wip

## Problem

In the original PR[0], I've missed a couple of `release` occurrences
that should also be handled for `release-proxy` branch

- [0] https://github.com/neondatabase/neon/pull/6797

## Summary of changes
- Add handling for `release-proxy` branch to allure report
- Add handling for `release-proxy` branch to e2e tests malts.com

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

@@ -39,7 +39,7 @@ runs:
         PR_NUMBER=$(jq --raw-output .pull_request.number "$GITHUB_EVENT_PATH" || true)
         if [ "${PR_NUMBER}" != "null" ]; then
           BRANCH_OR_PR=pr-${PR_NUMBER}
-        elif [ "${GITHUB_REF_NAME}" = "main" ] || [ "${GITHUB_REF_NAME}" = "release" ]; then
+        elif [ "${GITHUB_REF_NAME}" = "main" ] || [ "${GITHUB_REF_NAME}" = "release" ] || [ "${GITHUB_REF_NAME}" = "release-proxy" ]; then
           # Shortcut for special branches
           BRANCH_OR_PR=${GITHUB_REF_NAME}
         else
@@ -59,7 +59,7 @@ runs:
         BUCKET: neon-github-public-dev
 
     # TODO: We can replace with a special docker image with Java and Allure pre-installed
-    - uses: actions/setup-java@v3
+    - uses: actions/setup-java@v4
       with:
         distribution: 'temurin'
         java-version: '17'
@@ -76,8 +76,8 @@ runs:
           rm -f ${ALLURE_ZIP}
         fi
       env:
-        ALLURE_VERSION: 2.24.0
-        ALLURE_ZIP_SHA256: 60b1d6ce65d9ef24b23cf9c2c19fd736a123487c38e54759f1ed1a7a77353c90
+        ALLURE_VERSION: 2.27.0
+        ALLURE_ZIP_SHA256: b071858fb2fa542c65d8f152c5c40d26267b2dfb74df1f1608a589ecca38e777
 
     # Potentially we could have several running build for the same key (for example, for the main branch), so we use improvised lock for this
     - name: Acquire lock
@@ -180,7 +180,7 @@ runs:
         fi
 
     - name: Cache poetry deps
-      uses: actions/cache@v3
+      uses: actions/cache@v4
       with:
         path: ~/.cache/pypoetry/virtualenvs
         key: v2-${{ runner.os }}-python-deps-${{ hashFiles('poetry.lock') }}
@@ -215,7 +215,7 @@ runs:
           rm -rf ${WORKDIR}
         fi
 
-    - uses: actions/github-script@v6
+    - uses: actions/github-script@v7
       if: always()
       env:
         REPORT_URL: ${{ steps.generate-report.outputs.report-url }}

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

@@ -19,7 +19,7 @@ runs:
         PR_NUMBER=$(jq --raw-output .pull_request.number "$GITHUB_EVENT_PATH" || true)
         if [ "${PR_NUMBER}" != "null" ]; then
           BRANCH_OR_PR=pr-${PR_NUMBER}
-        elif [ "${GITHUB_REF_NAME}" = "main" ] || [ "${GITHUB_REF_NAME}" = "release" ]; then
+        elif [ "${GITHUB_REF_NAME}" = "main" ] || [ "${GITHUB_REF_NAME}" = "release" ] || [ "${GITHUB_REF_NAME}" = "release-proxy" ]; then
           # Shortcut for special branches
           BRANCH_OR_PR=${GITHUB_REF_NAME}
         else

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

@@ -80,13 +80,13 @@ runs:
 
     - name: Checkout
       if: inputs.needs_postgres_source == 'true'
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
       with:
         submodules: true
         fetch-depth: 1
 
     - name: Cache poetry deps
-      uses: actions/cache@v3
+      uses: actions/cache@v4
       with:
         path: ~/.cache/pypoetry/virtualenvs
         key: v2-${{ runner.os }}-python-deps-${{ hashFiles('poetry.lock') }}

.github/workflows/actionlint.yml

@@ -16,8 +16,14 @@ concurrency:
   cancel-in-progress: ${{ github.event_name == 'pull_request' }}
 
 jobs:
-  actionlint:
+  check-permissions:
     if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
+    uses: ./.github/workflows/check-permissions.yml
+    with:
+      github-event-name: ${{ github.event_name}}
+
+  actionlint:
+    needs: [ check-permissions ]
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4

.github/workflows/approved-for-ci-run.yml

@@ -64,7 +64,7 @@ jobs:
     steps:
       - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"
 
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
         with:
           ref: main
           token: ${{ secrets.CI_ACCESS_TOKEN }}

.github/workflows/benchmarking.yml

@@ -66,7 +66,7 @@ jobs:
       options: --init
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -221,7 +221,7 @@ jobs:
     timeout-minutes: 480
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -366,7 +366,7 @@ jobs:
       options: --init
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -465,7 +465,7 @@ jobs:
       options: --init
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
     - name: Download Neon artifact
       uses: ./.github/actions/download
@@ -562,7 +562,7 @@ jobs:
       options: --init
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
     - name: Download Neon artifact
       uses: ./.github/actions/download

.github/workflows/build_and_test.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - main
       - release
+      - release-proxy
   pull_request:
 
 defaults:
@@ -27,24 +28,9 @@ env:
 jobs:
   check-permissions:
     if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
-    runs-on: ubuntu-latest
-    steps:
-    - name: Disallow PRs from forks
-      if: |
-        github.event_name == 'pull_request' &&
-        github.event.pull_request.head.repo.full_name != github.repository
-
-      run: |
-        if [ "${{ contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.pull_request.author_association) }}" = "true" ]; then
-          MESSAGE="Please create a PR from a branch of ${GITHUB_REPOSITORY} instead of a fork"
-        else
-          MESSAGE="The PR should be reviewed and labelled with 'approved-for-ci-run' to trigger a CI run"
-        fi
-
-        echo >&2 "We don't run CI for PRs from forks"
-        echo >&2 "${MESSAGE}"
-
-        exit 1
+    uses: ./.github/workflows/check-permissions.yml
+    with:
+      github-event-name: ${{ github.event_name}}
 
   cancel-previous-e2e-tests:
     needs: [ check-permissions ]
@@ -69,7 +55,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -82,6 +68,8 @@ jobs:
             echo "tag=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
             echo "tag=release-$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
+          elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
+            echo "tag=release-proxy-$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
           else
             echo "GITHUB_REF_NAME (value '$GITHUB_REF_NAME') is not set to either 'main' or 'release'"
             echo "tag=$GITHUB_RUN_ID" >> $GITHUB_OUTPUT
@@ -106,13 +94,13 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: false
           fetch-depth: 1
 
       - name: Cache poetry deps
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: ~/.cache/pypoetry/virtualenvs
           key: v2-${{ runner.os }}-python-deps-${{ hashFiles('poetry.lock') }}
@@ -138,7 +126,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -146,7 +134,7 @@ jobs:
 #      Disabled for now
 #      - name: Restore cargo deps cache
 #        id: cache_cargo
-#        uses: actions/cache@v3
+#        uses: actions/cache@v4
 #        with:
 #          path: |
 #            !~/.cargo/registry/src
@@ -231,7 +219,7 @@ jobs:
           done
 
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -253,7 +241,7 @@ jobs:
           done
 
           if [ "${FAILED}" = "true" ]; then
-            echo >&2 "Please update vendors/revisions.json if these changes are intentional"
+            echo >&2 "Please update vendor/revisions.json if these changes are intentional"
             exit 1
           fi
 
@@ -303,7 +291,7 @@ jobs:
       # compressed crates.
 #      - name: Cache cargo deps
 #        id: cache_cargo
-#        uses: actions/cache@v3
+#        uses: actions/cache@v4
 #        with:
 #          path: |
 #            ~/.cargo/registry/
@@ -317,21 +305,21 @@ jobs:
 
       - name: Cache postgres v14 build
         id: cache_pg_14
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v14
           key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v15
           key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v16
           key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
@@ -451,7 +439,7 @@ jobs:
         pg_version: [ v14, v15, v16 ]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -472,9 +460,14 @@ jobs:
           CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
           BUILD_TAG: ${{ needs.tag.outputs.build-tag }}
           PAGESERVER_VIRTUAL_FILE_IO_ENGINE: std-fs
+          PAGESERVER_GET_VECTORED_IMPL: vectored
 
+      # Temporary disable this step until we figure out why it's so flaky
+      # Ref https://github.com/neondatabase/neon/issues/4540
       - name: Merge and upload coverage data
-        if: matrix.build_type == 'debug' && matrix.pg_version == 'v14'
+        if: |
+          false &&
+          matrix.build_type == 'debug' && matrix.pg_version == 'v14'
         uses: ./.github/actions/save-coverage-data
 
   get-benchmarks-durations:
@@ -488,10 +481,10 @@ jobs:
     if: github.ref_name == 'main' || contains(github.event.pull_request.labels.*.name, 'run-benchmarks')
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Cache poetry deps
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: ~/.cache/pypoetry/virtualenvs
           key: v1-${{ runner.os }}-python-deps-${{ hashFiles('poetry.lock') }}
@@ -525,7 +518,7 @@ jobs:
         build_type: [ release ]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Pytest benchmarks
         uses: ./.github/actions/run-python-test-set
@@ -554,7 +547,7 @@ jobs:
       options: --init
 
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
 
       - name: Create Allure report
         if: ${{ !cancelled() }}
@@ -565,7 +558,7 @@ jobs:
         env:
           REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
 
-      - uses: actions/github-script@v6
+      - uses: actions/github-script@v7
         if: ${{ !cancelled() }}
         with:
           # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
@@ -605,7 +598,7 @@ jobs:
         coverage-json: ${{ steps.upload-coverage-report-new.outputs.summary-json }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 0
@@ -674,7 +667,7 @@ jobs:
           REPORT_URL=https://${BUCKET}.s3.amazonaws.com/code-coverage/${COMMIT_SHA}/lcov/summary.json
           echo "summary-json=${REPORT_URL}" >> $GITHUB_OUTPUT
 
-      - uses: actions/github-script@v6
+      - uses: actions/github-script@v7
         env:
           REPORT_URL_NEW: ${{ steps.upload-coverage-report-new.outputs.report-url }}
           COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
@@ -692,7 +685,7 @@ jobs:
             })
 
   trigger-e2e-tests:
-    if: ${{ !github.event.pull_request.draft || contains( github.event.pull_request.labels.*.name, 'run-e2e-tests-in-draft') || github.ref_name == 'main' || github.ref_name == 'release' }}
+    if: ${{ !github.event.pull_request.draft || contains( github.event.pull_request.labels.*.name, 'run-e2e-tests-in-draft') || github.ref_name == 'main' || github.ref_name == 'release' || github.ref_name == 'release-proxy' }}
     needs: [ check-permissions, promote-images, tag ]
     uses: ./.github/workflows/trigger-e2e-tests.yml
     secrets: inherit
@@ -700,158 +693,173 @@ jobs:
   neon-image:
     needs: [ check-permissions, build-buildtools-image, tag ]
     runs-on: [ self-hosted, gen3, large ]
-    container: gcr.io/kaniko-project/executor:v1.9.2-debug
-    defaults:
-      run:
-        shell: sh -eu {0}
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v1 # v3 won't work with kaniko
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 0
 
-      - name: Configure ECR and Docker Hub login
+      # Use custom DOCKER_CONFIG directory to avoid conflicts with default settings
+      # The default value is ~/.docker
+      - name: Set custom docker config directory
         run: |
-          DOCKERHUB_AUTH=$(echo -n "${{ secrets.NEON_DOCKERHUB_USERNAME }}:${{ secrets.NEON_DOCKERHUB_PASSWORD }}" | base64)
-          echo "::add-mask::${DOCKERHUB_AUTH}"
+          mkdir -p .docker-custom
+          echo DOCKER_CONFIG=$(pwd)/.docker-custom >> $GITHUB_ENV
+      - uses: docker/setup-buildx-action@v3
 
-          cat <<-EOF > /kaniko/.docker/config.json
-            {
-              "auths": {
-                "https://index.docker.io/v1/": {
-                  "auth": "${DOCKERHUB_AUTH}"
-                }
-              },
-              "credHelpers": {
-                "369495373322.dkr.ecr.eu-central-1.amazonaws.com": "ecr-login"
-              }
-            }
-          EOF
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
 
-      - name: Kaniko build neon
-        run:
-          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
-                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
-                           --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}}
+      - uses: docker/login-action@v3
+        with:
+          registry: 369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          username: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          password: ${{ secrets.AWS_SECRET_KEY_DEV }}
 
-      # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
-      - name: Cleanup ECR folder
-        run: rm -rf ~/.ecr
+      - uses: docker/build-push-action@v5
+        with:
+          context: .
+          build-args: |
+            GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
+            BUILD_TAG=${{ needs.tag.outputs.build-tag }}
+            TAG=${{ needs.build-buildtools-image.outputs.build-tools-tag }}
+            REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          provenance: false
+          push: true
+          pull: true
+          file: Dockerfile
+          cache-from: type=registry,ref=neondatabase/neon:cache
+          cache-to: type=registry,ref=neondatabase/neon:cache,mode=max
+          tags: |
+            369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}}
+            neondatabase/neon:${{needs.tag.outputs.build-tag}}
+
+      - name: Remove custom docker config directory
+        if: always()
+        run: |
+          rm -rf .docker-custom
 
   compute-tools-image:
     runs-on: [ self-hosted, gen3, large ]
     needs: [ check-permissions, build-buildtools-image, tag ]
-    container: gcr.io/kaniko-project/executor:v1.9.2-debug
-    defaults:
-      run:
-        shell: sh -eu {0}
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v1 # v3 won't work with kaniko
-
-      - name: Configure ECR and Docker Hub login
-        run: |
-          DOCKERHUB_AUTH=$(echo -n "${{ secrets.NEON_DOCKERHUB_USERNAME }}:${{ secrets.NEON_DOCKERHUB_PASSWORD }}" | base64)
-          echo "::add-mask::${DOCKERHUB_AUTH}"
-
-          cat <<-EOF > /kaniko/.docker/config.json
-            {
-              "auths": {
-                "https://index.docker.io/v1/": {
-                  "auth": "${DOCKERHUB_AUTH}"
-                }
-              },
-              "credHelpers": {
-                "369495373322.dkr.ecr.eu-central-1.amazonaws.com": "ecr-login"
-              }
-            }
-          EOF
-
-      - name: Kaniko build compute tools
-        run:
-          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
-                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
-                           --context .
-                           --build-arg GIT_VERSION=${{ 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:${{needs.tag.outputs.build-tag}}
-                           --destination neondatabase/compute-tools:${{needs.tag.outputs.build-tag}}
-
-      # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
-      - name: Cleanup ECR folder
-        run: rm -rf ~/.ecr
-
-  compute-node-image:
-    needs: [ check-permissions, build-buildtools-image, tag ]
-    runs-on: [ self-hosted, gen3, large ]
-    container:
-      image: gcr.io/kaniko-project/executor:v1.9.2-debug
-      # Workaround for "Resolving download.osgeo.org (download.osgeo.org)... failed: Temporary failure in name resolution.""
-      # Should be prevented by https://github.com/neondatabase/neon/issues/4281
-      options: --add-host=download.osgeo.org:140.211.15.30
-    strategy:
-      fail-fast: false
-      matrix:
-        version: [ v14, v15, v16 ]
-    defaults:
-      run:
-        shell: sh -eu {0}
-
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v1 # v3 won't work with kaniko
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 0
 
-      - name: Configure ECR and Docker Hub login
+      # Use custom DOCKER_CONFIG directory to avoid conflicts with default settings
+      # The default value is ~/.docker
+      - name: Set custom docker config directory
         run: |
-          DOCKERHUB_AUTH=$(echo -n "${{ secrets.NEON_DOCKERHUB_USERNAME }}:${{ secrets.NEON_DOCKERHUB_PASSWORD }}" | base64)
-          echo "::add-mask::${DOCKERHUB_AUTH}"
+          mkdir -p .docker-custom
+          echo DOCKER_CONFIG=$(pwd)/.docker-custom >> $GITHUB_ENV
+      - uses: docker/setup-buildx-action@v3
 
-          cat <<-EOF > /kaniko/.docker/config.json
-            {
-              "auths": {
-                "https://index.docker.io/v1/": {
-                  "auth": "${DOCKERHUB_AUTH}"
-                }
-              },
-              "credHelpers": {
-                "369495373322.dkr.ecr.eu-central-1.amazonaws.com": "ecr-login"
-              }
-            }
-          EOF
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
 
-      - name: Kaniko build compute node with extensions
-        run:
-          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
-                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
-                           --context .
-                           --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}}
-                           --destination neondatabase/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
-                           --cleanup
+      - uses: docker/login-action@v3
+        with:
+          registry: 369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          username: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          password: ${{ secrets.AWS_SECRET_KEY_DEV }}
 
-      # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
-      - name: Cleanup ECR folder
-        run: rm -rf ~/.ecr
+      - uses: docker/build-push-action@v5
+        with:
+          context: .
+          build-args: |
+            GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
+            BUILD_TAG=${{needs.tag.outputs.build-tag}}
+            TAG=${{needs.build-buildtools-image.outputs.build-tools-tag}}
+            REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          provenance: false
+          push: true
+          pull: true
+          file: Dockerfile.compute-tools
+          cache-from: type=registry,ref=neondatabase/compute-tools:cache
+          cache-to: type=registry,ref=neondatabase/compute-tools:cache,mode=max
+          tags: |
+            369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}}
+            neondatabase/compute-tools:${{needs.tag.outputs.build-tag}}
+
+      - name: Remove custom docker config directory
+        if: always()
+        run: |
+          rm -rf .docker-custom
+
+  compute-node-image:
+    needs: [ check-permissions, build-buildtools-image, tag ]
+    runs-on: [ self-hosted, gen3, large ]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        version: [ v14, v15, v16 ]
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          submodules: true
+          fetch-depth: 0
+
+      # Use custom DOCKER_CONFIG directory to avoid conflicts with default settings
+      # The default value is ~/.docker
+      - name: Set custom docker config directory
+        run: |
+          mkdir -p .docker-custom
+          echo DOCKER_CONFIG=$(pwd)/.docker-custom >> $GITHUB_ENV
+      - uses: docker/setup-buildx-action@v3
+        with:
+          # Disable parallelism for docker buildkit.
+          # As we already build everything with `make -j$(nproc)`, running it in additional level of parallelisam blows up the Runner.
+          config-inline: |
+            [worker.oci]
+              max-parallelism = 1
+
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+
+      - uses: docker/login-action@v3
+        with:
+          registry: 369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          username: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          password: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+      - uses: docker/build-push-action@v5
+        with:
+          context: .
+          build-args: |
+            GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
+            PG_VERSION=${{ matrix.version }}
+            BUILD_TAG=${{needs.tag.outputs.build-tag}}
+            TAG=${{needs.build-buildtools-image.outputs.build-tools-tag}}
+            REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          provenance: false
+          push: true
+          pull: true
+          file: Dockerfile.compute-node
+          cache-from: type=registry,ref=neondatabase/compute-node-${{ matrix.version }}:cache
+          cache-to: type=registry,ref=neondatabase/compute-node-${{ matrix.version }}:cache,mode=max
+          tags: |
+            369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+            neondatabase/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+
+      - name: Remove custom docker config directory
+        if: always()
+        run: |
+          rm -rf .docker-custom
 
   vm-compute-node-image:
     needs: [ check-permissions, tag, compute-node-image ]
@@ -900,7 +908,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -962,9 +970,7 @@ jobs:
           crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} vm-compute-node-v16
 
       - name: Add latest tag to images
-        if: |
-          (github.ref_name == 'main' || github.ref_name == 'release') &&
-           github.event_name != 'workflow_dispatch'
+        if: github.ref_name == 'main' || github.ref_name == 'release' || github.ref_name == 'release-proxy'
         run: |
           crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}} latest
           crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} latest
@@ -976,9 +982,7 @@ jobs:
           crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} latest
 
       - name: Push images to production ECR
-        if: |
-          (github.ref_name == 'main' || github.ref_name == 'release') &&
-           github.event_name != 'workflow_dispatch'
+        if: github.ref_name == 'main' || github.ref_name == 'release'|| github.ref_name == 'release-proxy'
         run: |
           crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/neon:latest
           crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:latest
@@ -1002,9 +1006,7 @@ jobs:
           crane push vm-compute-node-v16 neondatabase/vm-compute-node-v16:${{needs.tag.outputs.build-tag}}
 
       - name: Push latest tags to Docker Hub
-        if: |
-          (github.ref_name == 'main' || github.ref_name == 'release') &&
-          github.event_name != 'workflow_dispatch'
+        if: github.ref_name == 'main' || github.ref_name == 'release'|| github.ref_name == 'release-proxy'
         run: |
           crane tag neondatabase/neon:${{needs.tag.outputs.build-tag}} latest
           crane tag neondatabase/compute-tools:${{needs.tag.outputs.build-tag}} latest
@@ -1094,7 +1096,7 @@ jobs:
 
   deploy:
     needs: [ check-permissions, promote-images, tag, regress-tests, trigger-custom-extensions-build-and-wait ]
-    if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
+    if: github.ref_name == 'main' || github.ref_name == 'release'|| github.ref_name == 'release-proxy'
 
     runs-on: [ self-hosted, gen3, small ]
     container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/ansible:latest
@@ -1114,7 +1116,7 @@ jobs:
           done
 
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: false
           fetch-depth: 0
@@ -1129,15 +1131,29 @@ jobs:
             # TODO: move deployPreprodRegion to release (`"$GITHUB_REF_NAME" == "release"` block), once Staging support different compute tag prefixes for different regions
             gh workflow --repo neondatabase/aws run deploy-dev.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}} -f deployPreprodRegion=true
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
-            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main -f branch=main -f dockerTag=${{needs.tag.outputs.build-tag}}
+            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
+              -f deployPgSniRouter=false \
+              -f deployProxy=false \
+              -f deployStorage=true \
+              -f deployStorageBroker=true \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}}
+          elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
+            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
+              -f deployPgSniRouter=true \
+              -f deployProxy=true \
+              -f deployStorage=false \
+              -f deployStorageBroker=false \
+              -f branch=main \
+              -f dockerTag=${{needs.tag.outputs.build-tag}}
           else
             echo "GITHUB_REF_NAME (value '$GITHUB_REF_NAME') is not set to either 'main' or 'release'"
             exit 1
           fi
 
       - name: Create git tag
-        if: github.ref_name == 'release'
-        uses: actions/github-script@v6
+        if: github.ref_name == 'release' || github.ref_name == 'release-proxy'
+        uses: actions/github-script@v7
         with:
           # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
           retries: 5
@@ -1149,9 +1165,10 @@ jobs:
               sha: context.sha,
             })
 
+      # TODO: check how GitHub releases looks for proxy releases and enable it if it's ok
       - name: Create GitHub release
         if: github.ref_name == 'release'
-        uses: actions/github-script@v6
+        uses: actions/github-script@v7
         with:
           # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
           retries: 5

.github/workflows/check-permissions.yml

@@ -0,0 +1,36 @@
+name: Check Permissions
+
+on:
+  workflow_call:
+    inputs:
+      github-event-name:
+        required: true
+        type: string
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+# No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
+permissions: {}
+
+jobs:
+  check-permissions:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Disallow CI runs on PRs from forks
+      if: |
+        inputs.github-event-name  == 'pull_request' &&
+        github.event.pull_request.head.repo.full_name != github.repository
+      run: |
+        if [ "${{ contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.pull_request.author_association) }}" = "true" ]; then
+          MESSAGE="Please create a PR from a branch of ${GITHUB_REPOSITORY} instead of a fork"
+        else
+          MESSAGE="The PR should be reviewed and labelled with 'approved-for-ci-run' to trigger a CI run"
+        fi
+
+        # TODO: use actions/github-script to post this message as a PR comment
+        echo >&2 "We don't run CI for PRs from forks"
+        echo >&2 "${MESSAGE}"
+
+        exit 1

.github/workflows/neon_extra_builds.yml

@@ -20,7 +20,14 @@ env:
   COPT: '-Werror'
 
 jobs:
+  check-permissions:
+    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
+    uses: ./.github/workflows/check-permissions.yml
+    with:
+      github-event-name: ${{ github.event_name}}
+
   check-macos-build:
+    needs: [ check-permissions ]
     if: |
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos')  ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
@@ -57,21 +64,21 @@ jobs:
 
       - name: Cache postgres v14 build
         id: cache_pg_14
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v14
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v15
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v16
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
@@ -82,7 +89,7 @@ jobs:
           echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
 
       - name: Cache cargo deps
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: |
             ~/.cargo/registry
@@ -116,8 +123,8 @@ jobs:
         run: ./run_clippy.sh
 
   check-linux-arm-build:
+    needs: [ check-permissions ]
     timeout-minutes: 90
-    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: [ self-hosted, dev, arm64 ]
 
     env:
@@ -172,21 +179,21 @@ jobs:
 
       - name: Cache postgres v14 build
         id: cache_pg_14
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v14
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v15
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: pg_install/v16
           key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
@@ -237,8 +244,8 @@ jobs:
           cargo nextest run --package remote_storage --test test_real_azure
 
   check-codestyle-rust-arm:
+    needs: [ check-permissions ]
     timeout-minutes: 90
-    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: [ self-hosted, dev, arm64 ]
 
     container:
@@ -309,6 +316,7 @@ jobs:
         run: cargo deny check
 
   gather-rust-build-stats:
+    needs: [ check-permissions ]
     if: |
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-stats') ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
@@ -356,7 +364,7 @@ jobs:
           echo "report-url=${REPORT_URL}" >> $GITHUB_OUTPUT
 
       - name: Publish build stats report
-        uses: actions/github-script@v6
+        uses: actions/github-script@v7
         env:
           REPORT_URL: ${{ steps.upload-stats.outputs.report-url }}
           SHA: ${{ github.event.pull_request.head.sha || github.sha }}

.github/workflows/pg_clients.yml

@@ -28,7 +28,7 @@ jobs:
 
     steps:
     - name: Checkout
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
 
     - uses: actions/setup-python@v4
       with:
@@ -38,7 +38,7 @@ jobs:
       uses: snok/install-poetry@v1
 
     - name: Cache poetry deps
-      uses: actions/cache@v3
+      uses: actions/cache@v4
       with:
         path: ~/.cache/pypoetry/virtualenvs
         key: v2-${{ runner.os }}-python-deps-ubunutu-latest-${{ hashFiles('poetry.lock') }}
@@ -82,7 +82,7 @@ jobs:
     # It will be fixed after switching to gen2 runner
     - name: Upload python test logs
       if: always()
-      uses: actions/upload-artifact@v3
+      uses: actions/upload-artifact@v4
       with:
         retention-days: 7
         name: python-test-pg_clients-${{ runner.os }}-stage-logs

.github/workflows/release.yml

@@ -2,12 +2,31 @@ name: Create Release Branch
 
 on:
   schedule:
-    - cron: '0 6 * * 1'
+    # It should be kept in sync with if-condition in jobs
+    - cron: '0 6 * * MON' # Storage release
+    - cron: '0 6 * * THU' # Proxy release
   workflow_dispatch:
+    inputs:
+      create-storage-release-branch:
+        type: boolean
+        description: 'Create Storage release PR'
+        required: false
+      create-proxy-release-branch:
+        type: boolean
+        description: 'Create Proxy release PR'
+        required: false
+
+# No permission for GITHUB_TOKEN by default; the **minimal required** set of permissions should be granted in each job.
+permissions: {}
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
 
 jobs:
-  create_release_branch:
-    runs-on: [ ubuntu-latest ]
+  create-storage-release-branch:
+    if: ${{ github.event.schedule == '0 6 * * MON' || format('{0}', inputs.create-storage-release-branch) == 'true' }}
+    runs-on: ubuntu-latest
 
     permissions:
       contents: write # for `git push`
@@ -18,27 +37,67 @@ jobs:
       with:
         ref: main
 
-    - name: Get current date
-      id: date
-      run: echo "date=$(date +'%Y-%m-%d')" >> $GITHUB_OUTPUT
+    - name: Set environment variables
+      run: |
+        echo "RELEASE_DATE=$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
+        echo "RELEASE_BRANCH=rc/$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
 
     - name: Create release branch
-      run: git checkout -b releases/${{ steps.date.outputs.date }}
+      run: git checkout -b $RELEASE_BRANCH
 
     - name: Push new branch
-      run: git push origin releases/${{ steps.date.outputs.date }}
+      run: git push origin $RELEASE_BRANCH
 
     - name: Create pull request into release
       env:
         GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
       run: |
         cat << EOF > body.md
-          ## Release ${{ steps.date.outputs.date }}
+          ## Release ${RELEASE_DATE}
 
-          **Please merge this PR using 'Create a merge commit'!**
+          **Please merge this Pull Request using 'Create a merge commit' button**
         EOF
 
-        gh pr create --title "Release ${{ steps.date.outputs.date }}" \
+        gh pr create --title "Release ${RELEASE_DATE}" \
                      --body-file "body.md" \
-                     --head "releases/${{ steps.date.outputs.date }}" \
+                     --head "${RELEASE_BRANCH}" \
                      --base "release"
+
+  create-proxy-release-branch:
+    if: ${{ github.event.schedule == '0 6 * * THU' || format('{0}', inputs.create-proxy-release-branch) == 'true' }}
+    runs-on: ubuntu-latest
+
+    permissions:
+      contents: write # for `git push`
+
+    steps:
+    - name: Check out code
+      uses: actions/checkout@v4
+      with:
+        ref: main
+
+    - name: Set environment variables
+      run: |
+        echo "RELEASE_DATE=$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
+        echo "RELEASE_BRANCH=rc/proxy/$(date +'%Y-%m-%d')" | tee -a $GITHUB_ENV
+
+    - name: Create release branch
+      run: git checkout -b $RELEASE_BRANCH
+
+    - name: Push new branch
+      run: git push origin $RELEASE_BRANCH
+
+    - name: Create pull request into release
+      env:
+        GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+      run: |
+        cat << EOF > body.md
+          ## Proxy release ${RELEASE_DATE}
+
+          **Please merge this Pull Request using 'Create a merge commit' button**
+        EOF
+
+        gh pr create --title "Proxy release ${RELEASE_DATE}}" \
+                     --body-file "body.md" \
+                     --head "${RELEASE_BRANCH}" \
+                     --base "release-proxy"

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

@@ -9,7 +9,7 @@ on:
 defaults:
   run:
     shell: bash -euxo pipefail {0}
-    
+
 env:
   # A concurrency group that we use for e2e-tests runs, matches `concurrency.group` above with `github.repository` as a prefix
   E2E_CONCURRENCY_GROUP: ${{ github.repository }}-e2e-tests-${{ github.ref_name }}-${{ github.ref_name == 'main' && github.sha || 'anysha' }}
@@ -37,7 +37,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -51,6 +51,8 @@ jobs:
             echo "tag=$(git rev-list --count HEAD)" | tee -a $GITHUB_OUTPUT
           elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
             echo "tag=release-$(git rev-list --count HEAD)" | tee -a $GITHUB_OUTPUT
+          elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
+            echo "tag=release-proxy-$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
           else
             echo "GITHUB_REF_NAME (value '$GITHUB_REF_NAME') is not set to either 'main' or 'release'"
             BUILD_AND_TEST_RUN_ID=$(gh run list -b $CURRENT_BRANCH -c $CURRENT_SHA -w 'Build and Test' -L 1 --json databaseId --jq '.[].databaseId')
@@ -115,4 +117,3 @@ jobs:
                 \"concurrency_group\": \"${{ env.E2E_CONCURRENCY_GROUP }}\"
               }
             }"
- 

CODEOWNERS

@@ -1,10 +1,10 @@
 /compute_tools/ @neondatabase/control-plane @neondatabase/compute
-/control_plane/ @neondatabase/compute @neondatabase/storage
-/libs/pageserver_api/ @neondatabase/compute @neondatabase/storage
+/control_plane/attachment_service @neondatabase/storage
+/libs/pageserver_api/ @neondatabase/storage
 /libs/postgres_ffi/ @neondatabase/compute
 /libs/remote_storage/ @neondatabase/storage
 /libs/safekeeper_api/ @neondatabase/safekeepers
-/libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
+/libs/vm_monitor/ @neondatabase/autoscaling
 /pageserver/ @neondatabase/storage
 /pgxn/ @neondatabase/compute
 /proxy/ @neondatabase/proxy

Cargo.lock

@@ -284,8 +284,10 @@ dependencies = [
  "diesel_migrations",
  "futures",
  "git-version",
+ "humantime",
  "hyper",
  "metrics",
+ "once_cell",
  "pageserver_api",
  "pageserver_client",
  "postgres_connection",
@@ -1813,6 +1815,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e875f1719c16de097dee81ed675e2d9bb63096823ed3f0ca827b7dea3028bbbb"
 dependencies = [
  "enumset_derive",
+ "serde",
 ]
 
 [[package]]
@@ -2757,6 +2760,17 @@ version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "830d08ce1d1d941e6b30645f1a0eb5643013d835ce3779a5fc208261dbe10f55"
 
+[[package]]
+name = "leaky-bucket"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8eb491abd89e9794d50f93c8db610a29509123e3fbbc9c8c67a528e9391cd853"
+dependencies = [
+ "parking_lot 0.12.1",
+ "tokio",
+ "tracing",
+]
+
 [[package]]
 name = "libc"
 version = "0.2.150"
@@ -3448,6 +3462,7 @@ name = "pageserver"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "arc-swap",
  "async-compression",
  "async-stream",
  "async-trait",
@@ -3475,6 +3490,7 @@ dependencies = [
  "humantime-serde",
  "hyper",
  "itertools",
+ "leaky-bucket",
  "md5",
  "metrics",
  "nix 0.27.1",
@@ -3536,7 +3552,9 @@ dependencies = [
  "const_format",
  "enum-map",
  "hex",
+ "humantime",
  "humantime-serde",
+ "itertools",
  "postgres_ffi",
  "rand 0.8.5",
  "serde",
@@ -4436,6 +4454,7 @@ dependencies = [
  "futures",
  "futures-util",
  "http-types",
+ "humantime",
  "hyper",
  "itertools",
  "metrics",
@@ -4447,6 +4466,7 @@ dependencies = [
  "serde_json",
  "test-context",
  "tokio",
+ "tokio-stream",
  "tokio-util",
  "toml_edit",
  "tracing",
@@ -6345,6 +6365,7 @@ dependencies = [
  "hex-literal",
  "hyper",
  "jsonwebtoken",
+ "leaky-bucket",
  "metrics",
  "nix 0.27.1",
  "once_cell",

Cargo.toml

@@ -97,6 +97,7 @@ ipnet = "2.9.0"
 itertools = "0.10"
 jsonwebtoken = "9"
 lasso = "0.7"
+leaky-bucket = "1.0.1"
 libc = "0.2"
 md5 = "0.7.0"
 memoffset = "0.8"

Dockerfile

@@ -47,7 +47,7 @@ COPY --chown=nonroot . .
 # Show build caching stats to check if it was used in the end.
 # Has to be the part of the same RUN since cachepot daemon is killed in the end of this RUN, losing the compilation stats.
 RUN set -e \
-    && mold -run cargo build  \
+    && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment" cargo build  \
       --bin pg_sni_router  \
       --bin pageserver  \
       --bin pagectl  \

Dockerfile.compute-node

@@ -769,6 +769,40 @@ RUN wget https://github.com/eulerto/wal2json/archive/refs/tags/wal2json_2_5.tar.
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install
 
+#########################################################################################
+#
+# Layer "pg_ivm"
+# compile pg_ivm extension
+#
+#########################################################################################
+FROM build-deps AS pg-ivm-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/sraoss/pg_ivm/archive/refs/tags/v1.7.tar.gz -O pg_ivm.tar.gz && \
+    echo "ebfde04f99203c7be4b0e873f91104090e2e83e5429c32ac242d00f334224d5e pg_ivm.tar.gz" | sha256sum --check && \
+    mkdir pg_ivm-src && cd pg_ivm-src && tar xvzf ../pg_ivm.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_ivm.control
+
+#########################################################################################
+#
+# Layer "pg_partman"
+# compile pg_partman extension
+#
+#########################################################################################
+FROM build-deps AS pg-partman-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/pgpartman/pg_partman/archive/refs/tags/v5.0.1.tar.gz -O pg_partman.tar.gz && \
+    echo "75b541733a9659a6c90dbd40fccb904a630a32880a6e3044d0c4c5f4c8a65525 pg_partman.tar.gz" | sha256sum --check && \
+    mkdir pg_partman-src && cd pg_partman-src && tar xvzf ../pg_partman.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_partman.control
+
 #########################################################################################
 #
 # Layer "neon-pg-ext-build"
@@ -810,6 +844,8 @@ COPY --from=pg-semver-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-embedding-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
 COPY --from=pg-anon-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-ivm-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=pg-partman-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY pgxn/ pgxn/
 
 RUN make -j $(getconf _NPROCESSORS_ONLN) \
@@ -820,6 +856,10 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) \
         PG_CONFIG=/usr/local/pgsql/bin/pg_config \
         -C pgxn/neon_utils \
         -s install && \
+    make -j $(getconf _NPROCESSORS_ONLN) \
+        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
+        -C pgxn/neon_test_utils \
+        -s install && \
     make -j $(getconf _NPROCESSORS_ONLN) \
         PG_CONFIG=/usr/local/pgsql/bin/pg_config \
         -C pgxn/neon_rmgr \

Makefile

@@ -159,8 +159,8 @@ neon-pg-ext-%: postgres-%
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile install
 
-.PHONY: neon-pg-ext-clean-%
-neon-pg-ext-clean-%:
+.PHONY: neon-pg-clean-ext-%
+neon-pg-clean-ext-%:
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config \
 	-C $(POSTGRES_INSTALL_DIR)/build/neon-$* \
 	-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile clean
@@ -216,11 +216,11 @@ neon-pg-ext: \
 	neon-pg-ext-v15 \
 	neon-pg-ext-v16
 
-.PHONY: neon-pg-ext-clean
-neon-pg-ext-clean: \
-	neon-pg-ext-clean-v14 \
-	neon-pg-ext-clean-v15 \
-	neon-pg-ext-clean-v16
+.PHONY: neon-pg-clean-ext
+neon-pg-clean-ext: \
+	neon-pg-clean-ext-v14 \
+	neon-pg-clean-ext-v15 \
+	neon-pg-clean-ext-v16
 
 # shorthand to build all Postgres versions
 .PHONY: postgres
@@ -249,7 +249,7 @@ postgres-check: \
 
 # This doesn't remove the effects of 'configure'.
 .PHONY: clean
-clean: postgres-clean neon-pg-ext-clean
+clean: postgres-clean neon-pg-clean-ext
 	$(CARGO_CMD_PREFIX) cargo clean
 
 # This removes everything

README.md

@@ -5,7 +5,7 @@
 Neon is a serverless open-source alternative to AWS Aurora Postgres. It separates storage and compute and substitutes the PostgreSQL storage layer by redistributing data across a cluster of nodes.
 
