tests: extend test_sharding for splitting

refer #5508

.config/nextest.toml

@@ -0,0 +1,2 @@
+[profile.default]
+slow-timeout = "1m"

.github/workflows/benchmarking.yml

@@ -11,7 +11,7 @@ on:
     #          │ │ ┌───────────── day of the month (1 - 31)
     #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
     #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
-    - cron:  '0 3 * * *' # run once a day, timezone is utc
+    - cron:   '0 3 * * *' # run once a day, timezone is utc
 
   workflow_dispatch: # adds ability to run this manually
     inputs:
@@ -23,6 +23,21 @@ on:
         type: boolean
         description: 'Publish perf report. If not set, the report will be published only for the main branch'
         required: false
+      collect_olap_explain:
+        type: boolean
+        description: 'Collect EXPLAIN ANALYZE for OLAP queries. If not set, EXPLAIN ANALYZE will not be collected'
+        required: false
+        default: false
+      collect_pg_stat_statements:
+        type: boolean
+        description: 'Collect pg_stat_statements for OLAP queries. If not set, pg_stat_statements will not be collected'
+        required: false
+        default: false
+      run_AWS_RDS_AND_AURORA:
+        type: boolean
+        description: 'AWS-RDS and AWS-AURORA normally only run on Saturday. Set this to true to run them on every workflow_dispatch'
+        required: false
+        default: false
 
 defaults:
   run:
@@ -113,6 +128,8 @@ jobs:
     # - neon-captest-reuse: Reusing existing project
     # - rds-aurora: Aurora Postgres Serverless v2 with autoscaling from 0.5 to 2 ACUs
     # - rds-postgres: RDS Postgres db.m5.large instance (2 vCPU, 8 GiB) with gp3 EBS storage
+    env:
+      RUN_AWS_RDS_AND_AURORA: ${{ github.event.inputs.run_AWS_RDS_AND_AURORA || 'false' }}
     runs-on: ubuntu-latest
     outputs:
       pgbench-compare-matrix: ${{ steps.pgbench-compare-matrix.outputs.matrix }}
@@ -152,7 +169,7 @@ jobs:
           ]
         }'
 
-        if [ "$(date +%A)" = "Saturday" ]; then
+        if [ "$(date +%A)" = "Saturday" ] || [ ${RUN_AWS_RDS_AND_AURORA} = "true" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres" },
                                                    { "platform": "rds-aurora"   }]')
         fi
@@ -171,9 +188,9 @@ jobs:
           ]
         }'
 
-        if [ "$(date +%A)" = "Saturday" ]; then
+        if [ "$(date +%A)" = "Saturday" ] || [ ${RUN_AWS_RDS_AND_AURORA} = "true" ]; then
           matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres", "scale": "10" },
-                                                   { "platform": "rds-aurora",   "scale": "10" }]')
+                                                    { "platform": "rds-aurora",   "scale": "10" }]')
         fi
 
         echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT
@@ -337,6 +354,8 @@ jobs:
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
       DEFAULT_PG_VERSION: 14
       TEST_OUTPUT: /tmp/test_output
+      TEST_OLAP_COLLECT_EXPLAIN: ${{ github.event.inputs.collect_olap_explain }}
+      TEST_OLAP_COLLECT_PG_STAT_STATEMENTS: ${{ github.event.inputs.collect_pg_stat_statements }}
       BUILD_TYPE: remote
       SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
       PLATFORM: ${{ matrix.platform }}
@@ -399,6 +418,8 @@ jobs:
       env:
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        TEST_OLAP_COLLECT_EXPLAIN: ${{ github.event.inputs.collect_olap_explain || 'false' }}
+        TEST_OLAP_COLLECT_PG_STAT_STATEMENTS: ${{ github.event.inputs.collect_pg_stat_statements || 'false' }}
         BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
         TEST_OLAP_SCALE: 10
 

.github/workflows/build_and_push_docker_image.yml

@@ -0,0 +1,105 @@
+name: Build and Push Docker Image
+
+on:
+  workflow_call:
+    inputs:
+      dockerfile-path:
+        required: true
+        type: string
+      image-name:
+        required: true
+        type: string
+    outputs:
+      build-tools-tag:
+        description: "tag generated for build tools"
+        value: ${{ jobs.tag.outputs.build-tools-tag }}
+
+jobs:
+  check-if-build-tools-dockerfile-changed:
+    runs-on: ubuntu-latest
+    outputs:
+      docker_file_changed: ${{ steps.dockerfile.outputs.docker_file_changed }}
+    steps:
+      - name: Check if Dockerfile.buildtools has changed
+        id: dockerfile
+        run: |
+          if [[ "$GITHUB_EVENT_NAME" != "pull_request" ]]; then
+            echo "docker_file_changed=false" >> $GITHUB_OUTPUT
+            exit
+          fi
+          updated_files=$(gh pr --repo neondatabase/neon diff ${{ github.event.pull_request.number }} --name-only)
+          if [[ $updated_files == *"Dockerfile.buildtools"* ]]; then
+            echo "docker_file_changed=true" >> $GITHUB_OUTPUT
+          fi
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  tag:
+    runs-on: ubuntu-latest
+    needs: [ check-if-build-tools-dockerfile-changed ]
+    outputs:
+      build-tools-tag: ${{steps.buildtools-tag.outputs.image_tag}}
+
+    steps:
+      - name: Get buildtools tag
+        env:
+          DOCKERFILE_CHANGED: ${{ needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed }}
+        run: |
+          if [[ "$GITHUB_EVENT_NAME" == "pull_request" ]] && [[ "${DOCKERFILE_CHANGED}" == "true" ]]; then
+            IMAGE_TAG=$GITHUB_RUN_ID
+          else
+            IMAGE_TAG=pinned
+          fi
+
+          echo "image_tag=${IMAGE_TAG}" >> $GITHUB_OUTPUT
+        shell: bash
+        id: buildtools-tag
+
+  kaniko:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    needs: [ tag, check-if-build-tools-dockerfile-changed ]
+    runs-on: [ self-hosted, dev, x64 ]
+    container: gcr.io/kaniko-project/executor:v1.7.0-debug
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v1
+
+      - name: Configure ECR login
+        run: echo "{\"credsStore\":\"ecr-login\"}" > /kaniko/.docker/config.json
+
+      - name: Kaniko build
+        run: /kaniko/executor --reproducible --snapshotMode=redo --skip-unused-stages --dockerfile ${{ inputs.dockerfile-path }} --cache=true --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache  --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-amd64
+
+  kaniko-arm:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    needs: [ tag, check-if-build-tools-dockerfile-changed ]
+    runs-on: [ self-hosted, dev, arm64 ]
+    container: gcr.io/kaniko-project/executor:v1.7.0-debug
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v1
+
+      - name: Configure ECR login
+        run: echo "{\"credsStore\":\"ecr-login\"}" > /kaniko/.docker/config.json
+
+      - name: Kaniko build
+        run: /kaniko/executor --reproducible --snapshotMode=redo --skip-unused-stages --dockerfile ${{ inputs.dockerfile-path }} --cache=true --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-arm64
+
+  manifest:
+    if: needs.check-if-build-tools-dockerfile-changed.outputs.docker_file_changed == 'true'
+    name: 'manifest'
+    runs-on: [ self-hosted, dev, x64 ]
+    needs:
+      - tag
+      - kaniko
+      - kaniko-arm
+      - check-if-build-tools-dockerfile-changed
+
+    steps:
+      - name: Create manifest
+        run: docker manifest create 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }} --amend 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-amd64 --amend 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}-arm64
+
+      - name: Push manifest
+        run: docker manifest push 369495373322.dkr.ecr.eu-central-1.amazonaws.com/${{ inputs.image-name }}:${{ needs.tag.outputs.build-tools-tag }}

.github/workflows/build_and_test.yml

@@ -44,7 +44,6 @@ jobs:
 
         exit 1
 
-
   tag:
     needs: [ check-permissions ]
     runs-on: [ self-hosted, gen3, small ]
@@ -74,11 +73,19 @@ jobs:
         shell: bash
         id: build-tag
 
-  check-codestyle-python:
+  build-buildtools-image:
     needs: [ check-permissions ]
+    uses: ./.github/workflows/build_and_push_docker_image.yml
+    with:
+      dockerfile-path: Dockerfile.buildtools
+      image-name: build-tools
+    secrets: inherit
+
+  check-codestyle-python:
+    needs: [ check-permissions, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -108,10 +115,10 @@ jobs:
         run: poetry run mypy .
 
   check-codestyle-rust:
-    needs: [ check-permissions ]
+    needs: [ check-permissions, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -175,10 +182,10 @@ jobs:
         run: cargo deny check --hide-inclusion-graph
 
   build-neon:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, tag, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
     strategy:
       fail-fast: false
@@ -332,16 +339,16 @@ jobs:
         run: |
           ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
 
-      - name: Run cargo test
+      - name: Run rust tests
         run: |
-          ${cov_prefix} cargo test $CARGO_FLAGS $CARGO_FEATURES
+          ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES
 
           # Run separate tests for real S3
           export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
-          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
+          export REMOTE_STORAGE_S3_BUCKET=neon-github-ci-tests
           export REMOTE_STORAGE_S3_REGION=eu-central-1
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
+          ${cov_prefix} cargo nextest run $CARGO_FLAGS -E 'package(remote_storage)' -E 'test(test_real_s3)'
 
           # Run separate tests for real Azure Blob Storage
           # XXX: replace region with `eu-central-1`-like region
@@ -351,7 +358,7 @@ jobs:
           export REMOTE_STORAGE_AZURE_CONTAINER="${{ vars.REMOTE_STORAGE_AZURE_CONTAINER }}"
           export REMOTE_STORAGE_AZURE_REGION="${{ vars.REMOTE_STORAGE_AZURE_REGION }}"
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_azure
+          ${cov_prefix} cargo nextest run $CARGO_FLAGS -E 'package(remote_storage)' -E 'test(test_real_azure)'
 
       - name: Install rust binaries
         run: |
@@ -408,10 +415,10 @@ jobs:
         uses: ./.github/actions/save-coverage-data
 
   regress-tests:
-    needs: [ check-permissions, build-neon, tag ]
+    needs: [ check-permissions, build-neon, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       # Default shared memory is 64mb
       options: --init --shm-size=512mb
     strategy:
@@ -447,10 +454,10 @@ jobs:
         uses: ./.github/actions/save-coverage-data
 
   benchmarks:
-    needs: [ check-permissions, build-neon ]
+    needs: [ check-permissions, build-neon, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       # Default shared memory is 64mb
       options: --init --shm-size=512mb
     if: github.ref_name == 'main' || contains(github.event.pull_request.labels.*.name, 'run-benchmarks')
@@ -479,12 +486,12 @@ jobs:
       # while coverage is currently collected for the debug ones
 
   create-test-report:
-    needs: [ check-permissions, regress-tests, coverage-report, benchmarks ]
+    needs: [ check-permissions, regress-tests, coverage-report, benchmarks, build-buildtools-image ]
     if: ${{ !cancelled() && contains(fromJSON('["skipped", "success"]'), needs.check-permissions.result) }}
 
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
 
     steps:
@@ -526,11 +533,10 @@ jobs:
             })
 
   coverage-report:
-    needs: [ check-permissions, regress-tests ]
-
+    needs: [ check-permissions, regress-tests, build-buildtools-image ]
     runs-on: [ self-hosted, gen3, small ]
     container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools:${{ needs.build-buildtools-image.outputs.build-tools-tag }}
       options: --init
     strategy:
       fail-fast: false
@@ -694,7 +700,7 @@ jobs:
             }"
 
   neon-image:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container: gcr.io/kaniko-project/executor:v1.9.2-debug
     defaults:
@@ -733,6 +739,7 @@ jobs:
                            --context .
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg BUILD_TAG=${{ needs.tag.outputs.build-tag }}
+                           --build-arg TAG=${{ needs.build-buildtools-image.outputs.build-tools-tag }}
                            --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                            --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}}
                            --destination neondatabase/neon:${{needs.tag.outputs.build-tag}}
@@ -743,7 +750,7 @@ jobs:
 
   compute-tools-image:
     runs-on: [ self-hosted, gen3, large ]
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     container: gcr.io/kaniko-project/executor:v1.9.2-debug
     defaults:
       run:
@@ -778,6 +785,7 @@ jobs:
                            --context .
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
+                           --build-arg TAG=${{needs.build-buildtools-image.outputs.build-tools-tag}}
                            --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                            --dockerfile Dockerfile.compute-tools
                            --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}}
@@ -788,7 +796,7 @@ jobs:
         run: rm -rf ~/.ecr
 
   compute-node-image:
-    needs: [ check-permissions, tag ]
+    needs: [ check-permissions, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container:
       image: gcr.io/kaniko-project/executor:v1.9.2-debug
@@ -836,6 +844,7 @@ jobs:
                            --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                            --build-arg PG_VERSION=${{ matrix.version }}
                            --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
+                           --build-arg TAG=${{needs.build-buildtools-image.outputs.build-tools-tag}}
                            --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                            --dockerfile Dockerfile.compute-node
                            --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
@@ -857,7 +866,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.19.0
+      VM_BUILDER_VERSION: v0.21.0
 
     steps:
       - name: Checkout

.github/workflows/neon_extra_builds.yml

@@ -218,7 +218,7 @@ jobs:
 
           # Run separate tests for real S3
           export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
-          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
+          export REMOTE_STORAGE_S3_BUCKET=neon-github-ci-tests
           export REMOTE_STORAGE_S3_REGION=eu-central-1
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3

.github/workflows/update_build_tools_image.yml

@@ -0,0 +1,130 @@
+name: 'Update build tools image tag'
+
+# This workflow it used to update tag of build tools in ECR.
+# The most common use case is adding/moving `pinned` tag to `${GITHUB_RUN_IT}` image.
+
+on:
+  workflow_dispatch:
+    inputs:
+      from-tag:
+        description: 'Source tag'
+        required: true
+        type: string
+      to-tag:
+        description: 'Destination tag'
+        required: true
+        type: string
+        default: 'pinned'
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+env:
+  AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+  AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+permissions: {}
+
+jobs:
+  tag-image:
+    runs-on: [ self-hosted, gen3, small ]
+    container: golang:1.19-bullseye
+
+    env:
+      IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools
+      FROM_TAG: ${{ inputs.from-tag }}
+      TO_TAG: ${{ inputs.to-tag }}
+    outputs:
+      next-digest-buildtools: ${{ steps.next-digest.outputs.next-digest-buildtools }}
+      prev-digest-buildtools: ${{ steps.prev-digest.outputs.prev-digest-buildtools }}
+
+    steps:
+      - name: Install Crane & ECR helper
+        run: |
+          go install github.com/google/go-containerregistry/cmd/crane@a54d64203cffcbf94146e04069aae4a97f228ee2 # v0.16.1
+          go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@adf1bafd791ae7d4ff098108b1e91f36a4da5404 # v0.7.1
+
+      - name: Configure ECR login
+        run: |
+          mkdir /github/home/.docker/
+          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
+
+      - name: Get source image digest
+        id: next-digest
+        run: |
+          NEXT_DIGEST=$(crane digest ${IMAGE}:${FROM_TAG} || true)
+          if [ -z "${NEXT_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${FROM_TAG} does not exist"
+            exit 1
+          fi
+
+          echo "Current ${IMAGE}@${FROM_TAG} image is ${IMAGE}@${NEXT_DIGEST}"
+          echo "next-digest-buildtools=$NEXT_DIGEST" >> $GITHUB_OUTPUT
+
+      - name: Get destination image digest (if already exists)
+        id: prev-digest
+        run: |
+          PREV_DIGEST=$(crane digest ${IMAGE}:${TO_TAG} || true)
+          if [ -z "${PREV_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${TO_TAG} does not exist (it's ok)"
+          else
+            echo >&2 "Current ${IMAGE}@${TO_TAG} image is ${IMAGE}@${PREV_DIGEST}"
+
+            echo "prev-digest-buildtools=$PREV_DIGEST" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Tag image
+        run: |
+          crane tag "${IMAGE}:${FROM_TAG}" "${TO_TAG}"
+
+  rollback-tag-image:
+    needs:  tag-image
+    if: ${{ !success() }}
+
+    runs-on: [ self-hosted, gen3, small ]
+    container: golang:1.19-bullseye
+
+    env:
+      IMAGE: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/build-tools
+      FROM_TAG: ${{ inputs.from-tag }}
+      TO_TAG: ${{ inputs.to-tag }}
+
+    steps:
+      - name: Install Crane & ECR helper
+        run: |
+          go install github.com/google/go-containerregistry/cmd/crane@a54d64203cffcbf94146e04069aae4a97f228ee2 # v0.16.1
+          go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@adf1bafd791ae7d4ff098108b1e91f36a4da5404 # v0.7.1
+
+      - name: Configure ECR login
+        run: |
+          mkdir /github/home/.docker/
+          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
+
+      - name: Restore previous tag if needed
+        run: |
+          NEXT_DIGEST="${{ needs.tag-image.outputs.next-digest-buildtools }}"
+          PREV_DIGEST="${{ needs.tag-image.outputs.prev-digest-buildtools }}"
+
+          if [ -z "${NEXT_DIGEST}" ]; then
+            echo >&2 "Image ${IMAGE}:${FROM_TAG} does not exist, nothing to rollback"
+            exit 0
+          fi
+
+          if [ -z "${PREV_DIGEST}" ]; then
+            # I guess we should delete the tag here/untag the image, but crane does not support it
+            # - https://github.com/google/go-containerregistry/issues/999
+
+            echo >&2 "Image ${IMAGE}:${TO_TAG} did not exist, but it was created by the job, no need to rollback"
+
+            exit 0
+          fi
+
+          CURRENT_DIGEST=$(crane digest "${IMAGE}:${TO_TAG}")
+          if [ "${CURRENT_DIGEST}" == "${NEXT_DIGEST}" ]; then
+            crane tag "${IMAGE}@${PREV_DIGEST}" "${TO_TAG}"
+
+            echo >&2 "Successfully restored ${TO_TAG} tag from ${IMAGE}@${CURRENT_DIGEST} to ${IMAGE}@${PREV_DIGEST}"
+          else
+            echo >&2 "Image ${IMAGE}:${TO_TAG}@${CURRENT_DIGEST} is not required to be restored"
+          fi

.gitignore

@@ -6,6 +6,7 @@ __pycache__/
 test_output/
 .vscode
 .idea
+neon.iml
 /.neon
 /integration_tests/.neon
 

CONTRIBUTING.md

@@ -70,3 +70,17 @@ We're using the following approach to make it work:
 - The label gets removed automatically, so to run CI again with new changes, the label should be added again (after the review)
 
 For details see [`approved-for-ci-run.yml`](.github/workflows/approved-for-ci-run.yml)
+
+## How do I add the "pinned" tag to an buildtools image?
+We use the `pinned` tag for `Dockerfile.buildtools` build images in our CI/CD setup, currently adding the `pinned` tag is a manual operation.
+
+You can call it from GitHub UI: https://github.com/neondatabase/neon/actions/workflows/update_build_tools_image.yml,
+or using GitHub CLI:
+
+```bash
+gh workflow -R neondatabase/neon run update_build_tools_image.yml \
+            -f from-tag=6254913013 \
+            -f to-tag=pinned \
+
+# Default `-f to-tag` is `pinned`, so the parameter can be omitted.
+```

Cargo.lock

@@ -190,9 +190,9 @@ dependencies = [
 
 [[package]]
 name = "async-compression"
-version = "0.4.0"
+version = "0.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5b0122885821398cc923ece939e24d1056a2384ee719432397fa9db87230ff11"
+checksum = "bc2d0cfb2a7388d34f590e76686704c494ed7aaceed62ee1ba35cbf363abc2a5"
 dependencies = [
  "flate2",
  "futures-core",
@@ -256,6 +256,28 @@ dependencies = [
  "critical-section",
 ]
 
+[[package]]
+name = "attachment_service"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "control_plane",
+ "futures",
+ "hyper",
+ "pageserver_api",
+ "postgres_connection",
+ "reqwest",
+ "serde",
+ "serde_json",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "autocfg"
 version = "1.1.0"
@@ -1161,6 +1183,7 @@ dependencies = [
  "flate2",
  "futures",
  "hyper",
+ "nix 0.26.2",
  "notify",
  "num_cpus",
  "opentelemetry",
@@ -1168,8 +1191,10 @@ dependencies = [
  "regex",
  "remote_storage",
  "reqwest",
+ "rust-ini",
  "serde",
  "serde_json",
+ "signal-hook",
  "tar",
  "tokio",
  "tokio-postgres",
@@ -1201,6 +1226,26 @@ version = "0.9.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "28c122c3980598d243d63d9a704629a2d748d101f278052ff068be5a4423ab6f"
 
+[[package]]
+name = "const-random"
+version = "0.1.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5aaf16c9c2c612020bcfd042e170f6e32de9b9d75adb5277cdbbd2e2c8c8299a"
+dependencies = [
+ "const-random-macro",
+]
+
+[[package]]
+name = "const-random-macro"
+version = "0.1.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f9d839f2a20b0aee515dc581a6172f2321f96cab76c1a38a4c584a194955390e"
+dependencies = [
+ "getrandom 0.2.11",
+ "once_cell",
+ "tiny-keccak",
+]
+
 [[package]]
 name = "const_fn"
 version = "0.4.9"
@@ -1433,6 +1478,12 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "crunchy"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7a81dae078cea95a014a339291cec439d2f232ebe854a9d672b796c6afafa9b7"
+
 [[package]]
 name = "crypto-bigint"
 version = "0.4.9"
@@ -1575,6 +1626,15 @@ dependencies = [
  "syn 2.0.32",
 ]
 
+[[package]]
+name = "dlv-list"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "442039f5147480ba31067cb00ada1adae6892028e40e45fc5de7b7df6dcc1b5f"
+dependencies = [
+ "const-random",
+]
+
 [[package]]
 name = "dyn-clone"
 version = "1.0.14"
@@ -2106,6 +2166,20 @@ dependencies = [
  "hashbrown 0.13.2",
 ]
 
+[[package]]
+name = "hdrhistogram"
+version = "7.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "765c9198f173dd59ce26ff9f95ef0aafd0a0fe01fb9d72841bc5066a4c06511d"
+dependencies = [
+ "base64 0.21.1",
+ "byteorder",
+ "crossbeam-channel",
+ "flate2",
+ "nom",
+ "num-traits",
+]
+
 [[package]]
 name = "heapless"
 version = "0.8.0"
@@ -2487,13 +2561,14 @@ dependencies = [
 
 [[package]]
 name = "jsonwebtoken"
-version = "8.3.0"
+version = "9.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6971da4d9c3aa03c3d8f3ff0f4155b534aad021292003895a469716b2a230378"
+checksum = "5c7ea04a7c5c055c175f189b6dc6ba036fd62306b58c66c9f6389036c503a3f4"
 dependencies = [
  "base64 0.21.1",
- "pem 1.1.1",
- "ring 0.16.20",
+ "js-sys",
+ "pem 3.0.3",
+ "ring 0.17.6",
  "serde",
  "serde_json",
  "simple_asn1",
@@ -3028,6 +3103,16 @@ dependencies = [
  "tokio-stream",
 ]
 
+[[package]]
+name = "ordered-multimap"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a4d6a8c22fc714f0c2373e6091bf6f5e9b37b1bc0b1184874b7e0a4e303d318f"
+dependencies = [
+ "dlv-list",
+ "hashbrown 0.14.0",
+]
+
 [[package]]
 name = "os_info"
 version = "3.7.0"
@@ -3056,6 +3141,28 @@ dependencies = [
  "sha2",
 ]
 
+[[package]]
+name = "pagebench"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "clap",
+ "futures",
+ "hdrhistogram",
+ "humantime",
+ "humantime-serde",
+ "pageserver",
+ "pageserver_api",
+ "pageserver_client",
+ "rand 0.8.5",
+ "serde",
+ "serde_json",
+ "tokio",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "pagectl"
 version = "0.1.0"
@@ -3145,6 +3252,7 @@ dependencies = [
  "tokio",
  "tokio-io-timeout",
  "tokio-postgres",
+ "tokio-stream",
  "tokio-tar",
  "tokio-util",
  "toml_edit",
@@ -3182,11 +3290,19 @@ dependencies = [
 name = "pageserver_client"
 version = "0.1.0"
 dependencies = [
+ "anyhow",
  "async-trait",
+ "bytes",
+ "futures",
  "pageserver_api",
+ "postgres",
  "reqwest",
  "serde",
  "thiserror",
+ "tokio",
+ "tokio-postgres",
+ "tokio-stream",
+ "tokio-util",
  "utils",
  "workspace_hack",
 ]
@@ -3282,18 +3398,19 @@ checksum = "19b17cddbe7ec3f8bc800887bab5e717348c95ea2ca0b1bf0837fb964dc67099"
 
 [[package]]
 name = "pem"
-version = "1.1.1"
+version = "2.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a8835c273a76a90455d7344889b0964598e3316e2a79ede8e36f16bdcf2228b8"
+checksum = "6b13fe415cdf3c8e44518e18a7c95a13431d9bdf6d15367d82b23c377fdd441a"
 dependencies = [
- "base64 0.13.1",
+ "base64 0.21.1",
+ "serde",
 ]
 
 [[package]]
 name = "pem"
-version = "2.0.1"
+version = "3.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6b13fe415cdf3c8e44518e18a7c95a13431d9bdf6d15367d82b23c377fdd441a"
+checksum = "1b8fcc794035347fb64beda2d3b462595dd2753e3f268d89c5aae77e8cf2c310"
 dependencies = [
  "base64 0.21.1",
  "serde",
@@ -4169,6 +4286,16 @@ dependencies = [
  "unicode-ident",
 ]
 
+[[package]]
+name = "rust-ini"
+version = "0.20.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3e0698206bcb8882bf2a9ecb4c1e7785db57ff052297085a6efd4fe42302068a"
+dependencies = [
+ "cfg-if",
+ "ordered-multimap",
+]
+
 [[package]]
 name = "rustc-demangle"
 version = "0.1.23"
@@ -4300,12 +4427,14 @@ dependencies = [
  "async-stream",
  "aws-config",
  "aws-sdk-s3",
+ "aws-smithy-async",
  "bincode",
  "bytes",
  "chrono",
  "clap",
  "crc32c",
  "either",
+ "futures",
  "futures-util",
  "hex",
  "histogram",
@@ -4344,6 +4473,7 @@ dependencies = [
  "clap",
  "const_format",
  "crc32c",
+ "fail",
  "fs2",
  "futures",
  "git-version",
@@ -4419,12 +4549,12 @@ checksum = "d29ab0c6d3fc0ee92fe66e2d99f700eab17a8d57d1c1d3b748380fb20baa78cd"
 
 [[package]]
 name = "sct"
-version = "0.7.0"
+version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d53dcdb7c9f8158937a7981b48accfd39a43af418591a5d008c7b22b5e1b7ca4"
+checksum = "da046153aa2352493d6cb7da4b6e5c0c057d8a1d0a9aa8560baffdd945acd414"
 dependencies = [
- "ring 0.16.20",
- "untrusted 0.7.1",
+ "ring 0.17.6",
+ "untrusted 0.9.0",
 ]
 
 [[package]]
@@ -5123,6 +5253,15 @@ dependencies = [
  "time-core",
 ]
 
+[[package]]
+name = "tiny-keccak"
+version = "2.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2c9d3793400a45f954c52e73d068316d76b6f4e36977e3fcebb13a2721e80237"
+dependencies = [
+ "crunchy",
+]
+
 [[package]]
 name = "tinytemplate"
 version = "1.2.1"
@@ -5766,6 +5905,7 @@ dependencies = [
  "chrono",
  "const_format",
  "criterion",
+ "fail",
  "futures",
  "heapless",
  "hex",
@@ -6290,6 +6430,7 @@ dependencies = [
  "futures-io",
  "futures-sink",
  "futures-util",
+ "getrandom 0.2.11",
  "hex",
  "hmac",
  "hyper",
@@ -6301,6 +6442,7 @@ dependencies = [
  "num-bigint",
  "num-integer",
  "num-traits",
+ "once_cell",
  "prost",
  "rand 0.8.5",
  "regex",
@@ -6403,30 +6545,28 @@ checksum = "2a0956f1ba7c7909bfb66c2e9e4124ab6f6482560f6628b5aaeba39207c9aad9"
 
 [[package]]
 name = "zstd"
-version = "0.12.4"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a27595e173641171fc74a1232b7b1c7a7cb6e18222c11e9dfb9888fa424c53c"
+checksum = "bffb3309596d527cfcba7dfc6ed6052f1d39dfbd7c867aa2e865e4a449c10110"
 dependencies = [
  "zstd-safe",
 ]
 
 [[package]]
 name = "zstd-safe"
-version = "6.0.6"
+version = "7.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ee98ffd0b48ee95e6c5168188e44a54550b1564d9d530ee21d5f0eaed1069581"
+checksum = "43747c7422e2924c11144d5229878b98180ef8b06cca4ab5af37afc8a8d8ea3e"
 dependencies = [
- "libc",
  "zstd-sys",
 ]
 
 [[package]]
 name = "zstd-sys"
-version = "2.0.8+zstd.1.5.5"
+version = "2.0.9+zstd.1.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5556e6ee25d32df2586c098bbfa278803692a20d0ab9565e049480d52707ec8c"
+checksum = "9e16efa8a874a0481a574084d34cc26fdb3b99627480f785888deb6386506656"
 dependencies = [
  "cc",
- "libc",
  "pkg-config",
 ]

Cargo.toml

@@ -3,9 +3,11 @@ resolver = "2"
 members = [
     "compute_tools",
     "control_plane",
+    "control_plane/attachment_service",
     "pageserver",
     "pageserver/ctl",
     "pageserver/client",
+    "pageserver/pagebench",
     "proxy",
     "safekeeper",
     "storage_broker",
@@ -79,6 +81,7 @@ futures-util = "0.3"
 git-version = "0.3"
 hashbrown = "0.13"
 hashlink = "0.8.1"
+hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
@@ -91,7 +94,7 @@ hyper-tungstenite = "0.11"
 inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
-jsonwebtoken = "8"
+jsonwebtoken = "9"
 libc = "0.2"
 md5 = "0.7.0"
 memoffset = "0.8"

Dockerfile

@@ -3,7 +3,7 @@
 ### By default, the binaries inside the image have some mock parameters and can start, but are not intended to be used
 ### inside this image in the real deployments.
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 
 # Build Postgres

Dockerfile.buildtools

@@ -0,0 +1,166 @@
+FROM debian:bullseye-slim
+
+# Add nonroot user
+RUN useradd -ms /bin/bash nonroot -b /home
+SHELL ["/bin/bash", "-c"]
+
+# System deps
+RUN set -e \
+    && apt update \
+    && apt install -y \
+        autoconf \
+        automake \
+        bison \
+        build-essential \
+        ca-certificates \
+        cmake \
+        curl \
+        flex \
+        git \
+        gnupg \
+        gzip \
+        jq \
+        libcurl4-openssl-dev \
+        libbz2-dev \
+        libffi-dev \
+        liblzma-dev \
+        libncurses5-dev \
+        libncursesw5-dev \
+        libpq-dev \
+        libreadline-dev \
+        libseccomp-dev \
+        libsqlite3-dev \
+        libssl-dev \
+        libstdc++-10-dev \
+        libtool \
+        libxml2-dev \
+        libxmlsec1-dev \
+        libxxhash-dev \
+        lsof \
+        make \
+        netcat \
+        net-tools \
+        openssh-client \
+        parallel \
+        pkg-config \
+        unzip \
+        wget \
+        xz-utils \
+        zlib1g-dev \
+        zstd \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# protobuf-compiler (protoc)
+ENV PROTOC_VERSION 25.1
+RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
+    && unzip -q protoc.zip -d protoc \
+    && mv protoc/bin/protoc /usr/local/bin/protoc \
+    && mv protoc/include/google /usr/local/include/google \
+    && rm -rf protoc.zip protoc
+
+# LLVM
+ENV LLVM_VERSION=17
+RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
+    && echo "deb http://apt.llvm.org/bullseye/ llvm-toolchain-bullseye-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
+    && apt update \
+    && apt install -y clang-${LLVM_VERSION} llvm-${LLVM_VERSION} \
+    && bash -c 'for f in /usr/bin/clang*-${LLVM_VERSION} /usr/bin/llvm*-${LLVM_VERSION}; do ln -s "${f}" "${f%-${LLVM_VERSION}}"; done' \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# PostgreSQL 14
+RUN curl -fsSL 'https://www.postgresql.org/media/keys/ACCC4CF8.asc' | apt-key add - \
+    && echo 'deb http://apt.postgresql.org/pub/repos/apt bullseye-pgdg main' > /etc/apt/sources.list.d/pgdg.list \
+    && apt update \
+    && apt install -y postgresql-client-14 \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
+# AWS CLI
+RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-$(uname -m).zip" -o "awscliv2.zip" \
+    && unzip -q awscliv2.zip \
+    && ./aws/install \
+    && rm awscliv2.zip
+
+# Mold: A Modern Linker
+ENV MOLD_VERSION v2.4.0
+RUN set -e \
+    && git clone https://github.com/rui314/mold.git \
+    && mkdir mold/build \
+    && cd mold/build \
+    && git checkout ${MOLD_VERSION} \
+    && cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=clang++ .. \
+    && cmake --build . -j $(nproc) \
+    && cmake --install . \
+    && cd .. \
+    && rm -rf mold
+
+# LCOV
+# Build lcov from a fork:
+# It includes several bug fixes on top on v2.0 release (https://github.com/linux-test-project/lcov/compare/v2.0...master)
+# And patches from us:
+# - Generates json file with code coverage summary (https://github.com/neondatabase/lcov/commit/426e7e7a22f669da54278e9b55e6d8caabd00af0.tar.gz)
+RUN for package in Capture::Tiny DateTime Devel::Cover Digest::MD5 File::Spec JSON::XS Memory::Process Time::HiRes JSON; do yes | perl -MCPAN -e "CPAN::Shell->notest('install', '$package')"; done \
+    && wget https://github.com/neondatabase/lcov/archive/426e7e7a22f669da54278e9b55e6d8caabd00af0.tar.gz -O lcov.tar.gz \
+    && echo "61a22a62e20908b8b9e27d890bd0ea31f567a7b9668065589266371dcbca0992  lcov.tar.gz" | sha256sum --check \
+    && mkdir -p lcov && tar -xzf lcov.tar.gz -C lcov --strip-components=1 \
+    && cd lcov \
+    && make install \
+    && rm -rf ../lcov.tar.gz
+
+# Switch to nonroot user
+USER nonroot:nonroot
+WORKDIR /home/nonroot
+
+# Python
+ENV PYTHON_VERSION=3.9.2 \
+    PYENV_ROOT=/home/nonroot/.pyenv \
+    PATH=/home/nonroot/.pyenv/shims:/home/nonroot/.pyenv/bin:/home/nonroot/.poetry/bin:$PATH
+RUN set -e \
+    && cd $HOME \
+    && curl -sSO https://raw.githubusercontent.com/pyenv/pyenv-installer/master/bin/pyenv-installer \
+    && chmod +x pyenv-installer \
+    && ./pyenv-installer \
+    && export PYENV_ROOT=/home/nonroot/.pyenv \
+    && export PATH="$PYENV_ROOT/bin:$PATH" \
+    && export PATH="$PYENV_ROOT/shims:$PATH" \
+    && pyenv install ${PYTHON_VERSION} \
+    && pyenv global ${PYTHON_VERSION} \
+    && python --version \
+    && pip install --upgrade pip \
+    && pip --version \
+    && pip install pipenv wheel poetry
+
+# Switch to nonroot user (again)
+USER nonroot:nonroot
+WORKDIR /home/nonroot
+
+# Rust
+# Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
+ENV RUSTC_VERSION=1.74.0
+ENV RUSTUP_HOME="/home/nonroot/.rustup"
+ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && whoami && \
+	chmod +x rustup-init && \
+	./rustup-init -y --default-toolchain ${RUSTC_VERSION} && \
+	rm rustup-init && \
+    export PATH="$HOME/.cargo/bin:$PATH" && \
+    . "$HOME/.cargo/env" && \
+    cargo --version && rustup --version && \
+    rustup component add llvm-tools-preview rustfmt clippy && \
+    cargo install --git https://github.com/paritytech/cachepot && \
+    cargo install rustfilt && \
+    cargo install cargo-hakari && \
+    cargo install cargo-deny && \
+    cargo install cargo-hack && \
+    cargo install cargo-nextest && \
+    rm -rf /home/nonroot/.cargo/registry && \
+    rm -rf /home/nonroot/.cargo/git
+ENV RUSTC_WRAPPER=cachepot
+
+# Show versions
+RUN whoami \
+    && python --version \
+    && pip --version \
+    && cargo --version --verbose \
+    && rustup --version --verbose \
+    && rustc --version --verbose \
+    && clang --version

Dockerfile.compute-node

@@ -1,6 +1,6 @@
 ARG PG_VERSION
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
 
@@ -48,7 +48,29 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrowlocks.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgstattuple.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/refint.control && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control && \
+    # We need to grant EXECUTE on pg_stat_statements_reset() to neon_superuser.
+    # In vanilla postgres this function is limited to Postgres role superuser.
+    # In neon we have neon_superuser role that is not a superuser but replaces superuser in some cases.
+    # We could add the additional grant statements to the postgres repository but it would be hard to maintain, 
+    # whenever we need to pick up a new postgres version and we want to limit the changes in our postgres fork,
+    # so we do it here.
+    old_list="pg_stat_statements--1.0--1.1.sql pg_stat_statements--1.1--1.2.sql pg_stat_statements--1.2--1.3.sql pg_stat_statements--1.3--1.4.sql pg_stat_statements--1.4--1.5.sql pg_stat_statements--1.4.sql pg_stat_statements--1.5--1.6.sql"; \
+    # the first loop is for pg_stat_statement extension version <= 1.6
+    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
+        filename=$(basename "$file"); \
+        if echo "$old_list" | grep -q -F "$filename"; then \
+            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO neon_superuser;' >> $file; \
+        fi; \
+    done; \
+    # the second loop is for pg_stat_statement extension versions >= 1.7, 
+    # where pg_stat_statement_reset() got 3 additional arguments
+    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
+        filename=$(basename "$file"); \
+        if ! echo "$old_list" | grep -q -F "$filename"; then \
+            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO neon_superuser;' >> $file; \
+        fi; \
+    done      
 
 #########################################################################################
 #
@@ -569,6 +591,23 @@ RUN wget https://github.com/ChenHuajun/pg_roaringbitmap/archive/refs/tags/v0.5.4
     make -j $(getconf _NPROCESSORS_ONLN) install && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/roaringbitmap.control
 
+#########################################################################################
+#
+# Layer "pg-semver-pg-build"
+# compile pg_semver extension
+#
+#########################################################################################
+FROM build-deps AS pg-semver-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/theory/pg-semver/archive/refs/tags/v0.32.1.tar.gz -O pg_semver.tar.gz && \
+    echo "fbdaf7512026d62eec03fad8687c15ed509b6ba395bff140acd63d2e4fbe25d7 pg_semver.tar.gz" | sha256sum --check && \
+    mkdir pg_semver-src && cd pg_semver-src && tar xvzf ../pg_semver.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/semver.control
+
 #########################################################################################
 #
 # Layer "pg-embedding-pg-build"
@@ -768,6 +807,7 @@ COPY --from=pg-pgx-ulid-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-semver-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-embedding-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
 COPY pgxn/ pgxn/

Dockerfile.compute-tools

@@ -1,7 +1,7 @@
 # First transient image to build compute_tools binaries
 # NB: keep in sync with rust image version in .github/workflows/build_and_test.yml
 ARG REPOSITORY=neondatabase
-ARG IMAGE=rust
+ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
 

