Merge branch 'problame/batching-benchmark' into problame/merge-getpage-test

## Problem

We saw a scale test failure when one shard went
secondary->attached->secondary in a short period of time -- the metrics
for the shard failed a validation assertion that is meant to ensure the
size metric matches the sum of layer sizes in the SecondaryDetail
struct.

This appears to be due to two SecondaryTenants being alive at the same
time -- the first one was shut down but still had its contributions to
the metrics.

Closes: https://github.com/neondatabase/neon/issues/9628

## Summary of changes

- Refactor code for validating metrics and call it in shutdown as well
as during downloads
- Move code for dropping per-tenant secondary metrics from drop() into
shutdown(), so that once shutdown() completes it is definitely safe to
instantiate another SecondaryTenant for the same tenant.

.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

@@ -7,6 +7,10 @@ inputs:
     type: boolean
     required: false
     default: false
+  aws_oicd_role_arn:
+    description: 'the OIDC role arn to (re-)acquire for allure report upload - if not set call must acquire OIDC role'
+    required: false
+    default: ''
 
 outputs:
   base-url:
@@ -79,6 +83,14 @@ runs:
         ALLURE_VERSION: 2.27.0
         ALLURE_ZIP_SHA256: b071858fb2fa542c65d8f152c5c40d26267b2dfb74df1f1608a589ecca38e777
 
+    - name: (Re-)configure AWS credentials # necessary to upload reports to S3 after a long-running test
+      if: ${{ !cancelled() && (inputs.aws_oicd_role_arn != '') }}
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ inputs.aws_oicd_role_arn }}
+        role-duration-seconds: 3600 # 1 hour should be more than enough to upload report
+
     # Potentially we could have several running build for the same key (for example, for the main branch), so we use improvised lock for this
     - name: Acquire lock
       shell: bash -euxo pipefail {0}

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

@@ -8,6 +8,10 @@ inputs:
   unique-key:
     description: 'string to distinguish different results in the same run'
     required: true
+  aws_oicd_role_arn:
+    description: 'the OIDC role arn to (re-)acquire for allure report upload - if not set call must acquire OIDC role'
+    required: false
+    default: ''
 
 runs:
   using: "composite"
@@ -31,6 +35,14 @@ runs:
       env:
         REPORT_DIR: ${{ inputs.report-dir }}
 
+    - name: (Re-)configure AWS credentials # necessary to upload reports to S3 after a long-running test
+      if: ${{ !cancelled() && (inputs.aws_oicd_role_arn != '') }}
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ inputs.aws_oicd_role_arn }}
+        role-duration-seconds: 3600 # 1 hour should be more than enough to upload report
+
     - name: Upload test results
       shell: bash -euxo pipefail {0}
       run: |

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

@@ -48,6 +48,10 @@ inputs:
     description: 'benchmark durations JSON'
     required: false
     default: '{}'
+  aws_oicd_role_arn:
+    description: 'the OIDC role arn to (re-)acquire for allure report upload - if not set call must acquire OIDC role'
+    required: false
+    default: ''
 
 runs:
   using: "composite"
@@ -222,6 +226,13 @@ runs:
         # (for example if we didn't run the test for non build-and-test workflow)
         skip-if-does-not-exist: true
 
+    - name: (Re-)configure AWS credentials # necessary to upload reports to S3 after a long-running test
+      if: ${{ !cancelled() && (inputs.aws_oicd_role_arn != '') }}
+      uses: aws-actions/configure-aws-credentials@v4
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ inputs.aws_oicd_role_arn }}
+        role-duration-seconds: 3600 # 1 hour should be more than enough to upload report
     - name: Upload test results
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-store

.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/_create-release-pr.yml

@@ -0,0 +1,79 @@
+name: Create Release PR
+
+on:
+  workflow_call:
+    inputs:
+      component-name:
+        description: 'Component name'
+        required: true
+        type: string
+      release-branch:
+        description: 'Release branch'
+        required: true
+        type: string
+    secrets:
+      ci-access-token:
+        description: 'CI access token'
+        required: true
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+jobs:
+  create-storage-release-branch:
+    runs-on: ubuntu-22.04
+
+    permissions:
+      contents: write # for `git push`
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        ref: main
+
+    - name: Set variables
+      id: vars
+      env:
+        COMPONENT_NAME: ${{ inputs.component-name }}
+        RELEASE_BRANCH: ${{ inputs.release-branch }}
+      run: |
+        today=$(date +'%Y-%m-%d')
+        echo "title=${COMPONENT_NAME} release ${today}" | tee -a ${GITHUB_OUTPUT}
+        echo "rc-branch=rc/${RELEASE_BRANCH}/${today}"  | tee -a ${GITHUB_OUTPUT}
+
+    - name: Configure git
+      run: |
+        git config user.name "github-actions[bot]"
+        git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
+
+    - name: Create RC branch
+      env:
+        RC_BRANCH: ${{ steps.vars.outputs.rc-branch }}
+        TITLE: ${{ steps.vars.outputs.title }}
+      run: |
+        git checkout -b "${RC_BRANCH}"
+
+        # create an empty commit to distinguish workflow runs
+        # from other possible releases from the same commit
+        git commit --allow-empty -m "${TITLE}"
+
+        git push origin "${RC_BRANCH}"
+
+    - name: Create a PR into ${{ inputs.release-branch }}
+      env:
+        GH_TOKEN: ${{ secrets.ci-access-token }}
+        RC_BRANCH: ${{ steps.vars.outputs.rc-branch }}
+        RELEASE_BRANCH: ${{ inputs.release-branch }}
+        TITLE: ${{ steps.vars.outputs.title }}
+      run: |
+        cat << EOF > body.md
+          ## ${TITLE}
+
+          **Please merge this Pull Request using 'Create a merge commit' button**
+        EOF
+
+        gh pr create --title "${TITLE}" \
+                     --body-file "body.md" \
+                     --head "${RC_BRANCH}" \
+                     --base "${RELEASE_BRANCH}"

.github/workflows/benchmarking.yml

@@ -122,6 +122,7 @@ jobs:
         run_in_parallel: false
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
         # Set --sparse-ordering option of pytest-order plugin
         # to ensure tests are running in order of appears in the file.
         # It's important for test_perf_pgbench.py::test_pgbench_remote_* tests
@@ -133,6 +134,7 @@ jobs:
           --ignore test_runner/performance/test_perf_pgvector_queries.py
           --ignore test_runner/performance/test_logical_replication.py
           --ignore test_runner/performance/test_physical_replication.py
+          --ignore test_runner/performance/test_perf_ingest_using_pgcopydb.py
       env:
         BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -149,12 +151,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic perf testing: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
@@ -210,6 +214,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 5400
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -226,6 +231,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 5400
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -237,11 +243,13 @@ jobs:
       uses: ./.github/actions/allure-report-generate
       with:
         store-test-results-into-db: true
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
 
+    # Post both success and failure to the Slack channel
     - name: Post to a Slack channel
-      if: ${{ github.event.schedule && failure() }}
+      if: ${{ github.event.schedule }}
       uses: slackapi/slack-github-action@v1
       with:
         channel-id: "C06T9AMNDQQ" # on-call-compute-staging-stream
@@ -444,6 +452,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_pgbench_remote_init
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -458,6 +467,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_pgbench_remote_simple_update
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -472,6 +482,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_pgbench_remote_select_only
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -488,12 +499,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic perf testing on ${{ matrix.platform }}: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
@@ -545,12 +558,12 @@ jobs:
         arch=$(uname -m | sed 's/x86_64/amd64/g' | sed 's/aarch64/arm64/g')
 
         cd /home/nonroot
-        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-17/libpq5_17.0-1.pgdg110+1_${arch}.deb"
-        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb"
-        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.4-1.pgdg110+2_${arch}.deb"
-        dpkg -x libpq5_17.0-1.pgdg110+1_${arch}.deb pg
-        dpkg -x postgresql-16_16.4-1.pgdg110+2_${arch}.deb pg
-        dpkg -x postgresql-client-16_16.4-1.pgdg110+2_${arch}.deb pg
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-17/libpq5_17.1-1.pgdg110+1_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-client-16_16.5-1.pgdg110+1_${arch}.deb"
+        wget -q "https://apt.postgresql.org/pub/repos/apt/pool/main/p/postgresql-16/postgresql-16_16.5-1.pgdg110+1_${arch}.deb"
+        dpkg -x libpq5_17.1-1.pgdg110+1_${arch}.deb pg
+        dpkg -x postgresql-16_16.5-1.pgdg110+1_${arch}.deb pg
+        dpkg -x postgresql-client-16_16.5-1.pgdg110+1_${arch}.deb pg
 
         mkdir -p /tmp/neon/pg_install/v16/bin
         ln -s /home/nonroot/pg/usr/lib/postgresql/16/bin/pgbench /tmp/neon/pg_install/v16/bin/pgbench
@@ -598,6 +611,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_pgvector_indexing
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -612,6 +626,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -621,12 +636,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic perf testing on ${{ env.PLATFORM }}: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
@@ -722,6 +739,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 43200 -k test_clickbench
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -734,12 +752,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic OLAP perf testing on ${{ matrix.platform }}: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
@@ -836,6 +856,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_tpch
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -846,12 +867,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic TPC-H perf testing on ${{ matrix.platform }}: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
@@ -934,6 +957,7 @@ jobs:
         save_perf_report: ${{ env.SAVE_PERF_REPORT }}
         extra_params: -m remote_cluster --timeout 21600 -k test_user_examples
         pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
@@ -943,12 +967,14 @@ jobs:
       id: create-allure-report
       if: ${{ !cancelled() }}
       uses: ./.github/actions/allure-report-generate
+      with:
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
 
     - name: Post to a Slack channel
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: |
           Periodic TPC-H perf testing on ${{ matrix.platform }}: ${{ job.status }}
           <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>

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

@@ -2,18 +2,13 @@ name: Build build-tools image
 
 on:
   workflow_call:
-    inputs:
-      image-tag:
-        description: "build-tools image tag"
-        required: true
-        type: string
     outputs:
       image-tag:
         description: "build-tools tag"
-        value: ${{ inputs.image-tag }}
+        value: ${{ jobs.check-image.outputs.tag }}
       image:
         description: "build-tools image"
-        value: neondatabase/build-tools:${{ inputs.image-tag }}
+        value: neondatabase/build-tools:${{ jobs.check-image.outputs.tag }}
 
 defaults:
   run:
@@ -35,7 +30,36 @@ permissions: {}
 
 jobs:
   check-image:
-    uses: ./.github/workflows/check-build-tools-image.yml
+    runs-on: ubuntu-22.04
+    outputs:
+      tag: ${{ steps.get-build-tools-tag.outputs.image-tag }}
+      found: ${{ steps.check-image.outputs.found }}
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Get build-tools image tag for the current commit
+        id: get-build-tools-tag
+        env:
+          IMAGE_TAG: |
+            ${{ hashFiles('build-tools.Dockerfile',
+                          '.github/workflows/build-build-tools-image.yml') }}
+        run: |
+          echo "image-tag=${IMAGE_TAG}" | tee -a $GITHUB_OUTPUT
+
+      - name: Check if such tag found in the registry
+        id: check-image
+        env:
+          IMAGE_TAG: ${{ steps.get-build-tools-tag.outputs.image-tag }}
+        run: |
+          if docker manifest inspect neondatabase/build-tools:${IMAGE_TAG}; then
+            found=true
+          else
+            found=false
+          fi
+
+          echo "found=${found}" | tee -a $GITHUB_OUTPUT
+
 
   build-image:
     needs: [ check-image ]
@@ -48,20 +72,7 @@ jobs:
 
     runs-on: ${{ fromJson(format('["self-hosted", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}
 
-    env:
-      IMAGE_TAG: ${{ inputs.image-tag }}
-
     steps:
-      - name: Check `input.tag` is correct
-        env:
-          INPUTS_IMAGE_TAG: ${{ inputs.image-tag }}
-          CHECK_IMAGE_TAG : ${{ needs.check-image.outputs.image-tag }}
-        run: |
-          if [ "${INPUTS_IMAGE_TAG}" != "${CHECK_IMAGE_TAG}" ]; then
-            echo "'inputs.image-tag' (${INPUTS_IMAGE_TAG}) does not match the tag of the latest build-tools image 'inputs.image-tag' (${CHECK_IMAGE_TAG})"
-            exit 1
-          fi
-
       - uses: actions/checkout@v4
 
       - uses: neondatabase/dev-actions/set-docker-config-dir@6094485bf440001c94a94a3f9e221e81ff6b6193
@@ -92,10 +103,10 @@ jobs:
           cache-from: type=registry,ref=cache.neon.build/build-tools:cache-${{ matrix.debian-version }}-${{ matrix.arch }}
           cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/build-tools:cache-{0}-{1},mode=max', matrix.debian-version, matrix.arch) || '' }}
           tags: |
-            neondatabase/build-tools:${{ inputs.image-tag }}-${{ matrix.debian-version }}-${{ matrix.arch }}
+            neondatabase/build-tools:${{ needs.check-image.outputs.tag }}-${{ matrix.debian-version }}-${{ matrix.arch }}
 
   merge-images:
-    needs: [ build-image ]
+    needs: [ check-image, build-image ]
     runs-on: ubuntu-22.04
 
     steps:
@@ -107,7 +118,7 @@ jobs:
       - name: Create multi-arch image
         env:
           DEFAULT_DEBIAN_VERSION: bullseye
-          IMAGE_TAG: ${{ inputs.image-tag }}
+          IMAGE_TAG: ${{ needs.check-image.outputs.tag }}
         run: |
           for debian_version in bullseye bookworm; do
             tags=("-t" "neondatabase/build-tools:${IMAGE_TAG}-${debian_version}")

.github/workflows/build_and_test.yml

@@ -77,15 +77,9 @@ jobs:
         shell: bash
         id: build-tag
 
-  check-build-tools-image:
-    needs: [ check-permissions ]
-    uses: ./.github/workflows/check-build-tools-image.yml
-
   build-build-tools-image:
-    needs: [ check-build-tools-image ]
+    needs: [ check-permissions ]
     uses: ./.github/workflows/build-build-tools-image.yml
-    with:
-      image-tag: ${{ needs.check-build-tools-image.outputs.image-tag }}
     secrets: inherit
 
   check-codestyle-python:

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

@@ -1,51 +0,0 @@
-name: Check build-tools image
-
-on:
-  workflow_call:
-    outputs:
-      image-tag:
-        description: "build-tools image tag"
-        value: ${{ jobs.check-image.outputs.tag }}
-      found:
-        description: "Whether the image is found in the registry"
-        value: ${{ jobs.check-image.outputs.found }}
-
-defaults:
-  run:
-    shell: bash -euo pipefail {0}
-
-# No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
-permissions: {}
-
-jobs:
-  check-image:
-    runs-on: ubuntu-22.04
-    outputs:
-      tag: ${{ steps.get-build-tools-tag.outputs.image-tag }}
-      found: ${{ steps.check-image.outputs.found }}
-
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Get build-tools image tag for the current commit
-        id: get-build-tools-tag
-        env:
-          IMAGE_TAG: |
-            ${{ hashFiles('build-tools.Dockerfile',
-                          '.github/workflows/check-build-tools-image.yml',
-                          '.github/workflows/build-build-tools-image.yml') }}
-        run: |
-          echo "image-tag=${IMAGE_TAG}" | tee -a $GITHUB_OUTPUT
-
-      - name: Check if such tag found in the registry
-        id: check-image
-        env:
-          IMAGE_TAG: ${{ steps.get-build-tools-tag.outputs.image-tag }}
-        run: |
-          if docker manifest inspect neondatabase/build-tools:${IMAGE_TAG}; then
-            found=true
-          else
-            found=false
-          fi
-
-          echo "found=${found}" | tee -a $GITHUB_OUTPUT

.github/workflows/ingest_benchmark.yml

@@ -0,0 +1,159 @@
+name: benchmarking ingest
+
+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
+      if: ${{ matrix.target_project == 'new_empty_project' }}
+      env:
+          BENCHMARK_INGEST_TARGET_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}
+        ${PSQL} "${BENCHMARK_INGEST_TARGET_CONNSTR}" -c "CREATE EXTENSION IF NOT EXISTS neon; CREATE EXTENSION IF NOT EXISTS neon_utils;"
+        echo "BENCHMARK_INGEST_TARGET_CONNSTR=${BENCHMARK_INGEST_TARGET_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 
+      if: ${{ matrix.target_project == 'large_existing_project' }}
+      env:
+          BENCHMARK_INGEST_TARGET_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="${BENCHMARK_INGEST_TARGET_CONNSTR%/*}"
+        # Extract the query parameters (if any) after the database name
+        query_params="${BENCHMARK_INGEST_TARGET_CONNSTR#*\?}"
+        # Reconstruct the new connection string
+        if [ "$query_params" != "$BENCHMARK_INGEST_TARGET_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 [ "$query_params" != "$BENCHMARK_INGEST_TARGET_CONNSTR" ]; then
+          BENCHMARK_INGEST_TARGET_CONNSTR="${base_connstr}/ludicrous?${query_params}"
+        else
+          BENCHMARK_INGEST_TARGET_CONNSTR="${base_connstr}/ludicrous"
+        fi
+        ${PSQL} "${BENCHMARK_INGEST_TARGET_CONNSTR}" -c "CREATE EXTENSION IF NOT EXISTS neon; CREATE EXTENSION IF NOT EXISTS neon_utils;"
+        echo "BENCHMARK_INGEST_TARGET_CONNSTR=${BENCHMARK_INGEST_TARGET_CONNSTR}" >> $GITHUB_ENV
+
+    - name: Invoke pgcopydb  
+      uses: ./.github/actions/run-python-test-set
+      with:
+        build_type: remote
+        test_selection: performance/test_perf_ingest_using_pgcopydb.py
+        run_in_parallel: false
+        extra_params: -s -m remote_cluster --timeout 86400 -k test_ingest_performance_using_pgcopydb
+        pg_version: v16
+        save_perf_report: true
+        aws_oicd_role_arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        BENCHMARK_INGEST_SOURCE_CONNSTR: ${{ secrets.BENCHMARK_INGEST_SOURCE_CONNSTR }}
+        TARGET_PROJECT_TYPE: ${{ matrix.target_project }}
+        # we report PLATFORM in zenbenchmark NeonBenchmarker perf database and want to distinguish between new project and large tenant
+        PLATFORM: "${{ matrix.target_project }}-us-east-2-staging"
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+
+    - name: show tables sizes after ingest
+      run: |
+        export LD_LIBRARY_PATH=${PG_16_LIB_PATH}
+        ${PSQL} "${BENCHMARK_INGEST_TARGET_CONNSTR}" -c "\dt+"
+      
+    - 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

@@ -26,15 +26,9 @@ jobs:
     with:
       github-event-name: ${{ github.event_name}}
 
-  check-build-tools-image:
-    needs: [ check-permissions ]
-    uses: ./.github/workflows/check-build-tools-image.yml
-
   build-build-tools-image:
-    needs: [ check-build-tools-image ]
+    needs: [ check-permissions ]
     uses: ./.github/workflows/build-build-tools-image.yml
-    with:
-      image-tag: ${{ needs.check-build-tools-image.outputs.image-tag }}
     secrets: inherit
 
   check-macos-build:
@@ -44,7 +38,7 @@ jobs:
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
       github.ref_name == 'main'
     timeout-minutes: 90
-    runs-on: macos-14
+    runs-on: macos-15
 
     env:
       # Use release build only, to have less debug info around
@@ -58,7 +52,7 @@ jobs:
           submodules: true
 
       - name: Install macOS postgres dependencies
-        run: brew install flex bison openssl protobuf icu4c pkg-config
+        run: brew install flex bison openssl protobuf icu4c
 
       - name: Set pg 14 revision for caching
         id: pg_v14_rev

.github/workflows/periodic_pagebench.yml

@@ -72,7 +72,7 @@ jobs:
           echo "COMMIT_HASH=$INPUT_COMMIT_HASH" >> $GITHUB_ENV
         fi
 
-    - name: Start Bench with run_id   
+    - name: Start Bench with run_id
       run: |
         curl -k -X 'POST' \
         "${EC2_MACHINE_URL_US}/start_test/${GITHUB_RUN_ID}" \
@@ -116,7 +116,7 @@ jobs:
         -H 'accept: application/gzip' \
         -H "Authorization: Bearer $API_KEY" \
         --output "test_log_${GITHUB_RUN_ID}.gz"
-    
+
     - name: Unzip Test Log and Print it into this job's log
       if: always() && steps.poll_step.outputs.too_many_runs != 'true'
       run: |
@@ -134,13 +134,13 @@ jobs:
       if: ${{ github.event.schedule && failure() }}
       uses: slackapi/slack-github-action@v1
       with:
-        channel-id: "C033QLM5P7D" # dev-staging-stream
+        channel-id: "C06KHQVQ7U3" # on-call-qa-staging-stream
         slack-message: "Periodic pagebench testing on dedicated hardware: ${{ job.status }}\n${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}"
       env:
         SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
 
     - name: Cleanup Test Resources
-      if: always() 
+      if: always()
       run: |
         curl -k -X 'POST' \
         "${EC2_MACHINE_URL_US}/cleanup_test/${GITHUB_RUN_ID}" \

.github/workflows/pg-clients.yml

@@ -39,15 +39,9 @@ jobs:
     with:
       github-event-name: ${{ github.event_name }}
 
-  check-build-tools-image:
-    needs: [ check-permissions ]
-    uses: ./.github/workflows/check-build-tools-image.yml
-
   build-build-tools-image:
-    needs: [ check-build-tools-image ]
+    needs: [ check-permissions ]
     uses: ./.github/workflows/build-build-tools-image.yml
-    with:
-      image-tag: ${{ needs.check-build-tools-image.outputs.image-tag }}
     secrets: inherit
 
   test-logical-replication:

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

@@ -34,16 +34,10 @@ jobs:
         run: |
           echo "${PYTHON_CHANGED_FILES}"
 
-  check-build-tools-image:
+  build-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:

.github/workflows/release.yml

@@ -26,82 +26,26 @@ defaults:
 jobs:
   create-storage-release-branch:
     if: ${{ github.event.schedule == '0 6 * * MON' || format('{0}', inputs.create-storage-release-branch) == 'true' }}
-    runs-on: ubuntu-22.04
 
     permissions:
-      contents: write # for `git push`
+      contents: write
 
-    steps:
-    - name: Check out code
-      uses: actions/checkout@v4
-      with:
-        ref: main
-
-    - name: Set environment variables
-      run: |
-        echo "RELEASE_DATE=$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
-        echo "RELEASE_BRANCH=rc/$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
-
-    - name: Create release branch
-      run: git checkout -b $RELEASE_BRANCH
-
-    - name: Push new branch
-      run: git push origin $RELEASE_BRANCH
-
-    - name: Create pull request into release
-      env:
-        GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
-      run: |
-        TITLE="Storage & Compute release ${RELEASE_DATE}"
-
-        cat << EOF > body.md
-          ## ${TITLE}
-
-          **Please merge this Pull Request using 'Create a merge commit' button**
-        EOF
-
-        gh pr create --title "${TITLE}" \
-                     --body-file "body.md" \
-                     --head "${RELEASE_BRANCH}" \
-                     --base "release"
+    uses: ./.github/workflows/_create-release-pr.yml
+    with:
+      component-name: 'Storage & Compute'
+      release-branch: 'release'
+    secrets:
+      ci-access-token: ${{ secrets.CI_ACCESS_TOKEN }}
 
   create-proxy-release-branch:
     if: ${{ github.event.schedule == '0 6 * * THU' || format('{0}', inputs.create-proxy-release-branch) == 'true' }}
-    runs-on: ubuntu-22.04
 
     permissions:
-      contents: write # for `git push`
+      contents: write
 
-    steps:
-    - name: Check out code
-      uses: actions/checkout@v4
-      with:
-        ref: main
-
-    - name: Set environment variables
-      run: |
-        echo "RELEASE_DATE=$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
-        echo "RELEASE_BRANCH=rc/proxy/$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
-
-    - name: Create release branch
-      run: git checkout -b $RELEASE_BRANCH
-
-    - name: Push new branch
-      run: git push origin $RELEASE_BRANCH
-
-    - name: Create pull request into release
-      env:
-        GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
-      run: |
-        TITLE="Proxy release ${RELEASE_DATE}"
-
-        cat << EOF > body.md
-          ## ${TITLE}
-
-          **Please merge this Pull Request using 'Create a merge commit' button**
-        EOF
-
-        gh pr create --title "${TITLE}" \
-                     --body-file "body.md" \
-                     --head "${RELEASE_BRANCH}" \
-                     --base "release-proxy"
+    uses: ./.github/workflows/_create-release-pr.yml
+    with:
+      component-name: 'Proxy'
+      release-branch: 'release-proxy'
+    secrets:
+      ci-access-token: ${{ secrets.CI_ACCESS_TOKEN }}

.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

@@ -9,7 +9,6 @@ on:
     - Build and Test Locally
     - Build build-tools image
     - Check Permissions
-    - Check build-tools image
     - Check neon with extra platform builds
     - Cloud Regression Test
     - Create Release Branch

CODEOWNERS

@@ -1,6 +1,5 @@
+/.github/ @neondatabase/developer-productivity
 /compute_tools/ @neondatabase/control-plane @neondatabase/compute
-/storage_controller @neondatabase/storage
-/storage_scrubber @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/storage
 /libs/postgres_ffi/ @neondatabase/compute @neondatabase/storage
 /libs/remote_storage/ @neondatabase/storage
@@ -11,4 +10,6 @@
 /pgxn/neon/ @neondatabase/compute @neondatabase/storage
 /proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/storage
+/storage_controller @neondatabase/storage
+/storage_scrubber @neondatabase/storage
 /vendor/ @neondatabase/compute

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",
@@ -3575,7 +3578,6 @@ dependencies = [
  "thiserror",
  "tokio",
  "tokio-util",
- "toml_edit",
  "utils",
  "workspace_hack",
 ]
@@ -3639,6 +3641,7 @@ dependencies = [
  "serde_json",
  "serde_path_to_error",
  "serde_with",
+ "smallvec",
  "storage_broker",
  "strum",
  "strum_macros",
@@ -4006,7 +4009,7 @@ dependencies = [
 [[package]]
 name = "postgres"
 version = "0.19.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=20031d7a9ee1addeae6e0968e3899ae6bf01cee2#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#a130197713830a0ea0004b539b1f51a66b4c3e18"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -4019,7 +4022,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=20031d7a9ee1addeae6e0968e3899ae6bf01cee2#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#a130197713830a0ea0004b539b1f51a66b4c3e18"
 dependencies = [
  "base64 0.20.0",
  "byteorder",
@@ -4038,7 +4041,7 @@ dependencies = [
 [[package]]
 name = "postgres-types"
 version = "0.2.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=20031d7a9ee1addeae6e0968e3899ae6bf01cee2#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#a130197713830a0ea0004b539b1f51a66b4c3e18"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -5660,9 +5663,9 @@ dependencies = [
 
 [[package]]
 name = "smallvec"
-version = "1.13.1"
+version = "1.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e6ecd384b10a64542d77071bd64bd7b231f4ed5940fba55e98c3de13824cf3d7"
+checksum = "3c5e1a9a646d36c3599cd173a41282daf47c44583ad367b8e6837255952e5c67"
 
 [[package]]
 name = "smol_str"
@@ -6071,9 +6074,9 @@ dependencies = [
 
 [[package]]
 name = "tikv-jemalloc-ctl"
-version = "0.5.4"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "619bfed27d807b54f7f776b9430d4f8060e66ee138a28632ca898584d462c31c"
+checksum = "f21f216790c8df74ce3ab25b534e0718da5a1916719771d3fec23315c99e468b"
 dependencies = [
  "libc",
  "paste",
@@ -6082,9 +6085,9 @@ dependencies = [
 
 [[package]]
 name = "tikv-jemalloc-sys"
-version = "0.5.4+5.3.0-patched"
+version = "0.6.0+5.3.0-1-ge13ca993e8ccb9ba9847cc330696e02839f328f7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9402443cb8fd499b6f327e40565234ff34dbda27460c5b47db0db77443dd85d1"
+checksum = "cd3c60906412afa9c2b5b5a48ca6a5abe5736aec9eb48ad05037a677e52e4e2d"
 dependencies = [
  "cc",
  "libc",
@@ -6092,9 +6095,9 @@ dependencies = [
 
 [[package]]
 name = "tikv-jemallocator"
-version = "0.5.4"
+version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "965fe0c26be5c56c94e38ba547249074803efd52adfb66de62107d95aab3eaca"
+checksum = "4cec5ff18518d81584f477e9bfdf957f5bb0979b0bac3af4ca30b5b3ae2d2865"
 dependencies = [
  "libc",
  "tikv-jemalloc-sys",
@@ -6224,7 +6227,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.7"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=20031d7a9ee1addeae6e0968e3899ae6bf01cee2#20031d7a9ee1addeae6e0968e3899ae6bf01cee2"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#a130197713830a0ea0004b539b1f51a66b4c3e18"
 dependencies = [
  "async-trait",
  "byteorder",

Cargo.toml

@@ -168,8 +168,8 @@ sync_wrapper = "0.1.2"
 tar = "0.4"
 test-context = "0.3"
 thiserror = "1.0"
-tikv-jemallocator = "0.5"
-tikv-jemalloc-ctl = "0.5"
+tikv-jemallocator = { version = "0.6", features = ["stats"] }
+tikv-jemalloc-ctl = { version = "0.6", features = ["stats"] }
 tokio = { version = "1.17", features = ["macros"] }
 tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"
@@ -203,21 +203,10 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-
-# We want to use the 'neon' branch for these, but there's currently one
-# incompatible change on the branch. See:
-#
-# - PR #8076 which contained changes that depended on the new changes in
-#   the rust-postgres crate, and
-# - PR #8654 which reverted those changes and made the code in proxy incompatible
-#   with the tip of the 'neon' branch again.
-#
-# When those proxy changes are re-applied (see PR #8747), we can switch using
-# the tip of the 'neon' branch again.
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev = "20031d7a9ee1addeae6e0968e3899ae6bf01cee2" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev = "20031d7a9ee1addeae6e0968e3899ae6bf01cee2" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev = "20031d7a9ee1addeae6e0968e3899ae6bf01cee2" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev = "20031d7a9ee1addeae6e0968e3899ae6bf01cee2" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 
 ## Local libraries
 compute_api = { version = "0.1", path = "./libs/compute_api/" }
@@ -255,7 +244,7 @@ tonic-build = "0.12"
 [patch.crates-io]
 
 # Needed to get `tokio-postgres-rustls` to depend on our fork.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev = "20031d7a9ee1addeae6e0968e3899ae6bf01cee2" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 
 ################# Binary contents sections
 

compute/compute-node.Dockerfile

@@ -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/neon_collector.jsonnet

@@ -6,6 +6,7 @@
     import 'sql_exporter/compute_backpressure_throttling_seconds.libsonnet',
     import 'sql_exporter/compute_current_lsn.libsonnet',
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
+    import 'sql_exporter/compute_max_connections.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',
     import 'sql_exporter/compute_subscriptions_count.libsonnet',
     import 'sql_exporter/connection_counts.libsonnet',

compute/etc/sql_exporter/compute_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/etc/sql_exporter/compute_max_connections.libsonnet

