clippy

Otherwise an assertion in CONCURRENT_BACKGROUND_TASKS will
trip if you try to run the pageserver on a single core.

.config/nextest.toml

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

.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

@@ -1161,6 +1161,7 @@ dependencies = [
  "flate2",
  "futures",
  "hyper",
+ "nix 0.26.2",
  "notify",
  "num_cpus",
  "opentelemetry",
@@ -1168,8 +1169,10 @@ dependencies = [
  "regex",
  "remote_storage",
  "reqwest",
+ "rust-ini",
  "serde",
  "serde_json",
+ "signal-hook",
  "tar",
  "tokio",
  "tokio-postgres",
@@ -1201,6 +1204,26 @@ 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 +1456,12 @@ 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 +1604,15 @@ 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 +2144,20 @@ 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"
@@ -3029,6 +3081,16 @@ 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"
@@ -3057,6 +3119,28 @@ 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"
@@ -4180,6 +4264,16 @@ 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"
@@ -4311,12 +4405,14 @@ dependencies = [
  "async-stream",
  "aws-config",
  "aws-sdk-s3",
+ "aws-smithy-async",
  "bincode",
  "bytes",
  "chrono",
  "clap",
  "crc32c",
  "either",
+ "futures",
  "futures-util",
  "hex",
  "histogram",
@@ -4355,6 +4451,7 @@ dependencies = [
  "clap",
  "const_format",
  "crc32c",
+ "fail",
  "fs2",
  "futures",
  "git-version",
@@ -5134,6 +5231,15 @@ 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"
@@ -5777,6 +5883,7 @@ dependencies = [
  "chrono",
  "const_format",
  "criterion",
+ "fail",
  "futures",
  "heapless",
  "hex",
@@ -6301,6 +6408,7 @@ dependencies = [
  "futures-io",
  "futures-sink",
  "futures-util",
+ "getrandom 0.2.11",
  "hex",
  "hmac",
  "hyper",
@@ -6312,6 +6420,7 @@ dependencies = [
  "num-bigint",
  "num-integer",
  "num-traits",
+ "once_cell",
  "prost",
  "rand 0.8.5",
  "regex",

Cargo.toml

@@ -6,6 +6,7 @@ members = [
     "pageserver",
     "pageserver/ctl",
     "pageserver/client",
+    "pageserver/pagebench",
     "proxy",
     "safekeeper",
     "storage_broker",
@@ -79,6 +80,7 @@ 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"

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      
 
 #########################################################################################
 #

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
 

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"] }
@@ -39,3 +41,4 @@ remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
 vm_monitor = { version = "0.1", path = "../libs/vm_monitor/" }
 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

@@ -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/src/endpoint.rs

@@ -46,6 +46,8 @@ 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 utils::id::{NodeId, TenantId, TimelineId};
 
@@ -439,11 +441,14 @@ 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(())
     }
@@ -537,6 +542,7 @@ impl Endpoint {
             safekeeper_connstrings,
             storage_auth_token: auth_token.clone(),
             remote_extensions,
+            pgbouncer_settings: None,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -732,10 +738,15 @@ 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 '{}'",

deny.toml

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

libs/compute_api/src/spec.rs

@@ -73,6 +73,8 @@ pub struct ComputeSpec {
 
     // information about available remote extensions
     pub remote_extensions: Option<RemoteExtSpec>,
+
+    pub pgbouncer_settings: Option<HashMap<String, String>>,
 }
 
 /// 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

@@ -557,19 +557,6 @@ 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>,

libs/pageserver_api/src/shard.rs

@@ -81,6 +81,10 @@ 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)
+    }
 }
 
 /// Formatting helper
@@ -159,7 +163,7 @@ 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,
@@ -418,6 +422,21 @@ 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/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/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/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/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/src/mgmt_api.rs

@@ -5,6 +5,8 @@ use utils::{
     id::{TenantId, TimelineId},
 };
 
+pub mod util;
+
 #[derive(Debug)]
 pub struct Client {
     mgmt_api_endpoint: String,

pageserver/client/src/mgmt_api/util.rs

@@ -0,0 +1,49 @@
+//! Helpers to do common higher-level tasks with the [`Client`].
+
+use std::sync::Arc;
+
+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(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

@@ -115,15 +115,8 @@ impl PagestreamClient {
 
     pub async fn getpage(
         &mut self,
-        key: RelTagBlockNo,
-        lsn: Lsn,
+        req: PagestreamGetPageRequest,
     ) -> anyhow::Result<PagestreamGetPageResponse> {
-        let req = PagestreamGetPageRequest {
-            latest: false,
-            rel: key.rel_tag,
-            blkno: key.block_no,
-            lsn,
-        };
         let req = PagestreamFeMessage::GetPage(req);
         let req: bytes::Bytes = req.serialize();
         // let mut req = tokio_util::io::ReaderStream::new(&req);

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

@@ -25,6 +25,7 @@ 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};
@@ -66,9 +67,6 @@ 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.
@@ -308,6 +306,7 @@ 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,
         }
     }
 }
@@ -886,7 +885,9 @@ 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()
@@ -1290,34 +1291,6 @@ 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>,
@@ -1566,19 +1539,22 @@ 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) {
@@ -1589,7 +1565,7 @@ 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,
     };
 
@@ -1604,7 +1580,11 @@ 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;
 

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()),
             }

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

@@ -33,6 +33,7 @@ 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;
@@ -594,10 +595,9 @@ 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 +890,7 @@ 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 +1002,7 @@ 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 +1144,9 @@ 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 {
@@ -2839,9 +2838,7 @@ 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() {
@@ -3134,6 +3131,7 @@ 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,

pageserver/src/tenant/config.rs

@@ -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)]

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

@@ -130,7 +130,7 @@ 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 +140,27 @@ 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 +516,7 @@ 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:#}");
@@ -962,35 +961,27 @@ 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 +1093,71 @@ 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
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -1203,9 +1259,11 @@ 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 +1337,6 @@ 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}")]

pageserver/src/tenant/remote_timeline_client.rs

@@ -818,8 +818,25 @@ 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 +2209,6 @@ 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 +2217,10 @@ 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/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

@@ -878,6 +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)
                     }
                 };
@@ -926,21 +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

@@ -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?;
                     }
                 }
 

pageserver/src/walrecord.rs

@@ -21,10 +21,13 @@ pub enum NeonWalRecord {
     /// Native PostgreSQL WAL record
     Postgres { will_init: bool, rec: Bytes },
 
-    /// Clear bits in heap visibility map. ('flags' is bitmap of bits to clear)
+    /// Clear the bits specified in `flags` in the heap visibility map for the given heap blocks.
+    ///
+    /// For example, for `{ heap_blkno_1: None, heap_blkno_2: Some(23), flags: 0b0010_0000}`
+    /// redo will apply `&=0b0010_0000` on heap block 23's visibility map bitmask.
     ClearVisibilityMapFlags {
-        new_heap_blkno: Option<u32>,
-        old_heap_blkno: Option<u32>,
+        heap_blkno_1: Option<u32>,
+        heap_blkno_2: Option<u32>,
         flags: u8,
     },
     /// Mark transaction IDs as committed on a CLOG page

pageserver/src/walredo.rs

@@ -22,6 +22,7 @@ use anyhow::Context;
 use byteorder::{ByteOrder, LittleEndian};
 use bytes::{BufMut, Bytes, BytesMut};
 use nix::poll::*;
+use pageserver_api::shard::TenantShardId;
 use serde::Serialize;
 use std::collections::VecDeque;
 use std::io;
@@ -35,14 +36,11 @@ use std::sync::{Arc, Mutex, MutexGuard, RwLock};
 use std::time::Duration;
 use std::time::Instant;
 use tracing::*;
-use utils::{bin_ser::BeSer, id::TenantId, lsn::Lsn, nonblock::set_nonblock};
+use utils::{bin_ser::BeSer, lsn::Lsn, nonblock::set_nonblock};
 
 #[cfg(feature = "testing")]
 use std::sync::atomic::{AtomicUsize, Ordering};
 
-#[cfg(feature = "testing")]
-use pageserver_api::shard::TenantShardId;
-
 use crate::config::PageServerConf;
 use crate::metrics::{
     WalRedoKillCause, WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_PROCESS_COUNTERS,
@@ -92,7 +90,7 @@ struct ProcessOutput {
 /// records.
 ///
 pub struct PostgresRedoManager {
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     conf: &'static PageServerConf,
     last_redo_at: std::sync::Mutex<Option<Instant>>,
     redo_process: RwLock<Option<Arc<WalRedoProcess>>>,
@@ -186,10 +184,13 @@ impl PostgresRedoManager {
     ///
     /// Create a new PostgresRedoManager.
     ///
-    pub fn new(conf: &'static PageServerConf, tenant_id: TenantId) -> PostgresRedoManager {
+    pub fn new(
+        conf: &'static PageServerConf,
+        tenant_shard_id: TenantShardId,
+    ) -> PostgresRedoManager {
         // The actual process is launched lazily, on first request.
         PostgresRedoManager {
-            tenant_id,
+            tenant_shard_id,
             conf,
             last_redo_at: std::sync::Mutex::default(),
             redo_process: RwLock::new(None),
@@ -244,8 +245,12 @@ impl PostgresRedoManager {
                                 let timer =
                                     WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM.start_timer();
                                 let proc = Arc::new(
-                                    WalRedoProcess::launch(self.conf, self.tenant_id, pg_version)
-                                        .context("launch walredo process")?,
+                                    WalRedoProcess::launch(
+                                        self.conf,
+                                        self.tenant_shard_id,
+                                        pg_version,
+                                    )
+                                    .context("launch walredo process")?,
                                 );
                                 timer.observe_duration();
                                 *proc_guard = Some(Arc::clone(&proc));
@@ -398,8 +403,8 @@ impl PostgresRedoManager {
                 anyhow::bail!("tried to pass postgres wal record to neon WAL redo");
             }
             NeonWalRecord::ClearVisibilityMapFlags {
-                new_heap_blkno,
-                old_heap_blkno,
+                heap_blkno_1,
+                heap_blkno_2,
                 flags,
             } => {
                 // sanity check that this is modifying the correct relation
@@ -409,7 +414,8 @@ impl PostgresRedoManager {
                     "ClearVisibilityMapFlags record on unexpected rel {}",
                     rel
                 );
-                if let Some(heap_blkno) = *new_heap_blkno {
+                for heap_blkno in [heap_blkno_1, heap_blkno_2].into_iter().flatten() {
+                    let heap_blkno = *heap_blkno;
                     // Calculate the VM block and offset that corresponds to the heap block.
                     let map_block = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blkno);
                     let map_byte = pg_constants::HEAPBLK_TO_MAPBYTE(heap_blkno);
@@ -423,19 +429,6 @@ impl PostgresRedoManager {
 
                     map[map_byte as usize] &= !(flags << map_offset);
                 }
-
-                // Repeat for 'old_heap_blkno', if any
-                if let Some(heap_blkno) = *old_heap_blkno {
-                    let map_block = pg_constants::HEAPBLK_TO_MAPBLOCK(heap_blkno);
-                    let map_byte = pg_constants::HEAPBLK_TO_MAPBYTE(heap_blkno);
-                    let map_offset = pg_constants::HEAPBLK_TO_OFFSET(heap_blkno);
-
-                    assert!(map_block == blknum);
-
-                    let map = &mut page[pg_constants::MAXALIGN_SIZE_OF_PAGE_HEADER_DATA..];
-
-                    map[map_byte as usize] &= !(flags << map_offset);
-                }
             }
             // Non-relational WAL records are handled here, with custom code that has the
             // same effects as the corresponding Postgres WAL redo function.
@@ -638,7 +631,7 @@ impl<C: CommandExt> CloseFileDescriptors for C {
 struct WalRedoProcess {
     #[allow(dead_code)]
     conf: &'static PageServerConf,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     // Some() on construction, only becomes None on Drop.
     child: Option<NoLeakChild>,
     stdout: Mutex<ProcessOutput>,
@@ -652,10 +645,10 @@ impl WalRedoProcess {
     //
     // Start postgres binary in special WAL redo mode.
     //
-    #[instrument(skip_all,fields(tenant_id=%tenant_id, pg_version=pg_version))]
+    #[instrument(skip_all,fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), pg_version=pg_version))]
     fn launch(
         conf: &'static PageServerConf,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         pg_version: u32,
     ) -> anyhow::Result<Self> {
         let pg_bin_dir_path = conf.pg_bin_dir(pg_version).context("pg_bin_dir")?; // TODO these should be infallible.
@@ -680,7 +673,7 @@ impl WalRedoProcess {
             // as close-on-exec by default, but that's not enough, since we use
             // libraries that directly call libc open without setting that flag.
             .close_fds()
-            .spawn_no_leak_child(tenant_id)
+            .spawn_no_leak_child(tenant_shard_id)
             .context("spawn process")?;
         WAL_REDO_PROCESS_COUNTERS.started.inc();
         let mut child = scopeguard::guard(child, |child| {
@@ -741,12 +734,12 @@ impl WalRedoProcess {
                         error!(error=?e, "failed to read from walredo stderr");
                     }
                 }
-            }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_id, %pg_version))
+            }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %pg_version))
         );
 
         Ok(Self {
             conf,
-            tenant_id,
+            tenant_shard_id,
             child: Some(child),
             stdin: Mutex::new(ProcessInput {
                 stdin,
@@ -772,7 +765,7 @@ impl WalRedoProcess {
     // Apply given WAL records ('records') over an old page image. Returns
     // new page image.
     //
-    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%self.id()))]
+    #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), pid=%self.id()))]
     fn apply_wal_records(
         &self,
         tag: BufferTag,
@@ -966,11 +959,7 @@ impl WalRedoProcess {
         // these files will be collected to an allure report
         let filename = format!("walredo-{millis}-{}-{seq}.walredo", writebuf.len());
 
-        // TODO(sharding): update this call when WalRedoProcess gets a TenantShardId.
-        let path = self
-            .conf
-            .tenant_path(&TenantShardId::unsharded(self.tenant_id))
-            .join(&filename);
+        let path = self.conf.tenant_path(&self.tenant_shard_id).join(&filename);
 
         let res = std::fs::OpenOptions::new()
             .write(true)
@@ -1004,7 +993,7 @@ impl Drop for WalRedoProcess {
 /// Wrapper type around `std::process::Child` which guarantees that the child
 /// will be killed and waited-for by this process before being dropped.
 struct NoLeakChild {
-    tenant_id: TenantId,
+    tenant_id: TenantShardId,
     child: Option<Child>,
 }
 
@@ -1023,7 +1012,7 @@ impl DerefMut for NoLeakChild {
 }
 
 impl NoLeakChild {
-    fn spawn(tenant_id: TenantId, command: &mut Command) -> io::Result<Self> {
+    fn spawn(tenant_id: TenantShardId, command: &mut Command) -> io::Result<Self> {
         let child = command.spawn()?;
         Ok(NoLeakChild {
             tenant_id,
@@ -1078,7 +1067,7 @@ impl Drop for NoLeakChild {
             Some(child) => child,
             None => return,
         };
-        let tenant_id = self.tenant_id;
+        let tenant_shard_id = self.tenant_id;
         // Offload the kill+wait of the child process into the background.
         // If someone stops the runtime, we'll leak the child process.
         // We can ignore that case because we only stop the runtime on pageserver exit.
@@ -1086,7 +1075,11 @@ impl Drop for NoLeakChild {
             tokio::task::spawn_blocking(move || {
                 // Intentionally don't inherit the tracing context from whoever is dropping us.
                 // This thread here is going to outlive of our dropper.
-                let span = tracing::info_span!("walredo", %tenant_id);
+                let span = tracing::info_span!(
+                    "walredo",
+                    tenant_id = %tenant_shard_id.tenant_id,
+                    shard_id = %tenant_shard_id.shard_slug()
+                );
                 let _entered = span.enter();
                 Self::kill_and_wait_impl(child, WalRedoKillCause::NoLeakChildDrop);
             })
@@ -1096,11 +1089,11 @@ impl Drop for NoLeakChild {
 }
 
 trait NoLeakChildCommandExt {
-    fn spawn_no_leak_child(&mut self, tenant_id: TenantId) -> io::Result<NoLeakChild>;
+    fn spawn_no_leak_child(&mut self, tenant_id: TenantShardId) -> io::Result<NoLeakChild>;
 }
 
 impl NoLeakChildCommandExt for Command {
-    fn spawn_no_leak_child(&mut self, tenant_id: TenantId) -> io::Result<NoLeakChild> {
+    fn spawn_no_leak_child(&mut self, tenant_id: TenantShardId) -> io::Result<NoLeakChild> {
         NoLeakChild::spawn(tenant_id, self)
     }
 }
@@ -1155,6 +1148,7 @@ mod tests {
     use crate::repository::Key;
     use crate::{config::PageServerConf, walrecord::NeonWalRecord};
     use bytes::Bytes;
+    use pageserver_api::shard::TenantShardId;
     use std::str::FromStr;
     use utils::{id::TenantId, lsn::Lsn};
 
@@ -1264,9 +1258,9 @@ mod tests {
             let repo_dir = camino_tempfile::tempdir()?;
             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);
 
             Ok(RedoHarness {
                 _repo_dir: repo_dir,

pgxn/neon/Makefile

@@ -9,6 +9,7 @@ OBJS = \
 	libpagestore.o \
 	neon.o \
 	neon_utils.o \
+	neon_walreader.o \
 	pagestore_smgr.o \
 	relsize_cache.o \
 	walproposer.o \

pgxn/neon/libpagestore.c

@@ -35,7 +35,8 @@
 
 #define PageStoreTrace DEBUG5
 
-#define RECONNECT_INTERVAL_USEC 1000000
+#define MIN_RECONNECT_INTERVAL_USEC 1000
+#define MAX_RECONNECT_INTERVAL_USEC 1000000
 
 bool		connected = false;
 PGconn	   *pageserver_conn = NULL;
@@ -133,6 +134,11 @@ pageserver_connect(int elevel)
 	const char *values[3];
 	int			n;
 
+	static TimestampTz last_connect_time = 0;
+	static uint64_t delay_us = MIN_RECONNECT_INTERVAL_USEC;
+	TimestampTz now;
+        uint64_t us_since_last_connect;
+
 	Assert(!connected);
 
 	if (CheckConnstringUpdated())
@@ -140,6 +146,22 @@ pageserver_connect(int elevel)
 		ReloadConnstring();
 	}
 