README.md

@@ -29,13 +29,14 @@ See developer documentation in [SUMMARY.md](/docs/SUMMARY.md) for more informati
 ```bash
 apt install build-essential libtool libreadline-dev zlib1g-dev flex bison libseccomp-dev \
 libssl-dev clang pkg-config libpq-dev cmake postgresql-client protobuf-compiler \
-libcurl4-openssl-dev openssl python-poetry lsof libicu-dev
+libcurl4-openssl-dev openssl python3-poetry lsof libicu-dev
 ```
 * On Fedora, these packages are needed:
 ```bash
 dnf install flex bison readline-devel zlib-devel openssl-devel \
   libseccomp-devel perl clang cmake postgresql postgresql-contrib protobuf-compiler \
-  protobuf-devel libcurl-devel openssl poetry lsof libicu-devel
+  protobuf-devel libcurl-devel openssl poetry lsof libicu-devel libpq-devel python3-devel \
+  libffi-devel
 ```
 * On Arch based systems, these packages are needed:
 ```bash

compute_tools/Cargo.toml

@@ -13,6 +13,7 @@ clap.workspace = true
 flate2.workspace = true
 futures.workspace = true
 hyper = { workspace = true, features = ["full"] }
+nix.workspace = true
 notify.workspace = true
 num_cpus.workspace = true
 opentelemetry.workspace = true
@@ -20,6 +21,7 @@ postgres.workspace = true
 regex.workspace = true
 serde.workspace = true
 serde_json.workspace = true
+signal-hook.workspace = true
 tar.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
@@ -37,5 +39,6 @@ workspace_hack.workspace = true
 toml_edit.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
 vm_monitor = { version = "0.1", path = "../libs/vm_monitor/" }
-zstd = "0.12.4"
+zstd = "0.13"
 bytes = "1.0"
+rust-ini = "0.20.0"

compute_tools/src/bin/compute_ctl.rs

@@ -31,25 +31,31 @@
 //!             -C 'postgresql://cloud_admin@localhost/postgres' \
 //!             -S /var/db/postgres/specs/current.json \
 //!             -b /usr/local/bin/postgres \
-//!             -r http://pg-ext-s3-gateway
+//!             -r http://pg-ext-s3-gateway \
+//!             --pgbouncer-connstr 'host=localhost port=6432 dbname=pgbouncer user=cloud_admin sslmode=disable'
+//!             --pgbouncer-ini-path /etc/pgbouncer.ini \
 //! ```
 //!
 use std::collections::HashMap;
 use std::fs::File;
 use std::path::Path;
 use std::process::exit;
+use std::sync::atomic::Ordering;
 use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
 use std::{thread, time::Duration};
 
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Arg;
+use nix::sys::signal::{kill, Signal};
+use signal_hook::consts::{SIGQUIT, SIGTERM};
+use signal_hook::{consts::SIGINT, iterator::Signals};
 use tracing::{error, info};
 use url::Url;
 
 use compute_api::responses::ComputeStatus;
 
-use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec};
+use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec, PG_PID, SYNC_SAFEKEEPERS_PID};
 use compute_tools::configurator::launch_configurator;
 use compute_tools::extension_server::get_pg_version;
 use compute_tools::http::api::launch_http_server;
@@ -65,6 +71,13 @@ const BUILD_TAG_DEFAULT: &str = "latest";
 fn main() -> Result<()> {
     init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
 
+    let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
+    thread::spawn(move || {
+        for sig in signals.forever() {
+            handle_exit_signal(sig);
+        }
+    });
+
     let build_tag = option_env!("BUILD_TAG")
         .unwrap_or(BUILD_TAG_DEFAULT)
         .to_string();
@@ -99,6 +112,9 @@ fn main() -> Result<()> {
     let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
 
+    let pgbouncer_connstr = matches.get_one::<String>("pgbouncer-connstr");
+    let pgbouncer_ini_path = matches.get_one::<String>("pgbouncer-ini-path");
+
     // Extract OpenTelemetry context for the startup actions from the
     // TRACEPARENT and TRACESTATE env variables, and attach it to the current
     // tracing context.
@@ -209,6 +225,8 @@ fn main() -> Result<()> {
         ext_remote_storage: ext_remote_storage.map(|s| s.to_string()),
         ext_download_progress: RwLock::new(HashMap::new()),
         build_tag,
+        pgbouncer_connstr: pgbouncer_connstr.map(|s| s.to_string()),
+        pgbouncer_ini_path: pgbouncer_ini_path.map(|s| s.to_string()),
     };
     let compute = Arc::new(compute_node);
 
@@ -339,6 +357,7 @@ fn main() -> Result<()> {
         let ecode = pg
             .wait()
             .expect("failed to start waiting on Postgres process");
+        PG_PID.store(0, Ordering::SeqCst);
         info!("Postgres exited with code {}, shutting down", ecode);
         exit_code = ecode.code()
     }
@@ -493,6 +512,41 @@ fn cli() -> clap::Command {
                 )
                 .value_name("FILECACHE_CONNSTR"),
         )
+        .arg(
+            Arg::new("pgbouncer-connstr")
+                .long("pgbouncer-connstr")
+                .default_value(
+                    "host=localhost port=6432 dbname=pgbouncer user=cloud_admin sslmode=disable",
+                )
+                .value_name("PGBOUNCER_CONNSTR"),
+        )
+        .arg(
+            Arg::new("pgbouncer-ini-path")
+                .long("pgbouncer-ini-path")
+                // Note: this doesn't match current path for pgbouncer.ini.
+                // Until we fix it, we need to pass the path explicitly
+                // or this will be effectively no-op.
+                .default_value("/etc/pgbouncer.ini")
+                .value_name("PGBOUNCER_INI_PATH"),
+        )
+}
+
+/// When compute_ctl is killed, send also termination signal to sync-safekeepers
+/// to prevent leakage. TODO: it is better to convert compute_ctl to async and
+/// wait for termination which would be easy then.
+fn handle_exit_signal(sig: i32) {
+    info!("received {sig} termination signal");
+    let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
+    if ss_pid != 0 {
+        let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
+        kill(ss_pid, Signal::SIGTERM).ok();
+    }
+    let pg_pid = PG_PID.load(Ordering::SeqCst);
+    if pg_pid != 0 {
+        let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
+        kill(pg_pid, Signal::SIGTERM).ok();
+    }
+    exit(1);
 }
 
 #[test]

compute_tools/src/compute.rs

@@ -6,7 +6,10 @@ use std::os::unix::fs::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::thread;
 use std::time::Instant;
 
 use anyhow::{Context, Result};
@@ -33,6 +36,9 @@ use crate::spec::*;
 use crate::sync_sk::{check_if_synced, ping_safekeeper};
 use crate::{config, extension_server};
 
+pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
+pub static PG_PID: AtomicU32 = AtomicU32::new(0);
+
 /// Compute node info shared across several `compute_ctl` threads.
 pub struct ComputeNode {
     // Url type maintains proper escaping
@@ -64,6 +70,10 @@ pub struct ComputeNode {
     // key: ext_archive_name, value: started download time, download_completed?
     pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
     pub build_tag: String,
+    // connection string to pgbouncer to change settings
+    pub pgbouncer_connstr: Option<String>,
+    // path to pgbouncer.ini to change settings
+    pub pgbouncer_ini_path: Option<String>,
 }
 
 // store some metrics about download size that might impact startup time
@@ -496,6 +506,7 @@ impl ComputeNode {
             .stdout(Stdio::piped())
             .spawn()
             .expect("postgres --sync-safekeepers failed to start");
+        SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
 
         // `postgres --sync-safekeepers` will print all log output to stderr and
         // final LSN to stdout. So we pipe only stdout, while stderr will be automatically
@@ -503,6 +514,7 @@ impl ComputeNode {
         let sync_output = sync_handle
             .wait_with_output()
             .expect("postgres --sync-safekeepers failed");
+        SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
 
         if !sync_output.status.success() {
             anyhow::bail!(
@@ -657,6 +669,7 @@ impl ComputeNode {
             })
             .spawn()
             .expect("cannot start postgres process");
+        PG_PID.store(pg.id(), Ordering::SeqCst);
 
         wait_for_postgres(&mut pg, pgdata_path)?;
 
@@ -737,6 +750,31 @@ impl ComputeNode {
     pub fn reconfigure(&self) -> Result<()> {
         let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
 
+        if let Some(connstr) = &self.pgbouncer_connstr {
+            info!("tuning pgbouncer with connstr: {:?}", connstr);
+
+            let rt = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("failed to create rt");
+
+            // Spawn a thread to do the tuning,
+            // so that we don't block the main thread that starts Postgres.
+            let pgbouncer_settings = spec.pgbouncer_settings.clone();
+            let connstr_clone = connstr.clone();
+            let pgbouncer_ini_path = self.pgbouncer_ini_path.clone();
+            let _handle = thread::spawn(move || {
+                let res = rt.block_on(tune_pgbouncer(
+                    pgbouncer_settings,
+                    &connstr_clone,
+                    pgbouncer_ini_path,
+                ));
+                if let Err(err) = res {
+                    error!("error while tuning pgbouncer: {err:?}");
+                }
+            });
+        }
+
         // Write new config
         let pgdata_path = Path::new(&self.pgdata);
         let postgresql_conf_path = pgdata_path.join("postgresql.conf");
@@ -791,6 +829,32 @@ impl ComputeNode {
             pspec.timeline_id,
         );
 
+        // tune pgbouncer
+        if let Some(connstr) = &self.pgbouncer_connstr {
+            info!("tuning pgbouncer with connstr: {:?}", connstr);
+
+            let rt = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("failed to create rt");
+
+            // Spawn a thread to do the tuning,
+            // so that we don't block the main thread that starts Postgres.
+            let pgbouncer_settings = pspec.spec.pgbouncer_settings.clone();
+            let connstr_clone = connstr.clone();
+            let pgbouncer_ini_path = self.pgbouncer_ini_path.clone();
+            let _handle = thread::spawn(move || {
+                let res = rt.block_on(tune_pgbouncer(
+                    pgbouncer_settings,
+                    &connstr_clone,
+                    pgbouncer_ini_path,
+                ));
+                if let Err(err) = res {
+                    error!("error while tuning pgbouncer: {err:?}");
+                }
+            });
+        }
+
         info!(
             "start_compute spec.remote_extensions {:?}",
             pspec.spec.remote_extensions

compute_tools/src/pg_helpers.rs

@@ -9,9 +9,11 @@ use std::process::Child;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Result};
+use ini::Ini;
 use notify::{RecursiveMode, Watcher};
 use postgres::{Client, Transaction};
-use tracing::{debug, instrument};
+use tokio_postgres::NoTls;
+use tracing::{debug, error, info, instrument};
 
 use compute_api::spec::{Database, GenericOption, GenericOptions, PgIdent, Role};
 
@@ -359,3 +361,68 @@ pub fn create_pgdata(pgdata: &str) -> Result<()> {
 
     Ok(())
 }
+
+/// Update pgbouncer.ini with provided options
+pub fn update_pgbouncer_ini(
+    pgbouncer_config: HashMap<String, String>,
+    pgbouncer_ini_path: &str,
+) -> Result<()> {
+    let mut conf = Ini::load_from_file(pgbouncer_ini_path)?;
+    let section = conf.section_mut(Some("pgbouncer")).unwrap();
+
+    for (option_name, value) in pgbouncer_config.iter() {
+        section.insert(option_name, value);
+    }
+
+    conf.write_to_file(pgbouncer_ini_path)?;
+    Ok(())
+}
+
+/// Tune pgbouncer.
+/// 1. Apply new config using pgbouncer admin console
+/// 2. Add new values to pgbouncer.ini to preserve them after restart
+pub async fn tune_pgbouncer(
+    pgbouncer_settings: Option<HashMap<String, String>>,
+    pgbouncer_connstr: &str,
+    pgbouncer_ini_path: Option<String>,
+) -> Result<()> {
+    if let Some(pgbouncer_config) = pgbouncer_settings {
+        // Apply new config
+        let connect_result = tokio_postgres::connect(pgbouncer_connstr, NoTls).await;
+        let (client, connection) = connect_result.unwrap();
+        tokio::spawn(async move {
+            if let Err(e) = connection.await {
+                eprintln!("connection error: {}", e);
+            }
+        });
+
+        for (option_name, value) in pgbouncer_config.iter() {
+            info!(
+                "Applying pgbouncer setting change: {} = {}",
+                option_name, value
+            );
+            let query = format!("SET {} = {}", option_name, value);
+
+            let result = client.simple_query(&query).await;
+
+            info!("Applying pgbouncer setting change: {}", query);
+            info!("pgbouncer setting change result: {:?}", result);
+
+            if let Err(err) = result {
+                // Don't fail on error, just print it into log
+                error!(
+                    "Failed to apply pgbouncer setting change: {},  {}",
+                    query, err
+                );
+            };
+        }
+
+        // save values to pgbouncer.ini
+        // so that they are preserved after pgbouncer restart
+        if let Some(pgbouncer_ini_path) = pgbouncer_ini_path {
+            update_pgbouncer_ini(pgbouncer_config, &pgbouncer_ini_path)?;
+        }
+    }
+
+    Ok(())
+}

compute_tools/src/spec.rs

@@ -298,7 +298,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
                 // from neon_superuser.
                 let mut query: String = format!(
-                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB IN ROLE neon_superuser",
+                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -370,33 +370,49 @@ pub fn handle_role_deletions(spec: &ComputeSpec, connstr: &str, client: &mut Cli
     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);
-
-            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, &db.name, &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)?;
+            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(())
 }
 

control_plane/attachment_service/Cargo.toml

@@ -0,0 +1,25 @@
+[package]
+name = "attachment_service"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+clap.workspace = true
+futures.workspace = true
+hyper.workspace = true
+pageserver_api.workspace = true
+postgres_connection.workspace = true
+reqwest.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+thiserror.workspace = true
+tokio.workspace = true
+tokio-util.workspace = true
+tracing.workspace = true
+
+utils = { path = "../../libs/utils/" }
+control_plane = { path = ".." }
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+

control_plane/attachment_service/src/compute_hook.rs

@@ -0,0 +1,108 @@
+use std::collections::HashMap;
+
+use control_plane::endpoint::ComputeControlPlane;
+use control_plane::local_env::LocalEnv;
+use pageserver_api::shard::{ShardIndex, TenantShardId};
+use postgres_connection::parse_host_port;
+use utils::id::{NodeId, TenantId};
+
+pub(super) struct ComputeHookTenant {
+    shards: Vec<(ShardIndex, NodeId)>,
+}
+
+impl ComputeHookTenant {
+    pub(super) async fn maybe_reconfigure(&mut self, tenant_id: TenantId) -> anyhow::Result<()> {
+        // Find the highest shard count and drop any shards that aren't
+        // for that shard count.
+        let shard_count = self.shards.iter().map(|(k, _v)| k.shard_count).max();
+        let Some(shard_count) = shard_count else {
+            // No shards, nothing to do.
+            tracing::info!("ComputeHookTenant::maybe_reconfigure: no shards");
+            return Ok(());
+        };
+
+        self.shards.retain(|(k, _v)| k.shard_count == shard_count);
+        self.shards
+            .sort_by_key(|(shard, _node_id)| shard.shard_number);
+
+        if self.shards.len() == shard_count.0 as usize {
+            // We have pageservers for all the shards: proceed to reconfigure compute
+            let env = LocalEnv::load_config().expect("Error loading config");
+            let cplane = ComputeControlPlane::load(env.clone())
+                .expect("Error loading compute control plane");
+
+            let compute_pageservers = self
+                .shards
+                .iter()
+                .map(|(_shard, node_id)| {
+                    let ps_conf = env
+                        .get_pageserver_conf(*node_id)
+                        .expect("Unknown pageserver");
+                    let (pg_host, pg_port) = parse_host_port(&ps_conf.listen_pg_addr)
+                        .expect("Unable to parse listen_pg_addr");
+                    (pg_host, pg_port.unwrap_or(5432))
+                })
+                .collect::<Vec<_>>();
+
+            for (endpoint_name, endpoint) in &cplane.endpoints {
+                if endpoint.tenant_id == tenant_id && endpoint.status() == "running" {
+                    tracing::info!("🔁 Reconfiguring endpoint {}", endpoint_name,);
+                    endpoint.reconfigure(compute_pageservers.clone()).await?;
+                }
+            }
+        } else {
+            tracing::info!(
+                "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
+                self.shards.len(),
+                shard_count.0
+            );
+        }
+
+        Ok(())
+    }
+}
+
+/// The compute hook is a destination for notifications about changes to tenant:pageserver
+/// mapping.  It aggregates updates for the shards in a tenant, and when appropriate reconfigures
+/// the compute connection string.
+pub(super) struct ComputeHook {
+    state: tokio::sync::Mutex<HashMap<TenantId, ComputeHookTenant>>,
+}
+
+impl ComputeHook {
+    pub(super) fn new() -> Self {
+        Self {
+            state: Default::default(),
+        }
+    }
+
+    pub(super) async fn notify(
+        &self,
+        tenant_shard_id: TenantShardId,
+        node_id: NodeId,
+    ) -> anyhow::Result<()> {
+        tracing::info!("ComputeHook::notify: {}->{}", tenant_shard_id, node_id);
+        let mut locked = self.state.lock().await;
+        let entry = locked
+            .entry(tenant_shard_id.tenant_id)
+            .or_insert_with(|| ComputeHookTenant { shards: Vec::new() });
+
+        let shard_index = ShardIndex {
+            shard_count: tenant_shard_id.shard_count,
+            shard_number: tenant_shard_id.shard_number,
+        };
+
+        let mut set = false;
+        for (existing_shard, existing_node) in &mut entry.shards {
+            if *existing_shard == shard_index {
+                *existing_node = node_id;
+                set = true;
+            }
+        }
+        if !set {
+            entry.shards.push((shard_index, node_id));
+        }
+
+        entry.maybe_reconfigure(tenant_shard_id.tenant_id).await
+    }
+}

control_plane/attachment_service/src/http.rs

@@ -0,0 +1,195 @@
+use crate::reconciler::ReconcileError;
+use crate::service::Service;
+use hyper::StatusCode;
+use hyper::{Body, Request, Response};
+use pageserver_api::models::{TenantCreateRequest, TenantShardSplitRequest, TimelineCreateRequest};
+use pageserver_api::shard::TenantShardId;
+use std::sync::Arc;
+use utils::http::endpoint::request_span;
+use utils::http::request::parse_request_param;
+use utils::id::TenantId;
+
+use utils::{
+    http::{
+        endpoint::{self},
+        error::ApiError,
+        json::{json_request, json_response},
+        RequestExt, RouterBuilder,
+    },
+    id::NodeId,
+};
+
+use pageserver_api::control_api::{ReAttachRequest, ValidateRequest};
+
+use control_plane::attachment_service::{
+    AttachHookRequest, InspectRequest, NodeConfigureRequest, NodeRegisterRequest,
+    TenantShardMigrateRequest,
+};
+
+/// State available to HTTP request handlers
+#[derive(Clone)]
+pub struct HttpState {
+    service: Arc<crate::service::Service>,
+}
+
+impl HttpState {
+    pub fn new(service: Arc<crate::service::Service>) -> Self {
+        Self { service }
+    }
+}
+
+#[inline(always)]
+fn get_state(request: &Request<Body>) -> &HttpState {
+    request
+        .data::<Arc<HttpState>>()
+        .expect("unknown state type")
+        .as_ref()
+}
+
+/// Pageserver calls into this on startup, to learn which tenants it should attach
+async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let reattach_req = json_request::<ReAttachRequest>(&mut req).await?;
+    let state = get_state(&req);
+    json_response(StatusCode::OK, state.service.re_attach(reattach_req))
+}
+
+/// Pageserver calls into this before doing deletions, to confirm that it still
+/// holds the latest generation for the tenants with deletions enqueued
+async fn handle_validate(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let validate_req = json_request::<ValidateRequest>(&mut req).await?;
+    let state = get_state(&req);
+    json_response(StatusCode::OK, state.service.validate(validate_req))
+}
+
+/// Call into this before attaching a tenant to a pageserver, to acquire a generation number
+/// (in the real control plane this is unnecessary, because the same program is managing
+///  generation numbers and doing attachments).
+async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let attach_req = json_request::<AttachHookRequest>(&mut req).await?;
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.attach_hook(attach_req))
+}
+
+async fn handle_inspect(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let inspect_req = json_request::<InspectRequest>(&mut req).await?;
+
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.inspect(inspect_req))
+}
+
+async fn handle_tenant_create(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let create_req = json_request::<TenantCreateRequest>(&mut req).await?;
+    let state = get_state(&req);
+    json_response(
+        StatusCode::OK,
+        state.service.tenant_create(create_req).await?,
+    )
+}
+
+async fn handle_tenant_timeline_create(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let create_req = json_request::<TimelineCreateRequest>(&mut req).await?;
+
+    let state = get_state(&req);
+    json_response(
+        StatusCode::OK,
+        state
+            .service
+            .tenant_timeline_create(tenant_id, create_req)
+            .await?,
+    )
+}
+
+async fn handle_tenant_locate(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.tenant_locate(tenant_id)?)
+}
+
+async fn handle_node_register(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let register_req = json_request::<NodeRegisterRequest>(&mut req).await?;
+    let state = get_state(&req);
+    state.service.node_register(register_req);
+    json_response(StatusCode::OK, ())
+}
+
+async fn handle_node_configure(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let node_id: NodeId = parse_request_param(&req, "node_id")?;
+    let config_req = json_request::<NodeConfigureRequest>(&mut req).await?;
+    if node_id != config_req.node_id {
+        return Err(ApiError::BadRequest(anyhow::anyhow!(
+            "Path and body node_id differ"
+        )));
+    }
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.node_configure(config_req)?)
+}
+
+async fn handle_tenant_shard_split(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let split_req = json_request::<TenantShardSplitRequest>(&mut req).await?;
+    let state = get_state(&req);
+
+    json_response(
+        StatusCode::OK,
+        state
+            .service
+            .tenant_shard_split(tenant_id, split_req)
+            .await?,
+    )
+}
+
+async fn handle_tenant_shard_migrate(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&req, "tenant_shard_id")?;
+    let migrate_req = json_request::<TenantShardMigrateRequest>(&mut req).await?;
+    let state = get_state(&req);
+    json_response(
+        StatusCode::OK,
+        state
+            .service
+            .tenant_shard_migrate(tenant_shard_id, migrate_req)
+            .await?,
+    )
+}
+
+/// Status endpoint is just used for checking that our HTTP listener is up
+async fn handle_status(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    json_response(StatusCode::OK, ())
+}
+
+impl From<ReconcileError> for ApiError {
+    fn from(value: ReconcileError) -> Self {
+        ApiError::Conflict(format!("Reconciliation error: {}", value))
+    }
+}
+
+pub fn make_router(service: Arc<Service>) -> RouterBuilder<hyper::Body, ApiError> {
+    endpoint::make_router()
+        .data(Arc::new(HttpState { service }))
+        .get("/status", |r| request_span(r, handle_status))
+        .post("/re-attach", |r| request_span(r, handle_re_attach))
+        .post("/validate", |r| request_span(r, handle_validate))
+        .post("/attach-hook", |r| request_span(r, handle_attach_hook))
+        .post("/inspect", |r| request_span(r, handle_inspect))
+        .post("/node", |r| request_span(r, handle_node_register))
+        .put("/node/:node_id/config", |r| {
+            request_span(r, handle_node_configure)
+        })
+        .post("/tenant", |r| request_span(r, handle_tenant_create))
+        .post("/tenant/:tenant_id/timeline", |r| {
+            request_span(r, handle_tenant_timeline_create)
+        })
+        .get("/tenant/:tenant_id/locate", |r| {
+            request_span(r, handle_tenant_locate)
+        })
+        .put("/tenant/:tenant_id/shard_split", |r| {
+            request_span(r, handle_tenant_shard_split)
+        })
+        .put("/tenant/:tenant_shard_id/migrate", |r| {
+            request_span(r, handle_tenant_shard_migrate)
+        })
+}

control_plane/attachment_service/src/lib.rs

@@ -0,0 +1,49 @@
+use utils::seqwait::MonotonicCounter;
+
+mod compute_hook;
+pub mod http;
+mod node;
+mod reconciler;
+mod scheduler;
+pub mod service;
+mod tenant_state;
+
+#[derive(Clone)]
+enum PlacementPolicy {
+    /// Cheapest way to attach a tenant: just one pageserver, no secondary
+    Single,
+    /// Production-ready way to attach a tenant: one attached pageserver and
+    /// some number of secondaries.
+    Double(usize),
+}
+
+#[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone)]
+struct Sequence(u64);
+
+impl std::fmt::Display for Sequence {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
+impl MonotonicCounter<Sequence> for Sequence {
+    fn cnt_advance(&mut self, v: Sequence) {
+        assert!(*self <= v);
+        *self = v;
+    }
+    fn cnt_value(&self) -> Sequence {
+        *self
+    }
+}
+
+impl Sequence {
+    fn next(&self) -> Sequence {
+        Sequence(self.0 + 1)
+    }
+}
+
+impl Default for PlacementPolicy {
+    fn default() -> Self {
+        PlacementPolicy::Double(1)
+    }
+}

control_plane/attachment_service/src/main.rs

@@ -0,0 +1,65 @@
+/// The attachment service mimics the aspects of the control plane API
+/// that are required for a pageserver to operate.
+///
+/// This enables running & testing pageservers without a full-blown
+/// deployment of the Neon cloud platform.
+///
+use anyhow::anyhow;
+use attachment_service::http::make_router;
+use attachment_service::service::Service;
+use clap::Parser;
+use std::path::PathBuf;
+use utils::logging::{self, LogFormat};
+use utils::signals::{ShutdownSignals, Signal};
+
+use utils::tcp_listener;
+
+#[derive(Parser)]
+#[command(author, version, about, long_about = None)]
+#[command(arg_required_else_help(true))]
+struct Cli {
+    /// Host and port to listen on, like `127.0.0.1:1234`
+    #[arg(short, long)]
+    listen: std::net::SocketAddr,
+
+    /// Path to the .json file to store state (will be created if it doesn't exist)
+    #[arg(short, long)]
+    path: PathBuf,
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    logging::init(
+        LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::Disabled,
+        logging::Output::Stdout,
+    )?;
+
+    let args = Cli::parse();
+    tracing::info!(
+        "Starting, state at {}, listening on {}",
+        args.path.to_string_lossy(),
+        args.listen
+    );
+
+    let service = Service::spawn();
+
+    let http_listener = tcp_listener::bind(args.listen)?;
+    let router = make_router(service).build().map_err(|err| anyhow!(err))?;
+    let service = utils::http::RouterService::new(router).unwrap();
+    let server = hyper::Server::from_tcp(http_listener)?.serve(service);
+
+    tracing::info!("Serving on {0}", args.listen);
+
+    tokio::task::spawn(server);
+
+    ShutdownSignals::handle(|signal| match signal {
+        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
+            tracing::info!("Got {}. Terminating", signal.name());
+            // We're just a test helper: no graceful shutdown.
+            std::process::exit(0);
+        }
+    })?;
+
+    Ok(())
+}

control_plane/attachment_service/src/node.rs

@@ -0,0 +1,40 @@
+use control_plane::attachment_service::{NodeAvailability, NodeSchedulingPolicy};
+use utils::id::NodeId;
+
+#[derive(Clone)]
+pub(crate) struct Node {
+    pub(crate) id: NodeId,
+
+    pub(crate) availability: NodeAvailability,
+    pub(crate) scheduling: NodeSchedulingPolicy,
+
+    pub(crate) listen_http_addr: String,
+    pub(crate) listen_http_port: u16,
+
+    pub(crate) listen_pg_addr: String,
+    pub(crate) listen_pg_port: u16,
+}
+
+impl Node {
+    pub(crate) fn base_url(&self) -> String {
+        format!(
+            "http://{}:{}/v1",
+            self.listen_http_addr, self.listen_http_port
+        )
+    }
+
+    /// Is this node elegible to have work scheduled onto it?
+    pub(crate) fn may_schedule(&self) -> bool {
+        match self.availability {
+            NodeAvailability::Active => {}
+            NodeAvailability::Offline => return false,
+        }
+
+        match self.scheduling {
+            NodeSchedulingPolicy::Active => true,
+            NodeSchedulingPolicy::Draining => false,
+            NodeSchedulingPolicy::Filling => true,
+            NodeSchedulingPolicy::Pause => false,
+        }
+    }
+}

control_plane/attachment_service/src/reconciler.rs

@@ -0,0 +1,478 @@
+use control_plane::attachment_service::NodeAvailability;
+use control_plane::local_env::LocalEnv;
+use control_plane::pageserver::PageServerNode;
+use hyper::Method;
+use pageserver_api::models::{
+    LocationConfig, LocationConfigMode, LocationConfigSecondary, TenantConfig,
+    TenantLocationConfigRequest,
+};
+use pageserver_api::shard::{ShardIdentity, TenantShardId};
+use reqwest::Client;
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio_util::sync::CancellationToken;
+use utils::generation::Generation;
+use utils::id::{NodeId, TimelineId};
+use utils::lsn::Lsn;
+
+use crate::compute_hook::ComputeHook;
+use crate::node::Node;
+use crate::tenant_state::{IntentState, ObservedState, ObservedStateLocation};
+
+/// Object with the lifetime of the background reconcile task that is created
+/// for tenants which have a difference between their intent and observed states.
+pub(super) struct Reconciler {
+    /// See [`crate::tenant_state::TenantState`] for the meanings of these fields: they are a snapshot
+    /// of a tenant's state from when we spawned a reconcile task.
+    pub(super) tenant_shard_id: TenantShardId,
+    pub(crate) shard: ShardIdentity,
+    pub(crate) generation: Generation,
+    pub(crate) intent: IntentState,
+    pub(crate) config: TenantConfig,
+    pub(crate) observed: ObservedState,
+
+    /// A snapshot of the pageservers as they were when we were asked
+    /// to reconcile.
+    pub(crate) pageservers: Arc<HashMap<NodeId, Node>>,
+
+    /// A hook to notify the running postgres instances when we change the location
+    /// of a tenant
+    pub(crate) compute_hook: Arc<ComputeHook>,
+
+    /// A means to abort background reconciliation: it is essential to
+    /// call this when something changes in the original TenantState that
+    /// will make this reconciliation impossible or unnecessary, for
+    /// example when a pageserver node goes offline, or the PlacementPolicy for
+    /// the tenant is changed.
+    pub(crate) cancel: CancellationToken,
+}
+
+#[derive(thiserror::Error, Debug)]
+pub enum ReconcileError {
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
+impl Reconciler {
+    async fn location_config(
+        &mut self,
+        node_id: NodeId,
+        config: LocationConfig,
+    ) -> anyhow::Result<()> {
+        let node = self
+            .pageservers
+            .get(&node_id)
+            .expect("Pageserver may not be removed while referenced");
+
+        self.observed
+            .locations
+            .insert(node.id, ObservedStateLocation { conf: None });
+
+        let configure_request = TenantLocationConfigRequest {
+            tenant_id: self.tenant_shard_id,
+            config: config.clone(),
+        };
+
+        let client = Client::new();
+        let response = client
+            .request(
+                Method::PUT,
+                format!(
+                    "{}/tenant/{}/location_config",
+                    node.base_url(),
+                    self.tenant_shard_id
+                ),
+            )
+            .json(&configure_request)
+            .send()
+            .await?;
+
+        self.observed
+            .locations
+            .insert(node.id, ObservedStateLocation { conf: Some(config) });
+
+        response.error_for_status()?;
+
+        Ok(())
+    }
+
+    async fn maybe_live_migrate(&mut self) -> Result<(), ReconcileError> {
+        let destination = if let Some(node_id) = self.intent.attached {
+            match self.observed.locations.get(&node_id) {
+                Some(conf) => {
+                    // We will do a live migration only if the intended destination is not
+                    // currently in an attached state.
+                    match &conf.conf {
+                        Some(conf) if conf.mode == LocationConfigMode::Secondary => {
+                            // Fall through to do a live migration
+                            node_id
+                        }
+                        None | Some(_) => {
+                            // Attached or uncertain: don't do a live migration, proceed
+                            // with a general-case reconciliation
+                            tracing::info!("maybe_live_migrate: destination is None or attached");
+                            return Ok(());
+                        }
+                    }
+                }
+                None => {
+                    // Our destination is not attached: maybe live migrate if some other
+                    // node is currently attached.  Fall through.
+                    node_id
+                }
+            }
+        } else {
+            // No intent to be attached
+            tracing::info!("maybe_live_migrate: no attached intent");
+            return Ok(());
+        };
+
+        let mut origin = None;
+        for (node_id, state) in &self.observed.locations {
+            if let Some(observed_conf) = &state.conf {
+                if observed_conf.mode == LocationConfigMode::AttachedSingle {
+                    let node = self
+                        .pageservers
+                        .get(node_id)
+                        .expect("Nodes may not be removed while referenced");
+                    // We will only attempt live migration if the origin is not offline: this
+                    // avoids trying to do it while reconciling after responding to an HA failover.
+                    if !matches!(node.availability, NodeAvailability::Offline) {
+                        origin = Some(*node_id);
+                        break;
+                    }
+                }
+            }
+        }
+
+        let Some(origin) = origin else {
+            tracing::info!("maybe_live_migrate: no origin found");
+            return Ok(());
+        };
+
+        // We have an origin and a destination: proceed to do the live migration
+        let env = LocalEnv::load_config().expect("Error loading config");
+        let origin_ps = PageServerNode::from_env(
+            &env,
+            env.get_pageserver_conf(origin)
+                .expect("Conf missing pageserver"),
+        );
+        let destination_ps = PageServerNode::from_env(
+            &env,
+            env.get_pageserver_conf(destination)
+                .expect("Conf missing pageserver"),
+        );
+
+        tracing::info!(
+            "Live migrating {}->{}",
+            origin_ps.conf.id,
+            destination_ps.conf.id
+        );
+        self.live_migrate(origin_ps, destination_ps).await?;
+
+        Ok(())
+    }
+
+    pub async fn live_migrate(
+        &mut self,
+        origin_ps: PageServerNode,
+        dest_ps: PageServerNode,
+    ) -> anyhow::Result<()> {
+        // `maybe_live_migrate` is responsibble for sanity of inputs
+        assert!(origin_ps.conf.id != dest_ps.conf.id);
+
+        fn build_location_config(
+            shard: &ShardIdentity,
+            config: &TenantConfig,
+            mode: LocationConfigMode,
+            generation: Option<Generation>,
+            secondary_conf: Option<LocationConfigSecondary>,
+        ) -> LocationConfig {
+            LocationConfig {
+                mode,
+                generation: generation.map(|g| g.into().unwrap()),
+                secondary_conf,
+                tenant_conf: config.clone(),
+                shard_number: shard.number.0,
+                shard_count: shard.count.0,
+                shard_stripe_size: shard.stripe_size.0,
+            }
+        }
+
+        async fn get_lsns(
+            tenant_shard_id: TenantShardId,
+            pageserver: &PageServerNode,
+        ) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
+            let timelines = pageserver.timeline_list(&tenant_shard_id).await?;
+            Ok(timelines
+                .into_iter()
+                .map(|t| (t.timeline_id, t.last_record_lsn))
+                .collect())
+        }
+
+        async fn await_lsn(
+            tenant_shard_id: TenantShardId,
+            pageserver: &PageServerNode,
+            baseline: HashMap<TimelineId, Lsn>,
+        ) -> anyhow::Result<()> {
+            loop {
+                let latest = match get_lsns(tenant_shard_id, pageserver).await {
+                    Ok(l) => l,
+                    Err(e) => {
+                        println!(
+                            "🕑 Can't get LSNs on pageserver {} yet, waiting ({e})",
+                            pageserver.conf.id
+                        );
+                        std::thread::sleep(Duration::from_millis(500));
+                        continue;
+                    }
+                };
+
+                let mut any_behind: bool = false;
+                for (timeline_id, baseline_lsn) in &baseline {
+                    match latest.get(timeline_id) {
+                        Some(latest_lsn) => {
+                            println!("🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
+                            if latest_lsn < baseline_lsn {
+                                any_behind = true;
+                            }
+                        }
+                        None => {
+                            // Expected timeline isn't yet visible on migration destination.
+                            // (IRL we would have to account for timeline deletion, but this
+                            //  is just test helper)
+                            any_behind = true;
+                        }
+                    }
+                }
+
+                if !any_behind {
+                    println!("✅ LSN caught up.  Proceeding...");
+                    break;
+                } else {
+                    std::thread::sleep(Duration::from_millis(500));
+                }
+            }
+
+            Ok(())
+        }
+
+        tracing::info!(
+            "🔁 Switching origin pageserver {} to stale mode",
+            origin_ps.conf.id
+        );
+
+        // FIXME: it is incorrect to use self.generation here, we should use the generation
+        // from the ObservedState of the origin pageserver (it might be older than self.generation)
+        let stale_conf = build_location_config(
+            &self.shard,
+            &self.config,
+            LocationConfigMode::AttachedStale,
+            Some(self.generation),
+            None,
+        );
+        origin_ps
+            .location_config(
+                self.tenant_shard_id,
+                stale_conf,
+                Some(Duration::from_secs(10)),
+            )
+            .await?;
+
+        let baseline_lsns = Some(get_lsns(self.tenant_shard_id, &origin_ps).await?);
+
+        // Increment generation before attaching to new pageserver
+        self.generation = self.generation.next();
+
+        let dest_conf = build_location_config(
+            &self.shard,
+            &self.config,
+            LocationConfigMode::AttachedMulti,
+            Some(self.generation),
+            None,
+        );
+
+        tracing::info!("🔁 Attaching to pageserver {}", dest_ps.conf.id);
+        dest_ps
+            .location_config(self.tenant_shard_id, dest_conf, None)
+            .await?;
+
+        if let Some(baseline) = baseline_lsns {
+            tracing::info!("🕑 Waiting for LSN to catch up...");
+            await_lsn(self.tenant_shard_id, &dest_ps, baseline).await?;
+        }
+
+        tracing::info!("🔁 Notifying compute to use pageserver {}", dest_ps.conf.id);
+        self.compute_hook
+            .notify(self.tenant_shard_id, dest_ps.conf.id)
+            .await?;
+
+        // Downgrade the origin to secondary.  If the tenant's policy is PlacementPolicy::Single, then
+        // this location will be deleted in the general case reconciliation that runs after this.
+        let origin_secondary_conf = build_location_config(
+            &self.shard,
+            &self.config,
+            LocationConfigMode::Secondary,
+            None,
+            Some(LocationConfigSecondary { warm: true }),
+        );
+        origin_ps
+            .location_config(self.tenant_shard_id, origin_secondary_conf.clone(), None)
+            .await?;
+        // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
+        // partway through.  In fact, all location conf API calls should be in a wrapper that sets
+        // the observed state to None, then runs, then sets it to what we wrote.
+        self.observed.locations.insert(
+            origin_ps.conf.id,
+            ObservedStateLocation {
+                conf: Some(origin_secondary_conf),
+            },
+        );
+
+        println!(
+            "🔁 Switching to AttachedSingle mode on pageserver {}",
+            dest_ps.conf.id
+        );
+        let dest_final_conf = build_location_config(
+            &self.shard,
+            &self.config,
+            LocationConfigMode::AttachedSingle,
+            Some(self.generation),
+            None,
+        );
+        dest_ps
+            .location_config(self.tenant_shard_id, dest_final_conf.clone(), None)
+            .await?;
+        self.observed.locations.insert(
+            dest_ps.conf.id,
+            ObservedStateLocation {
+                conf: Some(dest_final_conf),
+            },
+        );
+
+        println!("✅ Migration complete");
+
+        Ok(())
+    }
+
+    /// Reconciling a tenant makes API calls to pageservers until the observed state
+    /// matches the intended state.
+    ///
+    /// First we apply special case handling (e.g. for live migrations), and then a
+    /// general case reconciliation where we walk through the intent by pageserver
+    /// and call out to the pageserver to apply the desired state.
+    pub(crate) async fn reconcile(&mut self) -> Result<(), ReconcileError> {
+        // TODO: if any of self.observed is None, call to remote pageservers
+        // to learn correct state.
+
+        // Special case: live migration
+        self.maybe_live_migrate().await?;
+
+        // If the attached pageserver is not attached, do so now.
+        if let Some(node_id) = self.intent.attached {
+            let mut wanted_conf =
+                attached_location_conf(self.generation, &self.shard, &self.config);
+            match self.observed.locations.get(&node_id) {
+                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
+                    // Nothing to do
+                    tracing::info!("Observed configuration already correct.")
+                }
+                Some(_) | None => {
+                    // If there is no observed configuration, or if its value does not equal our intent, then we must call out to the pageserver.
+                    self.generation = self.generation.next();
+                    wanted_conf.generation = self.generation.into();
+                    tracing::info!("Observed configuration requires update.");
+                    self.location_config(node_id, wanted_conf).await?;
+                    if let Err(e) = self
+                        .compute_hook
+                        .notify(self.tenant_shard_id, node_id)
+                        .await
+                    {
+                        tracing::warn!(
+                            "Failed to notify compute of newly attached pageserver {node_id}: {e}"
+                        );
+                    }
+                }
+            }
+        }
+
+        // Configure secondary locations: if these were previously attached this
+        // implicitly downgrades them from attached to secondary.
+        let mut changes = Vec::new();
+        for node_id in &self.intent.secondary {
+            let wanted_conf = secondary_location_conf(&self.shard, &self.config);
+            match self.observed.locations.get(node_id) {
+                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
+                    // Nothing to do
+                    tracing::info!(%node_id, "Observed configuration already correct.")
+                }
+                Some(_) | None => {
+                    // If there is no observed configuration, or if its value does not equal our intent, then we must call out to the pageserver.
+                    tracing::info!(%node_id, "Observed configuration requires update.");
+                    changes.push((*node_id, wanted_conf))
+                }
+            }
+        }
+
+        // Detach any extraneous pageservers that are no longer referenced
+        // by our intent.
+        let all_pageservers = self.intent.all_pageservers();
+        for node_id in self.observed.locations.keys() {
+            if all_pageservers.contains(node_id) {
+                // We are only detaching pageservers that aren't used at all.
+                continue;
+            }
+
+            changes.push((
+                *node_id,
+                LocationConfig {
+                    mode: LocationConfigMode::Detached,
+                    generation: None,
+                    secondary_conf: None,
+                    shard_number: self.shard.number.0,
+                    shard_count: self.shard.count.0,
+                    shard_stripe_size: self.shard.stripe_size.0,
+                    tenant_conf: self.config.clone(),
+                },
+            ));
+        }
+
+        for (node_id, conf) in changes {
+            self.location_config(node_id, conf).await?;
+        }
+
+        Ok(())
+    }
+}
+
+pub(crate) fn attached_location_conf(
+    generation: Generation,
+    shard: &ShardIdentity,
+    config: &TenantConfig,
+) -> LocationConfig {
+    LocationConfig {
+        mode: LocationConfigMode::AttachedSingle,
+        generation: generation.into(),
+        secondary_conf: None,
+        shard_number: shard.number.0,
+        shard_count: shard.count.0,
+        shard_stripe_size: shard.stripe_size.0,
+        tenant_conf: config.clone(),
+    }
+}
+
+pub(crate) fn secondary_location_conf(
+    shard: &ShardIdentity,
+    config: &TenantConfig,
+) -> LocationConfig {
+    LocationConfig {
+        mode: LocationConfigMode::Secondary,
+        generation: None,
+        secondary_conf: Some(LocationConfigSecondary { warm: true }),
+        shard_number: shard.number.0,
+        shard_count: shard.count.0,
+        shard_stripe_size: shard.stripe_size.0,
+        tenant_conf: config.clone(),
+    }
+}

control_plane/attachment_service/src/scheduler.rs

@@ -0,0 +1,89 @@
+use pageserver_api::shard::TenantShardId;
+use std::collections::{BTreeMap, HashMap};
+use utils::{http::error::ApiError, id::NodeId};
+
+use crate::{node::Node, tenant_state::TenantState};
+
+/// Scenarios in which we cannot find a suitable location for a tenant shard
+#[derive(thiserror::Error, Debug)]
+pub enum ScheduleError {
+    #[error("No pageservers found")]
+    NoPageservers,
+    #[error("No pageserver found matching constraint")]
+    ImpossibleConstraint,
+}
+
+impl From<ScheduleError> for ApiError {
+    fn from(value: ScheduleError) -> Self {
+        ApiError::Conflict(format!("Scheduling error: {}", value))
+    }
+}
+
+pub(crate) struct Scheduler {
+    tenant_counts: HashMap<NodeId, usize>,
+}
+
+impl Scheduler {
+    pub(crate) fn new(
+        tenants: &BTreeMap<TenantShardId, TenantState>,
+        nodes: &HashMap<NodeId, Node>,
+    ) -> Self {
+        let mut tenant_counts = HashMap::new();
+        for node_id in nodes.keys() {
+            tenant_counts.insert(*node_id, 0);
+        }
+
+        for tenant in tenants.values() {
+            if let Some(ps) = tenant.intent.attached {
+                let entry = tenant_counts.entry(ps).or_insert(0);
+                *entry += 1;
+            }
+        }
+
+        for (node_id, node) in nodes {
+            if !node.may_schedule() {
+                tenant_counts.remove(node_id);
+            }
+        }
+
+        Self { tenant_counts }
+    }
+
+    pub(crate) fn schedule_shard(
+        &mut self,
+        hard_exclude: &[NodeId],
+    ) -> Result<NodeId, ScheduleError> {
+        if self.tenant_counts.is_empty() {
+            return Err(ScheduleError::NoPageservers);
+        }
+
+        let mut tenant_counts: Vec<(NodeId, usize)> = self
+            .tenant_counts
+            .iter()
+            .filter_map(|(k, v)| {
+                if hard_exclude.contains(k) {
+                    None
+                } else {
+                    Some((*k, *v))
+                }
+            })
+            .collect();
+
+        // Sort by tenant count.  Nodes with the same tenant count are sorted by ID.
+        tenant_counts.sort_by_key(|i| (i.1, i.0));
+
+        if tenant_counts.is_empty() {
+            // After applying constraints, no pageservers were left
+            return Err(ScheduleError::ImpossibleConstraint);
+        }
+
+        for (node_id, count) in &tenant_counts {
+            tracing::info!("tenant_counts[{node_id}]={count}");
+        }
+
+        let node_id = tenant_counts.first().unwrap().0;
+        tracing::info!("scheduler selected node {node_id}");
+        *self.tenant_counts.get_mut(&node_id).unwrap() += 1;
+        Ok(node_id)
+    }
+}

control_plane/attachment_service/src/service.rs

@@ -0,0 +1,922 @@
+use std::{
+    collections::{BTreeMap, HashMap},
+    sync::Arc,
+    time::{Duration, Instant},
+};
+
+use control_plane::attachment_service::{
+    AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse, NodeAvailability,
+    NodeConfigureRequest, NodeRegisterRequest, NodeSchedulingPolicy, TenantCreateResponse,
+    TenantCreateResponseShard, TenantLocateResponse, TenantLocateResponseShard,
+    TenantShardMigrateRequest, TenantShardMigrateResponse,
+};
+use hyper::Method;
+use pageserver_api::{
+    control_api::{
+        ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
+        ValidateResponse, ValidateResponseTenant,
+    },
+    models::{
+        ShardParameters, TenantCreateRequest, TenantShardSplitRequest, TenantShardSplitResponse,
+        TimelineCreateRequest, TimelineInfo,
+    },
+    shard::{ShardCount, ShardIdentity, ShardNumber, TenantShardId},
+};
+use reqwest::Client;
+use utils::{
+    generation::Generation,
+    http::error::ApiError,
+    id::{NodeId, TenantId},
+};
+
+use crate::{
+    compute_hook::ComputeHook,
+    node::Node,
+    reconciler::attached_location_conf,
+    scheduler::Scheduler,
+    tenant_state::{
+        IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcilerWaiter,
+        TenantState,
+    },
+    PlacementPolicy,
+};
+
+const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
+
+// Top level state available to all HTTP handlers
+struct ServiceState {
+    tenants: BTreeMap<TenantShardId, TenantState>,
+
+    nodes: Arc<HashMap<NodeId, Node>>,
+
+    compute_hook: Arc<ComputeHook>,
+
+    result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
+}
+
+impl ServiceState {
+    fn new(result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>) -> Self {
+        Self {
+            tenants: BTreeMap::new(),
+            nodes: Arc::new(HashMap::new()),
+            compute_hook: Arc::new(ComputeHook::new()),
+            result_tx,
+        }
+    }
+}
+
+pub struct Service {
+    inner: Arc<std::sync::RwLock<ServiceState>>,
+}
+
+impl Service {
+    pub fn spawn() -> Arc<Self> {
+        let (result_tx, mut result_rx) = tokio::sync::mpsc::unbounded_channel();
+
+        let this = Arc::new(Self {
+            inner: Arc::new(std::sync::RwLock::new(ServiceState::new(result_tx))),
+        });
+
+        let result_task_this = this.clone();
+        tokio::task::spawn(async move {
+            while let Some(result) = result_rx.recv().await {
+                tracing::info!(
+                    "Reconcile result for sequence {}, ok={}",
+                    result.sequence,
+                    result.result.is_ok()
+                );
+                let mut locked = result_task_this.inner.write().unwrap();
+                if let Some(tenant) = locked.tenants.get_mut(&result.tenant_shard_id) {
+                    tenant.generation = result.generation;
+                    match result.result {
+                        Ok(()) => {
+                            for (node_id, loc) in &result.observed.locations {
+                                if let Some(conf) = &loc.conf {
+                                    tracing::info!(
+                                        "Updating observed location {}: {:?}",
+                                        node_id,
+                                        conf
+                                    );
+                                } else {
+                                    tracing::info!("Setting observed location {} to None", node_id,)
+                                }
+                            }
+                            tenant.observed = result.observed;
+                            tenant.waiter.advance(result.sequence);
+                        }
+                        Err(e) => {
+                            // TODO: some observability, record on teh tenant its last reconcile error
+                            tracing::warn!(
+                                "Reconcile error on tenant {}: {}",
+                                tenant.tenant_shard_id,
+                                e
+                            );
+
+                            for (node_id, o) in result.observed.locations {
+                                tenant.observed.locations.insert(node_id, o);
+                            }
+                        }
+                    }
+                }
+            }
+        });
+
+        this
+    }
+
+    pub(crate) fn attach_hook(&self, attach_req: AttachHookRequest) -> AttachHookResponse {
+        let mut locked = self.inner.write().unwrap();
+
+        let tenant_state = locked
+            .tenants
+            .entry(attach_req.tenant_shard_id)
+            .or_insert_with(|| {
+                TenantState::new(
+                    attach_req.tenant_shard_id,
+                    ShardIdentity::unsharded(),
+                    PlacementPolicy::Single,
+                )
+            });
+
+        if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
+            tenant_state.generation = tenant_state.generation.next();
+            tracing::info!(
+                tenant_id = %attach_req.tenant_shard_id,
+                ps_id = %attaching_pageserver,
+                generation = ?tenant_state.generation,
+                "issuing",
+            );
+        } else if let Some(ps_id) = tenant_state.intent.attached {
+            tracing::info!(
+                tenant_id = %attach_req.tenant_shard_id,
+                %ps_id,
+                generation = ?tenant_state.generation,
+                "dropping",
+            );
+        } else {
+            tracing::info!(
+            tenant_id = %attach_req.tenant_shard_id,
+            "no-op: tenant already has no pageserver");
+        }
+        tenant_state.intent.attached = attach_req.node_id;
+        let generation = tenant_state.generation;
+
+        tracing::info!(
+            "attach_hook: tenant {} set generation {:?}, pageserver {}",
+            attach_req.tenant_shard_id,
+            tenant_state.generation,
+            attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
+        );
+
+        AttachHookResponse {
