Merge pull request #9756 from neondatabase/rc/proxy/2024-11-14

Proxy release 2024-11-14

.github/actionlint.yml

@@ -20,3 +20,4 @@ config-variables:
   - REMOTE_STORAGE_AZURE_REGION
   - SLACK_UPCOMING_RELEASE_CHANNEL_ID
   - DEV_AWS_OIDC_ROLE_ARN
+  - BENCHMARK_INGEST_TARGET_PROJECTID

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

@@ -221,6 +221,8 @@ runs:
         REPORT_URL: ${{ steps.generate-report.outputs.report-url }}
         COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
       with:
+        # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+        retries: 5
         script: |
           const { REPORT_URL, COMMIT_SHA } = process.env
 

.github/actions/set-docker-config-dir/action.yml

@@ -1,36 +0,0 @@
-name: "Set custom docker config directory"
-description: "Create a directory for docker config and set DOCKER_CONFIG"
-
-# Use custom DOCKER_CONFIG directory to avoid conflicts with default settings
-runs:
-  using: "composite"
-  steps:
-  - name: Show warning on GitHub-hosted runners
-    if: runner.environment == 'github-hosted'
-    shell: bash -euo pipefail {0}
-    run: |
-      # Using the following environment variables to find a path to the workflow file
-      # ${GITHUB_WORKFLOW_REF} - octocat/hello-world/.github/workflows/my-workflow.yml@refs/heads/my_branch
-      # ${GITHUB_REPOSITORY}   - octocat/hello-world
-      # ${GITHUB_REF}          - refs/heads/my_branch
-      # From https://docs.github.com/en/actions/writing-workflows/choosing-what-your-workflow-does/variables
-
-      filename_with_ref=${GITHUB_WORKFLOW_REF#"$GITHUB_REPOSITORY/"}
-      filename=${filename_with_ref%"@$GITHUB_REF"}
-
-      # https://docs.github.com/en/actions/writing-workflows/choosing-what-your-workflow-does/workflow-commands-for-github-actions#setting-a-warning-message
-      title='Unnecessary usage of `.github/actions/set-docker-config-dir`'
-      message='No need to use `.github/actions/set-docker-config-dir` action on GitHub-hosted runners'
-      echo "::warning file=${filename},title=${title}::${message}"
-
-  - uses: pyTooling/Actions/with-post-step@74afc5a42a17a046c90c68cb5cfa627e5c6c5b6b # v1.0.7
-    env:
-      DOCKER_CONFIG: .docker-custom-${{ github.run_id }}-${{ github.run_attempt }}
-    with:
-      main: |
-        mkdir -p "${DOCKER_CONFIG}"
-        echo DOCKER_CONFIG=${DOCKER_CONFIG} | tee -a $GITHUB_ENV
-      post: |
-        if [ -d "${DOCKER_CONFIG}" ]; then
-          rm -r "${DOCKER_CONFIG}"
-        fi

.github/pull_request_template.md

@@ -1,14 +1,3 @@
 ## Problem
 
 ## Summary of changes
-
-## Checklist before requesting a review
-
-- [ ] I have performed a self-review of my code.
-- [ ] If it is a core feature, I have added thorough tests.
-- [ ] Do we need to implement analytics? if so did you add the relevant metrics to the dashboard?
-- [ ] If this PR requires public announcement, mark it with /release-notes label and add several sentences in this section.
-
-## Checklist before merging
-
-- [ ] Do not forget to reformat commit message to not include the above checklist

.github/workflows/_check-codestyle-python.yml

@@ -0,0 +1,37 @@
+name: Check Codestyle Python
+
+on:
+  workflow_call:
+    inputs:
+      build-tools-image:
+        description: 'build-tools image'
+        required: true
+        type: string
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+
+jobs:
+  check-codestyle-python:
+    runs-on: [ self-hosted, small ]
+    container:
+      image: ${{ inputs.build-tools-image }}
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+      options: --init
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/cache@v4
+        with:
+          path: ~/.cache/pypoetry/virtualenvs
+          key: v2-${{ runner.os }}-${{ runner.arch }}-python-deps-bookworm-${{ hashFiles('poetry.lock') }}
+
+      - run: ./scripts/pysync
+
+      - run: poetry run ruff check .
+      - run: poetry run ruff format --check .
+      - run: poetry run mypy .

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

@@ -64,7 +64,7 @@ jobs:
 
       - uses: actions/checkout@v4
 
-      - uses: ./.github/actions/set-docker-config-dir
+      - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
       - uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false

.github/workflows/build_and_test.yml

@@ -90,35 +90,10 @@ jobs:
 
   check-codestyle-python:
     needs: [ check-permissions, build-build-tools-image ]
-    runs-on: [ self-hosted, small ]
-    container:
-      image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
-      credentials:
-        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
-        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
-      options: --init
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-
-      - name: Cache poetry deps
-        uses: actions/cache@v4
-        with:
-          path: ~/.cache/pypoetry/virtualenvs
-          key: v2-${{ runner.os }}-${{ runner.arch }}-python-deps-bookworm-${{ hashFiles('poetry.lock') }}
-
-      - name: Install Python deps
-        run: ./scripts/pysync
-
-      - name: Run `ruff check` to ensure code format
-        run: poetry run ruff check .
-
-      - name: Run `ruff format` to ensure code format
-        run: poetry run ruff format --check .
-
-      - name: Run mypy to check types
-        run: poetry run mypy .
+    uses: ./.github/workflows/_check-codestyle-python.yml
+    with:
+      build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
+    secrets: inherit
 
   check-codestyle-jsonnet:
     needs: [ check-permissions, build-build-tools-image ]
@@ -141,6 +116,7 @@ jobs:
   # Check that the vendor/postgres-* submodules point to the
   # corresponding REL_*_STABLE_neon branches.
   check-submodules:
+    needs: [ check-permissions ]
     runs-on: ubuntu-22.04
     steps:
       - name: Checkout
@@ -521,6 +497,8 @@ jobs:
           REPORT_URL_NEW: ${{ steps.upload-coverage-report-new.outputs.report-url }}
           COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
           script: |
             const { REPORT_URL_NEW, COMMIT_SHA } = process.env
 
@@ -552,7 +530,7 @@ jobs:
         with:
           submodules: true
 
-      - uses: ./.github/actions/set-docker-config-dir
+      - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
       - uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
@@ -643,7 +621,7 @@ jobs:
         with:
           submodules: true
 
-      - uses: ./.github/actions/set-docker-config-dir
+      - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
       - uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
@@ -824,7 +802,7 @@ jobs:
           curl -fL https://github.com/neondatabase/autoscaling/releases/download/$VM_BUILDER_VERSION/vm-builder -o vm-builder
           chmod +x vm-builder
 
-      - uses: ./.github/actions/set-docker-config-dir
+      - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
       - uses: docker/login-action@v3
         with:
           username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
@@ -860,7 +838,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - uses: ./.github/actions/set-docker-config-dir
+      - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
       - uses: docker/login-action@v3
         with:
           username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}

.github/workflows/ingest_benchmark.yml

@@ -0,0 +1,372 @@
+name: Benchmarking
+
+on:
+  # uncomment to run on push for debugging your PR
+  # push:
+  #   branches: [ your branch ]
+  schedule:
+    # * is a special character in YAML so you have to quote this string
+    #          ┌───────────── minute (0 - 59)
+    #          │ ┌───────────── hour (0 - 23)
+    #          │ │ ┌───────────── day of the month (1 - 31)
+    #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    - cron:   '0 9 * * *' # 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 globally because we need dedicated resources which only exist once
+  group: ingest-bench-workflow
+  cancel-in-progress: true
+
+jobs:
+  ingest:
+    strategy:
+      matrix:
+        target_project: [new_empty_project, large_existing_project]  
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PG_CONFIG: /tmp/neon/pg_install/v16/bin/pg_config
+      PSQL: /tmp/neon/pg_install/v16/bin/psql
+      PG_16_LIB_PATH: /tmp/neon/pg_install/v16/lib
+      PGCOPYDB: /pgcopydb/bin/pgcopydb
+      PGCOPYDB_LIB_PATH: /pgcopydb/lib
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    container:
+      image: neondatabase/build-tools:pinned-bookworm
+      credentials:
+        username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+      options: --init
+    timeout-minutes: 1440
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Configure AWS credentials # necessary to download artefacts
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours is currently max associated with IAM role 
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+
+    - name: Create Neon Project
+      if: ${{ matrix.target_project == 'new_empty_project' }}
+      id: create-neon-project-ingest-target
+      uses: ./.github/actions/neon-project-create
+      with:
+        region_id: aws-us-east-2
+        postgres_version: 16
+        compute_units: '[7, 7]' # we want to test large compute here to avoid compute-side bottleneck
+        api_key: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Initialize Neon project and retrieve current backpressure seconds
+      if: ${{ matrix.target_project == 'new_empty_project' }}
+      env:
+          NEW_PROJECT_CONNSTR: ${{ steps.create-neon-project-ingest-target.outputs.dsn }}
+          NEW_PROJECT_ID: ${{ steps.create-neon-project-ingest-target.outputs.project_id }}
+      run: |
+        echo "Initializing Neon project with project_id: ${NEW_PROJECT_ID}"
+        export LD_LIBRARY_PATH=${PG_16_LIB_PATH}
+        ${PSQL} "${NEW_PROJECT_CONNSTR}" -c "CREATE EXTENSION IF NOT EXISTS neon; CREATE EXTENSION IF NOT EXISTS neon_utils;"
+        BACKPRESSURE_TIME_BEFORE_INGEST=$(${PSQL} "${NEW_PROJECT_CONNSTR}" -t -c "select backpressure_throttling_time()/1000000;")
+        echo "BACKPRESSURE_TIME_BEFORE_INGEST=${BACKPRESSURE_TIME_BEFORE_INGEST}" >> $GITHUB_ENV
+        echo "NEW_PROJECT_CONNSTR=${NEW_PROJECT_CONNSTR}" >> $GITHUB_ENV
+
+    - name: Create Neon Branch for large tenant
+      if: ${{ matrix.target_project == 'large_existing_project' }}
+      id: create-neon-branch-ingest-target
+      uses: ./.github/actions/neon-branch-create
+      with:
+        project_id: ${{ vars.BENCHMARK_INGEST_TARGET_PROJECTID }}
+        api_key: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Initialize Neon project and retrieve current backpressure seconds
+      if: ${{ matrix.target_project == 'large_existing_project' }}
+      env:
+          NEW_PROJECT_CONNSTR: ${{ steps.create-neon-branch-ingest-target.outputs.dsn }}
+          NEW_BRANCH_ID: ${{ steps.create-neon-branch-ingest-target.outputs.branch_id }}
+      run: |
+        echo "Initializing Neon branch with branch_id: ${NEW_BRANCH_ID}"
+        export LD_LIBRARY_PATH=${PG_16_LIB_PATH}
+        # Extract the part before the database name
+        base_connstr="${NEW_PROJECT_CONNSTR%/*}"
+        # Extract the query parameters (if any) after the database name
+        query_params="${NEW_PROJECT_CONNSTR#*\?}"
+        # Reconstruct the new connection string
+        if [ "$query_params" != "$NEW_PROJECT_CONNSTR" ]; then
+          new_connstr="${base_connstr}/neondb?${query_params}"
+        else
+          new_connstr="${base_connstr}/neondb"
+        fi
+        ${PSQL} "${new_connstr}" -c "drop database ludicrous;"
+        ${PSQL} "${new_connstr}" -c "CREATE DATABASE ludicrous;"
+        if [ "$query_params" != "$NEW_PROJECT_CONNSTR" ]; then
+          NEW_PROJECT_CONNSTR="${base_connstr}/ludicrous?${query_params}"
+        else
+          NEW_PROJECT_CONNSTR="${base_connstr}/ludicrous"
+        fi
+        ${PSQL} "${NEW_PROJECT_CONNSTR}" -c "CREATE EXTENSION IF NOT EXISTS neon; CREATE EXTENSION IF NOT EXISTS neon_utils;"
+        BACKPRESSURE_TIME_BEFORE_INGEST=$(${PSQL} "${NEW_PROJECT_CONNSTR}" -t -c "select backpressure_throttling_time()/1000000;")
+        echo "BACKPRESSURE_TIME_BEFORE_INGEST=${BACKPRESSURE_TIME_BEFORE_INGEST}" >> $GITHUB_ENV
+        echo "NEW_PROJECT_CONNSTR=${NEW_PROJECT_CONNSTR}" >> $GITHUB_ENV
+      
+        
+    - name: Create pgcopydb filter file
+      run: |
+        cat << EOF > /tmp/pgcopydb_filter.txt
+          [include-only-table]
+          public.events
+          public.emails
+          public.email_transmissions
+          public.payments
+          public.editions
+          public.edition_modules
+          public.sp_content
+          public.email_broadcasts
+          public.user_collections
+          public.devices
+          public.user_accounts
+          public.lessons
+          public.lesson_users
+          public.payment_methods
+          public.orders
+          public.course_emails
+          public.modules
+          public.users
+          public.module_users
+          public.courses
+          public.payment_gateway_keys
+          public.accounts
+          public.roles
+          public.payment_gateways
+          public.management
+          public.event_names
+        EOF
+
+    - name: Invoke pgcopydb
+      env:
+          BENCHMARK_INGEST_SOURCE_CONNSTR: ${{ secrets.BENCHMARK_INGEST_SOURCE_CONNSTR }}
+      run: |
+        export LD_LIBRARY_PATH=${PGCOPYDB_LIB_PATH}:${PG_16_LIB_PATH}
+        export PGCOPYDB_SOURCE_PGURI="${BENCHMARK_INGEST_SOURCE_CONNSTR}"
+        export PGCOPYDB_TARGET_PGURI="${NEW_PROJECT_CONNSTR}"
+        export PGOPTIONS="-c maintenance_work_mem=8388608 -c max_parallel_maintenance_workers=7"
+        ${PG_CONFIG} --bindir
+        ${PGCOPYDB} --version
+        ${PGCOPYDB} clone --skip-vacuum  --no-owner --no-acl --skip-db-properties --table-jobs 4 \
+          --index-jobs 4 --restore-jobs 4 --split-tables-larger-than 10GB --skip-extensions \
+          --use-copy-binary --filters /tmp/pgcopydb_filter.txt 2>&1 | tee /tmp/pgcopydb_${{ matrix.target_project }}.log
+
+    # create dummy pgcopydb log to test parsing
+    # - name: create dummy log for parser test
+    #   run: |
+    #     cat << EOF > /tmp/pgcopydb_${{ matrix.target_project }}.log
+    #     2024-11-04 18:00:53.433 500861 INFO   main.c:136                Running pgcopydb version 0.17.10.g8361a93 from "/usr/lib/postgresql/17/bin/pgcopydb"
+    #     2024-11-04 18:00:53.434 500861 INFO   cli_common.c:1225         [SOURCE] Copying database from "postgres://neondb_owner@ep-bitter-shape-w2c1ir0a.us-east-2.aws.neon.build/neondb?sslmode=require&keepalives=1&keepalives_idle=10&keepalives_interval=10&keepalives_count=60"
+    #     2024-11-04 18:00:53.434 500861 INFO   cli_common.c:1226         [TARGET] Copying database into "postgres://neondb_owner@ep-icy-union-w25qd5pj.us-east-2.aws.neon.build/ludicrous?sslmode=require&keepalives=1&keepalives_idle=10&keepalives_interval=10&keepalives_count=60"
+    #     2024-11-04 18:00:53.442 500861 INFO   copydb.c:105              Using work dir "/tmp/pgcopydb"
+    #     2024-11-04 18:00:53.541 500861 INFO   snapshot.c:107            Exported snapshot "00000008-00000033-1" from the source database
+    #     2024-11-04 18:00:53.556 500865 INFO   cli_clone_follow.c:543    STEP 1: fetch source database tables, indexes, and sequences
+    #     2024-11-04 18:00:54.570 500865 INFO   copydb_schema.c:716       Splitting source candidate tables larger than 10 GB
+    #     2024-11-04 18:00:54.570 500865 INFO   copydb_schema.c:829       Table public.events is 96 GB large which is larger than --split-tables-larger-than 10 GB, and does not have a unique column of type integer: splitting by CTID
+    #     2024-11-04 18:01:05.538 500865 INFO   copydb_schema.c:905       Table public.events is 96 GB large, 10 COPY processes will be used, partitioning on ctid.
+    #     2024-11-04 18:01:05.564 500865 INFO   copydb_schema.c:905       Table public.email_transmissions is 27 GB large, 4 COPY processes will be used, partitioning on id.
+    #     2024-11-04 18:01:05.584 500865 INFO   copydb_schema.c:905       Table public.lessons is 25 GB large, 4 COPY processes will be used, partitioning on id.
+    #     2024-11-04 18:01:05.605 500865 INFO   copydb_schema.c:905       Table public.lesson_users is 16 GB large, 3 COPY processes will be used, partitioning on id.
+    #     2024-11-04 18:01:05.605 500865 INFO   copydb_schema.c:761       Fetched information for 26 tables (including 4 tables split in 21 partitions total), with an estimated total of 907 million tuples and 175 GB on-disk
+    #     2024-11-04 18:01:05.687 500865 INFO   copydb_schema.c:968       Fetched information for 57 indexes (supporting 25 constraints)
+    #     2024-11-04 18:01:05.753 500865 INFO   sequences.c:78            Fetching information for 24 sequences
+    #     2024-11-04 18:01:05.903 500865 INFO   copydb_schema.c:1122      Fetched information for 4 extensions
+    #     2024-11-04 18:01:06.178 500865 INFO   copydb_schema.c:1538      Found 0 indexes (supporting 0 constraints) in the target database
+    #     2024-11-04 18:01:06.184 500865 INFO   cli_clone_follow.c:584    STEP 2: dump the source database schema (pre/post data)
+    #     2024-11-04 18:01:06.186 500865 INFO   pgcmd.c:468                /usr/lib/postgresql/16/bin/pg_dump -Fc --snapshot 00000008-00000033-1 --section=pre-data --section=post-data --file /tmp/pgcopydb/schema/schema.dump 'postgres://neondb_owner@ep-bitter-shape-w2c1ir0a.us-east-2.aws.neon.build/neondb?sslmode=require&keepalives=1&keepalives_idle=10&keepalives_interval=10&keepalives_count=60'
+    #     2024-11-04 18:01:06.952 500865 INFO   cli_clone_follow.c:592    STEP 3: restore the pre-data section to the target database
+    #     2024-11-04 18:01:07.004 500865 INFO   pgcmd.c:1001               /usr/lib/postgresql/16/bin/pg_restore --dbname 'postgres://neondb_owner@ep-icy-union-w25qd5pj.us-east-2.aws.neon.build/ludicrous?sslmode=require&keepalives=1&keepalives_idle=10&keepalives_interval=10&keepalives_count=60' --section pre-data --jobs 4 --no-owner --no-acl --use-list /tmp/pgcopydb/schema/pre-filtered.list /tmp/pgcopydb/schema/schema.dump
+    #     2024-11-04 18:01:07.438 500874 INFO   table-data.c:656          STEP 4: starting 4 table-data COPY processes
+    #     2024-11-04 18:01:07.451 500877 INFO   vacuum.c:139              STEP 8: skipping VACUUM jobs per --skip-vacuum
+    #     2024-11-04 18:01:07.457 500875 INFO   indexes.c:182             STEP 6: starting 4 CREATE INDEX processes
+    #     2024-11-04 18:01:07.457 500875 INFO   indexes.c:183             STEP 7: constraints are built by the CREATE INDEX processes
+    #     2024-11-04 18:01:07.507 500865 INFO   blobs.c:74                Skipping large objects: none found.
+    #     2024-11-04 18:01:07.509 500865 INFO   sequences.c:194           STEP 9: reset sequences values
+    #     2024-11-04 18:01:07.510 500886 INFO   sequences.c:290           Set sequences values on the target database
+    #     2024-11-04 20:49:00.587 500865 INFO   cli_clone_follow.c:608    STEP 10: restore the post-data section to the target database
+    #     2024-11-04 20:49:00.600 500865 INFO   pgcmd.c:1001               /usr/lib/postgresql/16/bin/pg_restore --dbname 'postgres://neondb_owner@ep-icy-union-w25qd5pj.us-east-2.aws.neon.build/ludicrous?sslmode=require&keepalives=1&keepalives_idle=10&keepalives_interval=10&keepalives_count=60' --section post-data --jobs 4 --no-owner --no-acl --use-list /tmp/pgcopydb/schema/post-filtered.list /tmp/pgcopydb/schema/schema.dump
+    #     2024-11-05 10:50:58.508 500865 INFO   cli_clone_follow.c:639    All step are now done, 16h49m elapsed
+    #     2024-11-05 10:50:58.508 500865 INFO   summary.c:3155            Printing summary for 26 tables and 57 indexes
+
+    #       OID | Schema |                 Name | Parts | copy duration | transmitted bytes | indexes | create index duration 
+    #     ------+--------+----------------------+-------+---------------+-------------------+---------+----------------------
+    #     24654 | public |               events |    10 |         1d11h |            878 GB |       1 |                 1h41m
+    #     24623 | public |  email_transmissions |     4 |         4h46m |             99 GB |       3 |                 2h04m
+    #     24665 | public |              lessons |     4 |         4h42m |            161 GB |       4 |                 1m11s
+    #     24661 | public |         lesson_users |     3 |         2h46m |             49 GB |       3 |                39m35s
+    #     24631 | public |               emails |     1 |        34m07s |             10 GB |       2 |                   17s
+    #     24739 | public |             payments |     1 |         5m47s |           1848 MB |       4 |                 4m40s
+    #     24681 | public |         module_users |     1 |         4m57s |           1610 MB |       3 |                 1m50s
+    #     24694 | public |               orders |     1 |         2m50s |            835 MB |       3 |                 1m05s
+    #     24597 | public |              devices |     1 |         1m45s |            498 MB |       2 |                   40s
+    #     24723 | public |      payment_methods |     1 |         1m24s |            548 MB |       2 |                   31s
+    #     24765 | public |     user_collections |     1 |         2m17s |           1005 MB |       2 |                 968ms
+    #     24774 | public |                users |     1 |           52s |            291 MB |       4 |                   27s
+    #     24760 | public |        user_accounts |     1 |           16s |            172 MB |       3 |                   16s
+    #     24606 | public |      edition_modules |     1 |         8s983 |             46 MB |       3 |                 4s749
+    #     24583 | public |        course_emails |     1 |         8s526 |             26 MB |       2 |                 996ms
+    #     24685 | public |              modules |     1 |         1s592 |             21 MB |       3 |                 1s696
+    #     24610 | public |             editions |     1 |         2s199 |           7483 kB |       2 |                 1s032
+    #     24755 | public |           sp_content |     1 |         1s555 |           4177 kB |       0 |                   0ms
+    #     24619 | public |     email_broadcasts |     1 |         744ms |           2645 kB |       2 |                 677ms
+    #     24590 | public |              courses |     1 |         387ms |           1540 kB |       2 |                 367ms
+    #     24704 | public | payment_gateway_keys |     1 |         1s972 |            164 kB |       2 |                  27ms
+    #     24576 | public |             accounts |     1 |          58ms |             24 kB |       1 |                  14ms
+    #     24647 | public |          event_names |     1 |          32ms |             397 B |       1 |                   8ms
+    #     24716 | public |     payment_gateways |     1 |         1s675 |             117 B |       1 |                  11ms
+    #     24748 | public |                roles |     1 |          71ms |             173 B |       1 |                   8ms
+    #     24676 | public |           management |     1 |          33ms |              40 B |       1 |                  19ms
+
+
+    #                                                   Step   Connection    Duration    Transfer   Concurrency
+    #     --------------------------------------------------   ----------  ----------  ----------  ------------
+    #       Catalog Queries (table ordering, filtering, etc)       source         12s                         1
+    #                                             Dump Schema       source       765ms                         1
+    #                                         Prepare Schema       target       466ms                         1
+    #           COPY, INDEX, CONSTRAINTS, VACUUM (wall clock)         both       2h47m                        12
+    #                                       COPY (cumulative)         both       7h46m     1225 GB             4
+    #                               CREATE INDEX (cumulative)       target       4h36m                         4
+    #                               CONSTRAINTS (cumulative)       target       8s493                         4
+    #                                     VACUUM (cumulative)       target         0ms                         4
+    #                                         Reset Sequences         both        60ms                         1
+    #                             Large Objects (cumulative)       (null)         0ms                         0
+    #                                         Finalize Schema         both      14h01m                         4
+    #     --------------------------------------------------   ----------  ----------  ----------  ------------
+    #                               Total Wall Clock Duration         both      16h49m                        20
+
+
+    #     EOF
+
+
+    - name: show tables sizes and retrieve current backpressure seconds
+      run: |
+        export LD_LIBRARY_PATH=${PG_16_LIB_PATH}
+        ${PSQL} "${NEW_PROJECT_CONNSTR}" -c "\dt+"
+        BACKPRESSURE_TIME_AFTER_INGEST=$(${PSQL} "${NEW_PROJECT_CONNSTR}" -t -c "select backpressure_throttling_time()/1000000;")
+        echo "BACKPRESSURE_TIME_AFTER_INGEST=${BACKPRESSURE_TIME_AFTER_INGEST}" >> $GITHUB_ENV
+
+    - name: Parse pgcopydb log and report performance metrics
+      env:
+        PERF_TEST_RESULT_CONNSTR: ${{ secrets.PERF_TEST_RESULT_CONNSTR }}
+      run: |
+        export LD_LIBRARY_PATH=${PG_16_LIB_PATH}
+
+        # Define the log file path
+        LOG_FILE="/tmp/pgcopydb_${{ matrix.target_project }}.log"
+        
+        # Get the current git commit hash
+        git config --global --add safe.directory /__w/neon/neon
+        COMMIT_HASH=$(git rev-parse --short HEAD)
+        
+        # Define the platform and test suite
+        PLATFORM="pg16-${{ matrix.target_project }}-us-east-2-staging"
+        SUIT="pgcopydb_ingest_bench"
+        
+        # Function to convert time (e.g., "2h47m", "4h36m", "118ms", "8s493") to seconds
+        convert_to_seconds() {
+          local duration=$1
+          local total_seconds=0
+    
+          # Check for hours (h)
+          if [[ "$duration" =~ ([0-9]+)h ]]; then
+            total_seconds=$((total_seconds + ${BASH_REMATCH[1]#0} * 3600))
+          fi
+    
+          # Check for seconds (s)
+          if [[ "$duration" =~ ([0-9]+)s ]]; then
+            total_seconds=$((total_seconds + ${BASH_REMATCH[1]#0}))
+          fi
+    
+          # Check for milliseconds (ms) (if applicable)
+          if [[ "$duration" =~ ([0-9]+)ms ]]; then
+            total_seconds=$((total_seconds + ${BASH_REMATCH[1]#0} / 1000))
+            duration=${duration/${BASH_REMATCH[0]}/} # need to remove it to avoid double counting with m 
+          fi
+
+          # Check for minutes (m) - must be checked after ms because m is contained in ms
+          if [[ "$duration" =~ ([0-9]+)m ]]; then
+            total_seconds=$((total_seconds + ${BASH_REMATCH[1]#0} * 60))
+          fi
+    
+          echo $total_seconds
+        }
+
+        # Calculate the backpressure difference in seconds
+        BACKPRESSURE_TIME_DIFF=$(awk "BEGIN {print $BACKPRESSURE_TIME_AFTER_INGEST - $BACKPRESSURE_TIME_BEFORE_INGEST}")
+
+        # Insert the backpressure time difference into the performance database
+        if [ -n "$BACKPRESSURE_TIME_DIFF" ]; then
+          PSQL_CMD="${PSQL} \"${PERF_TEST_RESULT_CONNSTR}\" -c \"
+          INSERT INTO public.perf_test_results (suit, revision, platform, metric_name, metric_value, metric_unit, metric_report_type, recorded_at_timestamp)
+          VALUES ('${SUIT}', '${COMMIT_HASH}', '${PLATFORM}', 'backpressure_time', ${BACKPRESSURE_TIME_DIFF}, 'seconds', 'lower_is_better', now());
+          \""
+          echo "Inserting backpressure time difference: ${BACKPRESSURE_TIME_DIFF} seconds"
+          eval $PSQL_CMD
+        fi
+
+        # Extract and process log lines
+        while IFS= read -r line; do
+          METRIC_NAME=""
+          # Match each desired line and extract the relevant information
+          if [[ "$line" =~ COPY,\ INDEX,\ CONSTRAINTS,\ VACUUM.* ]]; then
+            METRIC_NAME="COPY, INDEX, CONSTRAINTS, VACUUM (wall clock)"
+          elif [[ "$line" =~ COPY\ \(cumulative\).* ]]; then
+            METRIC_NAME="COPY (cumulative)"
+          elif [[ "$line" =~ CREATE\ INDEX\ \(cumulative\).* ]]; then
+            METRIC_NAME="CREATE INDEX (cumulative)"
+          elif [[ "$line" =~ CONSTRAINTS\ \(cumulative\).* ]]; then
+            METRIC_NAME="CONSTRAINTS (cumulative)"
+          elif [[ "$line" =~ Finalize\ Schema.* ]]; then
+            METRIC_NAME="Finalize Schema"
+          elif [[ "$line" =~ Total\ Wall\ Clock\ Duration.* ]]; then
+            METRIC_NAME="Total Wall Clock Duration"
+          fi
+          
+          # If a metric was matched, insert it into the performance database
+          if [ -n "$METRIC_NAME" ]; then
+            DURATION=$(echo "$line" | grep -oP '\d+h\d+m|\d+s|\d+ms|\d{1,2}h\d{1,2}m|\d+\.\d+s' | head -n 1)
+            METRIC_VALUE=$(convert_to_seconds "$DURATION")
+            PSQL_CMD="${PSQL} \"${PERF_TEST_RESULT_CONNSTR}\" -c \"
+            INSERT INTO public.perf_test_results (suit, revision, platform, metric_name, metric_value, metric_unit, metric_report_type, recorded_at_timestamp)
+            VALUES ('${SUIT}', '${COMMIT_HASH}', '${PLATFORM}', '${METRIC_NAME}', ${METRIC_VALUE}, 'seconds', 'lower_is_better', now());
+            \""
+            echo "Inserting ${METRIC_NAME} with value ${METRIC_VALUE} seconds"
+            eval $PSQL_CMD
+          fi
+        done < "$LOG_FILE"
+      
+    - name: Delete Neon Project
+      if: ${{ always() && matrix.target_project == 'new_empty_project' }}
+      uses: ./.github/actions/neon-project-delete
+      with:
+        project_id: ${{ steps.create-neon-project-ingest-target.outputs.project_id }}
+        api_key: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Delete Neon Branch for large tenant
+      if: ${{ always() && matrix.target_project == 'large_existing_project' }}
+      uses: ./.github/actions/neon-branch-delete
+      with:
+        project_id: ${{ vars.BENCHMARK_INGEST_TARGET_PROJECTID }}
+        branch_id: ${{ steps.create-neon-branch-ingest-target.outputs.branch_id }}
+        api_key: ${{ secrets.NEON_STAGING_API_KEY }}