+	now = GetCurrentTimestamp();
+        us_since_last_connect = now - last_connect_time;
+	if (us_since_last_connect < delay_us)
+	{
+		pg_usleep(delay_us - us_since_last_connect);
+		delay_us *= 2;
+		if (delay_us > MAX_RECONNECT_INTERVAL_USEC)
+			delay_us = MAX_RECONNECT_INTERVAL_USEC;
+		last_connect_time = GetCurrentTimestamp();
+	}
+	else
+	{
+		delay_us = MIN_RECONNECT_INTERVAL_USEC;
+		last_connect_time = now;
+	}
+
 	/*
 	 * Connect using the connection string we got from the
 	 * neon.pageserver_connstring GUC. If the NEON_AUTH_TOKEN environment
@@ -333,7 +355,6 @@ pageserver_send(NeonRequest *request)
 		{
 			HandleMainLoopInterrupts();
 			n_reconnect_attempts += 1;
-			pg_usleep(RECONNECT_INTERVAL_USEC);
 		}
 		n_reconnect_attempts = 0;
 	}

pgxn/neon/libpqwalproposer.h

@@ -0,0 +1,96 @@
+/*
+ * Interface to set of libpq wrappers walproposer and neon_walreader need.
+ * Similar to libpqwalreceiver, but it has blocking connection establishment and
+ * pqexec which don't fit us. Implementation is at walproposer_pg.c.
+ */
+#ifndef ___LIBPQWALPROPOSER_H__
+#define ___LIBPQWALPROPOSER_H__
+
+/* Re-exported and modified ExecStatusType */
+typedef enum
+{
+	/* We received a single CopyBoth result */
+	WP_EXEC_SUCCESS_COPYBOTH,
+
+	/*
+	 * Any success result other than a single CopyBoth was received. The
+	 * specifics of the result were already logged, but it may be useful to
+	 * provide an error message indicating which safekeeper messed up.
+	 *
+	 * Do not expect PQerrorMessage to be appropriately set.
+	 */
+	WP_EXEC_UNEXPECTED_SUCCESS,
+
+	/*
+	 * No result available at this time. Wait until read-ready, then call
+	 * again. Internally, this is returned when PQisBusy indicates that
+	 * PQgetResult would block.
+	 */
+	WP_EXEC_NEEDS_INPUT,
+	/* Catch-all failure. Check PQerrorMessage. */
+	WP_EXEC_FAILED,
+} WalProposerExecStatusType;
+
+/* Possible return values from walprop_async_read */
+typedef enum
+{
+	/* The full read was successful. buf now points to the data */
+	PG_ASYNC_READ_SUCCESS,
+
+	/*
+	 * The read is ongoing. Wait until the connection is read-ready, then try
+	 * again.
+	 */
+	PG_ASYNC_READ_TRY_AGAIN,
+	/* Reading failed. Check PQerrorMessage(conn) */
+	PG_ASYNC_READ_FAIL,
+} PGAsyncReadResult;
+
+/* Possible return values from walprop_async_write */
+typedef enum
+{
+	/* The write fully completed */
+	PG_ASYNC_WRITE_SUCCESS,
+
+	/*
+	 * The write started, but you'll need to call PQflush some more times to
+	 * finish it off. We just tried, so it's best to wait until the connection
+	 * is read- or write-ready to try again.
+	 *
+	 * If it becomes read-ready, call PQconsumeInput and flush again. If it
+	 * becomes write-ready, just call PQflush.
+	 */
+	PG_ASYNC_WRITE_TRY_FLUSH,
+	/* Writing failed. Check PQerrorMessage(conn) */
+	PG_ASYNC_WRITE_FAIL,
+} PGAsyncWriteResult;
+
+/*
+ * This header is included by walproposer.h to define walproposer_api; if we're
+ * building walproposer without pg, ignore libpq part, leaving only interface
+ * types.
+ */
+#ifndef WALPROPOSER_LIB
+
+#include "libpq-fe.h"
+
+/*
+ * Sometimes working directly with underlying PGconn is simpler, export the
+ * whole thing for simplicity.
+ */
+typedef struct WalProposerConn
+{
+	PGconn	   *pg_conn;
+	bool		is_nonblocking; /* whether the connection is non-blocking */
+	char	   *recvbuf;		/* last received CopyData message from
+								 * walprop_async_read */
+} WalProposerConn;
+
+extern WalProposerConn *libpqwp_connect_start(char *conninfo);
+extern bool libpqwp_send_query(WalProposerConn *conn, char *query);
+extern WalProposerExecStatusType libpqwp_get_query_result(WalProposerConn *conn);
+extern PGAsyncReadResult libpqwp_async_read(WalProposerConn *conn, char **buf, int *amount);
+extern void libpqwp_disconnect(WalProposerConn *conn);
+
+#endif							/* WALPROPOSER_LIB */
+#endif							/* ___LIBPQWALPROPOSER_H__ */

pgxn/neon/neon_walreader.c

@@ -0,0 +1,742 @@
+/*
+ * Like WALRead, but when WAL segment doesn't exist locally instead of throwing
+ * ERROR asynchronously tries to fetch it from the most advanced safekeeper.
+ *
+ * We can't use libpqwalreceiver as it blocks during connection establishment
+ * (and waiting for PQExec result), so use libpqwalproposer instead.
+ *
+ * TODO: keepalives are currently never sent, so the other side can close the
+ * connection prematurely.
+ *
+ * TODO: close conn if reading takes too long to prevent stuck connections.
+ */
+#include "postgres.h"
+
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "access/xlog_internal.h"
+#include "access/xlogdefs.h"
+#include "access/xlogreader.h"
+#include "libpq/pqformat.h"
+#include "storage/fd.h"
+#include "utils/wait_event.h"
+
+#include "libpq-fe.h"
+
+#include "neon_walreader.h"
+#include "walproposer.h"
+
+#define NEON_WALREADER_ERR_MSG_LEN 512
+
+/*
+ * Can be called where NeonWALReader *state is available in the context, adds log_prefix.
+ */
+#define nwr_log(elevel, fmt, ...) elog(elevel, "%s" fmt, state->log_prefix, ## __VA_ARGS__)
+
+static NeonWALReadResult NeonWALReadRemote(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+static NeonWALReadResult NeonWALReaderReadMsg(NeonWALReader *state);
+static void NeonWALReaderResetRemote(NeonWALReader *state);
+static bool NeonWALReadLocal(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+static bool neon_wal_segment_open(NeonWALReader *state, XLogSegNo nextSegNo, TimeLineID *tli_p);
+static void neon_wal_segment_close(NeonWALReader *state);
+static bool is_wal_segment_exists(XLogSegNo segno, int segsize,
+								  TimeLineID tli);
+
+/*
+ * State of connection to donor safekeeper.
+ */
+typedef enum
+{
+	/* no remote connection */
+	RS_NONE,
+	/* doing PQconnectPoll, need readable socket */
+	RS_CONNECTING_READ,
+	/* doing PQconnectPoll, need writable socket */
+	RS_CONNECTING_WRITE,
+	/* Waiting for START_REPLICATION result */
+	RS_WAIT_EXEC_RESULT,
+	/* replication stream established */
+	RS_ESTABLISHED,
+} NeonWALReaderRemoteState;
+
+struct NeonWALReader
+{
+	/*
+	 * LSN before which we assume WAL is not available locally. Exists because
+	 * though first segment after startup always exists, part before
+	 * basebackup LSN is filled with zeros.
+	 */
+	XLogRecPtr	available_lsn;
+	WALSegmentContext segcxt;
+	WALOpenSegment seg;
+	int			wre_errno;
+	/* Explains failure to read, static for simplicity. */
+	char		err_msg[NEON_WALREADER_ERR_MSG_LEN];
+
+	/*
+	 * Saved info about request in progress, used to check validity of
+	 * arguments after resume and remember how far we accomplished it. req_lsn
+	 * is 0 if there is no request in progress.
+	 */
+	XLogRecPtr	req_lsn;
+	Size		req_len;
+	Size		req_progress;
+	WalProposer *wp;			/* we learn donor through walproposer */
+	char		donor_name[64]; /* saved donor safekeeper name for logging */
+	/* state of connection to safekeeper */
+	NeonWALReaderRemoteState rem_state;
+	WalProposerConn *wp_conn;
+
+	/*
+	 * position in wp_conn recvbuf from which we'll copy WAL next time, or
+	 * NULL if there is no unprocessed message
+	 */
+	char	   *wal_ptr;
+	Size		wal_rem_len;	/* how many unprocessed bytes left in recvbuf */
+
+	/*
+	 * LSN of wal_ptr position according to walsender to cross check against
+	 * read request
+	 */
+	XLogRecPtr	rem_lsn;
+
+	/* prepended to lines logged by neon_walreader, if provided */
+	char		log_prefix[64];
+};
+
+/* palloc and initialize NeonWALReader */
+NeonWALReader *
+NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, WalProposer *wp, char *log_prefix)
+{
+	NeonWALReader *reader;
+
+	reader = (NeonWALReader *)
+		palloc_extended(sizeof(NeonWALReader),
+						MCXT_ALLOC_NO_OOM | MCXT_ALLOC_ZERO);
+	if (!reader)
+		return NULL;
+
+	reader->available_lsn = available_lsn;
+	reader->seg.ws_file = -1;
+	reader->seg.ws_segno = 0;
+	reader->seg.ws_tli = 0;
+	reader->segcxt.ws_segsize = wal_segment_size;
+
+	reader->wp = wp;
+
+	reader->rem_state = RS_NONE;
+
+	if (log_prefix)
+		strlcpy(reader->log_prefix, log_prefix, sizeof(reader->log_prefix));
+
+	return reader;
+}
+
+void
+NeonWALReaderFree(NeonWALReader *state)
+{
+	if (state->seg.ws_file != -1)
+		neon_wal_segment_close(state);
+	if (state->wp_conn)
+		libpqwp_disconnect(state->wp_conn);
+	pfree(state);
+}
+
+/*
+ * Like vanilla WALRead, but if requested position is before available_lsn or
+ * WAL segment doesn't exist on disk, it tries to fetch needed segment from the
+ * advanced safekeeper.
+ *
+ * Read 'count' bytes into 'buf', starting at location 'startptr', from WAL
+ * fetched from timeline 'tli'.
+ *
+ * Returns NEON_WALREAD_SUCCESS if succeeded, NEON_WALREAD_ERROR if an error
+ * occurs, in which case 'err' has the desciption. Error always closes remote
+ * connection, if there was any, so socket subscription should be removed.
+ *
+ * NEON_WALREAD_WOULDBLOCK means caller should obtain socket to wait for with
+ * NeonWALReaderSocket and call NeonWALRead again with exactly the same
+ * arguments when NeonWALReaderEvents happen on the socket. Note that per libpq
+ * docs during connection establishment (before first successful read) socket
+ * underneath might change.
+ *
+ * Also, eventually walreader should switch from remote to local read; caller
+ * should remove subscription to socket then by checking NeonWALReaderEvents
+ * after successful read (otherwise next read might reopen the connection with
+ * different socket).
+ *
+ * Reading not monotonically is not supported and will result in error.
+ *
+ * Caller should be sure that WAL up to requested LSN exists, otherwise
+ * NEON_WALREAD_WOULDBLOCK might be always returned.
+ */
+NeonWALReadResult
+NeonWALRead(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	/*
+	 * If requested data is before known available basebackup lsn or there is
+	 * already active remote state, do remote read.
+	 */
+	if (startptr < state->available_lsn || state->rem_state != RS_NONE)
+	{
+		return NeonWALReadRemote(state, buf, startptr, count, tli);
+	}
+	if (NeonWALReadLocal(state, buf, startptr, count, tli))
+	{
+		return NEON_WALREAD_SUCCESS;
+	}
+	else if (state->wre_errno == ENOENT)
+	{
+		nwr_log(LOG, "local read failed as segment at %X/%X doesn't exist, attempting remote",
+				LSN_FORMAT_ARGS(startptr));
+		return NeonWALReadRemote(state, buf, startptr, count, tli);
+	}
+	else
+	{
+		return NEON_WALREAD_ERROR;
+	}
+}
+
+/* Do the read from remote safekeeper. */
+static NeonWALReadResult
+NeonWALReadRemote(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	if (state->rem_state == RS_NONE)
+	{
+		XLogRecPtr	donor_lsn;
+
+		/* no connection yet; start one */
+		Safekeeper *donor = GetDonor(state->wp, &donor_lsn);
+
+		if (donor == NULL)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "failed to establish remote connection to fetch WAL: no donor available");
+			return NEON_WALREAD_ERROR;
+		}
+		snprintf(state->donor_name, sizeof(state->donor_name), "%s:%s", donor->host, donor->port);
+		nwr_log(LOG, "establishing connection to %s, flush_lsn %X/%X to fetch WAL",
+				state->donor_name, LSN_FORMAT_ARGS(donor_lsn));
+		state->wp_conn = libpqwp_connect_start(donor->conninfo);
+		if (PQstatus(state->wp_conn->pg_conn) == CONNECTION_BAD)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "failed to connect to %s to fetch WAL: immediately failed with %s",
+					 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+		/* we'll poll immediately */
+		state->rem_state = RS_CONNECTING_READ;
+	}
+
+	if (state->rem_state == RS_CONNECTING_READ || state->rem_state == RS_CONNECTING_WRITE)
+	{
+		switch (PQconnectPoll(state->wp_conn->pg_conn))
+		{
+			case PGRES_POLLING_FAILED:
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "failed to connect to %s to fetch WAL: poll error: %s",
+						 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+			case PGRES_POLLING_READING:
+				state->rem_state = RS_CONNECTING_READ;
+				return NEON_WALREAD_WOULDBLOCK;
+			case PGRES_POLLING_WRITING:
+				state->rem_state = RS_CONNECTING_WRITE;
+				return NEON_WALREAD_WOULDBLOCK;
+			case PGRES_POLLING_OK:
+				{
+					/* connection successfully established */
+					char		start_repl_query[128];
+
+					snprintf(start_repl_query, sizeof(start_repl_query),
+							 "START_REPLICATION PHYSICAL %X/%X (term='" UINT64_FORMAT "')",
+							 LSN_FORMAT_ARGS(startptr), state->wp->propTerm);
+					nwr_log(LOG, "connection to %s to fetch WAL succeeded, running %s",
+							state->donor_name, start_repl_query);
+					if (!libpqwp_send_query(state->wp_conn, start_repl_query))
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "failed to send %s query to %s: %s",
+								 start_repl_query, state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+						NeonWALReaderResetRemote(state);
+						return NEON_WALREAD_ERROR;
+					}
+					state->rem_state = RS_WAIT_EXEC_RESULT;
+					break;
+				}
+
+			default:			/* there is unused PGRES_POLLING_ACTIVE */
+				Assert(false);
+				return NEON_WALREAD_ERROR;	/* keep the compiler quiet */
+		}
+	}
+
+	if (state->rem_state == RS_WAIT_EXEC_RESULT)
+	{
+		switch (libpqwp_get_query_result(state->wp_conn))
+		{
+			case WP_EXEC_SUCCESS_COPYBOTH:
+				state->rem_state = RS_ESTABLISHED;
+				break;
+			case WP_EXEC_NEEDS_INPUT:
+				return NEON_WALREAD_WOULDBLOCK;
+			case WP_EXEC_FAILED:
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "get START_REPLICATION result from %s failed: %s",
+						 state->donor_name, PQerrorMessage(state->wp_conn->pg_conn));
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+			default:			/* can't happen */
+				snprintf(state->err_msg, sizeof(state->err_msg),
+						 "get START_REPLICATION result from %s: unexpected result",
+						 state->donor_name);
+				NeonWALReaderResetRemote(state);
+				return NEON_WALREAD_ERROR;
+		}
+	}
+
+	Assert(state->rem_state == RS_ESTABLISHED);
+
+	/*
+	 * If we had the request before, verify args are the same and advance the
+	 * result ptr according to the progress; otherwise register the request.
+	 */
+	if (state->req_lsn != InvalidXLogRecPtr)
+	{
+		if (state->req_lsn != startptr || state->req_len != count)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "args changed during request, was %X/%X %zu, now %X/%X %zu",
+					 LSN_FORMAT_ARGS(state->req_lsn), state->req_len, LSN_FORMAT_ARGS(startptr), count);
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+		nwr_log(DEBUG5, "continuing remote read at req_lsn=%X/%X len=%zu, req_progress=%zu",
+				LSN_FORMAT_ARGS(startptr),
+				count,
+				state->req_progress);
+		buf += state->req_progress;
+	}
+	else
+	{
+		state->req_lsn = startptr;
+		state->req_len = count;
+		state->req_progress = 0;
+		nwr_log(DEBUG5, "starting remote read req_lsn=%X/%X len=%zu",
+				LSN_FORMAT_ARGS(startptr),
+				count);
+	}
+
+	while (true)
+	{
+		Size		to_copy;
+
+		/*
+		 * If we have no ready data, receive new message.
+		 */
+		if (state->wal_rem_len == 0 &&
+
+		/*
+		 * check for the sake of 0 length reads; walproposer does these for
+		 * heartbeats, though generally they shouldn't hit remote source.
+		 */
+			state->req_len - state->req_progress > 0)
+		{
+			NeonWALReadResult read_msg_res = NeonWALReaderReadMsg(state);
+
+			if (read_msg_res != NEON_WALREAD_SUCCESS)
+				return read_msg_res;
+		}
+
+		if (state->req_lsn + state->req_progress != state->rem_lsn)
+		{
+			snprintf(state->err_msg, sizeof(state->err_msg),
+					 "expected remote WAL at %X/%X but got %X/%X. Non monotonic read requests could have caused this. req_lsn=%X/%X len=%zu",
+					 LSN_FORMAT_ARGS(state->req_lsn + state->req_progress),
+					 LSN_FORMAT_ARGS(state->rem_lsn),
+					 LSN_FORMAT_ARGS(state->req_lsn),
+					 state->req_len);
+			NeonWALReaderResetRemote(state);
+			return NEON_WALREAD_ERROR;
+		}
+
+		/* We can copy min of (available, requested) bytes. */
+		to_copy =
+			Min(state->req_len - state->req_progress, state->wal_rem_len);
+		memcpy(buf, state->wal_ptr, to_copy);
+		state->wal_ptr += to_copy;
+		state->wal_rem_len -= to_copy;
+		state->rem_lsn += to_copy;
+		if (state->wal_rem_len == 0)
+			state->wal_ptr = NULL;	/* freed by libpqwalproposer */
+		buf += to_copy;
+		state->req_progress += to_copy;
+		if (state->req_progress == state->req_len)
+		{
+			XLogSegNo	next_segno;
+			XLogSegNo	req_segno;
+
+			XLByteToSeg(state->req_lsn, req_segno, state->segcxt.ws_segsize);
+			XLByteToSeg(state->rem_lsn, next_segno, state->segcxt.ws_segsize);
+
+			/*
+			 * Request completed. If there is a chance of serving next one
+			 * locally, close the connection.
+			 */
+			if (state->req_lsn < state->available_lsn &&
+				state->rem_lsn >= state->available_lsn)
+			{
+				nwr_log(LOG, "closing remote connection as available_lsn %X/%X crossed and next read at %X/%X is likely to be served locally",
+						LSN_FORMAT_ARGS(state->available_lsn), LSN_FORMAT_ARGS(state->rem_lsn));
+				NeonWALReaderResetRemote(state);
+			}
+			else if (state->rem_lsn >= state->available_lsn && next_segno > req_segno &&
+					 is_wal_segment_exists(next_segno, state->segcxt.ws_segsize, tli))
+			{
+				nwr_log(LOG, "closing remote connection as WAL file at next lsn %X/%X exists",
+						LSN_FORMAT_ARGS(state->rem_lsn));
+				NeonWALReaderResetRemote(state);
+			}
+			state->req_lsn = InvalidXLogRecPtr;
+			state->req_len = 0;
+			state->req_progress = 0;
+			return NEON_WALREAD_SUCCESS;
+		}
+	}
+}
+
+/*
+ * Read one WAL message from the stream, sets state->wal_ptr in case of success.
+ * Resets remote state in case of failure.
+ */
+static NeonWALReadResult
+NeonWALReaderReadMsg(NeonWALReader *state)
+{
+	while (true)				/* loop until we get 'w' */
+	{
+		char	   *copydata_ptr;
+		int			copydata_size;
+		StringInfoData s;
+		char		msg_type;
+		int			hdrlen;
+
+		Assert(state->rem_state == RS_ESTABLISHED);
+		Assert(state->wal_ptr == NULL && state->wal_rem_len == 0);
+
+		switch (libpqwp_async_read(state->wp_conn,
+								   &copydata_ptr,
+								   &copydata_size))
+		{
+			case PG_ASYNC_READ_SUCCESS:
+				break;
+			case PG_ASYNC_READ_TRY_AGAIN:
+				return NEON_WALREAD_WOULDBLOCK;
+			case PG_ASYNC_READ_FAIL:
+				snprintf(state->err_msg,
+						 sizeof(state->err_msg),
+						 "req_lsn=%X/%X, req_len=%zu, req_progress=%zu, get copydata failed: %s",
+						 LSN_FORMAT_ARGS(state->req_lsn),
+						 state->req_len,
+						 state->req_progress,
+						 PQerrorMessage(state->wp_conn->pg_conn));
+				goto err;
+		}
+
+		/* put data on StringInfo to parse */
+		s.data = copydata_ptr;
+		s.len = copydata_size;
+		s.cursor = 0;
+		s.maxlen = -1;
+
+		if (copydata_size == 0)
+		{
+			snprintf(state->err_msg,
+					 sizeof(state->err_msg),
+					 "zero length copydata received");
+			goto err;
+		}
+		msg_type = pq_getmsgbyte(&s);
+		switch (msg_type)
+		{
+			case 'w':
+				{
+					XLogRecPtr	start_lsn;
+
+					hdrlen = sizeof(int64) + sizeof(int64) + sizeof(int64);
+					if (s.len - s.cursor < hdrlen)
+					{
+						snprintf(state->err_msg,
+								 sizeof(state->err_msg),
+								 "invalid WAL message received from primary");
+						goto err;
+					}
+
+					start_lsn = pq_getmsgint64(&s);
+					pq_getmsgint64(&s); /* XLogRecPtr	end_lsn; */
+					pq_getmsgint64(&s); /* TimestampTz send_time */
+
+					state->rem_lsn = start_lsn;
+					state->wal_rem_len = (Size) (s.len - s.cursor);
+					state->wal_ptr = (char *) pq_getmsgbytes(&s, s.len - s.cursor);
+					nwr_log(DEBUG5, "received WAL msg at %X/%X len %zu",
+							LSN_FORMAT_ARGS(state->rem_lsn), state->wal_rem_len);
+
+					return NEON_WALREAD_SUCCESS;
+				}
+			case 'k':
+				{
+					XLogRecPtr	end_lsn;
+					bool		reply_requested;
+
+					hdrlen = sizeof(int64) + sizeof(int64) + sizeof(char);
+					if (s.len - s.cursor < hdrlen)
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "invalid keepalive message received from primary");
+						goto err;
+					}
+
+					end_lsn = pq_getmsgint64(&s);
+					pq_getmsgint64(&s); /* TimestampTz timestamp; */
+					reply_requested = pq_getmsgbyte(&s);
+					nwr_log(DEBUG5, "received keepalive end_lsn=%X/%X reply_requested=%d",
+							LSN_FORMAT_ARGS(end_lsn),
+							reply_requested);
+					if (end_lsn < state->req_lsn + state->req_len)
+					{
+						snprintf(state->err_msg, sizeof(state->err_msg),
+								 "closing remote connection: requested WAL up to %X/%X, but current donor %s has only up to %X/%X",
+								 LSN_FORMAT_ARGS(state->req_lsn + state->req_len), state->donor_name, LSN_FORMAT_ARGS(end_lsn));
+						goto err;
+					}
+					continue;
+				}
+			default:
+				nwr_log(WARNING, "invalid replication message type %d", msg_type);
+				continue;
+		}
+	}
+err:
+	NeonWALReaderResetRemote(state);
+	return NEON_WALREAD_ERROR;
+}
+
+/* reset remote connection and request in progress */
+static void
+NeonWALReaderResetRemote(NeonWALReader *state)
+{
+	state->req_lsn = InvalidXLogRecPtr;
+	state->req_len = 0;
+	state->req_progress = 0;
+	state->rem_state = RS_NONE;
+	if (state->wp_conn)
+	{
+		libpqwp_disconnect(state->wp_conn);
+		state->wp_conn = NULL;
+	}
+	state->donor_name[0] = '\0';
+	state->wal_ptr = NULL;
+	state->wal_rem_len = 0;
+	state->rem_lsn = InvalidXLogRecPtr;
+}
+
+/*
+ * Return socket of connection to remote source. Must be called only when
+ * connection exists (NeonWALReaderEvents returns non zero).
+ */
+pgsocket
+NeonWALReaderSocket(NeonWALReader *state)
+{
+	if (!state->wp_conn)
+		nwr_log(FATAL, "NeonWALReaderSocket is called without active remote connection");
+	return PQsocket(state->wp_conn->pg_conn);
+}
+
+/*
+ * Whether remote connection is established. Once this is done, until successful
+ * local read or error socket is stable and user can update socket events
+ * instead of readding it each time.
+ */
+bool
+NeonWALReaderIsRemConnEstablished(NeonWALReader *state)
+{
+	return state->rem_state == RS_ESTABLISHED;
+}
+
+/*
+ * Returns events user should wait on connection socket or 0 if remote
+ * connection is not active.
+ */
+extern uint32
+NeonWALReaderEvents(NeonWALReader *state)
+{
+	switch (state->rem_state)
+	{
+		case RS_NONE:
+			return 0;
+		case RS_CONNECTING_READ:
+			return WL_SOCKET_READABLE;
+		case RS_CONNECTING_WRITE:
+			return WL_SOCKET_WRITEABLE;
+		case RS_WAIT_EXEC_RESULT:
+		case RS_ESTABLISHED:
+			return WL_SOCKET_READABLE;
+		default:
+			Assert(false);
+			return 0;			/* make compiler happy */
+	}
+}
+
+static bool
+NeonWALReadLocal(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli)
+{
+	char	   *p;
+	XLogRecPtr	recptr;
+	Size		nbytes;
+
+	p = buf;
+	recptr = startptr;
+	nbytes = count;
+
+	while (nbytes > 0)
+	{
+		uint32		startoff;
+		int			segbytes;
+		int			readbytes;
+
+		startoff = XLogSegmentOffset(recptr, state->segcxt.ws_segsize);
+
+		/*
+		 * If the data we want is not in a segment we have open, close what we
+		 * have (if anything) and open the next one, using the caller's
+		 * provided openSegment callback.
+		 */
+		if (state->seg.ws_file < 0 ||
+			!XLByteInSeg(recptr, state->seg.ws_segno, state->segcxt.ws_segsize) ||
+			tli != state->seg.ws_tli)
+		{
+			XLogSegNo	nextSegNo;
+
+			neon_wal_segment_close(state);
+
+			XLByteToSeg(recptr, nextSegNo, state->segcxt.ws_segsize);
+			if (!neon_wal_segment_open(state, nextSegNo, &tli))
+			{
+				char		fname[MAXFNAMELEN];
+
+				state->wre_errno = errno;
+
+				XLogFileName(fname, tli, nextSegNo, state->segcxt.ws_segsize);
+				snprintf(state->err_msg, sizeof(state->err_msg), "failed to open WAL segment %s while reading at %X/%X: %s",
+						 fname, LSN_FORMAT_ARGS(recptr), strerror(state->wre_errno));
+				return false;
+			}
+
+			/* This shouldn't happen -- indicates a bug in segment_open */
+			Assert(state->seg.ws_file >= 0);
+
+			/* Update the current segment info. */
+			state->seg.ws_tli = tli;
+			state->seg.ws_segno = nextSegNo;
+		}
+
+		/* How many bytes are within this segment? */
+		if (nbytes > (state->segcxt.ws_segsize - startoff))
+			segbytes = state->segcxt.ws_segsize - startoff;
+		else
+			segbytes = nbytes;
+
+#ifndef FRONTEND
+		pgstat_report_wait_start(WAIT_EVENT_WAL_READ);
+#endif
+
+		/* Reset errno first; eases reporting non-errno-affecting errors */
+		errno = 0;
+		readbytes = pg_pread(state->seg.ws_file, p, segbytes, (off_t) startoff);
+
+#ifndef FRONTEND
+		pgstat_report_wait_end();
+#endif
+
+		if (readbytes <= 0)
+		{
+			char		fname[MAXFNAMELEN];
+
+			XLogFileName(fname, state->seg.ws_tli, state->seg.ws_segno, state->segcxt.ws_segsize);
+
+			if (readbytes < 0)
+			{
+				state->wre_errno = errno;
+				snprintf(state->err_msg, sizeof(state->err_msg), "could not read from log segment %s, offset %d: %m: %s",
+						 fname, startoff, strerror(state->wre_errno));
+			}
+			else
+			{
+				snprintf(state->err_msg, sizeof(state->err_msg), "could not read from log segment %s, offset %d: %m: unexpected EOF",
+						 fname, startoff);
+			}
+			return false;
+		}
+
+		/* Update state for read */
+		recptr += readbytes;
+		nbytes -= readbytes;
+		p += readbytes;
+	}
+
+	return true;
+}
+
+/*
+ * Copy of vanilla wal_segment_open, but returns false in case of error instead
+ * of ERROR, with errno set.
+ *
+ * XLogReaderRoutine->segment_open callback for local pg_wal files
+ */
+static bool
+neon_wal_segment_open(NeonWALReader *state, XLogSegNo nextSegNo,
+					  TimeLineID *tli_p)
+{
+	TimeLineID	tli = *tli_p;
+	char		path[MAXPGPATH];
+
+	XLogFilePath(path, tli, nextSegNo, state->segcxt.ws_segsize);
+	nwr_log(DEBUG5, "opening %s", path);
+	state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
+	if (state->seg.ws_file >= 0)
+		return true;
+
+	return false;
+}
+
+static bool
+is_wal_segment_exists(XLogSegNo segno, int segsize, TimeLineID tli)
+{
+	struct stat stat_buffer;
+	char		path[MAXPGPATH];
+
+	XLogFilePath(path, tli, segno, segsize);
+	return stat(path, &stat_buffer) == 0;
+}
+
+/* copy of vanilla wal_segment_close with NeonWALReader */
+static void
+neon_wal_segment_close(NeonWALReader *state)
+{
+	if (state->seg.ws_file >= 0)
+	{
+		close(state->seg.ws_file);
+		/* need to check errno? */
+		state->seg.ws_file = -1;
+	}
+}
+
+char *
+NeonWALReaderErrMsg(NeonWALReader *state)
+{
+	return state->err_msg;
+}

