proxy: add benchmark for custom json logging vs official fmt logger

We don't want to depend on postgres_ffi in an API crate. If there is no
such dependency, we can compile stuff like `storcon_cli` without needing
a full working postgres build. Fixes regression of #12548 (before we
could compile it).

.config/hakari.toml

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

.dockerignore

@@ -27,4 +27,4 @@
 !storage_controller/
 !vendor/postgres-*/
 !workspace_hack/
-!build_tools/patches
+!build-tools/patches

.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
@@ -30,6 +31,7 @@ config-variables:
   - NEON_PROD_AWS_ACCOUNT_ID
   - PGREGRESS_PG16_PROJECT_ID
   - PGREGRESS_PG17_PROJECT_ID
+  - PREWARM_PGBENCH_SIZE
   - REMOTE_STORAGE_AZURE_CONTAINER
   - REMOTE_STORAGE_AZURE_REGION
   - SLACK_CICD_CHANNEL_ID

.github/actions/run-python-test-set/action.yml

@@ -176,7 +176,11 @@ runs:
         fi
 
         if [[ $BUILD_TYPE == "debug" && $RUNNER_ARCH == 'X64' ]]; then
-          cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          # We don't use code coverage for regression tests (the step is disabled),
+          # so there's no need to collect it.
+          # Ref https://github.com/neondatabase/neon/issues/4540
+          # cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          cov_prefix=()
         else
           cov_prefix=()
         fi

.github/workflows/_build-and-test-locally.yml

@@ -150,7 +150,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
@@ -162,7 +162,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
@@ -174,7 +174,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v16
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v17 build
         id: cache_pg_17
@@ -186,7 +186,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v17
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Build all
         # Note: the Makefile picks up BUILD_TYPE and CARGO_PROFILE from the env variables

.github/workflows/benchmarking.yml

@@ -219,6 +219,7 @@ jobs:
           --ignore test_runner/performance/test_cumulative_statistics_persistence.py
           --ignore test_runner/performance/test_perf_many_relations.py
           --ignore test_runner/performance/test_perf_oltp_large_tenant.py
+          --ignore test_runner/performance/test_lfc_prewarm.py
       env:
         BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -410,6 +411,77 @@ jobs:
       env:
         SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
 
+  prewarm-test:
+    if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PGBENCH_SIZE: ${{ vars.PREWARM_PGBENCH_SIZE }}
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 17
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
+      PLATFORM: "neon-staging"
+
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    container:
+      image: ghcr.io/neondatabase/build-tools:pinned-bookworm
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+
+    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
+
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+
+    - name: Run prewarm benchmark
+      uses: ./.github/actions/run-python-test-set
+      with:
+        build_type: ${{ env.BUILD_TYPE }}
+        test_selection: performance/test_lfc_prewarm.py
+        run_in_parallel: false
+        save_perf_report: ${{ env.SAVE_PERF_REPORT }}
+        extra_params: -m remote_cluster --timeout 5400
+        pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        NEON_API_KEY: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Create Allure report
+      id: create-allure-report
+      if: ${{ !cancelled() }}
+      uses: ./.github/actions/allure-report-generate
+      with:
+        store-test-results-into-db: true
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
+
   generate-matrices:
     if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
     # Create matrices for the benchmarking jobs, so we run benchmarks on rds only once a week (on Saturday)

.github/workflows/build-build-tools-image.yml

@@ -72,7 +72,7 @@ jobs:
           ARCHS: ${{ inputs.archs || '["x64","arm64"]' }}
           DEBIANS: ${{ inputs.debians || '["bullseye","bookworm"]' }}
           IMAGE_TAG: |
-            ${{ hashFiles('build-tools.Dockerfile',
+            ${{ hashFiles('build-tools/Dockerfile',
                           '.github/workflows/build-build-tools-image.yml') }}
         run: |
           echo "archs=${ARCHS}"           | tee -a ${GITHUB_OUTPUT}
@@ -144,7 +144,7 @@ jobs:
 
       - uses: docker/build-push-action@471d1dc4e07e5cdedd4c2171150001c434f0b7a4 # v6.15.0
         with:
-          file: build-tools.Dockerfile
+          file: build-tools/Dockerfile
           context: .
           provenance: false
           push: true

.github/workflows/build-macos.yml

@@ -32,162 +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$(sysctl -n hw.ncpu)
-
-      - name: Upload "pg_install/${{ 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
-
-      # `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
-        run: |
-          chmod +x pg_install/v*/bin/*
-
-      - 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) PG_INSTALL_CACHED=1
-
-      - 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
@@ -203,41 +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
@@ -249,17 +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

@@ -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/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/proxy-benchmark.yml

@@ -1,4 +1,4 @@
-name: Periodic proxy performance test on unit-perf hetzner runner
+name: Periodic proxy performance test on unit-perf-aws-arm runners
 
 on:
   push: # TODO: remove after testing
@@ -32,7 +32,7 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [self-hosted, unit-perf]
+    runs-on: [self-hosted, unit-perf-aws-arm]
     timeout-minutes: 60  # 1h timeout
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm

.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

.gitmodules

@@ -1,16 +1,16 @@
 [submodule "vendor/postgres-v14"]
 	path = vendor/postgres-v14
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_14_STABLE_neon
 [submodule "vendor/postgres-v15"]
 	path = vendor/postgres-v15
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_15_STABLE_neon
 [submodule "vendor/postgres-v16"]
 	path = vendor/postgres-v16
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_16_STABLE_neon
 [submodule "vendor/postgres-v17"]
 	path = vendor/postgres-v17
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_17_STABLE_neon

Cargo.lock

@@ -253,17 +253,6 @@ version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a8ab6b55fe97976e46f91ddbed8d147d966475dc29b2032757ba47e02376fbc3"
 
-[[package]]
-name = "atomic_enum"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99e1aca718ea7b89985790c94aad72d77533063fe00bc497bb79a7c2dae6a661"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn 2.0.100",
-]
-
 [[package]]
 name = "autocfg"
 version = "1.1.0"
@@ -698,40 +687,13 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum"
-version = "0.7.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
-dependencies = [
- "async-trait",
- "axum-core 0.4.5",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "itoa",
- "matchit 0.7.3",
- "memchr",
- "mime",
- "percent-encoding",
- "pin-project-lite",
- "rustversion",
- "serde",
- "sync_wrapper 1.0.1",
- "tower 0.5.2",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum"
 version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6d6fd624c75e18b3b4c6b9caf42b1afe24437daaee904069137d8bab077be8b8"
 dependencies = [
- "axum-core 0.5.0",
+ "axum-core",
  "base64 0.22.1",
  "bytes",
  "form_urlencoded",
@@ -739,10 +701,10 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "itoa",
- "matchit 0.8.4",
+ "matchit",
  "memchr",
  "mime",
  "percent-encoding",
@@ -762,26 +724,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum-core"
-version = "0.4.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09f2bd6146b97ae3359fa0cc6d6b376d9539582c7b4220f041a33ec24c226199"
-dependencies = [
- "async-trait",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "mime",
- "pin-project-lite",
- "rustversion",
- "sync_wrapper 1.0.1",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum-core"
 version = "0.5.0"
@@ -808,8 +750,8 @@ version = "0.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fc6f625a1f7705c6cf62d0d070794e94668988b1c38111baeec177c715f7b"
 dependencies = [
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "bytes",
  "form_urlencoded",
  "futures-util",
@@ -1143,12 +1085,12 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
 [[package]]
 name = "cbindgen"
-version = "0.28.0"
+version = "0.29.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eadd868a2ce9ca38de7eeafdcec9c7065ef89b42b32f0839278d55f35c54d1ff"
+checksum = "975982cdb7ad6a142be15bdf84aea7ec6a9e5d4d797c004d43185b24cfe4e684"
 dependencies = [
  "clap",
- "heck 0.4.1",
+ "heck",
  "indexmap 2.9.0",
  "log",
  "proc-macro2",
@@ -1286,7 +1228,7 @@ version = "4.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4ac6a0c7b1a9e9a5186361f67dfa1b88213572f427fb9ab038efb2bd8c582dab"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -1348,28 +1290,9 @@ dependencies = [
 name = "communicator"
 version = "0.1.0"
 dependencies = [
- "atomic_enum",
- "axum 0.8.1",
- "bytes",
  "cbindgen",
- "clashmap",
- "http 1.1.0",
- "libc",
- "metrics",
  "neon-shmem",
- "nix 0.30.1",
- "pageserver_client_grpc",
- "pageserver_page_api",
- "prometheus",
- "prost 0.13.5",
- "thiserror 1.0.69",
- "tokio",
- "tokio-pipe",
- "tonic 0.12.3",
- "tracing",
- "tracing-subscriber",
- "uring-common",
- "utils",
+ "workspace_hack",
 ]
 
 [[package]]
@@ -1398,7 +1321,7 @@ dependencies = [
  "aws-sdk-kms",
  "aws-sdk-s3",
  "aws-smithy-types",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "base64 0.22.1",
  "bytes",
@@ -1410,6 +1333,7 @@ dependencies = [
  "fail",
  "flate2",
  "futures",
+ "hostname-validator",
  "http 1.1.0",
  "indexmap 2.9.0",
  "itertools 0.10.5",
@@ -1424,6 +1348,7 @@ dependencies = [
  "p256 0.13.2",
  "pageserver_page_api",
  "postgres",
+ "postgres-types",
  "postgres_initdb",
  "postgres_versioninfo",
  "regex",
@@ -1701,9 +1626,9 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-utils"
-version = "0.8.21"
+version = "0.8.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
+checksum = "248e3bacc7dc6baa3b21e405ee045c3047101a49145e7e9eca583ab4c2ca5345"
 
 [[package]]
 name = "crossterm"
@@ -2043,7 +1968,7 @@ checksum = "0892a17df262a24294c382f0d5997571006e7a4348b4327557c4ff1cd4a8bccc"
 dependencies = [
  "darling",
  "either",
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -2157,7 +2082,7 @@ name = "endpoint_storage"
 version = "0.0.1"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "camino",
  "camino-tempfile",
@@ -2438,7 +2363,7 @@ dependencies = [
  "futures-core",
  "futures-sink",
  "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "pin-project",
  "rand 0.8.5",
@@ -2608,18 +2533,6 @@ dependencies = [
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "getrandom"
-version = "0.3.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "73fea8450eea4bac3940448fb7ae50d91f034f941199fcd9d909a5a07aa455f0"
-dependencies = [
- "cfg-if",
- "libc",
- "r-efi",
- "wasi 0.14.2+wasi-0.2.4",
-]
-
 [[package]]
 name = "gettid"
 version = "0.1.3"
@@ -2832,12 +2745,6 @@ dependencies = [
  "http 1.1.0",
 ]
 
-[[package]]
-name = "heck"
-version = "0.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
-
 [[package]]
 name = "heck"
 version = "0.5.0"
@@ -2894,6 +2801,12 @@ dependencies = [
  "windows",
 ]
 
+[[package]]
+name = "hostname-validator"
+version = "1.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f558a64ac9af88b5ba400d99b579451af0d39c6d360980045b91aac966d705e2"
+
 [[package]]
 name = "http"
 version = "0.2.9"
@@ -3009,9 +2922,9 @@ dependencies = [
 
 [[package]]
 name = "httparse"
-version = "1.10.1"
+version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
+checksum = "d897f394bad6a705d5f4104762e116a75639e470d80901eed05a860a95cb1904"
 
 [[package]]
 name = "httpdate"
@@ -3061,9 +2974,9 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "1.6.0"
+version = "1.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cc2b571658e38e0c01b1fdca3bbbe93c00d3d71693ff2770043f8c29bc7d6f80"
+checksum = "50dfd22e0e76d0f662d429a5f80fcaf3855009297eab6a0a9f8543834744ba05"
 dependencies = [
  "bytes",
  "futures-channel",
@@ -3103,7 +3016,7 @@ checksum = "a0bea761b46ae2b24eb4aef630d8d1c398157b6fc29e6350ecf090a0b70c952c"
 dependencies = [
  "futures-util",
  "http 1.1.0",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "rustls 0.22.4",
  "rustls-pki-types",
@@ -3118,7 +3031,7 @@ version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3203a961e5c83b6f5498933e78b6b263e208c197b63e9c6c53cc82ffd3f63793"
 dependencies = [
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "pin-project-lite",
  "tokio",
@@ -3127,20 +3040,20 @@ dependencies = [
 
 [[package]]
 name = "hyper-util"
-version = "0.1.12"
+version = "0.1.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cf9f1e950e0d9d1d3c47184416723cf29c0d1f93bd8cccf37e4beb6b44f31710"
+checksum = "cde7055719c54e36e95e8719f95883f22072a48ede39db7fc17a4e1d5281e9b9"
 dependencies = [
  "bytes",
  "futures-channel",
  "futures-util",
  "http 1.1.0",
  "http-body 1.0.0",
- "hyper 1.6.0",
- "libc",
+ "hyper 1.4.1",
  "pin-project-lite",
  "socket2",
  "tokio",
+ "tower 0.4.13",
  "tower-service",
  "tracing",
 ]
@@ -3577,6 +3490,15 @@ dependencies = [
  "wasm-bindgen",
 ]
 
+[[package]]
+name = "json"
+version = "0.1.0"
+dependencies = [
+ "futures",
+ "itoa",
+ "ryu",
+]
+
 [[package]]
 name = "json-structural-diff"
 version = "0.2.0"
@@ -3729,12 +3651,6 @@ dependencies = [
  "regex-automata 0.1.10",
 ]
 
-[[package]]
-name = "matchit"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"
-
 [[package]]
 name = "matchit"
 version = "0.8.4"
@@ -3780,7 +3696,7 @@ version = "0.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b9e6777fc80a575f9503d908c8b498782a6c3ee88a06cb416dc3941401e43b94"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -3842,7 +3758,7 @@ dependencies = [
  "procfs",
  "prometheus",
  "rand 0.8.5",
- "rand_distr 0.4.3",
+ "rand_distr",
  "twox-hash",
 ]
 
@@ -3931,25 +3847,11 @@ name = "neon-shmem"
 version = "0.1.0"
 dependencies = [
  "nix 0.30.1",
- "rand 0.9.1",
- "rand_distr 0.5.1",
- "spin",
  "tempfile",
  "thiserror 1.0.69",
  "workspace_hack",
 ]
 
-[[package]]
-name = "neonart"
-version = "0.1.0"
-dependencies = [
- "crossbeam-utils",
- "rand 0.9.1",
- "rand_distr 0.5.1",
- "spin",
- "tracing",
-]
-
 [[package]]
 name = "never-say-never"
 version = "6.6.666"
@@ -4383,20 +4285,18 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-trait",
- "axum 0.8.1",
  "bytes",
  "camino",
  "clap",
  "futures",
  "hdrhistogram",
- "http 1.1.0",
  "humantime",
  "humantime-serde",
- "metrics",
  "pageserver_api",
  "pageserver_client",
  "pageserver_client_grpc",
  "pageserver_page_api",
+ "pprof",
  "rand 0.8.5",
  "reqwest",
  "serde",
@@ -4425,6 +4325,7 @@ dependencies = [
  "pageserver_api",
  "postgres_ffi",
  "remote_storage",
+ "serde",
  "serde_json",
  "svg_fmt",
  "thiserror 1.0.69",
@@ -4442,6 +4343,7 @@ dependencies = [
  "arc-swap",
  "async-compression",
  "async-stream",
+ "base64 0.22.1",
  "bincode",
  "bit_field",
  "byteorder",
@@ -4479,7 +4381,6 @@ dependencies = [
  "pageserver_client",
  "pageserver_compaction",
  "pageserver_page_api",
- "peekable",
  "pem",
  "pin-project-lite",
  "postgres-protocol",
@@ -4493,7 +4394,6 @@ dependencies = [
  "pprof",
  "pq_proto",
  "procfs",
- "prost 0.13.5",
  "rand 0.8.5",
  "range-set-blaze",
  "regex",
@@ -4602,29 +4502,19 @@ name = "pageserver_client_grpc"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "async-trait",
+ "arc-swap",
  "bytes",
- "chrono",
- "dashmap 5.5.0",
+ "compute_api",
  "futures",
- "http 1.1.0",
- "hyper 1.6.0",
- "hyper-util",
- "metrics",
  "pageserver_api",
  "pageserver_page_api",
- "priority-queue",
- "rand 0.8.5",
- "scopeguard",
- "thiserror 1.0.69",
  "tokio",
  "tokio-stream",
  "tokio-util",
  "tonic 0.13.1",
- "tower 0.4.13",
  "tracing",
  "utils",
- "uuid",
+ "workspace_hack",
 ]
 
 [[package]]
@@ -4798,15 +4688,6 @@ dependencies = [
  "sha2",
 ]
 
-[[package]]
-name = "peekable"
-version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "225f9651e475709164f871dc2f5724956be59cb9edb055372ffeeab01ec2d20b"
-dependencies = [
- "smallvec",
-]
-
 [[package]]
 name = "pem"
 version = "3.0.3"
@@ -5240,17 +5121,6 @@ dependencies = [
  "elliptic-curve 0.13.8",
 ]
 