.github/workflows/neon_extra_builds.yml

@@ -201,6 +201,8 @@ jobs:
           REPORT_URL: ${{ steps.upload-stats.outputs.report-url }}
           SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
           script: |
             const { REPORT_URL, SHA } = process.env
 

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

@@ -0,0 +1,94 @@
+name: Pre-merge checks
+
+on:
+  merge_group:
+    branches:
+      - main
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+
+# No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
+permissions: {}
+
+jobs:
+  get-changed-files:
+    runs-on: ubuntu-22.04
+    outputs:
+      python-changed: ${{ steps.python-src.outputs.any_changed }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: tj-actions/changed-files@4edd678ac3f81e2dc578756871e4d00c19191daf # v45.0.4
+        id: python-src
+        with:
+          files: |
+            .github/workflows/pre-merge-checks.yml
+            **/**.py
+            poetry.lock
+            pyproject.toml
+
+      - name: PRINT ALL CHANGED FILES FOR DEBUG PURPOSES
+        env:
+          PYTHON_CHANGED_FILES: ${{ steps.python-src.outputs.all_changed_files }}
+        run: |
+          echo "${PYTHON_CHANGED_FILES}"
+
+  check-build-tools-image:
+    if: needs.get-changed-files.outputs.python-changed == 'true'
+    needs: [ get-changed-files ]
+    uses: ./.github/workflows/check-build-tools-image.yml
+
+  build-build-tools-image:
+    needs: [ check-build-tools-image ]
+    uses: ./.github/workflows/build-build-tools-image.yml
+    with:
+      image-tag: ${{ needs.check-build-tools-image.outputs.image-tag }}
+    secrets: inherit
+
+  check-codestyle-python:
+    if: needs.get-changed-files.outputs.python-changed == 'true'
+    needs: [ get-changed-files, build-build-tools-image ]
+    uses: ./.github/workflows/_check-codestyle-python.yml
+    with:
+      build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
+    secrets: inherit
+
+  # To get items from the merge queue merged into main we need to satisfy "Status checks that are required".
+  # Currently we require 2 jobs (checks with exact name):
+  # - conclusion
+  # - neon-cloud-e2e
+  conclusion:
+    if: always()
+    permissions:
+      statuses: write # for `github.repos.createCommitStatus(...)`
+    needs:
+      - get-changed-files
+      - check-codestyle-python
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Create fake `neon-cloud-e2e` check
+        uses: actions/github-script@v7
+        with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
+          script: |
+            const { repo, owner } = context.repo;
+            const targetUrl = `${context.serverUrl}/${owner}/${repo}/actions/runs/${context.runId}`;
+
+            await github.rest.repos.createCommitStatus({
+              owner: owner,
+              repo: repo,
+              sha: context.sha,
+              context: `neon-cloud-e2e`,
+              state: `success`,
+              target_url: targetUrl,
+              description: `fake check for merge queue`,
+            });
+
+      - name: Fail the job if any of the dependencies do not succeed or skipped
+        run: exit 1
+        if: |
+          (contains(needs.check-codestyle-python.result, 'skipped') && needs.get-changed-files.outputs.python-changed == 'true')
+          || contains(needs.*.result, 'failure')
+          || contains(needs.*.result, 'cancelled')

.github/workflows/report-workflow-stats-batch.yml

@@ -0,0 +1,29 @@
+name: Report Workflow Stats Batch
+
+on:
+  schedule:
+    - cron: '*/15 * * * *'
+    - cron: '25 0 * * *'
+
+jobs:
+  gh-workflow-stats-batch:
+    name: GitHub Workflow Stats Batch
+    runs-on: ubuntu-22.04
+    permissions:
+      actions: read
+    steps:
+    - name: Export Workflow Run for the past 2 hours
+      uses: neondatabase/gh-workflow-stats-action@v0.2.1
+      with:
+        db_uri: ${{ secrets.GH_REPORT_STATS_DB_RW_CONNSTR }}
+        db_table: "gh_workflow_stats_batch_neon"
+        gh_token: ${{ secrets.GITHUB_TOKEN }}
+        duration: '2h'
+    - name: Export Workflow Run for the past 24 hours
+      if: github.event.schedule == '25 0 * * *'
+      uses: neondatabase/gh-workflow-stats-action@v0.2.1
+      with:
+        db_uri: ${{ secrets.GH_REPORT_STATS_DB_RW_CONNSTR }}
+        db_table: "gh_workflow_stats_batch_neon"
+        gh_token: ${{ secrets.GITHUB_TOKEN }}
+        duration: '24h'

.github/workflows/report-workflow-stats.yml

@@ -23,6 +23,7 @@ on:
     - Test Postgres client libraries
     - Trigger E2E Tests
     - cleanup caches by a branch
+    - Pre-merge checks
     types: [completed]
 
 jobs:

Cargo.lock

@@ -1229,12 +1229,15 @@ dependencies = [
  "flate2",
  "futures",
  "hyper 0.14.30",
+ "metrics",
  "nix 0.27.1",
  "notify",
  "num_cpus",
+ "once_cell",
  "opentelemetry",
  "opentelemetry_sdk",
  "postgres",
+ "prometheus",
  "regex",
  "remote_storage",
  "reqwest 0.12.4",
@@ -4743,6 +4746,7 @@ dependencies = [
  "percent-encoding",
  "pin-project-lite",
  "rustls 0.22.4",
+ "rustls-native-certs 0.7.0",
  "rustls-pemfile 2.1.1",
  "rustls-pki-types",
  "serde",
@@ -5146,6 +5150,7 @@ dependencies = [
  "chrono",
  "clap",
  "crc32c",
+ "criterion",
  "desim",
  "fail",
  "futures",
@@ -5153,6 +5158,7 @@ dependencies = [
  "http 1.1.0",
  "humantime",
  "hyper 0.14.30",
+ "itertools 0.10.5",
  "metrics",
  "once_cell",
  "parking_lot 0.12.1",

build-tools.Dockerfile

@@ -1,12 +1,66 @@
 ARG DEBIAN_VERSION=bullseye
 
-FROM debian:${DEBIAN_VERSION}-slim
+FROM debian:bookworm-slim AS pgcopydb_builder
+ARG DEBIAN_VERSION
+
+RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
+        set -e && \
+        apt update && \
+        apt install -y --no-install-recommends \
+        ca-certificates wget gpg && \
+        wget -qO - https://www.postgresql.org/media/keys/ACCC4CF8.asc | gpg --dearmor -o /usr/share/keyrings/postgresql-keyring.gpg && \
+        echo "deb [signed-by=/usr/share/keyrings/postgresql-keyring.gpg] http://apt.postgresql.org/pub/repos/apt bookworm-pgdg main" > /etc/apt/sources.list.d/pgdg.list && \
+        apt-get update && \
+        apt install -y --no-install-recommends \
+        build-essential \
+        autotools-dev \
+        libedit-dev \
+        libgc-dev \
+        libpam0g-dev \
+        libreadline-dev \
+        libselinux1-dev \
+        libxslt1-dev \
+        libssl-dev \
+        libkrb5-dev \
+        zlib1g-dev \
+        liblz4-dev \
+        libpq5 \
+        libpq-dev \
+        libzstd-dev \
+        postgresql-16 \
+        postgresql-server-dev-16 \
+        postgresql-common  \
+        python3-sphinx && \
+        wget -O /tmp/pgcopydb.tar.gz https://github.com/dimitri/pgcopydb/archive/refs/tags/v0.17.tar.gz && \
+        mkdir /tmp/pgcopydb && \
+        tar -xzf /tmp/pgcopydb.tar.gz -C /tmp/pgcopydb --strip-components=1 && \
+        cd /tmp/pgcopydb && \
+        make -s clean && \
+        make -s -j12 install && \
+        libpq_path=$(find /lib /usr/lib -name "libpq.so.5" | head -n 1) && \
+        mkdir -p /pgcopydb/lib && \
+        cp "$libpq_path" /pgcopydb/lib/; \
+    else \
+        # copy command below will fail if we don't have dummy files, so we create them for other debian versions
+        mkdir -p /usr/lib/postgresql/16/bin && touch /usr/lib/postgresql/16/bin/pgcopydb && \
+        mkdir -p mkdir -p /pgcopydb/lib && touch /pgcopydb/lib/libpq.so.5; \
+    fi
+
+FROM debian:${DEBIAN_VERSION}-slim AS build_tools
 ARG DEBIAN_VERSION
 
 # Add nonroot user
 RUN useradd -ms /bin/bash nonroot -b /home
 SHELL ["/bin/bash", "-c"]
 
+RUN mkdir -p /pgcopydb/bin && \
+    mkdir -p /pgcopydb/lib && \
+    chmod -R 755 /pgcopydb && \
+    chown -R nonroot:nonroot /pgcopydb
+        
+COPY --from=pgcopydb_builder /usr/lib/postgresql/16/bin/pgcopydb /pgcopydb/bin/pgcopydb 
+COPY --from=pgcopydb_builder /pgcopydb/lib/libpq.so.5 /pgcopydb/lib/libpq.so.5 
+
 # System deps
 #
 # 'gdb' is included so that we get backtraces of core dumps produced in
@@ -38,7 +92,7 @@ RUN set -e \
         libseccomp-dev \
         libsqlite3-dev \
         libssl-dev \
-        $([[ "${DEBIAN_VERSION}" = "bullseye" ]] && libstdc++-10-dev || libstdc++-11-dev) \
+        $([[ "${DEBIAN_VERSION}" = "bullseye" ]] && echo libstdc++-10-dev || echo libstdc++-11-dev) \
         libtool \
         libxml2-dev \
         libxmlsec1-dev \
@@ -235,7 +289,13 @@ RUN whoami \
     && cargo --version --verbose \
     && rustup --version --verbose \
     && rustc --version --verbose \
-    && clang --version
+    && clang --version 
+
+RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
+    LD_LIBRARY_PATH=/pgcopydb/lib /pgcopydb/bin/pgcopydb --version; \
+else \
+    echo "pgcopydb is not available for ${DEBIAN_VERSION}"; \
+fi
 
 # Set following flag to check in Makefile if its running in Docker
 RUN touch /home/nonroot/.docker_build

compute/compute-node.Dockerfile

@@ -559,8 +559,8 @@ RUN case "${PG_VERSION}" in \
         export TIMESCALEDB_CHECKSUM=584a351c7775f0e067eaa0e7277ea88cab9077cc4c455cbbf09a5d9723dce95d \
         ;; \
       "v17") \
-        export TIMESCALEDB_VERSION=2.17.0 \
-        export TIMESCALEDB_CHECKSUM=155bf64391d3558c42f31ca0e523cfc6252921974f75298c9039ccad1c89811a \
+        export TIMESCALEDB_VERSION=2.17.1 \
+        export TIMESCALEDB_CHECKSUM=6277cf43f5695e23dae1c5cfeba00474d730b66ed53665a84b787a6bb1a57e28 \
         ;; \
     esac && \
     wget https://github.com/timescale/timescaledb/archive/refs/tags/${TIMESCALEDB_VERSION}.tar.gz -O timescaledb.tar.gz && \
@@ -624,16 +624,12 @@ FROM build-deps AS pg-cron-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-# 1.6.4 available, supports v17
 # This is an experimental extension that we do not support on prod yet.
 # !Do not remove!
 # We set it in shared_preload_libraries and computes will fail to start if library is not found.
 ENV PATH="/usr/local/pgsql/bin/:$PATH"
-RUN case "${PG_VERSION}" in "v17") \
-    echo "v17 extensions are not supported yet. Quit" && exit 0;; \
-    esac && \
-    wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.6.0.tar.gz -O pg_cron.tar.gz && \
-    echo "383a627867d730222c272bfd25cd5e151c578d73f696d32910c7db8c665cc7db pg_cron.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.6.4.tar.gz -O pg_cron.tar.gz && \
+    echo "52d1850ee7beb85a4cb7185731ef4e5a90d1de216709d8988324b0d02e76af61 pg_cron.tar.gz" | sha256sum --check && \
     mkdir pg_cron-src && cd pg_cron-src && tar xzf ../pg_cron.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -1475,6 +1471,8 @@ RUN mkdir -p /etc/local_proxy && chown postgres:postgres /etc/local_proxy
 COPY --from=postgres-exporter /bin/postgres_exporter /bin/postgres_exporter
 COPY --from=sql-exporter      /bin/sql_exporter      /bin/sql_exporter
 
+COPY --chown=postgres compute/etc/postgres_exporter.yml /etc/postgres_exporter.yml
+
 COPY --from=sql_exporter_preprocessor --chmod=0644 /home/nonroot/compute/etc/sql_exporter.yml               /etc/sql_exporter.yml
 COPY --from=sql_exporter_preprocessor --chmod=0644 /home/nonroot/compute/etc/neon_collector.yml             /etc/neon_collector.yml
 COPY --from=sql_exporter_preprocessor --chmod=0644 /home/nonroot/compute/etc/sql_exporter_autoscaling.yml   /etc/sql_exporter_autoscaling.yml

compute/etc/sql_exporter/compute_backpressure_throttling_seconds.sql

@@ -1 +1 @@
-SELECT neon.backpressure_throttling_time()::float8 / 1000 AS throttled;
+SELECT (neon.backpressure_throttling_time()::float8 / 1000000) AS throttled;

compute/patches/pg_anon.patch

@@ -1,3 +1,45 @@
+commit 00aa659afc9c7336ab81036edec3017168aabf40
+Author: Heikki Linnakangas <heikki@neon.tech>
+Date:   Tue Nov 12 16:59:19 2024 +0200
+
+    Temporarily disable test that depends on timezone
+
+diff --git a/tests/expected/generalization.out b/tests/expected/generalization.out
+index 23ef5fa..9e60deb 100644
+--- a/ext-src/pg_anon-src/tests/expected/generalization.out
++++ b/ext-src/pg_anon-src/tests/expected/generalization.out
+@@ -284,12 +284,9 @@ SELECT anon.generalize_tstzrange('19041107','century');
+  ["Tue Jan 01 00:00:00 1901 PST","Mon Jan 01 00:00:00 2001 PST")
+ (1 row)
+ 
+-SELECT anon.generalize_tstzrange('19041107','millennium');
+-                      generalize_tstzrange                       
+------------------------------------------------------------------
+- ["Thu Jan 01 00:00:00 1001 PST","Mon Jan 01 00:00:00 2001 PST")
+-(1 row)
+-
++-- temporarily disabled, see:
++-- https://gitlab.com/dalibo/postgresql_anonymizer/-/commit/199f0a392b37c59d92ae441fb8f037e094a11a52#note_2148017485
++--SELECT anon.generalize_tstzrange('19041107','millennium');
+ -- generalize_daterange
+ SELECT anon.generalize_daterange('19041107');
+   generalize_daterange   
+diff --git a/tests/sql/generalization.sql b/tests/sql/generalization.sql
+index b868344..b4fc977 100644
+--- a/ext-src/pg_anon-src/tests/sql/generalization.sql
++++ b/ext-src/pg_anon-src/tests/sql/generalization.sql
+@@ -61,7 +61,9 @@ SELECT anon.generalize_tstzrange('19041107','month');
+ SELECT anon.generalize_tstzrange('19041107','year');
+ SELECT anon.generalize_tstzrange('19041107','decade');
+ SELECT anon.generalize_tstzrange('19041107','century');
+-SELECT anon.generalize_tstzrange('19041107','millennium');
++-- temporarily disabled, see:
++-- https://gitlab.com/dalibo/postgresql_anonymizer/-/commit/199f0a392b37c59d92ae441fb8f037e094a11a52#note_2148017485
++--SELECT anon.generalize_tstzrange('19041107','millennium');
+ 
+ -- generalize_daterange
+ SELECT anon.generalize_daterange('19041107');
+
 commit 7dd414ee75f2875cffb1d6ba474df1f135a6fc6f
 Author: Alexey Masterov <alexeymasterov@neon.tech>
 Date:   Fri May 31 06:34:26 2024 +0000

compute/vm-image-spec-bookworm.yaml

@@ -26,7 +26,7 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter'
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
   - name: sql-exporter
     user: nobody
     sysvInitAction: respawn

compute/vm-image-spec-bullseye.yaml

@@ -26,7 +26,7 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter'
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
   - name: sql-exporter
     user: nobody
     sysvInitAction: respawn

compute_tools/Cargo.toml

@@ -18,9 +18,11 @@ clap.workspace = true
 flate2.workspace = true
 futures.workspace = true
 hyper0 = { workspace = true, features = ["full"] }
+metrics.workspace = true
 nix.workspace = true
 notify.workspace = true
 num_cpus.workspace = true
+once_cell.workspace = true
 opentelemetry.workspace = true
 opentelemetry_sdk.workspace = true
 postgres.workspace = true
@@ -39,6 +41,7 @@ tracing-subscriber.workspace = true
 tracing-utils.workspace = true
 thiserror.workspace = true
 url.workspace = true
+prometheus.workspace = true
 
 compute_api.workspace = true
 utils.workspace = true

compute_tools/src/compute.rs

@@ -364,11 +364,29 @@ impl ComputeNode {
         let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
 
         let basebackup_cmd = match lsn {
-            Lsn(0) => format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id),
-            _ => format!(
-                "basebackup {} {} {} --gzip",
-                spec.tenant_id, spec.timeline_id, lsn
-            ),
+            Lsn(0) => {
+                if spec.spec.mode != ComputeMode::Primary {
+                    format!(
+                        "basebackup {} {} --gzip --replica",
+                        spec.tenant_id, spec.timeline_id
+                    )
+                } else {
+                    format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
+                }
+            }
+            _ => {
+                if spec.spec.mode != ComputeMode::Primary {
+                    format!(
+                        "basebackup {} {} {} --gzip --replica",
+                        spec.tenant_id, spec.timeline_id, lsn
+                    )
+                } else {
+                    format!(
+                        "basebackup {} {} {} --gzip",
+                        spec.tenant_id, spec.timeline_id, lsn
+                    )
+                }
+            }
         };
 
         let copyreader = client.copy_out(basebackup_cmd.as_str())?;

compute_tools/src/config.rs

@@ -73,6 +73,19 @@ pub fn write_postgres_conf(
         )?;
     }
 
+    // Locales
+    if cfg!(target_os = "macos") {
+        writeln!(file, "lc_messages='C'")?;
+        writeln!(file, "lc_monetary='C'")?;
+        writeln!(file, "lc_time='C'")?;
+        writeln!(file, "lc_numeric='C'")?;
+    } else {
+        writeln!(file, "lc_messages='C.UTF-8'")?;
+        writeln!(file, "lc_monetary='C.UTF-8'")?;
+        writeln!(file, "lc_time='C.UTF-8'")?;
+        writeln!(file, "lc_numeric='C.UTF-8'")?;
+    }
+
     match spec.mode {
         ComputeMode::Primary => {}
         ComputeMode::Static(lsn) => {

compute_tools/src/http/api.rs

@@ -9,6 +9,7 @@ use crate::catalog::SchemaDumpError;
 use crate::catalog::{get_database_schema, get_dbs_and_roles};
 use crate::compute::forward_termination_signal;
 use crate::compute::{ComputeNode, ComputeState, ParsedSpec};
+use crate::installed_extensions;
 use compute_api::requests::{ConfigurationRequest, ExtensionInstallRequest, SetRoleGrantsRequest};
 use compute_api::responses::{
     ComputeStatus, ComputeStatusResponse, ExtensionInstallResult, GenericAPIError,
@@ -19,6 +20,8 @@ use anyhow::Result;
 use hyper::header::CONTENT_TYPE;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::{Body, Method, Request, Response, Server, StatusCode};
+use metrics::Encoder;
+use metrics::TextEncoder;
 use tokio::task;
 use tracing::{debug, error, info, warn};
 use tracing_utils::http::OtelName;
@@ -65,6 +68,28 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
             Response::new(Body::from(serde_json::to_string(&metrics).unwrap()))
         }
 
+        // Prometheus metrics
+        (&Method::GET, "/metrics") => {
+            debug!("serving /metrics GET request");
+
+            let mut buffer = vec![];
+            let metrics = installed_extensions::collect();
+            let encoder = TextEncoder::new();
+            encoder.encode(&metrics, &mut buffer).unwrap();
+
+            match Response::builder()
+                .status(StatusCode::OK)
+                .header(CONTENT_TYPE, encoder.format_type())
+                .body(Body::from(buffer))
+            {
+                Ok(response) => response,
+                Err(err) => {
+                    let msg = format!("error handling /metrics request: {err}");
+                    error!(msg);
+                    render_json_error(&msg, StatusCode::INTERNAL_SERVER_ERROR)
+                }
+            }
+        }
         // Collect Postgres current usage insights
         (&Method::GET, "/insights") => {
             info!("serving /insights GET request");

compute_tools/src/http/openapi_spec.yaml

@@ -37,6 +37,21 @@ paths:
               schema:
                 $ref: "#/components/schemas/ComputeMetrics"
 
+  /metrics
+    get:
+      tags:
+      - Info
+      summary: Get compute node metrics in text format.
+      description: ""
+      operationId: getComputeMetrics
+      responses:
+        200:
+          description: ComputeMetrics
+          content:
+            text/plain:
+              schema:
+                type: string
+                description: Metrics in text format.
   /insights:
     get:
       tags:

compute_tools/src/installed_extensions.rs

@@ -1,4 +1,5 @@
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use metrics::proto::MetricFamily;
 use std::collections::HashMap;
 use std::collections::HashSet;
 use tracing::info;
@@ -8,6 +9,10 @@ use anyhow::Result;
 use postgres::{Client, NoTls};
 use tokio::task;
 
+use metrics::core::Collector;
+use metrics::{register_uint_gauge_vec, UIntGaugeVec};
+use once_cell::sync::Lazy;
+
 /// We don't reuse get_existing_dbs() just for code clarity
 /// and to make database listing query here more explicit.
 ///
@@ -59,6 +64,12 @@ pub async fn get_installed_extensions(connstr: Url) -> Result<InstalledExtension
 
             for (extname, v) in extensions.iter() {
                 let version = v.to_string();
+
+                // increment the number of databases where the version of extension is installed
+                INSTALLED_EXTENSIONS
+                    .with_label_values(&[extname, &version])
+                    .inc();
+
                 extensions_map
                     .entry(extname.to_string())
                     .and_modify(|e| {
@@ -74,9 +85,11 @@ pub async fn get_installed_extensions(connstr: Url) -> Result<InstalledExtension
             }
         }
 
-        Ok(InstalledExtensions {
+        let res = InstalledExtensions {
             extensions: extensions_map.values().cloned().collect(),
-        })
+        };
+
+        Ok(res)
     })
     .await?
 }
@@ -97,6 +110,18 @@ pub fn get_installed_extensions_sync(connstr: Url) -> Result<()> {
         "[NEON_EXT_STAT] {}",
         serde_json::to_string(&result).expect("failed to serialize extensions list")
     );
-
     Ok(())
 }
+
+static INSTALLED_EXTENSIONS: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "installed_extensions",
+        "Number of databases where the version of extension is installed",
+        &["extension_name", "version"]
+    )
+    .expect("failed to define a metric")
+});
+
+pub fn collect() -> Vec<MetricFamily> {
+    INSTALLED_EXTENSIONS.collect()
+}

control_plane/src/bin/neon_local.rs

@@ -944,6 +944,9 @@ fn handle_init(args: &InitCmdArgs) -> anyhow::Result<LocalEnv> {
                         pg_auth_type: AuthType::Trust,
                         http_auth_type: AuthType::Trust,
                         other: Default::default(),
+                        // Typical developer machines use disks with slow fsync, and we don't care
+                        // about data integrity: disable disk syncs.
+                        no_sync: true,
                     }
                 })
                 .collect(),

control_plane/src/local_env.rs

@@ -225,6 +225,7 @@ pub struct PageServerConf {
     pub listen_http_addr: String,
     pub pg_auth_type: AuthType,
     pub http_auth_type: AuthType,
+    pub no_sync: bool,
 }
 
 impl Default for PageServerConf {
@@ -235,6 +236,7 @@ impl Default for PageServerConf {
             listen_http_addr: String::new(),
             pg_auth_type: AuthType::Trust,
             http_auth_type: AuthType::Trust,
+            no_sync: false,
         }
     }
 }
@@ -249,6 +251,8 @@ pub struct NeonLocalInitPageserverConf {
     pub listen_http_addr: String,
     pub pg_auth_type: AuthType,
     pub http_auth_type: AuthType,
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    pub no_sync: bool,
     #[serde(flatten)]
     pub other: HashMap<String, toml::Value>,
 }
@@ -261,6 +265,7 @@ impl From<&NeonLocalInitPageserverConf> for PageServerConf {
             listen_http_addr,
             pg_auth_type,
             http_auth_type,
+            no_sync,
             other: _,
         } = conf;
         Self {
@@ -269,6 +274,7 @@ impl From<&NeonLocalInitPageserverConf> for PageServerConf {
             listen_http_addr: listen_http_addr.clone(),
             pg_auth_type: *pg_auth_type,
             http_auth_type: *http_auth_type,
+            no_sync: *no_sync,
         }
     }
 }
@@ -569,6 +575,8 @@ impl LocalEnv {
                     listen_http_addr: String,
                     pg_auth_type: AuthType,
                     http_auth_type: AuthType,
+                    #[serde(default)]
+                    no_sync: bool,
                 }
                 let config_toml_path = dentry.path().join("pageserver.toml");
                 let config_toml: PageserverConfigTomlSubset = toml_edit::de::from_str(
@@ -591,6 +599,7 @@ impl LocalEnv {
                     listen_http_addr,
                     pg_auth_type,
                     http_auth_type,
+                    no_sync,
                 } = config_toml;
                 let IdentityTomlSubset {
                     id: identity_toml_id,
@@ -607,6 +616,7 @@ impl LocalEnv {
                     listen_http_addr,
                     pg_auth_type,
                     http_auth_type,
+                    no_sync,
                 };
                 pageservers.push(conf);
             }

control_plane/src/pageserver.rs

@@ -273,6 +273,7 @@ impl PageServerNode {
             )
         })?;
         let args = vec!["-D", datadir_path_str];
+
         background_process::start_process(
             "pageserver",
             &datadir,

docs/rfcs/038-aux-file-v2.md

@@ -91,7 +91,7 @@ generating the basebackup by scanning the `REPL_ORIGIN_KEY_PREFIX` keyspace.
 There are two places we need to read the aux files from the pageserver:
 
 * On the write path, when the compute node adds an aux file to the pageserver, we will retrieve the key from the storage, append the file to the hashed key, and write it back. The current `get` API already supports that.
-*  We use the vectored get API to retrieve all aux files during generating the basebackup. Because we need to scan a sparse keyspace, we slightly modified the vectored get path. The vectorized API will attempt to retrieve every single key within the requested key range, and therefore, we modified it in a way that keys within `NON_INHERITED_SPARSE_RANGE` will not trigger missing key error.
+*  We use the vectored get API to retrieve all aux files during generating the basebackup. Because we need to scan a sparse keyspace, we slightly modified the vectored get path. The vectorized API used to always attempt to retrieve every single key within the requested key range, and therefore, we modified it in a way that keys within `NON_INHERITED_SPARSE_RANGE` will not trigger missing key error. Furthermore, as aux file reads usually need all layer files intersecting with that key range within the branch and cover a big keyspace, it incurs large overhead for tracking keyspaces that have not been read. Therefore, for sparse keyspaces, we [do not track](https://github.com/neondatabase/neon/pull/9631) `ummapped_keyspace`.
 
