fix

## Problem

Safekeeper and pageserver metrics collection might time out. We've seen
this in both hadron and neon.

## Summary of changes

This PR moves metrics collection in PS/SK to the background so that we
will always get some metrics, despite there may be some delays. Will
leave it to the future work to reduce metrics collection time.

---------

Co-authored-by: Chen Luo <chen.luo@databricks.com>

.config/hakari.toml

@@ -33,6 +33,7 @@ workspace-members = [
     "compute_api",
     "consumption_metrics",
     "desim",
+    "json",
     "metrics",
     "pageserver_api",
     "postgres_backend",

.dockerignore

@@ -4,6 +4,7 @@
 !Cargo.lock
 !Cargo.toml
 !Makefile
+!postgres.mk
 !rust-toolchain.toml
 !scripts/ninstall.sh
 !docker-compose/run-tests.sh

.github/actionlint.yml

@@ -7,6 +7,7 @@ self-hosted-runner:
     - small-metal
     - small-arm64
     - unit-perf
+    - unit-perf-aws-arm
     - us-east-2
 config-variables:
   - AWS_ECR_REGION

.github/workflows/build-macos.yml

@@ -32,161 +32,14 @@ permissions:
   contents: read
 
 jobs:
-  build-pgxn:
-    if: |
-      inputs.pg_versions != '[]' || inputs.rebuild_everything ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos')  ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
-      github.ref_name == 'main'
-    timeout-minutes: 30
-    runs-on: macos-15
-    strategy:
-      matrix:
-        postgres-version: ${{ inputs.rebuild_everything && fromJSON('["v14", "v15", "v16", "v17"]') || fromJSON(inputs.pg_versions) }}
-    env:
-      # Use release build only, to have less debug info around
-      # Hence keeping target/ (and general cache size) smaller
-      BUILD_TYPE: release
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - name: Checkout main repo
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set pg ${{ matrix.postgres-version }} for caching
-        id: pg_rev
-        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-${{ matrix.postgres-version }}) | tee -a "${GITHUB_OUTPUT}"
-
-      - name: Cache postgres ${{ matrix.postgres-version }} build
-        id: cache_pg
-        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
-        with:
-          path: pg_install/${{ matrix.postgres-version }}
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ matrix.postgres-version }}-${{ steps.pg_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
-
-      - name: Checkout submodule vendor/postgres-${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          git submodule init vendor/postgres-${{ matrix.postgres-version }}
-          git submodule update --depth 1 --recursive
-
-      - name: Install build dependencies
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Build Postgres ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make postgres-${{ matrix.postgres-version }} -j$(sysctl -n hw.ncpu)
-
-      - name: Build Neon Pg Ext ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make "neon-pg-ext-${{ matrix.postgres-version }}" -j$(sysctl -n hw.ncpu)
-
-      - name: Get postgres headers ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make postgres-headers-${{ matrix.postgres-version }} -j$(sysctl -n hw.ncpu)
-
-      - name: Upload "pg_install/${{ matrix.postgres-version }}" artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: pg_install--${{ matrix.postgres-version }}
-          path: pg_install/${{ matrix.postgres-version }}
-          # The artifact is supposed to be used by the next job in the same workflow,
-          # so there’s no need to store it for too long.
-          retention-days: 1
-
-  build-walproposer-lib:
-    if: |
-      contains(inputs.pg_versions, 'v17') || inputs.rebuild_everything ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos')  ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
-      github.ref_name == 'main'
-    timeout-minutes: 30
-    runs-on: macos-15
-    needs: [build-pgxn]
-    env:
-      # Use release build only, to have less debug info around
-      # Hence keeping target/ (and general cache size) smaller
-      BUILD_TYPE: release
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - name: Checkout main repo
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set pg v17 for caching
-        id: pg_rev
-        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v17) | tee -a "${GITHUB_OUTPUT}"
-
-      - name: Download "pg_install/v17" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v17
-          path: pg_install/v17
-
-      - name: Cache walproposer-lib
-        id: cache_walproposer_lib
-        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
-        with:
-          path: build/walproposer-lib
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-walproposer_lib-v17-${{ steps.pg_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
-
-      - name: Checkout submodule vendor/postgres-v17
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          git submodule init vendor/postgres-v17
-          git submodule update --depth 1 --recursive
-
-      - name: Install build dependencies
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Build walproposer-lib (only for v17)
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run:
-          make walproposer-lib -j$(sysctl -n hw.ncpu)
-
-      - name: Upload "build/walproposer-lib" artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: build--walproposer-lib
-          path: build/walproposer-lib
-          # The artifact is supposed to be used by the next job in the same workflow,
-          # so there’s no need to store it for too long.
-          retention-days: 1
-
-  cargo-build:
+  make-all:
     if: |
       inputs.pg_versions != '[]' || inputs.rebuild_rust_code || inputs.rebuild_everything ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos') ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
       github.ref_name == 'main'
-    timeout-minutes: 30
+    timeout-minutes: 60
     runs-on: macos-15
-    needs: [build-pgxn, build-walproposer-lib]
     env:
       # Use release build only, to have less debug info around
       # Hence keeping target/ (and general cache size) smaller
@@ -202,41 +55,53 @@ jobs:
         with:
           submodules: true
 
-      - name: Download "pg_install/v14" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v14
-          path: pg_install/v14
-
-      - name: Download "pg_install/v15" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v15
-          path: pg_install/v15
-
-      - name: Download "pg_install/v16" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v16
-          path: pg_install/v16
-
-      - name: Download "pg_install/v17" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v17
-          path: pg_install/v17
-
-      - name: Download "build/walproposer-lib" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: build--walproposer-lib
-          path: build/walproposer-lib
-
-      # `actions/download-artifact` doesn't preserve permissions:
-      # https://github.com/actions/download-artifact?tab=readme-ov-file#permission-loss
-      - name: Make pg_install/v*/bin/* executable
+      - name: Install build dependencies
         run: |
-          chmod +x pg_install/v*/bin/*
+          brew install flex bison openssl protobuf icu4c
+
+      - name: Set extra env for macOS
+        run: |
+          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
+          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
+
+      - name: Restore "pg_install/" cache
+        id: cache_pg
+        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
+        with:
+          path: pg_install
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-install-v14-${{ hashFiles('Makefile', 'postgres.mk', 'vendor/revisions.json') }}
+
+      - name: Checkout vendor/postgres submodules
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          git submodule init
+          git submodule update --depth 1 --recursive
+
+      - name: Build Postgres
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          make postgres -j$(sysctl -n hw.ncpu)
+
+      # This isn't strictly necessary, but it makes the cached and non-cached builds more similar,
+      # When pg_install is restored from cache, there is no 'build/' directory. By removing it
+      # in a non-cached build too, we enforce that the rest of the steps don't depend on it,
+      # so that we notice any build caching bugs earlier.
+      - name: Remove build artifacts
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          rm -rf build
+
+      # Explicitly update the rust toolchain before running 'make'. The parallel make build can
+      # invoke 'cargo build' more than once in parallel, for different crates.  That's OK, 'cargo'
+      # does its own locking to prevent concurrent builds from stepping on each other's
+      # toes. However, it will first try to update the toolchain, and that step is not locked the
+      # same way. To avoid two toolchain updates running in parallel and stepping on each other's
+      # toes, ensure that the toolchain is up-to-date beforehand.
+      - name: Update rust toolchain
+        run: |
+          rustup --version &&
+          rustup update &&
+          rustup show
 
       - name: Cache cargo deps
         uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
@@ -248,17 +113,12 @@ jobs:
             target
           key: v1-${{ runner.os }}-${{ runner.arch }}-cargo-${{ hashFiles('./Cargo.lock') }}-${{ hashFiles('./rust-toolchain.toml') }}-rust
 
-      - name: Install build dependencies
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Run cargo build
-        run: cargo build --all --release -j$(sysctl -n hw.ncpu)
+      # Build the neon-specific postgres extensions, and all the Rust bits.
+      #
+      # Pass PG_INSTALL_CACHED=1 because PostgreSQL was already built and cached
+      # separately.
+      - name: Build all
+        run: PG_INSTALL_CACHED=1 BUILD_TYPE=release make -j$(sysctl -n hw.ncpu) all
 
       - name: Check that no warnings are produced
         run: ./run_clippy.sh

.github/workflows/build_and_test.yml

@@ -69,7 +69,7 @@ jobs:
           submodules: true
 
       - name: Check for file changes
-        uses: step-security/paths-filter@v3
+        uses: dorny/paths-filter@de90cc6fb38fc0963ad72b210f1f284cd68cea36  # v3.0.2
         id: files-changed
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
@@ -87,6 +87,24 @@ jobs:
     uses: ./.github/workflows/build-build-tools-image.yml
     secrets: inherit
 
+  lint-openapi-spec:
+    runs-on: ubuntu-22.04
+    needs: [ meta, check-permissions ]
+    # We do need to run this in `.*-rc-pr` because of hotfixes.
+    if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
+    steps:
+      - name: Harden the runner (Audit all outbound calls)
+        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
+        with:
+          egress-policy: audit
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - run: make lint-openapi-spec
+
   check-codestyle-python:
     needs: [ meta, check-permissions, build-build-tools-image ]
     # No need to run on `main` because we this in the merge queue. We do need to run this in `.*-rc-pr` because of hotfixes.
@@ -306,14 +324,14 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [ self-hosted, unit-perf ]
+    runs-on: [ self-hosted, unit-perf-aws-arm ]
     container:
       image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
       credentials:
         username: ${{ github.actor }}
         password: ${{ secrets.GITHUB_TOKEN }}
       # for changed limits, see comments on `options:` earlier in this file
-      options: --init --shm-size=512mb --ulimit memlock=67108864:67108864
+      options: --init --shm-size=512mb --ulimit memlock=67108864:67108864 --ulimit nofile=65536:65536 --security-opt seccomp=unconfined
     strategy:
       fail-fast: false
       matrix:
@@ -986,6 +1004,7 @@ jobs:
       - name: Verify docker-compose example and test extensions
         timeout-minutes: 60
         env:
+          PARALLEL_COMPUTES: 3
           TAG: >-
             ${{
               needs.meta.outputs.run-kind == 'compute-rc-pr'

.github/workflows/large_oltp_benchmark.yml

@@ -153,7 +153,7 @@ jobs:
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
 
     - name: Benchmark database maintenance
-      if: ${{ matrix.test_maintenance == 'true' }}
+      if: ${{ matrix.test_maintenance }}
       uses: ./.github/actions/run-python-test-set
       with:
         build_type: ${{ env.BUILD_TYPE }}

.github/workflows/neon_extra_builds.yml

@@ -53,7 +53,7 @@ jobs:
           submodules: true
 
       - name: Check for Postgres changes
-        uses: step-security/paths-filter@v3
+        uses: dorny/paths-filter@1441771bbfdd59dcd748680ee64ebd8faab1a242  #v3
         id: files_changed
         with:
           token: ${{ github.token }}

.github/workflows/periodic_pagebench.yml

@@ -1,4 +1,4 @@
-name: Periodic pagebench performance test on unit-perf hetzner runner
+name: Periodic pagebench performance test on unit-perf-aws-arm runners
 
 on:
   schedule:
@@ -40,7 +40,7 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [ self-hosted, unit-perf ]
+    runs-on: [ self-hosted, unit-perf-aws-arm ]
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm
       credentials:

.github/workflows/pre-merge-checks.yml

@@ -34,7 +34,7 @@ jobs:
 
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
-      - uses: step-security/changed-files@3dbe17c78367e7d60f00d78ae6781a35be47b4a1 # v45.0.1
+      - uses: tj-actions/changed-files@ed68ef82c095e0d48ec87eccea555d944a631a4c # v46.0.5
         id: python-src
         with:
           files: |
@@ -45,7 +45,7 @@ jobs:
             poetry.lock
             pyproject.toml
 
-      - uses: step-security/changed-files@3dbe17c78367e7d60f00d78ae6781a35be47b4a1 # v45.0.1
+      - uses: tj-actions/changed-files@ed68ef82c095e0d48ec87eccea555d944a631a4c # v46.0.5
         id: rust-src
         with:
           files: |

.github/workflows/proxy-benchmark.yml

@@ -0,0 +1,84 @@
+name: Periodic proxy performance test on unit-perf-aws-arm runners
+
+on:
+  push: # TODO: remove after testing
+    branches:
+      - test-proxy-bench # Runs on pushes to branches starting with test-proxy-bench
+  # schedule:
+    # * is a special character in YAML so you have to quote this string
+    #        ┌───────────── minute (0 - 59)
+    #        │ ┌───────────── hour (0 - 23)
+    #        │ │ ┌───────────── day of the month (1 - 31)
+    #        │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #        │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    # - cron: '0 5 * * *' # Runs at 5 UTC once a day
+  workflow_dispatch: # adds an ability to run this manually
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  run_periodic_proxybench_test:
+    permissions:
+      id-token: write # aws-actions/configure-aws-credentials
+      statuses: write
+      contents: write
+      pull-requests: write
+    runs-on: [self-hosted, unit-perf-aws-arm]
+    timeout-minutes: 60  # 1h timeout
+    container:
+      image: ghcr.io/neondatabase/build-tools:pinned-bookworm
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+    steps:
+    - name: Checkout proxy-bench Repo
+      uses: actions/checkout@v4
+      with:
+        repository: neondatabase/proxy-bench
+        path: proxy-bench
+
+    - name: Set up the environment which depends on $RUNNER_TEMP on nvme drive
+      id: set-env
+      shell: bash -euxo pipefail {0}
+      run: |
+        PROXY_BENCH_PATH=$(realpath ./proxy-bench)
+        {
+          echo "PROXY_BENCH_PATH=$PROXY_BENCH_PATH"
+          echo "NEON_DIR=${RUNNER_TEMP}/neon"
+          echo "TEST_OUTPUT=${PROXY_BENCH_PATH}/test_output"
+          echo ""
+        } >> "$GITHUB_ENV"
+
+    - name: Run proxy-bench
+      run: ${PROXY_BENCH_PATH}/run.sh
+
+    - name: Ingest Bench Results # neon repo script
+      if: always()
+      run: |
+        mkdir -p $TEST_OUTPUT
+        python $NEON_DIR/scripts/proxy_bench_results_ingest.py --out $TEST_OUTPUT
+
+    - name: Push Metrics to Proxy perf database
+      if: always()
+      env:
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PROXY_TEST_RESULT_CONNSTR }}"
+        REPORT_FROM: $TEST_OUTPUT
+      run: $NEON_DIR/scripts/generate_and_push_perf_report.sh
+
+    - name: Docker cleanup
+      if: always()
+      run: docker compose down
+
+    - name: Notify Failure
+      if: failure()
+      run: echo "Proxy bench job failed" && exit 1

.gitignore

@@ -6,6 +6,7 @@
 /tmp_check_cli
 __pycache__/
 test_output/
+neon_previous/
 .vscode
 .idea
 *.swp
@@ -14,7 +15,7 @@ neon.iml
 /.neon
 /integration_tests/.neon
 compaction-suite-results.*
-pgxn/neon/communicator/communicator_bindings.h
+docker-compose/docker-compose-parallel.yml
 
 # Coverage
 *.profraw

Cargo.toml

@@ -24,6 +24,7 @@ members = [
     "libs/pageserver_api",
     "libs/postgres_ffi",
     "libs/postgres_ffi_types",
+    "libs/postgres_versioninfo",
     "libs/safekeeper_api",
     "libs/desim",
     "libs/neon-shmem",
@@ -34,7 +35,6 @@ members = [
     "libs/pq_proto",
     "libs/tenant_size_model",
     "libs/metrics",
-    "libs/neonart",
     "libs/postgres_connection",
     "libs/remote_storage",
     "libs/tracing-utils",
@@ -43,6 +43,7 @@ members = [
     "libs/walproposer",
     "libs/wal_decoder",
     "libs/postgres_initdb",
+    "libs/proxy/json",
     "libs/proxy/postgres-protocol2",
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
@@ -91,7 +92,6 @@ clap = { version = "4.0", features = ["derive", "env"] }
 clashmap = { version = "1.0", features = ["raw-api"] }
 comfy-table = "7.1"
 const_format = "0.2"
-crossbeam-utils = "0.8.21"
 crc32c = "0.6"
 diatomic-waker = { version = "0.2.3" }
 either = "1.8"
@@ -150,12 +150,12 @@ parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
 pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pem = "3.0.3"
-peekable = "0.3.0"
 pin-project-lite = "0.2"
 pprof = { version = "0.14", features = ["criterion", "flamegraph", "frame-pointer", "prost-codec"] }
 procfs = "0.16"
 prometheus = {version = "0.13", default-features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.13.5"
+prost-types = "0.13.5"
 rand = "0.8"
 redis = { version = "0.29.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
@@ -179,13 +179,13 @@ serde_json = "1"
 serde_path_to_error = "0.1"
 serde_with = { version = "3", features = [ "base64" ] }
 serde_assert = "0.5.0"
+serde_repr = "0.1.20"
 sha2 = "0.10.2"
 signal-hook = "0.3"
 smallvec = "1.11"
 smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 spki = "0.7.3"
-spin = "0.9.8"
 strum = "0.26"
 strum_macros = "0.26"
 "subtle"  = "2.5.0"
@@ -197,6 +197,7 @@ thiserror = "1.0"
 tikv-jemallocator = { version = "0.6", features = ["profiling", "stats", "unprefixed_malloc_on_supported_platforms"] }
 tikv-jemalloc-ctl = { version = "0.6", features = ["stats"] }
 tokio = { version = "1.43.1", features = ["macros"] }
+tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.12.0"
 tokio-rustls = { version = "0.26.0", default-features = false, features = ["tls12", "ring"]}
@@ -238,9 +239,6 @@ x509-cert = { version = "0.2.5" }
 env_logger = "0.11"
 log = "0.4"
 
-tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
-uring-common = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
-
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
 postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
@@ -260,18 +258,17 @@ desim = { version = "0.1", path = "./libs/desim" }
 endpoint_storage = { version = "0.0.1", path = "./endpoint_storage/" }
 http-utils = { version = "0.1", path = "./libs/http-utils/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
-neonart = { version = "0.1", path = "./libs/neonart/" }
 neon-shmem = { version = "0.1", path = "./libs/neon-shmem/" }
 pageserver = { path = "./pageserver" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
 pageserver_client = { path = "./pageserver/client" }
-pageserver_client_grpc = { path = "./pageserver/client_grpc" }
 pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
 pageserver_page_api = { path = "./pageserver/page_api" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
 postgres_connection = { version = "0.1", path = "./libs/postgres_connection/" }
 postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
 postgres_ffi_types = { version = "0.1", path = "./libs/postgres_ffi_types/" }
+postgres_versioninfo = { version = "0.1", path = "./libs/postgres_versioninfo/" }
 postgres_initdb = { path = "./libs/postgres_initdb" }
 posthog_client_lite = { version = "0.1", path = "./libs/posthog_client_lite" }
 pq_proto = { version = "0.1", path = "./libs/pq_proto/" }
@@ -291,7 +288,7 @@ walproposer = { version = "0.1", path = "./libs/walproposer/" }
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
 
 ## Build dependencies
-cbindgen = "0.28.0"
+cbindgen = "0.29.0"
 criterion = "0.5.1"
 rcgen = "0.13"
 rstest = "0.18"

Dockerfile

@@ -30,7 +30,18 @@ ARG BASE_IMAGE_SHA=debian:${DEBIAN_FLAVOR}
 ARG BASE_IMAGE_SHA=${BASE_IMAGE_SHA/debian:bookworm-slim/debian@$BOOKWORM_SLIM_SHA}
 ARG BASE_IMAGE_SHA=${BASE_IMAGE_SHA/debian:bullseye-slim/debian@$BULLSEYE_SLIM_SHA}
 
-# Build Postgres
+# Naive way:
+#
+# 1. COPY . .
+# 1. make neon-pg-ext
+# 2. cargo build <storage binaries>
+#
+# But to enable docker to cache intermediate layers, we perform a few preparatory steps:
+#
+# - Build all postgres versions, depending on just the contents of vendor/
+# - Use cargo chef to build all rust dependencies
+
+# 1. Build all postgres versions
 FROM $REPOSITORY/$IMAGE:$TAG AS pg-build
 WORKDIR /home/nonroot
 
@@ -38,16 +49,15 @@ COPY --chown=nonroot vendor/postgres-v14 vendor/postgres-v14
 COPY --chown=nonroot vendor/postgres-v15 vendor/postgres-v15
 COPY --chown=nonroot vendor/postgres-v16 vendor/postgres-v16
 COPY --chown=nonroot vendor/postgres-v17 vendor/postgres-v17
-COPY --chown=nonroot pgxn pgxn
 COPY --chown=nonroot Makefile Makefile
+COPY --chown=nonroot postgres.mk postgres.mk
 COPY --chown=nonroot scripts/ninstall.sh scripts/ninstall.sh
 
 ENV BUILD_TYPE=release
 RUN set -e \
-    && mold -run make -j $(nproc) -s neon-pg-ext \
-    && tar -C pg_install -czf /home/nonroot/postgres_install.tar.gz .
+    && mold -run make -j $(nproc) -s postgres
 
-# Prepare cargo-chef recipe
+# 2. Prepare cargo-chef recipe
 FROM $REPOSITORY/$IMAGE:$TAG AS plan
 WORKDIR /home/nonroot
 
@@ -55,23 +65,22 @@ COPY --chown=nonroot . .
 
 RUN cargo chef prepare --recipe-path recipe.json
 
-# Build neon binaries
+# Main build image
 FROM $REPOSITORY/$IMAGE:$TAG AS build
 WORKDIR /home/nonroot
 ARG GIT_VERSION=local
 ARG BUILD_TAG
-
-COPY --from=pg-build /home/nonroot/pg_install/v14/include/postgresql/server pg_install/v14/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v15/include/postgresql/server pg_install/v15/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v16/include/postgresql/server pg_install/v16/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v17/include/postgresql/server pg_install/v17/include/postgresql/server
-COPY --from=plan     /home/nonroot/recipe.json                              recipe.json
-
 ARG ADDITIONAL_RUSTFLAGS=""
 
+# 3. Build cargo dependencies. Note that this step doesn't depend on anything else than
+# `recipe.json`, so the layer can be reused as long as none of the dependencies change.
+COPY --from=plan     /home/nonroot/recipe.json                              recipe.json
 RUN set -e \
     && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo chef cook --locked --release --recipe-path recipe.json
 
+# Perform the main build. We reuse the Postgres build artifacts from the intermediate 'pg-build'
+# layer, and the cargo dependencies built in the previous step.
+COPY --chown=nonroot --from=pg-build /home/nonroot/pg_install/ pg_install
 COPY --chown=nonroot . .
 
 RUN set -e \
@@ -86,10 +95,10 @@ RUN set -e \
       --bin endpoint_storage \
       --bin neon_local \
       --bin storage_scrubber \
-      --locked --release
+      --locked --release \
+    && mold -run make -j $(nproc) -s neon-pg-ext
 
-# Build final image
-#
+# Assemble the final image
 FROM $BASE_IMAGE_SHA
 WORKDIR /data
 
@@ -129,12 +138,15 @@ COPY --from=build --chown=neon:neon /home/nonroot/target/release/proxy
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/endpoint_storage    /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/neon_local          /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_scrubber    /usr/local/bin
+COPY --from=build /home/nonroot/pg_install/v14 /usr/local/v14/
+COPY --from=build /home/nonroot/pg_install/v15 /usr/local/v15/
+COPY --from=build /home/nonroot/pg_install/v16 /usr/local/v16/
+COPY --from=build /home/nonroot/pg_install/v17 /usr/local/v17/
 
-COPY --from=pg-build /home/nonroot/pg_install/v14 /usr/local/v14/
-COPY --from=pg-build /home/nonroot/pg_install/v15 /usr/local/v15/
-COPY --from=pg-build /home/nonroot/pg_install/v16 /usr/local/v16/
-COPY --from=pg-build /home/nonroot/pg_install/v17 /usr/local/v17/
-COPY --from=pg-build /home/nonroot/postgres_install.tar.gz /data/
+# Deprecated: Old deployment scripts use this tarball which contains all the Postgres binaries.
+# That's obsolete, since all the same files are also present under /usr/local/v*. But to keep the
+# old scripts working for now, create the tarball.
+RUN tar -C /usr/local -cvzf /data/postgres_install.tar.gz v14 v15 v16 v17
 
 # By default, pageserver uses `.neon/` working directory in WORKDIR, so create one and fill it with the dummy config.
 # Now, when `docker run ... pageserver` is run, it can start without errors, yet will have some default dummy values.

Makefile

@@ -4,11 +4,14 @@ ROOT_PROJECT_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 # managers.
 POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install/
 
+# Supported PostgreSQL versions
+POSTGRES_VERSIONS = v17 v16 v15 v14
+
 # CARGO_BUILD_FLAGS: Extra flags to pass to `cargo build`. `--locked`
 # and `--features testing` are popular examples.
 #
-# CARGO_PROFILE: You can also set to override the cargo profile to
-# use. By default, it is derived from BUILD_TYPE.
+# CARGO_PROFILE: Set to override the cargo profile to use. By default,
+# it is derived from BUILD_TYPE.
 
 # All intermediate build artifacts are stored here.
 BUILD_DIR := build
@@ -102,13 +105,13 @@ CACHEDIR_TAG_CONTENTS := "Signature: 8a477f597d28d172789f06886806bc55"
 # Top level Makefile to build Neon and PostgreSQL
 #
 .PHONY: all
-all: neon postgres neon-pg-ext
+all: neon postgres-install neon-pg-ext
 
 ### Neon Rust bits
 #
-# The 'postgres_ffi' depends on the Postgres headers.
+# The 'postgres_ffi' crate depends on the Postgres headers.
 .PHONY: neon
-neon: postgres-headers walproposer-lib cargo-target-dir
+neon: postgres-headers-install walproposer-lib cargo-target-dir
 	+@echo "Compiling Neon"
 	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS) $(CARGO_PROFILE)
 
@@ -118,78 +121,8 @@ cargo-target-dir:
 	mkdir -p target
 	test -e target/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > target/CACHEDIR.TAG
 
-### PostgreSQL parts
-# Some rules are duplicated for Postgres v14 and 15. We may want to refactor
-# to avoid the duplication in the future, but it's tolerable for now.
-#
-$(BUILD_DIR)/%/config.status:
-	mkdir -p $(BUILD_DIR)
-	test -e $(BUILD_DIR)/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > $(BUILD_DIR)/CACHEDIR.TAG
-
-	+@echo "Configuring Postgres $* build"
-	@test -s $(ROOT_PROJECT_DIR)/vendor/postgres-$*/configure || { \
-		echo "\nPostgres submodule not found in $(ROOT_PROJECT_DIR)/vendor/postgres-$*/, execute "; \
-		echo "'git submodule update --init --recursive --depth 2 --progress .' in project root.\n"; \
-		exit 1; }
-	mkdir -p $(BUILD_DIR)/$*
-
-	VERSION=$*; \
-	EXTRA_VERSION=$$(cd $(ROOT_PROJECT_DIR)/vendor/postgres-$$VERSION && git rev-parse HEAD); \
-	(cd $(BUILD_DIR)/$$VERSION && \
-	env PATH="$(EXTRA_PATH_OVERRIDES):$$PATH" $(ROOT_PROJECT_DIR)/vendor/postgres-$$VERSION/configure \
-		CFLAGS='$(PG_CFLAGS)' LDFLAGS='$(PG_LDFLAGS)' \
-		$(PG_CONFIGURE_OPTS) --with-extra-version=" ($$EXTRA_VERSION)" \
-		--prefix=$(abspath $(POSTGRES_INSTALL_DIR))/$$VERSION > configure.log)
-
-# nicer alias to run 'configure'
-# Note: I've been unable to use templates for this part of our configuration.
-# I'm not sure why it wouldn't work, but this is the only place (apart from
-# the "build-all-versions" entry points) where direct mention of PostgreSQL
-# versions is used.
-.PHONY: postgres-configure-v17
-postgres-configure-v17: $(BUILD_DIR)/v17/config.status
-.PHONY: postgres-configure-v16
-postgres-configure-v16: $(BUILD_DIR)/v16/config.status
-.PHONY: postgres-configure-v15
-postgres-configure-v15: $(BUILD_DIR)/v15/config.status
-.PHONY: postgres-configure-v14
-postgres-configure-v14: $(BUILD_DIR)/v14/config.status
-
-# Install the PostgreSQL header files into $(POSTGRES_INSTALL_DIR)/<version>/include
-.PHONY: postgres-headers-%
-postgres-headers-%: postgres-configure-%
-	+@echo "Installing PostgreSQL $* headers"
-	$(MAKE) -C $(BUILD_DIR)/$*/src/include MAKELEVEL=0 install
-
-# Compile and install PostgreSQL
-.PHONY: postgres-%
-postgres-%: postgres-configure-% \
-		  postgres-headers-% # to prevent `make install` conflicts with neon's `postgres-headers`
-	+@echo "Compiling PostgreSQL $*"
-	$(MAKE) -C $(BUILD_DIR)/$* MAKELEVEL=0 install
-	+@echo "Compiling libpq $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/src/interfaces/libpq install
-	+@echo "Compiling pg_prewarm $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_prewarm install
-	+@echo "Compiling pg_buffercache $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_buffercache install
-	+@echo "Compiling pg_visibility $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_visibility install
-	+@echo "Compiling pageinspect $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pageinspect install
-	+@echo "Compiling pg_trgm $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_trgm install
-	+@echo "Compiling amcheck $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/amcheck install
-	+@echo "Compiling test_decoding $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/test_decoding install
-
-.PHONY: postgres-check-%
-postgres-check-%: postgres-%
-	$(MAKE) -C $(BUILD_DIR)/$* MAKELEVEL=0 check
-
 .PHONY: neon-pg-ext-%
-neon-pg-ext-%: postgres-% cargo-target-dir
+neon-pg-ext-%: postgres-install-% cargo-target-dir
 	+@echo "Compiling neon-specific Postgres extensions for $*"
 	mkdir -p $(BUILD_DIR)/pgxn-$*
 	$(MAKE) PG_CONFIG="$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config" COPT='$(COPT)' \
@@ -231,39 +164,14 @@ ifeq ($(UNAME_S),Linux)
 		pg_crc32c.o
 endif
 
+# Shorthand to call neon-pg-ext-% target for all Postgres versions
 .PHONY: neon-pg-ext
-neon-pg-ext: \
-	neon-pg-ext-v14 \
-	neon-pg-ext-v15 \
-	neon-pg-ext-v16 \
-	neon-pg-ext-v17
-
-# shorthand to build all Postgres versions
-.PHONY: postgres
-postgres: \
-	postgres-v14 \
-	postgres-v15 \
-	postgres-v16 \
-	postgres-v17
-
-.PHONY: postgres-headers
-postgres-headers: \
-	postgres-headers-v14 \
-	postgres-headers-v15 \
-	postgres-headers-v16 \
-	postgres-headers-v17
-
-.PHONY: postgres-check
-postgres-check: \
-	postgres-check-v14 \
-	postgres-check-v15 \
-	postgres-check-v16 \
-	postgres-check-v17
+neon-pg-ext: $(foreach pg_version,$(POSTGRES_VERSIONS),neon-pg-ext-$(pg_version))
 
 # This removes everything
 .PHONY: distclean
 distclean:
-	$(RM) -r $(POSTGRES_INSTALL_DIR)
+	$(RM) -r $(POSTGRES_INSTALL_DIR) $(BUILD_DIR)
 	$(CARGO_CMD_PREFIX) cargo clean
 
 .PHONY: fmt
@@ -311,3 +219,28 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 .PHONY: setup-pre-commit-hook
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit
+
+.PHONY: lint-openapi-spec
+lint-openapi-spec:
+	# operation-2xx-response: pageserver timeline delete returns 404 on success
+	find . -iname "openapi_spec.y*ml" -exec\
+		docker run --rm -v ${PWD}:/spec ghcr.io/redocly/cli:1.34.4\
+			--skip-rule=operation-operationId --skip-rule=operation-summary --extends=minimal\
+			--skip-rule=no-server-example.com --skip-rule=operation-2xx-response\
+			lint {} \+
+
+# Targets for building PostgreSQL are defined in postgres.mk.
+#
+# But if the caller has indicated that PostgreSQL is already
+# installed, by setting the PG_INSTALL_CACHED variable, skip it.
+ifdef PG_INSTALL_CACHED
+postgres-install: skip-install
+$(foreach pg_version,$(POSTGRES_VERSIONS),postgres-install-$(pg_version)): skip-install
+postgres-headers-install:
+	+@echo "Skipping installation of PostgreSQL headers because PG_INSTALL_CACHED is set"
+skip-install:
+	+@echo "Skipping PostgreSQL installation because PG_INSTALL_CACHED is set"
+
+else
+include postgres.mk
+endif

build-tools.Dockerfile

@@ -165,6 +165,7 @@ RUN curl -fsSL \
     && rm sql_exporter.tar.gz
 
 # protobuf-compiler (protoc)
+# Keep the version the same as in compute/compute-node.Dockerfile
 ENV PROTOC_VERSION=25.1
 RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
     && unzip -q protoc.zip -d protoc \
@@ -179,7 +180,7 @@ RUN curl -sL "https://github.com/peak/s5cmd/releases/download/v${S5CMD_VERSION}/
     && mv s5cmd /usr/local/bin/s5cmd
 
 # LLVM
-ENV LLVM_VERSION=19
+ENV LLVM_VERSION=20
 RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && echo "deb http://apt.llvm.org/${DEBIAN_VERSION}/ llvm-toolchain-${DEBIAN_VERSION}-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
     && apt update \
@@ -292,7 +293,7 @@ WORKDIR /home/nonroot
 
 # Rust
 # Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
-ENV RUSTC_VERSION=1.87.0
+ENV RUSTC_VERSION=1.88.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
 ARG RUSTFILT_VERSION=0.2.1

clippy.toml

@@ -1,9 +1,12 @@
 disallowed-methods = [
     "tokio::task::block_in_place",
+
     # Allow this for now, to deny it later once we stop using Handle::block_on completely
     # "tokio::runtime::Handle::block_on",
-    # use tokio_epoll_uring_ext instead
-    "tokio_epoll_uring::thread_local_system",
+
+    # tokio-epoll-uring:
+    # - allow-invalid because the method doesn't exist on macOS
+    { path = "tokio_epoll_uring::thread_local_system", replacement = "tokio_epoll_uring_ext module inside pageserver crate", allow-invalid = true }
 ]
 
 disallowed-macros = [

compute/Makefile

@@ -22,7 +22,7 @@ sql_exporter.yml: $(jsonnet_files)
 		--output-file etc/$@ \
 		--tla-str collector_name=neon_collector \
 		--tla-str collector_file=neon_collector.yml \
-		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter' \
+		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter&pgaudit.log=none' \
 		etc/sql_exporter.jsonnet
 
 sql_exporter_autoscaling.yml: $(jsonnet_files)
@@ -30,7 +30,7 @@ sql_exporter_autoscaling.yml: $(jsonnet_files)
 		--output-file etc/$@ \
 		--tla-str collector_name=neon_collector_autoscaling \
 		--tla-str collector_file=neon_collector_autoscaling.yml \
-		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter_autoscaling' \
+		--tla-str 'connection_string=postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter_autoscaling&pgaudit.log=none' \
 		etc/sql_exporter.jsonnet
 
 .PHONY: clean

compute/compute-node.Dockerfile

@@ -115,6 +115,9 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION
 
+# Keep in sync with build-tools.Dockerfile
+ENV PROTOC_VERSION=25.1
+
 # Use strict mode for bash to catch errors early
 SHELL ["/bin/bash", "-euo", "pipefail", "-c"]
 
@@ -149,8 +152,14 @@ RUN case $DEBIAN_VERSION in \
     libclang-dev \
     jsonnet \
     $VERSION_INSTALLS \
-    && apt clean && rm -rf /var/lib/apt/lists/* && \
-    useradd -ms /bin/bash nonroot -b /home
+    && apt clean && rm -rf /var/lib/apt/lists/* \
+    && useradd -ms /bin/bash nonroot -b /home \
+    # Install protoc from binary release, since Debian's versions are too old.
+    && curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
+    && unzip -q protoc.zip -d protoc \
+    && mv protoc/bin/protoc /usr/local/bin/protoc \
+    && mv protoc/include/google /usr/local/include/google \
+    && rm -rf protoc.zip protoc
 
 #########################################################################################
 #
@@ -171,9 +180,6 @@ RUN cd postgres && \
     eval $CONFIGURE_CMD && \
     make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s install && \
     make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s -C contrib/ install && \
-    # Install headers
-    make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s -C src/include install && \
-    make MAKELEVEL=0 -j $(getconf _NPROCESSORS_ONLN) -s -C src/interfaces/libpq install && \
     # Enable some of contrib extensions
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/autoinc.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/dblink.control && \
@@ -1173,7 +1179,7 @@ COPY --from=pgrag-src /ext-src/ /ext-src/
 # Install it using virtual environment, because Python 3.11 (the default version on Debian 12 (Bookworm)) complains otherwise
 WORKDIR /ext-src/onnxruntime-src
 RUN apt update && apt install --no-install-recommends --no-install-suggests -y \
-    python3 python3-pip python3-venv protobuf-compiler && \
+    python3 python3-pip python3-venv && \
     apt clean && rm -rf /var/lib/apt/lists/* && \
     python3 -m venv venv && \
     . venv/bin/activate && \
@@ -1566,29 +1572,31 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) && \
 FROM build-deps AS pgaudit-src
 ARG PG_VERSION
 WORKDIR /ext-src
+COPY "compute/patches/pgaudit-parallel_workers-${PG_VERSION}.patch" .
 RUN case "${PG_VERSION}" in \
     "v14") \
-    export PGAUDIT_VERSION=1.6.2 \
-    export PGAUDIT_CHECKSUM=1f350d70a0cbf488c0f2b485e3a5c9b11f78ad9e3cbb95ef6904afa1eb3187eb \
+    export PGAUDIT_VERSION=1.6.3 \
+    export PGAUDIT_CHECKSUM=37a8f5a7cc8d9188e536d15cf0fdc457fcdab2547caedb54442c37f124110919 \
     ;; \
     "v15") \
-    export PGAUDIT_VERSION=1.7.0 \
-    export PGAUDIT_CHECKSUM=8f4a73e451c88c567e516e6cba7dc1e23bc91686bb6f1f77f8f3126d428a8bd8 \
+    export PGAUDIT_VERSION=1.7.1 \
+    export PGAUDIT_CHECKSUM=e9c8e6e092d82b2f901d72555ce0fe7780552f35f8985573796cd7e64b09d4ec \
     ;; \
     "v16") \
-    export PGAUDIT_VERSION=16.0 \
-    export PGAUDIT_CHECKSUM=d53ef985f2d0b15ba25c512c4ce967dce07b94fd4422c95bd04c4c1a055fe738 \
+    export PGAUDIT_VERSION=16.1 \
+    export PGAUDIT_CHECKSUM=3bae908ab70ba0c6f51224009dbcfff1a97bd6104c6273297a64292e1b921fee \
     ;; \
     "v17") \
-    export PGAUDIT_VERSION=17.0 \
-    export PGAUDIT_CHECKSUM=7d0d08d030275d525f36cd48b38c6455f1023da863385badff0cec44965bfd8c \
+    export PGAUDIT_VERSION=17.1 \
+    export PGAUDIT_CHECKSUM=9c5f37504d393486cc75d2ced83f75f5899be64fa85f689d6babb833b4361e6c \
     ;; \
     *) \
     echo "pgaudit is not supported on this PostgreSQL version" && exit 1;; \
     esac && \
     wget https://github.com/pgaudit/pgaudit/archive/refs/tags/${PGAUDIT_VERSION}.tar.gz -O pgaudit.tar.gz && \
     echo "${PGAUDIT_CHECKSUM} pgaudit.tar.gz" | sha256sum --check && \
-    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C .
+    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C . && \
+    patch -p1 < "/ext-src/pgaudit-parallel_workers-${PG_VERSION}.patch"
 
 FROM pg-build AS pgaudit-build
 COPY --from=pgaudit-src /ext-src/ /ext-src/
@@ -1628,11 +1636,14 @@ RUN make install USE_PGXS=1 -j $(getconf _NPROCESSORS_ONLN)
 # compile neon extensions
 #
 #########################################################################################
-FROM pg-build AS neon-ext-build
+FROM pg-build-with-cargo AS neon-ext-build
 ARG PG_VERSION
 
-COPY pgxn/ pgxn/
-RUN make -j $(getconf _NPROCESSORS_ONLN) -C pgxn -s install-compute
+USER root
+COPY . .
+
+RUN make -j $(getconf _NPROCESSORS_ONLN) -C pgxn -s install-compute \
+      BUILD_TYPE=release CARGO_BUILD_FLAGS="--locked --release" NEON_CARGO_ARTIFACT_TARGET_DIR="$(pwd)/target/release"
 
 #########################################################################################
 #
@@ -1904,10 +1915,10 @@ RUN cd /ext-src/pg_repack-src && patch -p1 </ext-src/pg_repack.patch && rm -f /e
 
 COPY --chmod=755 docker-compose/run-tests.sh /run-tests.sh
 RUN echo /usr/local/pgsql/lib > /etc/ld.so.conf.d/00-neon.conf && /sbin/ldconfig
-RUN apt-get update && apt-get install -y libtap-parser-sourcehandler-pgtap-perl jq \
+RUN apt-get update && apt-get install -y libtap-parser-sourcehandler-pgtap-perl jq parallel \
    && apt clean && rm -rf /ext-src/*.tar.gz /ext-src/*.patch /var/lib/apt/lists/*
 ENV PATH=/usr/local/pgsql/bin:$PATH
-ENV PGHOST=compute
+ENV PGHOST=compute1
 ENV PGPORT=55433
 ENV PGUSER=cloud_admin
 ENV PGDATABASE=postgres
@@ -1977,7 +1988,7 @@ RUN apt update && \
         locales \
         lsof \
         procps \
-        rsyslog \
+        rsyslog-gnutls \
         screen \
         tcpdump \
         $VERSION_INSTALLS && \

compute/etc/neon_collector.jsonnet

@@ -8,6 +8,8 @@
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
     import 'sql_exporter/compute_logical_snapshots_bytes.libsonnet',
     import 'sql_exporter/compute_max_connections.libsonnet',
+    import 'sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet',
+    import 'sql_exporter/compute_pg_oldest_mxid_age.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',
     import 'sql_exporter/compute_subscriptions_count.libsonnet',
     import 'sql_exporter/connection_counts.libsonnet',

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet

@@ -0,0 +1,13 @@
+{
+  metric_name: 'compute_pg_oldest_frozen_xid_age',
+  type: 'gauge',
+  help: 'Age of oldest XIDs that have not been frozen by VACUUM. An indicator of how long it has been since VACUUM last ran.',
+  key_labels: [
+    'database_name',
+  ],
+  value_label: 'metric',
+  values: [
+    'frozen_xid_age',
+  ],
+  query: importstr 'sql_exporter/compute_pg_oldest_frozen_xid_age.sql',
+}

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.sql

@@ -0,0 +1,4 @@
+SELECT datname database_name,
+  age(datfrozenxid) frozen_xid_age
+FROM pg_database
+ORDER BY frozen_xid_age DESC LIMIT 10;

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.libsonnet

@@ -0,0 +1,13 @@
+{
+  metric_name: 'compute_pg_oldest_mxid_age',
+  type: 'gauge',
+  help: 'Age of oldest MXIDs that have not been replaced by VACUUM. An indicator of how long it has been since VACUUM last ran.',
+  key_labels: [
+    'database_name',
+  ],
+  value_label: 'metric',
+  values: [
+    'min_mxid_age',
+  ],
+  query: importstr 'sql_exporter/compute_pg_oldest_mxid_age.sql',
+}

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.sql

@@ -0,0 +1,4 @@
+SELECT datname database_name,
+  mxid_age(datminmxid) min_mxid_age
+FROM pg_database
+ORDER BY min_mxid_age DESC LIMIT 10;

compute/patches/anon_v2.patch

@@ -1,8 +1,8 @@
 diff --git a/sql/anon.sql b/sql/anon.sql
-index 0cdc769..f6cc950 100644
+index 0cdc769..b450327 100644
 --- a/sql/anon.sql
 +++ b/sql/anon.sql
-@@ -1141,3 +1141,8 @@ $$
+@@ -1141,3 +1141,15 @@ $$
  -- TODO : https://en.wikipedia.org/wiki/L-diversity
  
  -- TODO : https://en.wikipedia.org/wiki/T-closeness
@@ -11,6 +11,13 @@ index 0cdc769..f6cc950 100644
 +
 +GRANT ALL ON SCHEMA anon to neon_superuser;
 +GRANT ALL ON ALL TABLES IN SCHEMA anon TO neon_superuser;
++
++DO $$
++BEGIN
++    IF current_setting('server_version_num')::int >= 150000 THEN
++        GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO neon_superuser;
++    END IF;
++END $$;
 diff --git a/sql/init.sql b/sql/init.sql
 index 7da6553..9b6164b 100644
 --- a/sql/init.sql

compute/patches/pgaudit-parallel_workers-v14.patch

@@ -0,0 +1,143 @@
+commit 7220bb3a3f23fa27207d77562dcc286f9a123313
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index baa8011..a601375 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2563,6 +2563,37 @@ COMMIT;
+ NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
+ DROP TABLE part_test;
+ NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 5e6fd38..ac9ded2 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index cc1374a..1870a60 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1612,6 +1612,36 @@ COMMIT;
+ 
+ DROP TABLE part_test;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v15.patch