 ## Quick start
-Try the [Neon Free Tier](https://neon.tech/docs/introduction/technical-preview-free-tier/) to create a serverless Postgres instance. Then connect to it with your preferred Postgres client (psql, dbeaver, etc) or use the online [SQL Editor](https://neon.tech/docs/get-started-with-neon/query-with-neon-sql-editor/). See [Connect from any application](https://neon.tech/docs/connect/connect-from-any-app/) for connection instructions.
+Try the [Neon Free Tier](https://neon.tech) to create a serverless Postgres instance. Then connect to it with your preferred Postgres client (psql, dbeaver, etc) or use the online [SQL Editor](https://neon.tech/docs/get-started-with-neon/query-with-neon-sql-editor/). See [Connect from any application](https://neon.tech/docs/connect/connect-from-any-app/) for connection instructions.
 
 Alternatively, compile and run the project [locally](#running-local-installation).
 
@@ -249,6 +249,16 @@ testing locally, it is convenient to run just one set of permutations, like this
 DEFAULT_PG_VERSION=15 BUILD_TYPE=release ./scripts/pytest
 ```
 
+## Flamegraphs
+
+You may find yourself in need of flamegraphs for software in this repository.
+You can use [`flamegraph-rs`](https://github.com/flamegraph-rs/flamegraph) or the original [`flamegraph.pl`](https://github.com/brendangregg/FlameGraph). Your choice!
+
+>[!IMPORTANT]
+> If you're using `lld` or `mold`, you need the `--no-rosegment` linker argument.
+> It's a [general thing with Rust / lld / mold](https://crbug.com/919499#c16), not specific to this repository.
+> See [this PR for further instructions](https://github.com/neondatabase/neon/pull/6764).
+
 ## Documentation
 
 [docs](/docs) Contains a top-level overview of all available markdown documentation.

compute_tools/src/bin/compute_ctl.rs

@@ -45,7 +45,6 @@ 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};
@@ -53,7 +52,9 @@ use url::Url;
 
 use compute_api::responses::ComputeStatus;
 
-use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec, PG_PID, SYNC_SAFEKEEPERS_PID};
+use compute_tools::compute::{
+    forward_termination_signal, ComputeNode, ComputeState, ParsedSpec, PG_PID,
+};
 use compute_tools::configurator::launch_configurator;
 use compute_tools::extension_server::get_pg_version;
 use compute_tools::http::api::launch_http_server;
@@ -394,6 +395,15 @@ fn main() -> Result<()> {
         info!("synced safekeepers at lsn {lsn}");
     }
 
+    let mut state = compute.state.lock().unwrap();
+    if state.status == ComputeStatus::TerminationPending {
+        state.status = ComputeStatus::Terminated;
+        compute.state_changed.notify_all();
+        // we were asked to terminate gracefully, don't exit to avoid restart
+        delay_exit = true
+    }
+    drop(state);
+
     if let Err(err) = compute.check_for_core_dumps() {
         error!("error while checking for core dumps: {err:?}");
     }
@@ -523,16 +533,7 @@ fn cli() -> clap::Command {
 /// 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();
-    }
+    forward_termination_signal();
     exit(1);
 }
 

compute_tools/src/compute.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 use std::env;
 use std::fs;
 use std::io::BufRead;
-use std::os::unix::fs::PermissionsExt;
+use std::os::unix::fs::{symlink, PermissionsExt};
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
@@ -28,6 +28,8 @@ use compute_api::responses::{ComputeMetrics, ComputeStatus};
 use compute_api::spec::{ComputeFeature, ComputeMode, ComputeSpec};
 use utils::measured_stream::MeasuredReader;
 
+use nix::sys::signal::{kill, Signal};
+
 use remote_storage::{DownloadError, RemotePath};
 
 use crate::checker::create_availability_check_data;
@@ -324,7 +326,8 @@ impl ComputeNode {
         let spec = compute_state.pspec.as_ref().expect("spec must be set");
         let start_time = Instant::now();
 
-        let mut config = postgres::Config::from_str(&spec.pageserver_connstr)?;
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let mut config = postgres::Config::from_str(shard0_connstr)?;
 
         // Use the storage auth token from the config file, if given.
         // Note: this overrides any password set in the connection string.
@@ -634,6 +637,48 @@ impl ComputeNode {
         // Update pg_hba.conf received with basebackup.
         update_pg_hba(pgdata_path)?;
 
+        // Place pg_dynshmem under /dev/shm. This allows us to use
+        // 'dynamic_shared_memory_type = mmap' so that the files are placed in
+        // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
+        //
+        // Why on earth don't we just stick to the 'posix' default, you might
+        // ask.  It turns out that making large allocations with 'posix' doesn't
+        // work very well with autoscaling. The behavior we want is that:
+        //
+        // 1. You can make large DSM allocations, larger than the current RAM
+        //    size of the VM, without errors
+        //
+        // 2. If the allocated memory is really used, the VM is scaled up
+        //    automatically to accommodate that
+        //
+        // We try to make that possible by having swap in the VM. But with the
+        // default 'posix' DSM implementation, we fail step 1, even when there's
+        // plenty of swap available. PostgreSQL uses posix_fallocate() to create
+        // the shmem segment, which is really just a file in /dev/shm in Linux,
+        // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
+        // than available RAM.
+        //
+        // Using 'dynamic_shared_memory_type = mmap' works around that, because
+        // the Postgres 'mmap' DSM implementation doesn't use
+        // posix_fallocate(). Instead, it uses repeated calls to write(2) to
+        // fill the file with zeros. It's weird that that differs between
+        // 'posix' and 'mmap', but we take advantage of it. When the file is
+        // filled slowly with write(2), the kernel allows it to grow larger, as
+        // long as there's swap available.
+        //
+        // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
+        // segment to be larger than currently available RAM. But because we
+        // don't want to store it on a real file, which the kernel would try to
+        // flush to disk, so symlink pg_dynshm to /dev/shm.
+        //
+        // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
+        // control plane control that option. If 'mmap' is not used, this
+        // symlink doesn't affect anything.
+        //
+        // See https://github.com/neondatabase/autoscaling/issues/800
+        std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
+        symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
+
         match spec.mode {
             ComputeMode::Primary => {}
             ComputeMode::Replica | ComputeMode::Static(..) => {
@@ -1279,3 +1324,17 @@ LIMIT 100",
         Ok(remote_ext_metrics)
     }
 }
+
+pub fn forward_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);
+        // use 'immediate' shutdown (SIGQUIT): https://www.postgresql.org/docs/current/server-shutdown.html
+        kill(pg_pid, Signal::SIGQUIT).ok();
+    }
+}

compute_tools/src/config.rs

@@ -51,6 +51,9 @@ pub fn write_postgres_conf(
     if let Some(s) = &spec.pageserver_connstring {
         writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
     }
+    if let Some(stripe_size) = spec.shard_stripe_size {
+        writeln!(file, "neon.stripe_size={stripe_size}")?;
+    }
     if !spec.safekeeper_connstrings.is_empty() {
         writeln!(
             file,
@@ -79,6 +82,12 @@ pub fn write_postgres_conf(
         ComputeMode::Replica => {
             // hot_standby is 'on' by default, but let's be explicit
             writeln!(file, "hot_standby=on")?;
+
+            // Inform the replica about the primary state
+            // Default is 'false'
+            if let Some(primary_is_running) = spec.primary_is_running {
+                writeln!(file, "neon.primary_is_running={}", primary_is_running)?;
+            }
         }
     }
 

compute_tools/src/http/api.rs

@@ -5,6 +5,7 @@ use std::net::SocketAddr;
 use std::sync::Arc;
 use std::thread;
 
+use crate::compute::forward_termination_signal;
 use crate::compute::{ComputeNode, ComputeState, ParsedSpec};
 use compute_api::requests::ConfigurationRequest;
 use compute_api::responses::{ComputeStatus, ComputeStatusResponse, GenericAPIError};
@@ -123,6 +124,17 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
             }
         }
 
+        (&Method::POST, "/terminate") => {
+            info!("serving /terminate POST request");
+            match handle_terminate_request(compute).await {
+                Ok(()) => Response::new(Body::empty()),
+                Err((msg, code)) => {
+                    error!("error handling /terminate request: {msg}");
+                    render_json_error(&msg, code)
+                }
+            }
+        }
+
         // download extension files from remote extension storage on demand
         (&Method::POST, route) if route.starts_with("/extension_server/") => {
             info!("serving {:?} POST request", route);
@@ -297,6 +309,49 @@ fn render_json_error(e: &str, status: StatusCode) -> Response<Body> {
         .unwrap()
 }
 
+async fn handle_terminate_request(compute: &Arc<ComputeNode>) -> Result<(), (String, StatusCode)> {
+    {
+        let mut state = compute.state.lock().unwrap();
+        if state.status == ComputeStatus::Terminated {
+            return Ok(());
+        }
+        if state.status != ComputeStatus::Empty && state.status != ComputeStatus::Running {
+            let msg = format!(
+                "invalid compute status for termination request: {:?}",
+                state.status.clone()
+            );
+            return Err((msg, StatusCode::PRECONDITION_FAILED));
+        }
+        state.status = ComputeStatus::TerminationPending;
+        compute.state_changed.notify_all();
+        drop(state);
+    }
+    forward_termination_signal();
+    info!("sent signal and notified waiters");
+
+    // Spawn a blocking thread to wait for compute to become Terminated.
+    // This is needed to do not block the main pool of workers and
+    // be able to serve other requests while some particular request
+    // is waiting for compute to finish configuration.
+    let c = compute.clone();
+    task::spawn_blocking(move || {
+        let mut state = c.state.lock().unwrap();
+        while state.status != ComputeStatus::Terminated {
+            state = c.state_changed.wait(state).unwrap();
+            info!(
+                "waiting for compute to become Terminated, current status: {:?}",
+                state.status
+            );
+        }
+
+        Ok(())
+    })
+    .await
+    .unwrap()?;
+    info!("terminated Postgres");
+    Ok(())
+}
+
 // Main Hyper HTTP server function that runs it and blocks waiting on it forever.
 #[tokio::main]
 async fn serve(port: u16, state: Arc<ComputeNode>) {

compute_tools/src/http/openapi_spec.yaml

@@ -168,6 +168,29 @@ paths:
               schema:
                 $ref: "#/components/schemas/GenericError"
 
+  /terminate:
+    post:
+      tags:
+      - Terminate
+      summary: Terminate Postgres and wait for it to exit
+      description: ""
+      operationId: terminate
+      responses:
+        200:
+          description: Result
+        412:
+          description: "wrong state"
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
+        500:
+          description: "Unexpected error"
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
+
 components:
   securitySchemes:
     JWT:

compute_tools/src/spec.rs

@@ -655,6 +655,9 @@ pub fn handle_grants(
         // remove this code if possible. The worst thing that could happen is that
         // user won't be able to use public schema in NEW databases created in the
         // very OLD project.
+        //
+        // Also, alter default permissions so that relations created by extensions can be
+        // used by neon_superuser without permission issues.
         let grant_query = "DO $$\n\
                 BEGIN\n\
                     IF EXISTS(\n\
@@ -673,6 +676,8 @@ pub fn handle_grants(
                             GRANT CREATE ON SCHEMA public TO web_access;\n\
                         END IF;\n\
                     END IF;\n\
+                    ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser WITH GRANT OPTION;\n\
+                    ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser WITH GRANT OPTION;\n\
                 END\n\
             $$;"
         .to_string();
@@ -777,6 +782,12 @@ BEGIN
 END
 $$;"#,
         "GRANT pg_monitor TO neon_superuser WITH ADMIN OPTION",
+        // Don't remove: these are some SQLs that we originally applied in migrations but turned out to execute somewhere else.
+        "",
+        "",
+        "",
+        "",
+        // Add new migrations below.
     ];
 
     let mut query = "CREATE SCHEMA IF NOT EXISTS neon_migration";
@@ -803,8 +814,13 @@ $$;"#,
     client.simple_query(query)?;
 
     while current_migration < migrations.len() {
-        info!("Running migration:\n{}\n", migrations[current_migration]);
-        client.simple_query(migrations[current_migration])?;
+        let migration = &migrations[current_migration];
+        if migration.is_empty() {
+            info!("Skip migration id={}", current_migration);
+        } else {
+            info!("Running migration:\n{}\n", migration);
+            client.simple_query(migration)?;
+        }
         current_migration += 1;
     }
     let setval = format!(

control_plane/attachment_service/Cargo.toml

@@ -4,6 +4,11 @@ version = "0.1.0"
 edition.workspace = true
 license.workspace = true
 
+[features]
+default = []
+# Enables test-only APIs and behaviors
+testing = []
+
 [dependencies]
 anyhow.workspace = true
 aws-config.workspace = true
@@ -13,6 +18,8 @@ clap.workspace = true
 futures.workspace = true
 git-version.workspace = true
 hyper.workspace = true
+humantime.workspace = true
+once_cell.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 postgres_connection.workspace = true

control_plane/attachment_service/src/compute_hook.rs

@@ -3,7 +3,7 @@ use std::{collections::HashMap, time::Duration};
 use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
 use control_plane::local_env::LocalEnv;
 use hyper::{Method, StatusCode};
-use pageserver_api::shard::{ShardCount, ShardIndex, ShardNumber, TenantShardId};
+use pageserver_api::shard::{ShardIndex, ShardNumber, TenantShardId};
 use postgres_connection::parse_host_port;
 use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
@@ -77,7 +77,7 @@ impl ComputeHookTenant {
         self.shards
             .sort_by_key(|(shard, _node_id)| shard.shard_number);
 
-        if self.shards.len() == shard_count.0 as usize || shard_count == ShardCount(0) {
+        if self.shards.len() == shard_count.count() as usize || shard_count.is_unsharded() {
             // We have pageservers for all the shards: emit a configuration update
             return Some(ComputeHookNotifyRequest {
                 tenant_id,
@@ -94,7 +94,7 @@ impl ComputeHookTenant {
             tracing::info!(
                 "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
                 self.shards.len(),
-                shard_count.0
+                shard_count.count()
             );
         }
 
@@ -155,7 +155,7 @@ impl ComputeHook {
 
         for (endpoint_name, endpoint) in &cplane.endpoints {
             if endpoint.tenant_id == tenant_id && endpoint.status() == EndpointStatus::Running {
-                tracing::info!("🔁 Reconfiguring endpoint {}", endpoint_name,);
+                tracing::info!("Reconfiguring endpoint {}", endpoint_name,);
                 endpoint.reconfigure(compute_pageservers.clone()).await?;
             }
         }
@@ -177,7 +177,7 @@ impl ComputeHook {
             req
         };
 
-        tracing::debug!(
+        tracing::info!(
             "Sending notify request to {} ({:?})",
             url,
             reconfigure_request
@@ -266,7 +266,7 @@ impl ComputeHook {
     /// periods, but we don't retry forever.  The **caller** is responsible for handling failures and
     /// ensuring that they eventually call again to ensure that the compute is eventually notified of
     /// the proper pageserver nodes for a tenant.
-    #[tracing::instrument(skip_all, fields(tenant_shard_id, node_id))]
+    #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), node_id))]
     pub(super) async fn notify(
         &self,
         tenant_shard_id: TenantShardId,
@@ -298,7 +298,7 @@ impl ComputeHook {
         let Some(reconfigure_request) = reconfigure_request else {
             // The tenant doesn't yet have pageservers for all its shards: we won't notify anything
             // until it does.
-            tracing::debug!("Tenant isn't yet ready to emit a notification",);
+            tracing::info!("Tenant isn't yet ready to emit a notification");
             return Ok(());
         };
 

control_plane/attachment_service/src/http.rs

@@ -4,7 +4,7 @@ use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
 use pageserver_api::models::{
     TenantCreateRequest, TenantLocationConfigRequest, TenantShardSplitRequest,
-    TimelineCreateRequest,
+    TenantTimeTravelRequest, TimelineCreateRequest,
 };
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api;
@@ -12,7 +12,7 @@ use std::sync::Arc;
 use std::time::{Duration, Instant};
 use utils::auth::SwappableJwtAuth;
 use utils::http::endpoint::{auth_middleware, request_span};
-use utils::http::request::parse_request_param;
+use utils::http::request::{must_get_query_param, parse_request_param};
 use utils::id::{TenantId, TimelineId};
 
 use utils::{
@@ -66,14 +66,7 @@ fn get_state(request: &Request<Body>) -> &HttpState {
 async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
     let reattach_req = json_request::<ReAttachRequest>(&mut req).await?;
     let state = get_state(&req);
-    json_response(
-        StatusCode::OK,
-        state
-            .service
-            .re_attach(reattach_req)
-            .await
-            .map_err(ApiError::InternalServerError)?,
-    )
+    json_response(StatusCode::OK, state.service.re_attach(reattach_req).await?)
 }
 
 /// Pageserver calls into this before doing deletions, to confirm that it still
@@ -114,7 +107,10 @@ async fn handle_tenant_create(
     mut req: Request<Body>,
 ) -> Result<Response<Body>, ApiError> {
     let create_req = json_request::<TenantCreateRequest>(&mut req).await?;
-    json_response(StatusCode::OK, service.tenant_create(create_req).await?)
+    json_response(
+        StatusCode::CREATED,
+        service.tenant_create(create_req).await?,
+    )
 }
 
 // For tenant and timeline deletions, which both implement an "initially return 202, then 404 once
@@ -177,6 +173,39 @@ async fn handle_tenant_location_config(
     )
 }
 
+async fn handle_tenant_time_travel_remote_storage(
+    service: Arc<Service>,
+    mut req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let time_travel_req = json_request::<TenantTimeTravelRequest>(&mut req).await?;
+
+    let timestamp_raw = must_get_query_param(&req, "travel_to")?;
+    let _timestamp = humantime::parse_rfc3339(&timestamp_raw).map_err(|_e| {
+        ApiError::BadRequest(anyhow::anyhow!(
+            "Invalid time for travel_to: {timestamp_raw:?}"
+        ))
+    })?;
+
+    let done_if_after_raw = must_get_query_param(&req, "done_if_after")?;
+    let _done_if_after = humantime::parse_rfc3339(&done_if_after_raw).map_err(|_e| {
+        ApiError::BadRequest(anyhow::anyhow!(
+            "Invalid time for done_if_after: {done_if_after_raw:?}"
+        ))
+    })?;
+
+    service
+        .tenant_time_travel_remote_storage(
+            &time_travel_req,
+            tenant_id,
+            timestamp_raw,
+            done_if_after_raw,
+        )
+        .await?;
+
+    json_response(StatusCode::OK, ())
+}
+
 async fn handle_tenant_delete(
     service: Arc<Service>,
     req: Request<Body>,
@@ -196,7 +225,7 @@ async fn handle_tenant_timeline_create(
     let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
     let create_req = json_request::<TimelineCreateRequest>(&mut req).await?;
     json_response(
-        StatusCode::OK,
+        StatusCode::CREATED,
         service
             .tenant_timeline_create(tenant_id, create_req)
             .await?,
@@ -296,7 +325,10 @@ async fn handle_node_configure(mut req: Request<Body>) -> Result<Response<Body>,
     }
     let state = get_state(&req);
 
-    json_response(StatusCode::OK, state.service.node_configure(config_req)?)
+    json_response(
+        StatusCode::OK,
+        state.service.node_configure(config_req).await?,
+    )
 }
 
 async fn handle_tenant_shard_split(
@@ -333,6 +365,22 @@ async fn handle_tenant_drop(req: Request<Body>) -> Result<Response<Body>, ApiErr
     json_response(StatusCode::OK, state.service.tenant_drop(tenant_id).await?)
 }
 
+async fn handle_tenants_dump(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&req);
+    state.service.tenants_dump()
+}
+
+async fn handle_scheduler_dump(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&req);
+    state.service.scheduler_dump()
+}
+
+async fn handle_consistency_check(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.consistency_check().await?)
+}
+
 /// Status endpoint is just used for checking that our HTTP listener is up
 async fn handle_status(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
     json_response(StatusCode::OK, ())
@@ -421,6 +469,13 @@ pub fn make_router(
         .post("/debug/v1/node/:node_id/drop", |r| {
             request_span(r, handle_node_drop)
         })
+        .get("/debug/v1/tenant", |r| request_span(r, handle_tenants_dump))
+        .get("/debug/v1/scheduler", |r| {
+            request_span(r, handle_scheduler_dump)
+        })
+        .post("/debug/v1/consistency_check", |r| {
+            request_span(r, handle_consistency_check)
+        })
         .get("/control/v1/tenant/:tenant_id/locate", |r| {
             tenant_service_handler(r, handle_tenant_locate)
         })
@@ -451,6 +506,9 @@ pub fn make_router(
         .put("/v1/tenant/:tenant_id/location_config", |r| {
             tenant_service_handler(r, handle_tenant_location_config)
         })
+        .put("/v1/tenant/:tenant_id/time_travel_remote_storage", |r| {
+            tenant_service_handler(r, handle_tenant_time_travel_remote_storage)
+        })
         // Timeline operations
         .delete("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
             tenant_service_handler(r, handle_tenant_timeline_delete)

control_plane/attachment_service/src/lib.rs

@@ -3,6 +3,7 @@ use utils::seqwait::MonotonicCounter;
 
 mod compute_hook;
 pub mod http;
+pub mod metrics;
 mod node;
 pub mod persistence;
 mod reconciler;
@@ -11,7 +12,7 @@ mod schema;
 pub mod service;
 mod tenant_state;
 
-#[derive(Clone, Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize, Debug)]
 enum PlacementPolicy {
     /// Cheapest way to attach a tenant: just one pageserver, no secondary
     Single,
@@ -22,7 +23,7 @@ enum PlacementPolicy {
     Detached,
 }
 
-#[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone)]
+#[derive(Ord, PartialOrd, Eq, PartialEq, Copy, Clone, Serialize)]
 struct Sequence(u64);
 
 impl Sequence {
@@ -37,6 +38,12 @@ impl std::fmt::Display for Sequence {
     }
 }
 
+impl std::fmt::Debug for Sequence {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
 impl MonotonicCounter<Sequence> for Sequence {
     fn cnt_advance(&mut self, v: Sequence) {
         assert!(*self <= v);

control_plane/attachment_service/src/main.rs

@@ -6,6 +6,7 @@
 ///
 use anyhow::{anyhow, Context};
 use attachment_service::http::make_router;
+use attachment_service::metrics::preinitialize_metrics;
 use attachment_service::persistence::Persistence;
 use attachment_service::service::{Config, Service};
 use aws_config::{self, BehaviorVersion, Region};
@@ -15,6 +16,7 @@ use diesel::Connection;
 use metrics::launch_timestamp::LaunchTimestamp;
 use std::sync::Arc;
 use tokio::signal::unix::SignalKind;
+use tokio_util::sync::CancellationToken;
 use utils::auth::{JwtAuth, SwappableJwtAuth};
 use utils::logging::{self, LogFormat};
 
@@ -204,6 +206,8 @@ async fn async_main() -> anyhow::Result<()> {
         logging::Output::Stdout,
     )?;
 
+    preinitialize_metrics();
+
     let args = Cli::parse();
     tracing::info!(
         "version: {}, launch_timestamp: {}, build_tag {}, state at {}, listening on {}",
@@ -237,15 +241,23 @@ async fn async_main() -> anyhow::Result<()> {
     let auth = secrets
         .public_key
         .map(|jwt_auth| Arc::new(SwappableJwtAuth::new(jwt_auth)));
-    let router = make_router(service, auth)
+    let router = make_router(service.clone(), auth)
         .build()
         .map_err(|err| anyhow!(err))?;
     let router_service = utils::http::RouterService::new(router).unwrap();
-    let server = hyper::Server::from_tcp(http_listener)?.serve(router_service);
 
+    // Start HTTP server
+    let server_shutdown = CancellationToken::new();
+    let server = hyper::Server::from_tcp(http_listener)?
+        .serve(router_service)
+        .with_graceful_shutdown({
+            let server_shutdown = server_shutdown.clone();
+            async move {
+                server_shutdown.cancelled().await;
+            }
+        });
     tracing::info!("Serving on {0}", args.listen);
-
-    tokio::task::spawn(server);
+    let server_task = tokio::task::spawn(server);
 
     // Wait until we receive a signal
     let mut sigint = tokio::signal::unix::signal(SignalKind::interrupt())?;
@@ -266,5 +278,16 @@ async fn async_main() -> anyhow::Result<()> {
         }
     }
 
+    // Stop HTTP server first, so that we don't have to service requests
+    // while shutting down Service
+    server_shutdown.cancel();
+    if let Err(e) = server_task.await {
+        tracing::error!("Error joining HTTP server task: {e}")
+    }
+    tracing::info!("Joined HTTP server task");
+
+    service.shutdown().await;
+    tracing::info!("Service shutdown complete");
+
     std::process::exit(0);
 }

control_plane/attachment_service/src/metrics.rs

@@ -0,0 +1,32 @@
+use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
+use once_cell::sync::Lazy;
+
+pub(crate) struct ReconcilerMetrics {
+    pub(crate) spawned: IntCounter,
+    pub(crate) complete: IntCounterVec,
+}
+
+impl ReconcilerMetrics {
+    // Labels used on [`Self::complete`]
+    pub(crate) const SUCCESS: &'static str = "ok";
+    pub(crate) const ERROR: &'static str = "success";
+    pub(crate) const CANCEL: &'static str = "cancel";
+}
+
+pub(crate) static RECONCILER: Lazy<ReconcilerMetrics> = Lazy::new(|| ReconcilerMetrics {
+    spawned: register_int_counter!(
+        "storage_controller_reconcile_spawn",
+        "Count of how many times we spawn a reconcile task",
+    )
+    .expect("failed to define a metric"),
+    complete: register_int_counter_vec!(
+        "storage_controller_reconcile_complete",
+        "Reconciler tasks completed, broken down by success/failure/cancelled",
+        &["status"],
+    )
+    .expect("failed to define a metric"),
+});
+
+pub fn preinitialize_metrics() {
+    Lazy::force(&RECONCILER);
+}

control_plane/attachment_service/src/node.rs

@@ -1,9 +1,16 @@
 use control_plane::attachment_service::{NodeAvailability, NodeSchedulingPolicy};
+use serde::Serialize;
 use utils::id::NodeId;
 
 use crate::persistence::NodePersistence;
 
-#[derive(Clone)]
+/// Represents the in-memory description of a Node.
+///
+/// Scheduling statistics are maintened separately in [`crate::scheduler`].
+///
+/// The persistent subset of the Node is defined in [`crate::persistence::NodePersistence`]: the
+/// implementation of serialization on this type is only for debug dumps.
+#[derive(Clone, Serialize)]
 pub(crate) struct Node {
     pub(crate) id: NodeId,
 

control_plane/attachment_service/src/persistence.rs

@@ -6,7 +6,7 @@ use std::time::Duration;
 use self::split_state::SplitState;
 use camino::Utf8Path;
 use camino::Utf8PathBuf;
-use control_plane::attachment_service::{NodeAvailability, NodeSchedulingPolicy};
+use control_plane::attachment_service::NodeSchedulingPolicy;
 use diesel::pg::PgConnection;
 use diesel::prelude::*;
 use diesel::Connection;
@@ -130,24 +130,10 @@ impl Persistence {
     }
 
     /// At startup, populate the list of nodes which our shards may be placed on
-    pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<Node>> {
-        let nodes: Vec<Node> = self
+    pub(crate) async fn list_nodes(&self) -> DatabaseResult<Vec<NodePersistence>> {
+        let nodes: Vec<NodePersistence> = self
             .with_conn(move |conn| -> DatabaseResult<_> {
-                Ok(crate::schema::nodes::table
-                    .load::<NodePersistence>(conn)?
-                    .into_iter()
-                    .map(|n| Node {
-                        id: NodeId(n.node_id as u64),
-                        // At startup we consider a node offline until proven otherwise.
-                        availability: NodeAvailability::Offline,
-                        scheduling: NodeSchedulingPolicy::from_str(&n.scheduling_policy)
-                            .expect("Bad scheduling policy in DB"),
-                        listen_http_addr: n.listen_http_addr,
-                        listen_http_port: n.listen_http_port as u16,
-                        listen_pg_addr: n.listen_pg_addr,
-                        listen_pg_port: n.listen_pg_port as u16,
-                    })
-                    .collect::<Vec<Node>>())
+                Ok(crate::schema::nodes::table.load::<NodePersistence>(conn)?)
             })
             .await?;
 
@@ -156,6 +142,31 @@ impl Persistence {
         Ok(nodes)
     }
 
+    pub(crate) async fn update_node(
+        &self,
+        input_node_id: NodeId,
+        input_scheduling: NodeSchedulingPolicy,
+    ) -> DatabaseResult<()> {
+        use crate::schema::nodes::dsl::*;
+        let updated = self
+            .with_conn(move |conn| {
+                let updated = diesel::update(nodes)
+                    .filter(node_id.eq(input_node_id.0 as i64))
+                    .set((scheduling_policy.eq(String::from(input_scheduling)),))
+                    .execute(conn)?;
+                Ok(updated)
+            })
+            .await?;
+
+        if updated != 1 {
+            Err(DatabaseError::Logical(format!(
+                "Node {node_id:?} not found for update",
+            )))
+        } else {
+            Ok(())
+        }
+    }
+
     /// At startup, load the high level state for shards, such as their config + policy.  This will
     /// be enriched at runtime with state discovered on pageservers.
     pub(crate) async fn list_tenant_shards(&self) -> DatabaseResult<Vec<TenantShardPersistence>> {
@@ -222,7 +233,7 @@ impl Persistence {
             let tenant_shard_id = TenantShardId {
                 tenant_id: TenantId::from_str(tsp.tenant_id.as_str())?,
                 shard_number: ShardNumber(tsp.shard_number as u8),
-                shard_count: ShardCount(tsp.shard_count as u8),
+                shard_count: ShardCount::new(tsp.shard_count as u8),
             };
 
             tenants_map.insert(tenant_shard_id, tsp);
@@ -318,7 +329,7 @@ impl Persistence {
                 tenant_id: TenantId::from_str(tsp.tenant_id.as_str())
                     .map_err(|e| DatabaseError::Logical(format!("Malformed tenant id: {e}")))?,
                 shard_number: ShardNumber(tsp.shard_number as u8),
-                shard_count: ShardCount(tsp.shard_count as u8),
+                shard_count: ShardCount::new(tsp.shard_count as u8),
             };
             result.insert(tenant_shard_id, Generation::new(tsp.generation as u32));
         }
@@ -340,7 +351,7 @@ impl Persistence {
                 let updated = diesel::update(tenant_shards)
                     .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
                     .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
-                    .filter(shard_count.eq(tenant_shard_id.shard_count.0 as i32))
+                    .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
                     .set((
                         generation.eq(generation + 1),
                         generation_pageserver.eq(node_id.0 as i64),
@@ -362,7 +373,7 @@ impl Persistence {
             let updated = diesel::update(tenant_shards)
                 .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
                 .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
-                .filter(shard_count.eq(tenant_shard_id.shard_count.0 as i32))
+                .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
                 .set((
                     generation_pageserver.eq(i64::MAX),
                     placement_policy.eq(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
@@ -381,7 +392,6 @@ impl Persistence {
     //
     // We create the child shards here, so that they will be available for increment_generation calls
     // if some pageserver holding a child shard needs to restart before the overall tenant split is complete.
-    #[allow(dead_code)]
     pub(crate) async fn begin_shard_split(
         &self,
         old_shard_count: ShardCount,
@@ -393,21 +403,19 @@ impl Persistence {
             conn.transaction(|conn| -> DatabaseResult<()> {
                 // Mark parent shards as splitting
 
-                let expect_parent_records = std::cmp::max(1, old_shard_count.0);
-
                 let updated = diesel::update(tenant_shards)
                     .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .filter(shard_count.eq(old_shard_count.0 as i32))
+                    .filter(shard_count.eq(old_shard_count.literal() as i32))
                     .set((splitting.eq(1),))
                     .execute(conn)?;
                 if u8::try_from(updated)
                     .map_err(|_| DatabaseError::Logical(
                         format!("Overflow existing shard count {} while splitting", updated))
-                    )? != expect_parent_records {
+                    )? != old_shard_count.count() {
                     // Perhaps a deletion or another split raced with this attempt to split, mutating
                     // the parent shards that we intend to split. In this case the split request should fail.
                     return Err(DatabaseError::Logical(
-                        format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {expect_parent_records})")
+                        format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {})", old_shard_count.count())
                     ));
                 }
 
@@ -419,7 +427,7 @@ impl Persistence {
                     let mut parent = crate::schema::tenant_shards::table
                         .filter(tenant_id.eq(parent_shard_id.tenant_id.to_string()))
                         .filter(shard_number.eq(parent_shard_id.shard_number.0 as i32))
-                        .filter(shard_count.eq(parent_shard_id.shard_count.0 as i32))
+                        .filter(shard_count.eq(parent_shard_id.shard_count.literal() as i32))
                         .load::<TenantShardPersistence>(conn)?;
                     let parent = if parent.len() != 1 {
                         return Err(DatabaseError::Logical(format!(
@@ -449,7 +457,6 @@ impl Persistence {
 
     // When we finish shard splitting, we must atomically clean up the old shards
     // and insert the new shards, and clear the splitting marker.
-    #[allow(dead_code)]
     pub(crate) async fn complete_shard_split(
         &self,
         split_tenant_id: TenantId,
@@ -461,7 +468,7 @@ impl Persistence {
                 // Drop parent shards
                 diesel::delete(tenant_shards)
                     .filter(tenant_id.eq(split_tenant_id.to_string()))
-                    .filter(shard_count.eq(old_shard_count.0 as i32))
+                    .filter(shard_count.eq(old_shard_count.literal() as i32))
                     .execute(conn)?;
 
                 // Clear sharding flag
@@ -481,7 +488,7 @@ impl Persistence {
 }
 
 /// Parts of [`crate::tenant_state::TenantState`] that are stored durably
-#[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone)]
+#[derive(Queryable, Selectable, Insertable, Serialize, Deserialize, Clone, Eq, PartialEq)]
 #[diesel(table_name = crate::schema::tenant_shards)]
 pub(crate) struct TenantShardPersistence {
     #[serde(default)]
@@ -510,7 +517,7 @@ pub(crate) struct TenantShardPersistence {
 }
 
 /// Parts of [`crate::node::Node`] that are stored durably
-#[derive(Serialize, Deserialize, Queryable, Selectable, Insertable)]
+#[derive(Serialize, Deserialize, Queryable, Selectable, Insertable, Eq, PartialEq)]
 #[diesel(table_name = crate::schema::nodes)]
 pub(crate) struct NodePersistence {
     pub(crate) node_id: i64,

control_plane/attachment_service/src/reconciler.rs

@@ -13,6 +13,7 @@ use tokio_util::sync::CancellationToken;
 use utils::generation::Generation;
 use utils::id::{NodeId, TimelineId};
 use utils::lsn::Lsn;
+use utils::sync::gate::GateGuard;
 
 use crate::compute_hook::{ComputeHook, NotifyError};
 use crate::node::Node;
@@ -26,7 +27,7 @@ pub(super) struct Reconciler {
     pub(super) tenant_shard_id: TenantShardId,
     pub(crate) shard: ShardIdentity,
     pub(crate) generation: Generation,
-    pub(crate) intent: IntentState,
+    pub(crate) intent: TargetState,
     pub(crate) config: TenantConfig,
     pub(crate) observed: ObservedState,
 
@@ -53,14 +54,46 @@ pub(super) struct Reconciler {
     /// the tenant is changed.
     pub(crate) cancel: CancellationToken,
 
+    /// Reconcilers are registered with a Gate so that during a graceful shutdown we
+    /// can wait for all the reconcilers to respond to their cancellation tokens.
+    pub(crate) _gate_guard: GateGuard,
+
     /// Access to persistent storage for updating generation numbers
     pub(crate) persistence: Arc<Persistence>,
 }
 
+/// This is a snapshot of [`crate::tenant_state::IntentState`], but it does not do any
+/// reference counting for Scheduler.  The IntentState is what the scheduler works with,
+/// and the TargetState is just the instruction for a particular Reconciler run.
+#[derive(Debug)]
+pub(crate) struct TargetState {
+    pub(crate) attached: Option<NodeId>,
+    pub(crate) secondary: Vec<NodeId>,
+}
+
+impl TargetState {
+    pub(crate) fn from_intent(intent: &IntentState) -> Self {
+        Self {
+            attached: *intent.get_attached(),
+            secondary: intent.get_secondary().clone(),
+        }
+    }
+
+    fn all_pageservers(&self) -> Vec<NodeId> {
+        let mut result = self.secondary.clone();
+        if let Some(node_id) = &self.attached {
+            result.push(*node_id);
+        }
+        result
+    }
+}
+
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum ReconcileError {
     #[error(transparent)]
     Notify(#[from] NotifyError),
+    #[error("Cancelled")]
+    Cancel,
     #[error(transparent)]
     Other(#[from] anyhow::Error),
 }
@@ -263,7 +296,7 @@ impl Reconciler {
                 secondary_conf,
                 tenant_conf: config.clone(),
                 shard_number: shard.number.0,
-                shard_count: shard.count.0,
+                shard_count: shard.count.literal(),
                 shard_stripe_size: shard.stripe_size.0,
             }
         }
@@ -405,18 +438,49 @@ impl Reconciler {
             match self.observed.locations.get(&node_id) {
                 Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
                     // Nothing to do
-                    tracing::info!("Observed configuration already correct.")
+                    tracing::info!(%node_id, "Observed configuration already correct.")
                 }
-                _ => {
+                observed => {
                     // In all cases other than a matching observed configuration, we will
                     // reconcile this location.  This includes locations with different configurations, as well
                     // as locations with unknown (None) observed state.
-                    self.generation = self
-                        .persistence
-                        .increment_generation(self.tenant_shard_id, node_id)
-                        .await?;
-                    wanted_conf.generation = self.generation.into();
-                    tracing::info!("Observed configuration requires update.");
+
+                    // The general case is to increment the generation.  However, there are cases
+                    // where this is not necessary:
+                    // - if we are only updating the TenantConf part of the location
+                    // - if we are only changing the attachment mode (e.g. going to attachedmulti or attachedstale)
+                    //   and the location was already in the correct generation
+                    let increment_generation = match observed {
+                        None => true,
+                        Some(ObservedStateLocation { conf: None }) => true,
+                        Some(ObservedStateLocation {
+                            conf: Some(observed),
+                        }) => {
+                            // If mode and generation are the same, it follows that only the configuration has changed