 ## Compaction and Image Layer Generation
 

libs/pageserver_api/src/config.rs

@@ -64,6 +64,7 @@ pub struct ConfigToml {
     #[serde(with = "humantime_serde")]
     pub wal_redo_timeout: Duration,
     pub superuser: String,
+    pub locale: String,
     pub page_cache_size: usize,
     pub max_file_descriptors: usize,
     pub pg_distrib_dir: Option<Utf8PathBuf>,
@@ -106,6 +107,8 @@ pub struct ConfigToml {
     pub ephemeral_bytes_per_memory_kb: usize,
     pub l0_flush: Option<crate::models::L0FlushConfig>,
     pub virtual_file_io_mode: Option<crate::models::virtual_file::IoMode>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub no_sync: Option<bool>,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -274,6 +277,11 @@ pub mod defaults {
     pub const DEFAULT_WAL_REDO_TIMEOUT: &str = "60 s";
 
     pub const DEFAULT_SUPERUSER: &str = "cloud_admin";
+    pub const DEFAULT_LOCALE: &str = if cfg!(target_os = "macos") {
+        "C"
+    } else {
+        "C.UTF-8"
+    };
 
     pub const DEFAULT_PAGE_CACHE_SIZE: usize = 8192;
     pub const DEFAULT_MAX_FILE_DESCRIPTORS: usize = 100;
@@ -324,6 +332,7 @@ impl Default for ConfigToml {
             wal_redo_timeout: (humantime::parse_duration(DEFAULT_WAL_REDO_TIMEOUT)
                 .expect("cannot parse default wal redo timeout")),
             superuser: (DEFAULT_SUPERUSER.to_string()),
+            locale: DEFAULT_LOCALE.to_string(),
             page_cache_size: (DEFAULT_PAGE_CACHE_SIZE),
             max_file_descriptors: (DEFAULT_MAX_FILE_DESCRIPTORS),
             pg_distrib_dir: None, // Utf8PathBuf::from("./pg_install"), // TODO: formely, this was std::env::current_dir()
@@ -389,6 +398,7 @@ impl Default for ConfigToml {
             l0_flush: None,
             virtual_file_io_mode: None,
             tenant_config: TenantConfigToml::default(),
+            no_sync: None,
         }
     }
 }

libs/pageserver_api/src/record.rs

@@ -80,18 +80,18 @@ impl NeonWalRecord {
     }
 
     #[cfg(feature = "testing")]
-    pub fn wal_clear() -> Self {
+    pub fn wal_clear(s: impl AsRef<str>) -> Self {
         Self::Test {
-            append: "".to_string(),
+            append: s.as_ref().to_string(),
             clear: true,
             will_init: false,
         }
     }
 
     #[cfg(feature = "testing")]
-    pub fn wal_init() -> Self {
+    pub fn wal_init(s: impl AsRef<str>) -> Self {
         Self::Test {
-            append: "".to_string(),
+            append: s.as_ref().to_string(),
             clear: true,
             will_init: true,
         }

libs/postgres_ffi/src/wal_generator.rs

@@ -1,10 +1,10 @@
-use std::ffi::CStr;
+use std::ffi::{CStr, CString};
 
 use bytes::{Bytes, BytesMut};
 use crc32c::crc32c_append;
 use utils::lsn::Lsn;
 
-use super::bindings::{XLogLongPageHeaderData, XLogPageHeaderData, XLOG_PAGE_MAGIC};
+use super::bindings::{RmgrId, XLogLongPageHeaderData, XLogPageHeaderData, XLOG_PAGE_MAGIC};
 use super::xlog_utils::{
     XlLogicalMessage, XLOG_RECORD_CRC_OFFS, XLOG_SIZE_OF_XLOG_RECORD, XLP_BKP_REMOVABLE,
     XLP_FIRST_IS_CONTRECORD,
@@ -16,11 +16,65 @@ use crate::pg_constants::{
 };
 use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
 
-/// Generates binary WAL records for use in tests and benchmarks. Currently only generates logical
-/// messages (effectively noops) with a fixed payload. It is used as an iterator which yields
-/// encoded bytes for a single WAL record, including internal page headers if it spans pages.
-/// Concatenating the bytes will yield a complete, well-formed WAL, which can be chunked at segment
-/// boundaries if desired. Not optimized for performance.
+/// A WAL record payload. Will be prefixed by an XLogRecord header when encoded.
+pub struct Record {
+    pub rmid: RmgrId,
+    pub info: u8,
+    pub data: Bytes,
+}
+
+impl Record {
+    /// Encodes the WAL record including an XLogRecord header. prev_lsn is the start position of
+    /// the previous record in the WAL -- this is ignored by the Safekeeper, but not Postgres.
+    pub fn encode(&self, prev_lsn: Lsn) -> Bytes {
+        // Prefix data with block ID and length.
+        let data_header = Bytes::from(match self.data.len() {
+            0 => vec![],
+            1..=255 => vec![XLR_BLOCK_ID_DATA_SHORT, self.data.len() as u8],
+            256.. => {
+                let len_bytes = (self.data.len() as u32).to_le_bytes();
+                [&[XLR_BLOCK_ID_DATA_LONG], len_bytes.as_slice()].concat()
+            }
+        });
+
+        // Construct the WAL record header.
+        let mut header = XLogRecord {
+            xl_tot_len: (XLOG_SIZE_OF_XLOG_RECORD + data_header.len() + self.data.len()) as u32,
+            xl_xid: 0,
+            xl_prev: prev_lsn.into(),
+            xl_info: self.info,
+            xl_rmid: self.rmid,
+            __bindgen_padding_0: [0; 2],
+            xl_crc: 0, // see below
+        };
+
+        // Compute the CRC checksum for the data, and the header up to the CRC field.
+        let mut crc = 0;
+        crc = crc32c_append(crc, &data_header);
+        crc = crc32c_append(crc, &self.data);
+        crc = crc32c_append(crc, &header.encode().unwrap()[0..XLOG_RECORD_CRC_OFFS]);
+        header.xl_crc = crc;
+
+        // Encode the final header and record.
+        let header = header.encode().unwrap();
+
+        [header, data_header, self.data.clone()].concat().into()
+    }
+}
+
+/// Generates WAL record payloads.
+///
+/// TODO: currently only provides LogicalMessageGenerator for trivial noop messages. Add a generator
+/// that creates a table and inserts rows.
+pub trait RecordGenerator: Iterator<Item = Record> {}
+
+impl<I: Iterator<Item = Record>> RecordGenerator for I {}
+
+/// Generates binary WAL for use in tests and benchmarks. The provided record generator constructs
+/// the WAL records. It is used as an iterator which yields encoded bytes for a single WAL record,
+/// including internal page headers if it spans pages. Concatenating the bytes will yield a
+/// complete, well-formed WAL, which can be chunked at segment boundaries if desired. Not optimized
+/// for performance.
 ///
 /// The WAL format is version-dependant (see e.g. `XLOG_PAGE_MAGIC`), so make sure to import this
 /// for the appropriate Postgres version (e.g. `postgres_ffi::v17::wal_generator::WalGenerator`).
@@ -31,10 +85,10 @@ use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
 /// |        Segment 1         |        Segment 2         |        Segment 3         |
 /// | Page 1 | Page 2 | Page 3 | Page 4 | Page 5 | Page 6 | Page 7 | Page 8 | Page 9 |
 /// | R1 |   R2  |R3|  R4  | R5  |  R6  |                 R7            | R8  |
-///
-/// TODO: support generating actual tables and rows.
 #[derive(Default)]
-pub struct WalGenerator {
+pub struct WalGenerator<R: RecordGenerator> {
+    /// Generates record payloads for the WAL.
+    pub record_generator: R,
     /// Current LSN to append the next record at.
     ///
     /// Callers can modify this (and prev_lsn) to restart generation at a different LSN, but should
@@ -46,73 +100,35 @@ pub struct WalGenerator {
     pub prev_lsn: Lsn,
 }
 
-impl WalGenerator {
-    // For now, hardcode the message payload.
-    // TODO: support specifying the payload size.
-    const PREFIX: &CStr = c"prefix";
-    const MESSAGE: &[u8] = b"message";
-
-    // Hardcode the sys, timeline, and DB IDs. We can make them configurable if we care about them.
+impl<R: RecordGenerator> WalGenerator<R> {
+    // Hardcode the sys and timeline ID. We can make them configurable if we care about them.
     const SYS_ID: u64 = 0;
     const TIMELINE_ID: u32 = 1;
-    const DB_ID: u32 = 0;
 
-    /// Creates a new WAL generator, which emits logical message records (noops).
-    pub fn new() -> Self {
-        Self::default()
+    /// Creates a new WAL generator with the given record generator.
+    pub fn new(record_generator: R) -> WalGenerator<R> {
+        Self {
+            record_generator,
+            lsn: Lsn(0),
+            prev_lsn: Lsn(0),
+        }
     }
 
-    /// Encodes a logical message (basically a noop), with the given prefix and message.
-    pub(crate) fn encode_logical_message(prefix: &CStr, message: &[u8]) -> Bytes {
-        let prefix = prefix.to_bytes_with_nul();
-        let header = XlLogicalMessage {
-            db_id: Self::DB_ID,
-            transactional: 0,
-            prefix_size: prefix.len() as u64,
-            message_size: message.len() as u64,
-        };
-        [&header.encode(), prefix, message].concat().into()
+    /// Appends a record with an arbitrary payload at the current LSN, then increments the LSN.
+    /// Returns the WAL bytes for the record, including page headers and padding, and the start LSN.
+    fn append_record(&mut self, record: Record) -> (Lsn, Bytes) {
+        let record = record.encode(self.prev_lsn);
+        let record = Self::insert_pages(record, self.lsn);
+        let record = Self::pad_record(record, self.lsn);
+        let lsn = self.lsn;
+        self.prev_lsn = self.lsn;
+        self.lsn += record.len() as u64;
+        (lsn, record)
     }
 
-    /// Encode a WAL record with the given payload data (e.g. a logical message).
-    pub(crate) fn encode_record(data: Bytes, rmid: u8, info: u8, prev_lsn: Lsn) -> Bytes {
-        // Prefix data with block ID and length.
-        let data_header = Bytes::from(match data.len() {
-            0 => vec![],
-            1..=255 => vec![XLR_BLOCK_ID_DATA_SHORT, data.len() as u8],
-            256.. => {
-                let len_bytes = (data.len() as u32).to_le_bytes();
-                [&[XLR_BLOCK_ID_DATA_LONG], len_bytes.as_slice()].concat()
-            }
-        });
-
-        // Construct the WAL record header.
-        let mut header = XLogRecord {
-            xl_tot_len: (XLOG_SIZE_OF_XLOG_RECORD + data_header.len() + data.len()) as u32,
-            xl_xid: 0,
-            xl_prev: prev_lsn.into(),
-            xl_info: info,
-            xl_rmid: rmid,
-            __bindgen_padding_0: [0; 2],
-            xl_crc: 0, // see below
-        };
-
-        // Compute the CRC checksum for the data, and the header up to the CRC field.
-        let mut crc = 0;
-        crc = crc32c_append(crc, &data_header);
-        crc = crc32c_append(crc, &data);
-        crc = crc32c_append(crc, &header.encode().unwrap()[0..XLOG_RECORD_CRC_OFFS]);
-        header.xl_crc = crc;
-
-        // Encode the final header and record.
-        let header = header.encode().unwrap();
-
-        [header, data_header, data].concat().into()
-    }
-
-    /// Injects page headers on 8KB page boundaries. Takes the current LSN position where the record
+    /// Inserts page headers on 8KB page boundaries. Takes the current LSN position where the record
     /// is to be appended.
-    fn encode_pages(record: Bytes, mut lsn: Lsn) -> Bytes {
+    fn insert_pages(record: Bytes, mut lsn: Lsn) -> Bytes {
         // Fast path: record fits in current page, and the page already has a header.
         if lsn.remaining_in_block() as usize >= record.len() && lsn.block_offset() > 0 {
             return record;
@@ -173,31 +189,71 @@ impl WalGenerator {
         }
         [record, Bytes::from(vec![0; padding])].concat().into()
     }
-
-    /// Generates a record with an arbitrary payload at the current LSN, then increments the LSN.
-    pub fn generate_record(&mut self, data: Bytes, rmid: u8, info: u8) -> Bytes {
-        let record = Self::encode_record(data, rmid, info, self.prev_lsn);
-        let record = Self::encode_pages(record, self.lsn);
-        let record = Self::pad_record(record, self.lsn);
-        self.prev_lsn = self.lsn;
-        self.lsn += record.len() as u64;
-        record
-    }
-
-    /// Generates a logical message at the current LSN. Can be used to construct arbitrary messages.
-    pub fn generate_logical_message(&mut self, prefix: &CStr, message: &[u8]) -> Bytes {
-        let data = Self::encode_logical_message(prefix, message);
-        self.generate_record(data, RM_LOGICALMSG_ID, XLOG_LOGICAL_MESSAGE)
-    }
 }
 
-/// Generate WAL records as an iterator.
-impl Iterator for WalGenerator {
+/// Generates WAL records as an iterator.
+impl<R: RecordGenerator> Iterator for WalGenerator<R> {
     type Item = (Lsn, Bytes);
 
     fn next(&mut self) -> Option<Self::Item> {
-        let lsn = self.lsn;
-        let record = self.generate_logical_message(Self::PREFIX, Self::MESSAGE);
-        Some((lsn, record))
+        let record = self.record_generator.next()?;
+        Some(self.append_record(record))
+    }
+}
+
+/// Generates logical message records (effectively noops) with a fixed message.
+pub struct LogicalMessageGenerator {
+    prefix: CString,
+    message: Vec<u8>,
+}
+
+impl LogicalMessageGenerator {
+    const DB_ID: u32 = 0; // hardcoded for now
+    const RM_ID: RmgrId = RM_LOGICALMSG_ID;
+    const INFO: u8 = XLOG_LOGICAL_MESSAGE;
+
+    /// Creates a new LogicalMessageGenerator.
+    pub fn new(prefix: &CStr, message: &[u8]) -> Self {
+        Self {
+            prefix: prefix.to_owned(),
+            message: message.to_owned(),
+        }
+    }
+
+    /// Encodes a logical message.
+    fn encode(prefix: &CStr, message: &[u8]) -> Bytes {
+        let prefix = prefix.to_bytes_with_nul();
+        let header = XlLogicalMessage {
+            db_id: Self::DB_ID,
+            transactional: 0,
+            prefix_size: prefix.len() as u64,
+            message_size: message.len() as u64,
+        };
+        [&header.encode(), prefix, message].concat().into()
+    }
+}
+
+impl Iterator for LogicalMessageGenerator {
+    type Item = Record;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        Some(Record {
+            rmid: Self::RM_ID,
+            info: Self::INFO,
+            data: Self::encode(&self.prefix, &self.message),
+        })
+    }
+}
+
+impl WalGenerator<LogicalMessageGenerator> {
+    /// Convenience method for appending a WAL record with an arbitrary logical message at the
+    /// current WAL LSN position. Returns the start LSN and resulting WAL bytes.
+    pub fn append_logical_message(&mut self, prefix: &CStr, message: &[u8]) -> (Lsn, Bytes) {
+        let record = Record {
+            rmid: LogicalMessageGenerator::RM_ID,
+            info: LogicalMessageGenerator::INFO,
+            data: LogicalMessageGenerator::encode(prefix, message),
+        };
+        self.append_record(record)
     }
 }

libs/postgres_ffi/src/xlog_utils.rs

@@ -12,9 +12,9 @@ use super::bindings::{
     CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID, TimestampTz,
     XLogLongPageHeaderData, XLogPageHeaderData, XLogRecPtr, XLogRecord, XLogSegNo, XLOG_PAGE_MAGIC,
 };
-use super::wal_generator::WalGenerator;
+use super::wal_generator::LogicalMessageGenerator;
 use super::PG_MAJORVERSION;
-use crate::pg_constants::{self, RM_LOGICALMSG_ID, XLOG_LOGICAL_MESSAGE};
+use crate::pg_constants;
 use crate::PG_TLI;
 use crate::{uint32, uint64, Oid};
 use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
@@ -493,12 +493,10 @@ pub fn encode_logical_message(prefix: &str, message: &str) -> Bytes {
     // This function can take untrusted input, so discard any NUL bytes in the prefix string.
     let prefix = CString::new(prefix.replace('\0', "")).expect("no NULs");
     let message = message.as_bytes();
-    WalGenerator::encode_record(
-        WalGenerator::encode_logical_message(&prefix, message),
-        RM_LOGICALMSG_ID,
-        XLOG_LOGICAL_MESSAGE,
-        Lsn(0),
-    )
+    LogicalMessageGenerator::new(&prefix, message)
+        .next()
+        .unwrap()
+        .encode(Lsn(0))
 }
 
 #[cfg(test)]

libs/remote_storage/src/error.rs

@@ -15,6 +15,9 @@ pub enum DownloadError {
     ///
     /// Concurrency control is not timed within timeout.
     Timeout,
+    /// Some integrity/consistency check failed during download. This is used during
+    /// timeline loads to cancel the load of a tenant if some timeline detects fatal corruption.
+    Fatal(String),
     /// The file was found in the remote storage, but the download failed.
     Other(anyhow::Error),
 }
@@ -29,6 +32,7 @@ impl std::fmt::Display for DownloadError {
             DownloadError::Unmodified => write!(f, "File was not modified"),
             DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
             DownloadError::Timeout => write!(f, "timeout"),
+            DownloadError::Fatal(why) => write!(f, "Fatal read error: {why}"),
             DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
         }
     }
@@ -41,7 +45,7 @@ impl DownloadError {
     pub fn is_permanent(&self) -> bool {
         use DownloadError::*;
         match self {
-            BadInput(_) | NotFound | Unmodified | Cancelled => true,
+            BadInput(_) | NotFound | Unmodified | Fatal(_) | Cancelled => true,
             Timeout | Other(_) => false,
         }
     }

libs/utils/scripts/restore_from_wal.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -euxo pipefail
 
@@ -6,9 +6,44 @@ PG_BIN=$1
 WAL_PATH=$2
 DATA_DIR=$3
 PORT=$4
+PG_VERSION=$5
 SYSID=$(od -A n -j 24 -N 8 -t d8 "$WAL_PATH"/000000010000000000000002* | cut -c 3-)
+
+# The way that initdb is invoked must match how the pageserver runs initdb.
+function initdb_with_args {
+    local cmd=(
+        "$PG_BIN"/initdb
+        -E utf8
+        -U cloud_admin
+        -D "$DATA_DIR"
+        --locale 'C.UTF-8'
+        --lc-collate 'C.UTF-8'
+        --lc-ctype 'C.UTF-8'
+        --lc-messages 'C.UTF-8'
+        --lc-monetary 'C.UTF-8'
+        --lc-numeric 'C.UTF-8'
+        --lc-time 'C.UTF-8'
+        --sysid="$SYSID"
+    )
+
+    case "$PG_VERSION" in
+        14)
+            # Postgres 14 and below didn't support --locale-provider
+            ;;
+        15 | 16)
+            cmd+=(--locale-provider 'libc')
+            ;;
+        *)
+            # Postgres 17 added the builtin provider
+            cmd+=(--locale-provider 'builtin')
+            ;;
+    esac
+
+    eval env -i LD_LIBRARY_PATH="$PG_BIN"/../lib "${cmd[*]}"
+}
+
 rm -fr "$DATA_DIR"
-env -i LD_LIBRARY_PATH="$PG_BIN"/../lib "$PG_BIN"/initdb -E utf8 -U cloud_admin -D "$DATA_DIR" --sysid="$SYSID"
+initdb_with_args
 echo "port=$PORT" >> "$DATA_DIR"/postgresql.conf
 echo "shared_preload_libraries='\$libdir/neon_rmgr.so'" >> "$DATA_DIR"/postgresql.conf
 REDO_POS=0x$("$PG_BIN"/pg_controldata -D "$DATA_DIR" | grep -F "REDO location"| cut -c 42-)

libs/utils/src/auth.rs

@@ -40,6 +40,11 @@ pub enum Scope {
     /// Allows access to storage controller APIs used by the scrubber, to interrogate the state
     /// of a tenant & post scrub results.
     Scrubber,
+
+    /// This scope is used for communication with other storage controller instances.
+    /// At the time of writing, this is only used for the step down request.
+    #[serde(rename = "controller_peer")]
+    ControllerPeer,
 }
 
 /// JWT payload. See docs/authentication.md for the format

libs/utils/src/crashsafe.rs

@@ -123,15 +123,27 @@ pub async fn fsync_async_opt(
     Ok(())
 }
 
-/// Like postgres' durable_rename, renames file issuing fsyncs do make it
-/// durable. After return, file and rename are guaranteed to be persisted.
+/// Like postgres' durable_rename, renames a file and issues fsyncs to make it durable. After
+/// returning, both the file and rename are guaranteed to be persisted. Both paths must be on the
+/// same file system.
 ///
-/// Unlike postgres, it only does fsyncs to 1) file to be renamed to make
-/// contents durable; 2) its directory entry to make rename durable 3) again to
-/// already renamed file, which is not required by standards but postgres does
-/// it, let's stick to that. Postgres additionally fsyncs newpath *before*
-/// rename if it exists to ensure that at least one of the files survives, but
-/// current callers don't need that.
+/// Unlike postgres, it only fsyncs 1) the file to make contents durable, and 2) the directory to
+/// make the rename durable. This sequence ensures the target file will never be incomplete.
+///
+/// Postgres also:
+///
+/// * Fsyncs the target file, if it exists, before the rename, to ensure either the new or existing
+///   file survives a crash. Current callers don't need this as it should already be fsynced if
+///   durability is needed.
+///
+/// * Fsyncs the file after the rename. This can be required with certain OSes or file systems (e.g.
+///   NFS), but not on Linux with most common file systems like ext4 (which we currently use).
+///
+/// An audit of 8 other databases found that none fsynced the file after a rename:
+/// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837180535>
+///
+/// eBPF probes confirmed that this is sufficient with ext4, XFS, and ZFS, but possibly not Btrfs:
+/// <https://github.com/neondatabase/neon/pull/9686#discussion_r1837926218>
 ///
 /// virtual_file.rs has similar code, but it doesn't use vfs.
 ///
@@ -149,9 +161,6 @@ pub async fn durable_rename(
     // Time to do the real deal.
     tokio::fs::rename(old_path.as_ref(), new_path.as_ref()).await?;
 
-    // Postgres'ish fsync of renamed file.
-    fsync_async_opt(new_path.as_ref(), do_fsync).await?;
-
     // Now fsync the parent
     let parent = match new_path.as_ref().parent() {
         Some(p) => p,

libs/utils/src/lsn.rs

@@ -138,6 +138,11 @@ impl Lsn {
         self.0.checked_sub(other).map(Lsn)
     }
 
+    /// Subtract a number, saturating at numeric bounds instead of overflowing.
+    pub fn saturating_sub<T: Into<u64>>(self, other: T) -> Lsn {
+        Lsn(self.0.saturating_sub(other.into()))
+    }
+
     /// Subtract a number, returning the difference as i128 to avoid overflow.
     pub fn widening_sub<T: Into<u64>>(self, other: T) -> i128 {
         let other: u64 = other.into();

pageserver/compaction/src/helpers.rs

@@ -35,6 +35,15 @@ pub fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
     !(a.end <= b.start || b.end <= a.start)
 }
 
+/// Whether a fully contains b, example as below
+/// ```plain
+/// |      a       |
+///       |  b  |
+/// ```
+pub fn fully_contains<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
+    a.start <= b.start && a.end >= b.end
+}
+
 pub fn union_to_keyspace<K: Ord>(a: &mut CompactionKeySpace<K>, b: CompactionKeySpace<K>) {
     let x = std::mem::take(a);
     let mut all_ranges_iter = [x.into_iter(), b.into_iter()]

pageserver/src/auth.rs

@@ -19,7 +19,8 @@ pub fn check_permission(claims: &Claims, tenant_id: Option<TenantId>) -> Result<
             | Scope::SafekeeperData
             | Scope::GenerationsApi
             | Scope::Infra
-            | Scope::Scrubber,
+            | Scope::Scrubber
+            | Scope::ControllerPeer,
             _,
         ) => Err(AuthError(
             format!(

pageserver/src/bin/pageserver.rs

@@ -154,13 +154,17 @@ fn main() -> anyhow::Result<()> {
             },
         };
 
-        let started = Instant::now();
-        syncfs(dirfd)?;
-        let elapsed = started.elapsed();
-        info!(
-            elapsed_ms = elapsed.as_millis(),
-            "made tenant directory contents durable"
-        );
+        if conf.no_sync {
+            info!("Skipping syncfs on startup");
+        } else {
+            let started = Instant::now();
+            syncfs(dirfd)?;
+            let elapsed = started.elapsed();
+            info!(
+                elapsed_ms = elapsed.as_millis(),
+                "made tenant directory contents durable"
+            );
+        }
     }
 
     // Initialize up failpoints support

pageserver/src/config.rs

@@ -69,6 +69,7 @@ pub struct PageServerConf {
     pub wal_redo_timeout: Duration,
 
     pub superuser: String,
+    pub locale: String,
 
     pub page_cache_size: usize,
     pub max_file_descriptors: usize,
@@ -178,6 +179,9 @@ pub struct PageServerConf {
 
     /// Direct IO settings
     pub virtual_file_io_mode: virtual_file::IoMode,
+
+    /// Optionally disable disk syncs (unsafe!)
+    pub no_sync: bool,
 }
 
 /// Token for authentication to safekeepers
@@ -298,6 +302,7 @@ impl PageServerConf {
             wait_lsn_timeout,
             wal_redo_timeout,
             superuser,
+            locale,
             page_cache_size,
             max_file_descriptors,
             pg_distrib_dir,
@@ -332,6 +337,7 @@ impl PageServerConf {
             concurrent_tenant_size_logical_size_queries,
             virtual_file_io_engine,
             tenant_config,
+            no_sync,
         } = config_toml;
 
         let mut conf = PageServerConf {
@@ -344,6 +350,7 @@ impl PageServerConf {
             wait_lsn_timeout,
             wal_redo_timeout,
             superuser,
+            locale,
             page_cache_size,
             max_file_descriptors,
             http_auth_type,
@@ -409,6 +416,7 @@ impl PageServerConf {
                 .map(crate::l0_flush::L0FlushConfig::from)
                 .unwrap_or_default(),
             virtual_file_io_mode: virtual_file_io_mode.unwrap_or(virtual_file::IoMode::preferred()),
+            no_sync: no_sync.unwrap_or(false),
         };
 
         // ------------------------------------------------------------

pageserver/src/http/routes.rs

@@ -2002,9 +2002,9 @@ async fn timeline_offload_handler(
                 "timeline has attached children".into(),
             ));
         }
-        if !timeline.can_offload() {
+        if let (false, reason) = timeline.can_offload() {
             return Err(ApiError::PreconditionFailed(
-                "Timeline::can_offload() returned false".into(),
+                format!("Timeline::can_offload() check failed: {}", reason) .into(),
             ));
         }
         offload_timeline(&tenant, &timeline)
@@ -2169,6 +2169,21 @@ async fn timeline_detach_ancestor_handler(
         let ctx = RequestContext::new(TaskKind::DetachAncestor, DownloadBehavior::Download);
         let ctx = &ctx;
 
+        // Flush the upload queues of all timelines before detaching ancestor. We do the same thing again
+        // during shutdown. This early upload ensures the pageserver does not need to upload too many
+        // things and creates downtime during timeline reloads.
+        for timeline in tenant.list_timelines() {
+            timeline
+                .remote_client
+                .wait_completion()
+                .await
+                .map_err(|e| {
+                    ApiError::PreconditionFailed(format!("cannot drain upload queue: {e}").into())
+                })?;
+        }
+
+        tracing::info!("all timeline upload queues are drained");
+
         let timeline = tenant.get_timeline(timeline_id, true)?;
 
         let progress = timeline

pageserver/src/pgdatadir_mapping.rs

@@ -45,7 +45,7 @@ use wal_decoder::serialized_batch::SerializedValueBatch;
 pub const MAX_AUX_FILE_DELTAS: usize = 1024;
 
 /// Max number of aux-file-related delta layers. The compaction will create a new image layer once this threshold is reached.
-pub const MAX_AUX_FILE_V2_DELTAS: usize = 64;
+pub const MAX_AUX_FILE_V2_DELTAS: usize = 16;
 
 #[derive(Debug)]
 pub enum LsnForTimestamp {

pageserver/src/tenant.rs

@@ -1433,6 +1433,12 @@ impl Tenant {
                     info!(%timeline_id, "index_part not found on remote");
                     continue;
                 }
+                Err(DownloadError::Fatal(why)) => {
+                    // If, while loading one remote timeline, we saw an indication that our generation
+                    // number is likely invalid, then we should not load the whole tenant.
+                    error!(%timeline_id, "Fatal error loading timeline: {why}");
+                    anyhow::bail!(why.to_string());
+                }
                 Err(e) => {
                     // Some (possibly ephemeral) error happened during index_part download.
                     // Pretend the timeline exists to not delete the timeline directory,
@@ -2493,7 +2499,8 @@ impl Tenant {
             timelines_to_compact_or_offload = timelines
                 .iter()
                 .filter_map(|(timeline_id, timeline)| {
-                    let (is_active, can_offload) = (timeline.is_active(), timeline.can_offload());
+                    let (is_active, (can_offload, _)) =
+                        (timeline.is_active(), timeline.can_offload());
                     let has_no_unoffloaded_children = {
                         !timelines
                             .iter()
@@ -4779,10 +4786,12 @@ async fn run_initdb(
 
     let _permit = INIT_DB_SEMAPHORE.acquire().await;
 
-    let initdb_command = tokio::process::Command::new(&initdb_bin_path)
+    let mut initdb_command = tokio::process::Command::new(&initdb_bin_path);
+    initdb_command
         .args(["--pgdata", initdb_target_dir.as_ref()])
         .args(["--username", &conf.superuser])
         .args(["--encoding", "utf8"])
+        .args(["--locale", &conf.locale])
         .arg("--no-instructions")
         .arg("--no-sync")
         .env_clear()
@@ -4792,15 +4801,27 @@ async fn run_initdb(
         // stdout invocation produces the same output every time, we don't need it
         .stdout(std::process::Stdio::null())
         // we would be interested in the stderr output, if there was any
-        .stderr(std::process::Stdio::piped())
-        .spawn()?;
+        .stderr(std::process::Stdio::piped());
+
+    // Before version 14, only the libc provide was available.
+    if pg_version > 14 {
+        // Version 17 brought with it a builtin locale provider which only provides
+        // C and C.UTF-8. While being safer for collation purposes since it is
+        // guaranteed to be consistent throughout a major release, it is also more
+        // performant.
+        let locale_provider = if pg_version >= 17 { "builtin" } else { "libc" };
+
+        initdb_command.args(["--locale-provider", locale_provider]);
+    }
+
+    let initdb_proc = initdb_command.spawn()?;
 
     // Ideally we'd select here with the cancellation token, but the problem is that
     // we can't safely terminate initdb: it launches processes of its own, and killing
     // initdb doesn't kill them. After we return from this function, we want the target
     // directory to be able to be cleaned up.
     // See https://github.com/neondatabase/neon/issues/6385
-    let initdb_output = initdb_command.wait_with_output().await?;
+    let initdb_output = initdb_proc.wait_with_output().await?;
     if !initdb_output.status.success() {
         return Err(InitdbError::Failed(
             initdb_output.status,
@@ -7742,13 +7763,13 @@ mod tests {
             (
                 get_key(3),
                 Lsn(0x20),
-                Value::WalRecord(NeonWalRecord::wal_clear()),
+                Value::WalRecord(NeonWalRecord::wal_clear("c")),
             ),
             (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
             (
                 get_key(4),
                 Lsn(0x20),
-                Value::WalRecord(NeonWalRecord::wal_init()),
+                Value::WalRecord(NeonWalRecord::wal_init("i")),
             ),
         ];
         let image1 = vec![(get_key(1), "0x10".into())];
@@ -7897,8 +7918,30 @@ mod tests {
 
     #[cfg(feature = "testing")]
     #[tokio::test]
-    async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
-        let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
+    async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
+        test_simple_bottom_most_compaction_deltas_helper(
+            "test_simple_bottom_most_compaction_deltas_1",
+            false,
+        )
+        .await
+    }
+
+    #[cfg(feature = "testing")]
+    #[tokio::test]
+    async fn test_simple_bottom_most_compaction_deltas_2() -> anyhow::Result<()> {
+        test_simple_bottom_most_compaction_deltas_helper(
+            "test_simple_bottom_most_compaction_deltas_2",
+            true,
+        )
+        .await
+    }
+
+    #[cfg(feature = "testing")]
+    async fn test_simple_bottom_most_compaction_deltas_helper(
+        test_name: &'static str,
+        use_delta_bottom_layer: bool,
+    ) -> anyhow::Result<()> {
+        let harness = TenantHarness::create(test_name).await?;
         let (tenant, ctx) = harness.load().await;
 
         fn get_key(id: u32) -> Key {
@@ -7929,6 +7972,16 @@ mod tests {
         let img_layer = (0..10)
             .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
             .collect_vec();
+        // or, delta layer at 0x10 if `use_delta_bottom_layer` is true
+        let delta4 = (0..10)
+            .map(|id| {
+                (
+                    get_key(id),
+                    Lsn(0x08),
+                    Value::WalRecord(NeonWalRecord::wal_init(format!("value {id}@0x10"))),
+                )
+            })
+            .collect_vec();
 
         let delta1 = vec![
             (
@@ -7982,21 +8035,61 @@ mod tests {
             ),
         ];
 