@@ -0,0 +1,143 @@
+commit 29dc2847f6255541992f18faf8a815dfab79631a
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index b22560b..73f0327 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2563,6 +2563,37 @@ COMMIT;
+ NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
+ DROP TABLE part_test;
+ NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 5e6fd38..ac9ded2 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index 8052426..7f0667b 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1612,6 +1612,36 @@ COMMIT;
+ 
+ DROP TABLE part_test;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v16.patch

@@ -0,0 +1,143 @@
+commit cc708dde7ef2af2a8120d757102d2e34c0463a0f
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index 8772054..9b66ac6 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2556,6 +2556,37 @@ DROP SERVER fdw_server;
+ NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server;,<not logged>
+ DROP EXTENSION postgres_fdw;
+ NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 004d1f9..f061164 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1339,7 +1340,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1420,7 +1421,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1475,7 +1476,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1495,7 +1496,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index 6aae88b..de6d7fd 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1631,6 +1631,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
+ DROP SERVER fdw_server;
+ DROP EXTENSION postgres_fdw;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v17.patch

@@ -0,0 +1,143 @@
+commit 8d02e4c6c5e1e8676251b0717a46054267091cb4
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index d696287..4b1059a 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2568,6 +2568,37 @@ DROP SERVER fdw_server;
+ NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server,<not logged>
+ DROP EXTENSION postgres_fdw;
+ NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 1764af1..0e48875 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1406,7 +1407,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit event onto the stack */
+         stackItem = stack_push();
+@@ -1489,7 +1490,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1544,7 +1545,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1564,7 +1565,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index e161f01..c873098 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1637,6 +1637,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
+ DROP SERVER fdw_server;
+ DROP EXTENSION postgres_fdw;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/vm-image-spec-bookworm.yaml

@@ -26,7 +26,7 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
   - name: pgbouncer-exporter
     user: postgres
     sysvInitAction: respawn
@@ -59,7 +59,7 @@ files:
       # the rules use ALL as the hostname. Avoid the pointless lookups and the "unable to
       # resolve host" log messages that they generate.
       Defaults !fqdn
-      
+
       # Allow postgres user (which is what compute_ctl runs as) to run /neonvm/bin/resize-swap
       # and /neonvm/bin/set-disk-quota as root without requiring entering a password (NOPASSWD),
       # regardless of hostname (ALL)

compute/vm-image-spec-bullseye.yaml

@@ -26,7 +26,7 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
   - name: pgbouncer-exporter
     user: postgres
     sysvInitAction: respawn
@@ -59,7 +59,7 @@ files:
       # the rules use ALL as the hostname. Avoid the pointless lookups and the "unable to
       # resolve host" log messages that they generate.
       Defaults !fqdn
-      
+
       # Allow postgres user (which is what compute_ctl runs as) to run /neonvm/bin/resize-swap
       # and /neonvm/bin/set-disk-quota as root without requiring entering a password (NOPASSWD),
       # regardless of hostname (ALL)

compute_tools/Cargo.toml

@@ -27,6 +27,7 @@ fail.workspace = true
 flate2.workspace = true
 futures.workspace = true
 http.workspace = true
+hostname-validator = "1.1"
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
@@ -65,7 +66,8 @@ url.workspace = true
 uuid.workspace = true
 walkdir.workspace = true
 x509-cert.workspace = true
-
+postgres-types.workspace = true
+postgres_versioninfo.workspace = true
 postgres_initdb.workspace = true
 compute_api.workspace = true
 utils.workspace = true

compute_tools/src/bin/compute_ctl.rs

@@ -36,6 +36,8 @@
 use std::ffi::OsString;
 use std::fs::File;
 use std::process::exit;
+use std::sync::Arc;
+use std::sync::atomic::AtomicU64;
 use std::sync::mpsc;
 use std::thread;
 use std::time::Duration;
@@ -190,7 +192,9 @@ fn main() -> Result<()> {
             cgroup: cli.cgroup,
             #[cfg(target_os = "linux")]
             vm_monitor_addr: cli.vm_monitor_addr,
-            installed_extensions_collection_interval: cli.installed_extensions_collection_interval,
+            installed_extensions_collection_interval: Arc::new(AtomicU64::new(
+                cli.installed_extensions_collection_interval,
+            )),
         },
         config,
     )?;

compute_tools/src/bin/fast_import.rs

@@ -29,7 +29,7 @@ use anyhow::{Context, bail};
 use aws_config::BehaviorVersion;
 use camino::{Utf8Path, Utf8PathBuf};
 use clap::{Parser, Subcommand};
-use compute_tools::extension_server::{PostgresMajorVersion, get_pg_version};
+use compute_tools::extension_server::get_pg_version;
 use nix::unistd::Pid;
 use std::ops::Not;
 use tracing::{Instrument, error, info, info_span, warn};
@@ -179,12 +179,8 @@ impl PostgresProcess {
             .await
             .context("create pgdata directory")?;
 
-        let pg_version = match get_pg_version(self.pgbin.as_ref()) {
-            PostgresMajorVersion::V14 => 14,
-            PostgresMajorVersion::V15 => 15,
-            PostgresMajorVersion::V16 => 16,
-            PostgresMajorVersion::V17 => 17,
-        };
+        let pg_version = get_pg_version(self.pgbin.as_ref());
+
         postgres_initdb::do_run_initdb(postgres_initdb::RunInitdbArgs {
             superuser: initdb_user,
             locale: DEFAULT_LOCALE, // XXX: this shouldn't be hard-coded,
@@ -486,10 +482,8 @@ async fn cmd_pgdata(
     };
 
     let superuser = "cloud_admin";
-    let destination_connstring = format!(
-        "host=localhost port={} user={} dbname=neondb",
-        pg_port, superuser
-    );
+    let destination_connstring =
+        format!("host=localhost port={pg_port} user={superuser} dbname=neondb");
 
     let pgdata_dir = workdir.join("pgdata");
     let mut proc = PostgresProcess::new(pgdata_dir.clone(), pg_bin_dir.clone(), pg_lib_dir.clone());

compute_tools/src/bin/fast_import/s3_uri.rs

@@ -69,7 +69,7 @@ impl clap::builder::TypedValueParser for S3Uri {
         S3Uri::from_str(value_str).map_err(|e| {
             clap::Error::raw(
                 clap::error::ErrorKind::InvalidValue,
-                format!("Failed to parse S3 URI: {}", e),
+                format!("Failed to parse S3 URI: {e}"),
             )
         })
     }

compute_tools/src/catalog.rs

@@ -22,7 +22,7 @@ pub async fn get_dbs_and_roles(compute: &Arc<ComputeNode>) -> anyhow::Result<Cat
 
     spawn(async move {
         if let Err(e) = connection.await {
-            eprintln!("connection error: {}", e);
+            eprintln!("connection error: {e}");
         }
     });
 
@@ -119,7 +119,7 @@ pub async fn get_database_schema(
         _ => {
             let mut lines = stderr_reader.lines();
             if let Some(line) = lines.next_line().await? {
-                if line.contains(&format!("FATAL:  database \"{}\" does not exist", dbname)) {
+                if line.contains(&format!("FATAL:  database \"{dbname}\" does not exist")) {
                     return Err(SchemaDumpError::DatabaseDoesNotExist);
                 }
                 warn!("pg_dump stderr: {}", line)

compute_tools/src/compute.rs

@@ -1,12 +1,12 @@
-use anyhow::{Context, Result, anyhow};
+use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
 use compute_api::privilege::Privilege;
 use compute_api::responses::{
     ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus, LfcOffloadState,
-    LfcPrewarmState, TlsConfig,
+    LfcPrewarmState, PromoteState, TlsConfig,
 };
 use compute_api::spec::{
-    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PgIdent,
+    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
 };
 use futures::StreamExt;
 use futures::future::join_all;
@@ -15,29 +15,28 @@ use itertools::Itertools;
 use nix::sys::signal::{Signal, kill};
 use nix::unistd::Pid;
 use once_cell::sync::Lazy;
-use pageserver_page_api as page_api;
+use pageserver_page_api::{self as page_api, BaseBackupCompression};
 use postgres;
 use postgres::NoTls;
 use postgres::error::SqlState;
 use remote_storage::{DownloadError, RemotePath};
 use std::collections::{HashMap, HashSet};
-use std::net::SocketAddr;
 use std::os::unix::fs::{PermissionsExt, symlink};
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
-use std::sync::atomic::{AtomicU32, Ordering};
+use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
 use std::sync::{Arc, Condvar, Mutex, RwLock};
 use std::time::{Duration, Instant};
 use std::{env, fs};
-use tokio::spawn;
-use tokio_util::io::StreamReader;
+use tokio::{spawn, sync::watch, task::JoinHandle, time};
 use tracing::{Instrument, debug, error, info, instrument, warn};
 use url::Url;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 use utils::measured_stream::MeasuredReader;
 use utils::pid_file;
+use utils::shard::{ShardCount, ShardIndex, ShardNumber};
 
 use crate::configurator::launch_configurator;
 use crate::disk_quota::set_disk_quota;
@@ -71,6 +70,7 @@ pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
         .unwrap_or(BUILD_TAG_DEFAULT)
         .to_string()
 });
+const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
 
 /// Static configuration params that don't change after startup. These mostly
 /// come from the CLI args, or are derived from them.
@@ -104,9 +104,11 @@ pub struct ComputeNodeParams {
     pub remote_ext_base_url: Option<Url>,
 
     /// Interval for installed extensions collection
-    pub installed_extensions_collection_interval: u64,
+    pub installed_extensions_collection_interval: Arc<AtomicU64>,
 }
 
+type TaskHandle = Mutex<Option<JoinHandle<()>>>;
+
 /// Compute node info shared across several `compute_ctl` threads.
 pub struct ComputeNode {
     pub params: ComputeNodeParams,
@@ -127,6 +129,10 @@ pub struct ComputeNode {
     // key: ext_archive_name, value: started download time, download_completed?
     pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
     pub compute_ctl_config: ComputeCtlConfig,
+
+    /// Handle to the extension stats collection task
+    extension_stats_task: TaskHandle,
+    lfc_offload_task: TaskHandle,
 }
 
 // store some metrics about download size that might impact startup time
@@ -168,6 +174,7 @@ pub struct ComputeState {
     /// WAL flush LSN that is set after terminating Postgres and syncing safekeepers if
     /// mode == ComputeMode::Primary. None otherwise
     pub terminate_flush_lsn: Option<Lsn>,
+    pub promote_state: Option<watch::Receiver<PromoteState>>,
 
     pub metrics: ComputeMetrics,
 }
@@ -185,6 +192,7 @@ impl ComputeState {
             lfc_prewarm_state: LfcPrewarmState::default(),
             lfc_offload_state: LfcOffloadState::default(),
             terminate_flush_lsn: None,
+            promote_state: None,
         }
     }
 
@@ -220,7 +228,8 @@ pub struct ParsedSpec {
     pub pageserver_connstr: String,
     pub safekeeper_connstrings: Vec<String>,
     pub storage_auth_token: Option<String>,
-    pub endpoint_storage_addr: Option<SocketAddr>,
+    /// k8s dns name and port
+    pub endpoint_storage_addr: Option<String>,
     pub endpoint_storage_token: Option<String>,
 }
 
@@ -252,8 +261,7 @@ impl ParsedSpec {
             // duplicate entry?
             if current == previous {
                 return Err(format!(
-                    "duplicate entry in safekeeper_connstrings: {}!",
-                    current,
+                    "duplicate entry in safekeeper_connstrings: {current}!",
                 ));
             }
 
@@ -316,13 +324,10 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
                 .or(Err("invalid timeline id"))?
         };
 
-        let endpoint_storage_addr: Option<SocketAddr> = spec
+        let endpoint_storage_addr: Option<String> = spec
             .endpoint_storage_addr
             .clone()
-            .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"))
-            .unwrap_or_default()
-            .parse()
-            .ok();
+            .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"));
         let endpoint_storage_token = spec
             .endpoint_storage_token
             .clone()
@@ -368,7 +373,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
 
 struct PostgresHandle {
     postgres: std::process::Child,
-    log_collector: tokio::task::JoinHandle<Result<()>>,
+    log_collector: JoinHandle<Result<()>>,
 }
 
 impl PostgresHandle {
@@ -382,7 +387,7 @@ struct StartVmMonitorResult {
     #[cfg(target_os = "linux")]
     token: tokio_util::sync::CancellationToken,
     #[cfg(target_os = "linux")]
-    vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
+    vm_monitor: Option<JoinHandle<Result<()>>>,
 }
 
 impl ComputeNode {
@@ -408,11 +413,11 @@ impl ComputeNode {
         // that can affect `compute_ctl` and prevent it from properly configuring the database schema.
         // Unset them via connection string options before connecting to the database.
         // N.B. keep it in sync with `ZENITH_OPTIONS` in `get_maintenance_client()`.
-        const EXTRA_OPTIONS: &str = "-c role=cloud_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0";
+        const EXTRA_OPTIONS: &str = "-c role=cloud_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0 -c pgaudit.log=none";
         let options = match conn_conf.get_options() {
             // Allow the control plane to override any options set by the
             // compute
-            Some(options) => format!("{} {}", EXTRA_OPTIONS, options),
+            Some(options) => format!("{EXTRA_OPTIONS} {options}"),
             None => EXTRA_OPTIONS.to_string(),
         };
         conn_conf.options(&options);
@@ -420,7 +425,7 @@ impl ComputeNode {
 
         let mut new_state = ComputeState::new();
         if let Some(spec) = config.spec {
-            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow!(msg))?;
+            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
             new_state.pspec = Some(pspec);
         }
 
@@ -432,6 +437,8 @@ impl ComputeNode {
             state_changed: Condvar::new(),
             ext_download_progress: RwLock::new(HashMap::new()),
             compute_ctl_config: config.compute_ctl_config,
+            extension_stats_task: Mutex::new(None),
+            lfc_offload_task: Mutex::new(None),
         })
     }
 
@@ -519,6 +526,9 @@ impl ComputeNode {
             None
         };
 
+        this.terminate_extension_stats_task();
+        this.terminate_lfc_offload_task();
+
         // Terminate the vm_monitor so it releases the file watcher on
         // /sys/fs/cgroup/neon-postgres.
         // Note: the vm-monitor only runs on linux because it requires cgroups.
@@ -755,10 +765,15 @@ impl ComputeNode {
         // Configure and start rsyslog for compliance audit logging
         match pspec.spec.audit_log_level {
             ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
-                let remote_endpoint =
+                let remote_tls_endpoint =
+                    std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
+                let remote_plain_endpoint =
                     std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
-                if remote_endpoint.is_empty() {
-                    anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
+
+                if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
+                    anyhow::bail!(
+                        "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
+                    );
                 }
 
                 let log_directory_path = Path::new(&self.params.pgdata).join("log");
@@ -774,7 +789,8 @@ impl ComputeNode {
                     log_directory_path.clone(),
                     endpoint_id,
                     project_id,
-                    &remote_endpoint,
+                    &remote_plain_endpoint,
+                    &remote_tls_endpoint,
                 )?;
 
                 // Launch a background task to clean up the audit logs
@@ -841,12 +857,15 @@ impl ComputeNode {
         // Log metrics so that we can search for slow operations in logs
         info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
 
-        // Spawn the extension stats background task
         self.spawn_extension_stats_task();
 
         if pspec.spec.autoprewarm {
+            info!("autoprewarming on startup as requested");
             self.prewarm_lfc(None);
         }
+        if let Some(seconds) = pspec.spec.offload_lfc_interval_seconds {
+            self.spawn_lfc_offload_task(Duration::from_secs(seconds.into()));
+        };
         Ok(())
     }
 
@@ -1001,82 +1020,80 @@ impl ComputeNode {
         Ok(())
     }
 
-    // Get basebackup from the libpq connection to pageserver using `connstr` and
-    // unarchive it to `pgdata` directory overriding all its previous content.
+    /// Fetches a basebackup from the Pageserver using the compute state's Pageserver connstring and
+    /// unarchives it to `pgdata` directory, replacing any existing contents.
     #[instrument(skip_all, fields(%lsn))]
     fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
         let spec = compute_state.pspec.as_ref().expect("spec must be set");
-        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
 
-        match Url::parse(shard0_connstr)?.scheme() {
-            "postgres" | "postgresql" => self.try_get_basebackup_libpq(spec, lsn),
-            "grpc" => self.try_get_basebackup_grpc(spec, lsn),
-            scheme => return Err(anyhow!("unknown URL scheme {scheme}")),
-        }
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let started = Instant::now();
+
+        let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
+            PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
+            PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
+        };
+
+        let mut state = self.state.lock().unwrap();
+        state.metrics.pageserver_connect_micros =
+            connected.duration_since(started).as_micros() as u64;
+        state.metrics.basebackup_bytes = size as u64;
+        state.metrics.basebackup_ms = started.elapsed().as_millis() as u64;
+
+        Ok(())
     }
 
-    fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<()> {
-        let start_time = Instant::now();
-
+    /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
+    /// the connection was established, and the (compressed) size of the basebackup.
+    fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
         let shard0_connstr = spec
             .pageserver_connstr
             .split(',')
             .next()
             .unwrap()
             .to_string();
+        let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
+            0 | 1 => ShardIndex::unsharded(),
+            count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
+        };
 
-        let chunks = tokio::runtime::Handle::current().block_on(async move {
-            let mut client = page_api::proto::PageServiceClient::connect(shard0_connstr).await?;
-
-            let req = page_api::proto::GetBaseBackupRequest {
-                lsn: lsn.0,
-                replica: false, // TODO: handle replicas, with LSN 0
-                full: false,
-            };
-            let mut req = tonic::Request::new(req);
-            let metadata = req.metadata_mut();
-            metadata.insert("neon-tenant-id", spec.tenant_id.to_string().parse()?);
-            metadata.insert("neon-timeline-id", spec.timeline_id.to_string().parse()?);
-            metadata.insert("neon-shard-id", "0000".to_string().parse()?); // TODO: shard count
-            if let Some(auth) = spec.storage_auth_token.as_ref() {
-                metadata.insert("authorization", format!("Bearer {auth}").parse()?);
-            }
-
-            let chunks = client.get_base_backup(req).await?.into_inner();
-            anyhow::Ok(chunks)
+        let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
+            let mut client = page_api::Client::connect(
+                shard0_connstr,
+                spec.tenant_id,
+                spec.timeline_id,
+                shard_index,
+                spec.storage_auth_token.clone(),
+                None, // NB: base backups use payload compression
+            )
+            .await?;
+            let connected = Instant::now();
+            let reader = client
+                .get_base_backup(page_api::GetBaseBackupRequest {
+                    lsn: (lsn != Lsn(0)).then_some(lsn),
+                    compression: BaseBackupCompression::Gzip,
+                    replica: spec.spec.mode != ComputeMode::Primary,
+                    full: false,
+                })
+                .await?;
+            anyhow::Ok((reader, connected))
         })?;
-        let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
 
-        // Convert the chunks stream into an AsyncRead
-        let stream_reader = StreamReader::new(
-            chunks.map(|chunk| chunk.map(|c| c.chunk).map_err(std::io::Error::other)),
-        );
+        let mut reader = MeasuredReader::new(tokio_util::io::SyncIoBridge::new(reader));
 
-        // Wrap the AsyncRead into a blocking reader for compatibility with tar::Archive
-        let reader = tokio_util::io::SyncIoBridge::new(stream_reader);
-        let mut measured_reader = MeasuredReader::new(reader);
-        let mut bufreader = std::io::BufReader::new(&mut measured_reader);
-
-        // Read the archive directly from the `CopyOutReader`
-        //
-        // Set `ignore_zeros` so that unpack() reads all the Copy data and
-        // doesn't stop at the end-of-archive marker. Otherwise, if the server
-        // sends an Error after finishing the tarball, we will not notice it.
-        let mut ar = tar::Archive::new(&mut bufreader);
+        // Set `ignore_zeros` so that unpack() reads the entire stream and doesn't just stop at the
+        // end-of-archive marker. If the server errors, the tar::Builder drop handler will write an
+        // end-of-archive marker before the error is emitted, and we would not see the error.
+        let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut reader));
         ar.set_ignore_zeros(true);
         ar.unpack(&self.params.pgdata)?;
 
-        // Report metrics
-        let mut state = self.state.lock().unwrap();
-        state.metrics.pageserver_connect_micros = pageserver_connect_micros;
-        state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
-        state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
-        Ok(())
+        Ok((connected, reader.get_byte_count()))
     }
 
-    fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<()> {
-        let start_time = Instant::now();
-
+    /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
+    /// when the connection was established, and the (compressed) size of the basebackup.
+    fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
         let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
         let mut config = postgres::Config::from_str(shard0_connstr)?;
 
@@ -1097,7 +1114,7 @@ impl ComputeNode {
 
         // Connect to pageserver
         let mut client = config.connect(NoTls)?;
-        let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
+        let connected = Instant::now();
 
         let basebackup_cmd = match lsn {
             Lsn(0) => {
@@ -1134,16 +1151,13 @@ impl ComputeNode {
         // Set `ignore_zeros` so that unpack() reads all the Copy data and
         // doesn't stop at the end-of-archive marker. Otherwise, if the server
         // sends an Error after finishing the tarball, we will not notice it.
+        // The tar::Builder drop handler will write an end-of-archive marker
+        // before emitting the error, and we would not see it otherwise.
         let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
         ar.set_ignore_zeros(true);
         ar.unpack(&self.params.pgdata)?;
 
-        // Report metrics
-        let mut state = self.state.lock().unwrap();
-        state.metrics.pageserver_connect_micros = pageserver_connect_micros;
-        state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
-        state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
-        Ok(())
+        Ok((connected, measured_reader.get_byte_count()))
     }
 
     // Gets the basebackup in a retry loop
@@ -1167,7 +1181,10 @@ impl ComputeNode {
                     return result;
                 }
                 Err(ref e) if attempts < max_attempts => {
-                    warn!("Failed to get basebackup: {e:?} (attempt {attempts}/{max_attempts})");
+                    warn!(
+                        "Failed to get basebackup: {} (attempt {}/{})",
+                        e, attempts, max_attempts
+                    );
                     std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
                     retry_period_ms *= 1.5;
                 }
@@ -1193,7 +1210,7 @@ impl ComputeNode {
         let sk_configs = sk_connstrs.into_iter().map(|connstr| {
             // Format connstr
             let id = connstr.clone();
-            let connstr = format!("postgresql://no_user@{}", connstr);
+            let connstr = format!("postgresql://no_user@{connstr}");
             let options = format!(
                 "-c timeline_id={} tenant_id={}",
                 pspec.timeline_id, pspec.tenant_id
@@ -1556,7 +1573,7 @@ impl ComputeNode {
                 let (mut client, connection) = conf.connect(NoTls).await?;
                 tokio::spawn(async move {
                     if let Err(e) = connection.await {
-                        eprintln!("connection error: {}", e);
+                        eprintln!("connection error: {e}");
                     }
                 });
 
@@ -1677,6 +1694,8 @@ impl ComputeNode {
             tls_config = self.compute_ctl_config.tls.clone();
         }
 
+        self.update_installed_extensions_collection_interval(&spec);
+
         let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
 
         // Merge-apply spec & changes to PostgreSQL state.
@@ -1699,7 +1718,7 @@ impl ComputeNode {
                 Ok((mut client, connection)) => {
                     tokio::spawn(async move {
                         if let Err(e) = connection.await {
-                            eprintln!("connection error: {}", e);
+                            eprintln!("connection error: {e}");
                         }
                     });
                     if let Err(e) = handle_migrations(&mut client).await {
@@ -1741,6 +1760,8 @@ impl ComputeNode {
 
         let tls_config = self.tls_config(&spec);
 
+        self.update_installed_extensions_collection_interval(&spec);
+
         if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
             info!("tuning pgbouncer");
 
@@ -2003,7 +2024,7 @@ impl ComputeNode {
         let (client, connection) = connect_result.unwrap();
         tokio::spawn(async move {
             if let Err(e) = connection.await {
-                eprintln!("connection error: {}", e);
+                eprintln!("connection error: {e}");
             }
         });
         let result = client
@@ -2045,7 +2066,7 @@ LIMIT 100",
             self.params
                 .remote_ext_base_url
                 .as_ref()
-                .ok_or(DownloadError::BadInput(anyhow!(
+                .ok_or(DownloadError::BadInput(anyhow::anyhow!(
                     "Remote extensions storage is not configured",
                 )))?;
 
@@ -2172,7 +2193,7 @@ LIMIT 100",
             db_client
                 .simple_query(&query)
                 .await
-                .with_context(|| format!("Failed to execute query: {}", query))?;
+                .with_context(|| format!("Failed to execute query: {query}"))?;
         }
 
         Ok(())
@@ -2199,7 +2220,7 @@ LIMIT 100",
         let version: Option<ExtVersion> = db_client
             .query_opt(version_query, &[&ext_name])
             .await
-            .with_context(|| format!("Failed to execute query: {}", version_query))?
+            .with_context(|| format!("Failed to execute query: {version_query}"))?
             .map(|row| row.get(0));
 
         // sanitize the inputs as postgres idents.
@@ -2214,14 +2235,14 @@ LIMIT 100",
             db_client
                 .simple_query(&query)
                 .await
-                .with_context(|| format!("Failed to execute query: {}", query))?;
+                .with_context(|| format!("Failed to execute query: {query}"))?;
         } else {
             let query =
                 format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
             db_client
                 .simple_query(&query)
                 .await
-                .with_context(|| format!("Failed to execute query: {}", query))?;
+                .with_context(|| format!("Failed to execute query: {query}"))?;
         }
 
         Ok(ext_version)
@@ -2241,7 +2262,7 @@ LIMIT 100",
         let remote_extensions = spec
             .remote_extensions
             .as_ref()
-            .ok_or(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();
@@ -2345,24 +2366,84 @@ LIMIT 100",
     }
 
     pub fn spawn_extension_stats_task(&self) {
+        self.terminate_extension_stats_task();
+
         let conf = self.tokio_conn_conf.clone();
-        let installed_extensions_collection_interval =
-            self.params.installed_extensions_collection_interval;
-        tokio::spawn(async move {
-            // An initial sleep is added to ensure that two collections don't happen at the same time.
-            // The first collection happens during compute startup.
-            tokio::time::sleep(tokio::time::Duration::from_secs(
-                installed_extensions_collection_interval,
-            ))
-            .await;
-            let mut interval = tokio::time::interval(tokio::time::Duration::from_secs(
-                installed_extensions_collection_interval,
-            ));
+        let atomic_interval = self.params.installed_extensions_collection_interval.clone();
+        let mut installed_extensions_collection_interval =
+            2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst);
+        info!(
+            "[NEON_EXT_SPAWN] Spawning background installed extensions worker with Timeout: {}",
+            installed_extensions_collection_interval
+        );
+        let handle = tokio::spawn(async move {
             loop {
-                interval.tick().await;
+                info!(
+                    "[NEON_EXT_INT_SLEEP]: Interval: {}",
+                    installed_extensions_collection_interval
+                );
+                // Sleep at the start of the loop to ensure that two collections don't happen at the same time.
+                // The first collection happens during compute startup.
+                tokio::time::sleep(tokio::time::Duration::from_secs(
+                    installed_extensions_collection_interval,
+                ))
+                .await;
                 let _ = installed_extensions(conf.clone()).await;
+                // Acquire a read lock on the compute spec and then update the interval if necessary
+                installed_extensions_collection_interval = std::cmp::max(
+                    installed_extensions_collection_interval,
+                    2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst),
+                );
             }
         });
+
+        // Store the new task handle
+        *self.extension_stats_task.lock().unwrap() = Some(handle);
+    }
+
+    fn terminate_extension_stats_task(&self) {
+        if let Some(h) = self.extension_stats_task.lock().unwrap().take() {
+            h.abort()
+        }
+    }
+
+    pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
+        self.terminate_lfc_offload_task();
+        let secs = interval.as_secs();
+        info!("spawning lfc offload worker with {secs}s interval");
+        let this = self.clone();
+        let handle = spawn(async move {
+            let mut interval = time::interval(interval);
+            interval.tick().await; // returns immediately
+            loop {
+                interval.tick().await;
+                this.offload_lfc_async().await;
+            }
+        });
+        *self.lfc_offload_task.lock().unwrap() = Some(handle);
+    }
+
+    fn terminate_lfc_offload_task(&self) {
+        if let Some(h) = self.lfc_offload_task.lock().unwrap().take() {
+            h.abort()
+        }
+    }
+
+    fn update_installed_extensions_collection_interval(&self, spec: &ComputeSpec) {
+        // Update the interval for collecting installed extensions statistics
+        // If the value is -1, we never suspend so set the value to default collection.
+        // If the value is 0, it means default, we will just continue to use the default.
+        if spec.suspend_timeout_seconds == -1 || spec.suspend_timeout_seconds == 0 {
+            self.params.installed_extensions_collection_interval.store(
+                DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL,
+                std::sync::atomic::Ordering::SeqCst,
+            );
+        } else {
+            self.params.installed_extensions_collection_interval.store(
+                spec.suspend_timeout_seconds as u64,
+                std::sync::atomic::Ordering::SeqCst,
+            );
+        }
     }
 }
 

compute_tools/src/compute_prewarm.rs

@@ -5,6 +5,7 @@ use compute_api::responses::LfcOffloadState;
 use compute_api::responses::LfcPrewarmState;
 use http::StatusCode;
 use reqwest::Client;
+use std::mem::replace;
 use std::sync::Arc;
 use tokio::{io::AsyncReadExt, spawn};
 use tracing::{error, info};
@@ -88,17 +89,15 @@ impl ComputeNode {
         self.state.lock().unwrap().lfc_offload_state.clone()
     }
 
-    /// Returns false if there is a prewarm request ongoing, true otherwise
+    /// If there is a prewarm request ongoing, return false, true otherwise
     pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
-        crate::metrics::LFC_PREWARM_REQUESTS.inc();
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
-            if let LfcPrewarmState::Prewarming =
-                std::mem::replace(state, LfcPrewarmState::Prewarming)
-            {
+            if let LfcPrewarmState::Prewarming = replace(state, LfcPrewarmState::Prewarming) {
                 return false;
             }
         }
+        crate::metrics::LFC_PREWARMS.inc();
 
         let cloned = self.clone();
         spawn(async move {
@@ -106,7 +105,8 @@ impl ComputeNode {
                 cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Completed;
                 return;
             };
-            error!(%err);
+            crate::metrics::LFC_PREWARM_ERRORS.inc();
+            error!(%err, "prewarming lfc");
             cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Failed {
                 error: err.to_string(),
             };
@@ -152,32 +152,42 @@ impl ComputeNode {
             .map(|_| ())
     }
 
-    /// Returns false if there is an offload request ongoing, true otherwise
+    /// If offload request is ongoing, return false, true otherwise
     pub fn offload_lfc(self: &Arc<Self>) -> bool {
-        crate::metrics::LFC_OFFLOAD_REQUESTS.inc();
         {
             let state = &mut self.state.lock().unwrap().lfc_offload_state;
-            if let LfcOffloadState::Offloading =
-                std::mem::replace(state, LfcOffloadState::Offloading)
-            {
+            if replace(state, LfcOffloadState::Offloading) == LfcOffloadState::Offloading {
                 return false;
             }
         }
-
         let cloned = self.clone();
-        spawn(async move {
-            let Err(err) = cloned.offload_lfc_impl().await else {
-                cloned.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
-                return;
-            };
-            error!(%err);
-            cloned.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
-                error: err.to_string(),
-            };
-        });
+        spawn(async move { cloned.offload_lfc_with_state_update().await });
         true
     }
 
+    pub async fn offload_lfc_async(self: &Arc<Self>) {
+        {
+            let state = &mut self.state.lock().unwrap().lfc_offload_state;
+            if replace(state, LfcOffloadState::Offloading) == LfcOffloadState::Offloading {
+                return;
+            }
+        }
+        self.offload_lfc_with_state_update().await
+    }
+
+    async fn offload_lfc_with_state_update(&self) {
+        crate::metrics::LFC_OFFLOADS.inc();
+        let Err(err) = self.offload_lfc_impl().await else {
+            self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
+            return;
+        };
+        crate::metrics::LFC_OFFLOAD_ERRORS.inc();
+        error!(%err, "offloading lfc");
+        self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
+            error: err.to_string(),
+        };
+    }
+
     async fn offload_lfc_impl(&self) -> Result<()> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
         info!(%url, "requesting LFC state from postgres");

compute_tools/src/compute_promote.rs

@@ -0,0 +1,132 @@
+use crate::compute::ComputeNode;
+use anyhow::{Context, Result, bail};
+use compute_api::{
+    responses::{LfcPrewarmState, PromoteState, SafekeepersLsn},
+    spec::ComputeMode,
+};
+use std::{sync::Arc, time::Duration};
+use tokio::time::sleep;
+use utils::lsn::Lsn;
+
+impl ComputeNode {
+    /// Returns only when promote fails or succeeds. If a network error occurs
+    /// and http client disconnects, this does not stop promotion, and subsequent
+    /// calls block until promote finishes.
+    /// Called by control plane on secondary after primary endpoint is terminated
+    pub async fn promote(self: &Arc<Self>, safekeepers_lsn: SafekeepersLsn) -> PromoteState {
+        let cloned = self.clone();
+        let start_promotion = || {
+            let (tx, rx) = tokio::sync::watch::channel(PromoteState::NotPromoted);
+            tokio::spawn(async move {
+                tx.send(match cloned.promote_impl(safekeepers_lsn).await {
+                    Ok(_) => PromoteState::Completed,
+                    Err(err) => {
+                        tracing::error!(%err, "promoting");
+                        PromoteState::Failed {
+                            error: err.to_string(),
+                        }
+                    }
+                })
+            });
+            rx
+        };
+
+        let mut task;
+        // self.state is unlocked after block ends so we lock it in promote_impl
+        // and task.changed() is reached
+        {
+            task = self
+                .state
+                .lock()
+                .unwrap()
+                .promote_state
+                .get_or_insert_with(start_promotion)
+                .clone()
+        }
+        task.changed().await.expect("promote sender dropped");
+        task.borrow().clone()
+    }
+
+    // Why do we have to supply safekeepers?
+    // For secondary we use primary_connection_conninfo so safekeepers field is empty
+    async fn promote_impl(&self, safekeepers_lsn: SafekeepersLsn) -> Result<()> {
+        {
+            let state = self.state.lock().unwrap();
+            let mode = &state.pspec.as_ref().unwrap().spec.mode;
+            if *mode != ComputeMode::Replica {
+                bail!("{} is not replica", mode.to_type_str());
+            }
+
+            // we don't need to query Postgres so not self.lfc_prewarm_state()
+            match &state.lfc_prewarm_state {
+                LfcPrewarmState::NotPrewarmed | LfcPrewarmState::Prewarming => {
+                    bail!("prewarm not requested or pending")
+                }
+                LfcPrewarmState::Failed { error } => {
+                    tracing::warn!(%error, "replica prewarm failed")
+                }
+                _ => {}
+            }
+        }
+
+        let client = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+            .await
+            .context("connecting to postgres")?;
+
+        let primary_lsn = safekeepers_lsn.wal_flush_lsn;
+        let mut last_wal_replay_lsn: Lsn = Lsn::INVALID;
+        const RETRIES: i32 = 20;
+        for i in 0..=RETRIES {
+            let row = client
+                .query_one("SELECT pg_last_wal_replay_lsn()", &[])
+                .await
+                .context("getting last replay lsn")?;
+            let lsn: u64 = row.get::<usize, postgres_types::PgLsn>(0).into();
+            last_wal_replay_lsn = lsn.into();
+            if last_wal_replay_lsn >= primary_lsn {
+                break;
+            }
+            tracing::info!("Try {i}, replica lsn {last_wal_replay_lsn}, primary lsn {primary_lsn}");
+            sleep(Duration::from_secs(1)).await;
+        }
+        if last_wal_replay_lsn < primary_lsn {
+            bail!("didn't catch up with primary in {RETRIES} retries");
+        }
+
+        // using $1 doesn't work with ALTER SYSTEM SET
+        let safekeepers_sql = format!(
+            "ALTER SYSTEM SET neon.safekeepers='{}'",
+            safekeepers_lsn.safekeepers
+        );
+        client
+            .query(&safekeepers_sql, &[])
+            .await
+            .context("setting safekeepers")?;
+        client
+            .query("SELECT pg_reload_conf()", &[])
+            .await
+            .context("reloading postgres config")?;
+        let row = client
+            .query_one("SELECT * FROM pg_promote()", &[])
+            .await
+            .context("pg_promote")?;
+        if !row.get::<usize, bool>(0) {
+            bail!("pg_promote() returned false");
+        }
+
+        let client = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+            .await
+            .context("connecting to postgres")?;
+        let row = client
+            .query_one("SHOW transaction_read_only", &[])
+            .await
+            .context("getting transaction_read_only")?;
+        if row.get::<usize, &str>(0) == "on" {
+            bail!("replica in read only mode after promotion");
+        }
+
+        let mut state = self.state.lock().unwrap();
+        state.pspec.as_mut().unwrap().spec.mode = ComputeMode::Primary;
+        Ok(())
+    }
+}

compute_tools/src/config.rs

@@ -51,7 +51,7 @@ pub fn write_postgres_conf(
 
     // Write the postgresql.conf content from the spec file as is.
     if let Some(conf) = &spec.cluster.postgresql_conf {
-        writeln!(file, "{}", conf)?;
+        writeln!(file, "{conf}")?;
     }
 
     // Add options for connecting to storage
@@ -70,7 +70,7 @@ pub fn write_postgres_conf(
         );
         // If generation is given, prepend sk list with g#number:
         if let Some(generation) = spec.safekeepers_generation {
-            write!(neon_safekeepers_value, "g#{}:", generation)?;
+            write!(neon_safekeepers_value, "g#{generation}:")?;
         }
         neon_safekeepers_value.push_str(&spec.safekeeper_connstrings.join(","));
         writeln!(
@@ -109,8 +109,8 @@ pub fn write_postgres_conf(
         tls::update_key_path_blocking(pgdata_path, tls_config);
 
         // these are the default, but good to be explicit.
-        writeln!(file, "ssl_cert_file = '{}'", SERVER_CRT)?;
-        writeln!(file, "ssl_key_file = '{}'", SERVER_KEY)?;
+        writeln!(file, "ssl_cert_file = '{SERVER_CRT}'")?;
+        writeln!(file, "ssl_key_file = '{SERVER_KEY}'")?;
     }
 
     // Locales
@@ -191,8 +191,7 @@ pub fn write_postgres_conf(
                 }
                 writeln!(
                     file,
-                    "shared_preload_libraries='{}{}'",
-                    libs, extra_shared_preload_libraries
+                    "shared_preload_libraries='{libs}{extra_shared_preload_libraries}'"
                 )?;
             } else {
                 // Typically, this should be unreacheable,
@@ -244,8 +243,7 @@ pub fn write_postgres_conf(
                 }
                 writeln!(
                     file,
-                    "shared_preload_libraries='{}{}'",
-                    libs, extra_shared_preload_libraries
+                    "shared_preload_libraries='{libs}{extra_shared_preload_libraries}'"
                 )?;
             } else {
                 // Typically, this should be unreacheable,
@@ -263,7 +261,7 @@ pub fn write_postgres_conf(
         }
     }
 
-    writeln!(file, "neon.extension_server_port={}", extension_server_port)?;
+    writeln!(file, "neon.extension_server_port={extension_server_port}")?;
 
     if spec.drop_subscriptions_before_start {
         writeln!(file, "neon.disable_logical_replication_subscribers=true")?;
@@ -291,7 +289,7 @@ where
 {
     let path = pgdata_path.join("compute_ctl_temp_override.conf");
     let mut file = File::create(path)?;
-    write!(file, "{}", options)?;
+    write!(file, "{options}")?;
 
     let res = exec();
 

compute_tools/src/config_template/compute_audit_rsyslog_template.conf

@@ -10,7 +10,13 @@ input(type="imfile" File="{log_directory}/*.log"
   startmsg.regex="^[[:digit:]]{{4}}-[[:digit:]]{{2}}-[[:digit:]]{{2}} [[:digit:]]{{2}}:[[:digit:]]{{2}}:[[:digit:]]{{2}}.[[:digit:]]{{3}} GMT,")
 
 # the directory to store rsyslog state files
-global(workDirectory="/var/log/rsyslog")
+global(
+  workDirectory="/var/log/rsyslog"
+  DefaultNetstreamDriverCAFile="/etc/ssl/certs/ca-certificates.crt"
+)
+
+# Whether the remote syslog receiver uses tls
+set $.remote_syslog_tls = "{remote_syslog_tls}";
 
 # Construct json, endpoint_id and project_id as additional metadata
 set $.json_log!endpoint_id = "{endpoint_id}";
@@ -21,5 +27,29 @@ set $.json_log!msg = $msg;
 template(name="PgAuditLog" type="string"
     string="<%PRI%>1 %TIMESTAMP:::date-rfc3339% %HOSTNAME% - - - - %$.json_log%")
 
-# Forward to remote syslog receiver (@@<hostname>:<port>;format
-local5.info @@{remote_endpoint};PgAuditLog
+# Forward to remote syslog receiver (over TLS)
+if ( $syslogtag == 'pgaudit_log' ) then {{
+  if ( $.remote_syslog_tls == 'true' ) then {{
+    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
+      template="PgAuditLog"
+      queue.type="linkedList"
+      queue.size="1000"
+      action.ResumeRetryCount="10"
+      StreamDriver="gtls"
+      StreamDriverMode="1"
+      StreamDriverAuthMode="x509/name"
+      StreamDriverPermittedPeers="{remote_syslog_host}"
+      StreamDriver.CheckExtendedKeyPurpose="on"
+      StreamDriver.PermitExpiredCerts="off"
+    )
+    stop
+  }} else {{
+    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
+      template="PgAuditLog"
+      queue.type="linkedList"
+      queue.size="1000"
+      action.ResumeRetryCount="10"
+    )
+    stop
+  }}
+}}

compute_tools/src/extension_server.rs

@@ -74,9 +74,11 @@ More specifically, here is an example ext_index.json
 use std::path::Path;
 use std::str;
 
+use crate::metrics::{REMOTE_EXT_REQUESTS_TOTAL, UNKNOWN_HTTP_STATUS};
 use anyhow::{Context, Result, bail};
 use bytes::Bytes;
 use compute_api::spec::RemoteExtSpec;
+use postgres_versioninfo::PgMajorVersion;
 use regex::Regex;
 use remote_storage::*;
 use reqwest::StatusCode;
@@ -86,8 +88,6 @@ use tracing::log::warn;
 use url::Url;
 use zstd::stream::read::Decoder;
 
-use crate::metrics::{REMOTE_EXT_REQUESTS_TOTAL, UNKNOWN_HTTP_STATUS};
-
 fn get_pg_config(argument: &str, pgbin: &str) -> String {
     // gives the result of `pg_config [argument]`
     // where argument is a flag like `--version` or `--sharedir`
@@ -106,7 +106,7 @@ fn get_pg_config(argument: &str, pgbin: &str) -> String {
         .to_string()
 }
 