+            gen: attach_req.node_id.map(|_| generation.into().unwrap()),
+        }
+    }
+
+    pub(crate) fn inspect(&self, inspect_req: InspectRequest) -> InspectResponse {
+        let locked = self.inner.read().unwrap();
+
+        let tenant_state = locked.tenants.get(&inspect_req.tenant_shard_id);
+
+        InspectResponse {
+            attachment: tenant_state.and_then(|s| {
+                s.intent
+                    .attached
+                    .map(|ps| (s.generation.into().unwrap(), ps))
+            }),
+        }
+    }
+
+    pub(crate) fn re_attach(&self, reattach_req: ReAttachRequest) -> ReAttachResponse {
+        let mut locked = self.inner.write().unwrap();
+
+        let mut response = ReAttachResponse {
+            tenants: Vec::new(),
+        };
+        for (t, state) in &mut locked.tenants {
+            if state.intent.attached == Some(reattach_req.node_id) {
+                state.generation = state.generation.next();
+                response.tenants.push(ReAttachResponseTenant {
+                    id: *t,
+                    gen: state.generation.into().unwrap(),
+                });
+            }
+        }
+        response
+    }
+
+    pub(crate) fn validate(&self, validate_req: ValidateRequest) -> ValidateResponse {
+        let locked = self.inner.read().unwrap();
+
+        let mut response = ValidateResponse {
+            tenants: Vec::new(),
+        };
+
+        for req_tenant in validate_req.tenants {
+            if let Some(tenant_state) = locked.tenants.get(&req_tenant.id) {
+                let valid = tenant_state.generation == Generation::new(req_tenant.gen);
+                tracing::info!(
+                    "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
+                    req_tenant.id,
+                    req_tenant.gen,
+                    tenant_state.generation
+                );
+                response.tenants.push(ValidateResponseTenant {
+                    id: req_tenant.id,
+                    valid,
+                });
+            }
+        }
+        response
+    }
+
+    pub(crate) async fn tenant_create(
+        &self,
+        create_req: TenantCreateRequest,
+    ) -> Result<TenantCreateResponse, ApiError> {
+        let (waiters, response_shards) = {
+            let mut locked = self.inner.write().unwrap();
+
+            tracing::info!(
+                "Creating tenant {}, shard_count={:?}, have {} pageservers",
+                create_req.new_tenant_id,
+                create_req.shard_parameters.count,
+                locked.nodes.len()
+            );
+
+            // This service expects to handle sharding itself: it is an error to try and directly create
+            // a particular shard here.
+            let tenant_id = if create_req.new_tenant_id.shard_count > ShardCount(1) {
+                return Err(ApiError::BadRequest(anyhow::anyhow!(
+                "Attempted to create a specific shard, this API is for creating the whole tenant"
+            )));
+            } else {
+                create_req.new_tenant_id.tenant_id
+            };
+
+            // Shard count 0 is valid: it means create a single shard (ShardCount(0) means "unsharded")
+            let literal_shard_count = if create_req.shard_parameters.is_unsharded() {
+                1
+            } else {
+                create_req.shard_parameters.count.0
+            };
+
+            let mut response_shards = Vec::new();
+
+            let mut scheduler = Scheduler::new(&locked.tenants, &locked.nodes);
+
+            for i in 0..literal_shard_count {
+                let shard_number = ShardNumber(i);
+
+                let tenant_shard_id = TenantShardId {
+                    tenant_id,
+                    shard_number,
+                    shard_count: create_req.shard_parameters.count,
+                };
+                tracing::info!("Creating shard {tenant_shard_id}...");
+
+                use std::collections::btree_map::Entry;
+                match locked.tenants.entry(tenant_shard_id) {
+                    Entry::Occupied(mut entry) => {
+                        tracing::info!(
+                            "Tenant shard {tenant_shard_id} already exists while creating"
+                        );
+
+                        // TODO: schedule() should take an anti-affinity expression that pushes
+                        // attached and secondary locations (independently) away frorm those
+                        // pageservers also holding a shard for this tenant.
+
+                        entry.get_mut().schedule(&mut scheduler).map_err(|e| {
+                            ApiError::Conflict(format!(
+                                "Failed to schedule shard {tenant_shard_id}: {e}"
+                            ))
+                        })?;
+
+                        response_shards.push(TenantCreateResponseShard {
+                            node_id: entry
+                                .get()
+                                .intent
+                                .attached
+                                .expect("We just set pageserver if it was None"),
+                            generation: entry.get().generation.into().unwrap(),
+                        });
+
+                        continue;
+                    }
+                    Entry::Vacant(entry) => {
+                        let mut state = TenantState::new(
+                            tenant_shard_id,
+                            ShardIdentity::from_params(shard_number, &create_req.shard_parameters),
+                            PlacementPolicy::Single,
+                        );
+
+                        if let Some(create_gen) = create_req.generation {
+                            state.generation = Generation::new(create_gen);
+                        }
+                        state.config = create_req.config.clone();
+
+                        state.schedule(&mut scheduler).map_err(|e| {
+                            ApiError::Conflict(format!(
+                                "Failed to schedule shard {tenant_shard_id}: {e}"
+                            ))
+                        })?;
+
+                        response_shards.push(TenantCreateResponseShard {
+                            node_id: state
+                                .intent
+                                .attached
+                                .expect("We just set pageserver if it was None"),
+                            generation: state.generation.into().unwrap(),
+                        });
+                        entry.insert(state)
+                    }
+                };
+            }
+
+            // Take a snapshot of pageservers
+            let pageservers = locked.nodes.clone();
+
+            let mut waiters = Vec::new();
+            let result_tx = locked.result_tx.clone();
+            let compute_hook = locked.compute_hook.clone();
+
+            for (_tenant_shard_id, shard) in locked
+                .tenants
+                .range_mut(TenantShardId::tenant_range(tenant_id))
+            {
+                if let Some(waiter) =
+                    shard.maybe_reconcile(result_tx.clone(), &pageservers, &compute_hook)
+                {
+                    waiters.push(waiter);
+                }
+            }
+            (waiters, response_shards)
+        };
+
+        let deadline = Instant::now().checked_add(Duration::from_secs(5)).unwrap();
+        for waiter in waiters {
+            let timeout = deadline.duration_since(Instant::now());
+            waiter.wait_timeout(timeout).await.map_err(|_| {
+                ApiError::Timeout(
+                    format!(
+                        "Timeout waiting for reconciliation of tenant shard {}",
+                        waiter.tenant_shard_id
+                    )
+                    .into(),
+                )
+            })?;
+        }
+        Ok(TenantCreateResponse {
+            shards: response_shards,
+        })
+    }
+
+    pub(crate) async fn tenant_timeline_create(
+        &self,
+        tenant_id: TenantId,
+        mut create_req: TimelineCreateRequest,
+    ) -> Result<TimelineInfo, ApiError> {
+        let mut timeline_info = None;
+        let ensure_waiters = {
+            let locked = self.inner.write().unwrap();
+
+            tracing::info!(
+                "Creating timeline {}/{}, have {} pageservers",
+                tenant_id,
+                create_req.new_timeline_id,
+                locked.nodes.len()
+            );
+
+            ensure_attached(locked, tenant_id).map_err(ApiError::InternalServerError)?
+        };
+
+        let deadline = Instant::now().checked_add(Duration::from_secs(5)).unwrap();
+        for waiter in ensure_waiters {
+            let timeout = deadline.duration_since(Instant::now());
+            waiter.wait_timeout(timeout).await.map_err(|_| {
+                ApiError::Timeout(
+                    format!(
+                        "Timeout waiting for reconciliation of tenant shard {}",
+                        waiter.tenant_shard_id
+                    )
+                    .into(),
+                )
+            })?;
+        }
+
+        let targets = {
+            let locked = self.inner.read().unwrap();
+            let mut targets = Vec::new();
+
+            for (tenant_shard_id, shard) in
+                locked.tenants.range(TenantShardId::tenant_range(tenant_id))
+            {
+                let node_id = shard.intent.attached.ok_or_else(|| {
+                    ApiError::InternalServerError(anyhow::anyhow!("Shard not scheduled"))
+                })?;
+                let node = locked
+                    .nodes
+                    .get(&node_id)
+                    .expect("Pageservers may not be deleted while referenced");
+
+                targets.push((*tenant_shard_id, node.clone()));
+            }
+            targets
+        };
+
+        if targets.is_empty() {
+            return Err(ApiError::NotFound(
+                anyhow::anyhow!("Tenant not found").into(),
+            ));
+        }
+
+        for (tenant_shard_id, node) in targets {
+            // TODO: issue shard timeline creates in parallel, once the 0th is done.
+
+            let shard_timeline_info = timeline_create(&tenant_shard_id, &node, &create_req)
+                .await
+                .map_err(|e| ApiError::Conflict(format!("Failed to create timeline: {e}")))?;
+
+            if timeline_info.is_none() {
+                // If the caller specified an ancestor but no ancestor LSN, we are responsible for
+                // propagating the LSN chosen by the first shard to the other shards: it is important
+                // that all shards end up with the same ancestor_start_lsn.
+                if create_req.ancestor_timeline_id.is_some()
+                    && create_req.ancestor_start_lsn.is_none()
+                {
+                    create_req.ancestor_start_lsn = shard_timeline_info.ancestor_lsn;
+                }
+
+                // We will return the TimelineInfo from the first shard
+                timeline_info = Some(shard_timeline_info);
+            }
+        }
+        Ok(timeline_info.expect("targets cannot be empty"))
+    }
+
+    pub(crate) fn tenant_locate(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<TenantLocateResponse, ApiError> {
+        let locked = self.inner.read().unwrap();
+        tracing::info!("Locating shards for tenant {tenant_id}");
+
+        // Take a snapshot of pageservers
+        let pageservers = locked.nodes.clone();
+
+        let mut result = Vec::new();
+        let mut shard_params: Option<ShardParameters> = None;
+
+        for (tenant_shard_id, shard) in locked.tenants.range(TenantShardId::tenant_range(tenant_id))
+        {
+            let node_id = shard
+                .intent
+                .attached
+                .ok_or(ApiError::BadRequest(anyhow::anyhow!(
+                    "Cannot locate a tenant that is not attached"
+                )))?;
+
+            let node = pageservers
+                .get(&node_id)
+                .expect("Pageservers may not be deleted while referenced");
+
+            result.push(TenantLocateResponseShard {
+                shard_id: *tenant_shard_id,
+                node_id,
+                listen_http_addr: node.listen_http_addr.clone(),
+                listen_http_port: node.listen_http_port,
+                listen_pg_addr: node.listen_pg_addr.clone(),
+                listen_pg_port: node.listen_pg_port,
+            });
+
+            match &shard_params {
+                None => {
+                    shard_params = Some(ShardParameters {
+                        stripe_size: Some(shard.shard.stripe_size),
+                        count: shard.shard.count,
+                    });
+                }
+                Some(params) => {
+                    if params.stripe_size != Some(shard.shard.stripe_size) {
+                        // This should never happen.  We enforce at runtime because it's simpler than
+                        // adding an extra per-tenant data structure to store the things that should be the same
+                        return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                            "Inconsistent shard stripe size parameters!"
+                        )));
+                    }
+                }
+            }
+        }
+
+        if result.is_empty() {
+            return Err(ApiError::NotFound(
+                anyhow::anyhow!("No shards for this tenant ID found").into(),
+            ));
+        }
+        let shard_params = shard_params.expect("result is non-empty, therefore this is set");
+        tracing::info!(
+            "Located tenant {} with params {:?} on shards {}",
+            tenant_id,
+            shard_params,
+            result
+                .iter()
+                .map(|s| format!("{:?}", s))
+                .collect::<Vec<_>>()
+                .join(",")
+        );
+
+        Ok(TenantLocateResponse {
+            shards: result,
+            shard_params,
+        })
+    }
+
+    pub(crate) async fn tenant_shard_split(
+        &self,
+        tenant_id: TenantId,
+        split_req: TenantShardSplitRequest,
+    ) -> Result<TenantShardSplitResponse, ApiError> {
+        let mut policy = None;
+        let (targets, compute_hook) = {
+            let mut locked = self.inner.write().unwrap();
+
+            let pageservers = locked.nodes.clone();
+
+            let mut targets = Vec::new();
+
+            for (tenant_shard_id, shard) in locked
+                .tenants
+                .range_mut(TenantShardId::tenant_range(tenant_id))
+            {
+                if policy.is_none() {
+                    policy = Some(shard.policy.clone());
+                }
+
+                if tenant_shard_id.shard_count == ShardCount(split_req.new_shard_count) {
+                    tracing::warn!(
+                        "Tenant shard {} already has shard count {}",
+                        tenant_shard_id,
+                        split_req.new_shard_count
+                    );
+                    continue;
+                }
+
+                let node_id =
+                    shard
+                        .intent
+                        .attached
+                        .ok_or(ApiError::BadRequest(anyhow::anyhow!(
+                            "Cannot split a tenant that is not attached"
+                        )))?;
+
+                let node = pageservers
+                    .get(&node_id)
+                    .expect("Pageservers may not be deleted while referenced");
+
+                // TODO: if any reconciliation is currently in progress for this shard, wait for it.
+
+                targets.push((*tenant_shard_id, node.clone()));
+            }
+            (targets, locked.compute_hook.clone())
+        };
+
+        let mut replacements = HashMap::new();
+        for (tenant_shard_id, node) in targets {
+            let client = Client::new();
+            let response = client
+                .request(
+                    Method::PUT,
+                    format!("{}/tenant/{}/shard_split", node.base_url(), tenant_shard_id),
+                )
+                .json(&TenantShardSplitRequest {
+                    new_shard_count: split_req.new_shard_count,
+                })
+                .send()
+                .await
+                .map_err(|e| {
+                    ApiError::Conflict(format!("Failed to split {}: {}", tenant_shard_id, e))
+                })?;
+            response.error_for_status_ref().map_err(|e| {
+                ApiError::Conflict(format!("Failed to split {}: {}", tenant_shard_id, e))
+            })?;
+            let response: TenantShardSplitResponse = response.json().await.map_err(|e| {
+                ApiError::InternalServerError(anyhow::anyhow!(
+                    "Malformed response from pageserver: {}",
+                    e
+                ))
+            })?;
+
+            tracing::info!(
+                "Split {} into {}",
+                tenant_shard_id,
+                response
+                    .new_shards
+                    .iter()
+                    .map(|s| format!("{:?}", s))
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
+
+            replacements.insert(tenant_shard_id, response.new_shards);
+        }
+
+        // TODO: concurrency: we're dropping the state lock while issuing split API calls.
+        //       We should add some marker to the TenantState that causes any other change
+        //       to refuse until the split is complete.  This will be related to a persistent
+        //       splitting marker that will ensure resume after crash.
+
+        // Replace all the shards we just split with their children
+        let mut response = TenantShardSplitResponse {
+            new_shards: Vec::new(),
+        };
+        let mut child_locations = Vec::new();
+        {
+            let mut locked = self.inner.write().unwrap();
+            for (replaced, children) in replacements.into_iter() {
+                let (pageserver, generation, shard_ident, config) = {
+                    let old_state = locked
+                        .tenants
+                        .remove(&replaced)
+                        .expect("It was present, we just split it");
+                    (
+                        old_state.intent.attached.unwrap(),
+                        old_state.generation,
+                        old_state.shard,
+                        old_state.config.clone(),
+                    )
+                };
+
+                locked.tenants.remove(&replaced);
+
+                for child in children {
+                    let mut child_shard = shard_ident;
+                    child_shard.number = child.shard_number;
+                    child_shard.count = child.shard_count;
+
+                    let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
+                    child_observed.insert(
+                        pageserver,
+                        ObservedStateLocation {
+                            conf: Some(attached_location_conf(generation, &child_shard, &config)),
+                        },
+                    );
+
+                    let mut child_state = TenantState::new(
+                        child,
+                        child_shard,
+                        policy
+                            .clone()
+                            .expect("We set this if any replacements are pushed"),
+                    );
+                    child_state.intent = IntentState::single(Some(pageserver));
+                    child_state.observed = ObservedState {
+                        locations: child_observed,
+                    };
+                    child_state.generation = generation;
+                    child_state.config = config.clone();
+
+                    child_locations.push((child, pageserver));
+
+                    locked.tenants.insert(child, child_state);
+                    response.new_shards.push(child);
+                }
+            }
+        }
+
+        for (child_id, child_ps) in child_locations {
+            if let Err(e) = compute_hook.notify(child_id, child_ps).await {
+                tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
+                        child_id, child_ps);
+            }
+        }
+        Ok(response)
+    }
+
+    pub(crate) async fn tenant_shard_migrate(
+        &self,
+        tenant_shard_id: TenantShardId,
+        migrate_req: TenantShardMigrateRequest,
+    ) -> Result<TenantShardMigrateResponse, ApiError> {
+        let waiter = {
+            let mut locked = self.inner.write().unwrap();
+
+            let result_tx = locked.result_tx.clone();
+            let pageservers = locked.nodes.clone();
+            let compute_hook = locked.compute_hook.clone();
+
+            let Some(shard) = locked.tenants.get_mut(&tenant_shard_id) else {
+                return Err(ApiError::NotFound(
+                    anyhow::anyhow!("Tenant shard not found").into(),
+                ));
+            };
+
+            if shard.intent.attached == Some(migrate_req.node_id) {
+                // No-op case: we will still proceed to wait for reconciliation in case it is
+                // incomplete from an earlier update to the intent.
+                tracing::info!("Migrating: intent is unchanged {:?}", shard.intent);
+            } else {
+                let old_attached = shard.intent.attached;
+
+                shard.intent.attached = Some(migrate_req.node_id);
+                match shard.policy {
+                    PlacementPolicy::Single => {
+                        shard.intent.secondary.clear();
+                    }
+                    PlacementPolicy::Double(_n) => {
+                        // If our new attached node was a secondary, it no longer should be.
+                        shard.intent.secondary.retain(|s| s != &migrate_req.node_id);
+
+                        // If we were already attached to something, demote that to a secondary
+                        if let Some(old_attached) = old_attached {
+                            shard.intent.secondary.push(old_attached);
+                        }
+                    }
+                }
+
+                tracing::info!("Migrating: new intent {:?}", shard.intent);
+                shard.sequence = shard.sequence.next();
+            }
+
+            shard.maybe_reconcile(result_tx, &pageservers, &compute_hook)
+        };
+
+        if let Some(waiter) = waiter {
+            waiter
+                .wait_timeout(RECONCILE_TIMEOUT)
+                .await
+                .map_err(|e| ApiError::Timeout(format!("{}", e).into()))?;
+        } else {
+            tracing::warn!("Migration is a no-op");
+        }
+
+        Ok(TenantShardMigrateResponse {})
+    }
+
+    pub(crate) fn node_register(&self, register_req: NodeRegisterRequest) {
+        let mut locked = self.inner.write().unwrap();
+
+        let mut new_nodes = (*locked.nodes).clone();
+
+        new_nodes.insert(
+            register_req.node_id,
+            Node {
+                id: register_req.node_id,
+                listen_http_addr: register_req.listen_http_addr,
+                listen_http_port: register_req.listen_http_port,
+                listen_pg_addr: register_req.listen_pg_addr,
+                listen_pg_port: register_req.listen_pg_port,
+                scheduling: NodeSchedulingPolicy::Filling,
+                // TODO: we shouldn't really call this Active until we've heartbeated it.
+                availability: NodeAvailability::Active,
+            },
+        );
+
+        locked.nodes = Arc::new(new_nodes);
+
+        tracing::info!(
+            "Registered pageserver {}, now have {} pageservers",
+            register_req.node_id,
+            locked.nodes.len()
+        );
+    }
+
+    pub(crate) fn node_configure(&self, config_req: NodeConfigureRequest) -> Result<(), ApiError> {
+        let mut locked = self.inner.write().unwrap();
+        let result_tx = locked.result_tx.clone();
+        let compute_hook = locked.compute_hook.clone();
+
+        let mut new_nodes = (*locked.nodes).clone();
+
+        let Some(node) = new_nodes.get_mut(&config_req.node_id) else {
+            return Err(ApiError::NotFound(
+                anyhow::anyhow!("Node not registered").into(),
+            ));
+        };
+
+        let mut offline_transition = false;
+        let mut active_transition = false;
+
+        if let Some(availability) = &config_req.availability {
+            match (availability, &node.availability) {
+                (NodeAvailability::Offline, NodeAvailability::Active) => {
+                    tracing::info!("Node {} transition to offline", config_req.node_id);
+                    offline_transition = true;
+                }
+                (NodeAvailability::Active, NodeAvailability::Offline) => {
+                    tracing::info!("Node {} transition to active", config_req.node_id);
+                    active_transition = true;
+                }
+                _ => {
+                    tracing::info!("Node {} no change during config", config_req.node_id);
+                    // No change
+                }
+            };
+            node.availability = *availability;
+        }
+
+        if let Some(scheduling) = config_req.scheduling {
+            node.scheduling = scheduling;
+
+            // TODO: once we have a background scheduling ticker for fill/drain, kick it
+            // to wake up and start working.
+        }
+
+        let new_nodes = Arc::new(new_nodes);
+
+        let mut scheduler = Scheduler::new(&locked.tenants, &new_nodes);
+        if offline_transition {
+            for (tenant_shard_id, tenant_state) in &mut locked.tenants {
+                if let Some(observed_loc) =
+                    tenant_state.observed.locations.get_mut(&config_req.node_id)
+                {
+                    // When a node goes offline, we set its observed configuration to None, indicating unknown: we will
+                    // not assume our knowledge of the node's configuration is accurate until it comes back online
+                    observed_loc.conf = None;
+                }
+
+                if tenant_state.intent.notify_offline(config_req.node_id) {
+                    tenant_state.sequence = tenant_state.sequence.next();
+                    match tenant_state.schedule(&mut scheduler) {
+                        Err(e) => {
+                            // It is possible that some tenants will become unschedulable when too many pageservers
+                            // go offline: in this case there isn't much we can do other than make the issue observable.
+                            // TODO: give TenantState a scheduling error attribute to be queried later.
+                            tracing::warn!(%tenant_shard_id, "Scheduling error when marking pageserver {} offline: {e}", config_req.node_id);
+                        }
+                        Ok(()) => {
+                            tenant_state.maybe_reconcile(
+                                result_tx.clone(),
+                                &new_nodes,
+                                &compute_hook,
+                            );
+                        }
+                    }
+                }
+            }
+        }
+
+        if active_transition {
+            // When a node comes back online, we must reconcile any tenant that has a None observed
+            // location on the node.
+            for tenant_state in locked.tenants.values_mut() {
+                if let Some(observed_loc) =
+                    tenant_state.observed.locations.get_mut(&config_req.node_id)
+                {
+                    if observed_loc.conf.is_none() {
+                        tenant_state.maybe_reconcile(result_tx.clone(), &new_nodes, &compute_hook);
+                    }
+                }
+            }
+
+            // TODO: in the background, we should balance work back onto this pageserver
+        }
+
+        locked.nodes = new_nodes;
+
+        Ok(())
+    }
+}
+
+// TODO: use mgmt_api client
+async fn timeline_create(
+    tenant_shard_id: &TenantShardId,
+    node: &Node,
+    req: &TimelineCreateRequest,
+) -> anyhow::Result<TimelineInfo> {
+    let client = Client::new();
+    let response = client
+        .request(
+            Method::POST,
+            format!("{}/tenant/{}/timeline", node.base_url(), tenant_shard_id),
+        )
+        .json(req)
+        .send()
+        .await?;
+    response.error_for_status_ref()?;
+
+    Ok(response.json().await?)
+}
+
+fn ensure_attached(
+    mut locked: std::sync::RwLockWriteGuard<'_, ServiceState>,
+    tenant_id: TenantId,
+) -> Result<Vec<ReconcilerWaiter>, anyhow::Error> {
+    let mut waiters = Vec::new();
+    let result_tx = locked.result_tx.clone();
+    let compute_hook = locked.compute_hook.clone();
+    let mut scheduler = Scheduler::new(&locked.tenants, &locked.nodes);
+    let pageservers = locked.nodes.clone();
+
+    for (_tenant_shard_id, shard) in locked
+        .tenants
+        .range_mut(TenantShardId::tenant_range(tenant_id))
+    {
+        shard.schedule(&mut scheduler)?;
+
+        if let Some(waiter) = shard.maybe_reconcile(result_tx.clone(), &pageservers, &compute_hook)
+        {
+            waiters.push(waiter);
+        }
+    }
+    Ok(waiters)
+}

control_plane/attachment_service/src/tenant_state.rs

@@ -0,0 +1,351 @@
+use std::{collections::HashMap, sync::Arc, time::Duration};
+
+use control_plane::attachment_service::NodeAvailability;
+use pageserver_api::{
+    models::{LocationConfig, TenantConfig},
+    shard::{ShardIdentity, TenantShardId},
+};
+use tokio::task::JoinHandle;
+use tokio_util::sync::CancellationToken;
+use utils::{
+    generation::Generation,
+    id::NodeId,
+    seqwait::{SeqWait, SeqWaitError},
+};
+
+use crate::{
+    compute_hook::ComputeHook,
+    node::Node,
+    reconciler::{attached_location_conf, secondary_location_conf, ReconcileError, Reconciler},
+    scheduler::{ScheduleError, Scheduler},
+    PlacementPolicy, Sequence,
+};
+
+pub(crate) struct TenantState {
+    pub(crate) tenant_shard_id: TenantShardId,
+
+    pub(crate) shard: ShardIdentity,
+
+    pub(crate) sequence: Sequence,
+
+    // Latest generation number: next time we attach, increment this
+    // and use the incremented number when attaching
+    pub(crate) generation: Generation,
+
+    // High level description of how the tenant should be set up.  Provided
+    // externally.
+    pub(crate) policy: PlacementPolicy,
+
+    // Low level description of exactly which pageservers should fulfil
+    // which role.  Generated by `Self::schedule`.
+    pub(crate) intent: IntentState,
+
+    // Low level description of how the tenant is configured on pageservers:
+    // if this does not match `Self::intent` then the tenant needs reconciliation
+    // with `Self::reconcile`.
+    pub(crate) observed: ObservedState,
+
+    pub(crate) config: TenantConfig,
+
+    /// If a reconcile task is currently in flight, it may be joined here (it is
+    /// only safe to join if either the result has been received or the reconciler's
+    /// cancellation token has been fired)
+    pub(crate) reconciler: Option<ReconcilerHandle>,
+
+    /// Optionally wait for reconciliation to complete up to a particular
+    /// sequence number.
+    pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
+}
+
+#[derive(Default, Clone, Debug)]
+pub(crate) struct IntentState {
+    pub(crate) attached: Option<NodeId>,
+    pub(crate) secondary: Vec<NodeId>,
+}
+
+#[derive(Default, Clone)]
+pub(crate) struct ObservedState {
+    pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
+}
+
+/// Our latest knowledge of how this tenant is configured in the outside world.
+///
+/// Meaning:
+///     * No instance of this type exists for a node: we are certain that we have nothing configured on that
+///       node for this shard.
+///     * Instance exists with conf==None: we *might* have some state on that node, but we don't know
+///       what it is (e.g. we failed partway through configuring it)
+///     * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
+///       and that configuration will still be present unless something external interfered.
+#[derive(Clone)]
+pub(crate) struct ObservedStateLocation {
+    /// If None, it means we do not know the status of this shard's location on this node, but
+    /// we know that we might have some state on this node.
+    pub(crate) conf: Option<LocationConfig>,
+}
+pub(crate) struct ReconcilerWaiter {
+    // For observability purposes, remember the ID of the shard we're
+    // waiting for.
+    pub(crate) tenant_shard_id: TenantShardId,
+
+    seq_wait: std::sync::Arc<SeqWait<Sequence, Sequence>>,
+    seq: Sequence,
+}
+
+impl ReconcilerWaiter {
+    pub(crate) async fn wait_timeout(&self, timeout: Duration) -> Result<(), SeqWaitError> {
+        self.seq_wait.wait_for_timeout(self.seq, timeout).await
+    }
+}
+
+/// Having spawned a reconciler task, the tenant shard's state will carry enough
+/// information to optionally cancel & await it later.
+pub(crate) struct ReconcilerHandle {
+    sequence: Sequence,
+    handle: JoinHandle<()>,
+    cancel: CancellationToken,
+}
+
+/// When a reconcile task completes, it sends this result object
+/// to be applied to the primary TenantState.
+pub(crate) struct ReconcileResult {
+    pub(crate) sequence: Sequence,
+    /// On errors, `observed` should be treated as an incompleted description
+    /// of state (i.e. any nodes present in the result should override nodes
+    /// present in the parent tenant state, but any unmentioned nodes should
+    /// not be removed from parent tenant state)
+    pub(crate) result: Result<(), ReconcileError>,
+
+    pub(crate) tenant_shard_id: TenantShardId,
+    pub(crate) generation: Generation,
+    pub(crate) observed: ObservedState,
+}
+
+impl IntentState {
+    pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
+        let mut result = Vec::new();
+        if let Some(p) = self.attached {
+            result.push(p)
+        }
+
+        result.extend(self.secondary.iter().copied());
+
+        result
+    }
+
+    pub(crate) fn single(node_id: Option<NodeId>) -> Self {
+        Self {
+            attached: node_id,
+            secondary: vec![],
+        }
+    }
+
+    /// When a node goes offline, we update intents to avoid using it
+    /// as their attached pageserver.
+    ///
+    /// Returns true if a change was made
+    pub(crate) fn notify_offline(&mut self, node_id: NodeId) -> bool {
+        if self.attached == Some(node_id) {
+            self.attached = None;
+            self.secondary.push(node_id);
+            true
+        } else {
+            false
+        }
+    }
+}
+
+impl TenantState {
+    pub(crate) fn new(
+        tenant_shard_id: TenantShardId,
+        shard: ShardIdentity,
+        policy: PlacementPolicy,
+    ) -> Self {
+        Self {
+            tenant_shard_id,
+            policy,
+            intent: IntentState::default(),
+            generation: Generation::new(0),
+            shard,
+            observed: ObservedState::default(),
+            config: TenantConfig::default(),
+            reconciler: None,
+            sequence: Sequence(1),
+            waiter: Arc::new(SeqWait::new(Sequence(0))),
+        }
+    }
+
+    pub(crate) fn schedule(&mut self, scheduler: &mut Scheduler) -> Result<(), ScheduleError> {
+        // TODO: before scheduling new nodes, check if any existing content in
+        // self.intent refers to pageservers that are offline, and pick other
+        // pageservers if so.
+
+        // Build the set of pageservers already in use by this tenant, to avoid scheduling
+        // more work on the same pageservers we're already using.
+        let mut used_pageservers = self.intent.all_pageservers();
+        let mut modified = false;
+
+        use PlacementPolicy::*;
+        match self.policy {
+            Single => {
+                // Should have exactly one attached, and zero secondaries
+                if self.intent.attached.is_none() {
+                    let node_id = scheduler.schedule_shard(&used_pageservers)?;
+                    self.intent.attached = Some(node_id);
+                    used_pageservers.push(node_id);
+                    modified = true;
+                }
+                if !self.intent.secondary.is_empty() {
+                    self.intent.secondary.clear();
+                    modified = true;
+                }
+            }
+            Double(secondary_count) => {
+                // Should have exactly one attached, and N secondaries
+                if self.intent.attached.is_none() {
+                    let node_id = scheduler.schedule_shard(&used_pageservers)?;
+                    self.intent.attached = Some(node_id);
+                    used_pageservers.push(node_id);
+                    modified = true;
+                }
+
+                while self.intent.secondary.len() < secondary_count {
+                    let node_id = scheduler.schedule_shard(&used_pageservers)?;
+                    self.intent.secondary.push(node_id);
+                    used_pageservers.push(node_id);
+                    modified = true;
+                }
+            }
+        }
+
+        if modified {
+            self.sequence.0 += 1;
+        }
+
+        Ok(())
+    }
+
+    fn dirty(&self) -> bool {
+        if let Some(node_id) = self.intent.attached {
+            let wanted_conf = attached_location_conf(self.generation, &self.shard, &self.config);
+            match self.observed.locations.get(&node_id) {
+                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
+                Some(_) | None => {
+                    return true;
+                }
+            }
+        }
+
+        for node_id in &self.intent.secondary {
+            let wanted_conf = secondary_location_conf(&self.shard, &self.config);
+            match self.observed.locations.get(node_id) {
+                Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
+                Some(_) | None => {
+                    return true;
+                }
+            }
+        }
+
+        false
+    }
+
+    pub(crate) fn maybe_reconcile(
+        &mut self,
+        result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
+        pageservers: &Arc<HashMap<NodeId, Node>>,
+        compute_hook: &Arc<ComputeHook>,
+    ) -> Option<ReconcilerWaiter> {
+        // If there are any ambiguous observed states, and the nodes they refer to are available,
+        // we should reconcile to clean them up.
+        let mut dirty_observed = false;
+        for (node_id, observed_loc) in &self.observed.locations {
+            let node = pageservers
+                .get(node_id)
+                .expect("Nodes may not be removed while referenced");
+            if observed_loc.conf.is_none()
+                && !matches!(node.availability, NodeAvailability::Offline)
+            {
+                dirty_observed = true;
+                break;
+            }
+        }
+
+        if !self.dirty() && !dirty_observed {
+            tracing::info!("Not dirty, no reconciliation needed.");
+            return None;
+        }
+
+        // Reconcile already in flight for the current sequence?
+        if let Some(handle) = &self.reconciler {
+            if handle.sequence == self.sequence {
+                return Some(ReconcilerWaiter {
+                    tenant_shard_id: self.tenant_shard_id,
+                    seq_wait: self.waiter.clone(),
+                    seq: self.sequence,
+                });
+            }
+        }
+
+        // Reconcile in flight for a stale sequence?  Our sequence's task will wait for it before
+        // doing our sequence's work.
+        let old_handle = self.reconciler.take();
+
+        let cancel = CancellationToken::new();
+        let mut reconciler = Reconciler {
+            tenant_shard_id: self.tenant_shard_id,
+            shard: self.shard,
+            generation: self.generation,
+            intent: self.intent.clone(),
+            config: self.config.clone(),
+            observed: self.observed.clone(),
+            pageservers: pageservers.clone(),
+            compute_hook: compute_hook.clone(),
+            cancel: cancel.clone(),
+        };
+
+        let reconcile_seq = self.sequence;
+
+        tracing::info!("Spawning Reconciler for sequence {}", self.sequence);
+        let join_handle = tokio::task::spawn(async move {
+            // Wait for any previous reconcile task to complete before we start
+            if let Some(old_handle) = old_handle {
+                old_handle.cancel.cancel();
+                if let Err(e) = old_handle.handle.await {
+                    // We can't do much with this other than log it: the task is done, so
+                    // we may proceed with our work.
+                    tracing::error!("Unexpected join error waiting for reconcile task: {e}");
+                }
+            }
+
+            // Early check for cancellation before doing any work
+            // TODO: wrap all remote API operations in cancellation check
+            // as well.
+            if reconciler.cancel.is_cancelled() {
+                return;
+            }
+
+            let result = reconciler.reconcile().await;
+            result_tx
+                .send(ReconcileResult {
+                    sequence: reconcile_seq,
+                    result,
+                    tenant_shard_id: reconciler.tenant_shard_id,
+                    generation: reconciler.generation,
+                    observed: reconciler.observed,
+                })
+                .ok();
+        });
+
+        self.reconciler = Some(ReconcilerHandle {
+            sequence: self.sequence,
+            handle: join_handle,
+            cancel,
+        });
+
+        Some(ReconcilerWaiter {
+            tenant_shard_id: self.tenant_shard_id,
+            seq_wait: self.waiter.clone(),
+            seq: self.sequence,
+        })
+    }
+}

control_plane/src/attachment_service.rs

@@ -1,8 +1,18 @@
 use crate::{background_process, local_env::LocalEnv};
 use anyhow::anyhow;
 use camino::Utf8PathBuf;
-use serde::{Deserialize, Serialize};
-use std::{path::PathBuf, process::Child};
+use hyper::{Method, StatusCode};
+use pageserver_api::{
+    models::{
+        ShardParameters, TenantCreateRequest, TenantShardSplitRequest, TenantShardSplitResponse,
+        TimelineCreateRequest, TimelineInfo,
+    },
+    shard::TenantShardId,
+};
+use postgres_connection::parse_host_port;
+use serde::{de::DeserializeOwned, Deserialize, Serialize};
+use std::{path::PathBuf, process::Child, str::FromStr};
+use tracing::instrument;
 use utils::id::{NodeId, TenantId};
 
 pub struct AttachmentService {
@@ -16,7 +26,7 @@ const COMMAND: &str = "attachment_service";
 
 #[derive(Serialize, Deserialize)]
 pub struct AttachHookRequest {
-    pub tenant_id: TenantId,
+    pub tenant_shard_id: TenantShardId,
     pub node_id: Option<NodeId>,
 }
 
@@ -27,7 +37,7 @@ pub struct AttachHookResponse {
 
 #[derive(Serialize, Deserialize)]
 pub struct InspectRequest {
-    pub tenant_id: TenantId,
+    pub tenant_shard_id: TenantShardId,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -35,6 +45,117 @@ pub struct InspectResponse {
     pub attachment: Option<(u32, NodeId)>,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantCreateResponseShard {
+    pub node_id: NodeId,
+    pub generation: u32,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantCreateResponse {
+    pub shards: Vec<TenantCreateResponseShard>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct NodeRegisterRequest {
+    pub node_id: NodeId,
+
+    pub listen_pg_addr: String,
+    pub listen_pg_port: u16,
+
+    pub listen_http_addr: String,
+    pub listen_http_port: u16,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct NodeConfigureRequest {
+    pub node_id: NodeId,
+
+    pub availability: Option<NodeAvailability>,
+    pub scheduling: Option<NodeSchedulingPolicy>,
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantLocateResponseShard {
+    pub shard_id: TenantShardId,
+    pub node_id: NodeId,
+
+    pub listen_pg_addr: String,
+    pub listen_pg_port: u16,
+
+    pub listen_http_addr: String,
+    pub listen_http_port: u16,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantLocateResponse {
+    pub shards: Vec<TenantLocateResponseShard>,
+    pub shard_params: ShardParameters,
+}
+
+/// Explicitly migrating a particular shard is a low level operation
+/// TODO: higher level "Reschedule tenant" operation where the request
+/// specifies some constraints, e.g. asking it to get off particular node(s)
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantShardMigrateRequest {
+    pub tenant_shard_id: TenantShardId,
+    pub node_id: NodeId,
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy)]
+pub enum NodeAvailability {
+    // Normal, happy state
+    Active,
+    // Offline: Tenants shouldn't try to attach here, but they may assume that their
+    // secondary locations on this node still exist.  Newly added nodes are in this
+    // state until we successfully contact them.
+    Offline,
+}
+
+impl FromStr for NodeAvailability {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self::Active),
+            "offline" => Ok(Self::Offline),
+            _ => Err(anyhow::anyhow!("Unknown availability state '{s}'")),
+        }
+    }
+}
+
+#[derive(Serialize, Deserialize, Clone, Copy)]
+pub enum NodeSchedulingPolicy {
+    // Normal, happy state
+    Active,
+
+    // A newly added node: gradually move some work here.
+    Filling,
+
+    // Do not schedule new work here, but leave configured locations in place.
+    Pause,
+
+    // Do not schedule work here.  Gracefully move work away, as resources allow.
+    Draining,
+}
+
+impl FromStr for NodeSchedulingPolicy {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self::Active),
+            "filling" => Ok(Self::Filling),
+            "pause" => Ok(Self::Pause),
+            "draining" => Ok(Self::Draining),
+            _ => Err(anyhow::anyhow!("Unknown scheduling state '{s}'")),
+        }
+    }
+}
+
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantShardMigrateResponse {}
+
 impl AttachmentService {
     pub fn from_env(env: &LocalEnv) -> Self {
         let path = env.base_data_dir.join("attachments.json");
@@ -67,31 +188,87 @@ impl AttachmentService {
     pub async fn start(&self) -> anyhow::Result<Child> {
         let path_str = self.path.to_string_lossy();
 
-        background_process::start_process(
+        let result = background_process::start_process(
             COMMAND,
             &self.env.base_data_dir,
             &self.env.attachment_service_bin(),
             ["-l", &self.listen, "-p", &path_str],
             [],
             background_process::InitialPidFile::Create(self.pid_file()),
-            // TODO: a real status check
-            || async move { anyhow::Ok(true) },
+            || async {
+                match self.status().await {
+                    Ok(_) => Ok(true),
+                    Err(_) => Ok(false),
+                }
+            },
         )
-        .await
+        .await;
+
+        for ps_conf in &self.env.pageservers {
+            let (pg_host, pg_port) =
+                parse_host_port(&ps_conf.listen_pg_addr).expect("Unable to parse listen_pg_addr");
+            let (http_host, http_port) = parse_host_port(&ps_conf.listen_http_addr)
+                .expect("Unable to parse listen_http_addr");
+            self.node_register(NodeRegisterRequest {
+                node_id: ps_conf.id,
+                listen_pg_addr: pg_host.to_string(),
+                listen_pg_port: pg_port.unwrap_or(5432),
+                listen_http_addr: http_host.to_string(),
+                listen_http_port: http_port.unwrap_or(80),
+            })
+            .await?;
+        }
+
+        result
     }
 
     pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
         background_process::stop_process(immediate, COMMAND, &self.pid_file())
     }
 
+    /// Simple HTTP request wrapper for calling into attachment service
+    async fn dispatch<RQ, RS>(
+        &self,
+        method: hyper::Method,
+        path: String,
+        body: Option<RQ>,
+    ) -> anyhow::Result<RS>
+    where
+        RQ: Serialize + Sized,
+        RS: DeserializeOwned + Sized,
+    {
+        let url = self
+            .env
+            .control_plane_api
+            .clone()
+            .unwrap()
+            .join(&path)
+            .unwrap();
+
+        let mut builder = self.client.request(method, url);
+        if let Some(body) = body {
+            builder = builder.json(&body)
+        }
+
+        let response = builder.send().await?;
+        if response.status() != StatusCode::OK {
+            return Err(anyhow!(
+                "Unexpected status {} on {}",
+                response.status(),
+                path
+            ));
+        }
+
+        Ok(response.json().await?)
+    }
+
     /// Call into the attach_hook API, for use before handing out attachments to pageservers
+    #[instrument(skip(self))]
     pub async fn attach_hook(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         pageserver_id: NodeId,
     ) -> anyhow::Result<Option<u32>> {
-        use hyper::StatusCode;
-
         let url = self
             .env
             .control_plane_api
@@ -101,7 +278,7 @@ impl AttachmentService {
             .unwrap();
 
         let request = AttachHookRequest {
-            tenant_id,
+            tenant_shard_id,
             node_id: Some(pageserver_id),
         };
 
@@ -114,9 +291,11 @@ impl AttachmentService {
         Ok(response.gen)
     }
 
-    pub async fn inspect(&self, tenant_id: TenantId) -> anyhow::Result<Option<(u32, NodeId)>> {
-        use hyper::StatusCode;
-
+    #[instrument(skip(self))]
+    pub async fn inspect(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> anyhow::Result<Option<(u32, NodeId)>> {
         let url = self
             .env
             .control_plane_api
@@ -125,7 +304,7 @@ impl AttachmentService {
             .join("inspect")
             .unwrap();
 
-        let request = InspectRequest { tenant_id };
+        let request = InspectRequest { tenant_shard_id };
 
         let response = self.client.post(url).json(&request).send().await?;
         if response.status() != StatusCode::OK {
@@ -135,4 +314,86 @@ impl AttachmentService {
         let response = response.json::<InspectResponse>().await?;
         Ok(response.attachment)
     }
+
+    #[instrument(skip(self))]
+    pub async fn tenant_create(
+        &self,
+        req: TenantCreateRequest,
+    ) -> anyhow::Result<TenantCreateResponse> {
+        self.dispatch(Method::POST, "tenant".to_string(), Some(req))
+            .await
+    }
+
+    #[instrument(skip(self))]
+    pub async fn tenant_locate(&self, tenant_id: TenantId) -> anyhow::Result<TenantLocateResponse> {
+        self.dispatch::<(), _>(Method::GET, format!("tenant/{tenant_id}/locate"), None)
+            .await
+    }
+
+    #[instrument(skip(self))]
+    pub async fn tenant_migrate(
+        &self,
+        tenant_shard_id: TenantShardId,
+        node_id: NodeId,
+    ) -> anyhow::Result<TenantShardMigrateResponse> {
+        self.dispatch(
+            Method::PUT,
+            format!("tenant/{tenant_shard_id}/migrate"),
+            Some(TenantShardMigrateRequest {
+                tenant_shard_id,
+                node_id,
+            }),
+        )
+        .await
+    }
+
+    #[instrument(skip(self), fields(%tenant_id, %new_shard_count))]
+    pub async fn tenant_split(
+        &self,
+        tenant_id: TenantId,
+        new_shard_count: u8,
+    ) -> anyhow::Result<TenantShardSplitResponse> {
+        self.dispatch(
+            Method::PUT,
+            format!("tenant/{tenant_id}/shard_split"),
+            Some(TenantShardSplitRequest { new_shard_count }),
+        )
+        .await
+    }
+
+    #[instrument(skip_all, fields(node_id=%req.node_id))]
+    pub async fn node_register(&self, req: NodeRegisterRequest) -> anyhow::Result<()> {
+        self.dispatch::<_, ()>(Method::POST, "node".to_string(), Some(req))
+            .await
+    }
+
+    #[instrument(skip_all, fields(node_id=%req.node_id))]
+    pub async fn node_configure(&self, req: NodeConfigureRequest) -> anyhow::Result<()> {
+        self.dispatch::<_, ()>(
+            Method::PUT,
+            format!("node/{}/config", req.node_id),
+            Some(req),
+        )
+        .await
+    }
+
+    #[instrument(skip(self))]
+    pub async fn status(&self) -> anyhow::Result<()> {
+        self.dispatch::<(), ()>(Method::GET, "status".to_string(), None)
+            .await
+    }
+
+    #[instrument(skip_all, fields(%tenant_id, timeline_id=%req.new_timeline_id))]
+    pub async fn tenant_timeline_create(
+        &self,
+        tenant_id: TenantId,
+        req: TimelineCreateRequest,
+    ) -> anyhow::Result<TimelineInfo> {
+        self.dispatch(
+            Method::POST,
+            format!("tenant/{tenant_id}/timeline"),
+            Some(req),
+        )
+        .await
+    }
 }

control_plane/src/bin/attachment_service.rs

@@ -1,337 +0,0 @@
-/// The attachment service mimics the aspects of the control plane API
-/// that are required for a pageserver to operate.
-///
-/// This enables running & testing pageservers without a full-blown
-/// deployment of the Neon cloud platform.
-///
-use anyhow::anyhow;
-use clap::Parser;
-use hex::FromHex;
-use hyper::StatusCode;
-use hyper::{Body, Request, Response};
-use pageserver_api::shard::TenantShardId;
-use serde::{Deserialize, Serialize};
-use std::path::{Path, PathBuf};
-use std::{collections::HashMap, sync::Arc};
-use utils::http::endpoint::request_span;
-use utils::logging::{self, LogFormat};
-use utils::signals::{ShutdownSignals, Signal};
-
-use utils::{
-    http::{
-        endpoint::{self},
-        error::ApiError,
-        json::{json_request, json_response},
-        RequestExt, RouterBuilder,
-    },
-    id::{NodeId, TenantId},
-    tcp_listener,
-};
-
-use pageserver_api::control_api::{
-    ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest, ValidateResponse,
-    ValidateResponseTenant,
-};
-
-use control_plane::attachment_service::{
-    AttachHookRequest, AttachHookResponse, InspectRequest, InspectResponse,
-};
-
-#[derive(Parser)]
-#[command(author, version, about, long_about = None)]
-#[command(arg_required_else_help(true))]
-struct Cli {
-    /// Host and port to listen on, like `127.0.0.1:1234`
-    #[arg(short, long)]
-    listen: std::net::SocketAddr,
-
-    /// Path to the .json file to store state (will be created if it doesn't exist)
-    #[arg(short, long)]
-    path: PathBuf,
-}
-
-// The persistent state of each Tenant
-#[derive(Serialize, Deserialize, Clone)]
-struct TenantState {
-    // Currently attached pageserver
-    pageserver: Option<NodeId>,
-
-    // Latest generation number: next time we attach, increment this
-    // and use the incremented number when attaching
-    generation: u32,
-}
-
-fn to_hex_map<S, V>(input: &HashMap<TenantId, V>, serializer: S) -> Result<S::Ok, S::Error>
-where
-    S: serde::Serializer,
-    V: Clone + Serialize,
-{
-    let transformed = input.iter().map(|(k, v)| (hex::encode(k), v.clone()));
-
-    transformed
-        .collect::<HashMap<String, V>>()
-        .serialize(serializer)
-}
-
-fn from_hex_map<'de, D, V>(deserializer: D) -> Result<HashMap<TenantId, V>, D::Error>
-where
-    D: serde::de::Deserializer<'de>,
-    V: Deserialize<'de>,
-{
-    let hex_map = HashMap::<String, V>::deserialize(deserializer)?;
-    hex_map
-        .into_iter()
-        .map(|(k, v)| {
-            TenantId::from_hex(k)
-                .map(|k| (k, v))
-                .map_err(serde::de::Error::custom)
-        })
-        .collect()
-}
-
-// Top level state available to all HTTP handlers
-#[derive(Serialize, Deserialize)]
-struct PersistentState {
-    #[serde(serialize_with = "to_hex_map", deserialize_with = "from_hex_map")]
-    tenants: HashMap<TenantId, TenantState>,
-
-    #[serde(skip)]
-    path: PathBuf,
-}
-
-impl PersistentState {
-    async fn save(&self) -> anyhow::Result<()> {
-        let bytes = serde_json::to_vec(self)?;
-        tokio::fs::write(&self.path, &bytes).await?;
-
-        Ok(())
-    }
-
-    async fn load(path: &Path) -> anyhow::Result<Self> {
-        let bytes = tokio::fs::read(path).await?;
-        let mut decoded = serde_json::from_slice::<Self>(&bytes)?;
-        decoded.path = path.to_owned();
-        Ok(decoded)
-    }
-
-    async fn load_or_new(path: &Path) -> Self {
-        match Self::load(path).await {
-            Ok(s) => {
-                tracing::info!("Loaded state file at {}", path.display());