-[[package]]
-name = "priority-queue"
-version = "2.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef08705fa1589a1a59aa924ad77d14722cb0cd97b67dd5004ed5f4a4873fce8d"
-dependencies = [
- "autocfg",
- "equivalent",
- "indexmap 2.9.0",
-]
-
 [[package]]
 name = "proc-macro2"
 version = "1.0.94"
@@ -5329,7 +5199,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "22505a5c94da8e3b7c2996394d1c933236c4d743e81a410bcca4e6989fc066a4"
 dependencies = [
  "bytes",
- "heck 0.5.0",
+ "heck",
  "itertools 0.12.1",
  "log",
  "multimap",
@@ -5350,7 +5220,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0c1318b19085f08681016926435853bbf7858f9c082d0999b80550ff5d9abe15"
 dependencies = [
  "bytes",
- "heck 0.5.0",
+ "heck",
  "itertools 0.12.1",
  "log",
  "multimap",
@@ -5420,6 +5290,7 @@ dependencies = [
  "async-trait",
  "atomic-take",
  "aws-config",
+ "aws-credential-types",
  "aws-sdk-iam",
  "aws-sigv4",
  "base64 0.22.1",
@@ -5432,6 +5303,7 @@ dependencies = [
  "clashmap",
  "compute_api",
  "consumption_metrics",
+ "criterion",
  "ecdsa 0.16.9",
  "ed25519-dalek",
  "env_logger",
@@ -5451,7 +5323,7 @@ dependencies = [
  "humantime",
  "humantime-serde",
  "hyper 0.14.30",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "indexmap 2.9.0",
  "ipnet",
@@ -5459,6 +5331,7 @@ dependencies = [
  "itoa",
  "jose-jwa",
  "jose-jwk",
+ "json",
  "lasso",
  "measured",
  "metrics",
@@ -5475,7 +5348,7 @@ dependencies = [
  "postgres_backend",
  "pq_proto",
  "rand 0.8.5",
- "rand_distr 0.4.3",
+ "rand_distr",
  "rcgen",
  "redis",
  "regex",
@@ -5579,12 +5452,6 @@ dependencies = [
  "proc-macro2",
 ]
 
-[[package]]
-name = "r-efi"
-version = "5.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "74765f6d916ee2faa39bc8e68e4f3ed8949b48cccdac59983d287a7cb71ce9c5"
-
 [[package]]
 name = "rand"
 version = "0.7.3"
@@ -5609,16 +5476,6 @@ dependencies = [
  "rand_core 0.6.4",
 ]
 
-[[package]]
-name = "rand"
-version = "0.9.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9fbfd9d094a40bf3ae768db9361049ace4c0e04a4fd6b359518bd7b73a73dd97"
-dependencies = [
- "rand_chacha 0.9.0",
- "rand_core 0.9.3",
-]
-
 [[package]]
 name = "rand_chacha"
 version = "0.2.2"
@@ -5639,16 +5496,6 @@ dependencies = [
  "rand_core 0.6.4",
 ]
 
-[[package]]
-name = "rand_chacha"
-version = "0.9.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
-dependencies = [
- "ppv-lite86",
- "rand_core 0.9.3",
-]
-
 [[package]]
 name = "rand_core"
 version = "0.5.1"
@@ -5667,15 +5514,6 @@ dependencies = [
  "getrandom 0.2.11",
 ]
 
-[[package]]
-name = "rand_core"
-version = "0.9.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "99d9a13982dcf210057a8a78572b2217b667c3beacbf3a0d8b454f6f82837d38"
-dependencies = [
- "getrandom 0.3.2",
-]
-
 [[package]]
 name = "rand_distr"
 version = "0.4.3"
@@ -5686,16 +5524,6 @@ dependencies = [
  "rand 0.8.5",
 ]
 
-[[package]]
-name = "rand_distr"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a8615d50dcf34fa31f7ab52692afec947c4dd0ab803cc87cb3b0b4570ff7463"
-dependencies = [
- "num-traits",
- "rand 0.9.1",
-]
-
 [[package]]
 name = "rand_hc"
 version = "0.2.0"
@@ -5884,6 +5712,8 @@ dependencies = [
  "azure_identity",
  "azure_storage",
  "azure_storage_blobs",
+ "base64 0.22.1",
+ "byteorder",
  "bytes",
  "camino",
  "camino-tempfile",
@@ -5892,7 +5722,7 @@ dependencies = [
  "http-body-util",
  "http-types",
  "humantime-serde",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "itertools 0.10.5",
  "metrics",
  "once_cell",
@@ -5932,7 +5762,7 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-rustls 0.26.0",
  "hyper-util",
  "ipnet",
@@ -5989,7 +5819,7 @@ dependencies = [
  "futures",
  "getrandom 0.2.11",
  "http 1.1.0",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "parking_lot 0.11.2",
  "reqwest",
  "reqwest-middleware",
@@ -6010,7 +5840,7 @@ dependencies = [
  "async-trait",
  "getrandom 0.2.11",
  "http 1.1.0",
- "matchit 0.8.4",
+ "matchit",
  "opentelemetry",
  "reqwest",
  "reqwest-middleware",
@@ -6382,6 +6212,7 @@ dependencies = [
  "postgres-protocol",
  "postgres_backend",
  "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
  "pprof",
  "pq_proto",
@@ -6426,7 +6257,7 @@ dependencies = [
  "anyhow",
  "const_format",
  "pageserver_api",
- "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
  "pq_proto",
  "serde",
@@ -6958,12 +6789,12 @@ dependencies = [
 
 [[package]]
 name = "socket2"
-version = "0.5.9"
+version = "0.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4f5fd57c80058a56cf5c777ab8a126398ece8e442983605d280a44ce79d0edef"
+checksum = "7b5fac59a5cb5dd637972e5fca70daf0523c9067fcdc4842f053dae04a18f8e9"
 dependencies = [
  "libc",
- "windows-sys 0.52.0",
+ "windows-sys 0.48.0",
 ]
 
 [[package]]
@@ -6971,9 +6802,6 @@ name = "spin"
 version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6980e8d7511241f8acf4aebddbb1ff938df5eebe98691418c4468d0b72a96a67"
-dependencies = [
- "lock_api",
-]
 
 [[package]]
 name = "spinning_top"
@@ -7032,7 +6860,7 @@ dependencies = [
  "http-body-util",
  "http-utils",
  "humantime",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "metrics",
  "once_cell",
@@ -7168,6 +6996,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "reqwest",
+ "safekeeper_api",
  "serde_json",
  "storage_controller_client",
  "tokio",
@@ -7216,7 +7045,7 @@ version = "0.26.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4c6bee85a5a24955dc440386795aa378cd9cf82acd5f764469152d2270e581be"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "rustversion",
@@ -7641,16 +7470,6 @@ dependencies = [
  "syn 2.0.100",
 ]
 
-[[package]]
-name = "tokio-pipe"
-version = "0.2.12"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f213a84bffbd61b8fa0ba8a044b4bbe35d471d0b518867181e82bd5c15542784"
-dependencies = [
- "libc",
- "tokio",
-]
-
 [[package]]
 name = "tokio-postgres"
 version = "0.7.10"
@@ -7747,6 +7566,7 @@ dependencies = [
  "futures-core",
  "pin-project-lite",
  "tokio",
+ "tokio-util",
 ]
 
 [[package]]
@@ -7845,25 +7665,16 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
 dependencies = [
- "async-stream",
  "async-trait",
- "axum 0.7.9",
  "base64 0.22.1",
  "bytes",
- "h2 0.4.4",
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.6.0",
- "hyper-timeout",
- "hyper-util",
  "percent-encoding",
  "pin-project",
  "prost 0.13.5",
- "socket2",
- "tokio",
  "tokio-stream",
- "tower 0.4.13",
  "tower-layer",
  "tower-service",
  "tracing",
@@ -7876,7 +7687,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7e581ba15a835f4d9ea06c55ab1bd4dce26fc53752c69a04aac00703bfb49ba9"
 dependencies = [
  "async-trait",
- "axum 0.8.1",
+ "axum",
  "base64 0.22.1",
  "bytes",
  "flate2",
@@ -7884,7 +7695,7 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-timeout",
  "hyper-util",
  "percent-encoding",
@@ -7937,16 +7748,11 @@ checksum = "b8fa9be0de6cf49e536ce1851f987bd21a43b771b09473c3549a6c853db37c1c"
 dependencies = [
  "futures-core",
  "futures-util",
- "indexmap 1.9.3",
  "pin-project",
  "pin-project-lite",
- "rand 0.8.5",
- "slab",
  "tokio",
- "tokio-util",
  "tower-layer",
  "tower-service",
- "tracing",
 ]
 
 [[package]]
@@ -8430,7 +8236,7 @@ name = "vm_monitor"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "cgroups-rs",
  "clap",
  "futures",
@@ -8542,15 +8348,6 @@ version = "0.11.0+wasi-snapshot-preview1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423"
 
-[[package]]
-name = "wasi"
-version = "0.14.2+wasi-0.2.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9683f9a5a998d873c0d21fcbe3c083009670149a8fab228644b8bd36b2c48cb3"
-dependencies = [
- "wit-bindgen-rt",
-]
-
 [[package]]
 name = "wasite"
 version = "0.1.0"
@@ -8908,15 +8705,6 @@ dependencies = [
  "windows-sys 0.48.0",
 ]
 
-[[package]]
-name = "wit-bindgen-rt"
-version = "0.39.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f42320e61fe2cfd34354ecb597f86f413484a798ba44a8ca1165c58d42da6c1"
-dependencies = [
- "bitflags 2.8.0",
-]
-
 [[package]]
 name = "workspace_hack"
 version = "0.1.0"
@@ -8924,8 +8712,8 @@ dependencies = [
  "ahash",
  "anstream",
  "anyhow",
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "base64 0.21.7",
  "base64ct",
  "bytes",
@@ -8947,8 +8735,10 @@ dependencies = [
  "fail",
  "form_urlencoded",
  "futures-channel",
+ "futures-core",
  "futures-executor",
  "futures-io",
+ "futures-sink",
  "futures-util",
  "generic-array",
  "getrandom 0.2.11",
@@ -8957,7 +8747,7 @@ dependencies = [
  "hex",
  "hmac",
  "hyper 0.14.30",
- "hyper 1.6.0",
+ "hyper 1.4.1",
  "hyper-util",
  "indexmap 2.9.0",
  "itertools 0.12.1",
@@ -8975,6 +8765,7 @@ dependencies = [
  "num-iter",
  "num-rational",
  "num-traits",
+ "once_cell",
  "p256 0.13.2",
  "parquet",
  "prettyplease",

Cargo.toml

@@ -35,7 +35,6 @@ members = [
     "libs/pq_proto",
     "libs/tenant_size_model",
     "libs/metrics",
-    "libs/neonart",
     "libs/postgres_connection",
     "libs/remote_storage",
     "libs/tracing-utils",
@@ -44,6 +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",
@@ -92,7 +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"
@@ -151,7 +150,6 @@ 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"
@@ -188,7 +186,6 @@ 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"
@@ -200,10 +197,11 @@ 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"]}
-tokio-stream = "0.1"
+tokio-stream = { version = "0.1", features = ["sync"] }
 tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
@@ -241,9 +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" }
@@ -263,7 +258,6 @@ 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/" }
@@ -295,7 +289,7 @@ 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
 
@@ -38,17 +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
 
@@ -56,23 +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 \
@@ -87,10 +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
 
@@ -130,12 +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

@@ -109,7 +109,7 @@ 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-install walproposer-lib cargo-target-dir
 	+@echo "Compiling Neon"
@@ -122,7 +122,7 @@ cargo-target-dir:
 	test -e target/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > target/CACHEDIR.TAG
 
 .PHONY: neon-pg-ext-%
-neon-pg-ext-%: postgres-install-%
+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)' \
@@ -220,6 +220,15 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 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

build-tools/Dockerfile

@@ -35,7 +35,7 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
     echo -e "retry_connrefused=on\ntimeout=15\ntries=5\nretry-on-host-error=on\n" > /root/.wgetrc && \
     echo -e "--retry-connrefused\n--connect-timeout 15\n--retry 5\n--max-time 300\n" > /root/.curlrc
 
-COPY build_tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
+COPY build-tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
 
 RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         set -e && \

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/compute-node.Dockerfile

@@ -9,7 +9,7 @@
 #
 # build-tools:   This contains Rust compiler toolchain and other tools needed at compile
 #                time. This is also used for the storage builds. This image is defined in
-#                build-tools.Dockerfile.
+#                build-tools/Dockerfile.
 #
 # build-deps:    Contains C compiler, other build tools, and compile-time dependencies
 #                needed to compile PostgreSQL and most extensions. (Some extensions need
@@ -115,7 +115,7 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION
 
-# Keep in sync with build-tools.Dockerfile
+# Keep in sync with build-tools/Dockerfile
 ENV PROTOC_VERSION=25.1
 
 # Use strict mode for bash to catch errors early
@@ -1572,6 +1572,7 @@ 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 "${PG_VERSION}" in \
     "v14") \
     export PGAUDIT_VERSION=1.6.3 \
@@ -1594,7 +1595,8 @@ RUN case "${PG_VERSION}" in \
     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/
@@ -1634,11 +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"
 
 #########################################################################################
 #
@@ -1785,7 +1790,7 @@ RUN set -e \
 #########################################################################################
 FROM build-deps AS exporters
 ARG TARGETARCH
-# Keep sql_exporter version same as in build-tools.Dockerfile and
+# Keep sql_exporter version same as in build-tools/Dockerfile and
 # test_runner/regress/test_compute_metrics.py
 # See comment on the top of the file regading `echo`, `-e` and `\n`
 RUN if [ "$TARGETARCH" = "amd64" ]; then\
@@ -1910,10 +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
@@ -1983,7 +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_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 +66,7 @@ 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

