set thp:never during build

Microsoft exposes JWKs without the alg header. It's only included on the
tokens. Not a problem.

Also noticed that wrt the `typ` header:
> It will typically not be used by applications when it is already known
that the object is a JWT. This parameter is ignored by JWT
implementations; any processing of this parameter is performed by the
JWT application.

Since we know we are expecting JWTs only, I've followed the guidance and
removed the validation.

.dockerignore

@@ -13,6 +13,7 @@
 # Directories
 !.cargo/
 !.config/
+!compute/
 !compute_tools/
 !control_plane/
 !libs/

.github/workflows/_benchmarking_preparation.yml

@@ -3,19 +3,23 @@ name: Prepare benchmarking databases by restoring dumps
 on:
   workflow_call:
     # no inputs needed
-    
+
 defaults:
   run:
     shell: bash -euxo pipefail {0}
 
 jobs:
   setup-databases:
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
-        platform: [ aws-rds-postgres, aws-aurora-serverless-v2-postgres, neon ] 
+        platform: [ aws-rds-postgres, aws-aurora-serverless-v2-postgres, neon ]
         database: [ clickbench, tpch, userexample ]
-  
+
     env:
       LD_LIBRARY_PATH: /tmp/neon/pg_install/v16/lib
       PLATFORM: ${{ matrix.platform }}
@@ -23,7 +27,10 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -32,13 +39,13 @@ jobs:
       run: |
         case "${PLATFORM}" in
           neon)
-            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_CONNSTR }}
             ;;
           aws-rds-postgres)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_POSTGRES_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_RDS_POSTGRES_CONNSTR }}
             ;;
           aws-aurora-serverless-v2-postgres)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_AURORA_CONNSTR }} 
+            CONNSTR=${{ secrets.BENCHMARK_RDS_AURORA_CONNSTR }}
             ;;
           *)
             echo >&2 "Unknown PLATFORM=${PLATFORM}"
@@ -46,10 +53,17 @@ jobs:
             ;;
         esac
 
-        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT  
+        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      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:
@@ -57,23 +71,23 @@ jobs:
         path: /tmp/neon/
         prefix: latest
 
-    # we create a table that has one row for each database that we want to restore with the status whether the restore is done    
+    # we create a table that has one row for each database that we want to restore with the status whether the restore is done
     - name: Create benchmark_restore_status table if it does not exist
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-prep-connstr.outputs.connstr }}
         DATABASE_NAME: ${{ matrix.database }}
-      # to avoid a race condition of multiple jobs trying to create the table at the same time, 
+      # to avoid a race condition of multiple jobs trying to create the table at the same time,
       # we use an advisory lock
       run: |
         ${PG_BINARIES}/psql "${{ env.BENCHMARK_CONNSTR }}" -c "
-        SELECT pg_advisory_lock(4711);  
+        SELECT pg_advisory_lock(4711);
         CREATE TABLE IF NOT EXISTS benchmark_restore_status (
         databasename text primary key,
         restore_done boolean
         );
         SELECT pg_advisory_unlock(4711);
         "
-    
+
     - name: Check if restore is already done
       id: check-restore-done
       env:
@@ -107,7 +121,7 @@ jobs:
         DATABASE_NAME: ${{ matrix.database }}
       run: |
         mkdir -p /tmp/dumps
-        aws s3 cp s3://neon-github-dev/performance/pgdumps/$DATABASE_NAME/$DATABASE_NAME.pg_dump /tmp/dumps/ 
+        aws s3 cp s3://neon-github-dev/performance/pgdumps/$DATABASE_NAME/$DATABASE_NAME.pg_dump /tmp/dumps/
 
     - name: Replace database name in connection string
       if: steps.check-restore-done.outputs.skip != 'true'
@@ -126,17 +140,17 @@ jobs:
         else
           new_connstr="${base_connstr}/${DATABASE_NAME}"
         fi
-        echo "database_connstr=${new_connstr}" >> $GITHUB_OUTPUT  
+        echo "database_connstr=${new_connstr}" >> $GITHUB_OUTPUT
 
     - name: Restore dump
       if: steps.check-restore-done.outputs.skip != 'true'
       env:
         DATABASE_NAME: ${{ matrix.database }}
         DATABASE_CONNSTR: ${{ steps.replace-dbname.outputs.database_connstr }}
-        # the following works only with larger computes: 
+        # the following works only with larger computes:
         # PGOPTIONS: "-c maintenance_work_mem=8388608 -c max_parallel_maintenance_workers=7"
         # we add the || true because:
-        # the dumps were created with Neon and contain neon extensions that are not 
+        # the dumps were created with Neon and contain neon extensions that are not
         # available in RDS, so we will always report an error, but we can ignore it
       run: |
         ${PG_BINARIES}/pg_restore --clean --if-exists --no-owner --jobs=4 \

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

@@ -236,9 +236,7 @@ jobs:
 
           # run pageserver tests with different settings
           for io_engine in std-fs tokio-epoll-uring ; do
-            for io_buffer_alignment in 0 1 512 ; do
-              NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine NEON_PAGESERVER_UNIT_TEST_IO_BUFFER_ALIGNMENT=$io_buffer_alignment ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
-            done
+            NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES  -E 'package(pageserver)'
           done
 
           # Run separate tests for real S3
@@ -257,7 +255,15 @@ jobs:
           ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES -E 'package(remote_storage)' -E 'test(test_real_azure)'
 
       - name: Install postgres binaries
-        run: cp -a pg_install /tmp/neon/pg_install
+        run: |
+          # Use tar to copy files matching the pattern, preserving the paths in the destionation
+          tar c \
+            pg_install/v* \
+            pg_install/build/*/src/test/regress/*.so \
+            pg_install/build/*/src/test/regress/pg_regress \
+            pg_install/build/*/src/test/isolation/isolationtester \
+            pg_install/build/*/src/test/isolation/pg_isolation_regress \
+            | tar  x -C /tmp/neon
 
       - name: Upload Neon artifact
         uses: ./.github/actions/upload

.github/workflows/benchmarking.yml

@@ -12,7 +12,6 @@ on:
     #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
     #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
     - cron:   '0 3 * * *' # run once a day, timezone is utc
-
   workflow_dispatch: # adds ability to run this manually
     inputs:
       region_id:
@@ -59,7 +58,7 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
@@ -68,12 +67,10 @@ jobs:
             PLATFORM: "neon-staging"
             region_id: ${{ github.event.inputs.region_id || 'aws-us-east-2' }}
             RUNNER: [ self-hosted, us-east-2, x64 ]
-            IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
           - DEFAULT_PG_VERSION: 16
             PLATFORM: "azure-staging"
             region_id: 'azure-eastus2'
             RUNNER: [ self-hosted, eastus2, x64 ]
-            IMAGE: neondatabase/build-tools:pinned
     env:
       TEST_PG_BENCH_DURATIONS_MATRIX: "300"
       TEST_PG_BENCH_SCALES_MATRIX: "10,100"
@@ -86,7 +83,10 @@ jobs:
 
     runs-on: ${{ matrix.RUNNER }}
     container:
-      image: ${{ matrix.IMAGE }}
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -164,6 +164,10 @@ jobs:
 
   replication-tests:
     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:
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
       DEFAULT_PG_VERSION: 16
@@ -174,12 +178,21 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      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
@@ -267,7 +280,7 @@ jobs:
         region_id_default=${{ env.DEFAULT_REGION_ID }}
         runner_default='["self-hosted", "us-east-2", "x64"]'
         runner_azure='["self-hosted", "eastus2", "x64"]'
-        image_default="369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned"
+        image_default="neondatabase/build-tools:pinned"
         matrix='{
           "pg_version" : [
             16
@@ -344,7 +357,7 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
 
     strategy:
       fail-fast: false
@@ -371,7 +384,7 @@ jobs:
     steps:
     - uses: actions/checkout@v4
 
-    - name: Configure AWS credentials # necessary on Azure runners
+    - name: Configure AWS credentials
       uses: aws-actions/configure-aws-credentials@v4
       with:
         aws-region: eu-central-1
@@ -492,17 +505,15 @@ jobs:
     permissions:
       contents: write
       statuses: write
-      id-token: write # Required for OIDC authentication in azure runners
+      id-token: write # aws-actions/configure-aws-credentials
     strategy:
       fail-fast: false
       matrix:
         include:
           - PLATFORM: "neonvm-captest-pgvector"
             RUNNER: [ self-hosted, us-east-2, x64 ]
-            IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
           - PLATFORM: "azure-captest-pgvector"
             RUNNER: [ self-hosted, eastus2, x64 ]
-            IMAGE: neondatabase/build-tools:pinned
 
     env:
       TEST_PG_BENCH_DURATIONS_MATRIX: "15m"
@@ -511,13 +522,16 @@ jobs:
       DEFAULT_PG_VERSION: 16
       TEST_OUTPUT: /tmp/test_output
       BUILD_TYPE: remote
-      LD_LIBRARY_PATH: /home/nonroot/pg/usr/lib/x86_64-linux-gnu
+
       SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
       PLATFORM: ${{ matrix.PLATFORM }}
 
     runs-on: ${{ matrix.RUNNER }}
     container:
-      image: ${{ matrix.IMAGE }}
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
@@ -527,17 +541,26 @@ jobs:
     # instead of using Neon artifacts containing pgbench
     - name: Install postgresql-16 where pytest expects it
       run: |
+        # Just to make it easier to test things locally on macOS (with arm64)
+        arch=$(uname -m | sed 's/x86_64/amd64/g' | sed 's/aarch64/arm64/g')
+
         cd /home/nonroot
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/libpq5_16.4-1.pgdg110%2B1_amd64.deb
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.4-1.pgdg110%2B1_amd64.deb
-        wget -q https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.4-1.pgdg110%2B1_amd64.deb 
-        dpkg -x libpq5_16.4-1.pgdg110+1_amd64.deb pg
-        dpkg -x postgresql-client-16_16.4-1.pgdg110+1_amd64.deb pg
-        dpkg -x postgresql-16_16.4-1.pgdg110+1_amd64.deb pg
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-17/libpq5_17.0-1.pgdg110+1_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.4-1.pgdg110+2_${arch}.deb"
+        dpkg -x libpq5_17.0-1.pgdg110+1_${arch}.deb pg
+        dpkg -x postgresql-16_16.4-1.pgdg110+2_${arch}.deb pg
+        dpkg -x postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb pg
+
         mkdir -p /tmp/neon/pg_install/v16/bin
-        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/pgbench /tmp/neon/pg_install/v16/bin/pgbench  
-        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/psql /tmp/neon/pg_install/v16/bin/psql  
-        ln -s /home/nonroot/pg/usr/lib/x86_64-linux-gnu /tmp/neon/pg_install/v16/lib 
+        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/pgbench /tmp/neon/pg_install/v16/bin/pgbench
+        ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/psql    /tmp/neon/pg_install/v16/bin/psql
+        ln -s /home/nonroot/pg/usr/lib/$(uname -m)-linux-gnu     /tmp/neon/pg_install/v16/lib
+
+        LD_LIBRARY_PATH="/home/nonroot/pg/usr/lib/$(uname -m)-linux-gnu:${LD_LIBRARY_PATH:-}"
+        export LD_LIBRARY_PATH
+        echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}" >> ${GITHUB_ENV}
+
         /tmp/neon/pg_install/v16/bin/pgbench --version
         /tmp/neon/pg_install/v16/bin/psql --version
 
@@ -559,7 +582,7 @@ jobs:
 
         echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT
 
-    - name: Configure AWS credentials # necessary on Azure runners to read/write from/to S3
+    - name: Configure AWS credentials
       uses: aws-actions/configure-aws-credentials@v4
       with:
         aws-region: eu-central-1
@@ -620,6 +643,10 @@ jobs:
     # *_CLICKBENCH_CONNSTR: Genuine ClickBench DB with ~100M rows
     # *_CLICKBENCH_10M_CONNSTR: DB with the first 10M rows of ClickBench DB
     if: ${{ !cancelled() && (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
     needs: [ generate-matrices, pgbench-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -638,12 +665,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      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:
@@ -714,6 +751,10 @@ jobs:
     #
     # *_TPCH_S10_CONNSTR: DB generated with scale factor 10 (~10 GB)
     if: ${{ !cancelled() && (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
     needs: [ generate-matrices, clickbench-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -731,12 +772,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      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:
@@ -806,6 +857,10 @@ jobs:
 
   user-examples-compare:
     if: ${{ !cancelled() && (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
     needs: [ generate-matrices, tpch-compare, prepare_AWS_RDS_databases ]
 
     strategy:
@@ -822,12 +877,22 @@ jobs:
 
     runs-on: [ self-hosted, us-east-2, x64 ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:pinned
+      image: neondatabase/build-tools:pinned
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
       options: --init
 
     steps:
     - uses: actions/checkout@v4
 
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@v4
+      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:

.github/workflows/build_and_test.yml

@@ -212,6 +212,10 @@ jobs:
       # This will catch compiler & clippy warnings in all feature combinations.
       # TODO: use cargo hack for build and test as well, but, that's quite expensive.
       # NB: keep clippy args in sync with ./run_clippy.sh
+      #
+      # The only difference between "clippy --debug" and "clippy --release" is that in --release mode,
+      # #[cfg(debug_assertions)] blocks are not built. It's not worth building everything for second
+      # time just for that, so skip "clippy --release".
       - run: |
           CLIPPY_COMMON_ARGS="$( source .neon_clippy_args; echo "$CLIPPY_COMMON_ARGS")"
           if [ "$CLIPPY_COMMON_ARGS" = "" ]; then
@@ -221,8 +225,6 @@ jobs:
           echo "CLIPPY_COMMON_ARGS=${CLIPPY_COMMON_ARGS}" >> $GITHUB_ENV
       - name: Run cargo clippy (debug)
         run: cargo hack --feature-powerset clippy $CLIPPY_COMMON_ARGS
-      - name: Run cargo clippy (release)
-        run: cargo hack --feature-powerset clippy --release $CLIPPY_COMMON_ARGS
 
       - name: Check documentation generation
         run: cargo doc --workspace --no-deps --document-private-items
@@ -600,7 +602,20 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version: [ v14, v15, v16, v17 ]
+        version:
+          # Much data was already generated on old PG versions with bullseye's
+          # libraries, the locales of which can cause data incompatibilities.
+          # However, new PG versions should check if they can be built on newer
+          # images, as that reduces the support burden of old and ancient
+          # distros.
+          - pg: v14
+            debian: bullseye-slim
+          - pg: v15
+            debian: bullseye-slim
+          - pg: v16
+            debian: bullseye-slim
+          - pg: v17
+            debian: bookworm-slim
         arch: [ x64, arm64 ]
 
     runs-on: ${{ fromJson(format('["self-hosted", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}
@@ -643,41 +658,46 @@ jobs:
           context: .
           build-args: |
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
-            PG_VERSION=${{ matrix.version }}
+            PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
-          cache-from: type=registry,ref=cache.neon.build/compute-node-${{ matrix.version }}:cache-${{ matrix.arch }}
-          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/compute-node-{0}:cache-{1},mode=max', matrix.version, matrix.arch) || '' }}
+          file: compute/Dockerfile.compute-node
+          cache-from: type=registry,ref=cache.neon.build/compute-node-${{ matrix.version.pg }}:cache-${{ matrix.arch }}
+          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/compute-node-{0}:cache-{1},mode=max', matrix.version.pg, matrix.arch) || '' }}
           tags: |
-            neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
+            neondatabase/compute-node-${{ matrix.version.pg }}:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
       - name: Build neon extensions test image
-        if: matrix.version == 'v16'
+        if: matrix.version.pg == 'v16'
         uses: docker/build-push-action@v6
         with:
           context: .
           build-args: |
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
-            PG_VERSION=${{ matrix.version }}
+            PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
+          file: compute/Dockerfile.compute-node
           target: neon-pg-ext-test
-          cache-from: type=registry,ref=cache.neon.build/neon-test-extensions-${{ matrix.version }}:cache-${{ matrix.arch }}
-          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/neon-test-extensions-{0}:cache-{1},mode=max', matrix.version, matrix.arch) || '' }}
+          cache-from: type=registry,ref=cache.neon.build/neon-test-extensions-${{ matrix.version.pg }}:cache-${{ matrix.arch }}
+          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/neon-test-extensions-{0}:cache-{1},mode=max', matrix.version.pg, matrix.arch) || '' }}
           tags: |
-            neondatabase/neon-test-extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}-${{ matrix.arch }}
+            neondatabase/neon-test-extensions-${{ matrix.version.pg }}:${{needs.tag.outputs.build-tag}}-${{ matrix.arch }}
 
       - name: Build compute-tools image
         # compute-tools are Postgres independent, so build it only once
-        if: matrix.version == 'v17'
+        # We pick 16, because that builds on debian 11 with older glibc (and is
+        # thus compatible with newer glibc), rather than 17 on Debian 12, as
+        # that isn't guaranteed to be compatible with Debian 11
+        if: matrix.version.pg == 'v16'
         uses: docker/build-push-action@v6
         with:
           target: compute-tools-image
@@ -686,10 +706,11 @@ jobs:
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
+          file: compute/Dockerfile.compute-node
           tags: |
             neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
@@ -752,7 +773,7 @@ jobs:
       matrix:
         version: [ v14, v15, v16, v17 ]
     env:
-      VM_BUILDER_VERSION: v0.29.3
+      VM_BUILDER_VERSION: v0.35.0
 
     steps:
       - uses: actions/checkout@v4
@@ -777,7 +798,7 @@ jobs:
       - name: Build vm image
         run: |
           ./vm-builder \
-            -spec=vm-image-spec.yaml \
+            -spec=compute/vm-image-spec.yaml \
             -src=neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }} \
             -dst=neondatabase/vm-compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
 
@@ -841,6 +862,9 @@ jobs:
     needs: [ check-permissions, tag, test-images, vm-compute-node-image ]
     runs-on: ubuntu-22.04
 
+    permissions:
+      id-token: write # for `aws-actions/configure-aws-credentials`
+
     env:
       VERSIONS: v14 v15 v16 v17
 
@@ -885,13 +909,19 @@ jobs:
           docker buildx imagetools create -t neondatabase/neon-test-extensions-v16:latest \
                                               neondatabase/neon-test-extensions-v16:${{ needs.tag.outputs.build-tag }}
 
+      - name: Configure AWS-prod credentials
+        if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          aws-region: eu-central-1
+          mask-aws-account-id: true
+          role-to-assume: ${{ secrets.PROD_GHA_OIDC_ROLE }}
+
       - name: Login to prod ECR
         uses: docker/login-action@v3
         if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
         with:
           registry: 093970136003.dkr.ecr.eu-central-1.amazonaws.com
-          username: ${{ secrets.PROD_GHA_RUNNER_LIMITED_AWS_ACCESS_KEY_ID }}
-          password: ${{ secrets.PROD_GHA_RUNNER_LIMITED_AWS_SECRET_ACCESS_KEY }}
 
       - name: Copy all images to prod ECR
         if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
@@ -1160,10 +1190,9 @@ jobs:
 
               files_to_promote+=("s3://${BUCKET}/${s3_key}")
 
-              # TODO Add v17
-              for pg_version in v14 v15 v16; do
+              for pg_version in v14 v15 v16 v17; do
                 # We run less tests for debug builds, so we don't need to promote them
-                if [ "${build_type}" == "debug" ] && { [ "${arch}" == "ARM64" ] || [ "${pg_version}" != "v16" ] ; }; then
+                if [ "${build_type}" == "debug" ] && { [ "${arch}" == "ARM64" ] || [ "${pg_version}" != "v17" ] ; }; then
                   continue
                 fi
 
@@ -1205,7 +1234,6 @@ jobs:
     # Usually we do `needs: [...]`
     needs:
       - build-and-test-locally
-      - check-submodules
       - check-codestyle-python
       - check-codestyle-rust
       - promote-images

.github/workflows/cloud-regress.yml

@@ -0,0 +1,102 @@
+name: Cloud Regression Test
+on:
+  schedule:
+    # * is a special character in YAML so you have to quote this string
+    #          ┌───────────── minute (0 - 59)
+    #          │ ┌───────────── hour (0 - 23)
+    #          │ │ ┌───────────── day of the month (1 - 31)
+    #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    - cron:  '45 1 * * *' # run once a day, timezone is utc
+  workflow_dispatch: # adds ability to run this manually
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+
+concurrency:
+  # Allow only one workflow
+  group: ${{ github.workflow }}
+  cancel-in-progress: true
+
+jobs:
+  regress:
+    env:
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 16
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+    runs-on: us-east-2
+    container:
+      image: neondatabase/build-tools:pinned
+      options: --init
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+
+      - name: Patch the test
+        run: |
+          cd "vendor/postgres-v${DEFAULT_PG_VERSION}"
+          patch -p1 < "../../compute/patches/cloud_regress_pg${DEFAULT_PG_VERSION}.patch"
+
+      - name: Generate a random password
+        id: pwgen
+        run: |
+          set +x
+          DBPASS=$(dd if=/dev/random bs=48 count=1 2>/dev/null | base64)
+          echo "::add-mask::${DBPASS//\//}"
+          echo DBPASS="${DBPASS//\//}" >> "${GITHUB_OUTPUT}"
+
+      - name: Change tests according to the generated password
+        env:
+          DBPASS: ${{ steps.pwgen.outputs.DBPASS }}
+        run: |
+          cd vendor/postgres-v"${DEFAULT_PG_VERSION}"/src/test/regress
+          for fname in sql/*.sql expected/*.out; do
+            sed -i.bak s/NEON_PASSWORD_PLACEHOLDER/"'${DBPASS}'"/ "${fname}"
+          done
+          for ph in $(grep NEON_MD5_PLACEHOLDER expected/password.out | awk '{print $3;}' | sort | uniq); do
+            USER=$(echo "${ph}" | cut -c 22-)
+            MD5=md5$(echo -n "${DBPASS}${USER}" | md5sum | awk '{print $1;}')
+            sed -i.bak "s/${ph}/${MD5}/" expected/password.out
+          done
+
+      - name: Download Neon artifact
+        uses: ./.github/actions/download
+        with:
+          name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+          path: /tmp/neon/
+          prefix: latest
+
+      - name: Run the regression tests
+        uses: ./.github/actions/run-python-test-set
+        with:
+          build_type: ${{ env.BUILD_TYPE }}
+          test_selection: cloud_regress
+          pg_version: ${{ env.DEFAULT_PG_VERSION }}
+          extra_params: -m remote_cluster
+        env:
+          BENCHMARK_CONNSTR: ${{ secrets.PG_REGRESS_CONNSTR }}
+
+      - name: Create Allure report
+        id: create-allure-report
+        if: ${{ !cancelled() }}
+        uses: ./.github/actions/allure-report-generate
+
+      - name: Post to a Slack channel
+        if: ${{ github.event.schedule && failure() }}
+        uses: slackapi/slack-github-action@v1
+        with:
+          channel-id: "C033QLM5P7D" # on-call-staging-stream
+          slack-message: |
+            Periodic pg_regress on staging: ${{ job.status }}
+            <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
+            <${{ steps.create-allure-report.outputs.report-url }}|Allure report>
+        env:
+          SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
+

.github/workflows/trigger-e2e-tests.yml

@@ -102,12 +102,12 @@ jobs:
           # Default set of platforms to run e2e tests on
           platforms='["docker", "k8s"]'
 
-          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
+          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or compute/Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
           # If the workflow run is not a pull request, add k8s-neonvm to the list.
           if [ "$GITHUB_EVENT_NAME" == "pull_request" ]; then
             for f in $(gh api "/repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" --paginate --jq '.[].filename'); do
               case "$f" in
-                vendor/*|pgxn/*|libs/vm_monitor/*|Dockerfile.compute-node)
+                vendor/*|pgxn/*|libs/vm_monitor/*|compute/Dockerfile.compute-node)
                   platforms=$(echo "${platforms}" | jq --compact-output '. += ["k8s-neonvm"] | unique')
                   ;;
                 *)

Cargo.toml

@@ -76,8 +76,6 @@ clap = { version = "4.0", features = ["derive"] }
 comfy-table = "7.1"
 const_format = "0.2"
 crc32c = "0.6"
-crossbeam-deque = "0.8.5"
-crossbeam-utils = "0.8.5"
 dashmap = { version = "5.5.0", features = ["raw-api"] }
 either = "1.8"
 enum-map = "2.4.2"
@@ -95,7 +93,7 @@ hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
-hostname = "0.3.1"
+hostname = "0.4"
 http = {version = "1.1.0", features = ["std"]}
 http-types = { version = "2", default-features = false }
 humantime = "2.1"
@@ -104,7 +102,6 @@ hyper = "0.14"
 tokio-tungstenite = "0.20.0"
 indexmap = "2"
 indoc = "2"
-inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
 jsonwebtoken = "9"
@@ -113,15 +110,16 @@ libc = "0.2"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
-memoffset = "0.8"
+memoffset = "0.9"
 nix = { version = "0.27", features = ["dir", "fs", "process", "socket", "signal", "poll"] }
 notify = "6.0.0"
 num_cpus = "1.15"
 num-traits = "0.2.15"
 once_cell = "1.13"
-opentelemetry = "0.20.0"
-opentelemetry-otlp = { version = "0.13.0", default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
-opentelemetry-semantic-conventions = "0.12.0"
+opentelemetry = "0.24"
+opentelemetry_sdk = "0.24"
+opentelemetry-otlp = { version = "0.17", default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry-semantic-conventions = "0.16"
 parking_lot = "0.12"
 parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
@@ -134,7 +132,7 @@ rand = "0.8"
 redis = { version = "0.25.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_20"] }
+reqwest-tracing = { version = "0.5", features = ["opentelemetry_0_24"] }
 reqwest-middleware = "0.3.0"
 reqwest-retry = "0.5"
 routerify = "3"
@@ -142,7 +140,6 @@ rpds = "0.13"
 rustc-hash = "1.1.0"
 rustls = "0.22"
 rustls-pemfile = "2"
-rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
 sd-notify = "0.4.1"
@@ -164,7 +161,6 @@ strum_macros = "0.26"
 svg_fmt = "0.4.3"
 sync_wrapper = "0.1.2"
 tar = "0.4"
-task-local-extensions = "0.1.4"
 test-context = "0.3"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"
@@ -182,8 +178,8 @@ toml_edit = "0.22"
 tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tower-service = "0.3.2"
 tracing = "0.1"
-tracing-error = "0.2.0"
-tracing-opentelemetry = "0.21.0"
+tracing-error = "0.2"
+tracing-opentelemetry = "0.25"
 tracing-subscriber = { version = "0.3", default-features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 try-lock = "0.2.5"
 twox-hash = { version = "1.6.3", default-features = false }

Dockerfile

@@ -42,6 +42,7 @@ COPY --from=pg-build /home/nonroot/pg_install/v17/lib                       pg_i
 COPY --chown=nonroot . .
 
 ARG ADDITIONAL_RUSTFLAGS
+ENV _RJEM_MALLOC_CONF="thp:never"
 RUN set -e \
     && PQ_LIB_DIR=$(pwd)/pg_install/v${STABLE_PG_VERSION}/lib RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment ${ADDITIONAL_RUSTFLAGS}" cargo build \
       --bin pg_sni_router  \

Dockerfile.build-tools

@@ -13,6 +13,9 @@ RUN useradd -ms /bin/bash nonroot -b /home
 SHELL ["/bin/bash", "-c"]
 
 # System deps
+#
+# 'gdb' is included so that we get backtraces of core dumps produced in
+# regression tests
 RUN set -e \
     && apt update \
     && apt install -y \
@@ -24,6 +27,7 @@ RUN set -e \
         cmake \
         curl \
         flex \
+        gdb \
         git \
         gnupg \
         gzip \

compute/Dockerfile.compute-node

@@ -3,13 +3,15 @@ ARG REPOSITORY=neondatabase
 ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
+ARG DEBIAN_FLAVOR=bullseye-slim
 
 #########################################################################################
 #
 # Layer "build-deps"
 #
 #########################################################################################
-FROM debian:bullseye-slim AS build-deps
+FROM debian:$DEBIAN_FLAVOR AS build-deps
+ARG DEBIAN_FLAVOR
 RUN apt update &&  \
     apt install -y git autoconf automake libtool build-essential bison flex libreadline-dev \
     zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget pkg-config libssl-dev \
@@ -280,7 +282,7 @@ FROM build-deps AS vector-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-COPY patches/pgvector.patch /pgvector.patch
+COPY compute/patches/pgvector.patch /pgvector.patch
 
 # By default, pgvector Makefile uses `-march=native`. We don't want that,
 # because we build the images on different machines than where we run them.
@@ -366,7 +368,7 @@ FROM build-deps AS rum-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-COPY patches/rum.patch /rum.patch
+COPY compute/patches/rum.patch /rum.patch
 
 RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
@@ -869,6 +871,28 @@ RUN case "${PG_VERSION}" in "v17") \
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/ulid.control
 
+#########################################################################################
+#
+# Layer "pg-session-jwt-build"
+# Compile "pg_session_jwt" extension
+#
+#########################################################################################
+
+FROM rust-extensions-build AS pg-session-jwt-build
+ARG PG_VERSION
+
+RUN case "${PG_VERSION}" in "v17") \
+    echo "pg_session_jwt does not yet have a release that supports pg17" && exit 0;; \
+    esac && \
+    wget https://github.com/neondatabase/pg_session_jwt/archive/ff0a72440e8ff584dab24b3f9b7c00c56c660b8e.tar.gz -O pg_session_jwt.tar.gz && \
+    echo "1fbb2b5a339263bcf6daa847fad8bccbc0b451cea6a62e6d3bf232b0087f05cb pg_session_jwt.tar.gz" | sha256sum --check && \
+    mkdir pg_session_jwt-src && cd pg_session_jwt-src && tar xzf ../pg_session_jwt.tar.gz --strip-components=1 -C . && \
+    sed -i 's/pgrx = "=0.11.3"/pgrx = { version = "=0.11.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release
+    # it's needed to enable extension because it uses untrusted C language
+    # sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_session_jwt.control && \
+    # echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_session_jwt.control
+
 #########################################################################################
 #
 # Layer "wal2json-build"
@@ -965,6 +989,7 @@ COPY --from=timescaledb-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-hint-plan-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-cron-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-pgx-ulid-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-session-jwt-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
@@ -1027,10 +1052,47 @@ RUN cd compute_tools && mold -run cargo build --locked --profile release-line-de
 #
 #########################################################################################
 
-FROM debian:bullseye-slim AS compute-tools-image
+FROM debian:$DEBIAN_FLAVOR AS compute-tools-image
+ARG DEBIAN_FLAVOR
 
 COPY --from=compute-tools /home/nonroot/target/release-line-debug-size-lto/compute_ctl /usr/local/bin/compute_ctl
 
+#########################################################################################
+#
+# Layer "pgbouncer"
+#
+#########################################################################################
+
+FROM debian:$DEBIAN_FLAVOR AS pgbouncer
+ARG DEBIAN_FLAVOR
+RUN set -e \
+    && apt-get update \
+    && apt-get install -y \
+        build-essential \
+        git \
+        libevent-dev \
+        libtool \
+        pkg-config
+
+# Use `dist_man_MANS=` to skip manpage generation (which requires python3/pandoc)
+ENV PGBOUNCER_TAG=pgbouncer_1_22_1
+RUN set -e \
+    && git clone --recurse-submodules --depth 1 --branch ${PGBOUNCER_TAG} https://github.com/pgbouncer/pgbouncer.git pgbouncer \
+    && cd pgbouncer \
+    && ./autogen.sh \
+    && LDFLAGS=-static ./configure --prefix=/usr/local/pgbouncer --without-openssl \
+    && make -j $(nproc) dist_man_MANS= \
+    && make install dist_man_MANS=
+
+#########################################################################################
+#
+# Layers "postgres-exporter" and "sql-exporter"
+#
+#########################################################################################
+
+FROM quay.io/prometheuscommunity/postgres-exporter:v0.12.1 AS postgres-exporter
+FROM burningalchemist/sql_exporter:0.13 AS sql-exporter
+
 #########################################################################################
 #
 # Clean up postgres folder before inclusion
@@ -1078,7 +1140,7 @@ COPY --from=pgjwt-pg-build /pgjwt.tar.gz /ext-src
 COPY --from=hypopg-pg-build /hypopg.tar.gz /ext-src
 COPY --from=pg-hashids-pg-build /pg_hashids.tar.gz /ext-src
 COPY --from=rum-pg-build /rum.tar.gz /ext-src
-COPY patches/rum.patch /ext-src
+COPY compute/patches/rum.patch /ext-src
 #COPY --from=pgtap-pg-build /pgtap.tar.gz /ext-src
 COPY --from=ip4r-pg-build /ip4r.tar.gz /ext-src
 COPY --from=prefix-pg-build /prefix.tar.gz /ext-src
@@ -1086,9 +1148,9 @@ COPY --from=hll-pg-build /hll.tar.gz /ext-src
 COPY --from=plpgsql-check-pg-build /plpgsql_check.tar.gz /ext-src
 #COPY --from=timescaledb-pg-build /timescaledb.tar.gz /ext-src
 COPY --from=pg-hint-plan-pg-build /pg_hint_plan.tar.gz /ext-src
-COPY patches/pg_hint_plan.patch /ext-src
+COPY compute/patches/pg_hint_plan.patch /ext-src
 COPY --from=pg-cron-pg-build /pg_cron.tar.gz /ext-src
-COPY patches/pg_cron.patch /ext-src
+COPY compute/patches/pg_cron.patch /ext-src
 #COPY --from=pg-pgx-ulid-build /home/nonroot/pgx_ulid.tar.gz /ext-src
 #COPY --from=rdkit-pg-build /rdkit.tar.gz /ext-src
 COPY --from=pg-uuidv7-pg-build /pg_uuidv7.tar.gz /ext-src
@@ -1097,7 +1159,7 @@ COPY --from=pg-semver-pg-build /pg_semver.tar.gz /ext-src
 #COPY --from=pg-embedding-pg-build /home/nonroot/pg_embedding-src/ /ext-src
 #COPY --from=wal2json-pg-build /wal2json_2_5.tar.gz /ext-src
 COPY --from=pg-anon-pg-build /pg_anon.tar.gz /ext-src
-COPY patches/pg_anon.patch /ext-src
+COPY compute/patches/pg_anon.patch /ext-src
 COPY --from=pg-ivm-build /pg_ivm.tar.gz /ext-src
 COPY --from=pg-partman-build /pg_partman.tar.gz /ext-src
 RUN case "${PG_VERSION}" in "v17") \
@@ -1144,7 +1206,9 @@ ENV PGDATABASE=postgres
 # Put it all together into the final image
 #
 #########################################################################################
-FROM debian:bullseye-slim
+FROM debian:$DEBIAN_FLAVOR
+ARG DEBIAN_FLAVOR
+ENV DEBIAN_FLAVOR=$DEBIAN_FLAVOR
 # Add user postgres
 RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
     echo "postgres:test_console_pass" | chpasswd && \
@@ -1160,23 +1224,50 @@ RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
 COPY --from=postgres-cleanup-layer --chown=postgres /usr/local/pgsql /usr/local
 COPY --from=compute-tools --chown=postgres /home/nonroot/target/release-line-debug-size-lto/compute_ctl /usr/local/bin/compute_ctl
 