-                s
-            }
-            Err(e)
-                if e.downcast_ref::<std::io::Error>()
-                    .map(|e| e.kind() == std::io::ErrorKind::NotFound)
-                    .unwrap_or(false) =>
-            {
-                tracing::info!("Will create state file at {}", path.display());
-                Self {
-                    tenants: HashMap::new(),
-                    path: path.to_owned(),
-                }
-            }
-            Err(e) => {
-                panic!("Failed to load state from '{}': {e:#} (maybe your .neon/ dir was written by an older version?)", path.display())
-            }
-        }
-    }
-}
-
-/// State available to HTTP request handlers
-#[derive(Clone)]
-struct State {
-    inner: Arc<tokio::sync::RwLock<PersistentState>>,
-}
-
-impl State {
-    fn new(persistent_state: PersistentState) -> State {
-        Self {
-            inner: Arc::new(tokio::sync::RwLock::new(persistent_state)),
-        }
-    }
-}
-
-#[inline(always)]
-fn get_state(request: &Request<Body>) -> &State {
-    request
-        .data::<Arc<State>>()
-        .expect("unknown state type")
-        .as_ref()
-}
-
-/// Pageserver calls into this on startup, to learn which tenants it should attach
-async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    let reattach_req = json_request::<ReAttachRequest>(&mut req).await?;
-
-    let state = get_state(&req).inner.clone();
-    let mut locked = state.write().await;
-
-    let mut response = ReAttachResponse {
-        tenants: Vec::new(),
-    };
-    for (t, state) in &mut locked.tenants {
-        if state.pageserver == Some(reattach_req.node_id) {
-            state.generation += 1;
-            response.tenants.push(ReAttachResponseTenant {
-                // TODO(sharding): make this shard-aware
-                id: TenantShardId::unsharded(*t),
-                gen: state.generation,
-            });
-        }
-    }
-
-    locked.save().await.map_err(ApiError::InternalServerError)?;
-
-    json_response(StatusCode::OK, response)
-}
-
-/// Pageserver calls into this before doing deletions, to confirm that it still
-/// holds the latest generation for the tenants with deletions enqueued
-async fn handle_validate(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    let validate_req = json_request::<ValidateRequest>(&mut req).await?;
-
-    let locked = get_state(&req).inner.read().await;
-
-    let mut response = ValidateResponse {
-        tenants: Vec::new(),
-    };
-
-    for req_tenant in validate_req.tenants {
-        // TODO(sharding): make this shard-aware
-        if let Some(tenant_state) = locked.tenants.get(&req_tenant.id.tenant_id) {
-            let valid = tenant_state.generation == req_tenant.gen;
-            tracing::info!(
-                "handle_validate: {}(gen {}): valid={valid} (latest {})",
-                req_tenant.id,
-                req_tenant.gen,
-                tenant_state.generation
-            );
-            response.tenants.push(ValidateResponseTenant {
-                id: req_tenant.id,
-                valid,
-            });
-        }
-    }
-
-    json_response(StatusCode::OK, response)
-}
-/// Call into this before attaching a tenant to a pageserver, to acquire a generation number
-/// (in the real control plane this is unnecessary, because the same program is managing
-///  generation numbers and doing attachments).
-async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    let attach_req = json_request::<AttachHookRequest>(&mut req).await?;
-
-    let state = get_state(&req).inner.clone();
-    let mut locked = state.write().await;
-
-    let tenant_state = locked
-        .tenants
-        .entry(attach_req.tenant_id)
-        .or_insert_with(|| TenantState {
-            pageserver: attach_req.node_id,
-            generation: 0,
-        });
-
-    if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
-        tenant_state.generation += 1;
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            ps_id = %attaching_pageserver,
-            generation = %tenant_state.generation,
-            "issuing",
-        );
-    } else if let Some(ps_id) = tenant_state.pageserver {
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            %ps_id,
-            generation = %tenant_state.generation,
-            "dropping",
-        );
-    } else {
-        tracing::info!(
-            tenant_id = %attach_req.tenant_id,
-            "no-op: tenant already has no pageserver");
-    }
-    tenant_state.pageserver = attach_req.node_id;
-    let generation = tenant_state.generation;
-
-    tracing::info!(
-        "handle_attach_hook: tenant {} set generation {}, pageserver {}",
-        attach_req.tenant_id,
-        tenant_state.generation,
-        attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
-    );
-
-    locked.save().await.map_err(ApiError::InternalServerError)?;
-
-    json_response(
-        StatusCode::OK,
-        AttachHookResponse {
-            gen: attach_req.node_id.map(|_| generation),
-        },
-    )
-}
-
-async fn handle_inspect(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    let inspect_req = json_request::<InspectRequest>(&mut req).await?;
-
-    let state = get_state(&req).inner.clone();
-    let locked = state.write().await;
-    let tenant_state = locked.tenants.get(&inspect_req.tenant_id);
-
-    json_response(
-        StatusCode::OK,
-        InspectResponse {
-            attachment: tenant_state.and_then(|s| s.pageserver.map(|ps| (s.generation, ps))),
-        },
-    )
-}
-
-fn make_router(persistent_state: PersistentState) -> RouterBuilder<hyper::Body, ApiError> {
-    endpoint::make_router()
-        .data(Arc::new(State::new(persistent_state)))
-        .post("/re-attach", |r| request_span(r, handle_re_attach))
-        .post("/validate", |r| request_span(r, handle_validate))
-        .post("/attach-hook", |r| request_span(r, handle_attach_hook))
-        .post("/inspect", |r| request_span(r, handle_inspect))
-}
-
-#[tokio::main]
-async fn main() -> anyhow::Result<()> {
-    logging::init(
-        LogFormat::Plain,
-        logging::TracingErrorLayerEnablement::Disabled,
-        logging::Output::Stdout,
-    )?;
-
-    let args = Cli::parse();
-    tracing::info!(
-        "Starting, state at {}, listening on {}",
-        args.path.to_string_lossy(),
-        args.listen
-    );
-
-    let persistent_state = PersistentState::load_or_new(&args.path).await;
-
-    let http_listener = tcp_listener::bind(args.listen)?;
-    let router = make_router(persistent_state)
-        .build()
-        .map_err(|err| anyhow!(err))?;
-    let service = utils::http::RouterService::new(router).unwrap();
-    let server = hyper::Server::from_tcp(http_listener)?.serve(service);
-
-    tracing::info!("Serving on {0}", args.listen);
-
-    tokio::task::spawn(server);
-
-    ShutdownSignals::handle(|signal| match signal {
-        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
-            tracing::info!("Got {}. Terminating", signal.name());
-            // We're just a test helper: no graceful shutdown.
-            std::process::exit(0);
-        }
-    })?;
-
-    Ok(())
-}

control_plane/src/bin/neon_local.rs

@@ -8,14 +8,18 @@
 use anyhow::{anyhow, bail, Context, Result};
 use clap::{value_parser, Arg, ArgAction, ArgMatches, Command};
 use compute_api::spec::ComputeMode;
-use control_plane::attachment_service::AttachmentService;
+use control_plane::attachment_service::{
+    AttachmentService, NodeAvailability, NodeConfigureRequest, NodeSchedulingPolicy,
+};
 use control_plane::endpoint::ComputeControlPlane;
 use control_plane::local_env::LocalEnv;
 use control_plane::pageserver::{PageServerNode, PAGESERVER_REMOTE_STORAGE_DIR};
 use control_plane::safekeeper::SafekeeperNode;
-use control_plane::tenant_migration::migrate_tenant;
 use control_plane::{broker, local_env};
-use pageserver_api::models::TimelineInfo;
+use pageserver_api::models::{
+    ShardParameters, TenantCreateRequest, TimelineCreateRequest, TimelineInfo,
+};
+use pageserver_api::shard::{ShardCount, ShardStripeSize, TenantShardId};
 use pageserver_api::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_PAGESERVER_HTTP_PORT,
     DEFAULT_PG_LISTEN_PORT as DEFAULT_PAGESERVER_PG_PORT,
@@ -30,6 +34,7 @@ use std::path::PathBuf;
 use std::process::exit;
 use std::str::FromStr;
 use storage_broker::DEFAULT_LISTEN_ADDR as DEFAULT_BROKER_ADDR;
+use url::Host;
 use utils::{
     auth::{Claims, Scope},
     id::{NodeId, TenantId, TenantTimelineId, TimelineId},
@@ -276,10 +281,10 @@ fn print_timeline(
 /// Connects to the pageserver to query this information.
 async fn get_timeline_infos(
     env: &local_env::LocalEnv,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> Result<HashMap<TimelineId, TimelineInfo>> {
     Ok(get_default_pageserver(env)
-        .timeline_list(tenant_id)
+        .timeline_list(tenant_shard_id)
         .await?
         .into_iter()
         .map(|timeline_info| (timeline_info.timeline_id, timeline_info))
@@ -297,6 +302,20 @@ fn get_tenant_id(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow::R
     }
 }
 
+// Helper function to parse --tenant_id option, for commands that accept a shard suffix
+fn get_tenant_shard_id(
+    sub_match: &ArgMatches,
+    env: &local_env::LocalEnv,
+) -> anyhow::Result<TenantShardId> {
+    if let Some(tenant_id_from_arguments) = parse_tenant_shard_id(sub_match).transpose() {
+        tenant_id_from_arguments
+    } else if let Some(default_id) = env.default_tenant_id {
+        Ok(TenantShardId::unsharded(default_id))
+    } else {
+        anyhow::bail!("No tenant shard id. Use --tenant-id, or set a default tenant");
+    }
+}
+
 fn parse_tenant_id(sub_match: &ArgMatches) -> anyhow::Result<Option<TenantId>> {
     sub_match
         .get_one::<String>("tenant-id")
@@ -305,6 +324,14 @@ fn parse_tenant_id(sub_match: &ArgMatches) -> anyhow::Result<Option<TenantId>> {
         .context("Failed to parse tenant id from the argument string")
 }
 
+fn parse_tenant_shard_id(sub_match: &ArgMatches) -> anyhow::Result<Option<TenantShardId>> {
+    sub_match
+        .get_one::<String>("tenant-id")
+        .map(|id_str| TenantShardId::from_str(id_str))
+        .transpose()
+        .context("Failed to parse tenant shard id from the argument string")
+}
+
 fn parse_timeline_id(sub_match: &ArgMatches) -> anyhow::Result<Option<TimelineId>> {
     sub_match
         .get_one::<String>("timeline-id")
@@ -393,47 +420,66 @@ async fn handle_tenant(
         Some(("create", create_match)) => {
             let tenant_conf: HashMap<_, _> = create_match
                 .get_many::<String>("config")
-                .map(|vals| vals.flat_map(|c| c.split_once(':')).collect())
+                .map(|vals: clap::parser::ValuesRef<'_, String>| {
+                    vals.flat_map(|c| c.split_once(':')).collect()
+                })
                 .unwrap_or_default();
 
+            let shard_count: u8 = create_match
+                .get_one::<u8>("shard-count")
+                .cloned()
+                .unwrap_or(0);
+
+            let shard_stripe_size: Option<u32> =
+                create_match.get_one::<u32>("shard-stripe-size").cloned();
+
+            let tenant_conf = PageServerNode::parse_config(tenant_conf)?;
+
             // If tenant ID was not specified, generate one
             let tenant_id = parse_tenant_id(create_match)?.unwrap_or_else(TenantId::generate);
 
-            let generation = if env.control_plane_api.is_some() {
-                // We must register the tenant with the attachment service, so
-                // that when the pageserver restarts, it will be re-attached.
-                let attachment_service = AttachmentService::from_env(env);
-                attachment_service
-                    .attach_hook(tenant_id, pageserver.conf.id)
-                    .await?
-            } else {
-                None
-            };
-
-            pageserver
-                .tenant_create(tenant_id, generation, tenant_conf)
+            // We must register the tenant with the attachment service, so
+            // that when the pageserver restarts, it will be re-attached.
+            let attachment_service = AttachmentService::from_env(env);
+            attachment_service
+                .tenant_create(TenantCreateRequest {
+                    // Note that ::unsharded here isn't actually because the tenant is unsharded, its because the
+                    // attachment service expecfs a shard-naive tenant_id in this attribute, and the TenantCreateRequest
+                    // type is used both in attachment service (for creating tenants) and in pageserver (for creating shards)
+                    new_tenant_id: TenantShardId::unsharded(tenant_id),
+                    generation: None,
+                    shard_parameters: ShardParameters {
+                        count: ShardCount(shard_count),
+                        stripe_size: shard_stripe_size.map(ShardStripeSize),
+                    },
+                    config: tenant_conf,
+                })
                 .await?;
             println!("tenant {tenant_id} successfully created on the pageserver");
 
             // Create an initial timeline for the new tenant
-            let new_timeline_id = parse_timeline_id(create_match)?;
+            let new_timeline_id =
+                parse_timeline_id(create_match)?.unwrap_or(TimelineId::generate());
             let pg_version = create_match
                 .get_one::<u32>("pg-version")
                 .copied()
                 .context("Failed to parse postgres version from the argument string")?;
 
-            let timeline_info = pageserver
-                .timeline_create(
+            // FIXME: passing None for ancestor_start_lsn is not kosher in a sharded world: we can't have
+            // different shards picking different start lsns.  Maybe we have to teach attachment service
+            // to let shard 0 branch first and then propagate the chosen LSN to other shards.
+            attachment_service
+                .tenant_timeline_create(
                     tenant_id,
-                    new_timeline_id,
-                    None,
-                    None,
-                    Some(pg_version),
-                    None,
+                    TimelineCreateRequest {
+                        new_timeline_id,
+                        ancestor_timeline_id: None,
+                        ancestor_start_lsn: None,
+                        existing_initdb_timeline_id: None,
+                        pg_version: Some(pg_version),
+                    },
                 )
                 .await?;
-            let new_timeline_id = timeline_info.timeline_id;
-            let last_record_lsn = timeline_info.last_record_lsn;
 
             env.register_branch_mapping(
                 DEFAULT_BRANCH_NAME.to_string(),
@@ -441,9 +487,7 @@ async fn handle_tenant(
                 new_timeline_id,
             )?;
 
-            println!(
-                "Created an initial timeline '{new_timeline_id}' at Lsn {last_record_lsn} for tenant: {tenant_id}",
-            );
+            println!("Created an initial timeline '{new_timeline_id}' for tenant: {tenant_id}",);
 
             if create_match.get_flag("set-default") {
                 println!("Setting tenant {tenant_id} as a default one");
@@ -470,12 +514,82 @@ async fn handle_tenant(
             println!("tenant {tenant_id} successfully configured on the pageserver");
         }
         Some(("migrate", matches)) => {
-            let tenant_id = get_tenant_id(matches, env)?;
+            let tenant_shard_id = get_tenant_shard_id(matches, env)?;
             let new_pageserver = get_pageserver(env, matches)?;
             let new_pageserver_id = new_pageserver.conf.id;
 
-            migrate_tenant(env, tenant_id, new_pageserver).await?;
-            println!("tenant {tenant_id} migrated to {}", new_pageserver_id);
+            let attachment_service = AttachmentService::from_env(env);
+            attachment_service
+                .tenant_migrate(tenant_shard_id, new_pageserver_id)
+                .await?;
+
+            println!("tenant {tenant_shard_id} migrated to {}", new_pageserver_id);
+        }
+        Some(("status", matches)) => {
+            let tenant_id = get_tenant_id(matches, env)?;
+
+            let mut shard_table = comfy_table::Table::new();
+            shard_table.set_header(["Shard", "Pageserver", "Physical Size"]);
+
+            let mut tenant_synthetic_size = None;
+
+            let attachment_service = AttachmentService::from_env(env);
+            for shard in attachment_service.tenant_locate(tenant_id).await?.shards {
+                let pageserver =
+                    PageServerNode::from_env(env, env.get_pageserver_conf(shard.node_id)?);
+
+                let size = pageserver
+                    .http_client
+                    .tenant_details(shard.shard_id)
+                    .await?
+                    .tenant_info
+                    .current_physical_size
+                    .unwrap();
+
+                shard_table.add_row([
+                    format!("{}", shard.shard_id.shard_slug()),
+                    format!("{}", shard.node_id.0),
+                    format!("{} MiB", size / (1024 * 1024)),
+                ]);
+
+                if shard.shard_id.is_zero() {
+                    tenant_synthetic_size =
+                        Some(pageserver.tenant_synthetic_size(shard.shard_id).await?);
+                }
+            }
+
+            let Some(synthetic_size) = tenant_synthetic_size else {
+                bail!("Shard 0 not found")
+            };
+
+            let mut tenant_table = comfy_table::Table::new();
+            tenant_table.add_row(["Tenant ID".to_string(), tenant_id.to_string()]);
+            tenant_table.add_row([
+                "Synthetic size".to_string(),
+                format!("{} MiB", synthetic_size.size.unwrap_or(0) / (1024 * 1024)),
+            ]);
+
+            println!("{tenant_table}");
+            println!("{shard_table}");
+        }
+        Some(("shard-split", matches)) => {
+            let tenant_id = get_tenant_id(matches, env)?;
+            let shard_count: u8 = matches.get_one::<u8>("shard-count").cloned().unwrap_or(0);
+
+            let attachment_service = AttachmentService::from_env(env);
+            let result = attachment_service
+                .tenant_split(tenant_id, shard_count)
+                .await?;
+            println!(
+                "Split tenant {} into shards {}",
+                tenant_id,
+                result
+                    .new_shards
+                    .iter()
+                    .map(|s| format!("{:?}", s))
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
         }
 
         Some((sub_name, _)) => bail!("Unexpected tenant subcommand '{}'", sub_name),
@@ -489,8 +603,10 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
 
     match timeline_match.subcommand() {
         Some(("list", list_match)) => {
-            let tenant_id = get_tenant_id(list_match, env)?;
-            let timelines = pageserver.timeline_list(&tenant_id).await?;
+            // TODO(sharding): this command shouldn't have to specify a shard ID: we should ask the attachment service
+            // where shard 0 is attached, and query there.
+            let tenant_shard_id = get_tenant_shard_id(list_match, env)?;
+            let timelines = pageserver.timeline_list(&tenant_shard_id).await?;
             print_timelines_tree(timelines, env.timeline_name_mappings())?;
         }
         Some(("create", create_match)) => {
@@ -505,18 +621,19 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
                 .context("Failed to parse postgres version from the argument string")?;
 
             let new_timeline_id_opt = parse_timeline_id(create_match)?;
+            let new_timeline_id = new_timeline_id_opt.unwrap_or(TimelineId::generate());
 
-            let timeline_info = pageserver
-                .timeline_create(
-                    tenant_id,
-                    new_timeline_id_opt,
-                    None,
-                    None,
-                    Some(pg_version),
-                    None,
-                )
+            let attachment_service = AttachmentService::from_env(env);
+            let create_req = TimelineCreateRequest {
+                new_timeline_id,
+                ancestor_timeline_id: None,
+                existing_initdb_timeline_id: None,
+                ancestor_start_lsn: None,
+                pg_version: Some(pg_version),
+            };
+            let timeline_info = attachment_service
+                .tenant_timeline_create(tenant_id, create_req)
                 .await?;
-            let new_timeline_id = timeline_info.timeline_id;
 
             let last_record_lsn = timeline_info.last_record_lsn;
             env.register_branch_mapping(new_branch_name.to_string(), tenant_id, new_timeline_id)?;
@@ -574,7 +691,6 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
                 None,
                 pg_version,
                 ComputeMode::Primary,
-                DEFAULT_PAGESERVER_ID,
             )?;
             println!("Done");
         }
@@ -598,17 +714,18 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
                 .map(|lsn_str| Lsn::from_str(lsn_str))
                 .transpose()
                 .context("Failed to parse ancestor start Lsn from the request")?;
-            let timeline_info = pageserver
-                .timeline_create(
-                    tenant_id,
-                    None,
-                    start_lsn,
-                    Some(ancestor_timeline_id),
-                    None,
-                    None,
-                )
+            let new_timeline_id = TimelineId::generate();
+            let attachment_service = AttachmentService::from_env(env);
+            let create_req = TimelineCreateRequest {
+                new_timeline_id,
+                ancestor_timeline_id: Some(ancestor_timeline_id),
+                existing_initdb_timeline_id: None,
+                ancestor_start_lsn: start_lsn,
+                pg_version: None,
+            };
+            let timeline_info = attachment_service
+                .tenant_timeline_create(tenant_id, create_req)
                 .await?;
-            let new_timeline_id = timeline_info.timeline_id;
 
             let last_record_lsn = timeline_info.last_record_lsn;
 
@@ -635,8 +752,10 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
 
     match sub_name {
         "list" => {
-            let tenant_id = get_tenant_id(sub_args, env)?;
-            let timeline_infos = get_timeline_infos(env, &tenant_id)
+            // TODO(sharding): this command shouldn't have to specify a shard ID: we should ask the attachment service
+            // where shard 0 is attached, and query there.
+            let tenant_shard_id = get_tenant_shard_id(sub_args, env)?;
+            let timeline_infos = get_timeline_infos(env, &tenant_shard_id)
                 .await
                 .unwrap_or_else(|e| {
                     eprintln!("Failed to load timeline info: {}", e);
@@ -661,7 +780,7 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
             for (endpoint_id, endpoint) in cplane
                 .endpoints
                 .iter()
-                .filter(|(_, endpoint)| endpoint.tenant_id == tenant_id)
+                .filter(|(_, endpoint)| endpoint.tenant_id == tenant_shard_id.tenant_id)
             {
                 let lsn_str = match endpoint.mode {
                     ComputeMode::Static(lsn) => {
@@ -680,7 +799,10 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 };
 
                 let branch_name = timeline_name_mappings
-                    .get(&TenantTimelineId::new(tenant_id, endpoint.timeline_id))
+                    .get(&TenantTimelineId::new(
+                        tenant_shard_id.tenant_id,
+                        endpoint.timeline_id,
+                    ))
                     .map(|name| name.as_str())
                     .unwrap_or("?");
 
@@ -728,13 +850,6 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .copied()
                 .unwrap_or(false);
 
-            let pageserver_id =
-                if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
-                    NodeId(id_str.parse().context("while parsing pageserver id")?)
-                } else {
-                    DEFAULT_PAGESERVER_ID
-                };
-
             let mode = match (lsn, hot_standby) {
                 (Some(lsn), false) => ComputeMode::Static(lsn),
                 (None, true) => ComputeMode::Replica,
@@ -762,7 +877,6 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 http_port,
                 pg_version,
                 mode,
-                pageserver_id,
             )?;
         }
         "start" => {
@@ -805,6 +919,22 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 endpoint.timeline_id,
             )?;
 
+            let attachment_service = AttachmentService::from_env(env);
+            let locate_result = attachment_service.tenant_locate(endpoint.tenant_id).await?;
+            let pageservers = locate_result
+                .shards
+                .into_iter()
+                .map(|shard| {
+                    (
+                        Host::parse(&shard.listen_pg_addr)
+                            .expect("Attachment service reported bad hostname"),
+                        shard.listen_pg_port,
+                    )
+                })
+                .collect::<Vec<_>>();
+            assert!(!pageservers.is_empty());
+            let stripe_size = locate_result.shard_params.stripe_size.map(|s| s.0 as usize);
+
             let ps_conf = env.get_pageserver_conf(pageserver_id)?;
             let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
                 let claims = Claims::new(Some(endpoint.tenant_id), Scope::Tenant);
@@ -816,7 +946,13 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
 
             println!("Starting existing endpoint {endpoint_id}...");
             endpoint
-                .start(&auth_token, safekeepers, remote_ext_config)
+                .start(
+                    &auth_token,
+                    safekeepers,
+                    pageservers,
+                    remote_ext_config,
+                    stripe_size,
+                )
                 .await?;
         }
         "reconfigure" => {
@@ -827,15 +963,31 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .endpoints
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let pageserver_id =
+            let pageservers =
                 if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
-                    Some(NodeId(
-                        id_str.parse().context("while parsing pageserver id")?,
-                    ))
+                    let ps_id = NodeId(id_str.parse().context("while parsing pageserver id")?);
+                    let pageserver = PageServerNode::from_env(env, env.get_pageserver_conf(ps_id)?);
+                    vec![(
+                        pageserver.pg_connection_config.host().clone(),
+                        pageserver.pg_connection_config.port(),
+                    )]
                 } else {
-                    None
+                    let attachment_service = AttachmentService::from_env(env);
+                    attachment_service
+                        .tenant_locate(endpoint.tenant_id)
+                        .await?
+                        .shards
+                        .into_iter()
+                        .map(|shard| {
+                            (
+                                Host::parse(&shard.listen_pg_addr)
+                                    .expect("Attachment service reported malformed host"),
+                                shard.listen_pg_port,
+                            )
+                        })
+                        .collect::<Vec<_>>()
                 };
-            endpoint.reconfigure(pageserver_id).await?;
+            endpoint.reconfigure(pageservers).await?;
         }
         "stop" => {
             let endpoint_id = sub_args
@@ -959,6 +1111,21 @@ async fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) ->
             }
         }
 
+        Some(("set-state", subcommand_args)) => {
+            let pageserver = get_pageserver(env, subcommand_args)?;
+            let scheduling = subcommand_args.get_one("scheduling");
+            let availability = subcommand_args.get_one("availability");
+
+            let attachment_service = AttachmentService::from_env(env);
+            attachment_service
+                .node_configure(NodeConfigureRequest {
+                    node_id: pageserver.conf.id,
+                    scheduling: scheduling.cloned(),
+                    availability: availability.cloned(),
+                })
+                .await?;
+        }
+
         Some(("status", subcommand_args)) => {
             match get_pageserver(env, subcommand_args)?.check_status().await {
                 Ok(_) => println!("Page server is up and running"),
@@ -1352,6 +1519,8 @@ fn cli() -> Command {
                 .arg(pg_version_arg.clone())
                 .arg(Arg::new("set-default").long("set-default").action(ArgAction::SetTrue).required(false)
                     .help("Use this tenant in future CLI commands where tenant_id is needed, but not specified"))
+                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
+                .arg(Arg::new("shard-stripe-size").value_parser(value_parser!(u32)).long("shard-stripe-size").action(ArgAction::Set).help("Sharding stripe size in pages"))
                 )
             .subcommand(Command::new("set-default").arg(tenant_id_arg.clone().required(true))
                 .about("Set a particular tenant as default in future CLI commands where tenant_id is needed, but not specified"))
@@ -1362,6 +1531,14 @@ fn cli() -> Command {
                 .about("Migrate a tenant from one pageserver to another")
                 .arg(tenant_id_arg.clone())
                 .arg(pageserver_id_arg.clone()))
+            .subcommand(Command::new("status")
+                .about("Human readable summary of the tenant's shards and attachment locations")
+                .arg(tenant_id_arg.clone()))
+            .subcommand(Command::new("shard-split")
+                .about("Increase the number of shards in the tenant")
+                .arg(tenant_id_arg.clone())
+                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
+                )
         )
         .subcommand(
             Command::new("pageserver")
@@ -1381,6 +1558,12 @@ fn cli() -> Command {
                     .about("Restart local pageserver")
                     .arg(pageserver_config_args.clone())
                 )
+                .subcommand(Command::new("set-state")
+                    .arg(Arg::new("availability").value_parser(value_parser!(NodeAvailability)).long("availability").action(ArgAction::Set).help("Availability state: offline,active"))
+                    .arg(Arg::new("scheduling").value_parser(value_parser!(NodeSchedulingPolicy)).long("scheduling").action(ArgAction::Set).help("Scheduling state: draining,pause,filling,active"))
+                    .about("Set scheduling or availability state of pageserver node")
+                    .arg(pageserver_config_args.clone())
+                )
         )
         .subcommand(
             Command::new("attachment_service")

control_plane/src/endpoint.rs

@@ -46,11 +46,14 @@ use std::time::Duration;
 
 use anyhow::{anyhow, bail, Context, Result};
 use compute_api::spec::RemoteExtSpec;
+use nix::sys::signal::kill;
+use nix::sys::signal::Signal;
 use serde::{Deserialize, Serialize};
+use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
 
+use crate::attachment_service::AttachmentService;
 use crate::local_env::LocalEnv;
-use crate::pageserver::PageServerNode;
 use crate::postgresql_conf::PostgresConf;
 
 use compute_api::responses::{ComputeState, ComputeStatus};
@@ -67,7 +70,6 @@ pub struct EndpointConf {
     http_port: u16,
     pg_version: u32,
     skip_pg_catalog_updates: bool,
-    pageserver_id: NodeId,
 }
 
 //
@@ -119,19 +121,14 @@ impl ComputeControlPlane {
         http_port: Option<u16>,
         pg_version: u32,
         mode: ComputeMode,
-        pageserver_id: NodeId,
     ) -> Result<Arc<Endpoint>> {
         let pg_port = pg_port.unwrap_or_else(|| self.get_port());
         let http_port = http_port.unwrap_or_else(|| self.get_port() + 1);
-        let pageserver =
-            PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
-
         let ep = Arc::new(Endpoint {
             endpoint_id: endpoint_id.to_owned(),
             pg_address: SocketAddr::new("127.0.0.1".parse().unwrap(), pg_port),
             http_address: SocketAddr::new("127.0.0.1".parse().unwrap(), http_port),
             env: self.env.clone(),
-            pageserver,
             timeline_id,
             mode,
             tenant_id,
@@ -157,7 +154,6 @@ impl ComputeControlPlane {
                 pg_port,
                 pg_version,
                 skip_pg_catalog_updates: true,
-                pageserver_id,
             })?,
         )?;
         std::fs::write(
@@ -216,7 +212,6 @@ pub struct Endpoint {
     // These are not part of the endpoint as such, but the environment
     // the endpoint runs in.
     pub env: LocalEnv,
-    pageserver: PageServerNode,
 
     // Optimizations
     skip_pg_catalog_updates: bool,
@@ -239,15 +234,11 @@ impl Endpoint {
         let conf: EndpointConf =
             serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;
 
-        let pageserver =
-            PageServerNode::from_env(env, env.get_pageserver_conf(conf.pageserver_id)?);
-
         Ok(Endpoint {
             pg_address: SocketAddr::new("127.0.0.1".parse().unwrap(), conf.pg_port),
             http_address: SocketAddr::new("127.0.0.1".parse().unwrap(), conf.http_port),
             endpoint_id,
             env: env.clone(),
-            pageserver,
             timeline_id: conf.timeline_id,
             mode: conf.mode,
             tenant_id: conf.tenant_id,
@@ -439,11 +430,14 @@ impl Endpoint {
         Ok(())
     }
 
-    fn wait_for_compute_ctl_to_exit(&self) -> Result<()> {
+    fn wait_for_compute_ctl_to_exit(&self, send_sigterm: bool) -> Result<()> {
         // TODO use background_process::stop_process instead
         let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
         let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
         let pid = nix::unistd::Pid::from_raw(pid as i32);
+        if send_sigterm {
+            kill(pid, Signal::SIGTERM).ok();
+        }
         crate::background_process::wait_until_stopped("compute_ctl", pid)?;
         Ok(())
     }
@@ -464,11 +458,21 @@ impl Endpoint {
         }
     }
 
+    fn build_pageserver_connstr(pageservers: &[(Host, u16)]) -> String {
+        pageservers
+            .iter()
+            .map(|(host, port)| format!("postgresql://no_user@{host}:{port}"))
+            .collect::<Vec<_>>()
+            .join(",")
+    }
+
     pub async fn start(
         &self,
         auth_token: &Option<String>,
         safekeepers: Vec<NodeId>,
+        pageservers: Vec<(Host, u16)>,
         remote_ext_config: Option<&String>,
+        shard_stripe_size: Option<usize>,
     ) -> Result<()> {
         if self.status() == "running" {
             anyhow::bail!("The endpoint is already running");
@@ -482,13 +486,9 @@ impl Endpoint {
             std::fs::remove_dir_all(self.pgdata())?;
         }
 
-        let pageserver_connstring = {
-            let config = &self.pageserver.pg_connection_config;
-            let (host, port) = (config.host(), config.port());
+        let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
+        assert!(!pageserver_connstring.is_empty());
 
-            // NOTE: avoid spaces in connection string, because it is less error prone if we forward it somewhere.
-            format!("postgresql://no_user@{host}:{port}")
-        };
         let mut safekeeper_connstrings = Vec::new();
         if self.mode == ComputeMode::Primary {
             for sk_id in safekeepers {
@@ -537,6 +537,8 @@ impl Endpoint {
             safekeeper_connstrings,
             storage_auth_token: auth_token.clone(),
             remote_extensions,
+            pgbouncer_settings: None,
+            shard_stripe_size,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -549,8 +551,11 @@ impl Endpoint {
 
         // Launch compute_ctl
         println!("Starting postgres node at '{}'", self.connstr());
-        let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
-        cmd.args(["--http-port", &self.http_address.port().to_string()])
+        let mut cmd = Command::new("/usr/bin/taskset");
+
+        cmd.args(["-c".to_string(), "8-11".to_string()])
+            .args([self.env.neon_distrib_dir.join("compute_ctl")])
+            .args(["--http-port", &self.http_address.port().to_string()])
             .args(["--pgdata", self.pgdata().to_str().unwrap()])
             .args(["--connstr", &self.connstr()])
             .args([
@@ -659,7 +664,7 @@ impl Endpoint {
         }
     }
 
-    pub async fn reconfigure(&self, pageserver_id: Option<NodeId>) -> Result<()> {
+    pub async fn reconfigure(&self, mut pageservers: Vec<(Host, u16)>) -> Result<()> {
         let mut spec: ComputeSpec = {
             let spec_path = self.endpoint_path().join("spec.json");
             let file = std::fs::File::open(spec_path)?;
@@ -669,25 +674,27 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        if let Some(pageserver_id) = pageserver_id {
-            let endpoint_config_path = self.endpoint_path().join("endpoint.json");
-            let mut endpoint_conf: EndpointConf = {
-                let file = std::fs::File::open(&endpoint_config_path)?;
-                serde_json::from_reader(file)?
-            };
-            endpoint_conf.pageserver_id = pageserver_id;
-            std::fs::write(
-                endpoint_config_path,
-                serde_json::to_string_pretty(&endpoint_conf)?,
-            )?;
-
-            let pageserver =
-                PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
-            let ps_http_conf = &pageserver.pg_connection_config;
-            let (host, port) = (ps_http_conf.host(), ps_http_conf.port());
-            spec.pageserver_connstring = Some(format!("postgresql://no_user@{host}:{port}"));
+        // If we weren't given explicit pageservers, query the attachment service
+        if pageservers.is_empty() {
+            let attachment_service = AttachmentService::from_env(&self.env);
+            let locate_result = attachment_service.tenant_locate(self.tenant_id).await?;
+            pageservers = locate_result
+                .shards
+                .into_iter()
+                .map(|shard| {
+                    (
+                        Host::parse(&shard.listen_pg_addr)
+                            .expect("Attachment service reported bad hostname"),
+                        shard.listen_pg_port,
+                    )
+                })
+                .collect::<Vec<_>>();
         }
 
+        let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
+        assert!(!pageserver_connstr.is_empty());
+        spec.pageserver_connstring = Some(pageserver_connstr);
+
         let client = reqwest::Client::new();
         let response = client
             .post(format!(
@@ -732,10 +739,15 @@ impl Endpoint {
             &None,
         )?;
 
-        // Also wait for the compute_ctl process to die. It might have some cleanup
-        // work to do after postgres stops, like syncing safekeepers, etc.
+        // Also wait for the compute_ctl process to die. It might have some
+        // cleanup work to do after postgres stops, like syncing safekeepers,
+        // etc.
         //
-        self.wait_for_compute_ctl_to_exit()?;
+        // If destroying, send it SIGTERM before waiting. Sometimes we do *not*
+        // want this cleanup: tests intentionally do stop when majority of
+        // safekeepers is down, so sync-safekeepers would hang otherwise. This
+        // could be a separate flag though.
+        self.wait_for_compute_ctl_to_exit(destroy)?;
         if destroy {
             println!(
                 "Destroying postgres data directory '{}'",

control_plane/src/lib.rs

@@ -14,4 +14,3 @@ pub mod local_env;
 pub mod pageserver;
 pub mod postgresql_conf;
 pub mod safekeeper;
-pub mod tenant_migration;

control_plane/src/pageserver.rs

@@ -12,12 +12,15 @@ use std::io::Write;
 use std::num::NonZeroU64;
 use std::path::PathBuf;
 use std::process::{Child, Command};
+use std::str::FromStr;
 use std::time::Duration;
 
 use anyhow::{bail, Context};
 use camino::Utf8PathBuf;
 use futures::SinkExt;
-use pageserver_api::models::{self, LocationConfig, TenantInfo, TimelineInfo};
+use pageserver_api::models::{
+    self, LocationConfig, ShardParameters, TenantHistorySize, TenantInfo, TimelineInfo,
+};
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api;
 use postgres_backend::AuthType;
@@ -216,11 +219,19 @@ impl PageServerNode {
         if update_config {
             args.push(Cow::Borrowed("--update-config"));
         }
+
+        let mut taskset_args = vec![
+            "-c".to_string(),
+            format!("{}", self.conf.id.0 - 1),
+            self.env.pageserver_bin().to_string_lossy().into(),
+        ];
+        taskset_args.extend(args.into_iter().map(|a| a.to_string()));
+
         background_process::start_process(
             "pageserver",
             &datadir,
-            &self.env.pageserver_bin(),
-            args.iter().map(Cow::as_ref),
+            &PathBuf::from_str("/usr/bin/taskset").unwrap(),
+            taskset_args,
             self.pageserver_env_variables()?,
             background_process::InitialPidFile::Expect(self.pid_file()),
             || async {
@@ -301,16 +312,8 @@ impl PageServerNode {
     pub async fn tenant_list(&self) -> mgmt_api::Result<Vec<TenantInfo>> {
         self.http_client.list_tenants().await
     }
-
-    pub async fn tenant_create(
-        &self,
-        new_tenant_id: TenantId,
-        generation: Option<u32>,
-        settings: HashMap<&str, &str>,
-    ) -> anyhow::Result<TenantId> {
-        let mut settings = settings.clone();
-
-        let config = models::TenantConfig {
+    pub fn parse_config(mut settings: HashMap<&str, &str>) -> anyhow::Result<models::TenantConfig> {
+        let result = models::TenantConfig {
             checkpoint_distance: settings
                 .remove("checkpoint_distance")
                 .map(|x| x.parse::<u64>())
@@ -371,11 +374,26 @@ impl PageServerNode {
                 .context("Failed to parse 'gc_feedback' as bool")?,
             heatmap_period: settings.remove("heatmap_period").map(|x| x.to_string()),
         };
+        if !settings.is_empty() {
+            bail!("Unrecognized tenant settings: {settings:?}")
+        } else {
+            Ok(result)
+        }
+    }
+
+    pub async fn tenant_create(
+        &self,
+        new_tenant_id: TenantId,
+        generation: Option<u32>,
+        settings: HashMap<&str, &str>,
+    ) -> anyhow::Result<TenantId> {
+        let config = Self::parse_config(settings.clone())?;
 
         let request = models::TenantCreateRequest {
             new_tenant_id: TenantShardId::unsharded(new_tenant_id),
             generation,
             config,
+            shard_parameters: ShardParameters::default(),
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -471,31 +489,32 @@ impl PageServerNode {
 
     pub async fn location_config(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         config: LocationConfig,
         flush_ms: Option<Duration>,
     ) -> anyhow::Result<()> {
         Ok(self
             .http_client
-            .location_config(tenant_id, config, flush_ms)
+            .location_config(tenant_shard_id, config, flush_ms)
             .await?)
     }
 
-    pub async fn timeline_list(&self, tenant_id: &TenantId) -> anyhow::Result<Vec<TimelineInfo>> {
-        Ok(self.http_client.list_timelines(*tenant_id).await?)
+    pub async fn timeline_list(
+        &self,
+        tenant_shard_id: &TenantShardId,
+    ) -> anyhow::Result<Vec<TimelineInfo>> {
+        Ok(self.http_client.list_timelines(*tenant_shard_id).await?)
     }
 
     pub async fn timeline_create(
         &self,
         tenant_id: TenantId,
-        new_timeline_id: Option<TimelineId>,
+        new_timeline_id: TimelineId,
         ancestor_start_lsn: Option<Lsn>,
         ancestor_timeline_id: Option<TimelineId>,
         pg_version: Option<u32>,
         existing_initdb_timeline_id: Option<TimelineId>,
     ) -> anyhow::Result<TimelineInfo> {
-        // If timeline ID was not specified, generate one
-        let new_timeline_id = new_timeline_id.unwrap_or(TimelineId::generate());
         let req = models::TimelineCreateRequest {
             new_timeline_id,
             ancestor_start_lsn,
@@ -581,4 +600,14 @@ impl PageServerNode {
 
         Ok(())
     }
+
+    pub async fn tenant_synthetic_size(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> anyhow::Result<TenantHistorySize> {
+        Ok(self
+            .http_client
+            .tenant_synthetic_size(tenant_shard_id)
+            .await?)
+    }
 }

control_plane/src/tenant_migration.rs

@@ -1,205 +0,0 @@
-//!
-//! Functionality for migrating tenants across pageservers: unlike most of neon_local, this code
-//! isn't scoped to a particular physical service, as it needs to update compute endpoints to
-//! point to the new pageserver.
-//!
-use crate::local_env::LocalEnv;
-use crate::{
-    attachment_service::AttachmentService, endpoint::ComputeControlPlane,
-    pageserver::PageServerNode,
-};
-use pageserver_api::models::{
-    LocationConfig, LocationConfigMode, LocationConfigSecondary, TenantConfig,
-};
-use std::collections::HashMap;
-use std::time::Duration;
-use utils::{
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-};
-
-/// Given an attached pageserver, retrieve the LSN for all timelines
-async fn get_lsns(
-    tenant_id: TenantId,
-    pageserver: &PageServerNode,
-) -> anyhow::Result<HashMap<TimelineId, Lsn>> {
-    let timelines = pageserver.timeline_list(&tenant_id).await?;
-    Ok(timelines
-        .into_iter()
-        .map(|t| (t.timeline_id, t.last_record_lsn))
-        .collect())
-}
-
-/// Wait for the timeline LSNs on `pageserver` to catch up with or overtake
-/// `baseline`.
-async fn await_lsn(
-    tenant_id: TenantId,
-    pageserver: &PageServerNode,
-    baseline: HashMap<TimelineId, Lsn>,
-) -> anyhow::Result<()> {
-    loop {
-        let latest = match get_lsns(tenant_id, pageserver).await {
-            Ok(l) => l,
-            Err(e) => {
-                println!(
-                    "🕑 Can't get LSNs on pageserver {} yet, waiting ({e})",
-                    pageserver.conf.id
-                );
-                std::thread::sleep(Duration::from_millis(500));
-                continue;
-            }
-        };
-
-        let mut any_behind: bool = false;
-        for (timeline_id, baseline_lsn) in &baseline {
-            match latest.get(timeline_id) {
-                Some(latest_lsn) => {
-                    println!("🕑 LSN origin {baseline_lsn} vs destination {latest_lsn}");
-                    if latest_lsn < baseline_lsn {
-                        any_behind = true;
-                    }
-                }
-                None => {
-                    // Expected timeline isn't yet visible on migration destination.
-                    // (IRL we would have to account for timeline deletion, but this
-                    //  is just test helper)
-                    any_behind = true;
-                }
-            }
-        }
-
-        if !any_behind {
-            println!("✅ LSN caught up.  Proceeding...");
-            break;
-        } else {
-            std::thread::sleep(Duration::from_millis(500));
-        }
-    }
-
-    Ok(())
-}
-
-/// This function spans multiple services, to demonstrate live migration of a tenant
-/// between pageservers:
-///  - Coordinate attach/secondary/detach on pageservers
-///  - call into attachment_service for generations
-///  - reconfigure compute endpoints to point to new attached pageserver
-pub async fn migrate_tenant(
-    env: &LocalEnv,
-    tenant_id: TenantId,
-    dest_ps: PageServerNode,
-) -> anyhow::Result<()> {
-    // Get a new generation
-    let attachment_service = AttachmentService::from_env(env);
-
-    fn build_location_config(
-        mode: LocationConfigMode,
-        generation: Option<u32>,
-        secondary_conf: Option<LocationConfigSecondary>,
-    ) -> LocationConfig {
-        LocationConfig {
-            mode,
-            generation,
-            secondary_conf,
-            tenant_conf: TenantConfig::default(),
-            shard_number: 0,
-            shard_count: 0,
-            shard_stripe_size: 0,
-        }
-    }
-
-    let previous = attachment_service.inspect(tenant_id).await?;
-    let mut baseline_lsns = None;
-    if let Some((generation, origin_ps_id)) = &previous {
-        let origin_ps = PageServerNode::from_env(env, env.get_pageserver_conf(*origin_ps_id)?);
-
-        if origin_ps_id == &dest_ps.conf.id {
-            println!("🔁 Already attached to {origin_ps_id}, freshening...");
-            let gen = attachment_service
-                .attach_hook(tenant_id, dest_ps.conf.id)
-                .await?;
-            let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
-            dest_ps.location_config(tenant_id, dest_conf, None).await?;
-            println!("✅ Migration complete");
-            return Ok(());
-        }
-
-        println!("🔁 Switching origin pageserver {origin_ps_id} to stale mode");
-
-        let stale_conf =
-            build_location_config(LocationConfigMode::AttachedStale, Some(*generation), None);
-        origin_ps
-            .location_config(tenant_id, stale_conf, Some(Duration::from_secs(10)))
-            .await?;
-
-        baseline_lsns = Some(get_lsns(tenant_id, &origin_ps).await?);
-    }
-
-    let gen = attachment_service
-        .attach_hook(tenant_id, dest_ps.conf.id)
-        .await?;
-    let dest_conf = build_location_config(LocationConfigMode::AttachedMulti, gen, None);
-
-    println!("🔁 Attaching to pageserver {}", dest_ps.conf.id);
-    dest_ps.location_config(tenant_id, dest_conf, None).await?;
-
-    if let Some(baseline) = baseline_lsns {
-        println!("🕑 Waiting for LSN to catch up...");
-        await_lsn(tenant_id, &dest_ps, baseline).await?;
-    }
-
-    let cplane = ComputeControlPlane::load(env.clone())?;
-    for (endpoint_name, endpoint) in &cplane.endpoints {
-        if endpoint.tenant_id == tenant_id {
-            println!(
-                "🔁 Reconfiguring endpoint {} to use pageserver {}",
-                endpoint_name, dest_ps.conf.id
-            );
-            endpoint.reconfigure(Some(dest_ps.conf.id)).await?;
-        }
-    }
-
-    for other_ps_conf in &env.pageservers {
-        if other_ps_conf.id == dest_ps.conf.id {
-            continue;
-        }
-
-        let other_ps = PageServerNode::from_env(env, other_ps_conf);
-        let other_ps_tenants = other_ps.tenant_list().await?;
-
-        // Check if this tenant is attached
-        let found = other_ps_tenants
-            .into_iter()
-            .map(|t| t.id)
-            .any(|i| i.tenant_id == tenant_id);
-        if !found {
-            continue;
-        }
-
-        // Downgrade to a secondary location
-        let secondary_conf = build_location_config(
-            LocationConfigMode::Secondary,
-            None,
-            Some(LocationConfigSecondary { warm: true }),
-        );
-
-        println!(
-            "💤 Switching to secondary mode on pageserver {}",
-            other_ps.conf.id
-        );
-        other_ps
-            .location_config(tenant_id, secondary_conf, None)
-            .await?;
-    }
-
-    println!(
-        "🔁 Switching to AttachedSingle mode on pageserver {}",
-        dest_ps.conf.id
-    );
-    let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
-    dest_ps.location_config(tenant_id, dest_conf, None).await?;
-
-    println!("✅ Migration complete");
-
-    Ok(())
-}

deny.toml

@@ -35,6 +35,7 @@ allow = [
     "Artistic-2.0",
     "BSD-2-Clause",
     "BSD-3-Clause",
+    "CC0-1.0",
     "ISC",
     "MIT",
     "MPL-2.0",

docs/rfcs/029-getpage-throttling.md

@@ -0,0 +1,197 @@
+# Per-Tenant GetPage@LSN Throttling
+
+Author: Christian Schwarz
+Date: Oct 24, 2023
+
+## Summary
+
+This RFC proposes per-tenant throttling of GetPage@LSN requests inside Pageserver
+and the interactions with its client, i.e., the neon_smgr component in Compute.
+
+The result of implementing & executing this RFC will be a fleet-wide upper limit for
+**"the highest GetPage/second that Pageserver can support for a single tenant/shard"**.
+
+## Background
+
+### GetPage@LSN Request Flow
+
+Pageserver exposes its `page_service.rs` as a libpq listener.
+The Computes' `neon_smgr` module connects to that libpq listener.
+Once a connection is established, the protocol allows Compute to request page images at a given LSN.
+We call these requests GetPage@LSN requests, or GetPage requests for short.
+Other request types can be sent, but these are low traffic compared to GetPage requests
+and are not the concern of this RFC.
+
+Pageserver associates one libpq connection with one tokio task.
+
+Per connection/task, the pq protocol is handled by the common `postgres_backend` crate.
+Its `run_message_loop` function invokes the `page_service` specific `impl<IO> postgres_backend::Handler<IO> for PageServerHandler`.
+Requests are processed in the order in which they arrive via the TCP-based pq protocol.
+So, there is no concurrent request processing within one connection/task.
+
+There is a degree of natural pipelining:
+Compute can "fill the pipe" by sending more than one GetPage request into the libpq TCP stream.
+And Pageserver can fill the pipe with responses in the other direction.
+Both directions are subject to the limit of tx/rx buffers, nodelay, TCP flow control, etc.
+
+### GetPage@LSN Access Pattern
+
+The Compute has its own hierarchy of caches, specifically `shared_buffers` and the `local file cache` (LFC).
+Compute only issues GetPage requests to Pageserver if it encounters a miss in these caches.
+
+If the working set stops fitting into Compute's caches, requests to Pageserver increase sharply -- the Compute starts *thrashing*.
+
+## Motivation
+
+In INC-69, a tenant issued 155k GetPage/second for a period of 10 minutes and 60k GetPage/second for a period of 3h,
+then dropping to ca 18k GetPage/second for a period of 9h.
+
+We noticed this because of an internal GetPage latency SLO burn rate alert, i.e.,
+the request latency profile during this period significantly exceeded what was acceptable according to the internal SLO.
+
+Sadly, we do not have the observability data to determine the impact of this tenant on other tenants on the same tenants.
+
+However, here are some illustrative data points for the 155k period:
+The tenant was responsible for >= 99% of the GetPage traffic and, frankly, the overall activity on this Pageserver instance.
+We were serving pages at 10 Gb/s (`155k x 8 kbyte (PAGE_SZ) per second is 1.12GiB/s = 9.4Gb/s.`)
+The CPU utilization of the instance was 75% user+system.
+Pageserver page cache served 1.75M accesses/second at a hit rate of ca 90%.
+The hit rate for materialized pages was ca. 40%.
+Curiously, IOPS to the Instance Store NVMe were very low, rarely exceeding 100.
+
+The fact that the IOPS were so low / the materialized page cache hit rate was so high suggests that **this tenant's compute's caches were thrashing**.
+The compute was of type `k8s-pod`; hence, auto-scaling could/would not have helped remediate the thrashing by provisioning more RAM.
+The consequence was that the **thrashing translated into excessive GetPage requests against Pageserver**.
+
+My claim is that it was **unhealthy to serve this workload at the pace we did**:
+* it is likely that other tenants were/would have experienced high latencies (again, we sadly don't have per-tenant latency data to confirm this)
+* more importantly, it was **unsustainable** to serve traffic at this pace for multiple reasons:
+    * **predictability of performance**: when the working set grows, the pageserver materialized page cache hit rate drops.
+      At some point, we're bound by the EC2 Instance Store NVMe drive's IOPS limit.
+      The result is an **uneven** performance profile from the Compute perspective.
+
+    * **economics**: Neon currently does not charge for IOPS, only capacity.