-pub fn get_pg_version(pgbin: &str) -> PostgresMajorVersion {
+pub fn get_pg_version(pgbin: &str) -> PgMajorVersion {
     // pg_config --version returns a (platform specific) human readable string
     // such as "PostgreSQL 15.4". We parse this to v14/v15/v16 etc.
     let human_version = get_pg_config("--version", pgbin);
@@ -114,25 +114,11 @@ pub fn get_pg_version(pgbin: &str) -> PostgresMajorVersion {
 }
 
 pub fn get_pg_version_string(pgbin: &str) -> String {
-    match get_pg_version(pgbin) {
-        PostgresMajorVersion::V14 => "v14",
-        PostgresMajorVersion::V15 => "v15",
-        PostgresMajorVersion::V16 => "v16",
-        PostgresMajorVersion::V17 => "v17",
-    }
-    .to_owned()
+    get_pg_version(pgbin).v_str()
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum PostgresMajorVersion {
-    V14,
-    V15,
-    V16,
-    V17,
-}
-
-fn parse_pg_version(human_version: &str) -> PostgresMajorVersion {
-    use PostgresMajorVersion::*;
+fn parse_pg_version(human_version: &str) -> PgMajorVersion {
+    use PgMajorVersion::*;
     // Normal releases have version strings like "PostgreSQL 15.4". But there
     // are also pre-release versions like "PostgreSQL 17devel" or "PostgreSQL
     // 16beta2" or "PostgreSQL 17rc1". And with the --with-extra-version
@@ -143,10 +129,10 @@ fn parse_pg_version(human_version: &str) -> PostgresMajorVersion {
         .captures(human_version)
     {
         Some(captures) if captures.len() == 2 => match &captures["major"] {
-            "14" => return V14,
-            "15" => return V15,
-            "16" => return V16,
-            "17" => return V17,
+            "14" => return PG14,
+            "15" => return PG15,
+            "16" => return PG16,
+            "17" => return PG17,
             _ => {}
         },
         _ => {}
@@ -310,10 +296,7 @@ async fn download_extension_tar(remote_ext_base_url: &Url, ext_path: &str) -> Re
 async fn do_extension_server_request(uri: Url) -> Result<Bytes, (String, String)> {
     let resp = reqwest::get(uri).await.map_err(|e| {
         (
-            format!(
-                "could not perform remote extensions server request: {:?}",
-                e
-            ),
+            format!("could not perform remote extensions server request: {e:?}"),
             UNKNOWN_HTTP_STATUS.to_string(),
         )
     })?;
@@ -323,7 +306,7 @@ async fn do_extension_server_request(uri: Url) -> Result<Bytes, (String, String)
         StatusCode::OK => match resp.bytes().await {
             Ok(resp) => Ok(resp),
             Err(e) => Err((
-                format!("could not read remote extensions server response: {:?}", e),
+                format!("could not read remote extensions server response: {e:?}"),
                 // It's fine to return and report error with status as 200 OK,
                 // because we still failed to read the response.
                 status.to_string(),
@@ -334,10 +317,7 @@ async fn do_extension_server_request(uri: Url) -> Result<Bytes, (String, String)
             status.to_string(),
         )),
         _ => Err((
-            format!(
-                "unexpected remote extensions server response status code: {}",
-                status
-            ),
+            format!("unexpected remote extensions server response status code: {status}"),
             status.to_string(),
         )),
     }
@@ -349,25 +329,25 @@ mod tests {
 
     #[test]
     fn test_parse_pg_version() {
-        use super::PostgresMajorVersion::*;
-        assert_eq!(parse_pg_version("PostgreSQL 15.4"), V15);
-        assert_eq!(parse_pg_version("PostgreSQL 15.14"), V15);
+        use postgres_versioninfo::PgMajorVersion::*;
+        assert_eq!(parse_pg_version("PostgreSQL 15.4"), PG15);
+        assert_eq!(parse_pg_version("PostgreSQL 15.14"), PG15);
         assert_eq!(
             parse_pg_version("PostgreSQL 15.4 (Ubuntu 15.4-0ubuntu0.23.04.1)"),
-            V15
+            PG15
         );
 
-        assert_eq!(parse_pg_version("PostgreSQL 14.15"), V14);
-        assert_eq!(parse_pg_version("PostgreSQL 14.0"), V14);
+        assert_eq!(parse_pg_version("PostgreSQL 14.15"), PG14);
+        assert_eq!(parse_pg_version("PostgreSQL 14.0"), PG14);
         assert_eq!(
             parse_pg_version("PostgreSQL 14.9 (Debian 14.9-1.pgdg120+1"),
-            V14
+            PG14
         );
 
-        assert_eq!(parse_pg_version("PostgreSQL 16devel"), V16);
-        assert_eq!(parse_pg_version("PostgreSQL 16beta1"), V16);
-        assert_eq!(parse_pg_version("PostgreSQL 16rc2"), V16);
-        assert_eq!(parse_pg_version("PostgreSQL 16extra"), V16);
+        assert_eq!(parse_pg_version("PostgreSQL 16devel"), PG16);
+        assert_eq!(parse_pg_version("PostgreSQL 16beta1"), PG16);
+        assert_eq!(parse_pg_version("PostgreSQL 16rc2"), PG16);
+        assert_eq!(parse_pg_version("PostgreSQL 16extra"), PG16);
     }
 
     #[test]

compute_tools/src/http/openapi_spec.yaml

@@ -83,6 +83,87 @@ paths:
               schema:
                 $ref: "#/components/schemas/DbsAndRoles"
 
+  /promote:
+    post:
+      tags:
+        - Promotion
+      summary: Promote secondary replica to primary
+      description: ""
+      operationId: promoteReplica
+      requestBody:
+        description: Promote requests data
+        required: true
+        content:
+          application/json:
+            schema:
+                $ref: "#/components/schemas/SafekeepersLsn"
+      responses:
+        200:
+          description: Promote succeeded or wasn't started
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PromoteState"
+        500:
+          description: Promote failed
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PromoteState"
+
+  /lfc/prewarm:
+    post:
+      summary: Request LFC Prewarm
+      parameters:
+        - name: from_endpoint
+          in: query
+          schema:
+            type: string
+      description: ""
+      operationId: lfcPrewarm
+      responses:
+        202:
+          description: LFC prewarm started
+        429:
+          description: LFC prewarm ongoing
+    get:
+      tags:
+        - Prewarm
+      summary: Get LFC prewarm state
+      description: ""
+      operationId: getLfcPrewarmState
+      responses:
+        200:
+          description: Prewarm state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/LfcPrewarmState"
+
+  /lfc/offload:
+    post:
+      summary: Request LFC offload
+      description: ""
+      operationId: lfcOffload
+      responses:
+        202:
+          description: LFC offload started
+        429:
+          description: LFC offload ongoing
+    get:
+      tags:
+        - Prewarm
+      summary: Get LFC offloading state
+      description: ""
+      operationId: getLfcOffloadState
+      responses:
+        200:
+          description: Offload state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/LfcOffloadState"
+
   /database_schema:
     get:
       tags:
@@ -335,15 +416,6 @@ components:
         total_startup_ms:
           type: integer
 
-    Info:
-      type: object
-      description: Information about VM/Pod.
-      required:
-        - num_cpus
-      properties:
-        num_cpus:
-          type: integer
-
     DbsAndRoles:
       type: object
       description: Databases and Roles
@@ -497,25 +569,69 @@ components:
           type: string
           example: "1.0.0"
 
-    InstalledExtensions:
+    SafekeepersLsn:
       type: object
+      required:
+        - safekeepers
+        - wal_flush_lsn
       properties:
-        extensions:
-          description: Contains list of installed extensions.
-          type: array
-          items:
-            type: object
-            properties:
-              extname:
-                type: string
-              version:
-                type: string
-                items:
-                  type: string
-              n_databases:
-                type: integer
-              owned_by_superuser:
-                type: integer
+        safekeepers:
+          description: Primary replica safekeepers
+          type: string
+        wal_flush_lsn:
+          description: Primary last WAL flush LSN
+          type: string
+
+    LfcPrewarmState:
+      type: object
+      required:
+        - status
+        - total
+        - prewarmed
+        - skipped
+      properties:
+        status:
+          description: Lfc prewarm status
+          enum: [not_prewarmed, prewarming, completed, failed]
+          type: string
+        error:
+          description: Lfc prewarm error, if any
+          type: string
+        total:
+          description: Total pages processed
+          type: integer
+        prewarmed:
+          description: Total pages prewarmed
+          type: integer
+        skipped:
+          description: Pages processed but not prewarmed
+          type: integer
+
+    LfcOffloadState:
+      type: object
+      required:
+        - status
+      properties:
+        status:
+          description: Lfc offload status
+          enum: [not_offloaded, offloading, completed, failed]
+          type: string
+        error:
+          description: Lfc offload error, if any
+          type: string
+
+    PromoteState:
+      type: object
+      required:
+        - status
+      properties:
+        status:
+          description: Promote result
+          enum: [not_promoted, completed, failed]
+          type: string
+        error:
+          description: Promote error, if any
+          type: string
 
     SetRoleGrantsRequest:
       type: object

compute_tools/src/http/routes/configure.rs

@@ -65,7 +65,7 @@ pub(in crate::http) async fn configure(
 
             if state.status == ComputeStatus::Failed {
                 let err = state.error.as_ref().map_or("unknown error", |x| x);
-                let msg = format!("compute configuration failed: {:?}", err);
+                let msg = format!("compute configuration failed: {err:?}");
                 return Err(msg);
             }
         }

compute_tools/src/http/routes/mod.rs

@@ -14,6 +14,7 @@ pub(in crate::http) mod insights;
 pub(in crate::http) mod lfc;
 pub(in crate::http) mod metrics;
 pub(in crate::http) mod metrics_json;
+pub(in crate::http) mod promote;
 pub(in crate::http) mod status;
 pub(in crate::http) mod terminate;
 

compute_tools/src/http/routes/promote.rs

@@ -0,0 +1,14 @@
+use crate::http::JsonResponse;
+use axum::Form;
+use http::StatusCode;
+
+pub(in crate::http) async fn promote(
+    compute: axum::extract::State<std::sync::Arc<crate::compute::ComputeNode>>,
+    Form(safekeepers_lsn): Form<compute_api::responses::SafekeepersLsn>,
+) -> axum::response::Response {
+    let state = compute.promote(safekeepers_lsn).await;
+    if let compute_api::responses::PromoteState::Failed { error } = state {
+        return JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, error);
+    }
+    JsonResponse::success(StatusCode::OK, state)
+}

compute_tools/src/http/server.rs

@@ -23,7 +23,7 @@ use super::{
     middleware::authorize::Authorize,
     routes::{
         check_writability, configure, database_schema, dbs_and_roles, extension_server, extensions,
-        grants, insights, lfc, metrics, metrics_json, status, terminate,
+        grants, insights, lfc, metrics, metrics_json, promote, status, terminate,
     },
 };
 use crate::compute::ComputeNode;
@@ -87,6 +87,7 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/lfc/prewarm", get(lfc::prewarm_state).post(lfc::prewarm))
                     .route("/lfc/offload", get(lfc::offload_state).post(lfc::offload))
+                    .route("/promote", post(promote::promote))
                     .route("/check_writability", post(check_writability::is_writable))
                     .route("/configure", post(configure::configure))
                     .route("/database_schema", get(database_schema::get_schema_dump))

compute_tools/src/installed_extensions.rs

@@ -43,7 +43,7 @@ pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExten
         let (mut client, connection) = conf.connect(NoTls).await?;
         tokio::spawn(async move {
             if let Err(e) = connection.await {
-                eprintln!("connection error: {}", e);
+                eprintln!("connection error: {e}");
             }
         });
 
@@ -57,7 +57,7 @@ pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExten
         let (client, connection) = conf.connect(NoTls).await?;
         tokio::spawn(async move {
             if let Err(e) = connection.await {
-                eprintln!("connection error: {}", e);
+                eprintln!("connection error: {e}");
             }
         });
 

compute_tools/src/lib.rs

@@ -12,6 +12,7 @@ pub mod logger;
 pub mod catalog;
 pub mod compute;
 pub mod compute_prewarm;
+pub mod compute_promote;
 pub mod disk_quota;
 pub mod extension_server;
 pub mod installed_extensions;

compute_tools/src/lsn_lease.rs

@@ -4,7 +4,9 @@ use std::thread;
 use std::time::{Duration, SystemTime};
 
 use anyhow::{Result, bail};
-use compute_api::spec::ComputeMode;
+use compute_api::spec::{ComputeMode, PageserverProtocol};
+use itertools::Itertools as _;
+use pageserver_page_api as page_api;
 use postgres::{NoTls, SimpleQueryMessage};
 use tracing::{info, warn};
 use utils::id::{TenantId, TimelineId};
@@ -76,25 +78,17 @@ fn acquire_lsn_lease_with_retry(
 
     loop {
         // Note: List of pageservers is dynamic, need to re-read configs before each attempt.
-        let configs = {
+        let (connstrings, auth) = {
             let state = compute.state.lock().unwrap();
-
             let spec = state.pspec.as_ref().expect("spec must be set");
-
-            let conn_strings = spec.pageserver_connstr.split(',');
-
-            conn_strings
-                .map(|connstr| {
-                    let mut config = postgres::Config::from_str(connstr).expect("Invalid connstr");
-                    if let Some(storage_auth_token) = &spec.storage_auth_token {
-                        config.password(storage_auth_token.clone());
-                    }
-                    config
-                })
-                .collect::<Vec<_>>()
+            (
+                spec.pageserver_connstr.clone(),
+                spec.storage_auth_token.clone(),
+            )
         };
 
-        let result = try_acquire_lsn_lease(tenant_id, timeline_id, lsn, &configs);
+        let result =
+            try_acquire_lsn_lease(&connstrings, auth.as_deref(), tenant_id, timeline_id, lsn);
         match result {
             Ok(Some(res)) => {
                 return Ok(res);
@@ -116,68 +110,104 @@ fn acquire_lsn_lease_with_retry(
     }
 }
 
-/// Tries to acquire an LSN lease through PS page_service API.
+/// Tries to acquire LSN leases on all Pageserver shards.
 fn try_acquire_lsn_lease(
+    connstrings: &str,
+    auth: Option<&str>,
     tenant_id: TenantId,
     timeline_id: TimelineId,
     lsn: Lsn,
-    configs: &[postgres::Config],
 ) -> Result<Option<SystemTime>> {
-    fn get_valid_until(
-        config: &postgres::Config,
-        tenant_shard_id: TenantShardId,
-        timeline_id: TimelineId,
-        lsn: Lsn,
-    ) -> Result<Option<SystemTime>> {
-        let mut client = config.connect(NoTls)?;
-        let cmd = format!("lease lsn {} {} {} ", tenant_shard_id, timeline_id, lsn);
-        let res = client.simple_query(&cmd)?;
-        let msg = match res.first() {
-            Some(msg) => msg,
-            None => bail!("empty response"),
-        };
-        let row = match msg {
-            SimpleQueryMessage::Row(row) => row,
-            _ => bail!("error parsing lsn lease response"),
+    let connstrings = connstrings.split(',').collect_vec();
+    let shard_count = connstrings.len();
+    let mut leases = Vec::new();
+
+    for (shard_number, &connstring) in connstrings.iter().enumerate() {
+        let tenant_shard_id = match shard_count {
+            0 | 1 => TenantShardId::unsharded(tenant_id),
+            shard_count => TenantShardId {
+                tenant_id,
+                shard_number: ShardNumber(shard_number as u8),
+                shard_count: ShardCount::new(shard_count as u8),
+            },
         };
 
-        // Note: this will be None if a lease is explicitly not granted.
-        let valid_until_str = row.get("valid_until");
-
-        let valid_until = valid_until_str.map(|s| {
-            SystemTime::UNIX_EPOCH
-                .checked_add(Duration::from_millis(u128::from_str(s).unwrap() as u64))
-                .expect("Time larger than max SystemTime could handle")
-        });
-        Ok(valid_until)
+        let lease = match PageserverProtocol::from_connstring(connstring)? {
+            PageserverProtocol::Libpq => {
+                acquire_lsn_lease_libpq(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
+            PageserverProtocol::Grpc => {
+                acquire_lsn_lease_grpc(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
+        };
+        leases.push(lease);
     }
 
-    let shard_count = configs.len();
+    Ok(leases.into_iter().min().flatten())
+}
 
-    let valid_until = if shard_count > 1 {
-        configs
-            .iter()
-            .enumerate()
-            .map(|(shard_number, config)| {
-                let tenant_shard_id = TenantShardId {
-                    tenant_id,
-                    shard_count: ShardCount::new(shard_count as u8),
-                    shard_number: ShardNumber(shard_number as u8),
-                };
-                get_valid_until(config, tenant_shard_id, timeline_id, lsn)
-            })
-            .collect::<Result<Vec<Option<SystemTime>>>>()?
-            .into_iter()
-            .min()
-            .unwrap()
-    } else {
-        get_valid_until(
-            &configs[0],
-            TenantShardId::unsharded(tenant_id),
-            timeline_id,
-            lsn,
-        )?
+/// Acquires an LSN lease on a single shard, using the libpq API. The connstring must use a
+/// postgresql:// scheme.
+fn acquire_lsn_lease_libpq(
+    connstring: &str,
+    auth: Option<&str>,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+    lsn: Lsn,
+) -> Result<Option<SystemTime>> {
+    let mut config = postgres::Config::from_str(connstring)?;
+    if let Some(auth) = auth {
+        config.password(auth);
+    }
+    let mut client = config.connect(NoTls)?;
+    let cmd = format!("lease lsn {tenant_shard_id} {timeline_id} {lsn} ");
+    let res = client.simple_query(&cmd)?;
+    let msg = match res.first() {
+        Some(msg) => msg,
+        None => bail!("empty response"),
+    };
+    let row = match msg {
+        SimpleQueryMessage::Row(row) => row,
+        _ => bail!("error parsing lsn lease response"),
     };
 
+    // Note: this will be None if a lease is explicitly not granted.
+    let valid_until_str = row.get("valid_until");
+
+    let valid_until = valid_until_str.map(|s| {
+        SystemTime::UNIX_EPOCH
+            .checked_add(Duration::from_millis(u128::from_str(s).unwrap() as u64))
+            .expect("Time larger than max SystemTime could handle")
+    });
     Ok(valid_until)
 }
+
+/// Acquires an LSN lease on a single shard, using the gRPC API. The connstring must use a
+/// grpc:// scheme.
+fn acquire_lsn_lease_grpc(
+    connstring: &str,
+    auth: Option<&str>,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+    lsn: Lsn,
+) -> Result<Option<SystemTime>> {
+    tokio::runtime::Handle::current().block_on(async move {
+        let mut client = page_api::Client::connect(
+            connstring.to_string(),
+            tenant_shard_id.tenant_id,
+            timeline_id,
+            tenant_shard_id.to_index(),
+            auth.map(String::from),
+            None,
+        )
+        .await?;
+
+        let req = page_api::LeaseLsnRequest { lsn };
+        match client.lease_lsn(req).await {
+            Ok(expires) => Ok(Some(expires)),
+            // Lease couldn't be acquired because the LSN has been garbage collected.
+            Err(err) if err.code() == tonic::Code::FailedPrecondition => Ok(None),
+            Err(err) => Err(err.into()),
+        }
+    })
+}

compute_tools/src/metrics.rs

@@ -97,20 +97,34 @@ pub(crate) static PG_TOTAL_DOWNTIME_MS: Lazy<GenericCounter<AtomicU64>> = Lazy::
     .expect("failed to define a metric")
 });
 
-/// Needed as neon.file_cache_prewarm_batch == 0 doesn't mean we never tried to prewarm.
-/// On the other hand, LFC_PREWARMED_PAGES is excessive as we can GET /lfc/prewarm
-pub(crate) static LFC_PREWARM_REQUESTS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static LFC_PREWARMS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
-        "compute_ctl_lfc_prewarm_requests_total",
-        "Total number of LFC prewarm requests made by compute_ctl",
+        "compute_ctl_lfc_prewarms_total",
+        "Total number of LFC prewarms requested by compute_ctl or autoprewarm option",
     )
     .expect("failed to define a metric")
 });
 
-pub(crate) static LFC_OFFLOAD_REQUESTS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static LFC_PREWARM_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
-        "compute_ctl_lfc_offload_requests_total",
-        "Total number of LFC offload requests made by compute_ctl",
+        "compute_ctl_lfc_prewarm_errors_total",
+        "Total number of LFC prewarm errors",
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static LFC_OFFLOADS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "compute_ctl_lfc_offloads_total",
+        "Total number of LFC offloads requested by compute_ctl or lfc_offload_period_seconds option",
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static LFC_OFFLOAD_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "compute_ctl_lfc_offload_errors_total",
+        "Total number of LFC offload errors",
     )
     .expect("failed to define a metric")
 });
@@ -124,7 +138,9 @@ pub fn collect() -> Vec<MetricFamily> {
     metrics.extend(AUDIT_LOG_DIR_SIZE.collect());
     metrics.extend(PG_CURR_DOWNTIME_MS.collect());
     metrics.extend(PG_TOTAL_DOWNTIME_MS.collect());
-    metrics.extend(LFC_PREWARM_REQUESTS.collect());
-    metrics.extend(LFC_OFFLOAD_REQUESTS.collect());
+    metrics.extend(LFC_PREWARMS.collect());
+    metrics.extend(LFC_PREWARM_ERRORS.collect());
+    metrics.extend(LFC_OFFLOADS.collect());
+    metrics.extend(LFC_OFFLOAD_ERRORS.collect());
     metrics
 }

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -0,0 +1 @@
+GRANT pg_signal_backend TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/tests/0004-grant_pg_monitor_to_neon_superuser.sql

@@ -7,13 +7,17 @@ BEGIN
         INTO monitor
         FROM pg_auth_members
         WHERE roleid = 'pg_monitor'::regrole
-            AND member = 'pg_monitor'::regrole;
+            AND member = 'neon_superuser'::regrole;
 
-    IF NOT monitor.member THEN
+    IF monitor IS NULL THEN
+        RAISE EXCEPTION 'no entry in pg_auth_members for neon_superuser and pg_monitor';
+    END IF;
+
+    IF monitor.admin IS NULL OR NOT monitor.member THEN
         RAISE EXCEPTION 'neon_superuser is not a member of pg_monitor';
     END IF;
 
-    IF NOT monitor.admin THEN
+    IF monitor.admin IS NULL OR NOT monitor.admin THEN
         RAISE EXCEPTION 'neon_superuser cannot grant pg_monitor';
     END IF;
 END $$;

compute_tools/src/migrations/tests/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -0,0 +1,23 @@
+DO $$
+DECLARE
+    signal_backend record;
+BEGIN
+    SELECT pg_has_role('neon_superuser', 'pg_signal_backend', 'member') AS member,
+            admin_option AS admin
+        INTO signal_backend
+        FROM pg_auth_members
+        WHERE roleid = 'pg_signal_backend'::regrole
+            AND member = 'neon_superuser'::regrole;
+
+    IF signal_backend IS NULL THEN
+        RAISE EXCEPTION 'no entry in pg_auth_members for neon_superuser and pg_signal_backend';
+    END IF;
+
+    IF signal_backend.member IS NULL OR NOT signal_backend.member THEN
+        RAISE EXCEPTION 'neon_superuser is not a member of pg_signal_backend';
+    END IF;
+
+    IF signal_backend.admin IS NULL OR NOT signal_backend.admin THEN
+        RAISE EXCEPTION 'neon_superuser cannot grant pg_signal_backend';
+    END IF;
+END $$;

compute_tools/src/pg_helpers.rs

@@ -36,9 +36,9 @@ pub fn escape_literal(s: &str) -> String {
     let res = s.replace('\'', "''").replace('\\', "\\\\");
 
     if res.contains('\\') {
-        format!("E'{}'", res)
+        format!("E'{res}'")
     } else {
-        format!("'{}'", res)
+        format!("'{res}'")
     }
 }
 
@@ -46,7 +46,7 @@ pub fn escape_literal(s: &str) -> String {
 /// with `'{}'` is not required, as it returns a ready-to-use config string.
 pub fn escape_conf_value(s: &str) -> String {
     let res = s.replace('\'', "''").replace('\\', "\\\\");
-    format!("'{}'", res)
+    format!("'{res}'")
 }
 
 pub trait GenericOptionExt {
@@ -446,7 +446,7 @@ pub async fn tune_pgbouncer(
         let mut pgbouncer_connstr =
             "host=localhost port=6432 dbname=pgbouncer user=postgres sslmode=disable".to_string();
         if let Ok(pass) = std::env::var("PGBOUNCER_PASSWORD") {
-            pgbouncer_connstr.push_str(format!(" password={}", pass).as_str());
+            pgbouncer_connstr.push_str(format!(" password={pass}").as_str());
         }
         pgbouncer_connstr
     };
@@ -464,7 +464,7 @@ pub async fn tune_pgbouncer(
             Ok((client, connection)) => {
                 tokio::spawn(async move {
                     if let Err(e) = connection.await {
-                        eprintln!("connection error: {}", e);
+                        eprintln!("connection error: {e}");
                     }
                 });
                 break client;

compute_tools/src/rsyslog.rs

@@ -4,8 +4,10 @@ use std::path::Path;
 use std::process::Command;
 use std::time::Duration;
 use std::{fs::OpenOptions, io::Write};
+use url::{Host, Url};
 
 use anyhow::{Context, Result, anyhow};
+use hostname_validator;
 use tracing::{error, info, instrument, warn};
 
 const POSTGRES_LOGS_CONF_PATH: &str = "/etc/rsyslog.d/postgres_logs.conf";
@@ -82,18 +84,84 @@ fn restart_rsyslog() -> Result<()> {
     Ok(())
 }
 
+fn parse_audit_syslog_address(
+    remote_plain_endpoint: &str,
+    remote_tls_endpoint: &str,
+) -> Result<(String, u16, String)> {
+    let tls;
+    let remote_endpoint = if !remote_tls_endpoint.is_empty() {
+        tls = "true".to_string();
+        remote_tls_endpoint
+    } else {
+        tls = "false".to_string();
+        remote_plain_endpoint
+    };
+    // Urlify the remote_endpoint, so parsing can be done with url::Url.
+    let url_str = format!("http://{remote_endpoint}");
+    let url = Url::parse(&url_str).map_err(|err| {
+        anyhow!("Error parsing {remote_endpoint}, expected host:port, got {err:?}")
+    })?;
+
+    let is_valid = url.scheme() == "http"
+        && url.path() == "/"
+        && url.query().is_none()
+        && url.fragment().is_none()
+        && url.username() == ""
+        && url.password().is_none();
+
+    if !is_valid {
+        return Err(anyhow!(
+            "Invalid address format {remote_endpoint}, expected host:port"
+        ));
+    }
+    let host = match url.host() {
+        Some(Host::Domain(h)) if hostname_validator::is_valid(h) => h.to_string(),
+        Some(Host::Ipv4(ip4)) => ip4.to_string(),
+        Some(Host::Ipv6(ip6)) => ip6.to_string(),
+        _ => return Err(anyhow!("Invalid host")),
+    };
+    let port = url
+        .port()
+        .ok_or_else(|| anyhow!("Invalid port in {remote_endpoint}"))?;
+
+    Ok((host, port, tls))
+}
+
+fn generate_audit_rsyslog_config(
+    log_directory: String,
+    endpoint_id: &str,
+    project_id: &str,
+    remote_syslog_host: &str,
+    remote_syslog_port: u16,
+    remote_syslog_tls: &str,
+) -> String {
+    format!(
+        include_str!("config_template/compute_audit_rsyslog_template.conf"),
+        log_directory = log_directory,
+        endpoint_id = endpoint_id,
+        project_id = project_id,
+        remote_syslog_host = remote_syslog_host,
+        remote_syslog_port = remote_syslog_port,
+        remote_syslog_tls = remote_syslog_tls
+    )
+}
+
 pub fn configure_audit_rsyslog(
     log_directory: String,
     endpoint_id: &str,
     project_id: &str,
     remote_endpoint: &str,
+    remote_tls_endpoint: &str,
 ) -> Result<()> {
-    let config_content: String = format!(
-        include_str!("config_template/compute_audit_rsyslog_template.conf"),
-        log_directory = log_directory,
-        endpoint_id = endpoint_id,
-        project_id = project_id,
-        remote_endpoint = remote_endpoint
+    let (remote_syslog_host, remote_syslog_port, remote_syslog_tls) =
+        parse_audit_syslog_address(remote_endpoint, remote_tls_endpoint).unwrap();
+    let config_content = generate_audit_rsyslog_config(
+        log_directory,
+        endpoint_id,
+        project_id,
+        &remote_syslog_host,
+        remote_syslog_port,
+        &remote_syslog_tls,
     );
 
     info!("rsyslog config_content: {}", config_content);
@@ -258,6 +326,8 @@ pub fn launch_pgaudit_gc(log_directory: String) {
 mod tests {
     use crate::rsyslog::PostgresLogsRsyslogConfig;
 
+    use super::{generate_audit_rsyslog_config, parse_audit_syslog_address};
+
     #[test]
     fn test_postgres_logs_config() {
         {
@@ -287,4 +357,146 @@ mod tests {
             assert!(res.is_err());
         }
     }
+
+    #[test]
+    fn test_parse_audit_syslog_address() {
+        {
+            // host:port format (plaintext)
+            let parsed = parse_audit_syslog_address("collector.host.tld:5555", "");
+            assert!(parsed.is_ok());
+            assert_eq!(
+                parsed.unwrap(),
+                (
+                    String::from("collector.host.tld"),
+                    5555,
+                    String::from("false")
+                )
+            );
+        }
+
+        {
+            // host:port format with ipv4 ip address (plaintext)
+            let parsed = parse_audit_syslog_address("10.0.0.1:5555", "");
+            assert!(parsed.is_ok());
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("10.0.0.1"), 5555, String::from("false"))
+            );
+        }
+
+        {
+            // host:port format with ipv6 ip address (plaintext)
+            let parsed =
+                parse_audit_syslog_address("[7e60:82ed:cb2e:d617:f904:f395:aaca:e252]:5555", "");
+            assert_eq!(
+                parsed.unwrap(),
+                (
+                    String::from("7e60:82ed:cb2e:d617:f904:f395:aaca:e252"),
+                    5555,
+                    String::from("false")
+                )
+            );
+        }
+
+        {
+            // Only TLS host:port defined
+            let parsed = parse_audit_syslog_address("", "tls.host.tld:5556");
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("tls.host.tld"), 5556, String::from("true"))
+            );
+        }
+
+        {
+            // tls host should take precedence, when both defined
+            let parsed = parse_audit_syslog_address("plaintext.host.tld:5555", "tls.host.tld:5556");
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("tls.host.tld"), 5556, String::from("true"))
+            );
+        }
+
+        {
+            // host without port (plaintext)
+            let parsed = parse_audit_syslog_address("collector.host.tld", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // port without host
+            let parsed = parse_audit_syslog_address(":5555", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // valid host with invalid port
+            let parsed = parse_audit_syslog_address("collector.host.tld:90001", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // invalid hostname with valid port
+            let parsed = parse_audit_syslog_address("-collector.host.tld:5555", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // parse error
+            let parsed = parse_audit_syslog_address("collector.host.tld:::5555", "");
+            assert!(parsed.is_err());
+        }
+    }
+
+    #[test]
+    fn test_generate_audit_rsyslog_config() {
+        {
+            // plaintext version
+            let log_directory = "/tmp/log".to_string();
+            let endpoint_id = "ep-test-endpoint-id";
+            let project_id = "test-project-id";
+            let remote_syslog_host = "collector.host.tld";
+            let remote_syslog_port = 5555;
+            let remote_syslog_tls = "false";
+
+            let conf_str = generate_audit_rsyslog_config(
+                log_directory,
+                endpoint_id,
+                project_id,
+                remote_syslog_host,
+                remote_syslog_port,
+                remote_syslog_tls,
+            );
+
+            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "false";"#));
+            assert!(conf_str.contains(r#"type="omfwd""#));
+            assert!(conf_str.contains(r#"target="collector.host.tld""#));
+            assert!(conf_str.contains(r#"port="5555""#));
+            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
+        }
+
+        {
+            // TLS version
+            let log_directory = "/tmp/log".to_string();
+            let endpoint_id = "ep-test-endpoint-id";
+            let project_id = "test-project-id";
+            let remote_syslog_host = "collector.host.tld";
+            let remote_syslog_port = 5556;
+            let remote_syslog_tls = "true";
+
+            let conf_str = generate_audit_rsyslog_config(
+                log_directory,
+                endpoint_id,
+                project_id,
+                remote_syslog_host,
+                remote_syslog_port,
+                remote_syslog_tls,
+            );
+
+            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "true";"#));
+            assert!(conf_str.contains(r#"type="omfwd""#));
+            assert!(conf_str.contains(r#"target="collector.host.tld""#));
+            assert!(conf_str.contains(r#"port="5556""#));
+            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
+        }
+    }
 }

compute_tools/src/spec.rs

@@ -23,12 +23,12 @@ fn do_control_plane_request(
 ) -> Result<ControlPlaneConfigResponse, (bool, String, String)> {
     let resp = reqwest::blocking::Client::new()
         .get(uri)
-        .header("Authorization", format!("Bearer {}", jwt))
+        .header("Authorization", format!("Bearer {jwt}"))
         .send()
         .map_err(|e| {
             (
                 true,
-                format!("could not perform request to control plane: {:?}", e),
+                format!("could not perform request to control plane: {e:?}"),
                 UNKNOWN_HTTP_STATUS.to_string(),
             )
         })?;
@@ -39,7 +39,7 @@ fn do_control_plane_request(
             Ok(spec_resp) => Ok(spec_resp),
             Err(e) => Err((
                 true,
-                format!("could not deserialize control plane response: {:?}", e),
+                format!("could not deserialize control plane response: {e:?}"),
                 status.to_string(),
             )),
         },
@@ -62,7 +62,7 @@ fn do_control_plane_request(
         // or some internal failure happened. Doesn't make much sense to retry in this case.
         _ => Err((
             false,
-            format!("unexpected control plane response status code: {}", status),
+            format!("unexpected control plane response status code: {status}"),
             status.to_string(),
         )),
     }
@@ -197,6 +197,7 @@ pub async fn handle_migrations(client: &mut Client) -> Result<()> {
         include_str!(
             "./migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql"
         ),
+        include_str!("./migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql"),
     ];
 
     MigrationRunner::new(client, &migrations)

compute_tools/src/spec_apply.rs

@@ -933,56 +933,53 @@ async fn get_operations<'a>(
                 PerDatabasePhase::DeleteDBRoleReferences => {
                     let ctx = ctx.read().await;
 
-                    let operations =
-                        spec.delta_operations
-                            .iter()
-                            .flatten()
-                            .filter(|op| op.action == "delete_role")
-                            .filter_map(move |op| {
-                                if db.is_owned_by(&op.name) {
-                                    return None;
-                                }
-                                if !ctx.roles.contains_key(&op.name) {
-                                    return None;
-                                }
-                                let quoted = op.name.pg_quote();
-                                let new_owner = match &db {
-                                    DB::SystemDB => PgIdent::from("cloud_admin").pg_quote(),
-                                    DB::UserDB(db) => db.owner.pg_quote(),
-                                };
-                                let (escaped_role, outer_tag) = op.name.pg_quote_dollar();
+                    let operations = spec
+                        .delta_operations
+                        .iter()
+                        .flatten()
+                        .filter(|op| op.action == "delete_role")
+                        .filter_map(move |op| {
+                            if db.is_owned_by(&op.name) {
+                                return None;
+                            }
+                            if !ctx.roles.contains_key(&op.name) {
+                                return None;
+                            }
+                            let quoted = op.name.pg_quote();
+                            let new_owner = match &db {
+                                DB::SystemDB => PgIdent::from("cloud_admin").pg_quote(),
+                                DB::UserDB(db) => db.owner.pg_quote(),
+                            };
+                            let (escaped_role, outer_tag) = op.name.pg_quote_dollar();
 
-                                Some(vec![
-                                    // This will reassign all dependent objects to the db owner
-                                    Operation {
-                                        query: format!(
-                                            "REASSIGN OWNED BY {} TO {}",
-                                            quoted, new_owner,
-                                        ),
-                                        comment: None,
-                                    },
-                                    // Revoke some potentially blocking privileges (Neon-specific currently)
-                                    Operation {
-                                        query: format!(
-                                            include_str!("sql/pre_drop_role_revoke_privileges.sql"),
-                                            // N.B. this has to be properly dollar-escaped with `pg_quote_dollar()`
-                                            role_name = escaped_role,
-                                            outer_tag = outer_tag,
-                                        ),
-                                        comment: None,
-                                    },
-                                    // This now will only drop privileges of the role
-                                    // TODO: this is obviously not 100% true because of the above case,
-                                    // there could be still some privileges that are not revoked. Maybe this
-                                    // only drops privileges that were granted *by this* role, not *to this* role,
-                                    // but this has to be checked.
-                                    Operation {
-                                        query: format!("DROP OWNED BY {}", quoted),
-                                        comment: None,
-                                    },
-                                ])
-                            })
-                            .flatten();
+                            Some(vec![
+                                // This will reassign all dependent objects to the db owner
+                                Operation {
+                                    query: format!("REASSIGN OWNED BY {quoted} TO {new_owner}",),
+                                    comment: None,
+                                },
+                                // Revoke some potentially blocking privileges (Neon-specific currently)
+                                Operation {
+                                    query: format!(
+                                        include_str!("sql/pre_drop_role_revoke_privileges.sql"),
+                                        // N.B. this has to be properly dollar-escaped with `pg_quote_dollar()`
+                                        role_name = escaped_role,
+                                        outer_tag = outer_tag,
+                                    ),
+                                    comment: None,
+                                },
+                                // This now will only drop privileges of the role
+                                // TODO: this is obviously not 100% true because of the above case,
+                                // there could be still some privileges that are not revoked. Maybe this
+                                // only drops privileges that were granted *by this* role, not *to this* role,
+                                // but this has to be checked.
+                                Operation {
+                                    query: format!("DROP OWNED BY {quoted}"),
+                                    comment: None,
+                                },
+                            ])
+                        })
+                        .flatten();
 
                     Ok(Box::new(operations))
                 }

compute_tools/src/sync_sk.rs

@@ -27,7 +27,7 @@ pub async fn ping_safekeeper(
     let (client, conn) = config.connect(tokio_postgres::NoTls).await?;
     tokio::spawn(async move {
         if let Err(e) = conn.await {
-            eprintln!("connection error: {}", e);
+            eprintln!("connection error: {e}");
         }
     });
 

compute_tools/tests/cluster_spec.json

@@ -3,7 +3,8 @@
 
   "timestamp": "2021-05-23T18:25:43.511Z",
   "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8b",
-
+  "suspend_timeout_seconds": 3600,
+  
   "cluster": {
     "cluster_id": "test-cluster-42",
     "name": "Zenith Test",

compute_tools/tests/pg_helpers_tests.rs

@@ -31,6 +31,7 @@ mod pg_helpers_tests {
 wal_level = logical
 hot_standby = on
 autoprewarm = off
+offload_lfc_interval_seconds = 20
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on
 log_connections = on

control_plane/src/bin/neon_local.rs

@@ -16,9 +16,9 @@ use std::time::Duration;
 use anyhow::{Context, Result, anyhow, bail};
 use clap::Parser;
 use compute_api::requests::ComputeClaimsScope;
-use compute_api::spec::ComputeMode;
+use compute_api::spec::{ComputeMode, PageserverProtocol};
 use control_plane::broker::StorageBroker;
-use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode, PageserverProtocol};
+use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
 use control_plane::local_env;
 use control_plane::local_env::{
@@ -48,7 +48,7 @@ use postgres_connection::parse_host_port;
 use safekeeper_api::membership::{SafekeeperGeneration, SafekeeperId};
 use safekeeper_api::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_SAFEKEEPER_HTTP_PORT,
-    DEFAULT_PG_LISTEN_PORT as DEFAULT_SAFEKEEPER_PG_PORT,
+    DEFAULT_PG_LISTEN_PORT as DEFAULT_SAFEKEEPER_PG_PORT, PgMajorVersion, PgVersionId,
 };
 use storage_broker::DEFAULT_LISTEN_ADDR as DEFAULT_BROKER_ADDR;
 use tokio::task::JoinSet;
@@ -64,7 +64,9 @@ const DEFAULT_PAGESERVER_ID: NodeId = NodeId(1);
 const DEFAULT_BRANCH_NAME: &str = "main";
 project_git_version!(GIT_VERSION);
 
-const DEFAULT_PG_VERSION: u32 = 17;
+#[allow(dead_code)]
+const DEFAULT_PG_VERSION: PgMajorVersion = PgMajorVersion::PG17;
+const DEFAULT_PG_VERSION_NUM: &str = "17";
 
 const DEFAULT_PAGESERVER_CONTROL_PLANE_API: &str = "http://127.0.0.1:1234/upcall/v1/";
 
@@ -167,9 +169,9 @@ struct TenantCreateCmdArgs {
     #[clap(short = 'c')]
     config: Vec<String>,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version to use for the initial timeline")]
-    pg_version: u32,
+    pg_version: PgMajorVersion,
 
     #[clap(
         long,
@@ -290,9 +292,9 @@ struct TimelineCreateCmdArgs {
     #[clap(long, help = "Human-readable alias for the new timeline")]
     branch_name: String,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version")]
-    pg_version: u32,
+    pg_version: PgMajorVersion,
 }
 
 #[derive(clap::Args)]
@@ -322,9 +324,9 @@ struct TimelineImportCmdArgs {
     #[clap(long, help = "Lsn the basebackup ends at")]
     end_lsn: Option<Lsn>,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version of the backup being imported")]
-    pg_version: u32,
+    pg_version: PgMajorVersion,
 }
 
 #[derive(clap::Subcommand)]
@@ -601,9 +603,9 @@ struct EndpointCreateCmdArgs {
     )]
     config_only: bool,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version")]
-    pg_version: u32,
+    pg_version: PgMajorVersion,
 
     /// Use gRPC to communicate with Pageservers, by generating grpc:// connstrings.
     ///
@@ -673,6 +675,16 @@ struct EndpointStartCmdArgs {
     #[arg(default_value = "90s")]
     start_timeout: Duration,
 
+    #[clap(
+        long,
+        help = "Download LFC cache from endpoint storage on endpoint startup",
+        default_value = "false"
+    )]
+    autoprewarm: bool,
+
+    #[clap(long, help = "Upload LFC cache to endpoint storage periodically")]
+    offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+
     #[clap(
         long,
         help = "Run in development mode, skipping VM-specific operations like process termination",
@@ -919,7 +931,7 @@ fn print_timeline(
             br_sym = "┗━";
         }
 
-        print!("{} @{}: ", br_sym, ancestor_lsn);
+        print!("{br_sym} @{ancestor_lsn}: ");
     }
 
     // Finally print a timeline id and name with new line
@@ -1295,7 +1307,7 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
                     },
                     new_members: None,
                 };
-                let pg_version = args.pg_version * 10000;
+                let pg_version = PgVersionId::from(args.pg_version);
                 let req = safekeeper_api::models::TimelineCreateRequest {
                     tenant_id,
                     timeline_id,
@@ -1583,22 +1595,24 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             let endpoint_storage_token = env.generate_auth_token(&claims)?;
             let endpoint_storage_addr = env.endpoint_storage.listen_addr.to_string();
 
+            let args = control_plane::endpoint::EndpointStartArgs {
+                auth_token,
+                endpoint_storage_token,
+                endpoint_storage_addr,
+                safekeepers_generation,
+                safekeepers,
+                pageservers,
+                remote_ext_base_url: remote_ext_base_url.clone(),
+                shard_stripe_size: stripe_size.0 as usize,
+                create_test_user: args.create_test_user,
+                start_timeout: args.start_timeout,
+                autoprewarm: args.autoprewarm,
+                offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
+                dev: args.dev,
+            };
+
             println!("Starting existing endpoint {endpoint_id}...");
-            endpoint
-                .start(
-                    &auth_token,
-                    endpoint_storage_token,
-                    endpoint_storage_addr,
-                    safekeepers_generation,
-                    safekeepers,
-                    pageservers,
-                    remote_ext_base_url.as_ref(),
-                    stripe_size.0 as usize,
-                    args.create_test_user,
-                    args.start_timeout,
-                    args.dev,
-                )
-                .await?;
+            endpoint.start(args).await?;
         }
         EndpointCmd::Reconfigure(args) => {
             let endpoint_id = &args.endpoint_id;
@@ -1649,7 +1663,9 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             // If --safekeepers argument is given, use only the listed
             // safekeeper nodes; otherwise all from the env.
             let safekeepers = parse_safekeepers(&args.safekeepers)?;
-            endpoint.reconfigure(pageservers, None, safekeepers).await?;
+            endpoint
+                .reconfigure(Some(pageservers), None, safekeepers, None)
+                .await?;
         }
         EndpointCmd::Stop(args) => {
             let endpoint_id = &args.endpoint_id;
@@ -1742,7 +1758,7 @@ async fn handle_pageserver(subcmd: &PageserverCmd, env: &local_env::LocalEnv) ->
                 StopMode::Immediate => true,
             };
             if let Err(e) = get_pageserver(env, args.pageserver_id)?.stop(immediate) {
-                eprintln!("pageserver stop failed: {}", e);
+                eprintln!("pageserver stop failed: {e}");
                 exit(1);
             }
         }
@@ -1751,7 +1767,7 @@ async fn handle_pageserver(subcmd: &PageserverCmd, env: &local_env::LocalEnv) ->
             let pageserver = get_pageserver(env, args.pageserver_id)?;
             //TODO what shutdown strategy should we use here?
             if let Err(e) = pageserver.stop(false) {
-                eprintln!("pageserver stop failed: {}", e);
+                eprintln!("pageserver stop failed: {e}");
                 exit(1);
             }
 