+                            let config_update = observed.generation == wanted_conf.generation
+                                && observed.mode == wanted_conf.mode;
+
+                            // Usually the short-lived attachment modes (multi and stale) are only used
+                            // in the case of [`Self::live_migrate`], but it is simple to handle them correctly
+                            // here too.  Locations are allowed to go Single->Stale and Multi->Single within the same generation.
+                            let mode_transition = observed.generation == wanted_conf.generation
+                                && ((observed.mode == LocationConfigMode::AttachedSingle
+                                    && wanted_conf.mode == LocationConfigMode::AttachedStale)
+                                    || (observed.mode == LocationConfigMode::AttachedMulti
+                                        && wanted_conf.mode == LocationConfigMode::AttachedSingle));
+
+                            !(config_update || mode_transition)
+                        }
+                    };
+
+                    if increment_generation {
+                        self.generation = self
+                            .persistence
+                            .increment_generation(self.tenant_shard_id, node_id)
+                            .await?;
+                        wanted_conf.generation = self.generation.into();
+                    }
+                    tracing::info!(%node_id, "Observed configuration requires update.");
                     self.location_config(node_id, wanted_conf, None).await?;
                     self.compute_notify().await?;
                 }
@@ -458,7 +522,7 @@ impl Reconciler {
                     generation: None,
                     secondary_conf: None,
                     shard_number: self.shard.number.0,
-                    shard_count: self.shard.count.0,
+                    shard_count: self.shard.count.literal(),
                     shard_stripe_size: self.shard.stripe_size.0,
                     tenant_conf: self.config.clone(),
                 },
@@ -466,6 +530,9 @@ impl Reconciler {
         }
 
         for (node_id, conf) in changes {
+            if self.cancel.is_cancelled() {
+                return Err(ReconcileError::Cancel);
+            }
             self.location_config(node_id, conf, None).await?;
         }
 
@@ -506,7 +573,7 @@ pub(crate) fn attached_location_conf(
         generation: generation.into(),
         secondary_conf: None,
         shard_number: shard.number.0,
-        shard_count: shard.count.0,
+        shard_count: shard.count.literal(),
         shard_stripe_size: shard.stripe_size.0,
         tenant_conf: config.clone(),
     }
@@ -521,7 +588,7 @@ pub(crate) fn secondary_location_conf(
         generation: None,
         secondary_conf: Some(LocationConfigSecondary { warm: true }),
         shard_number: shard.number.0,
-        shard_count: shard.count.0,
+        shard_count: shard.count.literal(),
         shard_stripe_size: shard.stripe_size.0,
         tenant_conf: config.clone(),
     }

control_plane/attachment_service/src/scheduler.rs

@@ -1,8 +1,7 @@
-use pageserver_api::shard::TenantShardId;
-use std::collections::{BTreeMap, HashMap};
-use utils::{http::error::ApiError, id::NodeId};
-
 use crate::{node::Node, tenant_state::TenantState};
+use serde::Serialize;
+use std::collections::HashMap;
+use utils::{http::error::ApiError, id::NodeId};
 
 /// Scenarios in which we cannot find a suitable location for a tenant shard
 #[derive(thiserror::Error, Debug)]
@@ -19,52 +18,203 @@ impl From<ScheduleError> for ApiError {
     }
 }
 
+#[derive(Serialize, Eq, PartialEq)]
+struct SchedulerNode {
+    /// How many shards are currently scheduled on this node, via their [`crate::tenant_state::IntentState`].
+    shard_count: usize,
+
+    /// Whether this node is currently elegible to have new shards scheduled (this is derived
+    /// from a node's availability state and scheduling policy).
+    may_schedule: bool,
+}
+
+/// This type is responsible for selecting which node is used when a tenant shard needs to choose a pageserver
+/// on which to run.
+///
+/// The type has no persistent state of its own: this is all populated at startup.  The Serialize
+/// impl is only for debug dumps.
+#[derive(Serialize)]
 pub(crate) struct Scheduler {
-    tenant_counts: HashMap<NodeId, usize>,
+    nodes: HashMap<NodeId, SchedulerNode>,
 }
 
 impl Scheduler {
-    pub(crate) fn new(
-        tenants: &BTreeMap<TenantShardId, TenantState>,
-        nodes: &HashMap<NodeId, Node>,
-    ) -> Self {
-        let mut tenant_counts = HashMap::new();
-        for node_id in nodes.keys() {
-            tenant_counts.insert(*node_id, 0);
+    pub(crate) fn new<'a>(nodes: impl Iterator<Item = &'a Node>) -> Self {
+        let mut scheduler_nodes = HashMap::new();
+        for node in nodes {
+            scheduler_nodes.insert(
+                node.id,
+                SchedulerNode {
+                    shard_count: 0,
+                    may_schedule: node.may_schedule(),
+                },
+            );
         }
 
-        for tenant in tenants.values() {
-            if let Some(ps) = tenant.intent.attached {
-                let entry = tenant_counts.entry(ps).or_insert(0);
-                *entry += 1;
-            }
+        Self {
+            nodes: scheduler_nodes,
         }
-
-        for (node_id, node) in nodes {
-            if !node.may_schedule() {
-                tenant_counts.remove(node_id);
-            }
-        }
-
-        Self { tenant_counts }
     }
 
-    pub(crate) fn schedule_shard(
-        &mut self,
-        hard_exclude: &[NodeId],
-    ) -> Result<NodeId, ScheduleError> {
-        if self.tenant_counts.is_empty() {
+    /// For debug/support: check that our internal statistics are in sync with the state of
+    /// the nodes & tenant shards.
+    ///
+    /// If anything is inconsistent, log details and return an error.
+    pub(crate) fn consistency_check<'a>(
+        &self,
+        nodes: impl Iterator<Item = &'a Node>,
+        shards: impl Iterator<Item = &'a TenantState>,
+    ) -> anyhow::Result<()> {
+        let mut expect_nodes: HashMap<NodeId, SchedulerNode> = HashMap::new();
+        for node in nodes {
+            expect_nodes.insert(
+                node.id,
+                SchedulerNode {
+                    shard_count: 0,
+                    may_schedule: node.may_schedule(),
+                },
+            );
+        }
+
+        for shard in shards {
+            if let Some(node_id) = shard.intent.get_attached() {
+                match expect_nodes.get_mut(node_id) {
+                    Some(node) => node.shard_count += 1,
+                    None => anyhow::bail!(
+                        "Tenant {} references nonexistent node {}",
+                        shard.tenant_shard_id,
+                        node_id
+                    ),
+                }
+            }
+
+            for node_id in shard.intent.get_secondary() {
+                match expect_nodes.get_mut(node_id) {
+                    Some(node) => node.shard_count += 1,
+                    None => anyhow::bail!(
+                        "Tenant {} references nonexistent node {}",
+                        shard.tenant_shard_id,
+                        node_id
+                    ),
+                }
+            }
+        }
+
+        for (node_id, expect_node) in &expect_nodes {
+            let Some(self_node) = self.nodes.get(node_id) else {
+                anyhow::bail!("Node {node_id} not found in Self")
+            };
+
+            if self_node != expect_node {
+                tracing::error!("Inconsistency detected in scheduling state for node {node_id}");
+                tracing::error!("Expected state: {}", serde_json::to_string(expect_node)?);
+                tracing::error!("Self state: {}", serde_json::to_string(self_node)?);
+
+                anyhow::bail!("Inconsistent state on {node_id}");
+            }
+        }
+
+        if expect_nodes.len() != self.nodes.len() {
+            // We just checked that all the expected nodes are present.  If the lengths don't match,
+            // it means that we have nodes in Self that are unexpected.
+            for node_id in self.nodes.keys() {
+                if !expect_nodes.contains_key(node_id) {
+                    anyhow::bail!("Node {node_id} found in Self but not in expected nodes");
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Increment the reference count of a node.  This reference count is used to guide scheduling
+    /// decisions, not for memory management: it represents one tenant shard whose IntentState targets
+    /// this node.
+    ///
+    /// It is an error to call this for a node that is not known to the scheduler (i.e. passed into
+    /// [`Self::new`] or [`Self::node_upsert`])
+    pub(crate) fn node_inc_ref(&mut self, node_id: NodeId) {
+        let Some(node) = self.nodes.get_mut(&node_id) else {
+            tracing::error!("Scheduler missing node {node_id}");
+            debug_assert!(false);
+            return;
+        };
+
+        node.shard_count += 1;
+    }
+
+    /// Decrement a node's reference count.  Inverse of [`Self::node_inc_ref`].
+    pub(crate) fn node_dec_ref(&mut self, node_id: NodeId) {
+        let Some(node) = self.nodes.get_mut(&node_id) else {
+            debug_assert!(false);
+            tracing::error!("Scheduler missing node {node_id}");
+            return;
+        };
+
+        node.shard_count -= 1;
+    }
+
+    pub(crate) fn node_upsert(&mut self, node: &Node) {
+        use std::collections::hash_map::Entry::*;
+        match self.nodes.entry(node.id) {
+            Occupied(mut entry) => {
+                entry.get_mut().may_schedule = node.may_schedule();
+            }
+            Vacant(entry) => {
+                entry.insert(SchedulerNode {
+                    shard_count: 0,
+                    may_schedule: node.may_schedule(),
+                });
+            }
+        }
+    }
+
+    pub(crate) fn node_remove(&mut self, node_id: NodeId) {
+        if self.nodes.remove(&node_id).is_none() {
+            tracing::warn!(node_id=%node_id, "Removed non-existent node from scheduler");
+        }
+    }
+
+    /// Where we have several nodes to choose from, for example when picking a secondary location
+    /// to promote to an attached location, this method may be used to pick the best choice based
+    /// on the scheduler's knowledge of utilization and availability.
+    ///
+    /// If the input is empty, or all the nodes are not elegible for scheduling, return None: the
+    /// caller can pick a node some other way.
+    pub(crate) fn node_preferred(&self, nodes: &[NodeId]) -> Option<NodeId> {
+        if nodes.is_empty() {
+            return None;
+        }
+
+        let node = nodes
+            .iter()
+            .map(|node_id| {
+                let may_schedule = self
+                    .nodes
+                    .get(node_id)
+                    .map(|n| n.may_schedule)
+                    .unwrap_or(false);
+                (*node_id, may_schedule)
+            })
+            .max_by_key(|(_n, may_schedule)| *may_schedule);
+
+        // If even the preferred node has may_schedule==false, return None
+        node.and_then(|(node_id, may_schedule)| if may_schedule { Some(node_id) } else { None })
+    }
+
+    pub(crate) fn schedule_shard(&self, hard_exclude: &[NodeId]) -> Result<NodeId, ScheduleError> {
+        if self.nodes.is_empty() {
             return Err(ScheduleError::NoPageservers);
         }
 
         let mut tenant_counts: Vec<(NodeId, usize)> = self
-            .tenant_counts
+            .nodes
             .iter()
             .filter_map(|(k, v)| {
-                if hard_exclude.contains(k) {
+                if hard_exclude.contains(k) || !v.may_schedule {
                     None
                 } else {
-                    Some((*k, *v))
+                    Some((*k, v.shard_count))
                 }
             })
             .collect();
@@ -73,17 +223,109 @@ impl Scheduler {
         tenant_counts.sort_by_key(|i| (i.1, i.0));
 
         if tenant_counts.is_empty() {
-            // After applying constraints, no pageservers were left
+            // After applying constraints, no pageservers were left.  We log some detail about
+            // the state of nodes to help understand why this happened.  This is not logged as an error because
+            // it is legitimately possible for enough nodes to be Offline to prevent scheduling a shard.
+            tracing::info!("Scheduling failure, while excluding {hard_exclude:?}, node states:");
+            for (node_id, node) in &self.nodes {
+                tracing::info!(
+                    "Node {node_id}: may_schedule={} shards={}",
+                    node.may_schedule,
+                    node.shard_count
+                );
+            }
+
             return Err(ScheduleError::ImpossibleConstraint);
         }
 
-        for (node_id, count) in &tenant_counts {
-            tracing::info!("tenant_counts[{node_id}]={count}");
-        }
-
         let node_id = tenant_counts.first().unwrap().0;
-        tracing::info!("scheduler selected node {node_id}");
-        *self.tenant_counts.get_mut(&node_id).unwrap() += 1;
+        tracing::info!(
+            "scheduler selected node {node_id} (elegible nodes {:?}, exclude: {hard_exclude:?})",
+            tenant_counts.iter().map(|i| i.0 .0).collect::<Vec<_>>()
+        );
+
+        // Note that we do not update shard count here to reflect the scheduling: that
+        // is IntentState's job when the scheduled location is used.
+
         Ok(node_id)
     }
 }
+
+#[cfg(test)]
+pub(crate) mod test_utils {
+
+    use crate::node::Node;
+    use control_plane::attachment_service::{NodeAvailability, NodeSchedulingPolicy};
+    use std::collections::HashMap;
+    use utils::id::NodeId;
+    /// Test helper: synthesize the requested number of nodes, all in active state.
+    ///
+    /// Node IDs start at one.
+    pub(crate) fn make_test_nodes(n: u64) -> HashMap<NodeId, Node> {
+        (1..n + 1)
+            .map(|i| {
+                (
+                    NodeId(i),
+                    Node {
+                        id: NodeId(i),
+                        availability: NodeAvailability::Active,
+                        scheduling: NodeSchedulingPolicy::Active,
+                        listen_http_addr: format!("httphost-{i}"),
+                        listen_http_port: 80 + i as u16,
+                        listen_pg_addr: format!("pghost-{i}"),
+                        listen_pg_port: 5432 + i as u16,
+                    },
+                )
+            })
+            .collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use utils::id::NodeId;
+
+    use crate::tenant_state::IntentState;
+    #[test]
+    fn scheduler_basic() -> anyhow::Result<()> {
+        let nodes = test_utils::make_test_nodes(2);
+
+        let mut scheduler = Scheduler::new(nodes.values());
+        let mut t1_intent = IntentState::new();
+        let mut t2_intent = IntentState::new();
+
+        let scheduled = scheduler.schedule_shard(&[])?;
+        t1_intent.set_attached(&mut scheduler, Some(scheduled));
+        let scheduled = scheduler.schedule_shard(&[])?;
+        t2_intent.set_attached(&mut scheduler, Some(scheduled));
+
+        assert_eq!(scheduler.nodes.get(&NodeId(1)).unwrap().shard_count, 1);
+        assert_eq!(scheduler.nodes.get(&NodeId(2)).unwrap().shard_count, 1);
+
+        let scheduled = scheduler.schedule_shard(&t1_intent.all_pageservers())?;
+        t1_intent.push_secondary(&mut scheduler, scheduled);
+
+        assert_eq!(scheduler.nodes.get(&NodeId(1)).unwrap().shard_count, 1);
+        assert_eq!(scheduler.nodes.get(&NodeId(2)).unwrap().shard_count, 2);
+
+        t1_intent.clear(&mut scheduler);
+        assert_eq!(scheduler.nodes.get(&NodeId(1)).unwrap().shard_count, 0);
+        assert_eq!(scheduler.nodes.get(&NodeId(2)).unwrap().shard_count, 1);
+
+        if cfg!(debug_assertions) {
+            // Dropping an IntentState without clearing it causes a panic in debug mode,
+            // because we have failed to properly update scheduler shard counts.
+            let result = std::panic::catch_unwind(move || {
+                drop(t2_intent);
+            });
+            assert!(result.is_err());
+        } else {
+            t2_intent.clear(&mut scheduler);
+            assert_eq!(scheduler.nodes.get(&NodeId(1)).unwrap().shard_count, 0);
+            assert_eq!(scheduler.nodes.get(&NodeId(2)).unwrap().shard_count, 0);
+        }
+
+        Ok(())
+    }
+}

control_plane/attachment_service/src/tenant_state.rs

@@ -1,27 +1,47 @@
 use std::{collections::HashMap, sync::Arc, time::Duration};
 
+use crate::{metrics, persistence::TenantShardPersistence};
 use control_plane::attachment_service::NodeAvailability;
 use pageserver_api::{
     models::{LocationConfig, LocationConfigMode, TenantConfig},
     shard::{ShardIdentity, TenantShardId},
 };
+use serde::Serialize;
 use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
+use tracing::{instrument, Instrument};
 use utils::{
     generation::Generation,
     id::NodeId,
     seqwait::{SeqWait, SeqWaitError},
+    sync::gate::Gate,
 };
 
 use crate::{
     compute_hook::ComputeHook,
     node::Node,
-    persistence::Persistence,
-    reconciler::{attached_location_conf, secondary_location_conf, ReconcileError, Reconciler},
+    persistence::{split_state::SplitState, Persistence},
+    reconciler::{
+        attached_location_conf, secondary_location_conf, ReconcileError, Reconciler, TargetState,
+    },
     scheduler::{ScheduleError, Scheduler},
     service, PlacementPolicy, Sequence,
 };
 
+/// Serialization helper
+fn read_mutex_content<S, T>(v: &std::sync::Mutex<T>, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::ser::Serializer,
+    T: Clone + std::fmt::Display,
+{
+    serializer.collect_str(&v.lock().unwrap())
+}
+
+/// In-memory state for a particular tenant shard.
+///
+/// This struct implement Serialize for debugging purposes, but is _not_ persisted
+/// itself: see [`crate::persistence`] for the subset of tenant shard state that is persisted.
+#[derive(Serialize)]
 pub(crate) struct TenantState {
     pub(crate) tenant_shard_id: TenantShardId,
 
@@ -56,20 +76,29 @@ pub(crate) struct TenantState {
     /// If a reconcile task is currently in flight, it may be joined here (it is
     /// only safe to join if either the result has been received or the reconciler's
     /// cancellation token has been fired)
+    #[serde(skip)]
     pub(crate) reconciler: Option<ReconcilerHandle>,
 
+    /// If a tenant is being split, then all shards with that TenantId will have a
+    /// SplitState set, this acts as a guard against other operations such as background
+    /// reconciliation, and timeline creation.
+    pub(crate) splitting: SplitState,
+
     /// Optionally wait for reconciliation to complete up to a particular
     /// sequence number.
+    #[serde(skip)]
     pub(crate) waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
 
     /// Indicates sequence number for which we have encountered an error reconciling.  If
     /// this advances ahead of [`Self::waiter`] then a reconciliation error has occurred,
     /// and callers should stop waiting for `waiter` and propagate the error.
+    #[serde(skip)]
     pub(crate) error_waiter: std::sync::Arc<SeqWait<Sequence, Sequence>>,
 
     /// The most recent error from a reconcile on this tenant
     /// TODO: generalize to an array of recent events
     /// TOOD: use a ArcSwap instead of mutex for faster reads?
+    #[serde(serialize_with = "read_mutex_content")]
     pub(crate) last_error: std::sync::Arc<std::sync::Mutex<String>>,
 
     /// If we have a pending compute notification that for some reason we weren't able to send,
@@ -79,13 +108,131 @@ pub(crate) struct TenantState {
     pub(crate) pending_compute_notification: bool,
 }
 
-#[derive(Default, Clone, Debug)]
+#[derive(Default, Clone, Debug, Serialize)]
 pub(crate) struct IntentState {
-    pub(crate) attached: Option<NodeId>,
-    pub(crate) secondary: Vec<NodeId>,
+    attached: Option<NodeId>,
+    secondary: Vec<NodeId>,
 }
 
-#[derive(Default, Clone)]
+impl IntentState {
+    pub(crate) fn new() -> Self {
+        Self {
+            attached: None,
+            secondary: vec![],
+        }
+    }
+    pub(crate) fn single(scheduler: &mut Scheduler, node_id: Option<NodeId>) -> Self {
+        if let Some(node_id) = node_id {
+            scheduler.node_inc_ref(node_id);
+        }
+        Self {
+            attached: node_id,
+            secondary: vec![],
+        }
+    }
+
+    pub(crate) fn set_attached(&mut self, scheduler: &mut Scheduler, new_attached: Option<NodeId>) {
+        if self.attached != new_attached {
+            if let Some(old_attached) = self.attached.take() {
+                scheduler.node_dec_ref(old_attached);
+            }
+            if let Some(new_attached) = &new_attached {
+                scheduler.node_inc_ref(*new_attached);
+            }
+            self.attached = new_attached;
+        }
+    }
+
+    /// Like set_attached, but the node is from [`Self::secondary`].  This swaps the node from
+    /// secondary to attached while maintaining the scheduler's reference counts.
+    pub(crate) fn promote_attached(
+        &mut self,
+        _scheduler: &mut Scheduler,
+        promote_secondary: NodeId,
+    ) {
+        // If we call this with a node that isn't in secondary, it would cause incorrect
+        // scheduler reference counting, since we assume the node is already referenced as a secondary.
+        debug_assert!(self.secondary.contains(&promote_secondary));
+
+        // TODO: when scheduler starts tracking attached + secondary counts separately, we will
+        // need to call into it here.
+        self.secondary.retain(|n| n != &promote_secondary);
+        self.attached = Some(promote_secondary);
+    }
+
+    pub(crate) fn push_secondary(&mut self, scheduler: &mut Scheduler, new_secondary: NodeId) {
+        debug_assert!(!self.secondary.contains(&new_secondary));
+        scheduler.node_inc_ref(new_secondary);
+        self.secondary.push(new_secondary);
+    }
+
+    /// It is legal to call this with a node that is not currently a secondary: that is a no-op
+    pub(crate) fn remove_secondary(&mut self, scheduler: &mut Scheduler, node_id: NodeId) {
+        let index = self.secondary.iter().position(|n| *n == node_id);
+        if let Some(index) = index {
+            scheduler.node_dec_ref(node_id);
+            self.secondary.remove(index);
+        }
+    }
+
+    pub(crate) fn clear_secondary(&mut self, scheduler: &mut Scheduler) {
+        for secondary in self.secondary.drain(..) {
+            scheduler.node_dec_ref(secondary);
+        }
+    }
+
+    pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
+        if let Some(old_attached) = self.attached.take() {
+            scheduler.node_dec_ref(old_attached);
+        }
+
+        self.clear_secondary(scheduler);
+    }
+
+    pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
+        let mut result = Vec::new();
+        if let Some(p) = self.attached {
+            result.push(p)
+        }
+
+        result.extend(self.secondary.iter().copied());
+
+        result
+    }
+
+    pub(crate) fn get_attached(&self) -> &Option<NodeId> {
+        &self.attached
+    }
+
+    pub(crate) fn get_secondary(&self) -> &Vec<NodeId> {
+        &self.secondary
+    }
+
+    /// When a node goes offline, we update intents to avoid using it
+    /// as their attached pageserver.
+    ///
+    /// Returns true if a change was made
+    pub(crate) fn notify_offline(&mut self, node_id: NodeId) -> bool {
+        if self.attached == Some(node_id) {
+            // TODO: when scheduler starts tracking attached + secondary counts separately, we will
+            // need to call into it here.
+            self.attached = None;
+            self.secondary.push(node_id);
+            true
+        } else {
+            false
+        }
+    }
+}
+
+impl Drop for IntentState {
+    fn drop(&mut self) {
+        // Must clear before dropping, to avoid leaving stale refcounts in the Scheduler
+        debug_assert!(self.attached.is_none() && self.secondary.is_empty());
+    }
+}
+
+#[derive(Default, Clone, Serialize)]
 pub(crate) struct ObservedState {
     pub(crate) locations: HashMap<NodeId, ObservedStateLocation>,
 }
@@ -99,7 +246,7 @@ pub(crate) struct ObservedState {
 ///       what it is (e.g. we failed partway through configuring it)
 ///     * Instance exists with conf==Some: this tells us what we last successfully configured on this node,
 ///       and that configuration will still be present unless something external interfered.
-#[derive(Clone)]
+#[derive(Clone, Serialize)]
 pub(crate) struct ObservedStateLocation {
     /// If None, it means we do not know the status of this shard's location on this node, but
     /// we know that we might have some state on this node.
@@ -175,46 +322,6 @@ pub(crate) struct ReconcileResult {
     pub(crate) pending_compute_notification: bool,
 }
 
-impl IntentState {
-    pub(crate) fn new() -> Self {
-        Self {
-            attached: None,
-            secondary: vec![],
-        }
-    }
-    pub(crate) fn all_pageservers(&self) -> Vec<NodeId> {
-        let mut result = Vec::new();
-        if let Some(p) = self.attached {
-            result.push(p)
-        }
-
-        result.extend(self.secondary.iter().copied());
-
-        result
-    }
-
-    pub(crate) fn single(node_id: Option<NodeId>) -> Self {
-        Self {
-            attached: node_id,
-            secondary: vec![],
-        }
-    }
-
-    /// When a node goes offline, we update intents to avoid using it
-    /// as their attached pageserver.
-    ///
-    /// Returns true if a change was made
-    pub(crate) fn notify_offline(&mut self, node_id: NodeId) -> bool {
-        if self.attached == Some(node_id) {
-            self.attached = None;
-            self.secondary.push(node_id);
-            true
-        } else {
-            false
-        }
-    }
-}
-
 impl ObservedState {
     pub(crate) fn new() -> Self {
         Self {
@@ -238,6 +345,7 @@ impl TenantState {
             observed: ObservedState::default(),
             config: TenantConfig::default(),
             reconciler: None,
+            splitting: SplitState::Idle,
             sequence: Sequence(1),
             waiter: Arc::new(SeqWait::new(Sequence(0))),
             error_waiter: Arc::new(SeqWait::new(Sequence(0))),
@@ -281,6 +389,9 @@ impl TenantState {
         // All remaining observed locations generate secondary intents.  This includes None
         // observations, as these may well have some local content on disk that is usable (this
         // is an edge case that might occur if we restarted during a migration or other change)
+        //
+        // We may leave intent.attached empty if we didn't find any attached locations: [`Self::schedule`]
+        // will take care of promoting one of these secondaries to be attached.
         self.observed.locations.keys().for_each(|node_id| {
             if Some(*node_id) != self.intent.attached {
                 self.intent.secondary.push(*node_id);
@@ -288,6 +399,33 @@ impl TenantState {
         });
     }
 
+    /// Part of [`Self::schedule`] that is used to choose exactly one node to act as the
+    /// attached pageserver for a shard.
+    ///
+    /// Returns whether we modified it, and the NodeId selected.
+    fn schedule_attached(
+        &mut self,
+        scheduler: &mut Scheduler,
+    ) -> Result<(bool, NodeId), ScheduleError> {
+        // No work to do if we already have an attached tenant
+        if let Some(node_id) = self.intent.attached {
+            return Ok((false, node_id));
+        }
+
+        if let Some(promote_secondary) = scheduler.node_preferred(&self.intent.secondary) {
+            // Promote a secondary
+            tracing::debug!("Promoted secondary {} to attached", promote_secondary);
+            self.intent.promote_attached(scheduler, promote_secondary);
+            Ok((true, promote_secondary))
+        } else {
+            // Pick a fresh node: either we had no secondaries or none were schedulable
+            let node_id = scheduler.schedule_shard(&self.intent.secondary)?;
+            tracing::debug!("Selected {} as attached", node_id);
+            self.intent.set_attached(scheduler, Some(node_id));
+            Ok((true, node_id))
+        }
+    }
+
     pub(crate) fn schedule(&mut self, scheduler: &mut Scheduler) -> Result<(), ScheduleError> {
         // TODO: before scheduling new nodes, check if any existing content in
         // self.intent refers to pageservers that are offline, and pick other
@@ -298,36 +436,29 @@ impl TenantState {
 
         // Build the set of pageservers already in use by this tenant, to avoid scheduling
         // more work on the same pageservers we're already using.
-        let mut used_pageservers = self.intent.all_pageservers();
         let mut modified = false;
 
         use PlacementPolicy::*;
         match self.policy {
             Single => {
                 // Should have exactly one attached, and zero secondaries
-                if self.intent.attached.is_none() {
-                    let node_id = scheduler.schedule_shard(&used_pageservers)?;
-                    self.intent.attached = Some(node_id);
-                    used_pageservers.push(node_id);
-                    modified = true;
-                }
+                let (modified_attached, _attached_node_id) = self.schedule_attached(scheduler)?;
+                modified |= modified_attached;
+
                 if !self.intent.secondary.is_empty() {
-                    self.intent.secondary.clear();
+                    self.intent.clear_secondary(scheduler);
                     modified = true;
                 }
             }
             Double(secondary_count) => {
                 // Should have exactly one attached, and N secondaries
-                if self.intent.attached.is_none() {
-                    let node_id = scheduler.schedule_shard(&used_pageservers)?;
-                    self.intent.attached = Some(node_id);
-                    used_pageservers.push(node_id);
-                    modified = true;
-                }
+                let (modified_attached, attached_node_id) = self.schedule_attached(scheduler)?;
+                modified |= modified_attached;
 
+                let mut used_pageservers = vec![attached_node_id];
                 while self.intent.secondary.len() < secondary_count {
                     let node_id = scheduler.schedule_shard(&used_pageservers)?;
-                    self.intent.secondary.push(node_id);
+                    self.intent.push_secondary(scheduler, node_id);
                     used_pageservers.push(node_id);
                     modified = true;
                 }
@@ -335,12 +466,12 @@ impl TenantState {
             Detached => {
                 // Should have no attached or secondary pageservers
                 if self.intent.attached.is_some() {
-                    self.intent.attached = None;
+                    self.intent.set_attached(scheduler, None);
                     modified = true;
                 }
 
                 if !self.intent.secondary.is_empty() {
-                    self.intent.secondary.clear();
+                    self.intent.clear_secondary(scheduler);
                     modified = true;
                 }
             }
@@ -406,6 +537,13 @@ impl TenantState {
             }
         }
 
+        for node_id in self.observed.locations.keys() {
+            if self.intent.attached != Some(*node_id) && !self.intent.secondary.contains(node_id) {
+                // We have observed state that isn't part of our intent: need to clean it up.
+                return true;
+            }
+        }
+
         // Even if there is no pageserver work to be done, if we have a pending notification to computes,
         // wake up a reconciler to send it.
         if self.pending_compute_notification {
@@ -415,6 +553,8 @@ impl TenantState {
         false
     }
 
+    #[allow(clippy::too_many_arguments)]
+    #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
     pub(crate) fn maybe_reconcile(
         &mut self,
         result_tx: tokio::sync::mpsc::UnboundedSender<ReconcileResult>,
@@ -422,6 +562,8 @@ impl TenantState {
         compute_hook: &Arc<ComputeHook>,
         service_config: &service::Config,
         persistence: &Arc<Persistence>,
+        gate: &Gate,
+        cancel: &CancellationToken,
     ) -> Option<ReconcilerWaiter> {
         // If there are any ambiguous observed states, and the nodes they refer to are available,
         // we should reconcile to clean them up.
@@ -443,6 +585,14 @@ impl TenantState {
             return None;
         }
 
+        // If we are currently splitting, then never start a reconciler task: the splitting logic
+        // requires that shards are not interfered with while it runs. Do this check here rather than
+        // up top, so that we only log this message if we would otherwise have done a reconciliation.
+        if !matches!(self.splitting, SplitState::Idle) {
+            tracing::info!("Refusing to reconcile, splitting in progress");
+            return None;
+        }
+
         // Reconcile already in flight for the current sequence?
         if let Some(handle) = &self.reconciler {
             if handle.sequence == self.sequence {
@@ -460,70 +610,105 @@ impl TenantState {
         // doing our sequence's work.
         let old_handle = self.reconciler.take();
 
-        let cancel = CancellationToken::new();
+        let Ok(gate_guard) = gate.enter() else {
+            // Shutting down, don't start a reconciler
+            return None;
+        };
+
+        // Advance the sequence before spawning a reconciler, so that sequence waiters
+        // can distinguish between before+after the reconcile completes.
+        self.sequence = self.sequence.next();
+
+        let reconciler_cancel = cancel.child_token();
         let mut reconciler = Reconciler {
             tenant_shard_id: self.tenant_shard_id,
             shard: self.shard,
             generation: self.generation,
-            intent: self.intent.clone(),
+            intent: TargetState::from_intent(&self.intent),
             config: self.config.clone(),
             observed: self.observed.clone(),
             pageservers: pageservers.clone(),
             compute_hook: compute_hook.clone(),
             service_config: service_config.clone(),
-            cancel: cancel.clone(),
+            _gate_guard: gate_guard,
+            cancel: reconciler_cancel.clone(),
             persistence: persistence.clone(),
             compute_notify_failure: false,
         };
 
         let reconcile_seq = self.sequence;
 
-        tracing::info!("Spawning Reconciler for sequence {}", self.sequence);
+        tracing::info!(seq=%reconcile_seq, "Spawning Reconciler for sequence {}", self.sequence);
         let must_notify = self.pending_compute_notification;
-        let join_handle = tokio::task::spawn(async move {
-            // Wait for any previous reconcile task to complete before we start
-            if let Some(old_handle) = old_handle {
-                old_handle.cancel.cancel();
-                if let Err(e) = old_handle.handle.await {
-                    // We can't do much with this other than log it: the task is done, so
-                    // we may proceed with our work.
-                    tracing::error!("Unexpected join error waiting for reconcile task: {e}");
+        let reconciler_span = tracing::info_span!(parent: None, "reconciler", seq=%reconcile_seq,
+                                                        tenant_id=%reconciler.tenant_shard_id.tenant_id,
+                                                        shard_id=%reconciler.tenant_shard_id.shard_slug());
+        metrics::RECONCILER.spawned.inc();
+        let join_handle = tokio::task::spawn(
+            async move {
+                // Wait for any previous reconcile task to complete before we start
+                if let Some(old_handle) = old_handle {
+                    old_handle.cancel.cancel();
+                    if let Err(e) = old_handle.handle.await {
+                        // We can't do much with this other than log it: the task is done, so
+                        // we may proceed with our work.
+                        tracing::error!("Unexpected join error waiting for reconcile task: {e}");
+                    }
                 }
+
+                // Early check for cancellation before doing any work
+                // TODO: wrap all remote API operations in cancellation check
+                // as well.
+                if reconciler.cancel.is_cancelled() {
+                    metrics::RECONCILER
+                        .complete
+                        .with_label_values(&[metrics::ReconcilerMetrics::CANCEL])
+                        .inc();
+                    return;
+                }
+
+                // Attempt to make observed state match intent state
+                let result = reconciler.reconcile().await;
+
+                // If we know we had a pending compute notification from some previous action, send a notification irrespective
+                // of whether the above reconcile() did any work
+                if result.is_ok() && must_notify {
+                    // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
+                    reconciler.compute_notify().await.ok();
+                }
+
+                // Update result counter
+                match &result {
+                    Ok(_) => metrics::RECONCILER
+                        .complete
+                        .with_label_values(&[metrics::ReconcilerMetrics::SUCCESS]),
+                    Err(ReconcileError::Cancel) => metrics::RECONCILER
+                        .complete
+                        .with_label_values(&[metrics::ReconcilerMetrics::CANCEL]),
+                    Err(_) => metrics::RECONCILER
+                        .complete
+                        .with_label_values(&[metrics::ReconcilerMetrics::ERROR]),
+                }
+                .inc();
+
+                result_tx
+                    .send(ReconcileResult {
+                        sequence: reconcile_seq,
+                        result,
+                        tenant_shard_id: reconciler.tenant_shard_id,
+                        generation: reconciler.generation,
+                        observed: reconciler.observed,
+                        pending_compute_notification: reconciler.compute_notify_failure,
+                    })
+                    .ok();
             }
-
-            // Early check for cancellation before doing any work
-            // TODO: wrap all remote API operations in cancellation check
-            // as well.
-            if reconciler.cancel.is_cancelled() {
-                return;
-            }
-
-            // Attempt to make observed state match intent state
-            let result = reconciler.reconcile().await;
-
-            // If we know we had a pending compute notification from some previous action, send a notification irrespective
-            // of whether the above reconcile() did any work
-            if result.is_ok() && must_notify {
-                // If this fails we will send the need to retry in [`ReconcileResult::pending_compute_notification`]
-                reconciler.compute_notify().await.ok();
-            }
-
-            result_tx
-                .send(ReconcileResult {
-                    sequence: reconcile_seq,
-                    result,
-                    tenant_shard_id: reconciler.tenant_shard_id,
-                    generation: reconciler.generation,
-                    observed: reconciler.observed,
-                    pending_compute_notification: reconciler.compute_notify_failure,
-                })
-                .ok();
-        });
+            .instrument(reconciler_span),
+        );
 
         self.reconciler = Some(ReconcilerHandle {
             sequence: self.sequence,
             handle: join_handle,
-            cancel,
+            cancel: reconciler_cancel,
         });
 
         Some(ReconcilerWaiter {
@@ -548,4 +733,103 @@ impl TenantState {
 
         debug_assert!(!self.intent.all_pageservers().contains(&node_id));
     }
+
+    pub(crate) fn to_persistent(&self) -> TenantShardPersistence {
+        TenantShardPersistence {
+            tenant_id: self.tenant_shard_id.tenant_id.to_string(),
+            shard_number: self.tenant_shard_id.shard_number.0 as i32,
+            shard_count: self.tenant_shard_id.shard_count.literal() as i32,
+            shard_stripe_size: self.shard.stripe_size.0 as i32,
+            generation: self.generation.into().unwrap_or(0) as i32,
+            generation_pageserver: self
+                .intent
+                .get_attached()
+                .map(|n| n.0 as i64)
+                .unwrap_or(i64::MAX),
+
+            placement_policy: serde_json::to_string(&self.policy).unwrap(),
+            config: serde_json::to_string(&self.config).unwrap(),