+# pgbouncer and its config
+COPY --from=pgbouncer         /usr/local/pgbouncer/bin/pgbouncer /usr/local/bin/pgbouncer
+COPY --chmod=0666 --chown=postgres compute/etc/pgbouncer.ini /etc/pgbouncer.ini
+
+# Metrics exporter binaries and  configuration files
+COPY --from=postgres-exporter /bin/postgres_exporter /bin/postgres_exporter
+COPY --from=sql-exporter      /bin/sql_exporter      /bin/sql_exporter
+
+COPY --chmod=0644 compute/etc/sql_exporter.yml               /etc/sql_exporter.yml
+COPY --chmod=0644 compute/etc/neon_collector.yml             /etc/neon_collector.yml
+COPY --chmod=0644 compute/etc/sql_exporter_autoscaling.yml   /etc/sql_exporter_autoscaling.yml
+COPY --chmod=0644 compute/etc/neon_collector_autoscaling.yml /etc/neon_collector_autoscaling.yml
+
 # Create remote extension download directory
 RUN mkdir /usr/local/download_extensions && chown -R postgres:postgres /usr/local/download_extensions
 
 # Install:
 # libreadline8 for psql
-# libicu67, locales for collations (including ICU and plpgsql_check)
 # liblz4-1 for lz4
 # libossp-uuid16 for extension ossp-uuid
-# libgeos, libgdal, libsfcgal1, libproj and libprotobuf-c1 for PostGIS
+# libgeos, libsfcgal1, and libprotobuf-c1 for PostGIS
 # libxml2, libxslt1.1 for xml2
 # libzstd1 for zstd
 # libboost* for rdkit
 # ca-certificates for communicating with s3 by compute_ctl
-RUN apt update &&  \
+
+
+RUN apt update && \
+    case $DEBIAN_FLAVOR in \
+      # Version-specific installs for Bullseye (PG14-PG16):
+      # libicu67, locales for collations (including ICU and plpgsql_check)
+      # libgdal28, libproj19 for PostGIS
+      bullseye*) \
+        VERSION_INSTALLS="libicu67 libgdal28 libproj19"; \
+      ;; \
+      # Version-specific installs for Bookworm (PG17):
+      # libicu72, locales for collations (including ICU and plpgsql_check)
+      # libgdal32, libproj25 for PostGIS
+      bookworm*) \
+        VERSION_INSTALLS="libicu72 libgdal32 libproj25"; \
+      ;; \
+    esac && \
     apt install --no-install-recommends -y \
         gdb \
-        libicu67 \
         liblz4-1 \
         libreadline8 \
         libboost-iostreams1.74.0 \
@@ -1185,17 +1276,16 @@ RUN apt update &&  \
         libboost-system1.74.0 \
         libossp-uuid16 \
         libgeos-c1v5 \
-        libgdal28 \
-        libproj19 \
         libprotobuf-c1 \
         libsfcgal1 \
         libxml2 \
         libxslt1.1 \
         libzstd1 \
-        libcurl4-openssl-dev \
+        libcurl4 \
         locales \
         procps \
-        ca-certificates && \
+        ca-certificates \
+        $VERSION_INSTALLS && \
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* && \
     localedef -i en_US -c -f UTF-8 -A /usr/share/locale/locale.alias en_US.UTF-8
 

compute/README.md

@@ -0,0 +1,21 @@
+This directory contains files that are needed to build the compute
+images, or included in the compute images.
+
+Dockerfile.compute-node
+	To build the compute image
+
+vm-image-spec.yaml
+	Instructions for vm-builder, to turn the compute-node image into
+	corresponding vm-compute-node image.
+
+etc/
+	Configuration files included in /etc in the compute image
+
+patches/
+	Some extensions need to be patched to work with Neon. This
+	directory contains such patches. They are applied to the extension
+	sources in Dockerfile.compute-node
+
+In addition to these, postgres itself, the neon postgres extension,
+and compute_ctl are built and copied into the compute image by
+Dockerfile.compute-node.

compute/etc/neon_collector.yml

@@ -0,0 +1,246 @@
+collector_name: neon_collector
+metrics:
+- metric_name: lfc_misses
+  type: gauge
+  help: 'lfc_misses'
+  key_labels:
+  values: [lfc_misses]
+  query: |
+    select lfc_value as lfc_misses from neon.neon_lfc_stats where lfc_key='file_cache_misses';
+
+- metric_name: lfc_used
+  type: gauge
+  help: 'LFC chunks used (chunk = 1MB)'
+  key_labels:
+  values: [lfc_used]
+  query: |
+    select lfc_value as lfc_used from neon.neon_lfc_stats where lfc_key='file_cache_used';
+
+- metric_name: lfc_hits
+  type: gauge
+  help: 'lfc_hits'
+  key_labels:
+  values: [lfc_hits]
+  query: |
+    select lfc_value as lfc_hits from neon.neon_lfc_stats where lfc_key='file_cache_hits';
+
+- metric_name: lfc_writes
+  type: gauge
+  help: 'lfc_writes'
+  key_labels:
+  values: [lfc_writes]
+  query: |
+    select lfc_value as lfc_writes from neon.neon_lfc_stats where lfc_key='file_cache_writes';
+
+- metric_name: lfc_cache_size_limit
+  type: gauge
+  help: 'LFC cache size limit in bytes'
+  key_labels:
+  values: [lfc_cache_size_limit]
+  query: |
+    select pg_size_bytes(current_setting('neon.file_cache_size_limit')) as lfc_cache_size_limit;
+
+- metric_name: connection_counts
+  type: gauge
+  help: 'Connection counts'
+  key_labels:
+    - datname
+    - state
+  values: [count]
+  query: |
+    select datname, state, count(*) as count from pg_stat_activity where state <> '' group by datname, state;
+
+- metric_name: pg_stats_userdb
+  type: gauge
+  help: 'Stats for several oldest non-system dbs'
+  key_labels:
+    - datname
+  value_label: kind
+  values:
+    - db_size
+    - deadlocks
+    # Rows
+    - inserted
+    - updated
+    - deleted
+  # We export stats for 10 non-system database. Without this limit
+  # it is too easy to abuse the system by creating lots of databases.
+  query: |
+    select pg_database_size(datname) as db_size, deadlocks,
+       tup_inserted as inserted, tup_updated as updated, tup_deleted as deleted,
+       datname
+     from pg_stat_database
+     where datname IN (
+       select datname
+       from pg_database
+       where datname <> 'postgres' and not datistemplate
+       order by oid
+       limit 10
+     );
+
+- metric_name: max_cluster_size
+  type: gauge
+  help: 'neon.max_cluster_size setting'
+  key_labels:
+  values: [max_cluster_size]
+  query: |
+    select setting::int as max_cluster_size from pg_settings where name = 'neon.max_cluster_size';
+
+- metric_name: db_total_size
+  type: gauge
+  help: 'Size of all databases'
+  key_labels:
+  values: [total]
+  query: |
+    select sum(pg_database_size(datname)) as total from pg_database;
+
+# DEPRECATED
+- metric_name: lfc_approximate_working_set_size
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels:
+  values: [approximate_working_set_size]
+  query: |
+    select neon.approximate_working_set_size(false) as approximate_working_set_size;
+
+- metric_name: lfc_approximate_working_set_size_windows
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels: [duration]
+  values: [size]
+  # NOTE: This is the "public" / "human-readable" version. Here, we supply a small selection
+  # of durations in a pretty-printed form.
+  query: |
+    select
+      x as duration,
+      neon.approximate_working_set_size_seconds(extract('epoch' from x::interval)::int) as size
+    from
+      (values ('5m'),('15m'),('1h')) as t (x);
+
+- metric_name: compute_current_lsn
+  type: gauge
+  help: 'Current LSN of the database'
+  key_labels:
+  values: [lsn]
+  query: |
+    select
+      case
+        when pg_catalog.pg_is_in_recovery()
+        then (pg_last_wal_replay_lsn() - '0/0')::FLOAT8
+        else (pg_current_wal_lsn() - '0/0')::FLOAT8
+      end as lsn;
+
+- metric_name: compute_receive_lsn
+  type: gauge
+  help: 'Returns the last write-ahead log location that has been received and synced to disk by streaming replication'
+  key_labels:
+  values: [lsn]
+  query: |
+    SELECT
+      CASE
+        WHEN pg_catalog.pg_is_in_recovery()
+        THEN (pg_last_wal_receive_lsn() - '0/0')::FLOAT8
+        ELSE 0
+      END AS lsn;
+
+- metric_name: replication_delay_bytes
+  type: gauge
+  help: 'Bytes between received and replayed LSN'
+  key_labels:
+  values: [replication_delay_bytes]
+  # We use a GREATEST call here because this calculation can be negative.
+  # The calculation is not atomic, meaning after we've gotten the receive
+  # LSN, the replay LSN may have advanced past the receive LSN we
+  # are using for the calculation.
+  query: |
+    SELECT GREATEST(0, pg_wal_lsn_diff(pg_last_wal_receive_lsn(), pg_last_wal_replay_lsn())) AS replication_delay_bytes;
+
+- metric_name: replication_delay_seconds
+  type: gauge
+  help: 'Time since last LSN was replayed'
+  key_labels:
+  values: [replication_delay_seconds]
+  query: |
+    SELECT
+      CASE
+        WHEN pg_last_wal_receive_lsn() = pg_last_wal_replay_lsn() THEN 0
+        ELSE GREATEST (0, EXTRACT (EPOCH FROM now() - pg_last_xact_replay_timestamp()))
+     END AS replication_delay_seconds;
+
+- metric_name: checkpoints_req
+  type: gauge
+  help: 'Number of requested checkpoints'
+  key_labels:
+  values: [checkpoints_req]
+  query: |
+    SELECT checkpoints_req FROM pg_stat_bgwriter;
+
+- metric_name: checkpoints_timed
+  type: gauge
+  help: 'Number of scheduled checkpoints'
+  key_labels:
+  values: [checkpoints_timed]
+  query: |
+    SELECT checkpoints_timed FROM pg_stat_bgwriter;
+
+- metric_name: compute_logical_snapshot_files
+  type: gauge
+  help: 'Number of snapshot files in pg_logical/snapshot'
+  key_labels:
+    - timeline_id
+  values: [num_logical_snapshot_files]
+  query: |
+    SELECT
+      (SELECT setting FROM pg_settings WHERE name = 'neon.timeline_id') AS timeline_id,
+      -- Postgres creates temporary snapshot files of the form %X-%X.snap.%d.tmp. These
+      -- temporary snapshot files are renamed to the actual snapshot files after they are
+      -- completely built. We only WAL-log the completely built snapshot files.
+      (SELECT COUNT(*) FROM pg_ls_dir('pg_logical/snapshots') AS name WHERE name LIKE '%.snap') AS num_logical_snapshot_files;
+
+# In all the below metrics, we cast LSNs to floats because Prometheus only supports floats.
+# It's probably fine because float64 can store integers from -2^53 to +2^53 exactly.
+
+# Number of slots is limited by max_replication_slots, so collecting position for all of them shouldn't be bad.
+- metric_name: logical_slot_restart_lsn
+  type: gauge
+  help: 'restart_lsn of logical slots'
+  key_labels:
+    - slot_name
+  values: [restart_lsn]
+  query: |
+    select slot_name, (restart_lsn - '0/0')::FLOAT8 as restart_lsn
+    from pg_replication_slots
+    where slot_type = 'logical';
+
+- metric_name: compute_subscriptions_count
+  type: gauge
+  help: 'Number of logical replication subscriptions grouped by enabled/disabled'
+  key_labels:
+    - enabled
+  values: [subscriptions_count]
+  query: |
+    select subenabled::text as enabled, count(*) as subscriptions_count
+    from pg_subscription
+    group by subenabled;
+
+- metric_name: retained_wal
+  type: gauge
+  help: 'Retained WAL in inactive replication slots'
+  key_labels:
+    - slot_name
+  values: [retained_wal]
+  query: |
+    SELECT slot_name, pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn)::FLOAT8 AS retained_wal
+    FROM pg_replication_slots
+    WHERE active = false;
+
+- metric_name: wal_is_lost
+  type: gauge
+  help: 'Whether or not the replication slot wal_status is lost'
+  key_labels:
+    - slot_name
+  values: [wal_is_lost]
+  query: |
+    SELECT slot_name,
+           CASE WHEN wal_status = 'lost' THEN 1 ELSE 0 END AS wal_is_lost
+    FROM pg_replication_slots;

compute/etc/neon_collector_autoscaling.yml

@@ -0,0 +1,55 @@
+collector_name: neon_collector_autoscaling
+metrics:
+- metric_name: lfc_misses
+  type: gauge
+  help: 'lfc_misses'
+  key_labels:
+  values: [lfc_misses]
+  query: |
+    select lfc_value as lfc_misses from neon.neon_lfc_stats where lfc_key='file_cache_misses';
+
+- metric_name: lfc_used
+  type: gauge
+  help: 'LFC chunks used (chunk = 1MB)'
+  key_labels:
+  values: [lfc_used]
+  query: |
+    select lfc_value as lfc_used from neon.neon_lfc_stats where lfc_key='file_cache_used';
+
+- metric_name: lfc_hits
+  type: gauge
+  help: 'lfc_hits'
+  key_labels:
+  values: [lfc_hits]
+  query: |
+    select lfc_value as lfc_hits from neon.neon_lfc_stats where lfc_key='file_cache_hits';
+
+- metric_name: lfc_writes
+  type: gauge
+  help: 'lfc_writes'
+  key_labels:
+  values: [lfc_writes]
+  query: |
+    select lfc_value as lfc_writes from neon.neon_lfc_stats where lfc_key='file_cache_writes';
+
+- metric_name: lfc_cache_size_limit
+  type: gauge
+  help: 'LFC cache size limit in bytes'
+  key_labels:
+  values: [lfc_cache_size_limit]
+  query: |
+    select pg_size_bytes(current_setting('neon.file_cache_size_limit')) as lfc_cache_size_limit;
+
+- metric_name: lfc_approximate_working_set_size_windows
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels: [duration_seconds]
+  values: [size]
+  # NOTE: This is the "internal" / "machine-readable" version. This outputs the working set
+  # size looking back 1..60 minutes, labeled with the number of minutes.
+  query: |
+    select
+      x::text as duration_seconds,
+      neon.approximate_working_set_size_seconds(x) as size
+    from
+      (select generate_series * 60 as x from generate_series(1, 60)) as t (x);

compute/etc/pgbouncer.ini

@@ -0,0 +1,17 @@
+[databases]
+*=host=localhost port=5432 auth_user=cloud_admin
+[pgbouncer]
+listen_port=6432
+listen_addr=0.0.0.0
+auth_type=scram-sha-256
+auth_user=cloud_admin
+auth_dbname=postgres
+client_tls_sslmode=disable
+server_tls_sslmode=disable
+pool_mode=transaction
+max_client_conn=10000
+default_pool_size=64
+max_prepared_statements=0
+admin_users=postgres
+unix_socket_dir=/tmp/
+unix_socket_mode=0777

compute/etc/sql_exporter.yml

@@ -0,0 +1,33 @@
+# Configuration for sql_exporter
+# Global defaults.
+global:
+  # If scrape_timeout <= 0, no timeout is set unless Prometheus provides one. The default is 10s.
+  scrape_timeout: 10s
+  # Subtracted from Prometheus' scrape_timeout to give us some headroom and prevent Prometheus from timing out first.
+  scrape_timeout_offset: 500ms
+  # Minimum interval between collector runs: by default (0s) collectors are executed on every scrape.
+  min_interval: 0s
+  # Maximum number of open connections to any one target. Metric queries will run concurrently on multiple connections,
+  # as will concurrent scrapes.
+  max_connections: 1
+  # Maximum number of idle connections to any one target. Unless you use very long collection intervals, this should
+  # always be the same as max_connections.
+  max_idle_connections: 1
+  # Maximum number of maximum amount of time a connection may be reused. Expired connections may be closed lazily before reuse.
+  # If 0, connections are not closed due to a connection's age.
+  max_connection_lifetime: 5m
+
+# The target to monitor and the collectors to execute on it.
+target:
+  # Data source name always has a URI schema that matches the driver name. In some cases (e.g. MySQL)
+  # the schema gets dropped or replaced to match the driver expected DSN format.
+  data_source_name: 'postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter'
+
+  # Collectors (referenced by name) to execute on the target.
+  # Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+  collectors: [neon_collector]
+
+# Collector files specifies a list of globs. One collector definition is read from each matching file.
+# Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+collector_files:
+  - "neon_collector.yml"

compute/etc/sql_exporter_autoscaling.yml

@@ -0,0 +1,33 @@
+# Configuration for sql_exporter for autoscaling-agent
+# Global defaults.
+global:
+  # If scrape_timeout <= 0, no timeout is set unless Prometheus provides one. The default is 10s.
+  scrape_timeout: 10s
+  # Subtracted from Prometheus' scrape_timeout to give us some headroom and prevent Prometheus from timing out first.
+  scrape_timeout_offset: 500ms
+  # Minimum interval between collector runs: by default (0s) collectors are executed on every scrape.
+  min_interval: 0s
+  # Maximum number of open connections to any one target. Metric queries will run concurrently on multiple connections,
+  # as will concurrent scrapes.
+  max_connections: 1
+  # Maximum number of idle connections to any one target. Unless you use very long collection intervals, this should
+  # always be the same as max_connections.
+  max_idle_connections: 1
+  # Maximum number of maximum amount of time a connection may be reused. Expired connections may be closed lazily before reuse.
+  # If 0, connections are not closed due to a connection's age.
+  max_connection_lifetime: 5m
+
+# The target to monitor and the collectors to execute on it.
+target:
+  # Data source name always has a URI schema that matches the driver name. In some cases (e.g. MySQL)
+  # the schema gets dropped or replaced to match the driver expected DSN format.
+  data_source_name: 'postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter_autoscaling'
+
+  # Collectors (referenced by name) to execute on the target.
+  # Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+  collectors: [neon_collector_autoscaling]
+
+# Collector files specifies a list of globs. One collector definition is read from each matching file.
+# Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+collector_files:
+  - "neon_collector_autoscaling.yml"

compute/vm-image-spec.yaml

@@ -0,0 +1,117 @@
+# Supplemental file for neondatabase/autoscaling's vm-builder, for producing the VM compute image.
+---
+commands:
+  - name: cgconfigparser
+    user: root
+    sysvInitAction: sysinit
+    shell: 'cgconfigparser -l /etc/cgconfig.conf -s 1664'
+  # restrict permissions on /neonvm/bin/resize-swap, because we grant access to compute_ctl for
+  # running it as root.
+  - name: chmod-resize-swap
+    user: root
+    sysvInitAction: sysinit
+    shell: 'chmod 711 /neonvm/bin/resize-swap'
+  - name: chmod-set-disk-quota
+    user: root
+    sysvInitAction: sysinit
+    shell: 'chmod 711 /neonvm/bin/set-disk-quota'
+  - name: pgbouncer
+    user: postgres
+    sysvInitAction: respawn
+    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini'
+  - name: postgres-exporter
+    user: nobody
+    sysvInitAction: respawn
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter'
+  - name: sql-exporter
+    user: nobody
+    sysvInitAction: respawn
+    shell: '/bin/sql_exporter -config.file=/etc/sql_exporter.yml -web.listen-address=:9399'
+  - name: sql-exporter-autoscaling
+    user: nobody
+    sysvInitAction: respawn
+    shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
+shutdownHook: |
+  su -p postgres --session-command '/usr/local/bin/pg_ctl stop -D /var/db/postgres/compute/pgdata -m fast --wait -t 10'
+files:
+  - filename: compute_ctl-sudoers
+    content: |
+      # Allow postgres user (which is what compute_ctl runs as) to run /neonvm/bin/resize-swap
+      # and /neonvm/bin/set-disk-quota as root without requiring entering a password (NOPASSWD),
+      # regardless of hostname (ALL)
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota
+  - filename: cgconfig.conf
+    content: |
+      # Configuration for cgroups in VM compute nodes
+      group neon-postgres {
+          perm {
+              admin {
+                  uid = postgres;
+              }
+              task {
+                  gid = users;
+              }
+          }
+          memory {}
+      }
+build: |
+  # Build cgroup-tools
+  #
+  # At time of writing (2023-03-14), debian bullseye has a version of cgroup-tools (technically
+  # libcgroup) that doesn't support cgroup v2 (version 0.41-11). Unfortunately, the vm-monitor
+  # requires cgroup v2, so we'll build cgroup-tools ourselves.
+  FROM debian:bullseye-slim as libcgroup-builder
+  ENV LIBCGROUP_VERSION=v2.0.3
+
+  RUN set -exu \
+      && apt update \
+      && apt install --no-install-recommends -y \
+          git \
+          ca-certificates \
+          automake \
+          cmake \
+          make \
+          gcc \
+          byacc \
+          flex \
+          libtool \
+          libpam0g-dev \
+      && git clone --depth 1 -b $LIBCGROUP_VERSION https://github.com/libcgroup/libcgroup \
+      && INSTALL_DIR="/libcgroup-install" \
+      && mkdir -p "$INSTALL_DIR/bin" "$INSTALL_DIR/include" \
+      && cd libcgroup \
+      # extracted from bootstrap.sh, with modified flags:
+      && (test -d m4 || mkdir m4) \
+      && autoreconf -fi \
+      && rm -rf autom4te.cache \
+      && CFLAGS="-O3" ./configure --prefix="$INSTALL_DIR" --sysconfdir=/etc --localstatedir=/var --enable-opaque-hierarchy="name=systemd" \
+      # actually build the thing...
+      && make install
+merge: |
+  # tweak nofile limits
+  RUN set -e \
+      && echo 'fs.file-max = 1048576' >>/etc/sysctl.conf \
+      && test ! -e /etc/security || ( \
+         echo '*    - nofile 1048576' >>/etc/security/limits.conf \
+      && echo 'root - nofile 1048576' >>/etc/security/limits.conf \
+         )
+
+  # Allow postgres user (compute_ctl) to run swap resizer.
+  # Need to install sudo in order to allow this.
+  #
+  # Also, remove the 'read' permission from group/other on /neonvm/bin/resize-swap, just to be safe.
+  RUN set -e \
+      && apt update \
+      && apt install --no-install-recommends -y \
+             sudo \
+      && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+  COPY compute_ctl-sudoers /etc/sudoers.d/compute_ctl-sudoers
+
+  COPY cgconfig.conf /etc/cgconfig.conf
+
+  RUN set -e \
+      && chmod 0644 /etc/cgconfig.conf
+
+  COPY --from=libcgroup-builder /libcgroup-install/bin/*  /usr/bin/
+  COPY --from=libcgroup-builder /libcgroup-install/lib/*  /usr/lib/
+  COPY --from=libcgroup-builder /libcgroup-install/sbin/* /usr/sbin/

compute_tools/Cargo.toml

@@ -21,6 +21,7 @@ nix.workspace = true
 notify.workspace = true
 num_cpus.workspace = true
 opentelemetry.workspace = true
+opentelemetry_sdk.workspace = true
 postgres.workspace = true
 regex.workspace = true
 serde_json.workspace = true

compute_tools/src/bin/compute_ctl.rs

@@ -44,6 +44,7 @@ use std::{thread, time::Duration};
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Arg;
+use compute_tools::disk_quota::set_disk_quota;
 use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
 use signal_hook::consts::{SIGQUIT, SIGTERM};
 use signal_hook::{consts::SIGINT, iterator::Signals};
@@ -151,6 +152,7 @@ fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
     let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
     let resize_swap_on_bind = matches.get_flag("resize-swap-on-bind");
+    let set_disk_quota_for_fs = matches.get_one::<String>("set-disk-quota-for-fs");
 
     Ok(ProcessCliResult {
         connstr,
@@ -161,6 +163,7 @@ fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
         spec_json,
         spec_path,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
     })
 }
 
@@ -173,6 +176,7 @@ struct ProcessCliResult<'clap> {
     spec_json: Option<&'clap String>,
     spec_path: Option<&'clap String>,
     resize_swap_on_bind: bool,
+    set_disk_quota_for_fs: Option<&'clap String>,
 }
 
 fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
@@ -214,7 +218,7 @@ fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
     }
     if !startup_tracing_carrier.is_empty() {
         use opentelemetry::propagation::TextMapPropagator;
-        use opentelemetry::sdk::propagation::TraceContextPropagator;
+        use opentelemetry_sdk::propagation::TraceContextPropagator;
         let guard = TraceContextPropagator::new()
             .extract(&startup_tracing_carrier)
             .attach();
@@ -293,6 +297,7 @@ fn wait_spec(
         pgbin,
         ext_remote_storage,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
         http_port,
         ..
     }: ProcessCliResult,
@@ -373,6 +378,7 @@ fn wait_spec(
         compute,
         http_port,
         resize_swap_on_bind,
+        set_disk_quota_for_fs: set_disk_quota_for_fs.cloned(),
     })
 }
 
@@ -381,6 +387,7 @@ struct WaitSpecResult {
     // passed through from ProcessCliResult
     http_port: u16,
     resize_swap_on_bind: bool,
+    set_disk_quota_for_fs: Option<String>,
 }
 
 fn start_postgres(
@@ -390,6 +397,7 @@ fn start_postgres(
         compute,
         http_port,
         resize_swap_on_bind,
+        set_disk_quota_for_fs,
     }: WaitSpecResult,
 ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
     // We got all we need, update the state.
@@ -403,6 +411,7 @@ fn start_postgres(
     );
     // before we release the mutex, fetch the swap size (if any) for later.
     let swap_size_bytes = state.pspec.as_ref().unwrap().spec.swap_size_bytes;
+    let disk_quota_bytes = state.pspec.as_ref().unwrap().spec.disk_quota_bytes;
     drop(state);
 
     // Launch remaining service threads
@@ -422,8 +431,8 @@ fn start_postgres(
         // OOM-killed during startup because swap wasn't available yet.
         match resize_swap(size_bytes) {
             Ok(()) => {
-                let size_gib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
-                info!(%size_bytes, %size_gib, "resized swap");
+                let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
+                info!(%size_bytes, %size_mib, "resized swap");
             }
             Err(err) => {
                 let err = err.context("failed to resize swap");
@@ -432,10 +441,29 @@ fn start_postgres(
                 // Mark compute startup as failed; don't try to start postgres, and report this
                 // error to the control plane when it next asks.
                 prestartup_failed = true;
-                let mut state = compute.state.lock().unwrap();
-                state.error = Some(format!("{err:?}"));
-                state.status = ComputeStatus::Failed;
-                compute.state_changed.notify_all();
+                compute.set_failed_status(err);
+                delay_exit = true;
+            }
+        }
+    }
+
+    // Set disk quota if the compute spec says so
+    if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) =
+        (disk_quota_bytes, set_disk_quota_for_fs)
+    {
+        match set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint) {
+            Ok(()) => {
+                let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
+                info!(%disk_quota_bytes, %size_mib, "set disk quota");
+            }
+            Err(err) => {
+                let err = err.context("failed to set disk quota");
+                error!("{err:#}");
+
+                // Mark compute startup as failed; don't try to start postgres, and report this
+                // error to the control plane when it next asks.
+                prestartup_failed = true;
+                compute.set_failed_status(err);
                 delay_exit = true;
             }
         }
@@ -450,16 +478,7 @@ fn start_postgres(
             Ok(pg) => Some(pg),
             Err(err) => {
                 error!("could not start the compute node: {:#}", err);
-                let mut state = compute.state.lock().unwrap();
-                state.error = Some(format!("{:?}", err));
-                state.status = ComputeStatus::Failed;
-                // Notify others that Postgres failed to start. In case of configuring the
-                // empty compute, it's likely that API handler is still waiting for compute
-                // state change. With this we will notify it that compute is in Failed state,
-                // so control plane will know about it earlier and record proper error instead
-                // of timeout.
-                compute.state_changed.notify_all();
-                drop(state); // unlock
+                compute.set_failed_status(err);
                 delay_exit = true;
                 None
             }
@@ -750,6 +769,11 @@ fn cli() -> clap::Command {
                 .long("resize-swap-on-bind")
                 .action(clap::ArgAction::SetTrue),
         )
+        .arg(
+            Arg::new("set-disk-quota-for-fs")
+                .long("set-disk-quota-for-fs")
+                .value_name("SET_DISK_QUOTA_FOR_FS")
+        )
 }
 
 /// When compute_ctl is killed, send also termination signal to sync-safekeepers

compute_tools/src/compute.rs

@@ -10,6 +10,7 @@ use std::sync::atomic::AtomicU32;
 use std::sync::atomic::Ordering;
 use std::sync::{Condvar, Mutex, RwLock};
 use std::thread;
+use std::time::Duration;
 use std::time::Instant;
 
 use anyhow::{Context, Result};
@@ -305,6 +306,13 @@ impl ComputeNode {
         self.state_changed.notify_all();
     }
 
+    pub fn set_failed_status(&self, err: anyhow::Error) {
+        let mut state = self.state.lock().unwrap();
+        state.error = Some(format!("{err:?}"));
+        state.status = ComputeStatus::Failed;
+        self.state_changed.notify_all();
+    }
+
     pub fn get_status(&self) -> ComputeStatus {
         self.state.lock().unwrap().status
     }
@@ -710,7 +718,7 @@ impl ComputeNode {
         info!("running initdb");
         let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
         Command::new(initdb_bin)
-            .args(["-D", pgdata])
+            .args(["--pgdata", pgdata])
             .output()
             .expect("cannot start initdb process");
 
@@ -1123,6 +1131,9 @@ impl ComputeNode {
     //
     // Use that as a default location and pattern, except macos where core dumps are written
     // to /cores/ directory by default.
+    //
+    // With default Linux settings, the core dump file is called just "core", so check for
+    // that too.
     pub fn check_for_core_dumps(&self) -> Result<()> {
         let core_dump_dir = match std::env::consts::OS {
             "macos" => Path::new("/cores/"),
@@ -1134,8 +1145,17 @@ impl ComputeNode {
         let files = fs::read_dir(core_dump_dir)?;
         let cores = files.filter_map(|entry| {
             let entry = entry.ok()?;
-            let _ = entry.file_name().to_str()?.strip_prefix("core.")?;
-            Some(entry.path())
+
+            let is_core_dump = match entry.file_name().to_str()? {
+                n if n.starts_with("core.") => true,
+                "core" => true,
+                _ => false,
+            };
+            if is_core_dump {
+                Some(entry.path())
+            } else {
+                None
+            }
         });
 
         // Print backtrace for each core dump
@@ -1386,6 +1406,36 @@ LIMIT 100",
         }
         Ok(remote_ext_metrics)
     }
+
+    /// Waits until current thread receives a state changed notification and
+    /// the pageserver connection strings has changed.
+    ///
+    /// 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_connstr = state
+            .pspec
+            .as_ref()
+            .expect("spec must be set")
+            .pageserver_connstr
+            .clone();
+        let mut unchanged = true;
+        let _ = self
+            .state_changed
+            .wait_timeout_while(state, duration, |s| {
+                let pageserver_connstr = &s
+                    .pspec
+                    .as_ref()
+                    .expect("spec must be set")
+                    .pageserver_connstr;
+                unchanged = pageserver_connstr == &old_pageserver_connstr;
+                unchanged
+            })
+            .unwrap();
+        if !unchanged {
+            info!("Pageserver config changed");
+        }
+    }
 }
 
 pub fn forward_termination_signal() {

compute_tools/src/configurator.rs

@@ -11,9 +11,17 @@ use crate::compute::ComputeNode;
 fn configurator_main_loop(compute: &Arc<ComputeNode>) {
     info!("waiting for reconfiguration requests");
     loop {
-        let state = compute.state.lock().unwrap();
-        let mut state = compute.state_changed.wait(state).unwrap();
+        let mut state = compute.state.lock().unwrap();
 