pgxn/neon/neon_walreader.h

@@ -0,0 +1,30 @@
+#ifndef __NEON_WALREADER_H__
+#define __NEON_WALREADER_H__
+
+#include "access/xlogdefs.h"
+
+/* forward declare so we don't have to expose the struct to the public */
+struct NeonWALReader;
+typedef struct NeonWALReader NeonWALReader;
+
+/* avoid including walproposer.h as it includes us */
+struct WalProposer;
+typedef struct WalProposer WalProposer;
+
+/* NeonWALRead return value */
+typedef enum
+{
+	NEON_WALREAD_SUCCESS,
+	NEON_WALREAD_WOULDBLOCK,
+	NEON_WALREAD_ERROR,
+} NeonWALReadResult;
+
+extern NeonWALReader *NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, WalProposer *wp, char *log_prefix);
+extern void NeonWALReaderFree(NeonWALReader *state);
+extern NeonWALReadResult NeonWALRead(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
+extern pgsocket NeonWALReaderSocket(NeonWALReader *state);
+extern uint32 NeonWALReaderEvents(NeonWALReader *state);
+extern bool NeonWALReaderIsRemConnEstablished(NeonWALReader *state);
+extern char *NeonWALReaderErrMsg(NeonWALReader *state);
+
+#endif							/* __NEON_WALREADER_H__ */

pgxn/neon/walproposer.h

@@ -8,6 +8,9 @@
 #include "replication/walreceiver.h"
 #include "utils/uuid.h"
 
+#include "libpqwalproposer.h"
+#include "neon_walreader.h"
+
 #define SK_MAGIC 0xCafeCeefu
 #define SK_PROTOCOL_VERSION 2
 
@@ -20,43 +23,9 @@
  */
 #define WL_NO_EVENTS 0
 
-struct WalProposerConn;			/* Defined in implementation (walprop_pg.c) */
+struct WalProposerConn;			/* Defined in libpqwalproposer.h */
 typedef struct WalProposerConn WalProposerConn;
 
-/* Possible return values from ReadPGAsync */
-typedef enum
-{
-	/* The full read was successful. buf now points to the data */
-	PG_ASYNC_READ_SUCCESS,
-
-	/*
-	 * The read is ongoing. Wait until the connection is read-ready, then try
-	 * again.
-	 */
-	PG_ASYNC_READ_TRY_AGAIN,
-	/* Reading failed. Check PQerrorMessage(conn) */
-	PG_ASYNC_READ_FAIL,
-} PGAsyncReadResult;
-
-/* Possible return values from WritePGAsync */
-typedef enum
-{
-	/* The write fully completed */
-	PG_ASYNC_WRITE_SUCCESS,
-
-	/*
-	 * The write started, but you'll need to call PQflush some more times to
-	 * finish it off. We just tried, so it's best to wait until the connection
-	 * is read- or write-ready to try again.
-	 *
-	 * If it becomes read-ready, call PQconsumeInput and flush again. If it
-	 * becomes write-ready, just call PQflush.
-	 */
-	PG_ASYNC_WRITE_TRY_FLUSH,
-	/* Writing failed. Check PQerrorMessage(conn) */
-	PG_ASYNC_WRITE_FAIL,
-} PGAsyncWriteResult;
-
 /*
  * WAL safekeeper state, which is used to wait for some event.
  *
@@ -133,6 +102,40 @@ typedef enum
 	SS_ACTIVE,
 } SafekeeperState;
 
+/*
+ * Sending WAL substates of SS_ACTIVE.
+ */
+typedef enum
+{
+	/*
+	 * We are ready to send more WAL, waiting for latch set to learn about
+	 * more WAL becoming available (or just a timeout to send heartbeat).
+	 */
+	SS_ACTIVE_SEND,
+
+	/*
+	 * Polling neon_walreader to receive chunk of WAL (probably remotely) to
+	 * send to this safekeeper.
+	 *
+	 * Note: socket management is done completely inside walproposer_pg for
+	 * simplicity, and thus simulation doesn't test it. Which is fine as
+	 * simulation is mainly aimed at consensus checks, not waiteventset
+	 * management.
+	 *
+	 * Also, while in this state we don't touch safekeeper socket, so in
+	 * theory it might close connection as inactive. This can be addressed if
+	 * needed; however, while fetching WAL we should regularly send it, so the
+	 * problem is unlikely. Vice versa is also true (SS_ACTIVE doesn't handle
+	 * walreader socket), but similarly shouldn't be a problem.
+	 */
+	SS_ACTIVE_READ_WAL,
+
+	/*
+	 * Waiting for write readiness to flush the socket.
+	 */
+	SS_ACTIVE_FLUSH,
+} SafekeeperActiveState;
+
 /* Consensus logical timestamp. */
 typedef uint64 term_t;
 
@@ -341,12 +344,11 @@ typedef struct Safekeeper
 	 */
 	XLogRecPtr	startStreamingAt;
 
-	bool		flushWrite;		/* set to true if we need to call AsyncFlush,*
-								 * to flush pending messages */
 	XLogRecPtr	streamingAt;	/* current streaming position */
 	AppendRequestHeader appendRequest;	/* request for sending to safekeeper */
 
 	SafekeeperState state;		/* safekeeper state machine state */
+	SafekeeperActiveState active_state;
 	TimestampTz latestMsgReceivedAt;	/* when latest msg is received */
 	AcceptorGreeting greetResponse; /* acceptor greeting */
 	VoteResponse voteResponse;	/* the vote */
@@ -367,12 +369,27 @@ typedef struct Safekeeper
 	/*
 	 * WAL reader, allocated for each safekeeper.
 	 */
-	XLogReaderState *xlogreader;
+	NeonWALReader *xlogreader;
 
 	/*
 	 * Position in wait event set. Equal to -1 if no event
 	 */
 	int			eventPos;
+
+	/*
+	 * Neon WAL reader position in wait event set, or -1 if no socket. Note
+	 * that event must be removed not only on error/failure, but also on
+	 * successful *local* read, as next read might again be remote, but with
+	 * different socket.
+	 */
+	int			nwrEventPos;
+
+	/*
+	 * Per libpq docs, during connection establishment socket might change,
+	 * remember here if it is stable to avoid readding to the event set if
+	 * possible. Must be reset whenever nwr event is deleted.
+	 */
+	bool		nwrConnEstablished;
 #endif
 
 
@@ -401,31 +418,6 @@ typedef enum
 	 */
 } WalProposerConnectPollStatusType;
 
-/* Re-exported and modified ExecStatusType */
-typedef enum
-{
-	/* We received a single CopyBoth result */
-	WP_EXEC_SUCCESS_COPYBOTH,
-
-	/*
-	 * Any success result other than a single CopyBoth was received. The
-	 * specifics of the result were already logged, but it may be useful to
-	 * provide an error message indicating which safekeeper messed up.
-	 *
-	 * Do not expect PQerrorMessage to be appropriately set.
-	 */
-	WP_EXEC_UNEXPECTED_SUCCESS,
-
-	/*
-	 * No result available at this time. Wait until read-ready, then call
-	 * again. Internally, this is returned when PQisBusy indicates that
-	 * PQgetResult would block.
-	 */
-	WP_EXEC_NEEDS_INPUT,
-	/* Catch-all failure. Check PQerrorMessage. */
-	WP_EXEC_FAILED,
-} WalProposerExecStatusType;
-
 /* Re-exported ConnStatusType */
 typedef enum
 {
@@ -486,7 +478,7 @@ typedef struct walproposer_api
 	/* Flush buffer to the network, aka PQflush. */
 	int			(*conn_flush) (Safekeeper *sk);
 
-	/* Close the connection, aka PQfinish. */
+	/* Reset sk state: close pq connection, deallocate xlogreader. */
 	void		(*conn_finish) (Safekeeper *sk);
 
 	/*
@@ -503,17 +495,20 @@ typedef struct walproposer_api
 	/* Blocking CopyData write, aka PQputCopyData + PQflush. */
 	bool		(*conn_blocking_write) (Safekeeper *sk, void const *buf, size_t size);
 
-	/* Download WAL from startpos to endpos and make it available locally. */
-	bool		(*recovery_download) (Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XLogRecPtr endpos);
-
-	/* Read WAL from disk to buf. */
-	void		(*wal_read) (Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count);
+	/*
+	 * Download WAL before basebackup for logical walsenders from sk, if
+	 * needed
+	 */
+	bool		(*recovery_download) (WalProposer *wp, Safekeeper *sk);
 
 	/* Allocate WAL reader. */
 	void		(*wal_reader_allocate) (Safekeeper *sk);
 
-	/* Deallocate event set. */
-	void		(*free_event_set) (WalProposer *wp);
+	/* Read WAL from disk to buf. */
+	NeonWALReadResult (*wal_read) (Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count, char **errmsg);
+
+	/* Returns events to be awaited on WAL reader, if any. */
+	uint32		(*wal_reader_events) (Safekeeper *sk);
 
 	/* Initialize event set. */
 	void		(*init_event_set) (WalProposer *wp);
@@ -521,9 +516,15 @@ typedef struct walproposer_api
 	/* Update events for an existing safekeeper connection. */
 	void		(*update_event_set) (Safekeeper *sk, uint32 events);
 
+	/* Configure wait event set for yield in SS_ACTIVE. */
+	void		(*active_state_update_event_set) (Safekeeper *sk);
+
 	/* Add a new safekeeper connection to the event set. */
 	void		(*add_safekeeper_event_set) (Safekeeper *sk, uint32 events);
 
+	/* Remove safekeeper connection from event set */
+	void		(*rm_safekeeper_event_set) (Safekeeper *sk);
+
 	/*
 	 * Wait until some event happens: - timeout is reached - socket event for
 	 * safekeeper connection - new WAL is available
@@ -556,26 +557,12 @@ typedef struct walproposer_api
 	 */
 	void		(*process_safekeeper_feedback) (WalProposer *wp, XLogRecPtr commitLsn);
 
-	/*
-	 * Called on peer_horizon_lsn updates. Used to advance replication slot
-	 * and to free up disk space by deleting unnecessary WAL.
-	 */
-	void		(*confirm_wal_streamed) (WalProposer *wp, XLogRecPtr lsn);
-
 	/*
 	 * Write a log message to the internal log processor. This is used only
 	 * when walproposer is compiled as a library. Otherwise, all logging is
 	 * handled by elog().
 	 */
 	void		(*log_internal) (WalProposer *wp, int level, const char *line);
-
-	/*
-	 * Called right after the proposer was elected, but before it started
-	 * recovery and sent ProposerElected message to the safekeepers.
-	 *
-	 * Used by logical replication to update truncateLsn.
-	 */
-	void		(*after_election) (WalProposer *wp);
 } walproposer_api;
 
 /*
@@ -709,15 +696,34 @@ extern void WalProposerBroadcast(WalProposer *wp, XLogRecPtr startpos, XLogRecPt
 extern void WalProposerPoll(WalProposer *wp);
 extern void WalProposerFree(WalProposer *wp);
 
+/*
+ * WaitEventSet API doesn't allow to remove socket, so walproposer_pg uses it to
+ * recreate set from scratch, hence the export.
+ */
+extern void SafekeeperStateDesiredEvents(Safekeeper *sk, uint32 *sk_events, uint32 *nwr_events);
+extern Safekeeper *GetDonor(WalProposer *wp, XLogRecPtr *donor_lsn);
+
 
 #define WPEVENT		1337		/* special log level for walproposer internal
 								 * events */
 
+#define WP_LOG_PREFIX "[WP] "
+
+/*
+ * wp_log is used in pure wp code (walproposer.c), allowing API callback to
+ * catch logging.
+ */
 #ifdef WALPROPOSER_LIB
 extern void WalProposerLibLog(WalProposer *wp, int elevel, char *fmt,...);
-#define walprop_log(elevel, ...) WalProposerLibLog(wp, elevel, __VA_ARGS__)
+#define wp_log(elevel, fmt, ...) WalProposerLibLog(wp, elevel, fmt, ## __VA_ARGS__)
 #else
-#define walprop_log(elevel, ...) elog(elevel, __VA_ARGS__)
+#define wp_log(elevel, fmt, ...) elog(elevel, WP_LOG_PREFIX fmt, ## __VA_ARGS__)
 #endif
 
+/*
+ * And wpg_log is used all other (postgres specific) walproposer code, just
+ * adding prefix.
+ */
+#define wpg_log(elevel, fmt, ...) elog(elevel, WP_LOG_PREFIX fmt, ## __VA_ARGS__)
+
 #endif							/* __NEON_WALPROPOSER_H__ */

pgxn/neon/walproposer_pg.c

@@ -12,6 +12,7 @@
 #include <unistd.h>
 #include <sys/stat.h>
 #include "access/xact.h"
+#include "access/xlog.h"
 #include "access/xlogdefs.h"
 #include "access/xlogutils.h"
 #include "access/xloginsert.h"
@@ -43,14 +44,19 @@
 #include "utils/ps_status.h"
 #include "utils/timestamp.h"
 
-#include "neon.h"
-#include "walproposer.h"
 #include "libpq-fe.h"
 
+#include "libpqwalproposer.h"
+#include "neon.h"
+#include "neon_walreader.h"
+#include "walproposer.h"
+
 #define XLOG_HDR_SIZE (1 + 8 * 3)	/* 'w' + startPos + walEnd + timestamp */
 #define XLOG_HDR_START_POS 1	/* offset of start position in wal sender*
 								 * message header */
 
