNK_SIZE};
+use camino::Utf8PathBuf;
+use consumption_metrics::EventType;
 use pageserver_api::models::TenantState;
 use reqwest::Url;
-use serde::Serialize;
-use serde_with::{serde_as, DisplayFromStr};
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::{Duration, SystemTime};
+use tokio::time::Instant;
 use tracing::*;
-use utils::id::{NodeId, TenantId, TimelineId};
-use utils::lsn::Lsn;
+use utils::id::NodeId;
+
+mod metrics;
+use metrics::MetricsKey;
+mod disk_cache;
+mod upload;
 
 const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
 
-#[serde_as]
-#[derive(Serialize, Debug, Clone, Copy)]
-struct Ids {
-    #[serde_as(as = "DisplayFromStr")]
-    tenant_id: TenantId,
-    #[serde_as(as = "Option<DisplayFromStr>")]
-    #[serde(skip_serializing_if = "Option::is_none")]
-    timeline_id: Option<TimelineId>,
-}
+/// Basically a key-value pair, but usually in a Vec except for [`Cache`].
+///
+/// This is as opposed to `consumption_metrics::Event` which is the externally communicated form.
+/// Difference is basically the missing idempotency key, which lives only for the duration of
+/// upload attempts.
+type RawMetric = (MetricsKey, (EventType, u64));
 
-/// Key that uniquely identifies the object, this metric describes.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-struct MetricsKey {
-    tenant_id: TenantId,
-    timeline_id: Option<TimelineId>,
-    metric: &'static str,
-}
-
-impl MetricsKey {
-    const fn absolute_values(self) -> AbsoluteValueFactory {
-        AbsoluteValueFactory(self)
-    }
-    const fn incremental_values(self) -> IncrementalValueFactory {
-        IncrementalValueFactory(self)
-    }
-}
-
-/// Helper type which each individual metric kind can return to produce only absolute values.
-struct AbsoluteValueFactory(MetricsKey);
-
-impl AbsoluteValueFactory {
-    fn at(self, time: DateTime<Utc>, val: u64) -> (MetricsKey, (EventType, u64)) {
-        let key = self.0;
-        (key, (EventType::Absolute { time }, val))
-    }
-}
-
-/// Helper type which each individual metric kind can return to produce only incremental values.
-struct IncrementalValueFactory(MetricsKey);
-
-impl IncrementalValueFactory {
-    #[allow(clippy::wrong_self_convention)]
-    fn from_previous_up_to(
-        self,
-        prev_end: DateTime<Utc>,
-        up_to: DateTime<Utc>,
-        val: u64,
-    ) -> (MetricsKey, (EventType, u64)) {
-        let key = self.0;
-        // cannot assert prev_end < up_to because these are realtime clock based
-        (
-            key,
-            (
-                EventType::Incremental {
-                    start_time: prev_end,
-                    stop_time: up_to,
-                },
-                val,
-            ),
-        )
-    }
-
-    fn key(&self) -> &MetricsKey {
-        &self.0
-    }
-}
-
-// the static part of a MetricsKey
-impl MetricsKey {
-    /// Absolute value of [`Timeline::get_last_record_lsn`].
-    ///
-    /// [`Timeline::get_last_record_lsn`]: crate::tenant::Timeline::get_last_record_lsn
-    const fn written_size(tenant_id: TenantId, timeline_id: TimelineId) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: Some(timeline_id),
-            metric: "written_size",
-        }
-        .absolute_values()
-    }
-
-    /// Values will be the difference of the latest [`MetricsKey::written_size`] to what we
-    /// previously sent, starting from the previously sent incremental time range ending at the
-    /// latest absolute measurement.
-    const fn written_size_delta(
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-    ) -> IncrementalValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: Some(timeline_id),
-            // the name here is correctly about data not size, because that is what is wanted by
-            // downstream pipeline
-            metric: "written_data_bytes_delta",
-        }
-        .incremental_values()
-    }
-
-    /// Exact [`Timeline::get_current_logical_size`].
-    ///
-    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
-    const fn timeline_logical_size(
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-    ) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: Some(timeline_id),
-            metric: "timeline_logical_size",
-        }
-        .absolute_values()
-    }
-
-    /// [`Tenant::remote_size`]
-    ///
-    /// [`Tenant::remote_size`]: crate::tenant::Tenant::remote_size
-    const fn remote_storage_size(tenant_id: TenantId) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: None,
-            metric: "remote_storage_size",
-        }
-        .absolute_values()
-    }
-
-    /// Sum of [`Timeline::resident_physical_size`] for each `Tenant`.
-    ///
-    /// [`Timeline::resident_physical_size`]: crate::tenant::Timeline::resident_physical_size
-    const fn resident_size(tenant_id: TenantId) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: None,
-            metric: "resident_size",
-        }
-        .absolute_values()
-    }
-
-    /// [`Tenant::cached_synthetic_size`] as refreshed by [`calculate_synthetic_size_worker`].
-    ///
-    /// [`Tenant::cached_synthetic_size`]: crate::tenant::Tenant::cached_synthetic_size
-    const fn synthetic_size(tenant_id: TenantId) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: None,
-            metric: "synthetic_storage_size",
-        }
-        .absolute_values()
-    }
-}
+/// Caches the [`RawMetric`]s
+///
+/// In practice, during startup, last sent values are stored here to be used in calculating new
+/// ones. After successful uploading, the cached values are updated to cache. This used to be used
+/// for deduplication, but that is no longer needed.
+type Cache = HashMap<MetricsKey, (EventType, u64)>;
 
 /// Main thread that serves metrics collection
 pub async fn collect_metrics(
     metric_collection_endpoint: &Url,
     metric_collection_interval: Duration,
-    cached_metric_collection_interval: Duration,
+    _cached_metric_collection_interval: Duration,
     synthetic_size_calculation_interval: Duration,
     node_id: NodeId,
+    local_disk_storage: Utf8PathBuf,
     ctx: RequestContext,
 ) -> anyhow::Result<()> {
-    let mut ticker = tokio::time::interval(metric_collection_interval);
-    info!("starting collect_metrics");
+    if _cached_metric_collection_interval != Duration::ZERO {
+        tracing::warn!(
+            "cached_metric_collection_interval is no longer used, please set it to zero."
+        )
+    }
 
     // spin up background worker that caclulates tenant sizes
     let worker_ctx =
@@ -202,543 +70,227 @@ pub async fn collect_metrics(
         },
     );
 
+    let path: Arc<Utf8PathBuf> = Arc::new(local_disk_storage);
+
+    let cancel = task_mgr::shutdown_token();
+
+    let restore_and_reschedule = restore_and_reschedule(&path, metric_collection_interval);
+
+    let mut cached_metrics = tokio::select! {
+        _ = cancel.cancelled() => return Ok(()),
+        ret = restore_and_reschedule => ret,
+    };
+
     // define client here to reuse it for all requests
     let client = reqwest::ClientBuilder::new()
         .timeout(DEFAULT_HTTP_REPORTING_TIMEOUT)
         .build()
         .expect("Failed to create http client with timeout");
-    let mut cached_metrics = HashMap::new();
-    let mut prev_iteration_time: std::time::Instant = std::time::Instant::now();
-
-    loop {
-        tokio::select! {
-            _ = task_mgr::shutdown_watcher() => {
-                info!("collect_metrics received cancellation request");
-                return Ok(());
-            },
-            tick_at = ticker.tick() => {
-
-                // send cached metrics every cached_metric_collection_interval
-                let send_cached = prev_iteration_time.elapsed() >= cached_metric_collection_interval;
-
-                if send_cached {
-                    prev_iteration_time = std::time::Instant::now();
-                }
-
-                collect_metrics_iteration(&client, &mut cached_metrics, metric_collection_endpoint, node_id, &ctx, send_cached).await;
-
-                crate::tenant::tasks::warn_when_period_overrun(
-                    tick_at.elapsed(),
-                    metric_collection_interval,
-                    "consumption_metrics_collect_metrics",
-                );
-            }
-        }
-    }
-}
-
-/// One iteration of metrics collection
-///
-/// Gather per-tenant and per-timeline metrics and send them to the `metric_collection_endpoint`.
-/// Cache metrics to avoid sending the same metrics multiple times.
-///
-/// This function handles all errors internally
-/// and doesn't break iteration if just one tenant fails.
-///
-/// TODO
-/// - refactor this function (chunking+sending part) to reuse it in proxy module;
-async fn collect_metrics_iteration(
-    client: &reqwest::Client,
-    cached_metrics: &mut HashMap<MetricsKey, (EventType, u64)>,
-    metric_collection_endpoint: &reqwest::Url,
-    node_id: NodeId,
-    ctx: &RequestContext,
-    send_cached: bool,
-) {
-    let mut current_metrics: Vec<(MetricsKey, (EventType, u64))> = Vec::new();
-    trace!(
-        "starting collect_metrics_iteration. metric_collection_endpoint: {}",
-        metric_collection_endpoint
-    );
-
-    // get list of tenants
-    let tenants = match mgr::list_tenants().await {
-        Ok(tenants) => tenants,
-        Err(err) => {
-            error!("failed to list tenants: {:?}", err);
-            return;
-        }
-    };
-
-    // iterate through list of Active tenants and collect metrics
-    for (tenant_id, tenant_state) in tenants {
-        if tenant_state != TenantState::Active {
-            continue;
-        }
-
-        let tenant = match mgr::get_tenant(tenant_id, true).await {
-            Ok(tenant) => tenant,
-            Err(err) => {
-                // It is possible that tenant was deleted between
-                // `list_tenants` and `get_tenant`, so just warn about it.
-                warn!("failed to get tenant {tenant_id:?}: {err:?}");
-                continue;
-            }
-        };
-
-        let mut tenant_resident_size = 0;
-
-        // iterate through list of timelines in tenant
-        for timeline in tenant.list_timelines() {
-            // collect per-timeline metrics only for active timelines
-
-            let timeline_id = timeline.timeline_id;
-
-            match TimelineSnapshot::collect(&timeline, ctx) {
-                Ok(Some(snap)) => {
-                    snap.to_metrics(
-                        tenant_id,
-                        timeline_id,
-                        Utc::now(),
-                        &mut current_metrics,
-                        cached_metrics,
-                    );
-                }
-                Ok(None) => {}
-                Err(e) => {
-                    error!(
-                        "failed to get metrics values for tenant {tenant_id} timeline {}: {e:#?}",
-                        timeline.timeline_id
-                    );
-                    continue;
-                }
-            }
-
-            tenant_resident_size += timeline.resident_physical_size();
-        }
-
-        current_metrics
-            .push(MetricsKey::remote_storage_size(tenant_id).at(Utc::now(), tenant.remote_size()));
-
-        current_metrics
-            .push(MetricsKey::resident_size(tenant_id).at(Utc::now(), tenant_resident_size));
-
-        // Note that this metric is calculated in a separate bgworker
-        // Here we only use cached value, which may lag behind the real latest one
-        let synthetic_size = tenant.cached_synthetic_size();
-
-        if synthetic_size != 0 {
-            // only send non-zeroes because otherwise these show up as errors in logs
-            current_metrics
-                .push(MetricsKey::synthetic_size(tenant_id).at(Utc::now(), synthetic_size));
-        }
-    }
-
-    // Filter metrics, unless we want to send all metrics, including cached ones.
-    // See: https://github.com/neondatabase/neon/issues/3485
-    if !send_cached {
-        current_metrics.retain(|(curr_key, (kind, curr_val))| {
-            if kind.is_incremental() {
-                // incremental values (currently only written_size_delta) should not get any cache
-                // deduplication because they will be used by upstream for "is still alive."
-                true
-            } else {
-                match cached_metrics.get(curr_key) {
-                    Some((_, val)) => val != curr_val,
-                    None => true,
-                }
-            }
-        });
-    }
-
-    if current_metrics.is_empty() {
-        trace!("no new metrics to send");
-        return;
-    }
-
-    // Send metrics.
-    // Split into chunks of 1000 metrics to avoid exceeding the max request size
-    let chunks = current_metrics.chunks(CHUNK_SIZE);
-
-    let mut chunk_to_send: Vec<Event<Ids>> = Vec::with_capacity(CHUNK_SIZE);
 
     let node_id = node_id.to_string();
 
-    for chunk in chunks {
-        chunk_to_send.clear();
+    loop {
+        let started_at = Instant::now();
 
-        // enrich metrics with type,timestamp and idempotency key before sending
-        chunk_to_send.extend(chunk.iter().map(|(curr_key, (when, curr_val))| Event {
-            kind: *when,
-            metric: curr_key.metric,
-            idempotency_key: idempotency_key(&node_id),
-            value: *curr_val,
-            extra: Ids {
-                tenant_id: curr_key.tenant_id,
-                timeline_id: curr_key.timeline_id,
-            },
-        }));
+        // these are point in time, with variable "now"
+        let metrics = metrics::collect_all_metrics(&cached_metrics, &ctx).await;
 
-        const MAX_RETRIES: u32 = 3;
+        let metrics = Arc::new(metrics);
 
-        for attempt in 0..MAX_RETRIES {
-            let res = client
-                .post(metric_collection_endpoint.clone())
-                .json(&EventChunk {
-                    events: (&chunk_to_send).into(),
-                })
-                .send()
-                .await;
+        // why not race cancellation here? because we are one of the last tasks, and if we are
+        // already here, better to try to flush the new values.
 
-            match res {
-                Ok(res) => {
-                    if res.status().is_success() {
-                        // update cached metrics after they were sent successfully
-                        for (curr_key, curr_val) in chunk.iter() {
-                            cached_metrics.insert(curr_key.clone(), *curr_val);
-                        }
-                    } else {
-                        error!("metrics endpoint refused the sent metrics: {:?}", res);
-                        for metric in chunk_to_send
-                            .iter()
-                            .filter(|metric| metric.value > (1u64 << 40))
-                        {
-                            // Report if the metric value is suspiciously large
-                            error!("potentially abnormal metric value: {:?}", metric);
-                        }
-                    }
-                    break;
+        let flush = async {
+            match disk_cache::flush_metrics_to_disk(&metrics, &path).await {
+                Ok(()) => {
+                    tracing::debug!("flushed metrics to disk");
                 }
-                Err(err) if err.is_timeout() => {
-                    error!(attempt, "timeout sending metrics, retrying immediately");
-                    continue;
-                }
-                Err(err) => {
-                    error!(attempt, ?err, "failed to send metrics");
-                    break;
+                Err(e) => {
+                    // idea here is that if someone creates a directory as our path, then they
+                    // might notice it from the logs before shutdown and remove it
+                    tracing::error!("failed to persist metrics to {path:?}: {e:#}");
                 }
             }
+        };
+
+        let upload = async {
+            let res = upload::upload_metrics(
+                &client,
+                metric_collection_endpoint,
+                &cancel,
+                &node_id,
+                &metrics,
+                &mut cached_metrics,
+            )
+            .await;
+            if let Err(e) = res {
+                // serialization error which should never happen
+                tracing::error!("failed to upload due to {e:#}");
+            }
+        };
+
+        // let these run concurrently
+        let (_, _) = tokio::join!(flush, upload);
+
+        crate::tenant::tasks::warn_when_period_overrun(
+            started_at.elapsed(),
+            metric_collection_interval,
+            BackgroundLoopKind::ConsumptionMetricsCollectMetrics,
+        );
+
+        let res = tokio::time::timeout_at(
+            started_at + metric_collection_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
+    }
+}
+
+/// Called on the first iteration in an attempt to join the metric uploading schedule from previous
+/// pageserver session. Pageserver is supposed to upload at intervals regardless of restarts.
+///
+/// Cancellation safe.
+async fn restore_and_reschedule(
+    path: &Arc<Utf8PathBuf>,
+    metric_collection_interval: Duration,
+) -> Cache {
+    let (cached, earlier_metric_at) = match disk_cache::read_metrics_from_disk(path.clone()).await {
+        Ok(found_some) => {
+            // there is no min needed because we write these sequentially in
+            // collect_all_metrics
+            let earlier_metric_at = found_some
+                .iter()
+                .map(|(_, (et, _))| et.recorded_at())
+                .copied()
+                .next();
+
+            let cached = found_some.into_iter().collect::<Cache>();
+
+            (cached, earlier_metric_at)
+        }
+        Err(e) => {
+            use std::io::{Error, ErrorKind};
+
+            let root = e.root_cause();
+            let maybe_ioerr = root.downcast_ref::<Error>();
+            let is_not_found = maybe_ioerr.is_some_and(|e| e.kind() == ErrorKind::NotFound);
+
+            if !is_not_found {
+                tracing::info!("failed to read any previous metrics from {path:?}: {e:#}");
+            }
+
+            (HashMap::new(), None)
+        }
+    };
+
+    if let Some(earlier_metric_at) = earlier_metric_at {
+        let earlier_metric_at: SystemTime = earlier_metric_at.into();
+
+        let error = reschedule(earlier_metric_at, metric_collection_interval).await;
+
+        if let Some(error) = error {
+            if error.as_secs() >= 60 {
+                tracing::info!(
+                    error_ms = error.as_millis(),
+                    "startup scheduling error due to restart"
+                )
+            }
         }
     }
+
+    cached
 }
 
-/// Internal type to make timeline metric production testable.
-///
-/// As this value type contains all of the information needed from a timeline to produce the
-/// metrics, it can easily be created with different values in test.
-struct TimelineSnapshot {
-    loaded_at: (Lsn, SystemTime),
-    last_record_lsn: Lsn,
-    current_exact_logical_size: Option<u64>,
-}
+async fn reschedule(
+    earlier_metric_at: SystemTime,
+    metric_collection_interval: Duration,
+) -> Option<Duration> {
+    let now = SystemTime::now();
+    match now.duration_since(earlier_metric_at) {
+        Ok(from_last_send) if from_last_send < metric_collection_interval => {
+            let sleep_for = metric_collection_interval - from_last_send;
 
-impl TimelineSnapshot {
-    /// Collect the metrics from an actual timeline.
-    ///
-    /// Fails currently only when [`Timeline::get_current_logical_size`] fails.
-    ///
-    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
-    fn collect(
-        t: &Arc<crate::tenant::Timeline>,
-        ctx: &RequestContext,
-    ) -> anyhow::Result<Option<Self>> {
-        use anyhow::Context;
+            let deadline = std::time::Instant::now() + sleep_for;
 
-        if !t.is_active() {
-            // no collection for broken or stopping needed, we will still keep the cached values
-            // though at the caller.
-            Ok(None)
-        } else {
-            let loaded_at = t.loaded_at;
-            let last_record_lsn = t.get_last_record_lsn();
+            tokio::time::sleep_until(deadline.into()).await;
 
-            let current_exact_logical_size = {
-                let span = info_span!("collect_metrics_iteration", tenant_id = %t.tenant_id, timeline_id = %t.timeline_id);
-                let res = span
-                    .in_scope(|| t.get_current_logical_size(ctx))
-                    .context("get_current_logical_size");
-                match res? {
-                    // Only send timeline logical size when it is fully calculated.
-                    (size, is_exact) if is_exact => Some(size),
-                    (_, _) => None,
-                }
-            };
+            let now = std::time::Instant::now();
 
-            Ok(Some(TimelineSnapshot {
-                loaded_at,
-                last_record_lsn,
-                current_exact_logical_size,
-            }))
-        }
-    }
-
-    /// Produce the timeline consumption metrics into the `metrics` argument.
-    fn to_metrics(
-        &self,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        now: DateTime<Utc>,
-        metrics: &mut Vec<(MetricsKey, (EventType, u64))>,
-        cache: &HashMap<MetricsKey, (EventType, u64)>,
-    ) {
-        let timeline_written_size = u64::from(self.last_record_lsn);
-
-        let (key, written_size_now) =
-            MetricsKey::written_size(tenant_id, timeline_id).at(now, timeline_written_size);
-
-        // last_record_lsn can only go up, right now at least, TODO: #2592 or related
-        // features might change this.
-
-        let written_size_delta_key = MetricsKey::written_size_delta(tenant_id, timeline_id);
-
-        // use this when available, because in a stream of incremental values, it will be
-        // accurate where as when last_record_lsn stops moving, we will only cache the last
-        // one of those.
-        let last_stop_time = cache
-            .get(written_size_delta_key.key())
-            .map(|(until, _val)| {
-                until
-                    .incremental_timerange()
-                    .expect("never create EventType::Absolute for written_size_delta")
-                    .end
-            });
-
-        // by default, use the last sent written_size as the basis for
-        // calculating the delta. if we don't yet have one, use the load time value.
-        let prev = cache
-            .get(&key)
-            .map(|(prev_at, prev)| {
-                // use the prev time from our last incremental update, or default to latest
-                // absolute update on the first round.
-                let prev_at = prev_at
-                    .absolute_time()
-                    .expect("never create EventType::Incremental for written_size");
-                let prev_at = last_stop_time.unwrap_or(prev_at);
-                (*prev_at, *prev)
-            })
-            .unwrap_or_else(|| {
-                // if we don't have a previous point of comparison, compare to the load time
-                // lsn.
-                let (disk_consistent_lsn, loaded_at) = &self.loaded_at;
-                (DateTime::from(*loaded_at), disk_consistent_lsn.0)
-            });
-
-        // written_size_bytes_delta
-        metrics.extend(
-            if let Some(delta) = written_size_now.1.checked_sub(prev.1) {
-                let up_to = written_size_now
-                    .0
-                    .absolute_time()
-                    .expect("never create EventType::Incremental for written_size");
-                let key_value = written_size_delta_key.from_previous_up_to(prev.0, *up_to, delta);
-                Some(key_value)
+            // executor threads might be busy, add extra measurements
+            Some(if now < deadline {
+                deadline - now
             } else {
-                None
-            },
-        );
-
-        // written_size
-        metrics.push((key, written_size_now));
-
-        if let Some(size) = self.current_exact_logical_size {
-            metrics.push(MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, size));
+                now - deadline
+            })
+        }
+        Ok(from_last_send) => Some(from_last_send.saturating_sub(metric_collection_interval)),
+        Err(_) => {
+            tracing::warn!(
+                ?now,
+                ?earlier_metric_at,
+                "oldest recorded metric is in future; first values will come out with inconsistent timestamps"
+            );
+            earlier_metric_at.duration_since(now).ok()
         }
     }
 }
 
 /// Caclculate synthetic size for each active tenant
-pub async fn calculate_synthetic_size_worker(
+async fn calculate_synthetic_size_worker(
     synthetic_size_calculation_interval: Duration,
     ctx: &RequestContext,
 ) -> anyhow::Result<()> {
     info!("starting calculate_synthetic_size_worker");
-
-    let mut ticker = tokio::time::interval(synthetic_size_calculation_interval);
-
-    loop {
-        tokio::select! {
-            _ = task_mgr::shutdown_watcher() => {
-                return Ok(());
-            },
-            tick_at = ticker.tick() => {
-
-                let tenants = match mgr::list_tenants().await {
-                    Ok(tenants) => tenants,
-                    Err(e) => {
-                        warn!("cannot get tenant list: {e:#}");
-                        continue;
-                    }
-                };
-                // iterate through list of Active tenants and collect metrics
-                for (tenant_id, tenant_state) in tenants {
-
-                    if tenant_state != TenantState::Active {
-                        continue;
-                    }
-
-                    if let Ok(tenant) = mgr::get_tenant(tenant_id, true).await
-                    {
-                        if let Err(e) = tenant.calculate_synthetic_size(
-                            LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize,
-                            ctx).await {
-                            error!("failed to calculate synthetic size for tenant {}: {}", tenant_id, e);
-                        }
-                    }
-
-                }
-
-                crate::tenant::tasks::warn_when_period_overrun(
-                    tick_at.elapsed(),
-                    synthetic_size_calculation_interval,
-                    "consumption_metrics_synthetic_size_worker",
-                );
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::collections::HashMap;
-
-    use std::time::SystemTime;
-    use utils::{
-        id::{TenantId, TimelineId},
-        lsn::Lsn,
+    scopeguard::defer! {
+        info!("calculate_synthetic_size_worker stopped");
     };
 