+            splitting: SplitState::default(),
+        }
+    }
+}
+
+#[cfg(test)]
+pub(crate) mod tests {
+    use pageserver_api::shard::{ShardCount, ShardNumber};
+    use utils::id::TenantId;
+
+    use crate::scheduler::test_utils::make_test_nodes;
+
+    use super::*;
+
+    fn make_test_tenant_shard(policy: PlacementPolicy) -> TenantState {
+        let tenant_id = TenantId::generate();
+        let shard_number = ShardNumber(0);
+        let shard_count = ShardCount::new(1);
+
+        let tenant_shard_id = TenantShardId {
+            tenant_id,
+            shard_number,
+            shard_count,
+        };
+        TenantState::new(
+            tenant_shard_id,
+            ShardIdentity::new(
+                shard_number,
+                shard_count,
+                pageserver_api::shard::ShardStripeSize(32768),
+            )
+            .unwrap(),
+            policy,
+        )
+    }
+
+    /// Test the scheduling behaviors used when a tenant configured for HA is subject
+    /// to nodes being marked offline.
+    #[test]
+    fn tenant_ha_scheduling() -> anyhow::Result<()> {
+        // Start with three nodes.  Our tenant will only use two.  The third one is
+        // expected to remain unused.
+        let mut nodes = make_test_nodes(3);
+
+        let mut scheduler = Scheduler::new(nodes.values());
+
+        let mut tenant_state = make_test_tenant_shard(PlacementPolicy::Double(1));
+        tenant_state
+            .schedule(&mut scheduler)
+            .expect("we have enough nodes, scheduling should work");
+
+        // Expect to initially be schedule on to different nodes
+        assert_eq!(tenant_state.intent.secondary.len(), 1);
+        assert!(tenant_state.intent.attached.is_some());
+
+        let attached_node_id = tenant_state.intent.attached.unwrap();
+        let secondary_node_id = *tenant_state.intent.secondary.iter().last().unwrap();
+        assert_ne!(attached_node_id, secondary_node_id);
+
+        // Notifying the attached node is offline should demote it to a secondary
+        let changed = tenant_state.intent.notify_offline(attached_node_id);
+        assert!(changed);
+
+        // Update the scheduler state to indicate the node is offline
+        nodes.get_mut(&attached_node_id).unwrap().availability = NodeAvailability::Offline;
+        scheduler.node_upsert(nodes.get(&attached_node_id).unwrap());
+
+        // Scheduling the node should promote the still-available secondary node to attached
+        tenant_state
+            .schedule(&mut scheduler)
+            .expect("active nodes are available");
+        assert_eq!(tenant_state.intent.attached.unwrap(), secondary_node_id);
+
+        // The original attached node should have been retained as a secondary
+        assert_eq!(
+            *tenant_state.intent.secondary.iter().last().unwrap(),
+            attached_node_id
+        );
+
+        tenant_state.intent.clear(&mut scheduler);
+
+        Ok(())
+    }
 }

control_plane/src/attachment_service.rs

@@ -113,7 +113,7 @@ pub struct TenantShardMigrateRequest {
     pub node_id: NodeId,
 }
 
-#[derive(Serialize, Deserialize, Clone, Copy)]
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq)]
 pub enum NodeAvailability {
     // Normal, happy state
     Active,
@@ -137,7 +137,7 @@ impl FromStr for NodeAvailability {
 
 /// FIXME: this is a duplicate of the type in the attachment_service crate, because the
 /// type needs to be defined with diesel traits in there.
-#[derive(Serialize, Deserialize, Clone, Copy)]
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq)]
 pub enum NodeSchedulingPolicy {
     Active,
     Filling,

control_plane/src/bin/neon_local.rs

@@ -450,7 +450,7 @@ async fn handle_tenant(
                     new_tenant_id: TenantShardId::unsharded(tenant_id),
                     generation: None,
                     shard_parameters: ShardParameters {
-                        count: ShardCount(shard_count),
+                        count: ShardCount::new(shard_count),
                         stripe_size: shard_stripe_size
                             .map(ShardStripeSize)
                             .unwrap_or(ShardParameters::DEFAULT_STRIPE_SIZE),
@@ -652,6 +652,10 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
             let name = import_match
                 .get_one::<String>("node-name")
                 .ok_or_else(|| anyhow!("No node name provided"))?;
+            let update_catalog = import_match
+                .get_one::<bool>("update-catalog")
+                .cloned()
+                .unwrap_or_default();
 
             // Parse base inputs
             let base_tarfile = import_match
@@ -694,6 +698,7 @@ async fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::Local
                 None,
                 pg_version,
                 ComputeMode::Primary,
+                !update_catalog,
             )?;
             println!("Done");
         }
@@ -831,6 +836,10 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .get_one::<String>("endpoint_id")
                 .map(String::to_string)
                 .unwrap_or_else(|| format!("ep-{branch_name}"));
+            let update_catalog = sub_args
+                .get_one::<bool>("update-catalog")
+                .cloned()
+                .unwrap_or_default();
 
             let lsn = sub_args
                 .get_one::<String>("lsn")
@@ -880,6 +889,7 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 http_port,
                 pg_version,
                 mode,
+                !update_catalog,
             )?;
         }
         "start" => {
@@ -918,6 +928,11 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .get(endpoint_id.as_str())
                 .ok_or_else(|| anyhow::anyhow!("endpoint {endpoint_id} not found"))?;
 
+            let create_test_user = sub_args
+                .get_one::<bool>("create-test-user")
+                .cloned()
+                .unwrap_or_default();
+
             cplane.check_conflicting_endpoints(
                 endpoint.mode,
                 endpoint.tenant_id,
@@ -972,6 +987,7 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                     pageservers,
                     remote_ext_config,
                     stripe_size.0 as usize,
+                    create_test_user,
                 )
                 .await?;
         }
@@ -1457,6 +1473,18 @@ fn cli() -> Command {
         .required(false)
         .default_value("1");
 
+    let update_catalog = Arg::new("update-catalog")
+        .value_parser(value_parser!(bool))
+        .long("update-catalog")
+        .help("If set, will set up the catalog for neon_superuser")
+        .required(false);
+
+    let create_test_user = Arg::new("create-test-user")
+        .value_parser(value_parser!(bool))
+        .long("create-test-user")
+        .help("If set, will create test user `user` and `neondb` database. Requires `update-catalog = true`")
+        .required(false);
+
     Command::new("Neon CLI")
         .arg_required_else_help(true)
         .version(GIT_VERSION)
@@ -1517,6 +1545,7 @@ fn cli() -> Command {
                 .arg(Arg::new("end-lsn").long("end-lsn")
                     .help("Lsn the basebackup ends at"))
                 .arg(pg_version_arg.clone())
+                .arg(update_catalog.clone())
             )
         ).subcommand(
             Command::new("tenant")
@@ -1630,6 +1659,7 @@ fn cli() -> Command {
                             .required(false))
                     .arg(pg_version_arg.clone())
                     .arg(hot_standby_arg.clone())
+                    .arg(update_catalog)
                 )
                 .subcommand(Command::new("start")
                     .about("Start postgres.\n If the endpoint doesn't exist yet, it is created.")
@@ -1637,6 +1667,7 @@ fn cli() -> Command {
                     .arg(endpoint_pageserver_id_arg.clone())
                     .arg(safekeepers_arg)
                     .arg(remote_ext_config_args)
+                    .arg(create_test_user)
                 )
                 .subcommand(Command::new("reconfigure")
                             .about("Reconfigure the endpoint")

control_plane/src/endpoint.rs

@@ -41,11 +41,15 @@ use std::net::SocketAddr;
 use std::net::TcpStream;
 use std::path::PathBuf;
 use std::process::Command;
+use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
 
 use anyhow::{anyhow, bail, Context, Result};
+use compute_api::spec::Database;
+use compute_api::spec::PgIdent;
 use compute_api::spec::RemoteExtSpec;
+use compute_api::spec::Role;
 use nix::sys::signal::kill;
 use nix::sys::signal::Signal;
 use serde::{Deserialize, Serialize};
@@ -122,6 +126,7 @@ impl ComputeControlPlane {
         http_port: Option<u16>,
         pg_version: u32,
         mode: ComputeMode,
+        skip_pg_catalog_updates: bool,
     ) -> Result<Arc<Endpoint>> {
         let pg_port = pg_port.unwrap_or_else(|| self.get_port());
         let http_port = http_port.unwrap_or_else(|| self.get_port() + 1);
@@ -140,7 +145,7 @@ impl ComputeControlPlane {
             // before and after start are the same. So, skip catalog updates,
             // with this we basically test a case of waking up an idle compute, where
             // we also skip catalog updates in the cloud.
-            skip_pg_catalog_updates: true,
+            skip_pg_catalog_updates,
             features: vec![],
         });
 
@@ -155,7 +160,7 @@ impl ComputeControlPlane {
                 http_port,
                 pg_port,
                 pg_version,
-                skip_pg_catalog_updates: true,
+                skip_pg_catalog_updates,
                 features: vec![],
             })?,
         )?;
@@ -500,6 +505,7 @@ impl Endpoint {
         pageservers: Vec<(Host, u16)>,
         remote_ext_config: Option<&String>,
         shard_stripe_size: usize,
+        create_test_user: bool,
     ) -> Result<()> {
         if self.status() == EndpointStatus::Running {
             anyhow::bail!("The endpoint is already running");
@@ -551,8 +557,26 @@ impl Endpoint {
                 cluster_id: None, // project ID: not used
                 name: None,       // project name: not used
                 state: None,
-                roles: vec![],
-                databases: vec![],
+                roles: if create_test_user {
+                    vec![Role {
+                        name: PgIdent::from_str("test").unwrap(),
+                        encrypted_password: None,
+                        options: None,
+                    }]
+                } else {
+                    Vec::new()
+                },
+                databases: if create_test_user {
+                    vec![Database {
+                        name: PgIdent::from_str("neondb").unwrap(),
+                        owner: PgIdent::from_str("test").unwrap(),
+                        options: None,
+                        restrict_conn: false,
+                        invalid: false,
+                    }]
+                } else {
+                    Vec::new()
+                },
                 settings: None,
                 postgresql_conf: Some(postgresql_conf),
             },
@@ -566,6 +590,7 @@ impl Endpoint {
             remote_extensions,
             pgbouncer_settings: None,
             shard_stripe_size: Some(shard_stripe_size),
+            primary_is_running: None,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -577,11 +602,16 @@ impl Endpoint {
             .open(self.endpoint_path().join("compute.log"))?;
 
         // Launch compute_ctl
-        println!("Starting postgres node at '{}'", self.connstr());
+        let conn_str = self.connstr("cloud_admin", "postgres");
+        println!("Starting postgres node at '{}'", conn_str);
+        if create_test_user {
+            let conn_str = self.connstr("user", "neondb");
+            println!("Also at '{}'", conn_str);
+        }
         let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
         cmd.args(["--http-port", &self.http_address.port().to_string()])
             .args(["--pgdata", self.pgdata().to_str().unwrap()])
-            .args(["--connstr", &self.connstr()])
+            .args(["--connstr", &conn_str])
             .args([
                 "--spec-path",
                 self.endpoint_path().join("spec.json").to_str().unwrap(),
@@ -652,7 +682,9 @@ impl Endpoint {
                         }
                         ComputeStatus::Empty
                         | ComputeStatus::ConfigurationPending
-                        | ComputeStatus::Configuration => {
+                        | ComputeStatus::Configuration
+                        | ComputeStatus::TerminationPending
+                        | ComputeStatus::Terminated => {
                             bail!("unexpected compute status: {:?}", state.status)
                         }
                     }
@@ -783,13 +815,13 @@ impl Endpoint {
         Ok(())
     }
 
-    pub fn connstr(&self) -> String {
+    pub fn connstr(&self, user: &str, db_name: &str) -> String {
         format!(
             "postgresql://{}@{}:{}/{}",
-            "cloud_admin",
+            user,
             self.pg_address.ip(),
             self.pg_address.port(),
-            "postgres"
+            db_name
         )
     }
 }

control_plane/src/pageserver.rs

@@ -210,6 +210,25 @@ impl PageServerNode {
         update_config: bool,
         register: bool,
     ) -> anyhow::Result<()> {
+        // Register the node with the storage controller before starting pageserver: pageserver must be registered to
+        // successfully call /re-attach and finish starting up.
+        if register {
+            let attachment_service = AttachmentService::from_env(&self.env);
+            let (pg_host, pg_port) =
+                parse_host_port(&self.conf.listen_pg_addr).expect("Unable to parse listen_pg_addr");
+            let (http_host, http_port) = parse_host_port(&self.conf.listen_http_addr)
+                .expect("Unable to parse listen_http_addr");
+            attachment_service
+                .node_register(NodeRegisterRequest {
+                    node_id: self.conf.id,
+                    listen_pg_addr: pg_host.to_string(),
+                    listen_pg_port: pg_port.unwrap_or(5432),
+                    listen_http_addr: http_host.to_string(),
+                    listen_http_port: http_port.unwrap_or(80),
+                })
+                .await?;
+        }
+
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
         print!(
@@ -248,23 +267,6 @@ impl PageServerNode {
         )
         .await?;
 
-        if register {
-            let attachment_service = AttachmentService::from_env(&self.env);
-            let (pg_host, pg_port) =
-                parse_host_port(&self.conf.listen_pg_addr).expect("Unable to parse listen_pg_addr");
-            let (http_host, http_port) = parse_host_port(&self.conf.listen_http_addr)
-                .expect("Unable to parse listen_http_addr");
-            attachment_service
-                .node_register(NodeRegisterRequest {
-                    node_id: self.conf.id,
-                    listen_pg_addr: pg_host.to_string(),
-                    listen_pg_port: pg_port.unwrap_or(5432),
-                    listen_http_addr: http_host.to_string(),
-                    listen_http_port: http_port.unwrap_or(80),
-                })
-                .await?;
-        }
-
         Ok(())
     }
 
@@ -400,6 +402,11 @@ impl PageServerNode {
                 .map(|x| x.parse::<bool>())
                 .transpose()
                 .context("Failed to parse 'lazy_slru_download' as bool")?,
+            timeline_get_throttle: settings
+                .remove("timeline_get_throttle")
+                .map(serde_json::from_str)
+                .transpose()
+                .context("parse `timeline_get_throttle` from json")?,
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -505,6 +512,11 @@ impl PageServerNode {
                     .map(|x| x.parse::<bool>())
                     .transpose()
                     .context("Failed to parse 'lazy_slru_download' as bool")?,
+                timeline_get_throttle: settings
+                    .remove("timeline_get_throttle")
+                    .map(serde_json::from_str)
+                    .transpose()
+                    .context("parse `timeline_get_throttle` from json")?,
             }
         };
 

libs/compute_api/src/responses.rs

@@ -52,6 +52,10 @@ pub enum ComputeStatus {
     // compute will exit soon or is waiting for
     // control-plane to terminate it.
     Failed,
+    // Termination requested
+    TerminationPending,
+    // Terminated Postgres
+    Terminated,
 }
 
 fn rfc3339_serialize<S>(x: &Option<DateTime<Utc>>, s: S) -> Result<S::Ok, S::Error>

libs/compute_api/src/spec.rs

@@ -79,6 +79,12 @@ pub struct ComputeSpec {
     // Stripe size for pageserver sharding, in pages
     #[serde(default)]
     pub shard_stripe_size: Option<usize>,
+
+    // When we are starting a new replica in hot standby mode,
+    // we need to know if the primary is running.
+    // This is used to determine if replica should wait for
+    // RUNNING_XACTS from primary or not.
+    pub primary_is_running: Option<bool>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.

libs/metrics/src/lib.rs

@@ -115,7 +115,6 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
 // performed by the process.
 // We know the size of the block, so we can determine the I/O bytes out of it.
 // The value might be not 100% exact, but should be fine for Prometheus metrics in this case.
-#[allow(clippy::unnecessary_cast)]
 fn update_rusage_metrics() {
     let rusage_stats = get_rusage_stats();
 
@@ -202,6 +201,11 @@ impl<P: Atomic> GenericCounterPairVec<P> {
     pub fn with_label_values(&self, vals: &[&str]) -> GenericCounterPair<P> {
         self.get_metric_with_label_values(vals).unwrap()
     }
+
+    pub fn remove_label_values(&self, res: &mut [Result<()>; 2], vals: &[&str]) {
+        res[0] = self.inc.remove_label_values(vals);
+        res[1] = self.dec.remove_label_values(vals);
+    }
 }
 
 impl<P: Atomic> GenericCounterPair<P> {
@@ -248,6 +252,15 @@ impl<P: Atomic> GenericCounterPair<P> {
     }
 }
 
+impl<P: Atomic> Clone for GenericCounterPair<P> {
+    fn clone(&self) -> Self {
+        Self {
+            inc: self.inc.clone(),
+            dec: self.dec.clone(),
+        }
+    }
+}
+
 /// Guard returned by [`GenericCounterPair::guard`]
 pub struct GenericCounterPairGuard<P: Atomic>(GenericCounter<P>);
 

libs/pageserver_api/Cargo.toml

@@ -18,9 +18,11 @@ enum-map.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
 hex.workspace = true
+humantime.workspace = true
 thiserror.workspace = true
 humantime-serde.workspace = true
 chrono.workspace = true
+itertools.workspace = true
 
 workspace_hack.workspace = true
 

libs/pageserver_api/src/keyspace.rs

@@ -2,6 +2,7 @@ use postgres_ffi::BLCKSZ;
 use std::ops::Range;
 
 use crate::key::Key;
+use itertools::Itertools;
 
 ///
 /// Represents a set of Keys, in a compact form.
@@ -63,9 +64,36 @@ impl KeySpace {
         KeyPartitioning { parts }
     }
 
-    /// Update the keyspace such that it doesn't contain any range
-    /// that is overlapping with `other`. This can involve splitting or
-    /// removing of existing ranges.
+    /// Merge another keyspace into the current one.
+    /// Note: the keyspaces must not ovelap (enforced via assertions)
+    pub fn merge(&mut self, other: &KeySpace) {
+        let all_ranges = self
+            .ranges
+            .iter()
+            .merge_by(other.ranges.iter(), |lhs, rhs| lhs.start < rhs.start);
+
+        let mut accum = KeySpaceAccum::new();
+        let mut prev: Option<&Range<Key>> = None;
+        for range in all_ranges {
+            if let Some(prev) = prev {
+                let overlap =
+                    std::cmp::max(range.start, prev.start) < std::cmp::min(range.end, prev.end);
+                assert!(
+                    !overlap,
+                    "Attempt to merge ovelapping keyspaces: {:?} overlaps {:?}",
+                    prev, range
+                );
+            }
+
+            accum.add_range(range.clone());
+            prev = Some(range);
+        }
+
+        self.ranges = accum.to_keyspace().ranges;
+    }
+
+    /// Remove all keys in `other` from `self`.
+    /// This can involve splitting or removing of existing ranges.
     pub fn remove_overlapping_with(&mut self, other: &KeySpace) {
         let (self_start, self_end) = match (self.start(), self.end()) {
             (Some(start), Some(end)) => (start, end),
@@ -220,16 +248,7 @@ impl KeySpaceAccum {
     }
 
     pub fn consume_keyspace(&mut self) -> KeySpace {
-        if let Some(accum) = self.accum.take() {
-            self.ranges.push(accum);
-        }
-
-        let mut prev_accum = KeySpaceAccum::new();
-        std::mem::swap(self, &mut prev_accum);
-
-        KeySpace {
-            ranges: prev_accum.ranges,
-        }
+        std::mem::take(self).to_keyspace()
     }
 
     pub fn size(&self) -> u64 {
@@ -279,6 +298,13 @@ impl KeySpaceRandomAccum {
         }
         KeySpace { ranges }
     }
+
+    pub fn consume_keyspace(&mut self) -> KeySpace {
+        let mut prev_accum = KeySpaceRandomAccum::new();
+        std::mem::swap(self, &mut prev_accum);
+
+        prev_accum.to_keyspace()
+    }
 }
 
 pub fn key_range_size(key_range: &Range<Key>) -> u32 {

libs/pageserver_api/src/models.rs

@@ -1,4 +1,7 @@
 pub mod partitioning;
+pub mod utilization;
+
+pub use utilization::PageserverUtilization;
 
 use std::{
     collections::HashMap,
@@ -180,7 +183,7 @@ pub enum TimelineState {
     Broken { reason: String, backtrace: String },
 }
 
-#[derive(Serialize, Deserialize)]
+#[derive(Serialize, Deserialize, Clone)]
 pub struct TimelineCreateRequest {
     pub new_timeline_id: TimelineId,
     #[serde(default)]
@@ -214,14 +217,14 @@ impl ShardParameters {
     pub const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
 
     pub fn is_unsharded(&self) -> bool {
-        self.count == ShardCount(0)
+        self.count.is_unsharded()
     }
 }
 
 impl Default for ShardParameters {
     fn default() -> Self {
         Self {
-            count: ShardCount(0),
+            count: ShardCount::new(0),
             stripe_size: Self::DEFAULT_STRIPE_SIZE,
         }
     }
@@ -283,6 +286,15 @@ pub struct TenantConfig {
     pub gc_feedback: Option<bool>,
     pub heatmap_period: Option<String>,
     pub lazy_slru_download: Option<bool>,
+    pub timeline_get_throttle: Option<ThrottleConfig>,
+}
+
+impl TenantConfig {
+    pub fn merge(&mut self, other: &TenantConfig) {
+        if let Some(max_lsn_wal_lag) = &other.max_lsn_wal_lag {
+            self.max_lsn_wal_lag = Some(*max_lsn_wal_lag);
+        }
+    }
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -290,6 +302,7 @@ pub struct TenantConfig {
 pub enum EvictionPolicy {
     NoEviction,
     LayerAccessThreshold(EvictionPolicyLayerAccessThreshold),
+    OnlyImitiate(EvictionPolicyLayerAccessThreshold),
 }
 
 impl EvictionPolicy {
@@ -297,6 +310,7 @@ impl EvictionPolicy {
         match self {
             EvictionPolicy::NoEviction => "NoEviction",
             EvictionPolicy::LayerAccessThreshold(_) => "LayerAccessThreshold",
+            EvictionPolicy::OnlyImitiate(_) => "OnlyImitiate",
         }
     }
 }
@@ -309,10 +323,39 @@ pub struct EvictionPolicyLayerAccessThreshold {
     pub threshold: Duration,
 }
 
+#[derive(Debug, Serialize, Deserialize, Clone, PartialEq, Eq)]
+pub struct ThrottleConfig {
+    pub task_kinds: Vec<String>, // TaskKind
+    pub initial: usize,
+    #[serde(with = "humantime_serde")]
+    pub refill_interval: Duration,
+    pub refill_amount: NonZeroUsize,
+    pub max: usize,
+    pub fair: bool,
+}
+
+impl ThrottleConfig {
+    pub fn disabled() -> Self {
+        Self {
+            task_kinds: vec![], // effectively disables the throttle
+            // other values don't matter with emtpy `task_kinds`.
+            initial: 0,
+            refill_interval: Duration::from_millis(1),
+            refill_amount: NonZeroUsize::new(1).unwrap(),
+            max: 1,
+            fair: true,
+        }
+    }
+    /// The requests per second allowed  by the given config.
+    pub fn steady_rps(&self) -> f64 {
+        (self.refill_amount.get() as f64) / (self.refill_interval.as_secs_f64())
+    }
+}
+
 /// A flattened analog of a `pagesever::tenant::LocationMode`, which
 /// lists out all possible states (and the virtual "Detached" state)
 /// in a flat form rather than using rust-style enums.
-#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq)]
+#[derive(Serialize, Deserialize, Debug, Clone, Copy, Eq, PartialEq)]
 pub enum LocationConfigMode {
     AttachedSingle,
     AttachedMulti,
@@ -376,6 +419,12 @@ pub struct TenantLocationConfigRequest {
     pub config: LocationConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+#[serde(deny_unknown_fields)]
+pub struct TenantTimeTravelRequest {
+    pub shard_counts: Vec<ShardCount>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct TenantShardLocation {

libs/pageserver_api/src/models/utilization.rs

@@ -0,0 +1,70 @@
+use std::time::SystemTime;
+
+/// Pageserver current utilization and scoring for how good candidate the pageserver would be for
+/// the next tenant.
+///
+/// See and maintain pageserver openapi spec for `/v1/utilization_score` as the truth.
+///
+/// `format: int64` fields must use `ser_saturating_u63` because openapi generated clients might
+/// not handle full u64 values properly.
+#[derive(serde::Serialize, Debug)]
+pub struct PageserverUtilization {
+    /// Used disk space
+    #[serde(serialize_with = "ser_saturating_u63")]
+    pub disk_usage_bytes: u64,
+    /// Free disk space
+    #[serde(serialize_with = "ser_saturating_u63")]
+    pub free_space_bytes: u64,
+    /// Lower is better score for how good candidate for a next tenant would this pageserver be.
+    #[serde(serialize_with = "ser_saturating_u63")]
+    pub utilization_score: u64,
+    /// When was this snapshot captured, pageserver local time.
+    ///
+    /// Use millis to give confidence that the value is regenerated often enough.
+    #[serde(serialize_with = "ser_rfc3339_millis")]
+    pub captured_at: SystemTime,
+}
+
+fn ser_rfc3339_millis<S: serde::Serializer>(
+    ts: &SystemTime,
+    serializer: S,
+) -> Result<S::Ok, S::Error> {
+    serializer.collect_str(&humantime::format_rfc3339_millis(*ts))
+}
+
+/// openapi knows only `format: int64`, so avoid outputting a non-parseable value by generated clients.
+///
+/// Instead of newtype, use this because a newtype would get require handling deserializing values
+/// with the highest bit set which is properly parsed by serde formats, but would create a
+/// conundrum on how to handle and again serialize such values at type level. It will be a few
+/// years until we can use more than `i64::MAX` bytes on a disk.
+fn ser_saturating_u63<S: serde::Serializer>(value: &u64, serializer: S) -> Result<S::Ok, S::Error> {
+    const MAX_FORMAT_INT64: u64 = i64::MAX as u64;
+
+    let value = (*value).min(MAX_FORMAT_INT64);
+
+    serializer.serialize_u64(value)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use super::*;
+
+    #[test]
+    fn u64_max_is_serialized_as_u63_max() {
+        let doc = PageserverUtilization {
+            disk_usage_bytes: u64::MAX,
+            free_space_bytes: 0,
+            utilization_score: u64::MAX,
+            captured_at: SystemTime::UNIX_EPOCH + Duration::from_secs(1708509779),
+        };
+
+        let s = serde_json::to_string(&doc).unwrap();
+
+        let expected = r#"{"disk_usage_bytes":9223372036854775807,"free_space_bytes":0,"utilization_score":9223372036854775807,"captured_at":"2024-02-21T10:02:59.000Z"}"#;
+
+        assert_eq!(s, expected);
+    }
+}

libs/pageserver_api/src/shard.rs

@@ -13,10 +13,41 @@ use utils::id::TenantId;
 pub struct ShardNumber(pub u8);
 
 #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
-pub struct ShardCount(pub u8);
+pub struct ShardCount(u8);
 
 impl ShardCount {
     pub const MAX: Self = Self(u8::MAX);
+
+    /// The internal value of a ShardCount may be zero, which means "1 shard, but use
+    /// legacy format for TenantShardId that excludes the shard suffix", also known
+    /// as `TenantShardId::unsharded`.
+    ///
+    /// This method returns the actual number of shards, i.e. if our internal value is
+    /// zero, we return 1 (unsharded tenants have 1 shard).
+    pub fn count(&self) -> u8 {
+        if self.0 > 0 {
+            self.0
+        } else {
+            1
+        }
+    }
+
+    /// The literal internal value: this is **not** the number of shards in the
+    /// tenant, as we have a special zero value for legacy unsharded tenants.  Use
+    /// [`Self::count`] if you want to know the cardinality of shards.
+    pub fn literal(&self) -> u8 {
+        self.0
+    }
+
+    pub fn is_unsharded(&self) -> bool {
+        self.0 == 0
+    }
+
+    /// `v` may be zero, or the number of shards in the tenant.  `v` is what
+    /// [`Self::literal`] would return.
+    pub fn new(val: u8) -> Self {
+        Self(val)
+    }
 }
 
 impl ShardNumber {
@@ -86,7 +117,7 @@ impl TenantShardId {
     }
 
     pub fn is_unsharded(&self) -> bool {
-        self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
+        self.shard_number == ShardNumber(0) && self.shard_count.is_unsharded()
     }
 
     /// Convenience for dropping the tenant_id and just getting the ShardIndex: this
@@ -471,10 +502,12 @@ impl ShardIdentity {
     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.
+            // A1: 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.
+            // A2: because key_is_shard0 also covers relation size keys, which are written
+            //     on all shards even though they're only maintained accurately on shard 0.
             false
         } else {
             !self.is_key_local(key)

libs/postgres_ffi/src/lib.rs

@@ -3,7 +3,7 @@
 #![allow(non_snake_case)]
 // bindgen creates some unsafe code with no doc comments.
 #![allow(clippy::missing_safety_doc)]
-// noted at 1.63 that in many cases there's a u32 -> u32 transmutes in bindgen code.
+// noted at 1.63 that in many cases there's u32 -> u32 transmutes in bindgen code.
 #![allow(clippy::useless_transmute)]
 // modules included with the postgres_ffi macro depend on the types of the specific version's
 // types, and trigger a too eager lint.

libs/postgres_ffi/src/pg_constants.rs

@@ -80,6 +80,9 @@ pub const XLOG_XACT_ABORT: u8 = 0x20;
 pub const XLOG_XACT_COMMIT_PREPARED: u8 = 0x30;
 pub const XLOG_XACT_ABORT_PREPARED: u8 = 0x40;
 
+// From standbydefs.h
+pub const XLOG_RUNNING_XACTS: u8 = 0x10;
+
 // From srlu.h
 pub const SLRU_PAGES_PER_SEGMENT: u32 = 32;
 pub const SLRU_SEG_SIZE: usize = BLCKSZ as usize * SLRU_PAGES_PER_SEGMENT as usize;

libs/postgres_ffi/src/xlog_utils.rs

@@ -119,11 +119,6 @@ pub fn generate_pg_control(
     // Generate new pg_control needed for bootstrap
     checkpoint.redo = normalize_lsn(lsn, WAL_SEGMENT_SIZE).0;
 
-    //reset some fields we don't want to preserve
-    //TODO Check this.
-    //We may need to determine the value from twophase data.
-    checkpoint.oldestActiveXid = 0;
-
     //save new values in pg_control
     pg_control.checkPoint = 0;
     pg_control.checkPointCopy = checkpoint;

libs/remote_storage/Cargo.toml

@@ -15,11 +15,13 @@ aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
 bytes.workspace = true
 camino.workspace = true
+humantime.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 futures.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tokio = { workspace = true, features = ["sync", "fs", "io-util"] }
+tokio-stream.workspace = true
 tokio-util = { workspace = true, features = ["compat"] }
 toml_edit.workspace = true
 tracing.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -22,16 +22,15 @@ use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerCl
 use bytes::Bytes;
 use futures::stream::Stream;
 use futures_util::StreamExt;
+use futures_util::TryStreamExt;
 use http_types::{StatusCode, Url};
-use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 use tracing::debug;
 
-use crate::s3_bucket::RequestKind;
-use crate::TimeTravelError;
 use crate::{
-    AzureConfig, ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath,
-    RemoteStorage, StorageMetadata,
+    error::Cancelled, s3_bucket::RequestKind, AzureConfig, ConcurrencyLimiter, Download,
+    DownloadError, Listing, ListingMode, RemotePath, RemoteStorage, StorageMetadata,
+    TimeTravelError, TimeoutOrCancel,
 };
 
 pub struct AzureBlobStorage {
@@ -39,10 +38,12 @@ pub struct AzureBlobStorage {
     prefix_in_container: Option<String>,
     max_keys_per_list_response: Option<NonZeroU32>,
     concurrency_limiter: ConcurrencyLimiter,
+    // Per-request timeout. Accessible for tests.
+    pub timeout: Duration,
 }
 
 impl AzureBlobStorage {
-    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+    pub fn new(azure_config: &AzureConfig, timeout: Duration) -> Result<Self> {
         debug!(
             "Creating azure remote storage for azure container {}",
             azure_config.container_name
@@ -79,6 +80,7 @@ impl AzureBlobStorage {
             prefix_in_container: azure_config.prefix_in_container.to_owned(),
             max_keys_per_list_response,
             concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+            timeout,
         })
     }
 
@@ -121,8 +123,11 @@ impl AzureBlobStorage {
     async fn download_for_builder(
         &self,
         builder: GetBlobBuilder,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
-        let mut response = builder.into_stream();
+        let kind = RequestKind::Get;
+
+        let _permit = self.permit(kind, cancel).await?;
 
         let mut etag = None;
         let mut last_modified = None;
@@ -130,39 +135,70 @@ impl AzureBlobStorage {
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
 
-        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()));
+        let download = async {
+            let response = builder
+                // convert to concrete Pageable
+                .into_stream()
+                // convert to TryStream
+                .into_stream()
+                .map_err(to_download_error);
+
+            // apply per request timeout
+            let response = tokio_stream::StreamExt::timeout(response, self.timeout);
+
+            // flatten
+            let response = response.map(|res| match res {
+                Ok(res) => res,
+                Err(_elapsed) => Err(DownloadError::Timeout),
+            });
+
+            let mut response = std::pin::pin!(response);
+
+            let mut bufs = Vec::new();
+            while let Some(part) = response.next().await {
+                let part = part?;
+                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())));
+                }
+                let data = part
+                    .data
+                    .collect()
+                    .await
+                    .map_err(|e| DownloadError::Other(e.into()))?;
+                bufs.push(data);
             }
-            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())));
-            }
-            let data = part
-                .data
-                .collect()
-                .await
-                .map_err(|e| DownloadError::Other(e.into()))?;
-            bufs.push(data);
+            Ok(Download {
+                download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
+                etag,
+                last_modified,
+                metadata: Some(StorageMetadata(metadata)),
+            })
+        };
+
+        tokio::select! {
+            bufs = download => bufs,
+            _ = cancel.cancelled() => Err(DownloadError::Cancelled),
         }
-        Ok(Download {
-            download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
-            etag,
-            last_modified,
-            metadata: Some(StorageMetadata(metadata)),
-        })
     }
 
-    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
-        self.concurrency_limiter
-            .acquire(kind)
-            .await
-            .expect("semaphore is never closed")
+    async fn permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
+        let acquire = self.concurrency_limiter.acquire(kind);
+
+        tokio::select! {
+            permit = acquire => Ok(permit.expect("never closed")),
+            _ = cancel.cancelled() => Err(Cancelled),
+        }
     }
 }
 
@@ -192,66 +228,87 @@ impl RemoteStorage for AzureBlobStorage {
         prefix: Option<&RemotePath>,
         mode: ListingMode,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<Listing, DownloadError> {
-        // get the passed prefix or if it is not set use prefix_in_bucket value
-        let list_prefix = prefix
-            .map(|p| self.relative_path_to_name(p))
-            .or_else(|| self.prefix_in_container.clone())
-            .map(|mut p| {
-                // required to end with a separator
-                // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
-                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
-                }
-                p
+        let _permit = self.permit(RequestKind::List, cancel).await?;
+
+        let op = async {
+            // get the passed prefix or if it is not set use prefix_in_bucket value
+            let list_prefix = prefix
+                .map(|p| self.relative_path_to_name(p))
+                .or_else(|| self.prefix_in_container.clone())
+                .map(|mut p| {
+                    // required to end with a separator
+                    // otherwise request will return only the entry of a prefix
+                    if matches!(mode, ListingMode::WithDelimiter)
+                        && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                    {
+                        p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                    }
+                    p
+                });
+
+            let mut builder = self.client.list_blobs();
+
+            if let ListingMode::WithDelimiter = mode {
+                builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+            }
+
+            if let Some(prefix) = list_prefix {
+                builder = builder.prefix(Cow::from(prefix.to_owned()));
+            }
+
+            if let Some(limit) = self.max_keys_per_list_response {
+                builder = builder.max_results(MaxResults::new(limit));
+            }
+
+            let response = builder.into_stream();
+            let response = response.into_stream().map_err(to_download_error);
+            let response = tokio_stream::StreamExt::timeout(response, self.timeout);
+            let response = response.map(|res| match res {
+                Ok(res) => res,
+                Err(_elapsed) => Err(DownloadError::Timeout),
             });
 
-        let mut builder = self.client.list_blobs();
+            let mut response = std::pin::pin!(response);
 
-        if let ListingMode::WithDelimiter = mode {
-            builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
-        }
+            let mut res = Listing::default();
 
-        if let Some(prefix) = list_prefix {
-            builder = builder.prefix(Cow::from(prefix.to_owned()));
-        }
+            let mut max_keys = max_keys.map(|mk| mk.get());
+            while let Some(entry) = response.next().await {
+                let entry = entry?;
+                let prefix_iter = entry
+                    .blobs
+                    .prefixes()
+                    .map(|prefix| self.name_to_relative_path(&prefix.name));
+                res.prefixes.extend(prefix_iter);
 
-        if let Some(limit) = self.max_keys_per_list_response {
-            builder = builder.max_results(MaxResults::new(limit));
-        }
+                let blob_iter = entry
+                    .blobs
+                    .blobs()
+                    .map(|k| self.name_to_relative_path(&k.name));
 
-        let mut response = builder.into_stream();
-        let mut res = Listing::default();
-        // NonZeroU32 doesn't support subtraction apparently
-        let mut max_keys = max_keys.map(|mk| mk.get());
-        while let Some(l) = response.next().await {
-            let entry = l.map_err(to_download_error)?;
-            let prefix_iter = entry
-                .blobs
-                .prefixes()
-                .map(|prefix| self.name_to_relative_path(&prefix.name));
-            res.prefixes.extend(prefix_iter);
+                for key in blob_iter {
+                    res.keys.push(key);
 
-            let blob_iter = entry
-                .blobs
-                .blobs()
-                .map(|k| self.name_to_relative_path(&k.name));
-
-            for key in blob_iter {
-                res.keys.push(key);
-                if let Some(mut mk) = max_keys {
-                    assert!(mk > 0);
-                    mk -= 1;
-                    if mk == 0 {
-                        return Ok(res); // limit reached
+                    if let Some(mut mk) = max_keys {
+                        assert!(mk > 0);
+                        mk -= 1;
+                        if mk == 0 {
+                            return Ok(res); // limit reached
+                        }