@@ -0,0 +1,10 @@
+{
+  metric_name: 'compute_max_connections',
+  type: 'gauge',
+  help: 'Max connections allowed for Postgres',
+  key_labels: null,
+  values: [
+    'max_connections',
+  ],
+  query: importstr 'sql_exporter/compute_max_connections.sql',
+}

compute/etc/sql_exporter/compute_max_connections.sql

@@ -0,0 +1 @@
+SELECT current_setting('max_connections') as max_connections;

compute/patches/cloud_regress_pg16.patch

@@ -147,7 +147,7 @@ index 542c2e098c..0062d3024f 100644
  ALTER TABLE ptnowner1 OWNER TO regress_ptnowner;
  ALTER TABLE ptnowner OWNER TO regress_ptnowner;
 diff --git a/src/test/regress/expected/collate.icu.utf8.out b/src/test/regress/expected/collate.icu.utf8.out
-index 97bbe53b64..eac3d42a79 100644
+index 3f9a8f539c..0a51b52940 100644
 --- a/src/test/regress/expected/collate.icu.utf8.out
 +++ b/src/test/regress/expected/collate.icu.utf8.out
 @@ -1016,7 +1016,7 @@ select * from collate_test1 where b ilike 'ABC';
@@ -309,7 +309,7 @@ index b48365ec98..a6ef910055 100644
  -- the wrong partition. This test is *not* guaranteed to trigger that bug, but
  -- does so when shared_buffers is small enough.  To test if we encountered the
 diff --git a/src/test/regress/expected/copy2.out b/src/test/regress/expected/copy2.out
-index faf1a4d1b0..a44c97db52 100644
+index 9a74820ee8..22400a5551 100644
 --- a/src/test/regress/expected/copy2.out
 +++ b/src/test/regress/expected/copy2.out
 @@ -553,8 +553,8 @@ select * from check_con_tbl;
@@ -573,7 +573,7 @@ index 93302a07ef..1a73f083ac 100644
  -- that does not match with what's expected.
  -- This checks all the object types that include schema qualifications.
 diff --git a/src/test/regress/expected/create_view.out b/src/test/regress/expected/create_view.out
-index f3f8c7b5a2..3e3e54ff4c 100644
+index f551624afb..57f1e432d4 100644
 --- a/src/test/regress/expected/create_view.out
 +++ b/src/test/regress/expected/create_view.out
 @@ -18,7 +18,8 @@ CREATE TABLE real_city (
@@ -700,12 +700,12 @@ index 6ed50fdcfa..caa00a345d 100644
  COMMENT ON FOREIGN DATA WRAPPER dummy IS 'useless';
  CREATE FOREIGN DATA WRAPPER postgresql VALIDATOR postgresql_fdw_validator;
 diff --git a/src/test/regress/expected/foreign_key.out b/src/test/regress/expected/foreign_key.out
-index 12e523c737..8872e23935 100644
+index 6b8c2f2414..8e13b7fa46 100644
 --- a/src/test/regress/expected/foreign_key.out
 +++ b/src/test/regress/expected/foreign_key.out
-@@ -1968,7 +1968,7 @@ ALTER TABLE fk_partitioned_fk ATTACH PARTITION fk_partitioned_fk_2
-   FOR VALUES IN (1600);
- -- leave these tables around intentionally
+@@ -1985,7 +1985,7 @@ ALTER TABLE fk_partitioned_fk_6 ATTACH PARTITION fk_partitioned_pk_6 FOR VALUES
+ ERROR:  cannot ALTER TABLE "fk_partitioned_pk_61" because it is being used by active queries in this session
+ DROP TABLE fk_partitioned_pk_6, fk_partitioned_fk_6;
  -- test the case when the referenced table is owned by a different user
 -create role regress_other_partitioned_fk_owner;
 +create role regress_other_partitioned_fk_owner PASSWORD NEON_PASSWORD_PLACEHOLDER;
@@ -713,7 +713,7 @@ index 12e523c737..8872e23935 100644
  set role regress_other_partitioned_fk_owner;
  create table other_partitioned_fk(a int, b int) partition by list (a);
 diff --git a/src/test/regress/expected/generated.out b/src/test/regress/expected/generated.out
-index 0f623f7119..b48588a54e 100644
+index 5881420388..4ae21aa43c 100644
 --- a/src/test/regress/expected/generated.out
 +++ b/src/test/regress/expected/generated.out
 @@ -534,7 +534,7 @@ CREATE TABLE gtest10a (a int PRIMARY KEY, b int GENERATED ALWAYS AS (a * 2) STOR
@@ -762,7 +762,7 @@ index a2036a1597..805d73b9d2 100644
  -- fields, leading to long bucket chains and lots of table expansion.
  -- this is therefore a stress test of the bucket overflow code (unlike
 diff --git a/src/test/regress/expected/identity.out b/src/test/regress/expected/identity.out
-index cc7772349f..98a08eb48d 100644
+index 1b74958de9..078187b542 100644
 --- a/src/test/regress/expected/identity.out
 +++ b/src/test/regress/expected/identity.out
 @@ -520,7 +520,7 @@ ALTER TABLE itest7 ALTER COLUMN a SET GENERATED BY DEFAULT;
@@ -775,10 +775,10 @@ index cc7772349f..98a08eb48d 100644
  GRANT SELECT, INSERT ON itest8 TO regress_identity_user1;
  SET ROLE regress_identity_user1;
 diff --git a/src/test/regress/expected/inherit.out b/src/test/regress/expected/inherit.out
-index 4943429e9b..0257f22b15 100644
+index 8f831c95c3..ec681b52af 100644
 --- a/src/test/regress/expected/inherit.out
 +++ b/src/test/regress/expected/inherit.out
-@@ -2606,7 +2606,7 @@ create index on permtest_parent (left(c, 3));
+@@ -2636,7 +2636,7 @@ create index on permtest_parent (left(c, 3));
  insert into permtest_parent
    select 1, 'a', left(fipshash(i::text), 5) from generate_series(0, 100) i;
  analyze permtest_parent;
@@ -1133,7 +1133,7 @@ index 8475231735..1afae5395f 100644
  SELECT rolname, rolpassword
      FROM pg_authid
 diff --git a/src/test/regress/expected/privileges.out b/src/test/regress/expected/privileges.out
-index fbb0489a4f..2905194e2c 100644
+index 5b9dba7b32..cc408dad42 100644
 --- a/src/test/regress/expected/privileges.out
 +++ b/src/test/regress/expected/privileges.out
 @@ -20,19 +20,19 @@ SELECT lo_unlink(oid) FROM pg_largeobject_metadata WHERE oid >= 1000 AND oid < 3
@@ -1185,7 +1185,7 @@ index fbb0489a4f..2905194e2c 100644
  GRANT pg_read_all_data TO regress_priv_user6;
  GRANT pg_write_all_data TO regress_priv_user7;
  GRANT pg_read_all_settings TO regress_priv_user8 WITH ADMIN OPTION;
-@@ -145,8 +145,8 @@ REVOKE pg_read_all_settings FROM regress_priv_user8;
+@@ -212,8 +212,8 @@ REVOKE pg_read_all_settings FROM regress_priv_user8;
  DROP USER regress_priv_user10;
  DROP USER regress_priv_user9;
  DROP USER regress_priv_user8;
@@ -1196,7 +1196,7 @@ index fbb0489a4f..2905194e2c 100644
  ALTER GROUP regress_priv_group1 ADD USER regress_priv_user4;
  GRANT regress_priv_group2 TO regress_priv_user2 GRANTED BY regress_priv_user1;
  SET SESSION AUTHORIZATION regress_priv_user1;
-@@ -172,12 +172,16 @@ GRANT regress_priv_role TO regress_priv_user1 WITH ADMIN OPTION GRANTED BY regre
+@@ -239,12 +239,16 @@ GRANT regress_priv_role TO regress_priv_user1 WITH ADMIN OPTION GRANTED BY regre
  ERROR:  permission denied to grant privileges as role "regress_priv_role"
  DETAIL:  The grantor must have the ADMIN option on role "regress_priv_role".
  GRANT regress_priv_role TO regress_priv_user1 WITH ADMIN OPTION GRANTED BY CURRENT_ROLE;
@@ -1213,7 +1213,7 @@ index fbb0489a4f..2905194e2c 100644
  DROP ROLE regress_priv_role;
  SET SESSION AUTHORIZATION regress_priv_user1;
  SELECT session_user, current_user;
-@@ -1709,7 +1713,7 @@ SELECT has_table_privilege('regress_priv_user1', 'atest4', 'SELECT WITH GRANT OP
+@@ -1776,7 +1780,7 @@ SELECT has_table_privilege('regress_priv_user1', 'atest4', 'SELECT WITH GRANT OP
  
  -- security-restricted operations
  \c -
@@ -1222,7 +1222,7 @@ index fbb0489a4f..2905194e2c 100644
  -- Check that index expressions and predicates are run as the table's owner
  -- A dummy index function checking current_user
  CREATE FUNCTION sro_ifun(int) RETURNS int AS $$
-@@ -2601,8 +2605,8 @@ drop cascades to function testns.priv_testagg(integer)
+@@ -2668,8 +2672,8 @@ drop cascades to function testns.priv_testagg(integer)
  drop cascades to function testns.priv_testproc(integer)
  -- Change owner of the schema & and rename of new schema owner
  \c -
@@ -1233,7 +1233,7 @@ index fbb0489a4f..2905194e2c 100644
  SET SESSION ROLE regress_schemauser1;
  CREATE SCHEMA testns;
  SELECT nspname, rolname FROM pg_namespace, pg_roles WHERE pg_namespace.nspname = 'testns' AND pg_namespace.nspowner = pg_roles.oid;
-@@ -2725,7 +2729,7 @@ DROP USER regress_priv_user7;
+@@ -2792,7 +2796,7 @@ DROP USER regress_priv_user7;
  DROP USER regress_priv_user8; -- does not exist
  ERROR:  role "regress_priv_user8" does not exist
  -- permissions with LOCK TABLE
@@ -1242,7 +1242,7 @@ index fbb0489a4f..2905194e2c 100644
  CREATE TABLE lock_table (a int);
  -- LOCK TABLE and SELECT permission
  GRANT SELECT ON lock_table TO regress_locktable_user;
-@@ -2807,7 +2811,7 @@ DROP USER regress_locktable_user;
+@@ -2874,7 +2878,7 @@ DROP USER regress_locktable_user;
  -- pg_backend_memory_contexts.
  -- switch to superuser
  \c -
@@ -1251,7 +1251,7 @@ index fbb0489a4f..2905194e2c 100644
  SELECT has_table_privilege('regress_readallstats','pg_backend_memory_contexts','SELECT'); -- no
   has_table_privilege 
  ---------------------
-@@ -2851,10 +2855,10 @@ RESET ROLE;
+@@ -2918,10 +2922,10 @@ RESET ROLE;
  -- clean up
  DROP ROLE regress_readallstats;
  -- test role grantor machinery
@@ -1266,7 +1266,7 @@ index fbb0489a4f..2905194e2c 100644
  GRANT regress_group TO regress_group_direct_manager WITH INHERIT FALSE, ADMIN TRUE;
  GRANT regress_group_direct_manager TO regress_group_indirect_manager;
  SET SESSION AUTHORIZATION regress_group_direct_manager;
-@@ -2883,9 +2887,9 @@ DROP ROLE regress_group_direct_manager;
+@@ -2950,9 +2954,9 @@ DROP ROLE regress_group_direct_manager;
  DROP ROLE regress_group_indirect_manager;
  DROP ROLE regress_group_member;
  -- test SET and INHERIT options with object ownership changes
@@ -1813,7 +1813,7 @@ index 5e6969b173..2c4d52237f 100644
  
  -- clean up roles
 diff --git a/src/test/regress/expected/rowsecurity.out b/src/test/regress/expected/rowsecurity.out
-index 97ca9bf72c..b2a7a6f710 100644
+index 218c0c2863..f7af0cfb12 100644
 --- a/src/test/regress/expected/rowsecurity.out
 +++ b/src/test/regress/expected/rowsecurity.out
 @@ -14,13 +14,13 @@ DROP ROLE IF EXISTS regress_rls_group2;
@@ -1917,6 +1917,19 @@ index b79fe9a1c0..e29fab88ab 100644
  ALTER DEFAULT PRIVILEGES FOR ROLE regress_selinto_user
  	  REVOKE INSERT ON TABLES FROM regress_selinto_user;
  GRANT ALL ON SCHEMA selinto_schema TO public;
+diff --git a/src/test/regress/expected/select_parallel.out b/src/test/regress/expected/select_parallel.out
+index afc6ab08c2..dfcd891af3 100644
+--- a/src/test/regress/expected/select_parallel.out
++++ b/src/test/regress/expected/select_parallel.out
+@@ -1220,7 +1220,7 @@ SELECT 1 FROM tenk1_vw_sec
+ 
+ rollback;
+ -- test that function option SET ROLE works in parallel workers.
+-create role regress_parallel_worker;
++create role regress_parallel_worker PASSWORD NEON_PASSWORD_PLACEHOLDER;
+ create function set_and_report_role() returns text as
+   $$ select current_setting('role') $$ language sql parallel safe
+   set role = regress_parallel_worker;
 diff --git a/src/test/regress/expected/select_views.out b/src/test/regress/expected/select_views.out
 index 1aeed8452b..7d9427d070 100644
 --- a/src/test/regress/expected/select_views.out
@@ -2369,7 +2382,7 @@ index 6cb9c926c0..5e689e4062 100644
  ALTER TABLE ptnowner1 OWNER TO regress_ptnowner;
  ALTER TABLE ptnowner OWNER TO regress_ptnowner;
 diff --git a/src/test/regress/sql/collate.icu.utf8.sql b/src/test/regress/sql/collate.icu.utf8.sql
-index 3db9e25913..c66d5aa2c2 100644
+index 8aa902d5ab..24bb823b86 100644
 --- a/src/test/regress/sql/collate.icu.utf8.sql
 +++ b/src/test/regress/sql/collate.icu.utf8.sql
 @@ -353,7 +353,7 @@ reset enable_seqscan;
@@ -2532,7 +2545,7 @@ index 43d2e906dd..6c993d70f0 100644
  -- An earlier bug (see commit b1ecb9b3fcf) could end up using a buffer from
  -- the wrong partition. This test is *not* guaranteed to trigger that bug, but
 diff --git a/src/test/regress/sql/copy2.sql b/src/test/regress/sql/copy2.sql
-index d759635068..d58e50dcc5 100644
+index cf3828c16e..cf3ca38175 100644
 --- a/src/test/regress/sql/copy2.sql
 +++ b/src/test/regress/sql/copy2.sql
 @@ -365,8 +365,8 @@ copy check_con_tbl from stdin;
@@ -2774,7 +2787,7 @@ index 1b7064247a..be5b662ce1 100644
  -- Cases where schema creation fails as objects are qualified with a schema
  -- that does not match with what's expected.
 diff --git a/src/test/regress/sql/create_view.sql b/src/test/regress/sql/create_view.sql
-index 3a78be1b0c..617d2dc8d6 100644
+index ae6841308b..47bc792e30 100644
 --- a/src/test/regress/sql/create_view.sql
 +++ b/src/test/regress/sql/create_view.sql
 @@ -23,7 +23,8 @@ CREATE TABLE real_city (
@@ -2901,11 +2914,11 @@ index aa147b14a9..370e0dd570 100644
  CREATE FOREIGN DATA WRAPPER dummy;
  COMMENT ON FOREIGN DATA WRAPPER dummy IS 'useless';
 diff --git a/src/test/regress/sql/foreign_key.sql b/src/test/regress/sql/foreign_key.sql
-index 22e177f89b..7138d5e1d4 100644
+index 45c7a534cb..32dd26b8cd 100644
 --- a/src/test/regress/sql/foreign_key.sql
 +++ b/src/test/regress/sql/foreign_key.sql
-@@ -1418,7 +1418,7 @@ ALTER TABLE fk_partitioned_fk ATTACH PARTITION fk_partitioned_fk_2
- -- leave these tables around intentionally
+@@ -1435,7 +1435,7 @@ ALTER TABLE fk_partitioned_fk_6 ATTACH PARTITION fk_partitioned_pk_6 FOR VALUES
+ DROP TABLE fk_partitioned_pk_6, fk_partitioned_fk_6;
  
  -- test the case when the referenced table is owned by a different user
 -create role regress_other_partitioned_fk_owner;
@@ -2963,7 +2976,7 @@ index 527024f710..de49c0b85f 100644
  -- the data in this file has a lot of duplicates in the index key
  -- fields, leading to long bucket chains and lots of table expansion.
 diff --git a/src/test/regress/sql/identity.sql b/src/test/regress/sql/identity.sql
-index 91d2e443b4..241c93f373 100644
+index 7537258a75..9041e35e34 100644
 --- a/src/test/regress/sql/identity.sql
 +++ b/src/test/regress/sql/identity.sql
 @@ -287,7 +287,7 @@ ALTER TABLE itest7 ALTER COLUMN a RESTART;
@@ -2976,10 +2989,10 @@ index 91d2e443b4..241c93f373 100644
  GRANT SELECT, INSERT ON itest8 TO regress_identity_user1;
  SET ROLE regress_identity_user1;
 diff --git a/src/test/regress/sql/inherit.sql b/src/test/regress/sql/inherit.sql
-index fe699c54d5..bdd5993f45 100644
+index b5b554a125..109889ad24 100644
 --- a/src/test/regress/sql/inherit.sql
 +++ b/src/test/regress/sql/inherit.sql
-@@ -950,7 +950,7 @@ create index on permtest_parent (left(c, 3));
+@@ -958,7 +958,7 @@ create index on permtest_parent (left(c, 3));
  insert into permtest_parent
    select 1, 'a', left(fipshash(i::text), 5) from generate_series(0, 100) i;
  analyze permtest_parent;
@@ -3218,7 +3231,7 @@ index 53e86b0b6c..f07cf1ec54 100644
  CREATE ROLE regress_passwd5 PASSWORD 'md5e73a4b11df52a6068f8b39f90be36023';
  
 diff --git a/src/test/regress/sql/privileges.sql b/src/test/regress/sql/privileges.sql
-index 3f68cafcd1..004b26831d 100644
+index 249df17a58..b258e7f26a 100644
 --- a/src/test/regress/sql/privileges.sql
 +++ b/src/test/regress/sql/privileges.sql
 @@ -24,18 +24,18 @@ RESET client_min_messages;
@@ -3269,7 +3282,7 @@ index 3f68cafcd1..004b26831d 100644
  
  GRANT pg_read_all_data TO regress_priv_user6;
  GRANT pg_write_all_data TO regress_priv_user7;
-@@ -130,8 +130,8 @@ DROP USER regress_priv_user10;
+@@ -163,8 +163,8 @@ DROP USER regress_priv_user10;
  DROP USER regress_priv_user9;
  DROP USER regress_priv_user8;
  
@@ -3280,7 +3293,7 @@ index 3f68cafcd1..004b26831d 100644
  
  ALTER GROUP regress_priv_group1 ADD USER regress_priv_user4;
  
-@@ -1124,7 +1124,7 @@ SELECT has_table_privilege('regress_priv_user1', 'atest4', 'SELECT WITH GRANT OP
+@@ -1157,7 +1157,7 @@ SELECT has_table_privilege('regress_priv_user1', 'atest4', 'SELECT WITH GRANT OP
  
  -- security-restricted operations
  \c -
@@ -3289,7 +3302,7 @@ index 3f68cafcd1..004b26831d 100644
  
  -- Check that index expressions and predicates are run as the table's owner
  
-@@ -1620,8 +1620,8 @@ DROP SCHEMA testns CASCADE;
+@@ -1653,8 +1653,8 @@ DROP SCHEMA testns CASCADE;
  -- Change owner of the schema & and rename of new schema owner
  \c -
  
@@ -3300,7 +3313,7 @@ index 3f68cafcd1..004b26831d 100644
  
  SET SESSION ROLE regress_schemauser1;
  CREATE SCHEMA testns;
-@@ -1715,7 +1715,7 @@ DROP USER regress_priv_user8; -- does not exist
+@@ -1748,7 +1748,7 @@ DROP USER regress_priv_user8; -- does not exist
  
  
  -- permissions with LOCK TABLE
@@ -3309,7 +3322,7 @@ index 3f68cafcd1..004b26831d 100644
  CREATE TABLE lock_table (a int);
  
  -- LOCK TABLE and SELECT permission
-@@ -1803,7 +1803,7 @@ DROP USER regress_locktable_user;
+@@ -1836,7 +1836,7 @@ DROP USER regress_locktable_user;
  -- switch to superuser
  \c -
  
@@ -3318,7 +3331,7 @@ index 3f68cafcd1..004b26831d 100644
  
  SELECT has_table_privilege('regress_readallstats','pg_backend_memory_contexts','SELECT'); -- no
  SELECT has_table_privilege('regress_readallstats','pg_shmem_allocations','SELECT'); -- no
-@@ -1823,10 +1823,10 @@ RESET ROLE;
+@@ -1856,10 +1856,10 @@ RESET ROLE;
  DROP ROLE regress_readallstats;
  
  -- test role grantor machinery
@@ -3333,7 +3346,7 @@ index 3f68cafcd1..004b26831d 100644
  
  GRANT regress_group TO regress_group_direct_manager WITH INHERIT FALSE, ADMIN TRUE;
  GRANT regress_group_direct_manager TO regress_group_indirect_manager;
-@@ -1848,9 +1848,9 @@ DROP ROLE regress_group_indirect_manager;
+@@ -1881,9 +1881,9 @@ DROP ROLE regress_group_indirect_manager;
  DROP ROLE regress_group_member;
  
  -- test SET and INHERIT options with object ownership changes
@@ -3625,7 +3638,7 @@ index c961b2d730..0859b89c4f 100644
  -- clean up roles
  DROP ROLE regress_test_def_superuser;
 diff --git a/src/test/regress/sql/rowsecurity.sql b/src/test/regress/sql/rowsecurity.sql
-index dec7340538..cdbc03a5cc 100644
+index d3bfd53e23..919ce1d0c6 100644
 --- a/src/test/regress/sql/rowsecurity.sql
 +++ b/src/test/regress/sql/rowsecurity.sql
 @@ -20,13 +20,13 @@ DROP SCHEMA IF EXISTS regress_rls_schema CASCADE;
@@ -3701,6 +3714,19 @@ index 689c448cc2..223ceb1d75 100644
  ALTER DEFAULT PRIVILEGES FOR ROLE regress_selinto_user
  	  REVOKE INSERT ON TABLES FROM regress_selinto_user;
  GRANT ALL ON SCHEMA selinto_schema TO public;
+diff --git a/src/test/regress/sql/select_parallel.sql b/src/test/regress/sql/select_parallel.sql
+index 33d78e16dc..cb193c9b27 100644
+--- a/src/test/regress/sql/select_parallel.sql
++++ b/src/test/regress/sql/select_parallel.sql
+@@ -464,7 +464,7 @@ SELECT 1 FROM tenk1_vw_sec
+ rollback;
+ 
+ -- test that function option SET ROLE works in parallel workers.
+-create role regress_parallel_worker;
++create role regress_parallel_worker PASSWORD NEON_PASSWORD_PLACEHOLDER;
+ 
+ create function set_and_report_role() returns text as
+   $$ select current_setting('role') $$ language sql parallel safe
 diff --git a/src/test/regress/sql/select_views.sql b/src/test/regress/sql/select_views.sql
 index e742f13699..7bd0255df8 100644
 --- a/src/test/regress/sql/select_views.sql

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/bin/compute_ctl.rs

@@ -105,6 +105,11 @@ fn main() -> Result<()> {
 fn init() -> Result<(String, clap::ArgMatches)> {
     init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
 
+    opentelemetry::global::set_error_handler(|err| {
+        tracing::info!("OpenTelemetry error: {err}");
+    })
+    .expect("global error handler lock poisoned");
+
     let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
     thread::spawn(move || {
         for sig in signals.forever() {

compute_tools/src/catalog.rs

@@ -1,38 +1,40 @@
-use compute_api::{
-    responses::CatalogObjects,
-    spec::{Database, Role},
-};
+use compute_api::responses::CatalogObjects;
 use futures::Stream;
-use postgres::{Client, NoTls};
+use postgres::NoTls;
 use std::{path::Path, process::Stdio, result::Result, sync::Arc};
 use tokio::{
     io::{AsyncBufReadExt, BufReader},
     process::Command,
-    task,
+    spawn,
 };
+use tokio_postgres::connect;
 use tokio_stream::{self as stream, StreamExt};
 use tokio_util::codec::{BytesCodec, FramedRead};
 use tracing::warn;
 
-use crate::{
-    compute::ComputeNode,
-    pg_helpers::{get_existing_dbs, get_existing_roles},
-};
+use crate::compute::ComputeNode;
+use crate::pg_helpers::{get_existing_dbs_async, get_existing_roles_async};
 
 pub async fn get_dbs_and_roles(compute: &Arc<ComputeNode>) -> anyhow::Result<CatalogObjects> {
     let connstr = compute.connstr.clone();
-    task::spawn_blocking(move || {
-        let mut client = Client::connect(connstr.as_str(), NoTls)?;
-        let roles: Vec<Role>;
-        {
-            let mut xact = client.transaction()?;
-            roles = get_existing_roles(&mut xact)?;
-        }
-        let databases: Vec<Database> = get_existing_dbs(&mut client)?.values().cloned().collect();
 
-        Ok(CatalogObjects { roles, databases })
-    })
-    .await?
+    let (client, connection): (tokio_postgres::Client, _) =
+        connect(connstr.as_str(), NoTls).await?;
+
+    spawn(async move {
+        if let Err(e) = connection.await {
+            eprintln!("connection error: {}", e);
+        }
+    });
+
+    let roles = get_existing_roles_async(&client).await?;
+
+    let databases = get_existing_dbs_async(&client)
+        .await?
+        .into_values()
+        .collect();
+
+    Ok(CatalogObjects { roles, databases })
 }
 
 #[derive(Debug, thiserror::Error)]

compute_tools/src/checker.rs

@@ -1,37 +1,9 @@
 use anyhow::{anyhow, Ok, Result};
-use postgres::Client;
 use tokio_postgres::NoTls;
 use tracing::{error, instrument, warn};
 
 use crate::compute::ComputeNode;
 
-/// Create a special service table for availability checks
-/// only if it does not exist already.
-pub fn create_availability_check_data(client: &mut Client) -> Result<()> {
-    let query = "
-        DO $$
-        BEGIN
-            IF NOT EXISTS(
-                SELECT 1
-                FROM pg_catalog.pg_tables
-                WHERE tablename = 'health_check'
-            )
-            THEN
-            CREATE TABLE health_check (
-                id serial primary key,
-                updated_at timestamptz default now()
-            );
-            INSERT INTO health_check VALUES (1, now())
-                ON CONFLICT (id) DO UPDATE
-                 SET updated_at = now();
-            END IF;
-        END
-        $$;";
-    client.execute(query, &[])?;
-
-    Ok(())
-}
-
 /// Update timestamp in a row in a special service table to check
 /// that we can actually write some data in this particular timeline.
 #[instrument(skip_all)]

compute_tools/src/compute.rs

@@ -1,20 +1,21 @@
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::env;
 use std::fs;
+use std::iter::once;
 use std::os::unix::fs::{symlink, PermissionsExt};
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
 use std::sync::atomic::AtomicU32;
 use std::sync::atomic::Ordering;
-use std::sync::{Condvar, Mutex, RwLock};
+use std::sync::{Arc, Condvar, Mutex, RwLock};
 use std::thread;
 use std::time::Duration;
 use std::time::Instant;
 
 use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
-use compute_api::spec::PgIdent;
+use compute_api::spec::{PgIdent, Role};
 use futures::future::join_all;
 use futures::stream::FuturesUnordered;
 use futures::StreamExt;
@@ -31,15 +32,23 @@ use compute_api::spec::{ComputeFeature, ComputeMode, ComputeSpec, ExtVersion};
 use utils::measured_stream::MeasuredReader;
 
 use nix::sys::signal::{kill, Signal};
-
 use remote_storage::{DownloadError, RemotePath};
+use tokio::spawn;
+use url::Url;
 
-use crate::checker::create_availability_check_data;
 use crate::installed_extensions::get_installed_extensions_sync;
 use crate::local_proxy;
-use crate::logger::inlinify;
 use crate::pg_helpers::*;
 use crate::spec::*;
+use crate::spec_apply::ApplySpecPhase::{
+    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreateSuperUser,
+    DropInvalidDatabases, DropRoles, HandleNeonExtension, HandleOtherExtensions,
+    RenameAndDeleteDatabases, RenameRoles, RunInEachDatabase,
+};
+use crate::spec_apply::PerDatabasePhase::{
+    ChangeSchemaPerms, DeleteDBRoleReferences, HandleAnonExtension,
+};
+use crate::spec_apply::{apply_operations, MutableApplyContext, DB};
 use crate::sync_sk::{check_if_synced, ping_safekeeper};
 use crate::{config, extension_server};
 
@@ -224,10 +233,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
     }
 }
 
-/// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
-/// that we give to customers
-#[instrument(skip_all)]
-fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
+pub(crate) fn construct_superuser_query(spec: &ComputeSpec) -> String {
     let roles = spec
         .cluster
         .roles
@@ -296,11 +302,8 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
             $$;"#,
         roles_decl, database_decl,
     );
-    info!("Neon superuser created: {}", inlinify(&query));
-    client
-        .simple_query(&query)
-        .map_err(|e| anyhow::anyhow!(e).context(query))?;
-    Ok(())
+
+    query
 }
 
 impl ComputeNode {
@@ -813,21 +816,14 @@ impl ComputeNode {
         Ok(())
     }
 
-    /// Do initial configuration of the already started Postgres.
-    #[instrument(skip_all)]
-    pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
-        // If connection fails,
-        // it may be the old node with `zenith_admin` superuser.
-        //
-        // In this case we need to connect with old `zenith_admin` name
-        // and create new user. We cannot simply rename connected user,
-        // but we can create a new one and grant it all privileges.
-        let mut connstr = self.connstr.clone();
+    async fn get_maintenance_client(url: &Url) -> Result<tokio_postgres::Client> {
+        let mut connstr = url.clone();
+
         connstr
             .query_pairs_mut()
             .append_pair("application_name", "apply_config");
 
-        let mut client = match Client::connect(connstr.as_str(), NoTls) {
+        let (client, conn) = match tokio_postgres::connect(connstr.as_str(), NoTls).await {
             Err(e) => match e.code() {
                 Some(&SqlState::INVALID_PASSWORD)
                 | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
@@ -845,8 +841,8 @@ impl ComputeNode {
                     let mut client =
                         Client::connect(zenith_admin_connstr.as_str(), NoTls)
                             .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
-                    // Disable forwarding so that users don't get a cloud_admin role
 
+                    // Disable forwarding so that users don't get a cloud_admin role
                     let mut func = || {
                         client.simple_query("SET neon.forward_ddl = false")?;
                         client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
@@ -858,49 +854,309 @@ impl ComputeNode {
                     drop(client);
 