-    use crate::consumption_metrics::MetricsKey;
+    let cause = LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize;
 
-    use super::TimelineSnapshot;
-    use chrono::{DateTime, Utc};
+    loop {
+        let started_at = Instant::now();
 
-    #[test]
-    fn startup_collected_timeline_metrics_before_advancing() {
-        let tenant_id = TenantId::generate();
-        let timeline_id = TimelineId::generate();
-
-        let mut metrics = Vec::new();
-        let cache = HashMap::new();
-
-        let initdb_lsn = Lsn(0x10000);
-        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
-
-        let snap = TimelineSnapshot {
-            loaded_at: (disk_consistent_lsn, SystemTime::now()),
-            last_record_lsn: disk_consistent_lsn,
-            current_exact_logical_size: Some(0x42000),
+        let tenants = match mgr::list_tenants().await {
+            Ok(tenants) => tenants,
+            Err(e) => {
+                warn!("cannot get tenant list: {e:#}");
+                continue;
+            }
         };
 
-        let now = DateTime::<Utc>::from(SystemTime::now());
+        for (tenant_id, tenant_state) in tenants {
+            if tenant_state != TenantState::Active {
+                continue;
+            }
 
-        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
-
-        assert_eq!(
-            metrics,
-            &[
-                MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
-                    snap.loaded_at.1.into(),
-                    now,
-                    0
-                ),
-                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
-            ]
-        );
-    }
-
-    #[test]
-    fn startup_collected_timeline_metrics_second_round() {
-        let tenant_id = TenantId::generate();
-        let timeline_id = TimelineId::generate();
-
-        let [now, before, init] = time_backwards();
-
-        let now = DateTime::<Utc>::from(now);
-        let before = DateTime::<Utc>::from(before);
-
-        let initdb_lsn = Lsn(0x10000);
-        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
-
-        let mut metrics = Vec::new();
-        let cache = HashMap::from([
-            MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0)
-        ]);
-
-        let snap = TimelineSnapshot {
-            loaded_at: (disk_consistent_lsn, init),
-            last_record_lsn: disk_consistent_lsn,
-            current_exact_logical_size: Some(0x42000),
-        };
-
-        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
-
-        assert_eq!(
-            metrics,
-            &[
-                MetricsKey::written_size_delta(tenant_id, timeline_id)
-                    .from_previous_up_to(before, now, 0),
-                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
-            ]
-        );
-    }
-
-    #[test]
-    fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
-        let tenant_id = TenantId::generate();
-        let timeline_id = TimelineId::generate();
-
-        let [now, just_before, before, init] = time_backwards();
-
-        let now = DateTime::<Utc>::from(now);
-        let just_before = DateTime::<Utc>::from(just_before);
-        let before = DateTime::<Utc>::from(before);
-
-        let initdb_lsn = Lsn(0x10000);
-        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
-
-        let mut metrics = Vec::new();
-        let cache = HashMap::from([
-            // at t=before was the last time the last_record_lsn changed
-            MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0),
-            // end time of this event is used for the next ones
-            MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
-                before,
-                just_before,
-                0,
-            ),
-        ]);
-
-        let snap = TimelineSnapshot {
-            loaded_at: (disk_consistent_lsn, init),
-            last_record_lsn: disk_consistent_lsn,
-            current_exact_logical_size: Some(0x42000),
-        };
-
-        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
-
-        assert_eq!(
-            metrics,
-            &[
-                MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
-                    just_before,
-                    now,
-                    0
-                ),
-                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
-            ]
-        );
-    }
-
-    fn time_backwards<const N: usize>() -> [std::time::SystemTime; N] {
-        let mut times = [std::time::SystemTime::UNIX_EPOCH; N];
-        times[0] = std::time::SystemTime::now();
-        for behind in 1..N {
-            times[behind] = times[0] - std::time::Duration::from_secs(behind as u64);
+            if let Ok(tenant) = mgr::get_tenant(tenant_id, true).await {
+                // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
+                // We can put in some prioritization for consumption metrics.
+                // Same for the loop that fetches computed metrics.
+                // By using the same limiter, we centralize metrics collection for "start" and "finished" counters,
+                // which turns out is really handy to understand the system.
+                if let Err(e) = tenant.calculate_synthetic_size(cause, ctx).await {
+                    error!("failed to calculate synthetic size for tenant {tenant_id}: {e:#}");
+                }
+            }
         }
 
-        times
+        crate::tenant::tasks::warn_when_period_overrun(
+            started_at.elapsed(),
+            synthetic_size_calculation_interval,
+            BackgroundLoopKind::ConsumptionMetricsSyntheticSizeWorker,
+        );
+
+        let res = tokio::time::timeout_at(
+            started_at + synthetic_size_calculation_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
     }
 }

pageserver/src/consumption_metrics/disk_cache.rs

@@ -0,0 +1,119 @@
+use anyhow::Context;
+use camino::{Utf8Path, Utf8PathBuf};
+use std::sync::Arc;
+
+use super::RawMetric;
+
+pub(super) async fn read_metrics_from_disk(
+    path: Arc<Utf8PathBuf>,
+) -> anyhow::Result<Vec<RawMetric>> {
+    // do not add context to each error, callsite will log with full path
+    let span = tracing::Span::current();
+    tokio::task::spawn_blocking(move || {
+        let _e = span.entered();
+
+        if let Some(parent) = path.parent() {
+            if let Err(e) = scan_and_delete_with_same_prefix(&path) {
+                tracing::info!("failed to cleanup temporary files in {parent:?}: {e:#}");
+            }
+        }
+
+        let mut file = std::fs::File::open(&*path)?;
+        let reader = std::io::BufReader::new(&mut file);
+        anyhow::Ok(serde_json::from_reader::<_, Vec<RawMetric>>(reader)?)
+    })
+    .await
+    .context("read metrics join error")
+    .and_then(|x| x)
+}
+
+fn scan_and_delete_with_same_prefix(path: &Utf8Path) -> std::io::Result<()> {
+    let it = std::fs::read_dir(path.parent().expect("caller checked"))?;
+
+    let prefix = path.file_name().expect("caller checked").to_string();
+
+    for entry in it {
+        let entry = entry?;
+        if !entry.metadata()?.is_file() {
+            continue;
+        }
+        let file_name = entry.file_name();
+
+        if path.file_name().unwrap() == file_name {
+            // do not remove our actual file
+            continue;
+        }
+
+        let file_name = file_name.to_string_lossy();
+
+        if !file_name.starts_with(&*prefix) {
+            continue;
+        }
+
+        let path = entry.path();
+
+        if let Err(e) = std::fs::remove_file(&path) {
+            tracing::warn!("cleaning up old tempfile {file_name:?} failed: {e:#}");
+        } else {
+            tracing::info!("cleaned up old tempfile {file_name:?}");
+        }
+    }
+
+    Ok(())
+}
+
+pub(super) async fn flush_metrics_to_disk(
+    current_metrics: &Arc<Vec<RawMetric>>,
+    path: &Arc<Utf8PathBuf>,
+) -> anyhow::Result<()> {
+    use std::io::Write;
+
+    anyhow::ensure!(path.parent().is_some(), "path must have parent: {path:?}");
+    anyhow::ensure!(
+        path.file_name().is_some(),
+        "path must have filename: {path:?}"
+    );
+
+    let span = tracing::Span::current();
+    tokio::task::spawn_blocking({
+        let current_metrics = current_metrics.clone();
+        let path = path.clone();
+        move || {
+            let _e = span.entered();
+
+            let parent = path.parent().expect("existence checked");
+            let file_name = path.file_name().expect("existence checked");
+            let mut tempfile = camino_tempfile::Builder::new()
+                .prefix(file_name)
+                .suffix(".tmp")
+                .tempfile_in(parent)?;
+
+            tracing::debug!("using tempfile {:?}", tempfile.path());
+
+            // write out all of the raw metrics, to be read out later on restart as cached values
+            {
+                let mut writer = std::io::BufWriter::new(&mut tempfile);
+                serde_json::to_writer(&mut writer, &*current_metrics)
+                    .context("serialize metrics")?;
+                writer
+                    .into_inner()
+                    .map_err(|_| anyhow::anyhow!("flushing metrics failed"))?;
+            }
+
+            tempfile.flush()?;
+            tempfile.as_file().sync_all()?;
+
+            fail::fail_point!("before-persist-last-metrics-collected");
+
+            drop(tempfile.persist(&*path).map_err(|e| e.error)?);
+
+            let f = std::fs::File::open(path.parent().unwrap())?;
+            f.sync_all()?;
+
+            anyhow::Ok(())
+        }
+    })
+    .await
+    .with_context(|| format!("write metrics to {path:?} join error"))
+    .and_then(|x| x.with_context(|| format!("write metrics to {path:?}")))
+}

pageserver/src/consumption_metrics/metrics.rs

@@ -0,0 +1,455 @@
+use crate::context::RequestContext;
+use anyhow::Context;
+use chrono::{DateTime, Utc};
+use consumption_metrics::EventType;
+use futures::stream::StreamExt;
+use serde_with::serde_as;
+use std::{sync::Arc, time::SystemTime};
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+};
+
+use super::{Cache, RawMetric};
+
+/// Name of the metric, used by `MetricsKey` factory methods and `deserialize_cached_events`
+/// instead of static str.
+// Do not rename any of these without first consulting with data team and partner
+// management.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
+pub(super) enum Name {
+    /// Timeline last_record_lsn, absolute
+    #[serde(rename = "written_size")]
+    WrittenSize,
+    /// Timeline last_record_lsn, incremental
+    #[serde(rename = "written_data_bytes_delta")]
+    WrittenSizeDelta,
+    /// Timeline logical size
+    #[serde(rename = "timeline_logical_size")]
+    LogicalSize,
+    /// Tenant remote size
+    #[serde(rename = "remote_storage_size")]
+    RemoteSize,
+    /// Tenant resident size
+    #[serde(rename = "resident_size")]
+    ResidentSize,
+    /// Tenant synthetic size
+    #[serde(rename = "synthetic_storage_size")]
+    SyntheticSize,
+}
+
+/// Key that uniquely identifies the object this metric describes.
+///
+/// This is a denormalization done at the MetricsKey const methods; these should not be constructed
+/// elsewhere.
+#[serde_with::serde_as]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
+pub(crate) struct MetricsKey {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub(super) tenant_id: TenantId,
+
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(super) timeline_id: Option<TimelineId>,
+
+    pub(super) metric: Name,
+}
+
+impl MetricsKey {
+    const fn absolute_values(self) -> AbsoluteValueFactory {
+        AbsoluteValueFactory(self)
+    }
+    const fn incremental_values(self) -> IncrementalValueFactory {
+        IncrementalValueFactory(self)
+    }
+}
+
+/// Helper type which each individual metric kind can return to produce only absolute values.
+struct AbsoluteValueFactory(MetricsKey);
+
+impl AbsoluteValueFactory {
+    const fn at(self, time: DateTime<Utc>, val: u64) -> RawMetric {
+        let key = self.0;
+        (key, (EventType::Absolute { time }, val))
+    }
+
+    fn key(&self) -> &MetricsKey {
+        &self.0
+    }
+}
+
+/// Helper type which each individual metric kind can return to produce only incremental values.
+struct IncrementalValueFactory(MetricsKey);
+
+impl IncrementalValueFactory {
+    #[allow(clippy::wrong_self_convention)]
+    const fn from_until(
+        self,
+        prev_end: DateTime<Utc>,
+        up_to: DateTime<Utc>,
+        val: u64,
+    ) -> RawMetric {
+        let key = self.0;
+        // cannot assert prev_end < up_to because these are realtime clock based
+        let when = EventType::Incremental {
+            start_time: prev_end,
+            stop_time: up_to,
+        };
+        (key, (when, val))
+    }
+
+    fn key(&self) -> &MetricsKey {
+        &self.0
+    }
+}
+
+// the static part of a MetricsKey
+impl MetricsKey {
+    /// Absolute value of [`Timeline::get_last_record_lsn`].
+    ///
+    /// [`Timeline::get_last_record_lsn`]: crate::tenant::Timeline::get_last_record_lsn
+    const fn written_size(tenant_id: TenantId, timeline_id: TimelineId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            metric: Name::WrittenSize,
+        }
+        .absolute_values()
+    }
+
+    /// Values will be the difference of the latest [`MetricsKey::written_size`] to what we
+    /// previously sent, starting from the previously sent incremental time range ending at the
+    /// latest absolute measurement.
+    const fn written_size_delta(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> IncrementalValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            metric: Name::WrittenSizeDelta,
+        }
+        .incremental_values()
+    }
+
+    /// Exact [`Timeline::get_current_logical_size`].
+    ///
+    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
+    const fn timeline_logical_size(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            metric: Name::LogicalSize,
+        }
+        .absolute_values()
+    }
+
+    /// [`Tenant::remote_size`]
+    ///
+    /// [`Tenant::remote_size`]: crate::tenant::Tenant::remote_size
+    const fn remote_storage_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: Name::RemoteSize,
+        }
+        .absolute_values()
+    }
+
+    /// Sum of [`Timeline::resident_physical_size`] for each `Tenant`.
+    ///
+    /// [`Timeline::resident_physical_size`]: crate::tenant::Timeline::resident_physical_size
+    const fn resident_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: Name::ResidentSize,
+        }
+        .absolute_values()
+    }
+
+    /// [`Tenant::cached_synthetic_size`] as refreshed by [`calculate_synthetic_size_worker`].
+    ///
+    /// [`Tenant::cached_synthetic_size`]: crate::tenant::Tenant::cached_synthetic_size
+    /// [`calculate_synthetic_size_worker`]: super::calculate_synthetic_size_worker
+    const fn synthetic_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: Name::SyntheticSize,
+        }
+        .absolute_values()
+    }
+}
+
+pub(super) async fn collect_all_metrics(
+    cached_metrics: &Cache,
+    ctx: &RequestContext,
+) -> Vec<RawMetric> {
+    use pageserver_api::models::TenantState;
+
+    let started_at = std::time::Instant::now();
+
+    let tenants = match crate::tenant::mgr::list_tenants().await {
+        Ok(tenants) => tenants,
+        Err(err) => {
+            tracing::error!("failed to list tenants: {:?}", err);
+            return vec![];
+        }
+    };
+
+    let tenants = futures::stream::iter(tenants).filter_map(|(id, state)| async move {
+        if state != TenantState::Active {
+            None
+        } else {
+            crate::tenant::mgr::get_tenant(id, true)
+                .await
+                .ok()
+                .map(|tenant| (id, tenant))
+        }
+    });
+
+    let res = collect(tenants, cached_metrics, ctx).await;
+
+    tracing::info!(
+        elapsed_ms = started_at.elapsed().as_millis(),
+        total = res.len(),
+        "collected metrics"
+    );
+
+    res
+}
+
+async fn collect<S>(tenants: S, cache: &Cache, ctx: &RequestContext) -> Vec<RawMetric>
+where
+    S: futures::stream::Stream<Item = (TenantId, Arc<crate::tenant::Tenant>)>,
+{
+    let mut current_metrics: Vec<RawMetric> = Vec::new();
+
+    let mut tenants = std::pin::pin!(tenants);
+
+    while let Some((tenant_id, tenant)) = tenants.next().await {
+        let mut tenant_resident_size = 0;
+
+        for timeline in tenant.list_timelines() {
+            let timeline_id = timeline.timeline_id;
+
+            match TimelineSnapshot::collect(&timeline, ctx) {
+                Ok(Some(snap)) => {
+                    snap.to_metrics(
+                        tenant_id,
+                        timeline_id,
+                        Utc::now(),
+                        &mut current_metrics,
+                        cache,
+                    );
+                }
+                Ok(None) => {}
+                Err(e) => {
+                    tracing::error!(
+                        "failed to get metrics values for tenant {tenant_id} timeline {}: {e:#?}",
+                        timeline.timeline_id
+                    );
+                    continue;
+                }
+            }
+
+            tenant_resident_size += timeline.resident_physical_size();
+        }
+
+        let snap = TenantSnapshot::collect(&tenant, tenant_resident_size);
+        snap.to_metrics(tenant_id, Utc::now(), cache, &mut current_metrics);
+    }
+
+    current_metrics
+}
+
+/// In-between abstraction to allow testing metrics without actual Tenants.
+struct TenantSnapshot {
+    resident_size: u64,
+    remote_size: u64,
+    synthetic_size: u64,
+}
+
+impl TenantSnapshot {
+    /// Collect tenant status to have metrics created out of it.
+    ///
+    /// `resident_size` is calculated of the timelines we had access to for other metrics, so we
+    /// cannot just list timelines here.
+    fn collect(t: &Arc<crate::tenant::Tenant>, resident_size: u64) -> Self {
+        TenantSnapshot {
+            resident_size,
+            remote_size: t.remote_size(),
+            // Note that this metric is calculated in a separate bgworker
+            // Here we only use cached value, which may lag behind the real latest one
+            synthetic_size: t.cached_synthetic_size(),
+        }
+    }
+
+    fn to_metrics(
+        &self,
+        tenant_id: TenantId,
+        now: DateTime<Utc>,
+        cached: &Cache,
+        metrics: &mut Vec<RawMetric>,
+    ) {
+        let remote_size = MetricsKey::remote_storage_size(tenant_id).at(now, self.remote_size);
+
+        let resident_size = MetricsKey::resident_size(tenant_id).at(now, self.resident_size);
+
+        let synthetic_size = {
+            let factory = MetricsKey::synthetic_size(tenant_id);
+            let mut synthetic_size = self.synthetic_size;
+
+            if synthetic_size == 0 {
+                if let Some((_, value)) = cached.get(factory.key()) {
+                    // use the latest value from previous session
+                    synthetic_size = *value;
+                }
+            }
+
+            if synthetic_size != 0 {
+                // only send non-zeroes because otherwise these show up as errors in logs
+                Some(factory.at(now, synthetic_size))
+            } else {
+                None
+            }
+        };
+
+        metrics.extend(
+            [Some(remote_size), Some(resident_size), synthetic_size]
+                .into_iter()
+                .flatten(),
+        );
+    }
+}
+
+/// Internal type to make timeline metric production testable.
+///
+/// As this value type contains all of the information needed from a timeline to produce the
+/// metrics, it can easily be created with different values in test.
+struct TimelineSnapshot {
+    loaded_at: (Lsn, SystemTime),
+    last_record_lsn: Lsn,
+    current_exact_logical_size: Option<u64>,
+}
+
+impl TimelineSnapshot {
+    /// Collect the metrics from an actual timeline.
+    ///
+    /// Fails currently only when [`Timeline::get_current_logical_size`] fails.
+    ///
+    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
+    fn collect(
+        t: &Arc<crate::tenant::Timeline>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Option<Self>> {
+        if !t.is_active() {
+            // no collection for broken or stopping needed, we will still keep the cached values
+            // though at the caller.
+            Ok(None)
+        } else {
+            let loaded_at = t.loaded_at;
+            let last_record_lsn = t.get_last_record_lsn();
+
+            let current_exact_logical_size = {
+                let span = tracing::info_span!("collect_metrics_iteration", tenant_id = %t.tenant_id, timeline_id = %t.timeline_id);
+                let res = span
+                    .in_scope(|| t.get_current_logical_size(ctx))
+                    .context("get_current_logical_size");
+                match res? {
+                    // Only send timeline logical size when it is fully calculated.
+                    (size, is_exact) if is_exact => Some(size),
+                    (_, _) => None,
+                }
+            };
+
+            Ok(Some(TimelineSnapshot {
+                loaded_at,
+                last_record_lsn,
+                current_exact_logical_size,
+            }))
+        }
+    }
+
+    /// Produce the timeline consumption metrics into the `metrics` argument.
+    fn to_metrics(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        now: DateTime<Utc>,
+        metrics: &mut Vec<RawMetric>,
+        cache: &Cache,
+    ) {
+        let timeline_written_size = u64::from(self.last_record_lsn);
+
+        let written_size_delta_key = MetricsKey::written_size_delta(tenant_id, timeline_id);
+
+        let last_stop_time = cache
+            .get(written_size_delta_key.key())
+            .map(|(until, _val)| {
+                until
+                    .incremental_timerange()
+                    .expect("never create EventType::Absolute for written_size_delta")
+                    .end
+            });
+
+        let (key, written_size_now) =
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, timeline_written_size);
+
+        // by default, use the last sent written_size as the basis for
+        // calculating the delta. if we don't yet have one, use the load time value.
+        let prev = cache
+            .get(&key)
+            .map(|(prev_at, prev)| {
+                // use the prev time from our last incremental update, or default to latest
+                // absolute update on the first round.
+                let prev_at = prev_at
+                    .absolute_time()
+                    .expect("never create EventType::Incremental for written_size");
+                let prev_at = last_stop_time.unwrap_or(prev_at);
+                (*prev_at, *prev)
+            })
+            .unwrap_or_else(|| {
+                // if we don't have a previous point of comparison, compare to the load time
+                // lsn.
+                let (disk_consistent_lsn, loaded_at) = &self.loaded_at;
+                (DateTime::from(*loaded_at), disk_consistent_lsn.0)
+            });
+
+        let up_to = now;
+
+        if let Some(delta) = written_size_now.1.checked_sub(prev.1) {
+            let key_value = written_size_delta_key.from_until(prev.0, up_to, delta);
+            // written_size_delta
+            metrics.push(key_value);
+            // written_size
+            metrics.push((key, written_size_now));
+        } else {
+            // the cached value was ahead of us, report zero until we've caught up
+            metrics.push(written_size_delta_key.from_until(prev.0, up_to, 0));
+            // the cached value was ahead of us, report the same until we've caught up
+            metrics.push((key, (written_size_now.0, prev.1)));
+        }
+
+        {
+            let factory = MetricsKey::timeline_logical_size(tenant_id, timeline_id);
+            let current_or_previous = self
+                .current_exact_logical_size
+                .or_else(|| cache.get(factory.key()).map(|(_, val)| *val));
+
+            if let Some(size) = current_or_previous {
+                metrics.push(factory.at(now, size));
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests;
+
+#[cfg(test)]
+pub(crate) use tests::metric_examples;