compute_tools/README.md

@@ -46,11 +46,14 @@ stateDiagram-v2
   Configuration --> Failed : Failed to configure the compute
   Configuration --> Running : Compute has been configured
   Empty --> Init : Compute spec is immediately available
-  Empty --> TerminationPending : Requested termination
+  Empty --> TerminationPendingFast : Requested termination
+  Empty --> TerminationPendingImmediate : Requested termination
   Init --> Failed : Failed to start Postgres
   Init --> Running : Started Postgres
-  Running --> TerminationPending : Requested termination
-  TerminationPending --> Terminated : Terminated compute
+  Running --> TerminationPendingFast : Requested termination
+  Running --> TerminationPendingImmediate : Requested termination
+  TerminationPendingFast --> Terminated compute with 30s delay for cplane to inspect status
+  TerminationPendingImmediate --> Terminated : Terminated compute immediately
   Failed --> [*] : Compute exited
   Terminated --> [*] : Compute exited
 ```

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/compute.rs

@@ -1,13 +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, PageserverConnectionInfo,
-    PageserverShardConnectionInfo, PgIdent,
+    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
 };
 use futures::StreamExt;
 use futures::future::join_all;
@@ -26,11 +25,11 @@ 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::{spawn, sync::watch, task::JoinHandle, time};
 use tracing::{Instrument, debug, error, info, instrument, warn};
 use url::Url;
 use utils::id::{TenantId, TimelineId};
@@ -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,
         }
     }
 
@@ -217,7 +225,7 @@ pub struct ParsedSpec {
     pub spec: ComputeSpec,
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
-    pub pageserver_conninfo: PageserverConnectionInfo,
+    pub pageserver_connstr: String,
     pub safekeeper_connstrings: Vec<String>,
     pub storage_auth_token: Option<String>,
     /// k8s dns name and port
@@ -264,27 +272,6 @@ impl ParsedSpec {
     }
 }
 
-fn extract_pageserver_conninfo_from_guc(
-    pageserver_connstring_guc: &str,
-) -> PageserverConnectionInfo {
-    PageserverConnectionInfo {
-        shards: pageserver_connstring_guc
-            .split(',')
-            .enumerate()
-            .map(|(i, connstr)| {
-                (
-                    i as u32,
-                    PageserverShardConnectionInfo {
-                        libpq_url: Some(connstr.to_string()),
-                        grpc_url: None,
-                    },
-                )
-            })
-            .collect(),
-        prefer_grpc: false,
-    }
-}
-
 impl TryFrom<ComputeSpec> for ParsedSpec {
     type Error = String;
     fn try_from(spec: ComputeSpec) -> Result<Self, String> {
@@ -294,17 +281,11 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
         // For backwards-compatibility, the top-level fields in the spec file
         // may be empty. In that case, we need to dig them from the GUCs in the
         // cluster.settings field.
-        let pageserver_conninfo = match &spec.pageserver_connection_info {
-            Some(x) => x.clone(),
-            None => {
-                if let Some(guc) = spec.cluster.settings.find("neon.pageserver_connstring") {
-                    extract_pageserver_conninfo_from_guc(&guc)
-                } else {
-                    return Err("pageserver connstr should be provided".to_string());
-                }
-            }
-        };
-
+        let pageserver_connstr = spec
+            .pageserver_connstring
+            .clone()
+            .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
+            .ok_or("pageserver connstr should be provided")?;
         let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
             if matches!(spec.mode, ComputeMode::Primary) {
                 spec.cluster
@@ -354,7 +335,7 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
 
         let res = ParsedSpec {
             spec,
-            pageserver_conninfo,
+            pageserver_connstr,
             safekeeper_connstrings,
             storage_auth_token,
             tenant_id,
@@ -392,7 +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 {
@@ -406,7 +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 {
@@ -444,7 +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);
         }
 
@@ -456,6 +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),
         })
     }
 
@@ -543,6 +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.
@@ -779,10 +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");
@@ -798,7 +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
@@ -865,12 +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(())
     }
 
@@ -961,14 +956,20 @@ impl ComputeNode {
             None
         };
 
-        let mut delay_exit = false;
         let mut state = self.state.lock().unwrap();
         state.terminate_flush_lsn = lsn;
-        if let ComputeStatus::TerminationPending { mode } = state.status {
+
+        let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
+        if state.status == ComputeStatus::TerminationPendingFast
+            || state.status == ComputeStatus::TerminationPendingImmediate
+        {
+            info!(
+                "Changing compute status from {} to {}",
+                state.status,
+                ComputeStatus::Terminated
+            );
             state.status = ComputeStatus::Terminated;
             self.state_changed.notify_all();
-            // we were asked to terminate gracefully, don't exit to avoid restart
-            delay_exit = mode == compute_api::responses::TerminateMode::Fast
         }
         drop(state);
 
@@ -1031,13 +1032,16 @@ impl ComputeNode {
     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();
         let started = Instant::now();
-        let (connected, size) = if spec.pageserver_conninfo.prefer_grpc {
-            self.try_get_basebackup_grpc(spec, lsn)?
-        } else {
-            self.try_get_basebackup_libpq(spec, lsn)?
+
+        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)?,
         };
 
+        self.fix_zenith_signal_neon_signal()?;
+
         let mut state = self.state.lock().unwrap();
         state.metrics.pageserver_connect_micros =
             connected.duration_since(started).as_micros() as u64;
@@ -1047,24 +1051,44 @@ impl ComputeNode {
         Ok(())
     }
 
+    /// Move the Zenith signal file to Neon signal file location.
+    /// This makes Compute compatible with older PageServers that don't yet
+    /// know about the Zenith->Neon rename.
+    fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
+        let datadir = Path::new(&self.params.pgdata);
+
+        let neonsig = datadir.join("neon.signal");
+
+        if neonsig.is_file() {
+            return Ok(());
+        }
+
+        let zenithsig = datadir.join("zenith.signal");
+
+        if zenithsig.is_file() {
+            fs::copy(zenithsig, neonsig)?;
+        }
+
+        Ok(())
+    }
+
     /// 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 = spec
-            .pageserver_conninfo
-            .shards
-            .get(&0)
-            .expect("shard 0 connection info missing");
-        let shard0_url = shard0.grpc_url.clone().expect("no grpc_url for shard 0");
-
-        let shard_index = match spec.pageserver_conninfo.shards.len() as u8 {
+        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 (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
             let mut client = page_api::Client::connect(
-                shard0_url,
+                shard0_connstr,
                 spec.tenant_id,
                 spec.timeline_id,
                 shard_index,
@@ -1099,13 +1123,8 @@ impl ComputeNode {
     /// 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 = spec
-            .pageserver_conninfo
-            .shards
-            .get(&0)
-            .expect("shard 0 connection info missing");
-        let shard0_connstr = shard0.libpq_url.clone().expect("no libpq_url for shard 0");
-        let mut config = postgres::Config::from_str(&shard0_connstr)?;
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let mut config = postgres::Config::from_str(shard0_connstr)?;
 
         // Use the storage auth token from the config file, if given.
         // Note: this overrides any password set in the connection string.
@@ -1191,7 +1210,10 @@ 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;
                 }
@@ -1264,9 +1286,7 @@ impl ComputeNode {
 
         // In case of error, log and fail the check, but don't crash.
         // We're playing it safe because these errors could be transient
-        // and we don't yet retry. Also being careful here allows us to
-        // be backwards compatible with safekeepers that don't have the
-        // TIMELINE_STATUS API yet.
+        // and we don't yet retry.
         if responses.len() < quorum {
             error!(
                 "failed sync safekeepers check {:?} {:?} {:?}",
@@ -1400,8 +1420,16 @@ impl ComputeNode {
             }
         };
 
-        self.get_basebackup(compute_state, lsn)
-            .with_context(|| format!("failed to get basebackup@{lsn}"))?;
+        info!(
+            "getting basebackup@{} from pageserver {}",
+            lsn, &pspec.pageserver_connstr
+        );
+        self.get_basebackup(compute_state, lsn).with_context(|| {
+            format!(
+                "failed to get basebackup@{} from pageserver {}",
+                lsn, &pspec.pageserver_connstr
+            )
+        })?;
 
         // Update pg_hba.conf received with basebackup.
         update_pg_hba(pgdata_path)?;
@@ -1693,6 +1721,8 @@ 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.
@@ -1757,6 +1787,8 @@ 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");
 
@@ -1800,6 +1832,8 @@ impl ComputeNode {
             tls_config,
         )?;
 
+        self.pg_reload_conf()?;
+
         if !spec.skip_pg_catalog_updates {
             let max_concurrent_connections = spec.reconfigure_concurrency;
             // Temporarily reset max_cluster_size in config
@@ -1819,10 +1853,9 @@ impl ComputeNode {
 
                 Ok(())
             })?;
+            self.pg_reload_conf()?;
         }
 
-        self.pg_reload_conf()?;
-
         let unknown_op = "unknown".to_string();
         let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
         info!(
@@ -1895,7 +1928,8 @@ impl ComputeNode {
 
                             // exit loop
                             ComputeStatus::Failed
-                            | ComputeStatus::TerminationPending { .. }
+                            | ComputeStatus::TerminationPendingFast
+                            | ComputeStatus::TerminationPendingImmediate
                             | ComputeStatus::Terminated => break 'cert_update,
 
                             // wait
@@ -2061,7 +2095,7 @@ 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",
                 )))?;
 
@@ -2257,7 +2291,7 @@ 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();
@@ -2336,22 +2370,22 @@ LIMIT 100",
     /// The operation will time out after a specified duration.
     pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
         let state = self.state.lock().unwrap();
-        let old_pageserver_conninfo = state
+        let old_pageserver_connstr = state
             .pspec
             .as_ref()
             .expect("spec must be set")
-            .pageserver_conninfo
+            .pageserver_connstr
             .clone();
         let mut unchanged = true;
         let _ = self
             .state_changed
             .wait_timeout_while(state, duration, |s| {
-                let pageserver_conninfo = &s
+                let pageserver_connstr = &s
                     .pspec
                     .as_ref()
                     .expect("spec must be set")
-                    .pageserver_conninfo;
-                unchanged = pageserver_conninfo == &old_pageserver_conninfo;
+                    .pageserver_connstr;
+                unchanged = pageserver_connstr == &old_pageserver_connstr;
                 unchanged
             })
             .unwrap();
@@ -2361,24 +2395,84 @@ 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,
+            );
+        }
     }
 }
 
@@ -2391,7 +2485,7 @@ pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
                 serde_json::to_string(&extensions).expect("failed to serialize extensions list")
             );
         }
-        Err(err) => error!("could not get installed extensions: {err:?}"),
+        Err(err) => error!("could not get installed extensions: {err}"),
     }
     Ok(())
 }

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};
@@ -69,7 +70,7 @@ impl ComputeNode {
             }
         };
         let row = match client
-            .query_one("select * from get_prewarm_info()", &[])
+            .query_one("select * from neon.get_prewarm_info()", &[])
             .await
         {
             Ok(row) => row,
@@ -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(),
             };
@@ -146,38 +146,48 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select prewarm_local_cache($1)", &[&uncompressed])
+            .query_one("select neon.prewarm_local_cache($1)", &[&uncompressed])
             .await
             .context("loading LFC state into postgres")
             .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");
@@ -186,7 +196,7 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select get_local_cache_state()", &[])
+            .query_one("select neon.get_local_cache_state()", &[])
             .await
             .context("querying LFC state")?
             .try_get::<usize, &[u8]>(0)

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

@@ -56,51 +56,9 @@ pub fn write_postgres_conf(
 
     // Add options for connecting to storage
     writeln!(file, "# Neon storage settings")?;
-
-    if let Some(conninfo) = &spec.pageserver_connection_info {
-        let mut libpq_urls: Option<Vec<String>> = Some(Vec::new());
-        let mut grpc_urls: Option<Vec<String>> = Some(Vec::new());
-
-        for shardno in 0..conninfo.shards.len() {
-            let info = conninfo.shards.get(&(shardno as u32)).ok_or_else(|| {
-                anyhow::anyhow!("shard {shardno} missing from pageserver_connection_info shard map")
-            })?;
-
-            if let Some(url) = &info.libpq_url {
-                if let Some(ref mut urls) = libpq_urls {
-                    urls.push(url.clone());
-                }
-            } else {
-                libpq_urls = None
-            }
-            if let Some(url) = &info.grpc_url {
-                if let Some(ref mut urls) = grpc_urls {
-                    urls.push(url.clone());
-                }
-            } else {
-                grpc_urls = None
-            }
-        }
-        if let Some(libpq_urls) = libpq_urls {
-            writeln!(
-                file,
-                "neon.pageserver_connstring={}",
-                escape_conf_value(&libpq_urls.join(","))
-            )?;
-        } else {
-            writeln!(file, "# no neon.pageserver_connstring")?;
-        }
-        if let Some(grpc_urls) = grpc_urls {
-            writeln!(
-                file,
-                "neon.pageserver_grpc_urls={}",
-                escape_conf_value(&grpc_urls.join(","))
-            )?;
-        } else {
-            writeln!(file, "# no neon.pageserver_grpc_urls")?;
-        }
+    if let Some(s) = &spec.pageserver_connstring {
+        writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
     }
-
     if let Some(stripe_size) = spec.shard_stripe_size {
         writeln!(file, "neon.stripe_size={stripe_size}")?;
     }

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/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:
@@ -290,9 +371,28 @@ paths:
       summary: Terminate Postgres and wait for it to exit
       description: ""
       operationId: terminate
+      parameters:
+        - name: mode
+          in: query
+          description: "Terminate mode: fast (wait 30s before returning) and immediate"
+          required: false
+          schema:
+            type: string
+            enum: ["fast", "immediate"]
+            default: fast
       responses:
         200:
           description: Result
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
+        201:
+          description: Result if compute is already terminated
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
         412:
           description: "wrong state"
           content:
@@ -335,15 +435,6 @@ 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
@@ -458,11 +549,14 @@ components:
       type: string
       enum:
         - empty
-        - init
-        - failed
-        - running
         - configuration_pending
+        - init
+        - running
         - configuration
+        - failed
+        - termination_pending_fast
+        - termination_pending_immediate
+        - terminated
       example: running
 
     ExtensionInstallRequest:
@@ -497,25 +591,69 @@ 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
@@ -544,6 +682,17 @@ components:
           description: Role name.
           example: "neon"
 
+    TerminateResponse:
+      type: object
+      required:
+        - lsn
+      properties:
+        lsn:
+          type: string
+          nullable: true
+          description: "last WAL flush LSN"
+          example: "0/028F10D8"
+
     SetRoleGrantsResponse:
       type: object
       required:

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/routes/terminate.rs

@@ -3,7 +3,7 @@ use crate::http::JsonResponse;
 use axum::extract::State;
 use axum::response::Response;
 use axum_extra::extract::OptionalQuery;
-use compute_api::responses::{ComputeStatus, TerminateResponse};
+use compute_api::responses::{ComputeStatus, TerminateMode, TerminateResponse};
 use http::StatusCode;
 use serde::Deserialize;
 use std::sync::Arc;
@@ -12,7 +12,7 @@ use tracing::info;
 
 #[derive(Deserialize, Default)]
 pub struct TerminateQuery {
-    mode: compute_api::responses::TerminateMode,
+    mode: TerminateMode,
 }
 
 /// Terminate the compute.
@@ -24,16 +24,16 @@ pub(in crate::http) async fn terminate(
     {
         let mut state = compute.state.lock().unwrap();
         if state.status == ComputeStatus::Terminated {
-            return JsonResponse::success(StatusCode::CREATED, state.terminate_flush_lsn);
+            let response = TerminateResponse {
+                lsn: state.terminate_flush_lsn,
+            };
+            return JsonResponse::success(StatusCode::CREATED, response);
         }
 
         if !matches!(state.status, ComputeStatus::Empty | ComputeStatus::Running) {
             return JsonResponse::invalid_status(state.status);
         }
-        state.set_status(
-            ComputeStatus::TerminationPending { mode },
-            &compute.state_changed,
-        );
+        state.set_status(mode.into(), &compute.state_changed);
     }
 
     forward_termination_signal(false);

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

@@ -2,6 +2,7 @@ use std::collections::HashMap;
 
 use anyhow::Result;
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use tokio_postgres::error::Error as PostgresError;