@@ -1768,7 +1784,7 @@ async fn handle_pageserver(subcmd: &PageserverCmd, env: &local_env::LocalEnv) ->
             {
                 Ok(_) => println!("Page server is up and running"),
                 Err(err) => {
-                    eprintln!("Page server is not available: {}", err);
+                    eprintln!("Page server is not available: {err}");
                     exit(1);
                 }
             }
@@ -1805,7 +1821,7 @@ async fn handle_storage_controller(
                 },
             };
             if let Err(e) = svc.stop(stop_args).await {
-                eprintln!("stop failed: {}", e);
+                eprintln!("stop failed: {e}");
                 exit(1);
             }
         }
@@ -1827,7 +1843,7 @@ async fn handle_safekeeper(subcmd: &SafekeeperCmd, env: &local_env::LocalEnv) ->
             let safekeeper = get_safekeeper(env, args.id)?;
 
             if let Err(e) = safekeeper.start(&args.extra_opt, &args.start_timeout).await {
-                eprintln!("safekeeper start failed: {}", e);
+                eprintln!("safekeeper start failed: {e}");
                 exit(1);
             }
         }
@@ -1839,7 +1855,7 @@ async fn handle_safekeeper(subcmd: &SafekeeperCmd, env: &local_env::LocalEnv) ->
                 StopMode::Immediate => true,
             };
             if let Err(e) = safekeeper.stop(immediate) {
-                eprintln!("safekeeper stop failed: {}", e);
+                eprintln!("safekeeper stop failed: {e}");
                 exit(1);
             }
         }
@@ -1852,12 +1868,12 @@ async fn handle_safekeeper(subcmd: &SafekeeperCmd, env: &local_env::LocalEnv) ->
             };
 
             if let Err(e) = safekeeper.stop(immediate) {
-                eprintln!("safekeeper stop failed: {}", e);
+                eprintln!("safekeeper stop failed: {e}");
                 exit(1);
             }
 
             if let Err(e) = safekeeper.start(&args.extra_opt, &args.start_timeout).await {
-                eprintln!("safekeeper start failed: {}", e);
+                eprintln!("safekeeper start failed: {e}");
                 exit(1);
             }
         }
@@ -2113,7 +2129,7 @@ async fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
 
     let storage = EndpointStorage::from_env(env);
     if let Err(e) = storage.stop(immediate) {
-        eprintln!("endpoint_storage stop failed: {:#}", e);
+        eprintln!("endpoint_storage stop failed: {e:#}");
     }
 
     for ps_conf in &env.pageservers {

control_plane/src/endpoint.rs

@@ -56,8 +56,8 @@ use compute_api::responses::{
     TlsConfig,
 };
 use compute_api::spec::{
-    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PgIdent,
-    RemoteExtSpec, Role,
+    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PageserverProtocol,
+    PgIdent, RemoteExtSpec, Role,
 };
 use jsonwebtoken::jwk::{
     AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
@@ -67,6 +67,7 @@ use nix::sys::signal::{Signal, kill};
 use pageserver_api::shard::ShardStripeSize;
 use pem::Pem;
 use reqwest::header::CONTENT_TYPE;
+use safekeeper_api::PgMajorVersion;
 use safekeeper_api::membership::SafekeeperGeneration;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
@@ -89,7 +90,7 @@ pub struct EndpointConf {
     pg_port: u16,
     external_http_port: u16,
     internal_http_port: u16,
-    pg_version: u32,
+    pg_version: PgMajorVersion,
     grpc: bool,
     skip_pg_catalog_updates: bool,
     reconfigure_concurrency: usize,
@@ -192,7 +193,7 @@ impl ComputeControlPlane {
         pg_port: Option<u16>,
         external_http_port: Option<u16>,
         internal_http_port: Option<u16>,
-        pg_version: u32,
+        pg_version: PgMajorVersion,
         mode: ComputeMode,
         grpc: bool,
         skip_pg_catalog_updates: bool,
@@ -312,7 +313,7 @@ pub struct Endpoint {
     pub internal_http_address: SocketAddr,
 
     // postgres major version in the format: 14, 15, etc.
-    pg_version: u32,
+    pg_version: PgMajorVersion,
 
     // These are not part of the endpoint as such, but the environment
     // the endpoint runs in.
@@ -372,27 +373,20 @@ impl std::fmt::Display for EndpointTerminateMode {
     }
 }
 
-/// Protocol used to connect to a Pageserver.
-#[derive(Clone, Copy, Debug)]
-pub enum PageserverProtocol {
-    Libpq,
-    Grpc,
-}
-
-impl PageserverProtocol {
-    /// Returns the URL scheme for the protocol, used in connstrings.
-    pub fn scheme(&self) -> &'static str {
-        match self {
-            Self::Libpq => "postgresql",
-            Self::Grpc => "grpc",
-        }
-    }
-}
-
-impl Display for PageserverProtocol {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.scheme())
-    }
+pub struct EndpointStartArgs {
+    pub auth_token: Option<String>,
+    pub endpoint_storage_token: String,
+    pub endpoint_storage_addr: String,
+    pub safekeepers_generation: Option<SafekeeperGeneration>,
+    pub safekeepers: Vec<NodeId>,
+    pub pageservers: Vec<(PageserverProtocol, Host, u16)>,
+    pub remote_ext_base_url: Option<String>,
+    pub shard_stripe_size: usize,
+    pub create_test_user: bool,
+    pub start_timeout: Duration,
+    pub autoprewarm: bool,
+    pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+    pub dev: bool,
 }
 
 impl Endpoint {
@@ -557,7 +551,7 @@ impl Endpoint {
                 conf.append("hot_standby", "on");
                 // prefetching of blocks referenced in WAL doesn't make sense for us
                 // Neon hot standby ignores pages that are not in the shared_buffers
-                if self.pg_version >= 15 {
+                if self.pg_version >= PgMajorVersion::PG15 {
                     conf.append("recovery_prefetch", "off");
                 }
             }
@@ -699,21 +693,7 @@ impl Endpoint {
         })
     }
 
-    #[allow(clippy::too_many_arguments)]
-    pub async fn start(
-        &self,
-        auth_token: &Option<String>,
-        endpoint_storage_token: String,
-        endpoint_storage_addr: String,
-        safekeepers_generation: Option<SafekeeperGeneration>,
-        safekeepers: Vec<NodeId>,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
-        remote_ext_base_url: Option<&String>,
-        shard_stripe_size: usize,
-        create_test_user: bool,
-        start_timeout: Duration,
-        dev: bool,
-    ) -> Result<()> {
+    pub async fn start(&self, args: EndpointStartArgs) -> Result<()> {
         if self.status() == EndpointStatus::Running {
             anyhow::bail!("The endpoint is already running");
         }
@@ -726,10 +706,10 @@ impl Endpoint {
             std::fs::remove_dir_all(self.pgdata())?;
         }
 
-        let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
+        let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
         assert!(!pageserver_connstring.is_empty());
 
-        let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
+        let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;
 
         // check for file remote_extensions_spec.json
         // if it is present, read it and pass to compute_ctl
@@ -757,7 +737,7 @@ impl Endpoint {
                     cluster_id: None, // project ID: not used
                     name: None,       // project name: not used
                     state: None,
-                    roles: if create_test_user {
+                    roles: if args.create_test_user {
                         vec![Role {
                             name: PgIdent::from_str("test").unwrap(),
                             encrypted_password: None,
@@ -766,7 +746,7 @@ impl Endpoint {
                     } else {
                         Vec::new()
                     },
-                    databases: if create_test_user {
+                    databases: if args.create_test_user {
                         vec![Database {
                             name: PgIdent::from_str("neondb").unwrap(),
                             owner: PgIdent::from_str("test").unwrap(),
@@ -788,20 +768,22 @@ impl Endpoint {
                 endpoint_id: Some(self.endpoint_id.clone()),
                 mode: self.mode,
                 pageserver_connstring: Some(pageserver_connstring),
-                safekeepers_generation: safekeepers_generation.map(|g| g.into_inner()),
+                safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
                 safekeeper_connstrings,
-                storage_auth_token: auth_token.clone(),
+                storage_auth_token: args.auth_token.clone(),
                 remote_extensions,
                 pgbouncer_settings: None,
-                shard_stripe_size: Some(shard_stripe_size),
+                shard_stripe_size: Some(args.shard_stripe_size),
                 local_proxy_config: None,
                 reconfigure_concurrency: self.reconfigure_concurrency,
                 drop_subscriptions_before_start: self.drop_subscriptions_before_start,
                 audit_log_level: ComputeAudit::Disabled,
                 logs_export_host: None::<String>,
-                endpoint_storage_addr: Some(endpoint_storage_addr),
-                endpoint_storage_token: Some(endpoint_storage_token),
-                autoprewarm: false,
+                endpoint_storage_addr: Some(args.endpoint_storage_addr),
+                endpoint_storage_token: Some(args.endpoint_storage_token),
+                autoprewarm: args.autoprewarm,
+                offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
+                suspend_timeout_seconds: -1, // Only used in neon_local.
             };
 
             // this strange code is needed to support respec() in tests
@@ -812,7 +794,7 @@ impl Endpoint {
                 debug!("spec.cluster {:?}", spec.cluster);
 
                 // fill missing fields again
-                if create_test_user {
+                if args.create_test_user {
                     spec.cluster.roles.push(Role {
                         name: PgIdent::from_str("test").unwrap(),
                         encrypted_password: None,
@@ -846,10 +828,10 @@ impl Endpoint {
 
         // Launch compute_ctl
         let conn_str = self.connstr("cloud_admin", "postgres");
-        println!("Starting postgres node at '{}'", conn_str);
-        if create_test_user {
+        println!("Starting postgres node at '{conn_str}'");
+        if args.create_test_user {
             let conn_str = self.connstr("test", "neondb");
-            println!("Also at '{}'", conn_str);
+            println!("Also at '{conn_str}'");
         }
         let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
         cmd.args([
@@ -879,11 +861,11 @@ impl Endpoint {
         .stderr(logfile.try_clone()?)
         .stdout(logfile);
 
-        if let Some(remote_ext_base_url) = remote_ext_base_url {
-            cmd.args(["--remote-ext-base-url", remote_ext_base_url]);
+        if let Some(remote_ext_base_url) = args.remote_ext_base_url {
+            cmd.args(["--remote-ext-base-url", &remote_ext_base_url]);
         }
 
-        if dev {
+        if args.dev {
             cmd.arg("--dev");
         }
 
@@ -915,10 +897,11 @@ impl Endpoint {
                 Ok(state) => {
                     match state.status {
                         ComputeStatus::Init => {
-                            if Instant::now().duration_since(start_at) > start_timeout {
+                            let timeout = args.start_timeout;
+                            if Instant::now().duration_since(start_at) > timeout {
                                 bail!(
                                     "compute startup timed out {:?}; still in Init state",
-                                    start_timeout
+                                    timeout
                                 );
                             }
                             // keep retrying
@@ -946,10 +929,10 @@ impl Endpoint {
                     }
                 }
                 Err(e) => {
-                    if Instant::now().duration_since(start_at) > start_timeout {
+                    if Instant::now().duration_since(start_at) > args.start_timeout {
                         return Err(e).context(format!(
                             "timed out {:?} waiting to connect to compute_ctl HTTP",
-                            start_timeout,
+                            args.start_timeout
                         ));
                     }
                 }
@@ -988,7 +971,7 @@ impl Endpoint {
             // reqwest does not export its error construction utility functions, so let's craft the message ourselves
             let url = response.url().to_owned();
             let msg = match response.text().await {
-                Ok(err_body) => format!("Error: {}", err_body),
+                Ok(err_body) => format!("Error: {err_body}"),
                 Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
             };
             Err(anyhow::anyhow!(msg))
@@ -997,12 +980,11 @@ impl Endpoint {
 
     pub async fn reconfigure(
         &self,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+        pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
         stripe_size: Option<ShardStripeSize>,
         safekeepers: Option<Vec<NodeId>>,
+        safekeeper_generation: Option<SafekeeperGeneration>,
     ) -> Result<()> {
-        anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
-
         let (mut spec, compute_ctl_config) = {
             let config_path = self.endpoint_path().join("config.json");
             let file = std::fs::File::open(config_path)?;
@@ -1014,16 +996,24 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
-        spec.pageserver_connstring = Some(pageserver_connstr);
-        if stripe_size.is_some() {
-            spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+        // If pageservers are not specified, don't change them.
+        if let Some(pageservers) = pageservers {
+            anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
+
+            let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
+            spec.pageserver_connstring = Some(pageserver_connstr);
+            if stripe_size.is_some() {
+                spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+            }
         }
 
         // If safekeepers are not specified, don't change them.
         if let Some(safekeepers) = safekeepers {
             let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
             spec.safekeeper_connstrings = safekeeper_connstrings;
+            if let Some(g) = safekeeper_generation {
+                spec.safekeepers_generation = Some(g.into_inner());
+            }
         }
 
         let client = reqwest::Client::builder()
@@ -1054,13 +1044,31 @@ impl Endpoint {
         } else {
             let url = response.url().to_owned();
             let msg = match response.text().await {
-                Ok(err_body) => format!("Error: {}", err_body),
+                Ok(err_body) => format!("Error: {err_body}"),
                 Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
             };
             Err(anyhow::anyhow!(msg))
         }
     }
 
+    pub async fn reconfigure_pageservers(
+        &self,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> Result<()> {
+        self.reconfigure(Some(pageservers), stripe_size, None, None)
+            .await
+    }
+
+    pub async fn reconfigure_safekeepers(
+        &self,
+        safekeepers: Vec<NodeId>,
+        generation: SafekeeperGeneration,
+    ) -> Result<()> {
+        self.reconfigure(None, None, Some(safekeepers), Some(generation))
+            .await
+    }
+
     pub async fn stop(
         &self,
         mode: EndpointTerminateMode,

control_plane/src/local_env.rs

@@ -12,9 +12,11 @@ use std::{env, fs};
 
 use anyhow::{Context, bail};
 use clap::ValueEnum;
+use pageserver_api::config::PostHogConfig;
 use pem::Pem;
 use postgres_backend::AuthType;
 use reqwest::{Certificate, Url};
+use safekeeper_api::PgMajorVersion;
 use serde::{Deserialize, Serialize};
 use utils::auth::encode_from_key_file;
 use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
@@ -210,7 +212,11 @@ pub struct NeonStorageControllerConf {
 
     pub use_local_compute_notifications: bool,
 
-    pub timeline_safekeeper_count: Option<i64>,
+    pub timeline_safekeeper_count: Option<usize>,
+
+    pub posthog_config: Option<PostHogConfig>,
+
+    pub kick_secondary_downloads: Option<bool>,
 }
 
 impl NeonStorageControllerConf {
@@ -242,6 +248,8 @@ impl Default for NeonStorageControllerConf {
             use_https_safekeeper_api: false,
             use_local_compute_notifications: true,
             timeline_safekeeper_count: None,
+            posthog_config: None,
+            kick_secondary_downloads: None,
         }
     }
 }
@@ -257,7 +265,7 @@ impl Default for EndpointStorageConf {
 impl NeonBroker {
     pub fn client_url(&self) -> Url {
         let url = if let Some(addr) = self.listen_https_addr {
-            format!("https://{}", addr)
+            format!("https://{addr}")
         } else {
             format!(
                 "http://{}",
@@ -421,25 +429,21 @@ impl LocalEnv {
         self.pg_distrib_dir.clone()
     }
 
-    pub fn pg_distrib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_distrib_dir(&self, pg_version: PgMajorVersion) -> anyhow::Result<PathBuf> {
         let path = self.pg_distrib_dir.clone();
 
-        #[allow(clippy::manual_range_patterns)]
-        match pg_version {
-            14 | 15 | 16 | 17 => Ok(path.join(format!("v{pg_version}"))),
-            _ => bail!("Unsupported postgres version: {}", pg_version),
-        }
+        Ok(path.join(pg_version.v_str()))
     }
 
-    pub fn pg_dir(&self, pg_version: u32, dir_name: &str) -> anyhow::Result<PathBuf> {
+    pub fn pg_dir(&self, pg_version: PgMajorVersion, dir_name: &str) -> anyhow::Result<PathBuf> {
         Ok(self.pg_distrib_dir(pg_version)?.join(dir_name))
     }
 
-    pub fn pg_bin_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_bin_dir(&self, pg_version: PgMajorVersion) -> anyhow::Result<PathBuf> {
         self.pg_dir(pg_version, "bin")
     }
 
-    pub fn pg_lib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
+    pub fn pg_lib_dir(&self, pg_version: PgMajorVersion) -> anyhow::Result<PathBuf> {
         self.pg_dir(pg_version, "lib")
     }
 
@@ -730,7 +734,7 @@ impl LocalEnv {
                 let config_toml_path = dentry.path().join("pageserver.toml");
                 let config_toml: PageserverConfigTomlSubset = toml_edit::de::from_str(
                     &std::fs::read_to_string(&config_toml_path)
-                        .with_context(|| format!("read {:?}", config_toml_path))?,
+                        .with_context(|| format!("read {config_toml_path:?}"))?,
                 )
                 .context("parse pageserver.toml")?;
                 let identity_toml_path = dentry.path().join("identity.toml");
@@ -740,7 +744,7 @@ impl LocalEnv {
                 }
                 let identity_toml: IdentityTomlSubset = toml_edit::de::from_str(
                     &std::fs::read_to_string(&identity_toml_path)
-                        .with_context(|| format!("read {:?}", identity_toml_path))?,
+                        .with_context(|| format!("read {identity_toml_path:?}"))?,
                 )
                 .context("parse identity.toml")?;
                 let PageserverConfigTomlSubset {

control_plane/src/pageserver.rs

@@ -22,6 +22,7 @@ use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api;
 use postgres_backend::AuthType;
 use postgres_connection::{PgConnectionConfig, parse_host_port};
+use safekeeper_api::PgMajorVersion;
 use utils::auth::{Claims, Scope};
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
@@ -121,7 +122,7 @@ impl PageServerNode {
                 .env
                 .generate_auth_token(&Claims::new(None, Scope::GenerationsApi))
                 .unwrap();
-            overrides.push(format!("control_plane_api_token='{}'", jwt_token));
+            overrides.push(format!("control_plane_api_token='{jwt_token}'"));
         }
 
         if !conf.other.contains_key("remote_storage") {
@@ -451,6 +452,12 @@ impl PageServerNode {
                 .map(|x| x.parse::<usize>())
                 .transpose()
                 .context("Failed to parse 'image_creation_threshold' as non zero integer")?,
+            // HADRON
+            image_layer_force_creation_period: settings
+                .remove("image_layer_force_creation_period")
+                .map(humantime::parse_duration)
+                .transpose()
+                .context("Failed to parse 'image_layer_force_creation_period' as duration")?,
             image_layer_creation_check_threshold: settings
                 .remove("image_layer_creation_check_threshold")
                 .map(|x| x.parse::<u8>())
@@ -607,7 +614,7 @@ impl PageServerNode {
         timeline_id: TimelineId,
         base: (Lsn, PathBuf),
         pg_wal: Option<(Lsn, PathBuf)>,
-        pg_version: u32,
+        pg_version: PgMajorVersion,
     ) -> anyhow::Result<()> {
         // Init base reader
         let (start_lsn, base_tarfile_path) = base;

control_plane/src/safekeeper.rs

@@ -143,7 +143,7 @@ impl SafekeeperNode {
         let id_string = id.to_string();
         // TODO: add availability_zone to the config.
         // Right now we just specify any value here and use it to check metrics in tests.
-        let availability_zone = format!("sk-{}", id_string);
+        let availability_zone = format!("sk-{id_string}");
 
         let mut args = vec![
             "-D".to_owned(),

control_plane/src/storage_controller.rs

@@ -6,6 +6,8 @@ use std::str::FromStr;
 use std::sync::OnceLock;
 use std::time::{Duration, Instant};
 
+use crate::background_process;
+use crate::local_env::{LocalEnv, NeonStorageControllerConf};
 use camino::{Utf8Path, Utf8PathBuf};
 use hyper0::Uri;
 use nix::unistd::Pid;
@@ -22,6 +24,7 @@ use pageserver_client::mgmt_api::ResponseErrorMessageExt;
 use pem::Pem;
 use postgres_backend::AuthType;
 use reqwest::{Method, Response};
+use safekeeper_api::PgMajorVersion;
 use serde::de::DeserializeOwned;
 use serde::{Deserialize, Serialize};
 use tokio::process::Command;
@@ -31,9 +34,6 @@ use utils::auth::{Claims, Scope, encode_from_key_file};
 use utils::id::{NodeId, TenantId};
 use whoami::username;
 
-use crate::background_process;
-use crate::local_env::{LocalEnv, NeonStorageControllerConf};
-
 pub struct StorageController {
     env: LocalEnv,
     private_key: Option<Pem>,
@@ -48,7 +48,7 @@ pub struct StorageController {
 
 const COMMAND: &str = "storage_controller";
 
-const STORAGE_CONTROLLER_POSTGRES_VERSION: u32 = 16;
+const STORAGE_CONTROLLER_POSTGRES_VERSION: PgMajorVersion = PgMajorVersion::PG16;
 
 const DB_NAME: &str = "storage_controller";
 
@@ -167,7 +167,7 @@ impl StorageController {
     fn storage_controller_instance_dir(&self, instance_id: u8) -> PathBuf {
         self.env
             .base_data_dir
-            .join(format!("storage_controller_{}", instance_id))
+            .join(format!("storage_controller_{instance_id}"))
     }
 
     fn pid_file(&self, instance_id: u8) -> Utf8PathBuf {
@@ -184,9 +184,15 @@ impl StorageController {
     /// to other versions if that one isn't found.  Some automated tests create circumstances
     /// where only one version is available in pg_distrib_dir, such as `test_remote_extensions`.
     async fn get_pg_dir(&self, dir_name: &str) -> anyhow::Result<Utf8PathBuf> {
-        let prefer_versions = [STORAGE_CONTROLLER_POSTGRES_VERSION, 16, 15, 14];
+        const PREFER_VERSIONS: [PgMajorVersion; 5] = [
+            STORAGE_CONTROLLER_POSTGRES_VERSION,
+            PgMajorVersion::PG16,
+            PgMajorVersion::PG15,
+            PgMajorVersion::PG14,
+            PgMajorVersion::PG17,
+        ];
 
-        for v in prefer_versions {
+        for v in PREFER_VERSIONS {
             let path = Utf8PathBuf::from_path_buf(self.env.pg_dir(v, dir_name)?).unwrap();
             if tokio::fs::try_exists(&path).await? {
                 return Ok(path);
@@ -220,7 +226,7 @@ impl StorageController {
             "-d",
             DB_NAME,
             "-p",
-            &format!("{}", postgres_port),
+            &format!("{postgres_port}"),
         ];
         let pg_lib_dir = self.get_pg_lib_dir().await.unwrap();
         let envs = [
@@ -263,7 +269,7 @@ impl StorageController {
                 "-h",
                 "localhost",
                 "-p",
-                &format!("{}", postgres_port),
+                &format!("{postgres_port}"),
                 "-U",
                 &username(),
                 "-O",
@@ -425,7 +431,7 @@ impl StorageController {
             // from `LocalEnv`'s config file (`.neon/config`).
             tokio::fs::write(
                 &pg_data_path.join("postgresql.conf"),
-                format!("port = {}\nfsync=off\n", postgres_port),
+                format!("port = {postgres_port}\nfsync=off\n"),
             )
             .await?;
 
@@ -477,7 +483,7 @@ impl StorageController {
             self.setup_database(postgres_port).await?;
         }
 
-        let database_url = format!("postgresql://localhost:{}/{DB_NAME}", postgres_port);
+        let database_url = format!("postgresql://localhost:{postgres_port}/{DB_NAME}");
 
         // We support running a startup SQL script to fiddle with the database before we launch storcon.
         // This is used by the test suite.
@@ -508,7 +514,7 @@ impl StorageController {
         drop(client);
         conn.await??;
 
-        let addr = format!("{}:{}", host, listen_port);
+        let addr = format!("{host}:{listen_port}");
         let address_for_peers = Uri::builder()
             .scheme(scheme)
             .authority(addr.clone())
@@ -557,6 +563,10 @@ impl StorageController {
             args.push("--use-local-compute-notifications".to_string());
         }
 
+        if let Some(value) = self.config.kick_secondary_downloads {
+            args.push(format!("--kick-secondary-downloads={value}"));
+        }
+
         if let Some(ssl_ca_file) = self.env.ssl_ca_cert_path() {
             args.push(format!("--ssl-ca-file={}", ssl_ca_file.to_str().unwrap()));
         }
@@ -628,10 +638,28 @@ impl StorageController {
             args.push("--timelines-onto-safekeepers".to_string());
         }
 
-        if let Some(sk_cnt) = self.config.timeline_safekeeper_count {
+        // neon_local is used in test environments where we often have less than 3 safekeepers.
+        if self.config.timeline_safekeeper_count.is_some() || self.env.safekeepers.len() < 3 {
+            let sk_cnt = self
+                .config
+                .timeline_safekeeper_count
+                .unwrap_or(self.env.safekeepers.len());
+
             args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
         }
 
+        let mut envs = vec![
+            ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
+            ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
+        ];
+
+        if let Some(posthog_config) = &self.config.posthog_config {
+            envs.push((
+                "POSTHOG_CONFIG".to_string(),
+                serde_json::to_string(posthog_config)?,
+            ));
+        }
+
         println!("Starting storage controller");
 
         background_process::start_process(
@@ -639,10 +667,7 @@ impl StorageController {
             &instance_dir,
             &self.env.storage_controller_bin(),
             args,
-            vec![
-                ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
-                ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
-            ],
+            envs,
             background_process::InitialPidFile::Create(self.pid_file(start_args.instance_id)),
             &start_args.start_timeout,
             || async {
@@ -806,9 +831,9 @@ impl StorageController {
             builder = builder.json(&body)
         }
         if let Some(private_key) = &self.private_key {
-            println!("Getting claims for path {}", path);
+            println!("Getting claims for path {path}");
             if let Some(required_claims) = Self::get_claims_for_path(&path)? {
-                println!("Got claims {:?} for path {}", required_claims, path);
+                println!("Got claims {required_claims:?} for path {path}");
                 let jwt_token = encode_from_key_file(&required_claims, private_key)?;
                 builder = builder.header(
                     reqwest::header::AUTHORIZATION,

control_plane/storcon_cli/src/main.rs

@@ -65,12 +65,27 @@ enum Command {
         #[arg(long)]
         scheduling: Option<NodeSchedulingPolicy>,
     },
-    // Set a node status as deleted.
+    /// Exists for backup usage and will be removed in future.
+    /// Use [`Command::NodeStartDelete`] instead, if possible.
     NodeDelete {
         #[arg(long)]
         node_id: NodeId,
     },
+    /// Start deletion of the specified pageserver.
+    NodeStartDelete {
+        #[arg(long)]
+        node_id: NodeId,
+    },
+    /// Cancel deletion of the specified pageserver and wait for `timeout`
+    /// for the operation to be canceled. May be retried.
+    NodeCancelDelete {
+        #[arg(long)]
+        node_id: NodeId,
+        #[arg(long)]
+        timeout: humantime::Duration,
+    },
     /// Delete a tombstone of node from the storage controller.
+    /// This is used when we want to allow the node to be re-registered.
     NodeDeleteTombstone {
         #[arg(long)]
         node_id: NodeId,
@@ -649,7 +664,7 @@ async fn main() -> anyhow::Result<()> {
                 response
                     .new_shards
                     .iter()
-                    .map(|s| format!("{:?}", s))
+                    .map(|s| format!("{s:?}"))
                     .collect::<Vec<_>>()
                     .join(",")
             );
@@ -771,8 +786,8 @@ async fn main() -> anyhow::Result<()> {
 
             println!("Tenant {tenant_id}");
             let mut table = comfy_table::Table::new();
-            table.add_row(["Policy", &format!("{:?}", policy)]);
-            table.add_row(["Stripe size", &format!("{:?}", stripe_size)]);
+            table.add_row(["Policy", &format!("{policy:?}")]);
+            table.add_row(["Stripe size", &format!("{stripe_size:?}")]);
             table.add_row(["Config", &serde_json::to_string_pretty(&config).unwrap()]);
             println!("{table}");
             println!("Shards:");
@@ -789,7 +804,7 @@ async fn main() -> anyhow::Result<()> {
                 let secondary = shard
                     .node_secondary
                     .iter()
-                    .map(|n| format!("{}", n))
+                    .map(|n| format!("{n}"))
                     .collect::<Vec<_>>()
                     .join(",");
 
@@ -863,7 +878,7 @@ async fn main() -> anyhow::Result<()> {
                 }
             } else {
                 // Make it obvious to the user that since they've omitted an AZ, we're clearing it
-                eprintln!("Clearing preferred AZ for tenant {}", tenant_id);
+                eprintln!("Clearing preferred AZ for tenant {tenant_id}");
             }
 
             // Construct a request that modifies all the tenant's shards
@@ -912,10 +927,43 @@ async fn main() -> anyhow::Result<()> {
                 .await?;
         }
         Command::NodeDelete { node_id } => {
+            eprintln!("Warning: This command is obsolete and will be removed in a future version");
+            eprintln!("Use `NodeStartDelete` instead, if possible");
             storcon_client
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
+        Command::NodeStartDelete { node_id } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::PUT,
+                    format!("control/v1/node/{node_id}/delete"),
+                    None,
+                )
+                .await?;
+            println!("Delete started for {node_id}");
+        }
+        Command::NodeCancelDelete { node_id, timeout } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::DELETE,
+                    format!("control/v1/node/{node_id}/delete"),
+                    None,
+                )
+                .await?;
+
+            println!("Waiting for node {node_id} to quiesce on scheduling policy ...");
+
+            let final_policy =
+                wait_for_scheduling_policy(storcon_client, node_id, *timeout, |sched| {
+                    !matches!(sched, NodeSchedulingPolicy::Deleting)
+                })
+                .await?;
+
+            println!(
+                "Delete was cancelled for node {node_id}. Schedulling policy is now {final_policy:?}"
+            );
+        }
         Command::NodeDeleteTombstone { node_id } => {
             storcon_client
                 .dispatch::<(), ()>(
@@ -1134,8 +1182,7 @@ async fn main() -> anyhow::Result<()> {
                     Err((tenant_shard_id, from, to, error)) => {
                         failure += 1;
                         println!(
-                            "Failed to migrate {} from node {} to node {}: {}",
-                            tenant_shard_id, from, to, error
+                            "Failed to migrate {tenant_shard_id} from node {from} to node {to}: {error}"
                         );
                     }
                 }
@@ -1277,8 +1324,7 @@ async fn main() -> anyhow::Result<()> {
             concurrency,
         } => {
             let mut path = format!(
-                "/v1/tenant/{}/timeline/{}/download_heatmap_layers",
-                tenant_shard_id, timeline_id,
+                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
             );
 
             if let Some(c) = concurrency {
@@ -1303,8 +1349,7 @@ async fn watch_tenant_shard(
 ) -> anyhow::Result<()> {
     if let Some(until_migrated_to) = until_migrated_to {
         println!(
-            "Waiting for tenant shard {} to be migrated to node {}",
-            tenant_shard_id, until_migrated_to
+            "Waiting for tenant shard {tenant_shard_id} to be migrated to node {until_migrated_to}"
         );
     }
 
@@ -1327,7 +1372,7 @@ async fn watch_tenant_shard(
             "attached: {} secondary: {} {}",
             shard
                 .node_attached
-                .map(|n| format!("{}", n))
+                .map(|n| format!("{n}"))
                 .unwrap_or("none".to_string()),
             shard
                 .node_secondary
@@ -1341,15 +1386,12 @@ async fn watch_tenant_shard(
                 "(reconciler idle)"
             }
         );
-        println!("{}", summary);
+        println!("{summary}");
 
         // Maybe drop out if we finished migration
         if let Some(until_migrated_to) = until_migrated_to {
             if shard.node_attached == Some(until_migrated_to) && !shard.is_reconciling {
-                println!(
-                    "Tenant shard {} is now on node {}",
-                    tenant_shard_id, until_migrated_to
-                );
+                println!("Tenant shard {tenant_shard_id} is now on node {until_migrated_to}");
                 break;
             }
         }

docker-compose/compute_wrapper/shell/compute.sh

@@ -54,14 +54,16 @@ else
     printf '%s\n' "${result}" | jq .
   fi
 
-  echo "Check if a timeline present"
-  PARAMS=(
-       -X GET
-       -H "Content-Type: application/json"
-       "http://pageserver:9898/v1/tenant/${tenant_id}/timeline"
-  )
-  timeline_id=$(curl "${PARAMS[@]}" | jq -r .[0].timeline_id)
-  if [[ -z "${timeline_id}" || "${timeline_id}" = null ]]; then
+  if [[ "${RUN_PARALLEL:-false}" != "true" ]]; then
+    echo "Check if a timeline present"
+    PARAMS=(
+         -X GET
+         -H "Content-Type: application/json"
+        "http://pageserver:9898/v1/tenant/${tenant_id}/timeline"
+    )
+    timeline_id=$(curl "${PARAMS[@]}" | jq -r .[0].timeline_id)
+  fi
+  if [[ -z "${timeline_id:-}" || "${timeline_id:-}" = null ]]; then
     generate_id timeline_id
     PARAMS=(
         -sbf

docker-compose/compute_wrapper/var/db/postgres/configs/config.json

@@ -4,6 +4,7 @@
 
         "timestamp": "2022-10-12T18:00:00.000Z",
         "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8c",
+        "suspend_timeout_seconds": -1,
 
         "cluster": {
             "cluster_id": "docker_compose",

docker-compose/docker-compose.yml

@@ -142,7 +142,7 @@ services:
       - "storage_broker"
       - "--listen-addr=0.0.0.0:50051"
 
-  compute:
+  compute1:
     restart: always
     build:
       context: ./compute_wrapper/
@@ -152,6 +152,7 @@ services:
         - TAG=${COMPUTE_TAG:-${TAG:-latest}}
         - http_proxy=${http_proxy:-}
         - https_proxy=${https_proxy:-}
+    image: built-compute
     environment:
       - PG_VERSION=${PG_VERSION:-16}
       - TENANT_ID=${TENANT_ID:-}
@@ -166,6 +167,11 @@ services:
       - 3080:3080 # http endpoints
     entrypoint:
       - "/shell/compute.sh"
+    # Ad an alias for compute1 for compatibility
+    networks:
+      default:
+        aliases:
+            - compute
     depends_on:
       - safekeeper1
       - safekeeper2
@@ -174,15 +180,20 @@ services:
 
   compute_is_ready:
     image: postgres:latest
+    environment:
+      - PARALLEL_COMPUTES=1
     entrypoint:
-      - "/bin/bash"
+      - "/bin/sh"
       - "-c"
     command:
-      - "until pg_isready -h compute -p 55433 -U cloud_admin ; do
-            echo 'Waiting to start compute...' && sleep 1;
-         done"
+      - "for i in $(seq 1 $${PARALLEL_COMPUTES}); do
+           until pg_isready -h compute$$i -p 55433 -U cloud_admin ; do
+             sleep 1;
+           done;
+         done;
+         echo All computes are started"
     depends_on:
-      - compute
+      - compute1
 
   neon-test-extensions:
     profiles: ["test-extensions"]
@@ -196,4 +207,4 @@ services:
     command:
       - sleep 3600
     depends_on:
-      - compute
+      - compute1

docker-compose/docker_compose_test.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # A basic test to ensure Docker images are built correctly.
 # Build a wrapper around the compute, start all services and runs a simple SQL query.
@@ -13,9 +13,36 @@
 #
 set -eux -o pipefail
 
+cd "$(dirname "${0}")"
 export COMPOSE_FILE='docker-compose.yml'
 export COMPOSE_PROFILES=test-extensions
-cd "$(dirname "${0}")"
+export PARALLEL_COMPUTES=${PARALLEL_COMPUTES:-1}
+READY_MESSAGE="All computes are started"
+COMPUTES=()
+for i in $(seq 1 "${PARALLEL_COMPUTES}"); do
+  COMPUTES+=("compute${i}")
+done
+CURRENT_TMPDIR=$(mktemp -d)
+trap 'rm -rf ${CURRENT_TMPDIR} docker-compose-parallel.yml' EXIT
+if [[ ${PARALLEL_COMPUTES} -gt 1 ]]; then
+  export COMPOSE_FILE=docker-compose-parallel.yml
+  cp docker-compose.yml docker-compose-parallel.yml
+  # Replace the environment variable PARALLEL_COMPUTES with the actual value
+  yq eval -i ".services.compute_is_ready.environment |=  map(select(. | test(\"^PARALLEL_COMPUTES=\") | not)) + [\"PARALLEL_COMPUTES=${PARALLEL_COMPUTES}\"]" ${COMPOSE_FILE}
+  for i in $(seq 2 "${PARALLEL_COMPUTES}"); do
+    # Duplicate compute1 as compute${i} for parallel execution
+    yq eval -i ".services.compute${i} = .services.compute1" ${COMPOSE_FILE}
+    # We don't need these sections, so delete them
+    yq eval -i "(del .services.compute${i}.build) | (del .services.compute${i}.ports) | (del .services.compute${i}.networks)" ${COMPOSE_FILE}
+    # Let the compute 1 be the only dependence
+    yq eval -i ".services.compute${i}.depends_on = [\"compute1\"]" ${COMPOSE_FILE}
+    # Set RUN_PARALLEL=true for compute2. They will generate tenant_id and timeline_id to avoid using the same as other computes
+    yq eval -i ".services.compute${i}.environment += [\"RUN_PARALLEL=true\"]" ${COMPOSE_FILE}
+    # Remove TENANT_ID and TIMELINE_ID from the environment variables of the generated computes
+    # They will create new TENANT_ID and TIMELINE_ID anyway.
+    yq eval -i ".services.compute${i}.environment |= map(select(. | (test(\"^TENANT_ID=\") or test(\"^TIMELINE_ID=\")) | not))" ${COMPOSE_FILE}
+  done
+fi
 PSQL_OPTION="-h localhost -U cloud_admin -p 55433 -d postgres"
 
 function cleanup() {
@@ -27,11 +54,11 @@ function cleanup() {
 
 for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
     pg_version=${pg_version/v/}
-    echo "clean up containers if exists"
+    echo "clean up containers if exist"
     cleanup
     PG_TEST_VERSION=$((pg_version < 16 ? 16 : pg_version))
-    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose up --quiet-pull --build -d
-
+    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose build compute1
+    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose up --quiet-pull -d
     echo "wait until the compute is ready. timeout after 60s. "
     cnt=0
     while sleep 3; do
@@ -41,45 +68,50 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
             echo "timeout before the compute is ready."
             exit 1
         fi
-        if docker compose logs "compute_is_ready" | grep -q "accepting connections"; then
+        if docker compose logs compute_is_ready | grep -q "${READY_MESSAGE}"; then
             echo "OK. The compute is ready to connect."
             echo "execute simple queries."
-            docker compose exec compute /bin/bash -c "psql ${PSQL_OPTION} -c 'SELECT 1'"
+            for compute in "${COMPUTES[@]}"; do
+              docker compose exec "${compute}" /bin/bash -c "psql ${PSQL_OPTION} -c 'SELECT 1'"
+            done
             break
         fi
     done
 
     if [[ ${pg_version} -ge 16 ]]; then
-        # This is required for the pg_hint_plan test, to prevent flaky log message causing the test to fail
-        # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
-        echo Adding dummy config
-        docker compose exec compute touch /var/db/postgres/compute/compute_ctl_temp_override.conf
-        # Prepare for the PostGIS test
-        docker compose exec compute mkdir -p /tmp/pgis_reg/pgis_reg_tmp
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/ext-src/postgis-src/raster/test "${TMPDIR}"
-        docker compose cp neon-test-extensions:/ext-src/postgis-src/regress/00-regress-install "${TMPDIR}"
-        docker compose exec compute mkdir -p /ext-src/postgis-src/raster /ext-src/postgis-src/regress /ext-src/postgis-src/regress/00-regress-install
-        docker compose cp "${TMPDIR}/test" compute:/ext-src/postgis-src/raster/test
-        docker compose cp "${TMPDIR}/00-regress-install" compute:/ext-src/postgis-src/regress
-        rm -rf "${TMPDIR}"
-        # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/ext-src/pg_hint_plan-src/data "${TMPDIR}/data"
-        docker compose cp "${TMPDIR}/data" compute:/ext-src/pg_hint_plan-src/
-        rm -rf "${TMPDIR}"
-        # The following block does the same for the contrib/file_fdw test
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/postgres/contrib/file_fdw/data "${TMPDIR}/data"
-        docker compose cp "${TMPDIR}/data" compute:/postgres/contrib/file_fdw/data
-        rm -rf "${TMPDIR}"
+        mkdir "${CURRENT_TMPDIR}"/{pg_hint_plan-src,file_fdw,postgis-src}
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/raster/test "${CURRENT_TMPDIR}/postgis-src/test"
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/regress/00-regress-install "${CURRENT_TMPDIR}/postgis-src/00-regress-install"
+        docker compose cp neon-test-extensions:/ext-src/pg_hint_plan-src/data "${CURRENT_TMPDIR}/pg_hint_plan-src/data"
+        docker compose cp neon-test-extensions:/postgres/contrib/file_fdw/data "${CURRENT_TMPDIR}/file_fdw/data"
+
+        for compute in "${COMPUTES[@]}"; do
+          # This is required for the pg_hint_plan test, to prevent flaky log message causing the test to fail
+          # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
+          echo Adding dummy config on "${compute}"
+          docker compose exec "${compute}" touch /var/db/postgres/compute/compute_ctl_temp_override.conf
+          # Prepare for the PostGIS test
+          docker compose exec "${compute}" mkdir -p /tmp/pgis_reg/pgis_reg_tmp /ext-src/postgis-src/raster /ext-src/postgis-src/regress /ext-src/postgis-src/regress/00-regress-install
+          docker compose cp "${CURRENT_TMPDIR}/postgis-src/test" "${compute}":/ext-src/postgis-src/raster/test
+          docker compose cp "${CURRENT_TMPDIR}/postgis-src/00-regress-install" "${compute}":/ext-src/postgis-src/regress
+          # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
+          docker compose cp "${CURRENT_TMPDIR}/pg_hint_plan-src/data" "${compute}":/ext-src/pg_hint_plan-src/
+          # The following block does the same for the contrib/file_fdw test
+          docker compose cp "${CURRENT_TMPDIR}/file_fdw/data" "${compute}":/postgres/contrib/file_fdw/data
+        done
         # Apply patches
         docker compose exec -T neon-test-extensions bash -c "(cd /postgres && patch -p1)" <"../compute/patches/contrib_pg${pg_version}.patch"
         # We are running tests now
         rm -f testout.txt testout_contrib.txt
+        # We want to run the longest tests first to better utilize parallelization and reduce overall test time.
+        # Tests listed in the RUN_FIRST variable will be run before others.
+        # If parallelization is not used, this environment variable will be ignored.
+
         docker compose exec -e USE_PGXS=1 -e SKIP=timescaledb-src,rdkit-src,pg_jsonschema-src,kq_imcx-src,wal2json_2_5-src,rag_jina_reranker_v1_tiny_en-src,rag_bge_small_en_v15-src \
+        -e RUN_FIRST=hll-src,postgis-src,pgtap-src -e PARALLEL_COMPUTES="${PARALLEL_COMPUTES}" \
         neon-test-extensions /run-tests.sh /ext-src | tee testout.txt && EXT_SUCCESS=1 || EXT_SUCCESS=0
         docker compose exec -e SKIP=start-scripts,postgres_fdw,ltree_plpython,jsonb_plpython,jsonb_plperl,hstore_plpython,hstore_plperl,dblink,bool_plperl \
+        -e PARALLEL_COMPUTES="${PARALLEL_COMPUTES}" \
         neon-test-extensions /run-tests.sh /postgres/contrib | tee testout_contrib.txt && CONTRIB_SUCCESS=1 || CONTRIB_SUCCESS=0
         if [[ ${EXT_SUCCESS} -eq 0 || ${CONTRIB_SUCCESS} -eq 0 ]]; then
             CONTRIB_FAILED=

docker-compose/run-tests.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -x
 
 if [[ -v BENCHMARK_CONNSTR ]]; then
@@ -26,8 +26,9 @@ if [[ -v BENCHMARK_CONNSTR ]]; then
   fi
 fi
 REGULAR_USER=false
-while getopts r arg; do
-  case $arg in
+PARALLEL_COMPUTES=${PARALLEL_COMPUTES:-1}
+while getopts pr arg; do
+  case ${arg} in
   r)
     REGULAR_USER=true
     shift $((OPTIND-1))
@@ -41,26 +42,49 @@ extdir=${1}
 
 cd "${extdir}" || exit 2
 FAILED=
-LIST=$( (echo -e "${SKIP//","/"\n"}"; ls) | sort | uniq -u)
-for d in ${LIST}; do
-    [ -d "${d}" ] || continue
-    if ! psql -w -c "select 1" >/dev/null; then
-      FAILED="${d} ${FAILED}"
-      break
-    fi
-    if [[ ${REGULAR_USER} = true ]] && [ -f "${d}"/regular-test.sh ]; then
-       "${d}/regular-test.sh" || FAILED="${d} ${FAILED}"
-       continue
-    fi
+export FAILED_FILE=/tmp/failed
+rm -f ${FAILED_FILE}
+mapfile -t LIST < <( (echo -e "${SKIP//","/"\n"}"; ls) | sort | uniq -u)
+if [[ ${PARALLEL_COMPUTES} -gt 1 ]]; then
+  # Avoid errors if RUN_FIRST is not defined
+  RUN_FIRST=${RUN_FIRST:-}
+  # Move entries listed in the RUN_FIRST variable to the beginning
+  ORDERED_LIST=$(printf "%s\n" "${LIST[@]}" | grep -x -Ff <(echo -e "${RUN_FIRST//,/$'\n'}"); printf "%s\n" "${LIST[@]}" | grep -vx -Ff <(echo -e "${RUN_FIRST//,/$'\n'}"))
+  parallel -j"${PARALLEL_COMPUTES}" "[[ -d {} ]] || exit 0
+                export PGHOST=compute{%}
+                if ! psql -c 'select 1'>/dev/null; then
+                  exit 1
+                fi
+                echo Running on \${PGHOST}
+                if [[ -f ${extdir}/{}/neon-test.sh ]]; then
+                  echo Running from script
+                  ${extdir}/{}/neon-test.sh || echo {} >> ${FAILED_FILE};
+                else
+                  echo Running using make;
+                  USE_PGXS=1 make -C {} installcheck || echo {} >> ${FAILED_FILE};
+                fi" ::: ${ORDERED_LIST}
+  [[ ! -f ${FAILED_FILE} ]] && exit 0
+else
+  for d in "${LIST[@]}"; do
+      [ -d "${d}" ] || continue
+      if ! psql -w -c "select 1" >/dev/null; then
+        FAILED="${d} ${FAILED}"
+        break
+      fi
+      if [[ ${REGULAR_USER} = true ]] && [ -f "${d}"/regular-test.sh ]; then
+        "${d}/regular-test.sh" || FAILED="${d} ${FAILED}"
+        continue
+      fi
 
-    if [ -f "${d}/neon-test.sh" ]; then
-       "${d}/neon-test.sh" || FAILED="${d} ${FAILED}"
-    else
-       USE_PGXS=1 make -C "${d}" installcheck || FAILED="${d} ${FAILED}"
-    fi
-done
-[ -z "${FAILED}" ] && exit 0
-for d in ${FAILED}; do
+      if [ -f "${d}/neon-test.sh" ]; then
+        "${d}/neon-test.sh" || FAILED="${d} ${FAILED}"
+      else
+        USE_PGXS=1 make -C "${d}" installcheck || FAILED="${d} ${FAILED}"
+      fi
+  done
+  [[ -z ${FAILED} ]]  && exit 0
+fi
+for d in ${FAILED} $([[ ! -f ${FAILED_FILE} ]] || cat ${FAILED_FILE}); do
   cat "$(find $d -name regression.diffs)"
 done
 for postgis_diff in /tmp/pgis_reg/*_diff; do
@@ -68,4 +92,5 @@ for postgis_diff in /tmp/pgis_reg/*_diff; do
   cat "${postgis_diff}"
 done
 echo "${FAILED}"
+cat ${FAILED_FILE}
 exit 1

docker-compose/test_extensions_upgrade.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -eux -o pipefail
 cd "$(dirname "${0}")"
 # Takes a variable name as argument. The result is stored in that variable.
@@ -60,8 +60,8 @@ function check_timeline() {
 # Restarts the compute node with the required compute tag and timeline.
 # Accepts the tag for the compute node and the timeline as parameters.
 function restart_compute() {
-  docker compose down compute compute_is_ready
-  COMPUTE_TAG=${1} TENANT_ID=${tenant_id} TIMELINE_ID=${2} docker compose up --quiet-pull -d --build compute compute_is_ready
+  docker compose down compute1 compute_is_ready
+  COMPUTE_TAG=${1} TENANT_ID=${tenant_id} TIMELINE_ID=${2} docker compose up --quiet-pull -d --build compute1 compute_is_ready
   wait_for_ready
   check_timeline ${2}
 }

docs/rfcs/035-safekeeper-dynamic-membership-change.md

@@ -20,7 +20,7 @@ In our case consensus leader is compute (walproposer), and we don't want to wake
 up all computes for the change. Neither we want to fully reimplement the leader
 logic second time outside compute. Because of that the proposed algorithm relies
 for issuing configurations on the external fault tolerant (distributed) strongly
-consisent storage with simple API: CAS (compare-and-swap) on the single key.
+consistent storage with simple API: CAS (compare-and-swap) on the single key.
 Properly configured postgres suits this.
 