pageserver/src/consumption_metrics/metrics/tests.rs

@@ -0,0 +1,297 @@
+use super::*;
+use std::collections::HashMap;
+use std::time::SystemTime;
+use utils::lsn::Lsn;
+
+#[test]
+fn startup_collected_timeline_metrics_before_advancing() {
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+
+    let mut metrics = Vec::new();
+    let cache = HashMap::new();
+
+    let initdb_lsn = Lsn(0x10000);
+    let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+    let snap = TimelineSnapshot {
+        loaded_at: (disk_consistent_lsn, SystemTime::now()),
+        last_record_lsn: disk_consistent_lsn,
+        current_exact_logical_size: Some(0x42000),
+    };
+
+    let now = DateTime::<Utc>::from(SystemTime::now());
+
+    snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(
+                snap.loaded_at.1.into(),
+                now,
+                0
+            ),
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+        ]
+    );
+}
+
+#[test]
+fn startup_collected_timeline_metrics_second_round() {
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+
+    let [now, before, init] = time_backwards();
+
+    let now = DateTime::<Utc>::from(now);
+    let before = DateTime::<Utc>::from(before);
+
+    let initdb_lsn = Lsn(0x10000);
+    let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+    let mut metrics = Vec::new();
+    let cache = HashMap::from([
+        MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0)
+    ]);
+
+    let snap = TimelineSnapshot {
+        loaded_at: (disk_consistent_lsn, init),
+        last_record_lsn: disk_consistent_lsn,
+        current_exact_logical_size: Some(0x42000),
+    };
+
+    snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, now, 0),
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+        ]
+    );
+}
+
+#[test]
+fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+
+    let [now, just_before, before, init] = time_backwards();
+
+    let now = DateTime::<Utc>::from(now);
+    let just_before = DateTime::<Utc>::from(just_before);
+    let before = DateTime::<Utc>::from(before);
+
+    let initdb_lsn = Lsn(0x10000);
+    let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+    let mut metrics = Vec::new();
+    let cache = HashMap::from([
+        // at t=before was the last time the last_record_lsn changed
+        MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0),
+        // end time of this event is used for the next ones
+        MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, just_before, 0),
+    ]);
+
+    let snap = TimelineSnapshot {
+        loaded_at: (disk_consistent_lsn, init),
+        last_record_lsn: disk_consistent_lsn,
+        current_exact_logical_size: Some(0x42000),
+    };
+
+    snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(just_before, now, 0),
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+        ]
+    );
+}
+
+#[test]
+fn post_restart_written_sizes_with_rolled_back_last_record_lsn() {
+    // it can happen that we lose the inmemorylayer but have previously sent metrics and we
+    // should never go backwards
+
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+
+    let [later, now, at_restart] = time_backwards();
+
+    // FIXME: tests would be so much easier if we did not need to juggle back and forth
+    // SystemTime and DateTime::<Utc> ... Could do the conversion only at upload time?
+    let now = DateTime::<Utc>::from(now);
+    let later = DateTime::<Utc>::from(later);
+    let before_restart = at_restart - std::time::Duration::from_secs(5 * 60);
+    let way_before = before_restart - std::time::Duration::from_secs(10 * 60);
+    let before_restart = DateTime::<Utc>::from(before_restart);
+    let way_before = DateTime::<Utc>::from(way_before);
+
+    let snap = TimelineSnapshot {
+        loaded_at: (Lsn(50), at_restart),
+        last_record_lsn: Lsn(50),
+        current_exact_logical_size: None,
+    };
+
+    let mut cache = HashMap::from([
+        MetricsKey::written_size(tenant_id, timeline_id).at(before_restart, 100),
+        MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(
+            way_before,
+            before_restart,
+            // not taken into account, but the timestamps are important
+            999_999_999,
+        ),
+    ]);
+
+    let mut metrics = Vec::new();
+    snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(
+                before_restart,
+                now,
+                0
+            ),
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, 100),
+        ]
+    );
+
+    // now if we cache these metrics, and re-run while "still in recovery"
+    cache.extend(metrics.drain(..));
+
+    // "still in recovery", because our snapshot did not change
+    snap.to_metrics(tenant_id, timeline_id, later, &mut metrics, &cache);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(now, later, 0),
+            MetricsKey::written_size(tenant_id, timeline_id).at(later, 100),
+        ]
+    );
+}
+
+#[test]
+fn post_restart_current_exact_logical_size_uses_cached() {
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+
+    let [now, at_restart] = time_backwards();
+
+    let now = DateTime::<Utc>::from(now);
+    let before_restart = at_restart - std::time::Duration::from_secs(5 * 60);
+    let before_restart = DateTime::<Utc>::from(before_restart);
+
+    let snap = TimelineSnapshot {
+        loaded_at: (Lsn(50), at_restart),
+        last_record_lsn: Lsn(50),
+        current_exact_logical_size: None,
+    };
+
+    let cache = HashMap::from([
+        MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(before_restart, 100)
+    ]);
+
+    let mut metrics = Vec::new();
+    snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+    metrics.retain(|(key, _)| key.metric == Name::LogicalSize);
+
+    assert_eq!(
+        metrics,
+        &[MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 100)]
+    );
+}
+
+#[test]
+fn post_restart_synthetic_size_uses_cached_if_available() {
+    let tenant_id = TenantId::generate();
+
+    let ts = TenantSnapshot {
+        resident_size: 1000,
+        remote_size: 1000,
+        // not yet calculated
+        synthetic_size: 0,
+    };
+
+    let now = SystemTime::now();
+    let before_restart = DateTime::<Utc>::from(now - std::time::Duration::from_secs(5 * 60));
+    let now = DateTime::<Utc>::from(now);
+
+    let cached = HashMap::from([MetricsKey::synthetic_size(tenant_id).at(before_restart, 1000)]);
+
+    let mut metrics = Vec::new();
+    ts.to_metrics(tenant_id, now, &cached, &mut metrics);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::remote_storage_size(tenant_id).at(now, 1000),
+            MetricsKey::resident_size(tenant_id).at(now, 1000),
+            MetricsKey::synthetic_size(tenant_id).at(now, 1000),
+        ]
+    );
+}
+
+#[test]
+fn post_restart_synthetic_size_is_not_sent_when_not_cached() {
+    let tenant_id = TenantId::generate();
+
+    let ts = TenantSnapshot {
+        resident_size: 1000,
+        remote_size: 1000,
+        // not yet calculated
+        synthetic_size: 0,
+    };
+
+    let now = SystemTime::now();
+    let now = DateTime::<Utc>::from(now);
+
+    let cached = HashMap::new();
+
+    let mut metrics = Vec::new();
+    ts.to_metrics(tenant_id, now, &cached, &mut metrics);
+
+    assert_eq!(
+        metrics,
+        &[
+            MetricsKey::remote_storage_size(tenant_id).at(now, 1000),
+            MetricsKey::resident_size(tenant_id).at(now, 1000),
+            // no synthetic size here
+        ]
+    );
+}
+
+fn time_backwards<const N: usize>() -> [std::time::SystemTime; N] {
+    let mut times = [std::time::SystemTime::UNIX_EPOCH; N];
+    times[0] = std::time::SystemTime::now();
+    for behind in 1..N {
+        times[behind] = times[0] - std::time::Duration::from_secs(behind as u64);
+    }
+
+    times
+}
+
+pub(crate) const fn metric_examples(
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+    now: DateTime<Utc>,
+    before: DateTime<Utc>,
+) -> [RawMetric; 6] {
+    [
+        MetricsKey::written_size(tenant_id, timeline_id).at(now, 0),
+        MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, now, 0),
+        MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0),
+        MetricsKey::remote_storage_size(tenant_id).at(now, 0),
+        MetricsKey::resident_size(tenant_id).at(now, 0),
+        MetricsKey::synthetic_size(tenant_id).at(now, 1),
+    ]
+}

pageserver/src/consumption_metrics/upload.rs

@@ -0,0 +1,443 @@
+use consumption_metrics::{Event, EventChunk, IdempotencyKey, CHUNK_SIZE};
+use serde_with::serde_as;
+use tokio_util::sync::CancellationToken;
+use tracing::Instrument;
+
+use super::{metrics::Name, Cache, MetricsKey, RawMetric};
+use utils::id::{TenantId, TimelineId};
+
+/// How the metrics from pageserver are identified.
+#[serde_with::serde_as]
+#[derive(serde::Serialize, serde::Deserialize, Debug, Clone, Copy, PartialEq)]
+struct Ids {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub(super) tenant_id: TenantId,
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(super) timeline_id: Option<TimelineId>,
+}
+
+#[tracing::instrument(skip_all, fields(metrics_total = %metrics.len()))]
+pub(super) async fn upload_metrics(
+    client: &reqwest::Client,
+    metric_collection_endpoint: &reqwest::Url,
+    cancel: &CancellationToken,
+    node_id: &str,
+    metrics: &[RawMetric],
+    cached_metrics: &mut Cache,
+) -> anyhow::Result<()> {
+    let mut uploaded = 0;
+    let mut failed = 0;
+
+    let started_at = std::time::Instant::now();
+
+    let mut iter = serialize_in_chunks(CHUNK_SIZE, metrics, node_id);
+
+    while let Some(res) = iter.next() {
+        let (chunk, body) = res?;
+
+        let event_bytes = body.len();
+
+        let is_last = iter.len() == 0;
+
+        let res = upload(client, metric_collection_endpoint, body, cancel, is_last)
+            .instrument(tracing::info_span!(
+                "upload",
+                %event_bytes,
+                uploaded,
+                total = metrics.len(),
+            ))
+            .await;
+
+        match res {
+            Ok(()) => {
+                for (curr_key, curr_val) in chunk {
+                    cached_metrics.insert(*curr_key, *curr_val);
+                }
+                uploaded += chunk.len();
+            }
+            Err(_) => {
+                // failure(s) have already been logged
+                //
+                // however this is an inconsistency: if we crash here, we will start with the
+                // values as uploaded. in practice, the rejections no longer happen.
+                failed += chunk.len();
+            }
+        }
+    }
+
+    let elapsed = started_at.elapsed();
+
+    tracing::info!(
+        uploaded,
+        failed,
+        elapsed_ms = elapsed.as_millis(),
+        "done sending metrics"
+    );
+
+    Ok(())
+}
+
+// The return type is quite ugly, but we gain testability in isolation
+fn serialize_in_chunks<'a, F>(
+    chunk_size: usize,
+    input: &'a [RawMetric],
+    factory: F,
+) -> impl ExactSizeIterator<Item = Result<(&'a [RawMetric], bytes::Bytes), serde_json::Error>> + 'a
+where
+    F: KeyGen<'a> + 'a,
+{
+    use bytes::BufMut;
+
+    struct Iter<'a, F> {
+        inner: std::slice::Chunks<'a, RawMetric>,
+        chunk_size: usize,
+
+        // write to a BytesMut so that we can cheaply clone the frozen Bytes for retries
+        buffer: bytes::BytesMut,
+        // chunk amount of events are reused to produce the serialized document
+        scratch: Vec<Event<Ids, Name>>,
+        factory: F,
+    }
+
+    impl<'a, F: KeyGen<'a>> Iterator for Iter<'a, F> {
+        type Item = Result<(&'a [RawMetric], bytes::Bytes), serde_json::Error>;
+
+        fn next(&mut self) -> Option<Self::Item> {
+            let chunk = self.inner.next()?;
+
+            if self.scratch.is_empty() {
+                // first round: create events with N strings
+                self.scratch.extend(
+                    chunk
+                        .iter()
+                        .map(|raw_metric| raw_metric.as_event(&self.factory.generate())),
+                );
+            } else {
+                // next rounds: update_in_place to reuse allocations
+                assert_eq!(self.scratch.len(), self.chunk_size);
+                self.scratch
+                    .iter_mut()
+                    .zip(chunk.iter())
+                    .for_each(|(slot, raw_metric)| {
+                        raw_metric.update_in_place(slot, &self.factory.generate())
+                    });
+            }
+
+            let res = serde_json::to_writer(
+                (&mut self.buffer).writer(),
+                &EventChunk {
+                    events: (&self.scratch[..chunk.len()]).into(),
+                },
+            );
+
+            match res {
+                Ok(()) => Some(Ok((chunk, self.buffer.split().freeze()))),
+                Err(e) => Some(Err(e)),
+            }
+        }
+
+        fn size_hint(&self) -> (usize, Option<usize>) {
+            self.inner.size_hint()
+        }
+    }
+
+    impl<'a, F: KeyGen<'a>> ExactSizeIterator for Iter<'a, F> {}
+
+    let buffer = bytes::BytesMut::new();
+    let inner = input.chunks(chunk_size);
+    let scratch = Vec::new();
+
+    Iter {
+        inner,
+        chunk_size,
+        buffer,
+        scratch,
+        factory,
+    }
+}
+
+trait RawMetricExt {
+    fn as_event(&self, key: &IdempotencyKey<'_>) -> Event<Ids, Name>;
+    fn update_in_place(&self, event: &mut Event<Ids, Name>, key: &IdempotencyKey<'_>);
+}
+
+impl RawMetricExt for RawMetric {
+    fn as_event(&self, key: &IdempotencyKey<'_>) -> Event<Ids, Name> {
+        let MetricsKey {
+            metric,
+            tenant_id,
+            timeline_id,
+        } = self.0;
+
+        let (kind, value) = self.1;
+
+        Event {
+            kind,
+            metric,
+            idempotency_key: key.to_string(),
+            value,
+            extra: Ids {
+                tenant_id,
+                timeline_id,
+            },
+        }
+    }
+
+    fn update_in_place(&self, event: &mut Event<Ids, Name>, key: &IdempotencyKey<'_>) {
+        use std::fmt::Write;
+
+        let MetricsKey {
+            metric,
+            tenant_id,
+            timeline_id,
+        } = self.0;
+
+        let (kind, value) = self.1;
+
+        *event = Event {
+            kind,
+            metric,
+            idempotency_key: {
+                event.idempotency_key.clear();
+                write!(event.idempotency_key, "{key}").unwrap();
+                std::mem::take(&mut event.idempotency_key)
+            },
+            value,
+            extra: Ids {
+                tenant_id,
+                timeline_id,
+            },
+        };
+    }
+}
+
+trait KeyGen<'a>: Copy {
+    fn generate(&self) -> IdempotencyKey<'a>;
+}
+
+impl<'a> KeyGen<'a> for &'a str {
+    fn generate(&self) -> IdempotencyKey<'a> {
+        IdempotencyKey::generate(self)
+    }
+}
+
+enum UploadError {
+    Rejected(reqwest::StatusCode),
+    Reqwest(reqwest::Error),
+    Cancelled,
+}
+
+impl std::fmt::Debug for UploadError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        // use same impl because backoff::retry will log this using both
+        std::fmt::Display::fmt(self, f)
+    }
+}
+
+impl std::fmt::Display for UploadError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        use UploadError::*;
+
+        match self {
+            Rejected(code) => write!(f, "server rejected the metrics with {code}"),
+            Reqwest(e) => write!(f, "request failed: {e}"),
+            Cancelled => write!(f, "cancelled"),
+        }
+    }
+}
+
+impl UploadError {
+    fn is_reject(&self) -> bool {
+        matches!(self, UploadError::Rejected(_))
+    }
+}
+
+// this is consumed by the test verifiers
+static LAST_IN_BATCH: reqwest::header::HeaderName =
+    reqwest::header::HeaderName::from_static("pageserver-metrics-last-upload-in-batch");
+
+async fn upload(
+    client: &reqwest::Client,
+    metric_collection_endpoint: &reqwest::Url,
+    body: bytes::Bytes,
+    cancel: &CancellationToken,
+    is_last: bool,
+) -> Result<(), UploadError> {
+    let warn_after = 3;
+    let max_attempts = 10;
+    let res = utils::backoff::retry(
+        move || {
+            let body = body.clone();
+            async move {
+                let res = client
+                    .post(metric_collection_endpoint.clone())
+                    .header(reqwest::header::CONTENT_TYPE, "application/json")
+                    .header(
+                        LAST_IN_BATCH.clone(),
+                        if is_last { "true" } else { "false" },
+                    )
+                    .body(body)
+                    .send()
+                    .await;
+
+                let res = res.and_then(|res| res.error_for_status());
+
+                // 10 redirects are normally allowed, so we don't need worry about 3xx
+                match res {
+                    Ok(_response) => Ok(()),
+                    Err(e) => {
+                        let status = e.status().filter(|s| s.is_client_error());
+                        if let Some(status) = status {
+                            // rejection used to be a thing when the server could reject a
+                            // whole batch of metrics if one metric was bad.
+                            Err(UploadError::Rejected(status))
+                        } else {
+                            Err(UploadError::Reqwest(e))
+                        }
+                    }
+                }
+            }
+        },
+        UploadError::is_reject,
+        warn_after,
+        max_attempts,
+        "upload consumption_metrics",
+        utils::backoff::Cancel::new(cancel.clone(), || UploadError::Cancelled),
+    )
+    .await;
+
+    match &res {
+        Ok(_) => {}
+        Err(e) if e.is_reject() => {
+            // permanent errors currently do not get logged by backoff::retry
+            // display alternate has no effect, but keeping it here for easier pattern matching.
+            tracing::error!("failed to upload metrics: {e:#}");
+        }
+        Err(_) => {
+            // these have been logged already
+        }
+    }
+
+    res
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use chrono::{DateTime, Utc};
+    use once_cell::sync::Lazy;
+
+    #[test]
+    fn chunked_serialization() {
+        let examples = metric_samples();
+        assert!(examples.len() > 1);
+
+        let factory = FixedGen::new(Utc::now(), "1", 42);
+
+        // need to use Event here because serde_json::Value uses default hashmap, not linked
+        // hashmap
+        #[derive(serde::Deserialize)]
+        struct EventChunk {
+            events: Vec<Event<Ids, Name>>,
+        }
+
+        let correct = serialize_in_chunks(examples.len(), &examples, factory)
+            .map(|res| res.unwrap().1)
+            .flat_map(|body| serde_json::from_slice::<EventChunk>(&body).unwrap().events)
+            .collect::<Vec<_>>();
+
+        for chunk_size in 1..examples.len() {
+            let actual = serialize_in_chunks(chunk_size, &examples, factory)
+                .map(|res| res.unwrap().1)
+                .flat_map(|body| serde_json::from_slice::<EventChunk>(&body).unwrap().events)
+                .collect::<Vec<_>>();
+
+            // if these are equal, it means that multi-chunking version works as well
+            assert_eq!(correct, actual);
+        }
+    }
+
+    #[derive(Clone, Copy)]
+    struct FixedGen<'a>(chrono::DateTime<chrono::Utc>, &'a str, u16);
+
+    impl<'a> FixedGen<'a> {
+        fn new(now: chrono::DateTime<chrono::Utc>, node_id: &'a str, nonce: u16) -> Self {
+            FixedGen(now, node_id, nonce)
+        }
+    }
+
+    impl<'a> KeyGen<'a> for FixedGen<'a> {
+        fn generate(&self) -> IdempotencyKey<'a> {
+            IdempotencyKey::for_tests(self.0, self.1, self.2)
+        }
+    }
+
+    static SAMPLES_NOW: Lazy<DateTime<Utc>> = Lazy::new(|| {
+        DateTime::parse_from_rfc3339("2023-09-15T00:00:00.123456789Z")
+            .unwrap()
+            .into()
+    });
+
+    #[test]
+    fn metric_image_stability() {
+        // it is important that these strings stay as they are
+
+        let examples = [
+            (
+                line!(),
+                r#"{"type":"absolute","time":"2023-09-15T00:00:00.123456789Z","metric":"written_size","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":0,"tenant_id":"00000000000000000000000000000000","timeline_id":"ffffffffffffffffffffffffffffffff"}"#,
+            ),
+            (
+                line!(),
+                r#"{"type":"incremental","start_time":"2023-09-14T00:00:00.123456789Z","stop_time":"2023-09-15T00:00:00.123456789Z","metric":"written_data_bytes_delta","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":0,"tenant_id":"00000000000000000000000000000000","timeline_id":"ffffffffffffffffffffffffffffffff"}"#,
+            ),
+            (
+                line!(),
+                r#"{"type":"absolute","time":"2023-09-15T00:00:00.123456789Z","metric":"timeline_logical_size","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":0,"tenant_id":"00000000000000000000000000000000","timeline_id":"ffffffffffffffffffffffffffffffff"}"#,
+            ),
+            (
+                line!(),
+                r#"{"type":"absolute","time":"2023-09-15T00:00:00.123456789Z","metric":"remote_storage_size","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":0,"tenant_id":"00000000000000000000000000000000"}"#,
+            ),
+            (
+                line!(),
+                r#"{"type":"absolute","time":"2023-09-15T00:00:00.123456789Z","metric":"resident_size","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":0,"tenant_id":"00000000000000000000000000000000"}"#,
+            ),
+            (
+                line!(),
+                r#"{"type":"absolute","time":"2023-09-15T00:00:00.123456789Z","metric":"synthetic_storage_size","idempotency_key":"2023-09-15 00:00:00.123456789 UTC-1-0000","value":1,"tenant_id":"00000000000000000000000000000000"}"#,
+            ),
+        ];
+
+        let idempotency_key = consumption_metrics::IdempotencyKey::for_tests(*SAMPLES_NOW, "1", 0);
+        let examples = examples.into_iter().zip(metric_samples());
+
+        for ((line, expected), (key, (kind, value))) in examples {
+            let e = consumption_metrics::Event {
+                kind,
+                metric: key.metric,
+                idempotency_key: idempotency_key.to_string(),
+                value,
+                extra: Ids {
+                    tenant_id: key.tenant_id,
+                    timeline_id: key.timeline_id,
+                },
+            };
+            let actual = serde_json::to_string(&e).unwrap();
+            assert_eq!(expected, actual, "example for {kind:?} from line {line}");
+        }
+    }
+
+    fn metric_samples() -> [RawMetric; 6] {
+        let tenant_id = TenantId::from_array([0; 16]);
+        let timeline_id = TimelineId::from_array([0xff; 16]);
+
+        let before = DateTime::parse_from_rfc3339("2023-09-14T00:00:00.123456789Z")
+            .unwrap()
+            .into();
+        let [now, before] = [*SAMPLES_NOW, before];
+
+        super::super::metrics::metric_examples(tenant_id, timeline_id, now, before)
+    }
+}

pageserver/src/context.rs

@@ -94,6 +94,18 @@ pub struct RequestContext {
     task_kind: TaskKind,
     download_behavior: DownloadBehavior,
     access_stats_behavior: AccessStatsBehavior,
+    page_content_kind: PageContentKind,
+}
+
+/// The kind of access to the page cache.
+#[derive(Clone, Copy, PartialEq, Eq, Debug, enum_map::Enum, strum_macros::IntoStaticStr)]
+pub enum PageContentKind {
+    Unknown,
+    DeltaLayerBtreeNode,
+    DeltaLayerValue,
+    ImageLayerBtreeNode,
+    ImageLayerValue,
+    InMemoryLayer,
 }
 
 /// Desired behavior if the operation requires an on-demand download
@@ -137,6 +149,7 @@ impl RequestContextBuilder {
                 task_kind,
                 download_behavior: DownloadBehavior::Download,
                 access_stats_behavior: AccessStatsBehavior::Update,
+                page_content_kind: PageContentKind::Unknown,
             },
         }
     }
@@ -149,6 +162,7 @@ impl RequestContextBuilder {
                 task_kind: original.task_kind,
                 download_behavior: original.download_behavior,
                 access_stats_behavior: original.access_stats_behavior,
+                page_content_kind: original.page_content_kind,
             },
         }
     }
@@ -167,6 +181,11 @@ impl RequestContextBuilder {
         self
     }
 
+    pub(crate) fn page_content_kind(mut self, k: PageContentKind) -> Self {
+        self.inner.page_content_kind = k;
+        self
+    }
+
     pub fn build(self) -> RequestContext {
         self.inner
     }
@@ -263,4 +282,8 @@ impl RequestContext {
     pub(crate) fn access_stats_behavior(&self) -> AccessStatsBehavior {
         self.access_stats_behavior
     }
+
+    pub(crate) fn page_content_kind(&self) -> PageContentKind {
+        self.page_content_kind
+    }
 }

pageserver/src/control_plane_client.rs

@@ -1,7 +1,9 @@
 use std::collections::HashMap;
 
-use hyper::StatusCode;
-use pageserver_api::control_api::{ReAttachRequest, ReAttachResponse};
+use pageserver_api::control_api::{
+    ReAttachRequest, ReAttachResponse, ValidateRequest, ValidateRequestTenant, ValidateResponse,
+};
+use serde::{de::DeserializeOwned, Serialize};
 use tokio_util::sync::CancellationToken;
 use url::Url;
 use utils::{
@@ -12,25 +14,34 @@ use utils::{
 
 use crate::config::PageServerConf;
 
-// Backoffs when control plane requests do not succeed: compromise between reducing load
-// on control plane, and retrying frequently when we are blocked on a control plane
-// response to make progress.
-const BACKOFF_INCREMENT: f64 = 0.1;
-const BACKOFF_MAX: f64 = 10.0;
-
 /// The Pageserver's client for using the control plane API: this is a small subset
 /// of the overall control plane API, for dealing with generations (see docs/rfcs/025-generation-numbers.md)
-pub(crate) struct ControlPlaneClient {
+pub struct ControlPlaneClient {
     http_client: reqwest::Client,
     base_url: Url,
     node_id: NodeId,
     cancel: CancellationToken,
 }
 