+      **We cannot afford to undercut the market in IOPS/$ this drastically; it leads to adverse selection and perverse incentives.**
+      For example, the 155k IOPS, which we served for 10min, would cost ca. 6.5k$/month when provisioned as an io2 EBS volume.
+      Even the 18k IOPS, which we served for 9h, would cost ca. 1.1k$/month when provisioned as an io2 EBS volume.
+      We charge 0$.
+      It could be economically advantageous to keep using a low-DRAM compute because Pageserver IOPS are fast enough and free.
+
+
+Note: It is helpful to think of Pageserver as a disk, because it's precisely where `neon_smgr` sits:
+vanilla Postgres gets its pages from disk, Neon Postgres gets them from Pageserver.
+So, regarding the above performance & economic arguments, it is fair to say that we currently provide an "as-fast-as-possible-IOPS" disk that we charge for only by capacity.
+
+## Solution: Throttling GetPage Requests
+
+**The consequence of the above analysis must be that Pageserver throttles GetPage@LSN requests**.
+That is, unless we want to start charging for provisioned GetPage@LSN/second.
+Throttling sets the correct incentive for a thrashing Compute to scale up its DRAM to the working set size.
+Neon Autoscaling will make this easy, [eventually](https://github.com/neondatabase/neon/pull/3913).
+
+## The Design Space
+
+What that remains is the question about *policy* and *mechanism*:
+
+**Policy** concerns itself with the question of what limit applies to a given connection|timeline|tenant.
+Candidates are:
+
+* hard limit, same limit value per connection|timeline|tenant
+    * Per-tenant will provide an upper bound for the impact of a tenant on a given Pageserver instance.
+      This is a major operational pain point / risk right now.
+* hard limit, configurable per connection|timeline|tenant
+    * This outsources policy to console/control plane, with obvious advantages for flexible structuring of what service we offer to customers.
+    * Note that this is not a mechanism to guarantee a minium provisioned rate, i.e., this is not a mechanism to guarantee a certain QoS for a tenant.
+* fair share among active connections|timelines|tenants per instance
+    * example: each connection|timeline|tenant gets a fair fraction of the machine's GetPage/second capacity
+    * NB: needs definition of "active", and knowledge of available GetPage/second capacity in advance
+* ...
+
+
+Regarding **mechanism**, it's clear that **backpressure** is the way to go.
+However, we must choose between
+* **implicit** backpressure through pq/TCP and
+* **explicit** rejection of requests + retries with exponential backoff
+
+Further, there is the question of how throttling GetPage@LSN will affect the **internal GetPage latency SLO**:
+where do we measure the SLI for Pageserver's internal getpage latency SLO? Before or after the throttling?
+
+And when we eventually move the measurement point into the Computes (to avoid coordinated omission),
+how do we avoid counting throttling-induced latency toward the internal getpage latency SLI/SLO?
+
+## Scope Of This RFC
+
+**This RFC proposes introducing a hard GetPage@LSN/second limit per tenant, with the same value applying to each tenant on a Pageserver**.
+
+This proposal is easy to implement and significantly de-risks operating large Pageservers,
+based on the assumption that extremely-high-GetPage-rate-episodes like the one from the "Motivation" section are uncorrelated between tenants.
+
+For example, suppose we pick a limit that allows up to 10 tenants to go at limit rate.
+Suppose our Pageserver can serve 100k GetPage/second total at a 100% page cache miss rate.
+If each tenant gets a hard limit of 10k GetPage/second, we can serve up to 10 tenants at limit speed without latency degradation.
+
+The mechanism for backpressure will be TCP-based implicit backpressure.
+The compute team isn't concerned about prefetch queue depth.
+Pageserver will implement it by delaying the reading of requests from the libpq connection(s).
+
+The rate limit will be implemented using a per-tenant token bucket.
+The bucket will be be shared among all connections to the tenant.
+The bucket implementation supports starvation-preventing `await`ing.
+The current candidate for the implementation is [`leaky_bucket`](https://docs.rs/leaky-bucket/).
+The getpage@lsn benchmark that's being added in https://github.com/neondatabase/neon/issues/5771
+can be used to evaluate the overhead of sharing the bucket among connections of a tenant.
+A possible technique to mitigate the impact of sharing the bucket would be to maintain a buffer of a few tokens per connection handler.
+
+Regarding metrics / the internal GetPage latency SLO:
+we will measure the GetPage latency SLO _after_ the throttler and introduce a new metric to measure the amount of throttling, quantified by:
+- histogram that records the tenants' observations of queue depth before they start waiting (one such histogram per pageserver)
+- histogram that records the tenants' observations of time spent waiting (one such histogram per pageserver)
+
+Further observability measures:
+- an INFO log message at frequency 1/min if the tenant/timeline/connection was throttled in that last minute.
+  The message will identify the tenant/timeline/connection to allow correlation with compute logs/stats.
+
+Rollout will happen as follows:
+- deploy 1: implementation + config: disabled by default, ability to enable it per tenant through tenant_conf
+- experimentation in staging and later production to study impact & interaction with auto-scaling
+- determination of a sensible global default value
+  - the value will be chosen as high as possible ...
+  - ... but low enough to work towards this RFC's goal that one tenant should not be able to dominate a pageserver instance.
+- deploy 2: implementation fixes if any + config: enabled by default with the aforementioned global default
+- reset of the experimental per-tenant overrides
+- gain experience & lower the limit over time
+  - we stop lowering the limit as soon as this RFC's goal is achieved, i.e.,
+    once we decide that in practice the chosen value sufficiently de-risks operating large pageservers
+
+The per-tenant override will remain for emergencies and testing.
+But since Console doesn't preserve it during tenant migrations, it isn't durably configurable for the tenant.
+
+Toward the upper layers of the Neon stack, the resulting limit will be
+**"the highest GetPage/second that Pageserver can support for a single tenant"**.
+
+### Rationale
+
+We decided against error + retry because of worries about starvation.
+
+## Future Work
+
+Enable per-tenant emergency override of the limit via Console.
+Should be part of a more general framework to specify tenant config overrides.
+**NB:** this is **not** the right mechanism to _sell_ different max GetPage/second levels to users,
+or _auto-scale_ the GetPage/second levels. Such functionality will require a separate RFC that
+concerns itself with GetPage/second capacity planning.
+
+Compute-side metrics for GetPage latency.
+
+Back-channel to inform Compute/Autoscaling/ControlPlane that the project is being throttled.
+
+Compute-side neon_smgr improvements to avoid sending the same GetPage request multiple times if multiple backends experience a cache miss.
+
+Dealing with read-only endpoints: users use read-only endpoints to scale reads for a single tenant.
+Possibly there are also assumptions around read-only endpoints not affecting the primary read-write endpoint's performance.
+With per-tenant rate limiting, we will not meet that expectation.
+However, we can currently only scale per tenant.
+Soon, we will have sharding (#5505), which will apply the throttling on a per-shard basis.
+But, that's orthogonal to scaling reads: if many endpoints hit one shard, they share the same throttling limit.
+To solve this properly, I think we'll need replicas for tenants / shard.
+To performance-isolate a tenant's endpoints from each other, we'd then route them to different replicas.

libs/compute_api/src/spec.rs

@@ -73,6 +73,11 @@ pub struct ComputeSpec {
 
     // information about available remote extensions
     pub remote_extensions: Option<RemoteExtSpec>,
+
+    pub pgbouncer_settings: Option<HashMap<String, String>>,
+
+    // Stripe size for pageserver sharding, in pages
+    pub shard_stripe_size: Option<usize>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.

libs/compute_api/tests/cluster_spec.json

@@ -243,5 +243,9 @@
         "public_extensions": [
           "postgis"
         ]
+      },
+      "pgbouncer_settings": {
+        "default_pool_size": "42",
+        "pool_mode": "session"
       }
 }

libs/pageserver_api/src/keyspace.rs

@@ -124,6 +124,9 @@ impl KeySpaceAccum {
                 if range.start == accum.end {
                     accum.end = range.end;
                 } else {
+                    // TODO: to efficiently support small sharding stripe sizes, we should avoid starting
+                    // a new range here if the skipped region was all keys that don't belong on this shard.
+                    // (https://github.com/neondatabase/neon/issues/6247)
                     assert!(range.start > accum.end);
                     self.ranges.push(accum.clone());
                     *accum = range;

libs/pageserver_api/src/models.rs

@@ -2,6 +2,7 @@ pub mod partitioning;
 
 use std::{
     collections::HashMap,
+    io::Read,
     num::{NonZeroU64, NonZeroUsize},
     time::SystemTime,
 };
@@ -17,9 +18,12 @@ use utils::{
     lsn::Lsn,
 };
 
-use crate::{reltag::RelTag, shard::TenantShardId};
+use crate::{
+    reltag::RelTag,
+    shard::{ShardCount, ShardStripeSize, TenantShardId},
+};
 use anyhow::bail;
-use bytes::{BufMut, Bytes, BytesMut};
+use bytes::{Buf, BufMut, Bytes, BytesMut};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -187,6 +191,41 @@ pub struct TimelineCreateRequest {
     pub pg_version: Option<u32>,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantShardSplitRequest {
+    pub new_shard_count: u8,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantShardSplitResponse {
+    pub new_shards: Vec<TenantShardId>,
+}
+
+/// Parameters that apply to all shards in a tenant.  Used during tenant creation.
+#[derive(Serialize, Deserialize, Debug)]
+#[serde(deny_unknown_fields)]
+pub struct ShardParameters {
+    pub count: ShardCount,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub stripe_size: Option<ShardStripeSize>,
+}
+
+impl ShardParameters {
+    pub fn is_unsharded(&self) -> bool {
+        self.count == ShardCount(0)
+    }
+}
+
+impl Default for ShardParameters {
+    fn default() -> Self {
+        Self {
+            count: ShardCount(0),
+            stripe_size: None,
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantCreateRequest {
@@ -194,6 +233,12 @@ pub struct TenantCreateRequest {
     #[serde(default)]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub generation: Option<u32>,
+
+    // If omitted, create a single shard with TenantShardId::unsharded()
+    #[serde(default)]
+    #[serde(skip_serializing_if = "ShardParameters::is_unsharded")]
+    pub shard_parameters: ShardParameters,
+
     #[serde(flatten)]
     pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
@@ -216,7 +261,7 @@ impl std::ops::Deref for TenantCreateRequest {
 
 /// An alternative representation of `pageserver::tenant::TenantConf` with
 /// simpler types.
-#[derive(Serialize, Deserialize, Debug, Default)]
+#[derive(Serialize, Deserialize, Debug, Default, Clone, Eq, PartialEq)]
 pub struct TenantConfig {
     pub checkpoint_distance: Option<u64>,
     pub checkpoint_timeout: Option<String>,
@@ -245,7 +290,7 @@ pub struct TenantConfig {
 /// A flattened analog of a `pagesever::tenant::LocationMode`, which
 /// lists out all possible states (and the virtual "Detached" state)
 /// in a flat form rather than using rust-style enums.
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq)]
 pub enum LocationConfigMode {
     AttachedSingle,
     AttachedMulti,
@@ -254,14 +299,14 @@ pub enum LocationConfigMode {
     Detached,
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq)]
 pub struct LocationConfigSecondary {
     pub warm: bool,
 }
 
 /// An alternative representation of `pageserver::tenant::LocationConf`,
 /// for use in external-facing APIs.
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq)]
 pub struct LocationConfig {
     pub mode: LocationConfigMode,
     /// If attaching, in what generation?
@@ -296,7 +341,7 @@ pub struct StatusResponse {
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantLocationConfigRequest {
-    pub tenant_id: TenantId,
+    pub tenant_id: TenantShardId,
     #[serde(flatten)]
     pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
@@ -369,6 +414,14 @@ pub struct TenantInfo {
     pub attachment_status: TenantAttachmentStatus,
 }
 
+#[derive(Serialize, Deserialize, Clone)]
+pub struct TenantDetails {
+    #[serde(flatten)]
+    pub tenant_info: TenantInfo,
+
+    pub timelines: Vec<TimelineId>,
+}
+
 /// This represents the output of the "timeline_detail" and "timeline_list" API calls.
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct TimelineInfo {
@@ -548,19 +601,6 @@ pub enum DownloadRemoteLayersTaskState {
     ShutDown,
 }
 
-pub type ConfigureFailpointsRequest = Vec<FailpointConfig>;
-
-/// Information for configuring a single fail point
-#[derive(Debug, Serialize, Deserialize)]
-pub struct FailpointConfig {
-    /// Name of the fail point
-    pub name: String,
-    /// List of actions to take, using the format described in `fail::cfg`
-    ///
-    /// We also support `actions = "exit"` to cause the fail point to immediately exit.
-    pub actions: String,
-}
-
 #[derive(Debug, Serialize, Deserialize)]
 pub struct TimelineGcRequest {
     pub gc_horizon: Option<u64>,
@@ -576,6 +616,7 @@ pub enum PagestreamFeMessage {
 }
 
 // Wrapped in libpq CopyData
+#[derive(strum_macros::EnumProperty)]
 pub enum PagestreamBeMessage {
     Exists(PagestreamExistsResponse),
     Nblocks(PagestreamNblocksResponse),
@@ -584,6 +625,29 @@ pub enum PagestreamBeMessage {
     DbSize(PagestreamDbSizeResponse),
 }
 
+// Keep in sync with `pagestore_client.h`
+#[repr(u8)]
+enum PagestreamBeMessageTag {
+    Exists = 100,
+    Nblocks = 101,
+    GetPage = 102,
+    Error = 103,
+    DbSize = 104,
+}
+impl TryFrom<u8> for PagestreamBeMessageTag {
+    type Error = u8;
+    fn try_from(value: u8) -> Result<Self, u8> {
+        match value {
+            100 => Ok(PagestreamBeMessageTag::Exists),
+            101 => Ok(PagestreamBeMessageTag::Nblocks),
+            102 => Ok(PagestreamBeMessageTag::GetPage),
+            103 => Ok(PagestreamBeMessageTag::Error),
+            104 => Ok(PagestreamBeMessageTag::DbSize),
+            _ => Err(value),
+        }
+    }
+}
+
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamExistsRequest {
     pub latest: bool,
@@ -638,6 +702,17 @@ pub struct PagestreamDbSizeResponse {
     pub db_size: i64,
 }
 
+// This is a cut-down version of TenantHistorySize from the pageserver crate, omitting fields
+// that require pageserver-internal types.  It is sufficient to get the total size.
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantHistorySize {
+    pub id: TenantId,
+    /// Size is a mixture of WAL and logical size, so the unit is bytes.
+    ///
+    /// Will be none if `?inputs_only=true` was given.
+    pub size: Option<u64>,
+}
+
 impl PagestreamFeMessage {
     pub fn serialize(&self) -> Bytes {
         let mut bytes = BytesMut::new();
@@ -739,35 +814,91 @@ impl PagestreamBeMessage {
     pub fn serialize(&self) -> Bytes {
         let mut bytes = BytesMut::new();
 
+        use PagestreamBeMessageTag as Tag;
         match self {
             Self::Exists(resp) => {
-                bytes.put_u8(100); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Exists as u8);
                 bytes.put_u8(resp.exists as u8);
             }
 
             Self::Nblocks(resp) => {
-                bytes.put_u8(101); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Nblocks as u8);
                 bytes.put_u32(resp.n_blocks);
             }
 
             Self::GetPage(resp) => {
-                bytes.put_u8(102); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::GetPage as u8);
                 bytes.put(&resp.page[..]);
             }
 
             Self::Error(resp) => {
-                bytes.put_u8(103); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::Error as u8);
                 bytes.put(resp.message.as_bytes());
                 bytes.put_u8(0); // null terminator
             }
             Self::DbSize(resp) => {
-                bytes.put_u8(104); /* tag from pagestore_client.h */
+                bytes.put_u8(Tag::DbSize as u8);
                 bytes.put_i64(resp.db_size);
             }
         }
 
         bytes.into()
     }
+
+    pub fn deserialize(buf: Bytes) -> anyhow::Result<Self> {
+        let mut buf = buf.reader();
+        let msg_tag = buf.read_u8()?;
+
+        use PagestreamBeMessageTag as Tag;
+        let ok =
+            match Tag::try_from(msg_tag).map_err(|tag: u8| anyhow::anyhow!("invalid tag {tag}"))? {
+                Tag::Exists => {
+                    let exists = buf.read_u8()?;
+                    Self::Exists(PagestreamExistsResponse {
+                        exists: exists != 0,
+                    })
+                }
+                Tag::Nblocks => {
+                    let n_blocks = buf.read_u32::<BigEndian>()?;
+                    Self::Nblocks(PagestreamNblocksResponse { n_blocks })
+                }
+                Tag::GetPage => {
+                    let mut page = vec![0; 8192]; // TODO: use MaybeUninit
+                    buf.read_exact(&mut page)?;
+                    PagestreamBeMessage::GetPage(PagestreamGetPageResponse { page: page.into() })
+                }
+                Tag::Error => {
+                    let buf = buf.get_ref();
+                    let cstr = std::ffi::CStr::from_bytes_until_nul(buf)?;
+                    let rust_str = cstr.to_str()?;
+                    PagestreamBeMessage::Error(PagestreamErrorResponse {
+                        message: rust_str.to_owned(),
+                    })
+                }
+                Tag::DbSize => {
+                    let db_size = buf.read_i64::<BigEndian>()?;
+                    Self::DbSize(PagestreamDbSizeResponse { db_size })
+                }
+            };
+        let remaining = buf.into_inner();
+        if !remaining.is_empty() {
+            anyhow::bail!(
+                "remaining bytes in msg with tag={msg_tag}: {}",
+                remaining.len()
+            );
+        }
+        Ok(ok)
+    }
+
+    pub fn kind(&self) -> &'static str {
+        match self {
+            Self::Exists(_) => "Exists",
+            Self::Nblocks(_) => "Nblocks",
+            Self::GetPage(_) => "GetPage",
+            Self::Error(_) => "Error",
+            Self::DbSize(_) => "DbSize",
+        }
+    }
 }
 
 #[cfg(test)]

libs/pageserver_api/src/shard.rs

@@ -1,6 +1,9 @@
 use std::{ops::RangeInclusive, str::FromStr};
 
-use crate::key::{is_rel_block_key, Key};
+use crate::{
+    key::{is_rel_block_key, Key},
+    models::ShardParameters,
+};
 use hex::FromHex;
 use serde::{Deserialize, Serialize};
 use thiserror;
@@ -81,6 +84,20 @@ impl TenantShardId {
     pub fn is_zero(&self) -> bool {
         self.shard_number == ShardNumber(0)
     }
+
+    pub fn is_unsharded(&self) -> bool {
+        self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
+    }
+
+    /// Convenience for dropping the tenant_id and just getting the ShardIndex: this
+    /// is useful when logging from code that is already in a span that includes tenant ID, to
+    /// keep messages reasonably terse.
+    pub fn to_index(&self) -> ShardIndex {
+        ShardIndex {
+            shard_number: self.shard_number,
+            shard_count: self.shard_count,
+        }
+    }
 }
 
 /// Formatting helper
@@ -159,7 +176,7 @@ impl From<[u8; 18]> for TenantShardId {
 /// shard we're dealing with, but do not need to know the full ShardIdentity (because
 /// we won't be doing any page->shard mapping), and do not need to know the fully qualified
 /// TenantShardId.
-#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy)]
+#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
 pub struct ShardIndex {
     pub shard_number: ShardNumber,
     pub shard_count: ShardCount,
@@ -329,7 +346,7 @@ const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
 pub struct ShardIdentity {
     pub number: ShardNumber,
     pub count: ShardCount,
-    stripe_size: ShardStripeSize,
+    pub stripe_size: ShardStripeSize,
     layout: ShardLayout,
 }
 
@@ -399,6 +416,17 @@ impl ShardIdentity {
         }
     }
 
+    /// For use when creating ShardIdentity instances for new shards, where a creation request
+    /// specifies the ShardParameters that apply to all shards.
+    pub fn from_params(number: ShardNumber, params: &ShardParameters) -> Self {
+        Self {
+            number,
+            count: params.count,
+            layout: LAYOUT_V1,
+            stripe_size: params.stripe_size.unwrap_or(DEFAULT_STRIPE_SIZE),
+        }
+    }
+
     fn is_broken(&self) -> bool {
         self.layout == LAYOUT_BROKEN
     }
@@ -418,6 +446,21 @@ impl ShardIdentity {
         }
     }
 
+    /// Return true if the key should be discarded if found in this shard's
+    /// data store, e.g. during compaction after a split
+    pub fn is_key_disposable(&self, key: &Key) -> bool {
+        if key_is_shard0(key) {
+            // Q: Why can't we dispose of shard0 content if we're not shard 0?
+            // A: because the WAL ingestion logic currently ingests some shard 0
+            //    content on all shards, even though it's only read on shard 0.  If we
+            //    dropped it, then subsequent WAL ingest to these keys would encounter
+            //    an error.
+            false
+        } else {
+            !self.is_key_local(key)
+        }
+    }
+
     pub fn shard_slug(&self) -> String {
         if self.count > ShardCount(0) {
             format!("-{:02x}{:02x}", self.number.0, self.count.0)

libs/remote_storage/src/azure_blob.rs

@@ -117,6 +117,8 @@ impl AzureBlobStorage {
     ) -> Result<Download, DownloadError> {
         let mut response = builder.into_stream();
 
+        let mut etag = None;
+        let mut last_modified = None;
         let mut metadata = HashMap::new();
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
@@ -124,6 +126,13 @@ impl AzureBlobStorage {
         let mut bufs = Vec::new();
         while let Some(part) = response.next().await {
             let part = part.map_err(to_download_error)?;
+            let etag_str: &str = part.blob.properties.etag.as_ref();
+            if etag.is_none() {
+                etag = Some(etag.unwrap_or_else(|| etag_str.to_owned()));
+            }
+            if last_modified.is_none() {
+                last_modified = Some(part.blob.properties.last_modified.into());
+            }
             if let Some(blob_meta) = part.blob.metadata {
                 metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
             }
@@ -136,6 +145,8 @@ impl AzureBlobStorage {
         }
         Ok(Download {
             download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
+            etag,
+            last_modified,
             metadata: Some(StorageMetadata(metadata)),
         })
     }

libs/remote_storage/src/lib.rs

@@ -14,7 +14,9 @@ mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
 
-use std::{collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::Arc};
+use std::{
+    collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::Arc, time::SystemTime,
+};
 
 use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
@@ -207,8 +209,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()>;
 }
 
+pub type DownloadStream = Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>;
 pub struct Download {
-    pub download_stream: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>,
+    pub download_stream: DownloadStream,
+    /// The last time the file was modified (`last-modified` HTTP header)
+    pub last_modified: Option<SystemTime>,
+    /// A way to identify this specific version of the resource (`etag` HTTP header)
+    pub etag: Option<String>,
     /// Extra key-value data, associated with the current remote file.
     pub metadata: Option<StorageMetadata>,
 }

libs/remote_storage/src/local_fs.rs

@@ -18,7 +18,7 @@ use tokio_util::io::ReaderStream;
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
-use crate::{Download, DownloadError, Listing, ListingMode, RemotePath};
+use crate::{Download, DownloadError, DownloadStream, Listing, ListingMode, RemotePath};
 
 use super::{RemoteStorage, StorageMetadata};
 
@@ -331,6 +331,8 @@ impl RemoteStorage for LocalFs {
                 .map_err(DownloadError::Other)?;
             Ok(Download {
                 metadata,
+                last_modified: None,
+                etag: None,
                 download_stream: Box::pin(source),
             })
         } else {
@@ -372,17 +374,17 @@ impl RemoteStorage for LocalFs {
                 .await
                 .map_err(DownloadError::Other)?;
 
-            Ok(match end_exclusive {
-                Some(end_exclusive) => Download {
-                    metadata,
-                    download_stream: Box::pin(ReaderStream::new(
-                        source.take(end_exclusive - start_inclusive),
-                    )),
-                },
-                None => Download {
-                    metadata,
-                    download_stream: Box::pin(ReaderStream::new(source)),
-                },
+            let download_stream: DownloadStream = match end_exclusive {
+                Some(end_exclusive) => Box::pin(ReaderStream::new(
+                    source.take(end_exclusive - start_inclusive),
+                )),
+                None => Box::pin(ReaderStream::new(source)),
+            };
+            Ok(Download {
+                metadata,
+                last_modified: None,
+                etag: None,
+                download_stream,
             })
         } else {
             Err(DownloadError::NotFound)

libs/remote_storage/src/s3_bucket.rs

@@ -16,6 +16,7 @@ use aws_config::{
     environment::credentials::EnvironmentVariableCredentialsProvider,
     imds::credentials::ImdsCredentialsProvider,
     meta::credentials::CredentialsProviderChain,
+    profile::ProfileFileCredentialsProvider,
     provider_config::ProviderConfig,
     retry::{RetryConfigBuilder, RetryMode},
     web_identity_token::WebIdentityTokenCredentialsProvider,
@@ -74,20 +75,29 @@ impl S3Bucket {
 
         let region = Some(Region::new(aws_config.bucket_region.clone()));
 
+        let provider_conf = ProviderConfig::without_region().with_region(region.clone());
+
         let credentials_provider = {
             // uses "AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"
             CredentialsProviderChain::first_try(
                 "env",
                 EnvironmentVariableCredentialsProvider::new(),
             )
+            // uses "AWS_PROFILE" / `aws sso login --profile <profile>`
+            .or_else(
+                "profile-sso",
+                ProfileFileCredentialsProvider::builder()
+                    .configure(&provider_conf)
+                    .build(),
+            )
             // uses "AWS_WEB_IDENTITY_TOKEN_FILE", "AWS_ROLE_ARN", "AWS_ROLE_SESSION_NAME"
             // needed to access remote extensions bucket
-            .or_else("token", {
-                let provider_conf = ProviderConfig::without_region().with_region(region.clone());
+            .or_else(
+                "token",
                 WebIdentityTokenCredentialsProvider::builder()
                     .configure(&provider_conf)
-                    .build()
-            })
+                    .build(),
+            )
             // uses imds v2
             .or_else("imds", ImdsCredentialsProvider::builder().build())
         };
@@ -218,17 +228,11 @@ impl S3Bucket {
 
         let started_at = ScopeGuard::into_inner(started_at);
 
-        if get_object.is_err() {
-            metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
-                kind,
-                AttemptOutcome::Err,
-                started_at,
-            );
-        }
-
         match get_object {
             Ok(object_output) => {
                 let metadata = object_output.metadata().cloned().map(StorageMetadata);
+                let etag = object_output.e_tag.clone();
+                let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
 
                 let body = object_output.body;
                 let body = ByteStreamAsStream::from(body);
@@ -237,15 +241,33 @@ impl S3Bucket {
 
                 Ok(Download {
                     metadata,
+                    etag,
+                    last_modified,
                     download_stream: Box::pin(body),
                 })
             }
             Err(SdkError::ServiceError(e)) if matches!(e.err(), GetObjectError::NoSuchKey(_)) => {
+                // Count this in the AttemptOutcome::Ok bucket, because 404 is not
+                // an error: we expect to sometimes fetch an object and find it missing,
+                // e.g. when probing for timeline indices.
+                metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
+                    kind,
+                    AttemptOutcome::Ok,
+                    started_at,
+                );
                 Err(DownloadError::NotFound)
             }
-            Err(e) => Err(DownloadError::Other(
-                anyhow::Error::new(e).context("download s3 object"),
-            )),
+            Err(e) => {
+                metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
+                    kind,
+                    AttemptOutcome::Err,
+                    started_at,
+                );
+
+                Err(DownloadError::Other(
+                    anyhow::Error::new(e).context("download s3 object"),
+                ))
+            }
         }
     }
 }

libs/remote_storage/tests/common/mod.rs

@@ -0,0 +1,200 @@
+use std::collections::HashSet;
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use anyhow::Context;
+use bytes::Bytes;
+use camino::Utf8Path;
+use futures::stream::Stream;
+use once_cell::sync::OnceCell;
+use remote_storage::{Download, GenericRemoteStorage, RemotePath};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+pub(crate) fn upload_stream(
+    content: std::borrow::Cow<'static, [u8]>,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    use std::borrow::Cow;
+
+    let content = match content {
+        Cow::Borrowed(x) => Bytes::from_static(x),
+        Cow::Owned(vec) => Bytes::from(vec),
+    };
+    wrap_stream(content)
+}
+
+pub(crate) fn wrap_stream(
+    content: bytes::Bytes,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    let len = content.len();
+    let content = futures::future::ready(Ok(content));
+
+    (futures::stream::once(content), len)
+}
+
+pub(crate) async fn download_to_vec(dl: Download) -> anyhow::Result<Vec<u8>> {
+    let mut buf = Vec::new();
+    tokio::io::copy_buf(
+        &mut tokio_util::io::StreamReader::new(dl.download_stream),
+        &mut buf,
+    )
+    .await?;
+    Ok(buf)
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+pub(crate) async fn upload_simple_remote_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} remote files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, len, &blob_path, None).await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}
+
+pub(crate) async fn cleanup(
+    client: &Arc<GenericRemoteStorage>,
+    objects_to_delete: HashSet<RemotePath>,
+) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+pub(crate) struct Uploads {
+    pub(crate) prefixes: HashSet<RemotePath>,
+    pub(crate) blobs: HashSet<RemotePath>,
+}
+
+pub(crate) async fn upload_remote_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} remote files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let (data, data_len) =
+                upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, data_len, &blob_path, None).await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+pub(crate) fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+            utils::logging::Output::Stdout,
+        )
+        .expect("logging init failed");
+    });
+}

libs/remote_storage/tests/test_real_azure.rs

@@ -2,23 +2,23 @@ use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
-use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
-use bytes::Bytes;
 use camino::Utf8Path;
-use futures::stream::Stream;
-use once_cell::sync::OnceCell;
 use remote_storage::{
-    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+    AzureConfig, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
 };
 use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
+use tracing::{debug, info};
 
-static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+mod common;
+
+use common::{
+    cleanup, download_to_vec, ensure_logging_ready, upload_remote_data, upload_simple_remote_data,
+    upload_stream, wrap_stream,
+};
 
 const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
 
@@ -30,7 +30,7 @@ const BASE_PREFIX: &str = "test";
 /// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
 /// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
 ///
-/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_remote_data`]
 /// where
 /// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
 /// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
@@ -97,7 +97,7 @@ async fn azure_pagination_should_work(
 /// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
 /// See `Azure_pagination_should_work` for more information.
 ///
-/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_remote_data`]
 /// Then performs the following queries:
 ///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
 ///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
@@ -218,18 +218,9 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
 
     ctx.client.upload(data, len, &path, None).await?;
 
-    async fn download_and_compare(dl: Download) -> anyhow::Result<Vec<u8>> {
-        let mut buf = Vec::new();
-        tokio::io::copy_buf(
-            &mut tokio_util::io::StreamReader::new(dl.download_stream),
-            &mut buf,
-        )
-        .await?;
-        Ok(buf)
-    }
     // Normal download request
     let dl = ctx.client.download(&path).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // Full range (end specified)
@@ -237,12 +228,12 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
         .client
         .download_byte_range(&path, 0, Some(len as u64))
         .await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // partial range (end specified)
     let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..10]);
 
     // partial range (end beyond real end)
@@ -250,17 +241,17 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
         .client
         .download_byte_range(&path, 8, Some(len as u64 * 100))
         .await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[8..]);
 
     // Partial range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 4, None).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..]);
 
     // Full range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 0, None).await?;
-    let buf = download_and_compare(dl).await?;
+    let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
@@ -272,17 +263,6 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
     Ok(())
 }
 
-fn ensure_logging_ready() {
-    LOGGING_DONE.get_or_init(|| {
-        utils::logging::init(
-            utils::logging::LogFormat::Test,
-            utils::logging::TracingErrorLayerEnablement::Disabled,
-            utils::logging::Output::Stdout,
-        )
-        .expect("logging init failed");
-    });
-}
-
 struct EnabledAzure {
     client: Arc<GenericRemoteStorage>,
     base_prefix: &'static str,
@@ -352,7 +332,7 @@ impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
 
         let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+        match upload_remote_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -414,7 +394,7 @@ impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
 
         let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+        match upload_simple_remote_data(&enabled.client, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -478,166 +458,3 @@ fn create_azure_client(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_azure_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} Azure files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
-            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}
-
-// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
-async fn upload_simple_azure_data(
-    client: &Arc<GenericRemoteStorage>,
-    upload_tasks_count: usize,
-) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
-    info!("Creating {upload_tasks_count} Azure files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
-            let blob_path = RemotePath::new(
-                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
-            )
-            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>(blob_path)
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok(upload_path) => {
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    if upload_tasks_failed {
-        ControlFlow::Break(uploaded_blobs)
-    } else {
-        ControlFlow::Continue(uploaded_blobs)
-    }
-}
-
-// FIXME: copypasted from test_real_s3, can't remember how to share a module which is not compiled
-// to binary
-fn upload_stream(
-    content: std::borrow::Cow<'static, [u8]>,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    use std::borrow::Cow;
-
-    let content = match content {
-        Cow::Borrowed(x) => Bytes::from_static(x),
-        Cow::Owned(vec) => Bytes::from(vec),
-    };
-    wrap_stream(content)
-}
-
-fn wrap_stream(
-    content: bytes::Bytes,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    let len = content.len();
-    let content = futures::future::ready(Ok(content));
-
-    (futures::stream::once(content), len)
-}

libs/remote_storage/tests/test_real_s3.rs

@@ -2,23 +2,23 @@ use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
-use std::path::PathBuf;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
-use bytes::Bytes;
 use camino::Utf8Path;
-use futures::stream::Stream;
-use once_cell::sync::OnceCell;
 use remote_storage::{
     GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
 };
 use test_context::{test_context, AsyncTestContext};
-use tokio::task::JoinSet;
-use tracing::{debug, error, info};
+use tracing::{debug, info};
 
-static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+mod common;
+
+use common::{
+    cleanup, download_to_vec, ensure_logging_ready, upload_remote_data, upload_simple_remote_data,
+    upload_stream, wrap_stream,
+};
 
 const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
 
@@ -30,7 +30,7 @@ const BASE_PREFIX: &str = "test";
 /// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
 /// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
 ///
-/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
+/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_remote_data`]
 /// where
 /// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
 /// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
@@ -95,7 +95,7 @@ async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3WithTestBlobs) -> any
 /// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified. Test will skip real code and pass if env vars not set.
 /// See `s3_pagination_should_work` for more information.
 ///
-/// First, create a set of S3 objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_s3_data`]
+/// First, create a set of S3 objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_remote_data`]
 /// Then performs the following queries:
 ///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
 ///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
@@ -198,15 +198,65 @@ async fn s3_delete_objects_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()>
     Ok(())
 }
 
-fn ensure_logging_ready() {
-    LOGGING_DONE.get_or_init(|| {
-        utils::logging::init(
-            utils::logging::LogFormat::Test,
-            utils::logging::TracingErrorLayerEnablement::Disabled,
-            utils::logging::Output::Stdout,
-        )
-        .expect("logging init failed");
-    });
+#[test_context(MaybeEnabledS3)]
+#[tokio::test]
+async fn s3_upload_download_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
+    let MaybeEnabledS3::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let orig = bytes::Bytes::from_static("remote blob data here".as_bytes());
+
+    let (data, len) = wrap_stream(orig.clone());
+
+    ctx.client.upload(data, len, &path, None).await?;
+
+    // Normal download request
+    let dl = ctx.client.download(&path).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(len as u64))
+        .await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    // partial range (end specified)
+    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[4..10]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(len as u64 * 100))
+        .await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[8..]);
+
+    // Partial range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig[4..]);
+
+    // Full range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let buf = download_to_vec(dl).await?;
+    assert_eq!(&buf, &orig);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
 }
 
 struct EnabledS3 {
@@ -278,7 +328,7 @@ impl AsyncTestContext for MaybeEnabledS3WithTestBlobs {
 
         let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_s3_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+        match upload_remote_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -340,7 +390,7 @@ impl AsyncTestContext for MaybeEnabledS3WithSimpleTestBlobs {
 
         let enabled = EnabledS3::setup(Some(max_keys_in_list_response)).await;
 
-        match upload_simple_s3_data(&enabled.client, upload_tasks_count).await {
+        match upload_simple_remote_data(&enabled.client, upload_tasks_count).await {
             ControlFlow::Continue(uploads) => {
                 info!("Remote objects created successfully");
 
@@ -403,166 +453,3 @@ fn create_s3_client(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
-
-struct Uploads {
-    prefixes: HashSet<RemotePath>,
-    blobs: HashSet<RemotePath>,
-}
-
-async fn upload_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    base_prefix_str: &'static str,
-    upload_tasks_count: usize,
-) -> ControlFlow<Uploads, Uploads> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
-            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
-                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
-            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, data_len) =
-                upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok((upload_prefix, upload_path)) => {
-                uploaded_prefixes.insert(upload_prefix);
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    let uploads = Uploads {
-        prefixes: uploaded_prefixes,
-        blobs: uploaded_blobs,
-    };
-    if upload_tasks_failed {
-        ControlFlow::Break(uploads)
-    } else {
-        ControlFlow::Continue(uploads)
-    }
-}
-
-async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
-    info!(
-        "Removing {} objects from the remote storage during cleanup",
-        objects_to_delete.len()
-    );
-    let mut delete_tasks = JoinSet::new();
-    for object_to_delete in objects_to_delete {
-        let task_client = Arc::clone(client);
-        delete_tasks.spawn(async move {
-            debug!("Deleting remote item at path {object_to_delete:?}");
-            task_client
-                .delete(&object_to_delete)
-                .await
-                .with_context(|| format!("{object_to_delete:?} removal"))
-        });
-    }
-
-    while let Some(task_run_result) = delete_tasks.join_next().await {
-        match task_run_result {
-            Ok(task_result) => match task_result {
-                Ok(()) => {}
-                Err(e) => error!("Delete task failed: {e:?}"),
-            },
-            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
-        }
-    }
-}
-
-// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
-async fn upload_simple_s3_data(
-    client: &Arc<GenericRemoteStorage>,
-    upload_tasks_count: usize,
-) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
-    info!("Creating {upload_tasks_count} S3 files");
-    let mut upload_tasks = JoinSet::new();
-    for i in 1..upload_tasks_count + 1 {
-        let task_client = Arc::clone(client);
-        upload_tasks.spawn(async move {
-            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
-            let blob_path = RemotePath::new(
-                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
-            )
-            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
-            debug!("Creating remote item {i} at path {blob_path:?}");
-
-            let (data, data_len) =
-                upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
-
-            Ok::<_, anyhow::Error>(blob_path)
-        });
-    }
-
-    let mut upload_tasks_failed = false;
-    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
-    while let Some(task_run_result) = upload_tasks.join_next().await {
-        match task_run_result
-            .context("task join failed")
-            .and_then(|task_result| task_result.context("upload task failed"))
-        {
-            Ok(upload_path) => {
-                uploaded_blobs.insert(upload_path);
-            }
-            Err(e) => {
-                error!("Upload task failed: {e:?}");
-                upload_tasks_failed = true;
-            }
-        }
-    }
-
-    if upload_tasks_failed {
-        ControlFlow::Break(uploaded_blobs)
-    } else {
-        ControlFlow::Continue(uploaded_blobs)
-    }
-}
-
-fn upload_stream(
-    content: std::borrow::Cow<'static, [u8]>,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    use std::borrow::Cow;
-
-    let content = match content {
-        Cow::Borrowed(x) => Bytes::from_static(x),
-        Cow::Owned(vec) => Bytes::from(vec),
-    };
-    wrap_stream(content)
-}
-
-fn wrap_stream(
-    content: bytes::Bytes,
-) -> (
-    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
-    usize,
-) {
-    let len = content.len();
-    let content = futures::future::ready(Ok(content));
-
-    (futures::stream::once(content), len)
-}

libs/utils/Cargo.toml

@@ -4,6 +4,12 @@ version = "0.1.0"
 edition.workspace = true
 license.workspace = true
 
+[features]
+default = []
+# Enables test-only APIs, incuding failpoints. In particular, enables the `fail_point!` macro,
+# which adds some runtime cost to run tests on outage conditions
+testing = ["fail/failpoints"]
+
 [dependencies]
 arc-swap.workspace = true
 sentry.workspace = true
@@ -16,6 +22,7 @@ chrono.workspace = true
 heapless.workspace = true
 hex = { workspace = true, features = ["serde"] }
 hyper = { workspace = true, features = ["full"] }
+fail.workspace = true
 futures = { workspace = true}
 jsonwebtoken.workspace = true
 nix.workspace = true

libs/utils/src/failpoint_support.rs

@@ -1,3 +1,14 @@
+//! Failpoint support code shared between pageserver and safekeepers.
+
+use crate::http::{
+    error::ApiError,
+    json::{json_request, json_response},
+};
+use hyper::{Body, Request, Response, StatusCode};
+use serde::{Deserialize, Serialize};
+use tokio_util::sync::CancellationToken;
+use tracing::*;
+
 /// use with fail::cfg("$name", "return(2000)")
 ///
 /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the
@@ -25,7 +36,7 @@ pub use __failpoint_sleep_millis_async as sleep_millis_async;
 // Helper function used by the macro. (A function has nicer scoping so we
 // don't need to decorate everything with "::")
 #[doc(hidden)]
-pub(crate) async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
+pub async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
     let millis = duration_str.parse::<u64>().unwrap();
     let d = std::time::Duration::from_millis(millis);
 
@@ -71,7 +82,7 @@ pub fn init() -> fail::FailScenario<'static> {
     scenario
 }
 
-pub(crate) fn apply_failpoint(name: &str, actions: &str) -> Result<(), String> {
+pub fn apply_failpoint(name: &str, actions: &str) -> Result<(), String> {
     if actions == "exit" {
         fail::cfg_callback(name, exit_failpoint)
     } else {
@@ -84,3 +95,45 @@ fn exit_failpoint() {
     tracing::info!("Exit requested by failpoint");
     std::process::exit(1);
 }
+
+pub type ConfigureFailpointsRequest = Vec<FailpointConfig>;
+
+/// Information for configuring a single fail point
+#[derive(Debug, Serialize, Deserialize)]
+pub struct FailpointConfig {
+    /// Name of the fail point
+    pub name: String,
+    /// List of actions to take, using the format described in `fail::cfg`
+    ///
+    /// We also support `actions = "exit"` to cause the fail point to immediately exit.
+    pub actions: String,
+}
+
+/// Configure failpoints through http.
+pub async fn failpoints_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    if !fail::has_failpoints() {
+        return Err(ApiError::BadRequest(anyhow::anyhow!(
+            "Cannot manage failpoints because storage was compiled without failpoints support"
+        )));
+    }
+
+    let failpoints: ConfigureFailpointsRequest = json_request(&mut request).await?;
+    for fp in failpoints {
+        info!("cfg failpoint: {} {}", fp.name, fp.actions);
+
+        // We recognize one extra "action" that's not natively recognized
+        // by the failpoints crate: exit, to immediately kill the process
+        let cfg_result = apply_failpoint(&fp.name, &fp.actions);
+
+        if let Err(err_msg) = cfg_result {
+            return Err(ApiError::BadRequest(anyhow::anyhow!(
+                "Failed to configure failpoints: {err_msg}"
+            )));
+        }
+    }
+
+    json_response(StatusCode::OK, ())
+}

libs/utils/src/http/error.rs

@@ -31,6 +31,9 @@ pub enum ApiError {
     #[error("Shutting down")]
     ShuttingDown,
 
+    #[error("Timeout")]
+    Timeout(Cow<'static, str>),
+
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -67,6 +70,10 @@ impl ApiError {
                 err.to_string(),
                 StatusCode::SERVICE_UNAVAILABLE,
             ),
+            ApiError::Timeout(err) => HttpErrorBody::response_from_msg_and_status(
+                err.to_string(),
+                StatusCode::REQUEST_TIMEOUT,
+            ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                 err.to_string(),
                 StatusCode::INTERNAL_SERVER_ERROR,

libs/utils/src/id.rs

@@ -1,3 +1,4 @@
+use std::num::ParseIntError;
 use std::{fmt, str::FromStr};
 
 use anyhow::Context;
@@ -374,6 +375,13 @@ impl fmt::Display for NodeId {
     }
 }
 
+impl FromStr for NodeId {
+    type Err = ParseIntError;
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        Ok(NodeId(u64::from_str(s)?))
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use serde_assert::{Deserializer, Serializer, Token, Tokens};

libs/utils/src/lib.rs

@@ -83,6 +83,8 @@ pub mod timeout;
 
 pub mod sync;
 
+pub mod failpoint_support;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:

libs/utils/src/lsn.rs

@@ -366,6 +366,49 @@ impl MonotonicCounter<Lsn> for RecordLsn {
     }
 }
 
+/// Implements  [`rand::distributions::uniform::UniformSampler`] so we can sample [`Lsn`]s.
+///
+/// This is used by the `pagebench` pageserver benchmarking tool.
+pub struct LsnSampler(<u64 as rand::distributions::uniform::SampleUniform>::Sampler);
+
+impl rand::distributions::uniform::SampleUniform for Lsn {
+    type Sampler = LsnSampler;
+}
+
+impl rand::distributions::uniform::UniformSampler for LsnSampler {
+    type X = Lsn;
+
+    fn new<B1, B2>(low: B1, high: B2) -> Self
+    where
+        B1: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+        B2: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+    {
+        Self(
+            <u64 as rand::distributions::uniform::SampleUniform>::Sampler::new(
+                low.borrow().0,
+                high.borrow().0,
+            ),
+        )
+    }
+
+    fn new_inclusive<B1, B2>(low: B1, high: B2) -> Self
+    where
+        B1: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+        B2: rand::distributions::uniform::SampleBorrow<Self::X> + Sized,
+    {
+        Self(
+            <u64 as rand::distributions::uniform::SampleUniform>::Sampler::new_inclusive(
+                low.borrow().0,
+                high.borrow().0,
+            ),
+        )
+    }
+
+    fn sample<R: rand::prelude::Rng + ?Sized>(&self, rng: &mut R) -> Self::X {
+        Lsn(self.0.sample(rng))
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use crate::bin_ser::BeSer;

libs/walproposer/bindgen_deps.h

@@ -1 +1,2 @@
+#include "postgres.h"
 #include "walproposer.h"

libs/walproposer/src/api_bindings.rs

@@ -8,12 +8,12 @@ use std::ffi::CString;
 
 use crate::bindings::uint32;
 use crate::bindings::walproposer_api;
+use crate::bindings::NeonWALReadResult;
 use crate::bindings::PGAsyncReadResult;
 use crate::bindings::PGAsyncWriteResult;
 use crate::bindings::Safekeeper;
 use crate::bindings::Size;
 use crate::bindings::StringInfoData;
-use crate::bindings::TimeLineID;
 use crate::bindings::TimestampTz;
 use crate::bindings::WalProposer;
 use crate::bindings::WalProposerConnStatusType;
@@ -178,31 +178,11 @@ extern "C" fn conn_blocking_write(
     }
 }
 
-extern "C" fn recovery_download(
-    sk: *mut Safekeeper,
-    _timeline: TimeLineID,
-    startpos: XLogRecPtr,
-    endpos: XLogRecPtr,
-) -> bool {
+extern "C" fn recovery_download(wp: *mut WalProposer, sk: *mut Safekeeper) -> bool {
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).recovery_download(&mut (*sk), startpos, endpos)
-    }
-}
-
-#[allow(clippy::unnecessary_cast)]
-extern "C" fn wal_read(
-    sk: *mut Safekeeper,
-    buf: *mut ::std::os::raw::c_char,
-    startptr: XLogRecPtr,
-    count: Size,
-) {
-    unsafe {
-        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
-        let callback_data = (*(*(*sk).wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).wal_read(&mut (*sk), buf, startptr)
+        (*api).recovery_download(&mut (*wp), &mut (*sk))
     }
 }
 
@@ -214,11 +194,28 @@ extern "C" fn wal_reader_allocate(sk: *mut Safekeeper) {
     }
 }
 
-extern "C" fn free_event_set(wp: *mut WalProposer) {
+#[allow(clippy::unnecessary_cast)]
+extern "C" fn wal_read(
+    sk: *mut Safekeeper,
+    buf: *mut ::std::os::raw::c_char,
+    startptr: XLogRecPtr,
+    count: Size,
+    _errmsg: *mut *mut ::std::os::raw::c_char,
+) -> NeonWALReadResult {
     unsafe {
-        let callback_data = (*(*wp).config).callback_data;
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).free_event_set(&mut (*wp));
+        // TODO: errmsg is not forwarded
+        (*api).wal_read(&mut (*sk), buf, startptr)
+    }
+}
+
+extern "C" fn wal_reader_events(sk: *mut Safekeeper) -> uint32 {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_reader_events(&mut (*sk))
     }
 }
 
@@ -238,6 +235,14 @@ extern "C" fn update_event_set(sk: *mut Safekeeper, events: uint32) {
     }
 }
 
+extern "C" fn active_state_update_event_set(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).active_state_update_event_set(&mut (*sk));
+    }
+}
+
 extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
@@ -246,6 +251,14 @@ extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
     }
 }
 
+extern "C" fn rm_safekeeper_event_set(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).rm_safekeeper_event_set(&mut (*sk));
+    }
+}
+
 extern "C" fn wait_event_set(
     wp: *mut WalProposer,
     timeout: ::std::os::raw::c_long,
@@ -313,14 +326,6 @@ extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLog
     }
 }
 
-extern "C" fn confirm_wal_streamed(wp: *mut WalProposer, lsn: XLogRecPtr) {
-    unsafe {
-        let callback_data = (*(*wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).confirm_wal_streamed(&mut (*wp), lsn)
-    }
-}
-
 extern "C" fn log_internal(
     wp: *mut WalProposer,
     level: ::std::os::raw::c_int,
@@ -335,14 +340,6 @@ extern "C" fn log_internal(
     }
 }
 
-extern "C" fn after_election(wp: *mut WalProposer) {
-    unsafe {
-        let callback_data = (*(*wp).config).callback_data;
-        let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).after_election(&mut (*wp))
-    }
-}
-
 #[derive(Debug)]
 pub enum Level {
     Debug5,
@@ -401,20 +398,20 @@ pub(crate) fn create_api() -> walproposer_api {