+        // We have to re-check the status after re-acquiring the lock because it could be that
+        // the status has changed while we were waiting for the lock, and we might not need to
+        // wait on the condition variable. Otherwise, we might end up in some soft-/deadlock, i.e.
+        // we are waiting for a condition variable that will never be signaled.
+        if state.status != ComputeStatus::ConfigurationPending {
+            state = compute.state_changed.wait(state).unwrap();
+        }
+
+        // Re-check the status after waking up
         if state.status == ComputeStatus::ConfigurationPending {
             info!("got configuration request");
             state.status = ComputeStatus::Configuration;

compute_tools/src/disk_quota.rs

@@ -0,0 +1,25 @@
+use anyhow::Context;
+
+pub const DISK_QUOTA_BIN: &str = "/neonvm/bin/set-disk-quota";
+
+/// If size_bytes is 0, it disables the quota. Otherwise, it sets filesystem quota to size_bytes.
+/// `fs_mountpoint` should point to the mountpoint of the filesystem where the quota should be set.
+pub fn set_disk_quota(size_bytes: u64, fs_mountpoint: &str) -> anyhow::Result<()> {
+    let size_kb = size_bytes / 1024;
+    // run `/neonvm/bin/set-disk-quota {size_kb} {mountpoint}`
+    let child_result = std::process::Command::new("/usr/bin/sudo")
+        .arg(DISK_QUOTA_BIN)
+        .arg(size_kb.to_string())
+        .arg(fs_mountpoint)
+        .spawn();
+
+    child_result
+        .context("spawn() failed")
+        .and_then(|mut child| child.wait().context("wait() failed"))
+        .and_then(|status| match status.success() {
+            true => Ok(()),
+            false => Err(anyhow::anyhow!("process exited with {status}")),
+        })
+        // wrap any prior error with the overall context that we couldn't run the command
+        .with_context(|| format!("could not run `/usr/bin/sudo {DISK_QUOTA_BIN}`"))
+}

compute_tools/src/lib.rs

@@ -10,6 +10,7 @@ pub mod http;
 pub mod logger;
 pub mod catalog;
 pub mod compute;
+pub mod disk_quota;
 pub mod extension_server;
 pub mod lsn_lease;
 mod migration;

compute_tools/src/logger.rs

@@ -1,4 +1,3 @@
-use tracing_opentelemetry::OpenTelemetryLayer;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::prelude::*;
 
@@ -23,8 +22,7 @@ pub fn init_tracing_and_logging(default_log_level: &str) -> anyhow::Result<()> {
         .with_writer(std::io::stderr);
 
     // Initialize OpenTelemetry
-    let otlp_layer =
-        tracing_utils::init_tracing_without_runtime("compute_ctl").map(OpenTelemetryLayer::new);
+    let otlp_layer = tracing_utils::init_tracing_without_runtime("compute_ctl");
 
     // Put it all together
     tracing_subscriber::registry()

compute_tools/src/lsn_lease.rs

@@ -57,10 +57,10 @@ fn lsn_lease_bg_task(
             .max(valid_duration / 2);
 
         info!(
-            "Succeeded, sleeping for {} seconds",
+            "Request succeeded, sleeping for {} seconds",
             sleep_duration.as_secs()
         );
-        thread::sleep(sleep_duration);
+        compute.wait_timeout_while_pageserver_connstr_unchanged(sleep_duration);
     }
 }
 
@@ -89,10 +89,7 @@ fn acquire_lsn_lease_with_retry(
                 .map(|connstr| {
                     let mut config = postgres::Config::from_str(connstr).expect("Invalid connstr");
                     if let Some(storage_auth_token) = &spec.storage_auth_token {
-                        info!("Got storage auth token from spec file");
                         config.password(storage_auth_token.clone());
-                    } else {
-                        info!("Storage auth token not set");
                     }
                     config
                 })
@@ -108,9 +105,11 @@ fn acquire_lsn_lease_with_retry(
                 bail!("Permanent error: lease could not be obtained, LSN is behind the GC cutoff");
             }
             Err(e) => {
-                warn!("Failed to acquire lsn lease: {e} (attempt {attempts}");
+                warn!("Failed to acquire lsn lease: {e} (attempt {attempts})");
 
-                thread::sleep(Duration::from_millis(retry_period_ms as u64));
+                compute.wait_timeout_while_pageserver_connstr_unchanged(Duration::from_millis(
+                    retry_period_ms as u64,
+                ));
                 retry_period_ms *= 1.5;
                 retry_period_ms = retry_period_ms.min(MAX_RETRY_PERIOD_MS);
             }

control_plane/Cargo.toml

@@ -9,7 +9,7 @@ anyhow.workspace = true
 camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
-git-version.workspace = true
+futures.workspace = true
 humantime.workspace = true
 nix.workspace = true
 once_cell.workspace = true

control_plane/src/branch_mappings.rs

@@ -0,0 +1,94 @@
+//! Branch mappings for convenience
+
+use std::collections::HashMap;
+use std::fs;
+use std::path::Path;
+
+use anyhow::{bail, Context};
+use serde::{Deserialize, Serialize};
+
+use utils::id::{TenantId, TenantTimelineId, TimelineId};
+
+/// Keep human-readable aliases in memory (and persist them to config XXX), to hide tenant/timeline hex strings from the user.
+#[derive(PartialEq, Eq, Clone, Debug, Default, Serialize, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+pub struct BranchMappings {
+    /// Default tenant ID to use with the 'neon_local' command line utility, when
+    /// --tenant_id is not explicitly specified. This comes from the branches.
+    pub default_tenant_id: Option<TenantId>,
+
+    // A `HashMap<String, HashMap<TenantId, TimelineId>>` would be more appropriate here,
+    // but deserialization into a generic toml object as `toml::Value::try_from` fails with an error.
+    // https://toml.io/en/v1.0.0 does not contain a concept of "a table inside another table".
+    pub mappings: HashMap<String, Vec<(TenantId, TimelineId)>>,
+}
+
+impl BranchMappings {
+    pub fn register_branch_mapping(
+        &mut self,
+        branch_name: String,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<()> {
+        let existing_values = self.mappings.entry(branch_name.clone()).or_default();
+
+        let existing_ids = existing_values
+            .iter()
+            .find(|(existing_tenant_id, _)| existing_tenant_id == &tenant_id);
+
+        if let Some((_, old_timeline_id)) = existing_ids {
+            if old_timeline_id == &timeline_id {
+                Ok(())
+            } else {
+                bail!("branch '{branch_name}' is already mapped to timeline {old_timeline_id}, cannot map to another timeline {timeline_id}");
+            }
+        } else {
+            existing_values.push((tenant_id, timeline_id));
+            Ok(())
+        }
+    }
+
+    pub fn get_branch_timeline_id(
+        &self,
+        branch_name: &str,
+        tenant_id: TenantId,
+    ) -> Option<TimelineId> {
+        // If it looks like a timeline ID, return it as it is
+        if let Ok(timeline_id) = branch_name.parse::<TimelineId>() {
+            return Some(timeline_id);
+        }
+
+        self.mappings
+            .get(branch_name)?
+            .iter()
+            .find(|(mapped_tenant_id, _)| mapped_tenant_id == &tenant_id)
+            .map(|&(_, timeline_id)| timeline_id)
+            .map(TimelineId::from)
+    }
+
+    pub fn timeline_name_mappings(&self) -> HashMap<TenantTimelineId, String> {
+        self.mappings
+            .iter()
+            .flat_map(|(name, tenant_timelines)| {
+                tenant_timelines.iter().map(|&(tenant_id, timeline_id)| {
+                    (TenantTimelineId::new(tenant_id, timeline_id), name.clone())
+                })
+            })
+            .collect()
+    }
+
+    pub fn persist(&self, path: &Path) -> anyhow::Result<()> {
+        let content = &toml::to_string_pretty(self)?;
+        fs::write(path, content).with_context(|| {
+            format!(
+                "Failed to write branch information into path '{}'",
+                path.display()
+            )
+        })
+    }
+
+    pub fn load(path: &Path) -> anyhow::Result<BranchMappings> {
+        let branches_file_contents = fs::read_to_string(path)?;
+        Ok(toml::from_str(branches_file_contents.as_str())?)
+    }
+}

control_plane/src/endpoint.rs

@@ -561,6 +561,7 @@ impl Endpoint {
             operation_uuid: None,
             features: self.features.clone(),
             swap_size_bytes: None,
+            disk_quota_bytes: None,
             cluster: Cluster {
                 cluster_id: None, // project ID: not used
                 name: None,       // project name: not used

control_plane/src/safekeeper.rs

@@ -113,7 +113,7 @@ impl SafekeeperNode {
 
     pub async fn start(
         &self,
-        extra_opts: Vec<String>,
+        extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
         print!(
@@ -196,7 +196,7 @@ impl SafekeeperNode {
             ]);
         }
 
-        args.extend(extra_opts);
+        args.extend_from_slice(extra_opts);
 
         background_process::start_process(
             &format!("safekeeper-{id}"),

control_plane/src/storage_controller.rs

@@ -346,7 +346,14 @@ impl StorageController {
             let pg_log_path = pg_data_path.join("postgres.log");
 
             if !tokio::fs::try_exists(&pg_data_path).await? {
-                let initdb_args = ["-D", pg_data_path.as_ref(), "--username", &username()];
+                let initdb_args = [
+                    "--pgdata",
+                    pg_data_path.as_ref(),
+                    "--username",
+                    &username(),
+                    "--no-sync",
+                    "--no-instructions",
+                ];
                 tracing::info!(
                     "Initializing storage controller database with args: {:?}",
                     initdb_args

control_plane/storcon_cli/src/main.rs

@@ -4,8 +4,8 @@ use std::{str::FromStr, time::Duration};
 use clap::{Parser, Subcommand};
 use pageserver_api::{
     controller_api::{
-        NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse, ShardSchedulingPolicy,
-        TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
+        AvailabilityZone, NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse,
+        ShardSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
     },
     models::{
         EvictionPolicy, EvictionPolicyLayerAccessThreshold, LocationConfigSecondary,
@@ -339,7 +339,7 @@ async fn main() -> anyhow::Result<()> {
                         listen_pg_port,
                         listen_http_addr,
                         listen_http_port,
-                        availability_zone_id,
+                        availability_zone_id: AvailabilityZone(availability_zone_id),
                     }),
                 )
                 .await?;

docker-compose/README.md

@@ -2,8 +2,8 @@
 # Example docker compose configuration
 
 The configuration in this directory is used for testing Neon docker images: it is
-not intended for deploying a usable system.  To run a development environment where
-you can experiment with a minature Neon system, use `cargo neon` rather than container images.
+not intended for deploying a usable system. To run a development environment where
+you can experiment with a miniature Neon system, use `cargo neon` rather than container images.
 
 This configuration does not start the storage controller, because the controller
 needs a way to reconfigure running computes, and no such thing exists in this setup.

docs/rfcs/038-independent-compute-release.md

@@ -0,0 +1,343 @@
+# Independent compute release
+
+Created at: 2024-08-30. Author: Alexey Kondratov (@ololobus)
+
+## Summary
+
+This document proposes an approach to fully independent compute release flow. It attempts to
+cover the following features:
+
+- Process is automated as much as possible to minimize human errors.
+- Compute<->storage protocol compatibility is ensured.
+- A transparent release history is available with an easy rollback strategy.
+- Although not in the scope of this document, there is a viable way to extend the proposed release
+  flow to achieve the canary and/or blue-green deployment strategies.
+
+## Motivation
+
+Previously, the compute release was tightly coupled to the storage release. This meant that once
+some storage nodes got restarted with a newer version, all new compute starts using these nodes
+automatically got a new version. Thus, two releases happen in parallel, which increases the blast
+radius and makes ownership fuzzy.
+
+Now, we practice a manual v0 independent compute release flow -- after getting a new compute release
+image and tag, we pin it region by region using Admin UI. It's better, but it still has its own flaws:
+
+1. It's a simple but fairly manual process, as you need to click through a few pages.
+2. It's prone to human errors, e.g., you could mistype or copy the wrong compute tag.
+3. We now require an additional approval in the Admin UI, which partially solves the 2.,
+   but also makes the whole process pretty annoying, as you constantly need to go back
+   and forth between two people.
+
+## Non-goals
+
+It's not the goal of this document to propose a design for some general-purpose release tool like Helm.
+The document considers how the current compute fleet is orchestrated at Neon. Even if we later
+decide to split the control plane further (e.g., introduce a separate compute controller), the proposed
+release process shouldn't change much, i.e., the releases table and API will reside in
+one of the parts.
+
+Achieving the canary and/or blue-green deploy strategies is out of the scope of this document. They
+were kept in mind, though, so it's expected that the proposed approach will lay down the foundation
+for implementing them in future iterations.
+
+## Impacted components
+
+Compute, control plane, CI, observability (some Grafana dashboards may require changes).
+
+## Prior art
+
+One of the very close examples is how Helm tracks [releases history](https://helm.sh/docs/helm/helm_history/).
+
+In the code:
+
+- [Release](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/release.go#L20-L43)
+- [Release info](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/info.go#L24-L40)
+- [Release status](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/status.go#L18-L42)
+
+TL;DR it has several important attributes:
+
+- Revision -- unique release ID/primary key. It is not the same as the application version,
+  because the same version can be deployed several times, e.g., after a newer version rollback.
+- App version -- version of the application chart/code.
+- Config -- set of overrides to the default config of the application.
+- Status -- current status of the release in the history.
+- Timestamps -- tracks when a release was created and deployed.
+
+## Proposed implementation
+
+### Separate release branch
+
+We will use a separate release branch, `release-compute`, to have a clean history for releases and commits.
+In order to avoid confusion with storage releases, we will use a different prefix for compute [git release
+tags](https://github.com/neondatabase/neon/releases) -- `release-compute-XXXX`. We will use the same tag for
+Docker images as well. The `neondatabase/compute-node-v16:release-compute-XXXX` looks longer and a bit redundant,
+but it's better to have image and git tags in sync.
+
+Currently, control plane relies on the numeric compute and storage release versions to decide on compute->storage
+compatibility. Once we implement this proposal, we should drop this code as release numbers will be completely
+independent. The only constraint we want is that it must monotonically increase within the same release branch.
+
+### Compute config/settings manifest
+
+We will create a new sub-directory `compute` and file `compute/manifest.yaml` with a structure:
+
+```yaml
+pg_settings:
+  # Common settings for primaries and secondaries of all versions.
+  common:
+    wal_log_hints: "off"
+    max_wal_size: "1024"
+
+  per_version:
+    14:
+      # Common settings for both replica and primary of version PG 14
+      common:
+        shared_preload_libraries: "neon,pg_stat_statements,extension_x"
+    15:
+      common:
+        shared_preload_libraries: "neon,pg_stat_statements,extension_x"
+      # Settings that should be applied only to
+      replica:
+        # Available only starting Postgres 15th
+        recovery_prefetch: "off"
+    # ...
+    17:
+      common:
+        # For example, if third-party `extension_x` is not yet available for PG 17
+        shared_preload_libraries: "neon,pg_stat_statements"
+      replica:
+        recovery_prefetch: "off"
+```
+
+**N.B.** Setting value should be a string with `on|off` for booleans and a number (as a string)
+without units for all numeric settings. That's how the control plane currently operates.
+
+The priority of settings will be (a higher number is a higher priority):
+
+1. Any static and hard-coded settings in the control plane
+2. `pg_settings->common`
+3. Per-version `common`
+4. Per-version `replica`
+5. Any per-user/project/endpoint overrides in the control plane
+6. Any dynamic setting calculated based on the compute size
+
+**N.B.** For simplicity, we do not do any custom logic for `shared_preload_libraries`, so it's completely
+overridden if specified on some level. Make sure that you include all necessary extensions in it when you
+do any overrides.
+
+**N.B.** There is a tricky question about what to do with custom compute image pinning we sometimes
+do for particular projects and customers. That's usually some ad-hoc work and images are based on
+the latest compute image, so it's relatively safe to assume that we could use settings from the latest compute
+release. If for some reason that's not true, and further overrides are needed, it's also possible to do
+on the project level together with pinning the image, so it's on-call/engineer/support responsibility to
+ensure that compute starts with the specified custom image. The only real risk is that compute image will get
+stale and settings from new releases will drift away, so eventually it will get something incompatible,
+but i) this is some operational issue, as we do not want stale images anyway, and ii) base settings
+receive something really new so rarely that the chance of this happening is very low. If we want to solve it completely,
+then together with pinning the image we could also pin the matching release revision in the control plane.
+
+The compute team will own the content of `compute/manifest.yaml`.
+
+### Control plane: releases table
+
+In order to store information about releases, the control plane will use a table `compute_releases` with the following
+schema:
+
+```sql
+CREATE TABLE compute_releases (
+  -- Unique release ID
+  -- N.B. Revision won't by synchronized across all regions, because all control planes are technically independent
+  -- services. We have the same situation with Helm releases as well because they could be deployed and rolled back
+  -- independently in different clusters.
+  revision BIGSERIAL PRIMARY KEY,
+  -- Numeric version of the compute image, e.g. 9057
+  version BIGINT NOT NULL,
+  -- Compute image tag, e.g. `release-9057`
+  tag TEXT NOT NULL,
+  -- Current release status. Currently, it will be a simple enum
+  -- * `deployed` -- release is deployed and used for new compute starts.
+  --                 Exactly one release can have this status at a time.
+  -- * `superseded` -- release has been replaced by a newer one.
+  -- But we can always extend it in the future when we need more statuses
+  -- for more complex deployment strategies.
+  status TEXT NOT NULL,
+  -- Any additional metadata for compute in the corresponding release
+  manifest JSONB NOT NULL,
+  -- Timestamp when release record was created in the control plane database
+  created_at TIMESTAMP NOT NULL DEFAULT now(),
+  -- Timestamp when release deployment was finished
+  deployed_at TIMESTAMP
+);
+```
+
+We keep track of the old releases not only for the sake of audit, but also because we usually have ~30% of
+old computes started using the image from one of the previous releases. Yet, when users want to reconfigure
+them without restarting, the control plane needs to know what settings are applicable to them, so we also need
+information about the previous releases that are readily available. There could be some other auxiliary info
+needed as well: supported extensions, compute flags, etc.
+
+**N.B.** Here, we can end up in an ambiguous situation when the same compute image is deployed twice, e.g.,
+it was deployed once, then rolled back, and then deployed again, potentially with a different manifest. Yet,
+we could've started some computes with the first deployment and some with the second. Thus, when we need to
+look up the manifest for the compute by its image tag, we will see two records in the table with the same tag,
+but different revision numbers. We can assume that this could happen only in case of rollbacks, so we
+can just take the latest revision for the given tag.
+
+### Control plane: management API
+
+The control plane will implement new API methods to manage releases:
+
+1. `POST /management/api/v2/compute_releases` to create a new release. With payload
+
+   ```json
+    {
+      "version": 9057,
+      "tag": "release-9057",
+      "manifest": {}
+    }
+   ```
+
+   and response
+
+   ```json
+    {
+      "revision": 53,
+      "version": 9057,
+      "tag": "release-9057",
+      "status": "deployed",
+      "manifest": {},
+      "created_at": "2024-08-15T15:52:01.0000Z",
+      "deployed_at": "2024-08-15T15:52:01.0000Z",
+    }
+   ```
+
+   Here, we can actually mix-in custom (remote) extensions metadata into the `manifest`, so that the control plane
+   will get information about all available extensions not bundled into compute image. The corresponding
+   workflow in `neondatabase/build-custom-extensions` should produce it as an artifact and make
+   it accessible to the workflow in the `neondatabase/infra`. See the complete release flow below. Doing that,
+   we put a constraint that new custom extension requires new compute release, which is good for the safety,
+   but is not exactly what we want operational-wise (we want to be able to deploy new extensions without new
+   images). Yet, it can be solved incrementally: v0 -- do not do anything with extensions at all;
+   v1 -- put them into the same manifest; v2 -- make them separate entities with their own lifecycle.
+
+   **N.B.** This method is intended to be used in CI workflows, and CI/network can be flaky. It's reasonable
+   to assume that we could retry the request several times, even though it's already succeeded. Although it's
+   not a big deal to create several identical releases one-by-one, it's better to avoid it, so the control plane
+   should check if the latest release is identical and just return `304 Not Modified` in this case.
+
+2. `POST /management/api/v2/compute_releases/rollback` to rollback to any previously deployed release. With payload
+   including the revision of the release to rollback to:
+
+   ```json
+   {
+      "revision": 52
+   }
+   ```
+
+   Rollback marks the current release as `superseded` and creates a new release with all the same data as the
+   requested revision, but with a new revision number.
+
+   This rollback API is not strictly needed, as we can just use `infra` repo workflow to deploy any
+   available tag. It's still nice to have for on-call and any urgent matters, for example, if we need
+   to rollback and GitHub is down. It's much easier to specify only the revision number vs. crafting
+   all the necessary data for the new release payload.
+
+### Compute->storage compatibility tests
+
+In order to safely release new compute versions independently from storage, we need to ensure that the currently
+deployed storage is compatible with the new compute version. Currently, we maintain backward compatibility
+in storage, but newer computes may require a newer storage version.
+
+Remote end-to-end (e2e) tests [already accept](https://github.com/neondatabase/cloud/blob/e3468d433e0d73d02b7d7e738d027f509b522408/.github/workflows/testing.yml#L43-L48)
+`storage_image_tag` and `compute_image_tag` as separate inputs. That means that we could reuse e2e tests to ensure
+compatibility between storage and compute:
+
+1. Pick the latest storage release tag and use it as `storage_image_tag`.
+2. Pick a new compute tag built in the current compute release PR and use it as `compute_image_tag`.
+   Here, we should use a temporary ECR image tag, because the final tag will be known only after the release PR is merged.
+3. Trigger e2e tests as usual.
+
+### Release flow
+
+```mermaid
+  sequenceDiagram
+
+  actor oncall as Compute on-call person
+  participant neon as neondatabase/neon
+
+  box private
+    participant cloud as neondatabase/cloud
+    participant exts as neondatabase/build-custom-extensions
+    participant infra as neondatabase/infra
+  end
+
+  box cloud
+    participant preprod as Pre-prod control plane
+    participant prod as Production control plane
+    participant k8s as Compute k8s
+  end
+
+  oncall ->> neon: Open release PR into release-compute
+
+  activate neon
+  neon ->> cloud: CI: trigger e2e compatibility tests
+  activate cloud
+  cloud -->> neon: CI: e2e tests pass
+  deactivate cloud
+  neon ->> neon: CI: pass PR checks, get approvals
+  deactivate neon
+
+  oncall ->> neon: Merge release PR into release-compute
+
+  activate neon
+  neon ->> neon: CI: pass checks, build and push images
+  neon ->> exts: CI: trigger extensions build
+  activate exts
+  exts -->> neon: CI: extensions are ready
+  deactivate exts
+  neon ->> neon: CI: create release tag
+  neon ->> infra: Trigger release workflow using the produced tag
+  deactivate neon
+
+  activate infra
+  infra ->> infra: CI: pass checks
+  infra ->> preprod: Release new compute image to pre-prod automatically <br/> POST /management/api/v2/compute_releases
+  activate preprod
+  preprod -->> infra: 200 OK
+  deactivate preprod
+
+  infra ->> infra: CI: wait for per-region production deploy approvals
+  oncall ->> infra: CI: approve deploys region by region
+  infra ->> k8s: Prewarm new compute image
+  infra ->> prod: POST /management/api/v2/compute_releases
+  activate prod
+  prod -->> infra: 200 OK
+  deactivate prod
+  deactivate infra
+```
+
+## Further work
+
+As briefly mentioned in other sections, eventually, we would like to use more complex deployment strategies.
+For example, we can pass a fraction of the total compute starts that should use the new release. Then we can
+mark the release as `partial` or `canary` and monitor its performance. If everything is fine, we can promote it
+to `deployed` status. If not, we can roll back to the previous one.
+
+## Alternatives
+
+In theory, we can try using Helm as-is:
+
+1. Write a compute Helm chart. That will actually have only some config map, which the control plane can access and read.
+   N.B. We could reuse the control plane chart as well, but then it's not a fully independent release again and even more fuzzy.
+2. The control plane will read it and start using the new compute version for new starts.
+
+Drawbacks:
+
+1. Helm releases work best if the workload is controlled by the Helm chart itself. Then you can have different
+   deployment strategies like rolling update or canary or blue/green deployments. At Neon, the compute starts are controlled
+   by control plane, so it makes it much more tricky.
+2. Releases visibility will suffer, i.e. instead of a nice table in the control plane and Admin UI, we would need to use
+   `helm` cli and/or K8s UIs like K8sLens.
+3. We do not restart all computes shortly after the new version release. This means that for some features and compatibility
+   purpose (see above) control plane may need some auxiliary info from the previous releases.

libs/compute_api/src/spec.rs

@@ -50,6 +50,16 @@ pub struct ComputeSpec {
     #[serde(default)]
     pub swap_size_bytes: Option<u64>,
 
+    /// If compute_ctl was passed `--set-disk-quota-for-fs`, a value of `Some(_)` instructs
+    /// compute_ctl to run `/neonvm/bin/set-disk-quota` with the given size and fs, when the
+    /// spec is first received.
+    ///
+    /// Both this field and `--set-disk-quota-for-fs` are required, so that the control plane's
+    /// spec generation doesn't need to be aware of the actual compute it's running on, while
+    /// guaranteeing gradual rollout of disk quota.
+    #[serde(default)]
+    pub disk_quota_bytes: Option<u64>,
+
     /// Expected cluster state at the end of transition process.
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,
@@ -268,6 +278,22 @@ pub struct GenericOption {
 /// declare a `trait` on it.
 pub type GenericOptions = Option<Vec<GenericOption>>;
 
+/// Configured the local-proxy application with the relevant JWKS and roles it should
+/// use for authorizing connect requests using JWT.
+#[derive(Clone, Debug, Deserialize, Serialize)]
+pub struct LocalProxySpec {
+    pub jwks: Vec<JwksSettings>,
+}
+
+#[derive(Clone, Debug, Deserialize, Serialize)]
+pub struct JwksSettings {
+    pub id: String,
+    pub role_names: Vec<String>,
+    pub jwks_url: String,
+    pub provider_name: String,
+    pub jwt_audience: Option<String>,
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/pageserver_api/src/config.rs

@@ -104,9 +104,6 @@ pub struct ConfigToml {
     pub image_compression: ImageCompressionAlgorithm,
     pub ephemeral_bytes_per_memory_kb: usize,
     pub l0_flush: Option<crate::models::L0FlushConfig>,
-    #[serde(skip_serializing)]
-    // TODO(https://github.com/neondatabase/neon/issues/8184): remove after this field is removed from all pageserver.toml's
-    pub compact_level0_phase1_value_access: serde::de::IgnoredAny,
     pub virtual_file_direct_io: crate::models::virtual_file::DirectIoMode,
     pub io_buffer_alignment: usize,
 }
@@ -384,7 +381,6 @@ impl Default for ConfigToml {
             image_compression: (DEFAULT_IMAGE_COMPRESSION),
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
             l0_flush: None,
-            compact_level0_phase1_value_access: Default::default(),
             virtual_file_direct_io: crate::models::virtual_file::DirectIoMode::default(),
 
             io_buffer_alignment: DEFAULT_IO_BUFFER_ALIGNMENT,

libs/pageserver_api/src/controller_api.rs

@@ -1,4 +1,5 @@
 use std::collections::{HashMap, HashSet};
+use std::fmt::Display;
 use std::str::FromStr;
 use std::time::{Duration, Instant};
 
@@ -57,7 +58,7 @@ pub struct NodeRegisterRequest {
     pub listen_http_addr: String,
     pub listen_http_port: u16,
 
-    pub availability_zone_id: String,
+    pub availability_zone_id: AvailabilityZone,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -74,10 +75,19 @@ pub struct TenantPolicyRequest {
     pub scheduling: Option<ShardSchedulingPolicy>,
 }
 
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Hash)]
+pub struct AvailabilityZone(pub String);
+
+impl Display for AvailabilityZone {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
 #[derive(Serialize, Deserialize)]
 pub struct ShardsPreferredAzsRequest {
     #[serde(flatten)]
-    pub preferred_az_ids: HashMap<TenantShardId, String>,
+    pub preferred_az_ids: HashMap<TenantShardId, AvailabilityZone>,
 }
 
 #[derive(Serialize, Deserialize)]

libs/pageserver_api/src/models.rs

@@ -37,14 +37,11 @@ use bytes::{Buf, BufMut, Bytes, BytesMut};
 /// ```mermaid
 /// stateDiagram-v2
 ///
-///     [*] --> Loading: spawn_load()
 ///     [*] --> Attaching: spawn_attach()
 ///
-///     Loading --> Activating: activate()
 ///     Attaching --> Activating: activate()
 ///     Activating --> Active: infallible
 ///
-///     Loading --> Broken: load() failure
 ///     Attaching --> Broken: attach() failure
 ///
 ///     Active --> Stopping: set_stopping(), part of shutdown & detach
@@ -68,10 +65,6 @@ use bytes::{Buf, BufMut, Bytes, BytesMut};
 )]
 #[serde(tag = "slug", content = "data")]
 pub enum TenantState {
-    /// This tenant is being loaded from local disk.
-    ///
-    /// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
-    Loading,
     /// This tenant is being attached to the pageserver.
     ///
     /// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
@@ -121,8 +114,6 @@ impl TenantState {
             // But, our attach task might still be fetching the remote timelines, etc.
             // So, return `Maybe` while Attaching, making Console wait for the attach task to finish.
             Self::Attaching | Self::Activating(ActivatingFrom::Attaching) => Maybe,
-            // tenant mgr startup distinguishes attaching from loading via marker file.
-            Self::Loading | Self::Activating(ActivatingFrom::Loading) => Attached,
             // We only reach Active after successful load / attach.
             // So, call atttachment status Attached.
             Self::Active => Attached,
@@ -191,10 +182,11 @@ impl LsnLease {
 }
 
 /// The only [`TenantState`] variants we could be `TenantState::Activating` from.
+///
+/// XXX: We used to have more variants here, but now it's just one, which makes this rather
+/// useless. Remove, once we've checked that there's no client code left that looks at this.
 #[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum ActivatingFrom {
-    /// Arrived to [`TenantState::Activating`] from [`TenantState::Loading`]
-    Loading,
     /// Arrived to [`TenantState::Activating`] from [`TenantState::Attaching`]
     Attaching,
 }
@@ -1562,11 +1554,8 @@ mod tests {
 
     #[test]
     fn tenantstatus_activating_serde() {
-        let states = [
-            TenantState::Activating(ActivatingFrom::Loading),
-            TenantState::Activating(ActivatingFrom::Attaching),
-        ];
-        let expected = "[{\"slug\":\"Activating\",\"data\":\"Loading\"},{\"slug\":\"Activating\",\"data\":\"Attaching\"}]";
+        let states = [TenantState::Activating(ActivatingFrom::Attaching)];
+        let expected = "[{\"slug\":\"Activating\",\"data\":\"Attaching\"}]";
 
         let actual = serde_json::to_string(&states).unwrap();
 
@@ -1581,13 +1570,7 @@ mod tests {
     fn tenantstatus_activating_strum() {
         // tests added, because we use these for metrics
         let examples = [
-            (line!(), TenantState::Loading, "Loading"),
             (line!(), TenantState::Attaching, "Attaching"),
-            (
-                line!(),
-                TenantState::Activating(ActivatingFrom::Loading),
-                "Activating",
-            ),
             (
                 line!(),
                 TenantState::Activating(ActivatingFrom::Attaching),

libs/postgres_backend/src/lib.rs

@@ -984,6 +984,7 @@ pub fn short_error(e: &QueryError) -> String {
 }
 
 fn log_query_error(query: &str, e: &QueryError) {
+    // If you want to change the log level of a specific error, also re-categorize it in `BasebackupQueryTimeOngoingRecording`.
     match e {
         QueryError::Disconnected(ConnectionError::Io(io_error)) => {
             if is_expected_io_error(io_error) {

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -93,9 +93,9 @@ impl Conf {
         );
         let output = self
             .new_pg_command("initdb")?
-            .arg("-D")
+            .arg("--pgdata")
             .arg(&self.datadir)
-            .args(["-U", "postgres", "--no-instructions", "--no-sync"])
+            .args(["--username", "postgres", "--no-instructions", "--no-sync"])
             .output()?;
         debug!("initdb output: {:?}", output);
         ensure!(

libs/tracing-utils/Cargo.toml

@@ -6,12 +6,14 @@ license.workspace = true
 
 [dependencies]
 hyper.workspace = true
-opentelemetry = { workspace = true, features=["rt-tokio"] }
-opentelemetry-otlp = { workspace = true, default-features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry = { workspace = true, features = ["trace"] }
+opentelemetry_sdk = { workspace = true, features = ["rt-tokio"] }
+opentelemetry-otlp = { workspace = true, default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
 opentelemetry-semantic-conventions.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tracing.workspace = true
 tracing-opentelemetry.workspace = true
+tracing-subscriber.workspace = true
 
 [dev-dependencies]
 tracing-subscriber.workspace = true    # For examples in docs

libs/tracing-utils/src/lib.rs

@@ -10,7 +10,6 @@
 //!
 //! ```rust,no_run
 //! use tracing_subscriber::prelude::*;
-//! use tracing_opentelemetry::OpenTelemetryLayer;
 //!
 //! #[tokio::main]
 //! async fn main() {
@@ -22,7 +21,7 @@
 //!         .with_writer(std::io::stderr);
 //!
 //!     // Initialize OpenTelemetry. Exports tracing spans as OpenTelemetry traces
-//!     let otlp_layer = tracing_utils::init_tracing("my_application").await.map(OpenTelemetryLayer::new);
+//!     let otlp_layer = tracing_utils::init_tracing("my_application").await;
 //!
 //!     // Put it all together
 //!     tracing_subscriber::registry()
@@ -35,15 +34,15 @@
 #![deny(unsafe_code)]
 #![deny(clippy::undocumented_unsafe_blocks)]
 
-use opentelemetry::sdk::Resource;
-use opentelemetry::KeyValue;
-use opentelemetry_otlp::WithExportConfig;
-use opentelemetry_otlp::{OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_EXPORTER_OTLP_TRACES_ENDPOINT};
-
-pub use tracing_opentelemetry::OpenTelemetryLayer;
-
 pub mod http;
 
+use opentelemetry::trace::TracerProvider;
+use opentelemetry::KeyValue;
+use opentelemetry_sdk::Resource;
+use tracing::Subscriber;
+use tracing_subscriber::registry::LookupSpan;
+use tracing_subscriber::Layer;
+
 /// Set up OpenTelemetry exporter, using configuration from environment variables.
 ///
 /// `service_name` is set as the OpenTelemetry 'service.name' resource (see
@@ -71,7 +70,10 @@ pub mod http;
 ///
 /// This doesn't block, but is marked as 'async' to hint that this must be called in
 /// asynchronous execution context.
-pub async fn init_tracing(service_name: &str) -> Option<opentelemetry::sdk::trace::Tracer> {
+pub async fn init_tracing<S>(service_name: &str) -> Option<impl Layer<S>>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
     if std::env::var("OTEL_SDK_DISABLED") == Ok("true".to_string()) {
         return None;
     };
@@ -80,9 +82,10 @@ pub async fn init_tracing(service_name: &str) -> Option<opentelemetry::sdk::trac
 
 /// Like `init_tracing`, but creates a separate tokio Runtime for the tracing
 /// tasks.
-pub fn init_tracing_without_runtime(
-    service_name: &str,
-) -> Option<opentelemetry::sdk::trace::Tracer> {
+pub fn init_tracing_without_runtime<S>(service_name: &str) -> Option<impl Layer<S>>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
     if std::env::var("OTEL_SDK_DISABLED") == Ok("true".to_string()) {
         return None;
     };
@@ -113,54 +116,36 @@ pub fn init_tracing_without_runtime(
     Some(init_tracing_internal(service_name.to_string()))
 }
 
-fn init_tracing_internal(service_name: String) -> opentelemetry::sdk::trace::Tracer {
-    // Set up exporter from the OTEL_EXPORTER_* environment variables
-    let mut exporter = opentelemetry_otlp::new_exporter().http().with_env();
+fn init_tracing_internal<S>(service_name: String) -> impl Layer<S>
+where
+    S: Subscriber + for<'span> LookupSpan<'span>,
+{
+    // Sets up exporter from the OTEL_EXPORTER_* environment variables.
+    let exporter = opentelemetry_otlp::new_exporter().http();
 
-    // XXX opentelemetry-otlp v0.18.0 has a bug in how it uses the
-    // OTEL_EXPORTER_OTLP_ENDPOINT env variable. According to the
-    // OpenTelemetry spec at
-    // <https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/protocol/exporter.md#endpoint-urls-for-otlphttp>,
-    // the full exporter URL is formed by appending "/v1/traces" to the value
-    // of OTEL_EXPORTER_OTLP_ENDPOINT. However, opentelemetry-otlp only does
-    // that with the grpc-tonic exporter. Other exporters, like the HTTP
-    // exporter, use the URL from OTEL_EXPORTER_OTLP_ENDPOINT as is, without
-    // appending "/v1/traces".
-    //
-    // See https://github.com/open-telemetry/opentelemetry-rust/pull/950
-    //
-    // Work around that by checking OTEL_EXPORTER_OTLP_ENDPOINT, and setting
-    // the endpoint url with the "/v1/traces" path ourselves. If the bug is
-    // fixed in a later version, we can remove this code. But if we don't
-    // remember to remove this, it won't do any harm either, as the crate will
-    // just ignore the OTEL_EXPORTER_OTLP_ENDPOINT setting when the endpoint
-    // is set directly with `with_endpoint`.
-    if std::env::var(OTEL_EXPORTER_OTLP_TRACES_ENDPOINT).is_err() {
-        if let Ok(mut endpoint) = std::env::var(OTEL_EXPORTER_OTLP_ENDPOINT) {
-            if !endpoint.ends_with('/') {
-                endpoint.push('/');
-            }
-            endpoint.push_str("v1/traces");
-            exporter = exporter.with_endpoint(endpoint);
-        }
-    }
+    // TODO: opentelemetry::global::set_error_handler() with custom handler that
+    //       bypasses default tracing layers, but logs regular looking log
+    //       messages.
 