-        let tline = tenant
-            .create_test_timeline_with_layers(
-                TIMELINE_ID,
-                Lsn(0x10),
-                DEFAULT_PG_VERSION,
-                &ctx,
-                vec![
-                    DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
-                    DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
-                    DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
-                ], // delta layers
-                vec![(Lsn(0x10), img_layer)], // image layers
-                Lsn(0x50),
-            )
-            .await?;
+        let tline = if use_delta_bottom_layer {
+            tenant
+                .create_test_timeline_with_layers(
+                    TIMELINE_ID,
+                    Lsn(0x08),
+                    DEFAULT_PG_VERSION,
+                    &ctx,
+                    vec![
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x08)..Lsn(0x10),
+                            delta4,
+                        ),
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x20)..Lsn(0x48),
+                            delta1,
+                        ),
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x20)..Lsn(0x48),
+                            delta2,
+                        ),
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x48)..Lsn(0x50),
+                            delta3,
+                        ),
+                    ], // delta layers
+                    vec![], // image layers
+                    Lsn(0x50),
+                )
+                .await?
+        } else {
+            tenant
+                .create_test_timeline_with_layers(
+                    TIMELINE_ID,
+                    Lsn(0x10),
+                    DEFAULT_PG_VERSION,
+                    &ctx,
+                    vec![
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x10)..Lsn(0x48),
+                            delta1,
+                        ),
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x10)..Lsn(0x48),
+                            delta2,
+                        ),
+                        DeltaLayerTestDesc::new_with_inferred_key_range(
+                            Lsn(0x48)..Lsn(0x50),
+                            delta3,
+                        ),
+                    ], // delta layers
+                    vec![(Lsn(0x10), img_layer)], // image layers
+                    Lsn(0x50),
+                )
+                .await?
+        };
         {
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
@@ -8106,7 +8199,7 @@ mod tests {
             (
                 key,
                 Lsn(0x10),
-                Value::Image(Bytes::copy_from_slice(b"0x10")),
+                Value::WalRecord(NeonWalRecord::wal_init("0x10")),
             ),
             (
                 key,
@@ -8168,7 +8261,7 @@ mod tests {
                     Lsn(0x20),
                     KeyLogAtLsn(vec![(
                         Lsn(0x20),
-                        Value::Image(Bytes::copy_from_slice(b"0x10;0x20")),
+                        Value::Image(Bytes::from_static(b"0x10;0x20")),
                     )]),
                 ),
                 (
@@ -9150,7 +9243,7 @@ mod tests {
 
             let will_init = will_init_keys.contains(&i);
             if will_init {
-                delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init())));
+                delta_layer_spec.push((key, lsn, Value::WalRecord(NeonWalRecord::wal_init(""))));
 
                 expected_key_values.insert(key, "".to_string());
             } else {
@@ -9208,6 +9301,23 @@ mod tests {
         Ok(())
     }
 
+    fn sort_layer_key(k1: &PersistentLayerKey, k2: &PersistentLayerKey) -> std::cmp::Ordering {
+        (
+            k1.is_delta,
+            k1.key_range.start,
+            k1.key_range.end,
+            k1.lsn_range.start,
+            k1.lsn_range.end,
+        )
+            .cmp(&(
+                k2.is_delta,
+                k2.key_range.start,
+                k2.key_range.end,
+                k2.lsn_range.start,
+                k2.lsn_range.end,
+            ))
+    }
+
     async fn inspect_and_sort(
         tline: &Arc<Timeline>,
         filter: Option<std::ops::Range<Key>>,
@@ -9216,25 +9326,30 @@ mod tests {
         if let Some(filter) = filter {
             all_layers.retain(|layer| overlaps_with(&layer.key_range, &filter));
         }
-        all_layers.sort_by(|k1, k2| {
-            (
-                k1.is_delta,
-                k1.key_range.start,
-                k1.key_range.end,
-                k1.lsn_range.start,
-                k1.lsn_range.end,
-            )
-                .cmp(&(
-                    k2.is_delta,
-                    k2.key_range.start,
-                    k2.key_range.end,
-                    k2.lsn_range.start,
-                    k2.lsn_range.end,
-                ))
-        });
+        all_layers.sort_by(sort_layer_key);
         all_layers
     }
 
+    #[cfg(feature = "testing")]
+    fn check_layer_map_key_eq(
+        mut left: Vec<PersistentLayerKey>,
+        mut right: Vec<PersistentLayerKey>,
+    ) {
+        left.sort_by(sort_layer_key);
+        right.sort_by(sort_layer_key);
+        if left != right {
+            eprintln!("---LEFT---");
+            for left in left.iter() {
+                eprintln!("{}", left);
+            }
+            eprintln!("---RIGHT---");
+            for right in right.iter() {
+                eprintln!("{}", right);
+            }
+            assert_eq!(left, right);
+        }
+    }
+
     #[cfg(feature = "testing")]
     #[tokio::test]
     async fn test_simple_partial_bottom_most_compaction() -> anyhow::Result<()> {
@@ -9327,127 +9442,206 @@ mod tests {
 
         let cancel = CancellationToken::new();
 
-        // Do a partial compaction on key range 0..4, we should generate a image layer; no other layers
-        // can be removed because they might be used for other key ranges.
+        // Do a partial compaction on key range 0..2
         tline
-            .partial_compact_with_gc(Some(get_key(0)..get_key(4)), &cancel, EnumSet::new(), &ctx)
+            .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
-        assert_eq!(
+        check_layer_map_key_eq(
             all_layers,
             vec![
+                // newly-generated image layer for the partial compaction range 0-2
                 PersistentLayerKey {
-                    key_range: get_key(0)..get_key(4),
+                    key_range: get_key(0)..get_key(2),
                     lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
+                    is_delta: false,
                 },
                 PersistentLayerKey {
                     key_range: get_key(0)..get_key(10),
                     lsn_range: Lsn(0x10)..Lsn(0x11),
-                    is_delta: false
+                    is_delta: false,
                 },
+                // delta1 is split and the second part is rewritten
                 PersistentLayerKey {
-                    key_range: get_key(1)..get_key(4),
+                    key_range: get_key(2)..get_key(4),
                     lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
+                    is_delta: true,
                 },
                 PersistentLayerKey {
                     key_range: get_key(5)..get_key(7),
                     lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
+                    is_delta: true,
                 },
                 PersistentLayerKey {
                     key_range: get_key(8)..get_key(10),
                     lsn_range: Lsn(0x48)..Lsn(0x50),
-                    is_delta: true
-                }
-            ]
+                    is_delta: true,
+                },
+            ],
         );
 
-        // Do a partial compaction on key range 4..10
+        // Do a partial compaction on key range 2..4
         tline
-            .partial_compact_with_gc(Some(get_key(4)..get_key(10)), &cancel, EnumSet::new(), &ctx)
+            .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
-        assert_eq!(
+        check_layer_map_key_eq(
             all_layers,
             vec![
                 PersistentLayerKey {
-                    key_range: get_key(0)..get_key(4),
+                    key_range: get_key(0)..get_key(2),
                     lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
+                    is_delta: false,
                 },
                 PersistentLayerKey {
-                    // if (in the future) GC kicks in, this layer will be removed
                     key_range: get_key(0)..get_key(10),
                     lsn_range: Lsn(0x10)..Lsn(0x11),
-                    is_delta: false
+                    is_delta: false,
                 },
+                // image layer generated for the compaction range 2-4
                 PersistentLayerKey {
-                    key_range: get_key(4)..get_key(10),
+                    key_range: get_key(2)..get_key(4),
                     lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
+                    is_delta: false,
                 },
+                // we have key2/key3 above the retain_lsn, so we still need this delta layer
                 PersistentLayerKey {
-                    key_range: get_key(1)..get_key(4),
+                    key_range: get_key(2)..get_key(4),
                     lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
+                    is_delta: true,
                 },
                 PersistentLayerKey {
                     key_range: get_key(5)..get_key(7),
                     lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
+                    is_delta: true,
                 },
                 PersistentLayerKey {
                     key_range: get_key(8)..get_key(10),
                     lsn_range: Lsn(0x48)..Lsn(0x50),
-                    is_delta: true
-                }
-            ]
+                    is_delta: true,
+                },
+            ],
+        );
+
+        // Do a partial compaction on key range 4..9
+        tline
+            .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
+            .await
+            .unwrap();
+        let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
+        check_layer_map_key_eq(
+            all_layers,
+            vec![
+                PersistentLayerKey {
+                    key_range: get_key(0)..get_key(2),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(0)..get_key(10),
+                    lsn_range: Lsn(0x10)..Lsn(0x11),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(2)..get_key(4),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(2)..get_key(4),
+                    lsn_range: Lsn(0x20)..Lsn(0x48),
+                    is_delta: true,
+                },
+                // image layer generated for this compaction range
+                PersistentLayerKey {
+                    key_range: get_key(4)..get_key(9),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(8)..get_key(10),
+                    lsn_range: Lsn(0x48)..Lsn(0x50),
+                    is_delta: true,
+                },
+            ],
+        );
+
+        // Do a partial compaction on key range 9..10
+        tline
+            .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
+            .await
+            .unwrap();
+        let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
+        check_layer_map_key_eq(
+            all_layers,
+            vec![
+                PersistentLayerKey {
+                    key_range: get_key(0)..get_key(2),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(0)..get_key(10),
+                    lsn_range: Lsn(0x10)..Lsn(0x11),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(2)..get_key(4),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(2)..get_key(4),
+                    lsn_range: Lsn(0x20)..Lsn(0x48),
+                    is_delta: true,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(4)..get_key(9),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                // image layer generated for the compaction range
+                PersistentLayerKey {
+                    key_range: get_key(9)..get_key(10),
+                    lsn_range: Lsn(0x20)..Lsn(0x21),
+                    is_delta: false,
+                },
+                PersistentLayerKey {
+                    key_range: get_key(8)..get_key(10),
+                    lsn_range: Lsn(0x48)..Lsn(0x50),
+                    is_delta: true,
+                },
+            ],
         );
 
         // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
         tline
-            .partial_compact_with_gc(Some(get_key(0)..get_key(10)), &cancel, EnumSet::new(), &ctx)
+            .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
-        assert_eq!(
+        check_layer_map_key_eq(
             all_layers,
             vec![
-                PersistentLayerKey {
-                    key_range: get_key(0)..get_key(4),
-                    lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
-                },
+                // aha, we removed all unnecessary image/delta layers and got a very clean layer map!
                 PersistentLayerKey {
                     key_range: get_key(0)..get_key(10),
                     lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
+                    is_delta: false,
                 },
                 PersistentLayerKey {
-                    key_range: get_key(4)..get_key(10),
-                    lsn_range: Lsn(0x20)..Lsn(0x21),
-                    is_delta: false
-                },
-                PersistentLayerKey {
-                    key_range: get_key(1)..get_key(4),
+                    key_range: get_key(2)..get_key(4),
                     lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
-                },
-                PersistentLayerKey {
-                    key_range: get_key(5)..get_key(7),
-                    lsn_range: Lsn(0x20)..Lsn(0x48),
-                    is_delta: true
+                    is_delta: true,
                 },
                 PersistentLayerKey {
                     key_range: get_key(8)..get_key(10),
                     lsn_range: Lsn(0x48)..Lsn(0x50),
-                    is_delta: true
-                }
-            ]
+                    is_delta: true,
+                },
+            ],
         );
 
         Ok(())

pageserver/src/tenant/mgr.rs

@@ -1959,7 +1959,7 @@ impl TenantManager {
             attempt.before_reset_tenant();
 
             let (_guard, progress) = utils::completion::channel();
-            match tenant.shutdown(progress, ShutdownMode::Hard).await {
+            match tenant.shutdown(progress, ShutdownMode::Flush).await {
                 Ok(()) => {
                     slot_guard.drop_old_value().expect("it was just shutdown");
                 }

pageserver/src/tenant/remote_timeline_client.rs

@@ -574,12 +574,18 @@ impl RemoteTimelineClient {
 
             if latest_index_generation > index_generation {
                 // Unexpected!  Why are we loading such an old index if a more recent one exists?
-                tracing::warn!(
+                // We will refuse to proceed, as there is no reasonable scenario where this should happen, but
+                // there _is_ a clear bug/corruption scenario where it would happen (controller sets the generation
+                // backwards).
+                tracing::error!(
                     ?index_generation,
                     ?latest_index_generation,
                     ?latest_index_mtime,
                     "Found a newer index while loading an old one"
                 );
+                return Err(DownloadError::Fatal(
+                    "Index age exceeds threshold and a newer index exists".into(),
+                ));
             }
         }
 
@@ -2201,6 +2207,18 @@ impl RemoteTimelineClient {
         inner.initialized_mut()?;
         Ok(UploadQueueAccessor { inner })
     }
+
+    pub(crate) fn no_pending_work(&self) -> bool {
+        let inner = self.upload_queue.lock().unwrap();
+        match &*inner {
+            UploadQueue::Uninitialized
+            | UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => true,
+            UploadQueue::Stopped(UploadQueueStopped::Deletable(x)) => {
+                x.upload_queue_for_deletion.no_pending_work()
+            }
+            UploadQueue::Initialized(x) => x.no_pending_work(),
+        }
+    }
 }
 
 pub(crate) struct UploadQueueAccessor<'a> {

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -653,6 +653,10 @@ impl DeltaLayerWriter {
         })
     }
 
+    pub fn is_empty(&self) -> bool {
+        self.inner.as_ref().unwrap().num_keys == 0
+    }
+
     ///
     /// Append a key-value pair to the file.
     ///

pageserver/src/tenant/storage_layer/filter_iterator.rs

@@ -1,4 +1,4 @@
-use std::ops::Range;
+use std::{ops::Range, sync::Arc};
 
 use anyhow::bail;
 use pageserver_api::{
@@ -9,7 +9,10 @@ use utils::lsn::Lsn;
 
 use pageserver_api::value::Value;
 
-use super::merge_iterator::MergeIterator;
+use super::{
+    merge_iterator::{MergeIterator, MergeIteratorItem},
+    PersistentLayerKey,
+};
 
 /// A filter iterator over merge iterators (and can be easily extended to other types of iterators).
 ///
@@ -48,10 +51,10 @@ impl<'a> FilterIterator<'a> {
         })
     }
 
-    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
-        while let Some(item) = self.inner.next().await? {
+    async fn next_inner<R: MergeIteratorItem>(&mut self) -> anyhow::Result<Option<R>> {
+        while let Some(item) = self.inner.next_inner::<R>().await? {
             while self.current_filter_idx < self.retain_key_filters.len()
-                && item.0 >= self.retain_key_filters[self.current_filter_idx].end
+                && item.key_lsn_value().0 >= self.retain_key_filters[self.current_filter_idx].end
             {
                 // [filter region]    [filter region]     [filter region]
                 //                                     ^ item
@@ -68,7 +71,7 @@ impl<'a> FilterIterator<'a> {
                 //                                                 ^ current filter (nothing)
                 return Ok(None);
             }
-            if self.retain_key_filters[self.current_filter_idx].contains(&item.0) {
+            if self.retain_key_filters[self.current_filter_idx].contains(&item.key_lsn_value().0) {
                 // [filter region]    [filter region]     [filter region]
                 //                                              ^ item
                 //                                        ^ current filter
@@ -81,6 +84,16 @@ impl<'a> FilterIterator<'a> {
         }
         Ok(None)
     }
+
+    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
+        self.next_inner().await
+    }
+
+    pub async fn next_with_trace(
+        &mut self,
+    ) -> anyhow::Result<Option<((Key, Lsn, Value), Arc<PersistentLayerKey>)>> {
+        self.next_inner().await
+    }
 }
 
 #[cfg(test)]

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -67,6 +67,8 @@ pub struct InMemoryLayer {
     /// The above fields never change, except for `end_lsn`, which is only set once.
     /// All other changing parts are in `inner`, and protected by a mutex.
     inner: RwLock<InMemoryLayerInner>,
+
+    estimated_in_mem_size: AtomicU64,
 }
 
 impl std::fmt::Debug for InMemoryLayer {
@@ -543,6 +545,10 @@ impl InMemoryLayer {
         Ok(inner.file.len())
     }
 
+    pub fn estimated_in_mem_size(&self) -> u64 {
+        self.estimated_in_mem_size.load(AtomicOrdering::Relaxed)
+    }
+
     /// Create a new, empty, in-memory layer
     pub async fn create(
         conf: &'static PageServerConf,
@@ -572,6 +578,7 @@ impl InMemoryLayer {
                 file,
                 resource_units: GlobalResourceUnits::new(),
             }),
+            estimated_in_mem_size: AtomicU64::new(0),
         })
     }
 
@@ -642,6 +649,12 @@ impl InMemoryLayer {
                 // because this case is unexpected, and we would like tests to fail if this happens.
                 warn!("Key {} at {} written twice at same LSN", key, lsn);
             }
+            self.estimated_in_mem_size.fetch_add(
+                (std::mem::size_of::<CompactKey>()
+                    + std::mem::size_of::<Lsn>()
+                    + std::mem::size_of::<IndexEntry>()) as u64,
+                AtomicOrdering::Relaxed,
+            );
         }
 
         inner.resource_units.maybe_publish_size(new_size);

pageserver/src/tenant/storage_layer/merge_iterator.rs

@@ -1,6 +1,7 @@
 use std::{
     cmp::Ordering,
     collections::{binary_heap, BinaryHeap},
+    sync::Arc,
 };
 
 use anyhow::bail;
@@ -13,10 +14,11 @@ use pageserver_api::value::Value;
 use super::{
     delta_layer::{DeltaLayerInner, DeltaLayerIterator},
     image_layer::{ImageLayerInner, ImageLayerIterator},
+    PersistentLayerDesc, PersistentLayerKey,
 };
 
 #[derive(Clone, Copy)]
-enum LayerRef<'a> {
+pub(crate) enum LayerRef<'a> {
     Image(&'a ImageLayerInner),
     Delta(&'a DeltaLayerInner),
 }
@@ -62,18 +64,20 @@ impl LayerIterRef<'_> {
 /// 1. Unified iterator for image and delta layers.
 /// 2. `Ord` for use in [`MergeIterator::heap`] (for the k-merge).
 /// 3. Lazy creation of the real delta/image iterator.
-enum IteratorWrapper<'a> {
+pub(crate) enum IteratorWrapper<'a> {
     NotLoaded {
         ctx: &'a RequestContext,
         first_key_lower_bound: (Key, Lsn),
         layer: LayerRef<'a>,
+        source_desc: Arc<PersistentLayerKey>,
     },
     Loaded {
         iter: PeekableLayerIterRef<'a>,
+        source_desc: Arc<PersistentLayerKey>,
     },
 }
 
-struct PeekableLayerIterRef<'a> {
+pub(crate) struct PeekableLayerIterRef<'a> {
     iter: LayerIterRef<'a>,
     peeked: Option<(Key, Lsn, Value)>, // None == end
 }
@@ -151,6 +155,12 @@ impl<'a> IteratorWrapper<'a> {
             layer: LayerRef::Image(image_layer),
             first_key_lower_bound: (image_layer.key_range().start, image_layer.lsn()),
             ctx,
+            source_desc: PersistentLayerKey {
+                key_range: image_layer.key_range().clone(),
+                lsn_range: PersistentLayerDesc::image_layer_lsn_range(image_layer.lsn()),
+                is_delta: false,
+            }
+            .into(),
         }
     }
 
@@ -162,12 +172,18 @@ impl<'a> IteratorWrapper<'a> {
             layer: LayerRef::Delta(delta_layer),
             first_key_lower_bound: (delta_layer.key_range().start, delta_layer.lsn_range().start),
             ctx,
+            source_desc: PersistentLayerKey {
+                key_range: delta_layer.key_range().clone(),
+                lsn_range: delta_layer.lsn_range().clone(),
+                is_delta: true,
+            }
+            .into(),
         }
     }
 
     fn peek_next_key_lsn_value(&self) -> Option<(&Key, Lsn, Option<&Value>)> {
         match self {
-            Self::Loaded { iter } => iter
+            Self::Loaded { iter, .. } => iter
                 .peek()
                 .as_ref()
                 .map(|(key, lsn, val)| (key, *lsn, Some(val))),
@@ -191,6 +207,7 @@ impl<'a> IteratorWrapper<'a> {
             ctx,
             first_key_lower_bound,
             layer,
+            source_desc,
         } = self
         else {
             unreachable!()
@@ -206,7 +223,10 @@ impl<'a> IteratorWrapper<'a> {
                 );
             }
         }
-        *self = Self::Loaded { iter };
+        *self = Self::Loaded {
+            iter,
+            source_desc: source_desc.clone(),
+        };
         Ok(())
     }
 
@@ -220,11 +240,19 @@ impl<'a> IteratorWrapper<'a> {
     /// The public interfaces to use are [`crate::tenant::storage_layer::delta_layer::DeltaLayerIterator`] and
     /// [`crate::tenant::storage_layer::image_layer::ImageLayerIterator`].
     async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
-        let Self::Loaded { iter } = self else {
+        let Self::Loaded { iter, .. } = self else {
             panic!("must load the iterator before using")
         };
         iter.next().await
     }
+
+    /// Get the persistent layer key corresponding to this iterator
+    fn trace_source(&self) -> Arc<PersistentLayerKey> {
+        match self {
+            Self::Loaded { source_desc, .. } => source_desc.clone(),
+            Self::NotLoaded { source_desc, .. } => source_desc.clone(),
+        }
+    }
 }
 
 /// A merge iterator over delta/image layer iterators.
@@ -242,6 +270,32 @@ pub struct MergeIterator<'a> {
     heap: BinaryHeap<IteratorWrapper<'a>>,
 }
 
+pub(crate) trait MergeIteratorItem {
+    fn new(item: (Key, Lsn, Value), iterator: &IteratorWrapper<'_>) -> Self;
+
+    fn key_lsn_value(&self) -> &(Key, Lsn, Value);
+}
+
+impl MergeIteratorItem for (Key, Lsn, Value) {
+    fn new(item: (Key, Lsn, Value), _: &IteratorWrapper<'_>) -> Self {
+        item
+    }
+
+    fn key_lsn_value(&self) -> &(Key, Lsn, Value) {
+        self
+    }
+}
+
+impl MergeIteratorItem for ((Key, Lsn, Value), Arc<PersistentLayerKey>) {
+    fn new(item: (Key, Lsn, Value), iter: &IteratorWrapper<'_>) -> Self {
+        (item, iter.trace_source().clone())
+    }
+
+    fn key_lsn_value(&self) -> &(Key, Lsn, Value) {
+        &self.0
+    }
+}
+
 impl<'a> MergeIterator<'a> {
     pub fn create(
         deltas: &[&'a DeltaLayerInner],
@@ -260,7 +314,7 @@ impl<'a> MergeIterator<'a> {
         }
     }
 
-    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
+    pub(crate) async fn next_inner<R: MergeIteratorItem>(&mut self) -> anyhow::Result<Option<R>> {
         while let Some(mut iter) = self.heap.peek_mut() {
             if !iter.is_loaded() {
                 // Once we load the iterator, we can know the real first key-value pair in the iterator.
@@ -275,10 +329,22 @@ impl<'a> MergeIterator<'a> {
                 binary_heap::PeekMut::pop(iter);
                 continue;
             };
-            return Ok(Some(item));
+            return Ok(Some(R::new(item, &iter)));
         }
         Ok(None)
     }
+
+    /// Get the next key-value pair from the iterator.
+    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
+        self.next_inner().await
+    }
+
+    /// Get the next key-value pair from the iterator, and trace where the key comes from.
+    pub async fn next_with_trace(
+        &mut self,
+    ) -> anyhow::Result<Option<((Key, Lsn, Value), Arc<PersistentLayerKey>)>> {
+        self.next_inner().await
+    }
 }
 
 #[cfg(test)]
@@ -496,7 +562,7 @@ mod tests {
             (
                 get_key(0),
                 Lsn(0x10),
-                Value::WalRecord(NeonWalRecord::wal_init()),
+                Value::WalRecord(NeonWalRecord::wal_init("")),
             ),
             (
                 get_key(0),
@@ -506,7 +572,7 @@ mod tests {
             (
                 get_key(5),
                 Lsn(0x10),
-                Value::WalRecord(NeonWalRecord::wal_init()),
+                Value::WalRecord(NeonWalRecord::wal_init("")),
             ),
             (
                 get_key(5),

pageserver/src/tenant/timeline.rs

@@ -23,6 +23,7 @@ use handle::ShardTimelineId;
 use offload::OffloadError;
 use once_cell::sync::Lazy;
 use pageserver_api::{
+    config::tenant_conf_defaults::DEFAULT_COMPACTION_THRESHOLD,
     key::{
         KEY_SIZE, METADATA_KEY_BEGIN_PREFIX, METADATA_KEY_END_PREFIX, NON_INHERITED_RANGE,
         NON_INHERITED_SPARSE_RANGE,
@@ -852,6 +853,10 @@ pub(crate) enum ShutdownMode {
     /// While we are flushing, we continue to accept read I/O for LSNs ingested before
     /// the call to [`Timeline::shutdown`].
     FreezeAndFlush,
+    /// Only flush the layers to the remote storage without freezing any open layers. This is the
+    /// mode used by ancestor detach and any other operations that reloads a tenant but not increasing
+    /// the generation number.
+    Flush,
     /// Shut down immediately, without waiting for any open layers to flush.
     Hard,
 }
@@ -1565,12 +1570,16 @@ impl Timeline {
     ///
     /// This is neccessary but not sufficient for offloading of the timeline as it might have
     /// child timelines that are not offloaded yet.
-    pub(crate) fn can_offload(&self) -> bool {
+    pub(crate) fn can_offload(&self) -> (bool, &'static str) {
         if self.remote_client.is_archived() != Some(true) {
-            return false;
+            return (false, "the timeline is not archived");
+        }
+        if !self.remote_client.no_pending_work() {
+            // if the remote client is still processing some work, we can't offload
+            return (false, "the upload queue is not drained yet");
         }
 
-        true
+        (true, "ok")
     }
 
     /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
@@ -1678,11 +1687,6 @@ impl Timeline {
     pub(crate) async fn shutdown(&self, mode: ShutdownMode) {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
-        let try_freeze_and_flush = match mode {
-            ShutdownMode::FreezeAndFlush => true,
-            ShutdownMode::Hard => false,
-        };
-
         // Regardless of whether we're going to try_freeze_and_flush
         // or not, stop ingesting any more data. Walreceiver only provides
         // cancellation but no "wait until gone", because it uses the Timeline::gate.
@@ -1704,7 +1708,7 @@ impl Timeline {
         // ... and inform any waiters for newer LSNs that there won't be any.
         self.last_record_lsn.shutdown();
 
-        if try_freeze_and_flush {
+        if let ShutdownMode::FreezeAndFlush = mode {
             if let Some((open, frozen)) = self
                 .layers
                 .read()
@@ -1746,6 +1750,20 @@ impl Timeline {
                     warn!("failed to freeze and flush: {e:#}");
                 }
             }
+
+            // `self.remote_client.shutdown().await` above should have already flushed everything from the queue, but
+            // we also do a final check here to ensure that the queue is empty.
+            if !self.remote_client.no_pending_work() {
+                warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
+            }
+        }
+
+        if let ShutdownMode::Flush = mode {
+            // drain the upload queue
+            self.remote_client.shutdown().await;
+            if !self.remote_client.no_pending_work() {
+                warn!("still have pending work in remote upload queue, but continuing shutting down anyways");
+            }
         }
 
         // Signal any subscribers to our cancellation token to drop out
@@ -3488,18 +3506,37 @@ impl Timeline {
 
                 let timer = self.metrics.flush_time_histo.start_timer();
 
+                let num_frozen_layers;
+                let frozen_layer_total_size;
                 let layer_to_flush = {
                     let guard = self.layers.read().await;
                     let Ok(lm) = guard.layer_map() else {
                         info!("dropping out of flush loop for timeline shutdown");
                         return;
                     };
+                    num_frozen_layers = lm.frozen_layers.len();
+                    frozen_layer_total_size = lm
+                        .frozen_layers
+                        .iter()
+                        .map(|l| l.estimated_in_mem_size())
+                        .sum::<u64>();
                     lm.frozen_layers.front().cloned()
                     // drop 'layers' lock to allow concurrent reads and writes
                 };
                 let Some(layer_to_flush) = layer_to_flush else {
                     break Ok(());
                 };
+                if num_frozen_layers
+                    > std::cmp::max(
+                        self.get_compaction_threshold(),
+                        DEFAULT_COMPACTION_THRESHOLD,
+                    )
+                    && frozen_layer_total_size >= /* 128 MB */ 128000000
+                {
+                    tracing::warn!(
+                        "too many frozen layers: {num_frozen_layers} layers with estimated in-mem size of {frozen_layer_total_size} bytes",
+                    );
+                }
                 match self.flush_frozen_layer(layer_to_flush, ctx).await {
                     Ok(this_layer_to_lsn) => {
                         flushed_to_lsn = std::cmp::max(flushed_to_lsn, this_layer_to_lsn);
@@ -4090,6 +4127,7 @@ impl Timeline {
     ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
         // Metadata keys image layer creation.
         let mut reconstruct_state = ValuesReconstructState::default();
+        let begin = Instant::now();
         let data = self
             .get_vectored_impl(partition.clone(), lsn, &mut reconstruct_state, ctx)
             .await?;
@@ -4106,14 +4144,11 @@ impl Timeline {
             (new_data, total_kb_retrieved / 1024, total_keys_retrieved)
         };
         let delta_files_accessed = reconstruct_state.get_delta_layers_visited();
+        let elapsed = begin.elapsed();
 
         let trigger_generation = delta_files_accessed as usize >= MAX_AUX_FILE_V2_DELTAS;
-        debug!(
-            trigger_generation,
-            delta_files_accessed,
-            total_kb_retrieved,
-            total_keys_retrieved,
-            "generate metadata images"
+        info!(
+            "metadata key compaction: trigger_generation={trigger_generation}, delta_files_accessed={delta_files_accessed}, total_kb_retrieved={total_kb_retrieved}, total_keys_retrieved={total_keys_retrieved}, read_time={}s", elapsed.as_secs_f64()
         );
 
         if !trigger_generation && mode == ImageLayerCreationMode::Try {

pageserver/src/tenant/timeline/compaction.rs

@@ -4,7 +4,7 @@
 //!
 //! The old legacy algorithm is implemented directly in `timeline.rs`.
 