                     // reconnect with connstring with expected name
-                    Client::connect(connstr.as_str(), NoTls)?
+                    tokio_postgres::connect(connstr.as_str(), NoTls).await?
                 }
                 _ => return Err(e.into()),
             },
-            Ok(client) => client,
+            Ok((client, conn)) => (client, conn),
         };
 
-        // Disable DDL forwarding because control plane already knows about these roles/databases.
+        spawn(async move {
+            if let Err(e) = conn.await {
+                error!("maintenance client connection error: {}", e);
+            }
+        });
+
+        // Disable DDL forwarding because control plane already knows about the roles/databases
+        // we're about to modify.
         client
             .simple_query("SET neon.forward_ddl = false")
+            .await
             .context("apply_config SET neon.forward_ddl = false")?;
 
-        // Proceed with post-startup configuration. Note, that order of operations is important.
-        let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
-        create_neon_superuser(spec, &mut client).context("apply_config create_neon_superuser")?;
-        cleanup_instance(&mut client).context("apply_config cleanup_instance")?;
-        handle_roles(spec, &mut client).context("apply_config handle_roles")?;
-        handle_databases(spec, &mut client).context("apply_config handle_databases")?;
-        handle_role_deletions(spec, connstr.as_str(), &mut client)
-            .context("apply_config handle_role_deletions")?;
-        handle_grants(
-            spec,
-            &mut client,
-            connstr.as_str(),
-            self.has_feature(ComputeFeature::AnonExtension),
-        )
-        .context("apply_config handle_grants")?;
-        handle_extensions(spec, &mut client).context("apply_config handle_extensions")?;
-        handle_extension_neon(&mut client).context("apply_config handle_extension_neon")?;
-        create_availability_check_data(&mut client)
-            .context("apply_config create_availability_check_data")?;
+        Ok(client)
+    }
 
-        // 'Close' connection
-        drop(client);
+    /// Apply the spec to the running PostgreSQL instance.
+    /// The caller can decide to run with multiple clients in parallel, or
+    /// single mode.  Either way, the commands executed will be the same, and
+    /// only commands run in different databases are parallelized.
+    #[instrument(skip_all)]
+    pub fn apply_spec_sql(
+        &self,
+        spec: Arc<ComputeSpec>,
+        url: Arc<Url>,
+        concurrency: usize,
+    ) -> Result<()> {
+        let rt = tokio::runtime::Builder::new_multi_thread()
+            .enable_all()
+            .build()?;
 
-        if let Some(ref local_proxy) = spec.local_proxy_config {
+        info!("Applying config with max {} concurrency", concurrency);
+        debug!("Config: {:?}", spec);
+
+        rt.block_on(async {
+            // Proceed with post-startup configuration. Note, that order of operations is important.
+            let client = Self::get_maintenance_client(&url).await?;
+            let spec = spec.clone();
+
+            let databases = get_existing_dbs_async(&client).await?;
+            let roles = get_existing_roles_async(&client)
+                .await?
+                .into_iter()
+                .map(|role| (role.name.clone(), role))
+                .collect::<HashMap<String, Role>>();
+
+            let jwks_roles = Arc::new(
+                spec.as_ref()
+                    .local_proxy_config
+                    .iter()
+                    .flat_map(|it| &it.jwks)
+                    .flatten()
+                    .flat_map(|setting| &setting.role_names)
+                    .cloned()
+                    .collect::<HashSet<_>>(),
+            );
+
+            let ctx = Arc::new(tokio::sync::RwLock::new(MutableApplyContext {
+                roles,
+                dbs: databases,
+            }));
+
+            for phase in [
+                CreateSuperUser,
+                DropInvalidDatabases,
+                RenameRoles,
+                CreateAndAlterRoles,
+                RenameAndDeleteDatabases,
+                CreateAndAlterDatabases,
+            ] {
+                debug!("Applying phase {:?}", &phase);
+                apply_operations(
+                    spec.clone(),
+                    ctx.clone(),
+                    jwks_roles.clone(),
+                    phase,
+                    || async { Ok(&client) },
+                )
+                .await?;
+            }
+
+            let concurrency_token = Arc::new(tokio::sync::Semaphore::new(concurrency));
+
+            let db_processes = spec
+                .cluster
+                .databases
+                .iter()
+                .map(|db| DB::new(db.clone()))
+                // include
+                .chain(once(DB::SystemDB))
+                .map(|db| {
+                    let spec = spec.clone();
+                    let ctx = ctx.clone();
+                    let jwks_roles = jwks_roles.clone();
+                    let mut url = url.as_ref().clone();
+                    let concurrency_token = concurrency_token.clone();
+                    let db = db.clone();
+
+                    debug!("Applying per-database phases for Database {:?}", &db);
+
+                    match &db {
+                        DB::SystemDB => {}
+                        DB::UserDB(db) => {
+                            url.set_path(db.name.as_str());
+                        }
+                    }
+
+                    let url = Arc::new(url);
+                    let fut = Self::apply_spec_sql_db(
+                        spec.clone(),
+                        url,
+                        ctx.clone(),
+                        jwks_roles.clone(),
+                        concurrency_token.clone(),
+                        db,
+                    );
+
+                    Ok(spawn(fut))
+                })
+                .collect::<Vec<Result<_, anyhow::Error>>>();
+
+            for process in db_processes.into_iter() {
+                let handle = process?;
+                handle.await??;
+            }
+
+            for phase in vec![
+                HandleOtherExtensions,
+                HandleNeonExtension,
+                CreateAvailabilityCheck,
+                DropRoles,
+            ] {
+                debug!("Applying phase {:?}", &phase);
+                apply_operations(
+                    spec.clone(),
+                    ctx.clone(),
+                    jwks_roles.clone(),
+                    phase,
+                    || async { Ok(&client) },
+                )
+                .await?;
+            }
+
+            Ok::<(), anyhow::Error>(())
+        })?;
+
+        Ok(())
+    }
+
+    /// Apply SQL migrations of the RunInEachDatabase phase.
+    ///
+    /// May opt to not connect to databases that don't have any scheduled
+    /// operations.  The function is concurrency-controlled with the provided
+    /// semaphore.  The caller has to make sure the semaphore isn't exhausted.
+    async fn apply_spec_sql_db(
+        spec: Arc<ComputeSpec>,
+        url: Arc<Url>,
+        ctx: Arc<tokio::sync::RwLock<MutableApplyContext>>,
+        jwks_roles: Arc<HashSet<String>>,
+        concurrency_token: Arc<tokio::sync::Semaphore>,
+        db: DB,
+    ) -> Result<()> {
+        let _permit = concurrency_token.acquire().await?;
+
+        let mut client_conn = None;
+
+        for subphase in [
+            DeleteDBRoleReferences,
+            ChangeSchemaPerms,
+            HandleAnonExtension,
+        ] {
+            apply_operations(
+                spec.clone(),
+                ctx.clone(),
+                jwks_roles.clone(),
+                RunInEachDatabase {
+                    db: db.clone(),
+                    subphase,
+                },
+                // Only connect if apply_operation actually wants a connection.
+                // It's quite possible this database doesn't need any queries,
+                // so by not connecting we save time and effort connecting to
+                // that database.
+                || async {
+                    if client_conn.is_none() {
+                        let db_client = Self::get_maintenance_client(&url).await?;
+                        client_conn.replace(db_client);
+                    }
+                    let client = client_conn.as_ref().unwrap();
+                    Ok(client)
+                },
+            )
+            .await?;
+        }
+
+        drop(client_conn);
+
+        Ok::<(), anyhow::Error>(())
+    }
+
+    /// Do initial configuration of the already started Postgres.
+    #[instrument(skip_all)]
+    pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
+        // If connection fails,
+        // it may be the old node with `zenith_admin` superuser.
+        //
+        // In this case we need to connect with old `zenith_admin` name
+        // and create new user. We cannot simply rename connected user,
+        // but we can create a new one and grant it all privileges.
+        let mut url = self.connstr.clone();
+        url.query_pairs_mut()
+            .append_pair("application_name", "apply_config");
+
+        let url = Arc::new(url);
+        let spec = Arc::new(
+            compute_state
+                .pspec
+                .as_ref()
+                .expect("spec must be set")
+                .spec
+                .clone(),
+        );
+
+        // Choose how many concurrent connections to use for applying the spec changes.
+        // If the cluster is not currently Running we don't have to deal with user connections,
+        // and can thus use all `max_connections` connection slots. However, that's generally not
+        // very efficient, so we generally still limit it to a smaller number.
+        let max_concurrent_connections = if compute_state.status != ComputeStatus::Running {
+            // If the settings contain 'max_connections', use that as template
+            if let Some(config) = spec.cluster.settings.find("max_connections") {
+                config.parse::<usize>().ok()
+            } else {
+                // Otherwise, try to find the setting in the postgresql_conf string
+                spec.cluster
+                    .postgresql_conf
+                    .iter()
+                    .flat_map(|conf| conf.split("\n"))
+                    .filter_map(|line| {
+                        if !line.contains("max_connections") {
+                            return None;
+                        }
+
+                        let (key, value) = line.split_once("=")?;
+                        let key = key
+                            .trim_start_matches(char::is_whitespace)
+                            .trim_end_matches(char::is_whitespace);
+
+                        let value = value
+                            .trim_start_matches(char::is_whitespace)
+                            .trim_end_matches(char::is_whitespace);
+
+                        if key != "max_connections" {
+                            return None;
+                        }
+
+                        value.parse::<usize>().ok()
+                    })
+                    .next()
+            }
+            // If max_connections is present, use at most 1/3rd of that.
+            // When max_connections is lower than 30, try to use at least 10 connections, but
+            // never more than max_connections.
+            .map(|limit| match limit {
+                0..10 => limit,
+                10..30 => 10,
+                30.. => limit / 3,
+            })
+            // If we didn't find max_connections, default to 10 concurrent connections.
+            .unwrap_or(10)
+        } else {
+            // state == Running
+            // Because the cluster is already in the Running state, we should assume users are
+            // already connected to the cluster, and high concurrency could negatively
+            // impact user connectivity. Therefore, we can limit concurrency to the number of
+            // reserved superuser connections, which users wouldn't be able to use anyway.
+            spec.cluster
+                .settings
+                .find("superuser_reserved_connections")
+                .iter()
+                .filter_map(|val| val.parse::<usize>().ok())
+                .map(|val| if val > 1 { val - 1 } else { 1 })
+                .last()
+                .unwrap_or(3)
+        };
+
+        // Merge-apply spec & changes to PostgreSQL state.
+        self.apply_spec_sql(spec.clone(), url.clone(), max_concurrent_connections)?;
+
+        if let Some(ref local_proxy) = &spec.clone().local_proxy_config {
             info!("configuring local_proxy");
             local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
         }
 
         // Run migrations separately to not hold up cold starts
         thread::spawn(move || {
-            let mut connstr = connstr.clone();
+            let mut connstr = url.as_ref().clone();
             connstr
                 .query_pairs_mut()
                 .append_pair("application_name", "migrations");
@@ -908,7 +1164,8 @@ impl ComputeNode {
             let mut client = Client::connect(connstr.as_str(), NoTls)?;
             handle_migrations(&mut client).context("apply_config handle_migrations")
         });
-        Ok(())
+
+        Ok::<(), anyhow::Error>(())
     }
 
     // Wrapped this around `pg_ctl reload`, but right now we don't use
@@ -971,32 +1228,16 @@ impl ComputeNode {
         config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
             self.pg_reload_conf()?;
 
-            let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
-
-            // Proceed with post-startup configuration. Note, that order of operations is important.
-            // Disable DDL forwarding because control plane already knows about these roles/databases.
             if spec.mode == ComputeMode::Primary {
-                client.simple_query("SET neon.forward_ddl = false")?;
-                cleanup_instance(&mut client)?;
-                handle_roles(&spec, &mut client)?;
-                handle_databases(&spec, &mut client)?;
-                handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-                handle_grants(
-                    &spec,
-                    &mut client,
-                    self.connstr.as_str(),
-                    self.has_feature(ComputeFeature::AnonExtension),
-                )?;
-                handle_extensions(&spec, &mut client)?;
-                handle_extension_neon(&mut client)?;
-                // We can skip handle_migrations here because a new migration can only appear
-                // if we have a new version of the compute_ctl binary, which can only happen
-                // if compute got restarted, in which case we'll end up inside of apply_config
-                // instead of reconfigure.
-            }
+                let mut url = self.connstr.clone();
+                url.query_pairs_mut()
+                    .append_pair("application_name", "apply_config");
+                let url = Arc::new(url);
 
-            // 'Close' connection
-            drop(client);
+                let spec = Arc::new(spec.clone());
+
+                self.apply_spec_sql(spec, url, 1)?;
+            }
 
             Ok(())
         })?;

compute_tools/src/config.rs

@@ -74,10 +74,17 @@ pub fn write_postgres_conf(
     }
 
     // Locales
-    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'")?;
+    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 => {}

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()
+}

compute_tools/src/lib.rs

@@ -23,5 +23,6 @@ pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
 pub mod spec;
+mod spec_apply;
 pub mod swap;
 pub mod sync_sk;

compute_tools/src/pg_helpers.rs

@@ -10,9 +10,9 @@ use std::thread::JoinHandle;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Result};
+use futures::StreamExt;
 use ini::Ini;
 use notify::{RecursiveMode, Watcher};
-use postgres::{Client, Transaction};
 use tokio::io::AsyncBufReadExt;
 use tokio::time::timeout;
 use tokio_postgres::NoTls;
@@ -197,27 +197,34 @@ impl Escaping for PgIdent {
 }
 
 /// Build a list of existing Postgres roles
-pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
-    let postgres_roles = xact
-        .query("SELECT rolname, rolpassword FROM pg_catalog.pg_authid", &[])?
-        .iter()
+pub async fn get_existing_roles_async(client: &tokio_postgres::Client) -> Result<Vec<Role>> {
+    let postgres_roles = client
+        .query_raw::<str, &String, &[String; 0]>(
+            "SELECT rolname, rolpassword FROM pg_catalog.pg_authid",
+            &[],
+        )
+        .await?
+        .filter_map(|row| async { row.ok() })
         .map(|row| Role {
             name: row.get("rolname"),
             encrypted_password: row.get("rolpassword"),
             options: None,
         })
-        .collect();
+        .collect()
+        .await;
 
     Ok(postgres_roles)
 }
 
 /// Build a list of existing Postgres databases
-pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>> {
+pub async fn get_existing_dbs_async(
+    client: &tokio_postgres::Client,
+) -> Result<HashMap<String, Database>> {
     // `pg_database.datconnlimit = -2` means that the database is in the
     // invalid state. See:
     //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
-    let postgres_dbs: Vec<Database> = client
-        .query(
+    let rowstream = client
+        .query_raw::<str, &String, &[String; 0]>(
             "SELECT
                 datname AS name,
                 datdba::regrole::text AS owner,
@@ -226,8 +233,11 @@ pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>
             FROM
                 pg_catalog.pg_database;",
             &[],
-        )?
-        .iter()
+        )
+        .await?;
+
+    let dbs_map = rowstream
+        .filter_map(|r| async { r.ok() })
         .map(|row| Database {
             name: row.get("name"),
             owner: row.get("owner"),
@@ -235,12 +245,9 @@ pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>
             invalid: row.get("invalid"),
             options: None,
         })
-        .collect();
-
-    let dbs_map = postgres_dbs
-        .iter()
         .map(|db| (db.name.clone(), db.clone()))
-        .collect::<HashMap<_, _>>();
+        .collect::<HashMap<_, _>>()
+        .await;
 
     Ok(dbs_map)
 }

compute_tools/src/spec.rs

@@ -1,22 +1,17 @@
-use std::collections::HashSet;
+use anyhow::{anyhow, bail, Result};
+use postgres::Client;
+use reqwest::StatusCode;
 use std::fs::File;
 use std::path::Path;
-use std::str::FromStr;
-
-use anyhow::{anyhow, bail, Context, Result};
-use postgres::config::Config;
-use postgres::{Client, NoTls};
-use reqwest::StatusCode;
-use tracing::{error, info, info_span, instrument, span_enabled, warn, Level};
+use tracing::{error, info, instrument, warn};
 
 use crate::config;
-use crate::logger::inlinify;
 use crate::migration::MigrationRunner;
 use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
-use compute_api::spec::{ComputeSpec, PgIdent, Role};
+use compute_api::spec::ComputeSpec;
 
 // Do control plane request and return response if any. In case of error it
 // returns a bool flag indicating whether it makes sense to retry the request
@@ -151,625 +146,6 @@ pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
-/// Compute could be unexpectedly shut down, for example, during the
-/// database dropping. This leaves the database in the invalid state,
-/// which prevents new db creation with the same name. This function
-/// will clean it up before proceeding with catalog updates. All
-/// possible future cleanup operations may go here too.
-#[instrument(skip_all)]
-pub fn cleanup_instance(client: &mut Client) -> Result<()> {
-    let existing_dbs = get_existing_dbs(client)?;
-
-    for (_, db) in existing_dbs {
-        if db.invalid {
-            // After recent commit in Postgres, interrupted DROP DATABASE
-            // leaves the database in the invalid state. According to the
-            // commit message, the only option for user is to drop it again.
-            // See:
-            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
-            //
-            // Postgres Neon extension is done the way, that db is de-registered
-            // in the control plane metadata only after it is dropped. So there is
-            // a chance that it still thinks that db should exist. This means
-            // that it will be re-created by `handle_databases()`. Yet, it's fine
-            // as user can just repeat drop (in vanilla Postgres they would need
-            // to do the same, btw).
-            let query = format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote());
-            info!("dropping invalid database {}", db.name);
-            client.execute(query.as_str(), &[])?;
-        }
-    }
-
-    Ok(())
-}
-
-/// Given a cluster spec json and open transaction it handles roles creation,
-/// deletion and update.
-#[instrument(skip_all)]
-pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let mut xact = client.transaction()?;
-    let existing_roles: Vec<Role> = get_existing_roles(&mut xact)?;
-
-    let mut jwks_roles = HashSet::new();
-    if let Some(local_proxy) = &spec.local_proxy_config {
-        for jwks_setting in local_proxy.jwks.iter().flatten() {
-            for role_name in &jwks_setting.role_names {
-                jwks_roles.insert(role_name.clone());
-            }
-        }
-    }
-
-    // Print a list of existing Postgres roles (only in debug mode)
-    if span_enabled!(Level::INFO) {
-        let mut vec = Vec::new();
-        for r in &existing_roles {
-            vec.push(format!(
-                "{}:{}",
-                r.name,
-                if r.encrypted_password.is_some() {
-                    "[FILTERED]"
-                } else {
-                    "(null)"
-                }
-            ));
-        }
-
-        info!("postgres roles (total {}): {:?}", vec.len(), vec);
-    }
-
-    // Process delta operations first
-    if let Some(ops) = &spec.delta_operations {
-        info!("processing role renames");
-        for op in ops {
-            match op.action.as_ref() {
-                "delete_role" => {
-                    // no-op now, roles will be deleted at the end of configuration
-                }
-                // Renaming role drops its password, since role name is
-                // used as a salt there.  It is important that this role
-                // is recorded with a new `name` in the `roles` list.
-                // Follow up roles update will set the new password.
-                "rename_role" => {
-                    let new_name = op.new_name.as_ref().unwrap();
-
-                    // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-                    if existing_roles.iter().any(|r| r.name == op.name) {
-                        let query: String = format!(
-                            "ALTER ROLE {} RENAME TO {}",
-                            op.name.pg_quote(),
-                            new_name.pg_quote()
-                        );
-
-                        warn!("renaming role '{}' to '{}'", op.name, new_name);
-                        xact.execute(query.as_str(), &[])?;
-                    }
-                }
-                _ => {}
-            }
-        }
-    }
-
-    // Refresh Postgres roles info to handle possible roles renaming
-    let existing_roles: Vec<Role> = get_existing_roles(&mut xact)?;
-
-    info!(
-        "handling cluster spec roles (total {})",
-        spec.cluster.roles.len()
-    );
-    for role in &spec.cluster.roles {
-        let name = &role.name;
-        // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-        let pg_role = existing_roles.iter().find(|r| r.name == *name);
-
-        enum RoleAction {
-            None,
-            Update,
-            Create,
-        }
-        let action = if let Some(r) = pg_role {
-            if (r.encrypted_password.is_none() && role.encrypted_password.is_some())
-                || (r.encrypted_password.is_some() && role.encrypted_password.is_none())
-            {
-                RoleAction::Update
-            } else if let Some(pg_pwd) = &r.encrypted_password {
-                // Check whether password changed or not (trim 'md5' prefix first if any)
-                //
-                // This is a backward compatibility hack, which comes from the times when we were using
-                // md5 for everyone and hashes were stored in the console db without md5 prefix. So when
-                // role comes from the control-plane (json spec) `Role.encrypted_password` doesn't have md5 prefix,
-                // but when role comes from Postgres (`get_existing_roles` / `existing_roles`) it has this prefix.
-                // Here is the only place so far where we compare hashes, so it seems to be the best candidate
-                // to place this compatibility layer.
-                let pg_pwd = if let Some(stripped) = pg_pwd.strip_prefix("md5") {
-                    stripped
-                } else {
-                    pg_pwd
-                };
-                if pg_pwd != *role.encrypted_password.as_ref().unwrap() {
-                    RoleAction::Update
-                } else {
-                    RoleAction::None
-                }
-            } else {
-                RoleAction::None
-            }
-        } else {
-            RoleAction::Create
-        };
-
-        match action {
-            RoleAction::None => {}
-            RoleAction::Update => {
-                // This can be run on /every/ role! Not just ones created through the console.
-                // This means that if you add some funny ALTER here that adds a permission,
-                // this will get run even on user-created roles! This will result in different
-                // behavior before and after a spec gets reapplied. The below ALTER as it stands
-                // now only grants LOGIN and changes the password. Please do not allow this branch
-                // to do anything silly.
-                let mut query: String = format!("ALTER ROLE {} ", name.pg_quote());
-                query.push_str(&role.to_pg_options());
-                xact.execute(query.as_str(), &[])?;
-            }
-            RoleAction::Create => {
-                // This branch only runs when roles are created through the console, so it is
-                // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
-                // from neon_superuser.
-                let mut query: String = format!(
-                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
-                    name.pg_quote()
-                );
-                if jwks_roles.contains(name.as_str()) {
-                    query = format!("CREATE ROLE {}", name.pg_quote());
-                }
-                info!("running role create query: '{}'", &query);
-                query.push_str(&role.to_pg_options());
-                xact.execute(query.as_str(), &[])?;
-            }
-        }
-
-        if span_enabled!(Level::INFO) {
-            let pwd = if role.encrypted_password.is_some() {
-                "[FILTERED]"
-            } else {
-                "(null)"
-            };
-            let action_str = match action {
-                RoleAction::None => "",
-                RoleAction::Create => " -> create",
-                RoleAction::Update => " -> update",
-            };
-            info!(" - {}:{}{}", name, pwd, action_str);
-        }
-    }
-
-    xact.commit()?;
-
-    Ok(())
-}
-
-/// Reassign all dependent objects and delete requested roles.
-#[instrument(skip_all)]
-pub fn handle_role_deletions(spec: &ComputeSpec, connstr: &str, client: &mut Client) -> Result<()> {
-    if let Some(ops) = &spec.delta_operations {
-        // First, reassign all dependent objects to db owners.
-        info!("reassigning dependent objects of to-be-deleted roles");
-
-        // Fetch existing roles. We could've exported and used `existing_roles` from
-        // `handle_roles()`, but we only make this list there before creating new roles.
-        // Which is probably fine as we never create to-be-deleted roles, but that'd
-        // just look a bit untidy. Anyway, the entire `pg_roles` should be in shared
-        // buffers already, so this shouldn't be a big deal.
-        let mut xact = client.transaction()?;
-        let existing_roles: Vec<Role> = get_existing_roles(&mut xact)?;
-        xact.commit()?;
-
-        for op in ops {
-            // Check that role is still present in Postgres, as this could be a
-            // restart with the same spec after role deletion.
-            if op.action == "delete_role" && existing_roles.iter().any(|r| r.name == op.name) {
-                reassign_owned_objects(spec, connstr, &op.name)?;
-            }
-        }
-
-        // Second, proceed with role deletions.
-        info!("processing role deletions");
-        let mut xact = client.transaction()?;
-        for op in ops {
-            // We do not check either role exists or not,
-            // Postgres will take care of it for us
-            if op.action == "delete_role" {
-                let query: String = format!("DROP ROLE IF EXISTS {}", &op.name.pg_quote());
-
-                warn!("deleting role '{}'", &op.name);
-                xact.execute(query.as_str(), &[])?;
-            }
-        }
-        xact.commit()?;
-    }
-
-    Ok(())
-}
-
-fn reassign_owned_objects_in_one_db(
-    conf: Config,
-    role_name: &PgIdent,
-    db_owner: &PgIdent,
-) -> Result<()> {
-    let mut client = conf.connect(NoTls)?;
-
-    // This will reassign all dependent objects to the db owner
-    let reassign_query = format!(
-        "REASSIGN OWNED BY {} TO {}",
-        role_name.pg_quote(),
-        db_owner.pg_quote()
-    );
-    info!(
-        "reassigning objects owned by '{}' in db '{}' to '{}'",
-        role_name,
-        conf.get_dbname().unwrap_or(""),
-        db_owner
-    );
-    client.simple_query(&reassign_query)?;
-
-    // This now will only drop privileges of the role
-    let drop_query = format!("DROP OWNED BY {}", role_name.pg_quote());
-    client.simple_query(&drop_query)?;
-    Ok(())
-}
-
-// Reassign all owned objects in all databases to the owner of the database.
-fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent) -> Result<()> {
-    for db in &spec.cluster.databases {
-        if db.owner != *role_name {
-            let mut conf = Config::from_str(connstr)?;
-            conf.dbname(&db.name);
-            reassign_owned_objects_in_one_db(conf, role_name, &db.owner)?;
-        }
-    }
-
-    // Also handle case when there are no databases in the spec.
-    // In this case we need to reassign objects in the default database.
-    let conf = Config::from_str(connstr)?;
-    let db_owner = PgIdent::from_str("cloud_admin")?;
-    reassign_owned_objects_in_one_db(conf, role_name, &db_owner)?;
-
-    Ok(())
-}
-
-/// It follows mostly the same logic as `handle_roles()` excepting that we
-/// does not use an explicit transactions block, since major database operations
-/// like `CREATE DATABASE` and `DROP DATABASE` do not support it. Statement-level
-/// atomicity should be enough here due to the order of operations and various checks,
-/// which together provide us idempotency.
-#[instrument(skip_all)]
-pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let existing_dbs = get_existing_dbs(client)?;
-
-    // Print a list of existing Postgres databases (only in debug mode)
-    if span_enabled!(Level::INFO) {
-        let mut vec = Vec::new();
-        for (dbname, db) in &existing_dbs {
-            vec.push(format!("{}:{}", dbname, db.owner));
-        }
-        info!("postgres databases (total {}): {:?}", vec.len(), vec);
-    }
-
-    // Process delta operations first
-    if let Some(ops) = &spec.delta_operations {
-        info!("processing delta operations on databases");
-        for op in ops {
-            match op.action.as_ref() {
-                // We do not check either DB exists or not,
-                // Postgres will take care of it for us
-                "delete_db" => {
-                    // In Postgres we can't drop a database if it is a template.
-                    // So we need to unset the template flag first, but it could
-                    // be a retry, so we could've already dropped the database.
-                    // Check that database exists first to make it idempotent.
-                    let unset_template_query: String = format!(
-                        "
-                        DO $$
-                        BEGIN
-                            IF EXISTS(
-                                SELECT 1
-                                FROM pg_catalog.pg_database
-                                WHERE datname = {}
-                            )
-                            THEN
-                            ALTER DATABASE {} is_template false;
-                            END IF;
-                        END
-                        $$;",
-                        escape_literal(&op.name),
-                        &op.name.pg_quote()
-                    );
-                    // Use FORCE to drop database even if there are active connections.
-                    // We run this from `cloud_admin`, so it should have enough privileges.
-                    // NB: there could be other db states, which prevent us from dropping
-                    // the database. For example, if db is used by any active subscription
-                    // or replication slot.
-                    // TODO: deal with it once we allow logical replication. Proper fix should
-                    // involve returning an error code to the control plane, so it could
-                    // figure out that this is a non-retryable error, return it to the user
-                    // and fail operation permanently.
-                    let drop_db_query: String = format!(
-                        "DROP DATABASE IF EXISTS {} WITH (FORCE)",
-                        &op.name.pg_quote()
-                    );
-
-                    warn!("deleting database '{}'", &op.name);
-                    client.execute(unset_template_query.as_str(), &[])?;
-                    client.execute(drop_db_query.as_str(), &[])?;
-                }
-                "rename_db" => {
-                    let new_name = op.new_name.as_ref().unwrap();
-
-                    if existing_dbs.contains_key(&op.name) {
-                        let query: String = format!(
-                            "ALTER DATABASE {} RENAME TO {}",
-                            op.name.pg_quote(),
-                            new_name.pg_quote()
-                        );
-
-                        warn!("renaming database '{}' to '{}'", op.name, new_name);
-                        client.execute(query.as_str(), &[])?;
-                    }
-                }
-                _ => {}
-            }
-        }
-    }
-
-    // Refresh Postgres databases info to handle possible renames
-    let existing_dbs = get_existing_dbs(client)?;
-
-    info!(
-        "handling cluster spec databases (total {})",
-        spec.cluster.databases.len()
-    );
-    for db in &spec.cluster.databases {
-        let name = &db.name;
-        let pg_db = existing_dbs.get(name);
-
-        enum DatabaseAction {
-            None,
-            Update,
-            Create,
-        }
-        let action = if let Some(r) = pg_db {
-            // XXX: db owner name is returned as quoted string from Postgres,
-            // when quoting is needed.
-            let new_owner = if r.owner.starts_with('"') {
-                db.owner.pg_quote()
-            } else {
-                db.owner.clone()
-            };
-
-            if new_owner != r.owner {
-                // Update the owner
-                DatabaseAction::Update
-            } else {
-                DatabaseAction::None
-            }
-        } else {
-            DatabaseAction::Create
-        };
-
-        match action {
-            DatabaseAction::None => {}
-            DatabaseAction::Update => {
-                let query: String = format!(
-                    "ALTER DATABASE {} OWNER TO {}",
-                    name.pg_quote(),
-                    db.owner.pg_quote()
-                );
-                let _guard = info_span!("executing", query).entered();
-                client.execute(query.as_str(), &[])?;
-            }
-            DatabaseAction::Create => {
-                let mut query: String = format!("CREATE DATABASE {} ", name.pg_quote());
-                query.push_str(&db.to_pg_options());
-                let _guard = info_span!("executing", query).entered();
-                client.execute(query.as_str(), &[])?;
-                let grant_query: String = format!(
-                    "GRANT ALL PRIVILEGES ON DATABASE {} TO neon_superuser",
-                    name.pg_quote()
-                );
-                client.execute(grant_query.as_str(), &[])?;
-            }
-        };
-
-        if span_enabled!(Level::INFO) {
-            let action_str = match action {
-                DatabaseAction::None => "",
-                DatabaseAction::Create => " -> create",
-                DatabaseAction::Update => " -> update",
-            };
-            info!(" - {}:{}{}", db.name, db.owner, action_str);
-        }
-    }
-
-    Ok(())
-}
-
-/// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
-/// to allow users creating trusted extensions and re-creating `public` schema, for example.
-#[instrument(skip_all)]
-pub fn handle_grants(
-    spec: &ComputeSpec,
-    client: &mut Client,
-    connstr: &str,
-    enable_anon_extension: bool,
-) -> Result<()> {
-    info!("modifying database permissions");
-    let existing_dbs = get_existing_dbs(client)?;
-
-    // Do some per-database access adjustments. We'd better do this at db creation time,
-    // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
-    // atomically.
-    for db in &spec.cluster.databases {
-        match existing_dbs.get(&db.name) {
-            Some(pg_db) => {
-                if pg_db.restrict_conn || pg_db.invalid {
-                    info!(
-                        "skipping grants for db {} (invalid: {}, connections not allowed: {})",
-                        db.name, pg_db.invalid, pg_db.restrict_conn
-                    );
-                    continue;
-                }
-            }
-            None => {
-                bail!(
-                    "database {} doesn't exist in Postgres after handle_databases()",
-                    db.name
-                );
-            }
-        }
-
-        let mut conf = Config::from_str(connstr)?;
-        conf.dbname(&db.name);
-
-        let mut db_client = conf.connect(NoTls)?;
-
-        // This will only change ownership on the schema itself, not the objects
-        // inside it. Without it owner of the `public` schema will be `cloud_admin`
-        // and database owner cannot do anything with it. SQL procedure ensures
-        // that it won't error out if schema `public` doesn't exist.
-        let alter_query = format!(
-            "DO $$\n\
-                DECLARE\n\
-                    schema_owner TEXT;\n\
-                BEGIN\n\
-                    IF EXISTS(\n\
-                        SELECT nspname\n\
-                        FROM pg_catalog.pg_namespace\n\
-                        WHERE nspname = 'public'\n\
-                    )\n\
-                    THEN\n\
-                        SELECT nspowner::regrole::text\n\
-                            FROM pg_catalog.pg_namespace\n\
-                            WHERE nspname = 'public'\n\
-                            INTO schema_owner;\n\
-                \n\
-                        IF schema_owner = 'cloud_admin' OR schema_owner = 'zenith_admin'\n\
-                        THEN\n\
-                            ALTER SCHEMA public OWNER TO {};\n\
-                        END IF;\n\
-                    END IF;\n\
-                END\n\
-            $$;",
-            db.owner.pg_quote()
-        );
-        db_client.simple_query(&alter_query)?;
-
-        // Explicitly grant CREATE ON SCHEMA PUBLIC to the web_access user.
-        // This is needed because since postgres 15 this privilege is removed by default.
-        // TODO: web_access isn't created for almost 1 year. It could be that we have
-        // active users of 1 year old projects, but hopefully not, so check it and
-        // remove this code if possible. The worst thing that could happen is that
-        // user won't be able to use public schema in NEW databases created in the
-        // very OLD project.
-        //
-        // Also, alter default permissions so that relations created by extensions can be
-        // used by neon_superuser without permission issues.
-        let grant_query = "DO $$\n\
-                BEGIN\n\
-                    IF EXISTS(\n\
-                        SELECT nspname\n\
-                        FROM pg_catalog.pg_namespace\n\
-                        WHERE nspname = 'public'\n\
-                    ) AND\n\
-                    current_setting('server_version_num')::int/10000 >= 15\n\
-                    THEN\n\
-                        IF EXISTS(\n\
-                            SELECT rolname\n\
-                            FROM pg_catalog.pg_roles\n\
-                            WHERE rolname = 'web_access'\n\
-                        )\n\
-                        THEN\n\
-                            GRANT CREATE ON SCHEMA public TO web_access;\n\
-                        END IF;\n\
-                    END IF;\n\
-                    IF EXISTS(\n\
-                        SELECT nspname\n\
-                        FROM pg_catalog.pg_namespace\n\
-                        WHERE nspname = 'public'\n\
-                    )\n\
-                    THEN\n\
-                        ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser WITH GRANT OPTION;\n\
-                        ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser WITH GRANT OPTION;\n\
-                    END IF;\n\
-                END\n\
-            $$;"
-        .to_string();
-
-        info!(
-            "grant query for db {} : {}",
-            &db.name,
-            inlinify(&grant_query)
-        );
-        db_client.simple_query(&grant_query)?;
-
-        // it is important to run this after all grants
-        if enable_anon_extension {
-            handle_extension_anon(spec, &db.owner, &mut db_client, false)
-                .context("handle_grants handle_extension_anon")?;
-        }
-    }
-
-    Ok(())
-}
-
-/// Create required system extensions
-#[instrument(skip_all)]
-pub fn handle_extensions(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
-        if libs.contains("pg_stat_statements") {
-            // Create extension only if this compute really needs it
-            let query = "CREATE EXTENSION IF NOT EXISTS pg_stat_statements";
-            info!("creating system extensions with query: {}", query);
-            client.simple_query(query)?;
-        }
-    }
-
-    Ok(())
-}
-
-/// Run CREATE and ALTER EXTENSION neon UPDATE for postgres database
-#[instrument(skip_all)]
-pub fn handle_extension_neon(client: &mut Client) -> Result<()> {
-    info!("handle extension neon");
-
-    let mut query = "CREATE SCHEMA IF NOT EXISTS neon";
-    client.simple_query(query)?;
-
-    query = "CREATE EXTENSION IF NOT EXISTS neon WITH SCHEMA neon";
-    info!("create neon extension with query: {}", query);
-    client.simple_query(query)?;
-
-    query = "UPDATE pg_extension SET extrelocatable = true WHERE extname = 'neon'";
-    client.simple_query(query)?;
-
-    query = "ALTER EXTENSION neon SET SCHEMA neon";
-    info!("alter neon extension schema with query: {}", query);
-    client.simple_query(query)?;
-
-    // this will be a no-op if extension is already up to date,
-    // which may happen in two cases:
-    // - extension was just installed
-    // - extension was already installed and is up to date
-    let query = "ALTER EXTENSION neon UPDATE";
-    info!("update neon extension version with query: {}", query);
-    if let Err(e) = client.simple_query(query) {
-        error!(
-            "failed to upgrade neon extension during `handle_extension_neon`: {}",
-            e
-        );
-    }
-
-    Ok(())
-}
-
 #[instrument(skip_all)]
 pub fn handle_neon_extension_upgrade(client: &mut Client) -> Result<()> {
     info!("handle neon extension upgrade");