+#define MB ((XLogRecPtr)1024 * 1024)
+
 #define WAL_PROPOSER_SLOT_NAME "wal_proposer_slot"
 
 char	   *wal_acceptors_list = "";
@@ -91,6 +97,12 @@ static void XLogBroadcastWalProposer(WalProposer *wp);
 static void XLogWalPropWrite(WalProposer *wp, char *buf, Size nbytes, XLogRecPtr recptr);
 static void XLogWalPropClose(XLogRecPtr recptr);
 
+static void add_nwr_event_set(Safekeeper *sk, uint32 events);
+static void update_nwr_event_set(Safekeeper *sk, uint32 events);
+static void rm_safekeeper_event_set(Safekeeper *to_remove, bool is_sk);
+
+static XLogRecPtr GetLogRepRestartLSN(WalProposer *wp);
+
 static void
 init_walprop_config(bool syncSafekeepers)
 {
@@ -214,7 +226,6 @@ backpressure_lag_impl(void)
 		XLogRecPtr	myFlushLsn = GetFlushRecPtr();
 #endif
 		replication_feedback_get_lsns(&writePtr, &flushPtr, &applyPtr);
-#define MB ((XLogRecPtr)1024 * 1024)
 
 		elog(DEBUG2, "current flushLsn %X/%X PageserverFeedback: write %X/%X flush %X/%X apply %X/%X",
 			 LSN_FORMAT_ARGS(myFlushLsn),
@@ -413,8 +424,8 @@ walprop_pg_start_streaming(WalProposer *wp, XLogRecPtr startpos)
 {
 	StartReplicationCmd cmd;
 
-	elog(LOG, "WAL proposer starts streaming at %X/%X",
-		 LSN_FORMAT_ARGS(startpos));
+	wpg_log(LOG, "WAL proposer starts streaming at %X/%X",
+			LSN_FORMAT_ARGS(startpos));
 	cmd.slotname = WAL_PROPOSER_SLOT_NAME;
 	cmd.timeline = wp->greetRequest.timeline;
 	cmd.startpoint = startpos;
@@ -538,17 +549,9 @@ walprop_pg_load_libpqwalreceiver(void)
 {
 	load_file("libpqwalreceiver", false);
 	if (WalReceiverFunctions == NULL)
-		elog(ERROR, "libpqwalreceiver didn't initialize correctly");
+		wpg_log(ERROR, "libpqwalreceiver didn't initialize correctly");
 }
 
-/* Header in walproposer.h -- Wrapper struct to abstract away the libpq connection */
-struct WalProposerConn
-{
-	PGconn	   *pg_conn;
-	bool		is_nonblocking; /* whether the connection is non-blocking */
-	char	   *recvbuf;		/* last received data from walprop_async_read */
-};
-
 /* Helper function */
 static bool
 ensure_nonblocking_status(WalProposerConn *conn, bool is_nonblocking)
@@ -586,16 +589,17 @@ walprop_status(Safekeeper *sk)
 	}
 }
 
-static void
-walprop_connect_start(Safekeeper *sk)
+WalProposerConn *
+libpqwp_connect_start(char *conninfo)
 {
+
 	PGconn	   *pg_conn;
+	WalProposerConn *conn;
 	const char *keywords[3];
 	const char *values[3];
 	int			n;
 	char	   *password = neon_auth_token;
 
-	Assert(sk->conn == NULL);
 
 	/*
 	 * Connect using the given connection string. If the NEON_AUTH_TOKEN
@@ -614,7 +618,7 @@ walprop_connect_start(Safekeeper *sk)
 		n++;
 	}
 	keywords[n] = "dbname";
-	values[n] = sk->conninfo;
+	values[n] = conninfo;
 	n++;
 	keywords[n] = NULL;
 	values[n] = NULL;
@@ -626,7 +630,7 @@ walprop_connect_start(Safekeeper *sk)
 	 * PGconn structure"
 	 */
 	if (!pg_conn)
-		elog(FATAL, "failed to allocate new PGconn object");
+		wpg_log(FATAL, "failed to allocate new PGconn object");
 
 	/*
 	 * And in theory this allocation can fail as well, but it's incredibly
@@ -635,11 +639,20 @@ walprop_connect_start(Safekeeper *sk)
 	 * palloc will exit on failure though, so there's not much we could do if
 	 * it *did* fail.
 	 */
-	sk->conn = palloc(sizeof(WalProposerConn));
-	sk->conn->pg_conn = pg_conn;
-	sk->conn->is_nonblocking = false;	/* connections always start in
-										 * blocking mode */
-	sk->conn->recvbuf = NULL;
+	conn = palloc(sizeof(WalProposerConn));
+	conn->pg_conn = pg_conn;
+	conn->is_nonblocking = false;	/* connections always start in blocking
+									 * mode */
+	conn->recvbuf = NULL;
+	return conn;
+}
+
+static void
+walprop_connect_start(Safekeeper *sk)
+{
+	Assert(sk->conn == NULL);
+	sk->conn = libpqwp_connect_start(sk->conninfo);
+
 }
 
 static WalProposerConnectPollStatusType
@@ -667,7 +680,7 @@ walprop_connect_poll(Safekeeper *sk)
 			 * unused. We'll expect it's never returned.
 			 */
 		case PGRES_POLLING_ACTIVE:
-			elog(FATAL, "Unexpected PGRES_POLLING_ACTIVE returned from PQconnectPoll");
+			wpg_log(FATAL, "unexpected PGRES_POLLING_ACTIVE returned from PQconnectPoll");
 
 			/*
 			 * This return is never actually reached, but it's here to make
@@ -683,26 +696,33 @@ walprop_connect_poll(Safekeeper *sk)
 	return return_val;
 }
 
-static bool
-walprop_send_query(Safekeeper *sk, char *query)
+extern bool
+libpqwp_send_query(WalProposerConn *conn, char *query)
 {
 	/*
 	 * We need to be in blocking mode for sending the query to run without
 	 * requiring a call to PQflush
 	 */
-	if (!ensure_nonblocking_status(sk->conn, false))
+	if (!ensure_nonblocking_status(conn, false))
 		return false;
 
 	/* PQsendQuery returns 1 on success, 0 on failure */
-	if (!PQsendQuery(sk->conn->pg_conn, query))
+	if (!PQsendQuery(conn->pg_conn, query))
 		return false;
 
 	return true;
 }
 
-static WalProposerExecStatusType
-walprop_get_query_result(Safekeeper *sk)
+static bool
+walprop_send_query(Safekeeper *sk, char *query)
 {
+	return libpqwp_send_query(sk->conn, query);
+}
+
+WalProposerExecStatusType
+libpqwp_get_query_result(WalProposerConn *conn)
+{
+
 	PGresult   *result;
 	WalProposerExecStatusType return_val;
 
@@ -710,14 +730,14 @@ walprop_get_query_result(Safekeeper *sk)
 	char	   *unexpected_success = NULL;
 
 	/* Consume any input that we might be missing */
-	if (!PQconsumeInput(sk->conn->pg_conn))
+	if (!PQconsumeInput(conn->pg_conn))
 		return WP_EXEC_FAILED;
 
-	if (PQisBusy(sk->conn->pg_conn))
+	if (PQisBusy(conn->pg_conn))
 		return WP_EXEC_NEEDS_INPUT;
 
 
-	result = PQgetResult(sk->conn->pg_conn);
+	result = PQgetResult(conn->pg_conn);
 
 	/*
 	 * PQgetResult returns NULL only if getting the result was successful &
@@ -725,7 +745,7 @@ walprop_get_query_result(Safekeeper *sk)
 	 */
 	if (!result)
 	{
-		elog(WARNING, "[libpqwalproposer] Unexpected successful end of command results");
+		wpg_log(WARNING, "[libpqwalproposer] Unexpected successful end of command results");
 		return WP_EXEC_UNEXPECTED_SUCCESS;
 	}
 
@@ -773,11 +793,17 @@ walprop_get_query_result(Safekeeper *sk)
 	}
 
 	if (unexpected_success)
-		elog(WARNING, "[libpqwalproposer] Unexpected successful %s", unexpected_success);
+		wpg_log(WARNING, "[libpqwalproposer] Unexpected successful %s", unexpected_success);
 
 	return return_val;
 }
 
+static WalProposerExecStatusType
+walprop_get_query_result(Safekeeper *sk)
+{
+	return libpqwp_get_query_result(sk->conn);
+}
+
 static pgsocket
 walprop_socket(Safekeeper *sk)
 {
@@ -790,42 +816,31 @@ walprop_flush(Safekeeper *sk)
 	return (PQflush(sk->conn->pg_conn));
 }
 
-static void
-walprop_finish(Safekeeper *sk)
+/* Like libpqrcv_receive. *buf is valid until the next call. */
+PGAsyncReadResult
+libpqwp_async_read(WalProposerConn *conn, char **buf, int *amount)
 {
-	if (!sk->conn)
-		return;
+	int			rawlen;
 
-	if (sk->conn->recvbuf != NULL)
-		PQfreemem(sk->conn->recvbuf);
-	PQfinish(sk->conn->pg_conn);
-	pfree(sk->conn);
-	sk->conn = NULL;
-}
-
-/*
- * Receive a message from the safekeeper.
- *
- * On success, the data is placed in *buf. It is valid until the next call
- * to this function.
- */
-static PGAsyncReadResult
-walprop_async_read(Safekeeper *sk, char **buf, int *amount)
-{
-	int			result;
-
-	if (sk->conn->recvbuf != NULL)
+	if (conn->recvbuf != NULL)
 	{
-		PQfreemem(sk->conn->recvbuf);
-		sk->conn->recvbuf = NULL;
+		PQfreemem(conn->recvbuf);
+		conn->recvbuf = NULL;
 	}
 
-	/* Call PQconsumeInput so that we have the data we need */
-	if (!PQconsumeInput(sk->conn->pg_conn))
+	/* Try to receive a CopyData message */
+	rawlen = PQgetCopyData(conn->pg_conn, &conn->recvbuf, true);
+	if (rawlen == 0)
 	{
-		*amount = 0;
-		*buf = NULL;
-		return PG_ASYNC_READ_FAIL;
+		/* Try consuming some data. */
+		if (!PQconsumeInput(conn->pg_conn))
+		{
+			*amount = 0;
+			*buf = NULL;
+			return PG_ASYNC_READ_FAIL;
+		}
+		/* Now that we've consumed some input, try again */
+		rawlen = PQgetCopyData(conn->pg_conn, &conn->recvbuf, true);
 	}
 
 	/*
@@ -839,7 +854,7 @@ walprop_async_read(Safekeeper *sk, char **buf, int *amount)
 	 * sometimes be triggered by the server returning an ErrorResponse (which
 	 * also happens to have the effect that the copy is done).
 	 */
-	switch (result = PQgetCopyData(sk->conn->pg_conn, &sk->conn->recvbuf, true))
+	switch (rawlen)
 	{
 		case 0:
 			*amount = 0;
@@ -854,10 +869,10 @@ walprop_async_read(Safekeeper *sk, char **buf, int *amount)
 				 * We can check PQgetResult to make sure that the server
 				 * failed; it'll always result in PGRES_FATAL_ERROR
 				 */
-				ExecStatusType status = PQresultStatus(PQgetResult(sk->conn->pg_conn));
+				ExecStatusType status = PQresultStatus(PQgetResult(conn->pg_conn));
 
 				if (status != PGRES_FATAL_ERROR)
-					elog(FATAL, "unexpected result status %d after failed PQgetCopyData", status);
+					wpg_log(FATAL, "unexpected result status %d after failed PQgetCopyData", status);
 
 				/*
 				 * If there was actually an error, it'll be properly reported
@@ -874,12 +889,24 @@ walprop_async_read(Safekeeper *sk, char **buf, int *amount)
 			return PG_ASYNC_READ_FAIL;
 		default:
 			/* Positive values indicate the size of the returned result */
-			*amount = result;
-			*buf = sk->conn->recvbuf;
+			*amount = rawlen;
+			*buf = conn->recvbuf;
 			return PG_ASYNC_READ_SUCCESS;
 	}
 }
 
+/*
+ * Receive a message from the safekeeper.
+ *
+ * On success, the data is placed in *buf. It is valid until the next call
+ * to this function.
+ */
+static PGAsyncReadResult
+walprop_async_read(Safekeeper *sk, char **buf, int *amount)
+{
+	return libpqwp_async_read(sk->conn, buf, amount);
+}
+
 static PGAsyncWriteResult
 walprop_async_write(Safekeeper *sk, void const *buf, size_t size)
 {
@@ -910,7 +937,7 @@ walprop_async_write(Safekeeper *sk, void const *buf, size_t size)
 		case -1:
 			return PG_ASYNC_WRITE_FAIL;
 		default:
-			elog(FATAL, "invalid return %d from PQputCopyData", result);
+			wpg_log(FATAL, "invalid return %d from PQputCopyData", result);
 	}
 
 	/*
@@ -931,7 +958,7 @@ walprop_async_write(Safekeeper *sk, void const *buf, size_t size)
 		case -1:
 			return PG_ASYNC_WRITE_FAIL;
 		default:
-			elog(FATAL, "invalid return %d from PQflush", result);
+			wpg_log(FATAL, "invalid return %d from PQflush", result);
 	}
 }
 
@@ -962,6 +989,33 @@ walprop_blocking_write(Safekeeper *sk, void const *buf, size_t size)
 	return true;
 }
 
+void
+libpqwp_disconnect(WalProposerConn *conn)
+{
+	if (conn->recvbuf != NULL)
+		PQfreemem(conn->recvbuf);
+	PQfinish(conn->pg_conn);
+	pfree(conn);
+}
+
+static void
+walprop_finish(Safekeeper *sk)
+{
+	if (sk->conn)
+	{
+		libpqwp_disconnect(sk->conn);
+		sk->conn = NULL;
+	}
+
+	/* free xlogreader */
+	if (sk->xlogreader)
+	{
+		NeonWALReaderFree(sk->xlogreader);
+		sk->xlogreader = NULL;
+	}
+	rm_safekeeper_event_set(sk, false);
+}
+
 /*
  * Subscribe for new WAL and stream it in the loop to safekeepers.
  *
@@ -1165,16 +1219,25 @@ XLogBroadcastWalProposer(WalProposer *wp)
 	}
 }
 
-/*
- * Receive WAL from most advanced safekeeper
- */
+/* Download WAL before basebackup for logical walsenders from sk, if needed */
 static bool
-WalProposerRecovery(Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XLogRecPtr endpos)
+WalProposerRecovery(WalProposer *wp, Safekeeper *sk)
 {
 	char	   *err;
 	WalReceiverConn *wrconn;
 	WalRcvStreamOptions options;
 	char		conninfo[MAXCONNINFO];
+	TimeLineID	timeline;
+	XLogRecPtr	startpos;
+	XLogRecPtr	endpos;
+	uint64		download_range_mb;
+
+	startpos = GetLogRepRestartLSN(wp);
+	if (startpos == InvalidXLogRecPtr)
+		return true;			/* recovery not needed */
+	endpos = wp->propEpochStartLsn;
+
+	timeline = wp->greetRequest.timeline;
 
 	if (!neon_auth_token)
 	{
@@ -1186,7 +1249,7 @@ WalProposerRecovery(Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XL
 
 		written = snprintf((char *) conninfo, MAXCONNINFO, "password=%s %s", neon_auth_token, sk->conninfo);
 		if (written > MAXCONNINFO || written < 0)
-			elog(FATAL, "could not append password to the safekeeper connection string");
+			wpg_log(FATAL, "could not append password to the safekeeper connection string");
 	}
 
 #if PG_MAJORVERSION_NUM < 16
@@ -1203,11 +1266,11 @@ WalProposerRecovery(Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XL
 						err)));
 		return false;
 	}
-	elog(LOG,
-		 "start recovery from %s:%s starting from %X/%08X till %X/%08X timeline "
-		 "%d",
-		 sk->host, sk->port, (uint32) (startpos >> 32),
-		 (uint32) startpos, (uint32) (endpos >> 32), (uint32) endpos, timeline);
+	wpg_log(LOG,
+			"start recovery for logical replication from %s:%s starting from %X/%08X till %X/%08X timeline "
+			"%d",
+			sk->host, sk->port, (uint32) (startpos >> 32),
+			(uint32) startpos, (uint32) (endpos >> 32), (uint32) endpos, timeline);
 
 	options.logical = false;
 	options.startpoint = startpos;
@@ -1400,28 +1463,54 @@ XLogWalPropClose(XLogRecPtr recptr)
 	walpropFile = -1;
 }
 
-static void
-walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count)
-{
-	WALReadError errinfo;
-
-	if (!WALRead(sk->xlogreader,
-				 buf,
-				 startptr,
-				 count,
-				 walprop_pg_get_timeline_id(),
-				 &errinfo))
-	{
-		WALReadRaiseError(&errinfo);
-	}
-}
-
 static void
 walprop_pg_wal_reader_allocate(Safekeeper *sk)
 {
-	sk->xlogreader = XLogReaderAllocate(wal_segment_size, NULL, XL_ROUTINE(.segment_open = wal_segment_open,.segment_close = wal_segment_close), NULL);
+	char		log_prefix[64];
+
+	snprintf(log_prefix, sizeof(log_prefix), WP_LOG_PREFIX "sk %s:%s nwr: ", sk->host, sk->port);
+	Assert(!sk->xlogreader);
+	sk->xlogreader = NeonWALReaderAllocate(wal_segment_size, sk->wp->propEpochStartLsn, sk->wp, log_prefix);
 	if (sk->xlogreader == NULL)
-		elog(FATAL, "Failed to allocate xlog reader");
+		wpg_log(FATAL, "failed to allocate xlog reader");
+}
+
+static NeonWALReadResult
+walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count, char **errmsg)
+{
+	NeonWALReadResult res;
+
+	res = NeonWALRead(sk->xlogreader,
+					  buf,
+					  startptr,
+					  count,
+					  walprop_pg_get_timeline_id());
+
+	if (res == NEON_WALREAD_SUCCESS)
+	{
+		/*
+		 * If we have the socket subscribed, but walreader doesn't need any
+		 * events, it must mean that remote connection just closed hoping to
+		 * do next read locally. Remove the socket then. It is important to do
+		 * as otherwise next read might open another connection and we won't
+		 * be able to distinguish whether we have correct socket added in wait
+		 * event set.
+		 */
+		if (NeonWALReaderEvents(sk->xlogreader) == 0)
+			rm_safekeeper_event_set(sk, false);
+	}
+	else if (res == NEON_WALREAD_ERROR)
+	{
+		*errmsg = NeonWALReaderErrMsg(sk->xlogreader);
+	}
+
+	return res;
+}
+
+static uint32
+walprop_pg_wal_reader_events(Safekeeper *sk)
+{
+	return NeonWALReaderEvents(sk->xlogreader);
 }
 
 static WaitEventSet *waitEvents;
@@ -1438,6 +1527,8 @@ walprop_pg_free_event_set(WalProposer *wp)
 	for (int i = 0; i < wp->n_safekeepers; i++)
 	{
 		wp->safekeeper[i].eventPos = -1;
+		wp->safekeeper[i].nwrEventPos = -1;
+		wp->safekeeper[i].nwrConnEstablished = false;
 	}
 }
 
@@ -1445,13 +1536,39 @@ static void
 walprop_pg_init_event_set(WalProposer *wp)
 {
 	if (waitEvents)
-		elog(FATAL, "double-initialization of event set");
+		wpg_log(FATAL, "double-initialization of event set");
 
-	waitEvents = CreateWaitEventSet(TopMemoryContext, 2 + wp->n_safekeepers);
+	/* for each sk, we have socket plus potentially socket for neon walreader */
+	waitEvents = CreateWaitEventSet(TopMemoryContext, 2 + 2 * wp->n_safekeepers);
 	AddWaitEventToSet(waitEvents, WL_LATCH_SET, PGINVALID_SOCKET,
 					  MyLatch, NULL);
 	AddWaitEventToSet(waitEvents, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
 					  NULL, NULL);
+
+	for (int i = 0; i < wp->n_safekeepers; i++)
+	{
+		wp->safekeeper[i].eventPos = -1;
+		wp->safekeeper[i].nwrEventPos = -1;
+		wp->safekeeper[i].nwrConnEstablished = false;
+	}
+}
+
+/* add safekeeper socket to wait event set */
+static void
+walprop_pg_add_safekeeper_event_set(Safekeeper *sk, uint32 events)
+{
+	Assert(sk->eventPos == -1);
+	sk->eventPos = AddWaitEventToSet(waitEvents, events, walprop_socket(sk), NULL, sk);
+}
+
+/* add neon wal reader socket to wait event set */
+static void
+add_nwr_event_set(Safekeeper *sk, uint32 events)
+{
+	Assert(sk->nwrEventPos == -1);
+	sk->nwrEventPos = AddWaitEventToSet(waitEvents, events, NeonWALReaderSocket(sk->xlogreader), NULL, sk);
+	sk->nwrConnEstablished = NeonWALReaderIsRemConnEstablished(sk->xlogreader);
+	wpg_log(DEBUG5, "sk %s:%s: added nwr socket events %d", sk->host, sk->port, events);
 }
 
 static void
@@ -1463,10 +1580,144 @@ walprop_pg_update_event_set(Safekeeper *sk, uint32 events)
 	ModifyWaitEvent(waitEvents, sk->eventPos, events, NULL);
 }
 