 use tokio_postgres::{Client, Config, NoTls};
 
 use crate::metrics::INSTALLED_EXTENSIONS;
@@ -10,7 +11,7 @@ use crate::metrics::INSTALLED_EXTENSIONS;
 /// and to make database listing query here more explicit.
 ///
 /// Limit the number of databases to 500 to avoid excessive load.
-async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
+async fn list_dbs(client: &mut Client) -> Result<Vec<String>, PostgresError> {
     // `pg_database.datconnlimit = -2` means that the database is in the
     // invalid state
     let databases = client
@@ -37,7 +38,9 @@ async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
 /// Same extension can be installed in multiple databases with different versions,
 /// so we report a separate metric (number of databases where it is installed)
 /// for each extension version.
-pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExtensions> {
+pub async fn get_installed_extensions(
+    mut conf: Config,
+) -> Result<InstalledExtensions, PostgresError> {
     conf.application_name("compute_ctl:get_installed_extensions");
     let databases: Vec<String> = {
         let (mut client, connection) = conf.connect(NoTls).await?;

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 +4,8 @@ use std::thread;
 use std::time::{Duration, SystemTime};
 
 use anyhow::{Result, bail};
-use compute_api::spec::{ComputeMode, PageserverConnectionInfo};
+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};
@@ -77,16 +78,17 @@ fn acquire_lsn_lease_with_retry(
 
     loop {
         // Note: List of pageservers is dynamic, need to re-read configs before each attempt.
-        let (conninfo, auth) = {
+        let (connstrings, auth) = {
             let state = compute.state.lock().unwrap();
             let spec = state.pspec.as_ref().expect("spec must be set");
             (
-                spec.pageserver_conninfo.clone(),
+                spec.pageserver_connstr.clone(),
                 spec.storage_auth_token.clone(),
             )
         };
 
-        let result = try_acquire_lsn_lease(conninfo, auth.as_deref(), tenant_id, timeline_id, lsn);
+        let result =
+            try_acquire_lsn_lease(&connstrings, auth.as_deref(), tenant_id, timeline_id, lsn);
         match result {
             Ok(Some(res)) => {
                 return Ok(res);
@@ -110,16 +112,17 @@ fn acquire_lsn_lease_with_retry(
 
 /// Tries to acquire LSN leases on all Pageserver shards.
 fn try_acquire_lsn_lease(
-    conninfo: PageserverConnectionInfo,
+    connstrings: &str,
     auth: Option<&str>,
     tenant_id: TenantId,
     timeline_id: TimelineId,
     lsn: Lsn,
 ) -> Result<Option<SystemTime>> {
-    let shard_count = conninfo.shards.len();
+    let connstrings = connstrings.split(',').collect_vec();
+    let shard_count = connstrings.len();
     let mut leases = Vec::new();
 
-    for (shard_number, shard) in conninfo.shards.into_iter() {
+    for (shard_number, &connstring) in connstrings.iter().enumerate() {
         let tenant_shard_id = match shard_count {
             0 | 1 => TenantShardId::unsharded(tenant_id),
             shard_count => TenantShardId {
@@ -129,22 +132,13 @@ fn try_acquire_lsn_lease(
             },
         };
 
-        let lease = if conninfo.prefer_grpc {
-            acquire_lsn_lease_grpc(
-                &shard.grpc_url.unwrap(),
-                auth,
-                tenant_shard_id,
-                timeline_id,
-                lsn,
-            )?
-        } else {
-            acquire_lsn_lease_libpq(
-                &shard.libpq_url.unwrap(),
-                auth,
-                tenant_shard_id,
-                timeline_id,
-                lsn,
-            )?
+        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);
     }

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/0002-alter_roles.sql

@@ -1,3 +1,16 @@
+-- On December 8th, 2023, an engineering escalation (INC-110) was opened after
+-- it was found that BYPASSRLS was being applied to all roles.
+--
+-- PR that introduced the issue: https://github.com/neondatabase/neon/pull/5657
+-- Subsequent commit on main: https://github.com/neondatabase/neon/commit/ad99fa5f0393e2679e5323df653c508ffa0ac072
+--
+-- NOBYPASSRLS and INHERIT are the defaults for a Postgres role, but because it
+-- isn't easy to know if a Postgres cluster is affected by the issue, we need to
+-- keep the migration around for a long time, if not indefinitely, so any
+-- cluster can be fixed.
+--
+-- Branching is the gift that keeps on giving...
+
 DO $$
 DECLARE
     role_name text;

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/monitor.rs

@@ -84,7 +84,8 @@ impl ComputeMonitor {
         if matches!(
             compute_status,
             ComputeStatus::Terminated
-                | ComputeStatus::TerminationPending { .. }
+                | ComputeStatus::TerminationPendingFast
+                | ComputeStatus::TerminationPendingImmediate
                 | ComputeStatus::Failed
         ) {
             info!(

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

@@ -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/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

@@ -16,7 +16,7 @@ use std::time::Duration;
 use anyhow::{Context, Result, anyhow, bail};
 use clap::Parser;
 use compute_api::requests::ComputeClaimsScope;
-use compute_api::spec::{ComputeMode, PageserverConnectionInfo, PageserverShardConnectionInfo};
+use compute_api::spec::{ComputeMode, PageserverProtocol};
 use control_plane::broker::StorageBroker;
 use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
@@ -64,7 +64,9 @@ const DEFAULT_PAGESERVER_ID: NodeId = NodeId(1);
 const DEFAULT_BRANCH_NAME: &str = "main";
 project_git_version!(GIT_VERSION);
 
+#[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/";
 
@@ -167,7 +169,7 @@ 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: PgMajorVersion,
 
@@ -290,7 +292,7 @@ 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: PgMajorVersion,
 }
@@ -322,7 +324,7 @@ 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: PgMajorVersion,
 }
@@ -601,7 +603,7 @@ 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: PgMajorVersion,
 
@@ -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",
@@ -1504,35 +1516,29 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 )?;
             }
 
-            let (shards, stripe_size) = if let Some(ps_id) = pageserver_id {
-                let conf = env.get_pageserver_conf(ps_id).unwrap();
-                let libpq_url = Some({
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    format!("postgres://no_user@{host}:{port}")
-                });
-                let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
+            let (pageservers, stripe_size) = if let Some(pageserver_id) = pageserver_id {
+                let conf = env.get_pageserver_conf(pageserver_id).unwrap();
+                // Use gRPC if requested.
+                let pageserver = if endpoint.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
                     let (host, port) = parse_host_port(grpc_addr)?;
                     let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    Some(format!("grpc://no_user@{host}:{port}"))
+                    (PageserverProtocol::Grpc, host, port)
                 } else {
-                    None
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
+                    let port = port.unwrap_or(5432);
+                    (PageserverProtocol::Libpq, host, port)
                 };
-                let pageserver = PageserverShardConnectionInfo {
-                    libpq_url,
-                    grpc_url,
-                };
-
                 // If caller is telling us what pageserver to use, this is not a tenant which is
                 // fully managed by storage controller, therefore not sharded.
-                (vec![(0, pageserver)], DEFAULT_STRIPE_SIZE)
+                (vec![pageserver], DEFAULT_STRIPE_SIZE)
             } else {
                 // Look up the currently attached location of the tenant, and its striping metadata,
                 // to pass these on to postgres.
                 let storage_controller = StorageController::from_env(env);
                 let locate_result = storage_controller.tenant_locate(endpoint.tenant_id).await?;
-                let shards = futures::future::try_join_all(locate_result.shards.into_iter().map(
-                    |shard| async move {
+                let pageservers = futures::future::try_join_all(
+                    locate_result.shards.into_iter().map(|shard| async move {
                         if let ComputeMode::Static(lsn) = endpoint.mode {
                             // Initialize LSN leases for static computes.
                             let conf = env.get_pageserver_conf(shard.node_id).unwrap();
@@ -1544,34 +1550,28 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                                 .await?;
                         }
 
-                        let libpq_host = Host::parse(&shard.listen_pg_addr)?;
-                        let libpq_port = shard.listen_pg_port;
-                        let libpq_url =
-                            Some(format!("postgres://no_user@{libpq_host}:{libpq_port}"));
-
-                        let grpc_url = if let Some(grpc_host) = shard.listen_grpc_addr {
-                            let grpc_port = shard.listen_grpc_port.expect("no gRPC port");
-                            Some(format!("grpc://no_user@{grpc_host}:{grpc_port}"))
+                        let pageserver = if endpoint.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))?,
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
                         } else {
-                            None
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr)?,
+                                shard.listen_pg_port,
+                            )
                         };
-                        let pageserver = PageserverShardConnectionInfo {
-                            libpq_url,
-                            grpc_url,
-                        };
-                        anyhow::Ok((shard.shard_id.shard_number.0 as u32, pageserver))
-                    },
-                ))
+                        anyhow::Ok(pageserver)
+                    }),
+                )
                 .await?;
                 let stripe_size = locate_result.shard_params.stripe_size;
 
-                (shards, stripe_size)
-            };
-            assert!(!shards.is_empty());
-            let pageserver_conninfo = PageserverConnectionInfo {
-                shards: shards.into_iter().collect(),
-                prefer_grpc: endpoint.grpc,
+                (pageservers, stripe_size)
             };
+            assert!(!pageservers.is_empty());
 
             let ps_conf = env.get_pageserver_conf(DEFAULT_PAGESERVER_ID)?;
             let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
@@ -1595,22 +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,
-                    pageserver_conninfo,
-                    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;
@@ -1618,27 +1620,20 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 .endpoints
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let shards = if let Some(ps_id) = args.endpoint_pageserver_id {
+            let pageservers = if let Some(ps_id) = args.endpoint_pageserver_id {
                 let conf = env.get_pageserver_conf(ps_id)?;
-                let libpq_url = Some({
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    format!("postgres://no_user@{host}:{port}")
-                });
-                let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
+                // Use gRPC if requested.
+                let pageserver = if endpoint.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
                     let (host, port) = parse_host_port(grpc_addr)?;
                     let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    Some(format!("grpc://no_user@{host}:{port}"))
+                    (PageserverProtocol::Grpc, host, port)
                 } else {
-                    None
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
+                    let port = port.unwrap_or(5432);
+                    (PageserverProtocol::Libpq, host, port)
                 };
-                let pageserver = PageserverShardConnectionInfo {
-                    libpq_url,
-                    grpc_url,
-                };
-                // If caller is telling us what pageserver to use, this is not a tenant which is
-                // fully managed by storage controller, therefore not sharded.
-                vec![(0, pageserver)]
+                vec![pageserver]
             } else {
                 let storage_controller = StorageController::from_env(env);
                 storage_controller
@@ -1648,36 +1643,28 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                     .into_iter()
                     .map(|shard| {
                         // Use gRPC if requested.
-                        let libpq_host = Host::parse(&shard.listen_pg_addr).expect("bad hostname");
-                        let libpq_port = shard.listen_pg_port;
-                        let libpq_url =
-                            Some(format!("postgres://no_user@{libpq_host}:{libpq_port}"));
-
-                        let grpc_url = if let Some(grpc_host) = shard.listen_grpc_addr {
-                            let grpc_port = shard.listen_grpc_port.expect("no gRPC port");
-                            Some(format!("grpc://no_user@{grpc_host}:{grpc_port}"))
+                        if endpoint.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))
+                                    .expect("bad hostname"),
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
                         } else {
-                            None
-                        };
-                        (
-                            shard.shard_id.shard_number.0 as u32,
-                            PageserverShardConnectionInfo {
-                                libpq_url,
-                                grpc_url,
-                            },
-                        )
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr).expect("bad hostname"),
+                                shard.listen_pg_port,
+                            )
+                        }
                     })
                     .collect::<Vec<_>>()
             };
-            let pageserver_conninfo = PageserverConnectionInfo {
-                shards: shards.into_iter().collect(),
-                prefer_grpc: endpoint.grpc,
-            };
             // If --safekeepers argument is given, use only the listed
             // safekeeper nodes; otherwise all from the env.
             let safekeepers = parse_safekeepers(&args.safekeepers)?;
             endpoint
-                .reconfigure(Some(pageserver_conninfo), None, safekeepers, None)
+                .reconfigure(Some(pageservers), None, safekeepers, None)
                 .await?;
         }
         EndpointCmd::Stop(args) => {

control_plane/src/broker.rs

@@ -36,7 +36,7 @@ impl StorageBroker {
     pub async fn start(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         let broker = &self.env.broker;
 
-        print!("Starting neon broker at {}", broker.client_url());
+        println!("Starting neon broker at {}", broker.client_url());
 
         let mut args = Vec::new();
 

control_plane/src/endpoint.rs

@@ -32,7 +32,8 @@
 //!     config.json                 - passed to `compute_ctl`
 //!     pgdata/
 //!         postgresql.conf       - copy of postgresql.conf created by `compute_ctl`
-//!         zenith.signal
+//!         neon.signal
+//!         zenith.signal         - copy of neon.signal, for backward compatibility
 //!         <other PostgreSQL files>
 //! ```
 //!
@@ -56,13 +57,9 @@ 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,
 };
-
-// re-export these, because they're used in the reconfigure() function
-pub use compute_api::spec::{PageserverConnectionInfo, PageserverShardConnectionInfo};
-
 use jsonwebtoken::jwk::{
     AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
     OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
@@ -78,6 +75,7 @@ use sha2::{Digest, Sha256};
 use spki::der::Decode;
 use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
 use tracing::debug;
+use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
 
 use crate::local_env::LocalEnv;
@@ -376,6 +374,22 @@ impl std::fmt::Display for EndpointTerminateMode {
     }
 }
 
+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 {
     fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
         if !entry.file_type()?.is_dir() {
@@ -450,7 +464,7 @@ impl Endpoint {
         conf.append("max_connections", "100");
         conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
-        // It also defines how often the walreciever will send a feedback message to the wal sender.
+        // It also defines how often the walreceiver will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");
         conf.append("listen_addresses", &self.pg_address.ip().to_string());
         conf.append("port", &self.pg_address.port().to_string());
@@ -640,6 +654,14 @@ impl Endpoint {
         }
     }
 
+    fn build_pageserver_connstr(pageservers: &[(PageserverProtocol, Host, u16)]) -> String {
+        pageservers
+            .iter()
+            .map(|(scheme, host, port)| format!("{scheme}://no_user@{host}:{port}"))
+            .collect::<Vec<_>>()
+            .join(",")
+    }
+
     /// Map safekeepers ids to the actual connection strings.
     fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
         let mut safekeeper_connstrings = Vec::new();
@@ -672,21 +694,7 @@ 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>,
-        pageserver_conninfo: PageserverConnectionInfo,
-        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");
         }
@@ -699,7 +707,10 @@ impl Endpoint {
             std::fs::remove_dir_all(self.pgdata())?;
         }
 
-        let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
+        let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
+        assert!(!pageserver_connstring.is_empty());
+
+        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
@@ -727,7 +738,7 @@ 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,
@@ -736,7 +747,7 @@ 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(),
@@ -757,21 +768,23 @@ impl Endpoint {
                 branch_id: None,
                 endpoint_id: Some(self.endpoint_id.clone()),
                 mode: self.mode,
-                pageserver_connection_info: Some(pageserver_conninfo),
-                safekeepers_generation: safekeepers_generation.map(|g| g.into_inner()),
+                pageserver_connstring: Some(pageserver_connstring),
+                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
@@ -782,7 +795,7 @@ 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,
@@ -817,7 +830,7 @@ impl Endpoint {
         // Launch compute_ctl
         let conn_str = self.connstr("cloud_admin", "postgres");
         println!("Starting postgres node at '{conn_str}'");