compute_tools/src/spec_apply.rs

@@ -0,0 +1,680 @@
+use std::collections::{HashMap, HashSet};
+use std::fmt::{Debug, Formatter};
+use std::future::Future;
+use std::iter::empty;
+use std::iter::once;
+use std::sync::Arc;
+
+use crate::compute::construct_superuser_query;
+use crate::pg_helpers::{escape_literal, DatabaseExt, Escaping, GenericOptionsSearch, RoleExt};
+use anyhow::{bail, Result};
+use compute_api::spec::{ComputeFeature, ComputeSpec, Database, PgIdent, Role};
+use futures::future::join_all;
+use tokio::sync::RwLock;
+use tokio_postgres::Client;
+use tracing::{debug, info_span, Instrument};
+
+#[derive(Clone)]
+pub enum DB {
+    SystemDB,
+    UserDB(Database),
+}
+
+impl DB {
+    pub fn new(db: Database) -> DB {
+        Self::UserDB(db)
+    }
+
+    pub fn is_owned_by(&self, role: &PgIdent) -> bool {
+        match self {
+            DB::SystemDB => false,
+            DB::UserDB(db) => &db.owner == role,
+        }
+    }
+}
+
+impl Debug for DB {
+    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+        match self {
+            DB::SystemDB => f.debug_tuple("SystemDB").finish(),
+            DB::UserDB(db) => f.debug_tuple("UserDB").field(&db.name).finish(),
+        }
+    }
+}
+
+#[derive(Copy, Clone, Debug)]
+pub enum PerDatabasePhase {
+    DeleteDBRoleReferences,
+    ChangeSchemaPerms,
+    HandleAnonExtension,
+}
+
+#[derive(Clone, Debug)]
+pub enum ApplySpecPhase {
+    CreateSuperUser,
+    DropInvalidDatabases,
+    RenameRoles,
+    CreateAndAlterRoles,
+    RenameAndDeleteDatabases,
+    CreateAndAlterDatabases,
+    RunInEachDatabase { db: DB, subphase: PerDatabasePhase },
+    HandleOtherExtensions,
+    HandleNeonExtension,
+    CreateAvailabilityCheck,
+    DropRoles,
+}
+
+pub struct Operation {
+    pub query: String,
+    pub comment: Option<String>,
+}
+
+pub struct MutableApplyContext {
+    pub roles: HashMap<String, Role>,
+    pub dbs: HashMap<String, Database>,
+}
+
+/// Appply the operations that belong to the given spec apply phase.
+///
+/// Commands within a single phase are executed in order of Iterator yield.
+/// Commands of ApplySpecPhase::RunInEachDatabase will execute in the database
+/// indicated by its `db` field, and can share a single client for all changes
+/// to that database.
+///
+/// Notes:
+/// - Commands are pipelined, and thus may cause incomplete apply if one
+///   command of many fails.
+/// - Failing commands will fail the phase's apply step once the return value
+///   is processed.
+/// - No timeouts have (yet) been implemented.
+/// - The caller is responsible for limiting and/or applying concurrency.
+pub async fn apply_operations<'a, Fut, F>(
+    spec: Arc<ComputeSpec>,
+    ctx: Arc<RwLock<MutableApplyContext>>,
+    jwks_roles: Arc<HashSet<String>>,
+    apply_spec_phase: ApplySpecPhase,
+    client: F,
+) -> Result<()>
+where
+    F: FnOnce() -> Fut,
+    Fut: Future<Output = Result<&'a Client>>,
+{
+    debug!("Starting phase {:?}", &apply_spec_phase);
+    let span = info_span!("db_apply_changes", phase=?apply_spec_phase);
+    let span2 = span.clone();
+    async move {
+        debug!("Processing phase {:?}", &apply_spec_phase);
+        let ctx = ctx;
+
+        let mut ops = get_operations(&spec, &ctx, &jwks_roles, &apply_spec_phase)
+            .await?
+            .peekable();
+
+        // Return (and by doing so, skip requesting the PostgreSQL client) if
+        // we don't have any operations scheduled.
+        if ops.peek().is_none() {
+            return Ok(());
+        }
+
+        let client = client().await?;
+
+        debug!("Applying phase {:?}", &apply_spec_phase);
+
+        let active_queries = ops
+            .map(|op| {
+                let Operation { comment, query } = op;
+                let inspan = match comment {
+                    None => span.clone(),
+                    Some(comment) => info_span!("phase {}: {}", comment),
+                };
+
+                async {
+                    let query = query;
+                    let res = client.simple_query(&query).await;
+                    debug!(
+                        "{} {}",
+                        if res.is_ok() {
+                            "successfully executed"
+                        } else {
+                            "failed to execute"
+                        },
+                        query
+                    );
+                    res
+                }
+                .instrument(inspan)
+            })
+            .collect::<Vec<_>>();
+
+        drop(ctx);
+
+        for it in join_all(active_queries).await {
+            drop(it?);
+        }
+
+        debug!("Completed phase {:?}", &apply_spec_phase);
+
+        Ok(())
+    }
+    .instrument(span2)
+    .await
+}
+
+/// Create a stream of operations to be executed for that phase of applying
+/// changes.
+///
+/// In the future we may generate a single stream of changes and then
+/// sort/merge/batch execution, but for now this is a nice way to improve
+/// batching behaviour of the commands.
+async fn get_operations<'a>(
+    spec: &'a ComputeSpec,
+    ctx: &'a RwLock<MutableApplyContext>,
+    jwks_roles: &'a HashSet<String>,
+    apply_spec_phase: &'a ApplySpecPhase,
+) -> Result<Box<dyn Iterator<Item = Operation> + 'a + Send>> {
+    match apply_spec_phase {
+        ApplySpecPhase::CreateSuperUser => {
+            let query = construct_superuser_query(spec);
+
+            Ok(Box::new(once(Operation {
+                query,
+                comment: None,
+            })))
+        }
+        ApplySpecPhase::DropInvalidDatabases => {
+            let mut ctx = ctx.write().await;
+            let databases = &mut ctx.dbs;
+
+            let keys: Vec<_> = databases
+                .iter()
+                .filter(|(_, db)| db.invalid)
+                .map(|(dbname, _)| dbname.clone())
+                .collect();
+
+            // After recent commit in Postgres, interrupted DROP DATABASE
+            // leaves the database in the invalid state. According to the
+            // commit message, the only option for user is to drop it again.
+            // See:
+            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+            //
+            // Postgres Neon extension is done the way, that db is de-registered
+            // in the control plane metadata only after it is dropped. So there is
+            // a chance that it still thinks that the db should exist. This means
+            // that it will be re-created by the `CreateDatabases` phase. This
+            // is fine, as user can just drop the table again (in vanilla
+            // Postgres they would need to do the same).
+            let operations = keys
+                .into_iter()
+                .filter_map(move |dbname| ctx.dbs.remove(&dbname))
+                .map(|db| Operation {
+                    query: format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote()),
+                    comment: Some(format!("Dropping invalid database {}", db.name)),
+                });
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::RenameRoles => {
+            let mut ctx = ctx.write().await;
+
+            let operations = spec
+                .delta_operations
+                .iter()
+                .flatten()
+                .filter(|op| op.action == "rename_role")
+                .filter_map(move |op| {
+                    let roles = &mut ctx.roles;
+
+                    if roles.contains_key(op.name.as_str()) {
+                        None
+                    } else {
+                        let new_name = op.new_name.as_ref().unwrap();
+                        let mut role = roles.remove(op.name.as_str()).unwrap();
+
+                        role.name = new_name.clone();
+                        role.encrypted_password = None;
+                        roles.insert(role.name.clone(), role);
+
+                        Some(Operation {
+                            query: format!(
+                                "ALTER ROLE {} RENAME TO {}",
+                                op.name.pg_quote(),
+                                new_name.pg_quote()
+                            ),
+                            comment: Some(format!("renaming role '{}' to '{}'", op.name, new_name)),
+                        })
+                    }
+                });
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::CreateAndAlterRoles => {
+            let mut ctx = ctx.write().await;
+
+            let operations = spec.cluster.roles
+                .iter()
+                .filter_map(move |role| {
+                    let roles = &mut ctx.roles;
+                    let db_role = roles.get(&role.name);
+
+                    match db_role {
+                        Some(db_role) => {
+                            if db_role.encrypted_password != role.encrypted_password {
+                                // This can be run on /every/ role! Not just ones created through the console.
+                                // This means that if you add some funny ALTER here that adds a permission,
+                                // this will get run even on user-created roles! This will result in different
+                                // behavior before and after a spec gets reapplied. The below ALTER as it stands
+                                // now only grants LOGIN and changes the password. Please do not allow this branch
+                                // to do anything silly.
+                                Some(Operation {
+                                    query: format!(
+                                        "ALTER ROLE {} {}",
+                                        role.name.pg_quote(),
+                                        role.to_pg_options(),
+                                    ),
+                                    comment: None,
+                                })
+                            } else {
+                                None
+                            }
+                        }
+                        None => {
+                            let query = if !jwks_roles.contains(role.name.as_str()) {
+                                format!(
+                                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser {}",
+                                    role.name.pg_quote(),
+                                    role.to_pg_options(),
+                                )
+                            } else {
+                                format!(
+                                    "CREATE ROLE {} {}",
+                                    role.name.pg_quote(),
+                                    role.to_pg_options(),
+                                )
+                            };
+                            Some(Operation {
+                                query,
+                                comment: Some(format!("creating role {}", role.name)),
+                            })
+                        }
+                    }
+                });
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::RenameAndDeleteDatabases => {
+            let mut ctx = ctx.write().await;
+
+            let operations = spec
+                .delta_operations
+                .iter()
+                .flatten()
+                .filter_map(move |op| {
+                    let databases = &mut ctx.dbs;
+                    match op.action.as_str() {
+                        // We do not check whether the DB exists or not,
+                        // Postgres will take care of it for us
+                        "delete_db" => {
+                            // In Postgres we can't drop a database if it is a template.
+                            // So we need to unset the template flag first, but it could
+                            // be a retry, so we could've already dropped the database.
+                            // Check that database exists first to make it idempotent.
+                            let unset_template_query: String = format!(
+                                include_str!("sql/unset_template_for_drop_dbs.sql"),
+                                datname_str = escape_literal(&op.name),
+                                datname = &op.name.pg_quote()
+                            );
+
+                            // Use FORCE to drop database even if there are active connections.
+                            // We run this from `cloud_admin`, so it should have enough privileges.
+                            // NB: there could be other db states, which prevent us from dropping
+                            // the database. For example, if db is used by any active subscription
+                            // or replication slot.
+                            // TODO: deal with it once we allow logical replication. Proper fix should
+                            // involve returning an error code to the control plane, so it could
+                            // figure out that this is a non-retryable error, return it to the user
+                            // and fail operation permanently.
+                            let drop_db_query: String = format!(
+                                "DROP DATABASE IF EXISTS {} WITH (FORCE)",
+                                &op.name.pg_quote()
+                            );
+
+                            databases.remove(&op.name);
+
+                            Some(vec![
+                                Operation {
+                                    query: unset_template_query,
+                                    comment: Some(format!(
+                                        "optionally clearing template flags for DB {}",
+                                        op.name,
+                                    )),
+                                },
+                                Operation {
+                                    query: drop_db_query,
+                                    comment: Some(format!("deleting database {}", op.name,)),
+                                },
+                            ])
+                        }
+                        "rename_db" => {
+                            if let Some(mut db) = databases.remove(&op.name) {
+                                // update state of known databases
+                                let new_name = op.new_name.as_ref().unwrap();
+                                db.name = new_name.clone();
+                                databases.insert(db.name.clone(), db);
+
+                                Some(vec![Operation {
+                                    query: format!(
+                                        "ALTER DATABASE {} RENAME TO {}",
+                                        op.name.pg_quote(),
+                                        new_name.pg_quote(),
+                                    ),
+                                    comment: Some(format!(
+                                        "renaming database '{}' to '{}'",
+                                        op.name, new_name
+                                    )),
+                                }])
+                            } else {
+                                None
+                            }
+                        }
+                        _ => None,
+                    }
+                })
+                .flatten();
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::CreateAndAlterDatabases => {
+            let mut ctx = ctx.write().await;
+
+            let operations = spec
+                .cluster
+                .databases
+                .iter()
+                .filter_map(move |db| {
+                    let databases = &mut ctx.dbs;
+                    if let Some(edb) = databases.get_mut(&db.name) {
+                        let change_owner = if edb.owner.starts_with('"') {
+                            db.owner.pg_quote() != edb.owner
+                        } else {
+                            db.owner != edb.owner
+                        };
+
+                        edb.owner = db.owner.clone();
+
+                        if change_owner {
+                            Some(vec![Operation {
+                                query: format!(
+                                    "ALTER DATABASE {} OWNER TO {}",
+                                    db.name.pg_quote(),
+                                    db.owner.pg_quote()
+                                ),
+                                comment: Some(format!(
+                                    "changing database owner of database {} to {}",
+                                    db.name, db.owner
+                                )),
+                            }])
+                        } else {
+                            None
+                        }
+                    } else {
+                        databases.insert(db.name.clone(), db.clone());
+
+                        Some(vec![
+                            Operation {
+                                query: format!(
+                                    "CREATE DATABASE {} {}",
+                                    db.name.pg_quote(),
+                                    db.to_pg_options(),
+                                ),
+                                comment: None,
+                            },
+                            Operation {
+                                query: format!(
+                                    "GRANT ALL PRIVILEGES ON DATABASE {} TO neon_superuser",
+                                    db.name.pg_quote()
+                                ),
+                                comment: None,
+                            },
+                        ])
+                    }
+                })
+                .flatten();
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::RunInEachDatabase { db, subphase } => {
+            match subphase {
+                PerDatabasePhase::DeleteDBRoleReferences => {
+                    let ctx = ctx.read().await;
+
+                    let operations =
+                        spec.delta_operations
+                            .iter()
+                            .flatten()
+                            .filter(|op| op.action == "delete_role")
+                            .filter_map(move |op| {
+                                if db.is_owned_by(&op.name) {
+                                    return None;
+                                }
+                                if !ctx.roles.contains_key(&op.name) {
+                                    return None;
+                                }
+                                let quoted = op.name.pg_quote();
+                                let new_owner = match &db {
+                                    DB::SystemDB => PgIdent::from("cloud_admin").pg_quote(),
+                                    DB::UserDB(db) => db.owner.pg_quote(),
+                                };
+
+                                Some(vec![
+                                    // This will reassign all dependent objects to the db owner
+                                    Operation {
+                                        query: format!(
+                                            "REASSIGN OWNED BY {} TO {}",
+                                            quoted, new_owner,
+                                        ),
+                                        comment: None,
+                                    },
+                                    // This now will only drop privileges of the role
+                                    Operation {
+                                        query: format!("DROP OWNED BY {}", quoted),
+                                        comment: None,
+                                    },
+                                ])
+                            })
+                            .flatten();
+
+                    Ok(Box::new(operations))
+                }
+                PerDatabasePhase::ChangeSchemaPerms => {
+                    let ctx = ctx.read().await;
+                    let databases = &ctx.dbs;
+
+                    let db = match &db {
+                        // ignore schema permissions on the system database
+                        DB::SystemDB => return Ok(Box::new(empty())),
+                        DB::UserDB(db) => db,
+                    };
+
+                    if databases.get(&db.name).is_none() {
+                        bail!("database {} doesn't exist in PostgreSQL", db.name);
+                    }
+
+                    let edb = databases.get(&db.name).unwrap();
+
+                    if edb.restrict_conn || edb.invalid {
+                        return Ok(Box::new(empty()));
+                    }
+
+                    let operations = vec![
+                        Operation {
+                            query: format!(
+                                include_str!("sql/set_public_schema_owner.sql"),
+                                db_owner = db.owner.pg_quote()
+                            ),
+                            comment: None,
+                        },
+                        Operation {
+                            query: String::from(include_str!("sql/default_grants.sql")),
+                            comment: None,
+                        },
+                    ]
+                    .into_iter();
+
+                    Ok(Box::new(operations))
+                }
+                PerDatabasePhase::HandleAnonExtension => {
+                    // Only install Anon into user databases
+                    let db = match &db {
+                        DB::SystemDB => return Ok(Box::new(empty())),
+                        DB::UserDB(db) => db,
+                    };
+                    // Never install Anon when it's not enabled as feature
+                    if !spec.features.contains(&ComputeFeature::AnonExtension) {
+                        return Ok(Box::new(empty()));
+                    }
+
+                    // Only install Anon when it's added in preload libraries
+                    let opt_libs = spec.cluster.settings.find("shared_preload_libraries");
+
+                    let libs = match opt_libs {
+                        Some(libs) => libs,
+                        None => return Ok(Box::new(empty())),
+                    };
+
+                    if !libs.contains("anon") {
+                        return Ok(Box::new(empty()));
+                    }
+
+                    let db_owner = db.owner.pg_quote();
+
+                    let operations = vec![
+                        // Create anon extension if this compute needs it
+                        // Users cannot create it themselves, because superuser is required.
+                        Operation {
+                            query: String::from("CREATE EXTENSION IF NOT EXISTS anon CASCADE"),
+                            comment: Some(String::from("creating anon extension")),
+                        },
+                        // Initialize anon extension
+                        // This also requires superuser privileges, so users cannot do it themselves.
+                        Operation {
+                            query: String::from("SELECT anon.init()"),
+                            comment: Some(String::from("initializing anon extension data")),
+                        },
+                        Operation {
+                            query: format!("GRANT ALL ON SCHEMA anon TO {}", db_owner),
+                            comment: Some(String::from(
+                                "granting anon extension schema permissions",
+                            )),
+                        },
+                        Operation {
+                            query: format!(
+                                "GRANT ALL ON ALL FUNCTIONS IN SCHEMA anon TO {}",
+                                db_owner
+                            ),
+                            comment: Some(String::from(
+                                "granting anon extension schema functions permissions",
+                            )),
+                        },
+                        // We need this, because some functions are defined as SECURITY DEFINER.
+                        // In Postgres SECURITY DEFINER functions are executed with the privileges
+                        // of the owner.
+                        // In anon extension this it is needed to access some GUCs, which are only accessible to
+                        // superuser. But we've patched postgres to allow db_owner to access them as well.
+                        // So we need to change owner of these functions to db_owner.
+                        Operation {
+                            query: format!(
+                                include_str!("sql/anon_ext_fn_reassign.sql"),
+                                db_owner = db_owner,
+                            ),
+                            comment: Some(String::from(
+                                "change anon extension functions owner to database_owner",
+                            )),
+                        },
+                        Operation {
+                            query: format!(
+                                "GRANT ALL ON ALL TABLES IN SCHEMA anon TO {}",
+                                db_owner,
+                            ),
+                            comment: Some(String::from(
+                                "granting anon extension tables permissions",
+                            )),
+                        },
+                        Operation {
+                            query: format!(
+                                "GRANT ALL ON ALL SEQUENCES IN SCHEMA anon TO {}",
+                                db_owner,
+                            ),
+                            comment: Some(String::from(
+                                "granting anon extension sequences permissions",
+                            )),
+                        },
+                    ]
+                    .into_iter();
+
+                    Ok(Box::new(operations))
+                }
+            }
+        }
+        // Interestingly, we only install p_s_s in the main database, even when
+        // it's preloaded.
+        ApplySpecPhase::HandleOtherExtensions => {
+            if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
+                if libs.contains("pg_stat_statements") {
+                    return Ok(Box::new(once(Operation {
+                        query: String::from("CREATE EXTENSION IF NOT EXISTS pg_stat_statements"),
+                        comment: Some(String::from("create system extensions")),
+                    })));
+                }
+            }
+            Ok(Box::new(empty()))
+        }
+        ApplySpecPhase::HandleNeonExtension => {
+            let operations = vec![
+                Operation {
+                    query: String::from("CREATE SCHEMA IF NOT EXISTS neon"),
+                    comment: Some(String::from("init: add schema for extension")),
+                },
+                Operation {
+                    query: String::from("CREATE EXTENSION IF NOT EXISTS neon WITH SCHEMA neon"),
+                    comment: Some(String::from(
+                        "init: install the extension if not already installed",
+                    )),
+                },
+                Operation {
+                    query: String::from(
+                        "UPDATE pg_extension SET extrelocatable = true WHERE extname = 'neon'",
+                    ),
+                    comment: Some(String::from("compat/fix: make neon relocatable")),
+                },
+                Operation {
+                    query: String::from("ALTER EXTENSION neon SET SCHEMA neon"),
+                    comment: Some(String::from("compat/fix: alter neon extension schema")),
+                },
+                Operation {
+                    query: String::from("ALTER EXTENSION neon UPDATE"),
+                    comment: Some(String::from("compat/update: update neon extension version")),
+                },
+            ]
+            .into_iter();
+
+            Ok(Box::new(operations))
+        }
+        ApplySpecPhase::CreateAvailabilityCheck => Ok(Box::new(once(Operation {
+            query: String::from(include_str!("sql/add_availabilitycheck_tables.sql")),
+            comment: None,
+        }))),
+        ApplySpecPhase::DropRoles => {
+            let operations = spec
+                .delta_operations
+                .iter()
+                .flatten()
+                .filter(|op| op.action == "delete_role")
+                .map(|op| Operation {
+                    query: format!("DROP ROLE IF EXISTS {}", op.name.pg_quote()),
+                    comment: None,
+                });
+
+            Ok(Box::new(operations))
+        }
+    }
+}

compute_tools/src/sql/add_availabilitycheck_tables.sql

@@ -0,0 +1,18 @@
+DO $$
+BEGIN
+    IF NOT EXISTS(
+        SELECT 1
+        FROM pg_catalog.pg_tables
+        WHERE tablename = 'health_check'
+    )
+    THEN
+    CREATE TABLE health_check (
+        id serial primary key,
+        updated_at timestamptz default now()
+    );
+    INSERT INTO health_check VALUES (1, now())
+        ON CONFLICT (id) DO UPDATE
+         SET updated_at = now();
+    END IF;
+END
+$$

compute_tools/src/sql/anon_ext_fn_reassign.sql

@@ -0,0 +1,12 @@
+DO $$
+DECLARE
+    query varchar;
+BEGIN
+    FOR query IN SELECT 'ALTER FUNCTION '||nsp.nspname||'.'||p.proname||'('||pg_get_function_identity_arguments(p.oid)||') OWNER TO {db_owner};'
+    FROM pg_proc p
+        JOIN pg_namespace nsp ON p.pronamespace = nsp.oid
+    WHERE nsp.nspname = 'anon' LOOP
+        EXECUTE query;
+    END LOOP;
+END
+$$;

compute_tools/src/sql/default_grants.sql

@@ -0,0 +1,30 @@
+DO
+$$
+    BEGIN
+        IF EXISTS(
+            SELECT nspname
+            FROM pg_catalog.pg_namespace
+            WHERE nspname = 'public'
+        ) AND
+           current_setting('server_version_num')::int / 10000 >= 15
+        THEN
+            IF EXISTS(
+                SELECT rolname
+                FROM pg_catalog.pg_roles
+                WHERE rolname = 'web_access'
+            )
+            THEN
+                GRANT CREATE ON SCHEMA public TO web_access;
+            END IF;
+        END IF;
+        IF EXISTS(
+            SELECT nspname
+            FROM pg_catalog.pg_namespace
+            WHERE nspname = 'public'
+        )
+        THEN
+            ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser WITH GRANT OPTION;
+            ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser WITH GRANT OPTION;
+        END IF;
+    END
+$$;

compute_tools/src/sql/set_public_schema_owner.sql

@@ -0,0 +1,23 @@
+DO
+$$
+    DECLARE
+        schema_owner TEXT;
+    BEGIN
+        IF EXISTS(
+            SELECT nspname
+            FROM pg_catalog.pg_namespace
+            WHERE nspname = 'public'
+        )
+        THEN
+            SELECT nspowner::regrole::text
+            FROM pg_catalog.pg_namespace
+            WHERE nspname = 'public'
+            INTO schema_owner;
+
+            IF schema_owner = 'cloud_admin' OR schema_owner = 'zenith_admin'
+            THEN
+                ALTER SCHEMA public OWNER TO {db_owner};
+            END IF;
+        END IF;
+    END
+$$;

compute_tools/src/sql/unset_template_for_drop_dbs.sql

@@ -0,0 +1,12 @@
+DO $$
+    BEGIN
+        IF EXISTS(
+            SELECT 1
+            FROM pg_catalog.pg_database
+            WHERE datname = {datname_str}
+        )
+        THEN
+            ALTER DATABASE {datname} is_template false;
+        END IF;
+    END
+$$;

deny.toml

@@ -37,6 +37,7 @@ allow = [
     "BSD-2-Clause",
     "BSD-3-Clause",
     "CC0-1.0",
+    "CDDL-1.0",
     "ISC",
     "MIT",
     "MPL-2.0",

libs/metrics/src/more_process_metrics.rs

@@ -2,14 +2,28 @@
 
 // This module has heavy inspiration from the prometheus crate's `process_collector.rs`.
 