     // Propagate trace information in the standard W3C TraceContext format.
     opentelemetry::global::set_text_map_propagator(
-        opentelemetry::sdk::propagation::TraceContextPropagator::new(),
+        opentelemetry_sdk::propagation::TraceContextPropagator::new(),
     );
 
-    opentelemetry_otlp::new_pipeline()
+    let tracer = opentelemetry_otlp::new_pipeline()
         .tracing()
         .with_exporter(exporter)
-        .with_trace_config(
-            opentelemetry::sdk::trace::config().with_resource(Resource::new(vec![KeyValue::new(
+        .with_trace_config(opentelemetry_sdk::trace::Config::default().with_resource(
+            Resource::new(vec![KeyValue::new(
                 opentelemetry_semantic_conventions::resource::SERVICE_NAME,
                 service_name,
-            )])),
-        )
-        .install_batch(opentelemetry::runtime::Tokio)
+            )]),
+        ))
+        .install_batch(opentelemetry_sdk::runtime::Tokio)
         .expect("could not initialize opentelemetry exporter")
+        .tracer("global");
+
+    tracing_opentelemetry::layer().with_tracer(tracer)
 }
 
 // Shutdown trace pipeline gracefully, so that it has a chance to send any

libs/utils/Cargo.toml

@@ -19,6 +19,7 @@ bincode.workspace = true
 bytes.workspace = true
 camino.workspace = true
 chrono.workspace = true
+git-version.workspace = true
 hex = { workspace = true, features = ["serde"] }
 humantime.workspace = true
 hyper = { workspace = true, features = ["full"] }

libs/utils/src/http/error.rs

@@ -82,7 +82,7 @@ impl ApiError {
                 StatusCode::INTERNAL_SERVER_ERROR,
             ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
-                err.to_string(),
+                format!("{err:#}"), // use alternative formatting so that we give the cause without backtrace
                 StatusCode::INTERNAL_SERVER_ERROR,
             ),
         }

libs/utils/src/leaky_bucket.rs

@@ -21,7 +21,13 @@
 //!
 //! Another explaination can be found here: <https://brandur.org/rate-limiting>
 
-use std::{sync::Mutex, time::Duration};
+use std::{
+    sync::{
+        atomic::{AtomicU64, Ordering},
+        Mutex,
+    },
+    time::Duration,
+};
 
 use tokio::{sync::Notify, time::Instant};
 
@@ -128,6 +134,7 @@ impl LeakyBucketState {
 
 pub struct RateLimiter {
     pub config: LeakyBucketConfig,
+    pub sleep_counter: AtomicU64,
     pub state: Mutex<LeakyBucketState>,
     /// a queue to provide this fair ordering.
     pub queue: Notify,
@@ -144,6 +151,7 @@ impl Drop for Requeue<'_> {
 impl RateLimiter {
     pub fn with_initial_tokens(config: LeakyBucketConfig, initial_tokens: f64) -> Self {
         RateLimiter {
+            sleep_counter: AtomicU64::new(0),
             state: Mutex::new(LeakyBucketState::with_initial_tokens(
                 &config,
                 initial_tokens,
@@ -163,15 +171,16 @@ impl RateLimiter {
 
     /// returns true if we did throttle
     pub async fn acquire(&self, count: usize) -> bool {
-        let mut throttled = false;
-
         let start = tokio::time::Instant::now();
 
+        let start_count = self.sleep_counter.load(Ordering::Acquire);
+        let mut end_count = start_count;
+
         // wait until we are the first in the queue
         let mut notified = std::pin::pin!(self.queue.notified());
         if !notified.as_mut().enable() {
-            throttled = true;
             notified.await;
+            end_count = self.sleep_counter.load(Ordering::Acquire);
         }
 
         // notify the next waiter in the queue when we are done.
@@ -184,9 +193,22 @@ impl RateLimiter {
                 .unwrap()
                 .add_tokens(&self.config, start, count as f64);
             match res {
-                Ok(()) => return throttled,
+                Ok(()) => return end_count > start_count,
                 Err(ready_at) => {
-                    throttled = true;
+                    struct Increment<'a>(&'a AtomicU64);
+
+                    impl Drop for Increment<'_> {
+                        fn drop(&mut self) {
+                            self.0.fetch_add(1, Ordering::AcqRel);
+                        }
+                    }
+
+                    // increment the counter after we finish sleeping (or cancel this task).
+                    // this ensures that tasks that have already started the acquire will observe
+                    // the new sleep count when they are allowed to resume on the notify.
+                    let _inc = Increment(&self.sleep_counter);
+                    end_count += 1;
+
                     tokio::time::sleep_until(ready_at).await;
                 }
             }

libs/utils/src/lib.rs

@@ -92,6 +92,10 @@ pub mod toml_edit_ext;
 
 pub mod circuit_breaker;
 
+// Re-export used in macro. Avoids adding git-version as dep in target crates.
+#[doc(hidden)]
+pub use git_version;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:
@@ -131,7 +135,7 @@ macro_rules! project_git_version {
     ($const_identifier:ident) => {
         // this should try GIT_VERSION first only then git_version::git_version!
         const $const_identifier: &::core::primitive::str = {
-            const __COMMIT_FROM_GIT: &::core::primitive::str = git_version::git_version! {
+            const __COMMIT_FROM_GIT: &::core::primitive::str = $crate::git_version::git_version! {
                 prefix = "",
                 fallback = "unknown",
                 args = ["--abbrev=40", "--always", "--dirty=-modified"] // always use full sha

pageserver/Cargo.toml

@@ -27,7 +27,6 @@ crc32c.workspace = true
 either.workspace = true
 fail.workspace = true
 futures.workspace = true
-git-version.workspace = true
 hex.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true

pageserver/benches/bench_walredo.rs

@@ -1,7 +1,7 @@
 //! Quantify a single walredo manager's throughput under N concurrent callers.
 //!
 //! The benchmark implementation ([`bench_impl`]) is parametrized by
-//! - `redo_work` => [`Request::short_request`] or [`Request::medium_request`]
+//! - `redo_work` => an async closure that takes a `PostgresRedoManager` and performs one redo
 //! - `n_redos` => number of times the benchmark shell execute the `redo_work`
 //! - `nclients` => number of clients (more on this shortly).
 //!
@@ -10,7 +10,7 @@
 //! Each task executes the `redo_work` `n_redos/nclients` times.
 //!
 //! We exercise the following combinations:
-//! - `redo_work = short / medium``
+//! - `redo_work = ping / short / medium``
 //! - `nclients = [1, 2, 4, 8, 16, 32, 64, 128]`
 //!
 //! We let `criterion` determine the `n_redos` using `iter_custom`.
@@ -27,33 +27,43 @@
 //!
 //! # Reference Numbers
 //!
-//! 2024-04-15 on i3en.3xlarge
+//! 2024-09-18 on im4gn.2xlarge
 //!
 //! ```text
-//! short/1           time:   [24.584 µs 24.737 µs 24.922 µs]
-//! short/2           time:   [33.479 µs 33.660 µs 33.888 µs]
-//! short/4           time:   [42.713 µs 43.046 µs 43.440 µs]
-//! short/8           time:   [71.814 µs 72.478 µs 73.240 µs]
-//! short/16          time:   [132.73 µs 134.45 µs 136.22 µs]
-//! short/32          time:   [258.31 µs 260.73 µs 263.27 µs]
-//! short/64          time:   [511.61 µs 514.44 µs 517.51 µs]
-//! short/128         time:   [992.64 µs 998.23 µs 1.0042 ms]
-//! medium/1          time:   [110.11 µs 110.50 µs 110.96 µs]
-//! medium/2          time:   [153.06 µs 153.85 µs 154.99 µs]
-//! medium/4          time:   [317.51 µs 319.92 µs 322.85 µs]
-//! medium/8          time:   [638.30 µs 644.68 µs 652.12 µs]
-//! medium/16         time:   [1.2651 ms 1.2773 ms 1.2914 ms]
-//! medium/32         time:   [2.5117 ms 2.5410 ms 2.5720 ms]
-//! medium/64         time:   [4.8088 ms 4.8555 ms 4.9047 ms]
-//! medium/128        time:   [8.8311 ms 8.9849 ms 9.1263 ms]
+//! ping/1                  time:   [21.789 µs 21.918 µs 22.078 µs]
+//! ping/2                  time:   [27.686 µs 27.812 µs 27.970 µs]
+//! ping/4                  time:   [35.468 µs 35.671 µs 35.926 µs]
+//! ping/8                  time:   [59.682 µs 59.987 µs 60.363 µs]
+//! ping/16                 time:   [101.79 µs 102.37 µs 103.08 µs]
+//! ping/32                 time:   [184.18 µs 185.15 µs 186.36 µs]
+//! ping/64                 time:   [349.86 µs 351.45 µs 353.47 µs]
+//! ping/128                time:   [684.53 µs 687.98 µs 692.17 µs]
+//! short/1                 time:   [31.833 µs 32.126 µs 32.428 µs]
+//! short/2                 time:   [35.558 µs 35.756 µs 35.992 µs]
+//! short/4                 time:   [44.850 µs 45.138 µs 45.484 µs]
+//! short/8                 time:   [65.985 µs 66.379 µs 66.853 µs]
+//! short/16                time:   [127.06 µs 127.90 µs 128.87 µs]
+//! short/32                time:   [252.98 µs 254.70 µs 256.73 µs]
+//! short/64                time:   [497.13 µs 499.86 µs 503.26 µs]
+//! short/128               time:   [987.46 µs 993.45 µs 1.0004 ms]
+//! medium/1                time:   [137.91 µs 138.55 µs 139.35 µs]
+//! medium/2                time:   [192.00 µs 192.91 µs 194.07 µs]
+//! medium/4                time:   [389.62 µs 391.55 µs 394.01 µs]
+//! medium/8                time:   [776.80 µs 780.33 µs 784.77 µs]
+//! medium/16               time:   [1.5323 ms 1.5383 ms 1.5459 ms]
+//! medium/32               time:   [3.0120 ms 3.0226 ms 3.0350 ms]
+//! medium/64               time:   [5.7405 ms 5.7787 ms 5.8166 ms]
+//! medium/128              time:   [10.412 ms 10.574 ms 10.718 ms]
 //! ```
 
 use anyhow::Context;
 use bytes::{Buf, Bytes};
 use criterion::{BenchmarkId, Criterion};
+use once_cell::sync::Lazy;
 use pageserver::{config::PageServerConf, walrecord::NeonWalRecord, walredo::PostgresRedoManager};
 use pageserver_api::{key::Key, shard::TenantShardId};
 use std::{
+    future::Future,
     sync::Arc,
     time::{Duration, Instant},
 };
@@ -61,40 +71,59 @@ use tokio::{sync::Barrier, task::JoinSet};
 use utils::{id::TenantId, lsn::Lsn};
 
 fn bench(c: &mut Criterion) {
-    {
-        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-        for nclients in nclients {
-            let mut group = c.benchmark_group("short");
-            group.bench_with_input(
-                BenchmarkId::from_parameter(nclients),
-                &nclients,
-                |b, nclients| {
-                    let redo_work = Arc::new(Request::short_input());
-                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
-                },
-            );
-        }
-    }
-    {
-        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-        for nclients in nclients {
-            let mut group = c.benchmark_group("medium");
-            group.bench_with_input(
-                BenchmarkId::from_parameter(nclients),
-                &nclients,
-                |b, nclients| {
-                    let redo_work = Arc::new(Request::medium_input());
-                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
-                },
-            );
-        }
+    macro_rules! bench_group {
+        ($name:expr, $redo_work:expr) => {{
+            let name: &str = $name;
+            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+            for nclients in nclients {
+                let mut group = c.benchmark_group(name);
+                group.bench_with_input(
+                    BenchmarkId::from_parameter(nclients),
+                    &nclients,
+                    |b, nclients| {
+                        b.iter_custom(|iters| bench_impl($redo_work, iters, *nclients));
+                    },
+                );
+            }
+        }};
     }
+    //
+    // benchmark the protocol implementation
+    //
+    let pg_version = 14;
+    bench_group!(
+        "ping",
+        Arc::new(move |mgr: Arc<PostgresRedoManager>| async move {
+            let _: () = mgr.ping(pg_version).await.unwrap();
+        })
+    );
+    //
+    // benchmarks with actual record redo
+    //
+    let make_redo_work = |req: &'static Request| {
+        Arc::new(move |mgr: Arc<PostgresRedoManager>| async move {
+            let page = req.execute(&mgr).await.unwrap();
+            assert_eq!(page.remaining(), 8192);
+        })
+    };
+    bench_group!("short", {
+        static REQUEST: Lazy<Request> = Lazy::new(Request::short_input);
+        make_redo_work(&REQUEST)
+    });
+    bench_group!("medium", {
+        static REQUEST: Lazy<Request> = Lazy::new(Request::medium_input);
+        make_redo_work(&REQUEST)
+    });
 }
 criterion::criterion_group!(benches, bench);
 criterion::criterion_main!(benches);
 
 // Returns the sum of each client's wall-clock time spent executing their share of the n_redos.
-fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration {
+fn bench_impl<F, Fut>(redo_work: Arc<F>, n_redos: u64, nclients: u64) -> Duration
+where
+    F: Fn(Arc<PostgresRedoManager>) -> Fut + Send + Sync + 'static,
+    Fut: Future<Output = ()> + Send + 'static,
+{
     let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
     let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
@@ -135,17 +164,20 @@ fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration
     })
 }
 
-async fn client(
+async fn client<F, Fut>(
     mgr: Arc<PostgresRedoManager>,
     start: Arc<Barrier>,
-    redo_work: Arc<Request>,
+    redo_work: Arc<F>,
     n_redos: u64,
-) -> Duration {
+) -> Duration
+where
+    F: Fn(Arc<PostgresRedoManager>) -> Fut + Send + Sync + 'static,
+    Fut: Future<Output = ()> + Send + 'static,
+{
     start.wait().await;
     let start = Instant::now();
     for _ in 0..n_redos {
-        let page = redo_work.execute(&mgr).await.unwrap();
-        assert_eq!(page.remaining(), 8192);
+        redo_work(Arc::clone(&mgr)).await;
         // The real pageserver will rarely if ever do 2 walredos in a row without
         // yielding to the executor.
         tokio::task::yield_now().await;

pageserver/client/src/mgmt_api.rs

@@ -432,7 +432,7 @@ impl Client {
             self.mgmt_api_endpoint
         );
 
-        self.request(Method::POST, &uri, req)
+        self.request(Method::PUT, &uri, req)
             .await?
             .json()
             .await
@@ -736,4 +736,22 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn timeline_init_lsn_lease(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        lsn: Lsn,
+    ) -> Result<LsnLease> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/lsn_lease",
+            self.mgmt_api_endpoint,
+        );
+
+        self.request(Method::POST, &uri, LsnLeaseRequest { lsn })
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/compaction/Cargo.toml

@@ -12,7 +12,6 @@ anyhow.workspace = true
 async-stream.workspace = true
 clap = { workspace = true, features = ["string"] }
 futures.workspace = true
-git-version.workspace = true
 itertools.workspace = true
 once_cell.workspace = true
 pageserver_api.workspace = true

pageserver/ctl/Cargo.toml

@@ -10,7 +10,6 @@ license.workspace = true
 anyhow.workspace = true
 camino.workspace = true
 clap = { workspace = true, features = ["string"] }
-git-version.workspace = true
 humantime.workspace = true
 pageserver = { path = ".." }
 pageserver_api.workspace = true

pageserver/src/bin/pageserver.rs

@@ -15,7 +15,7 @@ use clap::{Arg, ArgAction, Command};
 
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
 use pageserver::config::PageserverIdentity;
-use pageserver::control_plane_client::ControlPlaneClient;
+use pageserver::controller_upcall_client::ControllerUpcallClient;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::{COMPUTE_REQUEST_RUNTIME, WALRECEIVER_RUNTIME};
@@ -396,7 +396,7 @@ fn start_pageserver(
     // Set up deletion queue
     let (deletion_queue, deletion_workers) = DeletionQueue::new(
         remote_storage.clone(),
-        ControlPlaneClient::new(conf, &shutdown_pageserver),
+        ControllerUpcallClient::new(conf, &shutdown_pageserver),
         conf,
     );
     if let Some(deletion_workers) = deletion_workers {

pageserver/src/config.rs

@@ -324,7 +324,6 @@ impl PageServerConf {
             max_vectored_read_bytes,
             image_compression,
             ephemeral_bytes_per_memory_kb,
-            compact_level0_phase1_value_access: _,
             l0_flush,
             virtual_file_direct_io,
             concurrent_tenant_warmup,
@@ -535,16 +534,6 @@ mod tests {
             .expect("parse_and_validate");
     }
 
-    #[test]
-    fn test_compactl0_phase1_access_mode_is_ignored_silently() {
-        let input = indoc::indoc! {r#"
-            [compact_level0_phase1_value_access]
-            mode = "streaming-kmerge"
-            validate = "key-lsn-value"
-        "#};
-        toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(input).unwrap();
-    }
-
     /// If there's a typo in the pageserver config, we'd rather catch that typo
     /// and fail pageserver startup than silently ignoring the typo, leaving whoever
     /// made it in the believe that their config change is effective.

pageserver/src/controller_upcall_client.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 
 use futures::Future;
 use pageserver_api::{
-    controller_api::NodeRegisterRequest,
+    controller_api::{AvailabilityZone, NodeRegisterRequest},
     shard::TenantShardId,
     upcall_api::{
         ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
@@ -17,9 +17,12 @@ use utils::{backoff, failpoint_support, generation::Generation, id::NodeId};
 use crate::{config::PageServerConf, virtual_file::on_fatal_io_error};
 use pageserver_api::config::NodeMetadata;
 
-/// The Pageserver's client for using the control plane API: this is a small subset
-/// of the overall control plane API, for dealing with generations (see docs/rfcs/025-generation-numbers.md)
-pub struct ControlPlaneClient {
+/// The Pageserver's client for using the storage controller upcall API: this is a small API
+/// for dealing with generations (see docs/rfcs/025-generation-numbers.md).
+///
+/// The server presenting this API may either be the storage controller or some other
+/// service (such as the Neon control plane) providing a store of generation numbers.
+pub struct ControllerUpcallClient {
     http_client: reqwest::Client,
     base_url: Url,
     node_id: NodeId,
@@ -45,7 +48,7 @@ pub trait ControlPlaneGenerationsApi {
     ) -> impl Future<Output = Result<HashMap<TenantShardId, bool>, RetryForeverError>> + Send;
 }
 
-impl ControlPlaneClient {
+impl ControllerUpcallClient {
     /// A None return value indicates that the input `conf` object does not have control
     /// plane API enabled.
     pub fn new(conf: &'static PageServerConf, cancel: &CancellationToken) -> Option<Self> {
@@ -114,7 +117,7 @@ impl ControlPlaneClient {
     }
 }
 
-impl ControlPlaneGenerationsApi for ControlPlaneClient {
+impl ControlPlaneGenerationsApi for ControllerUpcallClient {
     /// Block until we get a successful response, or error out if we are shut down
     async fn re_attach(
         &self,
@@ -148,10 +151,10 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
                             .and_then(|jv| jv.as_str().map(|str| str.to_owned()));
 
                         match az_id_from_metadata {
-                            Some(az_id) => Some(az_id),
+                            Some(az_id) => Some(AvailabilityZone(az_id)),
                             None => {
                                 tracing::warn!("metadata.json does not contain an 'availability_zone_id' field");
-                                conf.availability_zone.clone()
+                                conf.availability_zone.clone().map(AvailabilityZone)
                             }
                         }
                     };
@@ -216,29 +219,38 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
             .join("validate")
             .expect("Failed to build validate path");
 
-        let request = ValidateRequest {
-            tenants: tenants
-                .into_iter()
-                .map(|(id, gen)| ValidateRequestTenant {
-                    id,
-                    gen: gen
-                        .into()
-                        .expect("Generation should always be valid for a Tenant doing deletions"),
-                })
-                .collect(),
-        };
+        // When sending validate requests, break them up into chunks so that we
+        // avoid possible edge cases of generating any HTTP requests that
+        // require database I/O across many thousands of tenants.
+        let mut result: HashMap<TenantShardId, bool> = HashMap::with_capacity(tenants.len());
+        for tenant_chunk in (tenants).chunks(128) {
+            let request = ValidateRequest {
+                tenants: tenant_chunk
+                    .iter()
+                    .map(|(id, generation)| ValidateRequestTenant {
+                        id: *id,
+                        gen: (*generation).into().expect(
+                            "Generation should always be valid for a Tenant doing deletions",
+                        ),
+                    })
+                    .collect(),
+            };
 
-        failpoint_support::sleep_millis_async!("control-plane-client-validate-sleep", &self.cancel);
-        if self.cancel.is_cancelled() {
-            return Err(RetryForeverError::ShuttingDown);
+            failpoint_support::sleep_millis_async!(
+                "control-plane-client-validate-sleep",
+                &self.cancel
+            );
+            if self.cancel.is_cancelled() {
+                return Err(RetryForeverError::ShuttingDown);
+            }
+
+            let response: ValidateResponse =
+                self.retry_http_forever(&re_attach_path, request).await?;
+            for rt in response.tenants {
+                result.insert(rt.id, rt.valid);
+            }
         }
 
-        let response: ValidateResponse = self.retry_http_forever(&re_attach_path, request).await?;
-
-        Ok(response
-            .tenants
-            .into_iter()
-            .map(|rt| (rt.id, rt.valid))
-            .collect())
+        Ok(result.into_iter().collect())
     }
 }

pageserver/src/deletion_queue.rs

@@ -6,7 +6,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
 
-use crate::control_plane_client::ControlPlaneGenerationsApi;
+use crate::controller_upcall_client::ControlPlaneGenerationsApi;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::remote_timeline_client::remote_timeline_path;
@@ -622,7 +622,7 @@ impl DeletionQueue {
     /// If remote_storage is None, then the returned workers will also be None.
     pub fn new<C>(
         remote_storage: GenericRemoteStorage,
-        control_plane_client: Option<C>,
+        controller_upcall_client: Option<C>,
         conf: &'static PageServerConf,
     ) -> (Self, Option<DeletionQueueWorkers<C>>)
     where
@@ -662,7 +662,7 @@ impl DeletionQueue {
                     conf,
                     backend_rx,
                     executor_tx,
-                    control_plane_client,
+                    controller_upcall_client,
                     lsn_table.clone(),
                     cancel.clone(),
                 ),
@@ -704,7 +704,7 @@ mod test {
     use tokio::task::JoinHandle;
 
     use crate::{
-        control_plane_client::RetryForeverError,
+        controller_upcall_client::RetryForeverError,
         repository::Key,
         tenant::{harness::TenantHarness, storage_layer::DeltaLayerName},
     };

pageserver/src/deletion_queue/validator.rs

@@ -25,8 +25,8 @@ use tracing::info;
 use tracing::warn;
 
 use crate::config::PageServerConf;
-use crate::control_plane_client::ControlPlaneGenerationsApi;
-use crate::control_plane_client::RetryForeverError;
+use crate::controller_upcall_client::ControlPlaneGenerationsApi;
+use crate::controller_upcall_client::RetryForeverError;
 use crate::metrics;
 use crate::virtual_file::MaybeFatalIo;
 
@@ -61,7 +61,7 @@ where
     tx: tokio::sync::mpsc::Sender<DeleterMessage>,
 
     // Client for calling into control plane API for validation of deletes
-    control_plane_client: Option<C>,
+    controller_upcall_client: Option<C>,
 
     // DeletionLists which are waiting generation validation.  Not safe to
     // execute until [`validate`] has processed them.
@@ -94,7 +94,7 @@ where
         conf: &'static PageServerConf,
         rx: tokio::sync::mpsc::Receiver<ValidatorQueueMessage>,
         tx: tokio::sync::mpsc::Sender<DeleterMessage>,
-        control_plane_client: Option<C>,
+        controller_upcall_client: Option<C>,
         lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
         cancel: CancellationToken,
     ) -> Self {
@@ -102,7 +102,7 @@ where
             conf,
             rx,
             tx,
-            control_plane_client,
+            controller_upcall_client,
             lsn_table,
             pending_lists: Vec::new(),
             validated_lists: Vec::new(),
@@ -145,8 +145,8 @@ where
             return Ok(());
         }
 
-        let tenants_valid = if let Some(control_plane_client) = &self.control_plane_client {
-            match control_plane_client
+        let tenants_valid = if let Some(controller_upcall_client) = &self.controller_upcall_client {
+            match controller_upcall_client
                 .validate(tenant_generations.iter().map(|(k, v)| (*k, *v)).collect())
                 .await
             {

pageserver/src/http/routes.rs

@@ -56,6 +56,7 @@ use utils::http::endpoint::request_span;
 use utils::http::request::must_parse_query_param;
 use utils::http::request::{get_request_param, must_get_query_param, parse_query_param};
 
+use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::DeletionQueueClient;
 use crate::pgdatadir_mapping::LsnForTimestamp;
@@ -80,7 +81,6 @@ use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::Timeline;
 use crate::tenant::GetTimelineError;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
-use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
 use pageserver_api::models::{
     StatusResponse, TenantConfigRequest, TenantInfo, TimelineCreateRequest, TimelineGcRequest,
@@ -589,6 +589,10 @@ async fn timeline_create_handler(
                 StatusCode::SERVICE_UNAVAILABLE,
                 HttpErrorBody::from_msg(e.to_string()),
             ),
+            Err(e @ tenant::CreateTimelineError::AncestorArchived) => json_response(
+                StatusCode::NOT_ACCEPTABLE,
+                HttpErrorBody::from_msg(e.to_string()),
+            ),
             Err(tenant::CreateTimelineError::ShuttingDown) => json_response(
                 StatusCode::SERVICE_UNAVAILABLE,
                 HttpErrorBody::from_msg("tenant shutting down".to_string()),
@@ -820,7 +824,7 @@ async fn get_lsn_by_timestamp_handler(
 
     let lease = if with_lease {
         timeline
-            .make_lsn_lease(lsn, timeline.get_lsn_lease_length_for_ts(), &ctx)
+            .init_lsn_lease(lsn, timeline.get_lsn_lease_length_for_ts(), &ctx)
             .inspect_err(|_| {
                 warn!("fail to grant a lease to {}", lsn);
             })
@@ -1688,9 +1692,18 @@ async fn lsn_lease_handler(
     let timeline =
         active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
             .await?;
-    let result = timeline
-        .make_lsn_lease(lsn, timeline.get_lsn_lease_length(), &ctx)
-        .map_err(|e| ApiError::InternalServerError(e.context("lsn lease http handler")))?;
+
+    let result = async {
+        timeline
+            .init_lsn_lease(lsn, timeline.get_lsn_lease_length(), &ctx)
+            .map_err(|e| {
+                ApiError::InternalServerError(
+                    e.context(format!("invalid lsn lease request at {lsn}")),
+                )
+            })
+    }
+    .instrument(info_span!("init_lsn_lease", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
+    .await?;
 
     json_response(StatusCode::OK, result)
 }
@@ -1706,8 +1719,13 @@ async fn timeline_gc_handler(
 
     let gc_req: TimelineGcRequest = json_request(&mut request).await?;
 
+    let state = get_state(&request);
+
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let gc_result = mgr::immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx).await?;
+    let gc_result = state
+        .tenant_manager
+        .immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx)
+        .await?;
 
     json_response(StatusCode::OK, gc_result)
 }
@@ -2955,7 +2973,7 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/preserve_initdb_archive",
             |r| api_handler(r, timeline_preserve_initdb_handler),
         )
-        .post(
+        .put(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/archival_config",
             |r| api_handler(r, timeline_archival_config_handler),
         )

pageserver/src/lib.rs

@@ -6,7 +6,7 @@ pub mod basebackup;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
-pub mod control_plane_client;
+pub mod controller_upcall_client;
 pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
 pub mod http;

pageserver/src/metrics.rs

@@ -8,6 +8,8 @@ use metrics::{
 };
 use once_cell::sync::Lazy;
 use pageserver_api::shard::TenantShardId;
+use postgres_backend::{is_expected_io_error, QueryError};
+use pq_proto::framed::ConnectionError;
 use strum::{EnumCount, VariantNames};
 use strum_macros::{IntoStaticStr, VariantNames};
 use tracing::warn;
@@ -1177,10 +1179,10 @@ pub(crate) mod virtual_file_io_engine {
 }
 
 struct GlobalAndPerTimelineHistogramTimer<'a, 'c> {
-    global_metric: &'a Histogram,
+    global_latency_histo: &'a Histogram,
 