-        if create_test_user {
+        if args.create_test_user {
             let conn_str = self.connstr("test", "neondb");
             println!("Also at '{conn_str}'");
         }
@@ -849,11 +862,11 @@ 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");
         }
 
@@ -885,10 +898,11 @@ 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
@@ -909,16 +923,18 @@ impl Endpoint {
                         ComputeStatus::Empty
                         | ComputeStatus::ConfigurationPending
                         | ComputeStatus::Configuration
-                        | ComputeStatus::TerminationPending { .. }
+                        | ComputeStatus::TerminationPendingFast
+                        | ComputeStatus::TerminationPendingImmediate
                         | ComputeStatus::Terminated => {
                             bail!("unexpected compute status: {:?}", state.status)
                         }
                     }
                 }
                 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 {start_timeout:?} waiting to connect to compute_ctl HTTP",
+                            "timed out {:?} waiting to connect to compute_ctl HTTP",
+                            args.start_timeout
                         ));
                     }
                 }
@@ -966,7 +982,7 @@ impl Endpoint {
 
     pub async fn reconfigure(
         &self,
-        pageserver_conninfo: Option<PageserverConnectionInfo>,
+        pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
         stripe_size: Option<ShardStripeSize>,
         safekeepers: Option<Vec<NodeId>>,
         safekeeper_generation: Option<SafekeeperGeneration>,
@@ -982,17 +998,15 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        if let Some(pageserver_conninfo) = pageserver_conninfo {
-            // If pageservers are provided, we need to ensure that they are not empty.
-            // This is a requirement for the compute_ctl configuration.
-            anyhow::ensure!(
-                !pageserver_conninfo.shards.is_empty(),
-                "no pageservers provided"
-            );
-            spec.pageserver_connection_info = Some(pageserver_conninfo);
-        }
-        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.
@@ -1041,7 +1055,7 @@ impl Endpoint {
 
     pub async fn reconfigure_pageservers(
         &self,
-        pageservers: PageserverConnectionInfo,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
         stripe_size: Option<ShardStripeSize>,
     ) -> Result<()> {
         self.reconfigure(Some(pageservers), stripe_size, None, None)

control_plane/src/local_env.rs

@@ -217,6 +217,9 @@ pub struct NeonStorageControllerConf {
     pub posthog_config: Option<PostHogConfig>,
 
     pub kick_secondary_downloads: Option<bool>,
+
+    #[serde(with = "humantime_serde")]
+    pub shard_split_request_timeout: Option<Duration>,
 }
 
 impl NeonStorageControllerConf {
@@ -250,6 +253,7 @@ impl Default for NeonStorageControllerConf {
             timeline_safekeeper_count: None,
             posthog_config: None,
             kick_secondary_downloads: None,
+            shard_split_request_timeout: None,
         }
     }
 }

control_plane/src/pageserver.rs

@@ -303,7 +303,7 @@ impl PageServerNode {
     async fn start_node(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
-        print!(
+        println!(
             "Starting pageserver node {} at '{}' in {:?}, retrying for {:?}",
             self.conf.id,
             self.pg_connection_config.raw_address(),
@@ -452,6 +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>())

control_plane/src/safekeeper.rs

@@ -127,7 +127,7 @@ impl SafekeeperNode {
         extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
-        print!(
+        println!(
             "Starting safekeeper at '{}' in '{}', retrying for {:?}",
             self.pg_connection_config.raw_address(),
             self.datadir_path().display(),

control_plane/src/storage_controller.rs

@@ -648,6 +648,13 @@ impl StorageController {
             args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
         }
 
+        if let Some(duration) = self.config.shard_split_request_timeout {
+            args.push(format!(
+                "--shard-split-request-timeout={}",
+                humantime::Duration::from(duration)
+            ));
+        }
+
         let mut envs = vec![
             ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
             ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
@@ -660,7 +667,7 @@ impl StorageController {
             ));
         }
 
-        println!("Starting storage controller");
+        println!("Starting storage controller at {scheme}://{host}:{listen_port}");
 
         background_process::start_process(
             COMMAND,

control_plane/storcon_cli/Cargo.toml

@@ -14,6 +14,7 @@ humantime.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 reqwest.workspace = true
+safekeeper_api.workspace=true
 serde_json = { workspace = true, features = ["raw_value"] }
 storage_controller_client.workspace = true
 tokio.workspace = true

control_plane/storcon_cli/src/main.rs

@@ -11,7 +11,7 @@ use pageserver_api::controller_api::{
     PlacementPolicy, SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest,
     ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
     SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TimelineSafekeeperMigrateRequest,
 };
 use pageserver_api::models::{
     EvictionPolicy, EvictionPolicyLayerAccessThreshold, ShardParameters, TenantConfig,
@@ -21,6 +21,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardStripeSize, TenantShardId};
 use pageserver_client::mgmt_api::{self};
 use reqwest::{Certificate, Method, StatusCode, Url};
+use safekeeper_api::models::TimelineLocateResponse;
 use storage_controller_client::control_api::Client;
 use utils::id::{NodeId, TenantId, TimelineId};
 
@@ -65,12 +66,27 @@ 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,
@@ -264,6 +280,23 @@ enum Command {
         #[arg(long)]
         concurrency: Option<usize>,
     },
+    /// Locate safekeepers for a timeline from the storcon DB.
+    TimelineLocate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+    },
+    /// Migrate a timeline to a new set of safekeepers
+    TimelineSafekeeperMigrate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+        /// Example: --new-sk-set 1,2,3
+        #[arg(long, required = true, value_delimiter = ',')]
+        new_sk_set: Vec<NodeId>,
+    },
 }
 
 #[derive(Parser)]
@@ -443,6 +476,7 @@ async fn main() -> anyhow::Result<()> {
                         listen_http_port,
                         listen_https_port,
                         availability_zone_id: AvailabilityZone(availability_zone_id),
+                        node_ip_addr: None,
                     }),
                 )
                 .await?;
@@ -912,10 +946,43 @@ 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::<(), ()>(
@@ -1276,7 +1343,7 @@ async fn main() -> anyhow::Result<()> {
             concurrency,
         } => {
             let mut path = format!(
-                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
+                "v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
             );
 
             if let Some(c) = concurrency {
@@ -1287,6 +1354,41 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::POST, path, None)
                 .await?;
         }
+        Command::TimelineLocate {
+            tenant_id,
+            timeline_id,
+        } => {
+            let path = format!("debug/v1/tenant/{tenant_id}/timeline/{timeline_id}/locate");
+
+            let resp = storcon_client
+                .dispatch::<(), TimelineLocateResponse>(Method::GET, path, None)
+                .await?;
+
+            let sk_set = resp.sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>();
+            let new_sk_set = resp
+                .new_sk_set
+                .as_ref()
+                .map(|ids| ids.iter().map(|id| id.0 as i64).collect::<Vec<_>>());
+
+            println!("generation = {}", resp.generation);
+            println!("sk_set = {sk_set:?}");
+            println!("new_sk_set = {new_sk_set:?}");
+        }
+        Command::TimelineSafekeeperMigrate {
+            tenant_id,
+            timeline_id,
+            new_sk_set,
+        } => {
+            let path = format!("v1/tenant/{tenant_id}/timeline/{timeline_id}/safekeeper_migrate");
+
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    path,
+                    Some(TimelineSafekeeperMigrateRequest { new_sk_set }),
+                )
+                .await?;
+        }
     }
 
     Ok(())

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/core_changes.md

@@ -129,9 +129,10 @@ segment to bootstrap the WAL writing, but it doesn't contain the checkpoint reco
 changes in xlog.c, to allow starting the compute node without reading the last checkpoint record
 from WAL.
 
-This includes code to read the `zenith.signal` file, which tells the startup code the LSN to start
-at. When the `zenith.signal` file is present, the startup uses that LSN instead of the last
-checkpoint's LSN. The system is known to be consistent at that LSN, without any WAL redo.
+This includes code to read the `neon.signal` (also `zenith.signal`) file, which tells the startup 
+code the LSN to start at. When the `neon.signal` file is present, the startup uses that LSN
+instead of the last checkpoint's LSN. The system is known to be consistent at that LSN, without 
+any WAL redo.
 
 
 ### How to get rid of the patch

docs/pageserver-services.md

@@ -75,7 +75,7 @@ CLI examples:
 * AWS S3  : `env AWS_ACCESS_KEY_ID='SOMEKEYAAAAASADSAH*#' AWS_SECRET_ACCESS_KEY='SOMEsEcReTsd292v' ${PAGESERVER_BIN} -c "remote_storage={bucket_name='some-sample-bucket',bucket_region='eu-north-1', prefix_in_bucket='/test_prefix/'}"`
 
 For Amazon AWS S3, a key id and secret access key could be located in `~/.aws/credentials` if awscli was ever configured to work with the desired bucket, on the AWS Settings page for a certain user. Also note, that the bucket names does not contain any protocols when used on AWS.
-For local S3 installations, refer to the their documentation for name format and credentials.
+For local S3 installations, refer to their documentation for name format and credentials.
 
 Similar to other pageserver settings, toml config file can be used to configure either of the storages as backup targets.
 Required sections are:

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/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}",

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/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")]
@@ -93,6 +121,15 @@ pub enum TerminateMode {
     Immediate,
 }
 
+impl From<TerminateMode> for ComputeStatus {
+    fn from(mode: TerminateMode) -> Self {
+        match mode {
+            TerminateMode::Fast => ComputeStatus::TerminationPendingFast,
+            TerminateMode::Immediate => ComputeStatus::TerminationPendingImmediate,
+        }
+    }
+}
+
 #[derive(Serialize, Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "snake_case")]
 pub enum ComputeStatus {
@@ -113,7 +150,9 @@ pub enum ComputeStatus {
     // control-plane to terminate it.
     Failed,
     // Termination requested
-    TerminationPending { mode: TerminateMode },
+    TerminationPendingFast,
+    // Termination requested, without waiting 30s before returning from /terminate
+    TerminationPendingImmediate,
     // Terminated Postgres
     Terminated,
 }
@@ -132,7 +171,10 @@ impl Display for ComputeStatus {
             ComputeStatus::Running => f.write_str("running"),
             ComputeStatus::Configuration => f.write_str("configuration"),
             ComputeStatus::Failed => f.write_str("failed"),
-            ComputeStatus::TerminationPending { .. } => f.write_str("termination-pending"),
+            ComputeStatus::TerminationPendingFast => f.write_str("termination-pending-fast"),
+            ComputeStatus::TerminationPendingImmediate => {
+                f.write_str("termination-pending-immediate")
+            }
             ComputeStatus::Terminated => f.write_str("terminated"),
         }
     }

libs/compute_api/src/spec.rs

@@ -105,11 +105,7 @@ pub struct ComputeSpec {
     // updated to fill these fields, we can make these non optional.
     pub tenant_id: Option<TenantId>,
     pub timeline_id: Option<TimelineId>,
-
-    // Pageserver information can be passed in two different ways:
-    // 1. Here
-    // 2. in cluster.settings. This is legacy, we are switching to method 1.
-    pub pageserver_connection_info: Option<PageserverConnectionInfo>,
+    pub pageserver_connstring: Option<String>,
 
     // More neon ids that we expose to the compute_ctl
     // and to postgres as neon extension GUCs.
@@ -185,9 +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.
@@ -209,20 +214,6 @@ pub enum ComputeFeature {
     UnknownFeature,
 }
 
-/// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
-#[derive(Clone, Debug, Default, Deserialize, Serialize, Eq, PartialEq)]
-pub struct PageserverConnectionInfo {
-    pub shards: HashMap<u32, PageserverShardConnectionInfo>,
-
-    pub prefer_grpc: bool,
-}
-
-#[derive(Clone, Debug, Default, Deserialize, Serialize, Eq, PartialEq)]
-pub struct PageserverShardConnectionInfo {
-    pub libpq_url: Option<String>,
-    pub grpc_url: Option<String>,
-}
-
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
 pub struct RemoteExtSpec {
     pub public_extensions: Option<Vec<String>>,
@@ -340,12 +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 {
@@ -457,7 +442,7 @@ pub struct JwksSettings {
 }
 
 /// Protocol used to connect to a Pageserver. Parsed from the connstring scheme.
-#[derive(Clone, Copy, Debug, Default)]
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub enum PageserverProtocol {
     /// The original protocol based on libpq and COPY. Uses postgresql:// or postgres:// scheme.
     #[default]

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/http-utils/src/endpoint.rs

@@ -20,6 +20,7 @@ use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::io::ReaderStream;
 use tracing::{Instrument, debug, 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};
@@ -250,9 +251,28 @@ 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 +297,18 @@ 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 +321,10 @@ 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 +333,7 @@ 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"
                 );
             }

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

@@ -6,13 +6,8 @@ license.workspace = true
 
 [dependencies]
 thiserror.workspace = true
-nix.workspace = true
-spin.workspace = true
+nix.workspace=true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
 
-[dev-dependencies]
-rand = "0.9.1"
-rand_distr = "0.5.1"
-
 [target.'cfg(target_os = "macos")'.dependencies]
 tempfile = "3.14.0"

libs/neon-shmem/src/hash.rs

@@ -1,367 +0,0 @@
-//! Hash table implementation on top of 'shmem'
-//!
-//! Features required in the long run by the communicator project:
-//!
-//! [X] Accessible from both Postgres processes and rust threads in the communicator process
-//! [X] Low latency
-//! [ ] Scalable to lots of concurrent accesses (currently uses a single spinlock)
-//! [ ] Resizable
-
-use std::fmt::Debug;
-use std::hash::Hash;
-use std::mem::MaybeUninit;
-use std::ops::Deref;
-
-use crate::shmem::ShmemHandle;
-
-use spin;
-
-mod core;
-
-#[cfg(test)]
-mod tests;
-
-use core::CoreHashMap;
-
-pub enum UpdateAction<V> {
-    Nothing,
-    Insert(V),
-    Remove,
-}
-
-#[derive(Debug)]
-pub struct OutOfMemoryError();
-
-pub struct HashMapInit<'a, K, V> {
-    // Hash table can be allocated in a fixed memory area, or in a resizeable ShmemHandle.
-    shmem_handle: Option<ShmemHandle>,
-    shared_ptr: *mut HashMapShared<'a, K, V>,
-}
-
-pub struct HashMapAccess<'a, K, V> {
-    shmem_handle: Option<ShmemHandle>,
-    shared_ptr: *mut HashMapShared<'a, K, V>,
-}
-
-unsafe impl<'a, K: Sync, V: Sync> Sync for HashMapAccess<'a, K, V> {}
-unsafe impl<'a, K: Send, V: Send> Send for HashMapAccess<'a, K, V> {}
-
-impl<'a, K, V> HashMapInit<'a, K, V> {
-    pub fn attach_writer(self) -> HashMapAccess<'a, K, V> {
-        HashMapAccess {
-            shmem_handle: self.shmem_handle,
-            shared_ptr: self.shared_ptr,
-        }
-    }
-
-    pub fn attach_reader(self) -> HashMapAccess<'a, K, V> {
-        // no difference to attach_writer currently
-        self.attach_writer()
-    }
-}
-
-// This is stored in the shared memory area
-struct HashMapShared<'a, K, V> {
-    inner: spin::RwLock<CoreHashMap<'a, K, V>>,
-}
-
-impl<'a, K, V> HashMapInit<'a, K, V>
-where
-    K: Clone + Hash + Eq,
-{
-    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
-    }
-
-    pub fn init_in_fixed_area(
-        num_buckets: u32,
-        area: &'a mut [MaybeUninit<u8>],
-    ) -> HashMapInit<'a, K, V> {
-        Self::init_common(num_buckets, None, area.as_mut_ptr().cast(), area.len())
-    }
-
-    /// Initialize a new hash map in the given shared memory area
-    pub fn init_in_shmem(num_buckets: u32, mut shmem: ShmemHandle) -> HashMapInit<'a, K, V> {
-        let size = Self::estimate_size(num_buckets);
-        shmem
-            .set_size(size)
-            .expect("could not resize shared memory area");
-
-        let ptr = unsafe { shmem.data_ptr.as_mut() };
-        Self::init_common(num_buckets, Some(shmem), ptr, size)
-    }
-
-    fn init_common(
-        num_buckets: u32,
-        shmem_handle: Option<ShmemHandle>,
-        area_ptr: *mut u8,
-        area_len: usize,
-    ) -> HashMapInit<'a, K, V> {
-        // carve out HashMapShared from the area. This does not include the hashmap's dictionary
-        // and buckets.
-        let mut ptr: *mut u8 = area_ptr;
-        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>>()) };
-
-        // the rest of the space is given to the hash map's dictionary and buckets
-        let remaining_area = unsafe {
-            std::slice::from_raw_parts_mut(ptr, area_len - ptr.offset_from(area_ptr) as usize)
-        };
-
-        let hashmap = CoreHashMap::new(num_buckets, remaining_area);
-        unsafe {
-            std::ptr::write(
-                shared_ptr,
-                HashMapShared {
-                    inner: spin::RwLock::new(hashmap),
-                },
-            );
-        }
-
-        HashMapInit {
-            shmem_handle,
-            shared_ptr,
-        }
-    }
-}
-
-impl<'a, K, V> HashMapAccess<'a, K, V>
-where
-    K: Clone + Hash + Eq,
-{
-    pub fn get<'e>(&'e self, key: &K) -> Option<ValueReadGuard<'e, K, V>> {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        let lock_guard = map.inner.read();
-
-        match lock_guard.get(key) {
-            None => None,
-            Some(val_ref) => {
-                let val_ptr = std::ptr::from_ref(val_ref);
-                Some(ValueReadGuard {
-                    _lock_guard: lock_guard,
-                    value: val_ptr,
-                })
-            }
-        }
-    }
-
-    /// Insert a value
-    pub fn insert(&self, key: &K, value: V) -> Result<bool, OutOfMemoryError> {