 In the system consensus is implemented at the timeline level, so algorithm below
@@ -34,7 +34,7 @@ A configuration is
 
 ```
 struct Configuration {
-    generation: Generation, // a number uniquely identifying configuration
+    generation: SafekeeperGeneration, // a number uniquely identifying configuration
     sk_set: Vec<NodeId>, // current safekeeper set
     new_sk_set: Optional<Vec<NodeId>>,
 }
@@ -81,11 +81,11 @@ configuration generation in them is less than its current one. Namely, it
 refuses to vote, to truncate WAL in `handle_elected` and to accept WAL. In
 response it sends its current configuration generation to let walproposer know.
 
-Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration`
-accepting `Configuration`. Safekeeper switches to the given conf it is higher than its
+Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/membership`
+accepting `Configuration`. Safekeeper switches to the given conf if it is higher than its
 current one and ignores it otherwise. In any case it replies with
 ```
-struct ConfigurationSwitchResponse {
+struct TimelineMembershipSwitchResponse {
     conf: Configuration,
     term: Term,
     last_log_term: Term,
@@ -108,7 +108,7 @@ establishes this configuration as its own and moves to voting.
 It should stop talking to safekeepers not listed in the configuration at this
 point, though it is not unsafe to continue doing so.
 
-To be elected it must receive votes from both majorites if `new_sk_set` is present.
+To be elected it must receive votes from both majorities if `new_sk_set` is present.
 Similarly, to commit WAL it must receive flush acknowledge from both majorities.
 
 If walproposer hears from safekeeper configuration higher than his own (i.e.
@@ -130,7 +130,7 @@ storage are reachable.
 1) Fetch current timeline configuration from the configuration storage.
 2) If it is already joint one and `new_set` is different from `desired_set`
    refuse to change. However, assign join conf to (in memory) var
-   `join_conf` and proceed to step 4 to finish the ongoing change.
+   `joint_conf` and proceed to step 4 to finish the ongoing change.
 3) Else, create joint `joint_conf: Configuration`: increment current conf number
    `n` and put `desired_set` to `new_sk_set`. Persist it in the configuration
    storage by doing CAS on the current generation: change happens only if
@@ -161,11 +161,11 @@ storage are reachable.
    because `pull_timeline` already includes it and plus additionally would be
    broadcast by compute. More importantly, we may proceed to the next step
    only when `<last_log_term, flush_lsn>` on the majority of the new set reached
-   `sync_position`. Similarly, on the happy path no waiting is not needed because
+   `sync_position`. Similarly, on the happy path no waiting is needed because
    `pull_timeline` already includes it. However, we should double
     check to be safe. For example, timeline could have been created earlier e.g.
     manually or after try-to-migrate, abort, try-to-migrate-again sequence.
-7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new
+7) Create `new_conf: Configuration` incrementing `joint_conf` generation and having new
    safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration
    storage under one more CAS.
 8) Call `PUT` `configuration` on safekeepers from the new set,
@@ -178,12 +178,12 @@ spec of it.
 
 Description above focuses on safety. To make the flow practical and live, here a few more
 considerations.
-1) It makes sense to ping new set to ensure it we are migrating to live node(s) before
+1) It makes sense to ping new set to ensure we are migrating to live node(s) before
   step 3.
 2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed
    it is safe to rollback to the old conf with one more CAS.
 3) On step 4 timeline might be already created on members of the new set for various reasons;
-   the simplest is the procedure restart. There are more complicated scenarious like mentioned
+   the simplest is the procedure restart. There are more complicated scenarios like mentioned
    in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving
    generations, so seems simpler to treat existing timeline as success. However, this also
    has a disadvantage: you might imagine an surpassingly unlikely schedule where condition in
@@ -192,7 +192,7 @@ considerations.
 4) In the end timeline should be locally deleted on the safekeeper(s) which are
    in the old set but not in the new one, unless they are unreachable. To be
    safe this also should be done under generation number (deletion proceeds only if
-   current configuration is <= than one in request and safekeeper is not memeber of it).
+   current configuration is <= than one in request and safekeeper is not member of it).
 5) If current conf fetched on step 1 is already not joint and members equal to `desired_set`,
    jump to step 7, using it as `new_conf`.
 
@@ -261,14 +261,14 @@ Timeline (branch) creation in cplane should call storage_controller POST
 Response should be augmented with `safekeepers_generation` and `safekeepers`
 fields like described in `/notify-safekeepers` above. Initially (currently)
 these fields may be absent; in this case cplane chooses safekeepers on its own
-like it currently does. The call should be retried until succeeds.
+like it currently does. The call should be retried until it succeeds.
 
 Timeline deletion and tenant deletion in cplane should call appropriate
 storage_controller endpoints like it currently does for sharded tenants. The
 calls should be retried until they succeed.
 
-When compute receives safekeepers list from control plane it needs to know the
-generation to checked whether it should be updated (note that compute may get
+When compute receives safekeeper list from control plane it needs to know the
+generation to check whether it should be updated (note that compute may get
 safekeeper list from either cplane or safekeepers). Currently `neon.safekeepers`
 GUC is just a comma separates list of `host:port`. Let's prefix it with
 `g#<generation>:` to this end, so it will look like
@@ -305,8 +305,8 @@ enum MigrationRequest {
 ```
 
 `FinishPending` requests to run the procedure to ensure state is clean: current
-configuration is not joint and majority of safekeepers are aware of it, but do
-not attempt to migrate anywhere. If current configuration fetched on step 1 is
+configuration is not joint and the majority of safekeepers are aware of it, but do
+not attempt to migrate anywhere. If the current configuration fetched on step 1 is
 not joint it jumps to step 7. It should be run at startup for all timelines (but
 similarly, in the first version it is ok to trigger it manually).
 
@@ -315,7 +315,7 @@ similarly, in the first version it is ok to trigger it manually).
 `safekeepers` table mirroring current `nodes` should be added, except that for
 `scheduling_policy`: it is enough to have at least in the beginning only 3
 fields: 1) `active` 2) `paused` (initially means only not assign new tlis there
-3) `decomissioned` (node is removed).
+3) `decommissioned` (node is removed).
 
 `timelines` table:
 ```
@@ -326,9 +326,10 @@ table! {
         tenant_id -> Varchar,
         start_lsn -> pg_lsn,
         generation -> Int4,
-        sk_set -> Array<Int4>, // list of safekeeper ids
+        sk_set -> Array<Int8>, // list of safekeeper ids
         new_sk_set -> Nullable<Array<Int8>>, // list of safekeeper ids, null if not joint conf
         cplane_notified_generation -> Int4,
+        sk_set_notified_generation -> Int4, // the generation a quorum of sk_set knows about
         deleted_at -> Nullable<Timestamptz>,
     }
 }
@@ -338,13 +339,23 @@ table! {
 might also want to add ancestor_timeline_id to preserve the hierarchy, but for
 this RFC it is not needed.
 
+`cplane_notified_generation` and `sk_set_notified_generation` fields are used to
+track the last stage of the algorithm, when we need to notify safekeeper set and cplane
+with the final configuration after it's already committed to DB.
+
+The timeline is up-to-date (no migration in progress) if `new_sk_set` is null and
+`*_notified_generation` fields are up to date with `generation`. 
+
+It's possible to replace `*_notified_generation` with one boolean field `migration_completed`,
+but for better observability it's nice to have them separately.
+
 #### API
 
 Node management is similar to pageserver:
-1) POST `/control/v1/safekeepers` inserts safekeeper.
-2) GET `/control/v1/safekeepers` lists safekeepers.
-3) GET `/control/v1/safekeepers/:node_id` gets safekeeper.
-4) PUT `/control/v1/safekepers/:node_id/status` changes status to e.g.
+1) POST `/control/v1/safekeeper` inserts safekeeper.
+2) GET `/control/v1/safekeeper` lists safekeepers.
+3) GET `/control/v1/safekeeper/:node_id` gets safekeeper.
+4) PUT `/control/v1/safekeper/:node_id/scheduling_policy` changes status to e.g.
    `offline` or `decomissioned`. Initially it is simpler not to schedule any
     migrations here.
 
@@ -368,8 +379,8 @@ Migration API: the first version is the simplest and the most imperative:
 all timelines from one safekeeper to another. It accepts json
 ```
 {
-    "src_sk": u32,
-    "dst_sk": u32,
+    "src_sk": NodeId,
+    "dst_sk": NodeId,
     "limit": Optional<u32>,
 }
 ```
@@ -379,12 +390,15 @@ Returns list of scheduled requests.
 2) PUT `/control/v1/tenant/:tenant_id/timeline/:timeline_id/safekeeper_migrate` schedules `MigrationRequest`
    to move single timeline to given set of safekeepers:
 ```
-{
-    "desired_set": Vec<u32>,
+struct TimelineSafekeeperMigrateRequest {
+    "new_sk_set": Vec<NodeId>,
 }
 ```
 
-Returns scheduled request.
+In the first version the handler migrates the timeline to `new_sk_set` synchronously.
+Should be retried until success.
+
+In the future we might change it to asynchronous API and return scheduled request.
 
 Similar call should be added for the tenant.
 