+use once_cell::sync::Lazy;
+use prometheus::Gauge;
+
 use crate::UIntGauge;
 
 pub struct Collector {
     descs: Vec<prometheus::core::Desc>,
     vmlck: crate::UIntGauge,
+    cpu_seconds_highres: Gauge,
 }
 
-const NMETRICS: usize = 1;
+const NMETRICS: usize = 2;
+
+static CLK_TCK_F64: Lazy<f64> = Lazy::new(|| {
+    let long = unsafe { libc::sysconf(libc::_SC_CLK_TCK) };
+    if long == -1 {
+        panic!("sysconf(_SC_CLK_TCK) failed");
+    }
+    let convertible_to_f64: i32 =
+        i32::try_from(long).expect("sysconf(_SC_CLK_TCK) is larger than i32");
+    convertible_to_f64 as f64
+});
 
 impl prometheus::core::Collector for Collector {
     fn desc(&self) -> Vec<&prometheus::core::Desc> {
@@ -27,6 +41,12 @@ impl prometheus::core::Collector for Collector {
                 mfs.extend(self.vmlck.collect())
             }
         }
+        if let Ok(stat) = myself.stat() {
+            let cpu_seconds = stat.utime + stat.stime;
+            self.cpu_seconds_highres
+                .set(cpu_seconds as f64 / *CLK_TCK_F64);
+            mfs.extend(self.cpu_seconds_highres.collect());
+        }
         mfs
     }
 }
@@ -43,7 +63,23 @@ impl Collector {
                 .cloned(),
         );
 
-        Self { descs, vmlck }
+        let cpu_seconds_highres = Gauge::new(
+            "libmetrics_process_cpu_seconds_highres",
+            "Total user and system CPU time spent in seconds.\
+             Sub-second resolution, hence better than `process_cpu_seconds_total`.",
+        )
+        .unwrap();
+        descs.extend(
+            prometheus::core::Collector::desc(&cpu_seconds_highres)
+                .into_iter()
+                .cloned(),
+        );
+
+        Self {
+            descs,
+            vmlck,
+            cpu_seconds_highres,
+        }
     }
 }
 

libs/pageserver_api/src/config.rs

@@ -109,6 +109,8 @@ pub struct ConfigToml {
     pub virtual_file_io_mode: Option<crate::models::virtual_file::IoMode>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub no_sync: Option<bool>,
+    #[serde(with = "humantime_serde")]
+    pub server_side_batch_timeout: Option<Duration>,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -277,7 +279,11 @@ pub mod defaults {
     pub const DEFAULT_WAL_REDO_TIMEOUT: &str = "60 s";
 
     pub const DEFAULT_SUPERUSER: &str = "cloud_admin";
-    pub const DEFAULT_LOCALE: &str = "C.UTF-8";
+    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;
@@ -313,6 +319,8 @@ pub mod defaults {
     pub const DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB: usize = 0;
 
     pub const DEFAULT_IO_BUFFER_ALIGNMENT: usize = 512;
+
+    pub const DEFAULT_SERVER_SIDE_BATCH_TIMEOUT: Option<&str> = None;
 }
 
 impl Default for ConfigToml {
@@ -393,6 +401,8 @@ impl Default for ConfigToml {
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
             l0_flush: None,
             virtual_file_io_mode: None,
+            server_side_batch_timeout: DEFAULT_SERVER_SIDE_BATCH_TIMEOUT
+                .map(|duration| humantime::parse_duration(duration).unwrap()),
             tenant_config: TenantConfigToml::default(),
             no_sync: None,
         }

libs/pageserver_api/src/key.rs

@@ -24,7 +24,7 @@ pub struct Key {
 
 /// When working with large numbers of Keys in-memory, it is more efficient to handle them as i128 than as
 /// a struct of fields.
-#[derive(Clone, Copy, Hash, PartialEq, Eq, Ord, PartialOrd)]
+#[derive(Clone, Copy, Hash, PartialEq, Eq, Ord, PartialOrd, Serialize, Deserialize)]
 pub struct CompactKey(i128);
 
 /// The storage key size.

libs/pageserver_api/src/record.rs

@@ -41,6 +41,11 @@ pub enum NeonWalRecord {
         file_path: String,
         content: Option<Bytes>,
     },
+    // Truncate visibility map page
+    TruncateVisibilityMap {
+        trunc_byte: usize,
+        trunc_offs: usize,
+    },
 
     /// A testing record for unit testing purposes. It supports append data to an existing image, or clear it.
     #[cfg(feature = "testing")]
@@ -80,18 +85,18 @@ 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/pageserver_api/src/reltag.rs

@@ -24,7 +24,7 @@ use postgres_ffi::Oid;
 // FIXME: should move 'forknum' as last field to keep this consistent with Postgres.
 // Then we could replace the custom Ord and PartialOrd implementations below with
 // deriving them. This will require changes in walredoproc.c.
-#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize)]
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize, Deserialize)]
 pub struct RelTag {
     pub forknum: u8,
     pub spcnode: Oid,

libs/postgres_backend/src/lib.rs

@@ -716,6 +716,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         Ok(())
     }
 
+    // Proto looks like this:
+    // FeMessage::Query("pagestream_v2{FeMessage::CopyData(PagesetreamFeMessage::GetPage(..))}")
+
     async fn process_message(
         &mut self,
         handler: &mut impl Handler<IO>,
@@ -831,7 +834,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         use CopyStreamHandlerEnd::*;
 
         let expected_end = match &end {
-            ServerInitiated(_) | CopyDone | CopyFail | Terminate | EOF => true,
+            ServerInitiated(_) | CopyDone | CopyFail | Terminate | EOF | Cancelled => true,
             CopyStreamHandlerEnd::Disconnected(ConnectionError::Io(io_error))
                 if is_expected_io_error(io_error) =>
             {
@@ -871,6 +874,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             // message from server' when it receives ErrorResponse (anything but
             // CopyData/CopyDone) back.
             CopyFail => Some((end.to_string(), SQLSTATE_SUCCESSFUL_COMPLETION)),
+
+            // When cancelled, send no response: we must not risk blocking on sending that response
+            Cancelled => None,
             _ => None,
         };
         if let Some((err, errcode)) = err_to_send_and_errcode {
@@ -1048,6 +1054,8 @@ pub enum CopyStreamHandlerEnd {
     /// The connection was lost
     #[error("connection error: {0}")]
     Disconnected(#[from] ConnectionError),
+    #[error("Shutdown")]
+    Cancelled,
     /// Some other error
     #[error(transparent)]
     Other(#[from] anyhow::Error),

libs/postgres_ffi/src/pg_constants.rs

@@ -243,8 +243,11 @@ const FSM_LEAF_NODES_PER_PAGE: usize = FSM_NODES_PER_PAGE - FSM_NON_LEAF_NODES_P
 pub const SLOTS_PER_FSM_PAGE: u32 = FSM_LEAF_NODES_PER_PAGE as u32;
 
 /* From visibilitymap.c */
-pub const VM_HEAPBLOCKS_PER_PAGE: u32 =
-    (BLCKSZ as usize - SIZEOF_PAGE_HEADER_DATA) as u32 * (8 / 2); // MAPSIZE * (BITS_PER_BYTE / BITS_PER_HEAPBLOCK)
+
+pub const VM_MAPSIZE: usize = BLCKSZ as usize - MAXALIGN_SIZE_OF_PAGE_HEADER_DATA;
+pub const VM_BITS_PER_HEAPBLOCK: usize = 2;
+pub const VM_HEAPBLOCKS_PER_BYTE: usize = 8 / VM_BITS_PER_HEAPBLOCK;
+pub const VM_HEAPBLOCKS_PER_PAGE: usize = VM_MAPSIZE * VM_HEAPBLOCKS_PER_BYTE;
 
 /* From origin.c */
 pub const REPLICATION_STATE_MAGIC: u32 = 0x1257DADE;

libs/postgres_ffi/src/walrecord.rs

@@ -16,7 +16,7 @@ use utils::bin_ser::DeserializeError;
 use utils::lsn::Lsn;
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlMultiXactCreate {
     pub mid: MultiXactId,
     /* new MultiXact's ID */
@@ -46,7 +46,7 @@ impl XlMultiXactCreate {
 }
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlMultiXactTruncate {
     pub oldest_multi_db: Oid,
     /* to-be-truncated range of multixact offsets */
@@ -72,7 +72,7 @@ impl XlMultiXactTruncate {
 }
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlRelmapUpdate {
     pub dbid: Oid,   /* database ID, or 0 for shared map */
     pub tsid: Oid,   /* database's tablespace, or pg_global */
@@ -90,7 +90,7 @@ impl XlRelmapUpdate {
 }
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlReploriginDrop {
     pub node_id: RepOriginId,
 }
@@ -104,7 +104,7 @@ impl XlReploriginDrop {
 }
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlReploriginSet {
     pub remote_lsn: Lsn,
     pub node_id: RepOriginId,
@@ -120,7 +120,7 @@ impl XlReploriginSet {
 }
 
 #[repr(C)]
-#[derive(Debug, Clone, Copy)]
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
 pub struct RelFileNode {
     pub spcnode: Oid, /* tablespace */
     pub dbnode: Oid,  /* database */
@@ -911,7 +911,7 @@ impl XlSmgrCreate {
 }
 
 #[repr(C)]
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlSmgrTruncate {
     pub blkno: BlockNumber,
     pub rnode: RelFileNode,
@@ -984,7 +984,7 @@ impl XlDropDatabase {
 /// xl_xact_parsed_abort structs in PostgreSQL, but we use the same
 /// struct for commits and aborts.
 ///
-#[derive(Debug)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct XlXactParsedRecord {
     pub xid: TransactionId,
     pub info: u8,

libs/pq_proto/src/lib.rs

@@ -184,6 +184,7 @@ pub struct CancelKeyData {
 
 impl fmt::Display for CancelKeyData {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // TODO: this is producing strange results, with 0xffffffff........ always in the logs.
         let hi = (self.backend_pid as u64) << 32;
         let lo = self.cancel_key as u64;
         let id = hi | lo;

libs/remote_storage/src/azure_blob.rs

@@ -97,10 +97,7 @@ impl AzureBlobStorage {
 
     pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
         assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
-        let path_string = path
-            .get_path()
-            .as_str()
-            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path.get_path().as_str();
         match &self.prefix_in_container {
             Some(prefix) => {
                 if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
@@ -277,19 +274,14 @@ impl RemoteStorage for AzureBlobStorage {
         cancel: &CancellationToken,
     ) -> impl Stream<Item = Result<Listing, DownloadError>> {
         // get the passed prefix or if it is not set use prefix_in_bucket value
-        let list_prefix = prefix
-            .map(|p| self.relative_path_to_name(p))
-            .or_else(|| self.prefix_in_container.clone())
-            .map(|mut p| {
-                // required to end with a separator
-                // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
-                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let list_prefix = prefix.map(|p| self.relative_path_to_name(p)).or_else(|| {
+            self.prefix_in_container.clone().map(|mut s| {
+                if !s.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    s.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
                 }
-                p
-            });
+                s
+            })
+        });
 
         async_stream::stream! {
             let _permit = self.permit(RequestKind::List, cancel).await?;

libs/remote_storage/src/config.rs

@@ -26,6 +26,16 @@ pub struct RemoteStorageConfig {
     pub timeout: Duration,
 }
 
+impl RemoteStorageKind {
+    pub fn bucket_name(&self) -> Option<&str> {
+        match self {
+            RemoteStorageKind::LocalFs { .. } => None,
+            RemoteStorageKind::AwsS3(config) => Some(&config.bucket_name),
+            RemoteStorageKind::AzureContainer(config) => Some(&config.container_name),
+        }
+    }
+}
+
 fn default_timeout() -> Duration {
     RemoteStorageConfig::DEFAULT_TIMEOUT
 }
@@ -178,6 +188,14 @@ impl RemoteStorageConfig {
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<RemoteStorageConfig> {
         Ok(utils::toml_edit_ext::deserialize_item(toml)?)
     }
+
+    pub fn from_toml_str(input: &str) -> anyhow::Result<RemoteStorageConfig> {
+        let toml_document = toml_edit::DocumentMut::from_str(input)?;
+        if let Some(item) = toml_document.get("remote_storage") {
+            return Self::from_toml(item);
+        }
+        Self::from_toml(toml_document.as_item())
+    }
 }
 
 #[cfg(test)]
@@ -185,8 +203,7 @@ mod tests {
     use super::*;
 
     fn parse(input: &str) -> anyhow::Result<RemoteStorageConfig> {
-        let toml = input.parse::<toml_edit::DocumentMut>().unwrap();
-        RemoteStorageConfig::from_toml(toml.as_item())
+        RemoteStorageConfig::from_toml_str(input)
     }
 
     #[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

@@ -50,8 +50,8 @@ REDO_POS=0x$("$PG_BIN"/pg_controldata -D "$DATA_DIR" | grep -F "REDO location"|
 declare -i WAL_SIZE=$REDO_POS+114
 "$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" start
 "$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" stop -m immediate
-cp "$DATA_DIR"/pg_wal/000000010000000000000001 .
+cp "$DATA_DIR"/pg_wal/000000010000000000000001 "$DATA_DIR"
 cp "$WAL_PATH"/* "$DATA_DIR"/pg_wal/
 for partial in "$DATA_DIR"/pg_wal/*.partial ; do mv "$partial" "${partial%.partial}" ; done
-dd if=000000010000000000000001 of="$DATA_DIR"/pg_wal/000000010000000000000001 bs=$WAL_SIZE count=1 conv=notrunc
-rm -f 000000010000000000000001
+dd if="$DATA_DIR"/000000010000000000000001 of="$DATA_DIR"/pg_wal/000000010000000000000001 bs=$WAL_SIZE count=1 conv=notrunc
+rm -f "$DATA_DIR"/000000010000000000000001

libs/utils/scripts/restore_from_wal_initdb.sh

@@ -14,8 +14,8 @@ REDO_POS=0x$("$PG_BIN"/pg_controldata -D "$DATA_DIR" | grep -F "REDO location"|
 declare -i WAL_SIZE=$REDO_POS+114
 "$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" start
 "$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" stop -m immediate
-cp "$DATA_DIR"/pg_wal/000000010000000000000001 .
+cp "$DATA_DIR"/pg_wal/000000010000000000000001 "$DATA_DIR"
 cp "$WAL_PATH"/* "$DATA_DIR"/pg_wal/
 for partial in "$DATA_DIR"/pg_wal/*.partial ; do mv "$partial" "${partial%.partial}" ; done
-dd if=000000010000000000000001 of="$DATA_DIR"/pg_wal/000000010000000000000001 bs=$WAL_SIZE count=1 conv=notrunc
-rm -f 000000010000000000000001
+dd if="$DATA_DIR"/000000010000000000000001 of="$DATA_DIR"/pg_wal/000000010000000000000001 bs=$WAL_SIZE count=1 conv=notrunc
+rm -f "$DATA_DIR"/000000010000000000000001

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/http/json.rs

@@ -5,6 +5,7 @@ use serde::{Deserialize, Serialize};
 
 use super::error::ApiError;
 
+/// Parse a json request body and deserialize it to the type `T`.
 pub async fn json_request<T: for<'de> Deserialize<'de>>(
     request: &mut Request<Body>,
 ) -> Result<T, ApiError> {
@@ -27,6 +28,27 @@ pub async fn json_request<T: for<'de> Deserialize<'de>>(
         .map_err(ApiError::BadRequest)
 }
 
+/// Parse a json request body and deserialize it to the type `T`. If the body is empty, return `T::default`.
+pub async fn json_request_maybe<T: for<'de> Deserialize<'de> + Default>(
+    request: &mut Request<Body>,
+) -> Result<T, ApiError> {
+    let body = hyper::body::aggregate(request.body_mut())
+        .await
+        .context("Failed to read request body")
+        .map_err(ApiError::BadRequest)?;
+
+    if body.remaining() == 0 {
+        return Ok(T::default());
+    }
+
+    let mut deser = serde_json::de::Deserializer::from_reader(body.reader());
+
+    serde_path_to_error::deserialize(&mut deser)
+        // intentionally stringify because the debug version is not helpful in python logs
+        .map_err(|e| anyhow::anyhow!("Failed to parse json request: {e}"))
+        .map_err(ApiError::BadRequest)
+}
+
 pub fn json_response<T: Serialize>(
     status: StatusCode,
     data: T,

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();

libs/wal_decoder/src/decoder.rs

@@ -19,7 +19,7 @@ impl InterpretedWalRecord {
     pub fn from_bytes_filtered(
         buf: Bytes,
         shard: &ShardIdentity,
-        record_end_lsn: Lsn,
+        next_record_lsn: Lsn,
         pg_version: u32,
     ) -> anyhow::Result<InterpretedWalRecord> {
         let mut decoded = DecodedWALRecord::default();
@@ -32,18 +32,18 @@ impl InterpretedWalRecord {
             FlushUncommittedRecords::No
         };
 
-        let metadata_record = MetadataRecord::from_decoded(&decoded, record_end_lsn, pg_version)?;
+        let metadata_record = MetadataRecord::from_decoded(&decoded, next_record_lsn, pg_version)?;
         let batch = SerializedValueBatch::from_decoded_filtered(
             decoded,
             shard,
-            record_end_lsn,
+            next_record_lsn,
             pg_version,
         )?;
 
         Ok(InterpretedWalRecord {
             metadata_record,
             batch,
-            end_lsn: record_end_lsn,
+            next_record_lsn,
             flush_uncommitted,
             xid,
         })
@@ -53,7 +53,7 @@ impl InterpretedWalRecord {
 impl MetadataRecord {
     fn from_decoded(
         decoded: &DecodedWALRecord,
-        record_end_lsn: Lsn,
+        next_record_lsn: Lsn,
         pg_version: u32,
     ) -> anyhow::Result<Option<MetadataRecord>> {
         // Note: this doesn't actually copy the bytes since
@@ -74,7 +74,9 @@ impl MetadataRecord {
                 Ok(None)
             }
             pg_constants::RM_CLOG_ID => Self::decode_clog_record(&mut buf, decoded, pg_version),
-            pg_constants::RM_XACT_ID => Self::decode_xact_record(&mut buf, decoded, record_end_lsn),
+            pg_constants::RM_XACT_ID => {
+                Self::decode_xact_record(&mut buf, decoded, next_record_lsn)
+            }
             pg_constants::RM_MULTIXACT_ID => {
                 Self::decode_multixact_record(&mut buf, decoded, pg_version)
             }
@@ -86,7 +88,9 @@ impl MetadataRecord {
             //
             // Alternatively, one can make the checkpoint part of the subscription protocol
             // to the pageserver. This should work fine, but can be done at a later point.
-            pg_constants::RM_XLOG_ID => Self::decode_xlog_record(&mut buf, decoded, record_end_lsn),
+            pg_constants::RM_XLOG_ID => {
+                Self::decode_xlog_record(&mut buf, decoded, next_record_lsn)
+            }
             pg_constants::RM_LOGICALMSG_ID => {
                 Self::decode_logical_message_record(&mut buf, decoded)
             }

libs/wal_decoder/src/models.rs

@@ -32,16 +32,19 @@ use postgres_ffi::walrecord::{
     XlSmgrTruncate, XlXactParsedRecord,
 };
 use postgres_ffi::{Oid, TransactionId};
+use serde::{Deserialize, Serialize};
 use utils::lsn::Lsn;
 
 use crate::serialized_batch::SerializedValueBatch;
 
+#[derive(Serialize, Deserialize)]
 pub enum FlushUncommittedRecords {
     Yes,
     No,
 }
 
 /// An interpreted Postgres WAL record, ready to be handled by the pageserver
+#[derive(Serialize, Deserialize)]
 pub struct InterpretedWalRecord {
     /// Optional metadata record - may cause writes to metadata keys
     /// in the storage engine
@@ -49,8 +52,10 @@ pub struct InterpretedWalRecord {
     /// A pre-serialized batch along with the required metadata for ingestion
     /// by the pageserver
     pub batch: SerializedValueBatch,
-    /// Byte offset within WAL for the end of the original PG WAL record
-    pub end_lsn: Lsn,
+    /// Byte offset within WAL for the start of the next PG WAL record.
+    /// Usually this is the end LSN of the current record, but in case of
+    /// XLOG SWITCH records it will be within the next segment.
+    pub next_record_lsn: Lsn,
     /// Whether to flush all uncommitted modifications to the storage engine
     /// before ingesting this record. This is currently only used for legacy PG
     /// database creations which read pages from a template database. Such WAL
@@ -62,6 +67,7 @@ pub struct InterpretedWalRecord {
 
 /// The interpreted part of the Postgres WAL record which requires metadata
 /// writes to the underlying storage engine.
+#[derive(Serialize, Deserialize)]
 pub enum MetadataRecord {
     Heapam(HeapamRecord),
     Neonrmgr(NeonrmgrRecord),
@@ -77,10 +83,12 @@ pub enum MetadataRecord {
     Replorigin(ReploriginRecord),
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum HeapamRecord {
     ClearVmBits(ClearVmBits),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct ClearVmBits {
     pub new_heap_blkno: Option<u32>,
     pub old_heap_blkno: Option<u32>,
@@ -88,24 +96,29 @@ pub struct ClearVmBits {
     pub flags: u8,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum NeonrmgrRecord {
     ClearVmBits(ClearVmBits),
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum SmgrRecord {
     Create(SmgrCreate),
     Truncate(XlSmgrTruncate),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct SmgrCreate {
     pub rel: RelTag,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum DbaseRecord {
     Create(DbaseCreate),
     Drop(DbaseDrop),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct DbaseCreate {
     pub db_id: Oid,
     pub tablespace_id: Oid,
@@ -113,27 +126,32 @@ pub struct DbaseCreate {
     pub src_tablespace_id: Oid,
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct DbaseDrop {
     pub db_id: Oid,
     pub tablespace_ids: Vec<Oid>,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum ClogRecord {
     ZeroPage(ClogZeroPage),
     Truncate(ClogTruncate),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct ClogZeroPage {
     pub segno: u32,
     pub rpageno: u32,
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct ClogTruncate {
     pub pageno: u32,
     pub oldest_xid: TransactionId,
     pub oldest_xid_db: Oid,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum XactRecord {
     Commit(XactCommon),
     Abort(XactCommon),
@@ -142,6 +160,7 @@ pub enum XactRecord {
     Prepare(XactPrepare),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct XactCommon {
     pub parsed: XlXactParsedRecord,
     pub origin_id: u16,
@@ -150,61 +169,73 @@ pub struct XactCommon {
     pub lsn: Lsn,
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct XactPrepare {
     pub xl_xid: TransactionId,
     pub data: Bytes,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum MultiXactRecord {
     ZeroPage(MultiXactZeroPage),
     Create(XlMultiXactCreate),
     Truncate(XlMultiXactTruncate),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct MultiXactZeroPage {
     pub slru_kind: SlruKind,
     pub segno: u32,
     pub rpageno: u32,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum RelmapRecord {
     Update(RelmapUpdate),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct RelmapUpdate {
     pub update: XlRelmapUpdate,
     pub buf: Bytes,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum XlogRecord {
     Raw(RawXlogRecord),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct RawXlogRecord {
     pub info: u8,
     pub lsn: Lsn,
     pub buf: Bytes,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum LogicalMessageRecord {
     Put(PutLogicalMessage),
     #[cfg(feature = "testing")]
     Failpoint,
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct PutLogicalMessage {
     pub path: String,
     pub buf: Bytes,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum StandbyRecord {
     RunningXacts(StandbyRunningXacts),
 }
 
+#[derive(Serialize, Deserialize)]
 pub struct StandbyRunningXacts {
     pub oldest_running_xid: TransactionId,
 }
 
+#[derive(Serialize, Deserialize)]
 pub enum ReploriginRecord {
     Set(XlReploriginSet),
     Drop(XlReploriginDrop),

libs/wal_decoder/src/serialized_batch.rs

@@ -16,6 +16,7 @@ use pageserver_api::shard::ShardIdentity;
 use pageserver_api::{key::CompactKey, value::Value};
 use postgres_ffi::walrecord::{DecodedBkpBlock, DecodedWALRecord};
 use postgres_ffi::{page_is_new, page_set_lsn, pg_constants, BLCKSZ};
+use serde::{Deserialize, Serialize};
 use utils::bin_ser::BeSer;
 use utils::lsn::Lsn;
 
@@ -29,6 +30,7 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 /// relation sizes. In the case of "observed" values, we only need to know
 /// the key and LSN, so two types of metadata are supported to save on network
 /// bandwidth.
+#[derive(Serialize, Deserialize)]
 pub enum ValueMeta {
     Serialized(SerializedValueMeta),
     Observed(ObservedValueMeta),
@@ -75,6 +77,7 @@ impl PartialEq for OrderedValueMeta {
 impl Eq for OrderedValueMeta {}
 
 /// Metadata for a [`Value`] serialized into the batch.
+#[derive(Serialize, Deserialize)]
 pub struct SerializedValueMeta {
     pub key: CompactKey,
     pub lsn: Lsn,
@@ -86,12 +89,14 @@ pub struct SerializedValueMeta {
 }
 
 /// Metadata for a [`Value`] observed by the batch
+#[derive(Serialize, Deserialize)]
 pub struct ObservedValueMeta {
     pub key: CompactKey,
     pub lsn: Lsn,
 }
 
 /// Batch of serialized [`Value`]s.
+#[derive(Serialize, Deserialize)]
 pub struct SerializedValueBatch {
     /// [`Value`]s serialized in EphemeralFile's native format,
     /// ready for disk write by the pageserver
@@ -132,7 +137,7 @@ impl SerializedValueBatch {
     pub(crate) fn from_decoded_filtered(
         decoded: DecodedWALRecord,
         shard: &ShardIdentity,
-        record_end_lsn: Lsn,
+        next_record_lsn: Lsn,
         pg_version: u32,
     ) -> anyhow::Result<SerializedValueBatch> {
         // First determine how big the buffer needs to be and allocate it up-front.
@@ -156,13 +161,17 @@ impl SerializedValueBatch {
             let key = rel_block_to_key(rel, blk.blkno);
 
             if !key.is_valid_key_on_write_path() {
-                anyhow::bail!("Unsupported key decoded at LSN {}: {}", record_end_lsn, key);
+                anyhow::bail!(
+                    "Unsupported key decoded at LSN {}: {}",
+                    next_record_lsn,
+                    key
+                );
             }
 
             let key_is_local = shard.is_key_local(&key);
 
             tracing::debug!(
-                lsn=%record_end_lsn,
+                lsn=%next_record_lsn,
                 key=%key,
                 "ingest: shard decision {}",
                 if !key_is_local { "drop" } else { "keep" },
@@ -174,7 +183,7 @@ impl SerializedValueBatch {
                     // its blkno in case it implicitly extends a relation.
                     metadata.push(ValueMeta::Observed(ObservedValueMeta {
                         key: key.to_compact(),
-                        lsn: record_end_lsn,
+                        lsn: next_record_lsn,
                     }))
                 }
 
@@ -205,7 +214,7 @@ impl SerializedValueBatch {
                 // that would corrupt the page.
                 //
                 if !page_is_new(&image) {
-                    page_set_lsn(&mut image, record_end_lsn)
+                    page_set_lsn(&mut image, next_record_lsn)
                 }
                 assert_eq!(image.len(), BLCKSZ as usize);
 
@@ -224,12 +233,12 @@ impl SerializedValueBatch {
 
             metadata.push(ValueMeta::Serialized(SerializedValueMeta {
                 key: key.to_compact(),
-                lsn: record_end_lsn,
+                lsn: next_record_lsn,
                 batch_offset: relative_off,
                 len: val_ser_size,
                 will_init: val.will_init(),
             }));
-            max_lsn = std::cmp::max(max_lsn, record_end_lsn);
+            max_lsn = std::cmp::max(max_lsn, next_record_lsn);
             len += 1;
         }
 

pageserver/Cargo.toml

@@ -84,6 +84,7 @@ enumset = { workspace = true, features = ["serde"]}
 strum.workspace = true
 strum_macros.workspace = true
 wal_decoder.workspace = true
+smallvec.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]
 procfs.workspace = true

pageserver/benches/bench_ingest.rs

@@ -167,6 +167,7 @@ fn criterion_benchmark(c: &mut Criterion) {
         16384,
         virtual_file::io_engine_for_bench(),
         conf.virtual_file_io_mode,
+        virtual_file::SyncMode::Sync,
     );
     page_cache::init(conf.page_cache_size);
 

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/ctl/Cargo.toml

@@ -18,7 +18,6 @@ postgres_ffi.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
 tokio-util.workspace = true
-toml_edit.workspace = true
 utils.workspace = true
 svg_fmt.workspace = true
 workspace_hack.workspace = true

pageserver/ctl/src/layer_map_analyzer.rs

@@ -138,6 +138,7 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
         10,
         virtual_file::api::IoEngineKind::StdFs,
         IoMode::preferred(),
+        virtual_file::SyncMode::Sync,
     );
     pageserver::page_cache::init(100);
 

pageserver/ctl/src/layers.rs

@@ -51,6 +51,7 @@ async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result
         10,
         virtual_file::api::IoEngineKind::StdFs,
         IoMode::preferred(),
+        virtual_file::SyncMode::Sync,
     );
     page_cache::init(100);
     let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
@@ -65,6 +66,7 @@ async fn read_image_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result
         10,
         virtual_file::api::IoEngineKind::StdFs,
         IoMode::preferred(),
+        virtual_file::SyncMode::Sync,
     );
     page_cache::init(100);
     let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
@@ -171,6 +173,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                 10,
                 virtual_file::api::IoEngineKind::StdFs,
                 IoMode::preferred(),
+                virtual_file::SyncMode::Sync,
             );
             pageserver::page_cache::init(100);
 

pageserver/ctl/src/main.rs

@@ -174,11 +174,7 @@ async fn main() -> anyhow::Result<()> {
                 println!("specified prefix '{}' failed validation", cmd.prefix);
                 return Ok(());
             };
-            let toml_document = toml_edit::DocumentMut::from_str(&cmd.config_toml_str)?;
-            let toml_item = toml_document
-                .get("remote_storage")
-                .expect("need remote_storage");
-            let config = RemoteStorageConfig::from_toml(toml_item)?;
+            let config = RemoteStorageConfig::from_toml_str(&cmd.config_toml_str)?;
             let storage = remote_storage::GenericRemoteStorage::from_config(&config).await;
             let cancel = CancellationToken::new();
             storage
@@ -209,6 +205,7 @@ async fn print_layerfile(path: &Utf8Path) -> anyhow::Result<()> {
         10,
         virtual_file::api::IoEngineKind::StdFs,
         IoMode::preferred(),
+        virtual_file::SyncMode::Sync,
     );
     page_cache::init(100);
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);

pageserver/src/bin/pageserver.rs

@@ -171,11 +171,18 @@ fn main() -> anyhow::Result<()> {
     let scenario = failpoint_support::init();
 