-        let mut success = None;
-
-        self.update_with_fn(key, |existing| {
-            if existing.is_some() {
-                success = Some(false);
-                UpdateAction::Nothing
-            } else {
-                success = Some(true);
-                UpdateAction::Insert(value)
-            }
-        })?;
-        Ok(success.expect("value_fn not called"))
-    }
-
-    /// Remove value. Returns true if it existed
-    pub fn remove(&self, key: &K) -> bool {
-        let mut result = false;
-        self.update_with_fn(key, |existing| match existing {
-            Some(_) => {
-                result = true;
-                UpdateAction::Remove
-            }
-            None => UpdateAction::Nothing,
-        })
-        .expect("out of memory while removing");
-        result
-    }
-
-    /// Update key using the given function. All the other modifying operations are based on this.
-    pub fn update_with_fn<F>(&self, key: &K, value_fn: F) -> Result<(), OutOfMemoryError>
-    where
-        F: FnOnce(Option<&V>) -> UpdateAction<V>,
-    {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        let mut lock_guard = map.inner.write();
-
-        let old_val = lock_guard.get(key);
-        let action = value_fn(old_val);
-        match (old_val, action) {
-            (_, UpdateAction::Nothing) => {}
-            (_, UpdateAction::Insert(new_val)) => {
-                let _ = lock_guard.insert(key, new_val);
-            }
-            (None, UpdateAction::Remove) => panic!("Remove action with no old value"),
-            (Some(_), UpdateAction::Remove) => {
-                let _ = lock_guard.remove(key);
-            }
-        }
-
-        Ok(())
-    }
-
-    /// Update key using the given function. All the other modifying operations are based on this.
-    pub fn update_with_fn_at_bucket<F>(
-        &self,
-        pos: usize,
-        value_fn: F,
-    ) -> Result<(), OutOfMemoryError>
-    where
-        F: FnOnce(Option<&V>) -> UpdateAction<V>,
-    {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        let mut lock_guard = map.inner.write();
-
-        let old_val = lock_guard.get_bucket(pos);
-        let action = value_fn(old_val.map(|(_k, v)| v));
-        match (old_val, action) {
-            (_, UpdateAction::Nothing) => {}
-            (_, UpdateAction::Insert(_new_val)) => panic!("cannot insert without key"),
-            (None, UpdateAction::Remove) => panic!("Remove action with no old value"),
-            (Some((key, _value)), UpdateAction::Remove) => {
-                let key = key.clone();
-                let _ = lock_guard.remove(&key);
-            }
-        }
-
-        Ok(())
-    }
-
-    pub fn get_num_buckets(&self) -> usize {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        map.inner.read().get_num_buckets()
-    }
-
-    /// Return the key and value stored in bucket with given index. This can be used to
-    /// iterate through the hash map. (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_bucket<'e>(&'e self, pos: usize) -> Option<ValueReadGuard<'e, K, V>> {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        let lock_guard = map.inner.read();
-
-        match lock_guard.get_bucket(pos) {
-            None => None,
-            Some((_key, val_ref)) => {
-                let val_ptr = std::ptr::from_ref(val_ref);
-                Some(ValueReadGuard {
-                    _lock_guard: lock_guard,
-                    value: val_ptr,
-                })
-            }
-        }
-    }
-
-    // for metrics
-    pub fn get_num_buckets_in_use(&self) -> usize {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        map.inner.read().buckets_in_use as usize
-    }
-
-    /// Grow
-    ///
-    /// 1. grow the underlying shared memory area
-    /// 2. Initialize new buckets. This overwrites the current dictionary
-    /// 3. Recalculate the dictionary
-    pub fn grow(&self, num_buckets: u32) -> Result<(), crate::shmem::Error> {
-        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
-        let mut lock_guard = map.inner.write();
-        let inner = &mut *lock_guard;
-        let old_num_buckets = inner.buckets.len() as u32;
-
-        if num_buckets < old_num_buckets {
-            panic!("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>::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 = inner.buckets.as_mut_ptr();
-        unsafe {
-            for i in old_num_buckets..num_buckets {
-                let bucket_ptr = buckets_ptr.add(i as usize);
-                bucket_ptr.write(core::Bucket {
-                    hash: 0,
-                    next: if i < num_buckets {
-                        i + 1
-                    } else {
-                        inner.free_head
-                    },
-                    inner: None,
-                });
-            }
-        }
-
-        // Recalculate the dictionary
-        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 item in dictionary.iter_mut() {
-            *item = core::INVALID_POS;
-        }
-
-        #[allow(clippy::needless_range_loop)]
-        for i in 0..old_num_buckets as usize {
-            if buckets[i].inner.is_none() {
-                continue;
-            }
-            let pos: usize = (buckets[i].hash % dictionary.len() as u64) as usize;
-            buckets[i].next = dictionary[pos];
-            dictionary[pos] = i as u32;
-        }
-
-        // Finally, update the CoreHashMap struct
-        inner.dictionary = dictionary;
-        inner.buckets = buckets;
-        inner.free_head = old_num_buckets;
-
-        Ok(())
-    }
-
-    // TODO: Shrinking is a multi-step process that requires co-operation from the caller
-    //
-    // 1. The caller must first call begin_shrink(). That forbids allocation of higher-numbered
-    // buckets.
-    //
-    // 2. Next, the caller must evict all entries in higher-numbered buckets.
-    //
-    // 3. Finally, call finish_shrink(). This recomputes the dictionary and shrinks the underlying
-    //    shmem area
-}
-
-pub struct ValueReadGuard<'a, K, V> {
-    _lock_guard: spin::RwLockReadGuard<'a, CoreHashMap<'a, K, V>>,
-    value: *const V,
-}
-
-impl<'a, K, V> Deref for ValueReadGuard<'a, K, V> {
-    type Target = V;
-
-    fn deref(&self) -> &Self::Target {
-        // SAFETY: The `lock_guard` ensures that the underlying map (and thus the value pointed to
-        // by `value`) remains valid for the lifetime `'a`. The `value` has been obtained from a
-        // valid reference within the map.
-        unsafe { &*self.value }
-    }
-}

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

@@ -1,233 +0,0 @@
-//! Simple hash table with chaining
-//!
-//! # Resizing
-//!
-
-use std::hash::{DefaultHasher, Hash, Hasher};
-use std::mem::MaybeUninit;
-
-pub(crate) const INVALID_POS: u32 = u32::MAX;
-
-// Bucket
-pub(crate) struct Bucket<K, V> {
-    pub(crate) hash: u64,
-    pub(crate) next: u32,
-    pub(crate) inner: Option<(K, V)>,
-}
-
-pub(crate) struct CoreHashMap<'a, K, V> {
-    pub(crate) dictionary: &'a mut [u32],
-    pub(crate) buckets: &'a mut [Bucket<K, V>],
-    pub(crate) free_head: u32,
-
-    // metrics
-    pub(crate) buckets_in_use: u32,
-}
-
-pub struct FullError();
-
-impl<'a, K, V> CoreHashMap<'a, K, V>
-where
-    K: Clone + Hash + Eq,
-{
-    const FILL_FACTOR: f32 = 0.60;
-
-    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(num_buckets: u32, area: &'a mut [u8]) -> CoreHashMap<'a, K, V> {
-        let len = area.len();
-
-        let mut ptr: *mut u8 = area.as_mut_ptr();
-        let end_ptr: *mut u8 = unsafe { area.as_mut_ptr().add(len) };
-
-        // carve out the buckets
-        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<Bucket<K, V>>())) };
-        let buckets_ptr = ptr;
-        ptr = unsafe { ptr.add(size_of::<Bucket<K, V>>() * num_buckets as usize) };
-
-        // use remaining space for the dictionary
-        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<u32>())) };
-        let dictionary_ptr = ptr;
-
-        assert!(ptr.addr() < end_ptr.addr());
-        let dictionary_size = unsafe { end_ptr.byte_offset_from(ptr) / size_of::<u32>() as isize };
-        assert!(dictionary_size > 0);
-
-        // Initialize the buckets
-        let buckets = {
-            let buckets_ptr: *mut MaybeUninit<Bucket<K, V>> = buckets_ptr.cast();
-            let buckets =
-                unsafe { std::slice::from_raw_parts_mut(buckets_ptr, num_buckets as usize) };
-            for i in 0..buckets.len() {
-                buckets[i].write(Bucket {
-                    hash: 0,
-                    next: if i < buckets.len() - 1 {
-                        i as u32 + 1
-                    } else {
-                        INVALID_POS
-                    },
-                    inner: None,
-                });
-            }
-            // TODO: use std::slice::assume_init_mut() once it stabilizes
-            unsafe { std::slice::from_raw_parts_mut(buckets_ptr.cast(), num_buckets as usize) }
-        };
-
-        // Initialize the dictionary
-        let dictionary = {
-            let dictionary_ptr: *mut MaybeUninit<u32> = dictionary_ptr.cast();
-            let dictionary =
-                unsafe { std::slice::from_raw_parts_mut(dictionary_ptr, dictionary_size as usize) };
-
-            for item in dictionary.iter_mut() {
-                item.write(INVALID_POS);
-            }
-            // TODO: use std::slice::assume_init_mut() once it stabilizes
-            unsafe {
-                std::slice::from_raw_parts_mut(dictionary_ptr.cast(), dictionary_size as usize)
-            }
-        };
-
-        CoreHashMap {
-            dictionary,
-            buckets,
-            free_head: 0,
-            buckets_in_use: 0,
-        }
-    }
-
-    pub fn get(&self, key: &K) -> Option<&V> {
-        let mut hasher = DefaultHasher::new();
-        key.hash(&mut hasher);
-        let hash = hasher.finish();
-
-        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;
-        }
-    }
-
-    pub fn insert(&mut self, key: &K, value: V) -> Result<(), FullError> {
-        let mut hasher = DefaultHasher::new();
-        key.hash(&mut hasher);
-        let hash = hasher.finish();
-
-        let first = self.dictionary[hash as usize % self.dictionary.len()];
-        if first == INVALID_POS {
-            // no existing entry
-            let pos = self.alloc_bucket(key.clone(), value, hash)?;
-            if pos == INVALID_POS {
-                return Err(FullError());
-            }
-            self.dictionary[hash as usize % self.dictionary.len()] = pos;
-            return Ok(());
-        }
-
-        let mut next = first;
-        loop {
-            let bucket = &mut self.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, update its value
-                *bucket_value = value;
-                return Ok(());
-            }
-
-            if bucket.next == INVALID_POS {
-                // No existing entry found. Append to the chain
-                let pos = self.alloc_bucket(key.clone(), value, hash)?;
-                if pos == INVALID_POS {
-                    return Err(FullError());
-                }
-                self.buckets[next as usize].next = pos;
-                return Ok(());
-            }
-            next = bucket.next;
-        }
-    }
-
-    pub fn remove(&mut self, key: &K) -> Result<(), FullError> {
-        let mut hasher = DefaultHasher::new();
-        key.hash(&mut hasher);
-        let hash = hasher.finish();
-
-        let mut next = self.dictionary[hash as usize % self.dictionary.len()];
-        let mut prev_pos: u32 = INVALID_POS;
-        loop {
-            if next == INVALID_POS {
-                // no existing entry
-                return Ok(());
-            }
-            let bucket = &mut self.buckets[next as usize];
-            let (bucket_key, _) = bucket.inner.as_mut().expect("entry is in use");
-            if bucket_key == key {
-                // found existing entry, unlink it from the chain
-                if prev_pos == INVALID_POS {
-                    self.dictionary[hash as usize % self.dictionary.len()] = bucket.next;
-                } else {
-                    self.buckets[prev_pos as usize].next = bucket.next;
-                }
-
-                // and add it to the freelist
-                let bucket = &mut self.buckets[next as usize];
-                bucket.hash = 0;
-                bucket.inner = None;
-                bucket.next = self.free_head;
-                self.free_head = next;
-                self.buckets_in_use -= 1;
-                return Ok(());
-            }
-            prev_pos = next;
-            next = bucket.next;
-        }
-    }
-
-    pub fn get_num_buckets(&self) -> usize {
-        self.buckets.len()
-    }
-
-    pub fn get_bucket(&self, pos: usize) -> Option<&(K, V)> {
-        if pos >= self.buckets.len() {
-            return None;
-        }
-
-        self.buckets[pos].inner.as_ref()
-    }
-
-    fn alloc_bucket(&mut self, key: K, value: V, hash: u64) -> Result<u32, FullError> {
-        let pos = self.free_head;
-        if pos == INVALID_POS {
-            return Err(FullError());
-        }
-
-        let bucket = &mut self.buckets[pos as usize];
-        self.free_head = bucket.next;
-        self.buckets_in_use += 1;
-
-        bucket.hash = hash;
-        bucket.next = INVALID_POS;
-        bucket.inner = Some((key, value));
-
-        Ok(pos)
-    }
-}

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

@@ -1,220 +0,0 @@
-use std::collections::BTreeMap;
-use std::collections::HashSet;
-use std::fmt::{Debug, Formatter};
-use std::sync::atomic::{AtomicUsize, Ordering};
-
-use crate::hash::HashMapAccess;
-use crate::hash::HashMapInit;
-use crate::hash::UpdateAction;
-use crate::shmem::ShmemHandle;
-
-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]) {
-    const MAX_MEM_SIZE: usize = 10000000;
-    let shmem = ShmemHandle::new("test_inserts", 0, MAX_MEM_SIZE).unwrap();
-
-    let init_struct = HashMapInit::<TestKey, usize>::init_in_shmem(100000, shmem);
-    let w = init_struct.attach_writer();
-
-    for (idx, k) in keys.iter().enumerate() {
-        let res = w.insert(&(*k).into(), 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));
-    }
-
-    //eprintln!("stats: {:?}", tree_writer.get_statistics());
-}
-
-#[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.contains(&key) {
-                continue;
-            }
-            used_keys.insert(key);
-            keys.push(key.into());
-            break;
-        }
-    }
-    test_inserts(&keys);
-}
-
-struct TestValue(AtomicUsize);
-
-impl TestValue {
-    fn new(val: usize) -> TestValue {
-        TestValue(AtomicUsize::new(val))
-    }
-
-    fn load(&self) -> usize {
-        self.0.load(Ordering::Relaxed)
-    }
-}
-
-impl Clone for TestValue {
-    fn clone(&self) -> TestValue {
-        TestValue::new(self.load())
-    }
-}
-
-impl Debug for TestValue {
-    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result<(), std::fmt::Error> {
-        write!(fmt, "{:?}", self.load())
-    }
-}
-
-#[derive(Clone, Debug)]
-struct TestOp(TestKey, Option<usize>);
-
-fn apply_op(
-    op: &TestOp,
-    sut: &HashMapAccess<TestKey, TestValue>,
-    shadow: &mut BTreeMap<TestKey, usize>,
-) {
-    eprintln!("applying op: {op:?}");
-
-    // 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)
-    };
-
-    // apply to Art tree
-    sut.update_with_fn(&op.0, |existing| {
-        assert_eq!(existing.map(TestValue::load), shadow_existing);
-
-        match (existing, op.1) {
-            (None, None) => UpdateAction::Nothing,
-            (None, Some(new_val)) => UpdateAction::Insert(TestValue::new(new_val)),
-            (Some(_old_val), None) => UpdateAction::Remove,
-            (Some(old_val), Some(new_val)) => {
-                old_val.0.store(new_val, Ordering::Relaxed);
-                UpdateAction::Nothing
-            }
-        }
-    })
-    .expect("out of memory");
-}
-
-#[test]
-fn random_ops() {
-    const MAX_MEM_SIZE: usize = 10000000;
-    let shmem = ShmemHandle::new("test_inserts", 0, MAX_MEM_SIZE).unwrap();
-
-    let init_struct = HashMapInit::<TestKey, TestValue>::init_in_shmem(100000, shmem);
-    let writer = init_struct.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, &writer, &mut shadow);
-
-        if i % 1000 == 0 {
-            eprintln!("{i} ops processed");
-            //eprintln!("stats: {:?}", tree_writer.get_statistics());
-            //test_iter(&tree_writer, &shadow);
-        }
-    }
-}
-
-#[test]
-fn test_grow() {
-    const MEM_SIZE: usize = 10000000;
-    let shmem = ShmemHandle::new("test_grow", 0, MEM_SIZE).unwrap();
-
-    let init_struct = HashMapInit::<TestKey, TestValue>::init_in_shmem(1000, shmem);
-    let writer = init_struct.attach_writer();
-
-    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
-
-    let mut rng = rand::rng();
-    for i in 0..10000 {
-        let key: TestKey = ((rng.next_u32() % 1000) as u128).into();
-
-        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
-
-        apply_op(&op, &writer, &mut shadow);
-
-        if i % 1000 == 0 {
-            eprintln!("{i} ops processed");
-            //eprintln!("stats: {:?}", tree_writer.get_statistics());
-            //test_iter(&tree_writer, &shadow);
-        }
-    }
-
-    writer.grow(1500).unwrap();
-
-    for i in 0..10000 {
-        let key: TestKey = ((rng.next_u32() % 1500) as u128).into();
-
-        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
-
-        apply_op(&op, &writer, &mut shadow);
-
-        if i % 1000 == 0 {
-            eprintln!("{i} ops processed");
-            //eprintln!("stats: {:?}", tree_writer.get_statistics());
-            //test_iter(&tree_writer, &shadow);
-        }
-    }
-}

libs/neon-shmem/src/lib.rs

@@ -1,4 +1,418 @@
 //! Shared memory utilities for neon communicator
 
-pub mod hash;
-pub mod shmem;
+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) -> 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/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

@@ -1,599 +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
-#[allow(clippy::only_used_in_recursion)]
-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),
-    }
-}
-
-#[allow(clippy::only_used_in_recursion)]
-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.is_empty() {
-        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
-#[allow(clippy::only_used_in_recursion)]
-#[allow(clippy::too_many_arguments)]
-pub(crate) fn update_recurse<'e, 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: &'_ 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..];
-    let level = level + prefix_match_len;
-
-    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();
-        }
-        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<V: Value + std::fmt::Debug>(
-    root: RootPtr<V>,
-    epoch_pin: &'_ 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
-#[allow(clippy::only_used_in_recursion)]
-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.is_empty() {
-        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<K: Key, V: Value, A: ArtAllocator<V>>(
-    key: &[u8],
-    value: V,