@@ -434,6 +448,9 @@ table! {
 }
 ```
 
+We load all pending ops from the table on startup into the memory.
+The table is needed only to preserve the state between restarts.
+
 `op_type` can be `include` (seed from peers and ensure generation is up to
 date), `exclude` (remove locally) and `delete`. Field is actually not strictly
 needed as it can be computed from current configuration, but gives more explicit
@@ -474,7 +491,7 @@ actions must be idempotent. Now, a tricky point here is timeline start LSN. For
 the initial (tenant creation) call cplane doesn't know it. However, setting
 start_lsn on safekeepers during creation is a good thing -- it provides a
 guarantee that walproposer can always find a common point in WAL histories of
-safekeeper and its own, and so absense of it would be a clear sign of
+safekeeper and its own, and so absence of it would be a clear sign of
 corruption. The following sequence works:
 1) Create timeline (or observe that it exists) on pageserver,
    figuring out last_record_lsn in response.
@@ -497,11 +514,9 @@ corruption. The following sequence works:
    retries the call until 200 response.
 
    There is a small question how request handler (timeline creation in this
-   case) would interact with per sk reconciler. As always I prefer to do the
-   simplest possible thing and here it seems to be just waking it up so it
-   re-reads the db for work to do. Passing work in memory is faster, but
-   that shouldn't matter, and path to scan db for work will exist anyway, 
-   simpler to reuse it.
+   case) would interact with per sk reconciler. In the current implementation
+   we first persist the request in the DB, and then send an in-memory request
+   to each safekeeper reconciler to process it.
 
 For pg version / wal segment size: while we may persist them in `timelines`
 table, it is not necessary as initial creation at step 3 can take them from
@@ -509,30 +524,40 @@ pageserver or cplane creation call and later pull_timeline will carry them
 around.
 
 Timeline migration.
-1) CAS to the db to create joint conf, and in the same transaction create
-   `safekeeper_timeline_pending_ops` `include` entries to initialize new members
-   as well as deliver this conf to current ones; poke per sk reconcilers to work
-   on it. Also any conf change should also poke cplane notifier task(s).
-2) Once it becomes possible per alg description above, get out of joint conf
-   with another CAS. Task should get wakeups from per sk reconcilers because 
-   conf switch is required for advancement; however retries should be sleep
-   based as well as LSN advancement might be needed, though in happy path 
-   it isn't. To see whether further transition is possible on wakup migration
-   executor polls safekeepers per the algorithm. CAS creating new conf with only
-   new members should again insert entries to `safekeeper_timeline_pending_ops`
-   to switch them there, as well as `exclude` rows to remove timeline from 
-   old members.
+1) CAS to the db to create joint conf. Since this moment the migration is considered to be 
+   "in progress". We can detect all "in-progress" migrations looking into the database.
+2) Do steps 4-6 from the algorithm, including `pull_timeline` onto `new_sk_set`, update membership
+   configuration on all safekeepers, notify cplane, etc. All operations are idempotent,
+   so we don't need to persist anything in the database at this stage. If any errors occur,
+   it's safe to retry or abort the migration.
+3) Once it becomes possible per alg description above, get out of joint conf
+   with another CAS. Also should insert `exclude` entries into `safekeeper_timeline_pending_ops`
+   in the same DB transaction. Adding `exclude` entries atomically is nesessary because after
+   CAS we don't have the list of excluded safekeepers in the `timelines` table anymore, but we
+   need to have them persisted somewhere in case the migration is interrupted right after the CAS.
+4) Finish the migration. The final membership configuration is committed to the DB at this stage.
+   So, the migration can not be aborted anymore. But it can still be retried if the migration fails
+   past stage 3. To finish the migration we need to send the new membership configuration to
+   a new quorum of safekeepers, notify cplane with the new safekeeper list and schedule the `exclude`
+   requests to in-memory queue for safekeeper reconciler. If the algrorithm is retried, it's
+   possible that we have already committed `exclude` requests to DB, but didn't send them to
+   the in-memory queue. In this case we need to read them from `safekeeper_timeline_pending_ops`
+   because it's the only place where they are persistent. The fields `sk_set_notified_generation`
+   and `cplane_notified_generation` are updated after each step. The migration is considered
+   fully completed when they match the `generation` field.
+
+In practice, we can report "success" after stage 3 and do the "finish" step in per-timeline
+reconciler (if we implement it). But it's wise to at least try to finish them synchronously,
+so the timeline is always in a "good state" and doesn't require an old quorum to commit
+WAL after the migration reported "success".
 
 Timeline deletion: just set `deleted_at` on the timeline row and insert
 `safekeeper_timeline_pending_ops` entries in the same xact, the rest is done by
 per sk reconcilers.
 
-When node is removed (set to `decomissioned`), `safekeeper_timeline_pending_ops`
+When node is removed (set to `decommissioned`), `safekeeper_timeline_pending_ops`
 for it must be cleared in the same transaction.
 
-One more task pool should infinitely retry notifying control plane about changed
-safekeeper sets (trying making `cplane_notified_generation` equal `generation`).
-
 #### Dealing with multiple instances of storage_controller
 
 Operations described above executed concurrently might create some errors but do
@@ -541,7 +566,7 @@ of storage_controller it is fine to have it temporarily, e.g. during redeploy.
 
 To harden against some controller instance creating some work in
 `safekeeper_timeline_pending_ops` and then disappearing without anyone pickup up
-the job per sk reconcilers apart from explicit wakups should scan for work
+the job per sk reconcilers apart from explicit wakeups should scan for work
 periodically. It is possible to remove that though if all db updates are
 protected with leadership token/term -- then such scans are needed only after
 leadership is acquired.
@@ -563,7 +588,7 @@ There should be following layers of tests:
    safekeeper communication and pull_timeline need to be mocked and main switch
    procedure wrapped to as a node (thread) in simulation tests, using these
    mocks. Test would inject migrations like it currently injects
-   safekeeper/walproposer restars. Main assert is the same -- committed WAL must
+   safekeeper/walproposer restarts. Main assert is the same -- committed WAL must
    not be lost.
 
 3) Since simulation testing injects at relatively high level points (not
@@ -613,7 +638,7 @@ Let's have the following implementation bits for gradual rollout:
   `notify-safekeepers`.
 
 Then the rollout for a region would be:
-- Current situation: safekeepers are choosen by control_plane.
+- Current situation: safekeepers are chosen by control_plane.
 - We manually migrate some timelines, test moving them around.
 - Then we enable `--set-safekeepers` so that all new timelines
   are on storage controller.

docs/rfcs/040-Endpoint-Persistent-Unlogged-Files-Storage.md

@@ -0,0 +1,396 @@
+# Memo: Endpoint Persistent Unlogged Files Storage
+Created on 2024-11-05
+Implemented on N/A
+
+## Summary
+A design for a storage system that allows storage of files required to make
+Neon's Endpoints have a better experience at or after a reboot.
+
+## Motivation
+Several systems inside PostgreSQL (and Neon) need some persistent storage for
+optimal workings across reboots and restarts, but still work without.
+Examples are the query-level statistics files of `pg_stat_statements` in
+`pg_stat/pg_stat_statements.stat`, and `pg_prewarm`'s `autoprewarm.blocks`.
+We need a storage system that can store and manage these files for each
+Endpoint, without necessarily granting users access to an unlimited storage
+device.
+
+## Goals
+- Store known files for Endpoints with reasonable persistence.  
+  _Data loss in this service, while annoying and bad for UX, won't lose any
+  customer's data._
+
+## Non Goals (if relevant)
+- This storage system does not need branching, file versioning, or other such
+  features. The files are as ephemeral to the timeline of the data as the
+  Endpoints that host the data.
+- This storage system does not need to store _all_ user files, only 'known'
+  user files.
+- This storage system does not need to be hosted fully inside Computes.  
+  _Instead, this will be a separate component similar to Pageserver,
+  SafeKeeper, the S3 proxy used for dynamically loaded extensions, etc._
+
+## Impacted components
+- Compute needs new code to load and store these files in its lifetime.
+- Control Plane needs to consider this new storage system when signalling
+  the deletion of an Endpoint, Timeline, or Tenant.
+- Control Plane needs to consider this new storage system when it resets
+  or re-assigns an endpoint's timeline/branch state.
+
+A new service is created: the Endpoint Persistent Unlogged Files Storage
+service.  This could be integrated in e.g. Pageserver or Control Plane, or a
+separately hosted service.
+
+## Proposed implementation
+Endpoint-related data files are managed by a newly designed service (which
+optionally is integrated in an existing service like Pageserver or Control
+Plane), which stores data directly into S3 or any blob storage of choice.
+
+Upon deletion of the Endpoint, or reassignment of the endpoint to a different
+branch, this ephemeral data is dropped: the data stored may not match the
+state of the branch's data after reassignment, and on endpoint deletion the
+data won't have any use to the user.
+
+Compute gets credentials (JWT token with Tenant, Timeline & Endpoint claims)
+which it can use to authenticate to this new service and retrieve and store
+data associated with this endpoint.  This limited scope reduces leaks of data
+across endpoints and timeline resets, and limits the ability of endpoints to
+mess with other endpoints' data.
+
+The path of this endpoint data in S3 is initially as follows:
+
+    s3://<regional-epufs-bucket>/
+      tenants/
+        <hex-tenant-id>/
+          tenants/
+            <hex-timeline-id>/
+              endpoints/
+                <endpoint-id>/
+                  pgdata/
+                    <file_path_in_pgdatadir>
+
+For other blob storages an equivalent or similar path can be constructed.
+
+### Reliability, failure modes and corner cases (if relevant)
+Reliability is important, but not critical to the workings of Neon.  The data
+stored in this service will, when lost, reduce performance, but won't be a
+cause of permanent data loss - only operational metadata is stored.
+
+Most, if not all, blob storage services have sufficiently high persistence
+guarantees to cater our need for persistence and uptime. The only concern with
+blob storages is that the access latency is generally higher than local disk,
+but for the object types stored (cache state, ...) I don't think this will be
+much of an issue.
+
+### Interaction/Sequence diagram (if relevant)
+
+In these diagrams you can replace S3 with any persistent storage device of
+choice, but S3 is chosen as representative name: The well-known and short name
+of AWS' blob storage. Azure Blob Storage should work too, but it has a much
+longer name making it less practical for the diagrams.
+
+Write data:
+
+```http
+POST /tenants/<tenant-id>/timelines/<tl-id>/endpoints/<endpoint-id>/pgdata/<the-pgdata-path>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "version": "<opaque>", # opaque file version token, changes when the file contents change
+  "size": <bytes>,
+}
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    co-->ep: Connect with credentials
+    co->>+ep: Store Unlogged Persistent File
+    opt is authenticated
+        ep->>s3: Write UPF to S3
+    end
+    ep->>-co: OK / Failure / Auth Failure
+    co-->ep: Cancel connection
+```
+
+Read data: (optional with cache-relevant request parameters, e.g. If-Modified-Since)
+```http
+GET /tenants/<tenant-id>/timelines/<tl-id>/endpoints/<endpoint-id>/pgdata/<the-pgdata-path>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+
+<file data>
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    co->>+ep: Read Unlogged Persistent File
+    opt is authenticated
+        ep->>+s3: Request UPF from storage
+        s3->>-ep: Receive UPF from storage
+    end
+    ep->>-co: OK(response) / Failure(storage, auth, ...)
+```
+
+Compute Startup:
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ps as Pageserver
+    participant ep as EPUFS
+    participant es as Extension server
+
+    note over co: Bind endpoint ep-xxx
+    par Get basebackup
+        co->>+ps: Request basebackup @ LSN
+        ps-)ps: Construct basebackup
+        ps->>-co: Receive basebackup TAR @ LSN
+    and Get startup-critical Unlogged Persistent Files
+        co->>+ep: Get all UPFs of endpoint ep-xxx
+        ep-)ep: Retrieve and gather all UPFs
+        ep->>-co: TAR of UPFs
+    and Get startup-critical extensions
+        loop For every startup-critical extension
+            co->>es: Get critical extension
+            es->>co: Receive critical extension
+        end
+    end
+    note over co: Start compute
+```
+
+CPlane ops:
+```http
+DELETE /tenants/<tenant-id>/timelines/<timeline-id>/endpoints/<endpoint-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "timeline": "<timeline-id>",
+  "endpoint": "<endpoint-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```http
+DELETE /tenants/<tenant-id>/timelines/<timeline-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "timeline": "<timeline-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```http
+DELETE /tenants/<tenant-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant cp as Control Plane
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    alt Tenant deleted
+        cp-)ep: Tenant deleted
+        loop For every object associated with removed tenant
+            ep->>s3: Remove data of deleted tenant from Storage
+        end
+        opt
+            ep-)cp: Tenant cleanup complete
+        end
+    alt Timeline deleted
+        cp-)ep: Timeline deleted
+        loop For every object associated with removed timeline
+            ep->>s3: Remove data of deleted timeline from Storage
+        end
+        opt
+            ep-)cp: Timeline cleanup complete
+        end
+    else Endpoint reassigned or removed
+        cp->>+ep: Endpoint reassigned
+        loop For every object associated with reassigned/removed endpoint
+            ep->>s3: Remove data from Storage
+        end
+        ep->>-cp: Cleanup complete
+    end
+```
+
+### Scalability (if relevant)
+
+Provisionally:  As this service is going to be part of compute startup, this
+service should be able to quickly respond to all requests.  Therefore this
+service is deployed to every AZ we host Computes in, and Computes communicate
+(generally) only to the EPUFS endpoint of the AZ they're hosted in.
+
+Local caching of frequently restarted endpoints' data or metadata may be
+needed for best performance.  However, due to the regional nature of stored
+data but zonal nature of the service deployment, we should be careful when we
+implement any local caching, as it is possible that computes in AZ 1 will
+update data originally written and thus cached by AZ 2.  Cache version tests
+and invalidation is therefore required if we want to roll out caching to this
+service, which is too broad a scope for an MVC.  This is why caching is left
+out of scope for this RFC, and should be considered separately after this RFC
+is implemented.
+
+### Security implications (if relevant)
+This service must be able to authenticate users at least by Tenant ID,
+Timeline ID and Endpoint ID. This will use the existing JWT infrastructure of
+Compute, which will be upgraded to the extent needed to support Timeline- and
+Endpoint-based claims.
+
+The service requires unlimited access to (a prefix of) a blob storage bucket,
+and thus must be hosted outside the Compute VM sandbox.
+
+A service that generates pre-signed request URLs for Compute to download the
+data from that URL is likely problematic, too:  Compute would be able to write
+unlimited data to the bucket, or exfiltrate this signed URL to get read/write
+access to specific objects in this bucket, which would still effectively give
+users access to the S3 bucket (but with improved access logging).
+
+There may be a use case for transferring data associated with one endpoint to
+another endpoint (e.g. to make one endpoint warm its caches with the state of
+another endpoint), but that's not currently in scope, and specific needs may
+be solved through out-of-line communication of data or pre-signed URLs.
+
+### Unresolved questions (if relevant)
+Caching of files is not in the implementation scope of the document, but
+should at some future point be considered to maximize performance.
+
+## Alternative implementation (if relevant)
+Several ideas have come up to solve this issue:
+
+### Use AUXfile
+One prevalent idea was to WAL-log the files using our AUXfile mechanism.
+
+Benefits:
+
++ We already have this storage mechanism
+
+Demerits:
+
+- It isn't available on read replicas
+- Additional WAL will be consumed during shutdown and after the shutdown
+  checkpoint, which needs PG modifications to work without panics.
+- It increases the data we need to manage in our versioned storage, thus
+  causing higher storage costs with higher retention due to duplication at
+  the storage layer.
+
+### Sign URLs for read/write operations, instead of proxying them
+
+Benefits:
+
++ The service can be implemented with a much reduced IO budget
+
+Demerits:
+
+- Users could get access to these signed credentials
+- Not all blob storage services may implement URL signing
+
+### Give endpoints each their own directly accessed block volume
+
+Benefits:
+
++ Easier to integrate for PostgreSQL
+
+Demerits:
+
+- Little control on data size and contents
+- Potentially problematic as we'd need to store data all across the pgdata
+  directory.
+- EBS is not a good candidate
+   - Attaches in 10s of seconds, if not more; i.e. too cold to start
+   - Shared EBS volumes are a no-go, as you'd have to schedule the endpoint
+     with users of the same EBS volumes, which can't work with VM migration
+   - EBS storage costs are very high (>80$/kilotenant when using a
+     volume/tenant)
+   - EBS volumes can't be mounted across AZ boundaries
+- Bucket per endpoint is unfeasible
+   - S3 buckets are priced at $20/month per 1k, which we could better spend
+     on developers.
+   - Allocating service accounts takes time (100s of ms), and service accounts
+     are a limited resource, too; so they're not a good candidate to allocate
+     on a per-endpoint basis.
+   - Giving credentials limited to prefix has similar issues as the pre-signed
+     URL approach.
+   - Bucket DNS lookup will fill DNS caches and put pressure on DNS lookup
+     much more than our current systems would.
+- Volumes bound by hypervisor are unlikely
+   - This requires significant investment and increased software on the
+     hypervisor.
+   - It is unclear if we can attach volumes after boot, i.e. for pooled
+     instances.
+
+### Put the files into a table
+
+Benefits:
+
+ + Mostly already available in PostgreSQL
+
+Demerits:
+
+ - Uses WAL
+   - Can't be used after shutdown checkpoint
+   - Needs a RW endpoint, and table & catalog access to write to this data
+ - Gets hit with DB size limitations
+ - Depending on user acces:
+   - Inaccessible:  
+     The user doesn't have control over database size caused by
+     these systems.
+   - Accessible:  
+     The user can corrupt these files and cause the system to crash while
+     user-corrupted files are present, thus increasing on-call overhead.
+
+## Definition of Done (if relevant)
+
+This project is done if we have:
+
+- One S3 bucket equivalent per region, which stores this per-endpoint data.
+- A new service endpoint in at least every AZ, which indirectly grants
+  endpoints access to the data stored for these endpoints in these buckets.
+- Compute writes & reads temp-data at shutdown and startup, respectively, for
+  at least the pg_prewarm or lfc_prewarm state files.
+- Cleanup of endpoint data is triggered when the endpoint is deleted or is
+  detached from its current timeline.

docs/rfcs/044-feature-flag.md

@@ -0,0 +1,179 @@
+# Storage Feature Flags
+
+In this RFC, we will describe how we will implement per-tenant feature flags.
+
+## PostHog as Feature Flag Service
+
+Before we start, let's talk about how current feature flag services work. PostHog is the feature flag service we are currently using across multiple user-facing components in the company. PostHog has two modes of operation: HTTP evaluation and server-side local evaluation.
+
+Let's assume we have a storage feature flag called gc-compaction and we want to roll it out to scale-tier users with resident size >= 10GB and <= 100GB.
+
+### Define User Profiles
+
+The first step is to synchronize our user profiles to the PostHog service. We can simply assume that each tenant is a user in PostHog. Each user profile has some properties associated with it. In our case, it will be: plan type (free, scale, enterprise, etc); resident size (in bytes); primary pageserver (string); region (string).
+
+### Define Feature Flags
+
+We would create a feature flag called gc-compaction in PostHog with 4 variants: disabled, stage-1, stage-2, fully-enabled. We will flip the feature flags from disabled to fully-enabled stage by stage for some percentage of our users.
+
+### Option 1: HTTP Evaluation Mode
+
+When using PostHog's HTTP evaluation mode, the client will make request to the PostHog service, asking for the value of a feature flag for a specific user.
+
+* Control plane will report the plan type to PostHog each time it attaches a tenant to the storcon or when the user upgrades/downgrades. It calls the PostHog profile API to associate tenant ID with the plan type. Assume we have X active tenants and such attach or plan change event happens each week, that would be 4X profile update requests per month.
+* Pageservers will report the resident size and the primary pageserver to the PostHog service. Assume we report resident size every 24 hours, that would be 30X requests per month.
+* Each tenant will request the state of the feature flag every 1 hour, that's 720X requests per month.
+* The Rust client would be easy to implement as we only need to call the `/decide` API on PostHog.
+
+Using the HTTP evaluation mode we will issue 754X requests a month.
+
+### Option 2: Local Evaluation Mode
+
+When using PostHog's HTTP evaluation mode, the client (usually the server in a browser/server architecture) will poll the feature flag configuration every 30s (default in the Python client) from PostHog. Such configuration contains data like:
+
+<details>
+
+<summary>Example JSON response from the PostHog local evaluation API</summary>
+
+```
+[
+    {
+        "id": 1,
+        "name": "Beta Feature",
+        "key": "person-flag",
+        "is_simple_flag": True,
+        "active": True,
+        "filters": {
+            "groups": [
+                {
+                    "properties": [
+                        {
+                            "key": "location",
+                            "operator": "exact",
+                            "value": ["Straße"],
+                            "type": "person",
+                        }
+                    ],
+                    "rollout_percentage": 100,
+                },
+                {
+                    "properties": [
+                        {
+                            "key": "star",
+                            "operator": "exact",
+                            "value": ["ſun"],
+                            "type": "person",
+                        }
+                    ],
+                    "rollout_percentage": 100,
+                },
+            ],
+        },
+    }
+]
+```
+
+</details>
+
+Note that the API only contains information like "under what condition => rollout percentage". The user is responsible to provide the properties required to the client for local evaluation, and the PostHog service (web UI) cannot know if a feature is enabled for the tenant or not until the client uses the `capture` API to report the result back. To control the rollout percentage, the user ID gets mapped to a float number in `[0, 1)` on a consistent hash ring. All values <= the percentage will get the feature enabled or set to the desired value.
+
+To use the local evaluation mode, the system needs:
+
+* Assume each pageserver will poll PostHog for the local evaluation JSON every 5 minutes (instead of the 30s default as it's too frequent). That's 8640Y per month, Y is the number of pageservers. Local evaluation requests cost 10x more than the normal decide request, so that's 86400Y request units to bill.
+* Storcon needs to store the plan type in the database and pass that information to the pageserver when attaching the tenant.
+* Storcon also needs to update PostHog with the active tenants, for example, when the tenant gets detached/attached. Assume each active tenant gets detached/attached every week, that would be 4X requests per month.
+* We do not need to update bill type or resident size to PostHog as all these are evaluated locally.
+* After each local evaluation of the feature flag, we need to call PostHog's capture event API to update the result of the evaluation that the feature is enabled. We can do this when the flag gets changed compared with the last cached state in memory. That would be at least 4X (assume we do deployment every week so the cache gets cleared) and maybe an additional multiplifier of 10 assume we have 10 active features.
+
+In this case, we will issue 86400Y + 40X requests per month.
+
+Assume X = 1,000,000 and Y = 100,
+
+|   | HTTP Evaluation  | Local Evaluation  |
+|---|---|---|
+| Latency of propagating the conditions/properties for feature flag  | 24 hours  | available locally  |
+| Latency of applying the feature flag  | 1 hour  | 5 minutes  |
+| Can properties be reported from different services |  Yes |  No  |
+| Do we need to sync billing info etc to pageserver |  No |  Yes  |
+| Cost | 75400$ / month | 4864$ / month |
+
+# Our Solution
+
+We will use PostHog _only_ as an UI to configure the feature flags. Whether a feature is enabled or not can only be queried through storcon/pageserver instead of using the PostHog UI. (We could report it back to PostHog via `capture_event` but it costs $$$.) This allows us to ramp up the feature flag functionality fast at first. At the same time, it would also give us the option to migrate to our own solution once we want to have more properties and more complex evaluation rules in our system.
+
+* We will create several fake users (tenants) in PostHog that contains all the properties we will use for evaluating a feature flag (i.e., resident size, billing type, pageserver id, etc.)
+* We will use PostHog's local evaluation API to poll the configuration of the feature flags and evaluate them locally on each of the pageserver.
+* The evaluation result will not be reported back to PostHog.
+* Storcon needs to pull some information from cplane database.
+* To know if a feature is currently enabled or not, we need to call the storcon/pageserver API; and we won't be able to know if a feature has been enabled on a tenant before easily: we need to look at the Grafana logs.
+
+We only need to pay for the 86400Y local evaluation requests (that would be setting Y=0 in solution 2 => $864/month, and even less if we proxy it through storcon).
+
+## Implementation
+
+* Pageserver: implement a PostHog local evaluation client. The client will be shared across all tenants on the pageserver with a single API: `evaluate(tenant_id, feature_flag, properties) -> json`.
+* Storcon: if we need plan type as the evaluation condition, pull it from cplane database.
+* Storcon/Pageserver: implement an HTTP API `:tenant_id/feature/:feature` to retrieve the current feature flag status.
+* Storcon/Pageserver: a loop to update the feature flag spec on both storcon and pageserver. Pageserver loop will only be activated if storcon does not push the specs to the pageserver.
+
+## Difference from Tenant Config
+
+* Feature flags can be modified by percentage, and the default config for each feature flag can be modified in UI without going through the release process.
+* Feature flags are more flexible and won't be persisted anywhere and will be passed as plain JSON over the wire so that do not need to handle backward/forward compatibility as in tenant config.
+* The expectation of tenant config is that once we add a flag we cannot remove it (or it will be hard to remove), but feature flags are more flexible.
+
+# Final Implementation
+
+* We added a new crate `posthog_lite_client` that supports local feature evaluations.
+* We set up two projects "Storage (staging)" and "Storage (production)" in the PostHog console.
+* Each pageserver reports 10 fake tenants to PostHog so that we can get all combinations of regions (and other properties) in the PostHog UI.
+* Supported properties: AZ, neon_region, pageserver, tenant_id.
+* You may use "Pageserver Feature Flags" dashboard to see the evaluation status.
+* The feature flag spec is polled on storcon every 30s (in each of the region) and storcon will propagate the spec to the pageservers.
+* The pageserver housekeeping loop updates the tenant-specific properties (e.g., remote size) for evaluation.
+
+Each tenant has a `feature_resolver` object. After you add a feature flag in the PostHog console, you can retrieve it with:
+
+```rust
+// Boolean flag
+self
+    .feature_resolver
+    .evaluate_boolean("flag")
+    .is_ok()
+// Multivariate flag
+self
+    .feature_resolver
+    .evaluate_multivariate("gc-comapction-strategy")
+    .ok();
+```
+
+The user needs to handle the case where the evaluation result is an error. This can occur in a variety of cases:
+
+* During the pageserver start, the feature flag spec has not been retrieved.
+* No condition group is matched.
+* The feature flag spec contains an operand/operation not supported by the lite PostHog library.
+
+For boolean flags, the return value is `Result<(), Error>`. `Ok(())` means the flag is evaluated to true. Otherwise,
+there is either an error in evaluation or it does not match any groups.
+
+For multivariate flags, the return value is `Result<String, Error>`. `Ok(variant)` indicates the flag is evaluated
+to a variant. Otherwise, there is either an error in evaluation or it does not match any groups.
+
+The evaluation logic is documented in the PostHog lite library. It compares the consistent hash of a flag key + tenant_id
+with the rollout percentage and determines which tenant to roll out a specific feature.
+
+Users can use the feature flag evaluation API to get the flag evaluation result of a specific tenant for debugging purposes.
+
+```
+curl http://localhost:9898/v1/tenant/:tenant_id/feature_flag?flag=:key&as=multivariate/boolean"
+```
+
+By default, the storcon pushes the feature flag specs to the pageservers every 30 seconds, which means that a change in feature flag in the
+PostHog UI will propagate to the pageservers within 30 seconds.
+
+# Future Works
+
+* Support dynamic tenant properties like logical size as the evaluation condition.
+* Support properties like `plan_type` (needs cplane to pass it down).
+* Report feature flag evaluation result back to PostHog (if the cost is okay).
+* Fast feature flag evaluation cache on critical paths (e.g., cache a feature flag result in `AtomicBool` and use it on the read path).

docs/rfcs/2025-03-17-compute-prewarm.md

@@ -0,0 +1,399 @@
+# Compute rolling restart with prewarm
+
+Created on 2025-03-17
+Implemented on _TBD_
+Author: Alexey Kondratov (@ololobus)
+
+## Summary
+
+This RFC describes an approach to reduce performance degradation due to missing caches after compute node restart, i.e.:
+
+1. Rolling restart of the running instance via 'warm' replica.
+2. Auto-prewarm compute caches after unplanned restart or scale-to-zero.
+
+## Motivation
+
+Neon currently implements several features that guarantee high uptime of compute nodes:
+
+1. Storage high-availability (HA), i.e. each tenant shard has a secondary pageserver location, so we can quickly switch over compute to it in case of primary pageserver failure.
+2. Fast compute provisioning, i.e. we have a fleet of pre-created empty computes, that are ready to serve workload, so restarting unresponsive compute is very fast.
+3. Preemptive NeonVM compute provisioning in case of k8s node unavailability.
+
+This helps us to be well-within the uptime SLO of 99.95% most of the time. Problems begin when we go up to multi-TB workloads and 32-64 CU computes.
+During restart, compute loses all caches: LFC, shared buffers, file system cache. Depending on the workload, it can take a lot of time to warm up the caches,
+so that performance could be degraded and might be even unacceptable for certain workloads. The latter means that although current approach works well for small to
+medium workloads, we still have to do some additional work to avoid performance degradation after restart of large instances.
+
+## Non Goals
+
+- Details of the persistence storage for prewarm data are out of scope, there is a separate RFC for that: <https://github.com/neondatabase/neon/pull/9661>.
+- Complete compute/Postgres HA setup and flow. Although it was originally in scope of this RFC, during preliminary research it appeared to be a rabbit hole, so it's worth of a separate RFC.
+- Low-level implementation details for Postgres replica-to-primary promotion. There are a lot of things to think and care about: how to start walproposer, [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html), and so on, but it's worth of at least a separate one-pager design document if not RFC.
+
+## Impacted components
+
+Postgres, compute_ctl, Control plane, Endpoint storage for unlogged storage of compute files.
+For the latter, we will need to implement a uniform abstraction layer on top of S3, ABS, etc., but
+S3 is used in text interchangeably with 'endpoint storage' for simplicity.
+
+## Proposed implementation
+
+### compute_ctl spec changes and auto-prewarm
+
+We are going to extend the current compute spec with the following attributes
+
+```rust
+struct ComputeSpec {
+    /// [All existing attributes]
+    ...
+    /// Whether to do auto-prewarm at start or not.
+    /// Default to `false`.
+    pub lfc_auto_prewarm: bool
+    /// Interval in seconds between automatic dumps of
+    /// LFC state into S3. Default `None`, which means 'off'.
+    pub lfc_dump_interval_sec: Option<i32>
+}
+```
+
+When `lfc_dump_interval_sec` is set to `N`, `compute_ctl` will periodically dump the LFC state
+and store it in S3, so that it could be used either for auto-prewarm after restart or by replica
+during the rolling restart. For enabling periodic dumping, we should consider the following value
+`lfc_dump_interval_sec=300` (5 minutes), same as in the upstream's `pg_prewarm.autoprewarm_interval`.
+
+When `lfc_auto_prewarm` is set to `true`, `compute_ctl` will start prewarming the LFC upon restart
+iif some of the previous states is present in S3.
+
+### compute_ctl API
+
+1. `POST /store_lfc_state` -- dump LFC state using Postgres SQL interface and store result in S3.
+    This has to be a blocking call, i.e. it will return only after the state is stored in S3.
+    If there is any concurrent request in progress, we should return `429 Too Many Requests`,
+    and let the caller to retry.
+
+2. `GET /dump_lfc_state` -- dump LFC state using Postgres SQL interface and return it as is
+    in text format suitable for the future restore/prewarm. This API is not strictly needed at
+    the end state, but could be useful for a faster prototyping of a complete rolling restart flow
+    with prewarm, as it doesn't require persistent for LFC state storage.
+
+3. `POST /restore_lfc_state` -- restore/prewarm LFC state with request
+
+    ```yaml
+    RestoreLFCStateRequest:
+      oneOf:
+        - type: object
+          required:
+            - lfc_state
+          properties:
+            lfc_state:
+              type: string
+              description: Raw LFC content dumped with GET `/dump_lfc_state`
+        - type: object
+          required:
+            - lfc_cache_key
+          properties:
+            lfc_cache_key:
+              type: string
+              description: |
+                endpoint_id of the source endpoint on the same branch
+                to use as a 'donor' for LFC content. Compute will look up
+                LFC content dump in S3 using this key and do prewarm.
+    ```
+
+    where `lfc_state` and `lfc_cache_key` are mutually exclusive.
+
+    The actual prewarming will happen asynchronously, so the caller need to check the
+    prewarm status using the compute's standard `GET /status` API.
+
+4. `GET /status` -- extend existing API with following attributes
+
+    ```rust
+    struct ComputeStatusResponse {
+        // [All existing attributes]
+        ...
+        pub prewarm_state: PrewarmState
+    }
+
+    /// Compute prewarm state. Will be stored in the shared Compute state
+    /// in compute_ctl
+    struct PrewarmState {
+        pub status: PrewarmStatus
+        /// Total number of pages to prewarm
+        pub pages_total: i64
+        /// Number of pages prewarmed so far
+        pub pages_processed: i64
+        /// Optional prewarm error
+        pub error: Option<String>
+    }
+
+    pub enum PrewarmStatus {
+        /// Prewarming was never requested on this compute
+        Off,
+        /// Prewarming was requested, but not started yet
+        Pending,
+        /// Prewarming is in progress. The caller should follow
+        /// `PrewarmState::progress`.
+        InProgress,
+        /// Prewarming has been successfully completed
+        Completed,
+        /// Prewarming failed. The caller should look at
+        /// `PrewarmState::error` for the reason.
+        Failed,
+        /// It is intended to be used by auto-prewarm if none of
+        /// the previous LFC states is available in S3.
+        /// This is a distinct state from the `Failed` because
+        /// technically it's not a failure and could happen if
+        /// compute was restart before it dumped anything into S3,
+        /// or just after the initial rollout of the feature.
+        Skipped,
+    }
+    ```
+
+5. `POST /promote` -- this is a **blocking** API call to promote compute replica into primary.
+    This API should be very similar to the existing `POST /configure` API, i.e. accept the
+    spec (primary spec, because originally compute was started as replica). It's a distinct
+    API method because semantics and response codes are different:
+
+    - If promotion is done successfully, it will return `200 OK`.
+    - If compute is already primary, the call will be no-op and `compute_ctl`
+      will return `412 Precondition Failed`.
+    - If, for some reason, second request reaches compute that is in progress of promotion,
+      it will respond with `429 Too Many Requests`.
+    - If compute hit any permanent failure during promotion `500 Internal Server Error`
+      will be returned.
+
+### Control plane operations
+
+The complete flow will be present as a sequence diagram in the next section, but here
+we just want to list some important steps that have to be done by control plane during
+the rolling restart via warm replica, but without much of low-level implementation details.
+
+1. Register the 'intent' of the instance restart, but not yet interrupt any workload at
+    primary and also accept new connections. This may require some endpoint state machine
+    changes, e.g. introduction of the `pending_restart` state. Being in this state also
+    **mustn't prevent any other operations except restart**: suspend, live-reconfiguration
+    (e.g. due to notify-attach call from the storage controller), deletion.
+
+2. Start new replica compute on the same timeline and start prewarming it. This process
+    may take quite a while, so the same concurrency considerations as in 1. should be applied
+    here as well.
+
+3. When warm replica is ready, control plane should:
+
+    3.1. Terminate the primary compute. Starting from here, **this is a critical section**,
+        if anything goes off, the only option is to start the primary normally and proceed
+        with auto-prewarm.
+
+    3.2. Send cache invalidation message to all proxies, notifying them that all new connections
+        should request and wait for the new connection details. At this stage, proxy has to also
+        drop any existing connections to the old primary, so they didn't do stale reads.
+
+    3.3. Attach warm replica compute to the primary endpoint inside control plane metadata
+        database.
+
+    3.4. Promote replica to primary.
+
+    3.5. When everything is done, finalize the endpoint state to be just `active`.
+
+### Complete rolling restart flow
+
+```mermaid
+  sequenceDiagram
+
+  autonumber
+
+  participant proxy as Neon proxy
+
+  participant cplane as Control plane
+
+  participant primary as Compute (primary)
+  box Compute (replica)
+    participant ctl as compute_ctl
+    participant pg as Postgres
+  end
+
+  box Endpoint unlogged storage
+    participant s3proxy as Endpoint storage service
+    participant s3 as S3/ABS/etc.
+  end
+
+
+  cplane ->> primary: POST /store_lfc_state
+  primary -->> cplane: 200 OK
+
+  cplane ->> ctl: POST /restore_lfc_state
+  activate ctl
+  ctl -->> cplane: 202 Accepted
+
+  activate cplane
+  cplane ->> ctl: GET /status: poll prewarm status
+  ctl ->> s3proxy: GET /read_file
+  s3proxy ->> s3: read file
+  s3 -->> s3proxy: file content
+  s3proxy -->> ctl: 200 OK: file content
+
+  proxy ->> cplane: GET /proxy_wake_compute
+  cplane -->> proxy: 200 OK: old primary conninfo
+
+  ctl ->> pg: prewarm LFC
+  activate pg
+  pg -->> ctl: prewarm is completed
+  deactivate pg
+
+  ctl -->> cplane: 200 OK: prewarm is completed
+  deactivate ctl
+  deactivate cplane
+
+  cplane -->> cplane: reassign replica compute to endpoint,<br>start terminating the old primary compute
+  activate cplane
+  cplane ->> proxy: invalidate caches
+
+  proxy ->> cplane: GET /proxy_wake_compute
+
+  cplane -x primary: POST /terminate
+  primary -->> cplane: 200 OK
+  note over primary: old primary<br>compute terminated
+
+  cplane ->> ctl: POST /promote
+  activate ctl
+  ctl ->> pg: pg_ctl promote
+  activate pg
+  pg -->> ctl: done
+  deactivate pg
+  ctl -->> cplane: 200 OK
+  deactivate ctl
+
+  cplane -->> cplane: finalize operation
+  cplane -->> proxy: 200 OK: new primary conninfo
+  deactivate cplane
+```
+
+### Network bandwidth and prewarm speed
+
+It's currently known that pageserver can sustain about 3000 RPS per shard for a few running computes.
+Large tenants are usually split into 8 shards, so the final formula may look like this:
+
+```text
+8 shards * 3000 RPS * 8 KB =~ 190 MB/s
+```
+
+so depending on the LFC size, prewarming will take at least:
+
+- ~5s for 1 GB
+- ~50s for 10 GB
+- ~5m for 100 GB
+- \>1h for 1 TB
+
+In total, one pageserver is normally capped by 30k RPS, so it obviously can't sustain many computes
+doing prewarm at the same time. Later, we may need an additional mechanism for computes to throttle
+the prewarming requests gracefully.
+
+### Reliability, failure modes and corner cases
+
+We consider following failures while implementing this RFC:
+
+1. Compute got interrupted/crashed/restarted during prewarm. The caller -- control plane -- should
+    detect that and start prewarm from the beginning.
+
+2. Control plane promotion request timed out or hit network issues. If it never reached the
+    compute, control plane should just repeat it. If it did reach the compute, then during
+    retry control plane can hit `409` as previous request triggered the promotion already.
+    In this case, control plane need to retry until either `200` or
+    permanent error `500` is returned.
+
+3. Compute got interrupted/crashed/restarted during promotion. At restart it will ask for
+    a spec from control plane, and its content should signal compute to start as **primary**,
+    so it's expected that control plane will continue polling for certain period of time and
+    will discover that compute is ready to accept connections if restart is fast enough.
+
+4. Any other unexpected failure or timeout during prewarming. This **failure mustn't be fatal**,
+    control plane has to report failure, terminate replica and keep primary running.
+
+5. Any other unexpected failure or timeout during promotion. Unfortunately, at this moment
+    we already have the primary node stopped, so the only option is to start primary again
+    and proceed with auto-prewarm.
+
+6. Any unexpected failure during auto-prewarm. This **failure mustn't be fatal**,
+    `compute_ctl` has to report the failure, but do not crash the compute.
+
+7. Control plane failed to confirm that old primary has terminated. This can happen, especially
+    in the future HA setup. In this case, control plane has to ensure that it sent VM deletion
+    and pod termination requests to k8s, so long-term we do not have two running primaries
+    on the same timeline.
+
+### Security implications
+
+There are two security implications to consider:
+
+1. Access to `compute_ctl` API. It has to be accessible from the outside of compute, so all
+    new API methods have to be exposed on the **external** HTTP port and **must** be authenticated
+    with JWT.
+
+2. Read/write only your own LFC state data in S3. Although it's not really a security concern,
+    since LFC state is just a mapping of blocks present in LFC at certain moment in time;
+    it still has to be highly restricted, so that i) only computes on the same timeline can
+    read S3 state; ii) each compute can only write to the path that contains it's `endpoint_id`.
+    Both of this must be validated by Endpoint storage service using the JWT token provided by `compute_ctl`.
+
+### Unresolved questions
+
+#### Billing, metrics and monitoring
+
+Currently, we only label computes with `endpoint_id` after attaching them to the endpoint.
+In this proposal, this means that temporary replica will remain unlabelled until it's promoted
+to primary. We can also hide it from users in the control plane API, but what to do with
+billing and monitoring is still unclear.
+
+We can probably mark it as 'billable' and tag with `project_id`, so it will be billed, but
+not interfere in any way with the current primary monitoring.
+
+Another thing to consider is how logs and metrics export will switch to the new compute.
+It's expected that OpenTelemetry collector will auto-discover the new compute and start
+scraping metrics from it.
+
+#### Auto-prewarm
+
+It's still an open question whether we need auto-prewarm at all. The author's gut-feeling is
+that yes, we need it, but might be not for all workloads, so it could end up exposed as a
+user-controllable knob on the endpoint. There are two arguments for that:
+
+1. Auto-prewarm existing in upstream's `pg_prewarm`, _probably for a reason_.
+
+2. There are still could be 2 flows when we cannot perform the rolling restart via the warm
+    replica: i) any failure or interruption during promotion; ii) wake up after scale-to-zero.
+    The latter might be challenged as well, i.e. one can argue that auto-prewarm may and will
+    compete with user-workload for storage resources. This is correct, but it might as well
+    reduce the time to get warm LFC and good performance.
+
+#### Low-level details of the replica promotion
+
+There are many things to consider here, but three items just off the top of my head:
+
+1. How to properly start the `walproposer` inside Postgres.
+
+2. What to do with logical replication. Currently, we do not include logical replication slots
+    inside basebackup, because nobody advances them at replica, so they just prevent the WAL
+    deletion. Yet, we do need to have them at primary after promotion. Starting with Postgres 17,
+    there is a new feature called
+    [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html)
+    and `synchronized_standby_slots` setting, but we need a plan for the older versions. Should we
+    request a new basebackup during promotion?
+
+3. How do we guarantee that replica will receive all the latest WAL from safekeepers? Do some
+    'shallow' version of sync safekeepers without data copying? Or just a standard version of
+    sync safekeepers?
+
+## Alternative implementation
+
+The proposal already assumes one of the alternatives -- do not have any persistent storage for
+LFC state. This is possible to implement faster with the proposed API, but it means that
+we do not implement auto-prewarm yet.
+
+## Definition of Done
+
+At the end of implementing this RFC we should have two high-level settings that enable:
+
+1. Auto-prewarm of user computes upon restart.
+2. Perform primary compute restart via the warm replica promotion.
+
+It also has to be decided what's the criteria for enabling one or both of these flows for
+certain clients.

endpoint_storage/src/app.rs

@@ -13,6 +13,8 @@ use utils::backoff::retry;
 pub fn app(state: Arc<Storage>) -> Router<()> {
     use axum::routing::{delete as _delete, get as _get};
     let delete_prefix = _delete(delete_prefix);
+    // NB: On any changes do not forget to update the OpenAPI spec
+    // in /endpoint_storage/src/openapi_spec.yml.
     Router::new()
         .route(
             "/{tenant_id}/{timeline_id}/{endpoint_id}/{*path}",
@@ -374,7 +376,7 @@ MC4CAQAwBQYDK2VwBCIEID/Drmc1AA6U/znNRWpF3zEGegOATQxfkdWxitcOMsIH
             let request = Request::builder()
                 .uri(format!("/{tenant}/{timeline}/{endpoint}/sub/path/key"))
                 .method(method)
-                .header("Authorization", format!("Bearer {}", token))
+                .header("Authorization", format!("Bearer {token}"))
                 .body(Body::empty())
                 .unwrap();
             let status = ServiceExt::ready(&mut app)

endpoint_storage/src/openapi_spec.yml

@@ -0,0 +1,146 @@
+openapi: "3.0.2"
+info:
+  title: Endpoint Storage API
+  description: Endpoint Storage API
+  version: "1.0"
+  license:
+    name: "Apache"
+    url: https://github.com/neondatabase/neon/blob/main/LICENSE
+servers:
+  - url: ""
+paths:
+  /status:
+    description: Healthcheck endpoint
+    get:
+      description: Healthcheck
+      security: []
+      responses:
+        "200":
+          description: OK
+
+  /{tenant_id}/{timeline_id}/{endpoint_id}/{key}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: endpoint_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: key
+        in: path
+        required: true
+        schema:
+          type: string
+    get:
+      description: Get file from blob storage
+      responses:
+        "200":
+          description: "File stream from blob storage"
+          content:
+            application/octet-stream:
+              schema:
+                type: string
+                format: binary
+        "400":
+          description: File was not found
+        "403":
+          description: JWT does not authorize request to this route
+    put:
+      description: Insert file into blob storage. If file exists, override it
+      requestBody:
+        content:
+          application/octet-stream:
+            schema:
+              type: string
+              format: binary
+      responses:
+        "200":
+          description: File was inserted successfully
+        "403":
+          description: JWT does not authorize request to this route
+    delete:
+      description: Delete file from blob storage
+      responses:
+        "200":
+          description: File was successfully deleted or not found
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}/{timeline_id}/{endpoint_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: endpoint_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete endpoint data from blob storage
+      responses:
+        "200":
+          description: Endpoint data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}/{timeline_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete timeline data from blob storage
+      responses:
+        "200":
+          description: Timeline data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete tenant data from blob storage
+      responses:
+        "200":
+          description: Tenant data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+components:
+  securitySchemes:
+    JWT:
+      type: http
+      scheme: bearer
+      bearerFormat: JWT
+
+security:
+  - JWT: []

libs/compute_api/Cargo.toml

@@ -12,6 +12,7 @@ jsonwebtoken.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 regex.workspace = true
+url.workspace = true
 
 utils = { path = "../utils" }
 remote_storage = { version = "0.1", path = "../remote_storage/" }

libs/compute_api/src/responses.rs

@@ -46,7 +46,7 @@ pub struct ExtensionInstallResponse {
     pub version: ExtVersion,
 }
 
-#[derive(Serialize, Default, Debug, Clone)]
+#[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcPrewarmState {
     #[default]
@@ -58,7 +58,18 @@ pub enum LfcPrewarmState {
     },
 }
 
-#[derive(Serialize, Default, Debug, Clone)]
+impl Display for LfcPrewarmState {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            LfcPrewarmState::NotPrewarmed => f.write_str("NotPrewarmed"),
+            LfcPrewarmState::Prewarming => f.write_str("Prewarming"),
+            LfcPrewarmState::Completed => f.write_str("Completed"),
+            LfcPrewarmState::Failed { error } => write!(f, "Error({error})"),
+        }
+    }
+}
+
+#[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcOffloadState {
     #[default]
@@ -70,6 +81,23 @@ pub enum LfcOffloadState {
     },
 }
 
+#[derive(Serialize, Debug, Clone, PartialEq)]
+#[serde(tag = "status", rename_all = "snake_case")]
+/// Response of /promote
+pub enum PromoteState {
+    NotPromoted,
+    Completed,
+    Failed { error: String },
+}
+
+#[derive(Deserialize, Serialize, Default, Debug, Clone)]
+#[serde(rename_all = "snake_case")]
+/// Result of /safekeepers_lsn
+pub struct SafekeepersLsn {
+    pub safekeepers: String,
+    pub wal_flush_lsn: utils::lsn::Lsn,
+}
+
 /// Response of the /status API
 #[derive(Serialize, Debug, Deserialize)]
 #[serde(rename_all = "snake_case")]

libs/compute_api/src/spec.rs

@@ -6,10 +6,12 @@
 use std::collections::HashMap;
 use std::fmt::Display;
 
+use anyhow::anyhow;
 use indexmap::IndexMap;
 use regex::Regex;
 use remote_storage::RemotePath;
 use serde::{Deserialize, Serialize};
+use url::Url;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -179,9 +181,18 @@ pub struct ComputeSpec {
     /// JWT for authorizing requests to endpoint storage service
     pub endpoint_storage_token: Option<String>,
 
-    /// Download LFC state from endpoint_storage and pass it to Postgres on startup
     #[serde(default)]
+    /// Download LFC state from endpoint storage and pass it to Postgres on compute startup
     pub autoprewarm: bool,
+
+    #[serde(default)]
+    /// Upload LFC state to endpoint storage periodically. Default value (None) means "don't upload"
+    pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+
+    /// Suspend timeout in seconds.
+    ///
+    /// We use this value to derive other values, such as the installed extensions metric.
+    pub suspend_timeout_seconds: i64,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -320,12 +331,6 @@ impl ComputeMode {
     }
 }
 
-impl Display for ComputeMode {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.to_type_str())
-    }
-}
-
 /// Log level for audit logging
 #[derive(Clone, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeAudit {
@@ -436,6 +441,47 @@ pub struct JwksSettings {
     pub jwt_audience: Option<String>,
 }
 
+/// Protocol used to connect to a Pageserver. Parsed from the connstring scheme.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub enum PageserverProtocol {
+    /// The original protocol based on libpq and COPY. Uses postgresql:// or postgres:// scheme.
+    #[default]
+    Libpq,
+    /// A newer, gRPC-based protocol. Uses grpc:// scheme.
+    Grpc,
+}
+
+impl PageserverProtocol {
+    /// Parses the protocol from a connstring scheme. Defaults to Libpq if no scheme is given.
+    /// Errors if the connstring is an invalid URL.
+    pub fn from_connstring(connstring: &str) -> anyhow::Result<Self> {
+        let scheme = match Url::parse(connstring) {
+            Ok(url) => url.scheme().to_lowercase(),
+            Err(url::ParseError::RelativeUrlWithoutBase) => return Ok(Self::default()),
+            Err(err) => return Err(anyhow!("invalid connstring URL: {err}")),
+        };
+        match scheme.as_str() {
+            "postgresql" | "postgres" => Ok(Self::Libpq),
+            "grpc" => Ok(Self::Grpc),
+            scheme => Err(anyhow!("invalid protocol scheme: {scheme}")),
+        }
+    }
+
+    /// Returns the URL scheme for the protocol, for use in connstrings.
+    pub fn scheme(&self) -> &'static str {
+        match self {
+            Self::Libpq => "postgresql",
+            Self::Grpc => "grpc",
+        }
+    }
+}
+