+                        max_keys = Some(mk);
                     }
-                    max_keys = Some(mk);
                 }
             }
+
+            Ok(res)
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(DownloadError::Cancelled),
         }
-        Ok(res)
     }
 
     async fn upload(
@@ -260,35 +317,52 @@ impl RemoteStorage for AzureBlobStorage {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Put).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+        let _permit = self.permit(RequestKind::Put, cancel).await?;
 
-        let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
-            Box::pin(from);
+        let op = async {
+            let blob_client = self.client.blob_client(self.relative_path_to_name(to));
 
-        let from = NonSeekableStream::new(from, data_size_bytes);
+            let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
+                Box::pin(from);
 
-        let body = azure_core::Body::SeekableStream(Box::new(from));
+            let from = NonSeekableStream::new(from, data_size_bytes);
 
-        let mut builder = blob_client.put_block_blob(body);
+            let body = azure_core::Body::SeekableStream(Box::new(from));
 
-        if let Some(metadata) = metadata {
-            builder = builder.metadata(to_azure_metadata(metadata));
+            let mut builder = blob_client.put_block_blob(body);
+
+            if let Some(metadata) = metadata {
+                builder = builder.metadata(to_azure_metadata(metadata));
+            }
+
+            let fut = builder.into_future();
+            let fut = tokio::time::timeout(self.timeout, fut);
+
+            match fut.await {
+                Ok(Ok(_response)) => Ok(()),
+                Ok(Err(azure)) => Err(azure.into()),
+                Err(_timeout) => Err(TimeoutOrCancel::Cancel.into()),
+            }
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(TimeoutOrCancel::Cancel.into()),
         }
-
-        let _response = builder.into_future().await?;
-
-        Ok(())
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
-        let _permit = self.permit(RequestKind::Get).await;
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
         let builder = blob_client.get();
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(builder, cancel).await
     }
 
     async fn download_byte_range(
@@ -296,8 +370,8 @@ impl RemoteStorage for AzureBlobStorage {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
-        let _permit = self.permit(RequestKind::Get).await;
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
         let mut builder = blob_client.get();
@@ -309,82 +383,113 @@ impl RemoteStorage for AzureBlobStorage {
         };
         builder = builder.range(range);
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(builder, cancel).await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Delete).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
+        self.delete_objects(std::array::from_ref(path), cancel)
+            .await
+    }
 
-        let builder = blob_client.delete();
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete, cancel).await?;
 
-        match builder.into_future().await {
-            Ok(_response) => Ok(()),
-            Err(e) => {
-                if let Some(http_err) = e.as_http_error() {
-                    if http_err.status() == StatusCode::NotFound {
-                        return Ok(());
+        let op = async {
+            // TODO batch requests are also not supported by the SDK
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+            for path in paths {
+                let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+                let request = blob_client.delete().into_future();
+
+                let res = tokio::time::timeout(self.timeout, request).await;
+
+                match res {
+                    Ok(Ok(_response)) => continue,
+                    Ok(Err(e)) => {
+                        if let Some(http_err) = e.as_http_error() {
+                            if http_err.status() == StatusCode::NotFound {
+                                continue;
+                            }
+                        }
+                        return Err(e.into());
                     }
+                    Err(_elapsed) => return Err(TimeoutOrCancel::Timeout.into()),
                 }
-                Err(anyhow::Error::new(e))
             }
+
+            Ok(())
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(TimeoutOrCancel::Cancel.into()),
         }
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
-        // Permit is already obtained by inner delete function
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Copy, cancel).await?;
 
-        // TODO batch requests are also not supported by the SDK
-        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
-        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
-        for path in paths {
-            self.delete(path).await?;
-        }
-        Ok(())
-    }
+        let timeout = tokio::time::sleep(self.timeout);
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Copy).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+        let mut copy_status = None;
 
-        let source_url = format!(
-            "{}/{}",
-            self.client.url()?,
-            self.relative_path_to_name(from)
-        );
-        let builder = blob_client.copy(Url::from_str(&source_url)?);
+        let op = async {
+            let blob_client = self.client.blob_client(self.relative_path_to_name(to));
 
-        let result = builder.into_future().await?;
+            let source_url = format!(
+                "{}/{}",
+                self.client.url()?,
+                self.relative_path_to_name(from)
+            );
 
-        let mut copy_status = result.copy_status;
-        let start_time = Instant::now();
-        const MAX_WAIT_TIME: Duration = Duration::from_secs(60);
-        loop {
-            match copy_status {
-                CopyStatus::Aborted => {
-                    anyhow::bail!("Received abort for copy from {from} to {to}.");
+            let builder = blob_client.copy(Url::from_str(&source_url)?);
+            let copy = builder.into_future();
+
+            let result = copy.await?;
+
+            copy_status = Some(result.copy_status);
+            loop {
+                match copy_status.as_ref().expect("we always set it to Some") {
+                    CopyStatus::Aborted => {
+                        anyhow::bail!("Received abort for copy from {from} to {to}.");
+                    }
+                    CopyStatus::Failed => {
+                        anyhow::bail!("Received failure response for copy from {from} to {to}.");
+                    }
+                    CopyStatus::Success => return Ok(()),
+                    CopyStatus::Pending => (),
                 }
-                CopyStatus::Failed => {
-                    anyhow::bail!("Received failure response for copy from {from} to {to}.");
-                }
-                CopyStatus::Success => return Ok(()),
-                CopyStatus::Pending => (),
+                // The copy is taking longer. Waiting a second and then re-trying.
+                // TODO estimate time based on copy_progress and adjust time based on that
+                tokio::time::sleep(Duration::from_millis(1000)).await;
+                let properties = blob_client.get_properties().into_future().await?;
+                let Some(status) = properties.blob.properties.copy_status else {
+                    tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
+                    return Ok(());
+                };
+                copy_status = Some(status);
             }
-            // The copy is taking longer. Waiting a second and then re-trying.
-            // TODO estimate time based on copy_progress and adjust time based on that
-            tokio::time::sleep(Duration::from_millis(1000)).await;
-            let properties = blob_client.get_properties().into_future().await?;
-            let Some(status) = properties.blob.properties.copy_status else {
-                tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
-                return Ok(());
-            };
-            if start_time.elapsed() > MAX_WAIT_TIME {
-                anyhow::bail!("Copy from from {from} to {to} took longer than limit MAX_WAIT_TIME={}s. copy_pogress={:?}.",
-                    MAX_WAIT_TIME.as_secs_f32(),
-                    properties.blob.properties.copy_progress,
-                );
-            }
-            copy_status = status;
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(anyhow::Error::new(TimeoutOrCancel::Cancel)),
+            _ = timeout => {
+                let e = anyhow::Error::new(TimeoutOrCancel::Timeout);
+                let e = e.context(format!("Timeout, last status: {copy_status:?}"));
+                Err(e)
+            },
         }
     }
 

libs/remote_storage/src/error.rs

@@ -0,0 +1,200 @@
+/// Reasons for downloads or listings to fail.
+#[derive(Debug)]
+pub enum DownloadError {
+    /// Validation or other error happened due to user input.
+    BadInput(anyhow::Error),
+    /// The file was not found in the remote storage.
+    NotFound,
+    /// A cancellation token aborted the download, typically during
+    /// tenant detach or process shutdown.
+    Cancelled,
+    /// A timeout happened while executing the request. Possible reasons:
+    /// - stuck tcp connection
+    ///
+    /// Concurrency control is not timed within timeout.
+    Timeout,
+    /// The file was found in the remote storage, but the download failed.
+    Other(anyhow::Error),
+}
+
+impl std::fmt::Display for DownloadError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            DownloadError::BadInput(e) => {
+                write!(f, "Failed to download a remote file due to user input: {e}")
+            }
+            DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
+            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
+            DownloadError::Timeout => write!(f, "timeout"),
+            DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
+        }
+    }
+}
+
+impl std::error::Error for DownloadError {}
+
+impl DownloadError {
+    /// Returns true if the error should not be retried with backoff
+    pub fn is_permanent(&self) -> bool {
+        use DownloadError::*;
+        match self {
+            BadInput(_) | NotFound | Cancelled => true,
+            Timeout | Other(_) => false,
+        }
+    }
+}
+
+impl From<std::io::Error> for DownloadError {
+    fn from(value: std::io::Error) -> Self {
+        let needs_unwrap = value.kind() == std::io::ErrorKind::Other
+            && value
+                .get_ref()
+                .and_then(|x| x.downcast_ref::<DownloadError>())
+                .is_some();
+
+        if needs_unwrap {
+            *value
+                .into_inner()
+                .expect("just checked")
+                .downcast::<DownloadError>()
+                .expect("just checked")
+        } else {
+            DownloadError::Other(value.into())
+        }
+    }
+}
+
+#[derive(Debug)]
+pub enum TimeTravelError {
+    /// Validation or other error happened due to user input.
+    BadInput(anyhow::Error),
+    /// The used remote storage does not have time travel recovery implemented
+    Unimplemented,
+    /// The number of versions/deletion markers is above our limit.
+    TooManyVersions,
+    /// A cancellation token aborted the process, typically during
+    /// request closure or process shutdown.
+    Cancelled,
+    /// Other errors
+    Other(anyhow::Error),
+}
+
+impl std::fmt::Display for TimeTravelError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            TimeTravelError::BadInput(e) => {
+                write!(
+                    f,
+                    "Failed to time travel recover a prefix due to user input: {e}"
+                )
+            }
+            TimeTravelError::Unimplemented => write!(
+                f,
+                "time travel recovery is not implemented for the current storage backend"
+            ),
+            TimeTravelError::Cancelled => write!(f, "Cancelled, shutting down"),
+            TimeTravelError::TooManyVersions => {
+                write!(f, "Number of versions/delete markers above limit")
+            }
+            TimeTravelError::Other(e) => write!(f, "Failed to time travel recover a prefix: {e:?}"),
+        }
+    }
+}
+
+impl std::error::Error for TimeTravelError {}
+
+/// Plain cancelled error.
+///
+/// By design this type does not not implement `std::error::Error` so it cannot be put as the root
+/// cause of `std::io::Error` or `anyhow::Error`. It should never need to be exposed out of this
+/// crate.
+///
+/// It exists to implement permit acquiring in `{Download,TimeTravel}Error` and `anyhow::Error` returning
+/// operations and ensuring that those get converted to proper versions with just `?`.
+#[derive(Debug)]
+pub(crate) struct Cancelled;
+
+impl From<Cancelled> for anyhow::Error {
+    fn from(_: Cancelled) -> Self {
+        anyhow::Error::new(TimeoutOrCancel::Cancel)
+    }
+}
+
+impl From<Cancelled> for TimeTravelError {
+    fn from(_: Cancelled) -> Self {
+        TimeTravelError::Cancelled
+    }
+}
+
+impl From<Cancelled> for TimeoutOrCancel {
+    fn from(_: Cancelled) -> Self {
+        TimeoutOrCancel::Cancel
+    }
+}
+
+impl From<Cancelled> for DownloadError {
+    fn from(_: Cancelled) -> Self {
+        DownloadError::Cancelled
+    }
+}
+
+/// This type is used at as the root cause for timeouts and cancellations with `anyhow::Error` returning
+/// RemoteStorage methods.
+///
+/// For use with `utils::backoff::retry` and `anyhow::Error` returning operations there is
+/// `TimeoutOrCancel::caused_by_cancel` method to query "proper form" errors.
+#[derive(Debug)]
+pub enum TimeoutOrCancel {
+    Timeout,
+    Cancel,
+}
+
+impl std::fmt::Display for TimeoutOrCancel {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        use TimeoutOrCancel::*;
+        match self {
+            Timeout => write!(f, "timeout"),
+            Cancel => write!(f, "cancel"),
+        }
+    }
+}
+
+impl std::error::Error for TimeoutOrCancel {}
+
+impl TimeoutOrCancel {
+    /// Returns true if the error was caused by [`TimeoutOrCancel::Cancel`].
+    pub fn caused_by_cancel(error: &anyhow::Error) -> bool {
+        error
+            .root_cause()
+            .downcast_ref::<Self>()
+            .is_some_and(Self::is_cancel)
+    }
+
+    pub fn is_cancel(&self) -> bool {
+        matches!(self, TimeoutOrCancel::Cancel)
+    }
+
+    pub fn is_timeout(&self) -> bool {
+        matches!(self, TimeoutOrCancel::Timeout)
+    }
+}
+
+/// This conversion is used when [`crate::support::DownloadStream`] notices a cancellation or
+/// timeout to wrap it in an `std::io::Error`.
+impl From<TimeoutOrCancel> for std::io::Error {
+    fn from(value: TimeoutOrCancel) -> Self {
+        let e = DownloadError::from(value);
+        std::io::Error::other(e)
+    }
+}
+
+impl From<TimeoutOrCancel> for DownloadError {
+    fn from(value: TimeoutOrCancel) -> Self {
+        use TimeoutOrCancel::*;
+
+        match value {
+            Timeout => DownloadError::Timeout,
+            Cancel => DownloadError::Cancelled,
+        }
+    }
+}

libs/remote_storage/src/lib.rs

@@ -10,6 +10,7 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 
 mod azure_blob;
+mod error;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
@@ -21,7 +22,7 @@ use std::{
     num::{NonZeroU32, NonZeroUsize},
     pin::Pin,
     sync::Arc,
-    time::SystemTime,
+    time::{Duration, SystemTime},
 };
 
 use anyhow::{bail, Context};
@@ -41,6 +42,8 @@ pub use self::{
 };
 use s3_bucket::RequestKind;
 
+pub use error::{DownloadError, TimeTravelError, TimeoutOrCancel};
+
 /// Currently, sync happens with AWS S3, that has two limits on requests per second:
 /// ~200 RPS for IAM services
 /// <https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html>
@@ -158,9 +161,10 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         let result = self
-            .list(prefix, ListingMode::WithDelimiter, None)
+            .list(prefix, ListingMode::WithDelimiter, None, cancel)
             .await?
             .prefixes;
         Ok(result)
@@ -182,9 +186,10 @@ pub trait RemoteStorage: Send + Sync + 'static {
         &self,
         prefix: Option<&RemotePath>,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         let result = self
-            .list(prefix, ListingMode::NoDelimiter, max_keys)
+            .list(prefix, ListingMode::NoDelimiter, max_keys, cancel)
             .await?
             .keys;
         Ok(result)
@@ -195,9 +200,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
         prefix: Option<&RemotePath>,
         _mode: ListingMode,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Listing, DownloadError>;
 
     /// Streams the local file contents into remote into the remote storage entry.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`.
     async fn upload(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
@@ -206,27 +215,61 @@ pub trait RemoteStorage: Send + Sync + 'static {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()>;
 
-    /// Streams the remote storage entry contents into the buffered writer given, returns the filled writer.
+    /// Streams the remote storage entry contents.
+    ///
+    /// The returned download stream will obey initial timeout and cancellation signal by erroring
+    /// on whichever happens first. Only one of the reasons will fail the stream, which is usually
+    /// enough for `tokio::io::copy_buf` usage. If needed the error can be filtered out.
+    ///
     /// Returns the metadata, if any was stored with the file previously.
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError>;
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError>;
 
-    /// Streams a given byte range of the remote storage entry contents into the buffered writer given, returns the filled writer.
+    /// Streams a given byte range of the remote storage entry contents.
+    ///
+    /// The returned download stream will obey initial timeout and cancellation signal by erroring
+    /// on whichever happens first. Only one of the reasons will fail the stream, which is usually
+    /// enough for `tokio::io::copy_buf` usage. If needed the error can be filtered out.
+    ///
     /// Returns the metadata, if any was stored with the file previously.
     async fn download_byte_range(
         &self,
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError>;
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()>;
+    /// Delete a single path from remote storage.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`. In such situation it is unknown if the deletion went through.
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()>;
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()>;
+    /// Delete a multiple paths from remote storage.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`. In such situation it is unknown which deletions, if any, went
+    /// through.
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()>;
 
     /// Copy a remote object inside a bucket from one path to another.
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()>;
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()>;
 
     /// Resets the content of everything with the given prefix to the given state
     async fn time_travel_recover(
@@ -238,7 +281,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
     ) -> Result<(), TimeTravelError>;
 }
 
-pub type DownloadStream = Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>;
+/// DownloadStream is sensitive to the timeout and cancellation used with the original
+/// [`RemoteStorage::download`] request. The type yields `std::io::Result<Bytes>` to be compatible
+/// with `tokio::io::copy_buf`.
+// This has 'static because safekeepers do not use cancellation tokens (yet)
+pub type DownloadStream =
+    Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>>;
+
 pub struct Download {
     pub download_stream: DownloadStream,
     /// The last time the file was modified (`last-modified` HTTP header)
@@ -257,86 +306,6 @@ impl Debug for Download {
     }
 }
 
-#[derive(Debug)]
-pub enum DownloadError {
-    /// Validation or other error happened due to user input.
-    BadInput(anyhow::Error),
-    /// The file was not found in the remote storage.
-    NotFound,
-    /// A cancellation token aborted the download, typically during
-    /// tenant detach or process shutdown.
-    Cancelled,
-    /// The file was found in the remote storage, but the download failed.
-    Other(anyhow::Error),
-}
-
-impl std::fmt::Display for DownloadError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            DownloadError::BadInput(e) => {
-                write!(f, "Failed to download a remote file due to user input: {e}")
-            }
-            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
-            DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
-            DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
-        }
-    }
-}
-
-impl std::error::Error for DownloadError {}
-
-impl DownloadError {
-    /// Returns true if the error should not be retried with backoff
-    pub fn is_permanent(&self) -> bool {
-        use DownloadError::*;
-        match self {
-            BadInput(_) => true,
-            NotFound => true,
-            Cancelled => true,
-            Other(_) => false,
-        }
-    }
-}
-
-#[derive(Debug)]
-pub enum TimeTravelError {
-    /// Validation or other error happened due to user input.
-    BadInput(anyhow::Error),
-    /// The used remote storage does not have time travel recovery implemented
-    Unimplemented,
-    /// The number of versions/deletion markers is above our limit.
-    TooManyVersions,
-    /// A cancellation token aborted the process, typically during
-    /// request closure or process shutdown.
-    Cancelled,
-    /// Other errors
-    Other(anyhow::Error),
-}
-
-impl std::fmt::Display for TimeTravelError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            TimeTravelError::BadInput(e) => {
-                write!(
-                    f,
-                    "Failed to time travel recover a prefix due to user input: {e}"
-                )
-            }
-            TimeTravelError::Unimplemented => write!(
-                f,
-                "time travel recovery is not implemented for the current storage backend"
-            ),
-            TimeTravelError::Cancelled => write!(f, "Cancelled, shutting down"),
-            TimeTravelError::TooManyVersions => {
-                write!(f, "Number of versions/delete markers above limit")
-            }
-            TimeTravelError::Other(e) => write!(f, "Failed to time travel recover a prefix: {e:?}"),
-        }
-    }
-}
-
-impl std::error::Error for TimeTravelError {}
-
 /// Every storage, currently supported.
 /// Serves as a simple way to pass around the [`RemoteStorage`] without dealing with generics.
 #[derive(Clone)]
@@ -354,12 +323,13 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         prefix: Option<&RemotePath>,
         mode: ListingMode,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<Listing, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list(prefix, mode, max_keys).await,
-            Self::AwsS3(s) => s.list(prefix, mode, max_keys).await,
-            Self::AzureBlob(s) => s.list(prefix, mode, max_keys).await,
-            Self::Unreliable(s) => s.list(prefix, mode, max_keys).await,
+            Self::LocalFs(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::AwsS3(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::AzureBlob(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::Unreliable(s) => s.list(prefix, mode, max_keys, cancel).await,
         }
     }
 
@@ -372,12 +342,13 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         &self,
         folder: Option<&RemotePath>,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list_files(folder, max_keys).await,
-            Self::AwsS3(s) => s.list_files(folder, max_keys).await,
-            Self::AzureBlob(s) => s.list_files(folder, max_keys).await,
-            Self::Unreliable(s) => s.list_files(folder, max_keys).await,
+            Self::LocalFs(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::AwsS3(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::AzureBlob(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::Unreliable(s) => s.list_files(folder, max_keys, cancel).await,
         }
     }
 
@@ -387,36 +358,43 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
     pub async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list_prefixes(prefix).await,
-            Self::AwsS3(s) => s.list_prefixes(prefix).await,
-            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
-            Self::Unreliable(s) => s.list_prefixes(prefix).await,
+            Self::LocalFs(s) => s.list_prefixes(prefix, cancel).await,
+            Self::AwsS3(s) => s.list_prefixes(prefix, cancel).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix, cancel).await,
+            Self::Unreliable(s) => s.list_prefixes(prefix, cancel).await,
         }
     }
 
+    /// See [`RemoteStorage::upload`]
     pub async fn upload(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
         }
     }
 
-    pub async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+    pub async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.download(from).await,
-            Self::AwsS3(s) => s.download(from).await,
-            Self::AzureBlob(s) => s.download(from).await,
-            Self::Unreliable(s) => s.download(from).await,
+            Self::LocalFs(s) => s.download(from, cancel).await,
+            Self::AwsS3(s) => s.download(from, cancel).await,
+            Self::AzureBlob(s) => s.download(from, cancel).await,
+            Self::Unreliable(s) => s.download(from, cancel).await,
         }
     }
 
@@ -425,54 +403,72 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         match self {
             Self::LocalFs(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::AwsS3(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::AzureBlob(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::Unreliable(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
         }
     }
 
-    pub async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::delete`]
+    pub async fn delete(
+        &self,
+        path: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.delete(path).await,
-            Self::AwsS3(s) => s.delete(path).await,
-            Self::AzureBlob(s) => s.delete(path).await,
-            Self::Unreliable(s) => s.delete(path).await,
+            Self::LocalFs(s) => s.delete(path, cancel).await,
+            Self::AwsS3(s) => s.delete(path, cancel).await,
+            Self::AzureBlob(s) => s.delete(path, cancel).await,
+            Self::Unreliable(s) => s.delete(path, cancel).await,
         }
     }
 
-    pub async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::delete_objects`]
+    pub async fn delete_objects(
+        &self,
+        paths: &[RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.delete_objects(paths).await,
-            Self::AwsS3(s) => s.delete_objects(paths).await,
-            Self::AzureBlob(s) => s.delete_objects(paths).await,
-            Self::Unreliable(s) => s.delete_objects(paths).await,
+            Self::LocalFs(s) => s.delete_objects(paths, cancel).await,
+            Self::AwsS3(s) => s.delete_objects(paths, cancel).await,
+            Self::AzureBlob(s) => s.delete_objects(paths, cancel).await,
+            Self::Unreliable(s) => s.delete_objects(paths, cancel).await,
         }
     }
 
-    pub async fn copy_object(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::copy`]
+    pub async fn copy_object(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.copy(from, to).await,
-            Self::AwsS3(s) => s.copy(from, to).await,
-            Self::AzureBlob(s) => s.copy(from, to).await,
-            Self::Unreliable(s) => s.copy(from, to).await,
+            Self::LocalFs(s) => s.copy(from, to, cancel).await,
+            Self::AwsS3(s) => s.copy(from, to, cancel).await,
+            Self::AzureBlob(s) => s.copy(from, to, cancel).await,
+            Self::Unreliable(s) => s.copy(from, to, cancel).await,
         }
     }
 
+    /// See [`RemoteStorage::time_travel_recover`].
     pub async fn time_travel_recover(
         &self,
         prefix: Option<&RemotePath>,
@@ -503,10 +499,11 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
 
 impl GenericRemoteStorage {
     pub fn from_config(storage_config: &RemoteStorageConfig) -> anyhow::Result<Self> {
+        let timeout = storage_config.timeout;
         Ok(match &storage_config.storage {
-            RemoteStorageKind::LocalFs(root) => {
-                info!("Using fs root '{root}' as a remote storage");
-                Self::LocalFs(LocalFs::new(root.clone())?)
+            RemoteStorageKind::LocalFs(path) => {
+                info!("Using fs root '{path}' as a remote storage");
+                Self::LocalFs(LocalFs::new(path.clone(), timeout)?)
             }
             RemoteStorageKind::AwsS3(s3_config) => {
                 // The profile and access key id are only printed here for debugging purposes,
@@ -516,12 +513,12 @@ impl GenericRemoteStorage {
                     std::env::var("AWS_ACCESS_KEY_ID").unwrap_or_else(|_| "<none>".into());
                 info!("Using s3 bucket '{}' in region '{}' as a remote storage, prefix in bucket: '{:?}', bucket endpoint: '{:?}', profile: {profile}, access_key_id: {access_key_id}",
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
-                Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
+                Self::AwsS3(Arc::new(S3Bucket::new(s3_config, timeout)?))
             }
             RemoteStorageKind::AzureContainer(azure_config) => {
                 info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
                       azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
-                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config, timeout)?))
             }
         })
     }
@@ -530,18 +527,15 @@ impl GenericRemoteStorage {
         Self::Unreliable(Arc::new(UnreliableWrapper::new(s, fail_first)))
     }
 
-    /// Takes storage object contents and its size and uploads to remote storage,
-    /// mapping `from_path` to the corresponding remote object id in the storage.
-    ///
-    /// The storage object does not have to be present on the `from_path`,
-    /// this path is used for the remote object id conversion only.
+    /// See [`RemoteStorage::upload`], which this method calls with `None` as metadata.
     pub async fn upload_storage_object(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         from_size_bytes: usize,
         to: &RemotePath,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
-        self.upload(from, from_size_bytes, to, None)
+        self.upload(from, from_size_bytes, to, None, cancel)
             .await
             .with_context(|| {
                 format!("Failed to upload data of length {from_size_bytes} to storage path {to:?}")
@@ -554,10 +548,11 @@ impl GenericRemoteStorage {
         &self,
         byte_range: Option<(u64, Option<u64>)>,
         from: &RemotePath,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         match byte_range {
-            Some((start, end)) => self.download_byte_range(from, start, end).await,
-            None => self.download(from).await,
+            Some((start, end)) => self.download_byte_range(from, start, end, cancel).await,
+            None => self.download(from, cancel).await,
         }
     }
 }
@@ -572,6 +567,9 @@ pub struct StorageMetadata(HashMap<String, String>);
 pub struct RemoteStorageConfig {
     /// The storage connection configuration.
     pub storage: RemoteStorageKind,
+    /// A common timeout enforced for all requests after concurrency limiter permit has been
+    /// acquired.
+    pub timeout: Duration,
 }
 
 /// A kind of a remote storage to connect to, with its connection configuration.
@@ -656,6 +654,8 @@ impl Debug for AzureConfig {
 }
 
 impl RemoteStorageConfig {
+    pub const DEFAULT_TIMEOUT: Duration = std::time::Duration::from_secs(120);
+
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
@@ -685,6 +685,27 @@ impl RemoteStorageConfig {
             .map(|endpoint| parse_toml_string("endpoint", endpoint))
             .transpose()?;
 
+        let timeout = toml
+            .get("timeout")
+            .map(|timeout| {
+                timeout
+                    .as_str()
+                    .ok_or_else(|| anyhow::Error::msg("timeout was not a string"))
+            })
+            .transpose()
+            .and_then(|timeout| {
+                timeout
+                    .map(humantime::parse_duration)
+                    .transpose()
+                    .map_err(anyhow::Error::new)
+            })
+            .context("parse timeout")?
+            .unwrap_or(Self::DEFAULT_TIMEOUT);
+
+        if timeout < Duration::from_secs(1) {
+            bail!("timeout was specified as {timeout:?} which is too low");
+        }
+
         let storage = match (
             local_path,
             bucket_name,
@@ -746,7 +767,7 @@ impl RemoteStorageConfig {
             }
         };
 
-        Ok(Some(RemoteStorageConfig { storage }))
+        Ok(Some(RemoteStorageConfig { storage, timeout }))
     }
 }
 
@@ -842,4 +863,24 @@ mod tests {
         let err = RemotePath::new(Utf8Path::new("/")).expect_err("Should fail on absolute paths");
         assert_eq!(err.to_string(), "Path \"/\" is not relative");
     }
+
+    #[test]
+    fn parse_localfs_config_with_timeout() {
+        let input = "local_path = '.'
+timeout = '5s'";
+
+        let toml = input.parse::<toml_edit::Document>().unwrap();
+
+        let config = RemoteStorageConfig::from_toml(toml.as_item())
+            .unwrap()
+            .expect("it exists");
+
+        assert_eq!(
+            config,
+            RemoteStorageConfig {
+                storage: RemoteStorageKind::LocalFs(Utf8PathBuf::from(".")),
+                timeout: Duration::from_secs(5)
+            }
+        );
+    }
 }

libs/remote_storage/src/local_fs.rs

@@ -5,7 +5,12 @@
 //! volume is mounted to the local FS.
 
 use std::{
-    borrow::Cow, future::Future, io::ErrorKind, num::NonZeroU32, pin::Pin, time::SystemTime,
+    borrow::Cow,
+    future::Future,
+    io::ErrorKind,
+    num::NonZeroU32,
+    pin::Pin,
+    time::{Duration, SystemTime},
 };
 
 use anyhow::{bail, ensure, Context};
@@ -20,7 +25,9 @@ use tokio_util::{io::ReaderStream, sync::CancellationToken};
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
-use crate::{Download, DownloadError, Listing, ListingMode, RemotePath, TimeTravelError};
+use crate::{
+    Download, DownloadError, Listing, ListingMode, RemotePath, TimeTravelError, TimeoutOrCancel,
+};
 
 use super::{RemoteStorage, StorageMetadata};
 
@@ -29,12 +36,13 @@ const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
 #[derive(Debug, Clone)]
 pub struct LocalFs {
     storage_root: Utf8PathBuf,
+    timeout: Duration,
 }
 
 impl LocalFs {
     /// Attempts to create local FS storage, along with its root directory.
     /// Storage root will be created (if does not exist) and transformed into an absolute path (if passed as relative).
-    pub fn new(mut storage_root: Utf8PathBuf) -> anyhow::Result<Self> {
+    pub fn new(mut storage_root: Utf8PathBuf, timeout: Duration) -> anyhow::Result<Self> {
         if !storage_root.exists() {
             std::fs::create_dir_all(&storage_root).with_context(|| {
                 format!("Failed to create all directories in the given root path {storage_root:?}")
@@ -46,7 +54,10 @@ impl LocalFs {
             })?;
         }
 
-        Ok(Self { storage_root })
+        Ok(Self {
+            storage_root,
+            timeout,
+        })
     }
 
     // mirrors S3Bucket::s3_object_to_relative_path
@@ -157,80 +168,14 @@ impl LocalFs {
 
         Ok(files)
     }
-}
 
-impl RemoteStorage for LocalFs {
-    async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-        max_keys: Option<NonZeroU32>,
-    ) -> Result<Listing, DownloadError> {
-        let mut result = Listing::default();
-
-        if let ListingMode::NoDelimiter = mode {
-            let keys = self
-                .list_recursive(prefix)
-                .await
-                .map_err(DownloadError::Other)?;
-
-            result.keys = keys
-                .into_iter()
-                .filter(|k| {
-                    let path = k.with_base(&self.storage_root);
-                    !path.is_dir()
-                })
-                .collect();
-            if let Some(max_keys) = max_keys {
-                result.keys.truncate(max_keys.get() as usize);
-            }
-
-            return Ok(result);
-        }
-
-        let path = match prefix {
-            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
-            None => Cow::Borrowed(&self.storage_root),
-        };
-
-        let prefixes_to_filter = get_all_files(path.as_ref(), false)
-            .await
-            .map_err(DownloadError::Other)?;
-
-        // filter out empty directories to mirror s3 behavior.
-        for prefix in prefixes_to_filter {
-            if prefix.is_dir()
-                && is_directory_empty(&prefix)
-                    .await
-                    .map_err(DownloadError::Other)?
-            {
-                continue;
-            }
-
-            let stripped = prefix
-                .strip_prefix(&self.storage_root)
-                .context("Failed to strip prefix")
-                .and_then(RemotePath::new)
-                .expect(
-                    "We list files for storage root, hence should be able to remote the prefix",
-                );
-
-            if prefix.is_dir() {
-                result.prefixes.push(stripped);
-            } else {
-                result.keys.push(stripped);
-            }
-        }
-
-        Ok(result)
-    }
-
-    async fn upload(
+    async fn upload0(
         &self,
         data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         let target_file_path = to.with_base(&self.storage_root);
         create_target_directory(&target_file_path).await?;
@@ -265,9 +210,26 @@ impl RemoteStorage for LocalFs {
         let mut buffer_to_read = data.take(from_size_bytes);
 
         // alternatively we could just write the bytes to a file, but local_fs is a testing utility
-        let bytes_read = io::copy_buf(&mut buffer_to_read, &mut destination)
-            .await
-            .with_context(|| {
+        let copy = io::copy_buf(&mut buffer_to_read, &mut destination);
+
+        let bytes_read = tokio::select! {
+            biased;
+            _ = cancel.cancelled() => {
+                let file = destination.into_inner();
+                // wait for the inflight operation(s) to complete so that there could be a next
+                // attempt right away and our writes are not directed to their file.
+                file.into_std().await;
+
+                // TODO: leave the temp or not? leaving is probably less racy. enabled truncate at
+                // least.
+                fs::remove_file(temp_file_path).await.context("remove temp_file_path after cancellation or timeout")?;
+                return Err(TimeoutOrCancel::Cancel.into());
+            }
+            read = copy => read,
+        };
+
+        let bytes_read =
+            bytes_read.with_context(|| {
                 format!(
                     "Failed to upload file (write temp) to the local storage at '{temp_file_path}'",
                 )
@@ -299,6 +261,9 @@ impl RemoteStorage for LocalFs {
             })?;
 
         if let Some(storage_metadata) = metadata {
+            // FIXME: we must not be using metadata much, since this would forget the old metadata
+            // for new writes? or perhaps metadata is sticky; could consider removing if it's never
+            // used.
             let storage_metadata_path = storage_metadata_path(&target_file_path);
             fs::write(
                 &storage_metadata_path,
@@ -315,8 +280,131 @@ impl RemoteStorage for LocalFs {
 
         Ok(())
     }
+}
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+impl RemoteStorage for LocalFs {
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Listing, DownloadError> {
+        let op = async {
+            let mut result = Listing::default();
+
+            if let ListingMode::NoDelimiter = mode {
+                let keys = self
+                    .list_recursive(prefix)
+                    .await
+                    .map_err(DownloadError::Other)?;
+
+                result.keys = keys
+                    .into_iter()
+                    .filter(|k| {
+                        let path = k.with_base(&self.storage_root);
+                        !path.is_dir()
+                    })
+                    .collect();
+
+                if let Some(max_keys) = max_keys {
+                    result.keys.truncate(max_keys.get() as usize);
+                }
+
+                return Ok(result);
+            }
+
+            let path = match prefix {