     // Basic initialization of things that don't change after startup
+    tracing::info!("Initializing virtual_file...");
     virtual_file::init(
         conf.max_file_descriptors,
         conf.virtual_file_io_engine,
         conf.virtual_file_io_mode,
+        if conf.no_sync {
+            virtual_file::SyncMode::UnsafeNoSync
+        } else {
+            virtual_file::SyncMode::Sync
+        },
     );
+    tracing::info!("Initializing page_cache...");
     page_cache::init(conf.page_cache_size);
 
     start_pageserver(launch_ts, conf).context("Failed to start pageserver")?;

pageserver/src/config.rs

@@ -182,6 +182,10 @@ pub struct PageServerConf {
 
     /// Optionally disable disk syncs (unsafe!)
     pub no_sync: bool,
+
+    /// Maximum amount of time for which a get page request request
+    /// might be held up for request merging.
+    pub server_side_batch_timeout: Option<Duration>,
 }
 
 /// Token for authentication to safekeepers
@@ -336,6 +340,7 @@ impl PageServerConf {
             concurrent_tenant_warmup,
             concurrent_tenant_size_logical_size_queries,
             virtual_file_io_engine,
+            server_side_batch_timeout,
             tenant_config,
             no_sync,
         } = config_toml;
@@ -377,6 +382,7 @@ impl PageServerConf {
             image_compression,
             timeline_offloading,
             ephemeral_bytes_per_memory_kb,
+            server_side_batch_timeout,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling

pageserver/src/http/routes.rs

@@ -83,6 +83,8 @@ use crate::tenant::storage_layer::LayerName;
 use crate::tenant::timeline::offload::offload_timeline;
 use crate::tenant::timeline::offload::OffloadError;
 use crate::tenant::timeline::CompactFlags;
+use crate::tenant::timeline::CompactOptions;
+use crate::tenant::timeline::CompactRange;
 use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::Timeline;
 use crate::tenant::GetTimelineError;
@@ -100,7 +102,7 @@ use utils::{
     http::{
         endpoint::{self, attach_openapi_ui, auth_middleware, check_permission_with},
         error::{ApiError, HttpErrorBody},
-        json::{json_request, json_response},
+        json::{json_request, json_request_maybe, json_response},
         request::parse_request_param,
         RequestExt, RouterBuilder,
     },
@@ -324,6 +326,7 @@ impl From<crate::tenant::DeleteTimelineError> for ApiError {
                     .into_boxed_str(),
             ),
             a @ AlreadyInProgress(_) => ApiError::Conflict(a.to_string()),
+            Cancelled => ApiError::ResourceUnavailable("shutting down".into()),
             Other(e) => ApiError::InternalServerError(e),
         }
     }
@@ -1926,13 +1929,15 @@ async fn timeline_gc_handler(
 
 // Run compaction immediately on given timeline.
 async fn timeline_compact_handler(
-    request: Request<Body>,
+    mut request: Request<Body>,
     cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
     let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
+    let compact_range = json_request_maybe::<Option<CompactRange>>(&mut request).await?;
+
     let state = get_state(&request);
 
     let mut flags = EnumSet::empty();
@@ -1956,11 +1961,16 @@ async fn timeline_compact_handler(
     let wait_until_uploaded =
         parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
 
+    let options = CompactOptions {
+        compact_range,
+        flags,
+    };
+
     async {
         let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
         let timeline = active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id).await?;
         timeline
-            .compact(&cancel, flags, &ctx)
+            .compact_with_options(&cancel, options, &ctx)
             .await
             .map_err(|e| ApiError::InternalServerError(e.into()))?;
         if wait_until_uploaded {

pageserver/src/metrics.rs

@@ -1187,6 +1187,7 @@ struct GlobalAndPerTimelineHistogramTimer<'a, 'c> {
     ctx: &'c RequestContext,
     start: std::time::Instant,
     op: SmgrQueryType,
+    count: usize,
 }
 
 impl Drop for GlobalAndPerTimelineHistogramTimer<'_, '_> {
@@ -1214,10 +1215,13 @@ impl Drop for GlobalAndPerTimelineHistogramTimer<'_, '_> {
                 elapsed
             }
         };
-        self.global_latency_histo
-            .observe(ex_throttled.as_secs_f64());
-        if let Some(per_timeline_getpage_histo) = self.per_timeline_latency_histo {
-            per_timeline_getpage_histo.observe(ex_throttled.as_secs_f64());
+
+        for _ in 0..self.count {
+            self.global_latency_histo
+                .observe(ex_throttled.as_secs_f64());
+            if let Some(per_timeline_getpage_histo) = self.per_timeline_latency_histo {
+                per_timeline_getpage_histo.observe(ex_throttled.as_secs_f64());
+            }
         }
     }
 }
@@ -1385,6 +1389,14 @@ impl SmgrQueryTimePerTimeline {
         &'a self,
         op: SmgrQueryType,
         ctx: &'c RequestContext,
+    ) -> Option<impl Drop + 'a> {
+        self.start_timer_many(op, 1, ctx)
+    }
+    pub(crate) fn start_timer_many<'c: 'a, 'a>(
+        &'a self,
+        op: SmgrQueryType,
+        count: usize,
+        ctx: &'c RequestContext,
     ) -> Option<impl Drop + 'a> {
         let start = Instant::now();
 
@@ -1422,6 +1434,7 @@ impl SmgrQueryTimePerTimeline {
             ctx,
             start,
             op,
+            count,
         })
     }
 }

pageserver/src/page_service.rs

@@ -7,13 +7,13 @@ use bytes::Buf;
 use futures::FutureExt;
 use itertools::Itertools;
 use once_cell::sync::OnceCell;
-use pageserver_api::models::TenantState;
+use pageserver_api::models::{self, TenantState};
 use pageserver_api::models::{
     PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
     PagestreamErrorResponse, PagestreamExistsRequest, PagestreamExistsResponse,
-    PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetPageResponse,
-    PagestreamGetSlruSegmentRequest, PagestreamGetSlruSegmentResponse, PagestreamNblocksRequest,
-    PagestreamNblocksResponse, PagestreamProtocolVersion,
+    PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetSlruSegmentRequest,
+    PagestreamGetSlruSegmentResponse, PagestreamNblocksRequest, PagestreamNblocksResponse,
+    PagestreamProtocolVersion,
 };
 use pageserver_api::shard::TenantShardId;
 use postgres_backend::{is_expected_io_error, AuthType, PostgresBackend, QueryError};
@@ -44,7 +44,7 @@ use crate::basebackup;
 use crate::basebackup::BasebackupError;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::metrics;
+use crate::metrics::{self};
 use crate::metrics::{ComputeCommandKind, COMPUTE_COMMANDS_COUNTERS, LIVE_CONNECTIONS};
 use crate::pgdatadir_mapping::Version;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
@@ -59,7 +59,7 @@ use crate::tenant::GetTimelineError;
 use crate::tenant::PageReconstructError;
 use crate::tenant::Timeline;
 use pageserver_api::key::rel_block_to_key;
-use pageserver_api::reltag::SlruKind;
+use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
 use postgres_ffi::BLCKSZ;
 
@@ -105,6 +105,7 @@ pub fn spawn(
             pg_auth,
             tcp_listener,
             conf.pg_auth_type,
+            conf.server_side_batch_timeout,
             libpq_ctx,
             cancel.clone(),
         )
@@ -153,6 +154,7 @@ pub async fn libpq_listener_main(
     auth: Option<Arc<SwappableJwtAuth>>,
     listener: tokio::net::TcpListener,
     auth_type: AuthType,
+    server_side_batch_timeout: Option<Duration>,
     listener_ctx: RequestContext,
     listener_cancel: CancellationToken,
 ) -> Connections {
@@ -183,6 +185,7 @@ pub async fn libpq_listener_main(
                     local_auth,
                     socket,
                     auth_type,
+                    server_side_batch_timeout,
                     connection_ctx,
                     connections_cancel.child_token(),
                 ));
@@ -210,6 +213,7 @@ async fn page_service_conn_main(
     auth: Option<Arc<SwappableJwtAuth>>,
     socket: tokio::net::TcpStream,
     auth_type: AuthType,
+    server_side_batch_timeout: Option<Duration>,
     connection_ctx: RequestContext,
     cancel: CancellationToken,
 ) -> ConnectionHandlerResult {
@@ -260,8 +264,13 @@ async fn page_service_conn_main(
     // and create a child per-query context when it invokes process_query.
     // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
     // and create the per-query context in process_query ourselves.
-    let mut conn_handler =
-        PageServerHandler::new(tenant_manager, auth, connection_ctx, cancel.clone());
+    let mut conn_handler = PageServerHandler::new(
+        tenant_manager,
+        auth,
+        server_side_batch_timeout,
+        connection_ctx,
+        cancel.clone(),
+    );
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
 
     match pgbackend.run(&mut conn_handler, &cancel).await {
@@ -304,6 +313,14 @@ struct PageServerHandler {
     cancel: CancellationToken,
 
     timeline_handles: TimelineHandles,
+
+    /// See [`PageServerConf::server_side_batch_timeout`]
+    server_side_batch_timeout: Option<Duration>,
+}
+
+struct Carry {
+    msg: BatchedFeMessage,
+    started_at: Instant,
 }
 
 struct TimelineHandles {
@@ -517,10 +534,47 @@ impl From<WaitLsnError> for QueryError {
     }
 }
 
+enum BatchedFeMessage {
+    Exists {
+        span: Span,
+        req: models::PagestreamExistsRequest,
+    },
+    Nblocks {
+        span: Span,
+        req: models::PagestreamNblocksRequest,
+    },
+    GetPage {
+        span: Span,
+        shard: timeline::handle::Handle<TenantManagerTypes>,
+        effective_request_lsn: Lsn,
+        pages: smallvec::SmallVec<[(RelTag, BlockNumber); 1]>,
+    },
+    DbSize {
+        span: Span,
+        req: models::PagestreamDbSizeRequest,
+    },
+    GetSlruSegment {
+        span: Span,
+        req: models::PagestreamGetSlruSegmentRequest,
+    },
+    RespondError {
+        span: Span,
+        error: PageStreamError,
+    },
+}
+
+enum BatchOrEof {
+    /// In the common case, this has one entry.
+    /// At most, it has two entries: the first is the leftover batch, the second is an error.
+    Batch(smallvec::SmallVec<[BatchedFeMessage; 1]>),
+    Eof,
+}
+
 impl PageServerHandler {
     pub fn new(
         tenant_manager: Arc<TenantManager>,
         auth: Option<Arc<SwappableJwtAuth>>,
+        server_side_batch_timeout: Option<Duration>,
         connection_ctx: RequestContext,
         cancel: CancellationToken,
     ) -> Self {
@@ -530,6 +584,7 @@ impl PageServerHandler {
             connection_ctx,
             timeline_handles: TimelineHandles::new(tenant_manager),
             cancel,
+            server_side_batch_timeout,
         }
     }
 
@@ -557,6 +612,258 @@ impl PageServerHandler {
         )
     }
 
+    #[instrument(skip_all, level = tracing::Level::TRACE)]
+    async fn read_batch_from_connection<IO>(
+        &mut self,
+        pgb: &mut PostgresBackend<IO>,
+        tenant_id: &TenantId,
+        timeline_id: &TimelineId,
+        maybe_carry: &mut Option<Carry>,
+        ctx: &RequestContext,
+    ) -> Result<BatchOrEof, QueryError>
+    where
+        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
+    {
+        debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
+
+        let mut batching_deadline_storage = None; // TODO: can this just be an unsync once_cell?
+
+        loop {
+            // Create a future that will become ready when we need to stop batching.
+            use futures::future::Either;
+            let batching_deadline = match (
+                &*maybe_carry as &Option<Carry>,
+                &mut batching_deadline_storage,
+            ) {
+                (None, None) => Either::Left(futures::future::pending()), // there's no deadline before we have something batched
+                (None, Some(_)) => unreachable!(),
+                (Some(_), Some(fut)) => Either::Right(fut), // below arm already ran
+                (Some(carry), None) => {
+                    match self.server_side_batch_timeout {
+                        None => {
+                            return Ok(BatchOrEof::Batch(smallvec::smallvec![
+                                maybe_carry
+                                    .take()
+                                    .expect("we already checked it's Some")
+                                    .msg
+                            ]))
+                        }
+                        Some(batch_timeout) => {
+                            // Take into consideration the time the carry spent waiting.
+                            let batch_timeout =
+                                batch_timeout.saturating_sub(carry.started_at.elapsed());
+                            if batch_timeout.is_zero() {
+                                // the timer doesn't support restarting with zero duration
+                                return Ok(BatchOrEof::Batch(smallvec::smallvec![
+                                    maybe_carry
+                                        .take()
+                                        .expect("we already checked it's Some")
+                                        .msg
+                                ]));
+                            } else {
+                                batching_deadline_storage = Some(Box::pin(async move {
+                                    tokio::time::sleep(batch_timeout).await;
+                                }));
+                                Either::Right(
+                                    batching_deadline_storage.as_mut().expect("we just set it"),
+                                )
+                            }
+                        }
+                    }
+                }
+            };
+            let msg = tokio::select! {
+                biased;
+                _ = self.cancel.cancelled() => {
+                    return Err(QueryError::Shutdown)
+                }
+                _ = batching_deadline => {
+                    return Ok(BatchOrEof::Batch(smallvec::smallvec![maybe_carry.take().expect("per construction of batching_deadline").msg]));
+                }
+                msg = pgb.read_message() => { msg }
+            };
+
+            let msg_start = Instant::now();
+
+            // Rest of this loop body is trying to batch `msg` into `batch`.
+            // If we can add msg to batch we continue into the next loop iteration.
+            // If we can't add msg to batch batch, we carry `msg` over to the next call.
+
+            let copy_data_bytes = match msg? {
+                Some(FeMessage::CopyData(bytes)) => bytes,
+                Some(FeMessage::Terminate) => {
+                    return Ok(BatchOrEof::Eof);
+                }
+                Some(m) => {
+                    return Err(QueryError::Other(anyhow::anyhow!(
+                        "unexpected message: {m:?} during COPY"
+                    )));
+                }
+                None => {
+                    return Ok(BatchOrEof::Eof);
+                } // client disconnected
+            };
+            trace!("query: {copy_data_bytes:?}");
+
+            fail::fail_point!("ps::handle-pagerequest-message");
+
+            // parse request
+            let neon_fe_msg = PagestreamFeMessage::parse(&mut copy_data_bytes.reader())?;
+
+            let this_msg = match neon_fe_msg {
+                PagestreamFeMessage::Exists(req) => BatchedFeMessage::Exists {
+                    span: tracing::info_span!("handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.request_lsn),
+                    req,
+                },
+                PagestreamFeMessage::Nblocks(req) => BatchedFeMessage::Nblocks {
+                    span: tracing::info_span!("handle_get_nblocks_request", rel = %req.rel, req_lsn = %req.request_lsn),
+                    req,
+                },
+                PagestreamFeMessage::DbSize(req) => BatchedFeMessage::DbSize {
+                    span: tracing::info_span!("handle_db_size_request", dbnode = %req.dbnode, req_lsn = %req.request_lsn),
+                    req,
+                },
+                PagestreamFeMessage::GetSlruSegment(req) => BatchedFeMessage::GetSlruSegment {
+                    span: tracing::info_span!("handle_get_slru_segment_request", kind = %req.kind, segno = %req.segno, req_lsn = %req.request_lsn),
+                    req,
+                },
+                PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
+                    request_lsn,
+                    not_modified_since,
+                    rel,
+                    blkno,
+                }) => {
+                    // shard_id is filled in by the handler
+                    let span = tracing::info_span!(
+                        "handle_get_page_at_lsn_request_batched",
+                        %tenant_id, %timeline_id, shard_id = tracing::field::Empty, req_lsn = %request_lsn,
+                        batch_size = tracing::field::Empty, batch_id = tracing::field::Empty
+                    );
+
+                    macro_rules! current_batch_and_error {
+                        ($error:expr) => {{
+                            let error = BatchedFeMessage::RespondError {
+                                span,
+                                error: $error,
+                            };
+                            let batch_and_error = match maybe_carry.take() {
+                                Some(carry) => smallvec::smallvec![carry.msg, error],
+                                None => smallvec::smallvec![error],
+                            };
+                            Ok(BatchOrEof::Batch(batch_and_error))
+                        }};
+                    }
+
+                    let key = rel_block_to_key(rel, blkno);
+                    let shard = match self
+                        .timeline_handles
+                        .get(*tenant_id, *timeline_id, ShardSelector::Page(key))
+                        .instrument(span.clone())
+                        .await
+                    {
+                        Ok(tl) => tl,
+                        Err(GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_))) => {
+                            // We already know this tenant exists in general, because we resolved it at
+                            // start of connection.  Getting a NotFound here indicates that the shard containing
+                            // the requested page is not present on this node: the client's knowledge of shard->pageserver
+                            // mapping is out of date.
+                            //
+                            // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
+                            // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
+                            // and talk to a different pageserver.
+                            return current_batch_and_error!(PageStreamError::Reconnect(
+                                "getpage@lsn request routed to wrong shard".into()
+                            ));
+                        }
+                        Err(e) => {
+                            return current_batch_and_error!(e.into());
+                        }
+                    };
+                    let effective_request_lsn = match Self::wait_or_get_last_lsn(
+                        &shard,
+                        request_lsn,
+                        not_modified_since,
+                        &shard.get_latest_gc_cutoff_lsn(),
+                        ctx,
+                    )
+                    // TODO: if we actually need to wait for lsn here, it delays the entire batch which doesn't need to wait
+                    .await
+                    {
+                        Ok(lsn) => lsn,
+                        Err(e) => {
+                            return current_batch_and_error!(e);
+                        }
+                    };
+                    BatchedFeMessage::GetPage {
+                        span,
+                        shard,
+                        effective_request_lsn,
+                        pages: smallvec::smallvec![(rel, blkno)],
+                    }
+                }
+            };
+
+            //
+            // batch
+            //
+            match (maybe_carry.as_mut(), this_msg) {
+                (None, this_msg) => {
+                    *maybe_carry = Some(Carry { msg: this_msg, started_at: msg_start });
+                }
+                (
+                    Some(Carry { msg: BatchedFeMessage::GetPage {
+                        span: _,
+                        shard: accum_shard,
+                        pages: ref mut accum_pages,
+                        effective_request_lsn: accum_lsn,
+                    }, started_at: _}),
+                    BatchedFeMessage::GetPage {
+                        span: _,
+                        shard: this_shard,
+                        pages: this_pages,
+                        effective_request_lsn: this_lsn,
+                    },
+                ) if async {
+                    assert_eq!(this_pages.len(), 1);
+                    if accum_pages.len() >= Timeline::MAX_GET_VECTORED_KEYS as usize {
+                        trace!(%accum_lsn, %this_lsn, "stopping batching because of batch size");
+                        assert_eq!(accum_pages.len(), Timeline::MAX_GET_VECTORED_KEYS as usize);
+                        return false;
+                    }
+                    if (accum_shard.tenant_shard_id, accum_shard.timeline_id)
+                        != (this_shard.tenant_shard_id, this_shard.timeline_id)
+                    {
+                        trace!(%accum_lsn, %this_lsn, "stopping batching because timeline object mismatch");
+                        // TODO: we _could_ batch & execute each shard seperately (and in parallel).
+                        // But the current logic for keeping responses in order does not support that.
+                        return false;
+                    }
+                    // the vectored get currently only supports a single LSN, so, bounce as soon
+                    // as the effective request_lsn changes
+                    if *accum_lsn != this_lsn {
+                        trace!(%accum_lsn, %this_lsn, "stopping batching because LSN changed");
+                        return false;
+                    }
+                    true
+                }
+                .await =>
+                {
+                    // ok to batch
+                    accum_pages.extend(this_pages);
+                }
+                (Some(carry), this_msg) => {
+                    // by default, don't continue batching
+                    let carry = std::mem::replace(carry,
+                        Carry {
+                            msg: this_msg,
+                            started_at: msg_start,
+                        });
+                    return Ok(BatchOrEof::Batch(smallvec::smallvec![carry.msg]));
+                }
+            }
+        }
+    }
+
     /// Pagestream sub-protocol handler.
     ///
     /// It is a simple request-response protocol inside a COPYBOTH session.
@@ -592,133 +899,161 @@ impl PageServerHandler {
             }
         }
 
+        let mut carry: Option<Carry> = None;
+
         loop {
-            // read request bytes (it's exactly 1 PagestreamFeMessage per CopyData)
-            let msg = tokio::select! {
-                biased;
-                _ = self.cancel.cancelled() => {
-                    return Err(QueryError::Shutdown)
-                }
-                msg = pgb.read_message() => { msg }
-            };
-            let copy_data_bytes = match msg? {
-                Some(FeMessage::CopyData(bytes)) => bytes,
-                Some(FeMessage::Terminate) => break,
-                Some(m) => {
-                    return Err(QueryError::Other(anyhow::anyhow!(
-                        "unexpected message: {m:?} during COPY"
-                    )));
-                }
-                None => break, // client disconnected
-            };
-
-            trace!("query: {copy_data_bytes:?}");
-            fail::fail_point!("ps::handle-pagerequest-message");
-
-            // parse request
-            let neon_fe_msg = PagestreamFeMessage::parse(&mut copy_data_bytes.reader())?;
-
-            // invoke handler function
-            let (handler_result, span) = match neon_fe_msg {
-                PagestreamFeMessage::Exists(req) => {
-                    fail::fail_point!("ps::handle-pagerequest-message::exists");
-                    let span = tracing::info_span!("handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.request_lsn);
-                    (
-                        self.handle_get_rel_exists_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
-                }
-                PagestreamFeMessage::Nblocks(req) => {
-                    fail::fail_point!("ps::handle-pagerequest-message::nblocks");
-                    let span = tracing::info_span!("handle_get_nblocks_request", rel = %req.rel, req_lsn = %req.request_lsn);
-                    (
-                        self.handle_get_nblocks_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
-                }
-                PagestreamFeMessage::GetPage(req) => {
-                    fail::fail_point!("ps::handle-pagerequest-message::getpage");
-                    // shard_id is filled in by the handler
-                    let span = tracing::info_span!("handle_get_page_at_lsn_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.request_lsn);
-                    (
-                        self.handle_get_page_at_lsn_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
-                }
-                PagestreamFeMessage::DbSize(req) => {
-                    fail::fail_point!("ps::handle-pagerequest-message::dbsize");
-                    let span = tracing::info_span!("handle_db_size_request", dbnode = %req.dbnode, req_lsn = %req.request_lsn);
-                    (
-                        self.handle_db_size_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
-                }
-                PagestreamFeMessage::GetSlruSegment(req) => {
-                    fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
-                    let span = tracing::info_span!("handle_get_slru_segment_request", kind = %req.kind, segno = %req.segno, req_lsn = %req.request_lsn);
-                    (
-                        self.handle_get_slru_segment_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
+            let maybe_batched = self
+                .read_batch_from_connection(pgb, &tenant_id, &timeline_id, &mut carry, &ctx)
+                .await?;
+            let batched = match maybe_batched {
+                BatchOrEof::Batch(b) => b,
+                BatchOrEof::Eof => {
+                    break;
                 }
             };
 
-            // Map handler result to protocol behavior.
-            // Some handler errors cause exit from pagestream protocol.
-            // Other handler errors are sent back as an error message and we stay in pagestream protocol.
-            let response_msg = match handler_result {
-                Err(e) => match &e {
-                    PageStreamError::Shutdown => {
-                        // If we fail to fulfil a request during shutdown, which may be _because_ of
-                        // shutdown, then do not send the error to the client.  Instead just drop the
-                        // connection.
-                        span.in_scope(|| info!("dropping connection due to shutdown"));
-                        return Err(QueryError::Shutdown);
+            for batch in batched {
+                // invoke handler function
+                let (handler_results, span): (
+                    Vec<Result<PagestreamBeMessage, PageStreamError>>,
+                    _,
+                ) = match batch {
+                    BatchedFeMessage::Exists { span, req } => {
+                        fail::fail_point!("ps::handle-pagerequest-message::exists");
+                        (
+                            vec![
+                                self.handle_get_rel_exists_request(
+                                    tenant_id,
+                                    timeline_id,
+                                    &req,
+                                    &ctx,
+                                )
+                                .instrument(span.clone())
+                                .await,
+                            ],
+                            span,
+                        )
                     }
-                    PageStreamError::Reconnect(reason) => {
-                        span.in_scope(|| info!("handler requested reconnect: {reason}"));
-                        return Err(QueryError::Reconnect);
+                    BatchedFeMessage::Nblocks { span, req } => {
+                        fail::fail_point!("ps::handle-pagerequest-message::nblocks");
+                        (
+                            vec![
+                                self.handle_get_nblocks_request(tenant_id, timeline_id, &req, &ctx)
+                                    .instrument(span.clone())
+                                    .await,
+                            ],
+                            span,
+                        )
                     }
-                    PageStreamError::Read(_)
-                    | PageStreamError::LsnTimeout(_)
-                    | PageStreamError::NotFound(_)
-                    | PageStreamError::BadRequest(_) => {
-                        // print the all details to the log with {:#}, but for the client the
-                        // error message is enough.  Do not log if shutting down, as the anyhow::Error
-                        // here includes cancellation which is not an error.
-                        let full = utils::error::report_compact_sources(&e);
-                        span.in_scope(|| {
-                            error!("error reading relation or page version: {full:#}")
-                        });
-                        PagestreamBeMessage::Error(PagestreamErrorResponse {
-                            message: e.to_string(),
-                        })
+                    BatchedFeMessage::GetPage {
+                        span,
+                        shard,
+                        effective_request_lsn,
+                        pages,
+                    } => {
+                        fail::fail_point!("ps::handle-pagerequest-message::getpage");
+                        (
+                            {
+                                let npages = pages.len();
+                                trace!(npages, "handling getpage request");
+                                let res = self
+                                    .handle_get_page_at_lsn_request_batched(
+                                        &shard,
+                                        effective_request_lsn,
+                                        pages,
+                                        &ctx,
+                                    )
+                                    .instrument(span.clone())
+                                    .await;
+                                assert_eq!(res.len(), npages);
+                                res
+                            },
+                            span,
+                        )
                     }
-                },
-                Ok(response_msg) => response_msg,
-            };
+                    BatchedFeMessage::DbSize { span, req } => {
+                        fail::fail_point!("ps::handle-pagerequest-message::dbsize");
+                        (
+                            vec![
+                                self.handle_db_size_request(tenant_id, timeline_id, &req, &ctx)
+                                    .instrument(span.clone())
+                                    .await,
+                            ],
+                            span,
+                        )
+                    }
+                    BatchedFeMessage::GetSlruSegment { span, req } => {
+                        fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
+                        (
+                            vec![
+                                self.handle_get_slru_segment_request(
+                                    tenant_id,
+                                    timeline_id,
+                                    &req,
+                                    &ctx,
+                                )
+                                .instrument(span.clone())
+                                .await,
+                            ],
+                            span,
+                        )
+                    }
+                    BatchedFeMessage::RespondError { span, error } => {
+                        // We've already decided to respond with an error, so we don't need to
+                        // call the handler.
+                        (vec![Err(error)], span)
+                    }
+                };
 
-            // marshal & transmit response message
-            pgb.write_message_noflush(&BeMessage::CopyData(&response_msg.serialize()))?;
-            tokio::select! {
-                biased;
-                _ = self.cancel.cancelled() => {
-                    // We were requested to shut down.
-                    info!("shutdown request received in page handler");
-                    return Err(QueryError::Shutdown)
+                // Map handler result to protocol behavior.
+                // Some handler errors cause exit from pagestream protocol.
+                // Other handler errors are sent back as an error message and we stay in pagestream protocol.
+                for handler_result in handler_results {
+                    let response_msg = match handler_result {
+                        Err(e) => match &e {
+                            PageStreamError::Shutdown => {
+                                // If we fail to fulfil a request during shutdown, which may be _because_ of
+                                // shutdown, then do not send the error to the client.  Instead just drop the
+                                // connection.
+                                span.in_scope(|| info!("dropping connection due to shutdown"));
+                                return Err(QueryError::Shutdown);
+                            }
+                            PageStreamError::Reconnect(reason) => {
+                                span.in_scope(|| info!("handler requested reconnect: {reason}"));
+                                return Err(QueryError::Reconnect);
+                            }
+                            PageStreamError::Read(_)
+                            | PageStreamError::LsnTimeout(_)
+                            | PageStreamError::NotFound(_)
+                            | PageStreamError::BadRequest(_) => {
+                                // print the all details to the log with {:#}, but for the client the
+                                // error message is enough.  Do not log if shutting down, as the anyhow::Error
+                                // here includes cancellation which is not an error.
+                                let full = utils::error::report_compact_sources(&e);
+                                span.in_scope(|| {
+                                    error!("error reading relation or page version: {full:#}")
+                                });
+                                PagestreamBeMessage::Error(PagestreamErrorResponse {
+                                    message: e.to_string(),
+                                })
+                            }
+                        },
+                        Ok(response_msg) => response_msg,
+                    };
+
+                    // marshal & transmit response message
+                    pgb.write_message_noflush(&BeMessage::CopyData(&response_msg.serialize()))?;
                 }
-                res = pgb.flush() => {
-                    res?;
+                tokio::select! {
+                    biased;
+                    _ = self.cancel.cancelled() => {
+                        // We were requested to shut down.
+                        info!("shutdown request received in page handler");
+                        return Err(QueryError::Shutdown)
+                    }
+                    res = pgb.flush() => {
+                        res?;
+                    }
                 }
             }
         }
@@ -964,60 +1299,30 @@ impl PageServerHandler {
         }))
     }
 
-    #[instrument(skip_all, fields(shard_id))]
-    async fn handle_get_page_at_lsn_request(
+    #[instrument(skip_all)]
+    async fn handle_get_page_at_lsn_request_batched(
         &mut self,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        req: &PagestreamGetPageRequest,
+        timeline: &Timeline,
+        effective_lsn: Lsn,
+        pages: smallvec::SmallVec<[(RelTag, BlockNumber); 1]>,
         ctx: &RequestContext,
-    ) -> Result<PagestreamBeMessage, PageStreamError> {
-        let timeline = match self
-            .timeline_handles
-            .get(
-                tenant_id,
-                timeline_id,
-                ShardSelector::Page(rel_block_to_key(req.rel, req.blkno)),
-            )
-            .await
-        {
-            Ok(tl) => tl,
-            Err(GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_))) => {
-                // We already know this tenant exists in general, because we resolved it at
-                // start of connection.  Getting a NotFound here indicates that the shard containing
-                // the requested page is not present on this node: the client's knowledge of shard->pageserver
-                // mapping is out of date.
-                //
-                // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
-                // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
-                // and talk to a different pageserver.
-                return Err(PageStreamError::Reconnect(
-                    "getpage@lsn request routed to wrong shard".into(),
-                ));
-            }
-            Err(e) => return Err(e.into()),
-        };
-
-        let _timer = timeline
-            .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetPageAtLsn, ctx);
-
-        let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
-        let lsn = Self::wait_or_get_last_lsn(
-            &timeline,
-            req.request_lsn,
-            req.not_modified_since,
-            &latest_gc_cutoff_lsn,
+    ) -> Vec<Result<PagestreamBeMessage, PageStreamError>> {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+        let _timer = timeline.query_metrics.start_timer_many(
+            metrics::SmgrQueryType::GetPageAtLsn,
+            pages.len(),
             ctx,
-        )
-        .await?;
+        );
 