     // Optional because not all op types are tracked per-timeline
-    timeline_metric: Option<&'a Histogram>,
+    per_timeline_latency_histo: Option<&'a Histogram>,
 
     ctx: &'c RequestContext,
     start: std::time::Instant,
@@ -1212,9 +1214,10 @@ impl<'a, 'c> Drop for GlobalAndPerTimelineHistogramTimer<'a, 'c> {
                 elapsed
             }
         };
-        self.global_metric.observe(ex_throttled.as_secs_f64());
-        if let Some(timeline_metric) = self.timeline_metric {
-            timeline_metric.observe(ex_throttled.as_secs_f64());
+        self.global_latency_histo
+            .observe(ex_throttled.as_secs_f64());
+        if let Some(per_timeline_getpage_histo) = self.per_timeline_latency_histo {
+            per_timeline_getpage_histo.observe(ex_throttled.as_secs_f64());
         }
     }
 }
@@ -1240,10 +1243,32 @@ pub enum SmgrQueryType {
 
 #[derive(Debug)]
 pub(crate) struct SmgrQueryTimePerTimeline {
-    global_metrics: [Histogram; SmgrQueryType::COUNT],
-    per_timeline_getpage: Histogram,
+    global_started: [IntCounter; SmgrQueryType::COUNT],
+    global_latency: [Histogram; SmgrQueryType::COUNT],
+    per_timeline_getpage_started: IntCounter,
+    per_timeline_getpage_latency: Histogram,
 }
 
+static SMGR_QUERY_STARTED_GLOBAL: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        // it's a counter, but, name is prepared to extend it to a histogram of queue depth
+        "pageserver_smgr_query_started_global_count",
+        "Number of smgr queries started, aggregated by query type.",
+        &["smgr_query_type"],
+    )
+    .expect("failed to define a metric")
+});
+
+static SMGR_QUERY_STARTED_PER_TENANT_TIMELINE: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        // it's a counter, but, name is prepared to extend it to a histogram of queue depth
+        "pageserver_smgr_query_started_count",
+        "Number of smgr queries started, aggregated by query type and tenant/timeline.",
+        &["smgr_query_type", "tenant_id", "shard_id", "timeline_id"],
+    )
+    .expect("failed to define a metric")
+});
+
 static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_smgr_query_seconds",
@@ -1319,14 +1344,20 @@ impl SmgrQueryTimePerTimeline {
         let tenant_id = tenant_shard_id.tenant_id.to_string();
         let shard_slug = format!("{}", tenant_shard_id.shard_slug());
         let timeline_id = timeline_id.to_string();
-        let global_metrics = std::array::from_fn(|i| {
+        let global_started = std::array::from_fn(|i| {
+            let op = SmgrQueryType::from_repr(i).unwrap();
+            SMGR_QUERY_STARTED_GLOBAL
+                .get_metric_with_label_values(&[op.into()])
+                .unwrap()
+        });
+        let global_latency = std::array::from_fn(|i| {
             let op = SmgrQueryType::from_repr(i).unwrap();
             SMGR_QUERY_TIME_GLOBAL
                 .get_metric_with_label_values(&[op.into()])
                 .unwrap()
         });
 
-        let per_timeline_getpage = SMGR_QUERY_TIME_PER_TENANT_TIMELINE
+        let per_timeline_getpage_started = SMGR_QUERY_STARTED_PER_TENANT_TIMELINE
             .get_metric_with_label_values(&[
                 SmgrQueryType::GetPageAtLsn.into(),
                 &tenant_id,
@@ -1334,18 +1365,32 @@ impl SmgrQueryTimePerTimeline {
                 &timeline_id,
             ])
             .unwrap();
+        let per_timeline_getpage_latency = SMGR_QUERY_TIME_PER_TENANT_TIMELINE
+            .get_metric_with_label_values(&[
+                SmgrQueryType::GetPageAtLsn.into(),
+                &tenant_id,
+                &shard_slug,
+                &timeline_id,
+            ])
+            .unwrap();
+
         Self {
-            global_metrics,
-            per_timeline_getpage,
+            global_started,
+            global_latency,
+            per_timeline_getpage_latency,
+            per_timeline_getpage_started,
         }
     }
     pub(crate) fn start_timer<'c: 'a, 'a>(
         &'a self,
         op: SmgrQueryType,
         ctx: &'c RequestContext,
-    ) -> Option<impl Drop + '_> {
-        let global_metric = &self.global_metrics[op as usize];
+    ) -> Option<impl Drop + 'a> {
         let start = Instant::now();
+
+        self.global_started[op as usize].inc();
+
+        // We subtract time spent throttled from the observed latency.
         match ctx.micros_spent_throttled.open() {
             Ok(()) => (),
             Err(error) => {
@@ -1364,15 +1409,16 @@ impl SmgrQueryTimePerTimeline {
             }
         }
 
-        let timeline_metric = if matches!(op, SmgrQueryType::GetPageAtLsn) {
-            Some(&self.per_timeline_getpage)
+        let per_timeline_latency_histo = if matches!(op, SmgrQueryType::GetPageAtLsn) {
+            self.per_timeline_getpage_started.inc();
+            Some(&self.per_timeline_getpage_latency)
         } else {
             None
         };
 
         Some(GlobalAndPerTimelineHistogramTimer {
-            global_metric,
-            timeline_metric,
+            global_latency_histo: &self.global_latency[op as usize],
+            per_timeline_latency_histo,
             ctx,
             start,
             op,
@@ -1423,9 +1469,12 @@ mod smgr_query_time_tests {
             let get_counts = || {
                 let global: u64 = ops
                     .iter()
-                    .map(|op| metrics.global_metrics[*op as usize].get_sample_count())
+                    .map(|op| metrics.global_latency[*op as usize].get_sample_count())
                     .sum();
-                (global, metrics.per_timeline_getpage.get_sample_count())
+                (
+                    global,
+                    metrics.per_timeline_getpage_latency.get_sample_count(),
+                )
             };
 
             let (pre_global, pre_per_tenant_timeline) = get_counts();
@@ -1461,6 +1510,7 @@ static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
 pub(crate) struct BasebackupQueryTime {
     ok: Histogram,
     error: Histogram,
+    client_error: Histogram,
 }
 
 pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
@@ -1474,6 +1524,7 @@ pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|
     BasebackupQueryTime {
         ok: vec.get_metric_with_label_values(&["ok"]).unwrap(),
         error: vec.get_metric_with_label_values(&["error"]).unwrap(),
+        client_error: vec.get_metric_with_label_values(&["client_error"]).unwrap(),
     }
 });
 
@@ -1487,7 +1538,7 @@ impl BasebackupQueryTime {
     pub(crate) fn start_recording<'c: 'a, 'a>(
         &'a self,
         ctx: &'c RequestContext,
-    ) -> BasebackupQueryTimeOngoingRecording<'_, '_> {
+    ) -> BasebackupQueryTimeOngoingRecording<'a, 'a> {
         let start = Instant::now();
         match ctx.micros_spent_throttled.open() {
             Ok(()) => (),
@@ -1510,7 +1561,7 @@ impl BasebackupQueryTime {
 }
 
 impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
-    pub(crate) fn observe<T, E>(self, res: &Result<T, E>) {
+    pub(crate) fn observe<T>(self, res: &Result<T, QueryError>) {
         let elapsed = self.start.elapsed();
         let ex_throttled = self
             .ctx
@@ -1529,10 +1580,15 @@ impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
                 elapsed
             }
         };
-        let metric = if res.is_ok() {
-            &self.parent.ok
-        } else {
-            &self.parent.error
+        // If you want to change categorize of a specific error, also change it in `log_query_error`.
+        let metric = match res {
+            Ok(_) => &self.parent.ok,
+            Err(QueryError::Disconnected(ConnectionError::Io(io_error)))
+                if is_expected_io_error(io_error) =>
+            {
+                &self.parent.client_error
+            }
+            Err(_) => &self.parent.error,
         };
         metric.observe(ex_throttled.as_secs_f64());
     }
@@ -2576,6 +2632,12 @@ impl TimelineMetrics {
             let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, shard_id, timeline_id]);
         }
 
+        let _ = SMGR_QUERY_STARTED_PER_TENANT_TIMELINE.remove_label_values(&[
+            SmgrQueryType::GetPageAtLsn.into(),
+            tenant_id,
+            shard_id,
+            timeline_id,
+        ]);
         let _ = SMGR_QUERY_TIME_PER_TENANT_TIMELINE.remove_label_values(&[
             SmgrQueryType::GetPageAtLsn.into(),
             tenant_id,
@@ -2592,6 +2654,8 @@ pub(crate) fn remove_tenant_metrics(tenant_shard_id: &TenantShardId) {
         let _ = TENANT_SYNTHETIC_SIZE_METRIC.remove_label_values(&[&tid]);
     }
 
+    tenant_throttling::remove_tenant_metrics(tenant_shard_id);
+
     // we leave the BROKEN_TENANTS_SET entry if any
 }
 
@@ -3055,41 +3119,173 @@ pub mod tokio_epoll_uring {
 pub(crate) mod tenant_throttling {
     use metrics::{register_int_counter_vec, IntCounter};
     use once_cell::sync::Lazy;
+    use utils::shard::TenantShardId;
 
     use crate::tenant::{self, throttle::Metric};
 
-    pub(crate) struct TimelineGet {
-        wait_time: IntCounter,
-        count: IntCounter,
+    struct GlobalAndPerTenantIntCounter {
+        global: IntCounter,
+        per_tenant: IntCounter,
     }
 
-    pub(crate) static TIMELINE_GET: Lazy<TimelineGet> = Lazy::new(|| {
-        static WAIT_USECS: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
-            register_int_counter_vec!(
-            "pageserver_tenant_throttling_wait_usecs_sum_global",
-            "Sum of microseconds that tenants spent waiting for a tenant throttle of a given kind.",
+    impl GlobalAndPerTenantIntCounter {
+        #[inline(always)]
+        pub(crate) fn inc(&self) {
+            self.inc_by(1)
+        }
+        #[inline(always)]
+        pub(crate) fn inc_by(&self, n: u64) {
+            self.global.inc_by(n);
+            self.per_tenant.inc_by(n);
+        }
+    }
+
+    pub(crate) struct TimelineGet {
+        count_accounted_start: GlobalAndPerTenantIntCounter,
+        count_accounted_finish: GlobalAndPerTenantIntCounter,
+        wait_time: GlobalAndPerTenantIntCounter,
+        count_throttled: GlobalAndPerTenantIntCounter,
+    }
+
+    static COUNT_ACCOUNTED_START: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count_accounted_start_global",
+            "Count of tenant throttling starts, by kind of throttle.",
             &["kind"]
         )
-            .unwrap()
-        });
-
-        static WAIT_COUNT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
-            register_int_counter_vec!(
-                "pageserver_tenant_throttling_count_global",
-                "Count of tenant throttlings, by kind of throttle.",
-                &["kind"]
-            )
-            .unwrap()
-        });
-
-        let kind = "timeline_get";
-        TimelineGet {
-            wait_time: WAIT_USECS.with_label_values(&[kind]),
-            count: WAIT_COUNT.with_label_values(&[kind]),
-        }
+        .unwrap()
+    });
+    static COUNT_ACCOUNTED_START_PER_TENANT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count_accounted_start",
+            "Count of tenant throttling starts, by kind of throttle.",
+            &["kind", "tenant_id", "shard_id"]
+        )
+        .unwrap()
+    });
+    static COUNT_ACCOUNTED_FINISH: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count_accounted_finish_global",
+            "Count of tenant throttling finishes, by kind of throttle.",
+            &["kind"]
+        )
+        .unwrap()
+    });
+    static COUNT_ACCOUNTED_FINISH_PER_TENANT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count_accounted_finish",
+            "Count of tenant throttling finishes, by kind of throttle.",
+            &["kind", "tenant_id", "shard_id"]
+        )
+        .unwrap()
+    });
+    static WAIT_USECS: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_wait_usecs_sum_global",
+            "Sum of microseconds that spent waiting throttle by kind of throttle.",
+            &["kind"]
+        )
+        .unwrap()
+    });
+    static WAIT_USECS_PER_TENANT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_wait_usecs_sum",
+            "Sum of microseconds that spent waiting throttle by kind of throttle.",
+            &["kind", "tenant_id", "shard_id"]
+        )
+        .unwrap()
     });
 
-    impl Metric for &'static TimelineGet {
+    static WAIT_COUNT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count_global",
+            "Count of tenant throttlings, by kind of throttle.",
+            &["kind"]
+        )
+        .unwrap()
+    });
+    static WAIT_COUNT_PER_TENANT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_tenant_throttling_count",
+            "Count of tenant throttlings, by kind of throttle.",
+            &["kind", "tenant_id", "shard_id"]
+        )
+        .unwrap()
+    });
+
+    const KIND: &str = "timeline_get";
+
+    impl TimelineGet {
+        pub(crate) fn new(tenant_shard_id: &TenantShardId) -> Self {
+            let per_tenant_label_values = &[
+                KIND,
+                &tenant_shard_id.tenant_id.to_string(),
+                &tenant_shard_id.shard_slug().to_string(),
+            ];
+            TimelineGet {
+                count_accounted_start: {
+                    GlobalAndPerTenantIntCounter {
+                        global: COUNT_ACCOUNTED_START.with_label_values(&[KIND]),
+                        per_tenant: COUNT_ACCOUNTED_START_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
+                    }
+                },
+                count_accounted_finish: {
+                    GlobalAndPerTenantIntCounter {
+                        global: COUNT_ACCOUNTED_FINISH.with_label_values(&[KIND]),
+                        per_tenant: COUNT_ACCOUNTED_FINISH_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
+                    }
+                },
+                wait_time: {
+                    GlobalAndPerTenantIntCounter {
+                        global: WAIT_USECS.with_label_values(&[KIND]),
+                        per_tenant: WAIT_USECS_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
+                    }
+                },
+                count_throttled: {
+                    GlobalAndPerTenantIntCounter {
+                        global: WAIT_COUNT.with_label_values(&[KIND]),
+                        per_tenant: WAIT_COUNT_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
+                    }
+                },
+            }
+        }
+    }
+
+    pub(crate) fn preinitialize_global_metrics() {
+        Lazy::force(&COUNT_ACCOUNTED_START);
+        Lazy::force(&COUNT_ACCOUNTED_FINISH);
+        Lazy::force(&WAIT_USECS);
+        Lazy::force(&WAIT_COUNT);
+    }
+
+    pub(crate) fn remove_tenant_metrics(tenant_shard_id: &TenantShardId) {
+        for m in &[
+            &COUNT_ACCOUNTED_START_PER_TENANT,
+            &COUNT_ACCOUNTED_FINISH_PER_TENANT,
+            &WAIT_USECS_PER_TENANT,
+            &WAIT_COUNT_PER_TENANT,
+        ] {
+            let _ = m.remove_label_values(&[
+                KIND,
+                &tenant_shard_id.tenant_id.to_string(),
+                &tenant_shard_id.shard_slug().to_string(),
+            ]);
+        }
+    }
+
+    impl Metric for TimelineGet {
+        #[inline(always)]
+        fn accounting_start(&self) {
+            self.count_accounted_start.inc();
+        }
+        #[inline(always)]
+        fn accounting_finish(&self) {
+            self.count_accounted_finish.inc();
+        }
         #[inline(always)]
         fn observe_throttling(
             &self,
@@ -3097,7 +3293,7 @@ pub(crate) mod tenant_throttling {
         ) {
             let val = u64::try_from(wait_time.as_micros()).unwrap();
             self.wait_time.inc_by(val);
-            self.count.inc();
+            self.count_throttled.inc();
         }
     }
 }
@@ -3227,11 +3423,14 @@ pub fn preinitialize_metrics() {
     }
 
     // countervecs
-    [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
-        .into_iter()
-        .for_each(|c| {
-            Lazy::force(c);
-        });
+    [
+        &BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT,
+        &SMGR_QUERY_STARTED_GLOBAL,
+    ]
+    .into_iter()
+    .for_each(|c| {
+        Lazy::force(c);
+    });
 
     // gauges
     WALRECEIVER_ACTIVE_MANAGERS.get();
@@ -3253,7 +3452,8 @@ pub fn preinitialize_metrics() {
 
     // Custom
     Lazy::force(&RECONSTRUCT_TIME);
-    Lazy::force(&tenant_throttling::TIMELINE_GET);
     Lazy::force(&BASEBACKUP_QUERY_TIME);
     Lazy::force(&COMPUTE_COMMANDS_COUNTERS);
+
+    tenant_throttling::preinitialize_global_metrics();
 }

pageserver/src/page_service.rs

@@ -273,10 +273,20 @@ async fn page_service_conn_main(
                 info!("Postgres client disconnected ({io_error})");
                 Ok(())
             } else {
-                Err(io_error).context("Postgres connection error")
+                let tenant_id = conn_handler.timeline_handles.tenant_id();
+                Err(io_error).context(format!(
+                    "Postgres connection error for tenant_id={:?} client at peer_addr={}",
+                    tenant_id, peer_addr
+                ))
             }
         }
-        other => other.context("Postgres query error"),
+        other => {
+            let tenant_id = conn_handler.timeline_handles.tenant_id();
+            other.context(format!(
+                "Postgres query error for tenant_id={:?} client peer_addr={}",
+                tenant_id, peer_addr
+            ))
+        }
     }
 }
 
@@ -340,6 +350,10 @@ impl TimelineHandles {
                 }
             })
     }
+
+    fn tenant_id(&self) -> Option<TenantId> {
+        self.wrapper.tenant_id.get().copied()
+    }
 }
 
 pub(crate) struct TenantManagerWrapper {
@@ -819,7 +833,7 @@ impl PageServerHandler {
         set_tracing_field_shard_id(&timeline);
 
         let lease = timeline
-            .make_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
+            .renew_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
             .inspect_err(|e| {
                 warn!("{e}");
             })

pageserver/src/pgdatadir_mapping.rs

@@ -840,6 +840,36 @@ impl Timeline {
         Ok(total_size * BLCKSZ as u64)
     }
 
+    /// Get a KeySpace that covers all the Keys that are in use at AND below the given LSN. This is only used
+    /// for gc-compaction.
+    ///
+    /// gc-compaction cannot use the same `collect_keyspace` function as the legacy compaction because it
+    /// processes data at multiple LSNs and needs to be aware of the fact that some key ranges might need to
+    /// be kept only for a specific range of LSN.
+    ///
+    /// Consider the case that the user created branches at LSN 10 and 20, where the user created a table A at
+    /// LSN 10 and dropped that table at LSN 20. `collect_keyspace` at LSN 10 will return the key range
+    /// corresponding to that table, while LSN 20 won't. The keyspace info at a single LSN is not enough to
+    /// determine which keys to retain/drop for gc-compaction.
+    ///
+    /// For now, it only drops AUX-v1 keys. But in the future, the function will be extended to return the keyspace
+    /// to be retained for each of the branch LSN.
+    ///
+    /// The return value is (dense keyspace, sparse keyspace).
+    pub(crate) async fn collect_gc_compaction_keyspace(
+        &self,
+    ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
+        let metadata_key_begin = Key::metadata_key_range().start;
+        let aux_v1_key = AUX_FILES_KEY;
+        let dense_keyspace = KeySpace {
+            ranges: vec![Key::MIN..aux_v1_key, aux_v1_key.next()..metadata_key_begin],
+        };
+        Ok((
+            dense_keyspace,
+            SparseKeySpace(KeySpace::single(Key::metadata_key_range())),
+        ))
+    }
+
     ///
     /// Get a KeySpace that covers all the Keys that are in use at the given LSN.
     /// Anything that's not listed maybe removed from the underlying storage (from

pageserver/src/tenant.rs

@@ -18,10 +18,10 @@ use camino::Utf8Path;
 use camino::Utf8PathBuf;
 use enumset::EnumSet;
 use futures::stream::FuturesUnordered;
-use futures::FutureExt;
 use futures::StreamExt;
 use pageserver_api::models;
 use pageserver_api::models::AuxFilePolicy;
+use pageserver_api::models::LsnLease;
 use pageserver_api::models::TimelineArchivalState;
 use pageserver_api::models::TimelineState;
 use pageserver_api::models::TopTenantShardItem;
@@ -34,6 +34,7 @@ use remote_storage::GenericRemoteStorage;
 use remote_storage::TimeoutOrCancel;
 use std::collections::BTreeMap;
 use std::fmt;
+use std::future::Future;
 use std::sync::Weak;
 use std::time::SystemTime;
 use storage_broker::BrokerClientChannel;
@@ -182,27 +183,54 @@ pub struct TenantSharedResources {
 pub(super) struct AttachedTenantConf {
     tenant_conf: TenantConfOpt,
     location: AttachedLocationConfig,
+    /// The deadline before which we are blocked from GC so that
+    /// leases have a chance to be renewed.
+    lsn_lease_deadline: Option<tokio::time::Instant>,
 }
 
 impl AttachedTenantConf {
     fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
+        // Sets a deadline before which we cannot proceed to GC due to lsn lease.
+        //
+        // We do this as the leases mapping are not persisted to disk. By delaying GC by lease
+        // length, we guarantee that all the leases we granted before will have a chance to renew
+        // when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
+        let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
+            Some(
+                tokio::time::Instant::now()
+                    + tenant_conf
+                        .lsn_lease_length
+                        .unwrap_or(LsnLease::DEFAULT_LENGTH),
+            )
+        } else {
+            // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
+            // because we don't do GC in these modes.
+            None
+        };
+
         Self {
             tenant_conf,
             location,
+            lsn_lease_deadline,
         }
     }
 
     fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
         match &location_conf.mode {
-            LocationMode::Attached(attach_conf) => Ok(Self {
-                tenant_conf: location_conf.tenant_conf,
-                location: *attach_conf,
-            }),
+            LocationMode::Attached(attach_conf) => {
+                Ok(Self::new(location_conf.tenant_conf, *attach_conf))
+            }
             LocationMode::Secondary(_) => {
                 anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
             }
         }
     }
+
+    fn is_gc_blocked_by_lsn_lease_deadline(&self) -> bool {
+        self.lsn_lease_deadline
+            .map(|d| tokio::time::Instant::now() < d)
+            .unwrap_or(false)
+    }
 }
 struct TimelinePreload {
     timeline_id: TimelineId,
@@ -302,7 +330,7 @@ pub struct Tenant {
     /// Throttle applied at the top of [`Timeline::get`].
     /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
     pub(crate) timeline_get_throttle:
-        Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
+        Arc<throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
 
     /// An ongoing timeline detach concurrency limiter.
     ///
@@ -563,6 +591,8 @@ pub enum CreateTimelineError {
     AncestorLsn(anyhow::Error),
     #[error("ancestor timeline is not active")]
     AncestorNotActive,
+    #[error("ancestor timeline is archived")]
+    AncestorArchived,
     #[error("tenant shutting down")]
     ShuttingDown,
     #[error(transparent)]
@@ -1031,13 +1061,9 @@ impl Tenant {
         }
 
         Ok(TenantPreload {
-            timelines: Self::load_timeline_metadata(
-                self,
-                remote_timeline_ids,
-                remote_storage,
-                cancel,
-            )
-            .await?,
+            timelines: self
+                .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
+                .await?,
         })
     }
 
@@ -1303,7 +1329,7 @@ impl Tenant {
         .await
     }
 
-    async fn load_timeline_metadata(
+    async fn load_timelines_metadata(
         self: &Arc<Tenant>,
         timeline_ids: HashSet<TimelineId>,
         remote_storage: &GenericRemoteStorage,
@@ -1311,33 +1337,10 @@ impl Tenant {
     ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
         let mut part_downloads = JoinSet::new();
         for timeline_id in timeline_ids {
-            let client = RemoteTimelineClient::new(
-                remote_storage.clone(),
-                self.deletion_queue_client.clone(),
-                self.conf,
-                self.tenant_shard_id,
-                timeline_id,
-                self.generation,
-            );
             let cancel_clone = cancel.clone();
             part_downloads.spawn(
-                async move {
-                    debug!("starting index part download");
-
-                    let index_part = client.download_index_file(&cancel_clone).await;
-
-                    debug!("finished index part download");
-
-                    Result::<_, anyhow::Error>::Ok(TimelinePreload {
-                        client,
-                        timeline_id,
-                        index_part,
-                    })
-                }
-                .map(move |res| {
-                    res.with_context(|| format!("download index part for timeline {timeline_id}"))
-                })
-                .instrument(info_span!("download_index_part", %timeline_id)),
+                self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
+                    .instrument(info_span!("download_index_part", %timeline_id)),
             );
         }
 
@@ -1348,8 +1351,7 @@ impl Tenant {
                 next = part_downloads.join_next() => {
                     match next {
                         Some(result) => {
-                            let preload_result = result.context("join preload task")?;
-                            let preload = preload_result?;
+                            let preload = result.context("join preload task")?;
                             timeline_preloads.insert(preload.timeline_id, preload);
                         },
                         None => {
@@ -1366,6 +1368,36 @@ impl Tenant {
         Ok(timeline_preloads)
     }
 
+    fn load_timeline_metadata(
+        self: &Arc<Tenant>,
+        timeline_id: TimelineId,
+        remote_storage: GenericRemoteStorage,
+        cancel: CancellationToken,
+    ) -> impl Future<Output = TimelinePreload> {
+        let client = RemoteTimelineClient::new(
+            remote_storage.clone(),
+            self.deletion_queue_client.clone(),
+            self.conf,
+            self.tenant_shard_id,
+            timeline_id,
+            self.generation,
+        );
+        async move {
+            debug_assert_current_span_has_tenant_and_timeline_id();
+            debug!("starting index part download");
+
+            let index_part = client.download_index_file(&cancel).await;
+
+            debug!("finished index part download");
+
+            TimelinePreload {
+                client,
+                timeline_id,
+                index_part,
+            }
+        }
+    }
+
     pub(crate) async fn apply_timeline_archival_config(
         &self,
         timeline_id: TimelineId,
@@ -1696,6 +1728,11 @@ impl Tenant {
                     return Err(CreateTimelineError::AncestorNotActive);
                 }
 
+                if ancestor_timeline.is_archived() == Some(true) {
+                    info!("tried to branch archived timeline");
+                    return Err(CreateTimelineError::AncestorArchived);
+                }
+
                 if let Some(lsn) = ancestor_start_lsn.as_mut() {
                     *lsn = lsn.align();
 
@@ -1813,6 +1850,11 @@ impl Tenant {
                 info!("Skipping GC in location state {:?}", conf.location);
                 return Ok(GcResult::default());
             }
+
+            if conf.is_gc_blocked_by_lsn_lease_deadline() {
+                info!("Skipping GC because lsn lease deadline is not reached");
+                return Ok(GcResult::default());
+            }
         }
 
         let _guard = match self.gc_block.start().await {
@@ -1966,9 +2008,6 @@ impl Tenant {
                 TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
                     panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
                 }
-                TenantState::Loading => {
-                    *current_state = TenantState::Activating(ActivatingFrom::Loading);
-                }
                 TenantState::Attaching => {
                     *current_state = TenantState::Activating(ActivatingFrom::Attaching);
                 }
@@ -2149,7 +2188,7 @@ impl Tenant {
     async fn set_stopping(
         &self,
         progress: completion::Barrier,
-        allow_transition_from_loading: bool,
+        _allow_transition_from_loading: bool,
         allow_transition_from_attaching: bool,
     ) -> Result<(), SetStoppingError> {
         let mut rx = self.state.subscribe();
@@ -2164,7 +2203,6 @@ impl Tenant {
                 );
                 false
             }
-            TenantState::Loading => allow_transition_from_loading,
             TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
         })
         .await
@@ -2183,13 +2221,6 @@ impl Tenant {
                 *current_state = TenantState::Stopping { progress };
                 true
             }
-            TenantState::Loading => {
-                if !allow_transition_from_loading {
-                    unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
-                };
-                *current_state = TenantState::Stopping { progress };
-                true
-            }
             TenantState::Active => {
                 // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
                 // are created after the transition to Stopping. That's harmless, as the Timelines
@@ -2245,7 +2276,7 @@ impl Tenant {
         // The load & attach routines own the tenant state until it has reached `Active`.
         // So, wait until it's done.
         rx.wait_for(|state| match state {
-            TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
+            TenantState::Activating(_) | TenantState::Attaching => {
                 info!(
                     "waiting for {} to turn Active|Broken|Stopping",
                     <&'static str>::from(state)
@@ -2265,7 +2296,7 @@ impl Tenant {
         let reason = reason.to_string();
         self.state.send_modify(|current_state| {
             match *current_state {
-                TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
+                TenantState::Activating(_) | TenantState::Attaching => {
                     unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
                 }
                 TenantState::Active => {
@@ -2309,7 +2340,7 @@ impl Tenant {
         loop {
             let current_state = receiver.borrow_and_update().clone();
             match current_state {
-                TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
+                TenantState::Attaching | TenantState::Activating(_) => {
                     // in these states, there's a chance that we can reach ::Active
                     self.activate_now();
                     match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
@@ -2632,6 +2663,8 @@ impl Tenant {
             Arc::new(AttachedTenantConf {
                 tenant_conf: new_tenant_conf.clone(),
                 location: inner.location,
+                // Attached location is not changed, no need to update lsn lease deadline.
+                lsn_lease_deadline: inner.lsn_lease_deadline,
             })
         });
 
@@ -2831,7 +2864,7 @@ impl Tenant {
             gate: Gate::default(),
             timeline_get_throttle: Arc::new(throttle::Throttle::new(
                 Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
-                &crate::metrics::tenant_throttling::TIMELINE_GET,
+                crate::metrics::tenant_throttling::TimelineGet::new(&tenant_shard_id),
             )),
             tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
             ongoing_timeline_detach: std::sync::Mutex::default(),
@@ -3625,7 +3658,7 @@ impl Tenant {
         start_lsn: Lsn,
         ancestor: Option<Arc<Timeline>>,
         last_aux_file_policy: Option<AuxFilePolicy>,
-    ) -> anyhow::Result<UninitializedTimeline> {
+    ) -> anyhow::Result<UninitializedTimeline<'a>> {
         let tenant_shard_id = self.tenant_shard_id;
 
         let resources = self.build_timeline_resources(new_timeline_id);
@@ -3889,9 +3922,9 @@ async fn run_initdb(
     let _permit = INIT_DB_SEMAPHORE.acquire().await;
 
     let initdb_command = tokio::process::Command::new(&initdb_bin_path)
-        .args(["-D", initdb_target_dir.as_ref()])
-        .args(["-U", &conf.superuser])
-        .args(["-E", "utf8"])
+        .args(["--pgdata", initdb_target_dir.as_ref()])
+        .args(["--username", &conf.superuser])
+        .args(["--encoding", "utf8"])
         .arg("--no-instructions")
         .arg("--no-sync")
         .env_clear()
@@ -4142,7 +4175,7 @@ pub(crate) mod harness {
             let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
 
             let tenant = Arc::new(Tenant::new(
-                TenantState::Loading,
+                TenantState::Attaching,
                 self.conf,
                 AttachedTenantConf::try_from(LocationConf::attached_single(
                     TenantConfOpt::from(self.tenant_conf.clone()),
@@ -4463,13 +4496,17 @@ mod tests {
         tline.freeze_and_flush().await.map_err(|e| e.into())
     }
 
-    #[tokio::test]
+    #[tokio::test(start_paused = true)]
     async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
         let (tenant, ctx) =
             TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
                 .await?
                 .load()
                 .await;
+        // Advance to the lsn lease deadline so that GC is not blocked by
+        // initial transition into AttachedSingle.
+        tokio::time::advance(tenant.get_lsn_lease_length()).await;
+        tokio::time::resume();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -7246,9 +7283,17 @@ mod tests {
         Ok(())
     }
 
-    #[tokio::test]
+    #[tokio::test(start_paused = true)]
     async fn test_lsn_lease() -> anyhow::Result<()> {
-        let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
+        let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
+            .await
+            .unwrap()
+            .load()
+            .await;
+        // Advance to the lsn lease deadline so that GC is not blocked by
+        // initial transition into AttachedSingle.
+        tokio::time::advance(tenant.get_lsn_lease_length()).await;
+        tokio::time::resume();
         let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
 
         let end_lsn = Lsn(0x100);
@@ -7276,24 +7321,33 @@ mod tests {
 
         let leased_lsns = [0x30, 0x50, 0x70];
         let mut leases = Vec::new();
-        let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
-            leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
-            Ok(())
+        leased_lsns.iter().for_each(|n| {
+            leases.push(
+                timeline
+                    .init_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)
+                    .expect("lease request should succeed"),
+            );
         });
 
-        // Renewing with shorter lease should not change the lease.
-        let updated_lease_0 =
-            timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
-        assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
+        let updated_lease_0 = timeline
+            .renew_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)
+            .expect("lease renewal should succeed");
+        assert_eq!(
+            updated_lease_0.valid_until, leases[0].valid_until,
+            " Renewing with shorter lease should not change the lease."
+        );
 
-        // Renewing with a long lease should renew lease with later expiration time.
-        let updated_lease_1 = timeline.make_lsn_lease(
-            Lsn(leased_lsns[1]),
-            timeline.get_lsn_lease_length() * 2,
-            &ctx,
-        )?;
-
-        assert!(updated_lease_1.valid_until > leases[1].valid_until);
+        let updated_lease_1 = timeline
+            .renew_lsn_lease(
+                Lsn(leased_lsns[1]),
+                timeline.get_lsn_lease_length() * 2,
+                &ctx,
+            )
+            .expect("lease renewal should succeed");
+        assert!(
+            updated_lease_1.valid_until > leases[1].valid_until,
+            "Renewing with a long lease should renew lease with later expiration time."
+        );
 
         // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
         info!(
@@ -7310,7 +7364,8 @@ mod tests {
                 &CancellationToken::new(),
                 &ctx,
             )
-            .await?;
+            .await
+            .unwrap();
 
         // Keeping everything <= Lsn(0x80) b/c leases:
         // 0/10: initdb layer
@@ -7324,13 +7379,16 @@ mod tests {
         // Make lease on a already GC-ed LSN.
         // 0/80 does not have a valid lease + is below latest_gc_cutoff
         assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
-        let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
-        assert!(res.is_err());
+        timeline
+            .init_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx)
+            .expect_err("lease request on GC-ed LSN should fail");
 
         // Should still be able to renew a currently valid lease
         // Assumption: original lease to is still valid for 0/50.
-        let _ =
-            timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
+        // (use `Timeline::init_lsn_lease` for testing so it always does validation)
+        timeline
+            .init_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)
+            .expect("lease renewal with validation should succeed");
 
         Ok(())
     }

pageserver/src/tenant/checks.rs

@@ -5,6 +5,7 @@ use itertools::Itertools;
 use super::storage_layer::LayerName;
 
 /// Checks whether a layer map is valid (i.e., is a valid result of the current compaction algorithm if nothing goes wrong).
+///
 /// The function checks if we can split the LSN range of a delta layer only at the LSNs of the delta layers. For example,
 ///
 /// ```plain

pageserver/src/tenant/gc_block.rs

@@ -1,29 +1,12 @@
-use std::{collections::HashMap, time::Duration};
+use std::collections::HashMap;
 
-use super::remote_timeline_client::index::GcBlockingReason;
-use tokio::time::Instant;
 use utils::id::TimelineId;
 
-type TimelinesBlocked = HashMap<TimelineId, enumset::EnumSet<GcBlockingReason>>;
+use super::remote_timeline_client::index::GcBlockingReason;
 
-#[derive(Default)]
-struct Storage {
-    timelines_blocked: TimelinesBlocked,
-    /// The deadline before which we are blocked from GC so that
-    /// leases have a chance to be renewed.
-    lsn_lease_deadline: Option<Instant>,
-}
+type Storage = HashMap<TimelineId, enumset::EnumSet<GcBlockingReason>>;
 
-impl Storage {
-    fn is_blocked_by_lsn_lease_deadline(&self) -> bool {
-        self.lsn_lease_deadline
-            .map(|d| Instant::now() < d)
-            .unwrap_or(false)
-    }
-}
-
-/// GcBlock provides persistent (per-timeline) gc blocking and facilitates transient time based gc
-/// blocking.
+/// GcBlock provides persistent (per-timeline) gc blocking.
 #[derive(Default)]
 pub(crate) struct GcBlock {
     /// The timelines which have current reasons to block gc.
@@ -66,17 +49,6 @@ impl GcBlock {
         }
     }
 
-    /// Sets a deadline before which we cannot proceed to GC due to lsn lease.
-    ///
-    /// We do this as the leases mapping are not persisted to disk. By delaying GC by lease
-    /// length, we guarantee that all the leases we granted before will have a chance to renew
-    /// when we run GC for the first time after restart / transition from AttachedMulti to AttachedSingle.
-    pub(super) fn set_lsn_lease_deadline(&self, lsn_lease_length: Duration) {
-        let deadline = Instant::now() + lsn_lease_length;
-        let mut g = self.reasons.lock().unwrap();
-        g.lsn_lease_deadline = Some(deadline);
-    }
-
     /// Describe the current gc blocking reasons.
     ///
     /// TODO: make this json serializable.
@@ -102,7 +74,7 @@ impl GcBlock {
     ) -> anyhow::Result<bool> {
         let (added, uploaded) = {
             let mut g = self.reasons.lock().unwrap();
-            let set = g.timelines_blocked.entry(timeline.timeline_id).or_default();
+            let set = g.entry(timeline.timeline_id).or_default();
             let added = set.insert(reason);
 
             // LOCK ORDER: intentionally hold the lock, see self.reasons.
@@ -133,7 +105,7 @@ impl GcBlock {
 
         let (remaining_blocks, uploaded) = {
             let mut g = self.reasons.lock().unwrap();
-            match g.timelines_blocked.entry(timeline.timeline_id) {
+            match g.entry(timeline.timeline_id) {
                 Entry::Occupied(mut oe) => {
                     let set = oe.get_mut();
                     set.remove(reason);
@@ -147,7 +119,7 @@ impl GcBlock {
                 }
             }
 
-            let remaining_blocks = g.timelines_blocked.len();
+            let remaining_blocks = g.len();
 
             // LOCK ORDER: intentionally hold the lock while scheduling; see self.reasons
             let uploaded = timeline
@@ -172,11 +144,11 @@ impl GcBlock {
     pub(crate) fn before_delete(&self, timeline: &super::Timeline) {
         let unblocked = {
             let mut g = self.reasons.lock().unwrap();
-            if g.timelines_blocked.is_empty() {
+            if g.is_empty() {
                 return;
             }
 
-            g.timelines_blocked.remove(&timeline.timeline_id);
+            g.remove(&timeline.timeline_id);
 
             BlockingReasons::clean_and_summarize(g).is_none()
         };
@@ -187,11 +159,10 @@ impl GcBlock {
     }
 
     /// Initialize with the non-deleted timelines of this tenant.
-    pub(crate) fn set_scanned(&self, scanned: TimelinesBlocked) {
+    pub(crate) fn set_scanned(&self, scanned: Storage) {
         let mut g = self.reasons.lock().unwrap();
-        assert!(g.timelines_blocked.is_empty());
-        g.timelines_blocked
-            .extend(scanned.into_iter().filter(|(_, v)| !v.is_empty()));
+        assert!(g.is_empty());
+        g.extend(scanned.into_iter().filter(|(_, v)| !v.is_empty()));
 
         if let Some(reasons) = BlockingReasons::clean_and_summarize(g) {
             tracing::info!(summary=?reasons, "initialized with gc blocked");
@@ -205,7 +176,6 @@ pub(super) struct Guard<'a> {
 
 #[derive(Debug)]
 pub(crate) struct BlockingReasons {
-    tenant_blocked_by_lsn_lease_deadline: bool,
     timelines: usize,
     reasons: enumset::EnumSet<GcBlockingReason>,
 }
@@ -214,8 +184,8 @@ impl std::fmt::Display for BlockingReasons {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(
             f,
-            "tenant_blocked_by_lsn_lease_deadline: {}, {} timelines block for {:?}",
-            self.tenant_blocked_by_lsn_lease_deadline, self.timelines, self.reasons
+            "{} timelines block for {:?}",
+            self.timelines, self.reasons
         )
     }
 }
@@ -223,15 +193,13 @@ impl std::fmt::Display for BlockingReasons {
 impl BlockingReasons {
     fn clean_and_summarize(mut g: std::sync::MutexGuard<'_, Storage>) -> Option<Self> {
         let mut reasons = enumset::EnumSet::empty();
-        g.timelines_blocked.retain(|_key, value| {
+        g.retain(|_key, value| {
             reasons = reasons.union(*value);
             !value.is_empty()
         });
-        let blocked_by_lsn_lease_deadline = g.is_blocked_by_lsn_lease_deadline();
-        if !g.timelines_blocked.is_empty() || blocked_by_lsn_lease_deadline {
+        if !g.is_empty() {
             Some(BlockingReasons {
-                tenant_blocked_by_lsn_lease_deadline: blocked_by_lsn_lease_deadline,
-                timelines: g.timelines_blocked.len(),
+                timelines: g.len(),
                 reasons,
             })
         } else {
@@ -240,17 +208,14 @@ impl BlockingReasons {
     }
 
     fn summarize(g: &std::sync::MutexGuard<'_, Storage>) -> Option<Self> {
-        let blocked_by_lsn_lease_deadline = g.is_blocked_by_lsn_lease_deadline();
-        if g.timelines_blocked.is_empty() && !blocked_by_lsn_lease_deadline {
+        if g.is_empty() {
             None
         } else {
             let reasons = g
-                .timelines_blocked
                 .values()
                 .fold(enumset::EnumSet::empty(), |acc, next| acc.union(*next));
             Some(BlockingReasons {
-                tenant_blocked_by_lsn_lease_deadline: blocked_by_lsn_lease_deadline,
-                timelines: g.timelines_blocked.len(),
+                timelines: g.len(),
                 reasons,
             })
         }

pageserver/src/tenant/mgr.rs

@@ -30,8 +30,8 @@ use utils::{backoff, completion, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::control_plane_client::{
-    ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
+use crate::controller_upcall_client::{
+    ControlPlaneGenerationsApi, ControllerUpcallClient, RetryForeverError,
 };
 use crate::deletion_queue::DeletionQueueClient;
 use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
@@ -122,7 +122,7 @@ pub(crate) enum ShardSelector {
     Known(ShardIndex),
 }
 
-/// A convenience for use with the re_attach ControlPlaneClient function: rather
+/// A convenience for use with the re_attach ControllerUpcallClient function: rather
 /// than the serializable struct, we build this enum that encapsulates
 /// the invariant that attached tenants always have generations.
 ///
@@ -219,7 +219,11 @@ async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<U
         + TEMP_FILE_SUFFIX;
     let tmp_path = path_with_suffix_extension(&path, &rand_suffix);
     fs::rename(path.as_ref(), &tmp_path).await?;
-    fs::File::open(parent).await?.sync_all().await?;
+    fs::File::open(parent)
+        .await?
+        .sync_all()
+        .await
+        .maybe_fatal_err("safe_rename_tenant_dir")?;
     Ok(tmp_path)
 }
 
@@ -341,7 +345,7 @@ async fn init_load_generations(
             "Emergency mode!  Tenants will be attached unsafely using their last known generation"
         );
         emergency_generations(tenant_confs)
-    } else if let Some(client) = ControlPlaneClient::new(conf, cancel) {
+    } else if let Some(client) = ControllerUpcallClient::new(conf, cancel) {
         info!("Calling control plane API to re-attach tenants");
         // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
         match client.re_attach(conf).await {
@@ -949,12 +953,6 @@ impl TenantManager {
                 (LocationMode::Attached(attach_conf), Some(TenantSlot::Attached(tenant))) => {
                     match attach_conf.generation.cmp(&tenant.generation) {
                         Ordering::Equal => {
-                            if attach_conf.attach_mode == AttachmentMode::Single {
-                                tenant
-                                    .gc_block
-                                    .set_lsn_lease_deadline(tenant.get_lsn_lease_length());
-                            }
-
                             // A transition from Attached to Attached in the same generation, we may
                             // take our fast path and just provide the updated configuration
                             // to the tenant.
@@ -2199,6 +2197,82 @@ impl TenantManager {
 
         Ok((wanted_bytes, shard_count as u32))
     }
+
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
+    pub(crate) async fn immediate_gc(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        gc_req: TimelineGcRequest,
+        cancel: CancellationToken,
+        ctx: &RequestContext,
+    ) -> Result<GcResult, ApiError> {
+        let tenant = {
+            let guard = self.tenants.read().unwrap();
+            guard
+                .get(&tenant_shard_id)
+                .cloned()
+                .with_context(|| format!("tenant {tenant_shard_id}"))
+                .map_err(|e| ApiError::NotFound(e.into()))?
+        };
+
+        let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
+        // Use tenant's pitr setting
+        let pitr = tenant.get_pitr_interval();
+
+        tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
+        // Run in task_mgr to avoid race with tenant_detach operation
+        let ctx: RequestContext =
+            ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
+
+        let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
+
+        fail::fail_point!("immediate_gc_task_pre");
+
+        #[allow(unused_mut)]
+        let mut result = tenant
+            .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
+            .await;
+        // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
+        // better once the types support it.
+
+        #[cfg(feature = "testing")]
+        {
+            // we need to synchronize with drop completion for python tests without polling for
+            // log messages
+            if let Ok(result) = result.as_mut() {
+                let mut js = tokio::task::JoinSet::new();
+                for layer in std::mem::take(&mut result.doomed_layers) {
+                    js.spawn(layer.wait_drop());
+                }
+                tracing::info!(
+                    total = js.len(),
+                    "starting to wait for the gc'd layers to be dropped"
+                );
+                while let Some(res) = js.join_next().await {
+                    res.expect("wait_drop should not panic");
+                }
+            }
+
+            let timeline = tenant.get_timeline(timeline_id, false).ok();
+            let rtc = timeline.as_ref().map(|x| &x.remote_client);
+
+            if let Some(rtc) = rtc {
+                // layer drops schedule actions on remote timeline client to actually do the
+                // deletions; don't care about the shutdown error, just exit fast
+                drop(rtc.wait_completion().await);
+            }
+        }
+
+        result.map_err(|e| match e {
+            GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
+            GcError::TimelineNotFound => {
+                ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
+            }
+            other => ApiError::InternalServerError(anyhow::anyhow!(other)),
+        })
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -2343,7 +2417,7 @@ enum TenantSlotDropError {
 /// Errors that can happen any time we are walking the tenant map to try and acquire
 /// the TenantSlot for a particular tenant.
 #[derive(Debug, thiserror::Error)]
-pub enum TenantMapError {
+pub(crate) enum TenantMapError {
     // Tried to read while initializing
     #[error("tenant map is still initializing")]
     StillInitializing,
@@ -2373,7 +2447,7 @@ pub enum TenantMapError {
 /// The `old_value` may be dropped before the SlotGuard is dropped, by calling
 /// `drop_old_value`.  It is an error to call this without shutting down
 /// the conents of `old_value`.
-pub struct SlotGuard {
+pub(crate) struct SlotGuard {
     tenant_shard_id: TenantShardId,
     old_value: Option<TenantSlot>,
     upserted: bool,
@@ -2766,81 +2840,6 @@ use {
     utils::http::error::ApiError,
 };
 
-#[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
-pub(crate) async fn immediate_gc(
-    tenant_shard_id: TenantShardId,
-    timeline_id: TimelineId,
-    gc_req: TimelineGcRequest,
-    cancel: CancellationToken,
-    ctx: &RequestContext,
-) -> Result<GcResult, ApiError> {
-    let tenant = {
-        let guard = TENANTS.read().unwrap();
-        guard
-            .get(&tenant_shard_id)
-            .cloned()
-            .with_context(|| format!("tenant {tenant_shard_id}"))
-            .map_err(|e| ApiError::NotFound(e.into()))?
-    };
-
-    let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
-    // Use tenant's pitr setting
-    let pitr = tenant.get_pitr_interval();
-
-    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
-
-    // Run in task_mgr to avoid race with tenant_detach operation
-    let ctx: RequestContext =
-        ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
-
-    let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
-
-    fail::fail_point!("immediate_gc_task_pre");
-
-    #[allow(unused_mut)]
-    let mut result = tenant
-        .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
-        .await;
-    // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
-    // better once the types support it.
-
-    #[cfg(feature = "testing")]
-    {
-        // we need to synchronize with drop completion for python tests without polling for
-        // log messages
-        if let Ok(result) = result.as_mut() {
-            let mut js = tokio::task::JoinSet::new();
-            for layer in std::mem::take(&mut result.doomed_layers) {
-                js.spawn(layer.wait_drop());
-            }
-            tracing::info!(
-                total = js.len(),
-                "starting to wait for the gc'd layers to be dropped"
-            );
-            while let Some(res) = js.join_next().await {
-                res.expect("wait_drop should not panic");
-            }
-        }
-
-        let timeline = tenant.get_timeline(timeline_id, false).ok();
-        let rtc = timeline.as_ref().map(|x| &x.remote_client);
-
-        if let Some(rtc) = rtc {
-            // layer drops schedule actions on remote timeline client to actually do the
-            // deletions; don't care about the shutdown error, just exit fast
-            drop(rtc.wait_completion().await);
-        }
-    }
-
-    result.map_err(|e| match e {
-        GcError::TenantCancelled | GcError::TimelineCancelled => ApiError::ShuttingDown,
-        GcError::TimelineNotFound => {
-            ApiError::NotFound(anyhow::anyhow!("Timeline not found").into())
-        }
-        other => ApiError::InternalServerError(anyhow::anyhow!(other)),
-    })
-}
-
 #[cfg(test)]
 mod tests {
     use std::collections::BTreeMap;

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -178,6 +178,7 @@ async fn download_object<'a>(
                 destination_file
                     .flush()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("flush source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 
@@ -185,6 +186,7 @@ async fn download_object<'a>(
                 destination_file
                     .sync_all()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("failed to fsync source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 
@@ -232,6 +234,7 @@ async fn download_object<'a>(
                 destination_file
                     .sync_all()
                     .await
+                    .maybe_fatal_err("download_object sync_all")
                     .with_context(|| format!("failed to fsync source file at {dst_path}"))
                     .map_err(DownloadError::Other)?;
 

pageserver/src/tenant/storage_layer.rs

@@ -1,13 +1,13 @@
 //! Common traits and structs for layers
 
 pub mod delta_layer;
+pub mod filter_iterator;
 pub mod image_layer;
 pub mod inmemory_layer;
 pub(crate) mod layer;
 mod layer_desc;
 mod layer_name;
 pub mod merge_iterator;
-
 pub mod split_writer;
 
 use crate::context::{AccessStatsBehavior, RequestContext};

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -39,12 +39,12 @@ use crate::tenant::disk_btree::{
 use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
-    BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
-    VectoredReadCoalesceMode, VectoredReadPlanner,
+    BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
+    VectoredReadPlanner,
 };
 use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
-use crate::virtual_file::{self, VirtualFile};
+use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::{walrecord, TEMP_FILE_SUFFIX};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{anyhow, bail, ensure, Context, Result};
@@ -589,7 +589,9 @@ impl DeltaLayerWriterInner {
         );
 
         // fsync the file
-        file.sync_all().await?;
+        file.sync_all()
+            .await
+            .maybe_fatal_err("delta_layer sync_all")?;
 
         trace!("created delta layer {}", self.path);
 
@@ -1021,13 +1023,30 @@ impl DeltaLayerInner {
                     continue;
                 }
             };
-
+            let view = BufView::new_slice(&blobs_buf.buf);
             for meta in blobs_buf.blobs.iter().rev() {
                 if Some(meta.meta.key) == ignore_key_with_err {
                     continue;
                 }
+                let blob_read = meta.read(&view).await;
+                let blob_read = match blob_read {
+                    Ok(buf) => buf,
+                    Err(e) => {
+                        reconstruct_state.on_key_error(
+                            meta.meta.key,
+                            PageReconstructError::Other(anyhow!(e).context(format!(
+                                "Failed to decompress blob from virtual file {}",
+                                self.file.path,
+                            ))),
+                        );
+
+                        ignore_key_with_err = Some(meta.meta.key);
+                        continue;
+                    }
+                };
+
+                let value = Value::des(&blob_read);
 
-                let value = Value::des(&blobs_buf.buf[meta.start..meta.end]);
                 let value = match value {
                     Ok(v) => v,
                     Err(e) => {
@@ -1116,7 +1135,7 @@ impl DeltaLayerInner {
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         use crate::tenant::vectored_blob_io::{
-            BlobMeta, VectoredReadBuilder, VectoredReadExtended,
+            BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
         };
         use futures::stream::TryStreamExt;
 
@@ -1166,8 +1185,8 @@ impl DeltaLayerInner {
 
         let mut prev: Option<(Key, Lsn, BlobRef)> = None;
 
-        let mut read_builder: Option<VectoredReadBuilder> = None;
-        let read_mode = VectoredReadCoalesceMode::get();
+        let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
+        let align = virtual_file::get_io_buffer_alignment();
 
         let max_read_size = self
             .max_vectored_read_bytes
@@ -1211,12 +1230,12 @@ impl DeltaLayerInner {
                 {
                     None
                 } else {
-                    read_builder.replace(VectoredReadBuilder::new(
+                    read_builder.replace(ChunkedVectoredReadBuilder::new(
                         offsets.start.pos(),
                         offsets.end.pos(),
                         meta,
                         max_read_size,
-                        read_mode,
+                        align,
                     ))
                 }
             } else {
@@ -1243,21 +1262,21 @@ impl DeltaLayerInner {
                 buf.reserve(read.size());
                 let res = reader.read_blobs(&read, buf, ctx).await?;
 
+                let view = BufView::new_slice(&res.buf);
+
                 for blob in res.blobs {
                     let key = blob.meta.key;
                     let lsn = blob.meta.lsn;
-                    let data = &res.buf[blob.start..blob.end];
+
+                    let data = blob.read(&view).await?;
 
                     #[cfg(debug_assertions)]
-                    Value::des(data)
+                    Value::des(&data)
                         .with_context(|| {
                             format!(
-                                "blob failed to deserialize for {}@{}, {}..{}: {:?}",
-                                blob.meta.key,
-                                blob.meta.lsn,
-                                blob.start,
-                                blob.end,
-                                utils::Hex(data)
+                                "blob failed to deserialize for {}: {:?}",
+                                blob,
+                                utils::Hex(&data)
                             )
                         })
                         .unwrap();
@@ -1265,15 +1284,15 @@ impl DeltaLayerInner {
                     // is it an image or will_init walrecord?
                     // FIXME: this could be handled by threading the BlobRef to the
                     // VectoredReadBuilder
-                    let will_init = crate::repository::ValueBytes::will_init(data)
+                    let will_init = crate::repository::ValueBytes::will_init(&data)
                         .inspect_err(|_e| {
                             #[cfg(feature = "testing")]
-                            tracing::error!(data=?utils::Hex(data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
+                            tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
                         })
                         .unwrap_or(false);
 
                     per_blob_copy.clear();
-                    per_blob_copy.extend_from_slice(data);
+                    per_blob_copy.extend_from_slice(&data);
 
                     let (tmp, res) = writer
                         .put_value_bytes(
@@ -1538,8 +1557,11 @@ impl<'a> DeltaLayerIterator<'a> {
             .read_blobs(&plan, buf, self.ctx)
             .await?;
         let frozen_buf = blobs_buf.buf.freeze();
+        let view = BufView::new_bytes(frozen_buf);
         for meta in blobs_buf.blobs.iter() {
-            let value = Value::des(&frozen_buf[meta.start..meta.end])?;
+            let blob_read = meta.read(&view).await?;
+            let value = Value::des(&blob_read)?;
+
             next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
         }
         self.key_values_batch = next_batch;
@@ -1916,9 +1938,13 @@ pub(crate) mod test {
                 let blobs_buf = vectored_blob_reader
                     .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
                     .await?;
+                let view = BufView::new_slice(&blobs_buf.buf);
                 for meta in blobs_buf.blobs.iter() {
-                    let value = &blobs_buf.buf[meta.start..meta.end];
-                    assert_eq!(value, entries_meta.index[&(meta.meta.key, meta.meta.lsn)]);
+                    let value = meta.read(&view).await?;
+                    assert_eq!(
+                        &value[..],
+                        &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
+                    );
                 }
 
                 buf = Some(blobs_buf.buf);

pageserver/src/tenant/storage_layer/filter_iterator.rs

@@ -0,0 +1,205 @@
+use std::ops::Range;
+
+use anyhow::bail;
+use pageserver_api::{
+    key::Key,
+    keyspace::{KeySpace, SparseKeySpace},
+};
+use utils::lsn::Lsn;
+
+use crate::repository::Value;
+
+use super::merge_iterator::MergeIterator;
+
+/// A filter iterator over merge iterators (and can be easily extended to other types of iterators).
+///
+/// The iterator will skip any keys not included in the keyspace filter. In other words, the keyspace filter contains the keys
+/// to be retained.
+pub struct FilterIterator<'a> {
+    inner: MergeIterator<'a>,
+    retain_key_filters: Vec<Range<Key>>,
+    current_filter_idx: usize,
+}
+
+impl<'a> FilterIterator<'a> {
+    pub fn create(
+        inner: MergeIterator<'a>,
+        dense_keyspace: KeySpace,
+        sparse_keyspace: SparseKeySpace,
+    ) -> anyhow::Result<Self> {
+        let mut retain_key_filters = Vec::new();
+        retain_key_filters.extend(dense_keyspace.ranges);
+        retain_key_filters.extend(sparse_keyspace.0.ranges);
+        retain_key_filters.sort_by(|a, b| a.start.cmp(&b.start));
+        // Verify key filters are non-overlapping and sorted
+        for window in retain_key_filters.windows(2) {
+            if window[0].end > window[1].start {
+                bail!(
+                    "Key filters are overlapping: {:?} and {:?}",
+                    window[0],
+                    window[1]
+                );
+            }
+        }
+        Ok(Self {
+            inner,
+            retain_key_filters,
+            current_filter_idx: 0,
+        })
+    }
+
+    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
+        while let Some(item) = self.inner.next().await? {
+            while self.current_filter_idx < self.retain_key_filters.len()
+                && item.0 >= self.retain_key_filters[self.current_filter_idx].end
+            {
+                // [filter region]    [filter region]     [filter region]
+                //                                     ^ item
+                //                    ^ current filter
+                self.current_filter_idx += 1;
+                // [filter region]    [filter region]     [filter region]
+                //                                     ^ item
+                //                                        ^ current filter
+            }
+            if self.current_filter_idx >= self.retain_key_filters.len() {
+                // We already exhausted all filters, so we should return now
+                // [filter region] [filter region] [filter region]
+                //                                                    ^ item
+                //                                                 ^ current filter (nothing)
+                return Ok(None);
+            }
+            if self.retain_key_filters[self.current_filter_idx].contains(&item.0) {
+                // [filter region]    [filter region]     [filter region]
+                //                                              ^ item
+                //                                        ^ current filter
+                return Ok(Some(item));
+            }
+            // If the key is not contained in the key retaining filters, continue to the next item.
+            // [filter region]    [filter region]     [filter region]
+            //                                     ^ item
+            //                                        ^ current filter
+        }
+        Ok(None)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use itertools::Itertools;
+    use pageserver_api::key::Key;
+    use utils::lsn::Lsn;
+
+    use crate::{
+        tenant::{
+            harness::{TenantHarness, TIMELINE_ID},
+            storage_layer::delta_layer::test::produce_delta_layer,
+        },
+        DEFAULT_PG_VERSION,
+    };
+
+    async fn assert_filter_iter_equal(
+        filter_iter: &mut FilterIterator<'_>,
+        expect: &[(Key, Lsn, Value)],
+    ) {
+        let mut expect_iter = expect.iter();
+        loop {
+            let o1 = filter_iter.next().await.unwrap();
+            let o2 = expect_iter.next();
+            assert_eq!(o1.is_some(), o2.is_some());
+            if o1.is_none() && o2.is_none() {
+                break;
+            }
+            let (k1, l1, v1) = o1.unwrap();
+            let (k2, l2, v2) = o2.unwrap();
+            assert_eq!(&k1, k2);
+            assert_eq!(l1, *l2);
+            assert_eq!(&v1, v2);
+        }
+    }
+
+    #[tokio::test]
+    async fn filter_keyspace_iterator() {
+        use crate::repository::Value;
+        use bytes::Bytes;
+
+        let harness = TenantHarness::create("filter_iterator_filter_keyspace_iterator")
+            .await
+            .unwrap();
+        let (tenant, ctx) = harness.load().await;
+
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await
+            .unwrap();
+
+        fn get_key(id: u32) -> Key {
+            let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
+            key.field6 = id;
+            key
+        }
+        const N: usize = 100;
+        let test_deltas1 = (0..N)
+            .map(|idx| {
+                (
+                    get_key(idx as u32),
+                    Lsn(0x20 * ((idx as u64) % 10 + 1)),
+                    Value::Image(Bytes::from(format!("img{idx:05}"))),
+                )
+            })
+            .collect_vec();
+        let resident_layer_1 = produce_delta_layer(&tenant, &tline, test_deltas1.clone(), &ctx)
+            .await
+            .unwrap();
+
+        let merge_iter = MergeIterator::create(
+            &[resident_layer_1.get_as_delta(&ctx).await.unwrap()],
+            &[],
+            &ctx,
+        );
+
+        let mut filter_iter = FilterIterator::create(
+            merge_iter,
+            KeySpace {
+                ranges: vec![
+                    get_key(5)..get_key(10),
+                    get_key(20)..get_key(30),
+                    get_key(90)..get_key(110),
+                    get_key(1000)..get_key(2000),
+                ],
+            },
+            SparseKeySpace(KeySpace::default()),
+        )
+        .unwrap();
+        let mut result = Vec::new();
+        result.extend(test_deltas1[5..10].iter().cloned());
+        result.extend(test_deltas1[20..30].iter().cloned());
+        result.extend(test_deltas1[90..100].iter().cloned());
+        assert_filter_iter_equal(&mut filter_iter, &result).await;
+
+        let merge_iter = MergeIterator::create(
+            &[resident_layer_1.get_as_delta(&ctx).await.unwrap()],
+            &[],
+            &ctx,
+        );
+
+        let mut filter_iter = FilterIterator::create(
+            merge_iter,
+            KeySpace {
+                ranges: vec![
+                    get_key(0)..get_key(10),
+                    get_key(20)..get_key(30),
+                    get_key(90)..get_key(95),
+                ],
+            },
+            SparseKeySpace(KeySpace::default()),
+        )
+        .unwrap();
+        let mut result = Vec::new();
+        result.extend(test_deltas1[0..10].iter().cloned());
+        result.extend(test_deltas1[20..30].iter().cloned());
+        result.extend(test_deltas1[90..95].iter().cloned());
+        assert_filter_iter_equal(&mut filter_iter, &result).await;
+    }
+}

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -36,11 +36,12 @@ use crate::tenant::disk_btree::{
 };
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
-    BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead, VectoredReadPlanner,
+    BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
+    VectoredReadPlanner,
 };
 use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
-use crate::virtual_file::{self, VirtualFile};
+use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
 use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::{Bytes, BytesMut};
@@ -547,15 +548,15 @@ impl ImageLayerInner {
 
             let buf = BytesMut::with_capacity(buf_size);
             let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
-
             let frozen_buf = blobs_buf.buf.freeze();
+            let view = BufView::new_bytes(frozen_buf);
 
             for meta in blobs_buf.blobs.iter() {
-                let img_buf = frozen_buf.slice(meta.start..meta.end);
+                let img_buf = meta.read(&view).await?;
 
                 key_count += 1;
                 writer
-                    .put_image(meta.meta.key, img_buf, ctx)
+                    .put_image(meta.meta.key, img_buf.into_bytes(), ctx)
                     .await
                     .context(format!("Storing key {}", meta.meta.key))?;
             }
@@ -602,13 +603,28 @@ impl ImageLayerInner {
             match res {
                 Ok(blobs_buf) => {
                     let frozen_buf = blobs_buf.buf.freeze();
-
+                    let view = BufView::new_bytes(frozen_buf);
                     for meta in blobs_buf.blobs.iter() {
-                        let img_buf = frozen_buf.slice(meta.start..meta.end);
+                        let img_buf = meta.read(&view).await;
+
+                        let img_buf = match img_buf {
+                            Ok(img_buf) => img_buf,
+                            Err(e) => {
+                                reconstruct_state.on_key_error(
+                                    meta.meta.key,
+                                    PageReconstructError::Other(anyhow!(e).context(format!(
+                                        "Failed to decompress blob from virtual file {}",
+                                        self.file.path,
+                                    ))),
+                                );
+
+                                continue;
+                            }
+                        };
                         reconstruct_state.update_key(
                             &meta.meta.key,
                             self.lsn,
-                            Value::Image(img_buf),
+                            Value::Image(img_buf.into_bytes()),
                         );
                     }
                 }
@@ -873,7 +889,9 @@ impl ImageLayerWriterInner {
         // set inner.file here. The first read will have to re-open it.
 