+/*
+ * Update neon_walreader event.
+ * Can be called when nwr socket doesn't exist, does nothing in this case.
+ */
 static void
-walprop_pg_add_safekeeper_event_set(Safekeeper *sk, uint32 events)
+update_nwr_event_set(Safekeeper *sk, uint32 events)
 {
-	sk->eventPos = AddWaitEventToSet(waitEvents, events, walprop_socket(sk), NULL, sk);
+	/* eventPos = -1 when we don't have an event */
+	if (sk->nwrEventPos != -1)
+		ModifyWaitEvent(waitEvents, sk->nwrEventPos, events, NULL);
+}
+
+
+static void
+walprop_pg_active_state_update_event_set(Safekeeper *sk)
+{
+	uint32		sk_events;
+	uint32		nwr_events;
+
+	Assert(sk->state == SS_ACTIVE);
+	SafekeeperStateDesiredEvents(sk, &sk_events, &nwr_events);
+
+	/*
+	 * If we need to wait for neon_walreader, ensure we have up to date socket
+	 * in the wait event set.
+	 */
+	if (sk->active_state == SS_ACTIVE_READ_WAL)
+	{
+		/*
+		 * If conn is established and socket is thus stable, update the event
+		 * directly; otherwise re-add it.
+		 */
+		if (sk->nwrConnEstablished)
+		{
+			Assert(sk->nwrEventPos != -1);
+			update_nwr_event_set(sk, nwr_events);
+		}
+		else
+		{
+			rm_safekeeper_event_set(sk, false);
+			add_nwr_event_set(sk, nwr_events);
+		}
+	}
+	else
+	{
+		/*
+		 * Hack: we should always set 0 here, but for random reasons
+		 * WaitEventSet (WaitEventAdjustEpoll) asserts that there is at least
+		 * some event. Since there is also no way to remove socket except
+		 * reconstructing the whole set, SafekeeperStateDesiredEvents instead
+		 * gives WL_SOCKET_CLOSED if socket exists. We never expect it to
+		 * trigger.
+		 *
+		 * On PG 14 which doesn't have WL_SOCKET_CLOSED resort to event
+		 * removal.
+		 */
+#if PG_VERSION_NUM >= 150000
+		Assert(nwr_events == WL_SOCKET_CLOSED || nwr_events == 0);
+		update_nwr_event_set(sk, WL_SOCKET_CLOSED);
+#else							/* pg 14 */
+		rm_safekeeper_event_set(sk, false);
+#endif
+	}
+	walprop_pg_update_event_set(sk, sk_events);
+}
+
+static void
+walprop_pg_rm_safekeeper_event_set(Safekeeper *to_remove)
+{
+	rm_safekeeper_event_set(to_remove, true);
+}
+
+/*
+ * A hacky way to remove single event from the event set. Can be called if event
+ * doesn't exist, does nothing in this case.
+ *
+ * Note: Internally, this completely reconstructs the event set. It should be
+ * avoided if possible.
+ *
+ * If is_sk is true, socket of connection to safekeeper is removed; otherwise
+ * socket of neon_walreader.
+ */
+static void
+rm_safekeeper_event_set(Safekeeper *to_remove, bool is_sk)
+{
+	WalProposer *wp = to_remove->wp;
+
+	wpg_log(DEBUG5, "sk %s:%s: removing event, is_sk %d",
+			to_remove->host, to_remove->port, is_sk);
+
+	/*
+	 * Shortpath for exiting if have nothing to do. We never call this
+	 * function with safekeeper socket not existing, but do that with neon
+	 * walreader socket.
+	 */
+	if ((is_sk && to_remove->eventPos == -1) ||
+		(!is_sk && to_remove->nwrEventPos == -1))
+	{
+		return;
+	}
+
+	/* Remove the existing event set, assign sk->eventPos = -1 */
+	walprop_pg_free_event_set(wp);
+
+	/* Re-initialize it without adding any safekeeper events */
+	wp->api.init_event_set(wp);
+
+	/*
+	 * loop through the existing safekeepers. If they aren't the one we're
+	 * removing, and if they have a socket we can use, re-add the applicable
+	 * events.
+	 */
+	for (int i = 0; i < wp->n_safekeepers; i++)
+	{
+		Safekeeper *sk = &wp->safekeeper[i];
+
+		/*
+		 * If this safekeeper isn't offline, add events for it, except for the
+		 * event requested to remove.
+		 */
+		if (sk->state != SS_OFFLINE)
+		{
+			uint32		sk_events;
+			uint32		nwr_events;
+
+			SafekeeperStateDesiredEvents(sk, &sk_events, &nwr_events);
+
+			if (sk != to_remove || !is_sk)
+			{
+				/* will set sk->eventPos */
+				wp->api.add_safekeeper_event_set(sk, sk_events);
+			}
+			if ((sk != to_remove || is_sk) && nwr_events)
+			{
+				add_nwr_event_set(sk, nwr_events);
+			}
+		}
+	}
 }
 
 static int
@@ -1484,8 +1735,8 @@ walprop_pg_wait_event_set(WalProposer *wp, long timeout, Safekeeper **sk, uint32
 		ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
 
 	/*
-	 * Now that we prepared the condvar, check flush ptr again -- it might have
-	 * changed before we subscribed to cv so we missed the wakeup.
+	 * Now that we prepared the condvar, check flush ptr again -- it might
+	 * have changed before we subscribed to cv so we missed the wakeup.
 	 *
 	 * Do that only when we're interested in new WAL: without sync-safekeepers
 	 * and if election already passed.
@@ -1548,7 +1799,7 @@ walprop_pg_finish_sync_safekeepers(WalProposer *wp, XLogRecPtr lsn)
 }
 
 /*
- * Get PageserverFeedback fields from the most advanced safekeeper
+ * Choose most advanced PageserverFeedback and set it to *rf.
  */
 static void
 GetLatestNeonFeedback(PageserverFeedback *rf, WalProposer *wp)
@@ -1571,15 +1822,13 @@ GetLatestNeonFeedback(PageserverFeedback *rf, WalProposer *wp)
 	rf->remote_consistent_lsn = wp->safekeeper[latest_safekeeper].appendResponse.rf.remote_consistent_lsn;
 	rf->replytime = wp->safekeeper[latest_safekeeper].appendResponse.rf.replytime;
 
-	elog(DEBUG2, "GetLatestNeonFeedback: currentClusterSize %lu,"
-		 " last_received_lsn %X/%X, disk_consistent_lsn %X/%X, remote_consistent_lsn %X/%X, replytime %lu",
-		 rf->currentClusterSize,
-		 LSN_FORMAT_ARGS(rf->last_received_lsn),
-		 LSN_FORMAT_ARGS(rf->disk_consistent_lsn),
-		 LSN_FORMAT_ARGS(rf->remote_consistent_lsn),
-		 rf->replytime);
-
-	replication_feedback_set(rf);
+	wpg_log(DEBUG2, "GetLatestNeonFeedback: currentClusterSize %lu,"
+			" last_received_lsn %X/%X, disk_consistent_lsn %X/%X, remote_consistent_lsn %X/%X, replytime %lu",
+			rf->currentClusterSize,
+			LSN_FORMAT_ARGS(rf->last_received_lsn),
+			LSN_FORMAT_ARGS(rf->disk_consistent_lsn),
+			LSN_FORMAT_ARGS(rf->remote_consistent_lsn),
+			rf->replytime);
 }
 
 /*
@@ -1619,63 +1868,69 @@ CombineHotStanbyFeedbacks(HotStandbyFeedback *hs, WalProposer *wp)
 		hs->catalog_xmin = InvalidFullTransactionId;
 }
 
+/*
+ * Based on commitLsn and safekeeper responses including pageserver feedback,
+ * 1) Propagate cluster size received from ps to ensure the limit.
+ * 2) Propagate pageserver LSN positions to ensure backpressure limits.
+ * 3) Advance walproposer slot to commitLsn (releasing WAL & waking up waiters).
+ * 4) Propagate hot standby feedback.
+ *
+ * None of that is functional in sync-safekeepers.
+ */
 static void
 walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
 {
 	HotStandbyFeedback hsFeedback;
-	XLogRecPtr	diskConsistentLsn;
+	XLogRecPtr	oldDiskConsistentLsn;
 
-	diskConsistentLsn = quorumFeedback.rf.disk_consistent_lsn;
+	if (wp->config->syncSafekeepers)
+		return;
 
-	if (!wp->config->syncSafekeepers)
+	oldDiskConsistentLsn = quorumFeedback.rf.disk_consistent_lsn;
+
+	/* Get PageserverFeedback fields from the most advanced safekeeper */
+	GetLatestNeonFeedback(&quorumFeedback.rf, wp);
+	replication_feedback_set(&quorumFeedback.rf);
+	SetZenithCurrentClusterSize(quorumFeedback.rf.currentClusterSize);
+
+	if (commitLsn > quorumFeedback.flushLsn || oldDiskConsistentLsn != quorumFeedback.rf.disk_consistent_lsn)
 	{
-		/* Get PageserverFeedback fields from the most advanced safekeeper */
-		GetLatestNeonFeedback(&quorumFeedback.rf, wp);
-		SetZenithCurrentClusterSize(quorumFeedback.rf.currentClusterSize);
-	}
-
-	if (commitLsn > quorumFeedback.flushLsn || diskConsistentLsn != quorumFeedback.rf.disk_consistent_lsn)
-	{
-
 		if (commitLsn > quorumFeedback.flushLsn)
 			quorumFeedback.flushLsn = commitLsn;
 
-		/* advance the replication slot */
-		if (!wp->config->syncSafekeepers)
-			ProcessStandbyReply(
-			/* write_lsn -  This is what durably stored in WAL service. */
-								quorumFeedback.flushLsn,
-			/* flush_lsn - This is what durably stored in WAL service. */
-								quorumFeedback.flushLsn,
+		/*
+		 * Advance the replication slot to commitLsn. WAL before it is
+		 * hardened and will be fetched from one of safekeepers by
+		 * neon_walreader if needed.
+		 *
+		 * Also wakes up syncrep waiters.
+		 */
+		ProcessStandbyReply(
+		/* write_lsn -  This is what durably stored in WAL service. */
+							quorumFeedback.flushLsn,
+		/* flush_lsn - This is what durably stored in WAL service. */
+							quorumFeedback.flushLsn,
 
-			/*
-			 * apply_lsn - This is what processed and durably saved at*
-			 * pageserver.
-			 */
-								quorumFeedback.rf.disk_consistent_lsn,
-								walprop_pg_get_current_timestamp(wp), false);
+		/*
+		 * apply_lsn - This is what processed and durably saved at*
+		 * pageserver.
+		 */
+							quorumFeedback.rf.disk_consistent_lsn,
+							walprop_pg_get_current_timestamp(wp), false);
 	}
 
 	CombineHotStanbyFeedbacks(&hsFeedback, wp);
 	if (hsFeedback.ts != 0 && memcmp(&hsFeedback, &quorumFeedback.hs, sizeof hsFeedback) != 0)
 	{
 		quorumFeedback.hs = hsFeedback;
-		if (!wp->config->syncSafekeepers)
-			ProcessStandbyHSFeedback(hsFeedback.ts,
-									 XidFromFullTransactionId(hsFeedback.xmin),
-									 EpochFromFullTransactionId(hsFeedback.xmin),
-									 XidFromFullTransactionId(hsFeedback.catalog_xmin),
-									 EpochFromFullTransactionId(hsFeedback.catalog_xmin));
+		ProcessStandbyHSFeedback(hsFeedback.ts,
+								 XidFromFullTransactionId(hsFeedback.xmin),
+								 EpochFromFullTransactionId(hsFeedback.xmin),
+								 XidFromFullTransactionId(hsFeedback.catalog_xmin),
+								 EpochFromFullTransactionId(hsFeedback.catalog_xmin));
 	}
 }
 
-static void
-walprop_pg_confirm_wal_streamed(WalProposer *wp, XLogRecPtr lsn)
-{
-	if (MyReplicationSlot)
-		PhysicalConfirmReceivedLocation(lsn);
-}
-
 static XLogRecPtr
 walprop_pg_get_redo_start_lsn(WalProposer *wp)
 {
@@ -1694,15 +1949,15 @@ walprop_pg_log_internal(WalProposer *wp, int level, const char *line)
 	elog(FATAL, "unexpected log_internal message at level %d: %s", level, line);
 }
 
-static void
-walprop_pg_after_election(WalProposer *wp)
+static XLogRecPtr
+GetLogRepRestartLSN(WalProposer *wp)
 {
 	FILE	   *f;
-	XLogRecPtr	lrRestartLsn;
+	XLogRecPtr	lrRestartLsn = InvalidXLogRecPtr;
 
 	/* We don't need to do anything in syncSafekeepers mode. */
 	if (wp->config->syncSafekeepers)
-		return;
+		return InvalidXLogRecPtr;
 
 	/*
 	 * If there are active logical replication subscription we need to provide
@@ -1710,22 +1965,40 @@ walprop_pg_after_election(WalProposer *wp)
 	 * replication slots.
 	 */
 	f = fopen("restart.lsn", "rb");
-	if (f != NULL && !wp->config->syncSafekeepers)
+	if (f != NULL)
 	{
-		fread(&lrRestartLsn, sizeof(lrRestartLsn), 1, f);
+		size_t		rc = fread(&lrRestartLsn, sizeof(lrRestartLsn), 1, f);
+
 		fclose(f);
-		if (lrRestartLsn != InvalidXLogRecPtr)
+		if (rc == 1 && lrRestartLsn != InvalidXLogRecPtr)
 		{
-			elog(LOG, "Logical replication restart LSN %X/%X", LSN_FORMAT_ARGS(lrRestartLsn));
+			uint64		download_range_mb;
+
+			wpg_log(LOG, "logical replication restart LSN %X/%X", LSN_FORMAT_ARGS(lrRestartLsn));
+
+			/*
+			 * If we need to download more than a max_slot_wal_keep_size,
+			 * don't do it to avoid risk of exploding pg_wal. Logical
+			 * replication won't work until recreated, but at least compute
+			 * would start; this also follows max_slot_wal_keep_size
+			 * semantics.
+			 */
+			download_range_mb = (wp->propEpochStartLsn - lrRestartLsn) / MB;
+			if (max_slot_wal_keep_size_mb > 0 && download_range_mb >= max_slot_wal_keep_size_mb)
+			{
+				wpg_log(WARNING, "not downloading WAL for logical replication since %X/%X as max_slot_wal_keep_size=%dMB",
+						LSN_FORMAT_ARGS(lrRestartLsn), max_slot_wal_keep_size_mb);
+				return InvalidXLogRecPtr;
+			}
 
 			/*
 			 * start from the beginning of the segment to fetch page headers
 			 * verifed by XLogReader
 			 */
 			lrRestartLsn = lrRestartLsn - XLogSegmentOffset(lrRestartLsn, wal_segment_size);
-			wp->truncateLsn = Min(wp->truncateLsn, lrRestartLsn);
 		}
 	}
+	return lrRestartLsn;
 }
 
 static const walproposer_api walprop_pg = {
@@ -1745,18 +2018,18 @@ static const walproposer_api walprop_pg = {
 	.conn_async_write = walprop_async_write,
 	.conn_blocking_write = walprop_blocking_write,
 	.recovery_download = WalProposerRecovery,
-	.wal_read = walprop_pg_wal_read,
 	.wal_reader_allocate = walprop_pg_wal_reader_allocate,
-	.free_event_set = walprop_pg_free_event_set,
+	.wal_read = walprop_pg_wal_read,
+	.wal_reader_events = walprop_pg_wal_reader_events,
 	.init_event_set = walprop_pg_init_event_set,
 	.update_event_set = walprop_pg_update_event_set,
+	.active_state_update_event_set = walprop_pg_active_state_update_event_set,
 	.add_safekeeper_event_set = walprop_pg_add_safekeeper_event_set,
+	.rm_safekeeper_event_set = walprop_pg_rm_safekeeper_event_set,
 	.wait_event_set = walprop_pg_wait_event_set,
 	.strong_random = walprop_pg_strong_random,
 	.get_redo_start_lsn = walprop_pg_get_redo_start_lsn,
 	.finish_sync_safekeepers = walprop_pg_finish_sync_safekeepers,
 	.process_safekeeper_feedback = walprop_pg_process_safekeeper_feedback,
-	.confirm_wal_streamed = walprop_pg_confirm_wal_streamed,
 	.log_internal = walprop_pg_log_internal,
-	.after_election = walprop_pg_after_election,
 };

poetry.lock

@@ -339,19 +339,19 @@ uvloop = ["uvloop (>=0.15.2)"]
 
 [[package]]
 name = "boto3"
-version = "1.26.16"
+version = "1.34.11"
 description = "The AWS SDK for Python"
 optional = false
-python-versions = ">= 3.7"
+python-versions = ">= 3.8"
 files = [
-    {file = "boto3-1.26.16-py3-none-any.whl", hash = "sha256:4f493a2aed71cee93e626de4f67ce58dd82c0473480a0fc45b131715cd8f4f30"},
-    {file = "boto3-1.26.16.tar.gz", hash = "sha256:31c0adf71e4bd19a5428580bb229d7ea3b5795eecaa0847a85385df00c026116"},
+    {file = "boto3-1.34.11-py3-none-any.whl", hash = "sha256:1af021e0c6e3040e8de66d403e963566476235bb70f9a8e3f6784813ac2d8026"},
+    {file = "boto3-1.34.11.tar.gz", hash = "sha256:31c130a40ec0631059b77d7e87f67ad03ff1685a5b37638ac0c4687026a3259d"},
 ]
 
 [package.dependencies]
-botocore = ">=1.29.16,<1.30.0"
+botocore = ">=1.34.11,<1.35.0"
 jmespath = ">=0.7.1,<2.0.0"
-s3transfer = ">=0.6.0,<0.7.0"
+s3transfer = ">=0.10.0,<0.11.0"
 
 [package.extras]
 crt = ["botocore[crt] (>=1.21.0,<2.0a0)"]
@@ -702,22 +702,25 @@ xray = ["mypy-boto3-xray (>=1.26.0,<1.27.0)"]
 
 [[package]]
 name = "botocore"
-version = "1.29.16"
+version = "1.34.11"
 description = "Low-level, data-driven core of boto 3."
 optional = false
-python-versions = ">= 3.7"
+python-versions = ">= 3.8"
 files = [
-    {file = "botocore-1.29.16-py3-none-any.whl", hash = "sha256:271b599e6cfe214405ed50d41cd967add1d5d469383dd81ff583bc818b47f59b"},
-    {file = "botocore-1.29.16.tar.gz", hash = "sha256:8cfcc10f2f1751608c3cec694f2d6b5e16ebcd50d0a104f9914d5616227c62e9"},
+    {file = "botocore-1.34.11-py3-none-any.whl", hash = "sha256:1ff1398b6ea670e1c01ac67a33af3da854f8e700d3528289c04f319c330d8250"},
+    {file = "botocore-1.34.11.tar.gz", hash = "sha256:51905c3d623c60df5dc5794387de7caf886d350180a01a3dfa762e903edb45a9"},
 ]
 
 [package.dependencies]
 jmespath = ">=0.7.1,<2.0.0"
 python-dateutil = ">=2.1,<3.0.0"
-urllib3 = ">=1.25.4,<1.27"
+urllib3 = [
+    {version = ">=1.25.4,<1.27", markers = "python_version < \"3.10\""},
+    {version = ">=1.25.4,<2.1", markers = "python_version >= \"3.10\""},
+]
 
 [package.extras]
-crt = ["awscrt (==0.14.0)"]
+crt = ["awscrt (==0.19.19)"]
 
 [[package]]
 name = "botocore-stubs"
@@ -1889,13 +1892,13 @@ files = [
 
 [[package]]
 name = "pytest"
-version = "7.3.1"
+version = "7.4.4"
 description = "pytest: simple powerful testing with Python"
 optional = false
 python-versions = ">=3.7"
 files = [
-    {file = "pytest-7.3.1-py3-none-any.whl", hash = "sha256:3799fa815351fea3a5e96ac7e503a96fa51cc9942c3753cda7651b93c1cfa362"},
-    {file = "pytest-7.3.1.tar.gz", hash = "sha256:434afafd78b1d78ed0addf160ad2b77a30d35d4bdf8af234fe621919d9ed15e3"},
+    {file = "pytest-7.4.4-py3-none-any.whl", hash = "sha256:b090cdf5ed60bf4c45261be03239c2c1c22df034fbffe691abe93cd80cea01d8"},
+    {file = "pytest-7.4.4.tar.gz", hash = "sha256:2cf0005922c6ace4a3e2ec8b4080eb0d9753fdc93107415332f50ce9e7994280"},
 ]
 