-        let page = timeline
-            .get_rel_page_at_lsn(req.rel, req.blkno, Version::Lsn(lsn), ctx)
-            .await?;
+        let pages = timeline
+            .get_rel_page_at_lsn_batched(pages, effective_lsn, ctx)
+            .await;
 
-        Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
-            page,
+        Vec::from_iter(pages.into_iter().map(|page| {
+            page.map(|page| {
+                PagestreamBeMessage::GetPage(models::PagestreamGetPageResponse { page })
+            })
+            .map_err(PageStreamError::from)
         }))
     }
 
@@ -1674,6 +1979,13 @@ fn set_tracing_field_shard_id(timeline: &Timeline) {
     debug_assert_current_span_has_tenant_and_timeline_id();
 }
 
+struct WaitedForLsn(Lsn);
+impl From<WaitedForLsn> for Lsn {
+    fn from(WaitedForLsn(lsn): WaitedForLsn) -> Self {
+        lsn
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use utils::shard::ShardCount;

pageserver/src/pgdatadir_mapping.rs

@@ -10,10 +10,15 @@ use super::tenant::{PageReconstructError, Timeline};
 use crate::aux_file;
 use crate::context::RequestContext;
 use crate::keyspace::{KeySpace, KeySpaceAccum};
-use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id;
+use crate::span::{
+    debug_assert_current_span_has_tenant_and_timeline_id,
+    debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
+};
+use crate::tenant::timeline::GetVectoredError;
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
+use itertools::Itertools;
 use pageserver_api::key::Key;
 use pageserver_api::key::{
     dbdir_key_range, rel_block_to_key, rel_dir_to_key, rel_key_range, rel_size_to_key,
@@ -30,7 +35,7 @@ use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
 use postgres_ffi::BLCKSZ;
 use postgres_ffi::{Oid, RepOriginId, TimestampTz, TransactionId};
 use serde::{Deserialize, Serialize};
-use std::collections::{hash_map, HashMap, HashSet};
+use std::collections::{hash_map, BTreeMap, HashMap, HashSet};
 use std::ops::ControlFlow;
 use std::ops::Range;
 use strum::IntoEnumIterator;
@@ -193,26 +198,195 @@ impl Timeline {
         version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<Bytes, PageReconstructError> {
-        if tag.relnode == 0 {
-            return Err(PageReconstructError::Other(
-                RelationError::InvalidRelnode.into(),
-            ));
+        match version {
+            Version::Lsn(effective_lsn) => {
+                let pages = smallvec::smallvec![(tag, blknum)];
+                let res = self
+                    .get_rel_page_at_lsn_batched(pages, effective_lsn, ctx)
+                    .await;
+                assert_eq!(res.len(), 1);
+                res.into_iter().next().unwrap()
+            }
+            Version::Modified(modification) => {
+                if tag.relnode == 0 {
+                    return Err(PageReconstructError::Other(
+                        RelationError::InvalidRelnode.into(),
+                    ));
+                }
+
+                let nblocks = self.get_rel_size(tag, version, ctx).await?;
+                if blknum >= nblocks {
+                    debug!(
+                        "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
+                        tag,
+                        blknum,
+                        version.get_lsn(),
+                        nblocks
+                    );
+                    return Ok(ZERO_PAGE.clone());
+                }
+
+                let key = rel_block_to_key(tag, blknum);
+                modification.get(key, ctx).await
+            }
+        }
+    }
+
+    /// Like [`Self::get_rel_page_at_lsn`], but returns a batch of pages.
+    ///
+    /// The ordering of the returned vec corresponds to the ordering of `pages`.
+    pub(crate) async fn get_rel_page_at_lsn_batched(
+        &self,
+        pages: smallvec::SmallVec<[(RelTag, BlockNumber); 1]>,
+        effective_lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Vec<Result<Bytes, PageReconstructError>> {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
+        let mut slots_filled = 0;
+        let page_count = pages.len();
+
+        // Would be nice to use smallvec here but it doesn't provide the spare_capacity_mut() API.
+        let mut result = Vec::with_capacity(pages.len());
+        let result_slots = result.spare_capacity_mut();
+
+        let mut keys_slots: BTreeMap<Key, smallvec::SmallVec<[usize; 1]>> = BTreeMap::default();
+        for (response_slot_idx, (tag, blknum)) in pages.into_iter().enumerate() {
+            if tag.relnode == 0 {
+                result_slots[response_slot_idx].write(Err(PageReconstructError::Other(
+                    RelationError::InvalidRelnode.into(),
+                )));
+
+                slots_filled += 1;
+                continue;
+            }
+
+            let nblocks = match self
+                .get_rel_size(tag, Version::Lsn(effective_lsn), ctx)
+                .await
+            {
+                Ok(nblocks) => nblocks,
+                Err(err) => {
+                    result_slots[response_slot_idx].write(Err(err));
+                    slots_filled += 1;
+                    continue;
+                }
+            };
+
+            if blknum >= nblocks {
+                debug!(
+                    "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
+                    tag, blknum, effective_lsn, nblocks
+                );
+                result_slots[response_slot_idx].write(Ok(ZERO_PAGE.clone()));
+                slots_filled += 1;
+                continue;
+            }
+
+            let key = rel_block_to_key(tag, blknum);
+
+            let key_slots = keys_slots.entry(key).or_default();
+            key_slots.push(response_slot_idx);
         }
 
-        let nblocks = self.get_rel_size(tag, version, ctx).await?;
-        if blknum >= nblocks {
-            debug!(
-                "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
-                tag,
-                blknum,
-                version.get_lsn(),
-                nblocks
-            );
-            return Ok(ZERO_PAGE.clone());
+        let keyspace = {
+            // add_key requires monotonicity
+            let mut acc = KeySpaceAccum::new();
+            for key in keys_slots
+                .keys()
+                // in fact it requires strong monotonicity
+                .dedup()
+            {
+                acc.add_key(*key);
+            }
+            acc.to_keyspace()
+        };
+
+        match self.get_vectored(keyspace, effective_lsn, ctx).await {
+            Ok(results) => {
+                for (key, res) in results {
+                    let mut key_slots = keys_slots.remove(&key).unwrap().into_iter();
+                    let first_slot = key_slots.next().unwrap();
+
+                    for slot in key_slots {
+                        let clone = match &res {
+                            Ok(buf) => Ok(buf.clone()),
+                            Err(err) => Err(match err {
+                                PageReconstructError::Cancelled => {
+                                    PageReconstructError::Cancelled
+                                }
+
+                                x @ PageReconstructError::Other(_) |
+                                x @ PageReconstructError::AncestorLsnTimeout(_) |
+                                x @ PageReconstructError::WalRedo(_) |
+                                x @ PageReconstructError::MissingKey(_) => {
+                                    PageReconstructError::Other(anyhow::anyhow!("there was more than one request for this key in the batch, error logged once: {x:?}"))
+                                },
+                            }),
+                        };
+
+                        result_slots[slot].write(clone);
+                        slots_filled += 1;
+                    }
+
+                    result_slots[first_slot].write(res);
+                    slots_filled += 1;
+                }
+            }
+            Err(err) => {
+                // this cannot really happen because get_vectored only errors globally on invalid LSN or too large batch size
+                // (We enforce the max batch size outside of this function, in the code that constructs the batch request.)
+                for slot in keys_slots.values().flatten() {
+                    // this whole `match` is a lot like `From<GetVectoredError> for PageReconstructError`
+                    // but without taking ownership of the GetVectoredError
+                    let err = match &err {
+                        GetVectoredError::Cancelled => {
+                            Err(PageReconstructError::Cancelled)
+                        }
+                        // TODO: restructure get_vectored API to make this error per-key
+                        GetVectoredError::MissingKey(err) => {
+                            Err(PageReconstructError::Other(anyhow::anyhow!("whole vectored get request failed because one or more of the requested keys were missing: {err:?}")))
+                        }
+                        // TODO: restructure get_vectored API to make this error per-key
+                        GetVectoredError::GetReadyAncestorError(err) => {
+                            Err(PageReconstructError::Other(anyhow::anyhow!("whole vectored get request failed because one or more key required ancestor that wasn't ready: {err:?}")))
+                        }
+                        // TODO: restructure get_vectored API to make this error per-key
+                        GetVectoredError::Other(err) => {
+                            Err(PageReconstructError::Other(
+                                anyhow::anyhow!("whole vectored get request failed: {err:?}"),
+                            ))
+                        }
+                        // TODO: we can prevent this error class by moving this check into the type system
+                        GetVectoredError::InvalidLsn(e) => {
+                            Err(anyhow::anyhow!("invalid LSN: {e:?}").into())
+                        }
+                        // NB: this should never happen in practice because we limit MAX_GET_VECTORED_KEYS
+                        // TODO: we can prevent this error class by moving this check into the type system
+                        GetVectoredError::Oversized(err) => {
+                            Err(anyhow::anyhow!(
+                                "batching oversized: {err:?}"
+                            )
+                            .into())
+                        }
+                    };
+
+                    result_slots[*slot].write(err);
+                }
+
+                slots_filled += keys_slots.values().map(|slots| slots.len()).sum::<usize>();
+            }
+        };
+
+        assert_eq!(slots_filled, page_count);
+        // SAFETY:
+        // 1. `result` and any of its uninint members are not read from until this point
+        // 2. The length below is tracked at run-time and matches the number of requested pages.
+        unsafe {
+            result.set_len(page_count);
         }
 
-        let key = rel_block_to_key(tag, blknum);
-        version.get(self, key, ctx).await
+        result
     }
 
     // Get size of a database in blocks

pageserver/src/tenant/mgr.rs

@@ -1719,10 +1719,11 @@ impl TenantManager {
                     parent_layers.push(relative_path.to_owned());
                 }
             }
-            debug_assert!(
-                !parent_layers.is_empty(),
-                "shutdown cannot empty the layermap"
-            );
+
+            if parent_layers.is_empty() {
+                tracing::info!("Ancestor shard has no resident layer to hard link");
+            }
+
             (parent_timelines, parent_layers)
         };
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -197,6 +197,7 @@ use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
 use utils::pausable_failpoint;
+use utils::shard::ShardNumber;
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
@@ -240,10 +241,11 @@ use utils::id::{TenantId, TimelineId};
 
 use self::index::IndexPart;
 
+use super::config::AttachedLocationConfig;
 use super::metadata::MetadataUpdate;
 use super::storage_layer::{Layer, LayerName, ResidentLayer};
 use super::upload_queue::{NotInitialized, SetDeletedFlagProgress};
-use super::Generation;
+use super::{DeleteTimelineError, Generation};
 
 pub(crate) use download::{
     download_index_part, download_tenant_manifest, is_temp_download_file,
@@ -301,6 +303,36 @@ pub enum WaitCompletionError {
 #[derive(Debug, thiserror::Error)]
 #[error("Upload queue either in unexpected state or hasn't downloaded manifest yet")]
 pub struct UploadQueueNotReadyError;
+/// Behavioral modes that enable seamless live migration.
+///
+/// See docs/rfcs/028-pageserver-migration.md to understand how these fit in.
+struct RemoteTimelineClientConfig {
+    /// If this is false, then update to remote_consistent_lsn are dropped rather
+    /// than being submitted to DeletionQueue for validation.  This behavior is
+    /// used when a tenant attachment is known to have a stale generation number,
+    /// such that validation attempts will always fail.  This is not necessary
+    /// for correctness, but avoids spamming error statistics with failed validations
+    /// when doing migrations of tenants.
+    process_remote_consistent_lsn_updates: bool,
+
+    /// If this is true, then object deletions are held in a buffer in RemoteTimelineClient
+    /// rather than being submitted to the DeletionQueue.  This behavior is used when a tenant
+    /// is known to be multi-attached, in order to avoid disrupting other attached tenants
+    /// whose generations' metadata refers to the deleted objects.
+    block_deletions: bool,
+}
+
+/// RemoteTimelineClientConfig's state is entirely driven by LocationConf, but we do
+/// not carry the entire LocationConf structure: it's much more than we need.  The From
+/// impl extracts the subset of the LocationConf that is interesting to RemoteTimelineClient.
+impl From<&AttachedLocationConfig> for RemoteTimelineClientConfig {
+    fn from(lc: &AttachedLocationConfig) -> Self {
+        Self {
+            block_deletions: !lc.may_delete_layers_hint(),
+            process_remote_consistent_lsn_updates: lc.may_upload_layers_hint(),
+        }
+    }
+}
 
 /// A client for accessing a timeline's data in remote storage.
 ///
@@ -321,7 +353,7 @@ pub struct UploadQueueNotReadyError;
 /// in the index part file, whenever timeline metadata is uploaded.
 ///
 /// Downloads are not queued, they are performed immediately.
-pub struct RemoteTimelineClient {
+pub(crate) struct RemoteTimelineClient {
     conf: &'static PageServerConf,
 
     runtime: tokio::runtime::Handle,
@@ -338,6 +370,9 @@ pub struct RemoteTimelineClient {
 
     deletion_queue_client: DeletionQueueClient,
 
+    /// Subset of tenant configuration used to control upload behaviors during migrations
+    config: std::sync::RwLock<RemoteTimelineClientConfig>,
+
     cancel: CancellationToken,
 }
 
@@ -348,13 +383,14 @@ impl RemoteTimelineClient {
     /// Note: the caller must initialize the upload queue before any uploads can be scheduled,
     /// by calling init_upload_queue.
     ///
-    pub fn new(
+    pub(crate) fn new(
         remote_storage: GenericRemoteStorage,
         deletion_queue_client: DeletionQueueClient,
         conf: &'static PageServerConf,
         tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         generation: Generation,
+        location_conf: &AttachedLocationConfig,
     ) -> RemoteTimelineClient {
         RemoteTimelineClient {
             conf,
@@ -374,6 +410,7 @@ impl RemoteTimelineClient {
                 &tenant_shard_id,
                 &timeline_id,
             )),
+            config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(location_conf)),
             cancel: CancellationToken::new(),
         }
     }
@@ -429,6 +466,43 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    /// Notify this client of a change to its parent tenant's config, as this may cause us to
+    /// take action (unblocking deletions when transitioning from AttachedMulti to AttachedSingle)
+    pub(super) fn update_config(&self, location_conf: &AttachedLocationConfig) {
+        let new_conf = RemoteTimelineClientConfig::from(location_conf);
+        let unblocked = !new_conf.block_deletions;
+
+        // Update config before draining deletions, so that we don't race with more being
+        // inserted.  This can result in deletions happening our of order, but that does not
+        // violate any invariants: deletions only need to be ordered relative to upload of the index
+        // that dereferences the deleted objects, and we are not changing that order.
+        *self.config.write().unwrap() = new_conf;
+
+        if unblocked {
+            // If we may now delete layers, drain any that were blocked in our old
+            // configuration state
+            let mut queue_locked = self.upload_queue.lock().unwrap();
+
+            if let Ok(queue) = queue_locked.initialized_mut() {
+                let blocked_deletions = std::mem::take(&mut queue.blocked_deletions);
+                for d in blocked_deletions {
+                    if let Err(e) = self.deletion_queue_client.push_layers_sync(
+                        self.tenant_shard_id,
+                        self.timeline_id,
+                        self.generation,
+                        d.layers,
+                    ) {
+                        // This could happen if the pageserver is shut down while a tenant
+                        // is transitioning from a deletion-blocked state: we will leak some
+                        // S3 objects in this case.
+                        warn!("Failed to drain blocked deletions: {}", e);
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
     /// Returns `None` if nothing is yet uplodaded, `Some(disk_consistent_lsn)` otherwise.
     pub fn remote_consistent_lsn_projected(&self) -> Option<Lsn> {
         match &mut *self.upload_queue.lock().unwrap() {
@@ -574,12 +648,18 @@ 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(),
+                ));
             }
         }
 
@@ -1544,15 +1624,17 @@ impl RemoteTimelineClient {
     /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
     /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
     /// deletes leaked files if any and proceeds with deletion of index file at the end.
-    pub(crate) async fn delete_all(self: &Arc<Self>) -> anyhow::Result<()> {
+    pub(crate) async fn delete_all(self: &Arc<Self>) -> Result<(), DeleteTimelineError> {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
         let layers: Vec<RemotePath> = {
             let mut locked = self.upload_queue.lock().unwrap();
-            let stopped = locked.stopped_mut()?;
+            let stopped = locked.stopped_mut().map_err(DeleteTimelineError::Other)?;
 
             if !matches!(stopped.deleted_at, SetDeletedFlagProgress::Successful(_)) {
-                anyhow::bail!("deleted_at is not set")
+                return Err(DeleteTimelineError::Other(anyhow::anyhow!(
+                    "deleted_at is not set"
+                )));
             }
 
             debug_assert!(stopped.upload_queue_for_deletion.no_pending_work());
@@ -1587,7 +1669,10 @@ impl RemoteTimelineClient {
         };
 
         let layer_deletion_count = layers.len();
-        self.deletion_queue_client.push_immediate(layers).await?;
+        self.deletion_queue_client
+            .push_immediate(layers)
+            .await
+            .map_err(|_| DeleteTimelineError::Cancelled)?;
 
         // Delete the initdb.tar.zst, which is not always present, but deletion attempts of
         // inexistant objects are not considered errors.
@@ -1595,7 +1680,8 @@ impl RemoteTimelineClient {
             remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &self.timeline_id);
         self.deletion_queue_client
             .push_immediate(vec![initdb_path])
-            .await?;
+            .await
+            .map_err(|_| DeleteTimelineError::Cancelled)?;
 
         // Do not delete index part yet, it is needed for possible retry. If we remove it first
         // and retry will arrive to different pageserver there wont be any traces of it on remote storage
@@ -1603,7 +1689,9 @@ impl RemoteTimelineClient {
 
         // Execute all pending deletions, so that when we proceed to do a listing below, we aren't
         // taking the burden of listing all the layers that we already know we should delete.
-        self.flush_deletion_queue().await?;
+        self.flush_deletion_queue()
+            .await
+            .map_err(|_| DeleteTimelineError::Cancelled)?;
 
         let cancel = shutdown_token();
 
@@ -1666,28 +1754,32 @@ impl RemoteTimelineClient {
         if !remaining_layers.is_empty() {
             self.deletion_queue_client
                 .push_immediate(remaining_layers)
-                .await?;
+                .await
+                .map_err(|_| DeleteTimelineError::Cancelled)?;
         }
 
         fail::fail_point!("timeline-delete-before-index-delete", |_| {
-            Err(anyhow::anyhow!(
+            Err(DeleteTimelineError::Other(anyhow::anyhow!(
                 "failpoint: timeline-delete-before-index-delete"
-            ))?
+            )))?
         });
 
         debug!("enqueuing index part deletion");
         self.deletion_queue_client
             .push_immediate([latest_index].to_vec())
-            .await?;
+            .await
+            .map_err(|_| DeleteTimelineError::Cancelled)?;
 
         // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
         // for a flush to a persistent deletion list so that we may be sure deletion will occur.
-        self.flush_deletion_queue().await?;
+        self.flush_deletion_queue()
+            .await
+            .map_err(|_| DeleteTimelineError::Cancelled)?;
 
         fail::fail_point!("timeline-delete-after-index-delete", |_| {
-            Err(anyhow::anyhow!(
+            Err(DeleteTimelineError::Other(anyhow::anyhow!(
                 "failpoint: timeline-delete-after-index-delete"
-            ))?
+            )))?
         });
 
         info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
@@ -1894,16 +1986,24 @@ impl RemoteTimelineClient {
                     res
                 }
                 UploadOp::Delete(delete) => {
-                    pausable_failpoint!("before-delete-layer-pausable");
-                    self.deletion_queue_client
-                        .push_layers(
-                            self.tenant_shard_id,
-                            self.timeline_id,
-                            self.generation,
-                            delete.layers.clone(),
-                        )
-                        .await
-                        .map_err(|e| anyhow::anyhow!(e))
+                    if self.config.read().unwrap().block_deletions {
+                        let mut queue_locked = self.upload_queue.lock().unwrap();
+                        if let Ok(queue) = queue_locked.initialized_mut() {
+                            queue.blocked_deletions.push(delete.clone());
+                        }
+                        Ok(())
+                    } else {
+                        pausable_failpoint!("before-delete-layer-pausable");
+                        self.deletion_queue_client
+                            .push_layers(
+                                self.tenant_shard_id,
+                                self.timeline_id,
+                                self.generation,
+                                delete.layers.clone(),
+                            )
+                            .await
+                            .map_err(|e| anyhow::anyhow!(e))
+                    }
                 }
                 unexpected @ UploadOp::Barrier(_) | unexpected @ UploadOp::Shutdown => {
                     // unreachable. Barrier operations are handled synchronously in
@@ -2010,8 +2110,16 @@ impl RemoteTimelineClient {
                         // Legacy mode: skip validating generation
                         upload_queue.visible_remote_consistent_lsn.store(lsn);
                         None
-                    } else {
+                    } else if self
+                        .config
+                        .read()
+                        .unwrap()
+                        .process_remote_consistent_lsn_updates
+                    {
                         Some((lsn, upload_queue.visible_remote_consistent_lsn.clone()))
+                    } else {
+                        // Our config disables remote_consistent_lsn updates: drop it.
+                        None
                     }
                 }
                 UploadOp::Delete(_) => {
@@ -2148,6 +2256,7 @@ impl RemoteTimelineClient {
                         queued_operations: VecDeque::default(),
                         #[cfg(feature = "testing")]
                         dangling_files: HashMap::default(),
+                        blocked_deletions: Vec::new(),
                         shutting_down: false,
                         shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
                     };
@@ -2213,6 +2322,28 @@ impl RemoteTimelineClient {
             UploadQueue::Initialized(x) => x.no_pending_work(),
         }
     }
+
+    /// 'foreign' in the sense that it does not belong to this tenant shard.  This method
+    /// is used during GC for other shards to get the index of shard zero.
+    pub(crate) async fn download_foreign_index(
+        &self,
+        shard_number: ShardNumber,
+        cancel: &CancellationToken,
+    ) -> Result<(IndexPart, Generation, std::time::SystemTime), DownloadError> {
+        let foreign_shard_id = TenantShardId {
+            shard_number,
+            shard_count: self.tenant_shard_id.shard_count,
+            tenant_id: self.tenant_shard_id.tenant_id,
+        };
+        download_index_part(
+            &self.storage_impl,
+            &foreign_shard_id,
+            &self.timeline_id,
+            Generation::MAX,
+            cancel,
+        )
+        .await
+    }
 }
 
 pub(crate) struct UploadQueueAccessor<'a> {
@@ -2361,6 +2492,7 @@ mod tests {
     use crate::{
         context::RequestContext,
         tenant::{
+            config::AttachmentMode,
             harness::{TenantHarness, TIMELINE_ID},
             storage_layer::layer::local_layer_path,
             Tenant, Timeline,
@@ -2446,6 +2578,10 @@ mod tests {
 
         /// Construct a RemoteTimelineClient in an arbitrary generation
         fn build_client(&self, generation: Generation) -> Arc<RemoteTimelineClient> {
+            let location_conf = AttachedLocationConfig {
+                generation,
+                attach_mode: AttachmentMode::Single,
+            };
             Arc::new(RemoteTimelineClient {
                 conf: self.harness.conf,
                 runtime: tokio::runtime::Handle::current(),
@@ -2459,6 +2595,7 @@ mod tests {
                     &self.harness.tenant_shard_id,
                     &TIMELINE_ID,
                 )),
+                config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(&location_conf)),
                 cancel: CancellationToken::new(),
             })
         }

pageserver/src/tenant/secondary.rs

@@ -111,15 +111,6 @@ pub(crate) struct SecondaryTenant {
     pub(super) heatmap_total_size_metric: UIntGauge,
 }
 
-impl Drop for SecondaryTenant {
-    fn drop(&mut self) {
-        let tenant_id = self.tenant_shard_id.tenant_id.to_string();
-        let shard_id = format!("{}", self.tenant_shard_id.shard_slug());
-        let _ = SECONDARY_RESIDENT_PHYSICAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
-        let _ = SECONDARY_HEATMAP_TOTAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
-    }
-}
-
 impl SecondaryTenant {
     pub(crate) fn new(
         tenant_shard_id: TenantShardId,
@@ -167,6 +158,13 @@ impl SecondaryTenant {
 
         // Wait for any secondary downloader work to complete
         self.gate.close().await;
+
+        self.validate_metrics();
+
+        let tenant_id = self.tenant_shard_id.tenant_id.to_string();
+        let shard_id = format!("{}", self.tenant_shard_id.shard_slug());
+        let _ = SECONDARY_RESIDENT_PHYSICAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
+        let _ = SECONDARY_HEATMAP_TOTAL_SIZE.remove_label_values(&[&tenant_id, &shard_id]);
     }
 
     pub(crate) fn set_config(&self, config: &SecondaryLocationConfig) {
@@ -254,6 +252,20 @@ impl SecondaryTenant {
         .await
         .expect("secondary eviction should not have panicked");
     }
+
+    /// Exhaustive check that incrementally updated metrics match the actual state.
+    #[cfg(feature = "testing")]
+    fn validate_metrics(&self) {
+        let detail = self.detail.lock().unwrap();
+        let resident_size = detail.total_resident_size();
+
+        assert_eq!(resident_size, self.resident_size_metric.get());
+    }
+
+    #[cfg(not(feature = "testing"))]
+    fn validate_metrics(&self) {
+        // No-op in non-testing builds
+    }
 }
 
 /// The SecondaryController is a pseudo-rpc client for administrative control of secondary mode downloads,

pageserver/src/tenant/secondary/downloader.rs

@@ -242,6 +242,19 @@ impl SecondaryDetail {
         }
     }
 
+    #[cfg(feature = "testing")]
+    pub(crate) fn total_resident_size(&self) -> u64 {
+        self.timelines
+            .values()
+            .map(|tl| {
+                tl.on_disk_layers
+                    .values()
+                    .map(|v| v.metadata.file_size)
+                    .sum::<u64>()
+            })
+            .sum::<u64>()
+    }
+
     pub(super) fn evict_layer(
         &mut self,
         name: LayerName,
@@ -763,24 +776,7 @@ impl<'a> TenantDownloader<'a> {
         }
 
         // Metrics consistency check in testing builds
-        if cfg!(feature = "testing") {
-            let detail = self.secondary_state.detail.lock().unwrap();
-            let resident_size = detail
-                .timelines
-                .values()
-                .map(|tl| {
-                    tl.on_disk_layers
-                        .values()
-                        .map(|v| v.metadata.file_size)
-                        .sum::<u64>()
-                })
-                .sum::<u64>();
-            assert_eq!(
-                resident_size,
-                self.secondary_state.resident_size_metric.get()
-            );
-        }
-
+        self.secondary_state.validate_metrics();
         // Only update last_etag after a full successful download: this way will not skip
         // the next download, even if the heatmap's actual etag is unchanged.
         self.secondary_state.detail.lock().unwrap().last_download = Some(DownloadSummary {

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/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

@@ -38,6 +38,7 @@ use pageserver_api::{
     shard::{ShardIdentity, ShardNumber, TenantShardId},
 };
 use rand::Rng;
+use remote_storage::DownloadError;
 use serde_with::serde_as;
 use storage_broker::BrokerClientChannel;
 use tokio::{
@@ -272,7 +273,7 @@ pub struct Timeline {
 
     /// Remote storage client.
     /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
-    pub remote_client: Arc<RemoteTimelineClient>,
+    pub(crate) remote_client: Arc<RemoteTimelineClient>,
 
     // What page versions do we hold in the repository? If we get a
     // request > last_record_lsn, we need to wait until we receive all
@@ -477,8 +478,31 @@ impl GcInfo {
         self.retain_lsns.sort_by_key(|i| i.0);
     }
 
-    pub(super) fn remove_child(&mut self, child_id: TimelineId) {
-        self.retain_lsns.retain(|i| i.1 != child_id);
+    pub(super) fn remove_child_maybe_offloaded(
+        &mut self,
+        child_id: TimelineId,
+        maybe_offloaded: MaybeOffloaded,
+    ) -> bool {
+        // Remove at most one element. Needed for correctness if there is two live `Timeline` objects referencing
+        // the same timeline. Shouldn't but maybe can occur when Arc's live longer than intended.
+        let mut removed = false;
+        self.retain_lsns.retain(|i| {
+            if removed {
+                return true;
+            }
+            let remove = i.1 == child_id && i.2 == maybe_offloaded;
+            removed |= remove;
+            !remove
+        });
+        removed
+    }
+
+    pub(super) fn remove_child_not_offloaded(&mut self, child_id: TimelineId) -> bool {
+        self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::No)
+    }
+
+    pub(super) fn remove_child_offloaded(&mut self, child_id: TimelineId) -> bool {
+        self.remove_child_maybe_offloaded(child_id, MaybeOffloaded::Yes)
     }
 }
 
@@ -751,6 +775,21 @@ pub(crate) enum CompactFlags {
     DryRun,
 }
 
+#[serde_with::serde_as]
+#[derive(Debug, Clone, serde::Deserialize)]
+pub(crate) struct CompactRange {
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub start: Key,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub end: Key,
+}
+
+#[derive(Clone, Default)]
+pub(crate) struct CompactOptions {
+    pub flags: EnumSet<CompactFlags>,
+    pub compact_range: Option<CompactRange>,
+}
+
 impl std::fmt::Debug for Timeline {
     fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
         write!(f, "Timeline<{}>", self.timeline_id)
@@ -1589,6 +1628,25 @@ impl Timeline {
         cancel: &CancellationToken,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
+    ) -> Result<bool, CompactionError> {
+        self.compact_with_options(
+            cancel,
+            CompactOptions {
+                flags,
+                compact_range: None,
+            },
+            ctx,
+        )
+        .await
+    }
+
+    /// Outermost timeline compaction operation; downloads needed layers. Returns whether we have pending
+    /// compaction tasks.
+    pub(crate) async fn compact_with_options(
+        self: &Arc<Self>,
+        cancel: &CancellationToken,
+        options: CompactOptions,
+        ctx: &RequestContext,
     ) -> Result<bool, CompactionError> {
         // most likely the cancellation token is from background task, but in tests it could be the
         // request task as well.
@@ -1626,7 +1684,7 @@ impl Timeline {
                 self.compact_tiered(cancel, ctx).await?;
                 Ok(false)
             }
-            CompactionAlgorithm::Legacy => self.compact_legacy(cancel, flags, ctx).await,
+            CompactionAlgorithm::Legacy => self.compact_legacy(cancel, options, ctx).await,
         }
     }
 
@@ -2114,14 +2172,14 @@ impl Timeline {
             )
     }
 
-    pub(super) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
+    pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
         // NB: Most tenant conf options are read by background loops, so,
         // changes will automatically be picked up.
 