+/// Represent operations which internally retry on all errors other than
+/// cancellation token firing: the only way they can fail is ShuttingDown.
+pub enum RetryForeverError {
+    ShuttingDown,
+}
+
+#[async_trait::async_trait]
+pub trait ControlPlaneGenerationsApi {
+    async fn re_attach(&self) -> Result<HashMap<TenantId, Generation>, RetryForeverError>;
+    async fn validate(
+        &self,
+        tenants: Vec<(TenantId, Generation)>,
+    ) -> Result<HashMap<TenantId, bool>, RetryForeverError>;
+}
+
 impl ControlPlaneClient {
     /// A None return value indicates that the input `conf` object does not have control
     /// plane API enabled.
-    pub(crate) fn new(conf: &'static PageServerConf, cancel: &CancellationToken) -> Option<Self> {
+    pub fn new(conf: &'static PageServerConf, cancel: &CancellationToken) -> Option<Self> {
         let mut url = match conf.control_plane_api.as_ref() {
             Some(u) => u.clone(),
             None => return None,
@@ -42,39 +53,78 @@ impl ControlPlaneClient {
             segs.pop_if_empty().push("");
         }
 
-        let client = reqwest::ClientBuilder::new()
-            .build()
-            .expect("Failed to construct http client");
+        let mut client = reqwest::ClientBuilder::new();
+
+        if let Some(jwt) = &conf.control_plane_api_token {
+            let mut headers = hyper::HeaderMap::new();
+            headers.insert("Authorization", jwt.get_contents().parse().unwrap());
+            client = client.default_headers(headers);
+        }
 
         Some(Self {
-            http_client: client,
+            http_client: client.build().expect("Failed to construct HTTP client"),
             base_url: url,
             node_id: conf.id,
             cancel: cancel.clone(),
         })
     }
 
-    async fn try_re_attach(
+    async fn retry_http_forever<R, T>(
         &self,
-        url: Url,
-        request: &ReAttachRequest,
-    ) -> anyhow::Result<ReAttachResponse> {
-        match self.http_client.post(url).json(request).send().await {
-            Err(e) => Err(anyhow::Error::from(e)),
-            Ok(r) => {
-                if r.status() == StatusCode::OK {
-                    r.json::<ReAttachResponse>()
-                        .await
-                        .map_err(anyhow::Error::from)
-                } else {
-                    Err(anyhow::anyhow!("Unexpected status {}", r.status()))
-                }
+        url: &url::Url,
+        request: R,
+    ) -> Result<T, RetryForeverError>
+    where
+        R: Serialize,
+        T: DeserializeOwned,
+    {
+        #[derive(thiserror::Error, Debug)]
+        enum RemoteAttemptError {
+            #[error("shutdown")]
+            Shutdown,
+            #[error("remote: {0}")]
+            Remote(reqwest::Error),
+        }
+
+        match backoff::retry(
+            || async {
+                let response = self
+                    .http_client
+                    .post(url.clone())
+                    .json(&request)
+                    .send()
+                    .await
+                    .map_err(RemoteAttemptError::Remote)?;
+
+                response
+                    .error_for_status_ref()
+                    .map_err(RemoteAttemptError::Remote)?;
+                response
+                    .json::<T>()
+                    .await
+                    .map_err(RemoteAttemptError::Remote)
+            },
+            |_| false,
+            3,
+            u32::MAX,
+            "calling control plane generation validation API",
+            backoff::Cancel::new(self.cancel.clone(), || RemoteAttemptError::Shutdown),
+        )
+        .await
+        {
+            Err(RemoteAttemptError::Shutdown) => Err(RetryForeverError::ShuttingDown),
+            Err(RemoteAttemptError::Remote(_)) => {
+                panic!("We retry forever, this should never be reached");
             }
+            Ok(r) => Ok(r),
         }
     }
+}
 
-    /// Block until we get a successful response
-    pub(crate) async fn re_attach(&self) -> anyhow::Result<HashMap<TenantId, Generation>> {
+#[async_trait::async_trait]
+impl ControlPlaneGenerationsApi for ControlPlaneClient {
+    /// Block until we get a successful response, or error out if we are shut down
+    async fn re_attach(&self) -> Result<HashMap<TenantId, Generation>, RetryForeverError> {
         let re_attach_path = self
             .base_url
             .join("re-attach")
@@ -83,37 +133,51 @@ impl ControlPlaneClient {
             node_id: self.node_id,
         };
 
-        let mut attempt = 0;
-        loop {
-            let result = self.try_re_attach(re_attach_path.clone(), &request).await;
-            match result {
-                Ok(res) => {
-                    tracing::info!(
-                        "Received re-attach response with {} tenants",
-                        res.tenants.len()
-                    );
+        fail::fail_point!("control-plane-client-re-attach");
 
-                    return Ok(res
-                        .tenants
-                        .into_iter()
-                        .map(|t| (t.id, Generation::new(t.generation)))
-                        .collect::<HashMap<_, _>>());
-                }
-                Err(e) => {
-                    tracing::error!("Error re-attaching tenants, retrying: {e:#}");
-                    backoff::exponential_backoff(
-                        attempt,
-                        BACKOFF_INCREMENT,
-                        BACKOFF_MAX,
-                        &self.cancel,
-                    )
-                    .await;
-                    if self.cancel.is_cancelled() {
-                        return Err(anyhow::anyhow!("Shutting down"));
-                    }
-                    attempt += 1;
-                }
-            }
-        }
+        let response: ReAttachResponse = self.retry_http_forever(&re_attach_path, request).await?;
+        tracing::info!(
+            "Received re-attach response with {} tenants",
+            response.tenants.len()
+        );
+
+        Ok(response
+            .tenants
+            .into_iter()
+            .map(|t| (t.id, Generation::new(t.generation)))
+            .collect::<HashMap<_, _>>())
+    }
+
+    /// Block until we get a successful response, or error out if we are shut down
+    async fn validate(
+        &self,
+        tenants: Vec<(TenantId, Generation)>,
+    ) -> Result<HashMap<TenantId, bool>, RetryForeverError> {
+        let re_attach_path = self
+            .base_url
+            .join("validate")
+            .expect("Failed to build validate path");
+
+        let request = ValidateRequest {
+            tenants: tenants
+                .into_iter()
+                .map(|(id, gen)| ValidateRequestTenant {
+                    id,
+                    gen: gen
+                        .into()
+                        .expect("Generation should always be valid for a Tenant doing deletions"),
+                })
+                .collect(),
+        };
+
+        fail::fail_point!("control-plane-client-validate");
+
+        let response: ValidateResponse = self.retry_http_forever(&re_attach_path, request).await?;
+
+        Ok(response
+            .tenants
+            .into_iter()
+            .map(|rt| (rt.id, rt.valid))
+            .collect())
     }
 }

pageserver/src/deletion_queue/deleter.rs

@@ -0,0 +1,172 @@
+//! The deleter is the final stage in the deletion queue.  It accumulates remote
+//! paths to delete, and periodically executes them in batches of up to 1000
+//! using the DeleteObjects request.
+//!
+//! Its purpose is to increase efficiency of remote storage I/O by issuing a smaller
+//! number of full-sized DeleteObjects requests, rather than a larger number of
+//! smaller requests.
+
+use remote_storage::GenericRemoteStorage;
+use remote_storage::RemotePath;
+use remote_storage::MAX_KEYS_PER_DELETE;
+use std::time::Duration;
+use tokio_util::sync::CancellationToken;
+use tracing::info;
+use tracing::warn;
+use utils::backoff;
+
+use crate::metrics;
+
+use super::DeletionQueueError;
+use super::FlushOp;
+
+const AUTOFLUSH_INTERVAL: Duration = Duration::from_secs(10);
+
+pub(super) enum DeleterMessage {
+    Delete(Vec<RemotePath>),
+    Flush(FlushOp),
+}
+
+/// Non-persistent deletion queue, for coalescing multiple object deletes into
+/// larger DeleteObjects requests.
+pub(super) struct Deleter {
+    // Accumulate up to 1000 keys for the next deletion operation
+    accumulator: Vec<RemotePath>,
+
+    rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
+
+    cancel: CancellationToken,
+    remote_storage: GenericRemoteStorage,
+}
+
+impl Deleter {
+    pub(super) fn new(
+        remote_storage: GenericRemoteStorage,
+        rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
+        cancel: CancellationToken,
+    ) -> Self {
+        Self {
+            remote_storage,
+            rx,
+            cancel,
+            accumulator: Vec::new(),
+        }
+    }
+
+    /// Wrap the remote `delete_objects` with a failpoint
+    async fn remote_delete(&self) -> Result<(), anyhow::Error> {
+        fail::fail_point!("deletion-queue-before-execute", |_| {
+            info!("Skipping execution, failpoint set");
+            metrics::DELETION_QUEUE
+                .remote_errors
+                .with_label_values(&["failpoint"])
+                .inc();
+            Err(anyhow::anyhow!("failpoint hit"))
+        });
+
+        // A backoff::retry is used here for two reasons:
+        // - To provide a backoff rather than busy-polling the API on errors
+        // - To absorb transient 429/503 conditions without hitting our error
+        //   logging path for issues deleting objects.
+        backoff::retry(
+            || async { self.remote_storage.delete_objects(&self.accumulator).await },
+            |_| false,
+            3,
+            10,
+            "executing deletion batch",
+            backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Shutting down")),
+        )
+        .await
+    }
+
+    /// Block until everything in accumulator has been executed
+    async fn flush(&mut self) -> Result<(), DeletionQueueError> {
+        while !self.accumulator.is_empty() && !self.cancel.is_cancelled() {
+            match self.remote_delete().await {
+                Ok(()) => {
+                    // Note: we assume that the remote storage layer returns Ok(()) if some
+                    // or all of the deleted objects were already gone.
+                    metrics::DELETION_QUEUE
+                        .keys_executed
+                        .inc_by(self.accumulator.len() as u64);
+                    info!(
+                        "Executed deletion batch {}..{}",
+                        self.accumulator
+                            .first()
+                            .expect("accumulator should be non-empty"),
+                        self.accumulator
+                            .last()
+                            .expect("accumulator should be non-empty"),
+                    );
+                    self.accumulator.clear();
+                }
+                Err(e) => {
+                    if self.cancel.is_cancelled() {
+                        return Err(DeletionQueueError::ShuttingDown);
+                    }
+                    warn!("DeleteObjects request failed: {e:#}, will continue trying");
+                    metrics::DELETION_QUEUE
+                        .remote_errors
+                        .with_label_values(&["execute"])
+                        .inc();
+                }
+            };
+        }
+        if self.cancel.is_cancelled() {
+            // Expose an error because we may not have actually flushed everything
+            Err(DeletionQueueError::ShuttingDown)
+        } else {
+            Ok(())
+        }
+    }
+
+    pub(super) async fn background(&mut self) -> Result<(), DeletionQueueError> {
+        self.accumulator.reserve(MAX_KEYS_PER_DELETE);
+
+        loop {
+            if self.cancel.is_cancelled() {
+                return Err(DeletionQueueError::ShuttingDown);
+            }
+
+            let msg = match tokio::time::timeout(AUTOFLUSH_INTERVAL, self.rx.recv()).await {
+                Ok(Some(m)) => m,
+                Ok(None) => {
+                    // All queue senders closed
+                    info!("Shutting down");
+                    return Err(DeletionQueueError::ShuttingDown);
+                }
+                Err(_) => {
+                    // Timeout, we hit deadline to execute whatever we have in hand.  These functions will
+                    // return immediately if no work is pending
+                    self.flush().await?;
+
+                    continue;
+                }
+            };
+
+            match msg {
+                DeleterMessage::Delete(mut list) => {
+                    while !list.is_empty() || self.accumulator.len() == MAX_KEYS_PER_DELETE {
+                        if self.accumulator.len() == MAX_KEYS_PER_DELETE {
+                            self.flush().await?;
+                            // If we have received this number of keys, proceed with attempting to execute
+                            assert_eq!(self.accumulator.len(), 0);
+                        }
+
+                        let available_slots = MAX_KEYS_PER_DELETE - self.accumulator.len();
+                        let take_count = std::cmp::min(available_slots, list.len());
+                        for path in list.drain(list.len() - take_count..) {
+                            self.accumulator.push(path);
+                        }
+                    }
+                }
+                DeleterMessage::Flush(flush_op) => {
+                    // If flush() errors, we drop the flush_op and the caller will get
+                    // an error recv()'ing their oneshot channel.
+                    self.flush().await?;
+                    flush_op.notify();
+                }
+            }
+        }
+    }
+}

pageserver/src/deletion_queue/list_writer.rs

@@ -0,0 +1,479 @@
+//! The list writer is the first stage in the deletion queue.  It accumulates
+//! layers to delete, and periodically writes out these layers into a persistent
+//! DeletionList.
+//!
+//! The purpose of writing DeletionLists is to decouple the decision to
+//! delete an object from the validation required to execute it: even if
+//! validation is not possible, e.g. due to a control plane outage, we can
+//! still persist our intent to delete an object, in a way that would
+//! survive a restart.
+//!
+//! DeletionLists are passed onwards to the Validator.
+
+use super::DeletionHeader;
+use super::DeletionList;
+use super::FlushOp;
+use super::ValidatorQueueMessage;
+
+use std::collections::HashMap;
+use std::fs::create_dir_all;
+use std::time::Duration;
+
+use regex::Regex;
+use remote_storage::RemotePath;
+use tokio_util::sync::CancellationToken;
+use tracing::debug;
+use tracing::info;
+use tracing::warn;
+use utils::generation::Generation;
+use utils::id::TenantId;
+use utils::id::TimelineId;
+
+use crate::config::PageServerConf;
+use crate::deletion_queue::TEMP_SUFFIX;
+use crate::metrics;
+use crate::tenant::remote_timeline_client::remote_layer_path;
+use crate::tenant::storage_layer::LayerFileName;
+
+// The number of keys in a DeletionList before we will proactively persist it
+// (without reaching a flush deadline).  This aims to deliver objects of the order
+// of magnitude 1MB when we are under heavy delete load.
+const DELETION_LIST_TARGET_SIZE: usize = 16384;
+
+// Ordinarily, we only flush to DeletionList periodically, to bound the window during
+// which we might leak objects from not flushing a DeletionList after
+// the objects are already unlinked from timeline metadata.
+const FRONTEND_DEFAULT_TIMEOUT: Duration = Duration::from_millis(10000);
+
+// If someone is waiting for a flush to DeletionList, only delay a little to accumulate
+// more objects before doing the flush.
+const FRONTEND_FLUSHING_TIMEOUT: Duration = Duration::from_millis(100);
+
+#[derive(Debug)]
+pub(super) struct DeletionOp {
+    pub(super) tenant_id: TenantId,
+    pub(super) timeline_id: TimelineId,
+    // `layers` and `objects` are both just lists of objects.  `layers` is used if you do not
+    // have a config object handy to project it to a remote key, and need the consuming worker
+    // to do it for you.
+    pub(super) layers: Vec<(LayerFileName, Generation)>,
+    pub(super) objects: Vec<RemotePath>,
+
+    /// The _current_ generation of the Tenant attachment in which we are enqueuing
+    /// this deletion.
+    pub(super) generation: Generation,
+}
+
+#[derive(Debug)]
+pub(super) struct RecoverOp {
+    pub(super) attached_tenants: HashMap<TenantId, Generation>,
+}
+
+#[derive(Debug)]
+pub(super) enum ListWriterQueueMessage {
+    Delete(DeletionOp),
+    // Wait until all prior deletions make it into a persistent DeletionList
+    Flush(FlushOp),
+    // Wait until all prior deletions have been executed (i.e. objects are actually deleted)
+    FlushExecute(FlushOp),
+    // Call once after re-attaching to control plane, to notify the deletion queue about
+    // latest attached generations & load any saved deletion lists from disk.
+    Recover(RecoverOp),
+}
+
+pub(super) struct ListWriter {
+    conf: &'static PageServerConf,
+
+    // Incoming frontend requests to delete some keys
+    rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
+
+    // Outbound requests to the backend to execute deletion lists we have composed.
+    tx: tokio::sync::mpsc::Sender<ValidatorQueueMessage>,
+
+    // The list we are currently building, contains a buffer of keys to delete
+    // and our next sequence number
+    pending: DeletionList,
+
+    // These FlushOps should notify the next time we flush
+    pending_flushes: Vec<FlushOp>,
+
+    // Worker loop is torn down when this fires.
+    cancel: CancellationToken,
+
+    // Safety guard to do recovery exactly once
+    recovered: bool,
+}
+
+impl ListWriter {
+    // Initially DeletionHeader.validated_sequence is zero.  The place we start our
+    // sequence numbers must be higher than that.
+    const BASE_SEQUENCE: u64 = 1;
+
+    pub(super) fn new(
+        conf: &'static PageServerConf,
+        rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
+        tx: tokio::sync::mpsc::Sender<ValidatorQueueMessage>,
+        cancel: CancellationToken,
+    ) -> Self {
+        Self {
+            pending: DeletionList::new(Self::BASE_SEQUENCE),
+            conf,
+            rx,
+            tx,
+            pending_flushes: Vec::new(),
+            cancel,
+            recovered: false,
+        }
+    }
+
+    /// Try to flush `list` to persistent storage
+    ///
+    /// This does not return errors, because on failure to flush we do not lose
+    /// any state: flushing will be retried implicitly on the next deadline
+    async fn flush(&mut self) {
+        if self.pending.is_empty() {
+            for f in self.pending_flushes.drain(..) {
+                f.notify();
+            }
+            return;
+        }
+
+        match self.pending.save(self.conf).await {
+            Ok(_) => {
+                info!(sequence = self.pending.sequence, "Stored deletion list");
+
+                for f in self.pending_flushes.drain(..) {
+                    f.notify();
+                }
+
+                // Take the list we've accumulated, replace it with a fresh list for the next sequence
+                let next_list = DeletionList::new(self.pending.sequence + 1);
+                let list = std::mem::replace(&mut self.pending, next_list);
+
+                if let Err(e) = self.tx.send(ValidatorQueueMessage::Delete(list)).await {
+                    // This is allowed to fail: it will only happen if the backend worker is shut down,
+                    // so we can just drop this on the floor.
+                    info!("Deletion list dropped, this is normal during shutdown ({e:#})");
+                }
+            }
+            Err(e) => {
+                metrics::DELETION_QUEUE.unexpected_errors.inc();
+                warn!(
+                    sequence = self.pending.sequence,
+                    "Failed to write deletion list, will retry later ({e:#})"
+                );
+            }
+        }
+    }
+
+    /// Load the header, to learn the sequence number up to which deletions
+    /// have been validated.  We will apply validated=true to DeletionLists
+    /// <= this sequence when loading them.
+    ///
+    /// It is not an error for the header to not exist: we return None, and
+    /// the caller should act as if validated_sequence is 0
+    async fn load_validated_sequence(&self) -> Result<Option<u64>, anyhow::Error> {
+        let header_path = self.conf.deletion_header_path();
+        match tokio::fs::read(&header_path).await {
+            Ok(header_bytes) => {
+                match serde_json::from_slice::<DeletionHeader>(&header_bytes) {
+                    Ok(h) => Ok(Some(h.validated_sequence)),
+                    Err(e) => {
+                        warn!(
+                            "Failed to deserialize deletion header, ignoring {header_path}: {e:#}",
+                        );
+                        // This should never happen unless we make a mistake with our serialization.
+                        // Ignoring a deletion header is not consequential for correctnes because all deletions
+                        // are ultimately allowed to fail: worst case we leak some objects for the scrubber to clean up.
+                        metrics::DELETION_QUEUE.unexpected_errors.inc();
+                        Ok(None)
+                    }
+                }
+            }
+            Err(e) => {
+                if e.kind() == std::io::ErrorKind::NotFound {
+                    debug!("Deletion header {header_path} not found, first start?");
+                    Ok(None)
+                } else {
+                    Err(anyhow::anyhow!(e))
+                }
+            }
+        }
+    }
+
+    async fn recover(
+        &mut self,
+        attached_tenants: HashMap<TenantId, Generation>,
+    ) -> Result<(), anyhow::Error> {
+        debug!(
+            "recovering with {} attached tenants",
+            attached_tenants.len()
+        );
+
+        // Load the header
+        let validated_sequence = self.load_validated_sequence().await?.unwrap_or(0);
+
+        self.pending.sequence = validated_sequence + 1;
+
+        let deletion_directory = self.conf.deletion_prefix();
+        let mut dir = match tokio::fs::read_dir(&deletion_directory).await {
+            Ok(d) => d,
+            Err(e) => {
+                warn!("Failed to open deletion list directory {deletion_directory}: {e:#}");
+
+                // Give up: if we can't read the deletion list directory, we probably can't
+                // write lists into it later, so the queue won't work.
+                return Err(e.into());
+            }
+        };
+
+        let list_name_pattern =
+            Regex::new("(?<sequence>[a-zA-Z0-9]{16})-(?<version>[a-zA-Z0-9]{2}).list").unwrap();
+
+        let temp_extension = format!(".{TEMP_SUFFIX}");
+        let header_path = self.conf.deletion_header_path();
+        let mut seqs: Vec<u64> = Vec::new();
+        while let Some(dentry) = dir.next_entry().await? {
+            let file_name = dentry.file_name();
+            let dentry_str = file_name.to_string_lossy();
+
+            if file_name == header_path.file_name().unwrap_or("") {
+                // Don't try and parse the header's name like a list
+                continue;
+            }
+
+            if dentry_str.ends_with(&temp_extension) {
+                info!("Cleaning up temporary file {dentry_str}");
+                let absolute_path =
+                    deletion_directory.join(dentry.file_name().to_str().expect("non-Unicode path"));
+                if let Err(e) = tokio::fs::remove_file(&absolute_path).await {
+                    // Non-fatal error: we will just leave the file behind but not
+                    // try and load it.
+                    warn!("Failed to clean up temporary file {absolute_path}: {e:#}");
+                }
+
+                continue;
+            }
+
+            let file_name = dentry.file_name().to_owned();
+            let basename = file_name.to_string_lossy();
+            let seq_part = if let Some(m) = list_name_pattern.captures(&basename) {
+                m.name("sequence")
+                    .expect("Non optional group should be present")
+                    .as_str()
+            } else {
+                warn!("Unexpected key in deletion queue: {basename}");
+                metrics::DELETION_QUEUE.unexpected_errors.inc();
+                continue;
+            };
+
+            let seq: u64 = match u64::from_str_radix(seq_part, 16) {
+                Ok(s) => s,
+                Err(e) => {
+                    warn!("Malformed key '{basename}': {e}");
+                    metrics::DELETION_QUEUE.unexpected_errors.inc();
+                    continue;
+                }
+            };
+            seqs.push(seq);
+        }
+        seqs.sort();
+
+        // Start our next deletion list from after the last location validated by
+        // previous process lifetime, or after the last location found (it is updated
+        // below after enumerating the deletion lists)
+        self.pending.sequence = validated_sequence + 1;
+        if let Some(max_list_seq) = seqs.last() {
+            self.pending.sequence = std::cmp::max(self.pending.sequence, max_list_seq + 1);
+        }
+
+        for s in seqs {
+            let list_path = self.conf.deletion_list_path(s);
+
+            let list_bytes = tokio::fs::read(&list_path).await?;
+
+            let mut deletion_list = match serde_json::from_slice::<DeletionList>(&list_bytes) {
+                Ok(l) => l,
+                Err(e) => {
+                    // Drop the list on the floor: any objects it referenced will be left behind
+                    // for scrubbing to clean up.  This should never happen unless we have a serialization bug.
+                    warn!(sequence = s, "Failed to deserialize deletion list: {e}");
+                    metrics::DELETION_QUEUE.unexpected_errors.inc();
+                    continue;
+                }
+            };
+
+            if deletion_list.sequence <= validated_sequence {
+                // If the deletion list falls below valid_seq, we may assume that it was
+                // already validated the last time this pageserver ran.  Otherwise, we still
+                // load it, as it may still contain content valid in this generation.
+                deletion_list.validated = true;
+            } else {
+                // Special case optimization: if a tenant is still attached, and no other
+                // generation was issued to another node in the interval while we restarted,
+                // then we may treat deletion lists from the previous generation as if they
+                // belong to our currently attached generation, and proceed to validate & execute.
+                for (tenant_id, tenant_list) in &mut deletion_list.tenants {
+                    if let Some(attached_gen) = attached_tenants.get(tenant_id) {
+                        if attached_gen.previous() == tenant_list.generation {
+                            tenant_list.generation = *attached_gen;
+                        }
+                    }
+                }
+            }
+
+            info!(
+                validated = deletion_list.validated,
+                sequence = deletion_list.sequence,
+                "Recovered deletion list"
+            );
+
+            // We will drop out of recovery if this fails: it indicates that we are shutting down
+            // or the backend has panicked
+            metrics::DELETION_QUEUE
+                .keys_submitted
+                .inc_by(deletion_list.len() as u64);
+            self.tx
+                .send(ValidatorQueueMessage::Delete(deletion_list))
+                .await?;
+        }
+
+        info!(next_sequence = self.pending.sequence, "Replay complete");
+
+        Ok(())
+    }
+
+    /// This is the front-end ingest, where we bundle up deletion requests into DeletionList
+    /// and write them out, for later validation by the backend and execution by the executor.
+    pub(super) async fn background(&mut self) {
+        info!("Started deletion frontend worker");
+
+        // Synchronous, but we only do it once per process lifetime so it's tolerable
+        if let Err(e) = create_dir_all(&self.conf.deletion_prefix()) {
+            tracing::error!(
+                "Failed to create deletion list directory {}, deletions will not be executed ({e})",
+                self.conf.deletion_prefix(),
+            );
+            metrics::DELETION_QUEUE.unexpected_errors.inc();
+            return;
+        }
+
+        while !self.cancel.is_cancelled() {
+            let timeout = if self.pending_flushes.is_empty() {
+                FRONTEND_DEFAULT_TIMEOUT
+            } else {
+                FRONTEND_FLUSHING_TIMEOUT
+            };
+
+            let msg = match tokio::time::timeout(timeout, self.rx.recv()).await {
+                Ok(Some(msg)) => msg,
+                Ok(None) => {
+                    // Queue sender destroyed, shutting down
+                    break;
+                }
+                Err(_) => {
+                    // Hit deadline, flush.
+                    self.flush().await;
+                    continue;
+                }
+            };
+
+            match msg {
+                ListWriterQueueMessage::Delete(op) => {
+                    assert!(
+                        self.recovered,
+                        "Cannot process deletions before recovery.  This is a bug."
+                    );
+
+                    debug!(
+                        "Delete: ingesting {} layers, {} other objects",
+                        op.layers.len(),
+                        op.objects.len()
+                    );
+
+                    let mut layer_paths = Vec::new();
+                    for (layer, generation) in op.layers {
+                        layer_paths.push(remote_layer_path(
+                            &op.tenant_id,
+                            &op.timeline_id,
+                            &layer,
+                            generation,
+                        ));
+                    }
+                    layer_paths.extend(op.objects);
+
+                    if !self.pending.push(
+                        &op.tenant_id,
+                        &op.timeline_id,
+                        op.generation,
+                        &mut layer_paths,
+                    ) {
+                        self.flush().await;
+                        let retry_succeeded = self.pending.push(
+                            &op.tenant_id,
+                            &op.timeline_id,
+                            op.generation,
+                            &mut layer_paths,
+                        );
+                        if !retry_succeeded {
+                            // Unexpected: after we flush, we should have
+                            // drained self.pending, so a conflict on
+                            // generation numbers should be impossible.
+                            tracing::error!(
+                                "Failed to enqueue deletions, leaking objects.  This is a bug."
+                            );
+                            metrics::DELETION_QUEUE.unexpected_errors.inc();
+                        }
+                    }
+                }
+                ListWriterQueueMessage::Flush(op) => {
+                    if self.pending.is_empty() {
+                        // Execute immediately
+                        debug!("Flush: No pending objects, flushing immediately");
+                        op.notify()
+                    } else {
+                        // Execute next time we flush
+                        debug!("Flush: adding to pending flush list for next deadline flush");
+                        self.pending_flushes.push(op);
+                    }
+                }
+                ListWriterQueueMessage::FlushExecute(op) => {
+                    debug!("FlushExecute: passing through to backend");
+                    // We do not flush to a deletion list here: the client sends a Flush before the FlushExecute
+                    if let Err(e) = self.tx.send(ValidatorQueueMessage::Flush(op)).await {
+                        info!("Can't flush, shutting down ({e})");
+                        // Caller will get error when their oneshot sender was dropped.
+                    }
+                }
+                ListWriterQueueMessage::Recover(op) => {
+                    if self.recovered {
+                        tracing::error!(
+                            "Deletion queue recovery called more than once.  This is a bug."
+                        );
+                        metrics::DELETION_QUEUE.unexpected_errors.inc();
+                        // Non-fatal: although this is a bug, since we did recovery at least once we may proceed.
+                        continue;
+                    }
+
+                    if let Err(e) = self.recover(op.attached_tenants).await {
+                        // This should only happen in truly unrecoverable cases, like the recovery finding that the backend
+                        // queue receiver has been dropped, or something is critically broken with
+                        // the local filesystem holding deletion lists.
+                        info!(
+                            "Deletion queue recover aborted, deletion queue will not proceed ({e})"
+                        );
+                        metrics::DELETION_QUEUE.unexpected_errors.inc();
+                        return;
+                    } else {
+                        self.recovered = true;
+                    }
+                }
+            }
+
+            if self.pending.len() > DELETION_LIST_TARGET_SIZE || !self.pending_flushes.is_empty() {
+                self.flush().await;
+            }
+        }
+        info!("Deletion queue shut down.");
+    }
+}