-    node: &mut WriteLockedNodeRef<V>,
-    parent: &mut WriteLockedNodeRef<V>,
-    parent_key: u8,
-    guard: &'_ 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<K: Key, V: Value, A: ArtAllocator<V>>(
-    wnode: &mut WriteLockedNodeRef<V>,
-    key: &[u8],
-    value: V,
-    guard: &'_ 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 {
-    version + 2
-}
-
-fn is_obsolete(version: u64) -> bool {
-    (version & 1) == 1
-}
-
-fn is_locked(version: u64) -> bool {
-    (version & 2) == 2
-}

libs/neonart/src/algorithm/node_ref.rs

@@ -1,349 +0,0 @@
-use std::fmt::Debug;
-use std::marker::PhantomData;
-
-use super::node_ptr;
-use super::node_ptr::NodePtr;
-use crate::EpochPin;
-use crate::Value;
-use crate::algorithm::lock_and_version::AtomicLockAndVersion;
-use crate::algorithm::lock_and_version::ConcurrentUpdateError;
-use crate::allocator::ArtAllocator;
-use crate::allocator::OutOfMemoryError;
-
-pub struct NodeRef<'e, V> {
-    ptr: NodePtr<V>,
-
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V> Debug for NodeRef<'e, V> {
-    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
-        write!(fmt, "{:?}", self.ptr)
-    }
-}
-
-impl<'e, V: Value> NodeRef<'e, V> {
-    pub(crate) fn from_root_ptr(root_ptr: NodePtr<V>) -> NodeRef<'e, V> {
-        NodeRef {
-            ptr: root_ptr,
-            phantom: PhantomData,
-        }
-    }
-
-    pub(crate) fn read_lock_or_restart(
-        &self,
-    ) -> Result<ReadLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        let version = self.lockword().read_lock_or_restart()?;
-        Ok(ReadLockedNodeRef {
-            ptr: self.ptr,
-            version,
-            phantom: self.phantom,
-        })
-    }
-
-    pub(crate) fn write_lock_or_restart(
-        &self,
-    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        self.lockword().write_lock_or_restart()?;
-        Ok(WriteLockedNodeRef {
-            ptr: self.ptr,
-            phantom: self.phantom,
-        })
-    }
-
-    fn lockword(&self) -> &AtomicLockAndVersion {
-        self.ptr.lockword()
-    }
-}
-
-/// A reference to a node that has been optimistically read-locked. The functions re-check
-/// the version after each read.
-pub struct ReadLockedNodeRef<'e, V> {
-    ptr: NodePtr<V>,
-    version: u64,
-
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V: Value> ReadLockedNodeRef<'e, V> {
-    pub(crate) fn is_leaf(&self) -> bool {
-        self.ptr.is_leaf()
-    }
-
-    pub(crate) fn is_full(&self) -> bool {
-        self.ptr.is_full()
-    }
-
-    pub(crate) fn get_prefix(&self) -> &[u8] {
-        self.ptr.get_prefix()
-    }
-
-    /// Note: because we're only holding a read lock, the prefix can change concurrently.
-    /// You must be prepared to restart, if read_unlock() returns error later.
-    ///
-    /// Returns the length of the prefix, or None if it's not a match
-    pub(crate) fn prefix_matches(&self, key: &[u8]) -> Option<usize> {
-        self.ptr.prefix_matches(key)
-    }
-
-    pub(crate) fn find_child_or_restart(
-        &self,
-        key_byte: u8,
-    ) -> Result<Option<NodeRef<'e, V>>, ConcurrentUpdateError> {
-        let child_or_value = self.ptr.find_child(key_byte);
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        match child_or_value {
-            None => Ok(None),
-            Some(child_ptr) => Ok(Some(NodeRef {
-                ptr: child_ptr,
-                phantom: self.phantom,
-            })),
-        }
-    }
-
-    pub(crate) fn find_next_child_or_restart(
-        &self,
-        min_key_byte: u8,
-    ) -> Result<Option<(u8, NodeRef<'e, V>)>, ConcurrentUpdateError> {
-        let child_or_value = self.ptr.find_next_child(min_key_byte);
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        match child_or_value {
-            None => Ok(None),
-            Some((k, child_ptr)) => Ok(Some((
-                k,
-                NodeRef {
-                    ptr: child_ptr,
-                    phantom: self.phantom,
-                },
-            ))),
-        }
-    }
-
-    pub(crate) fn get_leaf_value_ptr(&self) -> Result<*const V, ConcurrentUpdateError> {
-        let result = self.ptr.get_leaf_value();
-        self.ptr.lockword().check_or_restart(self.version)?;
-
-        // Extend the lifetime.
-        let result = std::ptr::from_ref(result);
-
-        Ok(result)
-    }
-
-    pub(crate) fn upgrade_to_write_lock_or_restart(
-        self,
-    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
-        self.ptr
-            .lockword()
-            .upgrade_to_write_lock_or_restart(self.version)?;
-
-        Ok(WriteLockedNodeRef {
-            ptr: self.ptr,
-            phantom: self.phantom,
-        })
-    }
-
-    pub(crate) fn read_unlock_or_restart(self) -> Result<(), ConcurrentUpdateError> {
-        self.ptr.lockword().check_or_restart(self.version)?;
-        Ok(())
-    }
-
-    pub(crate) fn check_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
-        self.ptr.lockword().check_or_restart(self.version)?;
-        Ok(())
-    }
-}
-
-/// A reference to a node that has been optimistically read-locked. The functions re-check
-/// the version after each read.
-pub struct WriteLockedNodeRef<'e, V> {
-    ptr: NodePtr<V>,
-    phantom: PhantomData<&'e EpochPin<'e>>,
-}
-
-impl<'e, V: Value> WriteLockedNodeRef<'e, V> {
-    pub(crate) fn can_shrink(&self) -> bool {
-        self.ptr.can_shrink()
-    }
-
-    pub(crate) fn num_children(&self) -> usize {
-        self.ptr.num_children()
-    }
-
-    pub(crate) fn write_unlock(mut self) {
-        self.ptr.lockword().write_unlock();
-        self.ptr = NodePtr::null();
-    }
-
-    pub(crate) fn write_unlock_obsolete(mut self) {
-        self.ptr.lockword().write_unlock_obsolete();
-        self.ptr = NodePtr::null();
-    }
-
-    pub(crate) fn get_prefix(&self) -> &[u8] {
-        self.ptr.get_prefix()
-    }
-
-    pub(crate) fn truncate_prefix(&mut self, new_prefix_len: usize) {
-        self.ptr.truncate_prefix(new_prefix_len)
-    }
-
-    pub(crate) fn prepend_prefix(&mut self, prefix: &[u8], prefix_byte: u8) {
-        self.ptr.prepend_prefix(prefix, prefix_byte)
-    }
-
-    pub(crate) fn insert_child(&mut self, key_byte: u8, child: NodePtr<V>) {
-        self.ptr.insert_child(key_byte, child)
-    }
-
-    pub(crate) fn get_leaf_value_mut(&mut self) -> &mut V {
-        self.ptr.get_leaf_value_mut()
-    }
-
-    pub(crate) fn grow<'a, A>(
-        &self,
-        allocator: &'a A,
-    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-    where
-        A: ArtAllocator<V>,
-    {
-        let new_node = self.ptr.grow(allocator)?;
-        Ok(NewNodeRef {
-            ptr: new_node,
-            allocator,
-            extra_nodes: Vec::new(),
-        })
-    }
-
-    pub(crate) fn shrink<'a, A>(
-        &self,
-        allocator: &'a A,
-    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-    where
-        A: ArtAllocator<V>,
-    {
-        let new_node = self.ptr.shrink(allocator)?;
-        Ok(NewNodeRef {
-            ptr: new_node,
-            allocator,
-            extra_nodes: Vec::new(),
-        })
-    }
-
-    pub(crate) fn as_ptr(&self) -> NodePtr<V> {
-        self.ptr
-    }
-
-    pub(crate) fn replace_child(&mut self, key_byte: u8, replacement: NodePtr<V>) {
-        self.ptr.replace_child(key_byte, replacement);
-    }
-
-    pub(crate) fn delete_child(&mut self, key_byte: u8) {
-        self.ptr.delete_child(key_byte);
-    }
-
-    pub(crate) fn find_remaining_child(&self) -> (u8, NodeRef<'e, V>) {
-        assert_eq!(self.num_children(), 1);
-        let child_or_value = self.ptr.find_next_child(0);
-
-        match child_or_value {
-            None => panic!("could not find only child in node"),
-            Some((k, child_ptr)) => (
-                k,
-                NodeRef {
-                    ptr: child_ptr,
-                    phantom: self.phantom,
-                },
-            ),
-        }
-    }
-}
-
-impl<'e, V> Drop for WriteLockedNodeRef<'e, V> {
-    fn drop(&mut self) {
-        if !self.ptr.is_null() {
-            self.ptr.lockword().write_unlock();
-        }
-    }
-}
-
-pub(crate) struct NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    ptr: NodePtr<V>,
-    allocator: &'a A,
-
-    extra_nodes: Vec<NodePtr<V>>,
-}
-
-impl<'a, V, A> NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    pub(crate) fn insert_old_child(&mut self, key_byte: u8, child: &WriteLockedNodeRef<V>) {
-        self.ptr.insert_child(key_byte, child.as_ptr())
-    }
-
-    pub(crate) fn into_ptr(mut self) -> NodePtr<V> {
-        let ptr = self.ptr;
-        self.ptr = NodePtr::null();
-        ptr
-    }
-
-    pub(crate) fn insert_new_child(&mut self, key_byte: u8, child: NewNodeRef<'a, V, A>) {
-        let child_ptr = child.into_ptr();
-        self.ptr.insert_child(key_byte, child_ptr);
-        self.extra_nodes.push(child_ptr);
-    }
-}
-
-impl<'a, V, A> Drop for NewNodeRef<'a, V, A>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    /// This drop implementation deallocates the newly allocated node, if into_ptr() was not called.
-    fn drop(&mut self) {
-        if !self.ptr.is_null() {
-            self.ptr.deallocate(self.allocator);
-            for p in self.extra_nodes.iter() {
-                p.deallocate(self.allocator);
-            }
-        }
-    }
-}
-
-pub(crate) fn new_internal<'a, V, A>(
-    prefix: &[u8],
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    Ok(NewNodeRef {
-        ptr: node_ptr::new_internal(prefix, allocator)?,
-        allocator,
-        extra_nodes: Vec::new(),
-    })
-}
-
-pub(crate) fn new_leaf<'a, V, A>(
-    prefix: &[u8],
-    value: V,
-    allocator: &'a A,
-) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
-where
-    V: Value,
-    A: ArtAllocator<V>,
-{
-    Ok(NewNodeRef {
-        ptr: node_ptr::new_leaf(prefix, value, allocator)?,
-        allocator,
-        extra_nodes: Vec::new(),
-    })
-}

libs/neonart/src/allocator.rs

@@ -1,156 +0,0 @@
-pub mod block;
-mod multislab;
-mod slab;
-pub mod r#static;
-
-use std::alloc::Layout;
-use std::marker::PhantomData;
-use std::mem::MaybeUninit;
-use std::sync::atomic::Ordering;
-
-use crate::allocator::multislab::MultiSlabAllocator;
-use crate::allocator::r#static::alloc_from_slice;
-
-use spin;
-
-use crate::Tree;
-pub use crate::algorithm::node_ptr::{
-    NodeInternal4, NodeInternal16, NodeInternal48, NodeInternal256, NodeLeaf,
-};
-
-#[derive(Debug)]
-pub struct OutOfMemoryError();
-
-pub trait ArtAllocator<V: crate::Value> {
-    fn alloc_tree(&self) -> *mut Tree<V>;
-
-    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V>;
-    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V>;
-    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V>;
-    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V>;
-    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V>;
-
-    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>);
-    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>);
-    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>);
-    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>);
-    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>);
-}
-
-pub struct ArtMultiSlabAllocator<'t, V>
-where
-    V: crate::Value,
-{
-    tree_area: spin::Mutex<Option<&'t mut MaybeUninit<Tree<V>>>>,
-
-    pub(crate) inner: MultiSlabAllocator<'t, 5>,
-
-    phantom_val: PhantomData<V>,
-}
-
-impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
-    const LAYOUTS: [Layout; 5] = [
-        Layout::new::<NodeInternal4<V>>(),
-        Layout::new::<NodeInternal16<V>>(),
-        Layout::new::<NodeInternal48<V>>(),
-        Layout::new::<NodeInternal256<V>>(),
-        Layout::new::<NodeLeaf<V>>(),
-    ];
-
-    pub fn new(area: &'t mut [MaybeUninit<u8>]) -> &'t mut ArtMultiSlabAllocator<'t, V> {
-        let (allocator_area, remain) = alloc_from_slice::<ArtMultiSlabAllocator<V>>(area);
-        let (tree_area, remain) = alloc_from_slice::<Tree<V>>(remain);
-
-        allocator_area.write(ArtMultiSlabAllocator {
-            tree_area: spin::Mutex::new(Some(tree_area)),
-            inner: MultiSlabAllocator::new(remain, &Self::LAYOUTS),
-            phantom_val: PhantomData,
-        })
-    }
-}
-
-impl<'t, V: crate::Value> ArtAllocator<V> for ArtMultiSlabAllocator<'t, V> {
-    fn alloc_tree(&self) -> *mut Tree<V> {
-        let mut t = self.tree_area.lock();
-        if let Some(tree_area) = t.take() {
-            return tree_area.as_mut_ptr().cast();
-        }
-        panic!("cannot allocate more than one tree");
-    }
-
-    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V> {
-        self.inner.alloc_slab(0).cast()
-    }
-    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V> {
-        self.inner.alloc_slab(1).cast()
-    }
-    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V> {
-        self.inner.alloc_slab(2).cast()
-    }
-    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V> {
-        self.inner.alloc_slab(3).cast()
-    }
-    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V> {
-        self.inner.alloc_slab(4).cast()
-    }
-
-    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>) {
-        self.inner.dealloc_slab(0, ptr.cast())
-    }
-
-    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>) {
-        self.inner.dealloc_slab(1, ptr.cast())
-    }
-    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>) {
-        self.inner.dealloc_slab(2, ptr.cast())
-    }
-    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>) {
-        self.inner.dealloc_slab(3, ptr.cast())
-    }
-    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>) {
-        self.inner.dealloc_slab(4, ptr.cast())
-    }
-}
-
-impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
-    pub(crate) fn get_statistics(&self) -> ArtMultiSlabStats {
-        ArtMultiSlabStats {
-            num_internal4: self.inner.slab_descs[0]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal16: self.inner.slab_descs[1]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal48: self.inner.slab_descs[2]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_internal256: self.inner.slab_descs[3]
-                .num_allocated
-                .load(Ordering::Relaxed),
-            num_leaf: self.inner.slab_descs[4]
-                .num_allocated
-                .load(Ordering::Relaxed),
-
-            num_blocks_internal4: self.inner.slab_descs[0].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal16: self.inner.slab_descs[1].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal48: self.inner.slab_descs[2].num_blocks.load(Ordering::Relaxed),
-            num_blocks_internal256: self.inner.slab_descs[3].num_blocks.load(Ordering::Relaxed),
-            num_blocks_leaf: self.inner.slab_descs[4].num_blocks.load(Ordering::Relaxed),
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct ArtMultiSlabStats {
-    pub num_internal4: u64,
-    pub num_internal16: u64,
-    pub num_internal48: u64,
-    pub num_internal256: u64,
-    pub num_leaf: u64,
-
-    pub num_blocks_internal4: u64,
-    pub num_blocks_internal16: u64,
-    pub num_blocks_internal48: u64,
-    pub num_blocks_internal256: u64,
-    pub num_blocks_leaf: u64,
-}

libs/neonart/src/allocator/block.rs

@@ -1,191 +0,0 @@
-//! Simple allocator of fixed-size blocks
-
-use std::mem::MaybeUninit;
-use std::sync::atomic::{AtomicU64, Ordering};
-
-use spin;
-
-pub const BLOCK_SIZE: usize = 16 * 1024;
-
-const INVALID_BLOCK: u64 = u64::MAX;
-
-pub(crate) struct BlockAllocator<'t> {
-    blocks_ptr: &'t [MaybeUninit<u8>],
-    num_blocks: u64,
-    num_initialized: AtomicU64,
-
-    freelist_head: spin::Mutex<u64>,
-}
-
-struct FreeListBlock {
-    inner: spin::Mutex<FreeListBlockInner>,
-}
-
-struct FreeListBlockInner {
-    next: u64,
-
-    num_free_blocks: u64,
-    free_blocks: [u64; 100], // FIXME: fill the rest of the block
-}
-
-impl<'t> BlockAllocator<'t> {
-    pub(crate) fn new(area: &'t mut [MaybeUninit<u8>]) -> Self {
-        // Use all the space for the blocks
-        let padding = area.as_ptr().align_offset(BLOCK_SIZE);
-        let remain = &mut area[padding..];
-
-        let num_blocks = (remain.len() / BLOCK_SIZE) as u64;
-
-        BlockAllocator {
-            blocks_ptr: remain,
-            num_blocks,
-            num_initialized: AtomicU64::new(0),
-            freelist_head: spin::Mutex::new(INVALID_BLOCK),
-        }
-    }
-
-    /// safety: you must hold a lock on the pointer to this block, otherwise it might get
-    /// reused for another kind of block
-    fn read_freelist_block(&self, blkno: u64) -> &FreeListBlock {
-        let ptr: *const FreeListBlock = self.get_block_ptr(blkno).cast();
-        unsafe { ptr.as_ref().unwrap() }
-    }
-
-    fn get_block_ptr(&self, blkno: u64) -> *mut u8 {
-        assert!(blkno < self.num_blocks);
-        unsafe {
-            self.blocks_ptr
-                .as_ptr()
-                .byte_offset(blkno as isize * BLOCK_SIZE as isize)
-        }
-        .cast_mut()
-        .cast()
-    }
-
-    #[allow(clippy::mut_from_ref)]
-    pub(crate) fn alloc_block(&self) -> &mut [MaybeUninit<u8>] {
-        // FIXME: handle OOM
-        let blkno = self.alloc_block_internal();
-        if blkno == INVALID_BLOCK {
-            panic!("out of memory");
-        }
-
-        let ptr: *mut MaybeUninit<u8> = self.get_block_ptr(blkno).cast();
-        unsafe { std::slice::from_raw_parts_mut(ptr, BLOCK_SIZE) }
-    }
-
-    fn alloc_block_internal(&self) -> u64 {
-        //  check the free list.
-        {
-            let mut freelist_head = self.freelist_head.lock();
-            if *freelist_head != INVALID_BLOCK {
-                let freelist_block = self.read_freelist_block(*freelist_head);
-
-                // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
-                let mut g = freelist_block.inner.lock();
-
-                if g.num_free_blocks > 0 {
-                    g.num_free_blocks -= 1;
-                    let result = g.free_blocks[g.num_free_blocks as usize];
-                    return result;
-                } else {
-                    // consume the freelist block itself
-                    let result = *freelist_head;
-                    *freelist_head = g.next;
-                    // This freelist block is now unlinked and can be repurposed
-                    drop(g);
-                    return result;
-                }
-            }
-        }
-
-        // If there are some blocks left that we've never used, pick next such block
-        let mut next_uninitialized = self.num_initialized.load(Ordering::Relaxed);
-        while next_uninitialized < self.num_blocks {
-            match self.num_initialized.compare_exchange(
-                next_uninitialized,
-                next_uninitialized + 1,
-                Ordering::Relaxed,
-                Ordering::Relaxed,
-            ) {
-                Ok(_) => {
-                    return next_uninitialized;
-                }
-                Err(old) => {
-                    next_uninitialized = old;
-                    continue;
-                }
-            }
-        }
-
-        // out of blocks
-        INVALID_BLOCK