-use std::collections::{BinaryHeap, HashSet};
+use std::collections::{BinaryHeap, HashMap, HashSet};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
 
@@ -56,7 +56,7 @@ use pageserver_api::value::Value;
 
 use utils::lsn::Lsn;
 
-use pageserver_compaction::helpers::overlaps_with;
+use pageserver_compaction::helpers::{fully_contains, overlaps_with};
 use pageserver_compaction::interface::*;
 
 use super::CompactionError;
@@ -64,6 +64,23 @@ use super::CompactionError;
 /// Maximum number of deltas before generating an image layer in bottom-most compaction.
 const COMPACTION_DELTA_THRESHOLD: usize = 5;
 
+pub struct GcCompactionJobDescription {
+    /// All layers to read in the compaction job
+    selected_layers: Vec<Layer>,
+    /// GC cutoff of the job
+    gc_cutoff: Lsn,
+    /// LSNs to retain for the job
+    retain_lsns_below_horizon: Vec<Lsn>,
+    /// Maximum layer LSN processed in this compaction
+    max_layer_lsn: Lsn,
+    /// Only compact layers overlapping with this range
+    compaction_key_range: Range<Key>,
+    /// When partial compaction is enabled, these layers need to be rewritten to ensure no overlap.
+    /// This field is here solely for debugging. The field will not be read once the compaction
+    /// description is generated.
+    rewrite_layers: Vec<Arc<PersistentLayerDesc>>,
+}
+
 /// The result of bottom-most compaction for a single key at each LSN.
 #[derive(Debug)]
 #[cfg_attr(test, derive(PartialEq))]
@@ -1722,7 +1739,8 @@ impl Timeline {
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        self.partial_compact_with_gc(None, cancel, flags, ctx).await
+        self.partial_compact_with_gc(Key::MIN..Key::MAX, cancel, flags, ctx)
+            .await
     }
 
     /// An experimental compaction building block that combines compaction with garbage collection.
@@ -1732,12 +1750,15 @@ impl Timeline {
     /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
     /// and create delta layers with all deltas >= gc horizon.
     ///
-    /// If `key_range`, it will only compact the keys within the range, aka partial compaction. This functionality
-    /// is not complete yet, and if it is set, only image layers will be generated.
-    ///
+    /// If `key_range` is provided, it will only compact the keys within the range, aka partial compaction.
+    /// Partial compaction will read and process all layers overlapping with the key range, even if it might
+    /// contain extra keys. After the gc-compaction phase completes, delta layers that are not fully contained
+    /// within the key range will be rewritten to ensure they do not overlap with the delta layers. Providing
+    /// Key::MIN..Key..MAX to the function indicates a full compaction, though technically, `Key::MAX` is not
+    /// part of the range.
     pub(crate) async fn partial_compact_with_gc(
         self: &Arc<Self>,
-        compaction_key_range: Option<Range<Key>>,
+        compaction_key_range: Range<Key>,
         cancel: &CancellationToken,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
@@ -1762,9 +1783,8 @@ impl Timeline {
         .await?;
 
         let dry_run = flags.contains(CompactFlags::DryRun);
-        let partial_compaction = compaction_key_range.is_some();
 
-        if let Some(ref compaction_key_range) = compaction_key_range {
+        if compaction_key_range == (Key::MIN..Key::MAX) {
             info!("running enhanced gc bottom-most compaction, dry_run={dry_run}, compaction_key_range={}..{}", compaction_key_range.start, compaction_key_range.end);
         } else {
             info!("running enhanced gc bottom-most compaction, dry_run={dry_run}");
@@ -1780,7 +1800,7 @@ impl Timeline {
         // The layer selection has the following properties:
         // 1. If a layer is in the selection, all layers below it are in the selection.
         // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
-        let (layer_selection, gc_cutoff, retain_lsns_below_horizon) = if !partial_compaction {
+        let job_desc = {
             let guard = self.layers.read().await;
             let layers = guard.layer_map()?;
             let gc_info = self.gc_info.read().unwrap();
@@ -1810,9 +1830,21 @@ impl Timeline {
             };
             // Then, pick all the layers that are below the max_layer_lsn. This is to ensure we can pick all single-key
             // layers to compact.
+            let mut rewrite_layers = Vec::new();
             for desc in layers.iter_historic_layers() {
-                if desc.get_lsn_range().end <= max_layer_lsn {
+                if desc.get_lsn_range().end <= max_layer_lsn
+                    && overlaps_with(&desc.get_key_range(), &compaction_key_range)
+                {
+                    // If the layer overlaps with the compaction key range, we need to read it to obtain all keys within the range,
+                    // even if it might contain extra keys
                     selected_layers.push(guard.get_from_desc(&desc));
+                    // If the layer is not fully contained within the key range, we need to rewrite it if it's a delta layer (it's fine
+                    // to overlap image layers)
+                    if desc.is_delta()
+                        && !fully_contains(&compaction_key_range, &desc.get_key_range())
+                    {
+                        rewrite_layers.push(desc);
+                    }
                 }
             }
             if selected_layers.is_empty() {
@@ -1820,82 +1852,59 @@ impl Timeline {
                 return Ok(());
             }
             retain_lsns_below_horizon.sort();
-            (selected_layers, gc_cutoff, retain_lsns_below_horizon)
-        } else {
-            // In case of partial compaction, we currently only support generating image layers, and therefore,
-            // we pick all layers that are below the lowest retain_lsn and does not intersect with any of the layers.
-            let guard = self.layers.read().await;
-            let layers = guard.layer_map()?;
-            let gc_info = self.gc_info.read().unwrap();
-            let mut min_lsn = gc_info.cutoffs.select_min();
-            for (lsn, _, _) in &gc_info.retain_lsns {
-                if lsn < &min_lsn {
-                    min_lsn = *lsn;
-                }
+            GcCompactionJobDescription {
+                selected_layers,
+                gc_cutoff,
+                retain_lsns_below_horizon,
+                max_layer_lsn,
+                compaction_key_range,
+                rewrite_layers,
             }
-            for lsn in gc_info.leases.keys() {
-                if lsn < &min_lsn {
-                    min_lsn = *lsn;
-                }
-            }
-            let mut selected_layers = Vec::new();
-            drop(gc_info);
-            // |-------| |-------| |-------|
-            // | Delta | | Delta | | Delta | -- min_lsn could be intersecting with the layers
-            // |-------| |-------| |-------| <- we want to pick all the layers below min_lsn, so that
-            // | Delta | | Delta | | Delta |    ...we can remove them after compaction
-            // |-------| |-------| |-------|
-            // Pick all the layers intersect or below the min_lsn, get the largest LSN in the selected layers.
-            let Some(compaction_key_range) = compaction_key_range.as_ref() else {
-                unreachable!()
-            };
-            for desc in layers.iter_historic_layers() {
-                if desc.get_lsn_range().end <= min_lsn
-                    && overlaps_with(&desc.key_range, compaction_key_range)
-                {
-                    selected_layers.push(guard.get_from_desc(&desc));
-                }
-            }
-            if selected_layers.is_empty() {
-                info!("no layers to compact with gc");
-                return Ok(());
-            }
-            (selected_layers, min_lsn, Vec::new())
         };
         let lowest_retain_lsn = if self.ancestor_timeline.is_some() {
-            if partial_compaction {
-                warn!("partial compaction cannot run on child branches (for now)");
-                return Ok(());
-            }
             Lsn(self.ancestor_lsn.0 + 1)
         } else {
-            let res = retain_lsns_below_horizon
+            let res = job_desc
+                .retain_lsns_below_horizon
                 .first()
                 .copied()
-                .unwrap_or(gc_cutoff);
+                .unwrap_or(job_desc.gc_cutoff);
             if cfg!(debug_assertions) {
                 assert_eq!(
                     res,
-                    retain_lsns_below_horizon
+                    job_desc
+                        .retain_lsns_below_horizon
                         .iter()
                         .min()
                         .copied()
-                        .unwrap_or(gc_cutoff)
+                        .unwrap_or(job_desc.gc_cutoff)
                 );
             }
             res
         };
         info!(
-            "picked {} layers for compaction with gc_cutoff={} lowest_retain_lsn={}",
-            layer_selection.len(),
-            gc_cutoff,
-            lowest_retain_lsn
+            "picked {} layers for compaction ({} layers need rewriting) with max_layer_lsn={} gc_cutoff={} lowest_retain_lsn={}, key_range={}..{}",
+            job_desc.selected_layers.len(),
+            job_desc.rewrite_layers.len(),
+            job_desc.max_layer_lsn,
+            job_desc.gc_cutoff,
+            lowest_retain_lsn,
+            job_desc.compaction_key_range.start,
+            job_desc.compaction_key_range.end
         );
 
-        self.check_compaction_space(&layer_selection).await?;
+        for layer in &job_desc.selected_layers {
+            debug!("read layer: {}", layer.layer_desc().key());
+        }
+        for layer in &job_desc.rewrite_layers {
+            debug!("rewrite layer: {}", layer.key());
+        }
+
+        self.check_compaction_space(&job_desc.selected_layers)
+            .await?;
 
         // Generate statistics for the compaction
-        for layer in &layer_selection {
+        for layer in &job_desc.selected_layers {
             let desc = layer.layer_desc();
             if desc.is_delta() {
                 stat.visit_delta_layer(desc.file_size());
@@ -1906,25 +1915,25 @@ impl Timeline {
 
         // Step 1: construct a k-merge iterator over all layers.
         // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
-        let layer_names: Vec<crate::tenant::storage_layer::LayerName> = layer_selection
+        let layer_names = job_desc
+            .selected_layers
             .iter()
             .map(|layer| layer.layer_desc().layer_name())
             .collect_vec();
         if let Some(err) = check_valid_layermap(&layer_names) {
-            bail!("cannot run gc-compaction because {}", err);
+            warn!("gc-compaction layer map check failed because {}, this is normal if partial compaction is not finished yet", err);
         }
         // The maximum LSN we are processing in this compaction loop
-        let end_lsn = layer_selection
+        let end_lsn = job_desc
+            .selected_layers
             .iter()
             .map(|l| l.layer_desc().lsn_range.end)
             .max()
             .unwrap();
-        // We don't want any of the produced layers to cover the full key range (i.e., MIN..MAX) b/c it will then be recognized
-        // as an L0 layer.
         let mut delta_layers = Vec::new();
         let mut image_layers = Vec::new();
         let mut downloaded_layers = Vec::new();
-        for layer in &layer_selection {
+        for layer in &job_desc.selected_layers {
             let resident_layer = layer.download_and_keep_resident().await?;
             downloaded_layers.push(resident_layer);
         }
@@ -1943,8 +1952,8 @@ impl Timeline {
             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.
+
+        // Step 2: Produce images+deltas.
         let mut accumulated_values = Vec::new();
         let mut last_key: Option<Key> = None;
 
@@ -1956,10 +1965,7 @@ impl Timeline {
                     self.conf,
                     self.timeline_id,
                     self.tenant_shard_id,
-                    compaction_key_range
-                        .as_ref()
-                        .map(|x| x.start)
-                        .unwrap_or(Key::MIN),
+                    job_desc.compaction_key_range.start,
                     lowest_retain_lsn,
                     self.get_compaction_target_size(),
                     ctx,
@@ -1979,6 +1985,13 @@ impl Timeline {
         )
         .await?;
 
+        #[derive(Default)]
+        struct RewritingLayers {
+            before: Option<DeltaLayerWriter>,
+            after: Option<DeltaLayerWriter>,
+        }
+        let mut delta_layer_rewriters = HashMap::<Arc<PersistentLayerKey>, RewritingLayers>::new();
+
         /// Returns None if there is no ancestor branch. Throw an error when the key is not found.
         ///
         /// Currently, we always get the ancestor image for each key in the child branch no matter whether the image
@@ -2004,10 +2017,51 @@ impl Timeline {
         // the key and LSN range are determined. However, to keep things simple here, we still
         // create this writer, and discard the writer in the end.
 
-        while let Some((key, lsn, val)) = merge_iter.next().await? {
+        while let Some(((key, lsn, val), desc)) = merge_iter.next_with_trace().await? {
             if cancel.is_cancelled() {
                 return Err(anyhow!("cancelled")); // TODO: refactor to CompactionError and pass cancel error
             }
+            if !job_desc.compaction_key_range.contains(&key) {
+                if !desc.is_delta {
+                    continue;
+                }
+                let rewriter = delta_layer_rewriters.entry(desc.clone()).or_default();
+                let rewriter = if key < job_desc.compaction_key_range.start {
+                    if rewriter.before.is_none() {
+                        rewriter.before = Some(
+                            DeltaLayerWriter::new(
+                                self.conf,
+                                self.timeline_id,
+                                self.tenant_shard_id,
+                                desc.key_range.start,
+                                desc.lsn_range.clone(),
+                                ctx,
+                            )
+                            .await?,
+                        );
+                    }
+                    rewriter.before.as_mut().unwrap()
+                } else if key >= job_desc.compaction_key_range.end {
+                    if rewriter.after.is_none() {
+                        rewriter.after = Some(
+                            DeltaLayerWriter::new(
+                                self.conf,
+                                self.timeline_id,
+                                self.tenant_shard_id,
+                                job_desc.compaction_key_range.end,
+                                desc.lsn_range.clone(),
+                                ctx,
+                            )
+                            .await?,
+                        );
+                    }
+                    rewriter.after.as_mut().unwrap()
+                } else {
+                    unreachable!()
+                };
+                rewriter.put_value(key, lsn, val, ctx).await?;
+                continue;
+            }
             match val {
                 Value::Image(_) => stat.visit_image_key(&val),
                 Value::WalRecord(_) => stat.visit_wal_key(&val),
@@ -2018,35 +2072,27 @@ impl Timeline {
                 }
                 accumulated_values.push((key, lsn, val));
             } else {
-                let last_key = last_key.as_mut().unwrap();
-                stat.on_unique_key_visited();
-                let skip_adding_key = if let Some(ref compaction_key_range) = compaction_key_range {
-                    !compaction_key_range.contains(last_key)
-                } else {
-                    false
-                };
-                if !skip_adding_key {
-                    let retention = self
-                        .generate_key_retention(
-                            *last_key,
-                            &accumulated_values,
-                            gc_cutoff,
-                            &retain_lsns_below_horizon,
-                            COMPACTION_DELTA_THRESHOLD,
-                            get_ancestor_image(self, *last_key, ctx).await?,
-                        )
-                        .await?;
-                    // Put the image into the image layer. Currently we have a single big layer for the compaction.
-                    retention
-                        .pipe_to(
-                            *last_key,
-                            &mut delta_layer_writer,
-                            image_layer_writer.as_mut(),
-                            &mut stat,
-                            ctx,
-                        )
-                        .await?;
-                }
+                let last_key: &mut Key = last_key.as_mut().unwrap();
+                stat.on_unique_key_visited(); // TODO: adjust statistics for partial compaction
+                let retention = self
+                    .generate_key_retention(
+                        *last_key,
+                        &accumulated_values,
+                        job_desc.gc_cutoff,
+                        &job_desc.retain_lsns_below_horizon,
+                        COMPACTION_DELTA_THRESHOLD,
+                        get_ancestor_image(self, *last_key, ctx).await?,
+                    )
+                    .await?;
+                retention
+                    .pipe_to(
+                        *last_key,
+                        &mut delta_layer_writer,
+                        image_layer_writer.as_mut(),
+                        &mut stat,
+                        ctx,
+                    )
+                    .await?;
                 accumulated_values.clear();
                 *last_key = key;
                 accumulated_values.push((key, lsn, val));
@@ -2057,35 +2103,43 @@ impl Timeline {
         let last_key = last_key.expect("no keys produced during compaction");
         stat.on_unique_key_visited();
 
-        let skip_adding_key = if let Some(ref compaction_key_range) = compaction_key_range {
-            !compaction_key_range.contains(&last_key)
-        } else {
-            false
-        };
-        if !skip_adding_key {
-            let retention = self
-                .generate_key_retention(
-                    last_key,
-                    &accumulated_values,
-                    gc_cutoff,
-                    &retain_lsns_below_horizon,
-                    COMPACTION_DELTA_THRESHOLD,
-                    get_ancestor_image(self, last_key, ctx).await?,
-                )
-                .await?;
-            // Put the image into the image layer. Currently we have a single big layer for the compaction.
-            retention
-                .pipe_to(
-                    last_key,
-                    &mut delta_layer_writer,
-                    image_layer_writer.as_mut(),
-                    &mut stat,
-                    ctx,
-                )
-                .await?;
-        }
+        let retention = self
+            .generate_key_retention(
+                last_key,
+                &accumulated_values,
+                job_desc.gc_cutoff,
+                &job_desc.retain_lsns_below_horizon,
+                COMPACTION_DELTA_THRESHOLD,
+                get_ancestor_image(self, last_key, ctx).await?,
+            )
+            .await?;
+        retention
+            .pipe_to(
+                last_key,
+                &mut delta_layer_writer,
+                image_layer_writer.as_mut(),
+                &mut stat,
+                ctx,
+            )
+            .await?;
         // end: move the above part to the loop body
 
+        let mut rewrote_delta_layers = Vec::new();
+        for (key, writers) in delta_layer_rewriters {
+            if let Some(delta_writer_before) = writers.before {
+                let (desc, path) = delta_writer_before
+                    .finish(job_desc.compaction_key_range.start, ctx)
+                    .await?;
+                let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
+                rewrote_delta_layers.push(layer);
+            }
+            if let Some(delta_writer_after) = writers.after {
+                let (desc, path) = delta_writer_after.finish(key.key_range.end, ctx).await?;
+                let layer = Layer::finish_creating(self.conf, self, desc, &path)?;
+                rewrote_delta_layers.push(layer);
+            }
+        }
+
         let discard = |key: &PersistentLayerKey| {
             let key = key.clone();
             async move { KeyHistoryRetention::discard_key(&key, self, dry_run).await }
@@ -2093,10 +2147,7 @@ impl Timeline {
 
         let produced_image_layers = if let Some(writer) = image_layer_writer {
             if !dry_run {
-                let end_key = compaction_key_range
-                    .as_ref()
-                    .map(|x| x.end)
-                    .unwrap_or(Key::MAX);
+                let end_key = job_desc.compaction_key_range.end;
                 writer
                     .finish_with_discard_fn(self, ctx, end_key, discard)
                     .await?
@@ -2117,10 +2168,8 @@ impl Timeline {
             Vec::new()
         };
 
-        if partial_compaction && !produced_delta_layers.is_empty() {
-            bail!("implementation error: partial compaction should not be producing delta layers (for now)");
-        }
-
+        // TODO: make image/delta/rewrote_delta layers generation atomic. At this point, we already generated resident layers, and if
+        // compaction is cancelled at this point, we might have some layers that are not cleaned up.
         let mut compact_to = Vec::new();
         let mut keep_layers = HashSet::new();
         let produced_delta_layers_len = produced_delta_layers.len();
@@ -2128,52 +2177,84 @@ impl Timeline {
         for action in produced_delta_layers {
             match action {
                 BatchWriterResult::Produced(layer) => {
+                    if cfg!(debug_assertions) {
+                        info!("produced delta layer: {}", layer.layer_desc().key());
+                    }
                     stat.produce_delta_layer(layer.layer_desc().file_size());
                     compact_to.push(layer);
                 }
                 BatchWriterResult::Discarded(l) => {
+                    if cfg!(debug_assertions) {
+                        info!("discarded delta layer: {}", l);
+                    }
                     keep_layers.insert(l);
                     stat.discard_delta_layer();
                 }
             }
         }
+        for layer in &rewrote_delta_layers {
+            debug!(
+                "produced rewritten delta layer: {}",
+                layer.layer_desc().key()
+            );
+        }
+        compact_to.extend(rewrote_delta_layers);
         for action in produced_image_layers {
             match action {
                 BatchWriterResult::Produced(layer) => {
+                    debug!("produced image layer: {}", layer.layer_desc().key());
                     stat.produce_image_layer(layer.layer_desc().file_size());
                     compact_to.push(layer);
                 }
                 BatchWriterResult::Discarded(l) => {
+                    debug!("discarded image layer: {}", l);
                     keep_layers.insert(l);
                     stat.discard_image_layer();
                 }
             }
         }
-        let mut layer_selection = layer_selection;
-        layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
-        if let Some(ref compaction_key_range) = compaction_key_range {
-            // Partial compaction might select more data than it processes, e.g., if
-            // the compaction_key_range only partially overlaps:
-            //
-            //         [---compaction_key_range---]
-            //   [---A----][----B----][----C----][----D----]
-            //
-            // A,B,C,D are all in the `layer_selection`. The created image layers contain
-            // whatever is needed from B, C, and from `----]` of A, and from  `[--` of D.
-            //
-            // In contrast, `[--A-` and `--D----]` have not been processed, so, we must
-            // keep that data.
-            //
-            // The solution for now is to keep A and D completely.
-            // (layer_selection is what we'll remove from the layer map, so,
-            //  retain what is _not_ fully covered by compaction_key_range).
-            layer_selection.retain(|x| {
-                let key_range = &x.layer_desc().key_range;
-                key_range.start >= compaction_key_range.start
-                    && key_range.end <= compaction_key_range.end
-            });
+
+        let mut layer_selection = job_desc.selected_layers;
+
+        // Partial compaction might select more data than it processes, e.g., if
+        // the compaction_key_range only partially overlaps:
+        //
+        //         [---compaction_key_range---]
+        //   [---A----][----B----][----C----][----D----]
+        //
+        // For delta layers, we will rewrite the layers so that it is cut exactly at
+        // the compaction key range, so we can always discard them. However, for image
+        // layers, as we do not rewrite them for now, we need to handle them differently.
+        // Assume image layers  A, B, C, D are all in the `layer_selection`.
+        //
+        // The created image layers contain whatever is needed from B, C, and from
+        // `----]` of A, and from  `[---` of D.
+        //
+        // In contrast, `[---A` and `D----]` have not been processed, so, we must
+        // keep that data.
+        //
+        // The solution for now is to keep A and D completely if they are image layers.
+        // (layer_selection is what we'll remove from the layer map, so, retain what
+        // is _not_ fully covered by compaction_key_range).
+        for layer in &layer_selection {
+            if !layer.layer_desc().is_delta() {
+                if !overlaps_with(
+                    &layer.layer_desc().key_range,
+                    &job_desc.compaction_key_range,
+                ) {
+                    bail!("violated constraint: image layer outside of compaction key range");
+                }
+                if !fully_contains(
+                    &job_desc.compaction_key_range,
+                    &layer.layer_desc().key_range,
+                ) {
+                    keep_layers.insert(layer.layer_desc().key());
+                }
+            }
         }
 
+        layer_selection.retain(|x| !keep_layers.contains(&x.layer_desc().key()));
+
         info!(
             "gc-compaction statistics: {}",
             serde_json::to_string(&stat)?
@@ -2192,6 +2273,7 @@ impl Timeline {
 
         // Step 3: Place back to the layer map.
         {
+            // TODO: sanity check if the layer map is valid (i.e., should not have overlaps)
             let mut guard = self.layers.write().await;
             guard
                 .open_mut()?

pageserver/src/tenant/timeline/offload.rs

@@ -47,21 +47,18 @@ pub(crate) async fn offload_timeline(
     match is_archived {
         Some(true) => (),
         Some(false) => {
-            tracing::warn!(?is_archived, "tried offloading a non-archived timeline");
+            tracing::warn!("tried offloading a non-archived timeline");
             return Err(OffloadError::NotArchived);
         }
         None => {
             // This is legal: calls to this function can race with the timeline shutting down
-            tracing::info!(
-                ?is_archived,
-                "tried offloading a timeline whose remote storage is not initialized"
-            );
+            tracing::info!("tried offloading a timeline whose remote storage is not initialized");
             return Err(OffloadError::Cancelled);
         }
     }
 
     // Now that the Timeline is in Stopping state, request all the related tasks to shut down.
-    timeline.shutdown(super::ShutdownMode::Hard).await;
+    timeline.shutdown(super::ShutdownMode::Flush).await;
 
     // TODO extend guard mechanism above with method
     // to make deletions possible while offloading is in progress

pageserver/src/walredo/apply_neon.rs

@@ -253,6 +253,10 @@ pub(crate) fn apply_in_neon(
             use bytes::BufMut;
             if *will_init {
                 assert!(*clear, "init record must be clear to ensure correctness");
+                assert!(
+                    page.is_empty(),
+                    "init record must be the first entry to ensure correctness"
+                );
             }
             if *clear {
                 page.clear();

pgxn/neon/logical_replication_monitor.c

@@ -1,7 +1,8 @@
+#include <dirent.h>
 #include <limits.h>
 #include <string.h>
-#include <dirent.h>
 #include <signal.h>
+#include <sys/stat.h>
 
 #include "postgres.h"
 
@@ -21,17 +22,35 @@
 
 static int	logical_replication_max_snap_files = 300;
 
+/*
+ * According to Chi (shyzh), the pageserver _should_ be good with 10 MB worth of
+ * snapshot files. Let's use 8 MB since 8 is a power of 2.
+ */
+static int	logical_replication_max_logicalsnapdir_size = 8000;
+
+/*
+ * A primitive description of a logical snapshot file including the LSN of the
+ * file and its size.
+ */
+typedef struct SnapDesc {
+	XLogRecPtr	lsn;
+	off_t		sz;
+} SnapDesc;
+
 PGDLLEXPORT void LogicalSlotsMonitorMain(Datum main_arg);
 
+/*
+ * Sorts an array of snapshot descriptors by their LSN.
+ */
 static int
-LsnDescComparator(const void *a, const void *b)
+SnapDescComparator(const void *a, const void *b)
 {
-	XLogRecPtr	lsn1 = *((const XLogRecPtr *) a);
-	XLogRecPtr	lsn2 = *((const XLogRecPtr *) b);
+	const SnapDesc	*desc1 = a;
+	const SnapDesc	*desc2 = b;
 
-	if (lsn1 < lsn2)
+	if (desc1->lsn < desc2->lsn)
 		return 1;
-	else if (lsn1 == lsn2)
+	else if (desc1->lsn == desc2->lsn)
 		return 0;
 	else
 		return -1;
@@ -43,28 +62,39 @@ LsnDescComparator(const void *a, const void *b)
  * slots having lower restart_lsn should be dropped.
  */
 static XLogRecPtr
-get_num_snap_files_lsn_threshold(void)
+get_snapshots_cutoff_lsn(void)
 {
-	DIR		   *dirdesc;
-	struct dirent *de;
-	char	   *snap_path = "pg_logical/snapshots/";
-	int			lsns_allocated = 1024;
-	int			lsns_num = 0;
-	XLogRecPtr *lsns;
-	XLogRecPtr	cutoff;
+/* PG 18 has a constant defined for this, PG_LOGICAL_SNAPSHOTS_DIR */
+#define SNAPDIR "pg_logical/snapshots"
 
-	if (logical_replication_max_snap_files < 0)
+	DIR		   *dirdesc;
+	int			dirdesc_fd;
+	struct dirent *de;
+	size_t		snapshot_index = 0;
+	SnapDesc   *snapshot_descriptors;
+	size_t		descriptors_allocated = 1024;
+	XLogRecPtr	cutoff = 0;
+	off_t		logicalsnapdir_size = 0;
+	const int	logical_replication_max_logicalsnapdir_size_bytes = logical_replication_max_logicalsnapdir_size * 1000;
+
+	if (logical_replication_max_snap_files < 0 && logical_replication_max_logicalsnapdir_size < 0)
 		return 0;
 
-	lsns = palloc(sizeof(XLogRecPtr) * lsns_allocated);
+	snapshot_descriptors = palloc(sizeof(*snapshot_descriptors) * descriptors_allocated);
+
+	dirdesc = AllocateDir(SNAPDIR);
+	dirdesc_fd = dirfd(dirdesc);
+	if (dirdesc_fd == -1)
+		ereport(ERROR, errmsg("failed to get a file descriptor for " SNAPDIR ": %m"));
 
 	/* find all .snap files and get their lsns */
-	dirdesc = AllocateDir(snap_path);
-	while ((de = ReadDir(dirdesc, snap_path)) != NULL)
+	while ((de = ReadDir(dirdesc, SNAPDIR)) != NULL)
 	{
-		XLogRecPtr	lsn;
 		uint32		hi;
 		uint32		lo;
+		struct stat	st;
+		XLogRecPtr	lsn;
+		SnapDesc   *desc;
 
 		if (strcmp(de->d_name, ".") == 0 ||
 			strcmp(de->d_name, "..") == 0)
@@ -79,28 +109,69 @@ get_num_snap_files_lsn_threshold(void)
 
 		lsn = ((uint64) hi) << 32 | lo;
 		elog(DEBUG5, "found snap file %X/%X", LSN_FORMAT_ARGS(lsn));
-		if (lsns_allocated == lsns_num)
+
+		if (fstatat(dirdesc_fd, de->d_name, &st, 0) == -1)
+			ereport(ERROR, errmsg("failed to get the size of " SNAPDIR "/%s: %m", de->d_name));
+
+		if (descriptors_allocated == snapshot_index)
 		{
-			lsns_allocated *= 2;
-			lsns = repalloc(lsns, sizeof(XLogRecPtr) * lsns_allocated);
+			descriptors_allocated *= 2;
+			snapshot_descriptors = repalloc(snapshot_descriptors, sizeof(*snapshot_descriptors) * descriptors_allocated);
 		}
-		lsns[lsns_num++] = lsn;
+
+		desc = &snapshot_descriptors[snapshot_index++];
+		desc->lsn = lsn;
+		desc->sz = st.st_size;
 	}
-	/* sort by lsn desc */
-	qsort(lsns, lsns_num, sizeof(XLogRecPtr), LsnDescComparator);
-	/* and take cutoff at logical_replication_max_snap_files */
-	if (logical_replication_max_snap_files > lsns_num)
-		cutoff = 0;
-	/* have less files than cutoff */
-	else
+
+	qsort(snapshot_descriptors, snapshot_index, sizeof(*snapshot_descriptors), SnapDescComparator);
+
+	/* Are there more snapshot files than specified? */
+	if (logical_replication_max_snap_files <= snapshot_index)
 	{
-		cutoff = lsns[logical_replication_max_snap_files - 1];
-		elog(LOG, "ls_monitor: dropping logical slots with restart_lsn lower %X/%X, found %d .snap files, limit is %d",
-			 LSN_FORMAT_ARGS(cutoff), lsns_num, logical_replication_max_snap_files);
+		cutoff = snapshot_descriptors[logical_replication_max_snap_files - 1].lsn;
+		elog(LOG,
+			"ls_monitor: dropping logical slots with restart_lsn lower %X/%X, found %zu snapshot files, limit is %d",
+			LSN_FORMAT_ARGS(cutoff), snapshot_index, logical_replication_max_snap_files);
 	}
-	pfree(lsns);
+
+	/* Is the size of the logical snapshots directory larger than specified?
+	 *
+	 * It's possible we could hit both thresholds, so remove any extra files
+	 * first, and then truncate based on size of the remaining files.
+	 */
+	if (logicalsnapdir_size > logical_replication_max_logicalsnapdir_size_bytes)
+	{
+		/* Unfortunately, iterating the directory does not guarantee any order
+		 * so we can't cache an index in the preceding loop.
+		 */
+
+		off_t		sz;
+		const XLogRecPtr original = cutoff;
+
+		sz = snapshot_descriptors[0].sz;
+		for (size_t i = 1; i < logical_replication_max_snap_files; ++i)
+		{
+			if (sz > logical_replication_max_logicalsnapdir_size_bytes)
+			{
+				cutoff = snapshot_descriptors[i - 1].lsn;
+				break;
+			}
+
+			sz += snapshot_descriptors[i].sz;
+		}
+
+		if (cutoff != original)
+			elog(LOG, "ls_monitor: dropping logical slots with restart_lsn lower than %X/%X, " SNAPDIR " is larger than %d KB",
+					LSN_FORMAT_ARGS(cutoff), logical_replication_max_logicalsnapdir_size);
+	}
+
+	pfree(snapshot_descriptors);
 	FreeDir(dirdesc);
+
 	return cutoff;
+
+#undef SNAPDIR
 }
 
 void
@@ -118,6 +189,16 @@ InitLogicalReplicationMonitor(void)
 							0,
 							NULL, NULL, NULL);
 