+                Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
+                None => Cow::Borrowed(&self.storage_root),
+            };
+
+            let prefixes_to_filter = get_all_files(path.as_ref(), false)
+                .await
+                .map_err(DownloadError::Other)?;
+
+            // filter out empty directories to mirror s3 behavior.
+            for prefix in prefixes_to_filter {
+                if prefix.is_dir()
+                    && is_directory_empty(&prefix)
+                        .await
+                        .map_err(DownloadError::Other)?
+                {
+                    continue;
+                }
+
+                let stripped = prefix
+                    .strip_prefix(&self.storage_root)
+                    .context("Failed to strip prefix")
+                    .and_then(RemotePath::new)
+                    .expect(
+                        "We list files for storage root, hence should be able to remote the prefix",
+                    );
+
+                if prefix.is_dir() {
+                    result.prefixes.push(stripped);
+                } else {
+                    result.keys.push(stripped);
+                }
+            }
+
+            Ok(result)
+        };
+
+        let timeout = async {
+            tokio::time::sleep(self.timeout).await;
+            Err(DownloadError::Timeout)
+        };
+
+        let cancelled = async {
+            cancel.cancelled().await;
+            Err(DownloadError::Cancelled)
+        };
+
+        tokio::select! {
+            res = op => res,
+            res = timeout => res,
+            res = cancelled => res,
+        }
+    }
+
+    async fn upload(
+        &self,
+        data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let cancel = cancel.child_token();
+
+        let op = self.upload0(data, data_size_bytes, to, metadata, &cancel);
+        let mut op = std::pin::pin!(op);
+
+        // race the upload0 to the timeout; if it goes over, do a graceful shutdown
+        let (res, timeout) = tokio::select! {
+            res = &mut op => (res, false),
+            _ = tokio::time::sleep(self.timeout) => {
+                cancel.cancel();
+                (op.await, true)
+            }
+        };
+
+        match res {
+            Err(e) if timeout && TimeoutOrCancel::caused_by_cancel(&e) => {
+                // we caused this cancel (or they happened simultaneously) -- swap it out to
+                // Timeout
+                Err(TimeoutOrCancel::Timeout.into())
+            }
+            res => res,
+        }
+    }
+
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let target_path = from.with_base(&self.storage_root);
         if file_exists(&target_path).map_err(DownloadError::BadInput)? {
             let source = ReaderStream::new(
@@ -334,6 +422,10 @@ impl RemoteStorage for LocalFs {
                 .read_storage_metadata(&target_path)
                 .await
                 .map_err(DownloadError::Other)?;
+
+            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
             Ok(Download {
                 metadata,
                 last_modified: None,
@@ -350,6 +442,7 @@ impl RemoteStorage for LocalFs {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         if let Some(end_exclusive) = end_exclusive {
             if end_exclusive <= start_inclusive {
@@ -391,6 +484,9 @@ impl RemoteStorage for LocalFs {
             let source = source.take(end_exclusive.unwrap_or(len) - start_inclusive);
             let source = ReaderStream::new(source);
 
+            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
             Ok(Download {
                 metadata,
                 last_modified: None,
@@ -402,7 +498,7 @@ impl RemoteStorage for LocalFs {
         }
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
         let file_path = path.with_base(&self.storage_root);
         match fs::remove_file(&file_path).await {
             Ok(()) => Ok(()),
@@ -414,14 +510,23 @@ impl RemoteStorage for LocalFs {
         }
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         for path in paths {
-            self.delete(path).await?
+            self.delete(path, cancel).await?
         }
         Ok(())
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        _cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         let from_path = from.with_base(&self.storage_root);
         let to_path = to.with_base(&self.storage_root);
         create_target_directory(&to_path).await?;
@@ -435,7 +540,6 @@ impl RemoteStorage for LocalFs {
         Ok(())
     }
 
-    #[allow(clippy::diverging_sub_expression)]
     async fn time_travel_recover(
         &self,
         _prefix: Option<&RemotePath>,
@@ -529,8 +633,9 @@ mod fs_tests {
         remote_storage_path: &RemotePath,
         expected_metadata: Option<&StorageMetadata>,
     ) -> anyhow::Result<String> {
+        let cancel = CancellationToken::new();
         let download = storage
-            .download(remote_storage_path)
+            .download(remote_storage_path, &cancel)
             .await
             .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
         ensure!(
@@ -545,16 +650,16 @@ mod fs_tests {
 
     #[tokio::test]
     async fn upload_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
 
-        let target_path_1 = upload_dummy_file(&storage, "upload_1", None).await?;
+        let target_path_1 = upload_dummy_file(&storage, "upload_1", None, &cancel).await?;
         assert_eq!(
             storage.list_all().await?,
             vec![target_path_1.clone()],
             "Should list a single file after first upload"
         );
 
-        let target_path_2 = upload_dummy_file(&storage, "upload_2", None).await?;
+        let target_path_2 = upload_dummy_file(&storage, "upload_2", None, &cancel).await?;
         assert_eq!(
             list_files_sorted(&storage).await?,
             vec![target_path_1.clone(), target_path_2.clone()],
@@ -566,7 +671,7 @@ mod fs_tests {
 
     #[tokio::test]
     async fn upload_file_negatives() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
 
         let id = RemotePath::new(Utf8Path::new("dummy"))?;
         let content = Bytes::from_static(b"12345");
@@ -575,34 +680,34 @@ mod fs_tests {
         // Check that you get an error if the size parameter doesn't match the actual
         // size of the stream.
         storage
-            .upload(content(), 0, &id, None)
+            .upload(content(), 0, &id, None, &cancel)
             .await
             .expect_err("upload with zero size succeeded");
         storage
-            .upload(content(), 4, &id, None)
+            .upload(content(), 4, &id, None, &cancel)
             .await
             .expect_err("upload with too short size succeeded");
         storage
-            .upload(content(), 6, &id, None)
+            .upload(content(), 6, &id, None, &cancel)
             .await
             .expect_err("upload with too large size succeeded");
 
         // Correct size is 5, this should succeed.
-        storage.upload(content(), 5, &id, None).await?;
+        storage.upload(content(), 5, &id, None, &cancel).await?;
 
         Ok(())
     }
 
-    fn create_storage() -> anyhow::Result<LocalFs> {
+    fn create_storage() -> anyhow::Result<(LocalFs, CancellationToken)> {
         let storage_root = tempdir()?.path().to_path_buf();
-        LocalFs::new(storage_root)
+        LocalFs::new(storage_root, Duration::from_secs(120)).map(|s| (s, CancellationToken::new()))
     }
 
     #[tokio::test]
     async fn download_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
         let contents = read_and_check_metadata(&storage, &upload_target, None).await?;
         assert_eq!(
@@ -612,7 +717,7 @@ mod fs_tests {
         );
 
         let non_existing_path = "somewhere/else";
-        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?).await {
+        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?, &cancel).await {
             Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
             other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
         }
@@ -621,9 +726,9 @@ mod fs_tests {
 
     #[tokio::test]
     async fn download_file_range_positive() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
         let full_range_download_contents =
             read_and_check_metadata(&storage, &upload_target, None).await?;
@@ -637,7 +742,12 @@ mod fs_tests {
         let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
 
         let first_part_download = storage
-            .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
+            .download_byte_range(
+                &upload_target,
+                0,
+                Some(first_part_local.len() as u64),
+                &cancel,
+            )
             .await?;
         assert!(
             first_part_download.metadata.is_none(),
@@ -655,6 +765,7 @@ mod fs_tests {
                 &upload_target,
                 first_part_local.len() as u64,
                 Some((first_part_local.len() + second_part_local.len()) as u64),
+                &cancel,
             )
             .await?;
         assert!(
@@ -669,7 +780,7 @@ mod fs_tests {
         );
 
         let suffix_bytes = storage
-            .download_byte_range(&upload_target, 13, None)
+            .download_byte_range(&upload_target, 13, None, &cancel)
             .await?
             .download_stream;
         let suffix_bytes = aggregate(suffix_bytes).await?;
@@ -677,7 +788,7 @@ mod fs_tests {
         assert_eq!(upload_name, suffix);
 
         let all_bytes = storage
-            .download_byte_range(&upload_target, 0, None)
+            .download_byte_range(&upload_target, 0, None, &cancel)
             .await?
             .download_stream;
         let all_bytes = aggregate(all_bytes).await?;
@@ -689,9 +800,9 @@ mod fs_tests {
 
     #[tokio::test]
     async fn download_file_range_negative() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
         let start = 1_000_000_000;
         let end = start + 1;
@@ -700,6 +811,7 @@ mod fs_tests {
                 &upload_target,
                 start,
                 Some(end), // exclusive end
+                &cancel,
             )
             .await
         {
@@ -716,7 +828,7 @@ mod fs_tests {
         let end = 234;
         assert!(start > end, "Should test an incorrect range");
         match storage
-            .download_byte_range(&upload_target, start, Some(end))
+            .download_byte_range(&upload_target, start, Some(end), &cancel)
             .await
         {
             Ok(_) => panic!("Should not allow downloading wrong ranges"),
@@ -733,15 +845,15 @@ mod fs_tests {
 
     #[tokio::test]
     async fn delete_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
-        storage.delete(&upload_target).await?;
+        storage.delete(&upload_target, &cancel).await?;
         assert!(storage.list_all().await?.is_empty());
 
         storage
-            .delete(&upload_target)
+            .delete(&upload_target, &cancel)
             .await
             .expect("Should allow deleting non-existing storage files");
 
@@ -750,14 +862,14 @@ mod fs_tests {
 
     #[tokio::test]
     async fn file_with_metadata() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
         let metadata = StorageMetadata(HashMap::from([
             ("one".to_string(), "1".to_string()),
             ("two".to_string(), "2".to_string()),
         ]));
         let upload_target =
-            upload_dummy_file(&storage, upload_name, Some(metadata.clone())).await?;
+            upload_dummy_file(&storage, upload_name, Some(metadata.clone()), &cancel).await?;
 
         let full_range_download_contents =
             read_and_check_metadata(&storage, &upload_target, Some(&metadata)).await?;
@@ -771,7 +883,12 @@ mod fs_tests {
         let (first_part_local, _) = uploaded_bytes.split_at(3);
 
         let partial_download_with_metadata = storage
-            .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
+            .download_byte_range(
+                &upload_target,
+                0,
+                Some(first_part_local.len() as u64),
+                &cancel,
+            )
             .await?;
         let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
         assert_eq!(
@@ -792,16 +909,20 @@ mod fs_tests {
     #[tokio::test]
     async fn list() -> anyhow::Result<()> {
         // No delimiter: should recursively list everything
-        let storage = create_storage()?;
-        let child = upload_dummy_file(&storage, "grandparent/parent/child", None).await?;
-        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None).await?;
+        let (storage, cancel) = create_storage()?;
+        let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
+        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
 
-        let listing = storage.list(None, ListingMode::NoDelimiter, None).await?;
+        let listing = storage
+            .list(None, ListingMode::NoDelimiter, None, &cancel)
+            .await?;
         assert!(listing.prefixes.is_empty());
         assert_eq!(listing.keys, [uncle.clone(), child.clone()].to_vec());
 
         // Delimiter: should only go one deep
-        let listing = storage.list(None, ListingMode::WithDelimiter, None).await?;
+        let listing = storage
+            .list(None, ListingMode::WithDelimiter, None, &cancel)
+            .await?;
 
         assert_eq!(
             listing.prefixes,
@@ -815,6 +936,7 @@ mod fs_tests {
                 Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
                 ListingMode::WithDelimiter,
                 None,
+                &cancel,
             )
             .await?;
         assert_eq!(
@@ -827,10 +949,75 @@ mod fs_tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn overwrite_shorter_file() -> anyhow::Result<()> {
+        let (storage, cancel) = create_storage()?;
+
+        let path = RemotePath::new("does/not/matter/file".into())?;
+
+        let body = Bytes::from_static(b"long file contents is long");
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(body, read);
+
+        let shorter = Bytes::from_static(b"shorter body");
+        {
+            let len = shorter.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(shorter.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(shorter, read);
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn cancelled_upload_can_later_be_retried() -> anyhow::Result<()> {
+        let (storage, cancel) = create_storage()?;
+
+        let path = RemotePath::new("does/not/matter/file".into())?;
+
+        let body = Bytes::from_static(b"long file contents is long");
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            let cancel = cancel.child_token();
+            cancel.cancel();
+            let e = storage
+                .upload(body, len, &path, None, &cancel)
+                .await
+                .unwrap_err();
+
+            assert!(TimeoutOrCancel::caused_by_cancel(&e));
+        }
+
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(body, read);
+
+        Ok(())
+    }
+
     async fn upload_dummy_file(
         storage: &LocalFs,
         name: &str,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<RemotePath> {
         let from_path = storage
             .storage_root
@@ -852,7 +1039,9 @@ mod fs_tests {
 
         let file = tokio_util::io::ReaderStream::new(file);
 
-        storage.upload(file, size, &relative_path, metadata).await?;
+        storage
+            .upload(file, size, &relative_path, metadata, cancel)
+            .await?;
         Ok(relative_path)
     }
 

libs/remote_storage/src/s3_bucket.rs

@@ -11,7 +11,7 @@ use std::{
     pin::Pin,
     sync::Arc,
     task::{Context, Poll},
-    time::SystemTime,
+    time::{Duration, SystemTime},
 };
 
 use anyhow::{anyhow, Context as _};
@@ -46,9 +46,9 @@ use utils::backoff;
 
 use super::StorageMetadata;
 use crate::{
-    support::PermitCarrying, ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode,
-    RemotePath, RemoteStorage, S3Config, TimeTravelError, MAX_KEYS_PER_DELETE,
-    REMOTE_STORAGE_PREFIX_SEPARATOR,
+    error::Cancelled, support::PermitCarrying, ConcurrencyLimiter, Download, DownloadError,
+    Listing, ListingMode, RemotePath, RemoteStorage, S3Config, TimeTravelError, TimeoutOrCancel,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
@@ -63,6 +63,8 @@ pub struct S3Bucket {
     prefix_in_bucket: Option<String>,
     max_keys_per_list_response: Option<i32>,
     concurrency_limiter: ConcurrencyLimiter,
+    // Per-request timeout. Accessible for tests.
+    pub timeout: Duration,
 }
 
 struct GetObjectRequest {
@@ -72,7 +74,7 @@ struct GetObjectRequest {
 }
 impl S3Bucket {
     /// Creates the S3 storage, errors if incorrect AWS S3 configuration provided.
-    pub fn new(aws_config: &S3Config) -> anyhow::Result<Self> {
+    pub fn new(aws_config: &S3Config, timeout: Duration) -> anyhow::Result<Self> {
         tracing::debug!(
             "Creating s3 remote storage for S3 bucket {}",
             aws_config.bucket_name
@@ -152,6 +154,7 @@ impl S3Bucket {
             max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
             concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
+            timeout,
         })
     }
 
@@ -185,40 +188,55 @@ impl S3Bucket {
         }
     }
 
-    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+    async fn permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
         let started_at = start_counting_cancelled_wait(kind);
-        let permit = self
-            .concurrency_limiter
-            .acquire(kind)
-            .await
-            .expect("semaphore is never closed");
+        let acquire = self.concurrency_limiter.acquire(kind);
+
+        let permit = tokio::select! {
+            permit = acquire => permit.expect("semaphore is never closed"),
+            _ = cancel.cancelled() => return Err(Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
             .wait_seconds
             .observe_elapsed(kind, started_at);
 
-        permit
+        Ok(permit)
     }
 
-    async fn owned_permit(&self, kind: RequestKind) -> tokio::sync::OwnedSemaphorePermit {
+    async fn owned_permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, Cancelled> {
         let started_at = start_counting_cancelled_wait(kind);
-        let permit = self
-            .concurrency_limiter
-            .acquire_owned(kind)
-            .await
-            .expect("semaphore is never closed");
+        let acquire = self.concurrency_limiter.acquire_owned(kind);
+
+        let permit = tokio::select! {
+            permit = acquire => permit.expect("semaphore is never closed"),
+            _ = cancel.cancelled() => return Err(Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
             .wait_seconds
             .observe_elapsed(kind, started_at);
-        permit
+        Ok(permit)
     }
 
-    async fn download_object(&self, request: GetObjectRequest) -> Result<Download, DownloadError> {
+    async fn download_object(
+        &self,
+        request: GetObjectRequest,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let kind = RequestKind::Get;
-        let permit = self.owned_permit(kind).await;
+
+        let permit = self.owned_permit(kind, cancel).await?;
 
         let started_at = start_measuring_requests(kind);
 
@@ -228,8 +246,13 @@ impl S3Bucket {
             .bucket(request.bucket)
             .key(request.key)
             .set_range(request.range)
-            .send()
-            .await;
+            .send();
+
+        let get_object = tokio::select! {
+            res = get_object => res,
+            _ = tokio::time::sleep(self.timeout) => return Err(DownloadError::Timeout),
+            _ = cancel.cancelled() => return Err(DownloadError::Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
 
@@ -259,6 +282,10 @@ impl S3Bucket {
             }
         };
 
+        // even if we would have no timeout left, continue anyways. the caller can decide to ignore
+        // the errors considering timeouts and cancellation.
+        let remaining = self.timeout.saturating_sub(started_at.elapsed());
+
         let metadata = object_output.metadata().cloned().map(StorageMetadata);
         let etag = object_output.e_tag;
         let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
@@ -268,6 +295,9 @@ impl S3Bucket {
         let body = PermitCarrying::new(permit, body);
         let body = TimedDownload::new(started_at, body);
 
+        let cancel_or_timeout = crate::support::cancel_or_timeout(remaining, cancel.clone());
+        let body = crate::support::DownloadStream::new(cancel_or_timeout, body);
+
         Ok(Download {
             metadata,
             etag,
@@ -278,33 +308,44 @@ impl S3Bucket {
 
     async fn delete_oids(
         &self,
-        kind: RequestKind,
+        _permit: &tokio::sync::SemaphorePermit<'_>,
         delete_objects: &[ObjectIdentifier],
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
+        let kind = RequestKind::Delete;
+        let mut cancel = std::pin::pin!(cancel.cancelled());
+
         for chunk in delete_objects.chunks(MAX_KEYS_PER_DELETE) {
             let started_at = start_measuring_requests(kind);
 
-            let resp = self
+            let req = self
                 .client
                 .delete_objects()
                 .bucket(self.bucket_name.clone())
                 .delete(
                     Delete::builder()
                         .set_objects(Some(chunk.to_vec()))
-                        .build()?,
+                        .build()
+                        .context("build request")?,
                 )
-                .send()
-                .await;
+                .send();
+
+            let resp = tokio::select! {
+                resp = req => resp,
+                _ = tokio::time::sleep(self.timeout) => return Err(TimeoutOrCancel::Timeout.into()),
+                _ = &mut cancel => return Err(TimeoutOrCancel::Cancel.into()),
+            };
 
             let started_at = ScopeGuard::into_inner(started_at);
             metrics::BUCKET_METRICS
                 .req_seconds
                 .observe_elapsed(kind, &resp, started_at);
 
-            let resp = resp?;
+            let resp = resp.context("request deletion")?;
             metrics::BUCKET_METRICS
                 .deleted_objects_total
                 .inc_by(chunk.len() as u64);
+
             if let Some(errors) = resp.errors {
                 // Log a bounded number of the errors within the response:
                 // these requests can carry 1000 keys so logging each one
@@ -320,9 +361,10 @@ impl S3Bucket {
                     );
                 }
 
-                return Err(anyhow::format_err!(
-                    "Failed to delete {} objects",
-                    errors.len()
+                return Err(anyhow::anyhow!(
+                    "Failed to delete {}/{} objects",
+                    errors.len(),
+                    chunk.len(),
                 ));
             }
         }
@@ -410,6 +452,7 @@ impl RemoteStorage for S3Bucket {
         prefix: Option<&RemotePath>,
         mode: ListingMode,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Listing, DownloadError> {
         let kind = RequestKind::List;
         // s3 sdk wants i32
@@ -431,10 +474,11 @@ impl RemoteStorage for S3Bucket {
                 p
             });
 
+        let _permit = self.permit(kind, cancel).await?;
+
         let mut continuation_token = None;
 
         loop {
-            let _guard = self.permit(kind).await;
             let started_at = start_measuring_requests(kind);
 
             // min of two Options, returning Some if one is value and another is
@@ -456,9 +500,15 @@ impl RemoteStorage for S3Bucket {
                 request = request.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
             }
 
-            let response = request
-                .send()
-                .await
+            let request = request.send();
+
+            let response = tokio::select! {
+                res = request => res,
+                _ = tokio::time::sleep(self.timeout) => return Err(DownloadError::Timeout),
+                _ = cancel.cancelled() => return Err(DownloadError::Cancelled),
+            };
+
+            let response = response
                 .context("Failed to list S3 prefixes")
                 .map_err(DownloadError::Other);
 
@@ -511,16 +561,17 @@ impl RemoteStorage for S3Bucket {
         from_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         let kind = RequestKind::Put;
-        let _guard = self.permit(kind).await;
+        let _permit = self.permit(kind, cancel).await?;
 
         let started_at = start_measuring_requests(kind);
 
         let body = Body::wrap_stream(from);
         let bytes_stream = ByteStream::new(SdkBody::from_body_0_4(body));
 
-        let res = self
+        let upload = self
             .client
             .put_object()
             .bucket(self.bucket_name.clone())
@@ -528,22 +579,40 @@ impl RemoteStorage for S3Bucket {
             .set_metadata(metadata.map(|m| m.0))
             .content_length(from_size_bytes.try_into()?)
             .body(bytes_stream)
-            .send()
-            .await;
+            .send();
 
-        let started_at = ScopeGuard::into_inner(started_at);
-        metrics::BUCKET_METRICS
-            .req_seconds
-            .observe_elapsed(kind, &res, started_at);
+        let upload = tokio::time::timeout(self.timeout, upload);
 
-        res?;
+        let res = tokio::select! {
+            res = upload => res,
+            _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
+        };
 
-        Ok(())
+        if let Ok(inner) = &res {
+            // do not incl. timeouts as errors in metrics but cancellations
+            let started_at = ScopeGuard::into_inner(started_at);
+            metrics::BUCKET_METRICS
+                .req_seconds
+                .observe_elapsed(kind, inner, started_at);
+        }
+
+        match res {
+            Ok(Ok(_put)) => Ok(()),
+            Ok(Err(sdk)) => Err(sdk.into()),
+            Err(_timeout) => Err(TimeoutOrCancel::Timeout.into()),
+        }
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         let kind = RequestKind::Copy;
-        let _guard = self.permit(kind).await;
+        let _permit = self.permit(kind, cancel).await?;
+
+        let timeout = tokio::time::sleep(self.timeout);
 
         let started_at = start_measuring_requests(kind);
 
@@ -554,14 +623,19 @@ impl RemoteStorage for S3Bucket {
             self.relative_path_to_s3_object(from)
         );
 
-        let res = self
+        let op = self
             .client
             .copy_object()
             .bucket(self.bucket_name.clone())
             .key(self.relative_path_to_s3_object(to))
             .copy_source(copy_source)
-            .send()
-            .await;
+            .send();
+
+        let res = tokio::select! {
+            res = op => res,
+            _ = timeout => return Err(TimeoutOrCancel::Timeout.into()),
+            _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
@@ -573,14 +647,21 @@ impl RemoteStorage for S3Bucket {
         Ok(())
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         // if prefix is not none then download file `prefix/from`
         // if prefix is none then download file `from`
-        self.download_object(GetObjectRequest {
-            bucket: self.bucket_name.clone(),
-            key: self.relative_path_to_s3_object(from),
-            range: None,
-        })
+        self.download_object(
+            GetObjectRequest {
+                bucket: self.bucket_name.clone(),
+                key: self.relative_path_to_s3_object(from),
+                range: None,
+            },
+            cancel,
+        )
         .await
     }
 
@@ -589,6 +670,7 @@ impl RemoteStorage for S3Bucket {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         // S3 accepts ranges as https://www.w3.org/Protocols/rfc2616/rfc2616-sec14.html#sec14.35
         // and needs both ends to be exclusive
@@ -598,31 +680,39 @@ impl RemoteStorage for S3Bucket {
             None => format!("bytes={start_inclusive}-"),
         });
 
-        self.download_object(GetObjectRequest {
-            bucket: self.bucket_name.clone(),
-            key: self.relative_path_to_s3_object(from),
-            range,
-        })
+        self.download_object(
+            GetObjectRequest {
+                bucket: self.bucket_name.clone(),
+                key: self.relative_path_to_s3_object(from),
+                range,
+            },
+            cancel,
+        )
         .await
     }
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
-        let kind = RequestKind::Delete;
-        let _guard = self.permit(kind).await;
 
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let kind = RequestKind::Delete;
+        let permit = self.permit(kind, cancel).await?;
         let mut delete_objects = Vec::with_capacity(paths.len());
         for path in paths {
             let obj_id = ObjectIdentifier::builder()
                 .set_key(Some(self.relative_path_to_s3_object(path)))
-                .build()?;
+                .build()
+                .context("convert path to oid")?;
             delete_objects.push(obj_id);
         }
 
-        self.delete_oids(kind, &delete_objects).await
+        self.delete_oids(&permit, &delete_objects, cancel).await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
         let paths = std::array::from_ref(path);
-        self.delete_objects(paths).await
+        self.delete_objects(paths, cancel).await
     }
 
     async fn time_travel_recover(
@@ -633,7 +723,7 @@ impl RemoteStorage for S3Bucket {
         cancel: &CancellationToken,
     ) -> Result<(), TimeTravelError> {
         let kind = RequestKind::TimeTravel;
-        let _guard = self.permit(kind).await;
+        let permit = self.permit(kind, cancel).await?;
 
         let timestamp = DateTime::from(timestamp);
         let done_if_after = DateTime::from(done_if_after);
@@ -647,7 +737,7 @@ impl RemoteStorage for S3Bucket {
 
         let warn_threshold = 3;
         let max_retries = 10;
-        let is_permanent = |_e: &_| false;
+        let is_permanent = |e: &_| matches!(e, TimeTravelError::Cancelled);
 
         let mut key_marker = None;
         let mut version_id_marker = None;
@@ -656,15 +746,19 @@ impl RemoteStorage for S3Bucket {
         loop {
             let response = backoff::retry(
                 || async {
-                    self.client
+                    let op = self
+                        .client
                         .list_object_versions()
                         .bucket(self.bucket_name.clone())
                         .set_prefix(prefix.clone())
                         .set_key_marker(key_marker.clone())
                         .set_version_id_marker(version_id_marker.clone())
-                        .send()
-                        .await
-                        .map_err(|e| TimeTravelError::Other(e.into()))
+                        .send();
+
+                    tokio::select! {
+                        res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
+                        _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
+                    }
                 },
                 is_permanent,
                 warn_threshold,
@@ -786,14 +880,18 @@ impl RemoteStorage for S3Bucket {
 
                         backoff::retry(
                             || async {
-                                self.client
+                                let op = self
+                                    .client
                                     .copy_object()
                                     .bucket(self.bucket_name.clone())
                                     .key(key)
                                     .copy_source(&source_id)
-                                    .send()
-                                    .await
-                                    .map_err(|e| TimeTravelError::Other(e.into()))
+                                    .send();
+
+                                tokio::select! {
+                                    res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
+                                    _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
+                                }
                             },
                             is_permanent,
                             warn_threshold,
@@ -824,10 +922,18 @@ impl RemoteStorage for S3Bucket {
                     let oid = ObjectIdentifier::builder()
                         .key(key.to_owned())
                         .build()
-                        .map_err(|e| TimeTravelError::Other(anyhow::Error::new(e)))?;
-                    self.delete_oids(kind, &[oid])
+                        .map_err(|e| TimeTravelError::Other(e.into()))?;
+
+                    self.delete_oids(&permit, &[oid], cancel)
                         .await
-                        .map_err(TimeTravelError::Other)?;
+                        .map_err(|e| {
+                            // delete_oid0 will use TimeoutOrCancel
+                            if TimeoutOrCancel::caused_by_cancel(&e) {
+                                TimeTravelError::Cancelled
+                            } else {
+                                TimeTravelError::Other(e)
+                            }
+                        })?;
                 }
             }
         }
@@ -963,7 +1069,8 @@ mod tests {
                 concurrency_limit: NonZeroUsize::new(100).unwrap(),
                 max_keys_per_list_response: Some(5),
             };
-            let storage = S3Bucket::new(&config).expect("remote storage init");
+            let storage =
+                S3Bucket::new(&config, std::time::Duration::ZERO).expect("remote storage init");
             for (test_path_idx, test_path) in all_paths.iter().enumerate() {
                 let result = storage.relative_path_to_s3_object(test_path);
                 let expected = expected_outputs[prefix_idx][test_path_idx];

libs/remote_storage/src/simulate_failures.rs

@@ -90,11 +90,16 @@ impl UnreliableWrapper {
         }
     }
 
-    async fn delete_inner(&self, path: &RemotePath, attempt: bool) -> anyhow::Result<()> {
+    async fn delete_inner(
+        &self,
+        path: &RemotePath,
+        attempt: bool,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         if attempt {
             self.attempt(RemoteOp::Delete(path.clone()))?;
         }
-        self.inner.delete(path).await
+        self.inner.delete(path, cancel).await
     }
 }
 
@@ -105,20 +110,22 @@ impl RemoteStorage for UnreliableWrapper {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))
             .map_err(DownloadError::Other)?;
-        self.inner.list_prefixes(prefix).await
+        self.inner.list_prefixes(prefix, cancel).await
     }
 
     async fn list_files(
         &self,
         folder: Option<&RemotePath>,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         self.attempt(RemoteOp::ListPrefixes(folder.cloned()))
             .map_err(DownloadError::Other)?;
-        self.inner.list_files(folder, max_keys).await
+        self.inner.list_files(folder, max_keys, cancel).await
     }
 
     async fn list(
@@ -126,10 +133,11 @@ impl RemoteStorage for UnreliableWrapper {
         prefix: Option<&RemotePath>,
         mode: ListingMode,
         max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Listing, DownloadError> {
         self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))
             .map_err(DownloadError::Other)?;
-        self.inner.list(prefix, mode, max_keys).await
+        self.inner.list(prefix, mode, max_keys, cancel).await
     }
 
     async fn upload(
@@ -140,15 +148,22 @@ impl RemoteStorage for UnreliableWrapper {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         self.attempt(RemoteOp::Upload(to.clone()))?;
-        self.inner.upload(data, data_size_bytes, to, metadata).await
+        self.inner
+            .upload(data, data_size_bytes, to, metadata, cancel)
+            .await
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         self.attempt(RemoteOp::Download(from.clone()))
             .map_err(DownloadError::Other)?;
-        self.inner.download(from).await
+        self.inner.download(from, cancel).await
     }
 
     async fn download_byte_range(
@@ -156,6 +171,7 @@ impl RemoteStorage for UnreliableWrapper {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         // Note: We treat any download_byte_range as an "attempt" of the same
         // operation. We don't pay attention to the ranges. That's good enough
@@ -163,20 +179,24 @@ impl RemoteStorage for UnreliableWrapper {
         self.attempt(RemoteOp::Download(from.clone()))
             .map_err(DownloadError::Other)?;
         self.inner
-            .download_byte_range(from, start_inclusive, end_exclusive)
+            .download_byte_range(from, start_inclusive, end_exclusive, cancel)
             .await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
-        self.delete_inner(path, true).await
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
+        self.delete_inner(path, true, cancel).await
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         self.attempt(RemoteOp::DeleteObjects(paths.to_vec()))?;
         let mut error_counter = 0;
         for path in paths {
             // Dont record attempt because it was already recorded above
-            if (self.delete_inner(path, false).await).is_err() {
+            if (self.delete_inner(path, false, cancel).await).is_err() {
                 error_counter += 1;
             }
         }
@@ -189,11 +209,16 @@ impl RemoteStorage for UnreliableWrapper {
         Ok(())
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         // copy is equivalent to download + upload
         self.attempt(RemoteOp::Download(from.clone()))?;
         self.attempt(RemoteOp::Upload(to.clone()))?;
-        self.inner.copy_object(from, to).await
+        self.inner.copy_object(from, to, cancel).await
     }
 
     async fn time_travel_recover(

libs/remote_storage/src/support.rs

@@ -1,9 +1,15 @@
 use std::{
+    future::Future,
     pin::Pin,
     task::{Context, Poll},
+    time::Duration,
 };
 
+use bytes::Bytes;
 use futures_util::Stream;
+use tokio_util::sync::CancellationToken;
+
+use crate::TimeoutOrCancel;
 
 pin_project_lite::pin_project! {
     /// An `AsyncRead` adapter which carries a permit for the lifetime of the value.
@@ -31,3 +37,139 @@ impl<S: Stream> Stream for PermitCarrying<S> {
         self.inner.size_hint()
     }
 }
+
+pin_project_lite::pin_project! {
+    pub(crate) struct DownloadStream<F, S> {
+        hit: bool,
+        #[pin]
+        cancellation: F,
+        #[pin]
+        inner: S,
+    }
+}
+
+impl<F, S> DownloadStream<F, S> {
+    pub(crate) fn new(cancellation: F, inner: S) -> Self {
+        Self {
+            cancellation,
+            hit: false,
+            inner,
+        }
+    }
+}
+
+/// See documentation on [`crate::DownloadStream`] on rationale why `std::io::Error` is used.
+impl<E, F, S> Stream for DownloadStream<F, S>
+where
+    std::io::Error: From<E>,
+    F: Future<Output = E>,
+    S: Stream<Item = std::io::Result<Bytes>>,
+{
+    type Item = <S as Stream>::Item;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.project();
+
+        if !*this.hit {
+            if let Poll::Ready(e) = this.cancellation.poll(cx) {
+                *this.hit = true;
+
+                // most likely this will be a std::io::Error wrapping a DownloadError
+                let e = Err(std::io::Error::from(e));
+                return Poll::Ready(Some(e));
+            }
+        }
+
+        this.inner.poll_next(cx)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+/// Fires only on the first cancel or timeout, not on both.
+pub(crate) async fn cancel_or_timeout(
+    timeout: Duration,
+    cancel: CancellationToken,
+) -> TimeoutOrCancel {
+    tokio::select! {
+        _ = tokio::time::sleep(timeout) => TimeoutOrCancel::Timeout,
+        _ = cancel.cancelled() => TimeoutOrCancel::Cancel,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::DownloadError;
+    use futures::stream::StreamExt;
+
+    #[tokio::test(start_paused = true)]
+    async fn cancelled_download_stream() {
+        let inner = futures::stream::pending();
+        let timeout = Duration::from_secs(120);
+        let cancel = CancellationToken::new();
+
+        let stream = DownloadStream::new(cancel_or_timeout(timeout, cancel.clone()), inner);
+        let mut stream = std::pin::pin!(stream);
+
+        let mut first = stream.next();
+
+        tokio::select! {
+            _ = &mut first => unreachable!("we haven't yet cancelled nor is timeout passed"),
+            _ = tokio::time::sleep(Duration::from_secs(1)) => {},
+        }
+
+        cancel.cancel();
+
+        let e = first.await.expect("there must be some").unwrap_err();
+        assert!(matches!(e.kind(), std::io::ErrorKind::Other), "{e:?}");
+        let inner = e.get_ref().expect("inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Cancelled)),
+            "{inner:?}"
+        );
+        let e = DownloadError::from(e);