         // fsync the file
-        file.sync_all().await?;
+        file.sync_all()
+            .await
+            .maybe_fatal_err("image_layer sync_all")?;
 
         trace!("created image layer {}", self.path);
 
@@ -1025,10 +1043,15 @@ impl<'a> ImageLayerIterator<'a> {
         let blobs_buf = vectored_blob_reader
             .read_blobs(&plan, buf, self.ctx)
             .await?;
-        let frozen_buf: Bytes = blobs_buf.buf.freeze();
+        let frozen_buf = blobs_buf.buf.freeze();
+        let view = BufView::new_bytes(frozen_buf);
         for meta in blobs_buf.blobs.iter() {
-            let img_buf = frozen_buf.slice(meta.start..meta.end);
-            next_batch.push_back((meta.meta.key, self.image_layer.lsn, Value::Image(img_buf)));
+            let img_buf = meta.read(&view).await?;
+            next_batch.push_back((
+                meta.meta.key,
+                self.image_layer.lsn,
+                Value::Image(img_buf.into_bytes()),
+            ));
         }
         self.key_values_batch = next_batch;
         Ok(())

pageserver/src/tenant/storage_layer/layer.rs

@@ -439,11 +439,30 @@ impl Layer {
 
     fn record_access(&self, ctx: &RequestContext) {
         if self.0.access_stats.record_access(ctx) {
-            // Visibility was modified to Visible
-            tracing::info!(
-                "Layer {} became visible as a result of access",
-                self.0.desc.key()
-            );
+            // Visibility was modified to Visible: maybe log about this
+            match ctx.task_kind() {
+                TaskKind::CalculateSyntheticSize
+                | TaskKind::GarbageCollector
+                | TaskKind::MgmtRequest => {
+                    // This situation is expected in code paths do binary searches of the LSN space to resolve
+                    // an LSN to a timestamp, which happens during GC, during GC cutoff calculations in synthetic size,
+                    // and on-demand for certain HTTP API requests.
+                }
+                _ => {
+                    // In all other contexts, it is unusual to do I/O involving layers which are not visible at
+                    // some branch tip, so we log the fact that we are accessing something that the visibility
+                    // calculation thought should not be visible.
+                    //
+                    // This case is legal in brief time windows: for example an in-flight getpage request can hold on to a layer object
+                    // which was covered by a concurrent compaction.
+                    tracing::info!(
+                        "Layer {} became visible as a result of access",
+                        self.0.desc.key()
+                    );
+                }
+            }
+
+            // Update the timeline's visible bytes count
             if let Some(tl) = self.0.timeline.upgrade() {
                 tl.metrics
                     .visible_physical_size_gauge

pageserver/src/tenant/storage_layer/layer/tests.rs

@@ -1025,6 +1025,15 @@ fn access_stats() {
     assert_eq!(access_stats.latest_activity(), lowres_time(atime));
     access_stats.set_visibility(LayerVisibilityHint::Visible);
     assert_eq!(access_stats.latest_activity(), lowres_time(atime));
+
+    // Recording access implicitly makes layer visible, if it wasn't already
+    let atime = UNIX_EPOCH + Duration::from_secs(2200000000);
+    access_stats.set_visibility(LayerVisibilityHint::Covered);
+    assert_eq!(access_stats.visibility(), LayerVisibilityHint::Covered);
+    assert!(access_stats.record_access_at(atime));
+    access_stats.set_visibility(LayerVisibilityHint::Visible);
+    assert!(!access_stats.record_access_at(atime));
+    access_stats.set_visibility(LayerVisibilityHint::Visible);
 }
 
 #[test]

pageserver/src/tenant/tasks.rs

@@ -163,8 +163,6 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
     // How many errors we have seen consequtively
     let mut error_run_count = 0;
 
-    let mut last_throttle_flag_reset_at = Instant::now();
-
     TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
     async {
         let ctx = RequestContext::todo_child(TaskKind::Compaction, DownloadBehavior::Download);
@@ -191,8 +189,6 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
             }
 
-
-
             let sleep_duration;
             if period == Duration::ZERO {
                 #[cfg(not(feature = "testing"))]
@@ -207,12 +203,18 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 };
 
                 // Run compaction
-                let IterationResult { output, elapsed } = iteration.run(tenant.compaction_iteration(&cancel, &ctx)).await;
+                let IterationResult { output, elapsed } = iteration
+                    .run(tenant.compaction_iteration(&cancel, &ctx))
+                    .await;
                 match output {
                     Ok(has_pending_task) => {
                         error_run_count = 0;
                         // schedule the next compaction immediately in case there is a pending compaction task
-                        sleep_duration = if has_pending_task { Duration::ZERO } else { period };
+                        sleep_duration = if has_pending_task {
+                            Duration::ZERO
+                        } else {
+                            period
+                        };
                     }
                     Err(e) => {
                         let wait_duration = backoff::exponential_backoff_duration_seconds(
@@ -233,38 +235,20 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
 
                 // the duration is recorded by performance tests by enabling debug in this function
-                tracing::debug!(elapsed_ms=elapsed.as_millis(), "compaction iteration complete");
+                tracing::debug!(
+                    elapsed_ms = elapsed.as_millis(),
+                    "compaction iteration complete"
+                );
             };
 
-
             // Perhaps we did no work and the walredo process has been idle for some time:
             // give it a chance to shut down to avoid leaving walredo process running indefinitely.
+            // TODO: move this to a separate task (housekeeping loop) that isn't affected by the back-off,
+            // so we get some upper bound guarantee on when walredo quiesce / this throttling reporting here happens.
             if let Some(walredo_mgr) = &tenant.walredo_mgr {
                 walredo_mgr.maybe_quiesce(period * 10);
             }
 
-            // TODO: move this (and walredo quiesce) to a separate task that isn't affected by the back-off,
-            // so we get some upper bound guarantee on when walredo quiesce / this throttling reporting here happens.
-            info_span!(parent: None, "timeline_get_throttle", tenant_id=%tenant.tenant_shard_id, shard_id=%tenant.tenant_shard_id.shard_slug()).in_scope(|| {
-                let now = Instant::now();
-                let prev = std::mem::replace(&mut last_throttle_flag_reset_at, now);
-                let Stats { count_accounted, count_throttled, sum_throttled_usecs } = tenant.timeline_get_throttle.reset_stats();
-                if count_throttled == 0 {
-                    return;
-                }
-                let allowed_rps = tenant.timeline_get_throttle.steady_rps();
-                let delta = now - prev;
-                info!(
-                    n_seconds=%format_args!("{:.3}",
-                    delta.as_secs_f64()),
-                    count_accounted,
-                    count_throttled,
-                    sum_throttled_usecs,
-                    allowed_rps=%format_args!("{allowed_rps:.0}"),
-                    "shard was throttled in the last n_seconds"
-                );
-            });
-
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
                 .await
@@ -346,7 +330,6 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             RequestContext::todo_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
 
         let mut first = true;
-        tenant.gc_block.set_lsn_lease_deadline(tenant.get_lsn_lease_length());
         loop {
             tokio::select! {
                 _ = cancel.cancelled() => {
@@ -437,6 +420,7 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 async fn ingest_housekeeping_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
     TENANT_TASK_EVENTS.with_label_values(&["start"]).inc();
     async {
+    let mut last_throttle_flag_reset_at = Instant::now();
         loop {
             tokio::select! {
                 _ = cancel.cancelled() => {
@@ -483,6 +467,28 @@ async fn ingest_housekeeping_loop(tenant: Arc<Tenant>, cancel: CancellationToken
                 kind: BackgroundLoopKind::IngestHouseKeeping,
             };
             iteration.run(tenant.ingest_housekeeping()).await;
+
+            // TODO: rename the background loop kind to something more generic, like, tenant housekeeping.
+            // Or just spawn another background loop for this throttle, it's not like it's super costly.
+            info_span!(parent: None, "timeline_get_throttle", tenant_id=%tenant.tenant_shard_id, shard_id=%tenant.tenant_shard_id.shard_slug()).in_scope(|| {
+                let now = Instant::now();
+                let prev = std::mem::replace(&mut last_throttle_flag_reset_at, now);
+                let Stats { count_accounted_start, count_accounted_finish, count_throttled, sum_throttled_usecs} = tenant.timeline_get_throttle.reset_stats();
+                if count_throttled == 0 {
+                    return;
+                }
+                let allowed_rps = tenant.timeline_get_throttle.steady_rps();
+                let delta = now - prev;
+                info!(
+                    n_seconds=%format_args!("{:.3}", delta.as_secs_f64()),
+                    count_accounted = count_accounted_finish,  // don't break existing log scraping
+                    count_throttled,
+                    sum_throttled_usecs,
+                    count_accounted_start, // log after pre-existing fields to not break existing log scraping
+                    allowed_rps=%format_args!("{allowed_rps:.0}"),
+                    "shard was throttled in the last n_seconds"
+                );
+            });
         }
     }
     .await;

pageserver/src/tenant/throttle.rs

@@ -24,8 +24,10 @@ use crate::{context::RequestContext, task_mgr::TaskKind};
 pub struct Throttle<M: Metric> {
     inner: ArcSwap<Inner>,
     metric: M,
-    /// will be turned into [`Stats::count_accounted`]
-    count_accounted: AtomicU64,
+    /// will be turned into [`Stats::count_accounted_start`]
+    count_accounted_start: AtomicU64,
+    /// will be turned into [`Stats::count_accounted_finish`]
+    count_accounted_finish: AtomicU64,
     /// will be turned into [`Stats::count_throttled`]
     count_throttled: AtomicU64,
     /// will be turned into [`Stats::sum_throttled_usecs`]
@@ -43,17 +45,21 @@ pub struct Observation {
     pub wait_time: Duration,
 }
 pub trait Metric {
+    fn accounting_start(&self);
+    fn accounting_finish(&self);
     fn observe_throttling(&self, observation: &Observation);
 }
 
 /// See [`Throttle::reset_stats`].
 pub struct Stats {
-    // Number of requests that were subject to throttling, i.e., requests of the configured [`Config::task_kinds`].
-    pub count_accounted: u64,
-    // Subset of the `accounted` requests that were actually throttled.
-    // Note that the numbers are stored as two independent atomics, so, there might be a slight drift.
+    /// Number of requests that started [`Throttle::throttle`] calls.
+    pub count_accounted_start: u64,
+    /// Number of requests that finished [`Throttle::throttle`] calls.
+    pub count_accounted_finish: u64,
+    /// Subset of the `accounted` requests that were actually throttled.
+    /// Note that the numbers are stored as two independent atomics, so, there might be a slight drift.
     pub count_throttled: u64,
-    // Sum of microseconds that throttled requests spent waiting for throttling.
+    /// Sum of microseconds that throttled requests spent waiting for throttling.
     pub sum_throttled_usecs: u64,
 }
 
@@ -65,7 +71,8 @@ where
         Self {
             inner: ArcSwap::new(Arc::new(Self::new_inner(config))),
             metric,
-            count_accounted: AtomicU64::new(0),
+            count_accounted_start: AtomicU64::new(0),
+            count_accounted_finish: AtomicU64::new(0),
             count_throttled: AtomicU64::new(0),
             sum_throttled_usecs: AtomicU64::new(0),
         }
@@ -117,11 +124,13 @@ where
     /// This method allows retrieving & resetting that flag.
     /// Useful for periodic reporting.
     pub fn reset_stats(&self) -> Stats {
-        let count_accounted = self.count_accounted.swap(0, Ordering::Relaxed);
+        let count_accounted_start = self.count_accounted_start.swap(0, Ordering::Relaxed);
+        let count_accounted_finish = self.count_accounted_finish.swap(0, Ordering::Relaxed);
         let count_throttled = self.count_throttled.swap(0, Ordering::Relaxed);
         let sum_throttled_usecs = self.sum_throttled_usecs.swap(0, Ordering::Relaxed);
         Stats {
-            count_accounted,
+            count_accounted_start,
+            count_accounted_finish,
             count_throttled,
             sum_throttled_usecs,
         }
@@ -139,9 +148,12 @@ where
         };
         let start = std::time::Instant::now();
 
+        self.metric.accounting_start();
+        self.count_accounted_start.fetch_add(1, Ordering::Relaxed);
         let did_throttle = inner.rate_limiter.acquire(key_count).await;
+        self.count_accounted_finish.fetch_add(1, Ordering::Relaxed);
+        self.metric.accounting_finish();
 
-        self.count_accounted.fetch_add(1, Ordering::Relaxed);
         if did_throttle {
             self.count_throttled.fetch_add(1, Ordering::Relaxed);
             let now = Instant::now();

pageserver/src/tenant/timeline.rs

@@ -66,6 +66,7 @@ use std::{
 use crate::{
     aux_file::AuxFileSizeEstimator,
     tenant::{
+        config::AttachmentMode,
         layer_map::{LayerMap, SearchResult},
         metadata::TimelineMetadata,
         storage_layer::{inmemory_layer::IndexEntry, PersistentLayerDesc},
@@ -112,7 +113,7 @@ use pageserver_api::reltag::RelTag;
 use pageserver_api::shard::ShardIndex;
 
 use postgres_connection::PgConnectionConfig;
-use postgres_ffi::to_pg_timestamp;
+use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
 use utils::{
     completion,
     generation::Generation,
@@ -196,9 +197,8 @@ fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
 /// The outward-facing resources required to build a Timeline
 pub struct TimelineResources {
     pub remote_client: RemoteTimelineClient,
-    pub timeline_get_throttle: Arc<
-        crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
-    >,
+    pub timeline_get_throttle:
+        Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
 }
 
@@ -406,9 +406,8 @@ pub struct Timeline {
     gc_lock: tokio::sync::Mutex<()>,
 
     /// Cloned from [`super::Tenant::timeline_get_throttle`] on construction.
-    timeline_get_throttle: Arc<
-        crate::tenant::throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>,
-    >,
+    timeline_get_throttle:
+        Arc<crate::tenant::throttle::Throttle<crate::metrics::tenant_throttling::TimelineGet>>,
 
     /// Keep aux directory cache to avoid it's reconstruction on each update
     pub(crate) aux_files: tokio::sync::Mutex<AuxFilesState>,
@@ -1326,27 +1325,53 @@ impl Timeline {
         Ok(())
     }
 
-    /// Obtains a temporary lease blocking garbage collection for the given LSN.
-    ///
-    /// This function will error if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is also
-    /// no existing lease to renew. If there is an existing lease in the map, the lease will be renewed only if
-    /// the request extends the lease. The returned lease is therefore the maximum between the existing lease and
-    /// the requesting lease.
-    pub(crate) fn make_lsn_lease(
+    /// Initializes an LSN lease. The function will return an error if the requested LSN is less than the `latest_gc_cutoff_lsn`.
+    pub(crate) fn init_lsn_lease(
         &self,
         lsn: Lsn,
         length: Duration,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<LsnLease> {
+        self.make_lsn_lease(lsn, length, true, ctx)
+    }
+
+    /// Renews a lease at a particular LSN. The requested LSN is not validated against the `latest_gc_cutoff_lsn` when we are in the grace period.
+    pub(crate) fn renew_lsn_lease(
+        &self,
+        lsn: Lsn,
+        length: Duration,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<LsnLease> {
+        self.make_lsn_lease(lsn, length, false, ctx)
+    }
+
+    /// Obtains a temporary lease blocking garbage collection for the given LSN.
+    ///
+    /// If we are in `AttachedSingle` mode and is not blocked by the lsn lease deadline, this function will error
+    /// if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is no existing request present.
+    ///
+    /// If there is an existing lease in the map, the lease will be renewed only if the request extends the lease.
+    /// The returned lease is therefore the maximum between the existing lease and the requesting lease.
+    fn make_lsn_lease(
+        &self,
+        lsn: Lsn,
+        length: Duration,
+        init: bool,
         _ctx: &RequestContext,
     ) -> anyhow::Result<LsnLease> {
         let lease = {
+            // Normalize the requested LSN to be aligned, and move to the first record
+            // if it points to the beginning of the page (header).
+            let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
+
             let mut gc_info = self.gc_info.write().unwrap();
 
             let valid_until = SystemTime::now() + length;
 
             let entry = gc_info.leases.entry(lsn);
 
-            let lease = {
-                if let Entry::Occupied(mut occupied) = entry {
+            match entry {
+                Entry::Occupied(mut occupied) => {
                     let existing_lease = occupied.get_mut();
                     if valid_until > existing_lease.valid_until {
                         existing_lease.valid_until = valid_until;
@@ -1358,20 +1383,28 @@ impl Timeline {
                     }
 
                     existing_lease.clone()
-                } else {
-                    // Reject already GC-ed LSN (lsn < latest_gc_cutoff)
-                    let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
-                    if lsn < *latest_gc_cutoff_lsn {
-                        bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
+                }
+                Entry::Vacant(vacant) => {
+                    // Reject already GC-ed LSN (lsn < latest_gc_cutoff) if we are in AttachedSingle and
+                    // not blocked by the lsn lease deadline.
+                    let validate = {
+                        let conf = self.tenant_conf.load();
+                        conf.location.attach_mode == AttachmentMode::Single
+                            && !conf.is_gc_blocked_by_lsn_lease_deadline()
+                    };
+
+                    if init || validate {
+                        let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
+                        if lsn < *latest_gc_cutoff_lsn {
+                            bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
+                        }
                     }
 
                     let dt: DateTime<Utc> = valid_until.into();
                     info!("lease created, valid until {}", dt);
-                    entry.or_insert(LsnLease { valid_until }).clone()
+                    vacant.insert(LsnLease { valid_until }).clone()
                 }
-            };
-
-            lease
+            }
         };
 
         Ok(lease)
@@ -1948,8 +1981,6 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
     }
 
-    // TODO(yuchen): remove unused flag after implementing https://github.com/neondatabase/neon/issues/8072
-    #[allow(unused)]
     pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
         let tenant_conf = self.tenant_conf.load();
         tenant_conf
@@ -3599,7 +3630,7 @@ impl Timeline {
                     ctx,
                 )
                 .await
-                .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
+                .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
 
             if self.cancel.is_cancelled() {
                 return Err(FlushLayerError::Cancelled);
@@ -3838,16 +3869,20 @@ impl Timeline {
         partition_size: u64,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<((KeyPartitioning, SparseKeyPartitioning), Lsn)> {
+    ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
         let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
             // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
             // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
             // and hence before the compaction task starts.
-            anyhow::bail!("repartition() called concurrently, this should not happen");
+            return Err(CompactionError::Other(anyhow!(
+                "repartition() called concurrently, this should not happen"
+            )));
         };
         let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
         if lsn < *partition_lsn {
-            anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
+            return Err(CompactionError::Other(anyhow!(
+                "repartition() called with LSN going backwards, this should not happen"
+            )));
         }
 
         let distance = lsn.0 - partition_lsn.0;
@@ -4316,7 +4351,9 @@ impl Timeline {
         timer.stop_and_record();
 
         // Creating image layers may have caused some previously visible layers to be covered
-        self.update_layer_visibility().await?;
+        if !image_layers.is_empty() {
+            self.update_layer_visibility().await?;
+        }
 
         Ok(image_layers)
     }
@@ -4447,6 +4484,12 @@ pub(crate) enum CompactionError {
     Other(anyhow::Error),
 }
 
+impl CompactionError {
+    pub fn is_cancelled(&self) -> bool {
+        matches!(self, CompactionError::ShuttingDown)
+    }
+}
+
 impl From<CollectKeySpaceError> for CompactionError {
     fn from(err: CollectKeySpaceError) -> Self {
         match err {

pageserver/src/tenant/timeline/compaction.rs

@@ -31,6 +31,7 @@ use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder}
 use crate::page_cache;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::remote_timeline_client::WaitCompletionError;
+use crate::tenant::storage_layer::filter_iterator::FilterIterator;
 use crate::tenant::storage_layer::merge_iterator::MergeIterator;
 use crate::tenant::storage_layer::split_writer::{
     SplitDeltaLayerWriter, SplitImageLayerWriter, SplitWriterResult,
@@ -389,7 +390,7 @@ impl Timeline {
                 // error but continue.
                 //
                 // Suppress error when it's due to cancellation
-                if !self.cancel.is_cancelled() {
+                if !self.cancel.is_cancelled() && !err.is_cancelled() {
                     tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
                 }
                 (1, false)
@@ -1772,6 +1773,7 @@ impl Timeline {
             gc_cutoff,
             lowest_retain_lsn
         );
+
         // Step 1: (In the future) construct a k-merge iterator over all layers. For now, simply collect all keys + LSNs.
         // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
         let mut lsn_split_point = BTreeSet::new(); // TODO: use a better data structure (range tree / range set?)
@@ -1820,7 +1822,12 @@ impl Timeline {
                 image_layers.push(layer);
             }
         }
-        let mut merge_iter = MergeIterator::create(&delta_layers, &image_layers, ctx);
+        let (dense_ks, sparse_ks) = self.collect_gc_compaction_keyspace().await?;
+        let mut merge_iter = FilterIterator::create(
+            MergeIterator::create(&delta_layers, &image_layers, ctx),
+            dense_ks,
+            sparse_ks,
+        )?;
         // Step 2: Produce images+deltas. TODO: ensure newly-produced delta does not overlap with other deltas.
         // Data of the same key.
         let mut accumulated_values = Vec::new();

pageserver/src/tenant/timeline/eviction_task.rs

@@ -30,8 +30,8 @@ use crate::{
     pgdatadir_mapping::CollectKeySpaceError,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     tenant::{
-        storage_layer::LayerVisibilityHint, tasks::BackgroundLoopKind, timeline::EvictionError,
-        LogicalSizeCalculationCause, Tenant,
+        size::CalculateSyntheticSizeError, storage_layer::LayerVisibilityHint,
+        tasks::BackgroundLoopKind, timeline::EvictionError, LogicalSizeCalculationCause, Tenant,
     },
 };
 
@@ -557,6 +557,8 @@ impl Timeline {
             gather_result = gather => {
                 match gather_result {
                     Ok(_) => {},
+                    // It can happen sometimes that we hit this instead of the cancellation token firing above
+                    Err(CalculateSyntheticSizeError::Cancelled) => {}
                     Err(e) => {
                         // We don't care about the result, but, if it failed, we should log it,
                         // since consumption metric might be hitting the cached value and

pageserver/src/tenant/vectored_blob_io.rs

@@ -16,8 +16,9 @@
 //! Note that the vectored blob api does *not* go through the page cache.
 
 use std::collections::BTreeMap;
+use std::ops::Deref;
 
-use bytes::BytesMut;
+use bytes::{Bytes, BytesMut};
 use pageserver_api::key::Key;
 use tokio::io::AsyncWriteExt;
 use tokio_epoll_uring::BoundedBuf;
@@ -35,11 +36,123 @@ pub struct BlobMeta {
     pub lsn: Lsn,
 }
 
-/// Blob offsets into [`VectoredBlobsBuf::buf`]
+/// A view into the vectored blobs read buffer.
+#[derive(Clone, Debug)]
+pub(crate) enum BufView<'a> {
+    Slice(&'a [u8]),
+    Bytes(bytes::Bytes),
+}
+
+impl<'a> BufView<'a> {
+    /// Creates a new slice-based view on the blob.
+    pub fn new_slice(slice: &'a [u8]) -> Self {
+        Self::Slice(slice)
+    }
+
+    /// Creates a new [`bytes::Bytes`]-based view on the blob.
+    pub fn new_bytes(bytes: bytes::Bytes) -> Self {
+        Self::Bytes(bytes)
+    }
+
+    /// Convert the view into `Bytes`.
+    ///
+    /// If using slice as the underlying storage, the copy will be an O(n) operation.
+    pub fn into_bytes(self) -> Bytes {
+        match self {
+            BufView::Slice(slice) => Bytes::copy_from_slice(slice),
+            BufView::Bytes(bytes) => bytes,
+        }
+    }
+
+    /// Creates a sub-view of the blob based on the range.
+    fn view(&self, range: std::ops::Range<usize>) -> Self {
+        match self {
+            BufView::Slice(slice) => BufView::Slice(&slice[range]),
+            BufView::Bytes(bytes) => BufView::Bytes(bytes.slice(range)),
+        }
+    }
+}
+
+impl<'a> Deref for BufView<'a> {
+    type Target = [u8];
+
+    fn deref(&self) -> &Self::Target {
+        match self {
+            BufView::Slice(slice) => slice,
+            BufView::Bytes(bytes) => bytes,
+        }
+    }
+}
+
+impl<'a> AsRef<[u8]> for BufView<'a> {
+    fn as_ref(&self) -> &[u8] {
+        match self {
+            BufView::Slice(slice) => slice,
+            BufView::Bytes(bytes) => bytes.as_ref(),
+        }
+    }
+}
+
+impl<'a> From<&'a [u8]> for BufView<'a> {
+    fn from(value: &'a [u8]) -> Self {
+        Self::new_slice(value)
+    }
+}
+
+impl From<Bytes> for BufView<'_> {
+    fn from(value: Bytes) -> Self {
+        Self::new_bytes(value)
+    }
+}
+
+/// Blob offsets into [`VectoredBlobsBuf::buf`]. The byte ranges is potentially compressed,
+/// subject to [`VectoredBlob::compression_bits`].
 pub struct VectoredBlob {
-    pub start: usize,
-    pub end: usize,
+    /// Blob metadata.
     pub meta: BlobMeta,
+    /// Start offset.
+    start: usize,
+    /// End offset.
+    end: usize,
+    /// Compression used on the the blob.
+    compression_bits: u8,
+}
+
+impl VectoredBlob {
+    /// Reads a decompressed view of the blob.
+    pub(crate) async fn read<'a>(&self, buf: &BufView<'a>) -> Result<BufView<'a>, std::io::Error> {
+        let view = buf.view(self.start..self.end);
+
+        match self.compression_bits {
+            BYTE_UNCOMPRESSED => Ok(view),
+            BYTE_ZSTD => {
+                let mut decompressed_vec = Vec::new();
+                let mut decoder =
+                    async_compression::tokio::write::ZstdDecoder::new(&mut decompressed_vec);
+                decoder.write_all(&view).await?;
+                decoder.flush().await?;
+                // Zero-copy conversion from `Vec` to `Bytes`
+                Ok(BufView::new_bytes(Bytes::from(decompressed_vec)))
+            }
+            bits => {
+                let error = std::io::Error::new(
+                    std::io::ErrorKind::InvalidData,
+                    format!("Failed to decompress blob for {}@{}, {}..{}: invalid compression byte {bits:x}", self.meta.key, self.meta.lsn, self.start, self.end),
+                );
+                Err(error)
+            }
+        }
+    }
+}
+
+impl std::fmt::Display for VectoredBlob {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "{}@{}, {}..{}",
+            self.meta.key, self.meta.lsn, self.start, self.end
+        )
+    }
 }
 
 /// Return type of [`VectoredBlobReader::read_blobs`]
@@ -72,171 +185,7 @@ pub(crate) enum VectoredReadExtended {
     No,
 }
 
-#[derive(Copy, Clone, Debug, PartialEq, Eq)]
-pub enum VectoredReadCoalesceMode {
-    /// Only coalesce exactly adjacent reads.
-    AdjacentOnly,
-    /// In addition to adjacent reads, also consider reads whose corresponding
-    /// `end` and `start` offsets reside at the same chunk.
-    Chunked(usize),
-}
-
-impl VectoredReadCoalesceMode {
-    /// [`AdjacentVectoredReadBuilder`] is used if alignment requirement is 0,
-    /// whereas [`ChunkedVectoredReadBuilder`] is used for alignment requirement 1 and higher.
-    pub(crate) fn get() -> Self {
-        let align = virtual_file::get_io_buffer_alignment_raw();
-        if align == 0 {
-            VectoredReadCoalesceMode::AdjacentOnly
-        } else {
-            VectoredReadCoalesceMode::Chunked(align)
-        }
-    }
-}
-
-pub(crate) enum VectoredReadBuilder {
-    Adjacent(AdjacentVectoredReadBuilder),
-    Chunked(ChunkedVectoredReadBuilder),
-}
-
-impl VectoredReadBuilder {
-    fn new_impl(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: Option<usize>,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        match mode {
-            VectoredReadCoalesceMode::AdjacentOnly => Self::Adjacent(
-                AdjacentVectoredReadBuilder::new(start_offset, end_offset, meta, max_read_size),
-            ),
-            VectoredReadCoalesceMode::Chunked(chunk_size) => {
-                Self::Chunked(ChunkedVectoredReadBuilder::new(
-                    start_offset,
-                    end_offset,
-                    meta,
-                    max_read_size,
-                    chunk_size,
-                ))
-            }
-        }
-    }
-
-    pub(crate) fn new(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: usize,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        Self::new_impl(start_offset, end_offset, meta, Some(max_read_size), mode)
-    }
-
-    pub(crate) fn new_streaming(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        mode: VectoredReadCoalesceMode,
-    ) -> Self {
-        Self::new_impl(start_offset, end_offset, meta, None, mode)
-    }
-
-    pub(crate) fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.extend(start, end, meta),
-            VectoredReadBuilder::Chunked(builder) => builder.extend(start, end, meta),
-        }
-    }
-
-    pub(crate) fn build(self) -> VectoredRead {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.build(),
-            VectoredReadBuilder::Chunked(builder) => builder.build(),
-        }
-    }
-
-    pub(crate) fn size(&self) -> usize {
-        match self {
-            VectoredReadBuilder::Adjacent(builder) => builder.size(),
-            VectoredReadBuilder::Chunked(builder) => builder.size(),
-        }
-    }
-}
-
-pub(crate) struct AdjacentVectoredReadBuilder {
-    /// Start offset of the read.
-    start: u64,
-    // End offset of the read.
-    end: u64,
-    /// Start offset and metadata for each blob in this read
-    blobs_at: VecMap<u64, BlobMeta>,
-    max_read_size: Option<usize>,
-}
-
-impl AdjacentVectoredReadBuilder {
-    /// Start building a new vectored read.
-    ///
-    /// Note that by design, this does not check against reading more than `max_read_size` to
-    /// support reading larger blobs than the configuration value. The builder will be single use
-    /// however after that.
-    pub(crate) fn new(
-        start_offset: u64,
-        end_offset: u64,
-        meta: BlobMeta,
-        max_read_size: Option<usize>,
-    ) -> Self {
-        let mut blobs_at = VecMap::default();
-        blobs_at
-            .append(start_offset, meta)
-            .expect("First insertion always succeeds");
-
-        Self {
-            start: start_offset,
-            end: end_offset,
-            blobs_at,
-            max_read_size,
-        }
-    }
-    /// Attempt to extend the current read with a new blob if the start
-    /// offset matches with the current end of the vectored read
-    /// and the resuting size is below the max read size
-    pub(crate) fn extend(&mut self, start: u64, end: u64, meta: BlobMeta) -> VectoredReadExtended {
-        tracing::trace!(start, end, "trying to extend");
-        let size = (end - start) as usize;
-        let not_limited_by_max_read_size = {
-            if let Some(max_read_size) = self.max_read_size {
-                self.size() + size <= max_read_size
-            } else {
-                true
-            }
-        };
-
-        if self.end == start && not_limited_by_max_read_size {
-            self.end = end;
-            self.blobs_at
-                .append(start, meta)
-                .expect("LSNs are ordered within vectored reads");
-
-            return VectoredReadExtended::Yes;
-        }
-
-        VectoredReadExtended::No
-    }
-
-    pub(crate) fn size(&self) -> usize {
-        (self.end - self.start) as usize
-    }
-
-    pub(crate) fn build(self) -> VectoredRead {
-        VectoredRead {
-            start: self.start,
-            end: self.end,
-            blobs_at: self.blobs_at,
-        }
-    }
-}
-
+/// A vectored read builder that tries to coalesce all reads that fits in a chunk.
 pub(crate) struct ChunkedVectoredReadBuilder {
     /// Start block number
     start_blk_no: usize,
@@ -260,7 +209,7 @@ impl ChunkedVectoredReadBuilder {
     /// Note that by design, this does not check against reading more than `max_read_size` to
     /// support reading larger blobs than the configuration value. The builder will be single use
     /// however after that.
-    pub(crate) fn new(
+    fn new_impl(
         start_offset: u64,
         end_offset: u64,
         meta: BlobMeta,
@@ -283,6 +232,25 @@ impl ChunkedVectoredReadBuilder {
         }
     }
 
+    pub(crate) fn new(
+        start_offset: u64,
+        end_offset: u64,
+        meta: BlobMeta,
+        max_read_size: usize,
+        align: usize,
+    ) -> Self {
+        Self::new_impl(start_offset, end_offset, meta, Some(max_read_size), align)
+    }
+
+    pub(crate) fn new_streaming(
+        start_offset: u64,
+        end_offset: u64,
+        meta: BlobMeta,
+        align: usize,
+    ) -> Self {
+        Self::new_impl(start_offset, end_offset, meta, None, align)
+    }
+
     /// Attempts to extend the current read with a new blob if the new blob resides in the same or the immediate next chunk.
     ///
     /// The resulting size also must be below the max read size.
@@ -361,17 +329,17 @@ pub struct VectoredReadPlanner {
 
     max_read_size: usize,
 
-    mode: VectoredReadCoalesceMode,
+    align: usize,
 }
 
 impl VectoredReadPlanner {
     pub fn new(max_read_size: usize) -> Self {
-        let mode = VectoredReadCoalesceMode::get();
+        let align = virtual_file::get_io_buffer_alignment();
         Self {
             blobs: BTreeMap::new(),
             prev: None,
             max_read_size,
-            mode,
+            align,
         }
     }
 
@@ -432,7 +400,7 @@ impl VectoredReadPlanner {
     }
 
     pub fn finish(self) -> Vec<VectoredRead> {
-        let mut current_read_builder: Option<VectoredReadBuilder> = None;
+        let mut current_read_builder: Option<ChunkedVectoredReadBuilder> = None;
         let mut reads = Vec::new();
 
         for (key, blobs_for_key) in self.blobs {
@@ -445,12 +413,12 @@ impl VectoredReadPlanner {
                 };
 
                 if extended == VectoredReadExtended::No {
-                    let next_read_builder = VectoredReadBuilder::new(
+                    let next_read_builder = ChunkedVectoredReadBuilder::new(
                         start_offset,
                         end_offset,
                         BlobMeta { key, lsn },
                         self.max_read_size,
-                        self.mode,
+                        self.align,
                     );
 
                     let prev_read_builder = current_read_builder.replace(next_read_builder);
@@ -514,7 +482,7 @@ impl<'a> VectoredBlobReader<'a> {
             );
         }
 
-        let mut buf = self
+        let buf = self
             .file
             .read_exact_at(buf.slice(0..read.size()), read.start, ctx)
             .await?
@@ -529,9 +497,6 @@ impl<'a> VectoredBlobReader<'a> {
         // of a blob is implicit: the start of the next blob if one exists
         // or the end of the read.
 