+impl Display for PageserverProtocol {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.scheme())
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::fs::File;

libs/compute_api/tests/cluster_spec.json

@@ -3,6 +3,7 @@
 
     "timestamp": "2021-05-23T18:25:43.511Z",
     "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8b",
+    "suspend_timeout_seconds": 3600,
 
     "cluster": {
         "cluster_id": "test-cluster-42",
@@ -89,6 +90,11 @@
                 "value": "off",
                 "vartype": "bool"
             },
+            {
+                "name": "offload_lfc_interval_seconds",
+                "value": "20",
+                "vartype": "integer"
+            },
             {
                 "name": "neon.safekeepers",
                 "value": "127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501",

libs/desim/src/executor.rs

@@ -71,7 +71,7 @@ impl Runtime {
                     debug!("thread panicked: {:?}", e);
                     let mut result = ctx.result.lock();
                     if result.0 == -1 {
-                        *result = (256, format!("thread panicked: {:?}", e));
+                        *result = (256, format!("thread panicked: {e:?}"));
                     }
                 });
             }

libs/desim/src/proto.rs

@@ -47,8 +47,8 @@ impl Debug for AnyMessage {
         match self {
             AnyMessage::None => write!(f, "None"),
             AnyMessage::InternalConnect => write!(f, "InternalConnect"),
-            AnyMessage::Just32(v) => write!(f, "Just32({})", v),
-            AnyMessage::ReplCell(v) => write!(f, "ReplCell({:?})", v),
+            AnyMessage::Just32(v) => write!(f, "Just32({v})"),
+            AnyMessage::ReplCell(v) => write!(f, "ReplCell({v:?})"),
             AnyMessage::Bytes(v) => write!(f, "Bytes({})", hex::encode(v)),
             AnyMessage::LSN(v) => write!(f, "LSN({})", Lsn(*v)),
         }

libs/http-utils/src/endpoint.rs

@@ -7,7 +7,7 @@ use anyhow::{Context, anyhow};
 use bytes::{Bytes, BytesMut};
 use hyper::header::{AUTHORIZATION, CONTENT_DISPOSITION, CONTENT_TYPE, HeaderName};
 use hyper::http::HeaderValue;
-use hyper::{Body, Method, Request, Response};
+use hyper::{Body, Request, Response};
 use jsonwebtoken::TokenData;
 use metrics::{Encoder, IntCounter, TextEncoder, register_int_counter};
 use once_cell::sync::Lazy;
@@ -18,8 +18,9 @@ use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
 use tokio::sync::{Mutex, Notify, mpsc};
 use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::io::ReaderStream;
-use tracing::{Instrument, debug, info, info_span, warn};
+use tracing::{Instrument, info, info_span, warn};
 use utils::auth::{AuthError, Claims, SwappableJwtAuth};
+use utils::metrics_collector::{METRICS_COLLECTOR, METRICS_STALE_MILLIS};
 
 use crate::error::{ApiError, api_error_handler, route_error_handler};
 use crate::request::{get_query_param, parse_query_param};
@@ -80,14 +81,10 @@ where
     let path = request.uri().path();
     let request_span = info_span!("request", %method, %path, %request_id);
 
-    let log_quietly = method == Method::GET;
     async move {
         let cancellation_guard = RequestCancelled::warn_when_dropped_without_responding();
-        if log_quietly {
-            debug!("Handling request");
-        } else {
-            info!("Handling request");
-        }
+
+        info!("Handling request");
 
         // No special handling for panics here. There's a `tracing_panic_hook` from another
         // module to do that globally.
@@ -108,11 +105,7 @@ where
         match res {
             Ok(response) => {
                 let response_status = response.status();
-                if log_quietly && response_status.is_success() {
-                    debug!("Request handled, status: {response_status}");
-                } else {
-                    info!("Request handled, status: {response_status}");
-                }
+                info!("Request handled, status: {response_status}");
                 Ok(response)
             }
             Err(err) => Ok(api_error_handler(err)),
@@ -250,9 +243,28 @@ impl std::io::Write for ChannelWriter {
     }
 }
 
-pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+pub async fn prometheus_metrics_handler(
+    req: Request<Body>,
+    force_metric_collection_on_scrape: bool,
+) -> Result<Response<Body>, ApiError> {
     SERVE_METRICS_COUNT.inc();
 
+    // HADRON
+    let requested_use_latest = parse_query_param(&req, "use_latest")?;
+
+    let use_latest = match requested_use_latest {
+        None => force_metric_collection_on_scrape,
+        Some(true) => true,
+        Some(false) => {
+            if force_metric_collection_on_scrape {
+                // We don't cache in this case
+                true
+            } else {
+                false
+            }
+        }
+    };
+
     let started_at = std::time::Instant::now();
 
     let (tx, rx) = mpsc::channel(1);
@@ -277,12 +289,18 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
 
         let _span = span.entered();
 
-        let metrics = metrics::gather();
+        // HADRON
+        let collected = if use_latest {
+            // Skip caching the results if we always force metric collection on scrape.
+            METRICS_COLLECTOR.run_once(!force_metric_collection_on_scrape)
+        } else {
+            METRICS_COLLECTOR.last_collected()
+        };
 
         let gathered_at = std::time::Instant::now();
 
         let res = encoder
-            .encode(&metrics, &mut writer)
+            .encode(&collected.metrics, &mut writer)
             .and_then(|_| writer.flush().map_err(|e| e.into()));
 
         // this instant is not when we finally got the full response sent, sending is done by hyper
@@ -295,6 +313,10 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
         let encoded_in = encoded_at - gathered_at - writer.wait_time();
         let total = encoded_at - started_at;
 
+        // HADRON
+        let staleness_ms = (encoded_at - collected.collected_at).as_millis();
+        METRICS_STALE_MILLIS.set(staleness_ms as i64);
+
         match res {
             Ok(()) => {
                 tracing::info!(
@@ -303,6 +325,7 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
                     spawning_ms = spawned_in.as_millis(),
                     collection_ms = collected_in.as_millis(),
                     encoding_ms = encoded_in.as_millis(),
+                    stalenss_ms = staleness_ms,
                     "responded /metrics"
                 );
             }
@@ -582,14 +605,14 @@ pub fn attach_openapi_ui(
                             deepLinking: true,
                             showExtensions: true,
                             showCommonExtensions: true,
-                            url: "{}",
+                            url: "{spec_mount_path}",
                         }})
                         window.ui = ui;
                     }};
                 </script>
                 </body>
                 </html>
-            "#, spec_mount_path))).unwrap())
+            "#))).unwrap())
              })
         )
 }
@@ -696,7 +719,7 @@ mod tests {
         let remote_addr = SocketAddr::new(IpAddr::from_str("127.0.0.1").unwrap(), 80);
         let mut service = builder.build(remote_addr);
         if let Err(e) = poll_fn(|ctx| service.poll_ready(ctx)).await {
-            panic!("request service is not ready: {:?}", e);
+            panic!("request service is not ready: {e:?}");
         }
 
         let mut req: Request<Body> = Request::default();
@@ -716,7 +739,7 @@ mod tests {
         let remote_addr = SocketAddr::new(IpAddr::from_str("127.0.0.1").unwrap(), 80);
         let mut service = builder.build(remote_addr);
         if let Err(e) = poll_fn(|ctx| service.poll_ready(ctx)).await {
-            panic!("request service is not ready: {:?}", e);
+            panic!("request service is not ready: {e:?}");
         }
 
         let req: Request<Body> = Request::default();

libs/http-utils/src/request.rs

@@ -41,17 +41,35 @@ pub fn get_query_param<'a>(
         Some(q) => q,
         None => return Ok(None),
     };
-    let mut values = url::form_urlencoded::parse(query.as_bytes())
+    let values = url::form_urlencoded::parse(query.as_bytes())
         .filter_map(|(k, v)| if k == param_name { Some(v) } else { None })
         // we call .next() twice below. If it's None the first time, .fuse() ensures it's None afterwards
         .fuse();
 
-    let value1 = values.next();
-    if values.next().is_some() {
-        return Err(ApiError::BadRequest(anyhow!(
-            "param {param_name} specified more than once"
-        )));
-    }
+    // Work around an issue with Alloy's pyroscope scrape where the "seconds"
+    // parameter is added several times. https://github.com/grafana/alloy/issues/3026
+    // TODO: revert after Alloy is fixed.
+    let value1 = values
+        .map(Ok)
+        .reduce(|acc, i| {
+            match acc {
+                Err(_) => acc,
+
+                // It's okay to have duplicates as along as they have the same value.
+                Ok(ref a) if a == &i.unwrap() => acc,
+
+                _ => Err(ApiError::BadRequest(anyhow!(
+                    "param {param_name} specified more than once"
+                ))),
+            }
+        })
+        .transpose()?;
+    // if values.next().is_some() {
+    //     return Err(ApiError::BadRequest(anyhow!(
+    //         "param {param_name} specified more than once"
+    //     )));
+    // }
+
     Ok(value1)
 }
 
@@ -92,3 +110,39 @@ pub async fn ensure_no_body(request: &mut Request<Body>) -> Result<(), ApiError>
         None => Ok(()),
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_get_query_param_duplicate() {
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert_eq!(value.unwrap(), "1");
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1&testparam=1")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert_eq!(value.unwrap(), "1");
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert!(value.is_none());
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1&testparam=2&testparam=3")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam");
+        assert!(value.is_err());
+    }
+}

libs/neon-shmem/Cargo.toml

@@ -6,27 +6,8 @@ license.workspace = true
 
 [dependencies]
 thiserror.workspace = true
-nix.workspace = true
+nix.workspace=true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
-rustc-hash = { version = "2.1.1" }
-rand = "0.9.1"
-libc.workspace = true
-lock_api = "0.4.13"
-
-[dev-dependencies]
-criterion = { workspace = true, features = ["html_reports"] }
-rand_distr = "0.5.1"
-xxhash-rust = { version = "0.8.15", features = ["xxh3"] }
-ahash.workspace = true
-twox-hash = { version = "2.1.1" }
-seahash = "4.1.0"
-hashbrown = { git = "https://github.com/quantumish/hashbrown.git", rev = "6610e6d" }
-foldhash = "0.1.5"
-
 
 [target.'cfg(target_os = "macos")'.dependencies]
 tempfile = "3.14.0"
-
-[[bench]]
-name = "hmap_resize"
-harness = false

libs/neon-shmem/benches/hmap_resize.rs

@@ -1,282 +0,0 @@
-use criterion::{criterion_group, criterion_main, BatchSize, Criterion, BenchmarkId};
-use neon_shmem::hash::HashMapAccess;
-use neon_shmem::hash::HashMapInit;
-use neon_shmem::hash::entry::Entry;
-use rand::prelude::*;
-use rand::distr::{Distribution, StandardUniform};
-use std::hash::BuildHasher;
-use std::default::Default;
-	
-// Taken from bindings to C code
-
-#[derive(Clone, Debug, Hash, Eq, PartialEq)]
-#[repr(C)]
-pub struct FileCacheKey {
-    pub _spc_id: u32,
-    pub _db_id: u32,
-    pub _rel_number: u32,
-    pub _fork_num: u32,
-    pub _block_num: u32,
-}
-
-impl Distribution<FileCacheKey> for StandardUniform {
-	// questionable, but doesn't need to be good randomness
-	fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> FileCacheKey {
-		FileCacheKey {
-			_spc_id: rng.random(),
-			_db_id: rng.random(),
-			_rel_number: rng.random(),
-			_fork_num: rng.random(),
-			_block_num: rng.random()
-		}
-    }
-}
-
-#[derive(Clone, Debug)]
-#[repr(C)]
-pub struct FileCacheEntry {
-    pub _offset: u32,
-    pub _access_count: u32,
-    pub _prev: *mut FileCacheEntry,
-    pub _next: *mut FileCacheEntry,
-    pub _state: [u32; 8],
-}
-
-impl FileCacheEntry {
-	fn dummy() -> Self {
-		Self {
-			_offset: 0,
-			_access_count: 0,
-			_prev: std::ptr::null_mut(),
-			_next: std::ptr::null_mut(),
-			_state: [0; 8]				
-		}
-	}
-}
-
-// Utilities for applying operations.
-
-#[derive(Clone, Debug)]
-struct TestOp<K,V>(K, Option<V>);
-
-fn apply_op<K: Clone + std::hash::Hash + Eq, V, S: std::hash::BuildHasher>(
-    op: TestOp<K,V>,
-    map: &mut HashMapAccess<K,V,S>,
-) {
-	let entry = map.entry(op.0);
-
-    match op.1 {
-		Some(new) => {
-			match entry {
-				Entry::Occupied(mut e) => Some(e.insert(new)),
-				Entry::Vacant(e) => { _ = e.insert(new).unwrap(); None },
-			}
-		},
-		None => {
-			match entry {
-				Entry::Occupied(e) => Some(e.remove()),
-				Entry::Vacant(_) => None,
-			}
-		},
-	};
-}
-
-// Hash utilities
-
-struct SeaRandomState {
-	k1: u64,
-	k2: u64,
-	k3: u64,
-	k4: u64 
-}
-
-impl std::hash::BuildHasher for SeaRandomState {
-	type Hasher = seahash::SeaHasher;
-	
-	fn build_hasher(&self) -> Self::Hasher {
-		seahash::SeaHasher::with_seeds(self.k1, self.k2, self.k3, self.k4)
-	}
-}
-
-impl SeaRandomState {
-	fn new() -> Self {
-		let mut rng = rand::rng();
-		Self { k1: rng.random(), k2: rng.random(), k3: rng.random(), k4: rng.random() }
-	}
-}
-
-fn small_benchs(c: &mut Criterion) {
-	let mut group = c.benchmark_group("Small maps");
-    group.sample_size(10);
-        
-	group.bench_function("small_rehash", |b| {
-		let ideal_filled = 4_000_000;
-		let size = 5_000_000;
-		let mut writer = HashMapInit::new_resizeable(size, size * 2).attach_writer();
-		let mut rng = rand::rng();		
-		while writer.get_num_buckets_in_use() < ideal_filled as usize {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			apply_op(TestOp(key, Some(val)), &mut writer);
-		}
-		b.iter(|| writer.shuffle());
-	});
-	
-
-	group.bench_function("small_rehash_xxhash", |b| {
-		let ideal_filled = 4_000_000;
-		let size = 5_000_000;
-		let mut writer = HashMapInit::new_resizeable(size, size * 2)
-			.with_hasher(twox_hash::xxhash64::RandomState::default())
-			.attach_writer();
-		let mut rng = rand::rng();		
-		while writer.get_num_buckets_in_use() < ideal_filled as usize {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			apply_op(TestOp(key, Some(val)), &mut writer);
-		}
-		b.iter(|| writer.shuffle());
-	});
-
-	
-	group.bench_function("small_rehash_ahash", |b| {
-		let ideal_filled = 4_000_000;
-		let size = 5_000_000;
-		let mut writer = HashMapInit::new_resizeable(size, size * 2)
-			.with_hasher(ahash::RandomState::default())
-			.attach_writer();
-		let mut rng = rand::rng();		
-		while writer.get_num_buckets_in_use() < ideal_filled as usize {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			apply_op(TestOp(key, Some(val)), &mut writer);
-		}
-		b.iter(|| writer.shuffle());
-	});
-
-	group.bench_function("small_rehash_seahash", |b| {
-		let ideal_filled = 4_000_000;
-		let size = 5_000_000;
-		let mut writer = HashMapInit::new_resizeable(size, size * 2)
-			.with_hasher(SeaRandomState::new())
-			.attach_writer();
-		let mut rng = rand::rng();
-		while writer.get_num_buckets_in_use() < ideal_filled as usize {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			apply_op(TestOp(key, Some(val)), &mut writer);
-		}
-		b.iter(|| writer.shuffle());
-	});
-
-	group.finish();	
-}
-
-fn real_benchs(c: &mut Criterion) {
-	let mut group = c.benchmark_group("Realistic workloads");
-	group.sample_size(10);	
-    group.bench_function("real_bulk_insert", |b| {
-		let size = 125_000_000;
-		let ideal_filled = 100_000_000;		
-		let mut rng = rand::rng();		
-		b.iter_batched(
-			|| HashMapInit::new_resizeable(size, size * 2).attach_writer(),
-			|writer| {
-				for _ in 0..ideal_filled {
-					let key: FileCacheKey = rng.random();
-					let val = FileCacheEntry::dummy();
-					let entry = writer.entry(key);
-					std::hint::black_box(match entry {
-						Entry::Occupied(mut e) => { e.insert(val); },
-						Entry::Vacant(e) => { _ = e.insert(val).unwrap(); },
-					})
-				}	
-			},
-			BatchSize::SmallInput,
-		)
-	});
-
-	group.bench_function("real_rehash", |b| {
-		let size = 125_000_000;
-		let ideal_filled = 100_000_000;		
-		let mut writer = HashMapInit::new_resizeable(size, size).attach_writer();
-		let mut rng = rand::rng();		
-		while writer.get_num_buckets_in_use() < ideal_filled {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			apply_op(TestOp(key, Some(val)), &mut writer);
-		}
-		b.iter(|| writer.shuffle());
-	});
-	
-	group.bench_function("real_rehash_hashbrown", |b| {
-		let size = 125_000_000;
-		let ideal_filled = 100_000_000;
-		let mut writer = hashbrown::raw::RawTable::new();
-		let mut rng = rand::rng();
-		let hasher = rustc_hash::FxBuildHasher::default();
-		unsafe {
-			writer.resize(size, |(k,_)| hasher.hash_one(&k),
-						  hashbrown::raw::Fallibility::Infallible).unwrap();
-		}
-		while writer.len() < ideal_filled as usize {
-			let key: FileCacheKey = rng.random();
-			let val = FileCacheEntry::dummy();
-			writer.insert(hasher.hash_one(&key), (key, val), |(k,_)| hasher.hash_one(&k));
-		}
-		b.iter(|| unsafe { writer.table.rehash_in_place(
-			&|table, index| hasher.hash_one(&table.bucket::<(FileCacheKey, FileCacheEntry)>(index).as_ref().0),
-			std::mem::size_of::<(FileCacheKey, FileCacheEntry)>(),
-            if std::mem::needs_drop::<(FileCacheKey, FileCacheEntry)>() {
-                Some(|ptr| std::ptr::drop_in_place(ptr as *mut (FileCacheKey, FileCacheEntry)))
-            } else {
-                None
-            },
-		) });
-	});
-
-	for elems in [2, 4, 8, 16, 32, 64, 96, 112] {
-		group.bench_with_input(BenchmarkId::new("real_rehash_varied", elems), &elems, |b, &size| {
-			let ideal_filled = size * 1_000_000;
-			let size = 125_000_000;
-			let mut writer = HashMapInit::new_resizeable(size, size).attach_writer();
-			let mut rng = rand::rng();		
-			while writer.get_num_buckets_in_use() < ideal_filled as usize {
-				let key: FileCacheKey = rng.random();
-				let val = FileCacheEntry::dummy();
-				apply_op(TestOp(key, Some(val)), &mut writer);
-			}
-			b.iter(|| writer.shuffle());
-		});
-		group.bench_with_input(BenchmarkId::new("real_rehash_varied_hashbrown", elems), &elems, |b, &size| {
-			let ideal_filled = size * 1_000_000;
-			let size = 125_000_000;
-			let mut writer = hashbrown::raw::RawTable::new();
-			let mut rng = rand::rng();
-			let hasher = rustc_hash::FxBuildHasher::default();
-			unsafe {
-				writer.resize(size, |(k,_)| hasher.hash_one(&k),
-							  hashbrown::raw::Fallibility::Infallible).unwrap();
-			}
-			while writer.len() < ideal_filled as usize {
-				let key: FileCacheKey = rng.random();
-				let val = FileCacheEntry::dummy();
-				writer.insert(hasher.hash_one(&key), (key, val), |(k,_)| hasher.hash_one(&k));
-			}
-			b.iter(|| unsafe { writer.table.rehash_in_place(
-				&|table, index| hasher.hash_one(&table.bucket::<(FileCacheKey, FileCacheEntry)>(index).as_ref().0),
-				std::mem::size_of::<(FileCacheKey, FileCacheEntry)>(),
-				if std::mem::needs_drop::<(FileCacheKey, FileCacheEntry)>() {
-					Some(|ptr| std::ptr::drop_in_place(ptr as *mut (FileCacheKey, FileCacheEntry)))
-				} else {
-					None
-				},
-			) });
-		});
-	}
-	
-	group.finish();
-}
-	
-criterion_group!(benches, small_benchs, real_benchs);
-criterion_main!(benches);

libs/neon-shmem/src/hash.rs

@@ -1,533 +0,0 @@
-//! Resizable hash table implementation on top of byte-level storage (either a [`ShmemHandle`] or a fixed byte array).
-//!
-//! This hash table has two major components: the bucket array and the dictionary. Each bucket within the
-//! bucket array contains a `Option<(K, V)>` and an index of another bucket. In this way there is both an 
-//! implicit freelist within the bucket array (`None` buckets point to other `None` entries) and various hash
-//! chains within the bucket array (a Some bucket will point to other Some buckets that had the same hash).
-//!
-//! Buckets are never moved unless they are within a region that is being shrunk, and so the actual hash-
-//! dependent component is done with the dictionary. When a new key is inserted into the map, a position
-//! within the dictionary is decided based on its hash, the data is inserted into an empty bucket based
-//! off of the freelist, and then the index of said bucket is placed in the dictionary.
-//!
-//! This map is resizable (if initialized on top of a [`ShmemHandle`]). Both growing and shrinking happen
-//! in-place and are at a high level achieved by expanding/reducing the bucket array and rebuilding the
-//! dictionary by rehashing all keys.
-
-use std::hash::{Hash, BuildHasher};
-use std::mem::MaybeUninit;
-
-use crate::{shmem, sync::*};
-use crate::shmem::ShmemHandle;
-
-mod core;
-pub mod entry;
-
-#[cfg(test)]
-mod tests;
-
-use core::{Bucket, CoreHashMap, INVALID_POS};
-use entry::{Entry, OccupiedEntry, VacantEntry, PrevPos};
-
-/// Builder for a [`HashMapAccess`].
-#[must_use]
-pub struct HashMapInit<'a, K, V, S = rustc_hash::FxBuildHasher> {
-    shmem_handle: Option<ShmemHandle>,
-    shared_ptr: *mut RwLock<HashMapShared<'a, K, V>>,
-	shared_size: usize,
-	hasher: S,
-	num_buckets: u32,
-}
-
-/// Accessor for a hash table. 
-pub struct HashMapAccess<'a, K, V, S = rustc_hash::FxBuildHasher> {
-    shmem_handle: Option<ShmemHandle>,
-    shared_ptr: *mut HashMapShared<'a, K, V>,
-	hasher: S,
-}
-
-unsafe impl<K: Sync, V: Sync, S> Sync for HashMapAccess<'_, K, V, S> {}
-unsafe impl<K: Send, V: Send, S> Send for HashMapAccess<'_, K, V, S> {}
-
-impl<'a, K: Clone + Hash + Eq, V, S> HashMapInit<'a, K, V, S> {
-	pub fn with_hasher<T: BuildHasher>(self, hasher: T) -> HashMapInit<'a, K, V, T> {
-		HashMapInit {
-			hasher,
-			shmem_handle: self.shmem_handle,
-			shared_ptr: self.shared_ptr,
-			shared_size: self.shared_size,
-			num_buckets: self.num_buckets,
-		}
-	}
-
-	/// Loosely (over)estimate the size needed to store a hash table with `num_buckets` buckets.
-	pub fn estimate_size(num_buckets: u32) -> usize {
-        // add some margin to cover alignment etc.
-        CoreHashMap::<K, V>::estimate_size(num_buckets) + size_of::<HashMapShared<K, V>>() + 1000
-    }
-
-	/// Initialize a table for writing.
-    pub fn attach_writer(self) -> HashMapAccess<'a, K, V, S> {
-        let mut ptr: *mut u8 = self.shared_ptr.cast();
-        let end_ptr: *mut u8 = unsafe { ptr.add(self.shared_size) };
-
-		// carve out area for the One Big Lock (TM) and the HashMapShared.
-		ptr = unsafe { ptr.add(ptr.align_offset(align_of::<libc::pthread_rwlock_t>())) };
-		let raw_lock_ptr = ptr;
-		ptr = unsafe { ptr.add(size_of::<libc::pthread_rwlock_t>()) };
-		ptr = unsafe { ptr.add(ptr.align_offset(align_of::<HashMapShared<K, V>>())) };
-		let shared_ptr: *mut HashMapShared<K, V> = ptr.cast();
-        ptr = unsafe { ptr.add(size_of::<HashMapShared<K, V>>()) };
-						
-        // carve out the buckets
-        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<core::Bucket<K, V>>())) };
-        let buckets_ptr = ptr;
-        ptr = unsafe { ptr.add(size_of::<core::Bucket<K, V>>() * self.num_buckets as usize) };
-
-        // use remaining space for the dictionary
-        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<u32>())) };
-        assert!(ptr.addr() < end_ptr.addr());
-        let dictionary_ptr = ptr;
-        let dictionary_size = unsafe { end_ptr.byte_offset_from(ptr) / size_of::<u32>() as isize };
-        assert!(dictionary_size > 0);
-
-        let buckets =
-            unsafe { std::slice::from_raw_parts_mut(buckets_ptr.cast(), self.num_buckets as usize) };
-        let dictionary = unsafe {
-            std::slice::from_raw_parts_mut(dictionary_ptr.cast(), dictionary_size as usize)
-        };
-        let hashmap = CoreHashMap::new(buckets, dictionary);
-		let lock = RwLock::from_raw(PthreadRwLock::new(raw_lock_ptr.cast()), hashmap);
-		unsafe {
-			std::ptr::write(shared_ptr, lock);
-		}
-		
-        HashMapAccess {
-            shmem_handle: self.shmem_handle,
-            shared_ptr,
-			hasher: self.hasher,
-        }
-    }
-
-	/// Initialize a table for reading. Currently identical to [`HashMapInit::attach_writer`].
-    pub fn attach_reader(self) -> HashMapAccess<'a, K, V, S> {
-        self.attach_writer()
-    }
-}
-
-/// Hash table data that is actually stored in the shared memory area.
-///
-/// NOTE: We carve out the parts from a contiguous chunk. Growing and shrinking the hash table
-/// relies on the memory layout! The data structures are laid out in the contiguous shared memory
-/// area as follows:
-///
-/// [`libc::pthread_rwlock_t`]
-/// [`HashMapShared`]
-/// [buckets]
-/// [dictionary]
-///
-/// In between the above parts, there can be padding bytes to align the parts correctly.
-type HashMapShared<'a, K, V> = RwLock<CoreHashMap<'a, K, V>>;
-
-impl<'a, K, V> HashMapInit<'a, K, V, rustc_hash::FxBuildHasher>
-where
-	K: Clone + Hash + Eq
-{
-	/// Place the hash table within a user-supplied fixed memory area.
-	pub fn with_fixed(
-		num_buckets: u32,
-        area: &'a mut [MaybeUninit<u8>],
-    ) -> Self {
-		Self {
-			num_buckets,
-			shmem_handle: None,
-			shared_ptr: area.as_mut_ptr().cast(),
-			shared_size: area.len(),
-			hasher: rustc_hash::FxBuildHasher,
-		}		
-    }
-
-    /// Place a new hash map in the given shared memory area
-	///
-	/// # Panics
-	/// Will panic on failure to resize area to expected map size.
-    pub fn with_shmem(num_buckets: u32, shmem: ShmemHandle) -> Self {
-		let size = Self::estimate_size(num_buckets);
-		shmem
-            .set_size(size)
-            .expect("could not resize shared memory area");
-		Self {
-			num_buckets,
-			shared_ptr: shmem.data_ptr.as_ptr().cast(),
-			shmem_handle: Some(shmem),
-			shared_size: size,
-			hasher: rustc_hash::FxBuildHasher
-		}
-    }
-
-	/// Make a resizable hash map within a new shared memory area with the given name.
-	pub fn new_resizeable_named(num_buckets: u32, max_buckets: u32, name: &str) -> Self {
-		let size = Self::estimate_size(num_buckets);
-		let max_size = Self::estimate_size(max_buckets);
-		let shmem = ShmemHandle::new(name, size, max_size)
-			.expect("failed to make shared memory area");
-		
-		Self {
-			num_buckets,
-			shared_ptr: shmem.data_ptr.as_ptr().cast(),
-			shmem_handle: Some(shmem),
-			shared_size: size,
-			hasher: rustc_hash::FxBuildHasher
-		}
-	}
-
-	/// Make a resizable hash map within a new anonymous shared memory area.
-	pub fn new_resizeable(num_buckets: u32, max_buckets: u32) -> Self {
-		use std::sync::atomic::{AtomicUsize, Ordering};
-		static COUNTER: AtomicUsize = AtomicUsize::new(0);
-		let val = COUNTER.fetch_add(1, Ordering::Relaxed);
-		let name = format!("neon_shmem_hmap{val}");
-		Self::new_resizeable_named(num_buckets, max_buckets, &name)
-	}
-}
-
-impl<'a, K, V, S: BuildHasher> HashMapAccess<'a, K, V, S>
-where
-    K: Clone + Hash + Eq,
-{
-	/// Hash a key using the map's hasher.
-	#[inline]
-    fn get_hash_value(&self, key: &K) -> u64 {
-		self.hasher.hash_one(key)        
-    }
-
-	fn entry_with_hash(&self, key: K, hash: u64) -> Entry<'a, '_, K, V> {
-		let mut map = unsafe { self.shared_ptr.as_ref() }.unwrap().write();
-        let dict_pos = hash as usize % map.dictionary.len();
-        let first = map.dictionary[dict_pos];
-        if first == INVALID_POS {
-            // no existing entry
-            return Entry::Vacant(VacantEntry {
-                map,
-                key,
-                dict_pos: dict_pos as u32,
-            });
-        }
-
-        let mut prev_pos = PrevPos::First(dict_pos as u32);
-        let mut next = first;
-        loop {
-            let bucket = &mut map.buckets[next as usize];
-            let (bucket_key, _bucket_value) = bucket.inner.as_mut().expect("entry is in use");
-            if *bucket_key == key {
-                // found existing entry
-                return Entry::Occupied(OccupiedEntry {
-                    map,
-                    _key: key,
-                    prev_pos,
-                    bucket_pos: next,
-                });
-            }
-
-            if bucket.next == INVALID_POS {
-                // No existing entry
-                return Entry::Vacant(VacantEntry {
-                    map,
-                    key,
-                    dict_pos: dict_pos as u32,
-                });
-            }
-            prev_pos = PrevPos::Chained(next);
-            next = bucket.next;
-        }
-	}
-	
-	/// Get a reference to the corresponding value for a key.
-    pub fn get<'e>(&'e self, key: &K) -> Option<ValueReadGuard<'e, V>> {
-		let hash = self.get_hash_value(key);
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
-		RwLockReadGuard::try_map(map, |m| m.get_with_hash(key, hash)).ok()
-    }
-
-	/// Get a reference to the entry containing a key.
-    pub fn entry(&self, key: K) -> Entry<'a, '_, K, V> {
-		let hash = self.get_hash_value(&key);
-		self.entry_with_hash(key, hash)
-    }
-
-	/// Remove a key given its hash. Returns the associated value if it existed.
-    pub fn remove(&self, key: &K) -> Option<V> {
-		let hash = self.get_hash_value(&key);
-        match self.entry_with_hash(key.clone(), hash) {
-            Entry::Occupied(e) => Some(e.remove()),
-            Entry::Vacant(_) => None
-        }
-    }
-
-	/// Insert/update a key. Returns the previous associated value if it existed.
-	///
-	/// # Errors
-	/// Will return [`core::FullError`] if there is no more space left in the map.
-    pub fn insert(&self, key: K, value: V) -> Result<Option<V>, core::FullError> {
-		let hash = self.get_hash_value(&key);
-        match self.entry_with_hash(key.clone(), hash) {
-            Entry::Occupied(mut e) => Ok(Some(e.insert(value))),
-            Entry::Vacant(e) => {
-				_ = e.insert(value)?;
-				Ok(None)
-			}
-        }
-    }
-	
-	/// Optionally return the entry for a bucket at a given index if it exists.
-	///
-	/// Has more overhead than one would intuitively expect: performs both a clone of the key
-	/// due to the [`OccupiedEntry`] type owning the key and also a hash of the key in order
-	/// to enable repairing the hash chain if the entry is removed.
-    pub fn entry_at_bucket(&self, pos: usize) -> Option<OccupiedEntry<'a, '_, K, V>> {
-        let map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-		if pos >= map.buckets.len() {
-			return None;
-		}
-
-		let entry = map.buckets[pos].inner.as_ref();
-		match entry {
-			Some((key, _)) => Some(OccupiedEntry {
-				_key: key.clone(),
-				bucket_pos: pos as u32,
-				prev_pos: entry::PrevPos::Unknown(
-					self.get_hash_value(&key)
-				),
-				map,
-			}),
-			_ => None,
-		}
-    }
-
-	/// Returns the number of buckets in the table.
-    pub fn get_num_buckets(&self) -> usize {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
-        map.get_num_buckets()
-    }
-
-    /// Return the key and value stored in bucket with given index. This can be used to
-    /// iterate through the hash map.
-	// TODO: An Iterator might be nicer. The communicator's clock algorithm needs to
-	// _slowly_ iterate through all buckets with its clock hand,  without holding a lock.
-	// If we switch to an Iterator, it must not hold the lock.
-    pub fn get_at_bucket(&self, pos: usize) -> Option<ValueReadGuard<(K, V)>> {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
-        if pos >= map.buckets.len() {
-            return None;
-        }
-		RwLockReadGuard::try_map(map, |m| m.buckets[pos].inner.as_ref()).ok()
-    }
-
-	/// Returns the index of the bucket a given value corresponds to.
-    pub fn get_bucket_for_value(&self, val_ptr: *const V) -> usize {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
-
-        let origin = map.buckets.as_ptr();
-        let idx = (val_ptr as usize - origin as usize) / size_of::<Bucket<K, V>>();
-        assert!(idx < map.buckets.len());
-
-        idx
-    }
-
-    /// Returns the number of occupied buckets in the table.
-    pub fn get_num_buckets_in_use(&self) -> usize {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
-        map.buckets_in_use as usize
-    }
-
-	/// Clears all entries in a table. Does not reset any shrinking operations.
-	pub fn clear(&self) {
-		let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-        map.clear();
-	}
-	
-	/// Perform an in-place rehash of some region (0..`rehash_buckets`) of the table and reset
-	/// the `buckets` and `dictionary` slices to be as long as `num_buckets`. Resets the freelist
-	/// in the process.
-	fn rehash_dict(
-		&self,
-		inner: &mut CoreHashMap<'a, K, V>,
-		buckets_ptr: *mut core::Bucket<K, V>,
-		end_ptr: *mut u8,
-		num_buckets: u32,
-		rehash_buckets: u32,
-	) {
-		inner.free_head = INVALID_POS;
-		
-        let buckets;
-        let dictionary;
-        unsafe {
-            let buckets_end_ptr = buckets_ptr.add(num_buckets as usize);
-            let dictionary_ptr: *mut u32 = buckets_end_ptr
-                .byte_add(buckets_end_ptr.align_offset(align_of::<u32>()))
-                .cast();
-            let dictionary_size: usize =
-                end_ptr.byte_offset_from(buckets_end_ptr) as usize / size_of::<u32>();
-
-            buckets = std::slice::from_raw_parts_mut(buckets_ptr, num_buckets as usize);
-            dictionary = std::slice::from_raw_parts_mut(dictionary_ptr, dictionary_size);
-        }		
-        for e in dictionary.iter_mut() {
-            *e = INVALID_POS;
-        }
-		
-        for (i, bucket) in buckets.iter_mut().enumerate().take(rehash_buckets as usize) {
-            if bucket.inner.is_none() {
-				bucket.next = inner.free_head;
-                inner.free_head = i as u32;
-				continue;
-            }
-
-			let hash = self.hasher.hash_one(&bucket.inner.as_ref().unwrap().0);
-            let pos: usize = (hash % dictionary.len() as u64) as usize;
-            bucket.next = dictionary[pos];
-            dictionary[pos] = i as u32;
-        }
-
-        inner.dictionary = dictionary;
-        inner.buckets = buckets;
-	}
-
-	/// Rehash the map without growing or shrinking. 
-	pub fn shuffle(&self) {
-		let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-		let num_buckets = map.get_num_buckets() as u32;
-		let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
-		let end_ptr: *mut u8 = unsafe { self.shared_ptr.byte_add(size_bytes).cast() };
-        let buckets_ptr = map.buckets.as_mut_ptr();
-		self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
-	}
-
-    /// Grow the number of buckets within the table. 
-    ///
-    /// 1. Grows the underlying shared memory area
-    /// 2. Initializes new buckets and overwrites the current dictionary
-    /// 3. Rehashes the dictionary
-	///
-	/// # Panics 
-	/// Panics if called on a map initialized with [`HashMapInit::with_fixed`].
-	///
-	/// # Errors
-	/// Returns an [`shmem::Error`] if any errors occur resizing the memory region.
-    pub fn grow(&self, num_buckets: u32) -> Result<(), shmem::Error> {
-        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-        let old_num_buckets = map.buckets.len() as u32;
-
-        assert!(num_buckets >= old_num_buckets, "grow called with a smaller number of buckets");
-        if num_buckets == old_num_buckets {
-            return Ok(());
-        }
-        let shmem_handle = self
-            .shmem_handle
-            .as_ref()
-            .expect("grow called on a fixed-size hash table");
-
-        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
-        shmem_handle.set_size(size_bytes)?;
-        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
-
-        // Initialize new buckets. The new buckets are linked to the free list.
-		// NB: This overwrites the dictionary!
-        let buckets_ptr = map.buckets.as_mut_ptr();
-        unsafe {
-            for i in old_num_buckets..num_buckets {
-                let bucket = buckets_ptr.add(i as usize);
-                bucket.write(core::Bucket {
-                    next: if i < num_buckets-1 {
-                        i + 1
-                    } else {
-                        map.free_head
-                    },
-                    inner: None,
-                });
-            }
-        }
-
-		self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, old_num_buckets);
-        map.free_head = old_num_buckets;
-
-        Ok(())
-    }
-
-	/// Begin a shrink, limiting all new allocations to be in buckets with index below `num_buckets`.
-	///
-	/// # Panics
-	/// Panics if called on a map initialized with [`HashMapInit::with_fixed`] or if `num_buckets` is
-	/// greater than the number of buckets in the map.
-	pub fn begin_shrink(&mut self, num_buckets: u32) {
-		let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-		assert!(
-			num_buckets <= map.get_num_buckets() as u32,
-            "shrink called with a larger number of buckets"
-        );
-		_ = self
-            .shmem_handle
-            .as_ref()
-            .expect("shrink called on a fixed-size hash table");
-		map.alloc_limit = num_buckets;
-	}
-
-	/// If a shrink operation is underway, returns the target size of the map. Otherwise, returns None.
-	pub fn shrink_goal(&self) -> Option<usize> {
-		let map = unsafe { self.shared_ptr.as_mut() }.unwrap().read();
-        let goal = map.alloc_limit;
-		if goal == INVALID_POS { None } else { Some(goal as usize) }
-	}
-	
-	/// Complete a shrink after caller has evicted entries, removing the unused buckets and rehashing.
-	///
-	/// # Panics
-	/// The following cases result in a panic: 
-	/// - Calling this function on a map initialized with [`HashMapInit::with_fixed`].
-	/// - Calling this function on a map when no shrink operation is in progress.
-	/// - Calling this function on a map with `shrink_mode` set to [`HashMapShrinkMode::Remap`] and
-	///   there are more buckets in use than the value returned by [`HashMapAccess::shrink_goal`].
-	///
-	/// # Errors
-	/// Returns an [`shmem::Error`] if any errors occur resizing the memory region.
-	pub fn finish_shrink(&self) -> Result<(), shmem::Error> {
-		let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
-		assert!(
-			map.alloc_limit != INVALID_POS,
-			"called finish_shrink when no shrink is in progress"
-		);
-
-		let num_buckets = map.alloc_limit; 
-
-		if map.get_num_buckets() == num_buckets as usize {
-            return Ok(());
-        }
-
-		assert!(
-			map.buckets_in_use <= num_buckets,
-			"called finish_shrink before enough entries were removed"
-		);
-		
-		for i in (num_buckets as usize)..map.buckets.len() {
-			if let Some((k, v)) = map.buckets[i].inner.take() {
-				// alloc_bucket increases count, so need to decrease since we're just moving
-				map.buckets_in_use -= 1;
-				map.alloc_bucket(k, v).unwrap();
-			}
-		}
-
-        let shmem_handle = self
-            .shmem_handle
-            .as_ref()
-            .expect("shrink called on a fixed-size hash table");
-
-		let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
-        shmem_handle.set_size(size_bytes)?;
-        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
-		let buckets_ptr = map.buckets.as_mut_ptr();
-		self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
-		map.alloc_limit = INVALID_POS;
-		
-		Ok(())
-	}
-}

libs/neon-shmem/src/hash/core.rs

@@ -1,177 +0,0 @@
-//! Simple hash table with chaining.
-
-use std::hash::Hash;
-use std::mem::MaybeUninit;
-
-use crate::hash::entry::*;
-
-/// Invalid position within the map (either within the dictionary or bucket array).
-pub(crate) const INVALID_POS: u32 = u32::MAX;
-
-/// Fundamental storage unit within the hash table. Either empty or contains a key-value pair.
-/// Always part of a chain of some kind (either a freelist if empty or a hash chain if full).
-pub(crate) struct Bucket<K, V> {
-	/// Index of next bucket in the chain.
-	pub(crate) next: u32,
-	/// Key-value pair contained within bucket.
-    pub(crate) inner: Option<(K, V)>,
-}
-
-/// Core hash table implementation.
-pub(crate) struct CoreHashMap<'a, K, V> {
-	/// Dictionary used to map hashes to bucket indices.
-    pub(crate) dictionary: &'a mut [u32],
-	/// Buckets containing key-value pairs.
-    pub(crate) buckets: &'a mut [Bucket<K, V>],
-	/// Head of the freelist.
-    pub(crate) free_head: u32,
-	/// Maximum index of a bucket allowed to be allocated. [`INVALID_POS`] if no limit.
-	pub(crate) alloc_limit: u32,
-    /// The number of currently occupied buckets.
-    pub(crate) buckets_in_use: u32,
-	// pub(crate) lock: libc::pthread_mutex_t,
-	// Unclear what the purpose of this is.
-    pub(crate) _user_list_head: u32,
-}
-
-/// Error for when there are no empty buckets left but one is needed.
-#[derive(Debug, PartialEq)]
-pub struct FullError();
-
-impl<'a, K: Clone + Hash + Eq, V> CoreHashMap<'a, K, V> {
-    const FILL_FACTOR: f32 = 0.60;
-
-	/// Estimate the size of data contained within the the hash map.
-    pub fn estimate_size(num_buckets: u32) -> usize {
-        let mut size = 0;
-
-        // buckets
-        size += size_of::<Bucket<K, V>>() * num_buckets as usize;
-
-        // dictionary
-        size += (f32::ceil((size_of::<u32>() * num_buckets as usize) as f32 / Self::FILL_FACTOR))
-            as usize;
-
-        size
-    }	
-
-    pub fn new(
-        buckets: &'a mut [MaybeUninit<Bucket<K, V>>],
-        dictionary: &'a mut [MaybeUninit<u32>],
-    ) -> Self {
-        // Initialize the buckets
-        for i in 0..buckets.len() {
-            buckets[i].write(Bucket {
-                next: if i < buckets.len() - 1 {
-                    i as u32 + 1
-                } else {
-                    INVALID_POS
-                },				
-                inner: None,
-            });
-        }
-
-        // Initialize the dictionary
-        for e in dictionary.iter_mut() {
-            e.write(INVALID_POS);
-        }
-
-        // TODO: use std::slice::assume_init_mut() once it stabilizes
-        let buckets =
-            unsafe { std::slice::from_raw_parts_mut(buckets.as_mut_ptr().cast(), buckets.len()) };
-        let dictionary = unsafe {
-            std::slice::from_raw_parts_mut(dictionary.as_mut_ptr().cast(), dictionary.len())
-        };
-
-        Self {
-            dictionary,
-            buckets,
-            free_head: 0,
-            buckets_in_use: 0,
-            _user_list_head: INVALID_POS,
-			alloc_limit: INVALID_POS,
-        }
-    }
-
-	/// Get the value associated with a key (if it exists) given its hash.
-    pub fn get_with_hash(&self, key: &K, hash: u64) -> Option<&V> {
-        let mut next = self.dictionary[hash as usize % self.dictionary.len()];
-        loop {
-            if next == INVALID_POS {
-                return None;
-            }
-
-            let bucket = &self.buckets[next as usize];
-            let (bucket_key, bucket_value) = bucket.inner.as_ref().expect("entry is in use");
-            if bucket_key == key {
-                return Some(bucket_value);
-            }
-            next = bucket.next;
-        }
-    }
-
-	/// Get number of buckets in map.
-    pub fn get_num_buckets(&self) -> usize {
-        self.buckets.len()
-    }
-
-	/// Clears all entries from the hashmap.
-	///
-	/// Does not reset any allocation limits, but does clear any entries beyond them.
-	pub fn clear(&mut self) {
-		for i in 0..self.buckets.len() {
-            self.buckets[i] = Bucket {
-                next: if i < self.buckets.len() - 1 {
-                    i as u32 + 1
-                } else {
-                    INVALID_POS
-                },				
-                inner: None,
-            }
-        }
-        for i in 0..self.dictionary.len() {
-            self.dictionary[i] = INVALID_POS;
-        }
-
-		self.free_head = 0;
-		self.buckets_in_use = 0;
-	}
-
-	/// Find the position of an unused bucket via the freelist and initialize it. 
-    pub(crate) fn alloc_bucket(&mut self, key: K, value: V) -> Result<u32, FullError> {
-        let mut pos = self.free_head;
-
-		// Find the first bucket we're *allowed* to use.
-		let mut prev = PrevPos::First(self.free_head);
-		while pos != INVALID_POS && pos >= self.alloc_limit {
-			let bucket = &mut self.buckets[pos as usize];
-			prev = PrevPos::Chained(pos);
-			pos = bucket.next;
-		}
-		if pos == INVALID_POS {
-			return Err(FullError());
-		}
-
-		// Repair the freelist.
-		match prev {
-			PrevPos::First(_) => {
-				let next_pos = self.buckets[pos as usize].next;
-				self.free_head = next_pos;				
-			}
-			PrevPos::Chained(p) => if p != INVALID_POS {
-				let next_pos = self.buckets[pos as usize].next;
-				self.buckets[p as usize].next = next_pos;
-			},
-			_ => unreachable!()
-		}
-
-		// Initialize the bucket.
-		let bucket = &mut self.buckets[pos as usize];
-		self.buckets_in_use += 1;
-        bucket.next = INVALID_POS;
-        bucket.inner = Some((key, value));
-
-        Ok(pos)
-    }
-}
-

libs/neon-shmem/src/hash/entry.rs

@@ -1,139 +0,0 @@
-//! Equivalent of [`std::collections::hash_map::Entry`] for this hashmap.
-
-use crate::hash::core::{CoreHashMap, FullError, INVALID_POS};
-use crate::sync::{RwLockWriteGuard, ValueWriteGuard};
-
-use std::hash::Hash;
-use std::mem;
-
-
-pub enum Entry<'a, 'b, K, V> {
-	Occupied(OccupiedEntry<'a, 'b, K, V>),
-    Vacant(VacantEntry<'a, 'b, K, V>),
-}
-
-/// Enum representing the previous position within a chain.
-#[derive(Clone, Copy)]
-pub(crate) enum PrevPos {
-	/// Starting index within the dictionary.  
-    First(u32),
-	/// Regular index within the buckets.
-    Chained(u32),
-	/// Unknown - e.g. the associated entry was retrieved by index instead of chain.
-	Unknown(u64),
-}
-
-pub struct OccupiedEntry<'a, 'b, K, V> {
-	/// Mutable reference to the map containing this entry.
-	pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
-	/// The key of the occupied entry
-    pub(crate) _key: K,
-	/// The index of the previous entry in the chain.
-    pub(crate) prev_pos: PrevPos,
-	/// The position of the bucket in the [`CoreHashMap`] bucket array.
-    pub(crate) bucket_pos: u32,
-}
-
-impl<K, V> OccupiedEntry<'_, '_, K, V> {
-    pub fn get(&self) -> &V {
-        &self.map.buckets[self.bucket_pos as usize]
-            .inner
-            .as_ref()
-            .unwrap()
-            .1
-    }
-
-    pub fn get_mut(&mut self) -> &mut V {
-        &mut self.map.buckets[self.bucket_pos as usize]
-            .inner
-            .as_mut()
-            .unwrap()
-            .1
-    }
-
-	/// Inserts a value into the entry, replacing (and returning) the existing value.
-    pub fn insert(&mut self, value: V) -> V {
-        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
-        // This assumes inner is Some, which it must be for an OccupiedEntry
-        mem::replace(&mut bucket.inner.as_mut().unwrap().1, value)
-    }
-
-	/// Removes the entry from the hash map, returning the value originally stored within it.
-	///
-	/// This may result in multiple bucket accesses if the entry was obtained by index as the
-	/// previous chain entry needs to be discovered in this case.
-	///
-	/// # Panics
-	/// Panics if the `prev_pos` field is equal to [`PrevPos::Unknown`]. In practice, this means
-	/// the entry was obtained via calling something like [`CoreHashMap::entry_at_bucket`].
-    pub fn remove(mut self) -> V {
-		// If this bucket was queried by index, go ahead and follow its chain from the start.
-		let prev = if let PrevPos::Unknown(hash) = self.prev_pos {
-			let dict_idx = hash as usize % self.map.dictionary.len();
-			let mut prev = PrevPos::First(dict_idx as u32);
-			let mut curr = self.map.dictionary[dict_idx];
-			while curr != self.bucket_pos {
-				curr = self.map.buckets[curr as usize].next;
-				prev = PrevPos::Chained(curr);
-			}
-			prev 	
-		} else {
-			self.prev_pos
-		};
-		
-        // CoreHashMap::remove returns Option<(K, V)>. We know it's Some for an OccupiedEntry.
-        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
-		
-        // unlink it from the chain
-        match prev {
-            PrevPos::First(dict_pos) => {
-				self.map.dictionary[dict_pos as usize] = bucket.next;
-			},
-            PrevPos::Chained(bucket_pos) => {
-				// println!("we think prev of {} is {bucket_pos}", self.bucket_pos);
-                self.map.buckets[bucket_pos as usize].next = bucket.next;
-            },
-			_ => unreachable!(),			
-        }
-
-        // and add it to the freelist
-		let free = self.map.free_head;
-        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
-        let old_value = bucket.inner.take();
-		bucket.next = free;
-        self.map.free_head = self.bucket_pos;
-        self.map.buckets_in_use -= 1;
-
-        old_value.unwrap().1
-    }
-}
-
-/// An abstract view into a vacant entry within the map.
-pub struct VacantEntry<'a, 'b, K, V> {
-	/// Mutable reference to the map containing this entry.
-	pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
-	/// The key to be inserted into this entry.
-    pub(crate) key: K,
-	/// The position within the dictionary corresponding to the key's hash.
-    pub(crate) dict_pos: u32,
-}
-
-impl<'b, K: Clone + Hash + Eq, V> VacantEntry<'_, 'b, K, V> {
-	/// Insert a value into the vacant entry, finding and populating an empty bucket in the process.
-	///
-	/// # Errors
-	/// Will return [`FullError`] if there are no unoccupied buckets in the map.
-    pub fn insert(mut self, value: V) -> Result<ValueWriteGuard<'b, V>, FullError> {
-        let pos = self.map.alloc_bucket(self.key, value)?;
-        if pos == INVALID_POS {
-            return Err(FullError());
-        }
-        self.map.buckets[pos as usize].next = self.map.dictionary[self.dict_pos as usize];
-        self.map.dictionary[self.dict_pos as usize] = pos;
-
-		Ok(RwLockWriteGuard::map(
-			self.map,
-			|m| &mut m.buckets[pos as usize].inner.as_mut().unwrap().1
-		))
-    }
-}

libs/neon-shmem/src/hash/tests.rs

@@ -1,426 +0,0 @@
-use std::collections::BTreeMap;
-use std::collections::HashSet;
-use std::fmt::Debug;
-use std::mem::MaybeUninit;
-
-use crate::hash::HashMapAccess;
-use crate::hash::HashMapInit;
-use crate::hash::Entry;
-use crate::hash::core::FullError;
-
-use rand::seq::SliceRandom;
-use rand::{Rng, RngCore};
-use rand_distr::Zipf;
-
-const TEST_KEY_LEN: usize = 16;
-
-#[derive(Clone, Copy, Debug, Hash, PartialEq, Eq, PartialOrd, Ord)]
-struct TestKey([u8; TEST_KEY_LEN]);
-
-impl From<&TestKey> for u128 {
-    fn from(val: &TestKey) -> u128 {
-        u128::from_be_bytes(val.0)
-    }
-}
-
-impl From<u128> for TestKey {
-    fn from(val: u128) -> TestKey {
-        TestKey(val.to_be_bytes())
-    }
-}
-
-impl<'a> From<&'a [u8]> for TestKey {
-    fn from(bytes: &'a [u8]) -> TestKey {
-        TestKey(bytes.try_into().unwrap())
-    }
-}
-
-fn test_inserts<K: Into<TestKey> + Copy>(keys: &[K]) {	
-    let w = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		100000, 120000, "test_inserts"
-	).attach_writer();
-
-    for (idx, k) in keys.iter().enumerate() {
-		let res = w.entry((*k).into());
-		match res {
-			Entry::Occupied(mut e) => { e.insert(idx); }
-			Entry::Vacant(e) => {
-				let res = e.insert(idx);
-				assert!(res.is_ok());
-			},
-		};
-    }
-
-    for (idx, k) in keys.iter().enumerate() {
-        let x = w.get(&(*k).into());
-        let value = x.as_deref().copied();
-        assert_eq!(value, Some(idx));
-    }
-}
-
-#[test]
-fn dense() {
-    // This exercises splitting a node with prefix
-    let keys: &[u128] = &[0, 1, 2, 3, 256];
-    test_inserts(keys);
-
-    // Dense keys
-    let mut keys: Vec<u128> = (0..10000).collect();
-    test_inserts(&keys);
-
-    // Do the same in random orders
-    for _ in 1..10 {
-        keys.shuffle(&mut rand::rng());
-        test_inserts(&keys);
-    }
-}
-
-#[test]
-fn sparse() {
-    // sparse keys
-    let mut keys: Vec<TestKey> = Vec::new();
-    let mut used_keys = HashSet::new();
-    for _ in 0..10000 {
-        loop {
-            let key = rand::random::<u128>();
-            if used_keys.get(&key).is_some() {
-                continue;
-            }
-            used_keys.insert(key);
-            keys.push(key.into());
-            break;
-        }
-    }
-    test_inserts(&keys);
-}
-
-#[derive(Clone, Debug)]
-struct TestOp(TestKey, Option<usize>);
-
-fn apply_op(
-    op: &TestOp,
-    map: &mut HashMapAccess<TestKey, usize>,
-    shadow: &mut BTreeMap<TestKey, usize>,
-) {
-    // apply the change to the shadow tree first
-    let shadow_existing = if let Some(v) = op.1 {
-        shadow.insert(op.0, v)
-    } else {
-        shadow.remove(&op.0)
-    };
-
-	let entry = map.entry(op.0);
-    let hash_existing = match op.1 {
-		Some(new) => {
-			match entry {
-				Entry::Occupied(mut e) => Some(e.insert(new)),
-				Entry::Vacant(e) => { _ = e.insert(new).unwrap(); None },
-			}
-		},
-		None => {
-			match entry {
-				Entry::Occupied(e) => Some(e.remove()),
-				Entry::Vacant(_) => None,
-			}
-		},
-	};
-
-	assert_eq!(shadow_existing, hash_existing);
-}
-
-fn do_random_ops(
-	num_ops: usize,
-	size: u32,
-	del_prob: f64,
-	writer: &mut HashMapAccess<TestKey, usize>,
-	shadow: &mut BTreeMap<TestKey, usize>,
-	rng: &mut rand::rngs::ThreadRng,
-) {
-	for i in 0..num_ops {
-        let key: TestKey = ((rng.next_u32() % size) as u128).into();
-        let op = TestOp(key, if rng.random_bool(del_prob) { Some(i) } else { None });
-        apply_op(&op, writer, shadow);
-    }
-}
-
-fn do_deletes(
-	num_ops: usize,
-	writer: &mut HashMapAccess<TestKey, usize>,
-	shadow: &mut BTreeMap<TestKey, usize>,
-) {
-	for _ in 0..num_ops {
-		let (k, _) = shadow.pop_first().unwrap();
-		writer.remove(&k);
-	}
-}
-
-fn do_shrink(
-	writer: &mut HashMapAccess<TestKey, usize>,
-	shadow: &mut BTreeMap<TestKey, usize>,
-	to: u32
-) {
-	assert!(writer.shrink_goal().is_none());
-	writer.begin_shrink(to);
-	assert_eq!(writer.shrink_goal(), Some(to as usize));
-	while writer.get_num_buckets_in_use() > to as usize {
-		let (k, _) = shadow.pop_first().unwrap();
-		let entry = writer.entry(k);
-		if let Entry::Occupied(e) = entry {
-			e.remove();
-		}
-	}
-	let old_usage = writer.get_num_buckets_in_use();
-	writer.finish_shrink().unwrap();
-	assert!(writer.shrink_goal().is_none());
-	assert_eq!(writer.get_num_buckets_in_use(), old_usage);
-}
-
-#[test]
-fn random_ops() {
-	let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		100000, 120000, "test_random"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-	
-    let distribution = Zipf::new(u128::MAX as f64, 1.1).unwrap();
-    let mut rng = rand::rng();
-    for i in 0..100000 {
-        let key: TestKey = (rng.sample(distribution) as u128).into();
-
-        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
-
-        apply_op(&op, &mut writer, &mut shadow);
-    }
-}
-
-
-#[test]
-fn test_shuffle() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1000, 1200, "test_shuf"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-
-    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
-    writer.shuffle();
-	do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
-}
-
-#[test]
-fn test_grow() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1000, 2000, "test_grow"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-
-    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
-	let old_usage = writer.get_num_buckets_in_use();
-    writer.grow(1500).unwrap();
-	assert_eq!(writer.get_num_buckets_in_use(), old_usage);
-	assert_eq!(writer.get_num_buckets(), 1500);
-	do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
-}
-
-#[test]
-fn test_clear() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_clear"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-    do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
-	writer.clear();
-	assert_eq!(writer.get_num_buckets_in_use(), 0);
-	assert_eq!(writer.get_num_buckets(), 1500);
-	while let Some((key, _)) = shadow.pop_first() {
-		assert!(writer.get(&key).is_none());
-	}
-	do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
-	for i in 0..(1500 - writer.get_num_buckets_in_use()) {
-		writer.insert((1500 + i as u128).into(), 0).unwrap();
-	}
-	assert_eq!(writer.insert(5000.into(), 0), Err(FullError {}));
-	writer.clear();
-	assert!(writer.insert(5000.into(), 0).is_ok());
-}
-
-#[test]
-fn test_idx_remove() {
-	let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_clear"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
-	for _ in 0..100 {
-		let idx = (rng.next_u32() % 1500) as usize;
-		if let Some(e) = writer.entry_at_bucket(idx) {
-			shadow.remove(&e._key);
-			e.remove();
-		}
-		
-	}
-	while let Some((key, val)) = shadow.pop_first() {
-		assert_eq!(*writer.get(&key).unwrap(), val);
-	}
-}
-
-#[test]
-fn test_idx_get() {
-	let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_clear"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
-	for _ in 0..100 {
-		let idx = (rng.next_u32() % 1500) as usize;
-		if let Some(pair) = writer.get_at_bucket(idx) {
-			{ 
-				let v: *const usize = &pair.1;
-				assert_eq!(writer.get_bucket_for_value(v), idx);
-			}
-			{
-				let v: *const usize = &pair.1;
-				assert_eq!(writer.get_bucket_for_value(v), idx);
-			}
-		}
-	}
-}
-
-#[test]
-fn test_shrink() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_shrink"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-	
-    do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);	
-	do_shrink(&mut writer, &mut shadow, 1000);	
-	assert_eq!(writer.get_num_buckets(), 1000);
-	do_deletes(500, &mut writer, &mut shadow);
-	do_random_ops(10000, 500, 0.75, &mut writer, &mut shadow, &mut rng);
-	assert!(writer.get_num_buckets_in_use() <= 1000);
-}
-
-#[test]
-fn test_shrink_grow_seq() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1000, 20000, "test_grow_seq"
-	).attach_writer();
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-    let mut rng = rand::rng();
-