+        assert!(matches!(e, DownloadError::Cancelled), "{e:?}");
+
+        tokio::select! {
+            _ = stream.next() => unreachable!("no timeout ever happens as we were already cancelled"),
+            _ = tokio::time::sleep(Duration::from_secs(121)) => {},
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn timeouted_download_stream() {
+        let inner = futures::stream::pending();
+        let timeout = Duration::from_secs(120);
+        let cancel = CancellationToken::new();
+
+        let stream = DownloadStream::new(cancel_or_timeout(timeout, cancel.clone()), inner);
+        let mut stream = std::pin::pin!(stream);
+
+        // because the stream uses 120s timeout and we are paused, we advance to 120s right away.
+        let first = stream.next();
+
+        let e = first.await.expect("there must be some").unwrap_err();
+        assert!(matches!(e.kind(), std::io::ErrorKind::Other), "{e:?}");
+        let inner = e.get_ref().expect("inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Timeout)),
+            "{inner:?}"
+        );
+        let e = DownloadError::from(e);
+        assert!(matches!(e, DownloadError::Timeout), "{e:?}");
+
+        cancel.cancel();
+
+        tokio::select! {
+            _ = stream.next() => unreachable!("no cancellation ever happens because we already timed out"),
+            _ = tokio::time::sleep(Duration::from_secs(121)) => {},
+        }
+    }
+}

libs/remote_storage/tests/common/mod.rs

@@ -10,6 +10,7 @@ use futures::stream::Stream;
 use once_cell::sync::OnceCell;
 use remote_storage::{Download, GenericRemoteStorage, RemotePath};
 use tokio::task::JoinSet;
+use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info};
 
 static LOGGING_DONE: OnceCell<()> = OnceCell::new();
@@ -58,8 +59,12 @@ pub(crate) async fn upload_simple_remote_data(
 ) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
     info!("Creating {upload_tasks_count} remote files");
     let mut upload_tasks = JoinSet::new();
+    let cancel = CancellationToken::new();
+
     for i in 1..upload_tasks_count + 1 {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
+
         upload_tasks.spawn(async move {
             let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
             let blob_path = RemotePath::new(
@@ -69,7 +74,9 @@ pub(crate) async fn upload_simple_remote_data(
             debug!("Creating remote item {i} at path {blob_path:?}");
 
             let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
+            task_client
+                .upload(data, len, &blob_path, None, &cancel)
+                .await?;
 
             Ok::<_, anyhow::Error>(blob_path)
         });
@@ -107,13 +114,15 @@ pub(crate) async fn cleanup(
         "Removing {} objects from the remote storage during cleanup",
         objects_to_delete.len()
     );
+    let cancel = CancellationToken::new();
     let mut delete_tasks = JoinSet::new();
     for object_to_delete in objects_to_delete {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
         delete_tasks.spawn(async move {
             debug!("Deleting remote item at path {object_to_delete:?}");
             task_client
-                .delete(&object_to_delete)
+                .delete(&object_to_delete, &cancel)
                 .await
                 .with_context(|| format!("{object_to_delete:?} removal"))
         });
@@ -141,8 +150,12 @@ pub(crate) async fn upload_remote_data(
 ) -> ControlFlow<Uploads, Uploads> {
     info!("Creating {upload_tasks_count} remote files");
     let mut upload_tasks = JoinSet::new();
+    let cancel = CancellationToken::new();
+
     for i in 1..upload_tasks_count + 1 {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
+
         upload_tasks.spawn(async move {
             let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
             let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
@@ -152,7 +165,9 @@ pub(crate) async fn upload_remote_data(
 
             let (data, data_len) =
                 upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
+            task_client
+                .upload(data, data_len, &blob_path, None, &cancel)
+                .await?;
 
             Ok::<_, anyhow::Error>((blob_prefix, blob_path))
         });

libs/remote_storage/tests/common/tests.rs

@@ -4,6 +4,7 @@ use remote_storage::RemotePath;
 use std::sync::Arc;
 use std::{collections::HashSet, num::NonZeroU32};
 use test_context::test_context;
+use tokio_util::sync::CancellationToken;
 use tracing::debug;
 
 use crate::common::{download_to_vec, upload_stream, wrap_stream};
@@ -45,13 +46,15 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
         }
     };
 
+    let cancel = CancellationToken::new();
+
     let test_client = Arc::clone(&ctx.enabled.client);
     let expected_remote_prefixes = ctx.remote_prefixes.clone();
 
     let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
         .context("common_prefix construction")?;
     let root_remote_prefixes = test_client
-        .list_prefixes(None)
+        .list_prefixes(None, &cancel)
         .await
         .context("client list root prefixes failure")?
         .into_iter()
@@ -62,7 +65,7 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
     );
 
     let nested_remote_prefixes = test_client
-        .list_prefixes(Some(&base_prefix))
+        .list_prefixes(Some(&base_prefix), &cancel)
         .await
         .context("client list nested prefixes failure")?
         .into_iter()
@@ -99,11 +102,12 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
             anyhow::bail!("S3 init failed: {e:?}")
         }
     };
+    let cancel = CancellationToken::new();
     let test_client = Arc::clone(&ctx.enabled.client);
     let base_prefix =
         RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
     let root_files = test_client
-        .list_files(None, None)
+        .list_files(None, None, &cancel)
         .await
         .context("client list root files failure")?
         .into_iter()
@@ -117,13 +121,13 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
     // Test that max_keys limit works. In total there are about 21 files (see
     // upload_simple_remote_data call in test_real_s3.rs).
     let limited_root_files = test_client
-        .list_files(None, Some(NonZeroU32::new(2).unwrap()))
+        .list_files(None, Some(NonZeroU32::new(2).unwrap()), &cancel)
         .await
         .context("client list root files failure")?;
     assert_eq!(limited_root_files.len(), 2);
 
     let nested_remote_files = test_client
-        .list_files(Some(&base_prefix), None)
+        .list_files(Some(&base_prefix), None, &cancel)
         .await
         .context("client list nested files failure")?
         .into_iter()
@@ -150,12 +154,17 @@ async fn delete_non_exising_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Resu
         MaybeEnabledStorage::Disabled => return Ok(()),
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(
         format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
     ))
     .with_context(|| "RemotePath conversion")?;
 
-    ctx.client.delete(&path).await.expect("should succeed");
+    ctx.client
+        .delete(&path, &cancel)
+        .await
+        .expect("should succeed");
 
     Ok(())
 }
@@ -168,6 +177,8 @@ async fn delete_objects_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<(
         MaybeEnabledStorage::Disabled => return Ok(()),
     };
 
+    let cancel = CancellationToken::new();
+
     let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
@@ -178,21 +189,21 @@ async fn delete_objects_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<(
         .with_context(|| "RemotePath conversion")?;
 
     let (data, len) = upload_stream("remote blob data1".as_bytes().into());
-    ctx.client.upload(data, len, &path1, None).await?;
+    ctx.client.upload(data, len, &path1, None, &cancel).await?;
 
     let (data, len) = upload_stream("remote blob data2".as_bytes().into());
-    ctx.client.upload(data, len, &path2, None).await?;
+    ctx.client.upload(data, len, &path2, None, &cancel).await?;
 
     let (data, len) = upload_stream("remote blob data3".as_bytes().into());
-    ctx.client.upload(data, len, &path3, None).await?;
+    ctx.client.upload(data, len, &path3, None, &cancel).await?;
 
-    ctx.client.delete_objects(&[path1, path2]).await?;
+    ctx.client.delete_objects(&[path1, path2], &cancel).await?;
 
-    let prefixes = ctx.client.list_prefixes(None).await?;
+    let prefixes = ctx.client.list_prefixes(None, &cancel).await?;
 
     assert_eq!(prefixes.len(), 1);
 
-    ctx.client.delete_objects(&[path3]).await?;
+    ctx.client.delete_objects(&[path3], &cancel).await?;
 
     Ok(())
 }
@@ -204,6 +215,8 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
         return Ok(());
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
@@ -211,47 +224,56 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
 
     let (data, len) = wrap_stream(orig.clone());
 
-    ctx.client.upload(data, len, &path, None).await?;
+    ctx.client.upload(data, len, &path, None, &cancel).await?;
 
     // Normal download request
-    let dl = ctx.client.download(&path).await?;
+    let dl = ctx.client.download(&path, &cancel).await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // Full range (end specified)
     let dl = ctx
         .client
-        .download_byte_range(&path, 0, Some(len as u64))
+        .download_byte_range(&path, 0, Some(len as u64), &cancel)
         .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // partial range (end specified)
-    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 4, Some(10), &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..10]);
 
     // partial range (end beyond real end)
     let dl = ctx
         .client
-        .download_byte_range(&path, 8, Some(len as u64 * 100))
+        .download_byte_range(&path, 8, Some(len as u64 * 100), &cancel)
         .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[8..]);
 
     // Partial range (end unspecified)
-    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 4, None, &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..]);
 
     // Full range (end unspecified)
-    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, None, &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
     ctx.client
-        .delete(&path)
+        .delete(&path, &cancel)
         .await
         .with_context(|| format!("{path:?} removal"))?;
 
@@ -265,6 +287,8 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
         return Ok(());
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(
         format!("{}/file_to_copy", ctx.base_prefix).as_str(),
     ))
@@ -278,18 +302,18 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
 
     let (data, len) = wrap_stream(orig.clone());
 
-    ctx.client.upload(data, len, &path, None).await?;
+    ctx.client.upload(data, len, &path, None, &cancel).await?;
 
     // Normal download request
-    ctx.client.copy_object(&path, &path_dest).await?;
+    ctx.client.copy_object(&path, &path_dest, &cancel).await?;
 
-    let dl = ctx.client.download(&path_dest).await?;
+    let dl = ctx.client.download(&path_dest, &cancel).await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
     ctx.client
-        .delete_objects(&[path.clone(), path_dest.clone()])
+        .delete_objects(&[path.clone(), path_dest.clone()], &cancel)
         .await
         .with_context(|| format!("{path:?} removal"))?;
 

libs/remote_storage/tests/test_real_azure.rs

@@ -1,9 +1,9 @@
-use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
+use std::{collections::HashSet, time::Duration};
 
 use anyhow::Context;
 use remote_storage::{
@@ -39,6 +39,17 @@ impl EnabledAzure {
             base_prefix: BASE_PREFIX,
         }
     }
+
+    #[allow(unused)] // this will be needed when moving the timeout integration tests back
+    fn configure_request_timeout(&mut self, timeout: Duration) {
+        match Arc::get_mut(&mut self.client).expect("outer Arc::get_mut") {
+            GenericRemoteStorage::AzureBlob(azure) => {
+                let azure = Arc::get_mut(azure).expect("inner Arc::get_mut");
+                azure.timeout = timeout;
+            }
+            _ => unreachable!(),
+        }
+    }
 }
 
 enum MaybeEnabledStorage {
@@ -213,6 +224,7 @@ fn create_azure_client(
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
         }),
+        timeout: Duration::from_secs(120),
     };
     Ok(Arc::new(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,

libs/remote_storage/tests/test_real_s3.rs

@@ -1,5 +1,6 @@
 use std::env;
 use std::fmt::{Debug, Display};
+use std::future::Future;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
 use std::sync::Arc;
@@ -9,9 +10,10 @@ use std::{collections::HashSet, time::SystemTime};
 use crate::common::{download_to_vec, upload_stream};
 use anyhow::Context;
 use camino::Utf8Path;
-use futures_util::Future;
+use futures_util::StreamExt;
 use remote_storage::{
-    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
+    DownloadError, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+    S3Config,
 };
 use test_context::test_context;
 use test_context::AsyncTestContext;
@@ -27,7 +29,6 @@ use common::{cleanup, ensure_logging_ready, upload_remote_data, upload_simple_re
 use utils::backoff;
 
 const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
-
 const BASE_PREFIX: &str = "test";
 
 #[test_context(MaybeEnabledStorage)]
@@ -69,8 +70,11 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
         ret
     }
 
-    async fn list_files(client: &Arc<GenericRemoteStorage>) -> anyhow::Result<HashSet<RemotePath>> {
-        Ok(retry(|| client.list_files(None, None))
+    async fn list_files(
+        client: &Arc<GenericRemoteStorage>,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<HashSet<RemotePath>> {
+        Ok(retry(|| client.list_files(None, None, cancel))
             .await
             .context("list root files failure")?
             .into_iter()
@@ -90,11 +94,11 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream("remote blob data1".as_bytes().into());
-        ctx.client.upload(data, len, &path1, None)
+        ctx.client.upload(data, len, &path1, None, &cancel)
     })
     .await?;
 
-    let t0_files = list_files(&ctx.client).await?;
+    let t0_files = list_files(&ctx.client, &cancel).await?;
     let t0 = time_point().await;
     println!("at t0: {t0_files:?}");
 
@@ -102,17 +106,17 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream(old_data.as_bytes().into());
-        ctx.client.upload(data, len, &path2, None)
+        ctx.client.upload(data, len, &path2, None, &cancel)
     })
     .await?;
 
-    let t1_files = list_files(&ctx.client).await?;
+    let t1_files = list_files(&ctx.client, &cancel).await?;
     let t1 = time_point().await;
     println!("at t1: {t1_files:?}");
 
     // A little check to ensure that our clock is not too far off from the S3 clock
     {
-        let dl = retry(|| ctx.client.download(&path2)).await?;
+        let dl = retry(|| ctx.client.download(&path2, &cancel)).await?;
         let last_modified = dl.last_modified.unwrap();
         let half_wt = WAIT_TIME.mul_f32(0.5);
         let t0_hwt = t0 + half_wt;
@@ -125,7 +129,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream("remote blob data3".as_bytes().into());
-        ctx.client.upload(data, len, &path3, None)
+        ctx.client.upload(data, len, &path3, None, &cancel)
     })
     .await?;
 
@@ -133,12 +137,12 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream(new_data.as_bytes().into());
-        ctx.client.upload(data, len, &path2, None)
+        ctx.client.upload(data, len, &path2, None, &cancel)
     })
     .await?;
 
-    retry(|| ctx.client.delete(&path1)).await?;
-    let t2_files = list_files(&ctx.client).await?;
+    retry(|| ctx.client.delete(&path1, &cancel)).await?;
+    let t2_files = list_files(&ctx.client, &cancel).await?;
     let t2 = time_point().await;
     println!("at t2: {t2_files:?}");
 
@@ -147,10 +151,10 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t2, t_final, &cancel)
         .await?;
-    let t2_files_recovered = list_files(&ctx.client).await?;
+    let t2_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t2: {t2_files_recovered:?}");
     assert_eq!(t2_files, t2_files_recovered);
-    let path2_recovered_t2 = download_to_vec(ctx.client.download(&path2).await?).await?;
+    let path2_recovered_t2 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
     assert_eq!(path2_recovered_t2, new_data.as_bytes());
 
     // after recovery to t1: path1 is back, path2 has the old content
@@ -158,10 +162,10 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t1, t_final, &cancel)
         .await?;
-    let t1_files_recovered = list_files(&ctx.client).await?;
+    let t1_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t1: {t1_files_recovered:?}");
     assert_eq!(t1_files, t1_files_recovered);
-    let path2_recovered_t1 = download_to_vec(ctx.client.download(&path2).await?).await?;
+    let path2_recovered_t1 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
     assert_eq!(path2_recovered_t1, old_data.as_bytes());
 
     // after recovery to t0: everything is gone except for path1
@@ -169,14 +173,14 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t0, t_final, &cancel)
         .await?;
-    let t0_files_recovered = list_files(&ctx.client).await?;
+    let t0_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t0: {t0_files_recovered:?}");
     assert_eq!(t0_files, t0_files_recovered);
 
     // cleanup
 
     let paths = &[path1, path2, path3];
-    retry(|| ctx.client.delete_objects(paths)).await?;
+    retry(|| ctx.client.delete_objects(paths, &cancel)).await?;
 
     Ok(())
 }
@@ -197,6 +201,16 @@ impl EnabledS3 {
             base_prefix: BASE_PREFIX,
         }
     }
+
+    fn configure_request_timeout(&mut self, timeout: Duration) {
+        match Arc::get_mut(&mut self.client).expect("outer Arc::get_mut") {
+            GenericRemoteStorage::AwsS3(s3) => {
+                let s3 = Arc::get_mut(s3).expect("inner Arc::get_mut");
+                s3.timeout = timeout;
+            }
+            _ => unreachable!(),
+        }
+    }
 }
 
 enum MaybeEnabledStorage {
@@ -370,8 +384,169 @@ fn create_s3_client(
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
         }),
+        timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
     };
     Ok(Arc::new(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
+
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn download_is_timeouted(ctx: &mut MaybeEnabledStorage) {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return;
+    };
+
+    let cancel = CancellationToken::new();
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/file_to_copy", ctx.base_prefix).as_str(),
+    ))
+    .unwrap();
+
+    let len = upload_large_enough_file(&ctx.client, &path, &cancel).await;
+
+    let timeout = std::time::Duration::from_secs(5);
+
+    ctx.configure_request_timeout(timeout);
+
+    let started_at = std::time::Instant::now();
+    let mut stream = ctx
+        .client
+        .download(&path, &cancel)
+        .await
+        .expect("download succeeds")
+        .download_stream;
+
+    if started_at.elapsed().mul_f32(0.9) >= timeout {
+        tracing::warn!(
+            elapsed_ms = started_at.elapsed().as_millis(),
+            "timeout might be too low, consumed most of it during headers"
+        );
+    }
+
+    let first = stream
+        .next()
+        .await
+        .expect("should have the first blob")
+        .expect("should have succeeded");
+
+    tracing::info!(len = first.len(), "downloaded first chunk");
+
+    assert!(
+        first.len() < len,
+        "uploaded file is too small, we downloaded all on first chunk"
+    );
+
+    tokio::time::sleep(timeout).await;
+
+    {
+        let started_at = std::time::Instant::now();
+        let next = stream
+            .next()
+            .await
+            .expect("stream should not have ended yet");
+
+        tracing::info!(
+            next.is_err = next.is_err(),
+            elapsed_ms = started_at.elapsed().as_millis(),
+            "received item after timeout"
+        );
+
+        let e = next.expect_err("expected an error, but got a chunk?");
+
+        let inner = e.get_ref().expect("std::io::Error::inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Timeout)),
+            "{inner:?}"
+        );
+    }
+
+    ctx.configure_request_timeout(RemoteStorageConfig::DEFAULT_TIMEOUT);
+
+    ctx.client.delete_objects(&[path], &cancel).await.unwrap()
+}
+
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn download_is_cancelled(ctx: &mut MaybeEnabledStorage) {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return;
+    };
+
+    let cancel = CancellationToken::new();
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/file_to_copy", ctx.base_prefix).as_str(),
+    ))
+    .unwrap();
+
+    let len = upload_large_enough_file(&ctx.client, &path, &cancel).await;
+
+    {
+        let mut stream = ctx
+            .client
+            .download(&path, &cancel)
+            .await
+            .expect("download succeeds")
+            .download_stream;
+
+        let first = stream
+            .next()
+            .await
+            .expect("should have the first blob")
+            .expect("should have succeeded");
+
+        tracing::info!(len = first.len(), "downloaded first chunk");
+
+        assert!(
+            first.len() < len,
+            "uploaded file is too small, we downloaded all on first chunk"
+        );
+
+        cancel.cancel();
+
+        let next = stream.next().await.expect("stream should have more");
+
+        let e = next.expect_err("expected an error, but got a chunk?");
+
+        let inner = e.get_ref().expect("std::io::Error::inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Cancelled)),
+            "{inner:?}"
+        );
+    }
+
+    let cancel = CancellationToken::new();
+
+    ctx.client.delete_objects(&[path], &cancel).await.unwrap();
+}
+
+/// Upload a long enough file so that we cannot download it in single chunk
+///
+/// For s3 the first chunk seems to be less than 10kB, so this has a bit of a safety margin
+async fn upload_large_enough_file(
+    client: &GenericRemoteStorage,
+    path: &RemotePath,
+    cancel: &CancellationToken,
+) -> usize {
+    let header = bytes::Bytes::from_static("remote blob data content".as_bytes());
+    let body = bytes::Bytes::from(vec![0u8; 1024]);
+    let contents = std::iter::once(header).chain(std::iter::repeat(body).take(128));
+
+    let len = contents.clone().fold(0, |acc, next| acc + next.len());
+
+    let contents = futures::stream::iter(contents.map(std::io::Result::Ok));
+
+    client
+        .upload(contents, len, path, None, cancel)
+        .await
+        .expect("upload succeeds");
+
+    len
+}

libs/utils/Cargo.toml

@@ -25,6 +25,7 @@ hyper = { workspace = true, features = ["full"] }
 fail.workspace = true
 futures = { workspace = true}
 jsonwebtoken.workspace = true
+leaky-bucket.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 pin-project-lite.workspace = true

libs/utils/benches/benchmarks.rs

@@ -1,5 +1,3 @@
-#![allow(unused)]
-
 use criterion::{criterion_group, criterion_main, Criterion};
 use utils::id;
 

libs/utils/src/auth.rs

@@ -29,6 +29,9 @@ pub enum Scope {
     // Should only be used e.g. for status check.
     // Currently also used for connection from any pageserver to any safekeeper.
     SafekeeperData,
+    // The scope used by pageservers in upcalls to storage controller and cloud control plane
+    #[serde(rename = "generations_api")]
+    GenerationsApi,
 }
 
 /// JWT payload. See docs/authentication.md for the format

pageserver/Cargo.toml

@@ -12,6 +12,7 @@ testing = ["fail/failpoints"]
 
 [dependencies]
 anyhow.workspace = true
+arc-swap.workspace = true
 async-compression.workspace = true
 async-stream.workspace = true
 async-trait.workspace = true
@@ -35,6 +36,7 @@ humantime.workspace = true
 humantime-serde.workspace = true
 hyper.workspace = true
 itertools.workspace = true
+leaky-bucket.workspace = true
 md5.workspace = true
 nix.workspace = true
 # hack to get the number of worker threads tokio uses
@@ -82,7 +84,7 @@ workspace_hack.workspace = true
 reqwest.workspace = true
 rpds.workspace = true
 enum-map.workspace = true
-enumset.workspace = true
+enumset = { workspace = true, features = ["serde"]}
 strum.workspace = true
 strum_macros.workspace = true
 

pageserver/benches/bench_walredo.rs

@@ -6,14 +6,28 @@
 //! There are two sets of inputs; `short` and `medium`. They were collected on postgres v14 by
 //! logging what happens when a sequential scan is requested on a small table, then picking out two
 //! suitable from logs.
+//!
+//!
+//! Reference data (git blame to see commit) on an i3en.3xlarge
+// ```text
+//! short/short/1           time:   [39.175 µs 39.348 µs 39.536 µs]
+//! short/short/2           time:   [51.227 µs 51.487 µs 51.755 µs]
+//! short/short/4           time:   [76.048 µs 76.362 µs 76.674 µs]
+//! short/short/8           time:   [128.94 µs 129.82 µs 130.74 µs]
+//! short/short/16          time:   [227.84 µs 229.00 µs 230.28 µs]
+//! short/short/32          time:   [455.97 µs 457.81 µs 459.90 µs]
+//! short/short/64          time:   [902.46 µs 904.84 µs 907.32 µs]
+//! short/short/128         time:   [1.7416 ms 1.7487 ms 1.7561 ms]
+//! ``
 
-use std::sync::{Arc, Barrier};
+use std::sync::Arc;
 
 use bytes::{Buf, Bytes};
 use pageserver::{
     config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
 };
 use pageserver_api::shard::TenantShardId;
+use tokio::task::JoinSet;
 use utils::{id::TenantId, lsn::Lsn};
 
 use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
@@ -39,11 +53,11 @@ fn redo_scenarios(c: &mut Criterion) {
             .build()
             .unwrap();
         tracing::info!("executing first");
-        short().execute(rt.handle(), &manager).unwrap();
+        rt.block_on(short().execute(&manager)).unwrap();
         tracing::info!("first executed");
     }
 
-    let thread_counts = [1, 2, 4, 8, 16];
+    let thread_counts = [1, 2, 4, 8, 16, 32, 64, 128];
 
     let mut group = c.benchmark_group("short");
     group.sampling_mode(criterion::SamplingMode::Flat);
@@ -74,114 +88,69 @@ fn redo_scenarios(c: &mut Criterion) {
     drop(group);
 }
 
-/// Sets up `threads` number of requesters to `request_redo`, with the given input.
+/// Sets up a multi-threaded tokio runtime with default worker thread count,
+/// then, spawn `requesters` tasks that repeatedly:
+/// - get input from `input_factor()`
+/// - call `manager.request_redo()` with their input
+///
+/// This stress-tests the scalability of a single walredo manager at high tokio-level concurrency.
+///
+/// Using tokio's default worker thread count means the results will differ on machines
+/// with different core countrs. We don't care about that, the performance will always
+/// be different on different hardware. To compare performance of different software versions,
+/// use the same hardware.
 fn add_multithreaded_walredo_requesters(
     b: &mut criterion::Bencher,
-    threads: u32,
+    nrequesters: usize,
     manager: &Arc<PostgresRedoManager>,
     input_factory: fn() -> Request,
 ) {
-    assert_ne!(threads, 0);
+    assert_ne!(nrequesters, 0);
 
-    if threads == 1 {
-        let rt = tokio::runtime::Builder::new_current_thread()
-            .enable_all()
-            .build()
-            .unwrap();
-        let handle = rt.handle();
-        b.iter_batched_ref(
-            || Some(input_factory()),
-            |input| execute_all(input.take(), handle, manager),
-            criterion::BatchSize::PerIteration,
-        );
-    } else {
-        let (work_tx, work_rx) = std::sync::mpsc::sync_channel(threads as usize);
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()
+        .unwrap();
 
-        let work_rx = std::sync::Arc::new(std::sync::Mutex::new(work_rx));
+    let barrier = Arc::new(tokio::sync::Barrier::new(nrequesters + 1));
 
-        let barrier = Arc::new(Barrier::new(threads as usize + 1));
-
-        let jhs = (0..threads)
-            .map(|_| {
-                std::thread::spawn({
-                    let manager = manager.clone();
-                    let barrier = barrier.clone();
-                    let work_rx = work_rx.clone();
-                    move || {
-                        let rt = tokio::runtime::Builder::new_current_thread()
-                            .enable_all()
-                            .build()
-                            .unwrap();
-                        let handle = rt.handle();
-                        loop {
-                            // queue up and wait if we want to go another round
-                            if work_rx.lock().unwrap().recv().is_err() {
-                                break;
-                            }
-
-                            let input = Some(input_factory());
-
-                            barrier.wait();
-
-                            execute_all(input, handle, &manager).unwrap();
-
-                            barrier.wait();
-                        }
-                    }
-                })
-            })
-            .collect::<Vec<_>>();
-
-        let _jhs = JoinOnDrop(jhs);
-
-        b.iter_batched(
-            || {
-                for _ in 0..threads {
-                    work_tx.send(()).unwrap()
-                }
-            },
-            |()| {
-                // start the work
-                barrier.wait();
-
-                // wait for work to complete
-                barrier.wait();
-            },
-            criterion::BatchSize::PerIteration,
-        );
-
-        drop(work_tx);
+    let mut requesters = JoinSet::new();
+    for _ in 0..nrequesters {
+        let _entered = rt.enter();
+        let manager = manager.clone();
+        let barrier = barrier.clone();
+        requesters.spawn(async move {
+            loop {
+                let input = input_factory();
+                barrier.wait().await;
+                let page = input.execute(&manager).await.unwrap();
+                assert_eq!(page.remaining(), 8192);
+                barrier.wait().await;
+            }
+        });
     }
-}
 
-struct JoinOnDrop(Vec<std::thread::JoinHandle<()>>);
+    let do_one_iteration = || {
+        rt.block_on(async {
+            barrier.wait().await;
+            // wait for work to complete
+            barrier.wait().await;
+        })
+    };
 
-impl Drop for JoinOnDrop {
-    // it's not really needless because we want join all then check for panicks
-    #[allow(clippy::needless_collect)]
-    fn drop(&mut self) {
-        // first join all
-        let results = self.0.drain(..).map(|jh| jh.join()).collect::<Vec<_>>();
-        // then check the results; panicking here is not great, but it does get the message across
-        // to the user, and sets an exit value.
-        results.into_iter().try_for_each(|res| res).unwrap();
-    }
-}
+    b.iter_batched(
+        || {
+            // warmup
+            do_one_iteration();
+        },
+        |()| {
+            // work loop
+            do_one_iteration();
+        },
+        criterion::BatchSize::PerIteration,
+    );
 
-fn execute_all<I>(
-    input: I,
-    handle: &tokio::runtime::Handle,
-    manager: &PostgresRedoManager,
-) -> anyhow::Result<()>
-where
-    I: IntoIterator<Item = Request>,
-{
-    // just fire all requests as fast as possible
-    input.into_iter().try_for_each(|req| {
-        let page = req.execute(handle, manager)?;
-        assert_eq!(page.remaining(), 8192);
-        anyhow::Ok(())
-    })
+    rt.block_on(requesters.shutdown());
 }
 
 criterion_group!(benches, redo_scenarios);
@@ -493,11 +462,7 @@ struct Request {
 }
 
 impl Request {
-    fn execute(
-        self,
-        rt: &tokio::runtime::Handle,
-        manager: &PostgresRedoManager,
-    ) -> anyhow::Result<Bytes> {
+    async fn execute(self, manager: &PostgresRedoManager) -> anyhow::Result<Bytes> {
         let Request {
             key,
             lsn,
@@ -506,6 +471,8 @@ impl Request {
             pg_version,
         } = self;
 
-        rt.block_on(manager.request_redo(key, lsn, base_img, records, pg_version))
+        manager
+            .request_redo(key, lsn, base_img, records, pg_version)
+            .await
     }
 }

pageserver/client/src/mgmt_api.rs

@@ -217,6 +217,20 @@ impl Client {
         }
     }
 
+    pub async fn tenant_time_travel_remote_storage(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timestamp: &str,
+        done_if_after: &str,
+    ) -> Result<()> {
+        let uri = format!(
+            "{}/v1/tenant/{tenant_shard_id}/time_travel_remote_storage?travel_to={timestamp}&done_if_after={done_if_after}",
+            self.mgmt_api_endpoint
+        );
+        self.request(Method::PUT, &uri, ()).await?;
+        Ok(())
+    }
+
     pub async fn tenant_config(&self, req: &TenantConfigRequest) -> Result<()> {
         let uri = format!("{}/v1/tenant/config", self.mgmt_api_endpoint);
         self.request(Method::PUT, &uri, req).await?;

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -1,6 +1,5 @@
 use anyhow::Context;
 use camino::Utf8PathBuf;
-use futures::future::join_all;
 use pageserver_api::key::{is_rel_block_key, key_to_rel_block, Key};
 use pageserver_api::keyspace::KeySpaceAccum;
 use pageserver_api::models::PagestreamGetPageRequest;
@@ -10,11 +9,10 @@ use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
 
 use rand::prelude::*;
-use tokio::sync::Barrier;
 use tokio::task::JoinSet;
-use tracing::{info, instrument};
+use tracing::info;
 
-use std::collections::{HashMap, HashSet};
+use std::collections::HashSet;
 use std::future::Future;
 use std::num::NonZeroUsize;
 use std::pin::Pin;
@@ -38,8 +36,12 @@ pub(crate) struct Args {
     num_clients: NonZeroUsize,
     #[clap(long)]
     runtime: Option<humantime::Duration>,
+    /// Each client sends requests at the given rate.
+    ///
+    /// If a request takes too long and we should be issuing a new request already,
+    /// we skip that request and account it as `MISSED`.
     #[clap(long)]
-    per_target_rate_limit: Option<usize>,
+    per_client_rate: Option<usize>,
     /// Probability for sending `latest=true` in the request (uniform distribution).
     #[clap(long, default_value = "1")]
     req_latest_probability: f64,
@@ -61,12 +63,16 @@ pub(crate) struct Args {
 #[derive(Debug, Default)]
 struct LiveStats {
     completed_requests: AtomicU64,
+    missed: AtomicU64,
 }
 
 impl LiveStats {
-    fn inc(&self) {
+    fn request_done(&self) {
         self.completed_requests.fetch_add(1, Ordering::Relaxed);
     }
+    fn missed(&self, n: u64) {
+        self.missed.fetch_add(n, Ordering::Relaxed);
+    }
 }
 
 #[derive(Clone, serde::Serialize, serde::Deserialize)]
@@ -220,13 +226,12 @@ async fn main_impl(
 
     let live_stats = Arc::new(LiveStats::default());
 
-    let num_client_tasks = args.num_clients.get() * timelines.len();
     let num_live_stats_dump = 1;
-    let num_work_sender_tasks = 1;
+    let num_work_sender_tasks = args.num_clients.get() * timelines.len();
     let num_main_impl = 1;
 
     let start_work_barrier = Arc::new(tokio::sync::Barrier::new(
-        num_client_tasks + num_live_stats_dump + num_work_sender_tasks + num_main_impl,
+        num_live_stats_dump + num_work_sender_tasks + num_main_impl,
     ));
 
     tokio::spawn({
@@ -238,10 +243,12 @@ async fn main_impl(
                 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 missed = stats.missed.swap(0, Ordering::Relaxed);
                 let elapsed = start.elapsed();
                 info!(
-                    "RPS: {:.0}",
-                    completed_requests as f64 / elapsed.as_secs_f64()
+                    "RPS: {:.0}   MISSED: {:.0}",
+                    completed_requests as f64 / elapsed.as_secs_f64(),
+                    missed as f64 / elapsed.as_secs_f64()
                 );
             }
         }
@@ -249,127 +256,105 @@ async fn main_impl(
 
     let cancel = CancellationToken::new();
 
-    let mut work_senders: HashMap<WorkerId, _> = HashMap::new();
-    let mut tasks = Vec::new();
+    let rps_period = args
+        .per_client_rate
+        .map(|rps_limit| Duration::from_secs_f64(1.0 / (rps_limit as f64)));
+    let make_worker: &dyn Fn(WorkerId) -> Pin<Box<dyn Send + Future<Output = ()>>> = &|worker_id| {
+        let live_stats = live_stats.clone();
+        let start_work_barrier = start_work_barrier.clone();
+        let ranges: Vec<KeyRange> = all_ranges
+            .iter()
+            .filter(|r| r.timeline == worker_id.timeline)
+            .cloned()
+            .collect();
+        let weights =
+            rand::distributions::weighted::WeightedIndex::new(ranges.iter().map(|v| v.len()))
+                .unwrap();
+
+        let cancel = cancel.clone();
+        Box::pin(async move {
+            let client =
+                pageserver_client::page_service::Client::new(args.page_service_connstring.clone())
+                    .await
+                    .unwrap();
+            let mut client = client
+                .pagestream(worker_id.timeline.tenant_id, worker_id.timeline.timeline_id)
+                .await
+                .unwrap();
+
+            start_work_barrier.wait().await;
+            let client_start = Instant::now();
+            let mut ticks_processed = 0;
+            while !cancel.is_cancelled() {
+                // Detect if a request took longer than the RPS rate
+                if let Some(period) = &rps_period {
+                    let periods_passed_until_now =
+                        usize::try_from(client_start.elapsed().as_micros() / period.as_micros())
+                            .unwrap();
+
+                    if periods_passed_until_now > ticks_processed {
+                        live_stats.missed((periods_passed_until_now - ticks_processed) as u64);
+                    }
+                    ticks_processed = periods_passed_until_now;
+                }
+
+                let start = Instant::now();
+                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 = Key::from_i128(key);
+                    assert!(is_rel_block_key(&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,
+                    }
+                };
+                client.getpage(req).await.unwrap();
+                let end = Instant::now();
+                live_stats.request_done();
+                ticks_processed += 1;
+                STATS.with(|stats| {
+                    stats
+                        .borrow()
+                        .lock()
+                        .unwrap()
+                        .observe(end.duration_since(start))
+                        .unwrap();
+                });
+
+                if let Some(period) = &rps_period {
+                    let next_at = client_start
+                        + Duration::from_micros(
+                            (ticks_processed) as u64 * u64::try_from(period.as_micros()).unwrap(),
+                        );
+                    tokio::time::sleep_until(next_at.into()).await;
+                }
+            }
+        })
+    };
+
+    info!("spawning workers");
+    let mut workers = JoinSet::new();
     for timeline in timelines.iter().cloned() {
         for num_client in 0..args.num_clients.get() {
-            let (sender, receiver) = tokio::sync::mpsc::channel(10); // TODO: not sure what the implications of this are
             let worker_id = WorkerId {
                 timeline,
                 num_client,
             };
-            work_senders.insert(worker_id, sender);
-            tasks.push(tokio::spawn(client(
-                args,
-                worker_id,
-                Arc::clone(&start_work_barrier),
-                receiver,
-                Arc::clone(&live_stats),
-                cancel.clone(),
-            )));
+            workers.spawn(make_worker(worker_id));
         }
     }
-
-    let work_sender: Pin<Box<dyn Send + Future<Output = ()>>> = {
-        let start_work_barrier = start_work_barrier.clone();
-        let cancel = cancel.clone();
-        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;
-
-                while !cancel.is_cancelled() {
-                    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 = Key::from_i128(key);
-                        let (rel_tag, block_no) =
-                            key_to_rel_block(key).expect("we filter non-rel-block keys out above");
-                        (
-                            WorkerId {
-                                timeline: r.timeline,
-                                num_client: rng.gen_range(0..args.num_clients.get()),
-                            },
-                            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?
-                    if sender.send(req).await.is_err() {
-                        assert!(cancel.is_cancelled(), "client has gone away unexpectedly");
-                    }
-                }
-            }),
-            Some(rps_limit) => Box::pin(async move {
-                let period = Duration::from_secs_f64(1.0 / (rps_limit as f64));
-                let make_task: &dyn Fn(WorkerId) -> Pin<Box<dyn Send + Future<Output = ()>>> =
-                    &|worker_id| {
-                        let sender = work_senders.get(&worker_id).unwrap();