         conn_async_write: Some(conn_async_write),
         conn_blocking_write: Some(conn_blocking_write),
         recovery_download: Some(recovery_download),
-        wal_read: Some(wal_read),
         wal_reader_allocate: Some(wal_reader_allocate),
-        free_event_set: Some(free_event_set),
+        wal_read: Some(wal_read),
+        wal_reader_events: Some(wal_reader_events),
         init_event_set: Some(init_event_set),
         update_event_set: Some(update_event_set),
+        active_state_update_event_set: Some(active_state_update_event_set),
         add_safekeeper_event_set: Some(add_safekeeper_event_set),
+        rm_safekeeper_event_set: Some(rm_safekeeper_event_set),
         wait_event_set: Some(wait_event_set),
         strong_random: Some(strong_random),
         get_redo_start_lsn: Some(get_redo_start_lsn),
         finish_sync_safekeepers: Some(finish_sync_safekeepers),
         process_safekeeper_feedback: Some(process_safekeeper_feedback),
-        confirm_wal_streamed: Some(confirm_wal_streamed),
         log_internal: Some(log_internal),
-        after_election: Some(after_election),
     }
 }
 

libs/walproposer/src/walproposer.rs

@@ -6,8 +6,8 @@ use utils::id::TenantTimelineId;
 use crate::{
     api_bindings::{create_api, take_vec_u8, Level},
     bindings::{
-        Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate, WalProposerFree,
-        WalProposerStart,
+        NeonWALReadResult, Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate,
+        WalProposerFree, WalProposerStart,
     },
 };
 
@@ -86,19 +86,19 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn recovery_download(&self, _sk: &mut Safekeeper, _startpos: u64, _endpos: u64) -> bool {
+    fn recovery_download(&self, _wp: &mut WalProposer, _sk: &mut Safekeeper) -> bool {
         todo!()
     }
 
-    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) {
+    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) -> NeonWALReadResult {
         todo!()
     }
 
-    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) {
+    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) -> NeonWALReadResult {
         todo!()
     }
 
-    fn free_event_set(&self, _wp: &mut WalProposer) {
+    fn wal_reader_events(&self, _sk: &mut Safekeeper) -> u32 {
         todo!()
     }
 
@@ -110,10 +110,18 @@ pub trait ApiImpl {
         todo!()
     }
 
+    fn active_state_update_event_set(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
     fn add_safekeeper_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
         todo!()
     }
 
+    fn rm_safekeeper_event_set(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
     fn wait_event_set(&self, _wp: &mut WalProposer, _timeout_millis: i64) -> WaitResult {
         todo!()
     }
@@ -134,10 +142,6 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn confirm_wal_streamed(&self, _wp: &mut WalProposer, _lsn: u64) {
-        todo!()
-    }
-
     fn log_internal(&self, _wp: &mut WalProposer, _level: Level, _msg: &str) {
         todo!()
     }
@@ -240,6 +244,7 @@ impl Drop for Wrapper {
 
 #[cfg(test)]
 mod tests {
+    use core::panic;
     use std::{
         cell::Cell,
         sync::{atomic::AtomicUsize, mpsc::sync_channel},
@@ -247,7 +252,7 @@ mod tests {
 
     use utils::id::TenantTimelineId;
 
-    use crate::{api_bindings::Level, walproposer::Wrapper};
+    use crate::{api_bindings::Level, bindings::NeonWALReadResult, walproposer::Wrapper};
 
     use super::ApiImpl;
 
@@ -355,12 +360,17 @@ mod tests {
             true
         }
 
-        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) {
-            println!("wal_reader_allocate")
+        fn recovery_download(
+            &self,
+            _wp: &mut crate::bindings::WalProposer,
+            _sk: &mut crate::bindings::Safekeeper,
+        ) -> bool {
+            true
         }
 
-        fn free_event_set(&self, _: &mut crate::bindings::WalProposer) {
-            println!("free_event_set")
+        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) -> NeonWALReadResult {
+            println!("wal_reader_allocate");
+            crate::bindings::NeonWALReadResult_NEON_WALREAD_SUCCESS
         }
 
         fn init_event_set(&self, _: &mut crate::bindings::WalProposer) {
@@ -383,6 +393,13 @@ mod tests {
             self.wait_events.set(WaitEventsData { sk, event_mask });
         }
 
+        fn rm_safekeeper_event_set(&self, sk: &mut crate::bindings::Safekeeper) {
+            println!(
+                "rm_safekeeper_event_set, sk={:?}",
+                sk as *mut crate::bindings::Safekeeper
+            );
+        }
+
         fn wait_event_set(
             &self,
             _: &mut crate::bindings::WalProposer,
@@ -408,7 +425,7 @@ mod tests {
         }
 
         fn log_internal(&self, _wp: &mut crate::bindings::WalProposer, level: Level, msg: &str) {
-            println!("walprop_log[{}] {}", level, msg);
+            println!("wp_log[{}] {}", level, msg);
         }
 
         fn after_election(&self, _wp: &mut crate::bindings::WalProposer) {

pageserver/Cargo.toml

@@ -63,6 +63,7 @@ thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
 tokio-io-timeout.workspace = true
 tokio-postgres.workspace = true
+tokio-stream.workspace = true
 tokio-util.workspace = true
 toml_edit = { workspace = true, features = [ "serde" ] }
 tracing.workspace = true

pageserver/benches/bench_walredo.rs

@@ -13,6 +13,7 @@ use bytes::{Buf, Bytes};
 use pageserver::{
     config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
 };
+use pageserver_api::shard::TenantShardId;
 use utils::{id::TenantId, lsn::Lsn};
 
 use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
@@ -26,9 +27,9 @@ fn redo_scenarios(c: &mut Criterion) {
 
     let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
     let conf = Box::leak(Box::new(conf));
-    let tenant_id = TenantId::generate();
+    let tenant_shard_id = TenantShardId::unsharded(TenantId::generate());
 
-    let manager = PostgresRedoManager::new(conf, tenant_id);
+    let manager = PostgresRedoManager::new(conf, tenant_shard_id);
 
     let manager = Arc::new(manager);
 

pageserver/client/Cargo.toml

@@ -12,3 +12,11 @@ reqwest.workspace = true
 utils.workspace = true
 serde.workspace = true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+tokio-postgres.workspace = true
+tokio-stream.workspace = true
+tokio.workspace = true
+futures.workspace = true
+tokio-util.workspace = true
+anyhow.workspace = true
+postgres.workspace = true
+bytes.workspace = true

pageserver/client/src/lib.rs

@@ -1 +1,2 @@
 pub mod mgmt_api;
+pub mod page_service;

pageserver/client/src/mgmt_api.rs

@@ -1,10 +1,12 @@
-use pageserver_api::models::*;
+use pageserver_api::{models::*, shard::TenantShardId};
 use reqwest::{IntoUrl, Method};
 use utils::{
     http::error::HttpErrorBody,
     id::{TenantId, TimelineId},
 };
 
+pub mod util;
+
 #[derive(Debug)]
 pub struct Client {
     mgmt_api_endpoint: String,
@@ -64,11 +66,26 @@ impl Client {
         resp.json().await.map_err(Error::ReceiveBody)
     }
 
+    pub async fn tenant_details(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> Result<pageserver_api::models::TenantDetails> {
+        let uri = format!("{}/v1/tenant/{tenant_shard_id}", self.mgmt_api_endpoint);
+        self.get(uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
     pub async fn list_timelines(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
     ) -> Result<Vec<pageserver_api::models::TimelineInfo>> {
-        let uri = format!("{}/v1/tenant/{tenant_id}/timeline", self.mgmt_api_endpoint);
+        let uri = format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline",
+            self.mgmt_api_endpoint
+        );
         self.get(&uri)
             .await?
             .json()
@@ -152,14 +169,17 @@ impl Client {
 
     pub async fn location_config(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         config: LocationConfig,
         flush_ms: Option<std::time::Duration>,
     ) -> Result<()> {
-        let req_body = TenantLocationConfigRequest { tenant_id, config };
+        let req_body = TenantLocationConfigRequest {
+            tenant_id: tenant_shard_id,
+            config,
+        };
         let path = format!(
             "{}/v1/tenant/{}/location_config",
-            self.mgmt_api_endpoint, tenant_id
+            self.mgmt_api_endpoint, tenant_shard_id
         );
         let path = if let Some(flush_ms) = flush_ms {
             format!("{}?flush_ms={}", path, flush_ms.as_millis())
@@ -185,4 +205,19 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn tenant_synthetic_size(
+        &self,
+        tenant_shard_id: TenantShardId,
+    ) -> Result<TenantHistorySize> {
+        let uri = format!(
+            "{}/v1/tenant/{}/synthetic_size",
+            self.mgmt_api_endpoint, tenant_shard_id
+        );
+        self.get(&uri)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/client/src/mgmt_api/util.rs

@@ -0,0 +1,53 @@
+//! Helpers to do common higher-level tasks with the [`Client`].
+
+use std::sync::Arc;
+
+use pageserver_api::shard::TenantShardId;
+use tokio::task::JoinSet;
+use utils::id::{TenantId, TenantTimelineId};
+
+use super::Client;
+
+/// Retrieve a list of all of the pageserver's timelines.
+///
+/// Fails if there are sharded tenants present on the pageserver.
+pub async fn get_pageserver_tenant_timelines_unsharded(
+    api_client: &Arc<Client>,
+) -> anyhow::Result<Vec<TenantTimelineId>> {
+    let mut timelines: Vec<TenantTimelineId> = Vec::new();
+    let mut tenants: Vec<TenantId> = Vec::new();
+    for ti in api_client.list_tenants().await? {
+        if !ti.id.is_unsharded() {
+            anyhow::bail!(
+                "only unsharded tenants are supported at this time: {}",
+                ti.id
+            );
+        }
+        tenants.push(ti.id.tenant_id)
+    }
+    let mut js = JoinSet::new();
+    for tenant_id in tenants {
+        js.spawn({
+            let mgmt_api_client = Arc::clone(api_client);
+            async move {
+                (
+                    tenant_id,
+                    mgmt_api_client
+                        .tenant_details(TenantShardId::unsharded(tenant_id))
+                        .await
+                        .unwrap(),
+                )
+            }
+        });
+    }
+    while let Some(res) = js.join_next().await {
+        let (tenant_id, details) = res.unwrap();
+        for timeline_id in details.timelines {
+            timelines.push(TenantTimelineId {
+                tenant_id,
+                timeline_id,
+            });
+        }
+    }
+    Ok(timelines)
+}

pageserver/client/src/page_service.rs

@@ -0,0 +1,144 @@
+use std::pin::Pin;
+
+use futures::SinkExt;
+use pageserver_api::{
+    models::{
+        PagestreamBeMessage, PagestreamFeMessage, PagestreamGetPageRequest,
+        PagestreamGetPageResponse,
+    },
+    reltag::RelTag,
+};
+use tokio::task::JoinHandle;
+use tokio_postgres::CopyOutStream;
+use tokio_stream::StreamExt;
+use tokio_util::sync::CancellationToken;
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+};
+
+pub struct Client {
+    client: tokio_postgres::Client,
+    cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
+    conn_task: JoinHandle<()>,
+}
+
+pub struct BasebackupRequest {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub lsn: Option<Lsn>,
+    pub gzip: bool,
+}
+
+impl Client {
+    pub async fn new(connstring: String) -> anyhow::Result<Self> {
+        let (client, connection) = tokio_postgres::connect(&connstring, postgres::NoTls).await?;
+
+        let conn_task_cancel = CancellationToken::new();
+        let conn_task = tokio::spawn({
+            let conn_task_cancel = conn_task_cancel.clone();
+            async move {
+                tokio::select! {
+                    _ = conn_task_cancel.cancelled() => { }
+                    res = connection => {
+                        res.unwrap();
+                    }
+                }
+            }
+        });
+        Ok(Self {
+            cancel_on_client_drop: Some(conn_task_cancel.drop_guard()),
+            conn_task,
+            client,
+        })
+    }
+
+    pub async fn pagestream(
+        self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<PagestreamClient> {
+        let copy_both: tokio_postgres::CopyBothDuplex<bytes::Bytes> = self
+            .client
+            .copy_both_simple(&format!("pagestream {tenant_id} {timeline_id}"))
+            .await?;
+        let Client {
+            cancel_on_client_drop,
+            conn_task,
+            client: _,
+        } = self;
+        Ok(PagestreamClient {
+            copy_both: Box::pin(copy_both),
+            conn_task,
+            cancel_on_client_drop,
+        })
+    }
+
+    pub async fn basebackup(&self, req: &BasebackupRequest) -> anyhow::Result<CopyOutStream> {
+        let BasebackupRequest {
+            tenant_id,
+            timeline_id,
+            lsn,
+            gzip,
+        } = req;
+        let mut args = Vec::with_capacity(5);
+        args.push("basebackup".to_string());
+        args.push(format!("{tenant_id}"));
+        args.push(format!("{timeline_id}"));
+        if let Some(lsn) = lsn {
+            args.push(format!("{lsn}"));
+        }
+        if *gzip {
+            args.push("--gzip".to_string())
+        }
+        Ok(self.client.copy_out(&args.join(" ")).await?)
+    }
+}
+
+/// Create using [`Client::pagestream`].
+pub struct PagestreamClient {
+    copy_both: Pin<Box<tokio_postgres::CopyBothDuplex<bytes::Bytes>>>,
+    cancel_on_client_drop: Option<tokio_util::sync::DropGuard>,
+    conn_task: JoinHandle<()>,
+}
+
+pub struct RelTagBlockNo {
+    pub rel_tag: RelTag,
+    pub block_no: u32,
+}
+
+impl PagestreamClient {
+    pub async fn shutdown(mut self) {
+        let _ = self.cancel_on_client_drop.take();
+        self.conn_task.await.unwrap();
+    }
+
+    pub async fn getpage(
+        &mut self,
+        req: PagestreamGetPageRequest,
+    ) -> anyhow::Result<PagestreamGetPageResponse> {
+        let req = PagestreamFeMessage::GetPage(req);
+        let req: bytes::Bytes = req.serialize();
+        // let mut req = tokio_util::io::ReaderStream::new(&req);
+        let mut req = tokio_stream::once(Ok(req));
+
+        self.copy_both.send_all(&mut req).await?;
+
+        let next: Option<Result<bytes::Bytes, _>> = self.copy_both.next().await;
+        let next: bytes::Bytes = next.unwrap()?;
+
+        let msg = PagestreamBeMessage::deserialize(next)?;
+        match msg {
+            PagestreamBeMessage::GetPage(p) => Ok(p),
+            PagestreamBeMessage::Error(e) => anyhow::bail!("Error: {:?}", e),
+            PagestreamBeMessage::Exists(_)
+            | PagestreamBeMessage::Nblocks(_)
+            | PagestreamBeMessage::DbSize(_) => {
+                anyhow::bail!(
+                    "unexpected be message kind in response to getpage request: {}",
+                    msg.kind()
+                )
+            }
+        }
+    }
+}

pageserver/pagebench/Cargo.toml

@@ -0,0 +1,26 @@
+[package]
+name = "pagebench"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow.workspace = true
+clap.workspace = true
+futures.workspace = true
+hdrhistogram.workspace = true
+humantime.workspace = true
+humantime-serde.workspace = true
+rand.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+tracing.workspace = true
+tokio.workspace = true
+
+pageserver = { path = ".." }
+pageserver_client.workspace = true
+pageserver_api.workspace = true
+utils = { path = "../../libs/utils/" }
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }

pageserver/pagebench/src/cmd/basebackup.rs

@@ -0,0 +1,272 @@
+use anyhow::Context;
+use pageserver_client::page_service::BasebackupRequest;
+
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+
+use rand::prelude::*;
+use tokio::sync::Barrier;
+use tokio::task::JoinSet;
+use tracing::{debug, info, instrument};
+
+use std::collections::HashMap;
+use std::num::NonZeroUsize;
+use std::ops::Range;
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Instant;
+
+use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
+use crate::util::{request_stats, tokio_thread_local_stats};
+
+/// basebackup@LatestLSN
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "localhost:64000")]
+    page_service_host_port: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(long, default_value = "1")]
+    num_clients: NonZeroUsize,
+    #[clap(long, default_value = "1.0")]
+    gzip_probability: f64,
+    #[clap(long)]
+    runtime: Option<humantime::Duration>,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+#[derive(Debug, Default)]
+struct LiveStats {
+    completed_requests: AtomicU64,
+}
+
+impl LiveStats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+struct Target {
+    timeline: TenantTimelineId,
+    lsn_range: Option<Range<Lsn>>,
+}
+
+#[derive(serde::Serialize)]
+struct Output {
+    total: request_stats::Output,
+}
+
+tokio_thread_local_stats::declare!(STATS: request_stats::Stats);
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    tokio_thread_local_stats::main!(STATS, move |thread_local_stats| {
+        main_impl(args, thread_local_stats)
+    })
+}
+
+async fn main_impl(
+    args: Args,
+    all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
+) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+    let mut js = JoinSet::new();
+    for timeline in &timelines {
+        js.spawn({
+            let timeline = *timeline;
+            // FIXME: this triggers initial logical size calculation
+            // https://github.com/neondatabase/neon/issues/6168
+            let info = mgmt_api_client
+                .timeline_info(timeline.tenant_id, timeline.timeline_id)
+                .await
+                .unwrap();
+            async move {
+                anyhow::Ok(Target {
+                    timeline,
+                    // TODO: support lsn_range != latest LSN
+                    lsn_range: Some(info.last_record_lsn..(info.last_record_lsn + 1)),
+                })
+            }
+        });
+    }
+    let mut all_targets: Vec<Target> = Vec::new();
+    while let Some(res) = js.join_next().await {
+        all_targets.push(res.unwrap().unwrap());
+    }
+
+    let live_stats = Arc::new(LiveStats::default());
+
+    let num_client_tasks = timelines.len();
+    let num_live_stats_dump = 1;
+    let num_work_sender_tasks = 1;
+
+    let start_work_barrier = Arc::new(tokio::sync::Barrier::new(
+        num_client_tasks + num_live_stats_dump + num_work_sender_tasks,
+    ));
+    let all_work_done_barrier = Arc::new(tokio::sync::Barrier::new(num_client_tasks));
+
+    tokio::spawn({
+        let stats = Arc::clone(&live_stats);
+        let start_work_barrier = Arc::clone(&start_work_barrier);
+        async move {
+            start_work_barrier.wait().await;
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                info!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut work_senders = HashMap::new();
+    let mut tasks = Vec::new();
+    for tl in &timelines {
+        let (sender, receiver) = tokio::sync::mpsc::channel(1); // TODO: not sure what the implications of this are
+        work_senders.insert(tl, sender);
+        tasks.push(tokio::spawn(client(
+            args,
+            *tl,
+            Arc::clone(&start_work_barrier),
+            receiver,
+            Arc::clone(&all_work_done_barrier),
+            Arc::clone(&live_stats),
+        )));
+    }
+
+    let work_sender = async move {
+        start_work_barrier.wait().await;
+        loop {
+            let (timeline, work) = {
+                let mut rng = rand::thread_rng();
+                let target = all_targets.choose(&mut rng).unwrap();
+                let lsn = target.lsn_range.clone().map(|r| rng.gen_range(r));
+                (
+                    target.timeline,
+                    Work {
+                        lsn,
+                        gzip: rng.gen_bool(args.gzip_probability),
+                    },
+                )
+            };
+            let sender = work_senders.get(&timeline).unwrap();
+            // TODO: what if this blocks?
+            sender.send(work).await.ok().unwrap();
+        }
+    };
+
+    if let Some(runtime) = args.runtime {
+        match tokio::time::timeout(runtime.into(), work_sender).await {
+            Ok(()) => unreachable!("work sender never terminates"),
+            Err(_timeout) => {
+                // this implicitly drops the work_senders, making all the clients exit
+            }
+        }
+    } else {
+        work_sender.await;
+        unreachable!("work sender never terminates");
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+
+    let output = Output {
+        total: {
+            let mut agg_stats = request_stats::Stats::new();
+            for stats in all_thread_local_stats.lock().unwrap().iter() {
+                let stats = stats.lock().unwrap();
+                agg_stats.add(&stats);
+            }
+            agg_stats.output()
+        },
+    };
+
+    let output = serde_json::to_string_pretty(&output).unwrap();
+    println!("{output}");
+
+    anyhow::Ok(())
+}
+
+#[derive(Copy, Clone)]
+struct Work {
+    lsn: Option<Lsn>,
+    gzip: bool,
+}
+
+#[instrument(skip_all)]
+async fn client(
+    args: &'static Args,
+    timeline: TenantTimelineId,
+    start_work_barrier: Arc<Barrier>,
+    mut work: tokio::sync::mpsc::Receiver<Work>,
+    all_work_done_barrier: Arc<Barrier>,
+    live_stats: Arc<LiveStats>,
+) {
+    start_work_barrier.wait().await;
+
+    let client = pageserver_client::page_service::Client::new(crate::util::connstring::connstring(
+        &args.page_service_host_port,
+        args.pageserver_jwt.as_deref(),
+    ))
+    .await
+    .unwrap();
+
+    while let Some(Work { lsn, gzip }) = work.recv().await {
+        let start = Instant::now();
+        let copy_out_stream = client
+            .basebackup(&BasebackupRequest {
+                tenant_id: timeline.tenant_id,
+                timeline_id: timeline.timeline_id,
+                lsn,
+                gzip,
+            })
+            .await
+            .with_context(|| format!("start basebackup for {timeline}"))
+            .unwrap();
+
+        use futures::StreamExt;
+        let size = Arc::new(AtomicUsize::new(0));
+        copy_out_stream
+            .for_each({
+                |r| {
+                    let size = Arc::clone(&size);
+                    async move {
+                        let size = Arc::clone(&size);
+                        size.fetch_add(r.unwrap().len(), Ordering::Relaxed);
+                    }
+                }
+            })
+            .await;
+        debug!("basebackup size is {} bytes", size.load(Ordering::Relaxed));
+        let elapsed = start.elapsed();
+        live_stats.inc();
+        STATS.with(|stats| {
+            stats.borrow().lock().unwrap().observe(elapsed).unwrap();
+        });
+    }
+
+    all_work_done_barrier.wait().await;
+}

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -0,0 +1,351 @@
+use anyhow::Context;
+use futures::future::join_all;
+use pageserver::pgdatadir_mapping::key_to_rel_block;
+use pageserver::repository;
+use pageserver_api::key::is_rel_block_key;
+use pageserver_api::models::PagestreamGetPageRequest;
+
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+
+use rand::prelude::*;
+use tokio::sync::Barrier;
+use tokio::task::JoinSet;
+use tracing::{info, instrument};
+
+use std::collections::HashMap;
+use std::future::Future;
+use std::num::NonZeroUsize;
+use std::pin::Pin;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
+use crate::util::{request_stats, tokio_thread_local_stats};
+
+/// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "postgres://postgres@localhost:64000")]
+    page_service_connstring: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(long, default_value = "1")]
+    num_clients: NonZeroUsize,
+    #[clap(long)]
+    runtime: Option<humantime::Duration>,
+    #[clap(long)]
+    per_target_rate_limit: Option<usize>,
+    /// Probability for sending `latest=true` in the request (uniform distribution).
+    #[clap(long, default_value = "1")]
+    req_latest_probability: f64,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+#[derive(Debug, Default)]
+struct LiveStats {
+    completed_requests: AtomicU64,
+}
+
+impl LiveStats {
+    fn inc(&self) {
+        self.completed_requests.fetch_add(1, Ordering::Relaxed);
+    }
+}
+
+#[derive(Clone)]
+struct KeyRange {
+    timeline: TenantTimelineId,
+    timeline_lsn: Lsn,
+    start: i128,
+    end: i128,
+}
+
+impl KeyRange {
+    fn len(&self) -> i128 {
+        self.end - self.start
+    }
+}
+
+#[derive(serde::Serialize)]
+struct Output {
+    total: request_stats::Output,
+}
+
+tokio_thread_local_stats::declare!(STATS: request_stats::Stats);
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    tokio_thread_local_stats::main!(STATS, move |thread_local_stats| {
+        main_impl(args, thread_local_stats)
+    })
+}
+
+async fn main_impl(
+    args: Args,
+    all_thread_local_stats: AllThreadLocalStats<request_stats::Stats>,
+) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+
+    let mut js = JoinSet::new();
+    for timeline in &timelines {
+        js.spawn({
+            let mgmt_api_client = Arc::clone(&mgmt_api_client);
+            let timeline = *timeline;
+            async move {
+                let partitioning = mgmt_api_client
+                    .keyspace(timeline.tenant_id, timeline.timeline_id)
+                    .await?;
+                let lsn = partitioning.at_lsn;
+
+                let ranges = partitioning
+                    .keys
+                    .ranges
+                    .iter()
+                    .filter_map(|r| {
+                        let start = r.start;
+                        let end = r.end;
+                        // filter out non-relblock keys
+                        match (is_rel_block_key(&start), is_rel_block_key(&end)) {
+                            (true, true) => Some(KeyRange {
+                                timeline,
+                                timeline_lsn: lsn,
+                                start: start.to_i128(),
+                                end: end.to_i128(),
+                            }),
+                            (true, false) | (false, true) => {
+                                unimplemented!("split up range")
+                            }
+                            (false, false) => None,
+                        }
+                    })
+                    .collect::<Vec<_>>();
+
+                anyhow::Ok(ranges)
+            }
+        });
+    }
+    let mut all_ranges: Vec<KeyRange> = Vec::new();
+    while let Some(res) = js.join_next().await {
+        all_ranges.extend(res.unwrap().unwrap());
+    }
+
+    let live_stats = Arc::new(LiveStats::default());
+
+    let num_client_tasks = timelines.len();
+    let num_live_stats_dump = 1;
+    let num_work_sender_tasks = 1;
+
+    let start_work_barrier = Arc::new(tokio::sync::Barrier::new(
+        num_client_tasks + num_live_stats_dump + num_work_sender_tasks,
+    ));
+    let all_work_done_barrier = Arc::new(tokio::sync::Barrier::new(num_client_tasks));
+
+    tokio::spawn({
+        let stats = Arc::clone(&live_stats);
+        let start_work_barrier = Arc::clone(&start_work_barrier);
+        async move {
+            start_work_barrier.wait().await;
+            loop {
+                let start = std::time::Instant::now();
+                tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+                let completed_requests = stats.completed_requests.swap(0, Ordering::Relaxed);
+                let elapsed = start.elapsed();
+                info!(
+                    "RPS: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64()
+                );
+            }
+        }
+    });
+
+    let mut work_senders = HashMap::new();
+    let mut tasks = Vec::new();
+    for tl in &timelines {
+        let (sender, receiver) = tokio::sync::mpsc::channel(10); // TODO: not sure what the implications of this are
+        work_senders.insert(tl, sender);
+        tasks.push(tokio::spawn(client(
+            args,
+            *tl,
+            Arc::clone(&start_work_barrier),
+            receiver,
+            Arc::clone(&all_work_done_barrier),
+            Arc::clone(&live_stats),
+        )));
+    }
+
+    let work_sender: Pin<Box<dyn Send + Future<Output = ()>>> = match args.per_target_rate_limit {
+        None => Box::pin(async move {
+            let weights = rand::distributions::weighted::WeightedIndex::new(
+                all_ranges.iter().map(|v| v.len()),
+            )
+            .unwrap();
+
+            start_work_barrier.wait().await;
+
+            loop {
+                let (timeline, req) = {
+                    let mut rng = rand::thread_rng();
+                    let r = &all_ranges[weights.sample(&mut rng)];
+                    let key: i128 = rng.gen_range(r.start..r.end);
+                    let key = repository::Key::from_i128(key);
+                    let (rel_tag, block_no) =
+                        key_to_rel_block(key).expect("we filter non-rel-block keys out above");
+                    (
+                        r.timeline,
+                        PagestreamGetPageRequest {
+                            latest: rng.gen_bool(args.req_latest_probability),
+                            lsn: r.timeline_lsn,
+                            rel: rel_tag,
+                            blkno: block_no,
+                        },
+                    )
+                };
+                let sender = work_senders.get(&timeline).unwrap();
+                // TODO: what if this blocks?
+                sender.send(req).await.ok().unwrap();
+            }
+        }),
+        Some(rps_limit) => Box::pin(async move {
+            let period = Duration::from_secs_f64(1.0 / (rps_limit as f64));
+
+            let make_timeline_task: &dyn Fn(
+                TenantTimelineId,
+            )
+                -> Pin<Box<dyn Send + Future<Output = ()>>> = &|timeline| {
+                let sender = work_senders.get(&timeline).unwrap();
+                let ranges: Vec<KeyRange> = all_ranges
+                    .iter()
+                    .filter(|r| r.timeline == timeline)
+                    .cloned()
+                    .collect();
+                let weights = rand::distributions::weighted::WeightedIndex::new(
+                    ranges.iter().map(|v| v.len()),
+                )
+                .unwrap();
+
+                Box::pin(async move {
+                    let mut ticker = tokio::time::interval(period);
+                    ticker.set_missed_tick_behavior(
+                        /* TODO review this choice */
+                        tokio::time::MissedTickBehavior::Burst,
+                    );
+                    loop {
+                        ticker.tick().await;
+                        let req = {
+                            let mut rng = rand::thread_rng();
+                            let r = &ranges[weights.sample(&mut rng)];
+                            let key: i128 = rng.gen_range(r.start..r.end);
+                            let key = repository::Key::from_i128(key);
+                            let (rel_tag, block_no) = key_to_rel_block(key)
+                                .expect("we filter non-rel-block keys out above");
+                            PagestreamGetPageRequest {
+                                latest: rng.gen_bool(args.req_latest_probability),
+                                lsn: r.timeline_lsn,
+                                rel: rel_tag,
+                                blkno: block_no,
+                            }
+                        };
+                        sender.send(req).await.ok().unwrap();
+                    }
+                })
+            };
+
+            let tasks: Vec<_> = work_senders
+                .keys()
+                .map(|tl| make_timeline_task(**tl))
+                .collect();
+
+            start_work_barrier.wait().await;
+
+            join_all(tasks).await;
+        }),
+    };
+
+    if let Some(runtime) = args.runtime {
+        match tokio::time::timeout(runtime.into(), work_sender).await {
+            Ok(()) => unreachable!("work sender never terminates"),
+            Err(_timeout) => {
+                // this implicitly drops the work_senders, making all the clients exit
+            }
+        }
+    } else {
+        work_sender.await;
+        unreachable!("work sender never terminates");
+    }
+
+    for t in tasks {
+        t.await.unwrap();
+    }
+
+    let output = Output {
+        total: {
+            let mut agg_stats = request_stats::Stats::new();
+            for stats in all_thread_local_stats.lock().unwrap().iter() {
+                let stats = stats.lock().unwrap();
+                agg_stats.add(&stats);
+            }
+            agg_stats.output()
+        },
+    };
+
+    let output = serde_json::to_string_pretty(&output).unwrap();
+    println!("{output}");
+
+    anyhow::Ok(())
+}
+
+#[instrument(skip_all)]
+async fn client(
+    args: &'static Args,
+    timeline: TenantTimelineId,
+    start_work_barrier: Arc<Barrier>,
+    mut work: tokio::sync::mpsc::Receiver<PagestreamGetPageRequest>,
+    all_work_done_barrier: Arc<Barrier>,
+    live_stats: Arc<LiveStats>,
+) {
+    start_work_barrier.wait().await;
+
+    let client = pageserver_client::page_service::Client::new(args.page_service_connstring.clone())
+        .await
+        .unwrap();
+    let mut client = client
+        .pagestream(timeline.tenant_id, timeline.timeline_id)
+        .await
+        .unwrap();
+
+    while let Some(req) = work.recv().await {
+        let start = Instant::now();
+        client
+            .getpage(req)
+            .await
+            .with_context(|| format!("getpage for {timeline}"))
+            .unwrap();
+        let elapsed = start.elapsed();
+        live_stats.inc();
+        STATS.with(|stats| {
+            stats.borrow().lock().unwrap().observe(elapsed).unwrap();
+        });
+    }
+
+    all_work_done_barrier.wait().await;
+}

pageserver/pagebench/src/cmd/trigger_initial_size_calculation.rs

@@ -0,0 +1,85 @@
+use std::sync::Arc;
+
+use humantime::Duration;
+use tokio::task::JoinSet;
+use utils::id::TenantTimelineId;
+
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "localhost:64000")]
+    page_service_host_port: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+    #[clap(
+        long,
+        help = "if specified, poll mgmt api to check whether init logical size calculation has completed"
+    )]
+    poll_for_completion: Option<Duration>,
+    #[clap(long)]
+    limit_to_first_n_targets: Option<usize>,
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()
+        .unwrap();
+
+    let main_task = rt.spawn(main_impl(args));
+    rt.block_on(main_task).unwrap()
+}
+
+async fn main_impl(args: Args) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: args.limit_to_first_n_targets,
+            targets: args.targets.clone(),
+        },
+    )
+    .await?;
+
+    // kick it off
+
+    let mut js = JoinSet::new();
+    for tl in timelines {
+        let mgmt_api_client = Arc::clone(&mgmt_api_client);
+        js.spawn(async move {
+            // TODO: API to explicitly trigger initial logical size computation.
+            // Should probably also avoid making it a side effect of timeline details to trigger initial logical size calculation.
+            // => https://github.com/neondatabase/neon/issues/6168
+            let info = mgmt_api_client
+                .timeline_info(tl.tenant_id, tl.timeline_id)
+                .await
+                .unwrap();
+
+            if let Some(period) = args.poll_for_completion {
+                let mut ticker = tokio::time::interval(period.into());
+                ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+                let mut info = info;
+                while !info.current_logical_size_is_accurate {
+                    ticker.tick().await;
+                    info = mgmt_api_client
+                        .timeline_info(tl.tenant_id, tl.timeline_id)
+                        .await
+                        .unwrap();
+                }
+            }
+        });
+    }
+    while let Some(res) = js.join_next().await {
+        let _: () = res.unwrap();
+    }
+    Ok(())
+}

pageserver/pagebench/src/main.rs

@@ -0,0 +1,48 @@
+use clap::Parser;
+use utils::logging;
+
+/// Re-usable pieces of code that aren't CLI-specific.
+mod util {
+    pub(crate) mod connstring;
+    pub(crate) mod request_stats;
+    #[macro_use]
+    pub(crate) mod tokio_thread_local_stats;
+    /// Re-usable pieces of CLI-specific code.
+    pub(crate) mod cli {
+        pub(crate) mod targets;
+    }
+}
+
+/// The pagebench CLI sub-commands, dispatched in [`main`] below.
+mod cmd {
+    pub(super) mod basebackup;
+    pub(super) mod getpage_latest_lsn;
+    pub(super) mod trigger_initial_size_calculation;
+}
+
+/// Component-level performance test for pageserver.
+#[derive(clap::Parser)]
+enum Args {
+    Basebackup(cmd::basebackup::Args),
+    GetPageLatestLsn(cmd::getpage_latest_lsn::Args),
+    TriggerInitialSizeCalculation(cmd::trigger_initial_size_calculation::Args),
+}
+
+fn main() {
+    logging::init(
+        logging::LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::Disabled,
+        logging::Output::Stderr,
+    )
+    .unwrap();
+
+    let args = Args::parse();
+    match args {
+        Args::Basebackup(args) => cmd::basebackup::main(args),
+        Args::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
+        Args::TriggerInitialSizeCalculation(args) => {
+            cmd::trigger_initial_size_calculation::main(args)
+        }
+    }
+    .unwrap()
+}

pageserver/pagebench/src/util/cli/targets.rs

@@ -0,0 +1,34 @@
+use std::sync::Arc;
+
+use pageserver_client::mgmt_api;
+use tracing::info;
+use utils::id::TenantTimelineId;
+
+pub(crate) struct Spec {
+    pub(crate) limit_to_first_n_targets: Option<usize>,
+    pub(crate) targets: Option<Vec<TenantTimelineId>>,
+}
+
+pub(crate) async fn discover(
+    api_client: &Arc<mgmt_api::Client>,
+    spec: Spec,
+) -> anyhow::Result<Vec<TenantTimelineId>> {
+    let mut timelines = if let Some(targets) = spec.targets {
+        targets
+    } else {
+        mgmt_api::util::get_pageserver_tenant_timelines_unsharded(api_client).await?
+    };
+
+    if let Some(limit) = spec.limit_to_first_n_targets {
+        timelines.sort(); // for determinism
+        timelines.truncate(limit);
+        if timelines.len() < limit {
+            anyhow::bail!("pageserver has less than limit_to_first_n_targets={limit} tenants");
+        }
+    }
+
+    info!("timelines:\n{:?}", timelines);
+    info!("number of timelines:\n{:?}", timelines.len());
+
+    Ok(timelines)
+}

pageserver/pagebench/src/util/connstring.rs

@@ -0,0 +1,8 @@
+pub(crate) fn connstring(host_port: &str, jwt: Option<&str>) -> String {
+    let colon_and_jwt = if let Some(jwt) = jwt {
+        format!(":{jwt}") // TODO: urlescape
+    } else {
+        String::new()
+    };
+    format!("postgres://postgres{colon_and_jwt}@{host_port}")
+}

pageserver/pagebench/src/util/request_stats.rs

@@ -0,0 +1,88 @@
+use std::time::Duration;
+
+use anyhow::Context;
+
+pub(crate) struct Stats {
+    latency_histo: hdrhistogram::Histogram<u64>,
+}
+
+impl Stats {
+    pub(crate) fn new() -> Self {
+        Self {
+            // Initialize with fixed bounds so that we panic at runtime instead of resizing the histogram,
+            // which would skew the benchmark results.
+            latency_histo: hdrhistogram::Histogram::new_with_bounds(1, 1_000_000_000, 3).unwrap(),
+        }
+    }
+    pub(crate) fn observe(&mut self, latency: Duration) -> anyhow::Result<()> {
+        let micros: u64 = latency
+            .as_micros()
+            .try_into()
+            .context("latency greater than u64")?;
+        self.latency_histo
+            .record(micros)
+            .context("add to histogram")?;
+        Ok(())
+    }
+    pub(crate) fn output(&self) -> Output {
+        let latency_percentiles = std::array::from_fn(|idx| {
+            let micros = self
+                .latency_histo
+                .value_at_percentile(LATENCY_PERCENTILES[idx]);
+            Duration::from_micros(micros)
+        });
+        Output {
+            request_count: self.latency_histo.len(),
+            latency_mean: Duration::from_micros(self.latency_histo.mean() as u64),
+            latency_percentiles: LatencyPercentiles {
+                latency_percentiles,
+            },
+        }
+    }
+    pub(crate) fn add(&mut self, other: &Self) {
+        let Self {
+            ref mut latency_histo,
+        } = self;
+        latency_histo.add(&other.latency_histo).unwrap();
+    }
+}
+
+impl Default for Stats {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+const LATENCY_PERCENTILES: [f64; 4] = [95.0, 99.00, 99.90, 99.99];
+
+struct LatencyPercentiles {
+    latency_percentiles: [Duration; 4],
+}
+
+impl serde::Serialize for LatencyPercentiles {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        use serde::ser::SerializeMap;
+        let mut ser = serializer.serialize_map(Some(LATENCY_PERCENTILES.len()))?;
+        for p in LATENCY_PERCENTILES {
+            ser.serialize_entry(
+                &format!("p{p}"),
+                &format!(
+                    "{}",
+                    &humantime::format_duration(self.latency_percentiles[0])
+                ),
+            )?;
+        }
+        ser.end()
+    }
+}
+
+#[derive(serde::Serialize)]
+pub(crate) struct Output {
+    request_count: u64,
+    #[serde(with = "humantime_serde")]
+    latency_mean: Duration,
+    latency_percentiles: LatencyPercentiles,
+}

pageserver/pagebench/src/util/tokio_thread_local_stats.rs

@@ -0,0 +1,45 @@
+pub(crate) type ThreadLocalStats<T> = Arc<Mutex<T>>;
+pub(crate) type AllThreadLocalStats<T> = Arc<Mutex<Vec<ThreadLocalStats<T>>>>;
+
+macro_rules! declare {
+    ($THREAD_LOCAL_NAME:ident: $T:ty) => {
+        thread_local! {
+            pub static $THREAD_LOCAL_NAME: std::cell::RefCell<crate::util::tokio_thread_local_stats::ThreadLocalStats<$T>> = std::cell::RefCell::new(
+                std::sync::Arc::new(std::sync::Mutex::new(Default::default()))
+            );
+        }
+    };
+}
+
+use std::sync::{Arc, Mutex};
+
+pub(crate) use declare;
+
+macro_rules! main {
+    ($THREAD_LOCAL_NAME:ident, $main_impl:expr) => {{
+        let main_impl = $main_impl;
+        let all = Arc::new(Mutex::new(Vec::new()));
+
+        let rt = tokio::runtime::Builder::new_multi_thread()
+            .on_thread_start({
+                let all = Arc::clone(&all);
+                move || {
+                    // pre-initialize the thread local stats by accessesing them
+                    // (some stats like requests_stats::Stats are quite costly to initialize,
+                    //  we don't want to pay that cost during the measurement period)
+                    $THREAD_LOCAL_NAME.with(|stats| {
+                        let stats: Arc<_> = Arc::clone(&*stats.borrow());
+                        all.lock().unwrap().push(stats);
+                    });
+                }
+            })
+            .enable_all()
+            .build()
+            .unwrap();
+
+        let main_task = rt.spawn(main_impl(all));
+        rt.block_on(main_task).unwrap()
+    }};
+}
+
+pub(crate) use main;

pageserver/src/basebackup.rs

@@ -23,6 +23,7 @@ use tracing::*;
 use tokio_tar::{Builder, EntryType, Header};
 
 use crate::context::RequestContext;
+use crate::pgdatadir_mapping::Version;
 use crate::tenant::Timeline;
 use pageserver_api::reltag::{RelTag, SlruKind};
 
@@ -174,7 +175,7 @@ where
         ] {
             for segno in self
                 .timeline
-                .list_slru_segments(kind, self.lsn, self.ctx)
+                .list_slru_segments(kind, Version::Lsn(self.lsn), self.ctx)
                 .await?
             {
                 self.add_slru_segment(kind, segno).await?;
@@ -192,7 +193,7 @@ where
             // Otherwise only include init forks of unlogged relations.
             let rels = self
                 .timeline
-                .list_rels(spcnode, dbnode, self.lsn, self.ctx)
+                .list_rels(spcnode, dbnode, Version::Lsn(self.lsn), self.ctx)
                 .await?;
             for &rel in rels.iter() {
                 // Send init fork as main fork to provide well formed empty
@@ -267,7 +268,7 @@ where
     async fn add_rel(&mut self, src: RelTag, dst: RelTag) -> anyhow::Result<()> {
         let nblocks = self
             .timeline
-            .get_rel_size(src, self.lsn, false, self.ctx)
+            .get_rel_size(src, Version::Lsn(self.lsn), false, self.ctx)
             .await?;
 
         // If the relation is empty, create an empty file
@@ -288,7 +289,7 @@ where
             for blknum in startblk..endblk {
                 let img = self
                     .timeline
-                    .get_rel_page_at_lsn(src, blknum, self.lsn, false, self.ctx)
+                    .get_rel_page_at_lsn(src, blknum, Version::Lsn(self.lsn), false, self.ctx)
                     .await?;
                 segment_data.extend_from_slice(&img[..]);
             }
@@ -310,7 +311,7 @@ where
     async fn add_slru_segment(&mut self, slru: SlruKind, segno: u32) -> anyhow::Result<()> {
         let nblocks = self
             .timeline
-            .get_slru_segment_size(slru, segno, self.lsn, self.ctx)
+            .get_slru_segment_size(slru, segno, Version::Lsn(self.lsn), self.ctx)
             .await?;
 
         let mut slru_buf: Vec<u8> = Vec::with_capacity(nblocks as usize * BLCKSZ as usize);
@@ -352,7 +353,7 @@ where
         let relmap_img = if has_relmap_file {
             let img = self
                 .timeline
-                .get_relmap_file(spcnode, dbnode, self.lsn, self.ctx)
+                .get_relmap_file(spcnode, dbnode, Version::Lsn(self.lsn), self.ctx)
                 .await?;
 
             ensure!(
@@ -399,7 +400,7 @@ where
             if !has_relmap_file
                 && self
                     .timeline
-                    .list_rels(spcnode, dbnode, self.lsn, self.ctx)
+                    .list_rels(spcnode, dbnode, Version::Lsn(self.lsn), self.ctx)
                     .await?
                     .is_empty()
             {

pageserver/src/bin/pageserver.rs

@@ -31,6 +31,7 @@ use pageserver::{
     virtual_file,
 };
 use postgres_backend::AuthType;
+use utils::failpoint_support;
 use utils::logging::TracingErrorLayerEnablement;
 use utils::signals::ShutdownSignals;
 use utils::{
@@ -126,7 +127,7 @@ fn main() -> anyhow::Result<()> {
     }
 
     // Initialize up failpoints support
-    let scenario = pageserver::failpoint_support::init();
+    let scenario = failpoint_support::init();
 
     // Basic initialization of things that don't change after startup
     virtual_file::init(conf.max_file_descriptors);

pageserver/src/config.rs

@@ -76,6 +76,8 @@ pub mod defaults {
 
     pub const DEFAULT_HEATMAP_UPLOAD_CONCURRENCY: usize = 8;
 
+    pub const DEFAULT_INGEST_BATCH_SIZE: u64 = 100;
+
     ///
     /// Default built-in configuration file.
     ///
@@ -88,6 +90,7 @@ pub mod defaults {
 #wait_lsn_timeout = '{DEFAULT_WAIT_LSN_TIMEOUT}'
 #wal_redo_timeout = '{DEFAULT_WAL_REDO_TIMEOUT}'
 
+#page_cache_size = {DEFAULT_PAGE_CACHE_SIZE}
 #max_file_descriptors = {DEFAULT_MAX_FILE_DESCRIPTORS}
 
 # initial superuser role name to use when creating a new tenant
@@ -108,6 +111,8 @@ pub mod defaults {
 
 #background_task_maximum_delay = '{DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY}'
 
+#ingest_batch_size = {DEFAULT_INGEST_BATCH_SIZE}
+
 [tenant_config]
 #checkpoint_distance = {DEFAULT_CHECKPOINT_DISTANCE} # in bytes
 #checkpoint_timeout = {DEFAULT_CHECKPOINT_TIMEOUT}
@@ -233,6 +238,9 @@ pub struct PageServerConf {
     /// How many heatmap uploads may be done concurrency: lower values implicitly deprioritize
     /// heatmap uploads vs. other remote storage operations.
     pub heatmap_upload_concurrency: usize,
+
+    /// Maximum number of WAL records to be ingested and committed at the same time
+    pub ingest_batch_size: u64,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -314,6 +322,8 @@ struct PageServerConfigBuilder {
     control_plane_emergency_mode: BuilderValue<bool>,
 
     heatmap_upload_concurrency: BuilderValue<usize>,
+
+    ingest_batch_size: BuilderValue<u64>,
 }
 
 impl Default for PageServerConfigBuilder {
@@ -386,6 +396,8 @@ impl Default for PageServerConfigBuilder {
             control_plane_emergency_mode: Set(false),
 
             heatmap_upload_concurrency: Set(DEFAULT_HEATMAP_UPLOAD_CONCURRENCY),
+
+            ingest_batch_size: Set(DEFAULT_INGEST_BATCH_SIZE),
         }
     }
 }
@@ -534,6 +546,10 @@ impl PageServerConfigBuilder {
         self.heatmap_upload_concurrency = BuilderValue::Set(value)
     }
 
+    pub fn ingest_batch_size(&mut self, ingest_batch_size: u64) {
+        self.ingest_batch_size = BuilderValue::Set(ingest_batch_size)
+    }
+
     pub fn build(self) -> anyhow::Result<PageServerConf> {
         let concurrent_tenant_warmup = self
             .concurrent_tenant_warmup
@@ -632,10 +648,12 @@ impl PageServerConfigBuilder {
             control_plane_emergency_mode: self
                 .control_plane_emergency_mode
                 .ok_or(anyhow!("missing control_plane_emergency_mode"))?,
-
             heatmap_upload_concurrency: self
                 .heatmap_upload_concurrency
                 .ok_or(anyhow!("missing heatmap_upload_concurrency"))?,
+            ingest_batch_size: self
+                .ingest_batch_size
+                .ok_or(anyhow!("missing ingest_batch_size"))?,
         })
     }
 }
@@ -878,6 +896,7 @@ impl PageServerConf {
                 "heatmap_upload_concurrency" => {
                     builder.heatmap_upload_concurrency(parse_toml_u64(key, item)? as usize)
                 },
+                "ingest_batch_size" => builder.ingest_batch_size(parse_toml_u64(key, item)?),
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -949,6 +968,7 @@ impl PageServerConf {
             control_plane_api_token: None,
             control_plane_emergency_mode: false,
             heatmap_upload_concurrency: defaults::DEFAULT_HEATMAP_UPLOAD_CONCURRENCY,
+            ingest_batch_size: defaults::DEFAULT_INGEST_BATCH_SIZE,
         }
     }
 }
@@ -1177,7 +1197,8 @@ background_task_maximum_delay = '334 s'
                 control_plane_api: None,
                 control_plane_api_token: None,
                 control_plane_emergency_mode: false,
-                heatmap_upload_concurrency: defaults::DEFAULT_HEATMAP_UPLOAD_CONCURRENCY
+                heatmap_upload_concurrency: defaults::DEFAULT_HEATMAP_UPLOAD_CONCURRENCY,
+                ingest_batch_size: defaults::DEFAULT_INGEST_BATCH_SIZE,
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1238,7 +1259,8 @@ background_task_maximum_delay = '334 s'
                 control_plane_api: None,
                 control_plane_api_token: None,
                 control_plane_emergency_mode: false,
-                heatmap_upload_concurrency: defaults::DEFAULT_HEATMAP_UPLOAD_CONCURRENCY
+                heatmap_upload_concurrency: defaults::DEFAULT_HEATMAP_UPLOAD_CONCURRENCY,
+                ingest_batch_size: 100,
             },
             "Should be able to parse all basic config values correctly"
         );
@@ -1468,6 +1490,7 @@ threshold = "20m"
                 period: Duration::from_secs(10),
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
+                eviction_order: crate::disk_usage_eviction_task::EvictionOrder::AbsoluteAccessed,
             })
         );
         match &conf.default_tenant_conf.eviction_policy {

pageserver/src/disk_usage_eviction_task.rs

@@ -74,6 +74,45 @@ pub struct DiskUsageEvictionTaskConfig {
     pub period: Duration,
     #[cfg(feature = "testing")]
     pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
+    /// Select sorting for evicted layers
+    #[serde(default)]
+    pub eviction_order: EvictionOrder,
+}
+
+/// Selects the sort order for eviction candidates *after* per tenant `min_resident_size`
+/// partitioning.
+#[derive(Default, Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", content = "args")]
+pub enum EvictionOrder {
+    /// Order the layers to be evicted by how recently they have been accessed in absolute
+    /// time.
+    ///
+    /// This strategy is unfair when some tenants grow faster than others towards the slower
+    /// growing.
+    #[default]
+    AbsoluteAccessed,
+
+    /// Order the layers to be evicted by how recently they have been accessed relatively within
+    /// the set of resident layers of a tenant.
+    ///
+    /// This strategy will evict layers more fairly but is untested.
+    RelativeAccessed {
+        #[serde(default)]
+        highest_layer_count_loses_first: bool,
+    },
+}
+
+impl EvictionOrder {
+    /// Return true, if with [`Self::RelativeAccessed`] order the tenants with the highest layer
+    /// counts should be the first ones to have their layers evicted.
+    fn highest_layer_count_loses_first(&self) -> bool {
+        match self {
+            EvictionOrder::AbsoluteAccessed => false,
+            EvictionOrder::RelativeAccessed {
+                highest_layer_count_loses_first,
+            } => *highest_layer_count_loses_first,
+        }
+    }
 }
 
 #[derive(Default)]
@@ -192,7 +231,14 @@ async fn disk_usage_eviction_task_iteration(
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
-    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
+    let res = disk_usage_eviction_task_iteration_impl(
+        state,
+        storage,
+        usage_pre,
+        task_config.eviction_order,
+        cancel,
+    )
+    .await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
@@ -278,6 +324,7 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     state: &State,
     _storage: &GenericRemoteStorage,
     usage_pre: U,
+    eviction_order: EvictionOrder,
     cancel: &CancellationToken,
 ) -> anyhow::Result<IterationOutcome<U>> {
     // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
@@ -297,7 +344,7 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
         "running disk usage based eviction due to pressure"
     );
 
-    let candidates = match collect_eviction_candidates(cancel).await? {
+    let candidates = match collect_eviction_candidates(eviction_order, cancel).await? {
         EvictionCandidates::Cancelled => {
             return Ok(IterationOutcome::Cancelled);
         }
@@ -307,16 +354,16 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     // Debug-log the list of candidates
     let now = SystemTime::now();
     for (i, (partition, candidate)) in candidates.iter().enumerate() {
+        let nth = i + 1;
         let desc = candidate.layer.layer_desc();
+        let total_candidates = candidates.len();
+        let size = desc.file_size;
+        let rel = candidate.relative_last_activity;
         debug!(
-            "cand {}/{}: size={}, no_access_for={}us, partition={:?}, {}/{}/{}",
-            i + 1,
-            candidates.len(),
-            desc.file_size,
+            "cand {nth}/{total_candidates}: size={size}, rel_last_activity={rel}, no_access_for={}us, partition={partition:?}, {}/{}/{}",
             now.duration_since(candidate.last_activity_ts)
                 .unwrap()
                 .as_micros(),
-            partition,
             desc.tenant_shard_id,
             desc.timeline_id,
             candidate.layer,
@@ -459,6 +506,7 @@ struct EvictionCandidate {
     timeline: Arc<Timeline>,
     layer: Layer,
     last_activity_ts: SystemTime,
+    relative_last_activity: finite_f32::FiniteF32,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
@@ -478,24 +526,24 @@ enum EvictionCandidates {
 /// order. A caller that evicts in that order, until pressure is relieved, implements
 /// the eviction policy outlined in the module comment.
 ///
-/// # Example
+/// # Example with EvictionOrder::AbsoluteAccessed
 ///
 /// Imagine that there are two tenants, A and B, with five layers each, a-e.
 /// Each layer has size 100, and both tenant's min_resident_size is 150.
 /// The eviction order would be
 ///
 /// ```text
-/// partition last_activity_ts    tenant/layer
-/// Above     18:30               A/c
-/// Above     19:00               A/b
-/// Above     18:29               B/c
-/// Above     19:05               B/b
-/// Above     20:00               B/a
-/// Above     20:03               A/a
-/// Below     20:30               A/d
-/// Below     20:40               B/d
-/// Below     20:45               B/e
-/// Below     20:58               A/e
+/// partition last_activity_ts tenant/layer
+/// Above     18:30            A/c
+/// Above     19:00            A/b