-    do_random_ops(500, 1000, 0.1, &mut writer, &mut shadow, &mut rng);
-	eprintln!("Shrinking to 750");
-    do_shrink(&mut writer, &mut shadow, 750);
-	do_random_ops(200, 1000, 0.5, &mut writer, &mut shadow, &mut rng);
-	eprintln!("Growing to 1500");
-	writer.grow(1500).unwrap();
-	do_random_ops(600, 1500, 0.1, &mut writer, &mut shadow, &mut rng);
-	eprintln!("Shrinking to 200");
-	while shadow.len() > 100 {
-		do_deletes(1, &mut writer, &mut shadow);
-	}
-	do_shrink(&mut writer, &mut shadow, 200);
-	do_random_ops(50, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
-	eprintln!("Growing to 10k");
-	writer.grow(10000).unwrap();
-	do_random_ops(10000, 5000, 0.25, &mut writer, &mut shadow, &mut rng);
-}
-
-#[test]
-fn test_bucket_ops() {
-	let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1000, 1200, "test_bucket_ops"
-	).attach_writer();
-	match writer.entry(1.into()) {
-		Entry::Occupied(mut e) => { e.insert(2); },
-		Entry::Vacant(e) => { _ = e.insert(2).unwrap(); },
-	}
-	assert_eq!(writer.get_num_buckets_in_use(), 1);
-	assert_eq!(writer.get_num_buckets(), 1000);
-	assert_eq!(*writer.get(&1.into()).unwrap(), 2);
-	let pos = match writer.entry(1.into()) {
-		Entry::Occupied(e) => {
-			assert_eq!(e._key, 1.into());
-			let pos = e.bucket_pos as usize;
-			pos
-		},
-		Entry::Vacant(_) => { panic!("Insert didn't affect entry"); },
-	};
-	assert_eq!(writer.entry_at_bucket(pos).unwrap()._key, 1.into());
-	assert_eq!(*writer.get_at_bucket(pos).unwrap(), (1.into(), 2));
-	{
-		let ptr: *const usize = &*writer.get(&1.into()).unwrap();
-		assert_eq!(writer.get_bucket_for_value(ptr), pos);
-	}
-	writer.remove(&1.into());
-	assert!(writer.get(&1.into()).is_none());
-}
-
-#[test]
-fn test_shrink_zero() {
-	let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_shrink_zero"
-	).attach_writer();
-	writer.begin_shrink(0);
-	for i in 0..1500 {
-		writer.entry_at_bucket(i).map(|x| x.remove());
-	}
-	writer.finish_shrink().unwrap();
-	assert_eq!(writer.get_num_buckets_in_use(), 0);
-	let entry = writer.entry(1.into());
-	if let Entry::Vacant(v) = entry {
-		assert!(v.insert(2).is_err());
-	} else {
-		panic!("Somehow got non-vacant entry in empty map.")
-	}
-	writer.grow(50).unwrap();
-	let entry = writer.entry(1.into());
-	if let Entry::Vacant(v) = entry {
-		assert!(v.insert(2).is_ok());
-	} else {
-		panic!("Somehow got non-vacant entry in empty map.")
-	}
-	assert_eq!(writer.get_num_buckets_in_use(), 1);
-}
-
-#[test]
-#[should_panic]
-fn test_grow_oom() {
-    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2000, "test_grow_oom"
-	).attach_writer();
-	writer.grow(20000).unwrap();
-}
-
-#[test]
-#[should_panic]
-fn test_shrink_bigger() {
-    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2500, "test_shrink_bigger"
-	).attach_writer();
-	writer.begin_shrink(2000);
-}
-
-#[test]
-#[should_panic]
-fn test_shrink_early_finish() {
-    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(
-		1500, 2500, "test_shrink_early_finish"
-	).attach_writer();
-	writer.finish_shrink().unwrap();
-}
-
-#[test]
-#[should_panic]
-fn test_shrink_fixed_size() {
-	let mut area = [MaybeUninit::uninit(); 10000];
-    let init_struct = HashMapInit::<TestKey, usize>::with_fixed(3, &mut area);
-    let mut writer = init_struct.attach_writer();
-	writer.begin_shrink(1);
-}

libs/neon-shmem/src/lib.rs

@@ -1,5 +1,418 @@
 //! Shared memory utilities for neon communicator
 
-pub mod hash;
-pub mod shmem;
-pub mod sync;
+use std::num::NonZeroUsize;
+use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
+use std::ptr::NonNull;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use nix::errno::Errno;
+use nix::sys::mman::MapFlags;
+use nix::sys::mman::ProtFlags;
+use nix::sys::mman::mmap as nix_mmap;
+use nix::sys::mman::munmap as nix_munmap;
+use nix::unistd::ftruncate as nix_ftruncate;
+
+/// ShmemHandle represents a shared memory area that can be shared by processes over fork().
+/// Unlike shared memory allocated by Postgres, this area is resizable, up to 'max_size' that's
+/// specified at creation.
+///
+/// The area is backed by an anonymous file created with memfd_create(). The full address space for
+/// 'max_size' is reserved up-front with mmap(), but whenever you call [`ShmemHandle::set_size`],
+/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
+/// will cause the file to be expanded, but we might use mprotect() etc. to enforce that in the
+/// future.
+pub struct ShmemHandle {
+    /// memfd file descriptor
+    fd: OwnedFd,
+
+    max_size: usize,
+
+    // Pointer to the beginning of the shared memory area. The header is stored there.
+    shared_ptr: NonNull<SharedStruct>,
+
+    // Pointer to the beginning of the user data
+    pub data_ptr: NonNull<u8>,
+}
+
+/// This is stored at the beginning in the shared memory area.
+struct SharedStruct {
+    max_size: usize,
+
+    /// Current size of the backing file. The high-order bit is used for the RESIZE_IN_PROGRESS flag
+    current_size: AtomicUsize,
+}
+
+const RESIZE_IN_PROGRESS: usize = 1 << 63;
+
+const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
+
+/// Error type returned by the ShmemHandle functions.
+#[derive(thiserror::Error, Debug)]
+#[error("{msg}: {errno}")]
+pub struct Error {
+    pub msg: String,
+    pub errno: Errno,
+}
+
+impl Error {
+    fn new(msg: &str, errno: Errno) -> Error {
+        Error {
+            msg: msg.to_string(),
+            errno,
+        }
+    }
+}
+
+impl ShmemHandle {
+    /// Create a new shared memory area. To communicate between processes, the processes need to be
+    /// fork()'d after calling this, so that the ShmemHandle is inherited by all processes.
+    ///
+    /// If the ShmemHandle is dropped, the memory is unmapped from the current process. Other
+    /// processes can continue using it, however.
+    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<ShmemHandle, Error> {
+        // create the backing anonymous file.
+        let fd = create_backing_file(name)?;
+
+        Self::new_with_fd(fd, initial_size, max_size)
+    }
+
+    fn new_with_fd(
+        fd: OwnedFd,
+        initial_size: usize,
+        max_size: usize,
+    ) -> Result<ShmemHandle, Error> {
+        // We reserve the high-order bit for the RESIZE_IN_PROGRESS flag, and the actual size
+        // is a little larger than this because of the SharedStruct header. Make the upper limit
+        // somewhat smaller than that, because with anything close to that, you'll run out of
+        // memory anyway.
+        if max_size >= 1 << 48 {
+            panic!("max size {max_size} too large");
+        }
+        if initial_size > max_size {
+            panic!("initial size {initial_size} larger than max size {max_size}");
+        }
+
+        // The actual initial / max size is the one given by the caller, plus the size of
+        // 'SharedStruct'.
+        let initial_size = HEADER_SIZE + initial_size;
+        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
+
+        // Reserve address space for it with mmap
+        //
+        // TODO: Use MAP_HUGETLB if possible
+        let start_ptr = unsafe {
+            nix_mmap(
+                None,
+                max_size,
+                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
+                MapFlags::MAP_SHARED,
+                &fd,
+                0,
+            )
+        }
+        .map_err(|e| Error::new("mmap failed: {e}", e))?;
+
+        // Reserve space for the initial size
+        enlarge_file(fd.as_fd(), initial_size as u64)?;
+
+        // Initialize the header
+        let shared: NonNull<SharedStruct> = start_ptr.cast();
+        unsafe {
+            shared.write(SharedStruct {
+                max_size: max_size.into(),
+                current_size: AtomicUsize::new(initial_size),
+            })
+        };
+
+        // The user data begins after the header
+        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
+
+        Ok(ShmemHandle {
+            fd,
+            max_size: max_size.into(),
+            shared_ptr: shared,
+            data_ptr,
+        })
+    }
+
+    // return reference to the header
+    fn shared(&self) -> &SharedStruct {
+        unsafe { self.shared_ptr.as_ref() }
+    }
+
+    /// Resize the shared memory area. 'new_size' must not be larger than the 'max_size' specified
+    /// when creating the area.
+    ///
+    /// This may only be called from one process/thread concurrently. We detect that case
+    /// and return an Error.
+    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
+        let new_size = new_size + HEADER_SIZE;
+        let shared = self.shared();
+
+        if new_size > self.max_size {
+            panic!(
+                "new size ({} is greater than max size ({})",
+                new_size, self.max_size
+            );
+        }
+        assert_eq!(self.max_size, shared.max_size);
+
+        // Lock the area by setting the bit in 'current_size'
+        //
+        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
+        // and the posix_fallocate/ftruncate call is surely a synchronization point anyway. But
+        // since this is not performance-critical, better safe than sorry .
+        let mut old_size = shared.current_size.load(Ordering::Acquire);
+        loop {
+            if (old_size & RESIZE_IN_PROGRESS) != 0 {
+                return Err(Error::new(
+                    "concurrent resize detected",
+                    Errno::UnknownErrno,
+                ));
+            }
+            match shared.current_size.compare_exchange(
+                old_size,
+                new_size,
+                Ordering::Acquire,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(x) => old_size = x,
+            }
+        }
+
+        // Ok, we got the lock.
+        //
+        // NB: If anything goes wrong, we *must* clear the bit!
+        let result = {
+            use std::cmp::Ordering::{Equal, Greater, Less};
+            match new_size.cmp(&old_size) {
+                Less => nix_ftruncate(&self.fd, new_size as i64).map_err(|e| {
+                    Error::new("could not shrink shmem segment, ftruncate failed: {e}", e)
+                }),
+                Equal => Ok(()),
+                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
+            }
+        };
+
+        // Unlock
+        shared.current_size.store(
+            if result.is_ok() { new_size } else { old_size },
+            Ordering::Release,
+        );
+
+        result
+    }
+
+    /// Returns the current user-visible size of the shared memory segment.
+    ///
+    /// NOTE: a concurrent set_size() call can change the size at any time. It is the caller's
+    /// responsibility not to access the area beyond the current size.
+    pub fn current_size(&self) -> usize {
+        let total_current_size =
+            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
+        total_current_size - HEADER_SIZE
+    }
+}
+
+impl Drop for ShmemHandle {
+    fn drop(&mut self) {
+        // SAFETY: The pointer was obtained from mmap() with the given size.
+        // We unmap the entire region.
+        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
+        // The fd is dropped automatically by OwnedFd.
+    }
+}
+
+/// Create a "backing file" for the shared memory area. On Linux, use memfd_create(), to create an
+/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
+/// development and testing, but in production we want the file to stay in memory.
+///
+/// disable 'unused_variables' warnings, because in the macos path, 'name' is unused.
+#[allow(unused_variables)]
+fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
+            .map_err(|e| Error::new("memfd_create failed: {e}", e))
+    }
+    #[cfg(target_os = "macos")]
+    {
+        let file = tempfile::tempfile().map_err(|e| {
+            Error::new(
+                "could not create temporary file to back shmem area: {e}",
+                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
+            )
+        })?;
+        Ok(OwnedFd::from(file))
+    }
+}
+
+fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
+    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
+    // we don't get a segfault later when trying to actually use it.
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::fcntl::posix_fallocate(fd, 0, size as i64).map_err(|e| {
+            Error::new(
+                "could not grow shmem segment, posix_fallocate failed: {e}",
+                e,
+            )
+        })
+    }
+    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
+    #[cfg(target_os = "macos")]
+    {
+        nix::unistd::ftruncate(fd, size as i64)
+            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed: {e}", e))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use nix::unistd::ForkResult;
+    use std::ops::Range;
+
+    /// check that all bytes in given range have the expected value.
+    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
+        for i in range {
+            let b = unsafe { *(ptr.add(i)) };
+            assert_eq!(expected, b, "unexpected byte at offset {i}");
+        }
+    }
+
+    /// Write 'b' to all bytes in the given range
+    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
+        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
+    }
+
+    // simple single-process test of growing and shrinking
+    #[test]
+    fn test_shmem_resize() -> Result<(), Error> {
+        let max_size = 1024 * 1024;
+        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
+
+        assert_eq!(init_struct.current_size(), 0);
+
+        // Initial grow
+        let size1 = 10000;
+        init_struct.set_size(size1).unwrap();
+        assert_eq!(init_struct.current_size(), size1);
+
+        // Write some data
+        let data_ptr = init_struct.data_ptr.as_ptr();
+        write_range(data_ptr, 0xAA, 0..size1);
+        assert_range(data_ptr, 0xAA, 0..size1);
+
+        // Shrink
+        let size2 = 5000;
+        init_struct.set_size(size2).unwrap();
+        assert_eq!(init_struct.current_size(), size2);
+
+        // Grow again
+        let size3 = 20000;
+        init_struct.set_size(size3).unwrap();
+        assert_eq!(init_struct.current_size(), size3);
+
+        // Try to read it. The area that was shrunk and grown again should read as all zeros now
+        assert_range(data_ptr, 0xAA, 0..5000);
+        assert_range(data_ptr, 0, 5000..size1);
+
+        // Try to grow beyond max_size
+        //let size4 = max_size + 1;
+        //assert!(init_struct.set_size(size4).is_err());
+
+        // Dropping init_struct should unmap the memory
+        drop(init_struct);
+
+        Ok(())
+    }
+
+    /// This is used in tests to coordinate between test processes. It's like std::sync::Barrier,
+    /// but is stored in the shared memory area and works across processes. It's implemented by
+    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
+    struct SimpleBarrier {
+        num_procs: usize,
+        count: AtomicUsize,
+    }
+
+    impl SimpleBarrier {
+        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
+            unsafe {
+                *ptr = SimpleBarrier {
+                    num_procs,
+                    count: AtomicUsize::new(0),
+                }
+            }
+        }
+
+        pub fn wait(&self) {
+            let old = self.count.fetch_add(1, Ordering::Relaxed);
+
+            let generation = old / self.num_procs;
+
+            let mut current = old + 1;
+            while current < (generation + 1) * self.num_procs {
+                std::thread::sleep(std::time::Duration::from_millis(10));
+                current = self.count.load(Ordering::Relaxed);
+            }
+        }
+    }
+
+    #[test]
+    fn test_multi_process() {
+        // Initialize
+        let max_size = 1_000_000_000_000;
+        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
+        let ptr = init_struct.data_ptr.as_ptr();
+
+        // Store the SimpleBarrier in the first 1k of the area.
+        init_struct.set_size(10000).unwrap();
+        let barrier_ptr: *mut SimpleBarrier = unsafe {
+            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
+                .cast()
+        };
+        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
+        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
+
+        // Fork another test process. The code after this runs in both processes concurrently.
+        let fork_result = unsafe { nix::unistd::fork().unwrap() };
+
+        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, 1000..2000);
+        } else {
+            write_range(ptr, 0xBB, 2000..3000);
+        }
+        barrier.wait();
+        // Verify the contents. (in both processes)
+        assert_range(ptr, 0xAA, 1000..2000);
+        assert_range(ptr, 0xBB, 2000..3000);
+
+        // Grow, from the child this time
+        let size = 10_000_000;
+        if !fork_result.is_parent() {
+            init_struct.set_size(size).unwrap();
+        }
+        barrier.wait();
+
+        // make some writes at the end
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, (size - 10)..size);
+        } else {
+            write_range(ptr, 0xBB, (size - 20)..(size - 10));
+        }
+        barrier.wait();
+
+        // Verify the contents. (This runs in both processes)
+        assert_range(ptr, 0, (size - 1000)..(size - 20));
+        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
+        assert_range(ptr, 0xAA, (size - 10)..size);
+
+        if let ForkResult::Parent { child } = fork_result {
+            nix::sys::wait::waitpid(child, None).unwrap();
+        }
+    }
+}

libs/neon-shmem/src/shmem.rs

@@ -1,418 +0,0 @@
-//! Dynamically resizable contiguous chunk of shared memory
-
-use std::num::NonZeroUsize;
-use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
-use std::ptr::NonNull;
-use std::sync::atomic::{AtomicUsize, Ordering};
-
-use nix::errno::Errno;
-use nix::sys::mman::MapFlags;
-use nix::sys::mman::ProtFlags;
-use nix::sys::mman::mmap as nix_mmap;
-use nix::sys::mman::munmap as nix_munmap;
-use nix::unistd::ftruncate as nix_ftruncate;
-
-/// `ShmemHandle` represents a shared memory area that can be shared by processes over `fork()`.
-/// Unlike shared memory allocated by Postgres, this area is resizable, up to `max_size` that's
-/// specified at creation.
-///
-/// The area is backed by an anonymous file created with `memfd_create()`. The full address space for
-/// `max_size` is reserved up-front with `mmap()`, but whenever you call [`ShmemHandle::set_size`],
-/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
-/// will cause the file to be expanded, but we might use `mprotect()` etc. to enforce that in the
-/// future.
-pub struct ShmemHandle {
-    /// memfd file descriptor
-    fd: OwnedFd,
-
-    max_size: usize,
-
-    // Pointer to the beginning of the shared memory area. The header is stored there.
-    shared_ptr: NonNull<SharedStruct>,
-
-    // Pointer to the beginning of the user data
-    pub data_ptr: NonNull<u8>,
-}
-
-/// This is stored at the beginning in the shared memory area.
-struct SharedStruct {
-    max_size: usize,
-
-    /// Current size of the backing file. The high-order bit is used for the [`RESIZE_IN_PROGRESS`] flag.
-    current_size: AtomicUsize,
-}
-
-const RESIZE_IN_PROGRESS: usize = 1 << 63;
-
-const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
-
-/// Error type returned by the [`ShmemHandle`] functions.
-#[derive(thiserror::Error, Debug)]
-#[error("{msg}: {errno}")]
-pub struct Error {
-    pub msg: String,
-    pub errno: Errno,
-}
-
-impl Error {
-    fn new(msg: &str, errno: Errno) -> Self {
-        Self {
-            msg: msg.to_string(),
-            errno,
-        }
-    }
-}
-
-impl ShmemHandle {
-    /// Create a new shared memory area. To communicate between processes, the processes need to be
-    /// `fork()`'d after calling this, so that the `ShmemHandle` is inherited by all processes.
-    ///
-    /// If the `ShmemHandle` is dropped, the memory is unmapped from the current process. Other
-    /// processes can continue using it, however.
-    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<Self, Error> {
-        // create the backing anonymous file.
-        let fd = create_backing_file(name)?;
-
-        Self::new_with_fd(fd, initial_size, max_size)
-    }
-
-    fn new_with_fd(
-        fd: OwnedFd,
-        initial_size: usize,
-        max_size: usize,
-    ) -> Result<Self, Error> {
-        // We reserve the high-order bit for the `RESIZE_IN_PROGRESS` flag, and the actual size
-        // is a little larger than this because of the SharedStruct header. Make the upper limit
-        // somewhat smaller than that, because with anything close to that, you'll run out of
-        // memory anyway.
-        assert!(max_size < 1 << 48, "max size {max_size} too large");
-        
-        assert!(
-			initial_size <= max_size,
-            "initial size {initial_size} larger than max size {max_size}"
-        );
-
-        // The actual initial / max size is the one given by the caller, plus the size of
-        // 'SharedStruct'.
-        let initial_size = HEADER_SIZE + initial_size;
-        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
-
-        // Reserve address space for it with mmap
-        //
-        // TODO: Use MAP_HUGETLB if possible
-        let start_ptr = unsafe {
-            nix_mmap(
-                None,
-                max_size,
-                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
-                MapFlags::MAP_SHARED,
-                &fd,
-                0,
-            )
-        }
-        .map_err(|e| Error::new("mmap failed", e))?;
-
-        // Reserve space for the initial size
-        enlarge_file(fd.as_fd(), initial_size as u64)?;
-
-        // Initialize the header
-        let shared: NonNull<SharedStruct> = start_ptr.cast();
-        unsafe {
-            shared.write(SharedStruct {
-                max_size: max_size.into(),
-                current_size: AtomicUsize::new(initial_size),
-            });
-        }
-
-        // The user data begins after the header
-        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
-
-        Ok(Self {
-            fd,
-            max_size: max_size.into(),
-            shared_ptr: shared,
-            data_ptr,
-        })
-    }
-
-    // return reference to the header
-    fn shared(&self) -> &SharedStruct {
-        unsafe { self.shared_ptr.as_ref() }
-    }
-
-    /// Resize the shared memory area. `new_size` must not be larger than the `max_size` specified
-    /// when creating the area.
-    ///
-    /// This may only be called from one process/thread concurrently. We detect that case
-    /// and return an [`shmem::Error`](Error).
-    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
-        let new_size = new_size + HEADER_SIZE;
-        let shared = self.shared();
-
-        assert!(
-			new_size <= self.max_size,
-            "new size ({new_size}) is greater than max size ({})",
-			self.max_size
-        );
-
-		assert_eq!(self.max_size, shared.max_size);
-
-        // Lock the area by setting the bit in `current_size`
-        //
-        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
-        // and the `posix_fallocate`/`ftruncate` call is surely a synchronization point anyway. But
-        // since this is not performance-critical, better safe than sorry.
-        let mut old_size = shared.current_size.load(Ordering::Acquire);
-        loop {
-            if (old_size & RESIZE_IN_PROGRESS) != 0 {
-                return Err(Error::new(
-                    "concurrent resize detected",
-                    Errno::UnknownErrno,
-                ));
-            }
-            match shared.current_size.compare_exchange(
-                old_size,
-                new_size,
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            ) {
-                Ok(_) => break,
-                Err(x) => old_size = x,
-            }
-        }
-
-        // Ok, we got the lock.
-        //
-        // NB: If anything goes wrong, we *must* clear the bit!
-        let result = {
-            use std::cmp::Ordering::{Equal, Greater, Less};
-            match new_size.cmp(&old_size) {
-                Less => nix_ftruncate(&self.fd, new_size as i64).map_err(|e| {
-                    Error::new("could not shrink shmem segment, ftruncate failed", e)
-                }),
-                Equal => Ok(()),
-                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
-            }
-        };
-
-        // Unlock
-        shared.current_size.store(
-            if result.is_ok() { new_size } else { old_size },
-            Ordering::Release,
-        );
-
-        result
-    }
-
-    /// Returns the current user-visible size of the shared memory segment.
-    ///
-    /// NOTE: a concurrent [`ShmemHandle::set_size()`] call can change the size at any time.
-	/// It is the caller's responsibility not to access the area beyond the current size.
-    pub fn current_size(&self) -> usize {
-        let total_current_size =
-            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
-        total_current_size - HEADER_SIZE
-    }
-}
-
-impl Drop for ShmemHandle {
-    fn drop(&mut self) {
-        // SAFETY: The pointer was obtained from mmap() with the given size.
-        // We unmap the entire region.
-        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
-        // The fd is dropped automatically by OwnedFd.
-    }
-}
-
-/// Create a "backing file" for the shared memory area. On Linux, use `memfd_create()`, to create an
-/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
-/// development and testing, but in production we want the file to stay in memory.
-///
-/// Disable unused variables warnings because `name` is unused in the macos path.
-#[allow(unused_variables)]
-fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
-            .map_err(|e| Error::new("memfd_create failed", e))
-    }
-    #[cfg(target_os = "macos")]
-    {
-        let file = tempfile::tempfile().map_err(|e| {
-            Error::new(
-                "could not create temporary file to back shmem area",
-                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
-            )
-        })?;
-        Ok(OwnedFd::from(file))
-    }
-}
-
-fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
-    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
-    // we don't get a segfault later when trying to actually use it.
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::fcntl::posix_fallocate(fd, 0, size as i64).map_err(|e| {
-            Error::new(
-                "could not grow shmem segment, posix_fallocate failed",
-                e,
-            )
-        })
-    }
-    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
-    #[cfg(target_os = "macos")]
-    {
-        nix::unistd::ftruncate(fd, size as i64)
-            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed", e))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    use nix::unistd::ForkResult;
-    use std::ops::Range;
-
-    /// check that all bytes in given range have the expected value.
-    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
-        for i in range {
-            let b = unsafe { *(ptr.add(i)) };
-            assert_eq!(expected, b, "unexpected byte at offset {}", i);
-        }
-    }
-
-    /// Write 'b' to all bytes in the given range
-    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
-        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
-    }
-
-    // simple single-process test of growing and shrinking
-    #[test]
-    fn test_shmem_resize() -> Result<(), Error> {
-        let max_size = 1024 * 1024;
-        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
-
-        assert_eq!(init_struct.current_size(), 0);
-
-        // Initial grow
-        let size1 = 10000;
-        init_struct.set_size(size1).unwrap();
-        assert_eq!(init_struct.current_size(), size1);
-
-        // Write some data
-        let data_ptr = init_struct.data_ptr.as_ptr();
-        write_range(data_ptr, 0xAA, 0..size1);
-        assert_range(data_ptr, 0xAA, 0..size1);
-
-        // Shrink
-        let size2 = 5000;
-        init_struct.set_size(size2).unwrap();
-        assert_eq!(init_struct.current_size(), size2);
-
-        // Grow again
-        let size3 = 20000;
-        init_struct.set_size(size3).unwrap();
-        assert_eq!(init_struct.current_size(), size3);
-
-        // Try to read it. The area that was shrunk and grown again should read as all zeros now
-        assert_range(data_ptr, 0xAA, 0..5000);
-        assert_range(data_ptr, 0, 5000..size1);
-
-        // Try to grow beyond max_size
-        //let size4 = max_size + 1;
-        //assert!(init_struct.set_size(size4).is_err());
-
-        // Dropping init_struct should unmap the memory
-        drop(init_struct);
-
-        Ok(())
-    }
-
-    /// This is used in tests to coordinate between test processes. It's like `std::sync::Barrier`,
-    /// but is stored in the shared memory area and works across processes. It's implemented by
-    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
-    struct SimpleBarrier {
-        num_procs: usize,
-        count: AtomicUsize,
-    }
-
-    impl SimpleBarrier {
-        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
-            unsafe {
-                *ptr = SimpleBarrier {
-                    num_procs,
-                    count: AtomicUsize::new(0),
-                }
-            }
-        }
-
-        pub fn wait(&self) {
-            let old = self.count.fetch_add(1, Ordering::Relaxed);
-
-            let generation = old / self.num_procs;
-
-            let mut current = old + 1;
-            while current < (generation + 1) * self.num_procs {
-                std::thread::sleep(std::time::Duration::from_millis(10));
-                current = self.count.load(Ordering::Relaxed);
-            }
-        }
-    }
-
-    #[test]
-    fn test_multi_process() {
-        // Initialize
-        let max_size = 1_000_000_000_000;
-        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
-        let ptr = init_struct.data_ptr.as_ptr();
-
-        // Store the SimpleBarrier in the first 1k of the area.
-        init_struct.set_size(10000).unwrap();
-        let barrier_ptr: *mut SimpleBarrier = unsafe {
-            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
-                .cast()
-        };
-        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
-        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
-
-        // Fork another test process. The code after this runs in both processes concurrently.
-        let fork_result = unsafe { nix::unistd::fork().unwrap() };
-
-        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, 1000..2000);
-        } else {
-            write_range(ptr, 0xBB, 2000..3000);
-        }
-        barrier.wait();
-        // Verify the contents. (in both processes)
-        assert_range(ptr, 0xAA, 1000..2000);
-        assert_range(ptr, 0xBB, 2000..3000);
-
-        // Grow, from the child this time
-        let size = 10_000_000;
-        if !fork_result.is_parent() {
-            init_struct.set_size(size).unwrap();
-        }
-        barrier.wait();
-
-        // make some writes at the end
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, (size - 10)..size);
-        } else {
-            write_range(ptr, 0xBB, (size - 20)..(size - 10));
-        }
-        barrier.wait();
-
-        // Verify the contents. (This runs in both processes)
-        assert_range(ptr, 0, (size - 1000)..(size - 20));
-        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
-        assert_range(ptr, 0xAA, (size - 10)..size);
-
-        if let ForkResult::Parent { child } = fork_result {
-            nix::sys::wait::waitpid(child, None).unwrap();
-        }
-    }
-}

libs/neon-shmem/src/sync.rs

@@ -1,105 +0,0 @@
-//! Simple utilities akin to what's in [`std::sync`] but designed to work with shared memory.
-
-use std::mem::MaybeUninit;
-use std::ptr::NonNull;
-
-use nix::errno::Errno;
-
-pub type RwLock<T> = lock_api::RwLock<PthreadRwLock, T>;
-pub(crate) type RwLockReadGuard<'a, T> = lock_api::RwLockReadGuard<'a, PthreadRwLock, T>;
-pub type RwLockWriteGuard<'a, T> = lock_api::RwLockWriteGuard<'a, PthreadRwLock, T>;
-pub type ValueReadGuard<'a, T> = lock_api::MappedRwLockReadGuard<'a, PthreadRwLock, T>;
-pub type ValueWriteGuard<'a, T> = lock_api::MappedRwLockWriteGuard<'a, PthreadRwLock, T>;
-
-/// Shared memory read-write lock.
-pub struct PthreadRwLock(Option<NonNull<libc::pthread_rwlock_t>>);
-
-impl PthreadRwLock {
-	pub fn new(lock: *mut libc::pthread_rwlock_t) -> Self {
-		unsafe {
-			let mut attrs = MaybeUninit::uninit();
-			// Ignoring return value here - only possible error is OOM.
-			libc::pthread_rwlockattr_init(attrs.as_mut_ptr());
-			libc::pthread_rwlockattr_setpshared(
-				attrs.as_mut_ptr(),
-				libc::PTHREAD_PROCESS_SHARED
-			);
-			// TODO(quantumish): worth making this function return Result?
-			libc::pthread_rwlock_init(lock, attrs.as_mut_ptr());
-			// Safety: POSIX specifies that "any function affecting the attributes
-			// object (including destruction) shall not affect any previously
-			// initialized read-write locks". 
-			libc::pthread_rwlockattr_destroy(attrs.as_mut_ptr());
-			Self(Some(NonNull::new_unchecked(lock)))
-		}
-	}
-	
-	fn inner(&self) -> NonNull<libc::pthread_rwlock_t> {
-		match self.0 {
-			None => panic!("PthreadRwLock constructed badly - something likely used RawMutex::INIT"),
-			Some(x) => x,
-		}
-	}
-}
-
-unsafe impl lock_api::RawRwLock for PthreadRwLock {
-	type GuardMarker = lock_api::GuardSend;
-	const INIT: Self = Self(None);	
-	
-	fn lock_shared(&self) {
-		unsafe {
-			let res = libc::pthread_rwlock_rdlock(self.inner().as_ptr());
-			if res != 0 {
-				panic!("rdlock failed with {}", Errno::from_raw(res));
-			}
-		}
-	}
-
-	fn try_lock_shared(&self) -> bool {
-		unsafe {
-			let res = libc::pthread_rwlock_tryrdlock(self.inner().as_ptr());
-			match res {
-				0 => true,
-				libc::EAGAIN => false,
-				o => panic!("try_rdlock failed with {}", Errno::from_raw(res)),
-			}
-		}
-	}
-
-	fn lock_exclusive(&self) {
-		unsafe {
-			let res = libc::pthread_rwlock_wrlock(self.inner().as_ptr());
-			if res != 0 {
-				panic!("wrlock failed with {}", Errno::from_raw(res));
-			}
-		}
-	}
-
-	fn try_lock_exclusive(&self) -> bool {
-		unsafe {
-			let res = libc::pthread_rwlock_trywrlock(self.inner().as_ptr());
-			match res {
-				0 => true,
-				libc::EAGAIN => false,
-				o => panic!("try_wrlock failed with {}", Errno::from_raw(res)),
-			}
-		}
-	}
-
-	unsafe fn unlock_exclusive(&self) {
-		unsafe { 
-			let res = libc::pthread_rwlock_unlock(self.inner().as_ptr());
-			if res != 0 {
-				panic!("unlock failed with {}", Errno::from_raw(res));
-			}
-		}
-	}
-	unsafe fn unlock_shared(&self) {
-		unsafe {
-			let res = libc::pthread_rwlock_unlock(self.inner().as_ptr());
-			if res != 0 {
-				panic!("unlock failed with {}", Errno::from_raw(res));
-			}
-		}
-	}
-}

libs/neonart/Cargo.toml

@@ -1,14 +0,0 @@
-[package]
-name = "neonart"
-version = "0.1.0"
-edition.workspace = true
-license.workspace = true
-
-[dependencies]
-crossbeam-utils.workspace = true
-spin.workspace = true
-tracing.workspace = true
-
-[dev-dependencies]
-rand = "0.9.1"
-rand_distr = "0.5.1"

libs/neonart/src/algorithm.rs

@@ -1,594 +0,0 @@
-mod lock_and_version;
-pub(crate) mod node_ptr;
-mod node_ref;
-
-use std::vec::Vec;
-
-use crate::algorithm::lock_and_version::ConcurrentUpdateError;
-use crate::algorithm::node_ptr::MAX_PREFIX_LEN;
-use crate::algorithm::node_ref::{NewNodeRef, NodeRef, ReadLockedNodeRef, WriteLockedNodeRef};
-use crate::allocator::OutOfMemoryError;
-
-use crate::TreeWriteGuard;
-use crate::UpdateAction;
-use crate::allocator::ArtAllocator;
-use crate::epoch::EpochPin;
-use crate::{Key, Value};
-
-pub(crate) type RootPtr<V> = node_ptr::NodePtr<V>;
-
-#[derive(Debug)]
-pub enum ArtError {
-    ConcurrentUpdate, // need to retry
-    OutOfMemory,
-}
-
-impl From<ConcurrentUpdateError> for ArtError {
-    fn from(_: ConcurrentUpdateError) -> ArtError {
-        ArtError::ConcurrentUpdate
-    }
-}
-
-impl From<OutOfMemoryError> for ArtError {
-    fn from(_: OutOfMemoryError) -> ArtError {
-        ArtError::OutOfMemory
-    }
-}
-
-pub fn new_root<V: Value>(
-    allocator: &impl ArtAllocator<V>,
-) -> Result<RootPtr<V>, OutOfMemoryError> {
-    node_ptr::new_root(allocator)
-}
-
-pub(crate) fn search<'e, K: Key, V: Value>(
-    key: &K,
-    root: RootPtr<V>,
-    epoch_pin: &'e EpochPin,
-) -> Option<&'e V> {
-    loop {
-        let root_ref = NodeRef::from_root_ptr(root);
-        if let Ok(result) = lookup_recurse(key.as_bytes(), root_ref, None, epoch_pin) {
-            break result;
-        }
-        // retry
-    }
-}
-
-pub(crate) fn iter_next<'e, V: Value>(
-    key: &[u8],
-    root: RootPtr<V>,
-    epoch_pin: &'e EpochPin,
-) -> Option<(Vec<u8>, &'e V)> {
-    loop {
-        let mut path = Vec::new();
-        let root_ref = NodeRef::from_root_ptr(root);
-
-        match next_recurse(key, &mut path, root_ref, epoch_pin) {
-            Ok(Some(v)) => {
-                assert_eq!(path.len(), key.len());
-                break Some((path, v));
-            }
-            Ok(None) => break None,
-            Err(ConcurrentUpdateError()) => {
-                // retry
-                continue;
-            }
-        }
-    }
-}
-
-pub(crate) fn update_fn<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>, F>(
-    key: &K,
-    value_fn: F,
-    root: RootPtr<V>,
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), OutOfMemoryError>
-where
-    F: FnOnce(Option<&V>) -> UpdateAction<V>,
-{
-    let value_fn_cell = std::cell::Cell::new(Some(value_fn));
-    loop {
-        let root_ref = NodeRef::from_root_ptr(root);
-        let this_value_fn = |arg: Option<&V>| value_fn_cell.take().unwrap()(arg);
-        let key_bytes = key.as_bytes();
-
-        match update_recurse(
-            key_bytes,
-            this_value_fn,
-            root_ref,
-            None,
-            None,
-            guard,
-            0,
-            key_bytes,
-        ) {
-            Ok(()) => break Ok(()),
-            Err(ArtError::ConcurrentUpdate) => {
-                continue; // retry
-            }
-            Err(ArtError::OutOfMemory) => break Err(OutOfMemoryError()),
-        }
-    }
-}
-
-// Error means you must retry.
-//
-// This corresponds to the 'lookupOpt' function in the paper
-fn lookup_recurse<'e, V: Value>(
-    key: &[u8],
-    node: NodeRef<'e, V>,
-    parent: Option<ReadLockedNodeRef<V>>,
-    epoch_pin: &'e EpochPin,
-) -> Result<Option<&'e V>, ConcurrentUpdateError> {
-    let rnode = node.read_lock_or_restart()?;
-    if let Some(parent) = parent {
-        parent.read_unlock_or_restart()?;
-    }
-
-    // check if the prefix matches, may increment level
-    let prefix_len = if let Some(prefix_len) = rnode.prefix_matches(key) {
-        prefix_len
-    } else {
-        rnode.read_unlock_or_restart()?;
-        return Ok(None);
-    };
-
-    if rnode.is_leaf() {
-        assert_eq!(key.len(), prefix_len);
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let v = unsafe { vptr.as_ref().unwrap() };
-        return Ok(Some(v));
-    }
-
-    let key = &key[prefix_len..];
-
-    // find child (or leaf value)
-    let next_node = rnode.find_child_or_restart(key[0])?;
-
-    match next_node {
-        None => Ok(None), // key not found
-        Some(child) => lookup_recurse(&key[1..], child, Some(rnode), epoch_pin),
-    }
-}
-
-fn next_recurse<'e, V: Value>(
-    min_key: &[u8],
-    path: &mut Vec<u8>,
-    node: NodeRef<'e, V>,
-    epoch_pin: &'e EpochPin,
-) -> Result<Option<&'e V>, ConcurrentUpdateError> {
-    let rnode = node.read_lock_or_restart()?;
-    let prefix = rnode.get_prefix();
-    if prefix.len() != 0 {
-        path.extend_from_slice(prefix);
-    }
-
-    use std::cmp::Ordering;
-    let comparison = path.as_slice().cmp(&min_key[0..path.len()]);
-    if comparison == Ordering::Less {
-        rnode.read_unlock_or_restart()?;
-        return Ok(None);
-    }
-
-    if rnode.is_leaf() {
-        assert_eq!(path.len(), min_key.len());
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let v = unsafe { vptr.as_ref().unwrap() };
-        return Ok(Some(v));
-    }
-
-    let mut min_key_byte = match comparison {
-        Ordering::Less => unreachable!(), // checked this above already
-        Ordering::Equal => min_key[path.len()],
-        Ordering::Greater => 0,
-    };
-
-    loop {
-        match rnode.find_next_child_or_restart(min_key_byte)? {
-            None => {
-                return Ok(None);
-            }
-            Some((key_byte, child_ref)) => {
-                let path_len = path.len();
-                path.push(key_byte);
-                let result = next_recurse(min_key, path, child_ref, epoch_pin)?;
-                if result.is_some() {
-                    return Ok(result);
-                }
-                if key_byte == u8::MAX {
-                    return Ok(None);
-                }
-                path.truncate(path_len);
-                min_key_byte = key_byte + 1;
-            }
-        }
-    }
-}
-
-// This corresponds to the 'insertOpt' function in the paper
-pub(crate) fn update_recurse<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>, F>(
-    key: &[u8],
-    value_fn: F,
-    node: NodeRef<'e, V>,
-    rparent: Option<(ReadLockedNodeRef<V>, u8)>,
-    rgrandparent: Option<(ReadLockedNodeRef<V>, u8)>,
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-    level: usize,
-    orig_key: &[u8],
-) -> Result<(), ArtError>
-where
-    F: FnOnce(Option<&V>) -> UpdateAction<V>,
-{
-    let rnode = node.read_lock_or_restart()?;
-
-    let prefix_match_len = rnode.prefix_matches(key);
-    if prefix_match_len.is_none() {
-        let (rparent, parent_key) = rparent.expect("direct children of the root have no prefix");
-        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-        match value_fn(None) {
-            UpdateAction::Nothing => {}
-            UpdateAction::Insert(new_value) => {
-                insert_split_prefix(key, new_value, &mut wnode, &mut wparent, parent_key, guard)?;
-            }
-            UpdateAction::Remove => {
-                panic!("unexpected Remove action on insertion");
-            }
-        }
-        wnode.write_unlock();
-        wparent.write_unlock();
-        return Ok(());
-    }
-    let prefix_match_len = prefix_match_len.unwrap();
-    let key = &key[prefix_match_len as usize..];
-    let level = level + prefix_match_len as usize;
-
-    if rnode.is_leaf() {
-        assert_eq!(key.len(), 0);
-        let (rparent, parent_key) = rparent.expect("root cannot be leaf");
-        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-        // safety: Now that we have acquired the write lock, we have exclusive access to the
-        // value. XXX: There might be concurrent reads though?
-        let value_mut = wnode.get_leaf_value_mut();
-
-        match value_fn(Some(value_mut)) {
-            UpdateAction::Nothing => {
-                wparent.write_unlock();
-                wnode.write_unlock();
-            }
-            UpdateAction::Insert(_) => panic!("cannot insert over existing value"),
-            UpdateAction::Remove => {
-                guard.remember_obsolete_node(wnode.as_ptr());
-                wparent.delete_child(parent_key);
-                wnode.write_unlock_obsolete();
-
-                if let Some(rgrandparent) = rgrandparent {
-                    // FIXME: Ignore concurrency error. It doesn't lead to
-                    // corruption, but it means we might leak something. Until
-                    // another update cleans it up.
-                    let _ = cleanup_parent(wparent, rgrandparent, guard);
-                }
-            }
-        }
-
-        return Ok(());
-    }
-
-    let next_node = rnode.find_child_or_restart(key[0])?;
-
-    if next_node.is_none() {
-        if rnode.is_full() {
-            let (rparent, parent_key) = rparent.expect("root node cannot become full");
-            let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
-            let wnode = rnode.upgrade_to_write_lock_or_restart()?;
-
-            match value_fn(None) {
-                UpdateAction::Nothing => {
-                    wnode.write_unlock();
-                    wparent.write_unlock();
-                }
-                UpdateAction::Insert(new_value) => {
-                    insert_and_grow(key, new_value, wnode, &mut wparent, parent_key, guard)?;
-                    wparent.write_unlock();
-                }
-                UpdateAction::Remove => {
-                    panic!("unexpected Remove action on insertion");
-                }
-            };
-        } else {
-            let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
-            if let Some((rparent, _)) = rparent {
-                rparent.read_unlock_or_restart()?;
-            }
-            match value_fn(None) {
-                UpdateAction::Nothing => {}
-                UpdateAction::Insert(new_value) => {
-                    insert_to_node(&mut wnode, key, new_value, guard)?;
-                }
-                UpdateAction::Remove => {
-                    panic!("unexpected Remove action on insertion");
-                }
-            };
-            wnode.write_unlock();
-        }
-        return Ok(());
-    } else {
-        let next_child = next_node.unwrap(); // checked above it's not None
-        if let Some((ref rparent, _)) = rparent {
-            rparent.check_or_restart()?;
-        }
-
-        // recurse to next level
-        update_recurse(
-            &key[1..],
-            value_fn,
-            next_child,
-            Some((rnode, key[0])),
-            rparent,
-            guard,
-            level + 1,
-            orig_key,
-        )
-    }
-}
-
-#[derive(Clone)]
-enum PathElement {
-    Prefix(Vec<u8>),
-    KeyByte(u8),
-}
-
-impl std::fmt::Debug for PathElement {
-    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
-        match self {
-            PathElement::Prefix(prefix) => write!(fmt, "{:?}", prefix),
-            PathElement::KeyByte(key_byte) => write!(fmt, "{}", key_byte),
-        }
-    }
-}
-
-pub(crate) fn dump_tree<'e, V: Value + std::fmt::Debug>(
-    root: RootPtr<V>,
-    epoch_pin: &'e EpochPin,
-    dst: &mut dyn std::io::Write,
-) {
-    let root_ref = NodeRef::from_root_ptr(root);
-
-    let _ = dump_recurse(&[], root_ref, &epoch_pin, 0, dst);
-}
-
-// TODO: return an Err if writeln!() returns error, instead of unwrapping
-fn dump_recurse<'e, V: Value + std::fmt::Debug>(
-    path: &[PathElement],
-    node: NodeRef<'e, V>,
-    epoch_pin: &'e EpochPin,
-    level: usize,
-    dst: &mut dyn std::io::Write,
-) -> Result<(), ConcurrentUpdateError> {
-    let indent = str::repeat(" ", level);
-
-    let rnode = node.read_lock_or_restart()?;
-    let mut path = Vec::from(path);
-    let prefix = rnode.get_prefix();
-    if prefix.len() != 0 {
-        path.push(PathElement::Prefix(Vec::from(prefix)));
-    }
-
-    if rnode.is_leaf() {
-        let vptr = rnode.get_leaf_value_ptr()?;
-        // safety: It's OK to return a ref of the pointer because we checked the version
-        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
-        // as long as the epoch is pinned.
-        let val = unsafe { vptr.as_ref().unwrap() };
-        writeln!(dst, "{} {:?}: {:?}", indent, path, val).unwrap();
-        return Ok(());
-    }
-
-    for key_byte in 0..=u8::MAX {
-        match rnode.find_child_or_restart(key_byte)? {
-            None => continue,
-            Some(child_ref) => {
-                let rchild = child_ref.read_lock_or_restart()?;
-                writeln!(
-                    dst,
-                    "{} {:?}, {}: prefix {:?}",
-                    indent,
-                    &path,
-                    key_byte,
-                    rchild.get_prefix()
-                )
-                .unwrap();
-
-                let mut child_path = path.clone();
-                child_path.push(PathElement::KeyByte(key_byte));
-
-                dump_recurse(&child_path, child_ref, epoch_pin, level + 1, dst)?;
-            }
-        }
-    }
-
-    Ok(())
-}
-
-///```text
-///        [fooba]r -> value
-///
-/// [foo]b -> [a]r  -> value
-///      e -> [ls]e -> value
-///```
-fn insert_split_prefix<'e, K: Key, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    node: &mut WriteLockedNodeRef<V>,
-    parent: &mut WriteLockedNodeRef<V>,
-    parent_key: u8,
-    guard: &'e TreeWriteGuard<K, V, A>,
-) -> Result<(), OutOfMemoryError> {
-    let old_node = node;
-    let old_prefix = old_node.get_prefix();
-    let common_prefix_len = common_prefix(key, old_prefix);
-
-    // Allocate a node for the new value.
-    let new_value_node = allocate_node_for_value(
-        &key[common_prefix_len + 1..],
-        value,
-        guard.tree_writer.allocator,
-    )?;
-
-    // Allocate a new internal node with the common prefix
-    // FIXME: deallocate 'new_value_node' on OOM
-    let mut prefix_node =
-        node_ref::new_internal(&key[..common_prefix_len], guard.tree_writer.allocator)?;
-
-    // Add the old node and the new nodes to the new internal node
-    prefix_node.insert_old_child(old_prefix[common_prefix_len], old_node);
-    prefix_node.insert_new_child(key[common_prefix_len], new_value_node);
-
-    // Modify the prefix of the old child in place
-    old_node.truncate_prefix(old_prefix.len() - common_prefix_len - 1);
-
-    // replace the pointer in the parent
-    parent.replace_child(parent_key, prefix_node.into_ptr());
-
-    Ok(())
-}
-
-fn insert_to_node<'e, K: Key, V: Value, A: ArtAllocator<V>>(
-    wnode: &mut WriteLockedNodeRef<V>,
-    key: &[u8],
-    value: V,
-    guard: &'e TreeWriteGuard<K, V, A>,
-) -> Result<(), OutOfMemoryError> {
-    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
-    wnode.insert_child(key[0], value_child.into_ptr());
-    Ok(())
-}
-
-// On entry: 'parent' and 'node' are locked
-fn insert_and_grow<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    wnode: WriteLockedNodeRef<V>,
-    parent: &mut WriteLockedNodeRef<V>,
-    parent_key_byte: u8,
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), ArtError> {
-    let mut bigger_node = wnode.grow(guard.tree_writer.allocator)?;
-
-    // FIXME: deallocate 'bigger_node' on OOM
-    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
-    bigger_node.insert_new_child(key[0], value_child);
-
-    // Replace the pointer in the parent
-    parent.replace_child(parent_key_byte, bigger_node.into_ptr());
-
-    guard.remember_obsolete_node(wnode.as_ptr());
-    wnode.write_unlock_obsolete();
-
-    Ok(())
-}
-
-fn cleanup_parent<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
-    wparent: WriteLockedNodeRef<V>,
-    rgrandparent: (ReadLockedNodeRef<V>, u8),
-    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
-) -> Result<(), ArtError> {
-    let (rgrandparent, grandparent_key_byte) = rgrandparent;
-
-    // If the parent becomes completely empty after the deletion, remove the parent from the
-    // grandparent. (This case is possible because we reserve only 8 bytes for the prefix.)
-    // TODO: not implemented.
-
-    // If the parent has only one child, replace the parent with the remaining child. (This is not
-    // possible if the child's prefix field cannot absorb the parent's)
-    if wparent.num_children() == 1 {
-        // Try to lock the remaining child. This can fail if the child is updated
-        // concurrently.
-        let (key_byte, remaining_child) = wparent.find_remaining_child();
-
-        let mut wremaining_child = remaining_child.write_lock_or_restart()?;
-
-        if 1 + wremaining_child.get_prefix().len() + wparent.get_prefix().len() <= MAX_PREFIX_LEN {
-            let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
-
-            // Ok, we have locked the leaf, the parent, the grandparent, and the parent's only
-            // remaining leaf. Proceed with the updates.
-
-            // Update the prefix on the remaining leaf
-            wremaining_child.prepend_prefix(wparent.get_prefix(), key_byte);
-
-            // Replace the pointer in the grandparent to point directly to the remaining leaf
-            wgrandparent.replace_child(grandparent_key_byte, wremaining_child.as_ptr());
-
-            // Mark the parent as deleted.
-            guard.remember_obsolete_node(wparent.as_ptr());
-            wparent.write_unlock_obsolete();
-            return Ok(());
-        }
-    }
-
-    // If the parent's children would fit on a smaller node type after the deletion, replace it with
-    // a smaller node.
-    if wparent.can_shrink() {
-        let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
-        let smaller_node = wparent.shrink(guard.tree_writer.allocator)?;
-
-        // Replace the pointer in the grandparent
-        wgrandparent.replace_child(grandparent_key_byte, smaller_node.into_ptr());
-
-        guard.remember_obsolete_node(wparent.as_ptr());
-        wparent.write_unlock_obsolete();
-        return Ok(());
-    }
-
-    // nothing to do
-    wparent.write_unlock();
-    Ok(())
-}
-
-// Allocate a new leaf node to hold 'value'. If the key is long, we
-// may need to allocate new internal nodes to hold it too
-fn allocate_node_for_value<'a, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError> {
-    let mut prefix_off = key.len().saturating_sub(MAX_PREFIX_LEN);
-
-    let leaf_node = node_ref::new_leaf(&key[prefix_off..key.len()], value, allocator)?;
-
-    let mut node = leaf_node;
-    while prefix_off > 0 {
-        // Need another internal node
-        let remain_prefix = &key[0..prefix_off];
-
-        prefix_off = remain_prefix.len().saturating_sub(MAX_PREFIX_LEN + 1);
-        let mut internal_node = node_ref::new_internal(
-            &remain_prefix[prefix_off..remain_prefix.len() - 1],
-            allocator,
-        )?;
-        internal_node.insert_new_child(*remain_prefix.last().unwrap(), node);
-        node = internal_node;
-    }
-
-    Ok(node)
-}
-
-fn common_prefix(a: &[u8], b: &[u8]) -> usize {
-    for i in 0..MAX_PREFIX_LEN {
-        if a[i] != b[i] {
-            return i;
-        }
-    }
-    panic!("prefixes are equal");
-}

libs/neonart/src/algorithm/lock_and_version.rs

@@ -1,117 +0,0 @@
-//! Each node in the tree has contains one atomic word that stores three things:
-//!
-//! Bit 0: set if the node is "obsolete". An obsolete node has been removed from the tree,
-//!        but might still be accessed by concurrent readers until the epoch expires.
-//! Bit 1: set if the node is currently write-locked. Used as a spinlock.
-//! Bits 2-63: Version number, incremented every time the node is modified.
-//!
-//! AtomicLockAndVersion represents that.
-
-use std::sync::atomic::{AtomicU64, Ordering};
-
-pub(crate) struct ConcurrentUpdateError();
-
-pub(crate) struct AtomicLockAndVersion {
-    inner: AtomicU64,
-}
-
-impl AtomicLockAndVersion {
-    pub(crate) fn new() -> AtomicLockAndVersion {
-        AtomicLockAndVersion {
-            inner: AtomicU64::new(0),
-        }
-    }
-}
-
-impl AtomicLockAndVersion {
-    pub(crate) fn read_lock_or_restart(&self) -> Result<u64, ConcurrentUpdateError> {
-        let version = self.await_node_unlocked();
-        if is_obsolete(version) {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(version)
-    }
-
-    pub(crate) fn check_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
-        self.read_unlock_or_restart(version)
-    }
-
-    pub(crate) fn read_unlock_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
-        if self.inner.load(Ordering::Acquire) != version {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn upgrade_to_write_lock_or_restart(
-        &self,
-        version: u64,
-    ) -> Result<(), ConcurrentUpdateError> {
-        if self
-            .inner
-            .compare_exchange(
-                version,
-                set_locked_bit(version),
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            )
-            .is_err()
-        {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn write_lock_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
-        let old = self.inner.load(Ordering::Relaxed);
-        if is_obsolete(old) || is_locked(old) {
-            return Err(ConcurrentUpdateError());
-        }
-        if self
-            .inner
-            .compare_exchange(
-                old,
-                set_locked_bit(old),
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            )
-            .is_err()
-        {
-            return Err(ConcurrentUpdateError());
-        }
-        Ok(())
-    }
-
-    pub(crate) fn write_unlock(&self) {
-        // reset locked bit and overflow into version
-        self.inner.fetch_add(2, Ordering::Release);
-    }
-
-    pub(crate) fn write_unlock_obsolete(&self) {
-        // set obsolete, reset locked, overflow into version
-        self.inner.fetch_add(3, Ordering::Release);
-    }
-
-    // Helper functions
-    fn await_node_unlocked(&self) -> u64 {
-        let mut version = self.inner.load(Ordering::Acquire);
-        while is_locked(version) {
-            // spinlock
-            std::thread::yield_now();
-            version = self.inner.load(Ordering::Acquire)
-        }
-        version
-    }
-}
-
-fn set_locked_bit(version: u64) -> u64 {
-    return version + 2;
-}
-
-fn is_obsolete(version: u64) -> bool {
-    return (version & 1) == 1;
-}
-
-fn is_locked(version: u64) -> bool {
-    return (version & 2) == 2;
-}