+	DefineCustomIntVariable(
+							"neon.logical_replication_max_logicalsnapdir_size",
+							"Maximum allowed size of the pg_logical/snapshots directory (KB). When exceeded, slots are dropped until the limit is met. -1 disables the limit.",
+							NULL,
+							&logical_replication_max_logicalsnapdir_size,
+							8000, -1, INT_MAX,
+							PGC_SIGHUP,
+							GUC_UNIT_KB,
+							NULL, NULL, NULL);
+
 	memset(&bgw, 0, sizeof(bgw));
 	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
 	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
@@ -162,7 +243,7 @@ LogicalSlotsMonitorMain(Datum main_arg)
 		 * If there are too many .snap files, just drop all logical slots to
 		 * prevent aux files bloat.
 		 */
-		cutoff_lsn = get_num_snap_files_lsn_threshold();
+		cutoff_lsn = get_snapshots_cutoff_lsn();
 		if (cutoff_lsn > 0)
 		{
 			for (int i = 0; i < max_replication_slots; i++)

pgxn/neon/neon_walreader.c

@@ -611,6 +611,17 @@ NeonWALReadLocal(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size coun
 	recptr = startptr;
 	nbytes = count;
 
+/* Try to read directly from WAL buffers first. */
+#if PG_MAJORVERSION_NUM >= 17
+	{
+		Size	rbytes;
+		rbytes = WALReadFromBuffers(p, recptr, nbytes, tli);
+		recptr += rbytes;
+		nbytes -= rbytes;
+		p += rbytes;
+	}
+#endif
+
 	while (nbytes > 0)
 	{
 		uint32		startoff;

pgxn/neon/walproposer.c

@@ -1361,29 +1361,35 @@ SendAppendRequests(Safekeeper *sk)
 		if (sk->active_state == SS_ACTIVE_READ_WAL)
 		{
 			char	   *errmsg;
+			int			req_len;
 
 			req = &sk->appendRequest;
+			req_len = req->endLsn - req->beginLsn;
 
-			switch (wp->api.wal_read(sk,
-									 &sk->outbuf.data[sk->outbuf.len],
-									 req->beginLsn,
-									 req->endLsn - req->beginLsn,
-									 &errmsg))
+			/* We send zero sized AppenRequests as heartbeats; don't wal_read for these. */
+			if (req_len > 0)
 			{
-				case NEON_WALREAD_SUCCESS:
-					break;
-				case NEON_WALREAD_WOULDBLOCK:
-					return true;
-				case NEON_WALREAD_ERROR:
-					wp_log(WARNING, "WAL reading for node %s:%s failed: %s",
-						   sk->host, sk->port, errmsg);
-					ShutdownConnection(sk);
-					return false;
-				default:
-					Assert(false);
+				switch (wp->api.wal_read(sk,
+										&sk->outbuf.data[sk->outbuf.len],
+										req->beginLsn,
+										req_len,
+										&errmsg))
+				{
+					case NEON_WALREAD_SUCCESS:
+						break;
+					case NEON_WALREAD_WOULDBLOCK:
+						return true;
+					case NEON_WALREAD_ERROR:
+						wp_log(WARNING, "WAL reading for node %s:%s failed: %s",
+							sk->host, sk->port, errmsg);
+						ShutdownConnection(sk);
+						return false;
+					default:
+						Assert(false);
+				}
 			}
 
-			sk->outbuf.len += req->endLsn - req->beginLsn;
+			sk->outbuf.len += req_len;
 
 			writeResult = wp->api.conn_async_write(sk, sk->outbuf.data, sk->outbuf.len);
 

pgxn/neon/walproposer_pg.c

@@ -1489,33 +1489,11 @@ walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count,
 {
 	NeonWALReadResult res;
 
-#if PG_MAJORVERSION_NUM >= 17
-	if (!sk->wp->config->syncSafekeepers)
-	{
-		Size	rbytes;
-		rbytes = WALReadFromBuffers(buf, startptr, count,
-									walprop_pg_get_timeline_id());
-
-		startptr += rbytes;
-		count -= rbytes;
-	}
-#endif
-
-	if (count == 0)
-	{
-		res = NEON_WALREAD_SUCCESS;
-	}
-	else
-	{
-		Assert(count > 0);
-
-		/* Now read the remaining WAL from the WAL file */
-		res = NeonWALRead(sk->xlogreader,
-						  buf,
-						  startptr,
-						  count,
-						  walprop_pg_get_timeline_id());
-	}
+	res = NeonWALRead(sk->xlogreader,
+					  buf,
+					  startptr,
+					  count,
+					  walprop_pg_get_timeline_id());
 
 	if (res == NEON_WALREAD_SUCCESS)
 	{

proxy/Cargo.toml

@@ -60,7 +60,7 @@ prometheus.workspace = true
 rand.workspace = true
 regex.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
-reqwest.workspace = true
+reqwest = { workspace = true, features = ["rustls-tls-native-roots"] }
 reqwest-middleware = { workspace = true, features = ["json"] }
 reqwest-retry.workspace = true
 reqwest-tracing.workspace = true

proxy/src/auth/backend/jwt.rs

@@ -7,8 +7,11 @@ use arc_swap::ArcSwapOption;
 use dashmap::DashMap;
 use jose_jwk::crypto::KeyInfo;
 use reqwest::{redirect, Client};
+use reqwest_retry::policies::ExponentialBackoff;
+use reqwest_retry::RetryTransientMiddleware;
 use serde::de::Visitor;
 use serde::{Deserialize, Deserializer};
+use serde_json::value::RawValue;
 use signature::Verifier;
 use thiserror::Error;
 use tokio::time::Instant;
@@ -16,7 +19,7 @@ use tokio::time::Instant;
 use crate::auth::backend::ComputeCredentialKeys;
 use crate::context::RequestMonitoring;
 use crate::control_plane::errors::GetEndpointJwksError;
-use crate::http::parse_json_body_with_limit;
+use crate::http::read_body_with_limit;
 use crate::intern::RoleNameInt;
 use crate::types::{EndpointId, RoleName};
 
@@ -28,6 +31,10 @@ const MAX_RENEW: Duration = Duration::from_secs(3600);
 const MAX_JWK_BODY_SIZE: usize = 64 * 1024;
 const JWKS_USER_AGENT: &str = "neon-proxy";
 
+const JWKS_CONNECT_TIMEOUT: Duration = Duration::from_secs(2);
+const JWKS_FETCH_TIMEOUT: Duration = Duration::from_secs(5);
+const JWKS_FETCH_RETRIES: u32 = 3;
+
 /// How to get the JWT auth rules
 pub(crate) trait FetchAuthRules: Clone + Send + Sync + 'static {
     fn fetch_auth_rules(
@@ -55,7 +62,7 @@ pub(crate) struct AuthRule {
 }
 
 pub struct JwkCache {
-    client: reqwest::Client,
+    client: reqwest_middleware::ClientWithMiddleware,
 
     map: DashMap<(EndpointId, RoleName), Arc<JwkCacheEntryLock>>,
 }
@@ -117,6 +124,14 @@ impl Default for JwkCacheEntryLock {
     }
 }
 
+#[derive(Deserialize)]
+struct JwkSet<'a> {
+    /// we parse into raw-value because not all keys in a JWKS are ones
+    /// we can parse directly, so we parse them lazily.
+    #[serde(borrow)]
+    keys: Vec<&'a RawValue>,
+}
+
 impl JwkCacheEntryLock {
     async fn acquire_permit<'a>(self: &'a Arc<Self>) -> JwkRenewalPermit<'a> {
         JwkRenewalPermit::acquire_permit(self).await
@@ -130,7 +145,7 @@ impl JwkCacheEntryLock {
         &self,
         _permit: JwkRenewalPermit<'_>,
         ctx: &RequestMonitoring,
-        client: &reqwest::Client,
+        client: &reqwest_middleware::ClientWithMiddleware,
         endpoint: EndpointId,
         auth_rules: &F,
     ) -> Result<Arc<JwkCacheEntry>, JwtError> {
@@ -154,22 +169,73 @@ impl JwkCacheEntryLock {
             let req = client.get(rule.jwks_url.clone());
             // TODO(conrad): eventually switch to using reqwest_middleware/`new_client_with_timeout`.
             // TODO(conrad): We need to filter out URLs that point to local resources. Public internet only.
-            match req.send().await.and_then(|r| r.error_for_status()) {
+            match req.send().await.and_then(|r| {
+                r.error_for_status()
+                    .map_err(reqwest_middleware::Error::Reqwest)
+            }) {
                 // todo: should we re-insert JWKs if we want to keep this JWKs URL?
                 // I expect these failures would be quite sparse.
                 Err(e) => tracing::warn!(url=?rule.jwks_url, error=?e, "could not fetch JWKs"),
                 Ok(r) => {
                     let resp: http::Response<reqwest::Body> = r.into();
-                    match parse_json_body_with_limit::<jose_jwk::JwkSet>(
-                        resp.into_body(),
-                        MAX_JWK_BODY_SIZE,
-                    )
-                    .await
+
+                    let bytes = match read_body_with_limit(resp.into_body(), MAX_JWK_BODY_SIZE)
+                        .await
                     {
+                        Ok(bytes) => bytes,
+                        Err(e) => {
+                            tracing::warn!(url=?rule.jwks_url, error=?e, "could not decode JWKs");
+                            continue;
+                        }
+                    };
+
+                    match serde_json::from_slice::<JwkSet>(&bytes) {
                         Err(e) => {
                             tracing::warn!(url=?rule.jwks_url, error=?e, "could not decode JWKs");
                         }
                         Ok(jwks) => {
+                            // size_of::<&RawValue>() == 16
+                            // size_of::<jose_jwk::Jwk>() == 288
+                            // better to not pre-allocate this as it might be pretty large - especially if it has many
+                            // keys we don't want or need.
+                            // trivial 'attack': `{"keys":[` + repeat(`0`).take(30000).join(`,`) + `]}`
+                            // this would consume 8MiB just like that!
+                            let mut keys = vec![];
+                            let mut failed = 0;
+                            for key in jwks.keys {
+                                match serde_json::from_str::<jose_jwk::Jwk>(key.get()) {
+                                    Ok(key) => {
+                                        // if `use` (called `cls` in rust) is specified to be something other than signing,
+                                        // we can skip storing it.
+                                        if key
+                                            .prm
+                                            .cls
+                                            .as_ref()
+                                            .is_some_and(|c| *c != jose_jwk::Class::Signing)
+                                        {
+                                            continue;
+                                        }
+
+                                        keys.push(key);
+                                    }
+                                    Err(e) => {
+                                        tracing::debug!(url=?rule.jwks_url, failed=?e, "could not decode JWK");
+                                        failed += 1;
+                                    }
+                                }
+                            }
+                            keys.shrink_to_fit();
+
+                            if failed > 0 {
+                                tracing::warn!(url=?rule.jwks_url, failed, "could not decode JWKs");
+                            }
+
+                            if keys.is_empty() {
+                                tracing::warn!(url=?rule.jwks_url, "no valid JWKs found inside the response body");
+                                continue;
+                            }
+
+                            let jwks = jose_jwk::JwkSet { keys };
                             key_sets.insert(
                                 rule.id,
                                 KeySet {
@@ -179,7 +245,7 @@ impl JwkCacheEntryLock {
                                 },
                             );
                         }
-                    }
+                    };
                 }
             }
         }
@@ -196,7 +262,7 @@ impl JwkCacheEntryLock {
     async fn get_or_update_jwk_cache<F: FetchAuthRules>(
         self: &Arc<Self>,
         ctx: &RequestMonitoring,
-        client: &reqwest::Client,
+        client: &reqwest_middleware::ClientWithMiddleware,
         endpoint: EndpointId,
         fetch: &F,
     ) -> Result<Arc<JwkCacheEntry>, JwtError> {
@@ -250,7 +316,7 @@ impl JwkCacheEntryLock {
         self: &Arc<Self>,
         ctx: &RequestMonitoring,
         jwt: &str,
-        client: &reqwest::Client,
+        client: &reqwest_middleware::ClientWithMiddleware,
         endpoint: EndpointId,
         role_name: &RoleName,
         fetch: &F,
@@ -369,8 +435,19 @@ impl Default for JwkCache {
         let client = Client::builder()
             .user_agent(JWKS_USER_AGENT)
             .redirect(redirect::Policy::none())
+            .tls_built_in_native_certs(true)
+            .connect_timeout(JWKS_CONNECT_TIMEOUT)
+            .timeout(JWKS_FETCH_TIMEOUT)
             .build()
-            .expect("using &str and standard redirect::Policy");
+            .expect("client config should be valid");
+
+        // Retry up to 3 times with increasing intervals between attempts.
+        let retry_policy = ExponentialBackoff::builder().build_with_max_retries(JWKS_FETCH_RETRIES);
+
+        let client = reqwest_middleware::ClientBuilder::new(client)
+            .with(RetryTransientMiddleware::new_with_policy(retry_policy))
+            .build();
+
         JwkCache {
             client,
             map: DashMap::default(),
@@ -1209,4 +1286,63 @@ X0n5X2/pBLJzxZc62ccvZYVnctBiFs6HbSnxpuMQCfkt/BcR/ttIepBQQIW86wHL
             }
         }
     }
+
+    #[tokio::test]
+    async fn check_jwk_keycloak_regression() {
+        let (rs, valid_jwk) = new_rsa_jwk(RS1, "rs1".into());
+        let valid_jwk = serde_json::to_value(valid_jwk).unwrap();
+
+        // This is valid, but we cannot parse it as we have no support for encryption JWKs, only signature based ones.
+        // This is taken directly from keycloak.
+        let invalid_jwk = serde_json::json! {
+            {
+                "kid": "U-Jc9xRli84eNqRpYQoIPF-GNuRWV3ZvAIhziRW2sbQ",
+                "kty": "RSA",
+                "alg": "RSA-OAEP",
+                "use": "enc",
+                "n": "yypYWsEKmM_wWdcPnSGLSm5ytw1WG7P7EVkKSulcDRlrM6HWj3PR68YS8LySYM2D9Z-79oAdZGKhIfzutqL8rK1vS14zDuPpAM-RWY3JuQfm1O_-1DZM8-07PmVRegP5KPxsKblLf_My8ByH6sUOIa1p2rbe2q_b0dSTXYu1t0dW-cGL5VShc400YymvTwpc-5uYNsaVxZajnB7JP1OunOiuCJ48AuVp3PqsLzgoXqlXEB1ZZdch3xT3bxaTtNruGvG4xmLZY68O_T3yrwTCNH2h_jFdGPyXdyZToCMSMK2qSbytlfwfN55pT9Vv42Lz1YmoB7XRjI9aExKPc5AxFw",
+                "e": "AQAB",
+                "x5c": [
+                    "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"
+                ],
+                "x5t": "QhfzMMnuAfkReTgZ1HtrfyOeeZs",
+                "x5t#S256": "cmHDUdKgLiRCEN28D5FBy9IJLFmR7QWfm77SLhGTCTU"
+            }
+        };
+
+        let jwks = serde_json::json! {{ "keys": [invalid_jwk, valid_jwk ] }};
+        let jwks_addr = jwks_server(move |path| match path {
+            "/" => Some(serde_json::to_vec(&jwks).unwrap()),
+            _ => None,
+        })
+        .await;
+
+        let role_name = RoleName::from("anonymous");
+        let role = RoleNameInt::from(&role_name);
+
+        let rules = vec![AuthRule {
+            id: "foo".to_owned(),
+            jwks_url: format!("http://{jwks_addr}/").parse().unwrap(),
+            audience: None,
+            role_names: vec![role],
+        }];
+
+        let fetch = Fetch(rules);
+        let jwk_cache = JwkCache::default();
+
+        let endpoint = EndpointId::from("ep");
+
+        let token = new_rsa_jwt("rs1".into(), rs);
+
+        jwk_cache
+            .check_jwt(
+                &RequestMonitoring::test(),
+                endpoint.clone(),
+                &role_name,
+                &fetch,
+                &token,
+            )
+            .await
+            .unwrap();
+    }
 }

proxy/src/http/mod.rs

@@ -6,7 +6,6 @@ pub mod health_server;
 
 use std::time::Duration;
 
-use anyhow::bail;
 use bytes::Bytes;
 use http::Method;
 use http_body_util::BodyExt;
@@ -16,7 +15,7 @@ use reqwest_middleware::RequestBuilder;
 pub(crate) use reqwest_middleware::{ClientWithMiddleware, Error};
 pub(crate) use reqwest_retry::policies::ExponentialBackoff;
 pub(crate) use reqwest_retry::RetryTransientMiddleware;
-use serde::de::DeserializeOwned;
+use thiserror::Error;
 
 use crate::metrics::{ConsoleRequest, Metrics};
 use crate::url::ApiUrl;
@@ -122,10 +121,19 @@ impl Endpoint {
     }
 }
 
-pub(crate) async fn parse_json_body_with_limit<D: DeserializeOwned>(
+#[derive(Error, Debug)]
+pub(crate) enum ReadBodyError {
+    #[error("Content length exceeds limit of {limit} bytes")]
+    BodyTooLarge { limit: usize },
+
+    #[error(transparent)]
+    Read(#[from] reqwest::Error),
+}
+
+pub(crate) async fn read_body_with_limit(
     mut b: impl Body<Data = Bytes, Error = reqwest::Error> + Unpin,
     limit: usize,
-) -> anyhow::Result<D> {
+) -> Result<Vec<u8>, ReadBodyError> {
     // We could use `b.limited().collect().await.to_bytes()` here
     // but this ends up being slightly more efficient as far as I can tell.
 
@@ -133,20 +141,20 @@ pub(crate) async fn parse_json_body_with_limit<D: DeserializeOwned>(
     // in reqwest, this value is influenced by the Content-Length header.
     let lower_bound = match usize::try_from(b.size_hint().lower()) {
         Ok(bound) if bound <= limit => bound,
-        _ => bail!("Content length exceeds limit of {limit} bytes"),
+        _ => return Err(ReadBodyError::BodyTooLarge { limit }),
     };
     let mut bytes = Vec::with_capacity(lower_bound);
 
     while let Some(frame) = b.frame().await.transpose()? {
         if let Ok(data) = frame.into_data() {
             if bytes.len() + data.len() > limit {
-                bail!("Content length exceeds limit of {limit} bytes")
+                return Err(ReadBodyError::BodyTooLarge { limit });
             }
             bytes.extend_from_slice(&data);
         }
     }
 
-    Ok(serde_json::from_slice::<D>(&bytes)?)
+    Ok(bytes)
 }
 
 #[cfg(test)]

proxy/src/lib.rs

@@ -1,12 +1,6 @@
 // rustc lints/lint groups
 // https://doc.rust-lang.org/rustc/lints/groups.html
-#![deny(
-    deprecated,
-    future_incompatible,
-    let_underscore,
-    nonstandard_style,
-    rust_2024_compatibility
-)]
+#![deny(deprecated, future_incompatible, let_underscore, nonstandard_style)]
 #![warn(clippy::all, clippy::pedantic, clippy::cargo)]
 // List of denied lints from the clippy::restriction group.
 // https://rust-lang.github.io/rust-clippy/master/index.html#?groups=restriction

proxy/src/serverless/conn_pool_lib.rs

@@ -16,8 +16,7 @@ use super::http_conn_pool::ClientDataHttp;
 use super::local_conn_pool::ClientDataLocal;
 use crate::auth::backend::ComputeUserInfo;
 use crate::context::RequestMonitoring;
-use crate::control_plane::messages::ColdStartInfo;
-use crate::control_plane::messages::MetricsAuxInfo;
+use crate::control_plane::messages::{ColdStartInfo, MetricsAuxInfo};
 use crate::metrics::{HttpEndpointPoolsGuard, Metrics};
 use crate::types::{DbName, EndpointCacheKey, RoleName};
 use crate::usage_metrics::{Ids, MetricCounter, USAGE_METRICS};

proxy/src/serverless/http_conn_pool.rs

@@ -7,7 +7,6 @@ use hyper::client::conn::http2;
 use hyper_util::rt::{TokioExecutor, TokioIo};
 use parking_lot::RwLock;
 use rand::Rng;
-use std::result::Result::Ok;
 use tokio::net::TcpStream;
 use tracing::{debug, error, info, info_span, Instrument};
 

safekeeper/Cargo.toml

@@ -61,8 +61,14 @@ utils.workspace = true
 workspace_hack.workspace = true
 
 [dev-dependencies]
+criterion.workspace = true
+itertools.workspace = true
 walproposer.workspace = true
 rand.workspace = true
 desim.workspace = true
 tracing.workspace = true
 tracing-subscriber = { workspace = true, features = ["json"] }
+
+[[bench]]
+name = "receive_wal"
+harness = false

safekeeper/benches/README.md

@@ -0,0 +1,22 @@
+## Safekeeper Benchmarks
+
+To run benchmarks:
+
+```sh
+# All benchmarks.
+cargo bench --package safekeeper
+
+# Specific file.
+cargo bench --package safekeeper --bench receive_wal
+
+# Specific benchmark.
+cargo bench --package safekeeper --bench receive_wal process_msg/fsync=false
+
+# List available benchmarks.
+cargo bench --package safekeeper --benches -- --list
+```
+
+Additional charts and statistics are available in `target/criterion/report/index.html`.
+
+Benchmarks are automatically compared against the previous run. To compare against other runs, see
+`--baseline` and `--save-baseline`.

safekeeper/benches/benchutils.rs

@@ -0,0 +1,102 @@
+use std::sync::Arc;
+
+use camino_tempfile::Utf8TempDir;
+use safekeeper::rate_limit::RateLimiter;
+use safekeeper::safekeeper::{ProposerAcceptorMessage, ProposerElected, SafeKeeper, TermHistory};
+use safekeeper::state::{TimelinePersistentState, TimelineState};
+use safekeeper::timeline::{get_timeline_dir, SharedState, StateSK, Timeline};
+use safekeeper::timelines_set::TimelinesSet;
+use safekeeper::wal_backup::remote_timeline_path;
+use safekeeper::{control_file, wal_storage, SafeKeeperConf};
+use tokio::fs::create_dir_all;
+use utils::id::{NodeId, TenantTimelineId};
+use utils::lsn::Lsn;
+
+/// A Safekeeper benchmarking environment. Uses a tempdir for storage, removed on drop.
+pub struct Env {
+    /// Whether to enable fsync.
+    pub fsync: bool,
+    /// Benchmark directory. Deleted when dropped.
+    pub tempdir: Utf8TempDir,
+}
+
+impl Env {
+    /// Creates a new benchmarking environment in a temporary directory. fsync controls whether to
+    /// enable fsyncing.
+    pub fn new(fsync: bool) -> anyhow::Result<Self> {
+        let tempdir = camino_tempfile::tempdir()?;
+        Ok(Self { fsync, tempdir })
+    }
+
+    /// Constructs a Safekeeper config for the given node ID.
+    fn make_conf(&self, node_id: NodeId) -> SafeKeeperConf {
+        let mut conf = SafeKeeperConf::dummy();
+        conf.my_id = node_id;
+        conf.no_sync = !self.fsync;
+        conf.workdir = self.tempdir.path().join(format!("safekeeper-{node_id}"));
+        conf
+    }
+
+    /// Constructs a Safekeeper with the given node and tenant/timeline ID.
+    ///
+    /// TODO: we should support using in-memory storage, to measure non-IO costs. This would be
+    /// easier if SafeKeeper used trait objects for storage rather than generics. It's also not
+    /// currently possible to construct a timeline using non-file storage since StateSK only accepts
+    /// SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>.
+    pub async fn make_safekeeper(
+        &self,
+        node_id: NodeId,
+        ttid: TenantTimelineId,
+    ) -> anyhow::Result<SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>> {
+        let conf = self.make_conf(node_id);
+
+        let timeline_dir = get_timeline_dir(&conf, &ttid);
+        create_dir_all(&timeline_dir).await?;
+
+        let mut pstate = TimelinePersistentState::empty();
+        pstate.tenant_id = ttid.tenant_id;
+        pstate.timeline_id = ttid.timeline_id;
+
+        let wal = wal_storage::PhysicalStorage::new(&ttid, &timeline_dir, &pstate, conf.no_sync)?;
+        let ctrl =
+            control_file::FileStorage::create_new(&timeline_dir, pstate, conf.no_sync).await?;
+        let state = TimelineState::new(ctrl);
+        let mut safekeeper = SafeKeeper::new(state, wal, conf.my_id)?;
+
+        // Emulate an initial election.
+        safekeeper
+            .process_msg(&ProposerAcceptorMessage::Elected(ProposerElected {
+                term: 1,
+                start_streaming_at: Lsn(0),
+                term_history: TermHistory(vec![(1, Lsn(0)).into()]),
+                timeline_start_lsn: Lsn(0),
+            }))
+            .await?;
+
+        Ok(safekeeper)
+    }
+
+    /// Constructs a timeline, including a new Safekeeper with the given node ID, and spawns its
+    /// manager task.
+    pub async fn make_timeline(
+        &self,
+        node_id: NodeId,
+        ttid: TenantTimelineId,
+    ) -> anyhow::Result<Arc<Timeline>> {
+        let conf = self.make_conf(node_id);
+        let timeline_dir = get_timeline_dir(&conf, &ttid);
+        let remote_path = remote_timeline_path(&ttid)?;
+
+        let safekeeper = self.make_safekeeper(node_id, ttid).await?;
+        let shared_state = SharedState::new(StateSK::Loaded(safekeeper));
+
+        let timeline = Timeline::new(ttid, &timeline_dir, &remote_path, shared_state);
+        timeline.bootstrap(
+            &mut timeline.write_shared_state().await,
+            &conf,
+            Arc::new(TimelinesSet::default()), // ignored for now
+            RateLimiter::new(0, 0),
+        );
+        Ok(timeline)
+    }
+}