-    }
-
-    // TODO: this is currently unused. The slab allocator never releases blocks
-    #[allow(dead_code)]
-    pub(crate) fn release_block(&self, block_ptr: *mut u8) {
-        let blockno = unsafe { block_ptr.byte_offset_from(self.blocks_ptr) / BLOCK_SIZE as isize };
-        self.release_block_internal(blockno as u64);
-    }
-
-    fn release_block_internal(&self, blockno: u64) {
-        let mut freelist_head = self.freelist_head.lock();
-        if *freelist_head != INVALID_BLOCK {
-            let freelist_block = self.read_freelist_block(*freelist_head);
-
-            // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
-            let mut g = freelist_block.inner.lock();
-
-            let num_free_blocks = g.num_free_blocks;
-            if num_free_blocks < g.free_blocks.len() as u64 {
-                g.free_blocks[num_free_blocks as usize] = blockno;
-                g.num_free_blocks += 1;
-                return;
-            }
-        }
-
-        // Convert the block into a new freelist block
-        let block_ptr: *mut FreeListBlock = self.get_block_ptr(blockno).cast();
-        let init = FreeListBlock {
-            inner: spin::Mutex::new(FreeListBlockInner {
-                next: *freelist_head,
-                num_free_blocks: 0,
-                free_blocks: [INVALID_BLOCK; 100],
-            }),
-        };
-        unsafe { (*block_ptr) = init };
-        *freelist_head = blockno;
-    }
-
-    // for debugging
-    pub(crate) fn get_statistics(&self) -> BlockAllocatorStats {
-        let mut num_free_blocks = 0;
-
-        let mut _prev_lock = None;
-        let head_lock = self.freelist_head.lock();
-        let mut next_blk = *head_lock;
-        let mut _head_lock = Some(head_lock);
-        while next_blk != INVALID_BLOCK {
-            let freelist_block = self.read_freelist_block(next_blk);
-            let lock = freelist_block.inner.lock();
-            num_free_blocks += lock.num_free_blocks;
-            next_blk = lock.next;
-            _prev_lock = Some(lock); // hold the lock until we've read the next block
-            _head_lock = None;
-        }
-
-        BlockAllocatorStats {
-            num_blocks: self.num_blocks,
-            num_initialized: self.num_initialized.load(Ordering::Relaxed),
-            num_free_blocks,
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct BlockAllocatorStats {
-    pub num_blocks: u64,
-    pub num_initialized: u64,
-    pub num_free_blocks: u64,
-}

libs/neonart/src/allocator/multislab.rs

@@ -1,33 +0,0 @@
-use std::alloc::Layout;
-use std::mem::MaybeUninit;
-
-use crate::allocator::block::BlockAllocator;
-use crate::allocator::slab::SlabDesc;
-
-pub struct MultiSlabAllocator<'t, const N: usize> {
-    pub(crate) block_allocator: BlockAllocator<'t>,
-
-    pub(crate) slab_descs: [SlabDesc; N],
-}
-
-impl<'t, const N: usize> MultiSlabAllocator<'t, N> {
-    pub(crate) fn new(
-        area: &'t mut [MaybeUninit<u8>],
-        layouts: &[Layout; N],
-    ) -> MultiSlabAllocator<'t, N> {
-        let block_allocator = BlockAllocator::new(area);
-        MultiSlabAllocator {
-            block_allocator,
-
-            slab_descs: std::array::from_fn(|i| SlabDesc::new(&layouts[i])),
-        }
-    }
-
-    pub(crate) fn alloc_slab(&self, slab_idx: usize) -> *mut u8 {
-        self.slab_descs[slab_idx].alloc_chunk(&self.block_allocator)
-    }
-
-    pub(crate) fn dealloc_slab(&self, slab_idx: usize, ptr: *mut u8) {
-        self.slab_descs[slab_idx].dealloc_chunk(ptr, &self.block_allocator)
-    }
-}

libs/neonart/src/allocator/slab.rs

@@ -1,433 +0,0 @@
-//! A slab allocator that carves out fixed-size chunks from larger blocks.
-//!
-//!
-
-use std::alloc::Layout;
-use std::mem::MaybeUninit;
-use std::ops::Deref;
-use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
-
-use spin;
-
-use super::alloc_from_slice;
-use super::block::BlockAllocator;
-
-use crate::allocator::block::BLOCK_SIZE;
-
-pub(crate) struct SlabDesc {
-    pub(crate) layout: Layout,
-
-    block_lists: spin::RwLock<BlockLists>,
-
-    pub(crate) num_blocks: AtomicU64,
-    pub(crate) num_allocated: AtomicU64,
-}
-
-// FIXME: Not sure if SlabDesc is really Sync or Send. It probably is when it's empty, but
-// 'block_lists' contains pointers when it's not empty. In the current use as part of the
-// the art tree, SlabDescs are only moved during initialization.
-unsafe impl Sync for SlabDesc {}
-unsafe impl Send for SlabDesc {}
-
-#[derive(Default, Debug)]
-struct BlockLists {
-    full_blocks: BlockList,
-    nonfull_blocks: BlockList,
-}
-
-impl BlockLists {
-    // Unlink a node. It must be in either one of the two lists.
-    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
-        let list = unsafe {
-            if (*elem).next.is_null() {
-                if self.full_blocks.tail == elem {
-                    Some(&mut self.full_blocks)
-                } else {
-                    Some(&mut self.nonfull_blocks)
-                }
-            } else if (*elem).prev.is_null() {
-                if self.full_blocks.head == elem {
-                    Some(&mut self.full_blocks)
-                } else {
-                    Some(&mut self.nonfull_blocks)
-                }
-            } else {
-                None
-            }
-        };
-        unsafe { unlink_slab_block(list, elem) };
-    }
-}
-
-unsafe fn unlink_slab_block(mut list: Option<&mut BlockList>, elem: *mut SlabBlockHeader) {
-    unsafe {
-        if (*elem).next.is_null() {
-            assert_eq!(list.as_ref().unwrap().tail, elem);
-            list.as_mut().unwrap().tail = (*elem).prev;
-        } else {
-            assert_eq!((*(*elem).next).prev, elem);
-            (*(*elem).next).prev = (*elem).prev;
-        }
-        if (*elem).prev.is_null() {
-            assert_eq!(list.as_ref().unwrap().head, elem);
-            list.as_mut().unwrap().head = (*elem).next;
-        } else {
-            assert_eq!((*(*elem).prev).next, elem);
-            (*(*elem).prev).next = (*elem).next;
-        }
-    }
-}
-
-#[derive(Debug)]
-struct BlockList {
-    head: *mut SlabBlockHeader,
-    tail: *mut SlabBlockHeader,
-}
-
-impl Default for BlockList {
-    fn default() -> Self {
-        BlockList {
-            head: std::ptr::null_mut(),
-            tail: std::ptr::null_mut(),
-        }
-    }
-}
-
-impl BlockList {
-    unsafe fn push_head(&mut self, elem: *mut SlabBlockHeader) {
-        unsafe {
-            if self.is_empty() {
-                self.tail = elem;
-                (*elem).next = std::ptr::null_mut();
-            } else {
-                (*elem).next = self.head;
-                (*self.head).prev = elem;
-            }
-            (*elem).prev = std::ptr::null_mut();
-            self.head = elem;
-        }
-    }
-
-    fn is_empty(&self) -> bool {
-        self.head.is_null()
-    }
-
-    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
-        unsafe { unlink_slab_block(Some(self), elem) }
-    }
-
-    #[cfg(test)]
-    fn dump(&self) {
-        let mut next = self.head;
-
-        while !next.is_null() {
-            let n = unsafe { next.as_ref() }.unwrap();
-            eprintln!(
-                "  blk {:?} (free {}/{})",
-                next,
-                n.num_free_chunks.load(Ordering::Relaxed),
-                n.num_chunks
-            );
-            next = n.next;
-        }
-    }
-}
-
-impl SlabDesc {
-    pub(crate) fn new(layout: &Layout) -> SlabDesc {
-        SlabDesc {
-            layout: *layout,
-            block_lists: spin::RwLock::new(BlockLists::default()),
-            num_allocated: AtomicU64::new(0),
-            num_blocks: AtomicU64::new(0),
-        }
-    }
-}
-
-#[derive(Debug)]
-struct SlabBlockHeader {
-    free_chunks_head: spin::Mutex<*mut FreeChunk>,
-    num_free_chunks: AtomicU32,
-    num_chunks: u32, // this is really a constant for a given Layout
-
-    // these fields are protected by the lock on the BlockLists
-    prev: *mut SlabBlockHeader,
-    next: *mut SlabBlockHeader,
-}
-
-struct FreeChunk {
-    next: *mut FreeChunk,
-}
-
-enum ReadOrWriteGuard<'a, T> {
-    Read(spin::RwLockReadGuard<'a, T>),
-    Write(spin::RwLockWriteGuard<'a, T>),
-}
-
-impl<'a, T> Deref for ReadOrWriteGuard<'a, T> {
-    type Target = T;
-
-    fn deref(&self) -> &<Self as Deref>::Target {
-        match self {
-            ReadOrWriteGuard::Read(g) => g.deref(),
-            ReadOrWriteGuard::Write(g) => g.deref(),
-        }
-    }
-}
-
-impl SlabDesc {
-    pub fn alloc_chunk(&self, block_allocator: &BlockAllocator) -> *mut u8 {
-        // Are there any free chunks?
-        let mut acquire_write = false;
-        'outer: loop {
-            let mut block_lists_guard = if acquire_write {
-                ReadOrWriteGuard::Write(self.block_lists.write())
-            } else {
-                ReadOrWriteGuard::Read(self.block_lists.read())
-            };
-            'inner: loop {
-                let block_ptr = block_lists_guard.nonfull_blocks.head;
-                if block_ptr.is_null() {
-                    break 'outer;
-                }
-                unsafe {
-                    let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
-                    if !(*free_chunks_head).is_null() {
-                        let result = *free_chunks_head;
-                        (*free_chunks_head) = (*result).next;
-                        let _old = (*block_ptr).num_free_chunks.fetch_sub(1, Ordering::Relaxed);
-
-                        self.num_allocated.fetch_add(1, Ordering::Relaxed);
-                        return result.cast();
-                    }
-                }
-
-                // The block at the head of the list was full. Grab write lock and retry
-                match block_lists_guard {
-                    ReadOrWriteGuard::Read(_) => {
-                        acquire_write = true;
-                        continue 'outer;
-                    }
-                    ReadOrWriteGuard::Write(ref mut g) => {
-                        // move the node to the list of full blocks
-                        unsafe {
-                            g.nonfull_blocks.unlink(block_ptr);
-                            g.full_blocks.push_head(block_ptr);
-                        };
-                        continue 'inner;
-                    }
-                }
-            }
-        }
-
-        // no free chunks. Allocate a new block (and the chunk from that)
-        let (new_block, new_chunk) = self.alloc_block_and_chunk(block_allocator);
-        self.num_blocks.fetch_add(1, Ordering::Relaxed);
-
-        // Add the block to the list in the SlabDesc
-        unsafe {
-            let mut block_lists_guard = self.block_lists.write();
-            block_lists_guard.nonfull_blocks.push_head(new_block);
-        }
-        self.num_allocated.fetch_add(1, Ordering::Relaxed);
-        new_chunk
-    }
-
-    pub fn dealloc_chunk(&self, chunk_ptr: *mut u8, _block_allocator: &BlockAllocator) {
-        // Find the block it belongs to. You can find the block from the address. (And knowing the
-        // layout, you could calculate the chunk number too.)
-        let block_ptr: *mut SlabBlockHeader = {
-            let block_addr = (chunk_ptr.addr() / BLOCK_SIZE) * BLOCK_SIZE;
-            chunk_ptr.with_addr(block_addr).cast()
-        };
-        let chunk_ptr: *mut FreeChunk = chunk_ptr.cast();
-
-        // Mark the chunk as free in 'freechunks' list
-        let num_chunks;
-        let num_free_chunks;
-        unsafe {
-            let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
-            (*chunk_ptr).next = *free_chunks_head;
-            *free_chunks_head = chunk_ptr;
-
-            num_free_chunks = (*block_ptr).num_free_chunks.fetch_add(1, Ordering::Relaxed) + 1;
-            num_chunks = (*block_ptr).num_chunks;
-        }
-
-        if num_free_chunks == 1 {
-            // If the block was full previously, add it to the nonfull blocks list. Note that
-            // we're not holding the lock anymore, so it can immediately become full again.
-            // That's harmless, it will be moved back to the full list again when a call
-            // to alloc_chunk() sees it.
-            let mut block_lists = self.block_lists.write();
-            unsafe {
-                block_lists.unlink(block_ptr);
-                block_lists.nonfull_blocks.push_head(block_ptr);
-            };
-        } else if num_free_chunks == num_chunks {
-            // If the block became completely empty, move it to the free list
-            // TODO
-            // FIXME: we're still holding the spinlock. It's not exactly safe to return it to
-            // the free blocks list, is it? Defer it as garbage to wait out concurrent updates?
-            //block_allocator.release_block()
-        }
-
-        // update stats
-        self.num_allocated.fetch_sub(1, Ordering::Relaxed);
-    }
-
-    fn alloc_block_and_chunk(
-        &self,
-        block_allocator: &BlockAllocator,
-    ) -> (*mut SlabBlockHeader, *mut u8) {
-        // fixme: handle OOM
-        let block_slice: &mut [MaybeUninit<u8>] = block_allocator.alloc_block();
-        let (block_header, remain) = alloc_from_slice::<SlabBlockHeader>(block_slice);
-
-        let padding = remain.as_ptr().align_offset(self.layout.align());
-
-        let num_chunks = (remain.len() - padding) / self.layout.size();
-
-        let first_chunk_ptr: *mut FreeChunk = remain[padding..].as_mut_ptr().cast();
-
-        unsafe {
-            let mut chunk_ptr = first_chunk_ptr;
-            for _ in 0..num_chunks - 1 {
-                let next_chunk_ptr = chunk_ptr.byte_add(self.layout.size());
-                (*chunk_ptr).next = next_chunk_ptr;
-                chunk_ptr = next_chunk_ptr;
-            }
-            (*chunk_ptr).next = std::ptr::null_mut();
-
-            let result_chunk = first_chunk_ptr;
-
-            let block_header = block_header.write(SlabBlockHeader {
-                free_chunks_head: spin::Mutex::new((*first_chunk_ptr).next),
-                prev: std::ptr::null_mut(),
-                next: std::ptr::null_mut(),
-                num_chunks: num_chunks as u32,
-                num_free_chunks: AtomicU32::new(num_chunks as u32 - 1),
-            });
-
-            (block_header, result_chunk.cast())
-        }
-    }
-
-    #[cfg(test)]
-    fn dump(&self) {
-        eprintln!(
-            "slab dump ({} blocks, {} allocated chunks)",
-            self.num_blocks.load(Ordering::Relaxed),
-            self.num_allocated.load(Ordering::Relaxed)
-        );
-        let lists = self.block_lists.read();
-
-        eprintln!("nonfull blocks:");
-        lists.nonfull_blocks.dump();
-        eprintln!("full blocks:");
-        lists.full_blocks.dump();
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    use rand::Rng;
-    use rand_distr::Zipf;
-
-    struct TestObject {
-        val: usize,
-        _dummy: [u8; BLOCK_SIZE / 4],
-    }
-
-    struct TestObjectSlab<'a>(SlabDesc, BlockAllocator<'a>);
-    impl<'a> TestObjectSlab<'a> {
-        fn new(block_allocator: BlockAllocator) -> TestObjectSlab {
-            TestObjectSlab(SlabDesc::new(&Layout::new::<TestObject>()), block_allocator)
-        }
-
-        fn alloc(&self, val: usize) -> *mut TestObject {
-            let obj: *mut TestObject = self.0.alloc_chunk(&self.1).cast();
-            unsafe { (*obj).val = val };
-            obj
-        }
-
-        fn dealloc(&self, obj: *mut TestObject) {
-            self.0.dealloc_chunk(obj.cast(), &self.1)
-        }
-    }
-
-    #[test]
-    fn test_slab_alloc() {
-        const MEM_SIZE: usize = 100000000;
-        let mut area = Box::new_uninit_slice(MEM_SIZE);
-        let block_allocator = BlockAllocator::new(&mut area);
-
-        let slab = TestObjectSlab::new(block_allocator);
-
-        let mut all: Vec<*mut TestObject> = Vec::new();
-        for i in 0..11 {
-            all.push(slab.alloc(i));
-        }
-        #[allow(clippy::needless_range_loop)]
-        for i in 0..11 {
-            assert!(unsafe { (*all[i]).val == i });
-        }
-
-        let distribution = Zipf::new(10.0, 1.1).unwrap();
-        let mut rng = rand::rng();
-        for _ in 0..100000 {
-            slab.0.dump();
-            let idx = rng.sample(distribution) as usize;
-            let ptr: *mut TestObject = all[idx];
-            if !ptr.is_null() {
-                assert_eq!(unsafe { (*ptr).val }, idx);
-                slab.dealloc(ptr);
-                all[idx] = std::ptr::null_mut();
-            } else {
-                all[idx] = slab.alloc(idx);
-            }
-        }
-    }
-
-    fn new_test_blk(i: u32) -> *mut SlabBlockHeader {
-        Box::into_raw(Box::new(SlabBlockHeader {
-            free_chunks_head: spin::Mutex::new(std::ptr::null_mut()),
-            num_free_chunks: AtomicU32::new(0),
-            num_chunks: i,
-            prev: std::ptr::null_mut(),
-            next: std::ptr::null_mut(),
-        }))
-    }
-
-    #[test]
-    fn test_block_linked_list() {
-        // note: these are leaked, but that's OK for tests
-        let a = new_test_blk(0);
-        let b = new_test_blk(1);
-
-        let mut list = BlockList::default();
-        assert!(list.is_empty());
-
-        unsafe {
-            list.push_head(a);
-            assert!(!list.is_empty());
-            list.unlink(a);
-        }
-        assert!(list.is_empty());
-
-        unsafe {
-            list.push_head(b);
-            list.push_head(a);
-            assert_eq!(list.head, a);
-            assert_eq!((*a).next, b);
-            assert_eq!((*b).prev, a);
-            assert_eq!(list.tail, b);
-
-            list.unlink(a);
-            list.unlink(b);
-            assert!(list.is_empty());
-        }
-    }
-}

libs/neonart/src/allocator/static.rs

@@ -1,44 +0,0 @@
-use std::mem::MaybeUninit;
-
-pub fn alloc_from_slice<T>(
-    area: &mut [MaybeUninit<u8>],
-) -> (&mut MaybeUninit<T>, &mut [MaybeUninit<u8>]) {
-    let layout = std::alloc::Layout::new::<T>();
-
-    let area_start = area.as_mut_ptr();
-
-    // pad to satisfy alignment requirements
-    let padding = area_start.align_offset(layout.align());
-    if padding + layout.size() > area.len() {
-        panic!("out of memory");
-    }
-    let area = &mut area[padding..];
-    let (result_area, remain) = area.split_at_mut(layout.size());
-
-    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
-    let result = unsafe { result_ptr.as_mut().unwrap() };
-
-    (result, remain)
-}
-
-pub fn alloc_array_from_slice<T>(
-    area: &mut [MaybeUninit<u8>],
-    len: usize,
-) -> (&mut [MaybeUninit<T>], &mut [MaybeUninit<u8>]) {
-    let layout = std::alloc::Layout::new::<T>();
-
-    let area_start = area.as_mut_ptr();
-
-    // pad to satisfy alignment requirements
-    let padding = area_start.align_offset(layout.align());
-    if padding + layout.size() * len > area.len() {
-        panic!("out of memory");
-    }
-    let area = &mut area[padding..];
-    let (result_area, remain) = area.split_at_mut(layout.size() * len);
-
-    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
-    let result = unsafe { std::slice::from_raw_parts_mut(result_ptr.as_mut().unwrap(), len) };
-
-    (result, remain)
-}