-        // Some scratch space, put here for reusing the allocation
-        let mut decompressed_vec = Vec::new();
-
         for (blob_start, meta) in blobs_at {
             let blob_start_in_buf = blob_start - start_offset;
             let first_len_byte = buf[blob_start_in_buf as usize];
@@ -557,35 +522,14 @@ impl<'a> VectoredBlobReader<'a> {
                 )
             };
 
-            let start_raw = blob_start_in_buf + size_length;
-            let end_raw = start_raw + blob_size;
-            let (start, end);
-            if compression_bits == BYTE_UNCOMPRESSED {
-                start = start_raw as usize;
-                end = end_raw as usize;
-            } else if compression_bits == BYTE_ZSTD {
-                let mut decoder =
-                    async_compression::tokio::write::ZstdDecoder::new(&mut decompressed_vec);
-                decoder
-                    .write_all(&buf[start_raw as usize..end_raw as usize])
-                    .await?;
-                decoder.flush().await?;
-                start = buf.len();
-                buf.extend_from_slice(&decompressed_vec);
-                end = buf.len();
-                decompressed_vec.clear();
-            } else {
-                let error = std::io::Error::new(
-                    std::io::ErrorKind::InvalidData,
-                    format!("invalid compression byte {compression_bits:x}"),
-                );
-                return Err(error);
-            }
+            let start = (blob_start_in_buf + size_length) as usize;
+            let end = start + blob_size as usize;
 
             metas.push(VectoredBlob {
                 start,
                 end,
                 meta: *meta,
+                compression_bits,
             });
         }
 
@@ -599,7 +543,7 @@ impl<'a> VectoredBlobReader<'a> {
 /// `handle` gets called and when the current key would just exceed the read_size and
 /// max_cnt constraints.
 pub struct StreamingVectoredReadPlanner {
-    read_builder: Option<VectoredReadBuilder>,
+    read_builder: Option<ChunkedVectoredReadBuilder>,
     // Arguments for previous blob passed into [`StreamingVectoredReadPlanner::handle`]
     prev: Option<(Key, Lsn, u64)>,
     /// Max read size per batch. This is not a strict limit. If there are [0, 100) and [100, 200), while the `max_read_size` is 150,
@@ -610,21 +554,21 @@ pub struct StreamingVectoredReadPlanner {
     /// Size of the current batch
     cnt: usize,
 
-    mode: VectoredReadCoalesceMode,
+    align: usize,
 }
 
 impl StreamingVectoredReadPlanner {
     pub fn new(max_read_size: u64, max_cnt: usize) -> Self {
         assert!(max_cnt > 0);
         assert!(max_read_size > 0);
-        let mode = VectoredReadCoalesceMode::get();
+        let align = virtual_file::get_io_buffer_alignment();
         Self {
             read_builder: None,
             prev: None,
             max_cnt,
             max_read_size,
             cnt: 0,
-            mode,
+            align,
         }
     }
 
@@ -673,11 +617,11 @@ impl StreamingVectoredReadPlanner {
             }
             None => {
                 self.read_builder = {
-                    Some(VectoredReadBuilder::new_streaming(
+                    Some(ChunkedVectoredReadBuilder::new_streaming(
                         start_offset,
                         end_offset,
                         BlobMeta { key, lsn },
-                        self.mode,
+                        self.align,
                     ))
                 };
             }
@@ -1003,7 +947,7 @@ mod tests {
         let reserved_bytes = blobs.iter().map(|bl| bl.len()).max().unwrap() * 2 + 16;
         let mut buf = BytesMut::with_capacity(reserved_bytes);
 
-        let mode = VectoredReadCoalesceMode::get();
+        let align = virtual_file::get_io_buffer_alignment();
         let vectored_blob_reader = VectoredBlobReader::new(&file);
         let meta = BlobMeta {
             key: Key::MIN,
@@ -1015,13 +959,19 @@ mod tests {
             if idx + 1 == offsets.len() {
                 continue;
             }
-            let read_builder = VectoredReadBuilder::new(*offset, *end, meta, 16 * 4096, mode);
+            let read_builder =
+                ChunkedVectoredReadBuilder::new(*offset, *end, meta, 16 * 4096, align);
             let read = read_builder.build();
             let result = vectored_blob_reader.read_blobs(&read, buf, &ctx).await?;
             assert_eq!(result.blobs.len(), 1);
             let read_blob = &result.blobs[0];
-            let read_buf = &result.buf[read_blob.start..read_blob.end];
-            assert_eq!(blob, read_buf, "mismatch for idx={idx} at offset={offset}");
+            let view = BufView::new_slice(&result.buf);
+            let read_buf = read_blob.read(&view).await?;
+            assert_eq!(
+                &blob[..],
+                &read_buf[..],
+                "mismatch for idx={idx} at offset={offset}"
+            );
             buf = result.buf;
         }
         Ok(())

pageserver/src/virtual_file.rs

@@ -466,6 +466,7 @@ impl VirtualFile {
                 &[]
             };
             utils::crashsafe::overwrite(&final_path, &tmp_path, content)
+                .maybe_fatal_err("crashsafe_overwrite")
         })
         .await
         .expect("blocking task is never aborted")
@@ -475,7 +476,7 @@ impl VirtualFile {
     pub async fn sync_all(&self) -> Result<(), Error> {
         with_file!(self, StorageIoOperation::Fsync, |file_guard| {
             let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
-            res
+            res.maybe_fatal_err("sync_all")
         })
     }
 
@@ -483,7 +484,7 @@ impl VirtualFile {
     pub async fn sync_data(&self) -> Result<(), Error> {
         with_file!(self, StorageIoOperation::Fsync, |file_guard| {
             let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
-            res
+            res.maybe_fatal_err("sync_data")
         })
     }
 
@@ -1147,7 +1148,9 @@ pub fn init(num_slots: usize, engine: IoEngineKind, io_buffer_alignment: usize)
         panic!("virtual_file::init called twice");
     }
     if set_io_buffer_alignment(io_buffer_alignment).is_err() {
-        panic!("IO buffer alignment ({io_buffer_alignment}) is not a power of two");
+        panic!(
+            "IO buffer alignment needs to be a power of two and greater than 512, got {io_buffer_alignment}"
+        );
     }
     io_engine::init(engine);
     crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
@@ -1174,14 +1177,16 @@ fn get_open_files() -> &'static OpenFiles {
 
 static IO_BUFFER_ALIGNMENT: AtomicUsize = AtomicUsize::new(DEFAULT_IO_BUFFER_ALIGNMENT);
 
-/// Returns true if `x` is zero or a power of two.
-fn is_zero_or_power_of_two(x: usize) -> bool {
-    (x == 0) || ((x & (x - 1)) == 0)
+/// Returns true if the alignment is a power of two and is greater or equal to 512.
+fn is_valid_io_buffer_alignment(align: usize) -> bool {
+    align.is_power_of_two() && align >= 512
 }
 
+/// Sets IO buffer alignment requirement. Returns error if the alignment requirement is
+/// not a power of two or less than 512 bytes.
 #[allow(unused)]
 pub(crate) fn set_io_buffer_alignment(align: usize) -> Result<(), usize> {
-    if is_zero_or_power_of_two(align) {
+    if is_valid_io_buffer_alignment(align) {
         IO_BUFFER_ALIGNMENT.store(align, std::sync::atomic::Ordering::Relaxed);
         Ok(())
     } else {
@@ -1189,19 +1194,19 @@ pub(crate) fn set_io_buffer_alignment(align: usize) -> Result<(), usize> {
     }
 }
 
-/// Gets the io buffer alignment requirement. Returns 0 if there is no requirement specified.
+/// Gets the io buffer alignment.
 ///
-/// This function should be used to check the raw config value.
-pub(crate) fn get_io_buffer_alignment_raw() -> usize {
+/// This function should be used for getting the actual alignment value to use.
+pub(crate) fn get_io_buffer_alignment() -> usize {
     let align = IO_BUFFER_ALIGNMENT.load(std::sync::atomic::Ordering::Relaxed);
 
     if cfg!(test) {
         let env_var_name = "NEON_PAGESERVER_UNIT_TEST_IO_BUFFER_ALIGNMENT";
         if let Some(test_align) = utils::env::var(env_var_name) {
-            if is_zero_or_power_of_two(test_align) {
+            if is_valid_io_buffer_alignment(test_align) {
                 test_align
             } else {
-                panic!("IO buffer alignment ({test_align}) is not a power of two");
+                panic!("IO buffer alignment needs to be a power of two and greater than 512, got {test_align}");
             }
         } else {
             align
@@ -1211,14 +1216,6 @@ pub(crate) fn get_io_buffer_alignment_raw() -> usize {
     }
 }
 
-/// Gets the io buffer alignment requirement. Returns 1 if the alignment config is set to zero.
-///
-/// This function should be used for getting the actual alignment value to use.
-pub(crate) fn get_io_buffer_alignment() -> usize {
-    let align = get_io_buffer_alignment_raw();
-    align.max(1)
-}
-
 #[cfg(test)]
 mod tests {
     use crate::context::DownloadBehavior;

pageserver/src/walredo.rs

@@ -205,6 +205,22 @@ impl PostgresRedoManager {
         }
     }
 
+    /// Do a ping request-response roundtrip.
+    ///
+    /// Not used in production, but by Rust benchmarks.
+    ///
+    /// # Cancel-Safety
+    ///
+    /// This method is cancellation-safe.
+    pub async fn ping(&self, pg_version: u32) -> Result<(), Error> {
+        self.do_with_walredo_process(pg_version, |proc| async move {
+            proc.ping(Duration::from_secs(1))
+                .await
+                .map_err(Error::Other)
+        })
+        .await
+    }
+
     pub fn status(&self) -> WalRedoManagerStatus {
         WalRedoManagerStatus {
             last_redo_at: {
@@ -297,6 +313,9 @@ impl PostgresRedoManager {
         }
     }
 
+    /// # Cancel-Safety
+    ///
+    /// This method is cancel-safe iff `closure` is cancel-safe.
     async fn do_with_walredo_process<
         F: FnOnce(Arc<Process>) -> Fut,
         Fut: Future<Output = Result<O, Error>>,
@@ -537,6 +556,17 @@ mod tests {
     use tracing::Instrument;
     use utils::{id::TenantId, lsn::Lsn};
 
+    #[tokio::test]
+    async fn test_ping() {
+        let h = RedoHarness::new().unwrap();
+
+        h.manager
+            .ping(14)
+            .instrument(h.span())
+            .await
+            .expect("ping should work");
+    }
+
     #[tokio::test]
     async fn short_v14_redo() {
         let expected = std::fs::read("test_data/short_v14_redo.page").unwrap();

pageserver/src/walredo/process.rs

@@ -6,6 +6,7 @@ use self::no_leak_child::NoLeakChild;
 use crate::{
     config::PageServerConf,
     metrics::{WalRedoKillCause, WAL_REDO_PROCESS_COUNTERS, WAL_REDO_RECORD_COUNTER},
+    page_cache::PAGE_SZ,
     span::debug_assert_current_span_has_tenant_id,
     walrecord::NeonWalRecord,
 };
@@ -237,6 +238,26 @@ impl WalRedoProcess {
         res
     }
 
+    /// Do a ping request-response roundtrip.
+    ///
+    /// Not used in production, but by Rust benchmarks.
+    pub(crate) async fn ping(&self, timeout: Duration) -> anyhow::Result<()> {
+        let mut writebuf: Vec<u8> = Vec::with_capacity(4);
+        protocol::build_ping_msg(&mut writebuf);
+        let Ok(res) = tokio::time::timeout(timeout, self.apply_wal_records0(&writebuf)).await
+        else {
+            anyhow::bail!("WAL redo ping timed out");
+        };
+        let response = res?;
+        if response.len() != PAGE_SZ {
+            anyhow::bail!(
+                "WAL redo ping response should respond with page-sized response: {}",
+                response.len()
+            );
+        }
+        Ok(())
+    }
+
     /// # Cancel-Safety
     ///
     /// When not polled to completion (e.g. because in `tokio::select!` another

pageserver/src/walredo/process/protocol.rs

@@ -55,3 +55,8 @@ pub(crate) fn build_get_page_msg(tag: BufferTag, buf: &mut Vec<u8>) {
     tag.ser_into(buf)
         .expect("serialize BufferTag should always succeed");
 }
+
+pub(crate) fn build_ping_msg(buf: &mut Vec<u8>) {
+    buf.put_u8(b'H');
+    buf.put_u32(4);
+}

pgxn/neon/Makefile

@@ -9,6 +9,8 @@ OBJS = \
 	hll.o \
 	libpagestore.o \
 	neon.o \
+	neon_pgversioncompat.o \
+	neon_perf_counters.o \
 	neon_utils.o \
 	neon_walreader.o \
 	pagestore_smgr.o \
@@ -23,7 +25,18 @@ SHLIB_LINK_INTERNAL = $(libpq)
 SHLIB_LINK = -lcurl
 
 EXTENSION = neon
-DATA = neon--1.0.sql neon--1.0--1.1.sql neon--1.1--1.2.sql neon--1.2--1.3.sql neon--1.3--1.2.sql neon--1.2--1.1.sql neon--1.1--1.0.sql  neon--1.3--1.4.sql neon--1.4--1.3.sql
+DATA = \
+	neon--1.0.sql \
+	neon--1.0--1.1.sql \
+	neon--1.1--1.2.sql \
+	neon--1.2--1.3.sql \
+	neon--1.3--1.4.sql \
+	neon--1.4--1.5.sql \
+	neon--1.5--1.4.sql \
+	neon--1.4--1.3.sql \
+	neon--1.3--1.2.sql \
+	neon--1.2--1.1.sql \
+	neon--1.1--1.0.sql
 PGFILEDESC = "neon - cloud storage for PostgreSQL"
 
 EXTRA_CLEAN = \

pgxn/neon/file_cache.c

@@ -109,6 +109,7 @@ typedef struct FileCacheControl
 								 * reenabling */
 	uint32		size;			/* size of cache file in chunks */
 	uint32		used;			/* number of used chunks */
+	uint32		used_pages;		/* number of used pages */
 	uint32		limit;			/* shared copy of lfc_size_limit */
 	uint64		hits;
 	uint64		misses;
@@ -905,6 +906,10 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 				/* Cache overflow: evict least recently used chunk */
 				FileCacheEntry *victim = dlist_container(FileCacheEntry, list_node, dlist_pop_head_node(&lfc_ctl->lru));
 	
+				for (int i = 0; i < BLOCKS_PER_CHUNK; i++)
+				{
+					lfc_ctl->used_pages -= (victim->bitmap[i >> 5] >> (i & 31)) & 1;
+				}
 				CriticalAssert(victim->access_count == 0);
 				entry->offset = victim->offset; /* grab victim's chunk */
 				hash_search_with_hash_value(lfc_hash, &victim->key, victim->hash, HASH_REMOVE, NULL);
@@ -959,6 +964,7 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 				for (int i = 0; i < blocks_in_chunk; i++)
 				{
+					lfc_ctl->used_pages += 1 - ((entry->bitmap[(chunk_offs + i) >> 5] >> ((chunk_offs + i) & 31)) & 1);
 					entry->bitmap[(chunk_offs + i) >> 5] |=
 						(1 << ((chunk_offs + i) & 31));
 				}
@@ -1051,6 +1057,11 @@ neon_get_lfc_stats(PG_FUNCTION_ARGS)
 			if (lfc_ctl)
 				value = lfc_ctl->size;
 			break;
+		case 5:
+			key = "file_cache_used_pages";
+			if (lfc_ctl)
+				value = lfc_ctl->used_pages;
+			break;
 		default:
 			SRF_RETURN_DONE(funcctx);
 	}

pgxn/neon/libpagestore.c

@@ -30,6 +30,7 @@
 #include "utils/guc.h"
 
 #include "neon.h"
+#include "neon_perf_counters.h"
 #include "neon_utils.h"
 #include "pagestore_client.h"
 #include "walproposer.h"
@@ -331,6 +332,7 @@ CLEANUP_AND_DISCONNECT(PageServer *shard)
 	}
 	if (shard->conn)
 	{
+		MyNeonCounters->pageserver_disconnects_total++;
 		PQfinish(shard->conn);
 		shard->conn = NULL;
 	}
@@ -737,6 +739,8 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 	PageServer *shard = &page_servers[shard_no];
 	PGconn	   *pageserver_conn;
 
+	MyNeonCounters->pageserver_requests_sent_total++;
+
 	/* If the connection was lost for some reason, reconnect */
 	if (shard->state == PS_Connected && PQstatus(shard->conn) == CONNECTION_BAD)
 	{
@@ -889,6 +893,7 @@ pageserver_flush(shardno_t shard_no)
 	}
 	else
 	{
+		MyNeonCounters->pageserver_send_flushes_total++;
 		if (PQflush(pageserver_conn))
 		{
 			char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
@@ -922,7 +927,7 @@ check_neon_id(char **newval, void **extra, GucSource source)
 static Size
 PagestoreShmemSize(void)
 {
-	return sizeof(PagestoreShmemState);
+	return add_size(sizeof(PagestoreShmemState), NeonPerfCountersShmemSize());
 }
 
 static bool
@@ -941,6 +946,9 @@ PagestoreShmemInit(void)
 		memset(&pagestore_shared->shard_map, 0, sizeof(ShardMap));
 		AssignPageserverConnstring(page_server_connstring, NULL);
 	}
+
+	NeonPerfCountersShmemInit();
+
 	LWLockRelease(AddinShmemInitLock);
 	return found;
 }

pgxn/neon/neon--1.4--1.5.sql

@@ -0,0 +1,39 @@
+\echo Use "ALTER EXTENSION neon UPDATE TO '1.5'" to load this file. \quit
+
+
+CREATE FUNCTION get_backend_perf_counters()
+RETURNS SETOF RECORD
+AS 'MODULE_PATHNAME', 'neon_get_backend_perf_counters'
+LANGUAGE C PARALLEL SAFE;
+
+CREATE FUNCTION get_perf_counters()
+RETURNS SETOF RECORD
+AS 'MODULE_PATHNAME', 'neon_get_perf_counters'
+LANGUAGE C PARALLEL SAFE;
+
+-- Show various metrics, for each backend. Note that the values are not reset
+-- when a backend exits. When a new backend starts with the backend ID, it will
+-- continue accumulating the values from where the old backend left. If you are
+-- only interested in the changes from your own session, store the values at the
+-- beginning of the session somewhere, and subtract them on subsequent calls.
+--
+-- For histograms, 'bucket_le' is the upper bound of the histogram bucket.
+CREATE VIEW neon_backend_perf_counters AS
+  SELECT P.procno, P.pid, P.metric, P.bucket_le, P.value
+  FROM get_backend_perf_counters() AS P (
+    procno integer,
+    pid integer,
+    metric text,
+    bucket_le float8,
+    value float8
+  );
+
+-- Summary across all backends. (This could also be implemented with
+-- an aggregate query over neon_backend_perf_counters view.)
+CREATE VIEW neon_perf_counters AS
+  SELECT P.metric, P.bucket_le, P.value
+  FROM get_perf_counters() AS P (
+    metric text,
+    bucket_le float8,
+    value float8
+  );

pgxn/neon/neon--1.5--1.4.sql

@@ -0,0 +1,4 @@
+DROP VIEW IF EXISTS neon_perf_counters;
+DROP VIEW IF EXISTS neon_backend_perf_counters;
+DROP FUNCTION IF EXISTS get_perf_counters();
+DROP FUNCTION IF EXISTS get_backend_perf_counters();

pgxn/neon/neon.control

@@ -1,6 +1,6 @@
 # neon extension
 comment = 'cloud storage for PostgreSQL'
-default_version = '1.4'
+default_version = '1.5'
 module_pathname = '$libdir/neon'
 relocatable = true
 trusted = true

pgxn/neon/neon_perf_counters.c

@@ -0,0 +1,261 @@
+/*-------------------------------------------------------------------------
+ *
+ * neon_perf_counters.c
+ *	  Collect statistics about Neon I/O
+ *
+ * Each backend has its own set of counters in shared memory.
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include <math.h>
+
+#include "funcapi.h"
+#include "miscadmin.h"
+#include "storage/proc.h"
+#include "storage/shmem.h"
+#include "utils/builtins.h"
+
+#include "neon_perf_counters.h"
+#include "neon_pgversioncompat.h"
+
+neon_per_backend_counters *neon_per_backend_counters_shared;
+
+Size
+NeonPerfCountersShmemSize(void)
+{
+	Size		size = 0;
+
+	size = add_size(size, mul_size(MaxBackends, sizeof(neon_per_backend_counters)));
+
+	return size;
+}
+
+void
+NeonPerfCountersShmemInit(void)
+{
+	bool		found;
+
+	neon_per_backend_counters_shared =
+		ShmemInitStruct("Neon perf counters",
+						mul_size(MaxBackends,
+								 sizeof(neon_per_backend_counters)),
+						&found);
+	Assert(found == IsUnderPostmaster);
+	if (!found)
+	{
+		/* shared memory is initialized to zeros, so nothing to do here */
+	}
+}
+
+/*
+ * Count a GetPage wait operation.
+ */
+void
+inc_getpage_wait(uint64 latency_us)
+{
+	int			lo = 0;
+	int			hi = NUM_GETPAGE_WAIT_BUCKETS - 1;
+
+	/* Find the right bucket with binary search */
+	while (lo < hi)
+	{
+		int			mid = (lo + hi) / 2;
+
+		if (latency_us < getpage_wait_bucket_thresholds[mid])
+			hi = mid;
+		else
+			lo = mid + 1;
+	}
+	MyNeonCounters->getpage_wait_us_bucket[lo]++;
+	MyNeonCounters->getpage_wait_us_sum += latency_us;
+	MyNeonCounters->getpage_wait_us_count++;
+}
+
+/*
+ * Support functions for the views, neon_backend_perf_counters and
+ * neon_perf_counters.
+ */
+
+typedef struct
+{
+	char	   *name;
+	bool		is_bucket;
+	double		bucket_le;
+	double		value;
+} metric_t;
+
+static metric_t *
+neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
+{
+#define NUM_METRICS (2 + NUM_GETPAGE_WAIT_BUCKETS + 8)
+	metric_t   *metrics = palloc((NUM_METRICS + 1) * sizeof(metric_t));
+	uint64		bucket_accum;
+	int			i = 0;
+	Datum		getpage_wait_str;
+
+	metrics[i].name = "getpage_wait_seconds_count";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_wait_us_count;
+	i++;
+	metrics[i].name = "getpage_wait_seconds_sum";
+	metrics[i].is_bucket = false;
+	metrics[i].value = ((double) counters->getpage_wait_us_sum) / 1000000.0;
+	i++;
+
+	bucket_accum = 0;
+	for (int bucketno = 0; bucketno < NUM_GETPAGE_WAIT_BUCKETS; bucketno++)
+	{
+		uint64		threshold = getpage_wait_bucket_thresholds[bucketno];
+
+		bucket_accum += counters->getpage_wait_us_bucket[bucketno];
+
+		metrics[i].name = "getpage_wait_seconds_bucket";
+		metrics[i].is_bucket = true;
+		metrics[i].bucket_le = (threshold == UINT64_MAX) ? INFINITY : ((double) threshold) / 1000000.0;
+		metrics[i].value = (double) bucket_accum;
+		i++;
+	}
+	metrics[i].name = "getpage_prefetch_requests_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_requests_total;
+	i++;
+	metrics[i].name = "getpage_sync_requests_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_sync_requests_total;
+	i++;
+	metrics[i].name = "getpage_prefetch_misses_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_misses_total;
+	i++;
+	metrics[i].name = "getpage_prefetch_discards_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_discards_total;
+	i++;
+	metrics[i].name = "pageserver_requests_sent_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_requests_sent_total;
+	i++;
+	metrics[i].name = "pageserver_requests_disconnects_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_disconnects_total;
+	i++;
+	metrics[i].name = "pageserver_send_flushes_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_send_flushes_total;
+	i++;
+	metrics[i].name = "file_cache_hits_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->file_cache_hits_total;
+	i++;
+
+	Assert(i == NUM_METRICS);
+
+	/* NULL entry marks end of array */
+	metrics[i].name = NULL;
+	metrics[i].value = 0;
+
+	return metrics;
+}
+
+/*
+ * Write metric to three output Datums
+ */
+static void
+metric_to_datums(metric_t *m, Datum *values, bool *nulls)
+{
+	values[0] = CStringGetTextDatum(m->name);
+	nulls[0] = false;
+	if (m->is_bucket)
+	{
+		values[1] = Float8GetDatum(m->bucket_le);
+		nulls[1] = false;
+	}
+	else
+	{
+		values[1] = (Datum) 0;
+		nulls[1] = true;
+	}
+	values[2] = Float8GetDatum(m->value);
+	nulls[2] = false;
+}
+
+PG_FUNCTION_INFO_V1(neon_get_backend_perf_counters);
+Datum
+neon_get_backend_perf_counters(PG_FUNCTION_ARGS)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Datum		values[5];
+	bool		nulls[5];
+
+	/* We put all the tuples into a tuplestore in one go. */
+	InitMaterializedSRF(fcinfo, 0);
+
+	for (int procno = 0; procno < MaxBackends; procno++)
+	{
+		PGPROC	   *proc = GetPGProcByNumber(procno);
+		int			pid = proc->pid;
+		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
+		metric_t   *metrics = neon_perf_counters_to_metrics(counters);
+
+		values[0] = Int32GetDatum(procno);
+		nulls[0] = false;
+		values[1] = Int32GetDatum(pid);
+		nulls[1] = false;
+
+		for (int i = 0; metrics[i].name != NULL; i++)
+		{
+			metric_to_datums(&metrics[i], &values[2], &nulls[2]);
+			tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+		}
+
+		pfree(metrics);
+	}
+
+	return (Datum) 0;
+}
+
+PG_FUNCTION_INFO_V1(neon_get_perf_counters);
+Datum
+neon_get_perf_counters(PG_FUNCTION_ARGS)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Datum		values[3];
+	bool		nulls[3];
+	Datum		getpage_wait_str;
+	neon_per_backend_counters totals = {0};
+	metric_t   *metrics;
+
+	/* We put all the tuples into a tuplestore in one go. */
+	InitMaterializedSRF(fcinfo, 0);
+
+	/* Aggregate the counters across all backends */
+	for (int procno = 0; procno < MaxBackends; procno++)
+	{
+		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
+
+		totals.getpage_wait_us_count += counters->getpage_wait_us_count;
+		totals.getpage_wait_us_sum += counters->getpage_wait_us_sum;
+		for (int bucketno = 0; bucketno < NUM_GETPAGE_WAIT_BUCKETS; bucketno++)
+			totals.getpage_wait_us_bucket[bucketno] += counters->getpage_wait_us_bucket[bucketno];
+		totals.getpage_prefetch_requests_total += counters->getpage_prefetch_requests_total;
+		totals.getpage_sync_requests_total += counters->getpage_sync_requests_total;
+		totals.getpage_prefetch_misses_total += counters->getpage_prefetch_misses_total;
+		totals.getpage_prefetch_discards_total += counters->getpage_prefetch_discards_total;
+		totals.pageserver_requests_sent_total += counters->pageserver_requests_sent_total;
+		totals.pageserver_disconnects_total += counters->pageserver_disconnects_total;
+		totals.pageserver_send_flushes_total += counters->pageserver_send_flushes_total;
+		totals.file_cache_hits_total += counters->file_cache_hits_total;
+	}
+
+	metrics = neon_perf_counters_to_metrics(&totals);
+	for (int i = 0; metrics[i].name != NULL; i++)
+	{
+		metric_to_datums(&metrics[i], &values[0], &nulls[0]);
+		tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+	}
+	pfree(metrics);
+
+	return (Datum) 0;
+}

pgxn/neon/neon_perf_counters.h

@@ -0,0 +1,111 @@
+/*-------------------------------------------------------------------------
+ *
+ * neon_perf_counters.h
+ *	  Performance counters for neon storage requests
+ *-------------------------------------------------------------------------
+ */
+
+#ifndef NEON_PERF_COUNTERS_H
+#define NEON_PERF_COUNTERS_H
+
+#if PG_VERSION_NUM >= 170000
+#include "storage/procnumber.h"
+#else
+#include "storage/backendid.h"
+#include "storage/proc.h"
+#endif
+
+static const uint64 getpage_wait_bucket_thresholds[] = {
+	      20,       30,       60,       100,  /* 0      -  100 us */
+	     200,      300,      600,	   1000,  /* 100 us - 1 ms */
+	    2000,     3000,     6000,     10000,  /* 1 ms   - 10 ms */
+	   20000,    30000,    60000,    100000,  /* 10 ms  - 100 ms */
+	  200000,   300000,   600000,   1000000,  /* 100 ms - 1 s */
+	 2000000,  3000000,  6000000,  10000000,  /* 1 s - 10 s */
+    20000000, 30000000, 60000000, 100000000,  /* 10 s - 100 s */
+	UINT64_MAX,
+};
+#define NUM_GETPAGE_WAIT_BUCKETS (lengthof(getpage_wait_bucket_thresholds))
+
+typedef struct
+{
+	/*
+	 * Histogram for how long an smgrread() request needs to wait for response
+	 * from pageserver. When prefetching is effective, these wait times can be
+	 * lower than the network latency to the pageserver, even zero, if the
+	 * page is already readily prefetched whenever we need to read a page.
+	 *
+	 * Note: we accumulate these in microseconds, because that's convenient in
+	 * the backend, but the 'neon_backend_perf_counters' view will convert
+	 * them to seconds, to make them more idiomatic as prometheus metrics.
+	 */
+	uint64		getpage_wait_us_count;
+	uint64		getpage_wait_us_sum;
+	uint64		getpage_wait_us_bucket[NUM_GETPAGE_WAIT_BUCKETS];
+
+	/*
+	 * Total number of speculative prefetch Getpage requests and synchronous
+	 * GetPage requests sent.
+	 */
+	uint64		getpage_prefetch_requests_total;
+	uint64		getpage_sync_requests_total;
+
+	/* XXX: It's not clear to me when these misses happen. */
+	uint64		getpage_prefetch_misses_total;
+
+	/*
+	 * Number of prefetched responses that were discarded becuase the
+	 * prefetched page was not needed or because it was concurrently fetched /
+	 * modified by another backend.
+	 */
+	uint64		getpage_prefetch_discards_total;
+
+	/*
+	 * Total number of requests send to pageserver. (prefetch_requests_total
+	 * and sync_request_total count only GetPage requests, this counts all
+	 * request types.)
+	 */
+	uint64		pageserver_requests_sent_total;
+
+	/*
+	 * Number of times the connection to the pageserver was lost and the
+	 * backend had to reconnect. Note that this doesn't count the first
+	 * connection in each backend, only reconnects.
+	 */
+	uint64		pageserver_disconnects_total;
+
+	/*
+	 * Number of network flushes to the pageserver. Synchronous requests are
+	 * flushed immediately, but when prefetching requests are sent in batches,
+	 * this can be smaller than pageserver_requests_sent_total.
+	 */
+	uint64		pageserver_send_flushes_total;
+
+	/*
+	 * Number of requests satisfied from the LFC.
+	 *
+	 * This is redundant with the server-wide file_cache_hits, but this gives
+	 * per-backend granularity, and it's handy to have this in the same place
+	 * as counters for requests that went to the pageserver. Maybe move all
+	 * the LFC stats to this struct in the future?
+	 */
+	uint64		file_cache_hits_total;
+
+} neon_per_backend_counters;
+
+/* Pointer to the shared memory array of neon_per_backend_counters structs */
+extern neon_per_backend_counters *neon_per_backend_counters_shared;
+
+#if PG_VERSION_NUM >= 170000
+#define MyNeonCounters (&neon_per_backend_counters_shared[MyProcNumber])
+#else
+#define MyNeonCounters (&neon_per_backend_counters_shared[MyProc->pgprocno])
+#endif
+
+extern void inc_getpage_wait(uint64 latency);
+
+extern Size NeonPerfCountersShmemSize(void);
+extern void NeonPerfCountersShmemInit(void);
+
+
+#endif							/* NEON_PERF_COUNTERS_H */

pgxn/neon/neon_pgversioncompat.c

@@ -0,0 +1,44 @@
+/*
+ * Support functions for the compatibility macros in neon_pgversioncompat.h
+ */
+#include "postgres.h"
+
+#include "funcapi.h"
+#include "miscadmin.h"
+#include "utils/tuplestore.h"
+
+#include "neon_pgversioncompat.h"
+
+#if PG_MAJORVERSION_NUM < 15
+void
+InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Tuplestorestate *tupstore;
+	MemoryContext old_context,
+				per_query_ctx;
+	TupleDesc	stored_tupdesc;
+
+	/* check to see if caller supports returning a tuplestore */
+	if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("set-valued function called in context that cannot accept a set")));
+
+	/*
+	 * Store the tuplestore and the tuple descriptor in ReturnSetInfo.  This
+	 * must be done in the per-query memory context.
+	 */
+	per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
+	old_context = MemoryContextSwitchTo(per_query_ctx);
+
+	if (get_call_result_type(fcinfo, NULL, &stored_tupdesc) != TYPEFUNC_COMPOSITE)
+		elog(ERROR, "return type must be a row type");
+
+	tupstore = tuplestore_begin_heap(false, false, work_mem);
+	rsinfo->returnMode = SFRM_Materialize;
+	rsinfo->setResult = tupstore;
+	rsinfo->setDesc = stored_tupdesc;
+	MemoryContextSwitchTo(old_context);
+}
+#endif

pgxn/neon/neon_pgversioncompat.h

@@ -6,6 +6,8 @@
 #ifndef NEON_PGVERSIONCOMPAT_H
 #define NEON_PGVERSIONCOMPAT_H
 
+#include "fmgr.h"
+
 #if PG_MAJORVERSION_NUM < 17
 #define NRelFileInfoBackendIsTemp(rinfo) (rinfo.backend != InvalidBackendId)
 #else
@@ -123,4 +125,8 @@
 #define AmAutoVacuumWorkerProcess() (IsAutoVacuumWorkerProcess())
 #endif
 
+#if PG_MAJORVERSION_NUM < 15
+extern void InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags);
+#endif
+
 #endif							/* NEON_PGVERSIONCOMPAT_H */