safekeeper/benches/receive_wal.rs

@@ -0,0 +1,341 @@
+//! WAL ingestion benchmarks.
+
+#[path = "benchutils.rs"]
+mod benchutils;
+
+use std::io::Write as _;
+
+use benchutils::Env;
+use camino_tempfile::tempfile;
+use criterion::{criterion_group, criterion_main, BatchSize, Bencher, Criterion};
+use itertools::Itertools as _;
+use postgres_ffi::v17::wal_generator::{LogicalMessageGenerator, WalGenerator};
+use safekeeper::receive_wal::{self, WalAcceptor};
+use safekeeper::safekeeper::{
+    AcceptorProposerMessage, AppendRequest, AppendRequestHeader, ProposerAcceptorMessage,
+};
+use tokio::io::AsyncWriteExt as _;
+use utils::id::{NodeId, TenantTimelineId};
+use utils::lsn::Lsn;
+
+const KB: usize = 1024;
+const MB: usize = 1024 * KB;
+const GB: usize = 1024 * MB;
+
+// Register benchmarks with Criterion.
+criterion_group!(
+    benches,
+    bench_process_msg,
+    bench_wal_acceptor,
+    bench_wal_acceptor_throughput,
+    bench_file_write
+);
+criterion_main!(benches);
+
+/// Benchmarks SafeKeeper::process_msg() as time per message and throughput. Each message is an
+/// AppendRequest with a single WAL record containing an XlLogicalMessage of varying size. When
+/// measuring throughput, only the logical message payload is considered, excluding
+/// segment/page/record headers.
+fn bench_process_msg(c: &mut Criterion) {
+    let mut g = c.benchmark_group("process_msg");
+    for fsync in [false, true] {
+        for commit in [false, true] {
+            for size in [8, KB, 8 * KB, 128 * KB, MB] {
+                // Kind of weird to change the group throughput per benchmark, but it's the only way
+                // to vary it per benchmark. It works.
+                g.throughput(criterion::Throughput::Bytes(size as u64));
+                g.bench_function(format!("fsync={fsync}/commit={commit}/size={size}"), |b| {
+                    run_bench(b, size, fsync, commit).unwrap()
+                });
+            }
+        }
+    }
+
+    // The actual benchmark. If commit is true, advance the commit LSN on every message.
+    fn run_bench(b: &mut Bencher, size: usize, fsync: bool, commit: bool) -> anyhow::Result<()> {
+        let runtime = tokio::runtime::Builder::new_current_thread() // single is fine, sync IO only
+            .enable_all()
+            .build()?;
+
+        // Construct the payload. The prefix counts towards the payload (including NUL terminator).
+        let prefix = c"p";
+        let prefixlen = prefix.to_bytes_with_nul().len();
+        assert!(size >= prefixlen);
+        let message = vec![0; size - prefixlen];
+
+        let walgen = &mut WalGenerator::new(LogicalMessageGenerator::new(prefix, &message));
+
+        // Set up the Safekeeper.
+        let env = Env::new(fsync)?;
+        let mut safekeeper =
+            runtime.block_on(env.make_safekeeper(NodeId(1), TenantTimelineId::generate()))?;
+
+        b.iter_batched_ref(
+            // Pre-construct WAL records and requests. Criterion will batch them.
+            || {
+                let (lsn, record) = walgen.next().expect("endless WAL");
+                ProposerAcceptorMessage::AppendRequest(AppendRequest {
+                    h: AppendRequestHeader {
+                        term: 1,
+                        term_start_lsn: Lsn(0),
+                        begin_lsn: lsn,
+                        end_lsn: lsn + record.len() as u64,
+                        commit_lsn: if commit { lsn } else { Lsn(0) }, // commit previous record
+                        truncate_lsn: Lsn(0),
+                        proposer_uuid: [0; 16],
+                    },
+                    wal_data: record,
+                })
+            },
+            // Benchmark message processing (time per message).
+            |msg| {
+                runtime
+                    .block_on(safekeeper.process_msg(msg))
+                    .expect("message failed")
+            },
+            BatchSize::SmallInput, // automatically determine a batch size
+        );
+        Ok(())
+    }
+}
+
+/// Benchmarks WalAcceptor message processing time by sending it a batch of WAL records and waiting
+/// for it to confirm that the last LSN has been flushed to storage. We pipeline a bunch of messages
+/// instead of measuring each individual message to amortize costs (e.g. fsync), which is more
+/// realistic. Records are XlLogicalMessage with a tiny payload (~64 bytes per record including
+/// headers). Records are pre-constructed to avoid skewing the benchmark.
+///
+/// TODO: add benchmarks with in-memory storage, see comment on `Env::make_safekeeper()`:
+fn bench_wal_acceptor(c: &mut Criterion) {
+    let mut g = c.benchmark_group("wal_acceptor");
+    for fsync in [false, true] {
+        for n in [1, 100, 10000] {
+            g.bench_function(format!("fsync={fsync}/n={n}"), |b| {
+                run_bench(b, n, fsync).unwrap()
+            });
+        }
+    }
+
+    /// The actual benchmark. n is the number of WAL records to send in a pipelined batch.
+    fn run_bench(b: &mut Bencher, n: usize, fsync: bool) -> anyhow::Result<()> {
+        let runtime = tokio::runtime::Runtime::new()?; // needs multithreaded
+
+        let env = Env::new(fsync)?;
+        let walgen = &mut WalGenerator::new(LogicalMessageGenerator::new(c"prefix", b"message"));
+
+        // Create buffered channels that can fit all requests, to avoid blocking on channels.
+        let (msg_tx, msg_rx) = tokio::sync::mpsc::channel(n);
+        let (reply_tx, mut reply_rx) = tokio::sync::mpsc::channel(n);
+
+        // Spawn the WalAcceptor task.
+        runtime.block_on(async {
+            // TODO: WalAcceptor doesn't actually need a full timeline, only
+            // Safekeeper::process_msg(). Consider decoupling them to simplify the setup.
+            let tli = env
+                .make_timeline(NodeId(1), TenantTimelineId::generate())
+                .await?
+                .wal_residence_guard()
+                .await?;
+            WalAcceptor::spawn(tli, msg_rx, reply_tx, Some(0));
+            anyhow::Ok(())
+        })?;
+
+        b.iter_batched(
+            // Pre-construct a batch of WAL records and requests.
+            || {
+                walgen
+                    .take(n)
+                    .map(|(lsn, record)| AppendRequest {
+                        h: AppendRequestHeader {
+                            term: 1,
+                            term_start_lsn: Lsn(0),
+                            begin_lsn: lsn,
+                            end_lsn: lsn + record.len() as u64,
+                            commit_lsn: Lsn(0),
+                            truncate_lsn: Lsn(0),
+                            proposer_uuid: [0; 16],
+                        },
+                        wal_data: record,
+                    })
+                    .collect_vec()
+            },
+            // Benchmark batch ingestion (time per batch).
+            |reqs| {
+                runtime.block_on(async {
+                    let final_lsn = reqs.last().unwrap().h.end_lsn;
+                    // Stuff all the messages into the buffered channel to pipeline them.
+                    for req in reqs {
+                        let msg = ProposerAcceptorMessage::AppendRequest(req);
+                        msg_tx.send(msg).await.expect("send failed");
+                    }
+                    // Wait for the last message to get flushed.
+                    while let Some(reply) = reply_rx.recv().await {
+                        if let AcceptorProposerMessage::AppendResponse(resp) = reply {
+                            if resp.flush_lsn >= final_lsn {
+                                return;
+                            }
+                        }
+                    }
+                    panic!("disconnected")
+                })
+            },
+            BatchSize::PerIteration, // only run one request batch at a time
+        );
+        Ok(())
+    }
+}
+
+/// Benchmarks WalAcceptor throughput by sending 1 GB of data with varying message sizes and waiting
+/// for the last LSN to be flushed to storage. Only the actual message payload counts towards
+/// throughput, headers are excluded and considered overhead. Records are XlLogicalMessage.
+///
+/// To avoid running out of memory, messages are constructed during the benchmark.
+fn bench_wal_acceptor_throughput(c: &mut Criterion) {
+    const VOLUME: usize = GB; // NB: excludes message/page/segment headers and padding
+
+    let mut g = c.benchmark_group("wal_acceptor_throughput");
+    g.sample_size(10);
+    g.throughput(criterion::Throughput::Bytes(VOLUME as u64));
+
+    for fsync in [false, true] {
+        for commit in [false, true] {
+            for size in [KB, 8 * KB, 128 * KB, MB] {
+                assert_eq!(VOLUME % size, 0, "volume must be divisible by size");
+                let count = VOLUME / size;
+                g.bench_function(format!("fsync={fsync}/commit={commit}/size={size}"), |b| {
+                    run_bench(b, count, size, fsync, commit).unwrap()
+                });
+            }
+        }
+    }
+
+    /// The actual benchmark. size is the payload size per message, count is the number of messages.
+    /// If commit is true, advance the commit LSN on each message.
+    fn run_bench(
+        b: &mut Bencher,
+        count: usize,
+        size: usize,
+        fsync: bool,
+        commit: bool,
+    ) -> anyhow::Result<()> {
+        let runtime = tokio::runtime::Runtime::new()?; // needs multithreaded
+
+        // Construct the payload. The prefix counts towards the payload (including NUL terminator).
+        let prefix = c"p";
+        let prefixlen = prefix.to_bytes_with_nul().len();
+        assert!(size >= prefixlen);
+        let message = vec![0; size - prefixlen];
+
+        let walgen = &mut WalGenerator::new(LogicalMessageGenerator::new(prefix, &message));
+
+        // Construct and spawn the WalAcceptor task.
+        let env = Env::new(fsync)?;
+
+        let (msg_tx, msg_rx) = tokio::sync::mpsc::channel(receive_wal::MSG_QUEUE_SIZE);
+        let (reply_tx, mut reply_rx) = tokio::sync::mpsc::channel(receive_wal::REPLY_QUEUE_SIZE);
+
+        runtime.block_on(async {
+            let tli = env
+                .make_timeline(NodeId(1), TenantTimelineId::generate())
+                .await?
+                .wal_residence_guard()
+                .await?;
+            WalAcceptor::spawn(tli, msg_rx, reply_tx, Some(0));
+            anyhow::Ok(())
+        })?;
+
+        // Ingest the WAL.
+        b.iter(|| {
+            runtime.block_on(async {
+                let reqgen = walgen.take(count).map(|(lsn, record)| AppendRequest {
+                    h: AppendRequestHeader {
+                        term: 1,
+                        term_start_lsn: Lsn(0),
+                        begin_lsn: lsn,
+                        end_lsn: lsn + record.len() as u64,
+                        commit_lsn: if commit { lsn } else { Lsn(0) }, // commit previous record
+                        truncate_lsn: Lsn(0),
+                        proposer_uuid: [0; 16],
+                    },
+                    wal_data: record,
+                });
+
+                // Send requests.
+                for req in reqgen {
+                    _ = reply_rx.try_recv(); // discard any replies, to avoid blocking
+                    let msg = ProposerAcceptorMessage::AppendRequest(req);
+                    msg_tx.send(msg).await.expect("send failed");
+                }
+
+                // Wait for last message to get flushed.
+                while let Some(reply) = reply_rx.recv().await {
+                    if let AcceptorProposerMessage::AppendResponse(resp) = reply {
+                        if resp.flush_lsn >= walgen.lsn {
+                            return;
+                        }
+                    }
+                }
+                panic!("disconnected")
+            })
+        });
+        Ok(())
+    }
+}
+
+/// Benchmarks OS write throughput by appending blocks of a given size to a file. This is intended
+/// to compare Tokio and stdlib writes, and give a baseline for optimal WAL throughput.
+fn bench_file_write(c: &mut Criterion) {
+    let mut g = c.benchmark_group("file_write");
+
+    for kind in ["stdlib", "tokio"] {
+        for fsync in [false, true] {
+            for size in [8, KB, 8 * KB, 128 * KB, MB] {
+                // Kind of weird to change the group throughput per benchmark, but it's the only way to
+                // vary it per benchmark. It works.
+                g.throughput(criterion::Throughput::Bytes(size as u64));
+                g.bench_function(
+                    format!("{kind}/fsync={fsync}/size={size}"),
+                    |b| match kind {
+                        "stdlib" => run_bench_stdlib(b, size, fsync).unwrap(),
+                        "tokio" => run_bench_tokio(b, size, fsync).unwrap(),
+                        name => panic!("unknown kind {name}"),
+                    },
+                );
+            }
+        }
+    }
+
+    fn run_bench_stdlib(b: &mut Bencher, size: usize, fsync: bool) -> anyhow::Result<()> {
+        let mut file = tempfile()?;
+        let buf = vec![0u8; size];
+
+        b.iter(|| {
+            file.write_all(&buf).unwrap();
+            file.flush().unwrap();
+            if fsync {
+                file.sync_data().unwrap();
+            }
+        });
+
+        Ok(())
+    }
+
+    fn run_bench_tokio(b: &mut Bencher, size: usize, fsync: bool) -> anyhow::Result<()> {
+        let runtime = tokio::runtime::Runtime::new()?; // needs multithreaded
+
+        let mut file = tokio::fs::File::from_std(tempfile()?);
+        let buf = vec![0u8; size];
+
+        b.iter(|| {
+            runtime.block_on(async {
+                file.write_all(&buf).await.unwrap();
+                file.flush().await.unwrap();
+                if fsync {
+                    file.sync_data().await.unwrap();
+                }
+            })
+        });
+
+        Ok(())
+    }
+}

safekeeper/src/auth.rs

@@ -20,7 +20,8 @@ pub fn check_permission(claims: &Claims, tenant_id: Option<TenantId>) -> Result<
             | Scope::PageServerApi
             | Scope::GenerationsApi
             | Scope::Infra
-            | Scope::Scrubber,
+            | Scope::Scrubber
+            | Scope::ControllerPeer,
             _,
         ) => Err(AuthError(
             format!(

safekeeper/src/lib.rs

@@ -112,9 +112,7 @@ impl SafeKeeperConf {
 }
 
 impl SafeKeeperConf {
-    #[cfg(test)]
-    #[allow(unused)]
-    fn dummy() -> Self {
+    pub fn dummy() -> Self {
         SafeKeeperConf {
             workdir: Utf8PathBuf::from("./"),
             no_sync: false,

safekeeper/src/metrics.rs

@@ -55,7 +55,7 @@ pub static WRITE_WAL_SECONDS: Lazy<Histogram> = Lazy::new(|| {
 pub static FLUSH_WAL_SECONDS: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "safekeeper_flush_wal_seconds",
-        "Seconds spent syncing WAL to a disk",
+        "Seconds spent syncing WAL to a disk (excluding segment initialization)",
         DISK_FSYNC_SECONDS_BUCKETS.to_vec()
     )
     .expect("Failed to register safekeeper_flush_wal_seconds histogram")

safekeeper/src/safekeeper.rs

@@ -979,7 +979,8 @@ where
             self.wal_store.flush_wal().await?;
         }
 
-        // Update commit_lsn.
+        // Update commit_lsn. It will be flushed to the control file regularly by the timeline
+        // manager, off of the WAL ingest hot path.
         if msg.h.commit_lsn != Lsn(0) {
             self.update_commit_lsn(msg.h.commit_lsn).await?;
         }
@@ -992,15 +993,6 @@ where
         self.state.inmem.peer_horizon_lsn =
             max(self.state.inmem.peer_horizon_lsn, msg.h.truncate_lsn);
 
-        // Update truncate and commit LSN in control file.
-        // To avoid negative impact on performance of extra fsync, do it only
-        // when commit_lsn delta exceeds WAL segment size.
-        if self.state.commit_lsn + (self.state.server.wal_seg_size as u64)
-            < self.state.inmem.commit_lsn
-        {
-            self.state.flush().await?;
-        }
-
         trace!(
             "processed AppendRequest of len {}, begin_lsn={}, end_lsn={:?}, commit_lsn={:?}, truncate_lsn={:?}, flushed={:?}",
             msg.wal_data.len(),

safekeeper/src/state.rs

@@ -4,6 +4,7 @@
 use std::{cmp::max, ops::Deref};
 
 use anyhow::{bail, Result};
+use postgres_ffi::WAL_SEGMENT_SIZE;
 use safekeeper_api::models::TimelineTermBumpResponse;
 use serde::{Deserialize, Serialize};
 use utils::{
@@ -138,14 +139,13 @@ impl TimelinePersistentState {
         })
     }
 
-    #[cfg(test)]
     pub fn empty() -> Self {
         TimelinePersistentState::new(
             &TenantTimelineId::empty(),
             ServerInfo {
                 pg_version: 170000, /* Postgres server version (major * 10000) */
                 system_id: 0,       /* Postgres system identifier */
-                wal_seg_size: 16 * 1024 * 1024,
+                wal_seg_size: WAL_SEGMENT_SIZE as u32,
             },
             vec![],
             Lsn::INVALID,

safekeeper/src/timeline.rs

@@ -2,7 +2,7 @@
 //! to glue together SafeKeeper and all other background services.
 
 use anyhow::{anyhow, bail, Result};
-use camino::Utf8PathBuf;
+use camino::{Utf8Path, Utf8PathBuf};
 use remote_storage::RemotePath;
 use safekeeper_api::models::TimelineTermBumpResponse;
 use serde::{Deserialize, Serialize};
@@ -108,16 +108,11 @@ pub type ReadGuardSharedState<'a> = RwLockReadGuard<'a, SharedState>;
 pub struct WriteGuardSharedState<'a> {
     tli: Arc<Timeline>,
     guard: RwLockWriteGuard<'a, SharedState>,
-    skip_update: bool,
 }
 
 impl<'a> WriteGuardSharedState<'a> {
     fn new(tli: Arc<Timeline>, guard: RwLockWriteGuard<'a, SharedState>) -> Self {
-        WriteGuardSharedState {
-            tli,
-            guard,
-            skip_update: false,
-        }
+        WriteGuardSharedState { tli, guard }
     }
 }
 
@@ -159,12 +154,10 @@ impl Drop for WriteGuardSharedState<'_> {
             }
         });
 
-        if !self.skip_update {
-            // send notification about shared state update
-            self.tli.shared_state_version_tx.send_modify(|old| {
-                *old += 1;
-            });
-        }
+        // send notification about shared state update
+        self.tli.shared_state_version_tx.send_modify(|old| {
+            *old += 1;
+        });
     }
 }
 
@@ -325,8 +318,17 @@ pub struct SharedState {
 }
 
 impl SharedState {
+    /// Creates a new SharedState.
+    pub fn new(sk: StateSK) -> Self {
+        Self {
+            sk,
+            peers_info: PeersInfo(vec![]),
+            wal_removal_on_hold: false,
+        }
+    }
+
     /// Restore SharedState from control file. If file doesn't exist, bails out.
-    fn restore(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Result<Self> {
+    pub fn restore(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Result<Self> {
         let timeline_dir = get_timeline_dir(conf, ttid);
         let control_store = control_file::FileStorage::restore_new(&timeline_dir, conf.no_sync)?;
         if control_store.server.wal_seg_size == 0 {
@@ -352,11 +354,7 @@ impl SharedState {
             }
         };
 
-        Ok(Self {
-            sk,
-            peers_info: PeersInfo(vec![]),
-            wal_removal_on_hold: false,
-        })
+        Ok(Self::new(sk))
     }
 
     pub(crate) fn get_wal_seg_size(&self) -> usize {
@@ -480,11 +478,13 @@ pub struct Timeline {
 }
 
 impl Timeline {
-    /// Load existing timeline from disk.
-    pub fn load_timeline(conf: &SafeKeeperConf, ttid: TenantTimelineId) -> Result<Arc<Timeline>> {
-        let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
-
-        let shared_state = SharedState::restore(conf, &ttid)?;
+    /// Constructs a new timeline.
+    pub fn new(
+        ttid: TenantTimelineId,
+        timeline_dir: &Utf8Path,
+        remote_path: &RemotePath,
+        shared_state: SharedState,
+    ) -> Arc<Self> {
         let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
             watch::channel(shared_state.sk.state().commit_lsn);
         let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
@@ -494,10 +494,11 @@ impl Timeline {
         let (shared_state_version_tx, shared_state_version_rx) = watch::channel(0);
 
         let walreceivers = WalReceivers::new();
-        let remote_path = remote_timeline_path(&ttid)?;
-        Ok(Arc::new(Timeline {
+
+        Arc::new(Self {
             ttid,
-            remote_path,
+            remote_path: remote_path.to_owned(),
+            timeline_dir: timeline_dir.to_owned(),
             commit_lsn_watch_tx,
             commit_lsn_watch_rx,
             term_flush_lsn_watch_tx,
@@ -508,13 +509,28 @@ impl Timeline {
             walsenders: WalSenders::new(walreceivers.clone()),
             walreceivers,
             cancel: CancellationToken::default(),
-            timeline_dir: get_timeline_dir(conf, &ttid),
             manager_ctl: ManagerCtl::new(),
             broker_active: AtomicBool::new(false),
             wal_backup_active: AtomicBool::new(false),
             last_removed_segno: AtomicU64::new(0),
             mgr_status: AtomicStatus::new(),
-        }))
+        })
+    }
+
+    /// Load existing timeline from disk.
+    pub fn load_timeline(conf: &SafeKeeperConf, ttid: TenantTimelineId) -> Result<Arc<Timeline>> {
+        let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
+
+        let shared_state = SharedState::restore(conf, &ttid)?;
+        let timeline_dir = get_timeline_dir(conf, &ttid);
+        let remote_path = remote_timeline_path(&ttid)?;
+
+        Ok(Timeline::new(
+            ttid,
+            &timeline_dir,
+            &remote_path,
+            shared_state,
+        ))
     }
 
     /// Initialize fresh timeline on disk and start background tasks. If init
@@ -1128,13 +1144,13 @@ async fn delete_dir(path: &Utf8PathBuf) -> Result<bool> {
 
 /// Get a path to the tenant directory. If you just need to get a timeline directory,
 /// use WalResidentTimeline::get_timeline_dir instead.
-pub(crate) fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
+pub fn get_tenant_dir(conf: &SafeKeeperConf, tenant_id: &TenantId) -> Utf8PathBuf {
     conf.workdir.join(tenant_id.to_string())
 }
 
 /// Get a path to the timeline directory. If you need to read WAL files from disk,
 /// use WalResidentTimeline::get_timeline_dir instead. This function does not check
 /// timeline eviction status and WAL files might not be present on disk.
-pub(crate) fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
+pub fn get_timeline_dir(conf: &SafeKeeperConf, ttid: &TenantTimelineId) -> Utf8PathBuf {
     get_tenant_dir(conf, &ttid.tenant_id).join(ttid.timeline_id.to_string())
 }

safekeeper/src/timeline_manager.rs

@@ -515,7 +515,12 @@ impl Manager {
             return;
         }
 
-        if state.cfile_last_persist_at.elapsed() > self.conf.control_file_save_interval {
+        if state.cfile_last_persist_at.elapsed() > self.conf.control_file_save_interval
+            // If the control file's commit_lsn lags more than one segment behind the current
+            // commit_lsn, flush immediately to limit recovery time in case of a crash. We don't do
+            // this on the WAL ingest hot path since it incurs fsync latency.
+            || state.commit_lsn.saturating_sub(state.cfile_commit_lsn).0 >= self.wal_seg_size as u64
+        {
             let mut write_guard = self.tli.write_shared_state().await;
             // it should be done in the background because it blocks manager task, but flush() should
             // be fast enough not to be a problem now

safekeeper/src/wal_storage.rs

@@ -31,7 +31,6 @@ use crate::state::TimelinePersistentState;
 use crate::wal_backup::{read_object, remote_timeline_path};
 use postgres_ffi::waldecoder::WalStreamDecoder;
 use postgres_ffi::XLogFileName;
-use postgres_ffi::XLOG_BLCKSZ;
 use pq_proto::SystemId;
 use utils::{id::TenantTimelineId, lsn::Lsn};
 
@@ -223,6 +222,15 @@ impl PhysicalStorage {
         )
     }
 
+    /// Call fsync if config requires so.
+    async fn fsync_file(&mut self, file: &File) -> Result<()> {
+        if !self.no_sync {
+            self.metrics
+                .observe_flush_seconds(time_io_closure(file.sync_all()).await?);
+        }
+        Ok(())
+    }
+
     /// Call fdatasync if config requires so.
     async fn fdatasync_file(&mut self, file: &File) -> Result<()> {
         if !self.no_sync {
@@ -249,6 +257,9 @@ impl PhysicalStorage {
             // Try to open existing partial file
             Ok((file, true))
         } else {
+            let _timer = WAL_STORAGE_OPERATION_SECONDS
+                .with_label_values(&["initialize_segment"])
+                .start_timer();
             // Create and fill new partial file
             //
             // We're using fdatasync during WAL writing, so file size must not
@@ -256,14 +267,16 @@ impl PhysicalStorage {
             // half initialized segment, first bake it under tmp filename and
             // then rename.
             let tmp_path = self.timeline_dir.join("waltmp");
-            let mut file = File::create(&tmp_path)
+            let file = File::create(&tmp_path)
                 .await
                 .with_context(|| format!("Failed to open tmp wal file {:?}", &tmp_path))?;
 
-            write_zeroes(&mut file, self.wal_seg_size).await?;
+            fail::fail_point!("sk-zero-segment", |_| {
+                info!("sk-zero-segment failpoint hit");
+                Err(anyhow::anyhow!("failpoint: sk-zero-segment"))
+            });
+            file.set_len(self.wal_seg_size as u64).await?;
 
-            // Note: this doesn't get into observe_flush_seconds metric. But
-            // segment init should be separate metric, if any.
             if let Err(e) = durable_rename(&tmp_path, &wal_file_partial_path, !self.no_sync).await {
                 // Probably rename succeeded, but fsync of it failed. Remove
                 // the file then to avoid using it.
@@ -486,12 +499,12 @@ impl Storage for PhysicalStorage {
         // Remove all segments after the given LSN.
         remove_segments_from_disk(&self.timeline_dir, self.wal_seg_size, |x| x > segno).await?;
 
-        let (mut file, is_partial) = self.open_or_create(segno).await?;
+        let (file, is_partial) = self.open_or_create(segno).await?;
 
         // Fill end with zeroes
-        file.seek(SeekFrom::Start(xlogoff as u64)).await?;
-        write_zeroes(&mut file, self.wal_seg_size - xlogoff).await?;
-        self.fdatasync_file(&file).await?;
+        file.set_len(xlogoff as u64).await?;
+        file.set_len(self.wal_seg_size as u64).await?;
+        self.fsync_file(&file).await?;
 
         if !is_partial {
             // Make segment partial once again
@@ -751,25 +764,6 @@ impl WalReader {
     }
 }
 
-/// Zero block for filling created WAL segments.
-const ZERO_BLOCK: &[u8] = &[0u8; XLOG_BLCKSZ];
-
-/// Helper for filling file with zeroes.
-async fn write_zeroes(file: &mut File, mut count: usize) -> Result<()> {
-    fail::fail_point!("sk-write-zeroes", |_| {
-        info!("write_zeroes hit failpoint");
-        Err(anyhow::anyhow!("failpoint: sk-write-zeroes"))
-    });
-
-    while count >= XLOG_BLCKSZ {
-        file.write_all(ZERO_BLOCK).await?;
-        count -= XLOG_BLCKSZ;
-    }
-    file.write_all(&ZERO_BLOCK[0..count]).await?;
-    file.flush().await?;
-    Ok(())
-}
-
 /// Helper function for opening WAL segment `segno` in `dir`. Returns file and
 /// whether it is .partial.
 pub(crate) async fn open_wal_file(

safekeeper/tests/walproposer_sim/walproposer_disk.rs

@@ -1,7 +1,7 @@
 use std::{ffi::CStr, sync::Arc};
 
 use parking_lot::{Mutex, MutexGuard};
-use postgres_ffi::v16::wal_generator::WalGenerator;
+use postgres_ffi::v16::wal_generator::{LogicalMessageGenerator, WalGenerator};
 use utils::lsn::Lsn;
 
 use super::block_storage::BlockStorage;
@@ -18,7 +18,7 @@ impl DiskWalProposer {
                 internal_available_lsn: Lsn(0),
                 prev_lsn: Lsn(0),
                 disk: BlockStorage::new(),
-                wal_generator: WalGenerator::new(),
+                wal_generator: WalGenerator::new(LogicalMessageGenerator::new(c"", &[])),
             }),
         })
     }
@@ -36,7 +36,7 @@ pub struct State {
     // actual WAL storage
     disk: BlockStorage,
     // WAL record generator
-    wal_generator: WalGenerator,
+    wal_generator: WalGenerator<LogicalMessageGenerator>,
 }
 
 impl State {
@@ -64,7 +64,7 @@ impl State {
 
     /// Inserts a logical record in the WAL at the current LSN.
     pub fn insert_logical_message(&mut self, prefix: &CStr, msg: &[u8]) {
-        let record = self.wal_generator.generate_logical_message(prefix, msg);
+        let (_, record) = self.wal_generator.append_logical_message(prefix, msg);
         self.disk.write(self.internal_available_lsn.into(), &record);
         self.prev_lsn = self.internal_available_lsn;
         self.internal_available_lsn += record.len() as u64;

storage_controller/src/http.rs

@@ -1033,7 +1033,7 @@ async fn handle_update_preferred_azs(req: Request<Body>) -> Result<Response<Body
 }
 
 async fn handle_step_down(req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    check_permissions(&req, Scope::Admin)?;
+    check_permissions(&req, Scope::ControllerPeer)?;
 
     let req = match maybe_forward(req).await {
         ForwardOutcome::Forwarded(res) => {

test_runner/fixtures/auth_tokens.py

@@ -45,3 +45,4 @@ class TokenScope(str, Enum):
     SAFEKEEPER_DATA = "safekeeperdata"
     TENANT = "tenant"
     SCRUBBER = "scrubber"
+    INFRA = "infra"

test_runner/fixtures/benchmark_fixture.py

@@ -80,7 +80,13 @@ class PgBenchRunResult:
     ):
         stdout_lines = stdout.splitlines()
 
+        number_of_clients = 0
+        number_of_threads = 0
+        number_of_transactions_actually_processed = 0
+        latency_average = 0.0
         latency_stddev = None
+        tps = 0.0
+        scale = 0
 
         # we know significant parts of these values from test input
         # but to be precise take them from output

test_runner/fixtures/compare_fixtures.py

@@ -8,7 +8,7 @@ from contextlib import _GeneratorContextManager, contextmanager
 
 # Type-related stuff
 from pathlib import Path
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, final
 
 import pytest
 from _pytest.fixtures import FixtureRequest
@@ -70,12 +70,12 @@ class PgCompare(ABC):
 
     @contextmanager
     @abstractmethod
-    def record_pageserver_writes(self, out_name: str):
+    def record_pageserver_writes(self, out_name: str) -> Iterator[None]:
         pass
 
     @contextmanager
     @abstractmethod
-    def record_duration(self, out_name: str):
+    def record_duration(self, out_name: str) -> Iterator[None]:
         pass
 
     @contextmanager
@@ -105,6 +105,7 @@ class PgCompare(ABC):
         return results
 
 
+@final
 class NeonCompare(PgCompare):
     """PgCompare interface for the neon stack."""
 
@@ -206,6 +207,7 @@ class NeonCompare(PgCompare):
         return self.zenbenchmark.record_duration(out_name)
 
 
+@final
 class VanillaCompare(PgCompare):
     """PgCompare interface for vanilla postgres."""
 