pageserver/src/deletion_queue/validator.rs

@@ -0,0 +1,416 @@
+//! The validator is responsible for validating DeletionLists for execution,
+//! based on whethe the generation in the DeletionList is still the latest
+//! generation for a tenant.
+//!
+//! The purpose of validation is to ensure split-brain safety in the cluster
+//! of pageservers: a deletion may only be executed if the tenant generation
+//! that originated it is still current.  See docs/rfcs/025-generation-numbers.md
+//! The purpose of accumulating lists before validating them is to reduce load
+//! on the control plane API by issuing fewer, larger requests.
+//!
+//! In addition to validating DeletionLists, the validator validates updates to remote_consistent_lsn
+//! for timelines: these are logically deletions because the safekeepers use remote_consistent_lsn
+//! to decide when old
+//!
+//! Deletions are passed onward to the Deleter.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+
+use camino::Utf8PathBuf;
+use tokio_util::sync::CancellationToken;
+use tracing::debug;
+use tracing::info;
+use tracing::warn;
+
+use crate::config::PageServerConf;
+use crate::control_plane_client::ControlPlaneGenerationsApi;
+use crate::control_plane_client::RetryForeverError;
+use crate::metrics;
+
+use super::deleter::DeleterMessage;
+use super::DeletionHeader;
+use super::DeletionList;
+use super::DeletionQueueError;
+use super::FlushOp;
+use super::VisibleLsnUpdates;
+
+// After this length of time, do any validation work that is pending,
+// even if we haven't accumulated many keys to delete.
+//
+// This also causes updates to remote_consistent_lsn to be validated, even
+// if there were no deletions enqueued.
+const AUTOFLUSH_INTERVAL: Duration = Duration::from_secs(10);
+
+// If we have received this number of keys, proceed with attempting to execute
+const AUTOFLUSH_KEY_COUNT: usize = 16384;
+
+#[derive(Debug)]
+pub(super) enum ValidatorQueueMessage {
+    Delete(DeletionList),
+    Flush(FlushOp),
+}
+pub(super) struct Validator<C>
+where
+    C: ControlPlaneGenerationsApi,
+{
+    conf: &'static PageServerConf,
+    rx: tokio::sync::mpsc::Receiver<ValidatorQueueMessage>,
+    tx: tokio::sync::mpsc::Sender<DeleterMessage>,
+
+    // Client for calling into control plane API for validation of deletes
+    control_plane_client: Option<C>,
+
+    // DeletionLists which are waiting generation validation.  Not safe to
+    // execute until [`validate`] has processed them.
+    pending_lists: Vec<DeletionList>,
+
+    // DeletionLists which have passed validation and are ready to execute.
+    validated_lists: Vec<DeletionList>,
+
+    // Sum of all the lengths of lists in pending_lists
+    pending_key_count: usize,
+
+    // Lsn validation state: we read projected LSNs and write back visible LSNs
+    // after validation.  This is the LSN equivalent of `pending_validation_lists`:
+    // it is drained in [`validate`]
+    lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
+
+    // If we failed to rewrite a deletion list due to local filesystem I/O failure,
+    // we must remember that and refuse to advance our persistent validated sequence
+    // number past the failure.
+    list_write_failed: Option<u64>,
+
+    cancel: CancellationToken,
+}
+
+impl<C> Validator<C>
+where
+    C: ControlPlaneGenerationsApi,
+{
+    pub(super) fn new(
+        conf: &'static PageServerConf,
+        rx: tokio::sync::mpsc::Receiver<ValidatorQueueMessage>,
+        tx: tokio::sync::mpsc::Sender<DeleterMessage>,
+        control_plane_client: Option<C>,
+        lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
+        cancel: CancellationToken,
+    ) -> Self {
+        Self {
+            conf,
+            rx,
+            tx,
+            control_plane_client,
+            lsn_table,
+            pending_lists: Vec::new(),
+            validated_lists: Vec::new(),
+            pending_key_count: 0,
+            list_write_failed: None,
+            cancel,
+        }
+    }
+    /// Process any outstanding validations of generations of pending LSN updates or pending
+    /// DeletionLists.
+    ///
+    /// Valid LSN updates propagate back to Timelines immediately, valid DeletionLists
+    /// go into the queue of ready-to-execute lists.
+    async fn validate(&mut self) -> Result<(), DeletionQueueError> {
+        let mut tenant_generations = HashMap::new();
+        for list in &self.pending_lists {
+            for (tenant_id, tenant_list) in &list.tenants {
+                // Note: DeletionLists are in logical time order, so generation always
+                // goes up.  By doing a simple insert() we will always end up with
+                // the latest generation seen for a tenant.
+                tenant_generations.insert(*tenant_id, tenant_list.generation);
+            }
+        }
+
+        let pending_lsn_updates = {
+            let mut lsn_table = self.lsn_table.write().expect("Lock should not be poisoned");
+            std::mem::take(&mut *lsn_table)
+        };
+        for (tenant_id, update) in &pending_lsn_updates.tenants {
+            let entry = tenant_generations
+                .entry(*tenant_id)
+                .or_insert(update.generation);
+            if update.generation > *entry {
+                *entry = update.generation;
+            }
+        }
+
+        if tenant_generations.is_empty() {
+            // No work to do
+            return Ok(());
+        }
+
+        let tenants_valid = if let Some(control_plane_client) = &self.control_plane_client {
+            match control_plane_client
+                .validate(tenant_generations.iter().map(|(k, v)| (*k, *v)).collect())
+                .await
+            {
+                Ok(tenants) => tenants,
+                Err(RetryForeverError::ShuttingDown) => {
+                    // The only way a validation call returns an error is when the cancellation token fires
+                    return Err(DeletionQueueError::ShuttingDown);
+                }
+            }
+        } else {
+            // Control plane API disabled.  In legacy mode we consider everything valid.
+            tenant_generations.keys().map(|k| (*k, true)).collect()
+        };
+
+        let mut validated_sequence: Option<u64> = None;
+
+        // Apply the validation results to the pending LSN updates
+        for (tenant_id, tenant_lsn_state) in pending_lsn_updates.tenants {
+            let validated_generation = tenant_generations
+                .get(&tenant_id)
+                .expect("Map was built from the same keys we're reading");
+
+            let valid = tenants_valid
+                .get(&tenant_id)
+                .copied()
+                // If the tenant was missing from the validation response, it has been deleted.
+                // The Timeline that requested the LSN update is probably already torn down,
+                // or will be torn down soon.  In this case, drop the update by setting valid=false.
+                .unwrap_or(false);
+
+            if valid && *validated_generation == tenant_lsn_state.generation {
+                for (_timeline_id, pending_lsn) in tenant_lsn_state.timelines {
+                    pending_lsn.result_slot.store(pending_lsn.projected);
+                }
+            } else {
+                // If we failed validation, then do not apply any of the projected updates
+                warn!("Dropped remote consistent LSN updates for tenant {tenant_id} in stale generation {:?}", tenant_lsn_state.generation);
+                metrics::DELETION_QUEUE.dropped_lsn_updates.inc();
+            }
+        }
+
+        // Apply the validation results to the pending deletion lists
+        for list in &mut self.pending_lists {
+            // Filter the list based on whether the server responded valid: true.
+            // If a tenant is omitted in the response, it has been deleted, and we should
+            // proceed with deletion.
+            let mut mutated = false;
+            list.tenants.retain(|tenant_id, tenant| {
+                let validated_generation = tenant_generations
+                    .get(tenant_id)
+                    .expect("Map was built from the same keys we're reading");
+
+                // If the tenant was missing from the validation response, it has been deleted.
+                // This means that a deletion is valid, but also redundant since the tenant's
+                // objects should have already been deleted.  Treat it as invalid to drop the
+                // redundant deletion.
+                let valid = tenants_valid.get(tenant_id).copied().unwrap_or(false);
+
+                // A list is valid if it comes from the current _or previous_ generation.
+                // - The previous generation case is permitted due to how we store deletion lists locally:
+                // if we see the immediately previous generation in a locally stored deletion list,
+                // it proves that this node's disk was used for both current & previous generations,
+                // and therefore no other node was involved in between: the two generations may be
+                // logically treated as the same.
+                // - In that previous generation case, we rewrote it to the current generation
+                // in recover(), so the comparison here is simply an equality.
+
+                let this_list_valid = valid
+                    && (tenant.generation == *validated_generation);
+
+                if !this_list_valid {
+                    warn!("Dropping stale deletions for tenant {tenant_id} in generation {:?}, objects may be leaked", tenant.generation);
+                    metrics::DELETION_QUEUE.keys_dropped.inc_by(tenant.len() as u64);
+                    mutated = true;
+                } else {
+                    metrics::DELETION_QUEUE.keys_validated.inc_by(tenant.len() as u64);
+                }
+                this_list_valid
+            });
+            list.validated = true;
+
+            if mutated {
+                // Save the deletion list if we had to make changes due to stale generations.  The
+                // saved list is valid for execution.
+                if let Err(e) = list.save(self.conf).await {
+                    // Highly unexpected.  Could happen if e.g. disk full.
+                    // If we didn't save the trimmed list, it is _not_ valid to execute.
+                    warn!("Failed to save modified deletion list {list}: {e:#}");
+                    metrics::DELETION_QUEUE.unexpected_errors.inc();
+
+                    // Rather than have a complex retry process, just drop it and leak the objects,
+                    // scrubber will clean up eventually.
+                    list.tenants.clear(); // Result is a valid-but-empty list, which is a no-op for execution.
+
+                    // We must remember this failure, to prevent later writing out a header that
+                    // would imply the unwritable list was valid on disk.
+                    if self.list_write_failed.is_none() {
+                        self.list_write_failed = Some(list.sequence);
+                    }
+                }
+            }
+
+            validated_sequence = Some(list.sequence);
+        }
+
+        if let Some(validated_sequence) = validated_sequence {
+            if let Some(list_write_failed) = self.list_write_failed {
+                // Rare error case: we failed to write out a deletion list to excise invalid
+                // entries, so we cannot advance the header's valid sequence number past that point.
+                //
+                // In this state we will continue to validate, execute and delete deletion lists,
+                // we just cannot update the header.  It should be noticed and fixed by a human due to
+                // the nonzero value of our unexpected_errors metric.
+                warn!(
+                    sequence_number = list_write_failed,
+                    "Cannot write header because writing a deletion list failed earlier",
+                );
+            } else {
+                // Write the queue header to record how far validation progressed.  This avoids having
+                // to rewrite each DeletionList to set validated=true in it.
+                let header = DeletionHeader::new(validated_sequence);
+
+                // Drop result because the validated_sequence is an optimization.  If we fail to save it,
+                // then restart, we will drop some deletion lists, creating work for scrubber.
+                // The save() function logs a warning on error.
+                if let Err(e) = header.save(self.conf).await {
+                    warn!("Failed to write deletion queue header: {e:#}");
+                    metrics::DELETION_QUEUE.unexpected_errors.inc();
+                }
+            }
+        }
+
+        // Transfer the validated lists to the validated queue, for eventual execution
+        self.validated_lists.append(&mut self.pending_lists);
+
+        Ok(())
+    }
+
+    async fn cleanup_lists(&mut self, list_paths: Vec<Utf8PathBuf>) {
+        for list_path in list_paths {
+            debug!("Removing deletion list {list_path}");
+
+            if let Err(e) = tokio::fs::remove_file(&list_path).await {
+                // Unexpected: we should have permissions and nothing else should
+                // be touching these files.  We will leave the file behind.  Subsequent
+                // pageservers will try and load it again: hopefully whatever storage
+                // issue (probably permissions) has been fixed by then.
+                tracing::error!("Failed to delete {list_path}: {e:#}");
+                metrics::DELETION_QUEUE.unexpected_errors.inc();
+                break;
+            }
+        }
+    }
+
+    async fn flush(&mut self) -> Result<(), DeletionQueueError> {
+        tracing::debug!("Flushing with {} pending lists", self.pending_lists.len());
+
+        // Issue any required generation validation calls to the control plane
+        self.validate().await?;
+
+        // After successful validation, nothing is pending: any lists that
+        // made it through validation will be in validated_lists.
+        assert!(self.pending_lists.is_empty());
+        self.pending_key_count = 0;
+
+        tracing::debug!(
+            "Validation complete, have {} validated lists",
+            self.validated_lists.len()
+        );
+
+        // Return quickly if we have no validated lists to execute.  This avoids flushing the
+        // executor when an idle backend hits its autoflush interval
+        if self.validated_lists.is_empty() {
+            return Ok(());
+        }
+
+        // Drain `validated_lists` into the executor
+        let mut executing_lists = Vec::new();
+        for list in self.validated_lists.drain(..) {
+            let list_path = self.conf.deletion_list_path(list.sequence);
+            let objects = list.into_remote_paths();
+            self.tx
+                .send(DeleterMessage::Delete(objects))
+                .await
+                .map_err(|_| DeletionQueueError::ShuttingDown)?;
+            executing_lists.push(list_path);
+        }
+
+        self.flush_executor().await?;
+
+        // Erase the deletion lists whose keys have all be deleted from remote storage
+        self.cleanup_lists(executing_lists).await;
+
+        Ok(())
+    }
+
+    async fn flush_executor(&mut self) -> Result<(), DeletionQueueError> {
+        // Flush the executor, so that all the keys referenced by these deletion lists
+        // are actually removed from remote storage.  This is a precondition to deleting
+        // the deletion lists themselves.
+        let (flush_op, rx) = FlushOp::new();
+        self.tx
+            .send(DeleterMessage::Flush(flush_op))
+            .await
+            .map_err(|_| DeletionQueueError::ShuttingDown)?;
+
+        rx.await.map_err(|_| DeletionQueueError::ShuttingDown)
+    }
+
+    pub(super) async fn background(&mut self) {
+        tracing::info!("Started deletion backend worker");
+
+        while !self.cancel.is_cancelled() {
+            let msg = match tokio::time::timeout(AUTOFLUSH_INTERVAL, self.rx.recv()).await {
+                Ok(Some(m)) => m,
+                Ok(None) => {
+                    // All queue senders closed
+                    info!("Shutting down");
+                    break;
+                }
+                Err(_) => {
+                    // Timeout, we hit deadline to execute whatever we have in hand.  These functions will
+                    // return immediately if no work is pending.
+                    match self.flush().await {
+                        Ok(()) => {}
+                        Err(DeletionQueueError::ShuttingDown) => {
+                            // If we are shutting down, then auto-flush can safely be skipped
+                        }
+                    }
+
+                    continue;
+                }
+            };
+
+            match msg {
+                ValidatorQueueMessage::Delete(list) => {
+                    if list.validated {
+                        // A pre-validated list may only be seen during recovery, if we are recovering
+                        // a DeletionList whose on-disk state has validated=true
+                        self.validated_lists.push(list)
+                    } else {
+                        self.pending_key_count += list.len();
+                        self.pending_lists.push(list);
+                    }
+
+                    if self.pending_key_count > AUTOFLUSH_KEY_COUNT {
+                        match self.flush().await {
+                            Ok(()) => {}
+                            Err(DeletionQueueError::ShuttingDown) => {
+                                // If we are shutting down, then auto-flush can safely be skipped
+                            }
+                        }
+                    }
+                }
+                ValidatorQueueMessage::Flush(op) => {
+                    match self.flush().await {
+                        Ok(()) => {
+                            op.notify();
+                        }
+                        Err(DeletionQueueError::ShuttingDown) => {
+                            // If we fail due to shutting down, we will just drop `op` to propagate that status.
+                        }
+                    }
+                }
+            }
+        }
+    }
+}

pageserver/src/disk_usage_eviction_task.rs

@@ -43,12 +43,12 @@
 
 use std::{
     collections::HashMap,
-    path::Path,
     sync::Arc,
     time::{Duration, SystemTime},
 };
 
 use anyhow::Context;
+use camino::Utf8Path;
 use remote_storage::GenericRemoteStorage;
 use serde::{Deserialize, Serialize};
 use tokio::time::Instant;
@@ -67,16 +67,40 @@ use crate::{
 pub struct DiskUsageEvictionTaskConfig {
     pub max_usage_pct: Percent,
     pub min_avail_bytes: u64,
+
+    // Control how far we will go when evicting: when usage exceeds max_usage_pct or min_avail_bytes,
+    // we will keep evicting layers until we reach the target.  The resulting disk usage should look
+    // like a sawtooth bouncing between the upper max/min line and the lower target line.
+    #[serde(default)]
+    pub target_usage_pct: Option<Percent>,
+    #[serde(default)]
+    pub target_avail_bytes: Option<u64>,
+
     #[serde(with = "humantime_serde")]
     pub period: Duration,
     #[cfg(feature = "testing")]
     pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
 }
 
+#[derive(Default)]
+enum Status {
+    /// We are within disk limits, and not currently doing any eviction
+    #[default]
+    Idle,
+    /// Disk limits have been exceeded: we will evict soon
+    UnderPressure,
+    /// We are currently doing an eviction pass.
+    Evicting,
+}
+
 #[derive(Default)]
 pub struct State {
     /// Exclude http requests and background task from running at the same time.
     mutex: tokio::sync::Mutex<()>,
+
+    /// Publish the current status of eviction work, for visibility to other subsystems
+    /// that modify their behavior if disk pressure is high or if eviction is going on.
+    status: std::sync::RwLock<Status>,
 }
 
 pub fn launch_disk_usage_global_eviction_task(
@@ -122,7 +146,7 @@ async fn disk_usage_eviction_task(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
     storage: GenericRemoteStorage,
-    tenants_dir: &Path,
+    tenants_dir: &Utf8Path,
     cancel: CancellationToken,
 ) {
     scopeguard::defer! {
@@ -176,7 +200,9 @@ async fn disk_usage_eviction_task(
 }
 
 pub trait Usage: Clone + Copy + std::fmt::Debug {
-    fn has_pressure(&self) -> bool;
+    fn pressure(&self) -> f64;
+    fn over_pressure(&self) -> bool;
+    fn no_pressure(&self) -> bool;
     fn add_available_bytes(&mut self, bytes: u64);
 }
 