+/// Above     18:29            B/c
+/// Above     19:05            B/b
+/// Above     20:00            B/a
+/// Above     20:03            A/a
+/// Below     20:30            A/d
+/// Below     20:40            B/d
+/// Below     20:45            B/e
+/// Below     20:58            A/e
 /// ```
 ///
 /// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
@@ -505,7 +553,77 @@ enum EvictionCandidates {
 /// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
 /// after exhauting the `Above` partition.
 /// So, we did not respect each tenant's min_resident_size.
+///
+/// # Example with EvictionOrder::RelativeAccessed
+///
+/// ```text
+/// partition relative_age last_activity_ts tenant/layer
+/// Above     0/4          18:30            A/c
+/// Above     0/4          18:29            B/c
+/// Above     1/4          19:00            A/b
+/// Above     1/4          19:05            B/b
+/// Above     2/4          20:00            B/a
+/// Above     2/4          20:03            A/a
+/// Below     3/4          20:30            A/d
+/// Below     3/4          20:40            B/d
+/// Below     4/4          20:45            B/e
+/// Below     4/4          20:58            A/e
+/// ```
+///
+/// With tenants having the same number of layers the picture does not change much. The same with
+/// A having many more layers **resident** (not all of them listed):
+///
+/// ```text
+/// Above       0/100      18:30            A/c
+/// Above       0/4        18:29            B/c
+/// Above       1/100      19:00            A/b
+/// Above       2/100      20:03            A/a
+/// Above       3/100      20:03            A/nth_3
+/// Above       4/100      20:03            A/nth_4
+///             ...
+/// Above       1/4        19:05            B/b
+/// Above      25/100      20:04            A/nth_25
+///             ...
+/// Above       2/4        20:00            B/a
+/// Above      50/100      20:10            A/nth_50
+///             ...
+/// Below       3/4        20:40            B/d
+/// Below      99/100      20:30            A/nth_99
+/// Below       4/4        20:45            B/e
+/// Below     100/100      20:58            A/nth_100
+/// ```
+///
+/// Now it's easier to see that because A has grown fast it has more layers to get evicted. What is
+/// difficult to see is what happens on the next round assuming the evicting 23 from the above list
+/// relieves the pressure (22 A layers gone, 1 B layers gone) but a new fast growing tenant C has
+/// appeared:
+///
+/// ```text
+/// Above       0/87       20:04            A/nth_23
+/// Above       0/3        19:05            B/b
+/// Above       0/50       20:59            C/nth_0
+/// Above       1/87       20:04            A/nth_24
+/// Above       1/50       21:00            C/nth_1
+/// Above       2/87       20:04            A/nth_25
+///             ...
+/// Above      16/50       21:02            C/nth_16
+/// Above       1/3        20:00            B/a
+/// Above      27/87       20:10            A/nth_50
+///             ...
+/// Below       2/3        20:40            B/d
+/// Below      49/50       21:05            C/nth_49
+/// Below      86/87       20:30            A/nth_99
+/// Below       3/3        20:45            B/e
+/// Below      50/50       21:05            C/nth_50
+/// Below      87/87       20:58            A/nth_100
+/// ```
+///
+/// Now relieving pressure with 23 layers would cost:
+/// - tenant A 14 layers
+/// - tenant B 1 layer
+/// - tenant C 8 layers
 async fn collect_eviction_candidates(
+    eviction_order: EvictionOrder,
     cancel: &CancellationToken,
 ) -> anyhow::Result<EvictionCandidates> {
     // get a snapshot of the list of tenants
@@ -591,12 +709,63 @@ async fn collect_eviction_candidates(
         tenant_candidates
             .sort_unstable_by_key(|(_, layer_info)| std::cmp::Reverse(layer_info.last_activity_ts));
         let mut cumsum: i128 = 0;
-        for (timeline, layer_info) in tenant_candidates.into_iter() {
+
+        // keeping the -1 or not decides if every tenant should lose their least recently accessed
+        // layer OR if this should happen in the order of having highest layer count:
+        let fudge = if eviction_order.highest_layer_count_loses_first() {
+            // relative_age vs. tenant layer count:
+            // - 0.1..=1.0 (10 layers)
+            // - 0.01..=1.0 (100 layers)
+            // - 0.001..=1.0 (1000 layers)
+            //
+            // leading to evicting less of the smallest tenants.
+            0
+        } else {
+            // use full 0.0..=1.0 range, which means even the smallest tenants could always lose a
+            // layer. the actual ordering is unspecified: for 10k tenants on a pageserver it could
+            // be that less than 10k layer evictions is enough, so we would not need to evict from
+            // all tenants.
+            //
+            // as the tenant ordering is now deterministic this could hit the same tenants
+            // disproportionetly on multiple invocations. alternative could be to remember how many
+            // layers did we evict last time from this tenant, and inject that as an additional
+            // fudge here.
+            1
+        };
+
+        let total = tenant_candidates
+            .len()
+            .checked_sub(fudge)
+            .filter(|&x| x > 0)
+            // support 0 or 1 resident layer tenants as well
+            .unwrap_or(1);
+        let divider = total as f32;
+
+        for (i, (timeline, layer_info)) in tenant_candidates.into_iter().enumerate() {
             let file_size = layer_info.file_size();
+
+            // as we iterate this reverse sorted list, the most recently accessed layer will always
+            // be 1.0; this is for us to evict it last.
+            let relative_last_activity = if matches!(
+                eviction_order,
+                EvictionOrder::RelativeAccessed { .. }
+            ) {
+                // another possibility: use buckets, like (256.0 * relative_last_activity) as u8 or
+                // similarly for u16. unsure how it would help.
+                finite_f32::FiniteF32::try_from_normalized((total - i) as f32 / divider)
+                    .unwrap_or_else(|val| {
+                        tracing::warn!(%fudge, "calculated invalid relative_last_activity for i={i}, total={total}: {val}");
+                        finite_f32::FiniteF32::ZERO
+                    })
+            } else {
+                finite_f32::FiniteF32::ZERO
+            };
+
             let candidate = EvictionCandidate {
                 timeline,
                 last_activity_ts: layer_info.last_activity_ts,
                 layer: layer_info.layer,
+                relative_last_activity,
             };
             let partition = if cumsum > min_resident_size as i128 {
                 MinResidentSizePartition::Above
@@ -610,8 +779,19 @@ async fn collect_eviction_candidates(
 
     debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
         "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first");
-    candidates
-        .sort_unstable_by_key(|(partition, candidate)| (*partition, candidate.last_activity_ts));
+
+    match eviction_order {
+        EvictionOrder::AbsoluteAccessed => {
+            candidates.sort_unstable_by_key(|(partition, candidate)| {
+                (*partition, candidate.last_activity_ts)
+            });
+        }
+        EvictionOrder::RelativeAccessed { .. } => {
+            candidates.sort_unstable_by_key(|(partition, candidate)| {
+                (*partition, candidate.relative_last_activity)
+            });
+        }
+    }
 
     Ok(EvictionCandidates::Finished(candidates))
 }
@@ -640,6 +820,66 @@ impl std::ops::Deref for TimelineKey {
     }
 }
 
+/// A totally ordered f32 subset we can use with sorting functions.
+mod finite_f32 {
+
+    /// A totally ordered f32 subset we can use with sorting functions.
+    #[derive(Clone, Copy, PartialEq)]
+    pub struct FiniteF32(f32);
+
+    impl std::fmt::Debug for FiniteF32 {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            std::fmt::Debug::fmt(&self.0, f)
+        }
+    }
+
+    impl std::fmt::Display for FiniteF32 {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            std::fmt::Display::fmt(&self.0, f)
+        }
+    }
+
+    impl std::cmp::Eq for FiniteF32 {}
+
+    impl std::cmp::PartialOrd for FiniteF32 {
+        fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+            Some(self.cmp(other))
+        }
+    }
+
+    impl std::cmp::Ord for FiniteF32 {
+        fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+            self.0.total_cmp(&other.0)
+        }
+    }
+
+    impl TryFrom<f32> for FiniteF32 {
+        type Error = f32;
+
+        fn try_from(value: f32) -> Result<Self, Self::Error> {
+            if value.is_finite() {
+                Ok(FiniteF32(value))
+            } else {
+                Err(value)
+            }
+        }
+    }
+
+    impl FiniteF32 {
+        pub const ZERO: FiniteF32 = FiniteF32(0.0);
+
+        pub fn try_from_normalized(value: f32) -> Result<Self, f32> {
+            if (0.0..=1.0).contains(&value) {
+                // -0.0 is within the range, make sure it is assumed 0.0..=1.0
+                let value = value.abs();
+                Ok(FiniteF32(value))
+            } else {
+                Err(value)
+            }
+        }
+    }
+}
+
 mod filesystem_level_usage {
     use anyhow::Context;
     use camino::Utf8Path;
@@ -721,6 +961,7 @@ mod filesystem_level_usage {
 
     #[test]
     fn max_usage_pct_pressure() {
+        use super::EvictionOrder;
         use super::Usage as _;
         use std::time::Duration;
         use utils::serde_percent::Percent;
@@ -732,6 +973,7 @@ mod filesystem_level_usage {
                 period: Duration::MAX,
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
+                eviction_order: EvictionOrder::default(),
             },
             total_bytes: 100_000,
             avail_bytes: 0,

pageserver/src/http/openapi_spec.yml

@@ -159,6 +159,12 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/ConflictError"
+        "412":
+          description: Deletion may not proceed, tenant is not in Active state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PreconditionFailedError"
         "500":
           description: Generic operation error
           content:

pageserver/src/http/routes.rs

@@ -14,16 +14,21 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
+use pageserver_api::models::TenantDetails;
+use pageserver_api::models::TenantShardSplitRequest;
+use pageserver_api::models::TenantShardSplitResponse;
 use pageserver_api::models::{
     DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
     TenantLoadRequest, TenantLocationConfigRequest,
 };
+use pageserver_api::shard::ShardCount;
 use pageserver_api::shard::TenantShardId;
 use remote_storage::GenericRemoteStorage;
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::auth::JwtAuth;
+use utils::failpoint_support::failpoints_handler;
 use utils::http::endpoint::request_span;
 use utils::http::json::json_request_or_empty_body;
 use utils::http::request::{get_request_param, must_get_query_param, parse_query_param};
@@ -65,9 +70,6 @@ use utils::{
     lsn::Lsn,
 };
 
-// Imports only used for testing APIs
-use pageserver_api::models::ConfigureFailpointsRequest;
-
 // For APIs that require an Active tenant, how long should we block waiting for that state?
 // This is not functionally necessary (clients will retry), but avoids generating a lot of
 // failed API calls while tenants are activating.
@@ -258,7 +260,7 @@ impl From<SetNewTenantConfigError> for ApiError {
             SetNewTenantConfigError::GetTenant(tid) => {
                 ApiError::NotFound(anyhow!("tenant {}", tid).into())
             }
-            e @ SetNewTenantConfigError::Persist(_) => {
+            e @ (SetNewTenantConfigError::Persist(_) | SetNewTenantConfigError::Other(_)) => {
                 ApiError::InternalServerError(anyhow::Error::new(e))
             }
         }
@@ -307,6 +309,7 @@ impl From<crate::tenant::delete::DeleteTenantError> for ApiError {
             SlotUpsertError(e) => e.into(),
             Other(o) => ApiError::InternalServerError(o),
             e @ InvalidState(_) => ApiError::PreconditionFailed(e.to_string().into_boxed_str()),
+            Cancelled => ApiError::ShuttingDown,
         }
     }
 }
@@ -592,8 +595,6 @@ async fn get_lsn_by_timestamp_handler(
         )));
     }
 
-    let version: Option<u8> = parse_query_param(&request, "version")?;
-
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     let timestamp_raw = must_get_query_param(&request, "timestamp")?;
     let timestamp = humantime::parse_rfc3339(&timestamp_raw)
@@ -606,31 +607,18 @@ async fn get_lsn_by_timestamp_handler(
     let result = timeline
         .find_lsn_for_timestamp(timestamp_pg, &cancel, &ctx)
         .await?;
-
-    if version.unwrap_or(0) > 1 {
-        #[derive(serde::Serialize)]
-        struct Result {
-            lsn: Lsn,
-            kind: &'static str,
-        }
-        let (lsn, kind) = match result {
-            LsnForTimestamp::Present(lsn) => (lsn, "present"),
-            LsnForTimestamp::Future(lsn) => (lsn, "future"),
-            LsnForTimestamp::Past(lsn) => (lsn, "past"),
-            LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
-        };
-        json_response(StatusCode::OK, Result { lsn, kind })
-    } else {
-        // FIXME: this is a temporary crutch not to break backwards compatibility
-        // See https://github.com/neondatabase/neon/pull/5608
-        let result = match result {
-            LsnForTimestamp::Present(lsn) => format!("{lsn}"),
-            LsnForTimestamp::Future(_lsn) => "future".into(),
-            LsnForTimestamp::Past(_lsn) => "past".into(),
-            LsnForTimestamp::NoData(_lsn) => "nodata".into(),
-        };
-        json_response(StatusCode::OK, result)
+    #[derive(serde::Serialize)]
+    struct Result {
+        lsn: Lsn,
+        kind: &'static str,
     }
+    let (lsn, kind) = match result {
+        LsnForTimestamp::Present(lsn) => (lsn, "present"),
+        LsnForTimestamp::Future(lsn) => (lsn, "future"),
+        LsnForTimestamp::Past(lsn) => (lsn, "past"),
+        LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
+    };
+    json_response(StatusCode::OK, Result { lsn, kind })
 }
 
 async fn get_timestamp_of_lsn_handler(
@@ -872,11 +860,14 @@ async fn tenant_status(
         }
 
         let state = tenant.current_state();
-        Result::<_, ApiError>::Ok(TenantInfo {
-            id: tenant_shard_id,
-            state: state.clone(),
-            current_physical_size: Some(current_physical_size),
-            attachment_status: state.attachment_status(),
+        Result::<_, ApiError>::Ok(TenantDetails {
+            tenant_info: TenantInfo {
+                id: tenant_shard_id,
+                state: state.clone(),
+                current_physical_size: Some(current_physical_size),
+                attachment_status: state.attachment_status(),
+            },
+            timelines: tenant.list_timeline_ids(),
         })
     }
     .instrument(info_span!("tenant_status_handler",
@@ -897,7 +888,9 @@ async fn tenant_delete_handler(
 
     let state = get_state(&request);
 
-    mgr::delete_tenant(state.conf, state.remote_storage.clone(), tenant_shard_id)
+    state
+        .tenant_manager
+        .delete_tenant(tenant_shard_id, ACTIVE_TENANT_TIMEOUT)
         .instrument(info_span!("tenant_delete_handler",
             tenant_id = %tenant_shard_id.tenant_id,
             shard = %tenant_shard_id.shard_slug()
@@ -997,6 +990,25 @@ async fn tenant_size_handler(
     )
 }
 
+async fn tenant_shard_split_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let req: TenantShardSplitRequest = json_request(&mut request).await?;
+
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let state = get_state(&request);
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+
+    let new_shards = state
+        .tenant_manager
+        .shard_split(tenant_shard_id, ShardCount(req.new_shard_count), &ctx)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, TenantShardSplitResponse { new_shards })
+}
+
 async fn layer_map_info_handler(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -1161,6 +1173,7 @@ async fn tenant_create_handler(
         state.conf,
         tenant_conf,
         target_tenant_id,
+        request_data.shard_parameters,
         generation,
         state.tenant_resources(),
         &ctx,
@@ -1301,34 +1314,6 @@ async fn handle_tenant_break(
     json_response(StatusCode::OK, ())
 }
 
-async fn failpoints_handler(
-    mut request: Request<Body>,
-    _cancel: CancellationToken,
-) -> Result<Response<Body>, ApiError> {
-    if !fail::has_failpoints() {
-        return Err(ApiError::BadRequest(anyhow!(
-            "Cannot manage failpoints because pageserver was compiled without failpoints support"
-        )));
-    }
-
-    let failpoints: ConfigureFailpointsRequest = json_request(&mut request).await?;
-    for fp in failpoints {
-        info!("cfg failpoint: {} {}", fp.name, fp.actions);
-
-        // We recognize one extra "action" that's not natively recognized
-        // by the failpoints crate: exit, to immediately kill the process
-        let cfg_result = crate::failpoint_support::apply_failpoint(&fp.name, &fp.actions);
-
-        if let Err(err_msg) = cfg_result {
-            return Err(ApiError::BadRequest(anyhow!(
-                "Failed to configure failpoints: {err_msg}"
-            )));
-        }
-    }
-
-    json_response(StatusCode::OK, ())
-}
-
 // Run GC immediately on given timeline.
 async fn timeline_gc_handler(
     mut request: Request<Body>,
@@ -1577,19 +1562,22 @@ async fn disk_usage_eviction_run(
     struct Config {
         /// How many bytes to evict before reporting that pressure is relieved.
         evict_bytes: u64,
+
+        #[serde(default)]
+        eviction_order: crate::disk_usage_eviction_task::EvictionOrder,
     }
 
     #[derive(Debug, Clone, Copy, serde::Serialize)]
     struct Usage {
         // remains unchanged after instantiation of the struct
-        config: Config,
+        evict_bytes: u64,
         // updated by `add_available_bytes`
         freed_bytes: u64,
     }
 
     impl crate::disk_usage_eviction_task::Usage for Usage {
         fn has_pressure(&self) -> bool {
-            self.config.evict_bytes > self.freed_bytes
+            self.evict_bytes > self.freed_bytes
         }
 
         fn add_available_bytes(&mut self, bytes: u64) {
@@ -1600,7 +1588,7 @@ async fn disk_usage_eviction_run(
     let config = json_request::<Config>(&mut r).await?;
 
     let usage = Usage {
-        config,
+        evict_bytes: config.evict_bytes,
         freed_bytes: 0,
     };
 
@@ -1615,7 +1603,11 @@ async fn disk_usage_eviction_run(
     let state = state.disk_usage_eviction_state.clone();
 
     let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
-        &state, storage, usage, &cancel,
+        &state,
+        storage,
+        usage,
+        config.eviction_order,
+        &cancel,
     )
     .await;
 
@@ -1827,6 +1819,9 @@ pub fn make_router(
         .put("/v1/tenant/config", |r| {
             api_handler(r, update_tenant_config_handler)
         })
+        .put("/v1/tenant/:tenant_shard_id/shard_split", |r| {
+            api_handler(r, tenant_shard_split_handler)
+        })
         .get("/v1/tenant/:tenant_shard_id/config", |r| {
             api_handler(r, get_tenant_config_handler)
         })

pageserver/src/import_datadir.rs

@@ -21,6 +21,7 @@ use tracing::*;
 use walkdir::WalkDir;
 
 use crate::context::RequestContext;
+use crate::metrics::WAL_INGEST;
 use crate::pgdatadir_mapping::*;
 use crate::tenant::remote_timeline_client::INITDB_PATH;
 use crate::tenant::Timeline;
@@ -312,13 +313,16 @@ async fn import_wal(
         waldecoder.feed_bytes(&buf);
 
         let mut nrecords = 0;
-        let mut modification = tline.begin_modification(endpoint);
+        let mut modification = tline.begin_modification(last_lsn);
         let mut decoded = DecodedWALRecord::default();
         while last_lsn <= endpoint {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 walingest
                     .ingest_record(recdata, lsn, &mut modification, &mut decoded, ctx)
                     .await?;
+                WAL_INGEST.records_committed.inc();
+
+                modification.commit(ctx).await?;
                 last_lsn = lsn;
 
                 nrecords += 1;
@@ -448,13 +452,14 @@ pub async fn import_wal_from_tar(
 
         waldecoder.feed_bytes(&bytes[offset..]);
 
-        let mut modification = tline.begin_modification(end_lsn);
+        let mut modification = tline.begin_modification(last_lsn);
         let mut decoded = DecodedWALRecord::default();
         while last_lsn <= end_lsn {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 walingest
                     .ingest_record(recdata, lsn, &mut modification, &mut decoded, ctx)
                     .await?;
+                modification.commit(ctx).await?;
                 last_lsn = lsn;
 
                 debug!("imported record at {} (end {})", lsn, end_lsn);

pageserver/src/lib.rs

@@ -25,8 +25,6 @@ pub mod walingest;
 pub mod walrecord;
 pub mod walredo;
 
-pub mod failpoint_support;
-
 use crate::task_mgr::TaskKind;
 use camino::Utf8Path;
 use deletion_queue::DeletionQueue;

pageserver/src/metrics.rs

@@ -522,14 +522,18 @@ pub(crate) mod initial_logical_size {
     impl StartCalculation {
         pub(crate) fn first(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
             let circumstances_label: &'static str = circumstances.into();
-            self.0.with_label_values(&["first", circumstances_label]);
+            self.0
+                .with_label_values(&["first", circumstances_label])
+                .inc();
             OngoingCalculationGuard {
                 inc_drop_calculation: Some(DROP_CALCULATION.first.clone()),
             }
         }
         pub(crate) fn retry(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
             let circumstances_label: &'static str = circumstances.into();
-            self.0.with_label_values(&["retry", circumstances_label]);
+            self.0
+                .with_label_values(&["retry", circumstances_label])
+                .inc();
             OngoingCalculationGuard {
                 inc_drop_calculation: Some(DROP_CALCULATION.retry.clone()),
             }
@@ -1019,12 +1023,62 @@ static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static SMGR_QUERY_TIME_GLOBAL_BUCKETS: Lazy<Vec<f64>> = Lazy::new(|| {
+    [
+        1,
+        10,
+        20,
+        40,
+        60,
+        80,
+        100,
+        200,
+        300,
+        400,
+        500,
+        600,
+        700,
+        800,
+        900,
+        1_000, // 1ms
+        2_000,
+        4_000,
+        6_000,
+        8_000,
+        10_000, // 10ms
+        20_000,
+        40_000,
+        60_000,
+        80_000,
+        100_000,
+        200_000,
+        400_000,
+        600_000,
+        800_000,
+        1_000_000, // 1s
+        2_000_000,
+        4_000_000,
+        6_000_000,
+        8_000_000,
+        10_000_000, // 10s
+        20_000_000,
+        50_000_000,
+        100_000_000,
+        200_000_000,
+        1_000_000_000, // 1000s
+    ]
+    .into_iter()
+    .map(Duration::from_micros)
+    .map(|d| d.as_secs_f64())
+    .collect()
+});
+
 static SMGR_QUERY_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_smgr_query_seconds_global",
         "Time spent on smgr query handling, aggregated by query type.",
         &["smgr_query_type"],
-        CRITICAL_OP_BUCKETS.into(),
+        SMGR_QUERY_TIME_GLOBAL_BUCKETS.clone(),
     )
     .expect("failed to define a metric")
 });

pageserver/src/page_service.rs

@@ -53,7 +53,7 @@ use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir::import_wal_from_tar;
 use crate::metrics;
 use crate::metrics::LIVE_CONNECTIONS_COUNT;
-use crate::pgdatadir_mapping::rel_block_to_key;
+use crate::pgdatadir_mapping::{rel_block_to_key, Version};
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
@@ -747,7 +747,7 @@ impl PageServerHandler {
                 .await?;
 
         let exists = timeline
-            .get_rel_exists(req.rel, lsn, req.latest, ctx)
+            .get_rel_exists(req.rel, Version::Lsn(lsn), req.latest, ctx)
             .await?;
 
         Ok(PagestreamBeMessage::Exists(PagestreamExistsResponse {
@@ -766,7 +766,9 @@ impl PageServerHandler {
             Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
-        let n_blocks = timeline.get_rel_size(req.rel, lsn, req.latest, ctx).await?;
+        let n_blocks = timeline
+            .get_rel_size(req.rel, Version::Lsn(lsn), req.latest, ctx)
+            .await?;
 
         Ok(PagestreamBeMessage::Nblocks(PagestreamNblocksResponse {
             n_blocks,
@@ -785,7 +787,13 @@ impl PageServerHandler {
                 .await?;
 
         let total_blocks = timeline
-            .get_db_size(DEFAULTTABLESPACE_OID, req.dbnode, lsn, req.latest, ctx)
+            .get_db_size(
+                DEFAULTTABLESPACE_OID,
+                req.dbnode,
+                Version::Lsn(lsn),
+                req.latest,
+                ctx,
+            )
             .await?;
         let db_size = total_blocks as i64 * BLCKSZ as i64;
 
@@ -794,7 +802,7 @@ impl PageServerHandler {
         }))
     }
 
-    async fn handle_get_page_at_lsn_request(
+    async fn do_handle_get_page_at_lsn_request(
         &self,
         timeline: &Timeline,
         req: &PagestreamGetPageRequest,
@@ -804,20 +812,25 @@ impl PageServerHandler {
         let lsn =
             Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
-        /*
-        // Add a 1s delay to some requests. The delay helps the requests to
-        // hit the race condition from github issue #1047 more easily.
-        use rand::Rng;
-        if rand::thread_rng().gen::<u8>() < 5 {
-            std::thread::sleep(std::time::Duration::from_millis(1000));
-        }
-        */
+        let page = timeline
+            .get_rel_page_at_lsn(req.rel, req.blkno, Version::Lsn(lsn), req.latest, ctx)
+            .await?;
 
+        Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
+            page,
+        }))
+    }
+
+    async fn handle_get_page_at_lsn_request(
+        &self,
+        timeline: &Timeline,
+        req: &PagestreamGetPageRequest,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<PagestreamBeMessage> {
         let key = rel_block_to_key(req.rel, req.blkno);
-        let page = if timeline.get_shard_identity().is_key_local(&key) {
-            timeline
-                .get_rel_page_at_lsn(req.rel, req.blkno, lsn, req.latest, ctx)
-                .await?
+        if timeline.get_shard_identity().is_key_local(&key) {
+            self.do_handle_get_page_at_lsn_request(timeline, req, ctx)
+                .await
         } else {
             // The Tenant shard we looked up at connection start does not hold this particular
             // key: look for other shards in this tenant.  This scenario occurs if a pageserver
@@ -844,6 +857,8 @@ impl PageServerHandler {
                     // informed yet.
                     //
                     // https://github.com/neondatabase/neon/issues/6038
+                    tracing::warn!("Page request routed to wrong shard: my identity {:?}, should go to shard {}, key {}",
+                        timeline.get_shard_identity(), timeline.get_shard_identity().get_shard_number(&key).0, key);
                     return Err(anyhow::anyhow!("Request routed to wrong shard"));
                 }
                 Err(e) => return Err(e.into()),
@@ -852,14 +867,10 @@ impl PageServerHandler {
             // Take a GateGuard for the duration of this request.  If we were using our main Timeline object,
             // the GateGuard was already held over the whole connection.
             let _timeline_guard = timeline.gate.enter().map_err(|_| QueryError::Shutdown)?;
-            timeline
-                .get_rel_page_at_lsn(req.rel, req.blkno, lsn, req.latest, ctx)
-                .await?
-        };
 
-        Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
-            page,
-        }))
+            self.do_handle_get_page_at_lsn_request(&timeline, req, ctx)
+                .await
+        }
     }
 
     #[allow(clippy::too_many_arguments)]

pageserver/src/pgdatadir_mapping.rs

@@ -11,7 +11,7 @@ use crate::context::RequestContext;
 use crate::keyspace::{KeySpace, KeySpaceAccum};
 use crate::repository::*;
 use crate::walrecord::NeonWalRecord;
-use anyhow::Context;
+use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes};
 use pageserver_api::key::is_rel_block_key;
 use pageserver_api::reltag::{RelTag, SlruKind};
@@ -147,6 +147,7 @@ impl Timeline {
     {
         DatadirModification {
             tline: self,
+            pending_lsns: Vec::new(),
             pending_updates: HashMap::new(),
             pending_deletions: Vec::new(),
             pending_nblocks: 0,
@@ -163,7 +164,7 @@ impl Timeline {
         &self,
         tag: RelTag,
         blknum: BlockNumber,
-        lsn: Lsn,
+        version: Version<'_>,
         latest: bool,
         ctx: &RequestContext,
     ) -> Result<Bytes, PageReconstructError> {
@@ -173,17 +174,20 @@ impl Timeline {
             ));
         }
 
-        let nblocks = self.get_rel_size(tag, lsn, latest, ctx).await?;
+        let nblocks = self.get_rel_size(tag, version, latest, ctx).await?;
         if blknum >= nblocks {
             debug!(
                 "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
-                tag, blknum, lsn, nblocks
+                tag,
+                blknum,
+                version.get_lsn(),
+                nblocks
             );
             return Ok(ZERO_PAGE.clone());
         }
 
         let key = rel_block_to_key(tag, blknum);
-        self.get(key, lsn, ctx).await
+        version.get(self, key, ctx).await
     }
 
     // Get size of a database in blocks
@@ -191,16 +195,16 @@ impl Timeline {
         &self,
         spcnode: Oid,
         dbnode: Oid,
-        lsn: Lsn,
+        version: Version<'_>,
         latest: bool,
         ctx: &RequestContext,
     ) -> Result<usize, PageReconstructError> {
         let mut total_blocks = 0;
 
-        let rels = self.list_rels(spcnode, dbnode, lsn, ctx).await?;
+        let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
 
         for rel in rels {
-            let n_blocks = self.get_rel_size(rel, lsn, latest, ctx).await?;
+            let n_blocks = self.get_rel_size(rel, version, latest, ctx).await?;
             total_blocks += n_blocks as usize;
         }
         Ok(total_blocks)
@@ -210,7 +214,7 @@ impl Timeline {
     pub async fn get_rel_size(
         &self,
         tag: RelTag,
-        lsn: Lsn,
+        version: Version<'_>,
         latest: bool,
         ctx: &RequestContext,
     ) -> Result<BlockNumber, PageReconstructError> {
@@ -220,12 +224,12 @@ impl Timeline {
             ));
         }
 
-        if let Some(nblocks) = self.get_cached_rel_size(&tag, lsn) {
+        if let Some(nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
             return Ok(nblocks);
         }
 
         if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
-            && !self.get_rel_exists(tag, lsn, latest, ctx).await?
+            && !self.get_rel_exists(tag, version, latest, ctx).await?
         {
             // FIXME: Postgres sometimes calls smgrcreate() to create
             // FSM, and smgrnblocks() on it immediately afterwards,
@@ -235,7 +239,7 @@ impl Timeline {
         }
 
         let key = rel_size_to_key(tag);
-        let mut buf = self.get(key, lsn, ctx).await?;
+        let mut buf = version.get(self, key, ctx).await?;
         let nblocks = buf.get_u32_le();
 
         if latest {
@@ -246,7 +250,7 @@ impl Timeline {
             // latest=true, then it can not cause cache corruption, because with latest=true
             // pageserver choose max(request_lsn, last_written_lsn) and so cached value will be
             // associated with most recent value of LSN.
-            self.update_cached_rel_size(tag, lsn, nblocks);
+            self.update_cached_rel_size(tag, version.get_lsn(), nblocks);
         }
         Ok(nblocks)
     }
@@ -255,7 +259,7 @@ impl Timeline {
     pub async fn get_rel_exists(
         &self,
         tag: RelTag,
-        lsn: Lsn,
+        version: Version<'_>,
         _latest: bool,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
@@ -266,12 +270,12 @@ impl Timeline {
         }
 
         // first try to lookup relation in cache
-        if let Some(_nblocks) = self.get_cached_rel_size(&tag, lsn) {
+        if let Some(_nblocks) = self.get_cached_rel_size(&tag, version.get_lsn()) {
             return Ok(true);
         }
         // fetch directory listing
         let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
-        let buf = self.get(key, lsn, ctx).await?;
+        let buf = version.get(self, key, ctx).await?;
 
         match RelDirectory::des(&buf).context("deserialization failure") {
             Ok(dir) => {
@@ -291,12 +295,12 @@ impl Timeline {
         &self,
         spcnode: Oid,
         dbnode: Oid,
-        lsn: Lsn,
+        version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<HashSet<RelTag>, PageReconstructError> {
         // fetch directory listing
         let key = rel_dir_to_key(spcnode, dbnode);
-        let buf = self.get(key, lsn, ctx).await?;
+        let buf = version.get(self, key, ctx).await?;
 
         match RelDirectory::des(&buf).context("deserialization failure") {
             Ok(dir) => {
@@ -332,11 +336,11 @@ impl Timeline {
         &self,
         kind: SlruKind,
         segno: u32,
-        lsn: Lsn,
+        version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<BlockNumber, PageReconstructError> {
         let key = slru_segment_size_to_key(kind, segno);
-        let mut buf = self.get(key, lsn, ctx).await?;
+        let mut buf = version.get(self, key, ctx).await?;
         Ok(buf.get_u32_le())
     }
 
@@ -345,12 +349,12 @@ impl Timeline {
         &self,
         kind: SlruKind,
         segno: u32,
-        lsn: Lsn,
+        version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
         // fetch directory listing
         let key = slru_dir_to_key(kind);
-        let buf = self.get(key, lsn, ctx).await?;
+        let buf = version.get(self, key, ctx).await?;
 
         match SlruSegmentDirectory::des(&buf).context("deserialization failure") {
             Ok(dir) => {
@@ -501,11 +505,11 @@ impl Timeline {
         mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
     ) -> Result<T, PageReconstructError> {
         for segno in self
-            .list_slru_segments(SlruKind::Clog, probe_lsn, ctx)
+            .list_slru_segments(SlruKind::Clog, Version::Lsn(probe_lsn), ctx)
             .await?
         {
             let nblocks = self
-                .get_slru_segment_size(SlruKind::Clog, segno, probe_lsn, ctx)
+                .get_slru_segment_size(SlruKind::Clog, segno, Version::Lsn(probe_lsn), ctx)
                 .await?;
             for blknum in (0..nblocks).rev() {
                 let clog_page = self
@@ -531,13 +535,13 @@ impl Timeline {
     pub async fn list_slru_segments(
         &self,
         kind: SlruKind,
-        lsn: Lsn,
+        version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<HashSet<u32>, PageReconstructError> {
         // fetch directory entry
         let key = slru_dir_to_key(kind);
 
-        let buf = self.get(key, lsn, ctx).await?;
+        let buf = version.get(self, key, ctx).await?;
         match SlruSegmentDirectory::des(&buf).context("deserialization failure") {
             Ok(dir) => Ok(dir.segments),
             Err(e) => Err(PageReconstructError::from(e)),
@@ -548,12 +552,12 @@ impl Timeline {
         &self,
         spcnode: Oid,
         dbnode: Oid,
-        lsn: Lsn,
+        version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<Bytes, PageReconstructError> {
         let key = relmap_file_key(spcnode, dbnode);
 
-        let buf = self.get(key, lsn, ctx).await?;
+        let buf = version.get(self, key, ctx).await?;
         Ok(buf)
     }
 
@@ -652,7 +656,10 @@ impl Timeline {
 
         let mut total_size: u64 = 0;
         for (spcnode, dbnode) in dbdir.dbdirs.keys() {
-            for rel in self.list_rels(*spcnode, *dbnode, lsn, ctx).await? {
+            for rel in self
+                .list_rels(*spcnode, *dbnode, Version::Lsn(lsn), ctx)
+                .await?
+            {
                 if self.cancel.is_cancelled() {
                     return Err(CalculateLogicalSizeError::Cancelled);
                 }
@@ -692,7 +699,7 @@ impl Timeline {
             result.add_key(rel_dir_to_key(spcnode, dbnode));
 
             let mut rels: Vec<RelTag> = self
-                .list_rels(spcnode, dbnode, lsn, ctx)
+                .list_rels(spcnode, dbnode, Version::Lsn(lsn), ctx)
                 .await?
                 .into_iter()
                 .collect();
@@ -799,18 +806,39 @@ pub struct DatadirModification<'a> {
     /// in the state in 'tline' yet.
     pub tline: &'a Timeline,
 
-    /// Lsn assigned by begin_modification
-    pub lsn: Lsn,
+    /// Current LSN of the modification
+    lsn: Lsn,
 
     // The modifications are not applied directly to the underlying key-value store.
     // The put-functions add the modifications here, and they are flushed to the
     // underlying key-value store by the 'finish' function.
-    pending_updates: HashMap<Key, Value>,
-    pending_deletions: Vec<Range<Key>>,
+    pending_lsns: Vec<Lsn>,
+    pending_updates: HashMap<Key, Vec<(Lsn, Value)>>,
+    pending_deletions: Vec<(Range<Key>, Lsn)>,
     pending_nblocks: i64,
 }
 
 impl<'a> DatadirModification<'a> {
+    /// Get the current lsn
+    pub(crate) fn get_lsn(&self) -> Lsn {
+        self.lsn
+    }
+
+    /// Set the current lsn
+    pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> anyhow::Result<()> {
+        ensure!(
+            lsn >= self.lsn,
+            "setting an older lsn {} than {} is not allowed",
+            lsn,
+            self.lsn
+        );
+        if lsn > self.lsn {
+            self.pending_lsns.push(self.lsn);
+            self.lsn = lsn;
+        }
+        Ok(())
+    }
+
     /// Initialize a completely new repository.
     ///
     /// This inserts the directory metadata entries that are assumed to
@@ -984,11 +1012,9 @@ impl<'a> DatadirModification<'a> {
         dbnode: Oid,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        let req_lsn = self.tline.get_last_record_lsn();
-
         let total_blocks = self
             .tline
-            .get_db_size(spcnode, dbnode, req_lsn, true, ctx)
+            .get_db_size(spcnode, dbnode, Version::Modified(self), true, ctx)
             .await?;
 
         // Remove entry from dbdir
@@ -1077,8 +1103,11 @@ impl<'a> DatadirModification<'a> {
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
-        let last_lsn = self.tline.get_last_record_lsn();
-        if self.tline.get_rel_exists(rel, last_lsn, true, ctx).await? {
+        if self
+            .tline
+            .get_rel_exists(rel, Version::Modified(self), true, ctx)
+            .await?
+        {
             let size_key = rel_size_to_key(rel);
             // Fetch the old size first
             let old_size = self.get(size_key, ctx).await?.get_u32_le();
@@ -1323,17 +1352,23 @@ impl<'a> DatadirModification<'a> {
         let writer = self.tline.writer().await;
 
         // Flush relation and  SLRU data blocks, keep metadata.
-        let mut retained_pending_updates = HashMap::new();
-        for (key, value) in self.pending_updates.drain() {
-            if is_rel_block_key(&key) || is_slru_block_key(key) {
-                // This bails out on first error without modifying pending_updates.
-                // That's Ok, cf this function's doc comment.
-                writer.put(key, self.lsn, &value, ctx).await?;
-            } else {
-                retained_pending_updates.insert(key, value);
+        let mut retained_pending_updates = HashMap::<_, Vec<_>>::new();
+        for (key, values) in self.pending_updates.drain() {
+            for (lsn, value) in values {
+                if is_rel_block_key(&key) || is_slru_block_key(key) {
+                    // This bails out on first error without modifying pending_updates.
+                    // That's Ok, cf this function's doc comment.
+                    writer.put(key, lsn, &value, ctx).await?;
+                } else {
+                    retained_pending_updates
+                        .entry(key)
+                        .or_default()
+                        .push((lsn, value));
+                }
             }
         }
-        self.pending_updates.extend(retained_pending_updates);
+
+        self.pending_updates = retained_pending_updates;
 
         if pending_nblocks != 0 {
             writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
@@ -1350,18 +1385,28 @@ impl<'a> DatadirModification<'a> {
     ///
     pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
         let writer = self.tline.writer().await;
-        let lsn = self.lsn;
+
         let pending_nblocks = self.pending_nblocks;
         self.pending_nblocks = 0;
 
-        for (key, value) in self.pending_updates.drain() {
-            writer.put(key, lsn, &value, ctx).await?;
-        }
-        for key_range in self.pending_deletions.drain(..) {
-            writer.delete(key_range, lsn).await?;
+        if !self.pending_updates.is_empty() {
+            writer.put_batch(&self.pending_updates, ctx).await?;
+            self.pending_updates.clear();
         }
 
-        writer.finish_write(lsn);
+        if !self.pending_deletions.is_empty() {
+            writer.delete_batch(&self.pending_deletions).await?;
+            self.pending_deletions.clear();
+        }
+
+        self.pending_lsns.push(self.lsn);
+        for pending_lsn in self.pending_lsns.drain(..) {
+            // Ideally, we should be able to call writer.finish_write() only once
+            // with the highest LSN. However, the last_record_lsn variable in the
+            // timeline keeps track of the latest LSN and the immediate previous LSN
+            // so we need to record every LSN to not leave a gap between them.
+            writer.finish_write(pending_lsn);
+        }
 
         if pending_nblocks != 0 {
             writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
@@ -1370,44 +1415,86 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
-    pub(crate) fn is_empty(&self) -> bool {
-        self.pending_updates.is_empty() && self.pending_deletions.is_empty()
+    pub(crate) fn len(&self) -> usize {
+        self.pending_updates.len() + self.pending_deletions.len()
     }
 
     // Internal helper functions to batch the modifications
 
     async fn get(&self, key: Key, ctx: &RequestContext) -> Result<Bytes, PageReconstructError> {
-        // Have we already updated the same key? Read the pending updated
+        // Have we already updated the same key? Read the latest pending updated
         // version in that case.
         //
         // Note: we don't check pending_deletions. It is an error to request a
         // value that has been removed, deletion only avoids leaking storage.
-        if let Some(value) = self.pending_updates.get(&key) {
-            if let Value::Image(img) = value {
-                Ok(img.clone())
-            } else {
-                // Currently, we never need to read back a WAL record that we
-                // inserted in the same "transaction". All the metadata updates
-                // work directly with Images, and we never need to read actual
-                // data pages. We could handle this if we had to, by calling
-                // the walredo manager, but let's keep it simple for now.
-                Err(PageReconstructError::from(anyhow::anyhow!(
-                    "unexpected pending WAL record"
-                )))
+        if let Some(values) = self.pending_updates.get(&key) {
+            if let Some((_, value)) = values.last() {
+                return if let Value::Image(img) = value {
+                    Ok(img.clone())
+                } else {
+                    // Currently, we never need to read back a WAL record that we
+                    // inserted in the same "transaction". All the metadata updates
+                    // work directly with Images, and we never need to read actual
+                    // data pages. We could handle this if we had to, by calling
+                    // the walredo manager, but let's keep it simple for now.
+                    Err(PageReconstructError::from(anyhow::anyhow!(
+                        "unexpected pending WAL record"
+                    )))
+                };
             }
-        } else {
-            let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
-            self.tline.get(key, lsn, ctx).await
         }
+        let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
+        self.tline.get(key, lsn, ctx).await
     }
 
     fn put(&mut self, key: Key, val: Value) {
-        self.pending_updates.insert(key, val);
+        let values = self.pending_updates.entry(key).or_default();
+        // Replace the previous value if it exists at the same lsn
+        if let Some((last_lsn, last_value)) = values.last_mut() {
+            if *last_lsn == self.lsn {
+                *last_value = val;
+                return;
+            }
+        }
+        values.push((self.lsn, val));
     }
 
     fn delete(&mut self, key_range: Range<Key>) {
         trace!("DELETE {}-{}", key_range.start, key_range.end);
-        self.pending_deletions.push(key_range);
+        self.pending_deletions.push((key_range, self.lsn));
+    }
+}
+
+/// This struct facilitates accessing either a committed key from the timeline at a
+/// specific LSN, or the latest uncommitted key from a pending modification.
+/// During WAL ingestion, the records from multiple LSNs may be batched in the same
+/// modification before being flushed to the timeline. Hence, the routines in WalIngest
+/// need to look up the keys in the modification first before looking them up in the
+/// timeline to not miss the latest updates.
+#[derive(Clone, Copy)]
+pub enum Version<'a> {
+    Lsn(Lsn),
+    Modified(&'a DatadirModification<'a>),
+}
+
+impl<'a> Version<'a> {
+    async fn get(
+        &self,
+        timeline: &Timeline,
+        key: Key,
+        ctx: &RequestContext,
+    ) -> Result<Bytes, PageReconstructError> {
+        match self {
+            Version::Lsn(lsn) => timeline.get(key, *lsn, ctx).await,
+            Version::Modified(modification) => modification.get(key, ctx).await,
+        }
+    }
+
+    fn get_lsn(&self) -> Lsn {
+        match self {
+            Version::Lsn(lsn) => *lsn,
+            Version::Modified(modification) => modification.lsn,
+        }
     }
 }
 
@@ -1776,6 +1863,7 @@ pub fn is_inherited_key(key: Key) -> bool {
     key != AUX_FILES_KEY
 }
 
+/// Guaranteed to return `Ok()` if [[is_rel_block_key]] returns `true` for `key`.
 pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     Ok(match key.field1 {
         0x00 => (
@@ -1790,7 +1878,6 @@ pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
         _ => anyhow::bail!("unexpected value kind 0x{:02x}", key.field1),
     })
 }
-
 pub fn is_rel_fsm_block_key(key: Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0 && key.field5 == FSM_FORKNUM && key.field6 != 0xffffffff
 }

pageserver/src/task_mgr.rs

@@ -147,7 +147,7 @@ pub(crate) static BACKGROUND_RUNTIME_WORKER_THREADS: Lazy<usize> = Lazy::new(||
     // else, but that has not been needed in a long time.
     std::env::var("TOKIO_WORKER_THREADS")
         .map(|s| s.parse::<usize>().unwrap())
-        .unwrap_or_else(|_e| usize::max(1, num_cpus::get()))
+        .unwrap_or_else(|_e| usize::max(2, num_cpus::get()))
 });
 
 #[derive(Debug, Clone, Copy)]

pageserver/src/tenant.rs

@@ -17,6 +17,7 @@ use enumset::EnumSet;
 use futures::stream::FuturesUnordered;
 use futures::FutureExt;
 use futures::StreamExt;
+use pageserver_api::models::ShardParameters;
 use pageserver_api::models::TimelineState;
 use pageserver_api::shard::ShardIdentity;
 use pageserver_api::shard::TenantShardId;
@@ -33,6 +34,7 @@ use tracing::*;
 use utils::backoff;
 use utils::completion;
 use utils::crashsafe::path_with_suffix_extension;
+use utils::failpoint_support;
 use utils::fs_ext;
 use utils::sync::gate::Gate;
 use utils::sync::gate::GateGuard;
@@ -49,6 +51,7 @@ use self::metadata::TimelineMetadata;
 use self::mgr::GetActiveTenantError;
 use self::mgr::GetTenantError;
 use self::mgr::TenantsMap;
+use self::remote_timeline_client::upload::upload_index_part;
 use self::remote_timeline_client::RemoteTimelineClient;
 use self::timeline::uninit::TimelineExclusionError;
 use self::timeline::uninit::TimelineUninitMark;
@@ -594,10 +597,9 @@ impl Tenant {
         mode: SpawnMode,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Tenant>> {
-        // TODO(sharding): make WalRedoManager shard-aware
         let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
             conf,
-            tenant_shard_id.tenant_id,
+            tenant_shard_id,
         )));
 
         let TenantSharedResources {
@@ -890,7 +892,7 @@ impl Tenant {
     ) -> anyhow::Result<()> {
         span::debug_assert_current_span_has_tenant_id();
 
-        crate::failpoint_support::sleep_millis_async!("before-attaching-tenant");
+        failpoint_support::sleep_millis_async!("before-attaching-tenant");
 
         let preload = match preload {
             Some(p) => p,
@@ -1002,7 +1004,7 @@ impl Tenant {
         // IndexPart is the source of truth.
         self.clean_up_timelines(&existent_timelines)?;
 
-        crate::failpoint_support::sleep_millis_async!("attach-before-activate");
+        failpoint_support::sleep_millis_async!("attach-before-activate");
 
         info!("Done");
 
@@ -1144,10 +1146,9 @@ impl Tenant {
         tenant_shard_id: TenantShardId,
         reason: String,
     ) -> Arc<Tenant> {
-        // TODO(sharding): make WalRedoManager shard-aware
         let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
             conf,
-            tenant_shard_id.tenant_id,
+            tenant_shard_id,
         )));
         Arc::new(Tenant::new(
             TenantState::Broken {
@@ -1531,6 +1532,7 @@ impl Tenant {
             })?;
 
         if active_only && !timeline.is_active() {
+            tracing::warn!("Timeline {} is not active", timeline.timeline_id);
             Err(GetTimelineError::NotActive {
                 tenant_id: self.tenant_shard_id.tenant_id,
                 timeline_id,
@@ -1552,6 +1554,10 @@ impl Tenant {
             .collect()
     }
 
+    pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
+        self.timelines.lock().unwrap().keys().cloned().collect()
+    }
+
     /// This is used to create the initial 'main' timeline during bootstrapping,
     /// or when importing a new base backup. The caller is expected to load an
     /// initial image of the datadir to the new timeline after this.
@@ -2300,6 +2306,66 @@ impl Tenant {
     pub(crate) fn get_generation(&self) -> Generation {
         self.generation
     }
+
+    pub(crate) async fn split_prepare(
+        &self,
+        child_shards: &Vec<TenantShardId>,
+    ) -> anyhow::Result<()> {
+        let timelines = self.timelines.lock().unwrap().clone();
+        for timeline in timelines.values() {
+            let Some(tl_client) = &timeline.remote_client else {
+                anyhow::bail!("Remote storage is mandatory");
+            };
+
+            let Some(remote_storage) = &self.remote_storage else {
+                anyhow::bail!("Remote storage is mandatory");
+            };
+
+            // TODO: some higher level should enforce that timeline creation/deletion does not
+            // happen concurrently with splits.  This is impossible to safely coordinate locally
+            // within one single pageserver's view of the world.
+
+            // Upload an index from the parent: this is partly to provide freshness for the
+            // child tenants that will copy it, and partly for general ease-of-debugging: there will
+            // always be a parent shard index in the same generation as we wrote the child shard index.
+            tl_client.schedule_index_upload_for_file_changes()?;
+            tl_client.wait_completion().await?;
+
+            // Shut down the timeline's remote client: this means that the indices we write
+            // for child shards will not be invalidated by the parent shard deleting layers.
+            tl_client.shutdown().await?;
+
+            // Download methods can still be used after shutdown, as they don't flow through the remote client's
+            // queue.
+            // TODO: create a way for remote timeline client to give us a copy of the last IndexPart it uploaded
+            //       without having to download it again.
+            // TODO: carry a cancellation token in here
+            let result = tl_client
+                .download_index_file(CancellationToken::new())
+                .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))
+                .await?;
+            let index_part = match result {
+                MaybeDeletedIndexPart::Deleted(_) => {
+                    anyhow::bail!("Timeline deletion happened concurrently with split")
+                }
+                MaybeDeletedIndexPart::IndexPart(p) => p,
+            };
+
+            for child_shard in child_shards {
+                upload_index_part(
+                    remote_storage,
+                    child_shard,
+                    &timeline.timeline_id,
+                    self.generation,
+                    &index_part,
+                    &self.cancel,
+                )
+                .await?;
+            }
+        }
+
+        Ok(())
+    }
 }
 
 /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
@@ -2643,10 +2709,11 @@ impl Tenant {
                 }
             }
 
-            // Legacy configs are implicitly in attached state
+            // Legacy configs are implicitly in attached state, and do not support sharding
             Ok(LocationConf::attached_single(
                 tenant_conf,
                 Generation::none(),
+                &ShardParameters::default(),
             ))
         } else {
             // FIXME If the config file is not found, assume that we're attaching
@@ -2835,9 +2902,7 @@ impl Tenant {
             }
         };
 
-        crate::failpoint_support::sleep_millis_async!(
-            "gc_iteration_internal_after_getting_gc_timelines"
-        );
+        failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
 
         // If there is nothing to GC, we don't want any messages in the INFO log.
         if !gc_timelines.is_empty() {
@@ -3130,6 +3195,7 @@ impl Tenant {
 
     /// For unit tests, make this visible so that other modules can directly create timelines
     #[cfg(test)]
+    #[tracing::instrument(fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
     pub(crate) async fn bootstrap_timeline_test(
         &self,
         timeline_id: TimelineId,
@@ -3220,43 +3286,45 @@ impl Tenant {
 
             // Upload the created data dir to S3
             if let Some(storage) = &self.remote_storage {
-                let temp_path = timelines_path.join(format!(
-                    "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
-                ));
+                if self.tenant_shard_id().is_zero() {
+                    let temp_path = timelines_path.join(format!(
+                        "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
+                    ));
 
-                let (pgdata_zstd, tar_zst_size) =
-                    import_datadir::create_tar_zst(&pgdata_path, &temp_path).await?;
-                backoff::retry(
-                    || async {
-                        self::remote_timeline_client::upload_initdb_dir(
-                            storage,
-                            &self.tenant_shard_id.tenant_id,
-                            &timeline_id,
-                            pgdata_zstd.try_clone().await?,
-                            tar_zst_size,
-                            &self.cancel,
-                        )
+                    let (pgdata_zstd, tar_zst_size) =
+                        import_datadir::create_tar_zst(&pgdata_path, &temp_path).await?;
+                    backoff::retry(
+                        || async {
+                            self::remote_timeline_client::upload_initdb_dir(