 [package.dependencies]
@@ -1907,7 +1910,7 @@ pluggy = ">=0.12,<2.0"
 tomli = {version = ">=1.0.0", markers = "python_version < \"3.11\""}
 
 [package.extras]
-testing = ["argcomplete", "attrs (>=19.2.0)", "hypothesis (>=3.56)", "mock", "nose", "pygments (>=2.7.2)", "requests", "xmlschema"]
+testing = ["argcomplete", "attrs (>=19.2.0)", "hypothesis (>=3.56)", "mock", "nose", "pygments (>=2.7.2)", "requests", "setuptools", "xmlschema"]
 
 [[package]]
 name = "pytest-asyncio"
@@ -2230,20 +2233,20 @@ files = [
 
 [[package]]
 name = "s3transfer"
-version = "0.6.0"
+version = "0.10.0"
 description = "An Amazon S3 Transfer Manager"
 optional = false
-python-versions = ">= 3.7"
+python-versions = ">= 3.8"
 files = [
-    {file = "s3transfer-0.6.0-py3-none-any.whl", hash = "sha256:06176b74f3a15f61f1b4f25a1fc29a4429040b7647133a463da8fa5bd28d5ecd"},
-    {file = "s3transfer-0.6.0.tar.gz", hash = "sha256:2ed07d3866f523cc561bf4a00fc5535827981b117dd7876f036b0c1aca42c947"},
+    {file = "s3transfer-0.10.0-py3-none-any.whl", hash = "sha256:3cdb40f5cfa6966e812209d0994f2a4709b561c88e90cf00c2696d2df4e56b2e"},
+    {file = "s3transfer-0.10.0.tar.gz", hash = "sha256:d0c8bbf672d5eebbe4e57945e23b972d963f07d82f661cabf678a5c88831595b"},
 ]
 
 [package.dependencies]
-botocore = ">=1.12.36,<2.0a.0"
+botocore = ">=1.33.2,<2.0a.0"
 
 [package.extras]
-crt = ["botocore[crt] (>=1.20.29,<2.0a.0)"]
+crt = ["botocore[crt] (>=1.33.2,<2.0a.0)"]
 
 [[package]]
 name = "sarif-om"
@@ -2740,4 +2743,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.9"
-content-hash = "c4e38082d246636903e15c02fbf8364c6afc1fd35d36a81c49f596ba68fc739b"
+content-hash = "8de8b05a9b35a6f76da7d7e3652ddbb521f1eca53fce7b933f537080a9d6eada"

proxy/src/bin/proxy.rs

@@ -11,6 +11,7 @@ use proxy::http;
 use proxy::rate_limiter::EndpointRateLimiter;
 use proxy::rate_limiter::RateBucketInfo;
 use proxy::rate_limiter::RateLimiterConfig;
+use proxy::serverless::GlobalConnPoolOptions;
 use proxy::usage_metrics;
 
 use anyhow::bail;
@@ -95,12 +96,8 @@ struct ProxyCliArgs {
     /// Allow self-signed certificates for compute nodes (for testing)
     #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
     allow_self_signed_compute: bool,
-    /// timeout for http connections
-    #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
-    sql_over_http_timeout: tokio::time::Duration,
-    /// Whether the SQL over http pool is opt-in
-    #[clap(long, default_value_t = true, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
-    sql_over_http_pool_opt_in: bool,
+    #[clap(flatten)]
+    sql_over_http: SqlOverHttpArgs,
     /// timeout for scram authentication protocol
     #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
     scram_protocol_timeout: tokio::time::Duration,
@@ -138,6 +135,36 @@ struct ProxyCliArgs {
     disable_ip_check_for_http: bool,
 }
 
+#[derive(clap::Args, Clone, Copy, Debug)]
+struct SqlOverHttpArgs {
+    /// timeout for http connection requests
+    #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
+    sql_over_http_timeout: tokio::time::Duration,
+
+    /// Whether the SQL over http pool is opt-in
+    #[clap(long, default_value_t = true, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
+    sql_over_http_pool_opt_in: bool,
+
+    /// How many connections to pool for each endpoint. Excess connections are discarded
+    #[clap(long, default_value_t = 20)]
+    sql_over_http_pool_max_conns_per_endpoint: usize,
+
+    /// How long pooled connections should remain idle for before closing
+    #[clap(long, default_value = "5m", value_parser = humantime::parse_duration)]
+    sql_over_http_idle_timeout: tokio::time::Duration,
+
+    /// Duration each shard will wait on average before a GC sweep.
+    /// A longer time will causes sweeps to take longer but will interfere less frequently.
+    #[clap(long, default_value = "10m", value_parser = humantime::parse_duration)]
+    sql_over_http_pool_gc_epoch: tokio::time::Duration,
+
+    /// How many shards should the global pool have. Must be a power of two.
+    /// More shards will introduce less contention for pool operations, but can
+    /// increase memory used by the pool
+    #[clap(long, default_value_t = 128)]
+    sql_over_http_pool_shards: usize,
+}
+
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
     let _logging_guard = proxy::logging::init().await?;
@@ -327,8 +354,14 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
         }
     };
     let http_config = HttpConfig {
-        timeout: args.sql_over_http_timeout,
-        pool_opt_in: args.sql_over_http_pool_opt_in,
+        request_timeout: args.sql_over_http.sql_over_http_timeout,
+        pool_options: GlobalConnPoolOptions {
+            max_conns_per_endpoint: args.sql_over_http.sql_over_http_pool_max_conns_per_endpoint,
+            gc_epoch: args.sql_over_http.sql_over_http_pool_gc_epoch,
+            pool_shards: args.sql_over_http.sql_over_http_pool_shards,
+            idle_timeout: args.sql_over_http.sql_over_http_idle_timeout,
+            opt_in: args.sql_over_http.sql_over_http_pool_opt_in,
+        },
     };
     let authentication_config = AuthenticationConfig {
         scram_protocol_timeout: args.scram_protocol_timeout,

proxy/src/config.rs

@@ -1,4 +1,4 @@
-use crate::{auth, rate_limiter::RateBucketInfo};
+use crate::{auth, rate_limiter::RateBucketInfo, serverless::GlobalConnPoolOptions};
 use anyhow::{bail, ensure, Context, Ok};
 use rustls::{sign, Certificate, PrivateKey};
 use sha2::{Digest, Sha256};
@@ -36,8 +36,8 @@ pub struct TlsConfig {
 }
 
 pub struct HttpConfig {
-    pub timeout: tokio::time::Duration,
-    pub pool_opt_in: bool,
+    pub request_timeout: tokio::time::Duration,
+    pub pool_options: GlobalConnPoolOptions,
 }
 
 pub struct AuthenticationConfig {

proxy/src/console/provider/neon.rs

@@ -11,7 +11,7 @@ use crate::{auth::backend::ComputeUserInfo, compute, http, scram};
 use async_trait::async_trait;
 use futures::TryFutureExt;
 use itertools::Itertools;
-use std::{net::SocketAddr, sync::Arc};
+use std::sync::Arc;
 use tokio::time::Instant;
 use tokio_postgres::config::SslMode;
 use tracing::{error, info, info_span, warn, Instrument};
@@ -141,7 +141,7 @@ impl Api {
             // We'll set username and such later using the startup message.
             // TODO: add more type safety (in progress).
             let mut config = compute::ConnCfg::new();
-            config.host(&host).port(port).ssl_mode(SslMode::Disable); // TLS is not configured on compute nodes.
+            config.host(host).port(port).ssl_mode(SslMode::Disable); // TLS is not configured on compute nodes.
 
             let node = NodeInfo {
                 config,
@@ -269,9 +269,10 @@ async fn parse_body<T: for<'a> serde::Deserialize<'a>>(
     Err(ApiError::Console { status, text })
 }
 
-fn parse_host_port(input: &str) -> Option<(String, u16)> {
-    let parsed: SocketAddr = input.parse().ok()?;
-    Some((parsed.ip().to_string(), parsed.port()))
+fn parse_host_port(input: &str) -> Option<(&str, u16)> {
+    let (host, port) = input.rsplit_once(':')?;
+    let ipv6_brackets: &[_] = &['[', ']'];
+    Some((host.trim_matches(ipv6_brackets), port.parse().ok()?))
 }
 
 #[cfg(test)]
@@ -279,9 +280,24 @@ mod tests {
     use super::*;
 
     #[test]
-    fn test_parse_host_port() {
+    fn test_parse_host_port_v4() {
         let (host, port) = parse_host_port("127.0.0.1:5432").expect("failed to parse");
         assert_eq!(host, "127.0.0.1");
         assert_eq!(port, 5432);
     }
+
+    #[test]
+    fn test_parse_host_port_v6() {
+        let (host, port) = parse_host_port("[2001:db8::1]:5432").expect("failed to parse");
+        assert_eq!(host, "2001:db8::1");
+        assert_eq!(port, 5432);
+    }
+
+    #[test]
+    fn test_parse_host_port_url() {
+        let (host, port) = parse_host_port("compute-foo-bar-1234.default.svc.cluster.local:5432")
+            .expect("failed to parse");
+        assert_eq!(host, "compute-foo-bar-1234.default.svc.cluster.local");
+        assert_eq!(port, 5432);
+    }
 }

proxy/src/serverless.rs

@@ -6,9 +6,13 @@ mod conn_pool;
 mod sql_over_http;
 mod websocket;
 
+pub use conn_pool::GlobalConnPoolOptions;
+
 use anyhow::bail;
 use hyper::StatusCode;
 use metrics::IntCounterPairGuard;
+use rand::rngs::StdRng;
+use rand::SeedableRng;
 pub use reqwest_middleware::{ClientWithMiddleware, Error};
 pub use reqwest_retry::{policies::ExponentialBackoff, RetryTransientMiddleware};
 use tokio_util::task::TaskTracker;
@@ -47,6 +51,11 @@ pub async fn task_main(
 
     let conn_pool = conn_pool::GlobalConnPool::new(config);
 
+    let conn_pool2 = Arc::clone(&conn_pool);
+    tokio::spawn(async move {
+        conn_pool2.gc_worker(StdRng::from_entropy()).await;
+    });
+
     // shutdown the connection pool
     tokio::spawn({
         let cancellation_token = cancellation_token.clone();

proxy/src/serverless/conn_pool.rs

@@ -1,15 +1,19 @@
 use anyhow::{anyhow, Context};
 use async_trait::async_trait;
 use dashmap::DashMap;
-use futures::future::poll_fn;
+use futures::{future::poll_fn, Future};
+use metrics::{register_int_counter_pair, IntCounterPair, IntCounterPairGuard};
+use once_cell::sync::Lazy;
 use parking_lot::RwLock;
 use pbkdf2::{
     password_hash::{PasswordHashString, PasswordHasher, PasswordVerifier, SaltString},
     Params, Pbkdf2,
 };
 use pq_proto::StartupMessageParams;
+use prometheus::{exponential_buckets, register_histogram, Histogram};
+use rand::Rng;
 use smol_str::SmolStr;
-use std::{collections::HashMap, net::IpAddr, sync::Arc};
+use std::{collections::HashMap, net::IpAddr, pin::pin, sync::Arc, sync::Weak, time::Duration};
 use std::{
     fmt,
     task::{ready, Poll},
@@ -18,7 +22,7 @@ use std::{
     ops::Deref,
     sync::atomic::{self, AtomicUsize},
 };
-use tokio::time;
+use tokio::time::{self, Instant};
 use tokio_postgres::{AsyncMessage, ReadyForQueryStatus};
 
 use crate::{
@@ -30,11 +34,10 @@ use crate::{
 };
 use crate::{compute, config};
 
-use tracing::{error, warn, Span};
+use tracing::{debug, error, warn, Span};
 use tracing::{info, info_span, Instrument};
 
 pub const APP_NAME: &str = "/sql_over_http";
-const MAX_CONNS_PER_ENDPOINT: usize = 20;
 
 #[derive(Debug, Clone)]
 pub struct ConnInfo {
@@ -69,6 +72,77 @@ struct ConnPoolEntry {
 pub struct EndpointConnPool {
     pools: HashMap<(SmolStr, SmolStr), DbUserConnPool>,
     total_conns: usize,
+    max_conns: usize,
+    _guard: IntCounterPairGuard,
+}
+
+impl EndpointConnPool {
+    fn get_conn_entry(&mut self, db_user: (SmolStr, SmolStr)) -> Option<ConnPoolEntry> {
+        let Self {
+            pools, total_conns, ..
+        } = self;
+        pools
+            .get_mut(&db_user)
+            .and_then(|pool_entries| pool_entries.get_conn_entry(total_conns))
+    }
+
+    fn remove_client(&mut self, db_user: (SmolStr, SmolStr), conn_id: uuid::Uuid) -> bool {
+        let Self {
+            pools, total_conns, ..
+        } = self;
+        if let Some(pool) = pools.get_mut(&db_user) {
+            let old_len = pool.conns.len();
+            pool.conns.retain(|conn| conn.conn.conn_id != conn_id);
+            let new_len = pool.conns.len();
+            let removed = old_len - new_len;
+            *total_conns -= removed;
+            removed > 0
+        } else {
+            false
+        }
+    }
+
+    fn put(pool: &RwLock<Self>, conn_info: &ConnInfo, client: ClientInner) -> anyhow::Result<()> {
+        let conn_id = client.conn_id;
+
+        if client.inner.is_closed() {
+            info!(%conn_id, "pool: throwing away connection '{conn_info}' because connection is closed");
+            return Ok(());
+        }
+
+        // return connection to the pool
+        let mut returned = false;
+        let mut per_db_size = 0;
+        let total_conns = {
+            let mut pool = pool.write();
+
+            if pool.total_conns < pool.max_conns {
+                // we create this db-user entry in get, so it should not be None
+                if let Some(pool_entries) = pool.pools.get_mut(&conn_info.db_and_user()) {
+                    pool_entries.conns.push(ConnPoolEntry {
+                        conn: client,
+                        _last_access: std::time::Instant::now(),
+                    });
+
+                    returned = true;
+                    per_db_size = pool_entries.conns.len();
+
+                    pool.total_conns += 1;
+                }
+            }
+
+            pool.total_conns
+        };
+
+        // do logging outside of the mutex
+        if returned {
+            info!(%conn_id, "pool: returning connection '{conn_info}' back to the pool, total_conns={total_conns}, for this (db, user)={per_db_size}");
+        } else {
+            info!(%conn_id, "pool: throwing away connection '{conn_info}' because pool is full, total_conns={total_conns}");
+        }
+
+        Ok(())
+    }
 }
 
 /// 4096 is the number of rounds that SCRAM-SHA-256 recommends.
@@ -87,6 +161,27 @@ pub struct DbUserConnPool {
     password_hash: Option<PasswordHashString>,
 }
 
+impl DbUserConnPool {
+    fn clear_closed_clients(&mut self, conns: &mut usize) {
+        let old_len = self.conns.len();
+
+        self.conns.retain(|conn| !conn.conn.inner.is_closed());
+
+        let new_len = self.conns.len();
+        let removed = old_len - new_len;
+        *conns -= removed;
+    }
+
+    fn get_conn_entry(&mut self, conns: &mut usize) -> Option<ConnPoolEntry> {
+        self.clear_closed_clients(conns);
+        let conn = self.conns.pop();
+        if conn.is_some() {
+            *conns -= 1;
+        }
+        conn
+    }
+}
+
 pub struct GlobalConnPool {
     // endpoint -> per-endpoint connection pool
     //
@@ -94,52 +189,127 @@ pub struct GlobalConnPool {
     // pool as early as possible and release the lock.
     global_pool: DashMap<SmolStr, Arc<RwLock<EndpointConnPool>>>,
 
+    /// Number of endpoint-connection pools
+    ///
     /// [`DashMap::len`] iterates over all inner pools and acquires a read lock on each.
     /// That seems like far too much effort, so we're using a relaxed increment counter instead.
     /// It's only used for diagnostics.
     global_pool_size: AtomicUsize,
 
+    proxy_config: &'static crate::config::ProxyConfig,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct GlobalConnPoolOptions {
     // Maximum number of connections per one endpoint.
     // Can mix different (dbname, username) connections.
     // When running out of free slots for a particular endpoint,
     // falls back to opening a new connection for each request.
-    max_conns_per_endpoint: usize,
+    pub max_conns_per_endpoint: usize,
 
-    proxy_config: &'static crate::config::ProxyConfig,
+    pub gc_epoch: Duration,
 
-    // Using a lock to remove any race conditions.
-    // Eg cleaning up connections while a new connection is returned
-    closed: RwLock<bool>,
+    pub pool_shards: usize,
+
+    pub idle_timeout: Duration,
+
+    pub opt_in: bool,
 }
 
+pub static GC_LATENCY: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "proxy_http_pool_reclaimation_lag_seconds",
+        "Time it takes to reclaim unused connection pools",
+        // 1us -> 65ms
+        exponential_buckets(1e-6, 2.0, 16).unwrap(),
+    )
+    .unwrap()
+});
+
+pub static ENDPOINT_POOLS: Lazy<IntCounterPair> = Lazy::new(|| {
+    register_int_counter_pair!(
+        "proxy_http_pool_endpoints_registered_total",
+        "Number of endpoints we have registered pools for",
+        "proxy_http_pool_endpoints_unregistered_total",
+        "Number of endpoints we have unregistered pools for",
+    )
+    .unwrap()
+});
+
 impl GlobalConnPool {
     pub fn new(config: &'static crate::config::ProxyConfig) -> Arc<Self> {
+        let shards = config.http_config.pool_options.pool_shards;
         Arc::new(Self {
-            global_pool: DashMap::new(),
+            global_pool: DashMap::with_shard_amount(shards),
             global_pool_size: AtomicUsize::new(0),
-            max_conns_per_endpoint: MAX_CONNS_PER_ENDPOINT,
             proxy_config: config,
-            closed: RwLock::new(false),
         })
     }
 
     pub fn shutdown(&self) {
-        *self.closed.write() = true;
+        // drops all strong references to endpoint-pools
+        self.global_pool.clear();
+    }
 
-        self.global_pool.retain(|_, endpoint_pool| {
-            let mut pool = endpoint_pool.write();
-            // by clearing this hashmap, we remove the slots that a connection can be returned to.
-            // when returning, it drops the connection if the slot doesn't exist
-            pool.pools.clear();
-            pool.total_conns = 0;
+    pub async fn gc_worker(&self, mut rng: impl Rng) {
+        let epoch = self.proxy_config.http_config.pool_options.gc_epoch;
+        let mut interval = tokio::time::interval(epoch / (self.global_pool.shards().len()) as u32);
+        loop {
+            interval.tick().await;
 
-            false
+            let shard = rng.gen_range(0..self.global_pool.shards().len());
+            self.gc(shard);
+        }
+    }
+
+    fn gc(&self, shard: usize) {
+        debug!(shard, "pool: performing epoch reclamation");
+
+        // acquire a random shard lock
+        let mut shard = self.global_pool.shards()[shard].write();
+
+        let timer = GC_LATENCY.start_timer();
+        let current_len = shard.len();
+        shard.retain(|endpoint, x| {
+            // if the current endpoint pool is unique (no other strong or weak references)
+            // then it is currently not in use by any connections.
+            if let Some(pool) = Arc::get_mut(x.get_mut()) {
+                let EndpointConnPool {
+                    pools, total_conns, ..
+                } = pool.get_mut();
+
+                // ensure that closed clients are removed
+                pools
+                    .iter_mut()
+                    .for_each(|(_, db_pool)| db_pool.clear_closed_clients(total_conns));
+
+                // we only remove this pool if it has no active connections
+                if *total_conns == 0 {
+                    info!("pool: discarding pool for endpoint {endpoint}");
+                    return false;
+                }
+            }
+
+            true
         });
+        let new_len = shard.len();
+        drop(shard);
+        timer.observe_duration();
+
+        let removed = current_len - new_len;
+
+        if removed > 0 {
+            let global_pool_size = self
+                .global_pool_size
+                .fetch_sub(removed, atomic::Ordering::Relaxed)
+                - removed;
+            info!("pool: performed global pool gc. size now {global_pool_size}");
+        }
     }
 
     pub async fn get(
         self: &Arc<Self>,
-        conn_info: &ConnInfo,
+        conn_info: ConnInfo,
         force_new: bool,
         session_id: uuid::Uuid,
         peer_addr: IpAddr,
@@ -147,15 +317,11 @@ impl GlobalConnPool {
         let mut client: Option<ClientInner> = None;
         let mut latency_timer = LatencyTimer::new("http");
 