@@ -271,6 +273,7 @@ class VanillaCompare(PgCompare):
         return self.zenbenchmark.record_duration(out_name)
 
 
+@final
 class RemoteCompare(PgCompare):
     """PgCompare interface for a remote postgres instance."""
 

test_runner/fixtures/endpoint/http.py

@@ -46,3 +46,8 @@ class EndpointHttpClient(requests.Session):
         )
         res.raise_for_status()
         return res.json()
+
+    def metrics(self) -> str:
+        res = self.get(f"http://localhost:{self.port}/metrics")
+        res.raise_for_status()
+        return res.text

test_runner/fixtures/h2server.py

@@ -4,11 +4,14 @@ https://python-hyper.org/projects/hyper-h2/en/stable/asyncio-example.html
 auth-broker -> local-proxy needs a h2 connection, so we need a h2 server :)
 """
 
+from __future__ import annotations
+
 import asyncio
 import collections
 import io
 import json
 from collections.abc import AsyncIterable
+from typing import TYPE_CHECKING, final
 
 import pytest_asyncio
 from h2.config import H2Configuration
@@ -25,34 +28,45 @@ from h2.events import (
 )
 from h2.exceptions import ProtocolError, StreamClosedError
 from h2.settings import SettingCodes
+from typing_extensions import override
+
+if TYPE_CHECKING:
+    from typing import Any, Optional
+
 
 RequestData = collections.namedtuple("RequestData", ["headers", "data"])
 
 
+@final
 class H2Server:
-    def __init__(self, host, port) -> None:
+    def __init__(self, host: str, port: int) -> None:
         self.host = host
         self.port = port
 
 
+@final
 class H2Protocol(asyncio.Protocol):
     def __init__(self):
         config = H2Configuration(client_side=False, header_encoding="utf-8")
         self.conn = H2Connection(config=config)
-        self.transport = None
-        self.stream_data = {}
-        self.flow_control_futures = {}
+        self.transport: Optional[asyncio.Transport] = None
+        self.stream_data: dict[int, RequestData] = {}
+        self.flow_control_futures: dict[int, asyncio.Future[Any]] = {}
 
-    def connection_made(self, transport: asyncio.Transport):  # type: ignore[override]
+    @override
+    def connection_made(self, transport: asyncio.BaseTransport):
+        assert isinstance(transport, asyncio.Transport)
         self.transport = transport
         self.conn.initiate_connection()
         self.transport.write(self.conn.data_to_send())
 
-    def connection_lost(self, _exc):
+    @override
+    def connection_lost(self, exc: Optional[Exception]):
         for future in self.flow_control_futures.values():
             future.cancel()
         self.flow_control_futures = {}
 
+    @override
     def data_received(self, data: bytes):
         assert self.transport is not None
         try:
@@ -77,7 +91,7 @@ class H2Protocol(asyncio.Protocol):
                     self.window_updated(event.stream_id, event.delta)
                 elif isinstance(event, RemoteSettingsChanged):
                     if SettingCodes.INITIAL_WINDOW_SIZE in event.changed_settings:
-                        self.window_updated(None, 0)
+                        self.window_updated(0, 0)
 
                 self.transport.write(self.conn.data_to_send())
 
@@ -123,7 +137,7 @@ class H2Protocol(asyncio.Protocol):
         else:
             stream_data.data.write(data)
 
-    def stream_reset(self, stream_id):
+    def stream_reset(self, stream_id: int):
         """
         A stream reset was sent. Stop sending data.
         """
@@ -131,7 +145,7 @@ class H2Protocol(asyncio.Protocol):
             future = self.flow_control_futures.pop(stream_id)
             future.cancel()
 
-    async def send_data(self, data, stream_id):
+    async def send_data(self, data: bytes, stream_id: int):
         """
         Send data according to the flow control rules.
         """
@@ -161,7 +175,7 @@ class H2Protocol(asyncio.Protocol):
             self.transport.write(self.conn.data_to_send())
             data = data[chunk_size:]
 
-    async def wait_for_flow_control(self, stream_id):
+    async def wait_for_flow_control(self, stream_id: int):
         """
         Waits for a Future that fires when the flow control window is opened.
         """
@@ -169,7 +183,7 @@ class H2Protocol(asyncio.Protocol):
         self.flow_control_futures[stream_id] = f
         await f
 
-    def window_updated(self, stream_id, delta):
+    def window_updated(self, stream_id: int, delta):
         """
         A window update frame was received. Unblock some number of flow control
         Futures.

test_runner/fixtures/neon_api.py

@@ -5,6 +5,8 @@ from typing import TYPE_CHECKING, cast, final
 
 import requests
 
+from fixtures.log_helper import log
+
 if TYPE_CHECKING:
     from typing import Any, Literal, Optional
 
@@ -30,7 +32,11 @@ class NeonAPI:
             kwargs["headers"] = {}
         kwargs["headers"]["Authorization"] = f"Bearer {self.__neon_api_key}"
 
-        return requests.request(method, f"{self.__neon_api_base_url}{endpoint}", **kwargs)
+        resp = requests.request(method, f"{self.__neon_api_base_url}{endpoint}", **kwargs)
+        log.debug("%s %s returned a %d: %s", method, endpoint, resp.status_code, resp.text)
+        resp.raise_for_status()
+
+        return resp
 
     def create_project(
         self,
@@ -66,8 +72,6 @@ class NeonAPI:
             json=data,
         )
 
-        assert resp.status_code == 201
-
         return cast("dict[str, Any]", resp.json())
 
     def get_project_details(self, project_id: str) -> dict[str, Any]:
@@ -79,7 +83,7 @@ class NeonAPI:
                 "Content-Type": "application/json",
             },
         )
-        assert resp.status_code == 200
+
         return cast("dict[str, Any]", resp.json())
 
     def delete_project(
@@ -95,8 +99,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def start_endpoint(
@@ -112,8 +114,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def suspend_endpoint(
@@ -129,8 +129,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def restart_endpoint(
@@ -146,8 +144,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def create_endpoint(
@@ -178,8 +174,6 @@ class NeonAPI:
             json=data,
         )
 
-        assert resp.status_code == 201
-
         return cast("dict[str, Any]", resp.json())
 
     def get_connection_uri(
@@ -206,8 +200,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def get_branches(self, project_id: str) -> dict[str, Any]:
@@ -219,8 +211,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def get_endpoints(self, project_id: str) -> dict[str, Any]:
@@ -232,8 +222,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def get_operations(self, project_id: str) -> dict[str, Any]:
@@ -246,8 +234,6 @@ class NeonAPI:
             },
         )
 
-        assert resp.status_code == 200
-
         return cast("dict[str, Any]", resp.json())
 
     def wait_for_operation_to_finish(self, project_id: str):

test_runner/fixtures/neon_fixtures.py

@@ -286,7 +286,7 @@ class PgProtocol:
         return self.safe_psql_many([query], **kwargs)[0]
 
     def safe_psql_many(
-        self, queries: Iterable[str], log_query=True, **kwargs: Any
+        self, queries: Iterable[str], log_query: bool = True, **kwargs: Any
     ) -> list[list[tuple[Any, ...]]]:
         """
         Execute queries against the node and return all rows.
@@ -306,7 +306,7 @@ class PgProtocol:
                         result.append(cur.fetchall())
         return result
 
-    def safe_psql_scalar(self, query, log_query=True) -> Any:
+    def safe_psql_scalar(self, query: str, log_query: bool = True) -> Any:
         """
         Execute query returning single row with single column.
         """
@@ -1065,6 +1065,9 @@ class NeonEnv:
                 "http_auth_type": http_auth_type,
                 # Default which can be overriden with `NeonEnvBuilder.pageserver_config_override`
                 "availability_zone": "us-east-2a",
+                # Disable pageserver disk syncs in tests: when running tests concurrently, this avoids
+                # the pageserver taking a long time to start up due to syncfs flushing other tests' data
+                "no_sync": True,
             }
             if self.pageserver_virtual_file_io_engine is not None:
                 ps_cfg["virtual_file_io_engine"] = self.pageserver_virtual_file_io_engine
@@ -1782,7 +1785,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         self.request(
             "PUT",
             f"{self.api}/control/v1/node/{node_id}/drain",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
 
     def cancel_node_drain(self, node_id):
@@ -1790,7 +1793,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         self.request(
             "DELETE",
             f"{self.api}/control/v1/node/{node_id}/drain",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
 
     def node_fill(self, node_id):
@@ -1798,7 +1801,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         self.request(
             "PUT",
             f"{self.api}/control/v1/node/{node_id}/fill",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
 
     def cancel_node_fill(self, node_id):
@@ -1806,14 +1809,14 @@ class NeonStorageController(MetricsGetter, LogUtils):
         self.request(
             "DELETE",
             f"{self.api}/control/v1/node/{node_id}/fill",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
 
     def node_status(self, node_id):
         response = self.request(
             "GET",
             f"{self.api}/control/v1/node/{node_id}",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
         return response.json()
 
@@ -1829,7 +1832,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         response = self.request(
             "GET",
             f"{self.api}/control/v1/node",
-            headers=self.headers(TokenScope.ADMIN),
+            headers=self.headers(TokenScope.INFRA),
         )
         return response.json()
 
@@ -1857,7 +1860,7 @@ class NeonStorageController(MetricsGetter, LogUtils):
         shard_count: Optional[int] = None,
         shard_stripe_size: Optional[int] = None,
         tenant_config: Optional[dict[Any, Any]] = None,
-        placement_policy: Optional[Union[dict[Any, Any] | str]] = None,
+        placement_policy: Optional[Union[dict[Any, Any], str]] = None,
     ):
         """
         Use this rather than pageserver_api() when you need to include shard parameters
@@ -4177,9 +4180,15 @@ class Safekeeper(LogUtils):
         return self
 
     def assert_no_errors(self):
-        assert not self.log_contains("manager task finished prematurely")
-        assert not self.log_contains("error while acquiring WalResidentTimeline guard")
-        assert not self.log_contains("timeout while acquiring WalResidentTimeline guard")
+        not_allowed = [
+            "manager task finished prematurely",
+            "error while acquiring WalResidentTimeline guard",
+            "timeout while acquiring WalResidentTimeline guard",
+            "invalid xlog page header:",
+            "WAL record crc mismatch at",
+        ]
+        for na in not_allowed:
+            assert not self.log_contains(na)
 
     def append_logical_message(
         self, tenant_id: TenantId, timeline_id: TimelineId, request: dict[str, Any]

test_runner/fixtures/pageserver/allowed_errors.py

@@ -93,6 +93,8 @@ DEFAULT_PAGESERVER_ALLOWED_ERRORS = (
     ".*WARN.*path=/v1/utilization .*request was dropped before completing",
     # Can happen during shutdown
     ".*scheduling deletion on drop failed: queue is in state Stopped.*",
+    # Too many frozen layers error is normal during intensive benchmarks
+    ".*too many frozen layers.*",
 )
 
 

test_runner/fixtures/pageserver/http.py

@@ -316,7 +316,7 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
     def tenant_location_conf(
         self,
         tenant_id: Union[TenantId, TenantShardId],
-        location_conf=dict[str, Any],
+        location_conf: dict[str, Any],
         flush_ms=None,
         lazy: Optional[bool] = None,
     ):

test_runner/fixtures/pageserver/utils.py

@@ -56,6 +56,8 @@ def wait_for_upload(
     lsn: Lsn,
 ):
     """waits for local timeline upload up to specified lsn"""
+
+    current_lsn = Lsn(0)
     for i in range(20):
         current_lsn = remote_consistent_lsn(pageserver_http, tenant, timeline)
         if current_lsn >= lsn:
@@ -203,6 +205,8 @@ def wait_for_last_record_lsn(
     lsn: Lsn,
 ) -> Lsn:
     """waits for pageserver to catch up to a certain lsn, returns the last observed lsn."""
+
+    current_lsn = Lsn(0)
     for i in range(1000):
         current_lsn = last_record_lsn(pageserver_http, tenant, timeline)
         if current_lsn >= lsn:

test_runner/fixtures/paths.py

@@ -112,7 +112,7 @@ def compatibility_snapshot_dir() -> Iterator[Path]:
 
 
 @pytest.fixture(scope="session")
-def compatibility_neon_binpath() -> Optional[Iterator[Path]]:
+def compatibility_neon_binpath() -> Iterator[Optional[Path]]:
     if os.getenv("REMOTE_ENV"):
         return
     comp_binpath = None
@@ -133,7 +133,7 @@ def pg_distrib_dir(base_dir: Path) -> Iterator[Path]:
 
 
 @pytest.fixture(scope="session")
-def compatibility_pg_distrib_dir() -> Optional[Iterator[Path]]:
+def compatibility_pg_distrib_dir() -> Iterator[Optional[Path]]:
     compat_distrib_dir = None
     if env_compat_postgres_bin := os.environ.get("COMPATIBILITY_POSTGRES_DISTRIB_DIR"):
         compat_distrib_dir = Path(env_compat_postgres_bin).resolve()

test_runner/fixtures/pg_version.py

@@ -1,10 +1,8 @@
 from __future__ import annotations
 
 import enum
-import os
 from typing import TYPE_CHECKING
 
-import pytest
 from typing_extensions import override
 
 if TYPE_CHECKING:
@@ -18,12 +16,15 @@ This fixture is used to determine which version of Postgres to use for tests.
 
 # Inherit PgVersion from str rather than int to make it easier to pass as a command-line argument
 # TODO: use enum.StrEnum for Python >= 3.11
-@enum.unique
 class PgVersion(str, enum.Enum):
     V14 = "14"
     V15 = "15"
     V16 = "16"
     V17 = "17"
+
+    # Default Postgres Version for tests that don't really depend on Postgres itself
+    DEFAULT = V16
+
     # Instead of making version an optional parameter in methods, we can use this fake entry
     # to explicitly rely on the default server version (could be different from pg_version fixture value)
     NOT_SET = "<-POSTRGRES VERSION IS NOT SET->"
@@ -59,27 +60,3 @@ class PgVersion(str, enum.Enum):
         # Make mypy happy
         # See https://github.com/python/mypy/issues/3974
         return None
-
-
-DEFAULT_VERSION: PgVersion = PgVersion.V16
-
-
-def skip_on_postgres(version: PgVersion, reason: str):
-    return pytest.mark.skipif(
-        PgVersion(os.environ.get("DEFAULT_PG_VERSION", DEFAULT_VERSION)) is version,
-        reason=reason,
-    )
-
-
-def xfail_on_postgres(version: PgVersion, reason: str):
-    return pytest.mark.xfail(
-        PgVersion(os.environ.get("DEFAULT_PG_VERSION", DEFAULT_VERSION)) is version,
-        reason=reason,
-    )
-
-
-def run_only_on_default_postgres(reason: str):
-    return pytest.mark.skipif(
-        PgVersion(os.environ.get("DEFAULT_PG_VERSION", DEFAULT_VERSION)) is not DEFAULT_VERSION,
-        reason=reason,
-    )

test_runner/fixtures/utils.py

@@ -25,6 +25,7 @@ from fixtures.pageserver.common_types import (
     parse_delta_layer,
     parse_image_layer,
 )
+from fixtures.pg_version import PgVersion
 
 if TYPE_CHECKING:
     from collections.abc import Iterable
@@ -37,6 +38,7 @@ if TYPE_CHECKING:
 
 
 Fn = TypeVar("Fn", bound=Callable[..., Any])
+
 COMPONENT_BINARIES = {
     "storage_controller": ("storage_controller",),
     "storage_broker": ("storage_broker",),
@@ -519,7 +521,7 @@ def assert_pageserver_backups_equal(left: Path, right: Path, skip_files: set[str
     This is essentially:
 
     lines=$(comm -3 \
-        <(mkdir left && cd left && tar xf "$left" && find . -type f -print0 | xargs sha256sum | sort -k2) \
+        <(mkdir left  && cd left  && tar xf "$left"  && find . -type f -print0 | xargs sha256sum | sort -k2) \
         <(mkdir right && cd right && tar xf "$right" && find . -type f -print0 | xargs sha256sum | sort -k2) \
         | wc -l)
     [ "$lines" = "0" ]
@@ -643,3 +645,40 @@ def allpairs_versions():
         )
         ids.append(f"combination_{''.join(cur_id)}")
     return {"argnames": "combination", "argvalues": tuple(argvalues), "ids": ids}
+
+
+def skip_on_postgres(version: PgVersion, reason: str):
+    return pytest.mark.skipif(
+        PgVersion(os.getenv("DEFAULT_PG_VERSION", PgVersion.DEFAULT)) is version,
+        reason=reason,
+    )
+
+
+def xfail_on_postgres(version: PgVersion, reason: str):
+    return pytest.mark.xfail(
+        PgVersion(os.getenv("DEFAULT_PG_VERSION", PgVersion.DEFAULT)) is version,
+        reason=reason,
+    )
+
+
+def run_only_on_default_postgres(reason: str):
+    return pytest.mark.skipif(
+        PgVersion(os.getenv("DEFAULT_PG_VERSION", PgVersion.DEFAULT)) is not PgVersion.DEFAULT,
+        reason=reason,
+    )
+
+
+def skip_in_debug_build(reason: str):
+    return pytest.mark.skipif(
+        os.getenv("BUILD_TYPE", "debug") == "debug",
+        reason=reason,
+    )
+
+
+def skip_on_ci(reason: str):
+    # `CI` variable is always set to `true` on GitHub
+    # Ref: https://docs.github.com/en/actions/writing-workflows/choosing-what-your-workflow-does/store-information-in-variables#default-environment-variables
+    return pytest.mark.skipif(
+        os.getenv("CI", "false") == "true",
+        reason=reason,
+    )

test_runner/performance/pageserver/pagebench/test_pageserver_max_throughput_getpage_at_latest_lsn.py

@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 import json
-import os
 from pathlib import Path
 from typing import TYPE_CHECKING
 
@@ -14,7 +13,7 @@ from fixtures.neon_fixtures import (
     PgBin,
     wait_for_last_flush_lsn,
 )
-from fixtures.utils import get_scale_for_db, humantime_to_ms
+from fixtures.utils import get_scale_for_db, humantime_to_ms, skip_on_ci
 
 from performance.pageserver.util import (
     setup_pageserver_with_tenants,
@@ -38,9 +37,8 @@ if TYPE_CHECKING:
 @pytest.mark.parametrize("pgbench_scale", [get_scale_for_db(200)])
 @pytest.mark.parametrize("n_tenants", [500])
 @pytest.mark.timeout(10000)
-@pytest.mark.skipif(
-    os.getenv("CI", "false") == "true",
-    reason="This test needs lot of resources and should run on dedicated HW, not in github action runners as part of CI",
+@skip_on_ci(
+    "This test needs lot of resources and should run on dedicated HW, not in github action runners as part of CI"
 )
 def test_pageserver_characterize_throughput_with_n_tenants(
     neon_env_builder: NeonEnvBuilder,
@@ -66,9 +64,8 @@ def test_pageserver_characterize_throughput_with_n_tenants(
 @pytest.mark.parametrize("n_clients", [1, 64])
 @pytest.mark.parametrize("n_tenants", [1])
 @pytest.mark.timeout(2400)
-@pytest.mark.skipif(
-    os.getenv("CI", "false") == "true",
-    reason="This test needs lot of resources and should run on dedicated HW, not in github action runners as part of CI",
+@skip_on_ci(
+    "This test needs lot of resources and should run on dedicated HW, not in github action runners as part of CI"
 )
 def test_pageserver_characterize_latencies_with_1_client_and_throughput_with_many_clients_one_tenant(
     neon_env_builder: NeonEnvBuilder,

test_runner/performance/test_copy.py

@@ -2,11 +2,13 @@ from __future__ import annotations
 
 from contextlib import closing
 from io import BufferedReader, RawIOBase
-from typing import Optional
+from typing import Optional, final
 
 from fixtures.compare_fixtures import PgCompare
+from typing_extensions import override
 
 
+@final
 class CopyTestData(RawIOBase):
     def __init__(self, rows: int):
         self.rows = rows
@@ -14,6 +16,7 @@ class CopyTestData(RawIOBase):
         self.linebuf: Optional[bytes] = None
         self.ptr = 0
 
+    @override
     def readable(self):
         return True
 

test_runner/performance/test_logical_replication.py

@@ -149,12 +149,16 @@ def test_subscriber_lag(
                 check_pgbench_still_running(pub_workload, "pub")
                 check_pgbench_still_running(sub_workload, "sub")
 
-                with (
-                    psycopg2.connect(pub_connstr) as pub_conn,
-                    psycopg2.connect(sub_connstr) as sub_conn,
-                ):
-                    with pub_conn.cursor() as pub_cur, sub_conn.cursor() as sub_cur:
-                        lag = measure_logical_replication_lag(sub_cur, pub_cur)
+                pub_conn = psycopg2.connect(pub_connstr)
+                sub_conn = psycopg2.connect(sub_connstr)
+                pub_conn.autocommit = True
+                sub_conn.autocommit = True
+
+                with pub_conn.cursor() as pub_cur, sub_conn.cursor() as sub_cur:
+                    lag = measure_logical_replication_lag(sub_cur, pub_cur)
+
+                pub_conn.close()
+                sub_conn.close()
 
                 log.info(f"Replica lagged behind master by {lag} seconds")
                 zenbenchmark.record("replica_lag", lag, "s", MetricReport.LOWER_IS_BETTER)
@@ -206,6 +210,7 @@ def test_publisher_restart(
     sub_conn = psycopg2.connect(sub_connstr)
     pub_conn.autocommit = True
     sub_conn.autocommit = True
+
     with pub_conn.cursor() as pub_cur, sub_conn.cursor() as sub_cur:
         pub_cur.execute("SELECT 1 FROM pg_catalog.pg_publication WHERE pubname = 'pub1'")
         pub_exists = len(pub_cur.fetchall()) != 0
@@ -222,6 +227,7 @@ def test_publisher_restart(
             sub_cur.execute(f"create subscription sub1 connection '{pub_connstr}' publication pub1")
 
         initial_sync_lag = measure_logical_replication_lag(sub_cur, pub_cur)
+
     pub_conn.close()
     sub_conn.close()
 
@@ -248,12 +254,17 @@ def test_publisher_restart(
                     ["pgbench", "-c10", pgbench_duration, "-Mprepared"],
                     env=pub_env,
                 )
-                with (
-                    psycopg2.connect(pub_connstr) as pub_conn,
-                    psycopg2.connect(sub_connstr) as sub_conn,
-                ):
-                    with pub_conn.cursor() as pub_cur, sub_conn.cursor() as sub_cur:
-                        lag = measure_logical_replication_lag(sub_cur, pub_cur)
+
+                pub_conn = psycopg2.connect(pub_connstr)
+                sub_conn = psycopg2.connect(sub_connstr)
+                pub_conn.autocommit = True
+                sub_conn.autocommit = True
+
+                with pub_conn.cursor() as pub_cur, sub_conn.cursor() as sub_cur:
+                    lag = measure_logical_replication_lag(sub_cur, pub_cur)
+
+                pub_conn.close()
+                sub_conn.close()
 
                 log.info(f"Replica lagged behind master by {lag} seconds")
                 zenbenchmark.record("replica_lag", lag, "s", MetricReport.LOWER_IS_BETTER)
@@ -288,58 +299,56 @@ def test_snap_files(
     env = benchmark_project_pub.pgbench_env
     connstr = benchmark_project_pub.connstr
 
-    with psycopg2.connect(connstr) as conn:
-        conn.autocommit = True
-        with conn.cursor() as cur:
-            cur.execute("SELECT rolsuper FROM pg_roles WHERE rolname = 'neondb_owner'")
-            is_super = cast("bool", cur.fetchall()[0][0])
-            assert is_super, "This benchmark won't work if we don't have superuser"
+    conn = psycopg2.connect(connstr)
+    conn.autocommit = True
+
+    with conn.cursor() as cur:
+        cur.execute("SELECT rolsuper FROM pg_roles WHERE rolname = 'neondb_owner'")
+        is_super = cast("bool", cur.fetchall()[0][0])
+        assert is_super, "This benchmark won't work if we don't have superuser"
+
+    conn.close()
 
     pg_bin.run_capture(["pgbench", "-i", "-I", "dtGvp", "-s100"], env=env)
 
     conn = psycopg2.connect(connstr)
     conn.autocommit = True
-    cur = conn.cursor()
-    cur.execute("ALTER SYSTEM SET neon.logical_replication_max_snap_files = -1")
 
-    with psycopg2.connect(connstr) as conn:
-        conn.autocommit = True
-        with conn.cursor() as cur:
-            cur.execute("SELECT pg_reload_conf()")
-
-    with psycopg2.connect(connstr) as conn:
-        conn.autocommit = True
-        with conn.cursor() as cur:
-            cur.execute(
-                """
-                DO $$
-                    BEGIN
-                    IF EXISTS (
-                        SELECT 1
-                        FROM pg_replication_slots
-                        WHERE slot_name = 'slotter'
-                    ) THEN
-                        PERFORM pg_drop_replication_slot('slotter');
-                    END IF;
-                END $$;
+    with conn.cursor() as cur:
+        cur.execute(
             """
-            )
-            cur.execute("SELECT pg_create_logical_replication_slot('slotter', 'test_decoding')")
+            DO $$
+                BEGIN
+                IF EXISTS (
+                    SELECT 1
+                    FROM pg_replication_slots
+                    WHERE slot_name = 'slotter'
+                ) THEN
+                    PERFORM pg_drop_replication_slot('slotter');
+                END IF;
+            END $$;
+        """
+        )
+        cur.execute("SELECT pg_create_logical_replication_slot('slotter', 'test_decoding')")
+
+    conn.close()
 
     workload = pg_bin.run_nonblocking(["pgbench", "-c10", pgbench_duration, "-Mprepared"], env=env)
     try:
         start = time.time()
         prev_measurement = time.time()
         while time.time() - start < test_duration_min * 60:
-            with psycopg2.connect(connstr) as conn:
-                with conn.cursor() as cur:
-                    cur.execute(
-                        "SELECT count(*) FROM (SELECT pg_log_standby_snapshot() FROM generate_series(1, 10000) g) s"
-                    )
-                    check_pgbench_still_running(workload)
-                    cur.execute(
-                        "SELECT pg_replication_slot_advance('slotter', pg_current_wal_lsn())"
-                    )
+            conn = psycopg2.connect(connstr)
+            conn.autocommit = True
+
+            with conn.cursor() as cur:
+                cur.execute(
+                    "SELECT count(*) FROM (SELECT pg_log_standby_snapshot() FROM generate_series(1, 10000) g) s"
+                )
+                check_pgbench_still_running(workload)
+                cur.execute("SELECT pg_replication_slot_advance('slotter', pg_current_wal_lsn())")
+
+            conn.close()
 
             # Measure storage
             if time.time() - prev_measurement > test_interval_min * 60:

test_runner/performance/test_physical_replication.py

@@ -102,15 +102,21 @@ def test_ro_replica_lag(
                     check_pgbench_still_running(master_workload)
                     check_pgbench_still_running(replica_workload)
                     time.sleep(sync_interval_min * 60)
+
+                    conn_master = psycopg2.connect(master_connstr)
+                    conn_replica = psycopg2.connect(replica_connstr)
+                    conn_master.autocommit = True
+                    conn_replica.autocommit = True
+
                     with (
-                        psycopg2.connect(master_connstr) as conn_master,
-                        psycopg2.connect(replica_connstr) as conn_replica,
+                        conn_master.cursor() as cur_master,
+                        conn_replica.cursor() as cur_replica,
                     ):
-                        with (
-                            conn_master.cursor() as cur_master,
-                            conn_replica.cursor() as cur_replica,
-                        ):
-                            lag = measure_replication_lag(cur_master, cur_replica)
+                        lag = measure_replication_lag(cur_master, cur_replica)
+
+                    conn_master.close()
+                    conn_replica.close()
+
                     log.info(f"Replica lagged behind master by {lag} seconds")
                     zenbenchmark.record("replica_lag", lag, "s", MetricReport.LOWER_IS_BETTER)
             finally:
@@ -219,11 +225,15 @@ def test_replication_start_stop(
         pg_bin.run_capture(["pgbench", "-i", "-I", "dtGvp", "-s10"], env=master_env)
 
         # Sync replicas
-        with psycopg2.connect(master_connstr) as conn_master:
-            with conn_master.cursor() as cur_master:
-                for i in range(num_replicas):
-                    conn_replica = psycopg2.connect(replica_connstr[i])
-                    measure_replication_lag(cur_master, conn_replica.cursor())
+        conn_master = psycopg2.connect(master_connstr)
+        conn_master.autocommit = True
+
+        with conn_master.cursor() as cur_master:
+            for i in range(num_replicas):
+                conn_replica = psycopg2.connect(replica_connstr[i])
+                measure_replication_lag(cur_master, conn_replica.cursor())
+
+        conn_master.close()
 
         master_pgbench = pg_bin.run_nonblocking(
             [
@@ -277,17 +287,22 @@ def test_replication_start_stop(
 
             time.sleep(configuration_test_time_sec)
 
-            with psycopg2.connect(master_connstr) as conn_master:
-                with conn_master.cursor() as cur_master:
-                    for ireplica in range(num_replicas):
-                        replica_conn = psycopg2.connect(replica_connstr[ireplica])
-                        lag = measure_replication_lag(cur_master, replica_conn.cursor())
-                        zenbenchmark.record(
-                            f"Replica {ireplica} lag", lag, "s", MetricReport.LOWER_IS_BETTER
-                        )
-                        log.info(
-                            f"Replica {ireplica} lagging behind master by {lag} seconds after configuration {iconfig:>b}"
-                        )
+            conn_master = psycopg2.connect(master_connstr)
+            conn_master.autocommit = True
+
+            with conn_master.cursor() as cur_master:
+                for ireplica in range(num_replicas):
+                    replica_conn = psycopg2.connect(replica_connstr[ireplica])
+                    lag = measure_replication_lag(cur_master, replica_conn.cursor())
+                    zenbenchmark.record(
+                        f"Replica {ireplica} lag", lag, "s", MetricReport.LOWER_IS_BETTER
+                    )
+                    log.info(
+                        f"Replica {ireplica} lagging behind master by {lag} seconds after configuration {iconfig:>b}"
+                    )
+
+            conn_master.close()
+
         master_pgbench.terminate()
     except Exception as e:
         error_occurred = True

test_runner/regress/test_branch_and_gc.py

@@ -8,7 +8,7 @@ from fixtures.common_types import Lsn, TimelineId
 from fixtures.log_helper import log
 from fixtures.neon_fixtures import NeonEnv
 from fixtures.pageserver.http import TimelineCreate406
-from fixtures.utils import query_scalar
+from fixtures.utils import query_scalar, skip_in_debug_build
 
 
 # Test the GC implementation when running with branching.
@@ -48,10 +48,8 @@ from fixtures.utils import query_scalar
 # Because the delta layer D covering lsn1 is corrupted, creating a branch
 # starting from lsn1 should return an error as follows:
 #     could not find data for key ... at LSN ..., for request at LSN ...
-def test_branch_and_gc(neon_simple_env: NeonEnv, build_type: str):
-    if build_type == "debug":
-        pytest.skip("times out in debug builds")
-
+@skip_in_debug_build("times out in debug builds")
+def test_branch_and_gc(neon_simple_env: NeonEnv):
     env = neon_simple_env
     pageserver_http_client = env.pageserver.http_client()
 

test_runner/regress/test_compaction.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import enum
 import json
-import os
 import time
 from typing import TYPE_CHECKING
 
@@ -13,7 +12,7 @@ from fixtures.neon_fixtures import (
     generate_uploads_and_deletions,
 )
 from fixtures.pageserver.http import PageserverApiException
-from fixtures.utils import wait_until
+from fixtures.utils import skip_in_debug_build, wait_until
 from fixtures.workload import Workload
 
 if TYPE_CHECKING:
@@ -32,7 +31,7 @@ AGGRESIVE_COMPACTION_TENANT_CONF = {
 }
 
 
-@pytest.mark.skipif(os.environ.get("BUILD_TYPE") == "debug", reason="only run with release build")
+@skip_in_debug_build("only run with release build")
 def test_pageserver_compaction_smoke(neon_env_builder: NeonEnvBuilder):
     """
     This is a smoke test that compaction kicks in. The workload repeatedly churns

test_runner/regress/test_compute_locales.py

@@ -0,0 +1,61 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, cast
+
+from fixtures.pg_version import PgVersion
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from fixtures.neon_fixtures import NeonEnv
+
+
+def test_default_locales(neon_simple_env: NeonEnv):
+    """
+    Test that the default locales for compute databases is C.UTF-8.
+    """
+    env = neon_simple_env
+
+    endpoint = env.endpoints.create_start("main")
+
+    domain_locales = cast(
+        "Sequence[str]",
+        endpoint.safe_psql(
+            "SELECT current_setting('lc_messages') AS lc_messages,"
+            + "current_setting('lc_monetary') AS lc_monetary,"
+            + "current_setting('lc_numeric') AS lc_numeric,"
+            + "current_setting('lc_time') AS lc_time"
+        )[0],
+    )
+    for dl in domain_locales:
+        assert dl == "C.UTF-8"
+
+    # Postgres 15 added the locale providers
+    if env.pg_version < PgVersion.V15:
+        results = cast(
+            "Sequence[str]",
+            endpoint.safe_psql(
+                "SELECT datcollate, datctype FROM pg_database WHERE datname = current_database()"
+            )[0],
+        )
+
+        datcollate = results[0]
+        datctype = results[1]
+    else:
+        results = cast(
+            "Sequence[str]",
+            endpoint.safe_psql(
+                "SELECT datlocprovider, datcollate, datctype FROM pg_database WHERE datname = current_database()"
+            )[0],
+        )
+        datlocprovider = results[0]
+        datcollate = results[1]
+        datctype = results[2]
+
+        if env.pg_version >= PgVersion.V17:
+            assert datlocprovider == "b", "The locale provider is not builtin"
+        else:
+            assert datlocprovider == "c", "The locale provider is not libc"
+
+    assert datcollate == "C.UTF-8"
+    assert datctype == "C.UTF-8"