+                                storage,
+                                &self.tenant_shard_id.tenant_id,
+                                &timeline_id,
+                                pgdata_zstd.try_clone().await?,
+                                tar_zst_size,
+                                &self.cancel,
+                            )
+                            .await
+                        },
+                        |_| false,
+                        3,
+                        u32::MAX,
+                        "persist_initdb_tar_zst",
+                        backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Cancelled")),
+                    )
+                    .await?;
+
+                    tokio::fs::remove_file(&temp_path)
                         .await
-                    },
-                    |_| false,
-                    3,
-                    u32::MAX,
-                    "persist_initdb_tar_zst",
-                    backoff::Cancel::new(self.cancel.clone(), || anyhow::anyhow!("Cancelled")),
-                )
-                .await?;
-
-                tokio::fs::remove_file(&temp_path)
-                    .await
-                    .or_else(|e| {
-                        if e.kind() == std::io::ErrorKind::NotFound {
-                            // If something else already removed the file, ignore the error
-                            Ok(())
-                        } else {
-                            Err(e)
-                        }
-                    })
-                    .with_context(|| format!("tempfile removal {temp_path}"))?;
+                        .or_else(|e| {
+                            if e.kind() == std::io::ErrorKind::NotFound {
+                                // If something else already removed the file, ignore the error
+                                Ok(())
+                            } else {
+                                Err(e)
+                            }
+                        })
+                        .with_context(|| format!("tempfile removal {temp_path}"))?;
+                }
             }
         }
         let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
@@ -3614,6 +3682,10 @@ impl Tenant {
 
         Ok(())
     }
+
+    pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
+        self.tenant_conf.read().unwrap().tenant_conf
+    }
 }
 
 fn remove_timeline_and_uninit_mark(
@@ -4055,6 +4127,7 @@ pub(crate) mod harness {
                 AttachedTenantConf::try_from(LocationConf::attached_single(
                     TenantConfOpt::from(self.tenant_conf),
                     self.generation,
+                    &ShardParameters::default(),
                 ))
                 .unwrap(),
                 // This is a legacy/test code path: sharding isn't supported here.

pageserver/src/tenant/config.rs

@@ -9,7 +9,7 @@
 //! may lead to a data loss.
 //!
 use anyhow::bail;
-use pageserver_api::models;
+use pageserver_api::models::{self, ShardParameters};
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
 use serde::de::IntoDeserializer;
 use serde::{Deserialize, Serialize};
@@ -46,6 +46,8 @@ pub mod defaults {
     pub const DEFAULT_WALRECEIVER_LAGGING_WAL_TIMEOUT: &str = "10 seconds";
     pub const DEFAULT_MAX_WALRECEIVER_LSN_WAL_LAG: u64 = 10 * 1024 * 1024;
     pub const DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD: &str = "24 hour";
+
+    pub const DEFAULT_INGEST_BATCH_SIZE: u64 = 100;
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -165,14 +167,17 @@ impl LocationConf {
     /// For use when loading from a legacy configuration: presence of a tenant
     /// implies it is in AttachmentMode::Single, which used to be the only
     /// possible state.  This function should eventually be removed.
-    pub(crate) fn attached_single(tenant_conf: TenantConfOpt, generation: Generation) -> Self {
+    pub(crate) fn attached_single(
+        tenant_conf: TenantConfOpt,
+        generation: Generation,
+        shard_params: &ShardParameters,
+    ) -> Self {
         Self {
             mode: LocationMode::Attached(AttachedLocationConfig {
                 generation,
                 attach_mode: AttachmentMode::Single,
             }),
-            // Legacy configuration loads are always from tenants created before sharding existed.
-            shard: ShardIdentity::unsharded(),
+            shard: ShardIdentity::from_params(ShardNumber(0), shard_params),
             tenant_conf,
         }
     }

pageserver/src/tenant/delete.rs

@@ -48,6 +48,9 @@ pub(crate) enum DeleteTenantError {
     #[error("Timeline {0}")]
     Timeline(#[from] DeleteTimelineError),
 
+    #[error("Cancelled")]
+    Cancelled,
+
     #[error(transparent)]
     Other(#[from] anyhow::Error),
 }

pageserver/src/tenant/mgr.rs

@@ -2,8 +2,10 @@
 //! page server.
 
 use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
+use itertools::Itertools;
 use pageserver_api::key::Key;
-use pageserver_api::shard::{ShardIdentity, ShardNumber, TenantShardId};
+use pageserver_api::models::ShardParameters;
+use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, TenantShardId};
 use rand::{distributions::Alphanumeric, Rng};
 use std::borrow::Cow;
 use std::collections::{BTreeMap, HashMap};
@@ -20,7 +22,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 
 use remote_storage::GenericRemoteStorage;
-use utils::crashsafe;
+use utils::{completion, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
@@ -130,7 +132,7 @@ impl TenantsMap {
 
     /// A page service client sends a TenantId, and to look up the correct Tenant we must
     /// resolve this to a fully qualified TenantShardId.
-    fn resolve_shard(
+    fn resolve_attached_shard(
         &self,
         tenant_id: &TenantId,
         selector: ShardSelector,
@@ -140,25 +142,27 @@ impl TenantsMap {
             TenantsMap::Initializing => None,
             TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
                 for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
+                    // Ignore all slots that don't contain an attached tenant
+                    let tenant = match &slot.1 {
+                        TenantSlot::Attached(t) => t,
+                        _ => continue,
+                    };
+
                     match selector {
                         ShardSelector::First => return Some(*slot.0),
                         ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
                             return Some(*slot.0)
                         }
                         ShardSelector::Page(key) => {
-                            if let Some(tenant) = slot.1.get_attached() {
-                                // First slot we see for this tenant, calculate the expected shard number
-                                // for the key: we will use this for checking if this and subsequent
-                                // slots contain the key, rather than recalculating the hash each time.
-                                if want_shard.is_none() {
-                                    want_shard = Some(tenant.shard_identity.get_shard_number(&key));
-                                }
+                            // First slot we see for this tenant, calculate the expected shard number
+                            // for the key: we will use this for checking if this and subsequent
+                            // slots contain the key, rather than recalculating the hash each time.
+                            if want_shard.is_none() {
+                                want_shard = Some(tenant.shard_identity.get_shard_number(&key));
+                            }
 
-                                if Some(tenant.shard_identity.number) == want_shard {
-                                    return Some(*slot.0);
-                                }
-                            } else {
-                                continue;
+                            if Some(tenant.shard_identity.number) == want_shard {
+                                return Some(*slot.0);
                             }
                         }
                         _ => continue,
@@ -514,10 +518,7 @@ pub async fn init_tenant_mgr(
             &ctx,
         ) {
             Ok(tenant) => {
-                tenants.insert(
-                    TenantShardId::unsharded(tenant.tenant_id()),
-                    TenantSlot::Attached(tenant),
-                );
+                tenants.insert(tenant_shard_id, TenantSlot::Attached(tenant));
             }
             Err(e) => {
                 error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
@@ -619,8 +620,6 @@ pub(crate) async fn shutdown_all_tenants() {
 }
 
 async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
-    use utils::completion;
-
     let mut join_set = JoinSet::new();
 
     // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
@@ -746,13 +745,21 @@ pub(crate) async fn create_tenant(
     conf: &'static PageServerConf,
     tenant_conf: TenantConfOpt,
     tenant_shard_id: TenantShardId,
+    shard_params: ShardParameters,
     generation: Generation,
     resources: TenantSharedResources,
     ctx: &RequestContext,
 ) -> Result<Arc<Tenant>, TenantMapInsertError> {
-    let location_conf = LocationConf::attached_single(tenant_conf, generation);
+    let location_conf = LocationConf::attached_single(tenant_conf, generation, &shard_params);
+
     info!("Creating tenant at location {location_conf:?}");
 
+    if shard_params.count != ShardCount(1) {
+        return Err(TenantMapInsertError::Other(anyhow::anyhow!(
+            "Only single-shard tenant creations may be serviced directly by a pageserver"
+        )));
+    }
+
     let slot_guard =
         tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
     let tenant_path = super::create_tenant_files(conf, &location_conf, &tenant_shard_id).await?;
@@ -787,6 +794,8 @@ pub(crate) enum SetNewTenantConfigError {
     GetTenant(#[from] GetTenantError),
     #[error(transparent)]
     Persist(anyhow::Error),
+    #[error(transparent)]
+    Other(anyhow::Error),
 }
 
 pub(crate) async fn set_new_tenant_config(
@@ -800,10 +809,21 @@ pub(crate) async fn set_new_tenant_config(
     info!("configuring tenant {tenant_id}");
     let tenant = get_tenant(tenant_shard_id, true)?;
 
+    if tenant.tenant_shard_id().shard_count > ShardCount(0) {
+        // Note that we use ShardParameters::default below.
+        return Err(SetNewTenantConfigError::Other(anyhow::anyhow!(
+            "This API may only be used on single-sharded tenants, use the /location_config API for sharded tenants"
+        )));
+    }
+
     // This is a legacy API that only operates on attached tenants: the preferred
     // API to use is the location_config/ endpoint, which lets the caller provide
     // the full LocationConf.
-    let location_conf = LocationConf::attached_single(new_tenant_conf, tenant.generation);
+    let location_conf = LocationConf::attached_single(
+        new_tenant_conf,
+        tenant.generation,
+        &ShardParameters::default(),
+    );
 
     Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf)
         .await
@@ -962,35 +982,27 @@ impl TenantManager {
         }
 
         let tenant_path = self.conf.tenant_path(&tenant_shard_id);
+        let timelines_path = self.conf.timelines_path(&tenant_shard_id);
+
+        // Directory structure is the same for attached and secondary modes:
+        // create it if it doesn't exist.  Timeline load/creation expects the
+        // timelines/ subdir to already exist.
+        //
+        // Does not need to be fsync'd because local storage is just a cache.
+        tokio::fs::create_dir_all(&timelines_path)
+            .await
+            .with_context(|| format!("Creating {timelines_path}"))?;
+
+        // Before activating either secondary or attached mode, persist the
+        // configuration, so that on restart we will re-attach (or re-start
+        // secondary) on the tenant.
+        Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
+            .await
+            .map_err(SetNewTenantConfigError::Persist)?;
 
         let new_slot = match &new_location_config.mode {
-            LocationMode::Secondary(_) => {
-                // Directory doesn't need to be fsync'd because if we crash it can
-                // safely be recreated next time this tenant location is configured.
-                tokio::fs::create_dir_all(&tenant_path)
-                    .await
-                    .with_context(|| format!("Creating {tenant_path}"))?;
-
-                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
-                    .await
-                    .map_err(SetNewTenantConfigError::Persist)?;
-
-                TenantSlot::Secondary
-            }
+            LocationMode::Secondary(_) => TenantSlot::Secondary,
             LocationMode::Attached(_attach_config) => {
-                let timelines_path = self.conf.timelines_path(&tenant_shard_id);
-
-                // Directory doesn't need to be fsync'd because we do not depend on
-                // it to exist after crashes: it may be recreated when tenant is
-                // re-attached, see https://github.com/neondatabase/neon/issues/5550
-                tokio::fs::create_dir_all(&tenant_path)
-                    .await
-                    .with_context(|| format!("Creating {timelines_path}"))?;
-
-                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
-                    .await
-                    .map_err(SetNewTenantConfigError::Persist)?;
-
                 let shard_identity = new_location_config.shard;
                 let tenant = tenant_spawn(
                     self.conf,
@@ -1102,6 +1114,228 @@ impl TenantManager {
                 .collect(),
         }
     }
+
+    pub(crate) async fn delete_tenant(
+        &self,
+        tenant_shard_id: TenantShardId,
+        activation_timeout: Duration,
+    ) -> Result<(), DeleteTenantError> {
+        // We acquire a SlotGuard during this function to protect against concurrent
+        // changes while the ::prepare phase of DeleteTenantFlow executes, but then
+        // have to return the Tenant to the map while the background deletion runs.
+        //
+        // TODO: refactor deletion to happen outside the lifetime of a Tenant.
+        // Currently, deletion requires a reference to the tenants map in order to
+        // keep the Tenant in the map until deletion is complete, and then remove
+        // it at the end.
+        //
+        // See https://github.com/neondatabase/neon/issues/5080
+
+        let slot_guard =
+            tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
+
+        // unwrap is safe because we used MustExist mode when acquiring
+        let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
+            TenantSlot::Attached(tenant) => tenant.clone(),
+            _ => {
+                // Express "not attached" as equivalent to "not found"
+                return Err(DeleteTenantError::NotAttached);
+            }
+        };
+
+        match tenant.current_state() {
+            TenantState::Broken { .. } | TenantState::Stopping { .. } => {
+                // If a tenant is broken or stopping, DeleteTenantFlow can
+                // handle it: broken tenants proceed to delete, stopping tenants
+                // are checked for deletion already in progress.
+            }
+            _ => {
+                tenant
+                    .wait_to_become_active(activation_timeout)
+                    .await
+                    .map_err(|e| match e {
+                        GetActiveTenantError::WillNotBecomeActive(_) => {
+                            DeleteTenantError::InvalidState(tenant.current_state())
+                        }
+                        GetActiveTenantError::Cancelled => DeleteTenantError::Cancelled,
+                        GetActiveTenantError::NotFound(_) => DeleteTenantError::NotAttached,
+                        GetActiveTenantError::WaitForActiveTimeout {
+                            latest_state: _latest_state,
+                            wait_time: _wait_time,
+                        } => DeleteTenantError::InvalidState(tenant.current_state()),
+                    })?;
+            }
+        }
+
+        let result = DeleteTenantFlow::run(
+            self.conf,
+            self.resources.remote_storage.clone(),
+            &TENANTS,
+            tenant,
+        )
+        .await;
+
+        // The Tenant goes back into the map in Stopping state, it will eventually be removed by DeleteTenantFLow
+        slot_guard.revert();
+        result
+    }
+
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), new_shard_count=%new_shard_count.0))]
+    pub(crate) async fn shard_split(
+        &self,
+        tenant_shard_id: TenantShardId,
+        new_shard_count: ShardCount,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Vec<TenantShardId>> {
+        let tenant = get_tenant(tenant_shard_id, true)?;
+
+        // Plan: identify what the new child shards will be
+        let effective_old_shard_count = std::cmp::max(tenant_shard_id.shard_count.0, 1);
+        if new_shard_count <= ShardCount(effective_old_shard_count) {
+            anyhow::bail!("Requested shard count is not an increase");
+        }
+        let expansion_factor = new_shard_count.0 / effective_old_shard_count;
+        if expansion_factor & (expansion_factor - 1) != 0 {
+            anyhow::bail!("Requested split is not a power of two");
+        }
+
+        // Key mapping is based on a round robin mapping of key hash modulo shard count,
+        // so our child shards are the ones which the same keys would map to.
+        let mut child_shards = Vec::new();
+        for shard_number in 0..ShardNumber(new_shard_count.0).0 {
+            if shard_number % effective_old_shard_count == tenant_shard_id.shard_number.0 {
+                child_shards.push(TenantShardId {
+                    tenant_id: tenant_shard_id.tenant_id,
+                    shard_number: ShardNumber(shard_number),
+                    shard_count: new_shard_count,
+                })
+            }
+        }
+
+        let parent_shard_identity = tenant.shard_identity;
+        let parent_tenant_conf = tenant.get_tenant_conf();
+        let parent_generation = tenant.generation;
+
+        // TODO: write a unit test for this
+        tracing::info!(
+            "Shard {} splits into: {}",
+            tenant_shard_id.to_index(),
+            child_shards
+                .iter()
+                .map(|id| format!("{}", id.to_index()))
+                .join(",")
+        );
+
+        // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
+        tenant.split_prepare(&child_shards).await?;
+
+        self.resources.deletion_queue_client.flush_advisory();
+
+        // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
+        drop(tenant);
+        let mut parent_slot_guard =
+            tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
+        let parent = match parent_slot_guard.get_old_value() {
+            Some(TenantSlot::Attached(t)) => t,
+            Some(TenantSlot::Secondary) => anyhow::bail!("Tenant location in secondary mode"),
+            Some(TenantSlot::InProgress(_)) => {
+                unreachable!()
+            }
+            None => {
+                // We don't actually need the parent shard to still be attached to do our work, but it's
+                // a weird enough situation that the caller probably didn't want us to continue working
+                // if they had detached the tenant they requested the split on.
+                anyhow::bail!("Detached parent shard in the middle of split!")
+            }
+        };
+
+        // TODO: hardlink layers from the parent into the child shard directories so that they don't immediately re-download
+        // TODO: erase the dentries from the parent
+
+        // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
+        // child shards to reach this point.
+        let mut target_lsns = HashMap::new();
+        for timeline in parent.timelines.lock().unwrap().clone().values() {
+            target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
+        }
+
+        // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
+        // and could slow down the children trying to catch up.
+
+        // Phase 3: Spawn the child shards
+        for child_shard in &child_shards {
+            let mut child_shard_identity = parent_shard_identity;
+            child_shard_identity.count = child_shard.shard_count;
+            child_shard_identity.number = child_shard.shard_number;
+
+            let child_location_conf = LocationConf {
+                mode: LocationMode::Attached(AttachedLocationConfig {
+                    generation: parent_generation,
+                    attach_mode: AttachmentMode::Single,
+                }),
+                shard: child_shard_identity,
+                tenant_conf: parent_tenant_conf,
+            };
+
+            self.upsert_location(*child_shard, child_location_conf, None, ctx)
+                .await?;
+        }
+
+        // Phase 4: wait for child chards WAL ingest to catch up to target LSN
+        for child_shard_id in &child_shards {
+            let child_shard = {
+                let locked = TENANTS.read().unwrap();
+                let peek_slot =
+                    tenant_map_peek_slot(&locked, child_shard_id, TenantSlotPeekMode::Read)?;
+                peek_slot.and_then(|s| s.get_attached()).cloned()
+            };
+            if let Some(t) = child_shard {
+                let timelines = t.timelines.lock().unwrap().clone();
+                for timeline in timelines.values() {
+                    let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
+                        continue;
+                    };
+
+                    tracing::info!(
+                        "Waiting for child shard {}/{} to reach target lsn {}...",
+                        child_shard_id,
+                        timeline.timeline_id,
+                        target_lsn
+                    );
+                    if let Err(e) = timeline.wait_lsn(*target_lsn, ctx).await {
+                        // Failure here might mean shutdown, in any case this part is an optimization
+                        // and we shouldn't hold up the split operation.
+                        tracing::warn!(
+                            "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
+                            timeline.timeline_id
+                        );
+                    } else {
+                        tracing::info!(
+                            "Child shard {}/{} reached target lsn {}",
+                            child_shard_id,
+                            timeline.timeline_id,
+                            target_lsn
+                        );
+                    }
+                }
+            }
+        }
+
+        // Phase 5: Shut down the parent shard.
+        let (_guard, progress) = completion::channel();
+        match parent.shutdown(progress, false).await {
+            Ok(()) => {}
+            Err(other) => {
+                other.wait().await;
+            }
+        }
+        parent_slot_guard.drop_old_value()?;
+
+        // Phase 6: Release the InProgress on the parent shard
+        drop(parent_slot_guard);
+
+        Ok(child_shards)
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -1203,9 +1437,11 @@ pub(crate) async fn get_active_tenant_with_timeout(
         let locked = TENANTS.read().unwrap();
 
         // Resolve TenantId to TenantShardId
-        let tenant_shard_id = locked.resolve_shard(&tenant_id, shard_selector).ok_or(
-            GetActiveTenantError::NotFound(GetTenantError::NotFound(tenant_id)),
-        )?;
+        let tenant_shard_id = locked
+            .resolve_attached_shard(&tenant_id, shard_selector)
+            .ok_or(GetActiveTenantError::NotFound(GetTenantError::NotFound(
+                tenant_id,
+            )))?;
 
         let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
             .map_err(GetTenantError::MapState)?;
@@ -1279,41 +1515,6 @@ pub(crate) async fn get_active_tenant_with_timeout(
     Ok(tenant)
 }
 
-pub(crate) async fn delete_tenant(
-    conf: &'static PageServerConf,
-    remote_storage: Option<GenericRemoteStorage>,
-    tenant_shard_id: TenantShardId,
-) -> Result<(), DeleteTenantError> {
-    // We acquire a SlotGuard during this function to protect against concurrent
-    // changes while the ::prepare phase of DeleteTenantFlow executes, but then
-    // have to return the Tenant to the map while the background deletion runs.
-    //
-    // TODO: refactor deletion to happen outside the lifetime of a Tenant.
-    // Currently, deletion requires a reference to the tenants map in order to
-    // keep the Tenant in the map until deletion is complete, and then remove
-    // it at the end.
-    //
-    // See https://github.com/neondatabase/neon/issues/5080
-
-    // TODO(sharding): make delete API sharding-aware
-    let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
-
-    // unwrap is safe because we used MustExist mode when acquiring
-    let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
-        TenantSlot::Attached(tenant) => tenant.clone(),
-        _ => {
-            // Express "not attached" as equivalent to "not found"
-            return Err(DeleteTenantError::NotAttached);
-        }
-    };
-
-    let result = DeleteTenantFlow::run(conf, remote_storage, &TENANTS, tenant).await;
-
-    // The Tenant goes back into the map in Stopping state, it will eventually be removed by DeleteTenantFLow
-    slot_guard.revert();
-    result
-}
-
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum DeleteTimelineError {
     #[error("Tenant {0}")]
@@ -1541,10 +1742,11 @@ pub(crate) async fn attach_tenant(
 ) -> Result<(), TenantMapInsertError> {
     // This is a legacy API (replaced by `/location_conf`).  It does not support sharding
     let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+    let shard_params = ShardParameters::default();
 
     let slot_guard =
         tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
-    let location_conf = LocationConf::attached_single(tenant_conf, generation);
+    let location_conf = LocationConf::attached_single(tenant_conf, generation, &shard_params);
     let tenant_dir = create_tenant_files(conf, &location_conf, &tenant_shard_id).await?;
     // TODO: tenant directory remains on disk if we bail out from here on.
     //       See https://github.com/neondatabase/neon/issues/4233
@@ -1987,8 +2189,6 @@ async fn remove_tenant_from_memory<V, F>(
 where
     F: std::future::Future<Output = anyhow::Result<V>>,
 {
-    use utils::completion;
-
     let mut slot_guard =
         tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -182,7 +182,7 @@
 
 pub(crate) mod download;
 pub mod index;
-mod upload;
+pub(crate) mod upload;
 
 use anyhow::Context;
 use camino::Utf8Path;
@@ -690,7 +690,10 @@ impl RemoteTimelineClient {
             .insert(layer.layer_desc().filename(), metadata.clone());
         upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
 
-        info!("scheduled layer file upload {layer}");
+        info!(
+            "scheduled layer file upload {layer} gen={:?} shard={:?}",
+            metadata.generation, metadata.shard
+        );
         let op = UploadOp::UploadLayer(layer, metadata);
         self.calls_unfinished_metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
@@ -818,8 +821,25 @@ impl RemoteTimelineClient {
     fn schedule_deletion_of_unlinked0(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        with_metadata: Vec<(LayerFileName, LayerFileMetadata)>,
+        mut with_metadata: Vec<(LayerFileName, LayerFileMetadata)>,
     ) {
+        // Filter out any layers which were not created by this tenant shard.  These are
+        // layers that originate from some ancestor shard after a split, and may still
+        // be referenced by other shards. We are free to delete them locally and remove
+        // them from our index (and would have already done so when we reach this point
+        // in the code), but we may not delete them remotely.
+        with_metadata.retain(|(name, meta)| {
+            let retain = meta.shard.shard_number == self.tenant_shard_id.shard_number
+                && meta.shard.shard_count == self.tenant_shard_id.shard_count;
+            if !retain {
+                tracing::debug!(
+                    "Skipping deletion of ancestor-shard layer {name}, from shard {}",
+                    meta.shard
+                );
+            }
+            retain
+        });
+
         for (name, meta) in &with_metadata {
             info!(
                 "scheduling deletion of layer {}{} (shard {})",
@@ -2192,15 +2212,6 @@ mod tests {
 
         let index_part_bytes = serde_json::to_vec(&example_index_part).unwrap();
 
-        let timeline_path = test_state.harness.timeline_path(&TIMELINE_ID);
-        let remote_timeline_dir = test_state.harness.remote_fs_dir.join(
-            timeline_path
-                .strip_prefix(&test_state.harness.conf.workdir)
-                .unwrap(),
-        );
-
-        std::fs::create_dir_all(remote_timeline_dir).expect("creating test dir should work");
-
         let index_path = test_state.harness.remote_fs_dir.join(
             remote_index_path(
                 &test_state.harness.tenant_shard_id,
@@ -2209,6 +2220,10 @@ mod tests {
             )
             .get_path(),
         );
+
+        std::fs::create_dir_all(index_path.parent().unwrap())
+            .expect("creating test dir should work");
+
         eprintln!("Writing {index_path}");
         std::fs::write(&index_path, index_part_bytes).unwrap();
         example_index_part

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -25,7 +25,7 @@ use super::index::LayerFileMetadata;
 use tracing::info;
 
 /// Serializes and uploads the given index part data to the remote storage.
-pub(super) async fn upload_index_part<'a>(
+pub(crate) async fn upload_index_part<'a>(
     storage: &'a GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -23,7 +23,7 @@ use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn, vec_map::VecMap};
 // while being able to use std::fmt::Write's methods
 use std::fmt::Write as _;
 use std::ops::Range;
-use tokio::sync::RwLock;
+use tokio::sync::{RwLock, RwLockWriteGuard};
 
 use super::{DeltaLayerWriter, ResidentLayer};
 
@@ -246,16 +246,43 @@ impl InMemoryLayer {
 
     /// Common subroutine of the public put_wal_record() and put_page_image() functions.
     /// Adds the page version to the in-memory tree
-    pub async fn put_value(
+    pub(crate) async fn put_value(
         &self,
         key: Key,
         lsn: Lsn,
         val: &Value,
         ctx: &RequestContext,
     ) -> Result<()> {
-        trace!("put_value key {} at {}/{}", key, self.timeline_id, lsn);
-        let inner: &mut _ = &mut *self.inner.write().await;
+        let mut inner = self.inner.write().await;
         self.assert_writable();
+        self.put_value_locked(&mut inner, key, lsn, val, ctx).await
+    }
+
+    pub(crate) async fn put_values(
+        &self,
+        values: &HashMap<Key, Vec<(Lsn, Value)>>,
+        ctx: &RequestContext,
+    ) -> Result<()> {
+        let mut inner = self.inner.write().await;
+        self.assert_writable();
+        for (key, vals) in values {
+            for (lsn, val) in vals {
+                self.put_value_locked(&mut inner, *key, *lsn, val, ctx)
+                    .await?;
+            }
+        }
+        Ok(())
+    }
+
+    async fn put_value_locked(
+        &self,
+        locked_inner: &mut RwLockWriteGuard<'_, InMemoryLayerInner>,
+        key: Key,
+        lsn: Lsn,
+        val: &Value,
+        ctx: &RequestContext,
+    ) -> Result<()> {
+        trace!("put_value key {} at {}/{}", key, self.timeline_id, lsn);
 
         let off = {
             // Avoid doing allocations for "small" values.
@@ -264,7 +291,7 @@ impl InMemoryLayer {
             let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
             buf.clear();
             val.ser_into(&mut buf)?;
-            inner
+            locked_inner
                 .file
                 .write_blob(
                     &buf,
@@ -275,7 +302,7 @@ impl InMemoryLayer {
                 .await?
         };
 
-        let vec_map = inner.index.entry(key).or_default();
+        let vec_map = locked_inner.index.entry(key).or_default();
         let old = vec_map.append_or_update_last(lsn, off).unwrap().0;
         if old.is_some() {
             // We already had an entry for this LSN. That's odd..
@@ -285,13 +312,11 @@ impl InMemoryLayer {
         Ok(())
     }
 
-    pub async fn put_tombstone(&self, _key_range: Range<Key>, _lsn: Lsn) -> Result<()> {
+    pub(crate) async fn put_tombstones(&self, _key_ranges: &[(Range<Key>, Lsn)]) -> Result<()> {
         // TODO: Currently, we just leak the storage for any deleted keys
-
         Ok(())
     }
 
-    /// Make the layer non-writeable. Only call once.
     /// Records the end_lsn for non-dropped layers.
     /// `end_lsn` is exclusive
     pub async fn freeze(&self, end_lsn: Lsn) {

pageserver/src/tenant/storage_layer/layer.rs

@@ -259,8 +259,9 @@ impl Layer {
 
         layer
             .get_value_reconstruct_data(key, lsn_range, reconstruct_data, &self.0, ctx)
-            .instrument(tracing::info_span!("get_value_reconstruct_data", layer=%self))
+            .instrument(tracing::debug_span!("get_value_reconstruct_data", layer=%self))
             .await
+            .with_context(|| format!("get_value_reconstruct_data for layer {self}"))
     }
 
     /// Download the layer if evicted.
@@ -654,7 +655,6 @@ impl LayerInner {
     }
 
     /// Cancellation safe.
-    #[tracing::instrument(skip_all, fields(layer=%self))]
     async fn get_or_maybe_download(
         self: &Arc<Self>,
         allow_download: bool,
@@ -663,95 +663,101 @@ impl LayerInner {
         let mut init_permit = None;
 
         loop {
-            let download = move |permit| async move {
-                // disable any scheduled but not yet running eviction deletions for this
-                let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
+            let download = move |permit| {
+                async move {
+                    // disable any scheduled but not yet running eviction deletions for this
+                    let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
 
-                // count cancellations, which currently remain largely unexpected
-                let init_cancelled =
-                    scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
+                    // count cancellations, which currently remain largely unexpected
+                    let init_cancelled =
+                        scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
 
-                // no need to make the evict_and_wait wait for the actual download to complete
-                drop(self.status.send(Status::Downloaded));
+                    // no need to make the evict_and_wait wait for the actual download to complete
+                    drop(self.status.send(Status::Downloaded));
 
-                let timeline = self
-                    .timeline
-                    .upgrade()
-                    .ok_or_else(|| DownloadError::TimelineShutdown)?;
+                    let timeline = self
+                        .timeline
+                        .upgrade()
+                        .ok_or_else(|| DownloadError::TimelineShutdown)?;
 
-                // FIXME: grab a gate
+                    // FIXME: grab a gate
 
-                let can_ever_evict = timeline.remote_client.as_ref().is_some();
+                    let can_ever_evict = timeline.remote_client.as_ref().is_some();
 
-                // check if we really need to be downloaded; could have been already downloaded by a
-                // cancelled previous attempt.
-                let needs_download = self
-                    .needs_download()
-                    .await
-                    .map_err(DownloadError::PreStatFailed)?;
+                    // check if we really need to be downloaded; could have been already downloaded by a
+                    // cancelled previous attempt.
+                    let needs_download = self
+                        .needs_download()
+                        .await
+                        .map_err(DownloadError::PreStatFailed)?;
 
-                let permit = if let Some(reason) = needs_download {
-                    if let NeedsDownload::NotFile(ft) = reason {
-                        return Err(DownloadError::NotFile(ft));
+                    let permit = if let Some(reason) = needs_download {
+                        if let NeedsDownload::NotFile(ft) = reason {
+                            return Err(DownloadError::NotFile(ft));
+                        }
+
+                        // only reset this after we've decided we really need to download. otherwise it'd
+                        // be impossible to mark cancelled downloads for eviction, like one could imagine
+                        // we would like to do for prefetching which was not needed.
+                        self.wanted_evicted.store(false, Ordering::Release);
+
+                        if !can_ever_evict {
+                            return Err(DownloadError::NoRemoteStorage);
+                        }
+
+                        if let Some(ctx) = ctx {
+                            self.check_expected_download(ctx)?;
+                        }
+
+                        if !allow_download {
+                            // this does look weird, but for LayerInner the "downloading" means also changing
+                            // internal once related state ...
+                            return Err(DownloadError::DownloadRequired);
+                        }
+
+                        tracing::info!(%reason, "downloading on-demand");
+
+                        self.spawn_download_and_wait(timeline, permit).await?
+                    } else {
+                        // the file is present locally, probably by a previous but cancelled call to
+                        // get_or_maybe_download. alternatively we might be running without remote storage.
+                        LAYER_IMPL_METRICS.inc_init_needed_no_download();
+
+                        permit
+                    };
+
+                    let since_last_eviction =
+                        self.last_evicted_at.lock().unwrap().map(|ts| ts.elapsed());
+                    if let Some(since_last_eviction) = since_last_eviction {
+                        // FIXME: this will not always be recorded correctly until #6028 (the no
+                        // download needed branch above)
+                        LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
                     }
 
-                    // only reset this after we've decided we really need to download. otherwise it'd
-                    // be impossible to mark cancelled downloads for eviction, like one could imagine
-                    // we would like to do for prefetching which was not needed.
-                    self.wanted_evicted.store(false, Ordering::Release);
+                    let res = Arc::new(DownloadedLayer {
+                        owner: Arc::downgrade(self),
+                        kind: tokio::sync::OnceCell::default(),
+                        version: next_version,
+                    });
 
-                    if !can_ever_evict {
-                        return Err(DownloadError::NoRemoteStorage);
+                    self.access_stats.record_residence_event(
+                        LayerResidenceStatus::Resident,
+                        LayerResidenceEventReason::ResidenceChange,
+                    );
+
+                    let waiters = self.inner.initializer_count();
+                    if waiters > 0 {
+                        tracing::info!(
+                            waiters,
+                            "completing the on-demand download for other tasks"
+                        );
                     }
 
-                    if let Some(ctx) = ctx {
-                        self.check_expected_download(ctx)?;
-                    }
+                    scopeguard::ScopeGuard::into_inner(init_cancelled);
 
-                    if !allow_download {
-                        // this does look weird, but for LayerInner the "downloading" means also changing
-                        // internal once related state ...
-                        return Err(DownloadError::DownloadRequired);
-                    }
-
-                    tracing::info!(%reason, "downloading on-demand");
-
-                    self.spawn_download_and_wait(timeline, permit).await?
-                } else {
-                    // the file is present locally, probably by a previous but cancelled call to
-                    // get_or_maybe_download. alternatively we might be running without remote storage.
-                    LAYER_IMPL_METRICS.inc_init_needed_no_download();
-
-                    permit
-                };
-
-                let since_last_eviction =
-                    self.last_evicted_at.lock().unwrap().map(|ts| ts.elapsed());
-                if let Some(since_last_eviction) = since_last_eviction {
-                    // FIXME: this will not always be recorded correctly until #6028 (the no
-                    // download needed branch above)
-                    LAYER_IMPL_METRICS.record_redownloaded_after(since_last_eviction);
+                    Ok((ResidentOrWantedEvicted::Resident(res), permit))
                 }
-
-                let res = Arc::new(DownloadedLayer {
-                    owner: Arc::downgrade(self),
-                    kind: tokio::sync::OnceCell::default(),
-                    version: next_version,
-                });
-
-                self.access_stats.record_residence_event(
-                    LayerResidenceStatus::Resident,
-                    LayerResidenceEventReason::ResidenceChange,
-                );
-
-                let waiters = self.inner.initializer_count();
-                if waiters > 0 {
-                    tracing::info!(waiters, "completing the on-demand download for other tasks");
-                }
-
-                scopeguard::ScopeGuard::into_inner(init_cancelled);
-
-                Ok((ResidentOrWantedEvicted::Resident(res), permit))
+                .instrument(tracing::info_span!("get_or_maybe_download", layer=%self))
             };
 
             if let Some(init_permit) = init_permit.take() {
@@ -872,6 +878,23 @@ impl LayerInner {
                         Ok(())
                     }
                     Err(e) => {
+                        let consecutive_failures =
+                            this.consecutive_failures.fetch_add(1, Ordering::Relaxed);
+
+                        let backoff = utils::backoff::exponential_backoff_duration_seconds(
+                            consecutive_failures.min(u32::MAX as usize) as u32,
+                            1.5,
+                            60.0,
+                        );
+
+                        let backoff = std::time::Duration::from_secs_f64(backoff);
+
+                        tokio::select! {
+                            _ = tokio::time::sleep(backoff) => {},
+                            _ = crate::task_mgr::shutdown_token().cancelled_owned() => {},
+                            _ = timeline.cancel.cancelled() => {},
+                        };
+
                         Err(e)
                     }
                 };
@@ -920,21 +943,9 @@ impl LayerInner {
                 Ok(permit)
             }
             Ok((Err(e), _permit)) => {
-                // FIXME: this should be with the spawned task and be cancellation sensitive
-                //
-                // while we should not need this, this backoff has turned out to be useful with
-                // a bug of unexpectedly deleted remote layer file (#5787).
-                let consecutive_failures =
-                    self.consecutive_failures.fetch_add(1, Ordering::Relaxed);
+                // sleep already happened in the spawned task, if it was not cancelled
+                let consecutive_failures = self.consecutive_failures.load(Ordering::Relaxed);
                 tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
-                let backoff = utils::backoff::exponential_backoff_duration_seconds(
-                    consecutive_failures.min(u32::MAX as usize) as u32,
-                    1.5,
-                    60.0,
-                );
-                let backoff = std::time::Duration::from_secs_f64(backoff);
-
-                tokio::time::sleep(backoff).await;
                 Err(DownloadError::DownloadFailed)
             }
             Err(_gone) => Err(DownloadError::DownloadCancelled),

pageserver/src/tenant/timeline.rs

@@ -496,6 +496,11 @@ impl Timeline {
             return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
         }
 
+        // This check is debug-only because of the cost of hashing, and because it's a double-check: we
+        // already checked the key against the shard_identity when looking up the Timeline from
+        // page_service.
+        debug_assert!(!self.shard_identity.is_key_disposable(&key));
+
         // XXX: structured stats collection for layer eviction here.
         trace!(
             "get page request for {}@{} from task kind {:?}",
@@ -1459,6 +1464,7 @@ impl Timeline {
                 max_lsn_wal_lag,
                 auth_token: crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
                 availability_zone: self.conf.availability_zone.clone(),
+                ingest_batch_size: self.conf.ingest_batch_size,
             },
             broker_client,
             ctx,
@@ -2223,13 +2229,13 @@ impl Timeline {
                     return Err(layer_traversal_error(
                         if cfg!(test) {
                             format!(
-                                "could not find data for key {} at LSN {}, for request at LSN {}\n{}",
-                                key, cont_lsn, request_lsn, std::backtrace::Backtrace::force_capture(),
+                                "could not find data for key {} (shard {:?}) at LSN {}, for request at LSN {}\n{}",
+                                key, self.shard_identity.get_shard_number(&key), cont_lsn, request_lsn, std::backtrace::Backtrace::force_capture(),
                             )
                         } else {
                             format!(
-                                "could not find data for key {} at LSN {}, for request at LSN {}",
-                                key, cont_lsn, request_lsn
+                                "could not find data for key {} (shard {:?}) at LSN {}, for request at LSN {}",
+                                key, self.shard_identity.get_shard_number(&key), cont_lsn, request_lsn
                             )
                         },
                         traversal_path,
@@ -2471,9 +2477,27 @@ impl Timeline {
         Ok(())
     }
 
-    async fn put_tombstone(&self, key_range: Range<Key>, lsn: Lsn) -> anyhow::Result<()> {
-        let layer = self.get_layer_for_write(lsn).await?;
-        layer.put_tombstone(key_range, lsn).await?;
+    async fn put_values(
+        &self,
+        values: &HashMap<Key, Vec<(Lsn, Value)>>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        // Pick the first LSN in the batch to get the layer to write to.
+        for lsns in values.values() {
+            if let Some((lsn, _)) = lsns.first() {
+                let layer = self.get_layer_for_write(*lsn).await?;
+                layer.put_values(values, ctx).await?;
+                break;
+            }
+        }
+        Ok(())
+    }
+
+    async fn put_tombstones(&self, tombstones: &[(Range<Key>, Lsn)]) -> anyhow::Result<()> {
+        if let Some((_, lsn)) = tombstones.first() {
+            let layer = self.get_layer_for_write(*lsn).await?;
+            layer.put_tombstones(tombstones).await?;
+        }
         Ok(())
     }
 
@@ -3035,6 +3059,15 @@ impl Timeline {
                 for range in &partition.ranges {
                     let mut key = range.start;
                     while key < range.end {
+                        if self.shard_identity.is_key_disposable(&key) {
+                            debug!(
+                                "Dropping key {} during compaction (it belongs on shard {:?})",
+                                key,
+                                self.shard_identity.get_shard_number(&key)
+                            );
+                            key = key.next();
+                            continue;
+                        }
                         let img = match self.get(key, lsn, ctx).await {
                             Ok(img) => img,
                             Err(err) => {
@@ -3061,6 +3094,7 @@ impl Timeline {
                                 }
                             }
                         };
+
                         image_layer_writer.put_image(key, &img).await?;
                         key = key.next();
                     }
@@ -3631,7 +3665,15 @@ impl Timeline {
                 )))
             });
 
-            writer.as_mut().unwrap().put_value(key, lsn, value).await?;
+            if !self.shard_identity.is_key_disposable(&key) {
+                writer.as_mut().unwrap().put_value(key, lsn, value).await?;
+            } else {
+                debug!(
+                    "Dropping key {} during compaction (it belongs on shard {:?})",
+                    key,
+                    self.shard_identity.get_shard_number(&key)
+                );
+            }
 
             if !new_layers.is_empty() {
                 fail_point!("after-timeline-compacted-first-L1");
@@ -4529,8 +4571,16 @@ impl<'a> TimelineWriter<'a> {
         self.tl.put_value(key, lsn, value, ctx).await
     }
 
-    pub async fn delete(&self, key_range: Range<Key>, lsn: Lsn) -> anyhow::Result<()> {
-        self.tl.put_tombstone(key_range, lsn).await
+    pub(crate) async fn put_batch(
+        &self,
+        batch: &HashMap<Key, Vec<(Lsn, Value)>>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        self.tl.put_values(batch, ctx).await
+    }
+
+    pub(crate) async fn delete_batch(&self, batch: &[(Range<Key>, Lsn)]) -> anyhow::Result<()> {
+        self.tl.put_tombstones(batch).await
     }
 
     /// Track the end of the latest digested WAL record.
@@ -4541,11 +4591,11 @@ impl<'a> TimelineWriter<'a> {
     /// 'lsn' must be aligned. This wakes up any wait_lsn() callers waiting for
     /// the 'lsn' or anything older. The previous last record LSN is stored alongside
     /// the latest and can be read.
-    pub fn finish_write(&self, new_lsn: Lsn) {
+    pub(crate) fn finish_write(&self, new_lsn: Lsn) {
         self.tl.finish_write(new_lsn);
     }
 
-    pub fn update_current_logical_size(&self, delta: i64) {
+    pub(crate) fn update_current_logical_size(&self, delta: i64) {
         self.tl.update_current_logical_size(delta)
     }
 }

pageserver/src/tenant/timeline/walreceiver.rs

@@ -58,6 +58,7 @@ pub struct WalReceiverConf {
     pub max_lsn_wal_lag: NonZeroU64,
     pub auth_token: Option<Arc<String>>,
     pub availability_zone: Option<String>,
+    pub ingest_batch_size: u64,
 }
 
 pub struct WalReceiver {

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

@@ -138,7 +138,7 @@ pub(super) async fn connection_manager_loop_step(
                     Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
                     Err(status) => {
                         match status.code() {
-                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") => {
+                            Code::Unknown if status.message().contains("stream closed because of a broken pipe") || status.message().contains("connection reset") => {
                                 // tonic's error handling doesn't provide a clear code for disconnections: we get
                                 // "h2 protocol error: error reading a body from connection: stream closed because of a broken pipe"
                                 info!("broker disconnected: {status}");
@@ -411,6 +411,7 @@ impl ConnectionManagerState {
 
         let node_id = new_sk.safekeeper_id;
         let connect_timeout = self.conf.wal_connect_timeout;
+        let ingest_batch_size = self.conf.ingest_batch_size;
         let timeline = Arc::clone(&self.timeline);
         let ctx = ctx.detached_child(
             TaskKind::WalReceiverConnectionHandler,
@@ -430,6 +431,7 @@ impl ConnectionManagerState {
                     connect_timeout,
                     ctx,
                     node_id,
+                    ingest_batch_size,
                 )
                 .await;
 
@@ -1345,6 +1347,7 @@ mod tests {
                 max_lsn_wal_lag: NonZeroU64::new(1024 * 1024).unwrap(),
                 auth_token: None,
                 availability_zone: None,
+                ingest_batch_size: 1,
             },
             wal_connection: None,
             wal_stream_candidates: HashMap::new(),

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

@@ -26,7 +26,7 @@ use tracing::{debug, error, info, trace, warn, Instrument};
 use super::TaskStateUpdate;
 use crate::{
     context::RequestContext,
-    metrics::{LIVE_CONNECTIONS_COUNT, WALRECEIVER_STARTED_CONNECTIONS},
+    metrics::{LIVE_CONNECTIONS_COUNT, WALRECEIVER_STARTED_CONNECTIONS, WAL_INGEST},
     task_mgr,
     task_mgr::TaskKind,
     task_mgr::WALRECEIVER_RUNTIME,
@@ -106,6 +106,7 @@ impl From<WalDecodeError> for WalReceiverError {
 
 /// Open a connection to the given safekeeper and receive WAL, sending back progress
 /// messages as we go.
+#[allow(clippy::too_many_arguments)]
 pub(super) async fn handle_walreceiver_connection(
     timeline: Arc<Timeline>,
     wal_source_connconf: PgConnectionConfig,
@@ -114,6 +115,7 @@ pub(super) async fn handle_walreceiver_connection(
     connect_timeout: Duration,
     ctx: RequestContext,
     node: NodeId,
+    ingest_batch_size: u64,
 ) -> Result<(), WalReceiverError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
@@ -305,7 +307,9 @@ pub(super) async fn handle_walreceiver_connection(
 
                 {
                     let mut decoded = DecodedWALRecord::default();
-                    let mut modification = timeline.begin_modification(endlsn);
+                    let mut modification = timeline.begin_modification(startlsn);
+                    let mut uncommitted_records = 0;
+                    let mut filtered_records = 0;
                     while let Some((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
@@ -314,14 +318,40 @@ pub(super) async fn handle_walreceiver_connection(
                             return Err(WalReceiverError::Other(anyhow!("LSN not aligned")));
                         }
 
-                        walingest
+                        // Ingest the records without immediately committing them.
+                        let ingested = walingest
                             .ingest_record(recdata, lsn, &mut modification, &mut decoded, &ctx)
                             .await
                             .with_context(|| format!("could not ingest record at {lsn}"))?;
+                        if !ingested {
+                            tracing::debug!("ingest: filtered out record @ LSN {lsn}");
+                            WAL_INGEST.records_filtered.inc();
+                            filtered_records += 1;
+                        }
 
                         fail_point!("walreceiver-after-ingest");
 
                         last_rec_lsn = lsn;
+
+                        // Commit every ingest_batch_size records. Even if we filtered out
+                        // all records, we still need to call commit to advance the LSN.
+                        uncommitted_records += 1;
+                        if uncommitted_records >= ingest_batch_size {
+                            WAL_INGEST
+                                .records_committed
+                                .inc_by(uncommitted_records - filtered_records);
+                            modification.commit(&ctx).await?;
+                            uncommitted_records = 0;
+                            filtered_records = 0;
+                        }
+                    }
+
+                    // Commit the remaining records.
+                    if uncommitted_records > 0 {
+                        WAL_INGEST
+                            .records_committed
+                            .inc_by(uncommitted_records - filtered_records);
+                        modification.commit(&ctx).await?;
                     }
                 }