@@ -184,18 +210,24 @@ async fn disk_usage_eviction_task_iteration(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
     storage: &GenericRemoteStorage,
-    tenants_dir: &Path,
+    tenants_dir: &Utf8Path,
     cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
+
+    if usage_pre.over_pressure() {
+        *state.status.write().unwrap() = Status::Evicting;
+    }
+
     let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
-            match outcome {
+            let new_status = match outcome {
                 IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
                     // nothing to do, select statement below will handle things
+                    Status::Idle
                 }
                 IterationOutcome::Finished(outcome) => {
                     // Verify with statvfs whether we made any real progress
@@ -205,21 +237,30 @@ async fn disk_usage_eviction_task_iteration(
 
                     debug!(?after, "disk usage");
 
-                    if after.has_pressure() {
+                    if after.over_pressure() {
                         // Don't bother doing an out-of-order iteration here now.
                         // In practice, the task period is set to a value in the tens-of-seconds range,
                         // which will cause another iteration to happen soon enough.
                         // TODO: deltas between the three different usages would be helpful,
                         // consider MiB, GiB, TiB
                         warn!(?outcome, ?after, "disk usage still high");
+                        Status::UnderPressure
                     } else {
                         info!(?outcome, ?after, "disk usage pressure relieved");
+                        Status::Idle
                     }
                 }
-            }
+            };
+
+            *state.status.write().unwrap() = new_status;
         }
         Err(e) => {
             error!("disk_usage_eviction_iteration failed: {:#}", e);
+            *state.status.write().unwrap() = if usage_pre.over_pressure() {
+                Status::UnderPressure
+            } else {
+                Status::Idle
+            };
         }
     }
 
@@ -285,8 +326,10 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
     debug!(?usage_pre, "disk usage");
 
-    if !usage_pre.has_pressure() {
+    if !usage_pre.over_pressure() {
         return Ok(IterationOutcome::NoPressure);
+    } else {
+        *state.status.write().unwrap() = Status::Evicting;
     }
 
     warn!(
@@ -334,7 +377,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     let mut warned = None;
     let mut usage_planned = usage_pre;
     for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
-        if !usage_planned.has_pressure() {
+        if usage_planned.no_pressure() {
             debug!(
                 no_candidates_evicted = i,
                 "took enough candidates for pressure to be relieved"
@@ -411,6 +454,11 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                                 evictions_failed.file_sizes += file_size;
                                 evictions_failed.count += 1;
                             }
+                            Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                                warn!(%layer, "failed to evict layer: {detail}");
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
                             None => {
                                 assert!(cancel.is_cancelled());
                                 return;
@@ -620,9 +668,8 @@ impl std::ops::Deref for TimelineKey {
 }
 
 mod filesystem_level_usage {
-    use std::path::Path;
-
     use anyhow::Context;
+    use camino::Utf8Path;
 
     use crate::statvfs::Statvfs;
 
@@ -640,22 +687,57 @@ mod filesystem_level_usage {
     }
 
     impl super::Usage for Usage<'_> {
-        fn has_pressure(&self) -> bool {
-            let usage_pct =
-                (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
+        /// Does the pressure exceed 1.0, i.e. has the disk usage exceeded upper bounds?
+        ///
+        /// This is the condition for starting eviction.
+        fn over_pressure(&self) -> bool {
+            self.pressure() >= 1.0
+        }
 
-            let pressures = [
-                (
-                    "min_avail_bytes",
-                    self.avail_bytes < self.config.min_avail_bytes,
-                ),
-                (
-                    "max_usage_pct",
-                    usage_pct >= self.config.max_usage_pct.get() as u64,
-                ),
-            ];
+        /// Is the pressure <0, ie.. has disk usage gone below the target bound?
+        ///
+        /// This is the condition for dropping out of eviction.
+        fn no_pressure(&self) -> bool {
+            self.pressure() <= 0.0
+        }
 
-            pressures.into_iter().any(|(_, has_pressure)| has_pressure)
+        fn pressure(&self) -> f64 {
+            let max_usage = std::cmp::min(
+                self.total_bytes - self.config.min_avail_bytes,
+                (self.total_bytes as f64 * (self.config.max_usage_pct.get() as f64 / 100.0)) as u64,
+            );
+
+            let mut target_usage = max_usage;
+            if let Some(target_avail_bytes) = self.config.target_avail_bytes {
+                target_usage = std::cmp::min(target_usage, self.total_bytes - target_avail_bytes);
+            }
+            if let Some(target_usage_pct) = self.config.target_usage_pct {
+                target_usage = std::cmp::min(
+                    target_usage,
+                    (self.total_bytes as f64 * (target_usage_pct.get() as f64 / 100.0)) as u64,
+                );
+            };
+
+            let usage = self.total_bytes - self.avail_bytes;
+            eprintln!(
+                "pressure: {} {}, current {}",
+                target_usage, max_usage, usage
+            );
+            if target_usage == max_usage {
+                // We are configured with a zero sized range: treat anything at+beyond limit as pressure 1.0, else 0.0
+                if usage >= max_usage {
+                    1.0
+                } else {
+                    0.0
+                }
+            } else if usage <= target_usage {
+                // No pressure.
+                0.0
+            } else {
+                // We are above target: pressure is the ratio of how much we exceed target to the size of the gap
+                let range_size = (max_usage - target_usage) as f64;
+                (usage - target_usage) as f64 / range_size
+            }
         }
 
         fn add_available_bytes(&mut self, bytes: u64) {
@@ -664,7 +746,7 @@ mod filesystem_level_usage {
     }
 
     pub fn get<'a>(
-        tenants_dir: &Path,
+        tenants_dir: &Utf8Path,
         config: &'a DiskUsageEvictionTaskConfig,
     ) -> anyhow::Result<Usage<'a>> {
         let mock_config = {
@@ -709,6 +791,8 @@ mod filesystem_level_usage {
             config: &DiskUsageEvictionTaskConfig {
                 max_usage_pct: Percent::new(85).unwrap(),
                 min_avail_bytes: 0,
+                target_avail_bytes: None,
+                target_usage_pct: None,
                 period: Duration::MAX,
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
@@ -717,24 +801,24 @@ mod filesystem_level_usage {
             avail_bytes: 0,
         };
 
-        assert!(usage.has_pressure(), "expected pressure at 100%");
+        assert!(usage.over_pressure(), "expected pressure at 100%");
 
         usage.add_available_bytes(14_000);
-        assert!(usage.has_pressure(), "expected pressure at 86%");
+        assert!(usage.over_pressure(), "expected pressure at 86%");
 
         usage.add_available_bytes(999);
-        assert!(usage.has_pressure(), "expected pressure at 85.001%");
+        assert!(usage.over_pressure(), "expected pressure at 85.001%");
 
         usage.add_available_bytes(1);
-        assert!(usage.has_pressure(), "expected pressure at precisely 85%");
+        assert!(usage.over_pressure(), "expected pressure at precisely 85%");
 
         usage.add_available_bytes(1);
-        assert!(!usage.has_pressure(), "no pressure at 84.999%");
+        assert!(!usage.over_pressure(), "no pressure at 84.999%");
 
         usage.add_available_bytes(999);
-        assert!(!usage.has_pressure(), "no pressure at 84%");
+        assert!(!usage.over_pressure(), "no pressure at 84%");
 
         usage.add_available_bytes(16_000);
-        assert!(!usage.has_pressure());
+        assert!(!usage.over_pressure());
     }
 }

pageserver/src/http/openapi_spec.yml

@@ -93,9 +93,16 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     delete:
       description: |
-        Attempts to delete specified tenant. 500 and 409 errors should be retried until 404 is retrieved.
+        Attempts to delete specified tenant. 500, 503 and 409 errors should be retried until 404 is retrieved.
         404 means that deletion successfully finished"
       responses:
         "400":
@@ -134,6 +141,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline:
     parameters:
@@ -178,6 +192,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline/{timeline_id}:
     parameters:
@@ -226,6 +247,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     delete:
       description: "Attempts to delete specified timeline. 500 and 409 errors should be retried"
       responses:
@@ -265,7 +293,74 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/PreconditionFailedError"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
 
+  /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_timestamp_of_lsn:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    get:
+      description: Get timestamp for a given LSN
+      parameters:
+        - name: lsn
+          in: query
+          required: true
+          schema:
+            type: integer
+          description: A LSN to get the timestamp
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                type: string
+                format: date-time
+        "400":
+          description: Error when no tenant id found in path, no timeline id or invalid timestamp
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "401":
+          description: Unauthorized Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/UnauthorizedError"
+        "403":
+          description: Forbidden Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ForbiddenError"
+        "404":
+          description: Timeline not found, or there is no timestamp information for the given lsn
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
         "500":
           description: Generic operation error
           content:
@@ -328,6 +423,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/do_gc:
     parameters:
       - name: tenant_id
@@ -375,6 +477,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/attach:
     parameters:
       - name: tenant_id
@@ -465,6 +574,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/detach:
     parameters:
@@ -518,6 +634,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/ignore:
     parameters:
@@ -560,6 +683,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/load:
     parameters:
@@ -604,6 +734,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/synthetic_size:
     parameters:
@@ -641,6 +778,12 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
 
   /v1/tenant/{tenant_id}/size:
     parameters:
@@ -704,6 +847,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/{tenant_id}/timeline/:
     parameters:
@@ -780,6 +930,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/:
     get:
       description: Get tenants list
@@ -810,6 +967,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
     post:
       description: |
         Create a tenant. Returns new tenant id on success.
@@ -860,6 +1024,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 
   /v1/tenant/config:
     put:
@@ -905,6 +1076,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
   /v1/tenant/{tenant_id}/config/:
     parameters:
       - name: tenant_id
@@ -954,6 +1132,13 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+        "503":
+          description: Temporarily unavailable, please retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
 components:
   securitySchemes:
     JWT:
@@ -1093,6 +1278,9 @@ components:
         remote_consistent_lsn:
           type: string
           format: hex
+        remote_consistent_lsn_visible:
+          type: string
+          format: hex
         ancestor_timeline_id:
           type: string
           format: hex
@@ -1217,6 +1405,13 @@ components:
       properties:
         msg:
           type: string
+    ServiceUnavailableError:
+      type: object
+      required:
+        - msg
+      properties:
+        msg:
+          type: string
     NotFoundError:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -2,14 +2,19 @@
 //! Management HTTP API
 //!
 use std::collections::HashMap;
+use std::str::FromStr;
 use std::sync::Arc;
 
 use anyhow::{anyhow, Context, Result};
+use futures::TryFutureExt;
+use humantime::format_rfc3339;
+use hyper::header::CONTENT_TYPE;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
 use pageserver_api::models::{
-    DownloadRemoteLayersTaskSpawnRequest, TenantAttachRequest, TenantLoadRequest,
+    DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
+    TenantLoadRequest, TenantLocationConfigRequest,
 };
 use remote_storage::GenericRemoteStorage;
 use tenant_size_model::{SizeResult, StorageModel};
@@ -24,17 +29,18 @@ use super::models::{
     TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
 };
 use crate::context::{DownloadBehavior, RequestContext};
+use crate::deletion_queue::DeletionQueueClient;
 use crate::metrics::{StorageTimeOperation, STORAGE_TIME_GLOBAL};
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
-use crate::tenant::config::TenantConfOpt;
+use crate::tenant::config::{LocationConf, TenantConfOpt};
 use crate::tenant::mgr::{
     GetTenantError, SetNewTenantConfigError, TenantMapInsertError, TenantStateError,
 };
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::timeline::Timeline;
-use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
+use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError, TenantSharedResources};
 use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
 use utils::{
@@ -61,6 +67,7 @@ pub struct State {
     remote_storage: Option<GenericRemoteStorage>,
     broker_client: storage_broker::BrokerClientChannel,
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
+    deletion_queue_client: DeletionQueueClient,
 }
 
 impl State {
@@ -70,8 +77,9 @@ impl State {
         remote_storage: Option<GenericRemoteStorage>,
         broker_client: storage_broker::BrokerClientChannel,
         disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
+        deletion_queue_client: DeletionQueueClient,
     ) -> anyhow::Result<Self> {
-        let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml"]
+        let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml", "/metrics"]
             .iter()
             .map(|v| v.parse().unwrap())
             .collect::<Vec<_>>();
@@ -82,8 +90,17 @@ impl State {
             remote_storage,
             broker_client,
             disk_usage_eviction_state,
+            deletion_queue_client,
         })
     }
+
+    fn tenant_resources(&self) -> TenantSharedResources {
+        TenantSharedResources {
+            broker_client: self.broker_client.clone(),
+            remote_storage: self.remote_storage.clone(),
+            deletion_queue_client: self.deletion_queue_client.clone(),
+        }
+    }
 }
 
 #[inline(always)]
@@ -119,11 +136,9 @@ impl From<PageReconstructError> for ApiError {
                 ApiError::InternalServerError(anyhow::anyhow!("request was cancelled"))
             }
             PageReconstructError::AncestorStopping(_) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
-            }
-            PageReconstructError::WalRedo(pre) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
+                ApiError::ResourceUnavailable(format!("{pre}").into())
             }
+            PageReconstructError::WalRedo(pre) => ApiError::InternalServerError(pre),
         }
     }
 }
@@ -132,12 +147,15 @@ impl From<TenantMapInsertError> for ApiError {
     fn from(tmie: TenantMapInsertError) -> ApiError {
         match tmie {
             TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
-                ApiError::InternalServerError(anyhow::Error::new(tmie))
+                ApiError::ResourceUnavailable(format!("{tmie}").into())
             }
             TenantMapInsertError::TenantAlreadyExists(id, state) => {
                 ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
             }
-            TenantMapInsertError::Closure(e) => ApiError::InternalServerError(e),
+            TenantMapInsertError::TenantExistsSecondary(id) => {
+                ApiError::Conflict(format!("tenant {id} already exists as secondary"))
+            }
+            TenantMapInsertError::Other(e) => ApiError::InternalServerError(e),
         }
     }
 }
@@ -146,6 +164,9 @@ impl From<TenantStateError> for ApiError {
     fn from(tse: TenantStateError) -> ApiError {
         match tse {
             TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
+            TenantStateError::IsStopping(_) => {
+                ApiError::ResourceUnavailable("Tenant is stopping".into())
+            }
             _ => ApiError::InternalServerError(anyhow::Error::new(tse)),
         }
     }
@@ -155,14 +176,17 @@ impl From<GetTenantError> for ApiError {
     fn from(tse: GetTenantError) -> ApiError {
         match tse {
             GetTenantError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
-            e @ GetTenantError::NotActive(_) => {
+            GetTenantError::Broken(reason) => {
+                ApiError::InternalServerError(anyhow!("tenant is broken: {}", reason))
+            }
+            GetTenantError::NotActive(_) => {
                 // Why is this not `ApiError::NotFound`?
                 // Because we must be careful to never return 404 for a tenant if it does
                 // in fact exist locally. If we did, the caller could draw the conclusion
                 // that it can attach the tenant to another PS and we'd be in split-brain.
                 //
                 // (We can produce this variant only in `mgr::get_tenant(..., active=true)` calls).
-                ApiError::InternalServerError(anyhow::Error::new(e))
+                ApiError::ResourceUnavailable("Tenant not yet active".into())
             }
         }
     }
@@ -283,7 +307,12 @@ async fn build_timeline_info_common(
     };
     let current_physical_size = Some(timeline.layer_size_sum().await);
     let state = timeline.current_state();
-    let remote_consistent_lsn = timeline.get_remote_consistent_lsn().unwrap_or(Lsn(0));
+    let remote_consistent_lsn_projected = timeline
+        .get_remote_consistent_lsn_projected()
+        .unwrap_or(Lsn(0));
+    let remote_consistent_lsn_visible = timeline
+        .get_remote_consistent_lsn_visible()
+        .unwrap_or(Lsn(0));
 
     let walreceiver_status = timeline.walreceiver_status();
 
@@ -293,7 +322,8 @@ async fn build_timeline_info_common(
         ancestor_timeline_id,
         ancestor_lsn,
         disk_consistent_lsn: timeline.get_disk_consistent_lsn(),
-        remote_consistent_lsn,
+        remote_consistent_lsn: remote_consistent_lsn_projected,
+        remote_consistent_lsn_visible,
         last_record_lsn,
         prev_record_lsn: Some(timeline.get_prev_record_lsn()),
         latest_gc_cutoff_lsn: *timeline.get_latest_gc_cutoff_lsn(),
@@ -363,6 +393,9 @@ async fn timeline_create_handler(
                     format!("{err:#}")
                 ))
             }
+            Err(e @ tenant::CreateTimelineError::AncestorNotActive) => {
+                json_response(StatusCode::SERVICE_UNAVAILABLE, HttpErrorBody::from_msg(e.to_string()))
+            }
             Err(tenant::CreateTimelineError::Other(err)) => Err(ApiError::InternalServerError(err)),
         }
     }
@@ -471,6 +504,33 @@ async fn get_lsn_by_timestamp_handler(
     json_response(StatusCode::OK, result)
 }
 
+async fn get_timestamp_of_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+
+    let lsn_str = must_get_query_param(&request, "lsn")?;
+    let lsn = Lsn::from_str(&lsn_str)
+        .with_context(|| format!("Invalid LSN: {lsn_str:?}"))
+        .map_err(ApiError::BadRequest)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+    let result = timeline.get_timestamp_for_lsn(lsn, &ctx).await?;
+
+    match result {
+        Some(time) => {
+            let time = format_rfc3339(postgres_ffi::from_pg_timestamp(time)).to_string();
+            json_response(StatusCode::OK, time)
+        }
+        None => json_response(StatusCode::NOT_FOUND, ()),
+    }
+}
+
 async fn tenant_attach_handler(
     mut request: Request<Body>,
     _cancel: CancellationToken,
@@ -492,24 +552,23 @@ async fn tenant_attach_handler(
 
     let generation = get_request_generation(state, maybe_body.as_ref().and_then(|r| r.generation))?;
 
-    if let Some(remote_storage) = &state.remote_storage {
-        mgr::attach_tenant(
-            state.conf,
-            tenant_id,
-            generation,
-            tenant_conf,
-            state.broker_client.clone(),
-            remote_storage.clone(),
-            &ctx,
-        )
-        .instrument(info_span!("tenant_attach", %tenant_id))
-        .await?;
-    } else {
+    if state.remote_storage.is_none() {
         return Err(ApiError::BadRequest(anyhow!(
             "attach_tenant is not possible because pageserver was configured without remote storage"
         )));
     }
 
+    mgr::attach_tenant(
+        state.conf,
+        tenant_id,
+        generation,
+        tenant_conf,
+        state.tenant_resources(),
+        &ctx,
+    )
+    .instrument(info_span!("tenant_attach", %tenant_id))
+    .await?;
+
     json_response(StatusCode::ACCEPTED, ())
 }
 
@@ -540,9 +599,14 @@ async fn tenant_detach_handler(
 
     let state = get_state(&request);
     let conf = state.conf;
-    mgr::detach_tenant(conf, tenant_id, detach_ignored.unwrap_or(false))
-        .instrument(info_span!("tenant_detach", %tenant_id))
-        .await?;
+    mgr::detach_tenant(
+        conf,
+        tenant_id,
+        detach_ignored.unwrap_or(false),
+        &state.deletion_queue_client,
+    )
+    .instrument(info_span!("tenant_detach", %tenant_id))
+    .await?;
 
     json_response(StatusCode::OK, ())
 }
@@ -570,6 +634,7 @@ async fn tenant_load_handler(
         generation,
         state.broker_client.clone(),
         state.remote_storage.clone(),
+        state.deletion_queue_client.clone(),
         &ctx,
     )
     .instrument(info_span!("load", %tenant_id))
@@ -603,8 +668,9 @@ async fn tenant_list_handler(
     let response_data = mgr::list_tenants()
         .instrument(info_span!("tenant_list"))
         .await
-        .map_err(anyhow::Error::new)
-        .map_err(ApiError::InternalServerError)?
+        .map_err(|_| {
+            ApiError::ResourceUnavailable("Tenant map is initializing or shutting down".into())
+        })?
         .iter()
         .map(|(id, state)| TenantInfo {
             id: *id,
@@ -911,8 +977,7 @@ async fn tenant_create_handler(
         tenant_conf,
         target_tenant_id,
         generation,
-        state.broker_client.clone(),
-        state.remote_storage.clone(),
+        state.tenant_resources(),
         &ctx,
     )
     .instrument(info_span!("tenant_create", tenant_id = %target_tenant_id))
@@ -983,6 +1048,56 @@ async fn update_tenant_config_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn put_tenant_location_config_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let request_data: TenantLocationConfigRequest = json_request(&mut request).await?;
+    let tenant_id = request_data.tenant_id;
+    check_permission(&request, Some(tenant_id))?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+    let state = get_state(&request);
+    let conf = state.conf;
+
+    // The `Detached` state is special, it doesn't upsert a tenant, it removes
+    // its local disk content and drops it from memory.
+    if let LocationConfigMode::Detached = request_data.config.mode {
+        if let Err(e) = mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
+            .instrument(info_span!("tenant_detach", %tenant_id))
+            .await
+        {
+            match e {
+                TenantStateError::NotFound(_) => {
+                    // This API is idempotent: a NotFound on a detach is fine.
+                }
+                _ => return Err(e.into()),
+            }
+        }
+        return json_response(StatusCode::OK, ());
+    }
+
+    let location_conf =
+        LocationConf::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
+
+    mgr::upsert_location(
+        state.conf,
+        tenant_id,
+        location_conf,
+        state.broker_client.clone(),
+        state.remote_storage.clone(),
+        state.deletion_queue_client.clone(),
+        &ctx,
+    )
+    .await
+    // TODO: badrequest assumes the caller was asking for something unreasonable, but in
+    // principle we might have hit something like concurrent API calls to the same tenant,
+    // which is not a 400 but a 409.
+    .map_err(ApiError::BadRequest)?;
+
+    json_response(StatusCode::OK, ())
+}
+
 /// Testing helper to transition a tenant to [`crate::tenant::TenantState::Broken`].
 async fn handle_tenant_break(
     r: Request<Body>,
@@ -1129,6 +1244,169 @@ async fn timeline_download_remote_layers_handler_get(
     json_response(StatusCode::OK, info)
 }
 