-        let pool = if force_new {
-            None
-        } else {
-            Some((conn_info.clone(), self.clone()))
-        };
-
         let mut hash_valid = false;
+        let mut endpoint_pool = Weak::new();
         if !force_new {
             let pool = self.get_or_create_endpoint_pool(&conn_info.hostname);
+            endpoint_pool = Arc::downgrade(&pool);
             let mut hash = None;
 
             // find a pool entry by (dbname, username) if exists
@@ -180,12 +346,8 @@ impl GlobalConnPool {
                 // we will continue with the regular connection flow
                 if validate.is_ok() {
                     hash_valid = true;
-                    let mut pool = pool.write();
-                    if let Some(pool_entries) = pool.pools.get_mut(&conn_info.db_and_user()) {
-                        if let Some(entry) = pool_entries.conns.pop() {
-                            client = Some(entry.conn);
-                            pool.total_conns -= 1;
-                        }
+                    if let Some(entry) = pool.write().get_conn_entry(conn_info.db_and_user()) {
+                        client = Some(entry.conn)
                     }
                 }
             }
@@ -198,11 +360,12 @@ impl GlobalConnPool {
                 info!(%conn_id, "pool: cached connection '{conn_info}' is closed, opening a new one");
                 connect_to_compute(
                     self.proxy_config,
-                    conn_info,
+                    &conn_info,
                     conn_id,
                     session_id,
                     latency_timer,
                     peer_addr,
+                    endpoint_pool.clone(),
                 )
                 .await
             } else {
@@ -214,18 +377,19 @@ impl GlobalConnPool {
                 );
                 latency_timer.pool_hit();
                 latency_timer.success();
-                return Ok(Client::new(client, pool).await);
+                return Ok(Client::new(client, conn_info, endpoint_pool).await);
             }
         } else {
             let conn_id = uuid::Uuid::new_v4();
             info!(%conn_id, "pool: opening a new connection '{conn_info}'");
             connect_to_compute(
                 self.proxy_config,
-                conn_info,
+                &conn_info,
                 conn_id,
                 session_id,
                 latency_timer,
                 peer_addr,
+                endpoint_pool.clone(),
             )
             .await
         };
@@ -269,59 +433,7 @@ impl GlobalConnPool {
             _ => {}
         }
         let new_client = new_client?;
-        Ok(Client::new(new_client, pool).await)
-    }
-
-    fn put(&self, conn_info: &ConnInfo, client: ClientInner) -> anyhow::Result<()> {
-        let conn_id = client.conn_id;
-
-        // We want to hold this open while we return. This ensures that the pool can't close
-        // while we are in the middle of returning the connection.
-        let closed = self.closed.read();
-        if *closed {
-            info!(%conn_id, "pool: throwing away connection '{conn_info}' because pool is closed");
-            return Ok(());
-        }
-
-        if client.inner.is_closed() {
-            info!(%conn_id, "pool: throwing away connection '{conn_info}' because connection is closed");
-            return Ok(());
-        }
-
-        let pool = self.get_or_create_endpoint_pool(&conn_info.hostname);
-
-        // return connection to the pool
-        let mut returned = false;
-        let mut per_db_size = 0;
-        let total_conns = {
-            let mut pool = pool.write();
-
-            if pool.total_conns < self.max_conns_per_endpoint {
-                // we create this db-user entry in get, so it should not be None
-                if let Some(pool_entries) = pool.pools.get_mut(&conn_info.db_and_user()) {
-                    pool_entries.conns.push(ConnPoolEntry {
-                        conn: client,
-                        _last_access: std::time::Instant::now(),
-                    });
-
-                    returned = true;
-                    per_db_size = pool_entries.conns.len();
-
-                    pool.total_conns += 1;
-                }
-            }
-
-            pool.total_conns
-        };
-
-        // do logging outside of the mutex
-        if returned {
-            info!(%conn_id, "pool: returning connection '{conn_info}' back to the pool, total_conns={total_conns}, for this (db, user)={per_db_size}");
-        } else {
-            info!(%conn_id, "pool: throwing away connection '{conn_info}' because pool is full, total_conns={total_conns}");
-        }
-
-        Ok(())
+        Ok(Client::new(new_client, conn_info, endpoint_pool).await)
     }
 
     fn get_or_create_endpoint_pool(&self, endpoint: &SmolStr) -> Arc<RwLock<EndpointConnPool>> {
@@ -334,6 +446,12 @@ impl GlobalConnPool {
         let new_pool = Arc::new(RwLock::new(EndpointConnPool {
             pools: HashMap::new(),
             total_conns: 0,
+            max_conns: self
+                .proxy_config
+                .http_config
+                .pool_options
+                .max_conns_per_endpoint,
+            _guard: ENDPOINT_POOLS.guard(),
         }));
 
         // find or create a pool for this endpoint
@@ -363,9 +481,11 @@ impl GlobalConnPool {
 }
 
 struct TokioMechanism<'a> {
+    pool: Weak<RwLock<EndpointConnPool>>,
     conn_info: &'a ConnInfo,
     session_id: uuid::Uuid,
     conn_id: uuid::Uuid,
+    idle: Duration,
 }
 
 #[async_trait]
@@ -385,6 +505,8 @@ impl ConnectMechanism for TokioMechanism<'_> {
             timeout,
             self.conn_id,
             self.session_id,
+            self.pool.clone(),
+            self.idle,
         )
         .await
     }
@@ -403,6 +525,7 @@ async fn connect_to_compute(
     session_id: uuid::Uuid,
     latency_timer: LatencyTimer,
     peer_addr: IpAddr,
+    pool: Weak<RwLock<EndpointConnPool>>,
 ) -> anyhow::Result<ClientInner> {
     let tls = config.tls_config.as_ref();
     let common_names = tls.and_then(|tls| tls.common_names.clone());
@@ -447,6 +570,8 @@ async fn connect_to_compute(
             conn_id,
             conn_info,
             session_id,
+            pool,
+            idle: config.http_config.pool_options.idle_timeout,
         },
         node_info,
         &extra,
@@ -462,6 +587,8 @@ async fn connect_to_compute_once(
     timeout: time::Duration,
     conn_id: uuid::Uuid,
     mut session: uuid::Uuid,
+    pool: Weak<RwLock<EndpointConnPool>>,
+    idle: Duration,
 ) -> Result<ClientInner, tokio_postgres::Error> {
     let mut config = (*node_info.config).clone();
 
@@ -490,13 +617,29 @@ async fn connect_to_compute_once(
         branch_id: node_info.aux.branch_id.clone(),
     };
 
+    let db_user = conn_info.db_and_user();
     tokio::spawn(
         async move {
             let _conn_gauge = conn_gauge;
+            let mut idle_timeout = pin!(tokio::time::sleep(idle));
             poll_fn(move |cx| {
                 if matches!(rx.has_changed(), Ok(true)) {
                     session = *rx.borrow_and_update();
                     info!(%session, "changed session");
+                    idle_timeout.as_mut().reset(Instant::now() + idle);
+                }
+
+                // 5 minute idle connection timeout
+                if idle_timeout.as_mut().poll(cx).is_ready() {
+                    idle_timeout.as_mut().reset(Instant::now() + idle);
+                    info!("connection idle");
+                    if let Some(pool) = pool.clone().upgrade() {
+                        // remove client from pool - should close the connection if it's idle.
+                        // does nothing if the client is currently checked-out and in-use
+                        if pool.write().remove_client(db_user.clone(), conn_id) {
+                            info!("idle connection removed");
+                        }
+                    }
                 }
 
                 loop {
@@ -514,15 +657,25 @@ async fn connect_to_compute_once(
                         }
                         Some(Err(e)) => {
                             error!(%session, "connection error: {}", e);
-                            return Poll::Ready(())
+                            break
                         }
                         None => {
                             info!("connection closed");
-                            return Poll::Ready(())
+                            break
                         }
                     }
                 }
-            }).await
+
+                // remove from connection pool
+                if let Some(pool) = pool.clone().upgrade() {
+                    if pool.write().remove_client(db_user.clone(), conn_id) {
+                        info!("closed connection removed");
+                    }
+                }
+
+                Poll::Ready(())
+            }).await;
+
         }
         .instrument(span)
     );
@@ -552,23 +705,27 @@ pub struct Client {
     conn_id: uuid::Uuid,
     span: Span,
     inner: Option<ClientInner>,
-    pool: Option<(ConnInfo, Arc<GlobalConnPool>)>,
+    conn_info: ConnInfo,
+    pool: Weak<RwLock<EndpointConnPool>>,
 }
 
 pub struct Discard<'a> {
     conn_id: uuid::Uuid,
-    pool: &'a mut Option<(ConnInfo, Arc<GlobalConnPool>)>,
+    conn_info: &'a ConnInfo,
+    pool: &'a mut Weak<RwLock<EndpointConnPool>>,
 }
 
 impl Client {
     pub(self) async fn new(
         inner: ClientInner,
-        pool: Option<(ConnInfo, Arc<GlobalConnPool>)>,
+        conn_info: ConnInfo,
+        pool: Weak<RwLock<EndpointConnPool>>,
     ) -> Self {
         Self {
             conn_id: inner.conn_id,
             inner: Some(inner),
             span: Span::current(),
+            conn_info,
             pool,
         }
     }
@@ -577,6 +734,7 @@ impl Client {
             inner,
             pool,
             conn_id,
+            conn_info,
             span: _,
         } = self;
         (
@@ -586,6 +744,7 @@ impl Client {
                 .inner,
             Discard {
                 pool,
+                conn_info,
                 conn_id: *conn_id,
             },
         )
@@ -601,14 +760,14 @@ impl Client {
 
 impl Discard<'_> {
     pub fn check_idle(&mut self, status: ReadyForQueryStatus) {
-        if status != ReadyForQueryStatus::Idle {
-            if let Some((conn_info, _)) = self.pool.take() {
-                info!(conn_id = %self.conn_id, "pool: throwing away connection '{conn_info}' because connection is not idle")
-            }
+        let conn_info = &self.conn_info;
+        if status != ReadyForQueryStatus::Idle && std::mem::take(self.pool).strong_count() > 0 {
+            info!(conn_id = %self.conn_id, "pool: throwing away connection '{conn_info}' because connection is not idle")
         }
     }
     pub fn discard(&mut self) {
-        if let Some((conn_info, _)) = self.pool.take() {
+        let conn_info = &self.conn_info;
+        if std::mem::take(self.pool).strong_count() > 0 {
             info!(conn_id = %self.conn_id, "pool: throwing away connection '{conn_info}' because connection is potentially in a broken state")
         }
     }
@@ -628,16 +787,17 @@ impl Deref for Client {
 
 impl Drop for Client {
     fn drop(&mut self) {
+        let conn_info = self.conn_info.clone();
         let client = self
             .inner
             .take()
             .expect("client inner should not be removed");
-        if let Some((conn_info, conn_pool)) = self.pool.take() {
+        if let Some(conn_pool) = std::mem::take(&mut self.pool).upgrade() {
             let current_span = self.span.clone();
             // return connection to the pool
             tokio::task::spawn_blocking(move || {
                 let _span = current_span.enter();
-                let _ = conn_pool.put(&conn_info, client);
+                let _ = EndpointConnPool::put(&conn_pool, &conn_info, client);
             });
         }
     }

proxy/src/serverless/sql_over_http.rs

@@ -206,7 +206,7 @@ pub async fn handle(
     config: &'static HttpConfig,
 ) -> Result<Response<Body>, ApiError> {
     let result = tokio::time::timeout(
-        config.timeout,
+        config.request_timeout,
         handle_inner(
             config,
             request,
@@ -278,7 +278,7 @@ pub async fn handle(
         Err(_) => {
             let message = format!(
                 "HTTP-Connection timed out, execution time exeeded {} seconds",
-                config.timeout.as_secs()
+                config.request_timeout.as_secs()
             );
             error!(message);
             json_response(
@@ -320,7 +320,8 @@ async fn handle_inner(
 
     // Allow connection pooling only if explicitly requested
     // or if we have decided that http pool is no longer opt-in
-    let allow_pool = !config.pool_opt_in || headers.get(&ALLOW_POOL) == Some(&HEADER_VALUE_TRUE);
+    let allow_pool =
+        !config.pool_options.opt_in || headers.get(&ALLOW_POOL) == Some(&HEADER_VALUE_TRUE);
 
     // isolation level, read only and deferrable
 
@@ -359,7 +360,7 @@ async fn handle_inner(
     let payload: Payload = serde_json::from_slice(&body)?;
 
     let mut client = conn_pool
-        .get(&conn_info, !allow_pool, session_id, peer_addr)
+        .get(conn_info, !allow_pool, session_id, peer_addr)
         .await?;
 
     let mut response = Response::builder()

pyproject.toml

@@ -6,7 +6,7 @@ authors = []
 
 [tool.poetry.dependencies]
 python = "^3.9"
-pytest = "^7.3.1"
+pytest = "^7.4.4"
 psycopg2-binary = "^2.9.6"
 typing-extensions = "^4.6.1"
 PyJWT = {version = "^2.1.0", extras = ["crypto"]}
@@ -17,7 +17,7 @@ aiopg = "^1.4.0"
 Jinja2 = "^3.0.2"
 types-requests = "^2.31.0.0"
 types-psycopg2 = "^2.9.21.10"
-boto3 = "^1.26.16"
+boto3 = "^1.34.11"
 boto3-stubs = {extras = ["s3"], version = "^1.26.16"}
 moto = {extras = ["server"], version = "^4.1.2"}
 backoff = "^2.2.1"

s3_scrubber/Cargo.toml

@@ -6,6 +6,7 @@ license.workspace = true
 
 [dependencies]
 aws-sdk-s3.workspace = true
+aws-smithy-async.workspace = true
 either.workspace = true
 tokio-rustls.workspace = true
 anyhow.workspace = true
@@ -39,3 +40,5 @@ tracing-subscriber.workspace = true
 clap.workspace = true
 tracing-appender = "0.2"
 histogram = "0.7"
+
+futures.workspace = true

s3_scrubber/src/checks.rs

@@ -1,9 +1,12 @@
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 
 use anyhow::Context;
 use aws_sdk_s3::{types::ObjectIdentifier, Client};
+use pageserver::tenant::remote_timeline_client::index::IndexLayerMetadata;
+use pageserver_api::shard::ShardIndex;
 use tracing::{error, info, warn};
 use utils::generation::Generation;
+use utils::id::TimelineId;
 
 use crate::cloud_admin_api::BranchData;
 use crate::metadata_stream::stream_listing;
@@ -40,7 +43,7 @@ impl TimelineAnalysis {
 
 pub(crate) fn branch_cleanup_and_check_errors(
     id: &TenantShardTimelineId,
-    s3_root: &RootTarget,
+    tenant_objects: &mut TenantObjectListing,
     s3_active_branch: Option<&BranchData>,
     console_branch: Option<BranchData>,
     s3_data: Option<S3TimelineBlobData>,
@@ -72,8 +75,8 @@ pub(crate) fn branch_cleanup_and_check_errors(
             match s3_data.blob_data {
                 BlobDataParseResult::Parsed {
                     index_part,
-                    index_part_generation,
-                    mut s3_layers,
+                    index_part_generation: _index_part_generation,
+                    s3_layers: _s3_layers,
                 } => {
                     if !IndexPart::KNOWN_VERSIONS.contains(&index_part.get_version()) {
                         result.errors.push(format!(
@@ -111,65 +114,19 @@ pub(crate) fn branch_cleanup_and_check_errors(
                             ))
                         }
 
-                        let layer_map_key = (layer, metadata.generation);
-                        if !s3_layers.remove(&layer_map_key) {
+                        if !tenant_objects.check_ref(id.timeline_id, &layer, &metadata) {
                             // FIXME: this will emit false positives if an index was
                             // uploaded concurrently with our scan.  To make this check
                             // correct, we need to try sending a HEAD request for the
                             // layer we think is missing.
                             result.errors.push(format!(
-                                "index_part.json contains a layer {}{} that is not present in remote storage",
-                                layer_map_key.0.file_name(),
-                                layer_map_key.1.get_suffix()
+                                "index_part.json contains a layer {}{} (shard {}) that is not present in remote storage",
+                                layer.file_name(),
+                                metadata.generation.get_suffix(),
+                                metadata.shard
                             ))
                         }
                     }
-
-                    let orphan_layers: Vec<(LayerFileName, Generation)> = s3_layers
-                        .into_iter()
-                        .filter(|(_layer_name, gen)|
-                            // A layer is only considered orphaned if it has a generation below
-                            // the index.  If the generation is >= the index, then the layer may
-                            // be an upload from a running pageserver, or even an upload from
-                            // a new generation that didn't upload an index yet.
-                            //
-                            // Even so, a layer that is not referenced by the index could just
-                            // be something enqueued for deletion, so while this check is valid
-                            // for indicating that a layer is garbage, it is not an indicator
-                            // of a problem.
-                            gen < &index_part_generation)
-                        .collect();
-
-                    if !orphan_layers.is_empty() {
-                        // An orphan layer is not an error: it's arguably not even a warning, but it is helpful to report
-                        // these as a hint that there is something worth cleaning up here.
-                        result.warnings.push(format!(
-                            "index_part.json does not contain layers from S3: {:?}",
-                            orphan_layers
-                                .iter()
-                                .map(|(layer_name, gen)| format!(
-                                    "{}{}",
-                                    layer_name.file_name(),
-                                    gen.get_suffix()
-                                ))
-                                .collect::<Vec<_>>(),
-                        ));
-                        result.garbage_keys.extend(orphan_layers.iter().map(
-                            |(layer_name, layer_gen)| {
-                                let mut key = s3_root.timeline_root(id).prefix_in_bucket;
-                                let delimiter = s3_root.delimiter();
-                                if !key.ends_with(delimiter) {
-                                    key.push_str(delimiter);
-                                }
-                                key.push_str(&format!(
-                                    "{}{}",
-                                    &layer_name.file_name(),
-                                    layer_gen.get_suffix()
-                                ));
-                                key
-                            },
-                        ));
-                    }
                 }
                 BlobDataParseResult::Relic => {}
                 BlobDataParseResult::Incorrect(parse_errors) => result.errors.extend(
@@ -204,6 +161,83 @@ pub(crate) fn branch_cleanup_and_check_errors(
     result
 }
 
+#[derive(Default)]
+pub(crate) struct LayerRef {
+    ref_count: usize,
+}
+
+/// Top-level index of objects in a tenant.  This may be used by any shard-timeline within
+/// the tenant to query whether an object exists.
+#[derive(Default)]
+pub(crate) struct TenantObjectListing {
+    shard_timelines:
+        HashMap<(ShardIndex, TimelineId), HashMap<(LayerFileName, Generation), LayerRef>>,
+}
+
+impl TenantObjectListing {
+    /// Having done an S3 listing of the keys within a timeline prefix, merge them into the overall
+    /// list of layer keys for the Tenant.
+    pub(crate) fn push(
+        &mut self,
+        ttid: TenantShardTimelineId,
+        layers: HashSet<(LayerFileName, Generation)>,
+    ) {
+        let shard_index = ShardIndex::new(
+            ttid.tenant_shard_id.shard_number,
+            ttid.tenant_shard_id.shard_count,
+        );
+        let replaced = self.shard_timelines.insert(
+            (shard_index, ttid.timeline_id),
+            layers
+                .into_iter()
+                .map(|l| (l, LayerRef::default()))
+                .collect(),
+        );
+
+        assert!(
+            replaced.is_none(),
+            "Built from an S3 object listing, which should never repeat a key"
+        );
+    }
+
+    /// Having loaded a timeline index, check if a layer referenced by the index exists.  If it does,
+    /// the layer's refcount will be incremented.  Later, after calling this for all references in all indices
+    /// in a tenant, orphan layers may be detected by their zero refcounts.
+    ///
+    /// Returns true if the layer exists
+    pub(crate) fn check_ref(
+        &mut self,
+        timeline_id: TimelineId,
+        layer_file: &LayerFileName,
+        metadata: &IndexLayerMetadata,
+    ) -> bool {
+        let Some(shard_tl) = self.shard_timelines.get_mut(&(metadata.shard, timeline_id)) else {
+            return false;
+        };
+
+        let Some(layer_ref) = shard_tl.get_mut(&(layer_file.clone(), metadata.generation)) else {
+            return false;
+        };
+
+        layer_ref.ref_count += 1;
+
+        true
+    }
+
+    pub(crate) fn get_orphans(&self) -> Vec<(ShardIndex, TimelineId, LayerFileName, Generation)> {
+        let mut result = Vec::new();
+        for ((shard_index, timeline_id), layers) in &self.shard_timelines {
+            for ((layer_file, generation), layer_ref) in layers {
+                if layer_ref.ref_count == 0 {
+                    result.push((*shard_index, *timeline_id, layer_file.clone(), *generation))
+                }
+            }
+        }
+
+        result
+    }
+}
+
 #[derive(Debug)]
 pub(crate) struct S3TimelineBlobData {
     pub(crate) blob_data: BlobDataParseResult,

s3_scrubber/src/lib.rs

@@ -15,10 +15,13 @@ use anyhow::Context;
 use aws_config::environment::EnvironmentVariableCredentialsProvider;
 use aws_config::imds::credentials::ImdsCredentialsProvider;
 use aws_config::meta::credentials::CredentialsProviderChain;
+use aws_config::profile::ProfileFileCredentialsProvider;
+use aws_config::retry::RetryConfig;
 use aws_config::sso::SsoCredentialsProvider;
 use aws_config::BehaviorVersion;
-use aws_sdk_s3::config::Region;
+use aws_sdk_s3::config::{AsyncSleep, Region, SharedAsyncSleep};
 use aws_sdk_s3::{Client, Config};
+use aws_smithy_async::rt::sleep::TokioSleep;
 
 use clap::ValueEnum;
 use pageserver::tenant::TENANTS_SEGMENT_NAME;
@@ -255,6 +258,11 @@ pub fn init_s3_client(account_id: Option<String>, bucket_region: Region) -> Clie
         let chain = CredentialsProviderChain::first_try(
             "env",
             EnvironmentVariableCredentialsProvider::new(),
+        )
+        // uses "AWS_PROFILE" / `aws sso login --profile <profile>`
+        .or_else(
+            "profile-sso",
+            ProfileFileCredentialsProvider::builder().build(),
         );
 