         // The threshold is embedded in the metric. So, we need to update it.
         {
             let new_threshold = Self::get_evictions_low_residence_duration_metric_threshold(
-                new_conf,
+                &new_conf.tenant_conf,
                 &self.conf.default_tenant_conf,
             );
 
@@ -2129,6 +2187,9 @@ impl Timeline {
             let shard_id_str = format!("{}", self.tenant_shard_id.shard_slug());
 
             let timeline_id_str = self.timeline_id.to_string();
+
+            self.remote_client.update_config(&new_conf.location);
+
             self.metrics
                 .evictions_with_low_residence_duration
                 .write()
@@ -4501,7 +4562,10 @@ impl Drop for Timeline {
             // This lock should never be poisoned, but in case it is we do a .map() instead of
             // an unwrap(), to avoid panicking in a destructor and thereby aborting the process.
             if let Ok(mut gc_info) = ancestor.gc_info.write() {
-                gc_info.remove_child(self.timeline_id)
+                if !gc_info.remove_child_not_offloaded(self.timeline_id) {
+                    tracing::error!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id,
+                        "Couldn't remove retain_lsn entry from offloaded timeline's parent: already removed");
+                }
             }
         }
     }
@@ -4761,6 +4825,86 @@ impl Timeline {
         Ok(())
     }
 
+    async fn find_gc_time_cutoff(
+        &self,
+        pitr: Duration,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
+    ) -> Result<Option<Lsn>, PageReconstructError> {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+        if self.shard_identity.is_shard_zero() {
+            // Shard Zero has SLRU data and can calculate the PITR time -> LSN mapping itself
+            let now = SystemTime::now();
+            let time_range = if pitr == Duration::ZERO {
+                humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
+            } else {
+                pitr
+            };
+
+            // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
+            let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
+            let timestamp = to_pg_timestamp(time_cutoff);
+
+            let time_cutoff = match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
+                LsnForTimestamp::Present(lsn) => Some(lsn),
+                LsnForTimestamp::Future(lsn) => {
+                    // The timestamp is in the future. That sounds impossible,
+                    // but what it really means is that there hasn't been
+                    // any commits since the cutoff timestamp.
+                    //
+                    // In this case we should use the LSN of the most recent commit,
+                    // which is implicitly the last LSN in the log.
+                    debug!("future({})", lsn);
+                    Some(self.get_last_record_lsn())
+                }
+                LsnForTimestamp::Past(lsn) => {
+                    debug!("past({})", lsn);
+                    None
+                }
+                LsnForTimestamp::NoData(lsn) => {
+                    debug!("nodata({})", lsn);
+                    None
+                }
+            };
+            Ok(time_cutoff)
+        } else {
+            // Shards other than shard zero cannot do timestamp->lsn lookups, and must instead learn their GC cutoff
+            // from shard zero's index.  The index doesn't explicitly tell us the time cutoff, but we may assume that
+            // the point up to which shard zero's last_gc_cutoff has advanced will either be the time cutoff, or a
+            // space cutoff that we would also have respected ourselves.
+            match self
+                .remote_client
+                .download_foreign_index(ShardNumber(0), cancel)
+                .await
+            {
+                Ok((index_part, index_generation, _index_mtime)) => {
+                    tracing::info!("GC loaded shard zero metadata (gen {index_generation:?}): latest_gc_cutoff_lsn: {}",
+                        index_part.metadata.latest_gc_cutoff_lsn());
+                    Ok(Some(index_part.metadata.latest_gc_cutoff_lsn()))
+                }
+                Err(DownloadError::NotFound) => {
+                    // This is unexpected, because during timeline creations shard zero persists to remote
+                    // storage before other shards are called, and during timeline deletion non-zeroth shards are
+                    // deleted before the zeroth one.  However, it should be harmless: if we somehow end up in this
+                    // state, then shard zero should _eventually_ write an index when it GCs.
+                    tracing::warn!("GC couldn't find shard zero's index for timeline");
+                    Ok(None)
+                }
+                Err(e) => {
+                    // TODO: this function should return a different error type than page reconstruct error
+                    Err(PageReconstructError::Other(anyhow::anyhow!(e)))
+                }
+            }
+
+            // TODO: after reading shard zero's GC cutoff, we should validate its generation with the storage
+            // controller.  Otherwise, it is possible that we see the GC cutoff go backwards while shard zero
+            // is going through a migration if we read the old location's index and it has GC'd ahead of the
+            // new location.  This is legal in principle, but problematic in practice because it might result
+            // in a timeline creation succeeding on shard zero ('s new location) but then failing on other shards
+            // because they have GC'd past the branch point.
+        }
+    }
+
     /// Find the Lsns above which layer files need to be retained on
     /// garbage collection.
     ///
@@ -4803,40 +4947,7 @@ impl Timeline {
         // - if PITR interval is set, then this is our cutoff.
         // - if PITR interval is not set, then we do a lookup
         //   based on DEFAULT_PITR_INTERVAL, so that size-based retention does not result in keeping history around permanently on idle databases.
-        let time_cutoff = {
-            let now = SystemTime::now();
-            let time_range = if pitr == Duration::ZERO {
-                humantime::parse_duration(DEFAULT_PITR_INTERVAL).expect("constant is invalid")
-            } else {
-                pitr
-            };
-
-            // If PITR is so large or `now` is so small that this underflows, we will retain no history (highly unexpected case)
-            let time_cutoff = now.checked_sub(time_range).unwrap_or(now);
-            let timestamp = to_pg_timestamp(time_cutoff);
-
-            match self.find_lsn_for_timestamp(timestamp, cancel, ctx).await? {
-                LsnForTimestamp::Present(lsn) => Some(lsn),
-                LsnForTimestamp::Future(lsn) => {
-                    // The timestamp is in the future. That sounds impossible,
-                    // but what it really means is that there hasn't been
-                    // any commits since the cutoff timestamp.
-                    //
-                    // In this case we should use the LSN of the most recent commit,
-                    // which is implicitly the last LSN in the log.
-                    debug!("future({})", lsn);
-                    Some(self.get_last_record_lsn())
-                }
-                LsnForTimestamp::Past(lsn) => {
-                    debug!("past({})", lsn);
-                    None
-                }
-                LsnForTimestamp::NoData(lsn) => {
-                    debug!("nodata({})", lsn);
-                    None
-                }
-            }
-        };
+        let time_cutoff = self.find_gc_time_cutoff(pitr, cancel, ctx).await?;
 
         Ok(match (pitr, time_cutoff) {
             (Duration::ZERO, Some(time_cutoff)) => {
@@ -5030,7 +5141,7 @@ impl Timeline {
 
             // 1. Is it newer than GC horizon cutoff point?
             if l.get_lsn_range().end > space_cutoff {
-                debug!(
+                info!(
                     "keeping {} because it's newer than space_cutoff {}",
                     l.layer_name(),
                     space_cutoff,
@@ -5041,7 +5152,7 @@ impl Timeline {
 
             // 2. It is newer than PiTR cutoff point?
             if l.get_lsn_range().end > time_cutoff {
-                debug!(
+                info!(
                     "keeping {} because it's newer than time_cutoff {}",
                     l.layer_name(),
                     time_cutoff,
@@ -5060,7 +5171,7 @@ impl Timeline {
             for retain_lsn in &retain_lsns {
                 // start_lsn is inclusive
                 if &l.get_lsn_range().start <= retain_lsn {
-                    debug!(
+                    info!(
                         "keeping {} because it's still might be referenced by child branch forked at {} is_dropped: xx is_incremental: {}",
                         l.layer_name(),
                         retain_lsn,
@@ -5075,7 +5186,7 @@ impl Timeline {
             if let Some(lsn) = &max_lsn_with_valid_lease {
                 // keep if layer start <= any of the lease
                 if &l.get_lsn_range().start <= lsn {
-                    debug!(
+                    info!(
                         "keeping {} because there is a valid lease preventing GC at {}",
                         l.layer_name(),
                         lsn,
@@ -5107,13 +5218,13 @@ impl Timeline {
             if !layers
                 .image_layer_exists(&l.get_key_range(), &(l.get_lsn_range().end..new_gc_cutoff))
             {
-                debug!("keeping {} because it is the latest layer", l.layer_name());
+                info!("keeping {} because it is the latest layer", l.layer_name());
                 result.layers_not_updated += 1;
                 continue 'outer;
             }
 
             // We didn't find any reason to keep this file, so remove it.
-            debug!(
+            info!(
                 "garbage collecting {} is_dropped: xx is_incremental: {}",
                 l.layer_name(),
                 l.is_incremental(),

pageserver/src/tenant/timeline/compaction.rs

@@ -4,13 +4,13 @@
 //!
 //! 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;
 
 use super::layer_manager::LayerManager;
 use super::{
-    CompactFlags, CreateImageLayersError, DurationRecorder, ImageLayerCreationMode,
+    CompactFlags, CompactOptions, CreateImageLayersError, DurationRecorder, ImageLayerCreationMode,
     RecordedDuration, Timeline,
 };
 
@@ -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))]
@@ -256,22 +273,32 @@ impl Timeline {
     pub(crate) async fn compact_legacy(
         self: &Arc<Self>,
         cancel: &CancellationToken,
-        flags: EnumSet<CompactFlags>,
+        options: CompactOptions,
         ctx: &RequestContext,
     ) -> Result<bool, CompactionError> {
-        if flags.contains(CompactFlags::EnhancedGcBottomMostCompaction) {
-            self.compact_with_gc(cancel, flags, ctx)
+        if options
+            .flags
+            .contains(CompactFlags::EnhancedGcBottomMostCompaction)
+        {
+            self.compact_with_gc(cancel, options, ctx)
                 .await
                 .map_err(CompactionError::Other)?;
             return Ok(false);
         }
 
-        if flags.contains(CompactFlags::DryRun) {
+        if options.flags.contains(CompactFlags::DryRun) {
             return Err(CompactionError::Other(anyhow!(
                 "dry-run mode is not supported for legacy compaction for now"
             )));
         }
 
+        if options.compact_range.is_some() {
+            // maybe useful in the future? could implement this at some point
+            return Err(CompactionError::Other(anyhow!(
+                "compaction range is not supported for legacy compaction for now"
+            )));
+        }
+
         // High level strategy for compaction / image creation:
         //
         // 1. First, calculate the desired "partitioning" of the
@@ -321,7 +348,7 @@ impl Timeline {
             .repartition(
                 self.get_last_record_lsn(),
                 self.get_compaction_target_size(),
-                flags,
+                options.flags,
                 ctx,
             )
             .await
@@ -337,7 +364,7 @@ impl Timeline {
                 let fully_compacted = self
                     .compact_level0(
                         target_file_size,
-                        flags.contains(CompactFlags::ForceL0Compaction),
+                        options.flags.contains(CompactFlags::ForceL0Compaction),
                         ctx,
                     )
                     .await?;
@@ -355,7 +382,10 @@ impl Timeline {
                         .create_image_layers(
                             &partitioning,
                             lsn,
-                            if flags.contains(CompactFlags::ForceImageLayerCreation) {
+                            if options
+                                .flags
+                                .contains(CompactFlags::ForceImageLayerCreation)
+                            {
                                 ImageLayerCreationMode::Force
                             } else {
                                 ImageLayerCreationMode::Try
@@ -1719,10 +1749,19 @@ impl Timeline {
     pub(crate) async fn compact_with_gc(
         self: &Arc<Self>,
         cancel: &CancellationToken,
-        flags: EnumSet<CompactFlags>,
+        options: CompactOptions,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        self.partial_compact_with_gc(None, cancel, flags, ctx).await
+        self.partial_compact_with_gc(
+            options
+                .compact_range
+                .map(|range| range.start..range.end)
+                .unwrap_or_else(|| Key::MIN..Key::MAX),
+            cancel,
+            options.flags,
+            ctx,
+        )
+        .await
     }
 
     /// An experimental compaction building block that combines compaction with garbage collection.
@@ -1732,12 +1771,15 @@ 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 +1804,8 @@ 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 +1821,7 @@ 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 +1851,21 @@ 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 +1873,59 @@ 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 +1936,25 @@ 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 +1973,8 @@ 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 +1986,7 @@ 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 +2006,13 @@ 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 +2038,59 @@ 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 self.shard_identity.is_key_disposable(&key) {
+                // If this shard does not need to store this key, simply skip it.
+                //
+                // This is not handled in the filter iterator because shard is determined by hash.
+                // Therefore, it does not give us any performance benefit to do things like skip
+                // a whole layer file as handling key spaces (ranges).
+                continue;
+            }
+            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 +2101,27 @@ 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 +2132,43 @@ 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 +2176,7 @@ 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 +2197,8 @@ 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 +2206,84 @@ 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 +2302,7 @@ 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/delete.rs

@@ -5,6 +5,7 @@ use std::{
 
 use anyhow::Context;
 use pageserver_api::{models::TimelineState, shard::TenantShardId};
+use remote_storage::DownloadError;
 use tokio::sync::OwnedMutexGuard;
 use tracing::{error, info, info_span, instrument, Instrument};
 use utils::{crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
@@ -16,7 +17,7 @@ use crate::{
         metadata::TimelineMetadata,
         remote_timeline_client::{PersistIndexPartWithDeletedFlagError, RemoteTimelineClient},
         CreateTimelineCause, DeleteTimelineError, MaybeDeletedIndexPart, Tenant,
-        TimelineOrOffloaded,
+        TenantManifestError, TimelineOrOffloaded,
     },
     virtual_file::MaybeFatalIo,
 };
@@ -110,13 +111,6 @@ pub(super) async fn delete_local_timeline_directory(
     info!("finished deleting layer files, releasing locks");
 }
 
-/// Removes remote layers and an index file after them.
-async fn delete_remote_layers_and_index(
-    remote_client: &Arc<RemoteTimelineClient>,
-) -> anyhow::Result<()> {
-    remote_client.delete_all().await.context("delete_all")
-}
-
 /// It is important that this gets called when DeletionGuard is being held.
 /// For more context see comments in [`DeleteTimelineFlow::prepare`]
 async fn remove_maybe_offloaded_timeline_from_tenant(
@@ -147,9 +141,10 @@ async fn remove_maybe_offloaded_timeline_from_tenant(
             );
         }
         TimelineOrOffloaded::Offloaded(timeline) => {
-            timelines_offloaded
+            let offloaded_timeline = timelines_offloaded
                 .remove(&timeline.timeline_id)
                 .expect("timeline that we were deleting was concurrently removed from 'timelines_offloaded' map");
+            offloaded_timeline.delete_from_ancestor_with_timelines(&timelines);
         }
     }
 
@@ -221,11 +216,24 @@ impl DeleteTimelineFlow {
             None => {
                 let remote_client = tenant
                     .build_timeline_client(timeline.timeline_id(), tenant.remote_storage.clone());
-                let result = remote_client
+                let result = match remote_client
                     .download_index_file(&tenant.cancel)
                     .instrument(info_span!("download_index_file"))
                     .await
-                    .map_err(|e| DeleteTimelineError::Other(anyhow::anyhow!("error: {:?}", e)))?;
+                {
+                    Ok(r) => r,
+                    Err(DownloadError::NotFound) => {
+                        // Deletion is already complete
+                        tracing::info!("Timeline already deleted in remote storage");
+                        return Ok(());
+                    }
+                    Err(e) => {
+                        return Err(DeleteTimelineError::Other(anyhow::anyhow!(
+                            "error: {:?}",
+                            e
+                        )));
+                    }
+                };
                 let index_part = match result {
                     MaybeDeletedIndexPart::Deleted(p) => {
                         tracing::info!("Timeline already set as deleted in remote index");
@@ -275,7 +283,7 @@ impl DeleteTimelineFlow {
 
     /// Shortcut to create Timeline in stopping state and spawn deletion task.
     #[instrument(skip_all, fields(%timeline_id))]
-    pub async fn resume_deletion(
+    pub(crate) async fn resume_deletion(
         tenant: Arc<Tenant>,
         timeline_id: TimelineId,
         local_metadata: &TimelineMetadata,
@@ -406,7 +414,12 @@ impl DeleteTimelineFlow {
             "timeline_delete",
             async move {
                 if let Err(err) = Self::background(guard, conf, &tenant, &timeline, remote_client).await {
-                    error!("Error: {err:#}");
+                    // Only log as an error if it's not a cancellation.
+                    if matches!(err, DeleteTimelineError::Cancelled) {
+                        info!("Shutdown during timeline deletion");
+                    }else {
+                        error!("Error: {err:#}");
+                    }
                     if let TimelineOrOffloaded::Timeline(timeline) = timeline {
                         timeline.set_broken(format!("{err:#}"))
                     }
@@ -438,7 +451,7 @@ impl DeleteTimelineFlow {
             Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm"))?
         });
 
-        delete_remote_layers_and_index(&remote_client).await?;
+        remote_client.delete_all().await?;
 
         pausable_failpoint!("in_progress_delete");
 
@@ -449,10 +462,10 @@ impl DeleteTimelineFlow {
         // So indeed, the tenant manifest might refer to an offloaded timeline which has already been deleted.
         // However, we handle this case in tenant loading code so the next time we attach, the issue is
         // resolved.
-        tenant
-            .store_tenant_manifest()
-            .await
-            .map_err(|e| DeleteTimelineError::Other(anyhow::anyhow!(e)))?;
+        tenant.store_tenant_manifest().await.map_err(|e| match e {
+            TenantManifestError::Cancelled => DeleteTimelineError::Cancelled,
+            _ => DeleteTimelineError::Other(e.into()),
+        })?;
 
         *guard = Self::Finished;
 

pageserver/src/tenant/timeline/offload.rs

@@ -66,7 +66,7 @@ pub(crate) async fn offload_timeline(
     let conf = &tenant.conf;
     delete_local_timeline_directory(conf, tenant.tenant_shard_id, &timeline).await;
 
-    remove_timeline_from_tenant(tenant, &timeline, &guard);
+    let remaining_refcount = remove_timeline_from_tenant(tenant, &timeline, &guard);
 
     {
         let mut offloaded_timelines = tenant.timelines_offloaded.lock().unwrap();
@@ -87,16 +87,20 @@ pub(crate) async fn offload_timeline(
     // not our actual state of offloaded timelines.
     tenant.store_tenant_manifest().await?;
 
+    tracing::info!("Timeline offload complete (remaining arc refcount: {remaining_refcount})");
+
     Ok(())
 }
 
 /// It is important that this gets called when DeletionGuard is being held.
 /// For more context see comments in [`DeleteTimelineFlow::prepare`]
+///
+/// Returns the strong count of the timeline `Arc`
 fn remove_timeline_from_tenant(
     tenant: &Tenant,
     timeline: &Timeline,
     _: &DeletionGuard, // using it as a witness
-) {
+) -> usize {
     // Remove the timeline from the map.
     let mut timelines = tenant.timelines.lock().unwrap();
     let children_exist = timelines
@@ -109,7 +113,9 @@ fn remove_timeline_from_tenant(
         panic!("Timeline grew children while we removed layer files");
     }
 
-    timelines
+    let timeline = timelines
         .remove(&timeline.timeline_id)
         .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
+
+    Arc::strong_count(&timeline)
 }

pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs

@@ -331,11 +331,11 @@ pub(super) async fn handle_walreceiver_connection(
                         Ok(())
                     }
 
-                    while let Some((record_end_lsn, recdata)) = waldecoder.poll_decode()? {
+                    while let Some((next_record_lsn, recdata)) = waldecoder.poll_decode()? {
                         // It is important to deal with the aligned records as lsn in getPage@LSN is
                         // aligned and can be several bytes bigger. Without this alignment we are
                         // at risk of hitting a deadlock.
-                        if !record_end_lsn.is_aligned() {
+                        if !next_record_lsn.is_aligned() {
                             return Err(WalReceiverError::Other(anyhow!("LSN not aligned")));
                         }
 
@@ -343,7 +343,7 @@ pub(super) async fn handle_walreceiver_connection(
                         let interpreted = InterpretedWalRecord::from_bytes_filtered(
                             recdata,
                             modification.tline.get_shard_identity(),
-                            record_end_lsn,
+                            next_record_lsn,
                             modification.tline.pg_version,
                         )?;
 
@@ -367,10 +367,10 @@ pub(super) async fn handle_walreceiver_connection(
                             .ingest_record(interpreted, &mut modification, &ctx)
                             .await
                             .with_context(|| {
-                                format!("could not ingest record at {record_end_lsn}")
+                                format!("could not ingest record at {next_record_lsn}")
                             })?;
                         if !ingested {
-                            tracing::debug!("ingest: filtered out record @ LSN {record_end_lsn}");
+                            tracing::debug!("ingest: filtered out record @ LSN {next_record_lsn}");
                             WAL_INGEST.records_filtered.inc();
                             filtered_records += 1;
                         }
@@ -380,7 +380,7 @@ pub(super) async fn handle_walreceiver_connection(
                         // to timeout the tests.
                         fail_point!("walreceiver-after-ingest");
 
-                        last_rec_lsn = record_end_lsn;
+                        last_rec_lsn = next_record_lsn;
 
                         // Commit every ingest_batch_size records. Even if we filtered out
                         // all records, we still need to call commit to advance the LSN.

pageserver/src/tenant/upload_queue.rs

@@ -88,6 +88,9 @@ pub(crate) struct UploadQueueInitialized {
     #[cfg(feature = "testing")]
     pub(crate) dangling_files: HashMap<LayerName, Generation>,
 
+    /// Deletions that are blocked by the tenant configuration
+    pub(crate) blocked_deletions: Vec<Delete>,
+
     /// Set to true when we have inserted the `UploadOp::Shutdown` into the `inprogress_tasks`.
     pub(crate) shutting_down: bool,
 
@@ -180,6 +183,7 @@ impl UploadQueue {
             queued_operations: VecDeque::new(),
             #[cfg(feature = "testing")]
             dangling_files: HashMap::new(),
+            blocked_deletions: Vec::new(),
             shutting_down: false,
             shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
         };
@@ -220,6 +224,7 @@ impl UploadQueue {
             queued_operations: VecDeque::new(),
             #[cfg(feature = "testing")]
             dangling_files: HashMap::new(),
+            blocked_deletions: Vec::new(),
             shutting_down: false,
             shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
         };
@@ -270,7 +275,7 @@ pub(crate) struct UploadTask {
 
 /// A deletion of some layers within the lifetime of a timeline.  This is not used
 /// for timeline deletion, which skips this queue and goes directly to DeletionQueue.
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub(crate) struct Delete {
     pub(crate) layers: Vec<(LayerName, LayerFileMetadata)>,
 }

pageserver/src/virtual_file.rs

@@ -175,10 +175,16 @@ impl VirtualFile {
     }
 
     pub async fn sync_all(&self) -> Result<(), Error> {
+        if SYNC_MODE.load(std::sync::atomic::Ordering::Relaxed) == SyncMode::UnsafeNoSync as u8 {
+            return Ok(());
+        }
         self.inner.sync_all().await
     }
 
     pub async fn sync_data(&self) -> Result<(), Error> {
+        if SYNC_MODE.load(std::sync::atomic::Ordering::Relaxed) == SyncMode::UnsafeNoSync as u8 {
+            return Ok(());
+        }
         self.inner.sync_data().await
     }
 
@@ -233,6 +239,27 @@ impl VirtualFile {
     }
 }
 
+/// Indicates whether to enable fsync, fdatasync, or O_SYNC/O_DSYNC when writing
+/// files. Switching this off is unsafe and only used for testing on machines
+/// with slow drives.
+#[repr(u8)]
+pub enum SyncMode {
+    Sync,
+    UnsafeNoSync,
+}
+
+impl TryFrom<u8> for SyncMode {
+    type Error = u8;
+
+    fn try_from(value: u8) -> Result<Self, Self::Error> {
+        Ok(match value {
+            v if v == (SyncMode::Sync as u8) => SyncMode::Sync,
+            v if v == (SyncMode::UnsafeNoSync as u8) => SyncMode::UnsafeNoSync,
+            x => return Err(x),
+        })
+    }
+}
+
 ///
 /// A virtual file descriptor. You can use this just like std::fs::File, but internally
 /// the underlying file is closed if the system is low on file descriptors,
@@ -1332,12 +1359,13 @@ impl OpenFiles {
 /// server startup.
 ///
 #[cfg(not(test))]
-pub fn init(num_slots: usize, engine: IoEngineKind, mode: IoMode) {
+pub fn init(num_slots: usize, engine: IoEngineKind, mode: IoMode, sync_mode: SyncMode) {
     if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
         panic!("virtual_file::init called twice");
     }
     set_io_mode(mode);
     io_engine::init(engine);
+    SYNC_MODE.store(sync_mode as u8, std::sync::atomic::Ordering::Relaxed);
     crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
 }
 
@@ -1379,6 +1407,9 @@ pub(crate) fn set_io_mode(mode: IoMode) {
 pub(crate) fn get_io_mode() -> IoMode {
     IoMode::try_from(IO_MODE.load(Ordering::Relaxed)).unwrap()
 }
+
+static SYNC_MODE: AtomicU8 = AtomicU8::new(SyncMode::Sync as u8);
+
 #[cfg(test)]
 mod tests {
     use crate::context::DownloadBehavior;

pageserver/src/walingest.rs

@@ -154,7 +154,7 @@ impl WalIngest {
         WAL_INGEST.records_received.inc();
         let prev_len = modification.len();
 
-        modification.set_lsn(interpreted.end_lsn)?;
+        modification.set_lsn(interpreted.next_record_lsn)?;
 
         if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes) {
             // Records of this type should always be preceded by a commit(), as they
@@ -587,11 +587,29 @@ impl WalIngest {
                 forknum: VISIBILITYMAP_FORKNUM,
             };
 
-            let mut vm_page_no = blkno / pg_constants::VM_HEAPBLOCKS_PER_PAGE;
-            if blkno % pg_constants::VM_HEAPBLOCKS_PER_PAGE != 0 {
-                // Tail of last remaining vm page has to be zeroed.
-                // We are not precise here and instead of digging in VM bitmap format just clear the whole page.
-                modification.put_rel_page_image_zero(rel, vm_page_no)?;
+            // last remaining block, byte, and bit
+            let mut vm_page_no = blkno / (pg_constants::VM_HEAPBLOCKS_PER_PAGE as u32);
+            let trunc_byte = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_PAGE
+                / pg_constants::VM_HEAPBLOCKS_PER_BYTE;
+            let trunc_offs = blkno as usize % pg_constants::VM_HEAPBLOCKS_PER_BYTE
+                * pg_constants::VM_BITS_PER_HEAPBLOCK;
+
+            // Unless the new size is exactly at a visibility map page boundary, the
+            // tail bits in the last remaining map page, representing truncated heap
+            // blocks, need to be cleared. This is not only tidy, but also necessary
+            // because we don't get a chance to clear the bits if the heap is extended
+            // again.
+            if (trunc_byte != 0 || trunc_offs != 0)
+                && self.shard.is_key_local(&rel_block_to_key(rel, vm_page_no))
+            {
+                modification.put_rel_wal_record(
+                    rel,
+                    vm_page_no,
+                    NeonWalRecord::TruncateVisibilityMap {
+                        trunc_byte,
+                        trunc_offs,
+                    },
+                )?;
                 vm_page_no += 1;
             }
             let nblocks = get_relsize(modification, rel, ctx).await?;
@@ -1510,6 +1528,11 @@ mod tests {
 
         assert_current_logical_size(&tline, Lsn(0x50));
 
+        let test_span = tracing::info_span!(parent: None, "test",
+                                            tenant_id=%tline.tenant_shard_id.tenant_id,
+                                            shard_id=%tline.tenant_shard_id.shard_slug(),
+                                            timeline_id=%tline.timeline_id);
+
         // The relation was created at LSN 2, not visible at LSN 1 yet.
         assert_eq!(
             tline
@@ -1544,6 +1567,7 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x20)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 2")
         );
@@ -1551,6 +1575,7 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x30)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
@@ -1558,12 +1583,14 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x40)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x40)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
         );
@@ -1571,18 +1598,21 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x50)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x50)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
         );
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 2 at 5")
         );
@@ -1605,12 +1635,14 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x60)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 0 at 3")
         );
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x60)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1 at 4")
         );
@@ -1625,6 +1657,7 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 2, Version::Lsn(Lsn(0x50)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 2 at 5")
         );
@@ -1657,12 +1690,14 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 0, Version::Lsn(Lsn(0x70)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             ZERO_PAGE
         );
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 1, Version::Lsn(Lsn(0x70)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1")
         );
@@ -1683,6 +1718,7 @@ mod tests {
             assert_eq!(
                 tline
                     .get_rel_page_at_lsn(TESTREL_A, blk, Version::Lsn(Lsn(0x80)), &ctx)
+                    .instrument(test_span.clone())
                     .await?,
                 ZERO_PAGE
             );
@@ -1690,6 +1726,7 @@ mod tests {
         assert_eq!(
             tline
                 .get_rel_page_at_lsn(TESTREL_A, 1500, Version::Lsn(Lsn(0x80)), &ctx)
+                .instrument(test_span.clone())
                 .await?,
             test_img("foo blk 1500")
         );
@@ -1797,6 +1834,11 @@ mod tests {
         }
         m.commit(&ctx).await?;
 
+        let test_span = tracing::info_span!(parent: None, "test",
+                                            tenant_id=%tline.tenant_shard_id.tenant_id,
+                                            shard_id=%tline.tenant_shard_id.shard_slug(),
+                                            timeline_id=%tline.timeline_id);
+
         // The relation was created at LSN 20, not visible at LSN 1 yet.
         assert_eq!(
             tline
@@ -1829,6 +1871,7 @@ mod tests {
             assert_eq!(
                 tline
                     .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(lsn), &ctx)
+                    .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
             );
@@ -1856,6 +1899,7 @@ mod tests {
             assert_eq!(
                 tline
                     .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x60)), &ctx)
+                    .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
             );
@@ -1874,6 +1918,7 @@ mod tests {
             assert_eq!(
                 tline
                     .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x50)), &ctx)
+                    .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
             );
@@ -1910,6 +1955,7 @@ mod tests {
             assert_eq!(
                 tline
                     .get_rel_page_at_lsn(TESTREL_A, blkno, Version::Lsn(Lsn(0x80)), &ctx)
+                    .instrument(test_span.clone())
                     .await?,
                 test_img(&data)
             );

pageserver/src/walredo/apply_neon.rs

@@ -42,6 +42,34 @@ pub(crate) fn apply_in_neon(
         } => {
             anyhow::bail!("tried to pass postgres wal record to neon WAL redo");
         }
+        //
+        // Code copied from PostgreSQL `visibilitymap_prepare_truncate` function in `visibilitymap.c`
+        //
+        NeonWalRecord::TruncateVisibilityMap {
+            trunc_byte,
+            trunc_offs,
+        } => {
+            // sanity check that this is modifying the correct relation
+            let (rel, _) = key.to_rel_block().context("invalid record")?;
+            assert!(
+                rel.forknum == VISIBILITYMAP_FORKNUM,
+                "TruncateVisibilityMap record on unexpected rel {}",
+                rel
+            );
+            let map = &mut page[pg_constants::MAXALIGN_SIZE_OF_PAGE_HEADER_DATA..];
+            map[*trunc_byte + 1..].fill(0u8);
+            /*----
+             * Mask out the unwanted bits of the last remaining byte.
+             *
+             * ((1 << 0) - 1) = 00000000
+             * ((1 << 1) - 1) = 00000001
+             * ...
+             * ((1 << 6) - 1) = 00111111
+             * ((1 << 7) - 1) = 01111111
+             *----
+             */
+            map[*trunc_byte] &= (1 << *trunc_offs) - 1;
+        }
         NeonWalRecord::ClearVisibilityMapFlags {
             new_heap_blkno,
             old_heap_blkno,
@@ -253,6 +281,10 @@ 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();