+async fn deletion_queue_flush(
+    r: Request<Body>,
+    cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&r);
+
+    if state.remote_storage.is_none() {
+        // Nothing to do if remote storage is disabled.
+        return json_response(StatusCode::OK, ());
+    }
+
+    let execute = parse_query_param(&r, "execute")?.unwrap_or(false);
+
+    let flush = async {
+        if execute {
+            state.deletion_queue_client.flush_execute().await
+        } else {
+            state.deletion_queue_client.flush().await
+        }
+    }
+    // DeletionQueueError's only case is shutting down.
+    .map_err(|_| ApiError::ShuttingDown);
+
+    tokio::select! {
+        res = flush => {
+            res.map(|()| json_response(StatusCode::OK, ()))?
+        }
+        _ = cancel.cancelled() => {
+            Err(ApiError::ShuttingDown)
+        }
+    }
+}
+
+/// Try if `GetPage@Lsn` is successful, useful for manual debugging.
+async fn getpage_at_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    struct Key(crate::repository::Key);
+
+    impl std::str::FromStr for Key {
+        type Err = anyhow::Error;
+
+        fn from_str(s: &str) -> std::result::Result<Self, Self::Err> {
+            crate::repository::Key::from_hex(s).map(Key)
+        }
+    }
+
+    let key: Key = parse_query_param(&request, "key")?
+        .ok_or_else(|| ApiError::BadRequest(anyhow!("missing 'key' query parameter")))?;
+    let lsn: Lsn = parse_query_param(&request, "lsn")?
+        .ok_or_else(|| ApiError::BadRequest(anyhow!("missing 'lsn' query parameter")))?;
+
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+
+        let page = timeline.get(key.0, lsn, &ctx).await?;
+
+        Result::<_, ApiError>::Ok(
+            Response::builder()
+                .status(StatusCode::OK)
+                .header(CONTENT_TYPE, "application/octet-stream")
+                .body(hyper::Body::from(page))
+                .unwrap(),
+        )
+    }
+    .instrument(info_span!("timeline_get", %tenant_id, %timeline_id))
+    .await
+}
+
+async fn timeline_collect_keyspace(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    struct Partitioning {
+        keys: crate::keyspace::KeySpace,
+
+        at_lsn: Lsn,
+    }
+
+    impl serde::Serialize for Partitioning {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeMap;
+            let mut map = serializer.serialize_map(Some(2))?;
+            map.serialize_key("keys")?;
+            map.serialize_value(&KeySpace(&self.keys))?;
+            map.serialize_key("at_lsn")?;
+            map.serialize_value(&WithDisplay(&self.at_lsn))?;
+            map.end()
+        }
+    }
+
+    struct WithDisplay<'a, T>(&'a T);
+
+    impl<'a, T: std::fmt::Display> serde::Serialize for WithDisplay<'a, T> {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            serializer.collect_str(&self.0)
+        }
+    }
+
+    struct KeySpace<'a>(&'a crate::keyspace::KeySpace);
+
+    impl<'a> serde::Serialize for KeySpace<'a> {
+        fn serialize<S>(&self, serializer: S) -> std::result::Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeSeq;
+            let mut seq = serializer.serialize_seq(Some(self.0.ranges.len()))?;
+            for kr in &self.0.ranges {
+                seq.serialize_element(&KeyRange(kr))?;
+            }
+            seq.end()
+        }
+    }
+
+    struct KeyRange<'a>(&'a std::ops::Range<crate::repository::Key>);
+
+    impl<'a> serde::Serialize for KeyRange<'a> {
+        fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+        where
+            S: serde::Serializer,
+        {
+            use serde::ser::SerializeTuple;
+            let mut t = serializer.serialize_tuple(2)?;
+            t.serialize_element(&WithDisplay(&self.0.start))?;
+            t.serialize_element(&WithDisplay(&self.0.end))?;
+            t.end()
+        }
+    }
+
+    let at_lsn: Option<Lsn> = parse_query_param(&request, "at_lsn")?;
+
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+        let at_lsn = at_lsn.unwrap_or_else(|| timeline.get_last_record_lsn());
+        let keys = timeline
+            .collect_keyspace(at_lsn, &ctx)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+
+        json_response(StatusCode::OK, Partitioning { keys, at_lsn })
+    }
+    .instrument(info_span!("timeline_collect_keyspace", %tenant_id, %timeline_id))
+    .await
+}
+
 async fn active_timeline_of_active_tenant(
     tenant_id: TenantId,
     timeline_id: TimelineId,
@@ -1174,10 +1452,22 @@ async fn disk_usage_eviction_run(
     }
 
     impl crate::disk_usage_eviction_task::Usage for Usage {
-        fn has_pressure(&self) -> bool {
+        fn over_pressure(&self) -> bool {
             self.config.evict_bytes > self.freed_bytes
         }
 
+        fn no_pressure(&self) -> bool {
+            !self.over_pressure()
+        }
+
+        fn pressure(&self) -> f64 {
+            if self.over_pressure() {
+                1.0
+            } else {
+                0.0
+            }
+        }
+
         fn add_available_bytes(&mut self, bytes: u64) {
             self.freed_bytes += bytes;
         }
@@ -1403,6 +1693,9 @@ pub fn make_router(
         .get("/v1/tenant/:tenant_id/config", |r| {
             api_handler(r, get_tenant_config_handler)
         })
+        .put("/v1/tenant/:tenant_id/location_config", |r| {
+            api_handler(r, put_tenant_location_config_handler)
+        })
         .get("/v1/tenant/:tenant_id/timeline", |r| {
             api_handler(r, timeline_list_handler)
         })
@@ -1428,6 +1721,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/get_lsn_by_timestamp",
             |r| api_handler(r, get_lsn_by_timestamp_handler),
         )
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/get_timestamp_of_lsn",
+            |r| api_handler(r, get_timestamp_of_lsn_handler),
+        )
         .put("/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc", |r| {
             api_handler(r, timeline_gc_handler)
         })
@@ -1463,6 +1760,9 @@ pub fn make_router(
         .put("/v1/disk_usage_eviction/run", |r| {
             api_handler(r, disk_usage_eviction_run)
         })
+        .put("/v1/deletion_queue/flush", |r| {
+            api_handler(r, deletion_queue_flush)
+        })
         .put("/v1/tenant/:tenant_id/break", |r| {
             testing_api_handler("set tenant state to broken", r, handle_tenant_break)
         })
@@ -1470,5 +1770,12 @@ pub fn make_router(
         .post("/v1/tracing/event", |r| {
             testing_api_handler("emit a tracing event", r, post_tracing_event_handler)
         })
+        .get("/v1/tenant/:tenant_id/timeline/:timeline_id/getpage", |r| {
+            testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler)
+        })
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/keyspace",
+            |r| testing_api_handler("read out the keyspace", r, timeline_collect_keyspace),
+        )
         .any(handler_404))
 }

pageserver/src/import_datadir.rs

@@ -6,6 +6,7 @@ use std::path::{Path, PathBuf};
 
 use anyhow::{bail, ensure, Context, Result};
 use bytes::Bytes;
+use camino::Utf8Path;
 use futures::StreamExt;
 use tokio::io::{AsyncRead, AsyncReadExt};
 use tokio_tar::Archive;
@@ -29,7 +30,7 @@ use postgres_ffi::{BLCKSZ, WAL_SEGMENT_SIZE};
 use utils::lsn::Lsn;
 
 // Returns checkpoint LSN from controlfile
-pub fn get_lsn_from_controlfile(path: &Path) -> Result<Lsn> {
+pub fn get_lsn_from_controlfile(path: &Utf8Path) -> Result<Lsn> {
     // Read control file to extract the LSN
     let controlfile_path = path.join("global").join("pg_control");
     let controlfile = ControlFileData::decode(&std::fs::read(controlfile_path)?)?;
@@ -46,7 +47,7 @@ pub fn get_lsn_from_controlfile(path: &Path) -> Result<Lsn> {
 /// cluster was not shut down cleanly.
 pub async fn import_timeline_from_postgres_datadir(
     tline: &Timeline,
-    pgdata_path: &Path,
+    pgdata_path: &Utf8Path,
     pgdata_lsn: Lsn,
     ctx: &RequestContext,
 ) -> Result<()> {
@@ -75,12 +76,12 @@ pub async fn import_timeline_from_postgres_datadir(
             {
                 pg_control = Some(control_file);
             }
-            modification.flush().await?;
+            modification.flush(ctx).await?;
         }
     }
 
     // We're done importing all the data files.
-    modification.commit().await?;
+    modification.commit(ctx).await?;
 
     // We expect the Postgres server to be shut down cleanly.
     let pg_control = pg_control.context("pg_control file not found")?;
@@ -256,7 +257,7 @@ async fn import_slru(
 /// Scan PostgreSQL WAL files in given directory and load all records between
 /// 'startpoint' and 'endpoint' into the repository.
 async fn import_wal(
-    walpath: &Path,
+    walpath: &Utf8Path,
     tline: &Timeline,
     startpoint: Lsn,
     endpoint: Lsn,
@@ -359,7 +360,7 @@ pub async fn import_basebackup_from_tar(
                     // We found the pg_control file.
                     pg_control = Some(res);
                 }
-                modification.flush().await?;
+                modification.flush(ctx).await?;
             }
             tokio_tar::EntryType::Directory => {
                 debug!("directory {:?}", file_path);
@@ -377,7 +378,7 @@ pub async fn import_basebackup_from_tar(
     // sanity check: ensure that pg_control is loaded
     let _pg_control = pg_control.context("pg_control file not found")?;
 
-    modification.commit().await?;
+    modification.commit(ctx).await?;
     Ok(())
 }
 

pageserver/src/lib.rs

@@ -3,7 +3,8 @@ pub mod basebackup;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
-mod control_plane_client;
+pub mod control_plane_client;
+pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;
@@ -24,9 +25,9 @@ pub mod walredo;
 
 pub mod failpoint_support;
 
-use std::path::Path;
-
 use crate::task_mgr::TaskKind;
+use camino::Utf8Path;
+use deletion_queue::DeletionQueue;
 use tracing::info;
 
 /// Current storage format version
@@ -48,8 +49,8 @@ static ZERO_PAGE: bytes::Bytes = bytes::Bytes::from_static(&[0u8; 8192]);
 
 pub use crate::metrics::preinitialize_metrics;
 
-#[tracing::instrument]
-pub async fn shutdown_pageserver(exit_code: i32) {
+#[tracing::instrument(skip_all, fields(%exit_code))]
+pub async fn shutdown_pageserver(deletion_queue: Option<DeletionQueue>, exit_code: i32) {
     use std::time::Duration;
     // Shut down the libpq endpoint task. This prevents new connections from
     // being accepted.
@@ -77,6 +78,11 @@ pub async fn shutdown_pageserver(exit_code: i32) {
     )
     .await;
 
+    // Best effort to persist any outstanding deletions, to avoid leaking objects
+    if let Some(mut deletion_queue) = deletion_queue {
+        deletion_queue.shutdown(Duration::from_secs(5)).await;
+    }
+
     // Shut down the HTTP endpoint last, so that you can still check the server's
     // status while it's shutting down.
     // FIXME: We should probably stop accepting commands like attach/detach earlier.
@@ -106,6 +112,10 @@ pub const METADATA_FILE_NAME: &str = "metadata";
 /// Full path: `tenants/<tenant_id>/config`.
 pub const TENANT_CONFIG_NAME: &str = "config";
 
+/// Per-tenant configuration file.
+/// Full path: `tenants/<tenant_id>/config`.
+pub const TENANT_LOCATION_CONFIG_NAME: &str = "config-v1";
+
 /// A suffix used for various temporary files. Any temporary files found in the
 /// data directory at pageserver startup can be automatically removed.
 pub const TEMP_FILE_SUFFIX: &str = "___temp";
@@ -125,25 +135,25 @@ pub const TIMELINE_DELETE_MARK_SUFFIX: &str = "___delete";
 /// Full path: `tenants/<tenant_id>/___ignored_tenant`.
 pub const IGNORED_TENANT_FILE_NAME: &str = "___ignored_tenant";
 
-pub fn is_temporary(path: &Path) -> bool {
+pub fn is_temporary(path: &Utf8Path) -> bool {
     match path.file_name() {
-        Some(name) => name.to_string_lossy().ends_with(TEMP_FILE_SUFFIX),
+        Some(name) => name.ends_with(TEMP_FILE_SUFFIX),
         None => false,
     }
 }
 
-fn ends_with_suffix(path: &Path, suffix: &str) -> bool {
+fn ends_with_suffix(path: &Utf8Path, suffix: &str) -> bool {
     match path.file_name() {
-        Some(name) => name.to_string_lossy().ends_with(suffix),
+        Some(name) => name.ends_with(suffix),
         None => false,
     }
 }
 
-pub fn is_uninit_mark(path: &Path) -> bool {
+pub fn is_uninit_mark(path: &Utf8Path) -> bool {
     ends_with_suffix(path, TIMELINE_UNINIT_MARK_SUFFIX)
 }
 
-pub fn is_delete_mark(path: &Path) -> bool {
+pub fn is_delete_mark(path: &Utf8Path) -> bool {
     ends_with_suffix(path, TIMELINE_DELETE_MARK_SUFFIX)
 }
 
@@ -163,6 +173,9 @@ fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
 /// delaying is needed.
 #[derive(Clone)]
 pub struct InitializationOrder {
+    /// Each initial tenant load task carries this until it is done loading timelines from remote storage
+    pub initial_tenant_load_remote: Option<utils::completion::Completion>,
+
     /// Each initial tenant load task carries this until completion.
     pub initial_tenant_load: Option<utils::completion::Completion>,
 

pageserver/src/metrics.rs

@@ -1,3 +1,4 @@
+use enum_map::EnumMap;
 use metrics::metric_vec_duration::DurationResultObserver;
 use metrics::{
     register_counter_vec, register_gauge_vec, register_histogram, register_histogram_vec,
@@ -93,15 +94,35 @@ pub(crate) static READ_NUM_FS_LAYERS: Lazy<Histogram> = Lazy::new(|| {
 });
 
 // Metrics collected on operations on the storage repository.
-pub(crate) static RECONSTRUCT_TIME: Lazy<Histogram> = Lazy::new(|| {
-    register_histogram!(
+
+pub(crate) struct ReconstructTimeMetrics {
+    ok: Histogram,
+    err: Histogram,
+}
+
+pub(crate) static RECONSTRUCT_TIME: Lazy<ReconstructTimeMetrics> = Lazy::new(|| {
+    let inner = register_histogram_vec!(
         "pageserver_getpage_reconstruct_seconds",
         "Time spent in reconstruct_value (reconstruct a page from deltas)",
+        &["result"],
         CRITICAL_OP_BUCKETS.into(),
     )
-    .expect("failed to define a metric")
+    .expect("failed to define a metric");
+    ReconstructTimeMetrics {
+        ok: inner.get_metric_with_label_values(&["ok"]).unwrap(),
+        err: inner.get_metric_with_label_values(&["err"]).unwrap(),
+    }
 });
 
+impl ReconstructTimeMetrics {
+    pub(crate) fn for_result<T, E>(&self, result: &Result<T, E>) -> &Histogram {
+        match result {
+            Ok(_) => &self.ok,
+            Err(_) => &self.err,
+        }
+    }
+}
+
 pub(crate) static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_materialized_cache_hits_direct_total",
@@ -127,22 +148,24 @@ pub(crate) static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounter> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub struct PageCacheMetrics {
+pub struct PageCacheMetricsForTaskKind {
     pub read_accesses_materialized_page: IntCounter,
-    pub read_accesses_ephemeral: IntCounter,
     pub read_accesses_immutable: IntCounter,
 
-    pub read_hits_ephemeral: IntCounter,
     pub read_hits_immutable: IntCounter,
     pub read_hits_materialized_page_exact: IntCounter,
     pub read_hits_materialized_page_older_lsn: IntCounter,
 }
 
+pub struct PageCacheMetrics {
+    map: EnumMap<TaskKind, EnumMap<PageContentKind, PageCacheMetricsForTaskKind>>,
+}
+
 static PAGE_CACHE_READ_HITS: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_page_cache_read_hits_total",
         "Number of read accesses to the page cache that hit",
-        &["key_kind", "hit_kind"]
+        &["task_kind", "key_kind", "content_kind", "hit_kind"]
     )
     .expect("failed to define a metric")
 });
@@ -151,55 +174,73 @@ static PAGE_CACHE_READ_ACCESSES: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_page_cache_read_accesses_total",
         "Number of read accesses to the page cache",
-        &["key_kind"]
+        &["task_kind", "key_kind", "content_kind"]
     )
     .expect("failed to define a metric")
 });
 
 pub static PAGE_CACHE: Lazy<PageCacheMetrics> = Lazy::new(|| PageCacheMetrics {
-    read_accesses_materialized_page: {
-        PAGE_CACHE_READ_ACCESSES
-            .get_metric_with_label_values(&["materialized_page"])
-            .unwrap()
-    },
+    map: EnumMap::from_array(std::array::from_fn(|task_kind| {
+        let task_kind = <TaskKind as enum_map::Enum>::from_usize(task_kind);
+        let task_kind: &'static str = task_kind.into();
+        EnumMap::from_array(std::array::from_fn(|content_kind| {
+            let content_kind = <PageContentKind as enum_map::Enum>::from_usize(content_kind);
+            let content_kind: &'static str = content_kind.into();
+            PageCacheMetricsForTaskKind {
+                read_accesses_materialized_page: {
+                    PAGE_CACHE_READ_ACCESSES
+                        .get_metric_with_label_values(&[
+                            task_kind,
+                            "materialized_page",
+                            content_kind,
+                        ])
+                        .unwrap()
+                },
 
-    read_accesses_ephemeral: {
-        PAGE_CACHE_READ_ACCESSES
-            .get_metric_with_label_values(&["ephemeral"])
-            .unwrap()
-    },
+                read_accesses_immutable: {
+                    PAGE_CACHE_READ_ACCESSES
+                        .get_metric_with_label_values(&[task_kind, "immutable", content_kind])
+                        .unwrap()
+                },
 
-    read_accesses_immutable: {
-        PAGE_CACHE_READ_ACCESSES
-            .get_metric_with_label_values(&["immutable"])
-            .unwrap()
-    },
+                read_hits_immutable: {
+                    PAGE_CACHE_READ_HITS
+                        .get_metric_with_label_values(&[task_kind, "immutable", content_kind, "-"])
+                        .unwrap()
+                },
 
-    read_hits_ephemeral: {
-        PAGE_CACHE_READ_HITS
-            .get_metric_with_label_values(&["ephemeral", "-"])
-            .unwrap()
-    },
+                read_hits_materialized_page_exact: {
+                    PAGE_CACHE_READ_HITS
+                        .get_metric_with_label_values(&[
+                            task_kind,
+                            "materialized_page",
+                            content_kind,
+                            "exact",
+                        ])
+                        .unwrap()
+                },
 
-    read_hits_immutable: {
-        PAGE_CACHE_READ_HITS
-            .get_metric_with_label_values(&["immutable", "-"])
-            .unwrap()
-    },
-
-    read_hits_materialized_page_exact: {
-        PAGE_CACHE_READ_HITS
-            .get_metric_with_label_values(&["materialized_page", "exact"])
-            .unwrap()
-    },
-
-    read_hits_materialized_page_older_lsn: {
-        PAGE_CACHE_READ_HITS
-            .get_metric_with_label_values(&["materialized_page", "older_lsn"])
-            .unwrap()
-    },
+                read_hits_materialized_page_older_lsn: {
+                    PAGE_CACHE_READ_HITS
+                        .get_metric_with_label_values(&[
+                            task_kind,
+                            "materialized_page",
+                            content_kind,
+                            "older_lsn",
+                        ])
+                        .unwrap()
+                },
+            }
+        }))
+    })),
 });
 
+impl PageCacheMetrics {
+    pub(crate) fn for_ctx(&self, ctx: &RequestContext) -> &PageCacheMetricsForTaskKind {
+        &self.map[ctx.task_kind()][ctx.page_content_kind()]
+    }
+}
+
 pub struct PageCacheSizeMetrics {
     pub max_bytes: UIntGauge,
 
@@ -243,6 +284,46 @@ pub static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> = Lazy::new(|| PageCacheS
     },
 });
 
+pub(crate) static PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "pageserver_page_cache_acquire_pinned_slot_seconds",
+        "Time spent acquiring a pinned slot in the page cache",
+        CRITICAL_OP_BUCKETS.into(),
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_page_cache_find_victim_iters_total",
+        "Counter for the number of iterations in the find_victim loop",
+    )
+    .expect("failed to define a metric")
+});
+
+static PAGE_CACHE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "page_cache_errors_total",
+        "Number of timeouts while acquiring a pinned slot in the page cache",
+        &["error_kind"]
+    )
+    .expect("failed to define a metric")
+});
+
+#[derive(IntoStaticStr)]
+#[strum(serialize_all = "kebab_case")]
+pub(crate) enum PageCacheErrorKind {
+    AcquirePinnedSlotTimeout,
+    EvictIterLimit,
+}
+
+pub(crate) fn page_cache_errors_inc(error_kind: PageCacheErrorKind) {
+    PAGE_CACHE_ERRORS
+        .get_metric_with_label_values(&[error_kind.into()])
+        .unwrap()
+        .inc();
+}
+
 pub(crate) static WAIT_LSN_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wait_lsn_seconds",
@@ -270,6 +351,14 @@ static RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+pub(crate) static RESIDENT_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_resident_physical_size_global",
+        "Like `pageserver_resident_physical_size`, but without tenant/timeline dimensions."
+    )
+    .expect("failed to define a metric")
+});
+
 static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_remote_physical_size",
@@ -280,6 +369,14 @@ static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static REMOTE_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_remote_physical_size_global",
+        "Like `pageserver_remote_physical_size`, but without tenant/timeline dimensions."
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_remote_ondemand_downloaded_layers_total",
@@ -594,10 +691,9 @@ impl StorageIoTime {
         .expect("failed to define a metric");
         let metrics = std::array::from_fn(|i| {
             let op = StorageIoOperation::from_repr(i).unwrap();
-            let metric = storage_io_histogram_vec
+            storage_io_histogram_vec
                 .get_metric_with_label_values(&[op.as_str()])
-                .unwrap();
-            metric
+                .unwrap()
         });
         Self { metrics }
     }
@@ -866,6 +962,61 @@ static REMOTE_TIMELINE_CLIENT_BYTES_FINISHED_COUNTER: Lazy<IntCounterVec> = Lazy
     .expect("failed to define a metric")
 });
 
+pub(crate) struct DeletionQueueMetrics {
+    pub(crate) keys_submitted: IntCounter,
+    pub(crate) keys_dropped: IntCounter,
+    pub(crate) keys_executed: IntCounter,
+    pub(crate) keys_validated: IntCounter,
+    pub(crate) dropped_lsn_updates: IntCounter,
+    pub(crate) unexpected_errors: IntCounter,
+    pub(crate) remote_errors: IntCounterVec,
+}
+pub(crate) static DELETION_QUEUE: Lazy<DeletionQueueMetrics> = Lazy::new(|| {
+    DeletionQueueMetrics{
+
+    keys_submitted: register_int_counter!(
+        "pageserver_deletion_queue_submitted_total",
+        "Number of objects submitted for deletion"
+    )
+    .expect("failed to define a metric"),
+
+    keys_dropped: register_int_counter!(
+        "pageserver_deletion_queue_dropped_total",
+        "Number of object deletions dropped due to stale generation."
+    )
+    .expect("failed to define a metric"),
+
+    keys_executed: register_int_counter!(
+        "pageserver_deletion_queue_executed_total",
+        "Number of objects deleted. Only includes objects that we actually deleted, sum with pageserver_deletion_queue_dropped_total for the total number of keys processed to completion"
+    )
+    .expect("failed to define a metric"),
+
+    keys_validated: register_int_counter!(
+        "pageserver_deletion_queue_validated_total",
+        "Number of keys validated for deletion.  Sum with pageserver_deletion_queue_dropped_total for the total number of keys that have passed through the validation stage."
+    )
+    .expect("failed to define a metric"),
+
+    dropped_lsn_updates: register_int_counter!(
+        "pageserver_deletion_queue_dropped_lsn_updates_total",
+        "Updates to remote_consistent_lsn dropped due to stale generation number."
+    )
+    .expect("failed to define a metric"),
+    unexpected_errors: register_int_counter!(
+        "pageserver_deletion_queue_unexpected_errors_total",
+        "Number of unexpected condiions that may stall the queue: any value above zero is unexpected."
+    )
+    .expect("failed to define a metric"),
+    remote_errors: register_int_counter_vec!(
+        "pageserver_deletion_queue_remote_errors_total",
+        "Retryable remote I/O errors while executing deletions, for example 503 responses to DeleteObjects",
+        &["op_kind"],
+    )
+    .expect("failed to define a metric")
+}
+});
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum RemoteOpKind {
     Upload,
@@ -916,6 +1067,26 @@ pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("Failed to register tenant_task_events metric")
 });
 