         // Use SSO if we were given an account ID
@@ -265,7 +273,7 @@ pub fn init_s3_client(account_id: Option<String>, bucket_region: Region) -> Clie
                     .account_id(sso_account)
                     .role_name("PowerUserAccess")
                     .start_url("https://neondb.awsapps.com/start")
-                    .region(Region::from_static("eu-central-1"))
+                    .region(bucket_region.clone())
                     .build(),
             ),
             None => chain,
@@ -277,9 +285,13 @@ pub fn init_s3_client(account_id: Option<String>, bucket_region: Region) -> Clie
         )
     };
 
+    let sleep_impl: Arc<dyn AsyncSleep> = Arc::new(TokioSleep::new());
+
     let mut builder = Config::builder()
         .behavior_version(BehaviorVersion::v2023_11_09())
         .region(bucket_region)
+        .retry_config(RetryConfig::adaptive().with_max_attempts(3))
+        .sleep_impl(SharedAsyncSleep::from(sleep_impl))
         .credentials_provider(credentials_provider);
 
     if let Ok(endpoint) = env::var("AWS_ENDPOINT_URL") {

s3_scrubber/src/main.rs

@@ -1,3 +1,4 @@
+use pageserver_api::shard::TenantShardId;
 use s3_scrubber::garbage::{find_garbage, purge_garbage, PurgeMode};
 use s3_scrubber::scan_metadata::scan_metadata;
 use s3_scrubber::{init_logging, BucketConfig, ConsoleConfig, NodeKind, TraversingDepth};
@@ -34,6 +35,8 @@ enum Command {
     ScanMetadata {
         #[arg(short, long, default_value_t = false)]
         json: bool,
+        #[arg(long = "tenant-id", num_args = 0..)]
+        tenant_ids: Vec<TenantShardId>,
     },
 }
 
@@ -57,35 +60,37 @@ async fn main() -> anyhow::Result<()> {
     ));
 
     match cli.command {
-        Command::ScanMetadata { json } => match scan_metadata(bucket_config.clone()).await {
-            Err(e) => {
-                tracing::error!("Failed: {e}");
-                Err(e)
-            }
-            Ok(summary) => {
-                if json {
-                    println!("{}", serde_json::to_string(&summary).unwrap())
-                } else {
-                    println!("{}", summary.summary_string());
+        Command::ScanMetadata { json, tenant_ids } => {
+            match scan_metadata(bucket_config.clone(), tenant_ids).await {
+                Err(e) => {
+                    tracing::error!("Failed: {e}");
+                    Err(e)
                 }
-                if summary.is_fatal() {
-                    Err(anyhow::anyhow!("Fatal scrub errors detected"))
-                } else if summary.is_empty() {
-                    // Strictly speaking an empty bucket is a valid bucket, but if someone ran the
-                    // scrubber they were likely expecting to scan something, and if we see no timelines
-                    // at all then it's likely due to some configuration issues like a bad prefix
-                    Err(anyhow::anyhow!(
-                        "No timelines found in bucket {} prefix {}",
-                        bucket_config.bucket,
-                        bucket_config
-                            .prefix_in_bucket
-                            .unwrap_or("<none>".to_string())
-                    ))
-                } else {
-                    Ok(())
+                Ok(summary) => {
+                    if json {
+                        println!("{}", serde_json::to_string(&summary).unwrap())
+                    } else {
+                        println!("{}", summary.summary_string());
+                    }
+                    if summary.is_fatal() {
+                        Err(anyhow::anyhow!("Fatal scrub errors detected"))
+                    } else if summary.is_empty() {
+                        // Strictly speaking an empty bucket is a valid bucket, but if someone ran the
+                        // scrubber they were likely expecting to scan something, and if we see no timelines
+                        // at all then it's likely due to some configuration issues like a bad prefix
+                        Err(anyhow::anyhow!(
+                            "No timelines found in bucket {} prefix {}",
+                            bucket_config.bucket,
+                            bucket_config
+                                .prefix_in_bucket
+                                .unwrap_or("<none>".to_string())
+                        ))
+                    } else {
+                        Ok(())
+                    }
                 }
             }
-        },
+        }
         Command::FindGarbage {
             node_kind,
             depth,

s3_scrubber/src/scan_metadata.rs

@@ -2,22 +2,25 @@ use std::collections::{HashMap, HashSet};
 
 use crate::checks::{
     branch_cleanup_and_check_errors, list_timeline_blobs, BlobDataParseResult, S3TimelineBlobData,
-    TimelineAnalysis,
+    TenantObjectListing, TimelineAnalysis,
 };
 use crate::metadata_stream::{stream_tenant_timelines, stream_tenants};
 use crate::{init_remote, BucketConfig, NodeKind, RootTarget, TenantShardTimelineId};
 use aws_sdk_s3::Client;
 use futures_util::{pin_mut, StreamExt, TryStreamExt};
 use histogram::Histogram;
+use pageserver::tenant::remote_timeline_client::remote_layer_path;
 use pageserver::tenant::IndexPart;
+use pageserver_api::shard::TenantShardId;
 use serde::Serialize;
+use utils::id::TenantId;
 
 #[derive(Serialize)]
 pub struct MetadataSummary {
     count: usize,
     with_errors: HashSet<TenantShardTimelineId>,
     with_warnings: HashSet<TenantShardTimelineId>,
-    with_garbage: HashSet<TenantShardTimelineId>,
+    with_orphans: HashSet<TenantShardTimelineId>,
     indices_by_version: HashMap<usize, usize>,
 
     layer_count: MinMaxHisto,
@@ -87,7 +90,7 @@ impl MetadataSummary {
             count: 0,
             with_errors: HashSet::new(),
             with_warnings: HashSet::new(),
-            with_garbage: HashSet::new(),
+            with_orphans: HashSet::new(),
             indices_by_version: HashMap::new(),
             layer_count: MinMaxHisto::new(),
             timeline_size_bytes: MinMaxHisto::new(),
@@ -141,6 +144,10 @@ impl MetadataSummary {
         }
     }
 
+    fn notify_timeline_orphan(&mut self, ttid: &TenantShardTimelineId) {
+        self.with_orphans.insert(*ttid);
+    }
+
     /// Long-form output for printing at end of a scan
     pub fn summary_string(&self) -> String {
         let version_summary: String = itertools::join(
@@ -154,7 +161,7 @@ impl MetadataSummary {
             "Timelines: {0}
 With errors: {1}
 With warnings: {2}
-With garbage: {3}
+With orphan layers: {3}
 Index versions: {version_summary}
 Timeline size bytes: {4}
 Layer size bytes: {5}
@@ -163,7 +170,7 @@ Timeline layer count: {6}
             self.count,
             self.with_errors.len(),
             self.with_warnings.len(),
-            self.with_garbage.len(),
+            self.with_orphans.len(),
             self.timeline_size_bytes.oneline(),
             self.layer_size_bytes.oneline(),
             self.layer_count.oneline(),
@@ -180,10 +187,17 @@ Timeline layer count: {6}
 }
 
 /// Scan the pageserver metadata in an S3 bucket, reporting errors and statistics.
-pub async fn scan_metadata(bucket_config: BucketConfig) -> anyhow::Result<MetadataSummary> {
+pub async fn scan_metadata(
+    bucket_config: BucketConfig,
+    tenant_ids: Vec<TenantShardId>,
+) -> anyhow::Result<MetadataSummary> {
     let (s3_client, target) = init_remote(bucket_config, NodeKind::Pageserver)?;
 
-    let tenants = stream_tenants(&s3_client, &target);
+    let tenants = if tenant_ids.is_empty() {
+        futures::future::Either::Left(stream_tenants(&s3_client, &target))
+    } else {
+        futures::future::Either::Right(futures::stream::iter(tenant_ids.into_iter().map(Ok)))
+    };
 
     // How many tenants to process in parallel.  We need to be mindful of pageservers
     // accessing the same per tenant prefixes, so use a lower setting than pageservers.
@@ -191,7 +205,7 @@ pub async fn scan_metadata(bucket_config: BucketConfig) -> anyhow::Result<Metada
 
     // Generate a stream of TenantTimelineId
     let timelines = tenants.map_ok(|t| stream_tenant_timelines(&s3_client, &target, t));
-    let timelines = timelines.try_buffer_unordered(CONCURRENCY);
+    let timelines = timelines.try_buffered(CONCURRENCY);
     let timelines = timelines.try_flatten();
 
     // Generate a stream of S3TimelineBlobData
@@ -204,17 +218,118 @@ pub async fn scan_metadata(bucket_config: BucketConfig) -> anyhow::Result<Metada
         Ok((ttid, data))
     }
     let timelines = timelines.map_ok(|ttid| report_on_timeline(&s3_client, &target, ttid));
-    let timelines = timelines.try_buffer_unordered(CONCURRENCY);
+    let timelines = timelines.try_buffered(CONCURRENCY);
 
+    // We must gather all the TenantShardTimelineId->S3TimelineBlobData for each tenant, because different
+    // shards in the same tenant might refer to one anothers' keys if a shard split has happened.
+
+    let mut tenant_id = None;
+    let mut tenant_objects = TenantObjectListing::default();
+    let mut tenant_timeline_results = Vec::new();
+
+    fn analyze_tenant(
+        tenant_id: TenantId,
+        summary: &mut MetadataSummary,
+        mut tenant_objects: TenantObjectListing,
+        timelines: Vec<(TenantShardTimelineId, S3TimelineBlobData)>,
+    ) {
+        let mut timeline_generations = HashMap::new();
+        for (ttid, data) in timelines {
+            // Stash the generation of each timeline, for later use identifying orphan layers
+            if let BlobDataParseResult::Parsed {
+                index_part: _index_part,
+                index_part_generation,
+                s3_layers: _s3_layers,
+            } = &data.blob_data
+            {
+                timeline_generations.insert(ttid, *index_part_generation);
+            }
+
+            // Apply checks to this timeline shard's metadata, and in the process update `tenant_objects`
+            // reference counts for layers across the tenant.
+            let analysis =
+                branch_cleanup_and_check_errors(&ttid, &mut tenant_objects, None, None, Some(data));
+            summary.update_analysis(&ttid, &analysis);
+        }
+
+        // Identifying orphan layers must be done on a tenant-wide basis, because individual
+        // shards' layers may be referenced by other shards.
+        //
+        // Orphan layers are not a corruption, and not an indication of a problem.  They are just
+        // consuming some space in remote storage, and may be cleaned up at leisure.
+        for (shard_index, timeline_id, layer_file, generation) in tenant_objects.get_orphans() {
+            let ttid = TenantShardTimelineId {
+                tenant_shard_id: TenantShardId {
+                    tenant_id,
+                    shard_count: shard_index.shard_count,
+                    shard_number: shard_index.shard_number,
+                },
+                timeline_id,
+            };
+
+            if let Some(timeline_generation) = timeline_generations.get(&ttid) {
+                if &generation >= timeline_generation {
+                    // Candidate orphan layer is in the current or future generation relative
+                    // to the index we read for this timeline shard, so its absence from the index
+                    // doesn't make it an orphan: more likely, it is a case where the layer was
+                    // uploaded, but the index referencing the layer wasn't written yet.
+                    continue;
+                }
+            }
+
+            let orphan_path = remote_layer_path(
+                &tenant_id,
+                &timeline_id,
+                shard_index,
+                &layer_file,
+                generation,
+            );
+
+            tracing::info!("Orphan layer detected: {orphan_path}");
+
+            summary.notify_timeline_orphan(&ttid);
+        }
+    }
+
+    // Iterate through  all the timeline results.  These are in key-order, so
+    // all results for the same tenant will be adjacent.  We accumulate these,
+    // and then call `analyze_tenant` to flush, when we see the next tenant ID.
     let mut summary = MetadataSummary::new();
     pin_mut!(timelines);
     while let Some(i) = timelines.next().await {
         let (ttid, data) = i?;
         summary.update_data(&data);
 
-        let analysis = branch_cleanup_and_check_errors(&ttid, &target, None, None, Some(data));
+        match tenant_id {
+            None => tenant_id = Some(ttid.tenant_shard_id.tenant_id),
+            Some(prev_tenant_id) => {
+                if prev_tenant_id != ttid.tenant_shard_id.tenant_id {
+                    let tenant_objects = std::mem::take(&mut tenant_objects);
+                    let timelines = std::mem::take(&mut tenant_timeline_results);
+                    analyze_tenant(prev_tenant_id, &mut summary, tenant_objects, timelines);
+                    tenant_id = Some(ttid.tenant_shard_id.tenant_id);
+                }
+            }
+        }
 
-        summary.update_analysis(&ttid, &analysis);
+        if let BlobDataParseResult::Parsed {
+            index_part: _index_part,
+            index_part_generation: _index_part_generation,
+            s3_layers,
+        } = &data.blob_data
+        {
+            tenant_objects.push(ttid, s3_layers.clone());
+        }
+        tenant_timeline_results.push((ttid, data));
+    }
+
+    if !tenant_timeline_results.is_empty() {
+        analyze_tenant(
+            tenant_id.expect("Must be set if results are present"),
+            &mut summary,
+            tenant_objects,
+            tenant_timeline_results,
+        );
     }
 
     Ok(summary)

safekeeper/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]
 async-stream.workspace = true
 anyhow.workspace = true
@@ -16,6 +22,7 @@ chrono.workspace = true
 clap = { workspace = true, features = ["derive"] }
 const_format.workspace = true
 crc32c.workspace = true
+fail.workspace = true
 fs2.workspace = true
 git-version.workspace = true
 hex.workspace = true

safekeeper/src/bin/safekeeper.rs

@@ -54,6 +54,19 @@ const ID_FILE_NAME: &str = "safekeeper.id";
 project_git_version!(GIT_VERSION);
 project_build_tag!(BUILD_TAG);
 
+const FEATURES: &[&str] = &[
+    #[cfg(feature = "testing")]
+    "testing",
+];
+
+fn version() -> String {
+    format!(
+        "{GIT_VERSION} failpoints: {}, features: {:?}",
+        fail::has_failpoints(),
+        FEATURES,
+    )
+}
+
 const ABOUT: &str = r#"
 A fleet of safekeepers is responsible for reliably storing WAL received from
 compute, passing it through consensus (mitigating potential computes brain
@@ -167,7 +180,9 @@ async fn main() -> anyhow::Result<()> {
     // getting 'argument cannot be used multiple times' error. This seems to be
     // impossible with pure Derive API, so convert struct to Command, modify it,
     // parse arguments, and then fill the struct back.
-    let cmd = <Args as clap::CommandFactory>::command().args_override_self(true);
+    let cmd = <Args as clap::CommandFactory>::command()
+        .args_override_self(true)
+        .version(version());
     let mut matches = cmd.get_matches();
     let mut args = <Args as clap::FromArgMatches>::from_arg_matches_mut(&mut matches)?;
 

safekeeper/src/http/routes.rs

@@ -12,6 +12,8 @@ use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
 use tokio::fs::File;
 use tokio::io::AsyncReadExt;
+use tokio_util::sync::CancellationToken;
+use utils::failpoint_support::failpoints_handler;
 
 use std::io::Write as _;
 use tokio::sync::mpsc;
@@ -444,6 +446,12 @@ pub fn make_router(conf: SafeKeeperConf) -> RouterBuilder<hyper::Body, ApiError>
         .data(Arc::new(conf))
         .data(auth)
         .get("/v1/status", |r| request_span(r, status_handler))
+        .put("/v1/failpoints", |r| {
+            request_span(r, move |r| async {
+                let cancel = CancellationToken::new();
+                failpoints_handler(r, cancel).await
+            })
+        })
         // Will be used in the future instead of implicit timeline creation
         .post("/v1/tenant/timeline", |r| {
             request_span(r, timeline_create_handler)

safekeeper/src/send_wal.rs

@@ -17,6 +17,7 @@ use postgres_ffi::{TimestampTz, MAX_SEND_SIZE};
 use pq_proto::{BeMessage, WalSndKeepAlive, XLogDataBody};
 use serde::{Deserialize, Serialize};
 use tokio::io::{AsyncRead, AsyncWrite};
+use utils::failpoint_support;
 use utils::id::TenantTimelineId;
 use utils::lsn::AtomicLsn;
 use utils::pageserver_feedback::PageserverFeedback;
@@ -391,15 +392,8 @@ impl SafekeeperPostgresHandler {
         // application_name: give only committed WAL (used by pageserver) or all
         // existing WAL (up to flush_lsn, used by walproposer or peer recovery).
         // The second case is always driven by a consensus leader which term
-        // must generally be also supplied. However we're sloppy to do this in
-        // walproposer recovery which will be removed soon. So TODO is to make
-        // it not Option'al then.
-        //
-        // Fetching WAL without term in recovery creates a small risk of this
-        // WAL getting concurrently garbaged if another compute rises which
-        // collects majority and starts fixing log on this safekeeper itself.
-        // That's ok as (old) proposer will never be able to commit such WAL.
-        let end_watch = if self.is_walproposer_recovery() {
+        // must be supplied.
+        let end_watch = if term.is_some() {
             EndWatch::Flush(tli.get_term_flush_lsn_watch_rx())
         } else {
             EndWatch::Commit(tli.get_commit_lsn_watch_rx())
@@ -535,12 +529,19 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
             );
 
             // try to send as much as available, capped by MAX_SEND_SIZE
-            let mut send_size = self
-                .end_pos
-                .checked_sub(self.start_pos)
-                .context("reading wal without waiting for it first")?
-                .0 as usize;
-            send_size = min(send_size, self.send_buf.len());
+            let mut chunk_end_pos = self.start_pos + MAX_SEND_SIZE as u64;
+            // if we went behind available WAL, back off
+            if chunk_end_pos >= self.end_pos {
+                chunk_end_pos = self.end_pos;
+            } else {
+                // If sending not up to end pos, round down to page boundary to
+                // avoid breaking WAL record not at page boundary, as protocol
+                // demands. See walsender.c (XLogSendPhysical).
+                chunk_end_pos = chunk_end_pos
+                    .checked_sub(chunk_end_pos.block_offset())
+                    .unwrap();
+            }
+            let send_size = (chunk_end_pos.0 - self.start_pos.0) as usize;
             let send_buf = &mut self.send_buf[..send_size];
             let send_size: usize;
             {
@@ -551,7 +552,8 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
                 } else {
                     None
                 };
-                // read wal into buffer
+                // Read WAL into buffer. send_size can be additionally capped to
+                // segment boundary here.
                 send_size = self.wal_reader.read(send_buf).await?
             };
             let send_buf = &send_buf[..send_size];
@@ -566,6 +568,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
                 }))
                 .await?;
 
+            if let Some(appname) = &self.appname {
+                if appname == "replica" {
+                    failpoint_support::sleep_millis_async!("sk-send-wal-replica-sleep");
+                }
+            }
             trace!(
                 "sent {} bytes of WAL {}-{}",
                 send_size,

safekeeper/src/wal_storage.rs

@@ -565,6 +565,9 @@ impl WalReader {
         })
     }
 
+    /// Read WAL at current position into provided buf, returns number of bytes
+    /// read. It can be smaller than buf size only if segment boundary is
+    /// reached.
     pub async fn read(&mut self, buf: &mut [u8]) -> Result<usize> {
         // If this timeline is new, we may not have a full segment yet, so
         // we pad the first bytes of the timeline's first WAL segment with 0s