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_start_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls started",
+            &["task"],
+        )
+        .unwrap()
+    });
+
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_finish_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls finished",
+            &["task"],
+        )
+        .unwrap()
+    });
+
 pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_background_loop_period_overrun_count",
@@ -1140,7 +1311,7 @@ pub struct TimelineMetrics {
     pub load_layer_map_histo: StorageTimeMetrics,
     pub garbage_collect_histo: StorageTimeMetrics,
     pub last_record_gauge: IntGauge,
-    pub resident_physical_size_gauge: UIntGauge,
+    resident_physical_size_gauge: UIntGauge,
     /// copy of LayeredTimeline.current_logical_size
     pub current_logical_size_gauge: UIntGauge,
     pub num_persistent_files_created: IntCounter,
@@ -1218,10 +1389,29 @@ impl TimelineMetrics {
     }
 
     pub fn record_new_file_metrics(&self, sz: u64) {
-        self.resident_physical_size_gauge.add(sz);
+        self.resident_physical_size_add(sz);
         self.num_persistent_files_created.inc_by(1);
         self.persistent_bytes_written.inc_by(sz);
     }
+
+    pub fn resident_physical_size_sub(&self, sz: u64) {
+        self.resident_physical_size_gauge.sub(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(sz);
+    }
+
+    pub fn resident_physical_size_add(&self, sz: u64) {
+        self.resident_physical_size_gauge.add(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
+    }
+
+    pub fn resident_physical_size_set(&self, sz: u64) {
+        self.resident_physical_size_gauge.set(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.set(sz);
+    }
+
+    pub fn resident_physical_size_get(&self) -> u64 {
+        self.resident_physical_size_gauge.get()
+    }
 }
 
 impl Drop for TimelineMetrics {
@@ -1229,7 +1419,10 @@ impl Drop for TimelineMetrics {
         let tenant_id = &self.tenant_id;
         let timeline_id = &self.timeline_id;
         let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
+        {
+            RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
+            let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
+        }
         let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
         let _ = NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, timeline_id]);
         let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, timeline_id]);
@@ -1280,10 +1473,46 @@ use std::sync::{Arc, Mutex};
 use std::task::{Context, Poll};
 use std::time::{Duration, Instant};
 
+use crate::context::{PageContentKind, RequestContext};
+use crate::task_mgr::TaskKind;
+
+/// Maintain a per timeline gauge in addition to the global gauge.
+struct PerTimelineRemotePhysicalSizeGauge {
+    last_set: u64,
+    gauge: UIntGauge,
+}
+
+impl PerTimelineRemotePhysicalSizeGauge {
+    fn new(per_timeline_gauge: UIntGauge) -> Self {
+        Self {
+            last_set: per_timeline_gauge.get(),
+            gauge: per_timeline_gauge,
+        }
+    }
+    fn set(&mut self, sz: u64) {
+        self.gauge.set(sz);
+        if sz < self.last_set {
+            REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set - sz);
+        } else {
+            REMOTE_PHYSICAL_SIZE_GLOBAL.add(sz - self.last_set);
+        };
+        self.last_set = sz;
+    }
+    fn get(&self) -> u64 {
+        self.gauge.get()
+    }
+}
+
+impl Drop for PerTimelineRemotePhysicalSizeGauge {
+    fn drop(&mut self) {
+        REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set);
+    }
+}
+
 pub struct RemoteTimelineClientMetrics {
     tenant_id: String,
     timeline_id: String,
-    remote_physical_size_gauge: Mutex<Option<UIntGauge>>,
+    remote_physical_size_gauge: Mutex<Option<PerTimelineRemotePhysicalSizeGauge>>,
     calls_unfinished_gauge: Mutex<HashMap<(&'static str, &'static str), IntGauge>>,
     bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
     bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
@@ -1301,18 +1530,24 @@ impl RemoteTimelineClientMetrics {
         }
     }
 
-    pub fn remote_physical_size_gauge(&self) -> UIntGauge {
+    pub(crate) fn remote_physical_size_set(&self, sz: u64) {
         let mut guard = self.remote_physical_size_gauge.lock().unwrap();
-        guard
-            .get_or_insert_with(|| {
+        let gauge = guard.get_or_insert_with(|| {
+            PerTimelineRemotePhysicalSizeGauge::new(
                 REMOTE_PHYSICAL_SIZE
                     .get_metric_with_label_values(&[
                         &self.tenant_id.to_string(),
                         &self.timeline_id.to_string(),
                     ])
-                    .unwrap()
-            })
-            .clone()
+                    .unwrap(),
+            )
+        });
+        gauge.set(sz);
+    }
+
+    pub(crate) fn remote_physical_size_get(&self) -> u64 {
+        let guard = self.remote_physical_size_gauge.lock().unwrap();
+        guard.as_ref().map(|gauge| gauge.get()).unwrap_or(0)
     }
 
     pub fn remote_operation_time(
@@ -1651,6 +1886,9 @@ pub fn preinitialize_metrics() {
         Lazy::force(c);
     });
 
+    // Deletion queue stats
+    Lazy::force(&DELETION_QUEUE);
+
     // countervecs
     [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
         .into_iter()
@@ -1664,7 +1902,6 @@ pub fn preinitialize_metrics() {
     // histograms
     [
         &READ_NUM_FS_LAYERS,
-        &RECONSTRUCT_TIME,
         &WAIT_LSN_TIME,
         &WAL_REDO_TIME,
         &WAL_REDO_WAIT_TIME,
@@ -1675,4 +1912,7 @@ pub fn preinitialize_metrics() {
     .for_each(|h| {
         Lazy::force(h);
     });
+
+    // Custom
+    Lazy::force(&RECONSTRUCT_TIME);
 }

pageserver/src/page_cache.rs

@@ -66,8 +66,7 @@
 //! 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
+//! callers with PageWriteGuard: 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.
 //!
@@ -75,7 +74,11 @@
 use std::{
     collections::{hash_map::Entry, HashMap},
     convert::TryInto,
-    sync::atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
+    sync::{
+        atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
+        Arc, Weak,
+    },
+    time::Duration,
 };
 
 use anyhow::Context;
@@ -85,7 +88,7 @@ use utils::{
     lsn::Lsn,
 };
 
-use crate::{metrics::PageCacheSizeMetrics, repository::Key};
+use crate::{context::RequestContext, metrics::PageCacheSizeMetrics, repository::Key};
 
 static PAGE_CACHE: OnceCell<PageCache> = OnceCell::new();
 const TEST_PAGE_CACHE_SIZE: usize = 50;
@@ -165,6 +168,8 @@ struct Slot {
 
 struct SlotInner {
     key: Option<CacheKey>,
+    // for `coalesce_readers_permit`
+    permit: std::sync::Mutex<Weak<PinnedSlotsPermit>>,
     buf: &'static mut [u8; PAGE_SZ],
 }
 
@@ -207,6 +212,22 @@ impl Slot {
     }
 }
 
+impl SlotInner {
+    /// If there is aready a reader, drop our permit and share its permit, just like we share read access.
+    fn coalesce_readers_permit(&self, permit: PinnedSlotsPermit) -> Arc<PinnedSlotsPermit> {
+        let mut guard = self.permit.lock().unwrap();
+        if let Some(existing_permit) = guard.upgrade() {
+            drop(guard);
+            drop(permit);
+            existing_permit
+        } else {
+            let permit = Arc::new(permit);
+            *guard = Arc::downgrade(&permit);
+            permit
+        }
+    }
+}
+
 pub struct PageCache {
     /// This contains the mapping from the cache key to buffer slot that currently
     /// contains the page, if any.
@@ -224,6 +245,8 @@ pub struct PageCache {
     /// The actual buffers with their metadata.
     slots: Box<[Slot]>,
 
+    pinned_slots: Arc<tokio::sync::Semaphore>,
+
     /// Index of the next candidate to evict, for the Clock replacement algorithm.
     /// This is interpreted modulo the page cache size.
     next_evict_slot: AtomicUsize,
@@ -231,23 +254,28 @@ pub struct PageCache {
     size_metrics: &'static PageCacheSizeMetrics,
 }
 
+struct PinnedSlotsPermit(tokio::sync::OwnedSemaphorePermit);
+
 ///
 /// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
 /// until the guard is dropped.
 ///
-pub struct PageReadGuard<'i>(tokio::sync::RwLockReadGuard<'i, SlotInner>);
+pub struct PageReadGuard<'i> {
+    _permit: Arc<PinnedSlotsPermit>,
+    slot_guard: tokio::sync::RwLockReadGuard<'i, SlotInner>,
+}
 
 impl std::ops::Deref for PageReadGuard<'_> {
     type Target = [u8; PAGE_SZ];
 
     fn deref(&self) -> &Self::Target {
-        self.0.buf
+        self.slot_guard.buf
     }
 }
 
 impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
     fn as_ref(&self) -> &[u8; PAGE_SZ] {
-        self.0.buf
+        self.slot_guard.buf
     }
 }
 
@@ -257,21 +285,25 @@ impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
 ///
 /// 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.
-///
+/// is expected to fill in the page contents and call mark_valid().
 pub struct PageWriteGuard<'i> {
-    inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,
+    state: PageWriteGuardState<'i>,
+}
 
-    // Are the page contents currently valid?
-    // Used to mark pages as invalid that are assigned but not yet filled with data.
-    valid: bool,
+enum PageWriteGuardState<'i> {
+    Invalid {
+        inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,
+        _permit: PinnedSlotsPermit,
+    },
+    Downgraded,
 }
 
 impl std::ops::DerefMut for PageWriteGuard<'_> {
     fn deref_mut(&mut self) -> &mut Self::Target {
-        self.inner.buf
+        match &mut self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => inner.buf,
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
@@ -279,25 +311,28 @@ impl std::ops::Deref for PageWriteGuard<'_> {
     type Target = [u8; PAGE_SZ];
 
     fn deref(&self) -> &Self::Target {
-        self.inner.buf
+        match &self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => inner.buf,
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
-impl AsMut<[u8; PAGE_SZ]> for PageWriteGuard<'_> {
-    fn as_mut(&mut self) -> &mut [u8; PAGE_SZ] {
-        self.inner.buf
-    }
-}
-
-impl PageWriteGuard<'_> {
+impl<'a> PageWriteGuard<'a> {
     /// 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;
+    #[must_use]
+    pub fn mark_valid(mut self) -> PageReadGuard<'a> {
+        let prev = std::mem::replace(&mut self.state, PageWriteGuardState::Downgraded);
+        match prev {
+            PageWriteGuardState::Invalid { inner, _permit } => {
+                assert!(inner.key.is_some());
+                PageReadGuard {
+                    _permit: Arc::new(_permit),
+                    slot_guard: inner.downgrade(),
+                }
+            }
+            PageWriteGuardState::Downgraded => unreachable!(),
+        }
     }
 }
 
@@ -308,11 +343,14 @@ impl Drop for PageWriteGuard<'_> {
     /// 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;
+        match &mut self.state {
+            PageWriteGuardState::Invalid { inner, _permit } => {
+                assert!(inner.key.is_some());
+                let self_key = inner.key.as_ref().unwrap();
+                PAGE_CACHE.get().unwrap().remove_mapping(self_key);
+                inner.key = None;
+            }
+            PageWriteGuardState::Downgraded => {}
         }
     }
 }
@@ -323,12 +361,6 @@ pub enum ReadBufResult<'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
@@ -346,8 +378,14 @@ impl PageCache {
         timeline_id: TimelineId,
         key: &Key,
         lsn: Lsn,
+        ctx: &RequestContext,
     ) -> Option<(Lsn, PageReadGuard)> {
+        let Ok(permit) = self.try_get_pinned_slot_permit().await else {
+            return None;
+        };
+
         crate::metrics::PAGE_CACHE
+            .for_ctx(ctx)
             .read_accesses_materialized_page
             .inc();
 
@@ -360,7 +398,10 @@ impl PageCache {
             lsn,
         };
 
-        if let Some(guard) = self.try_lock_for_read(&mut cache_key).await {
+        if let Some(guard) = self
+            .try_lock_for_read(&mut cache_key, &mut Some(permit))
+            .await
+        {
             if let CacheKey::MaterializedPage {
                 hash_key: _,
                 lsn: available_lsn,
@@ -368,10 +409,12 @@ impl PageCache {
             {
                 if available_lsn == lsn {
                     crate::metrics::PAGE_CACHE
+                        .for_ctx(ctx)
                         .read_hits_materialized_page_exact
                         .inc();
                 } else {
                     crate::metrics::PAGE_CACHE
+                        .for_ctx(ctx)
                         .read_hits_materialized_page_older_lsn
                         .inc();
                 }
@@ -404,20 +447,77 @@ impl PageCache {
             lsn,
         };
 
-        match self.lock_for_write(&cache_key).await? {
-            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);
+        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
+        loop {
+            // First check if the key already exists in the cache.
+            if let Some(slot_idx) = self.search_mapping_exact(&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().await;
+                if inner.key.as_ref() == Some(&cache_key) {
+                    slot.inc_usage_count();
+                    debug_assert!(
+                        {
+                            let guard = inner.permit.lock().unwrap();
+                            guard.upgrade().is_none()
+                        },
+                        "we hold a write lock, so, no one else should have a permit"
+                    );
+                    debug_assert_eq!(inner.buf.len(), img.len());
+                    // 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!(inner.buf == img);
+                    return Ok(());
+                }
             }
-            WriteBufResult::NotFound(mut write_guard) => {
-                write_guard.copy_from_slice(img);
-                write_guard.mark_valid();
-            }
-        }
+            debug_assert!(permit.is_some());
 
-        Ok(())
+            // Not found. Find a victim buffer
+            let (slot_idx, mut inner) = self
+                .find_victim(permit.as_ref().unwrap())
+                .await
+                .context("Failed to find evict 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());
+            slot.set_usage_count(1);
+            // Create a write guard for the slot so we go through the expected motions.
+            debug_assert!(
+                {
+                    let guard = inner.permit.lock().unwrap();
+                    guard.upgrade().is_none()
+                },
+                "we hold a write lock, so, no one else should have a permit"
+            );
+            let mut write_guard = PageWriteGuard {
+                state: PageWriteGuardState::Invalid {
+                    _permit: permit.take().unwrap(),
+                    inner,
+                },
+            };
+            write_guard.copy_from_slice(img);
+            let _ = write_guard.mark_valid();
+            return Ok(());
+        }
     }
 
     // Section 1.2: Public interface functions for working with immutable file pages.
@@ -426,10 +526,11 @@ impl PageCache {
         &self,
         file_id: FileId,
         blkno: u32,
+        ctx: &RequestContext,
     ) -> anyhow::Result<ReadBufResult> {
         let mut cache_key = CacheKey::ImmutableFilePage { file_id, blkno };
 
-        self.lock_for_read(&mut cache_key).await
+        self.lock_for_read(&mut cache_key, ctx).await
     }
 
     //
@@ -440,6 +541,29 @@ impl PageCache {
     // "mappings" after this section. But the routines in this section should
     // not require changes.
 
+    async fn try_get_pinned_slot_permit(&self) -> anyhow::Result<PinnedSlotsPermit> {
+        let timer = crate::metrics::PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME.start_timer();
+        match tokio::time::timeout(
+            // Choose small timeout, neon_smgr does its own retries.
+            // https://neondb.slack.com/archives/C04DGM6SMTM/p1694786876476869
+            Duration::from_secs(10),
+            Arc::clone(&self.pinned_slots).acquire_owned(),
+        )
+        .await
+        {
+            Ok(res) => Ok(PinnedSlotsPermit(
+                res.expect("this semaphore is never closed"),
+            )),
+            Err(_timeout) => {
+                timer.stop_and_discard();
+                crate::metrics::page_cache_errors_inc(
+                    crate::metrics::PageCacheErrorKind::AcquirePinnedSlotTimeout,
+                );
+                anyhow::bail!("timeout: there were page guards alive for all page cache slots")
+            }
+        }
+    }
+
     /// Look up a page in the cache.
     ///
     /// If the search criteria is not exact, *cache_key is updated with the key
@@ -449,7 +573,11 @@ impl PageCache {
     ///
     /// If no page is found, returns None and *cache_key is left unmodified.
     ///
-    async fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
+    async fn try_lock_for_read(
+        &self,
+        cache_key: &mut CacheKey,
+        permit: &mut Option<PinnedSlotsPermit>,
+    ) -> 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
@@ -459,7 +587,10 @@ impl PageCache {
             let inner = slot.inner.read().await;
             if inner.key.as_ref() == Some(cache_key) {
                 slot.inc_usage_count();
-                return Some(PageReadGuard(inner));
+                return Some(PageReadGuard {
+                    _permit: inner.coalesce_readers_permit(permit.take().unwrap()),
+                    slot_guard: inner,
+                });
             } else {
                 // search_mapping might have modified the search key; restore it.
                 *cache_key = cache_key_orig;
@@ -497,14 +628,22 @@ impl PageCache {
     /// }
     /// ```
     ///
-    async fn lock_for_read(&self, cache_key: &mut CacheKey) -> anyhow::Result<ReadBufResult> {
+    async fn lock_for_read(
+        &self,
+        cache_key: &mut CacheKey,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ReadBufResult> {
+        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
+
         let (read_access, hit) = match cache_key {
             CacheKey::MaterializedPage { .. } => {
                 unreachable!("Materialized pages use lookup_materialized_page")
             }
             CacheKey::ImmutableFilePage { .. } => (
-                &crate::metrics::PAGE_CACHE.read_accesses_immutable,
-                &crate::metrics::PAGE_CACHE.read_hits_immutable,
+                &crate::metrics::PAGE_CACHE
+                    .for_ctx(ctx)
+                    .read_accesses_immutable,
+                &crate::metrics::PAGE_CACHE.for_ctx(ctx).read_hits_immutable,
             ),
         };
         read_access.inc();
@@ -512,17 +651,21 @@ impl PageCache {
         let mut is_first_iteration = true;
         loop {
             // First check if the key already exists in the cache.
-            if let Some(read_guard) = self.try_lock_for_read(cache_key).await {
+            if let Some(read_guard) = self.try_lock_for_read(cache_key, &mut permit).await {
+                debug_assert!(permit.is_none());
                 if is_first_iteration {
                     hit.inc();
                 }
                 return Ok(ReadBufResult::Found(read_guard));
             }
+            debug_assert!(permit.is_some());
             is_first_iteration = false;
 
             // Not found. Find a victim buffer
-            let (slot_idx, mut inner) =
-                self.find_victim().context("Failed to find evict victim")?;
+            let (slot_idx, mut inner) = self
+                .find_victim(permit.as_ref().unwrap())
+                .await
+                .context("Failed to find evict victim")?;
 
             // Insert mapping for this. At this point, we may find that another
             // thread did the same thing concurrently. In that case, we evicted
@@ -544,70 +687,19 @@ impl PageCache {
             inner.key = Some(cache_key.clone());
             slot.set_usage_count(1);
 
+            debug_assert!(
+                {
+                    let guard = inner.permit.lock().unwrap();
+                    guard.upgrade().is_none()
+                },
+                "we hold a write lock, so, no one else should have a permit"
+            );
+
             return Ok(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".
-    async 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().await;
-            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.
-    async fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
-        loop {
-            // First check if the key already exists in the cache.
-            if let Some(write_guard) = self.try_lock_for_write(cache_key).await {
-                return Ok(WriteBufResult::Found(write_guard));
-            }
-
-            // Not found. Find a victim buffer
-            let (slot_idx, mut inner) =
-                self.find_victim().context("Failed to find evict 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());
-            slot.set_usage_count(1);
-
-            return Ok(WriteBufResult::NotFound(PageWriteGuard {
-                inner,
-                valid: false,
+                state: PageWriteGuardState::Invalid {
+                    _permit: permit.take().unwrap(),
+                    inner,
+                },
             }));
         }
     }
@@ -652,7 +744,7 @@ impl PageCache {
     ///
     /// Like 'search_mapping, but performs an "exact" search. Used for
     /// allocating a new buffer.
-    fn search_mapping_for_write(&self, key: &CacheKey) -> Option<usize> {
+    fn search_mapping_exact(&self, key: &CacheKey) -> Option<usize> {
         match key {
             CacheKey::MaterializedPage { hash_key, lsn } => {
                 let map = self.materialized_page_map.read().unwrap();
@@ -758,7 +850,10 @@ impl PageCache {
     /// Find a slot to evict.
     ///
     /// On return, the slot is empty and write-locked.
-    fn find_victim(&self) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
+    async fn find_victim(
+        &self,
+        _permit_witness: &PinnedSlotsPermit,
+    ) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
         let iter_limit = self.slots.len() * 10;
         let mut iters = 0;
         loop {
@@ -771,13 +866,40 @@ impl PageCache {
                 let mut inner = match slot.inner.try_write() {
                     Ok(inner) => inner,
                     Err(_err) => {
-                        // If we have looped through the whole buffer pool 10 times
-                        // and still haven't found a victim buffer, something's wrong.
-                        // Maybe all the buffers were in locked. That could happen in
-                        // theory, if you have more threads holding buffers locked than
-                        // there are buffers in the pool. In practice, with a reasonably
-                        // large buffer pool it really shouldn't happen.
                         if iters > iter_limit {
+                            // NB: Even with the permits, there's no hard guarantee that we will find a slot with
+                            // any particular number of iterations: other threads might race ahead and acquire and
+                            // release pins just as we're scanning the array.
+                            //
+                            // Imagine that nslots is 2, and as starting point, usage_count==1 on all
+                            // slots. There are two threads running concurrently, A and B. A has just
+                            // acquired the permit from the semaphore.
+                            //
+                            //   A: Look at slot 1. Its usage_count == 1, so decrement it to zero, and continue the search
+                            //   B: Acquire permit.
+                            //   B: Look at slot 2, decrement its usage_count to zero and continue the search
+                            //   B: Look at slot 1. Its usage_count is zero, so pin it and bump up its usage_count to 1.
+                            //   B: Release pin and permit again
+                            //   B: Acquire permit.
+                            //   B: Look at slot 2. Its usage_count is zero, so pin it and bump up its usage_count to 1.
+                            //   B: Release pin and permit again
+                            //
+                            // Now we're back in the starting situation that both slots have
+                            // usage_count 1, but A has now been through one iteration of the
+                            // find_victim() loop. This can repeat indefinitely and on each
+                            // iteration, A's iteration count increases by one.
+                            //
+                            // So, even though the semaphore for the permits is fair, the victim search
+                            // itself happens in parallel and is not fair.
+                            // Hence even with a permit, a task can theoretically be starved.
+                            // To avoid this, we'd need tokio to give priority to tasks that are holding
+                            // permits for longer.
+                            // Note that just yielding to tokio during iteration without such
+                            // priority boosting is likely counter-productive. We'd just give more opportunities
+                            // for B to bump usage count, further starving A.
+                            crate::metrics::page_cache_errors_inc(
+                                crate::metrics::PageCacheErrorKind::EvictIterLimit,
+                            );
                             anyhow::bail!("exceeded evict iter limit");
                         }
                         continue;
@@ -788,6 +910,7 @@ impl PageCache {
                     self.remove_mapping(old_key);
                     inner.key = None;
                 }
+                crate::metrics::PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL.inc_by(iters as u64);
                 return Ok((slot_idx, inner));
             }
         }
@@ -815,7 +938,11 @@ impl PageCache {
                 let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();
 
                 Slot {
-                    inner: tokio::sync::RwLock::new(SlotInner { key: None, buf }),
+                    inner: tokio::sync::RwLock::new(SlotInner {
+                        key: None,
+                        buf,
+                        permit: std::sync::Mutex::new(Weak::new()),
+                    }),
                     usage_count: AtomicU8::new(0),
                 }
             })
@@ -827,6 +954,7 @@ impl PageCache {
             slots,
             next_evict_slot: AtomicUsize::new(0),
             size_metrics,
+            pinned_slots: Arc::new(tokio::sync::Semaphore::new(num_pages)),
         }
     }
 }