-                        let ranges: Vec<KeyRange> = all_ranges
-                            .iter()
-                            .filter(|r| r.timeline == worker_id.timeline)
-                            .cloned()
-                            .collect();
-                        let weights = rand::distributions::weighted::WeightedIndex::new(
-                            ranges.iter().map(|v| v.len()),
-                        )
-                        .unwrap();
-
-                        let cancel = cancel.clone();
-                        Box::pin(async move {
-                            let mut ticker = tokio::time::interval(period);
-                            ticker.set_missed_tick_behavior(
-                                /* TODO review this choice */
-                                tokio::time::MissedTickBehavior::Burst,
-                            );
-                            while !cancel.is_cancelled() {
-                                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 = Key::from_i128(key);
-                                    assert!(is_rel_block_key(&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,
-                                    }
-                                };
-                                if sender.send(req).await.is_err() {
-                                    assert!(
-                                        cancel.is_cancelled(),
-                                        "client has gone away unexpectedly"
-                                    );
-                                }
-                            }
-                        })
-                    };
-
-                let tasks: Vec<_> = work_senders.keys().map(|tl| make_task(*tl)).collect();
-
-                start_work_barrier.wait().await;
-
-                join_all(tasks).await;
-            }),
+    let workers = async move {
+        while let Some(res) = workers.join_next().await {
+            res.unwrap();
         }
     };
 
-    let work_sender_task = tokio::spawn(work_sender);
-
     info!("waiting for everything to become ready");
     start_work_barrier.wait().await;
     info!("work started");
@@ -377,20 +362,13 @@ async fn main_impl(
         tokio::time::sleep(runtime.into()).await;
         info!("runtime over, signalling cancellation");
         cancel.cancel();
-        work_sender_task.await.unwrap();
+        workers.await;
         info!("work sender exited");
     } else {
-        work_sender_task.await.unwrap();
+        workers.await;
         unreachable!("work sender never terminates");
     }
 
-    info!("joining clients");
-    for t in tasks {
-        t.await.unwrap();
-    }
-
-    info!("all clients stopped");
-
     let output = Output {
         total: {
             let mut agg_stats = request_stats::Stats::new();
@@ -407,49 +385,3 @@ async fn main_impl(
 
     anyhow::Ok(())
 }
-
-#[instrument(skip_all)]
-async fn client(
-    args: &'static Args,
-    id: WorkerId,
-    start_work_barrier: Arc<Barrier>,
-    mut work: tokio::sync::mpsc::Receiver<PagestreamGetPageRequest>,
-    live_stats: Arc<LiveStats>,
-    cancel: CancellationToken,
-) {
-    let WorkerId {
-        timeline,
-        num_client: _,
-    } = id;
-    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();
-
-    let do_requests = async {
-        start_work_barrier.wait().await;
-        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();
-            });
-        }
-    };
-    tokio::select! {
-        res = do_requests => { res },
-        _ = cancel.cancelled() => {
-            // fallthrough to shutdown
-        }
-    }
-    client.shutdown().await;
-}

pageserver/src/auth.rs

@@ -14,8 +14,12 @@ pub fn check_permission(claims: &Claims, tenant_id: Option<TenantId>) -> Result<
         }
         (Scope::PageServerApi, None) => Ok(()), // access to management api for PageServerApi scope
         (Scope::PageServerApi, Some(_)) => Ok(()), // access to tenant api using PageServerApi scope
-        (Scope::SafekeeperData, _) => Err(AuthError(
-            "SafekeeperData scope makes no sense for Pageserver".into(),
+        (Scope::SafekeeperData | Scope::GenerationsApi, _) => Err(AuthError(
+            format!(
+                "JWT scope '{:?}' is ineligible for Pageserver auth",
+                claims.scope
+            )
+            .into(),
         )),
     }
 }

pageserver/src/basebackup.rs

@@ -261,10 +261,7 @@ where
             let mut slru_builder = SlruSegmentsBuilder::new(&mut self.ar);
 
             for part in slru_partitions.parts {
-                let blocks = self
-                    .timeline
-                    .get_vectored(&part.ranges, self.lsn, self.ctx)
-                    .await?;
+                let blocks = self.timeline.get_vectored(part, self.lsn, self.ctx).await?;
 
                 for (key, block) in blocks {
                     slru_builder.add_block(&key, block?).await?;

pageserver/src/config.rs

@@ -33,12 +33,13 @@ use utils::{
 use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
+use crate::tenant::timeline::GetVectoredImpl;
 use crate::tenant::{
     TENANTS_SEGMENT_NAME, TENANT_DELETED_MARKER_FILE_NAME, TIMELINES_SEGMENT_NAME,
 };
 use crate::virtual_file;
 use crate::{
-    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TENANT_HEATMAP_BASENAME,
+    IGNORED_TENANT_FILE_NAME, TENANT_CONFIG_NAME, TENANT_HEATMAP_BASENAME,
     TENANT_LOCATION_CONFIG_NAME, TIMELINE_DELETE_MARK_SUFFIX, TIMELINE_UNINIT_MARK_SUFFIX,
 };
 
@@ -84,6 +85,8 @@ pub mod defaults {
 
     pub const DEFAULT_VIRTUAL_FILE_IO_ENGINE: &str = "std-fs";
 
+    pub const DEFAULT_GET_VECTORED_IMPL: &str = "sequential";
+
     ///
     /// Default built-in configuration file.
     ///
@@ -121,6 +124,8 @@ pub mod defaults {
 
 #virtual_file_io_engine = '{DEFAULT_VIRTUAL_FILE_IO_ENGINE}'
 
+#get_vectored_impl = '{DEFAULT_GET_VECTORED_IMPL}'
+
 [tenant_config]
 #checkpoint_distance = {DEFAULT_CHECKPOINT_DISTANCE} # in bytes
 #checkpoint_timeout = {DEFAULT_CHECKPOINT_TIMEOUT}
@@ -256,6 +261,8 @@ pub struct PageServerConf {
     pub ingest_batch_size: u64,
 
     pub virtual_file_io_engine: virtual_file::IoEngineKind,
+
+    pub get_vectored_impl: GetVectoredImpl,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -342,6 +349,8 @@ struct PageServerConfigBuilder {
     ingest_batch_size: BuilderValue<u64>,
 
     virtual_file_io_engine: BuilderValue<virtual_file::IoEngineKind>,
+
+    get_vectored_impl: BuilderValue<GetVectoredImpl>,
 }
 
 impl Default for PageServerConfigBuilder {
@@ -419,6 +428,8 @@ impl Default for PageServerConfigBuilder {
             ingest_batch_size: Set(DEFAULT_INGEST_BATCH_SIZE),
 
             virtual_file_io_engine: Set(DEFAULT_VIRTUAL_FILE_IO_ENGINE.parse().unwrap()),
+
+            get_vectored_impl: Set(DEFAULT_GET_VECTORED_IMPL.parse().unwrap()),
         }
     }
 }
@@ -579,6 +590,10 @@ impl PageServerConfigBuilder {
         self.virtual_file_io_engine = BuilderValue::Set(value);
     }
 
+    pub fn get_vectored_impl(&mut self, value: GetVectoredImpl) {
+        self.get_vectored_impl = BuilderValue::Set(value);
+    }
+
     pub fn build(self) -> anyhow::Result<PageServerConf> {
         let concurrent_tenant_warmup = self
             .concurrent_tenant_warmup
@@ -689,6 +704,9 @@ impl PageServerConfigBuilder {
             virtual_file_io_engine: self
                 .virtual_file_io_engine
                 .ok_or(anyhow!("missing virtual_file_io_engine"))?,
+            get_vectored_impl: self
+                .get_vectored_impl
+                .ok_or(anyhow!("missing get_vectored_impl"))?,
         })
     }
 }
@@ -808,17 +826,6 @@ impl PageServerConf {
             .join(connection_id.to_string())
     }
 
-    /// Points to a place in pageserver's local directory,
-    /// where certain timeline's metadata file should be located.
-    pub fn metadata_path(
-        &self,
-        tenant_shard_id: &TenantShardId,
-        timeline_id: &TimelineId,
-    ) -> Utf8PathBuf {
-        self.timeline_path(tenant_shard_id, timeline_id)
-            .join(METADATA_FILE_NAME)
-    }
-
     /// Turns storage remote path of a file into its local path.
     pub fn local_path(&self, remote_path: &RemotePath) -> Utf8PathBuf {
         remote_path.with_base(&self.workdir)
@@ -943,6 +950,9 @@ impl PageServerConf {
                 "virtual_file_io_engine" => {
                     builder.virtual_file_io_engine(parse_toml_from_str("virtual_file_io_engine", item)?)
                 }
+                "get_vectored_impl" => {
+                    builder.get_vectored_impl(parse_toml_from_str("get_vectored_impl", item)?)
+                }
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -1017,6 +1027,7 @@ impl PageServerConf {
             secondary_download_concurrency: defaults::DEFAULT_SECONDARY_DOWNLOAD_CONCURRENCY,
             ingest_batch_size: defaults::DEFAULT_INGEST_BATCH_SIZE,
             virtual_file_io_engine: DEFAULT_VIRTUAL_FILE_IO_ENGINE.parse().unwrap(),
+            get_vectored_impl: defaults::DEFAULT_GET_VECTORED_IMPL.parse().unwrap(),
         }
     }
 }
@@ -1250,6 +1261,7 @@ background_task_maximum_delay = '334 s'
                 secondary_download_concurrency: defaults::DEFAULT_SECONDARY_DOWNLOAD_CONCURRENCY,
                 ingest_batch_size: defaults::DEFAULT_INGEST_BATCH_SIZE,
                 virtual_file_io_engine: DEFAULT_VIRTUAL_FILE_IO_ENGINE.parse().unwrap(),
+                get_vectored_impl: defaults::DEFAULT_GET_VECTORED_IMPL.parse().unwrap(),
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1314,6 +1326,7 @@ background_task_maximum_delay = '334 s'
                 secondary_download_concurrency: defaults::DEFAULT_SECONDARY_DOWNLOAD_CONCURRENCY,
                 ingest_batch_size: 100,
                 virtual_file_io_engine: DEFAULT_VIRTUAL_FILE_IO_ENGINE.parse().unwrap(),
+                get_vectored_impl: defaults::DEFAULT_GET_VECTORED_IMPL.parse().unwrap(),
             },
             "Should be able to parse all basic config values correctly"
         );
@@ -1359,6 +1372,7 @@ broker_endpoint = '{broker_endpoint}'
                 parsed_remote_storage_config,
                 RemoteStorageConfig {
                     storage: RemoteStorageKind::LocalFs(local_storage_path.clone()),
+                    timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
                 },
                 "Remote storage config should correctly parse the local FS config and fill other storage defaults"
             );
@@ -1426,6 +1440,7 @@ broker_endpoint = '{broker_endpoint}'
                         concurrency_limit: s3_concurrency_limit,
                         max_keys_per_list_response: None,
                     }),
+                    timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
                 },
                 "Remote storage config should correctly parse the S3 config"
             );
@@ -1546,17 +1561,50 @@ threshold = "20m"
                 eviction_order: crate::disk_usage_eviction_task::EvictionOrder::AbsoluteAccessed,
             })
         );
+
         match &conf.default_tenant_conf.eviction_policy {
-            EvictionPolicy::NoEviction => panic!("Unexpected eviction opolicy tenant settings"),
-            EvictionPolicy::LayerAccessThreshold(eviction_thresold) => {
-                assert_eq!(eviction_thresold.period, Duration::from_secs(20 * 60));
-                assert_eq!(eviction_thresold.threshold, Duration::from_secs(20 * 60));
+            EvictionPolicy::LayerAccessThreshold(eviction_threshold) => {
+                assert_eq!(eviction_threshold.period, Duration::from_secs(20 * 60));
+                assert_eq!(eviction_threshold.threshold, Duration::from_secs(20 * 60));
             }
+            other => unreachable!("Unexpected eviction policy tenant settings: {other:?}"),
         }
 
         Ok(())
     }
 
+    #[test]
+    fn parse_imitation_only_pageserver_config() {
+        let tempdir = tempdir().unwrap();
+        let (workdir, pg_distrib_dir) = prepare_fs(&tempdir).unwrap();
+
+        let pageserver_conf_toml = format!(
+            r#"pg_distrib_dir = "{pg_distrib_dir}"
+metric_collection_endpoint = "http://sample.url"
+metric_collection_interval = "10min"
+id = 222
+
+[tenant_config]
+evictions_low_residence_duration_metric_threshold = "20m"
+
+[tenant_config.eviction_policy]
+kind = "OnlyImitiate"
+period = "20m"
+threshold = "20m"
+"#,
+        );
+        let toml: Document = pageserver_conf_toml.parse().unwrap();
+        let conf = PageServerConf::parse_and_validate(&toml, &workdir).unwrap();
+
+        match &conf.default_tenant_conf.eviction_policy {
+            EvictionPolicy::OnlyImitiate(t) => {
+                assert_eq!(t.period, Duration::from_secs(20 * 60));
+                assert_eq!(t.threshold, Duration::from_secs(20 * 60));
+            }
+            other => unreachable!("Unexpected eviction policy tenant settings: {other:?}"),
+        }
+    }
+
     fn prepare_fs(tempdir: &Utf8TempDir) -> anyhow::Result<(Utf8PathBuf, Utf8PathBuf)> {
         let tempdir_path = tempdir.path();
 

pageserver/src/deletion_queue.rs

@@ -834,7 +834,6 @@ mod test {
     }
 
     impl ControlPlaneGenerationsApi for MockControlPlane {
-        #[allow(clippy::diverging_sub_expression)] // False positive via async_trait
         async fn re_attach(&self) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
             unimplemented!()
         }
@@ -868,6 +867,7 @@ mod test {
         let remote_fs_dir = harness.conf.workdir.join("remote_fs").canonicalize_utf8()?;
         let storage_config = RemoteStorageConfig {
             storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
+            timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
         };
         let storage = GenericRemoteStorage::from_config(&storage_config).unwrap();
 
@@ -1171,6 +1171,7 @@ pub(crate) mod mock {
     pub struct ConsumerState {
         rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
         executor_rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
+        cancel: CancellationToken,
     }
 
     impl ConsumerState {
@@ -1184,7 +1185,7 @@ pub(crate) mod mock {
                 match msg {
                     DeleterMessage::Delete(objects) => {
                         for path in objects {
-                            match remote_storage.delete(&path).await {
+                            match remote_storage.delete(&path, &self.cancel).await {
                                 Ok(_) => {
                                     debug!("Deleted {path}");
                                 }
@@ -1217,7 +1218,7 @@ pub(crate) mod mock {
 
                         for path in objects {
                             info!("Executing deletion {path}");
-                            match remote_storage.delete(&path).await {
+                            match remote_storage.delete(&path, &self.cancel).await {
                                 Ok(_) => {
                                     debug!("Deleted {path}");
                                 }
@@ -1267,7 +1268,11 @@ pub(crate) mod mock {
                 executor_tx,
                 executed,
                 remote_storage,
-                consumer: std::sync::Mutex::new(ConsumerState { rx, executor_rx }),
+                consumer: std::sync::Mutex::new(ConsumerState {
+                    rx,
+                    executor_rx,
+                    cancel: CancellationToken::new(),
+                }),
                 lsn_table: Arc::new(std::sync::RwLock::new(VisibleLsnUpdates::new())),
             }
         }

pageserver/src/deletion_queue/deleter.rs

@@ -8,6 +8,7 @@
 
 use remote_storage::GenericRemoteStorage;
 use remote_storage::RemotePath;
+use remote_storage::TimeoutOrCancel;
 use remote_storage::MAX_KEYS_PER_DELETE;
 use std::time::Duration;
 use tokio_util::sync::CancellationToken;
@@ -71,9 +72,11 @@ impl Deleter {
                     Err(anyhow::anyhow!("failpoint: deletion-queue-before-execute"))
                 });
 
-                self.remote_storage.delete_objects(&self.accumulator).await
+                self.remote_storage
+                    .delete_objects(&self.accumulator, &self.cancel)
+                    .await
             },
-            |_| false,
+            TimeoutOrCancel::caused_by_cancel,
             3,
             10,
             "executing deletion batch",

pageserver/src/disk_usage_eviction_task.rs

@@ -351,7 +351,6 @@ pub enum IterationOutcome<U> {
     Finished(IterationOutcomeFinished<U>),
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Serialize)]
 pub struct IterationOutcomeFinished<U> {
     /// The actual usage observed before we started the iteration.
@@ -366,7 +365,6 @@ pub struct IterationOutcomeFinished<U> {
 }
 
 #[derive(Debug, Serialize)]
-#[allow(dead_code)]
 struct AssumedUsage<U> {
     /// The expected value for `after`, after phase 2.
     projected_after: U,
@@ -374,14 +372,12 @@ struct AssumedUsage<U> {
     failed: LayerCount,
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Serialize)]
 struct PlannedUsage<U> {
     respecting_tenant_min_resident_size: U,
     fallback_to_global_lru: Option<U>,
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Default, Serialize)]
 struct LayerCount {
     file_sizes: u64,
@@ -565,7 +561,6 @@ pub(crate) struct EvictionSecondaryLayer {
 #[derive(Clone)]
 pub(crate) enum EvictionLayer {
     Attached(Layer),
-    #[allow(dead_code)]
     Secondary(EvictionSecondaryLayer),
 }
 
@@ -1105,7 +1100,6 @@ mod filesystem_level_usage {
     use super::DiskUsageEvictionTaskConfig;
 
     #[derive(Debug, Clone, Copy)]
-    #[allow(dead_code)]
     pub struct Usage<'a> {
         config: &'a DiskUsageEvictionTaskConfig,
 

pageserver/src/http/openapi_spec.yml

@@ -1379,6 +1379,25 @@ paths:
               schema:
                 $ref: "#/components/schemas/ServiceUnavailableError"
 
+  /v1/utilization:
+    get:
+      description: |
+        Returns the pageservers current utilization and fitness score for new tenants.
+
+      responses:
+        "200":
+            description: Pageserver utilization and fitness score
+            content:
+              application/json:
+                schema:
+                  $ref: "#/components/schemas/PageserverUtilization"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+
 components:
   securitySchemes:
     JWT:
@@ -1691,6 +1710,33 @@ components:
           type: string
           enum: [past, present, future, nodata]
 
+    PageserverUtilization:
+      type: object
+      required:
+        - disk_usage_bytes
+        - free_space_bytes
+        - utilization_score
+      properties:
+        disk_usage_bytes:
+          type: integer
+          format: int64
+          minimum: 0
+          description: The amount of disk space currently utilized by layer files.
+        free_space_bytes:
+          type: integer
+          format: int64
+          minimum: 0
+          description: The amount of usable disk space left.
+        utilization_score:
+          type: integer
+          format: int64
+          minimum: 0
+          maximum: 9223372036854775807
+          default: 9223372036854775807
+          description: |
+            Lower is better score for how good this pageserver would be for the next tenant.
+            The default or maximum value can be returned in situations when a proper score cannot (yet) be calculated.
+
     Error:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -83,12 +83,12 @@ use utils::{
 // This is not functionally necessary (clients will retry), but avoids generating a lot of
 // failed API calls while tenants are activating.
 #[cfg(not(feature = "testing"))]
-const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
+pub(crate) const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
 
 // Tests run on slow/oversubscribed nodes, and may need to wait much longer for tenants to
 // finish attaching, if calls to remote storage are slow.
 #[cfg(feature = "testing")]
-const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
+pub(crate) const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
 
 pub struct State {
     conf: &'static PageServerConf,
@@ -100,6 +100,7 @@ pub struct State {
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     deletion_queue_client: DeletionQueueClient,
     secondary_controller: SecondaryController,
+    latest_utilization: tokio::sync::Mutex<Option<(std::time::Instant, bytes::Bytes)>>,
 }
 
 impl State {
@@ -128,6 +129,7 @@ impl State {
             disk_usage_eviction_state,
             deletion_queue_client,
             secondary_controller,
+            latest_utilization: Default::default(),
         })
     }
 }
@@ -571,10 +573,16 @@ async fn timeline_list_handler(
         parse_query_param(&request, "force-await-initial-logical-size")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
+    let state = get_state(&request);
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
 
     let response_data = async {
-        let tenant = mgr::get_tenant(tenant_shard_id, true)?;
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id, false)?;
+
+        tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
         let timelines = tenant.list_timelines();
 
         let mut response_data = Vec::with_capacity(timelines.len());
@@ -616,7 +624,7 @@ async fn timeline_preserve_initdb_handler(
     // location where timeline recreation cand find it.
 
     async {
-        let tenant = mgr::get_tenant(tenant_shard_id, true)?;
+        let tenant = mgr::get_tenant(tenant_shard_id, false)?;
 
         let timeline = tenant
             .get_timeline(timeline_id, false)
@@ -1136,7 +1144,7 @@ async fn tenant_shard_split_handler(
 
     let new_shards = state
         .tenant_manager
-        .shard_split(tenant_shard_id, ShardCount(req.new_shard_count), &ctx)
+        .shard_split(tenant_shard_id, ShardCount::new(req.new_shard_count), &ctx)
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -1951,11 +1959,60 @@ async fn put_io_engine_handler(
     mut r: Request<Body>,
     _cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
     let kind: crate::virtual_file::IoEngineKind = json_request(&mut r).await?;
     crate::virtual_file::io_engine::set(kind);
     json_response(StatusCode::OK, ())
 }
 
+/// Polled by control plane.
+///
+/// See [`crate::utilization`].
+async fn get_utilization(
+    r: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    // this probably could be completely public, but lets make that change later.
+    check_permission(&r, None)?;
+
+    let state = get_state(&r);
+    let mut g = state.latest_utilization.lock().await;
+
+    let regenerate_every = Duration::from_secs(1);
+    let still_valid = g
+        .as_ref()
+        .is_some_and(|(captured_at, _)| captured_at.elapsed() < regenerate_every);
+
+    // avoid needless statvfs calls even though those should be non-blocking fast.
+    // regenerate at most 1Hz to allow polling at any rate.
+    if !still_valid {
+        let path = state.conf.tenants_path();
+        let doc = crate::utilization::regenerate(path.as_std_path())
+            .map_err(ApiError::InternalServerError)?;
+
+        let mut buf = Vec::new();
+        serde_json::to_writer(&mut buf, &doc)
+            .context("serialize")
+            .map_err(ApiError::InternalServerError)?;
+
+        let body = bytes::Bytes::from(buf);
+
+        *g = Some((std::time::Instant::now(), body));
+    }
+
+    // hyper 0.14 doesn't yet have Response::clone so this is a bit of extra legwork
+    let cached = g.as_ref().expect("just set").1.clone();
+
+    Response::builder()
+        .header(hyper::http::header::CONTENT_TYPE, "application/json")
+        // thought of using http date header, but that is second precision which does not give any
+        // debugging aid
+        .status(StatusCode::OK)
+        .body(hyper::Body::from(cached))
+        .context("build response")
+        .map_err(ApiError::InternalServerError)
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -2214,8 +2271,9 @@ pub fn make_router(
         )
         .get(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/keyspace",
-            |r| testing_api_handler("read out the keyspace", r, timeline_collect_keyspace),
+            |r| api_handler(r, timeline_collect_keyspace),
         )
         .put("/v1/io_engine", |r| api_handler(r, put_io_engine_handler))
+        .get("/v1/utilization", |r| api_handler(r, get_utilization))
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -22,6 +22,7 @@ pub(crate) mod statvfs;
 pub mod task_mgr;
 pub mod tenant;
 pub mod trace;
+pub mod utilization;
 pub mod virtual_file;
 pub mod walingest;
 pub mod walrecord;
@@ -168,15 +169,6 @@ pub fn is_delete_mark(path: &Utf8Path) -> bool {
     ends_with_suffix(path, TIMELINE_DELETE_MARK_SUFFIX)
 }
 
-fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
-    if let Some(e) = e.io_error() {
-        if e.kind() == std::io::ErrorKind::NotFound {
-            return true;
-        }
-    }
-    false
-}
-
 /// During pageserver startup, we need to order operations not to exhaust tokio worker threads by
 /// blocking.
 ///

pageserver/src/metrics.rs

@@ -4,8 +4,8 @@ use metrics::{
     register_counter_vec, register_gauge_vec, register_histogram, register_histogram_vec,
     register_int_counter, register_int_counter_pair_vec, register_int_counter_vec,
     register_int_gauge, register_int_gauge_vec, register_uint_gauge, register_uint_gauge_vec,
-    Counter, CounterVec, GaugeVec, Histogram, HistogramVec, IntCounter, IntCounterPairVec,
-    IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
+    Counter, CounterVec, GaugeVec, Histogram, HistogramVec, IntCounter, IntCounterPair,
+    IntCounterPairVec, IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
 use pageserver_api::shard::TenantShardId;
@@ -1266,13 +1266,12 @@ pub(crate) static LIVE_CONNECTIONS_COUNT: Lazy<IntGaugeVec> = Lazy::new(|| {
 
 // remote storage metrics
 
-/// NB: increment _after_ recording the current value into [`REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST`].
-static REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE: Lazy<IntGaugeVec> = Lazy::new(|| {
-    register_int_gauge_vec!(
-        "pageserver_remote_timeline_client_calls_unfinished",
-        "Number of ongoing calls to remote timeline client. \
-         Used to populate pageserver_remote_timeline_client_calls_started. \
-         This metric is not useful for sampling from Prometheus, but useful in tests.",
+static REMOTE_TIMELINE_CLIENT_CALLS: Lazy<IntCounterPairVec> = Lazy::new(|| {
+    register_int_counter_pair_vec!(
+        "pageserver_remote_timeline_client_calls_started",
+        "Number of started calls to remote timeline client.",
+        "pageserver_remote_timeline_client_calls_finished",
+        "Number of finshed calls to remote timeline client.",
         &[
             "tenant_id",
             "shard_id",
@@ -1281,23 +1280,7 @@ static REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE: Lazy<IntGaugeVec> = Lazy::
             "op_kind"
         ],
     )
-    .expect("failed to define a metric")
-});
-
-static REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
-        "pageserver_remote_timeline_client_calls_started",
-        "When calling a remote timeline client method, we record the current value \
-         of the calls_unfinished gauge in this histogram. Plot the histogram \
-         over time in a heatmap to visualize how many operations were ongoing \
-         at a given instant. It gives you a better idea of the queue depth \
-         than plotting the gauge directly, since operations may complete faster \
-         than the sampling interval.",
-        &["file_kind", "op_kind"],
-        // The calls_unfinished gauge is an integer gauge, hence we have integer buckets.
-        vec![0.0, 1.0, 2.0, 4.0, 6.0, 8.0, 10.0, 15.0, 20.0, 40.0, 60.0, 80.0, 100.0, 500.0],
-    )
-    .expect("failed to define a metric")
+    .unwrap()
 });
 
 static REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER: Lazy<IntCounterVec> =
@@ -2078,7 +2061,7 @@ pub(crate) struct RemoteTimelineClientMetrics {
     shard_id: String,
     timeline_id: String,
     remote_physical_size_gauge: Mutex<Option<PerTimelineRemotePhysicalSizeGauge>>,
-    calls_unfinished_gauge: Mutex<HashMap<(&'static str, &'static str), IntGauge>>,
+    calls: Mutex<HashMap<(&'static str, &'static str), IntCounterPair>>,
     bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
     bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
 }
@@ -2089,7 +2072,7 @@ impl RemoteTimelineClientMetrics {
             tenant_id: tenant_shard_id.tenant_id.to_string(),
             shard_id: format!("{}", tenant_shard_id.shard_slug()),
             timeline_id: timeline_id.to_string(),
-            calls_unfinished_gauge: Mutex::new(HashMap::default()),
+            calls: Mutex::new(HashMap::default()),
             bytes_started_counter: Mutex::new(HashMap::default()),
             bytes_finished_counter: Mutex::new(HashMap::default()),
             remote_physical_size_gauge: Mutex::new(None),
@@ -2129,15 +2112,15 @@ impl RemoteTimelineClientMetrics {
             .unwrap()
     }
 
-    fn calls_unfinished_gauge(
+    fn calls_counter_pair(
         &self,
         file_kind: &RemoteOpFileKind,
         op_kind: &RemoteOpKind,
-    ) -> IntGauge {
-        let mut guard = self.calls_unfinished_gauge.lock().unwrap();
+    ) -> IntCounterPair {
+        let mut guard = self.calls.lock().unwrap();
         let key = (file_kind.as_str(), op_kind.as_str());
         let metric = guard.entry(key).or_insert_with(move || {
-            REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE
+            REMOTE_TIMELINE_CLIENT_CALLS
                 .get_metric_with_label_values(&[
                     &self.tenant_id,
                     &self.shard_id,
@@ -2150,17 +2133,6 @@ impl RemoteTimelineClientMetrics {
         metric.clone()
     }
 
-    fn calls_started_hist(
-        &self,
-        file_kind: &RemoteOpFileKind,
-        op_kind: &RemoteOpKind,
-    ) -> Histogram {
-        let key = (file_kind.as_str(), op_kind.as_str());
-        REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST
-            .get_metric_with_label_values(&[key.0, key.1])
-            .unwrap()
-    }
-
     fn bytes_started_counter(
         &self,
         file_kind: &RemoteOpFileKind,
@@ -2231,7 +2203,7 @@ impl RemoteTimelineClientMetrics {
 #[must_use]
 pub(crate) struct RemoteTimelineClientCallMetricGuard {
     /// Decremented on drop.
-    calls_unfinished_metric: Option<IntGauge>,
+    calls_counter_pair: Option<IntCounterPair>,
     /// If Some(), this references the bytes_finished metric, and we increment it by the given `u64` on drop.
     bytes_finished: Option<(IntCounter, u64)>,
 }
@@ -2241,10 +2213,10 @@ impl RemoteTimelineClientCallMetricGuard {
     /// The caller vouches to do the metric updates manually.
     pub fn will_decrement_manually(mut self) {
         let RemoteTimelineClientCallMetricGuard {
-            calls_unfinished_metric,
+            calls_counter_pair,
             bytes_finished,
         } = &mut self;
-        calls_unfinished_metric.take();
+        calls_counter_pair.take();
         bytes_finished.take();
     }
 }
@@ -2252,10 +2224,10 @@ impl RemoteTimelineClientCallMetricGuard {
 impl Drop for RemoteTimelineClientCallMetricGuard {
     fn drop(&mut self) {
         let RemoteTimelineClientCallMetricGuard {
-            calls_unfinished_metric,
+            calls_counter_pair,
             bytes_finished,
         } = self;
-        if let Some(guard) = calls_unfinished_metric.take() {
+        if let Some(guard) = calls_counter_pair.take() {
             guard.dec();
         }
         if let Some((bytes_finished_metric, value)) = bytes_finished {
@@ -2288,10 +2260,8 @@ impl RemoteTimelineClientMetrics {
         op_kind: &RemoteOpKind,
         size: RemoteTimelineClientMetricsCallTrackSize,
     ) -> RemoteTimelineClientCallMetricGuard {
-        let calls_unfinished_metric = self.calls_unfinished_gauge(file_kind, op_kind);
-        self.calls_started_hist(file_kind, op_kind)
-            .observe(calls_unfinished_metric.get() as f64);
-        calls_unfinished_metric.inc(); // NB: inc after the histogram, see comment on underlying metric
+        let calls_counter_pair = self.calls_counter_pair(file_kind, op_kind);
+        calls_counter_pair.inc();
 
         let bytes_finished = match size {
             RemoteTimelineClientMetricsCallTrackSize::DontTrackSize { reason: _reason } => {
@@ -2305,7 +2275,7 @@ impl RemoteTimelineClientMetrics {
             }
         };
         RemoteTimelineClientCallMetricGuard {
-            calls_unfinished_metric: Some(calls_unfinished_metric),
+            calls_counter_pair: Some(calls_counter_pair),
             bytes_finished,
         }
     }
@@ -2319,12 +2289,8 @@ impl RemoteTimelineClientMetrics {
         op_kind: &RemoteOpKind,
         size: RemoteTimelineClientMetricsCallTrackSize,
     ) {
-        let calls_unfinished_metric = self.calls_unfinished_gauge(file_kind, op_kind);
-        debug_assert!(
-            calls_unfinished_metric.get() > 0,
-            "begin and end should cancel out"
-        );
-        calls_unfinished_metric.dec();
+        let calls_counter_pair = self.calls_counter_pair(file_kind, op_kind);
+        calls_counter_pair.dec();
         match size {
             RemoteTimelineClientMetricsCallTrackSize::DontTrackSize { reason: _reason } => {}
             RemoteTimelineClientMetricsCallTrackSize::Bytes(size) => {
@@ -2341,18 +2307,15 @@ impl Drop for RemoteTimelineClientMetrics {
             shard_id,
             timeline_id,
             remote_physical_size_gauge,
-            calls_unfinished_gauge,
+            calls,
             bytes_started_counter,
             bytes_finished_counter,
         } = self;
-        for ((a, b), _) in calls_unfinished_gauge.get_mut().unwrap().drain() {
-            let _ = REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE.remove_label_values(&[
-                tenant_id,
-                shard_id,
-                timeline_id,
-                a,
-                b,
-            ]);
+        for ((a, b), _) in calls.get_mut().unwrap().drain() {
+            let mut res = [Ok(()), Ok(())];
+            REMOTE_TIMELINE_CLIENT_CALLS
+                .remove_label_values(&mut res, &[tenant_id, shard_id, timeline_id, a, b]);
+            // don't care about results
         }
         for ((a, b), _) in bytes_started_counter.get_mut().unwrap().drain() {
             let _ = REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER.remove_label_values(&[
@@ -2496,6 +2459,56 @@ pub mod tokio_epoll_uring {
     }
 }
 
+pub(crate) mod tenant_throttling {
+    use metrics::{register_int_counter_vec, IntCounter};
+    use once_cell::sync::Lazy;
+
+    use crate::tenant::{self, throttle::Metric};
+
+    pub(crate) struct TimelineGet {
+        wait_time: IntCounter,
+        count: IntCounter,
+    }
+
+    pub(crate) static TIMELINE_GET: Lazy<TimelineGet> = Lazy::new(|| {
+        static WAIT_USECS: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+            register_int_counter_vec!(
+            "pageserver_tenant_throttling_wait_usecs_sum_global",
+            "Sum of microseconds that tenants spent waiting for a tenant throttle of a given kind.",
+            &["kind"]
+        )
+            .unwrap()
+        });
+
+        static WAIT_COUNT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
+            register_int_counter_vec!(
+                "pageserver_tenant_throttling_count_global",
+                "Count of tenant throttlings, by kind of throttle.",
+                &["kind"]
+            )
+            .unwrap()
+        });
+
+        let kind = "timeline_get";
+        TimelineGet {
+            wait_time: WAIT_USECS.with_label_values(&[kind]),
+            count: WAIT_COUNT.with_label_values(&[kind]),
+        }
+    });
+
+    impl Metric for &'static TimelineGet {
+        #[inline(always)]
+        fn observe_throttling(
+            &self,
+            tenant::throttle::Observation { wait_time }: &tenant::throttle::Observation,
+        ) {
+            let val = u64::try_from(wait_time.as_micros()).unwrap();
+            self.wait_time.inc_by(val);
+            self.count.inc();
+        }
+    }
+}
+
 pub fn preinitialize_metrics() {
     // Python tests need these and on some we do alerting.
     //
@@ -2557,4 +2570,5 @@ pub fn preinitialize_metrics() {
 
     // Custom
     Lazy::force(&RECONSTRUCT_TIME);
+    Lazy::force(&tenant_throttling::TIMELINE_GET);
 }

pageserver/src/page_service.rs

@@ -26,7 +26,7 @@ use pageserver_api::models::{
     PagestreamNblocksResponse,
 };
 use pageserver_api::shard::ShardIndex;
-use pageserver_api::shard::{ShardCount, ShardNumber};
+use pageserver_api::shard::ShardNumber;
 use postgres_backend::{self, is_expected_io_error, AuthType, PostgresBackend, QueryError};
 use pq_proto::framed::ConnectionError;
 use pq_proto::FeStartupPacket;
@@ -998,7 +998,7 @@ impl PageServerHandler {
     ) -> Result<&Arc<Timeline>, Key> {
         let key = if let Some((first_idx, first_timeline)) = self.shard_timelines.iter().next() {
             // Fastest path: single sharded case
-            if first_idx.shard_count < ShardCount(2) {
+            if first_idx.shard_count.count() == 1 {
                 return Ok(&first_timeline.timeline);
             }
 

pageserver/src/pgdatadir_mapping.rs

@@ -15,6 +15,7 @@ use crate::walrecord::NeonWalRecord;
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
+use itertools::Itertools;
 use pageserver_api::key::{
     dbdir_key_range, is_rel_block_key, is_slru_block_key, rel_block_to_key, rel_dir_to_key,
     rel_key_range, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
@@ -156,6 +157,7 @@ impl Timeline {
             pending_updates: HashMap::new(),
             pending_deletions: Vec::new(),
             pending_nblocks: 0,
+            pending_aux_files: None,
             pending_directory_entries: Vec::new(),
             lsn,
         }
@@ -870,6 +872,14 @@ pub struct DatadirModification<'a> {
     pending_updates: HashMap<Key, Vec<(Lsn, Value)>>,
     pending_deletions: Vec<(Range<Key>, Lsn)>,
     pending_nblocks: i64,
+
+    // If we already wrote any aux file changes in this modification, stash the latest dir.  If set,
+    // [`Self::put_file`] may assume that it is safe to emit a delta rather than checking
+    // if AUX_FILES_KEY is already set.
+    pending_aux_files: Option<AuxFilesDirectory>,
+
+    /// For special "directory" keys that store key-value maps, track the size of the map
+    /// if it was updated in this modification.
     pending_directory_entries: Vec<(DirectoryKind, usize)>,
 }
 
@@ -1384,31 +1394,76 @@ impl<'a> DatadirModification<'a> {
         content: &[u8],
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
-            Ok(buf) => AuxFilesDirectory::des(&buf)?,
-            Err(e) => {
-                // This is expected: historical databases do not have the key.
-                debug!("Failed to get info about AUX files: {}", e);
-                AuxFilesDirectory {
-                    files: HashMap::new(),
+        let file_path = path.to_string();
+        let content = if content.is_empty() {
+            None
+        } else {
+            Some(Bytes::copy_from_slice(content))
+        };
+
+        let dir = if let Some(mut dir) = self.pending_aux_files.take() {
+            // We already updated aux files in `self`: emit a delta and update our latest value
+
+            self.put(
+                AUX_FILES_KEY,
+                Value::WalRecord(NeonWalRecord::AuxFile {
+                    file_path: file_path.clone(),
+                    content: content.clone(),
+                }),
+            );
+
+            dir.upsert(file_path, content);
+            dir
+        } else {
+            // Check if the AUX_FILES_KEY is initialized
+            match self.get(AUX_FILES_KEY, ctx).await {
+                Ok(dir_bytes) => {
+                    let mut dir = AuxFilesDirectory::des(&dir_bytes)?;
+                    // Key is already set, we may append a delta
+                    self.put(
+                        AUX_FILES_KEY,
+                        Value::WalRecord(NeonWalRecord::AuxFile {
+                            file_path: file_path.clone(),
+                            content: content.clone(),
+                        }),
+                    );
+                    dir.upsert(file_path, content);
+                    dir
+                }
+                Err(
+                    e @ (PageReconstructError::AncestorStopping(_)
+                    | PageReconstructError::Cancelled
+                    | PageReconstructError::AncestorLsnTimeout(_)),
+                ) => {
+                    // Important that we do not interpret a shutdown error as "not found" and thereby
+                    // reset the map.
+                    return Err(e.into());
+                }
+                // FIXME: PageReconstructError doesn't have an explicit variant for key-not-found, so
+                // we are assuming that all _other_ possible errors represents a missing key.  If some
+                // other error occurs, we may incorrectly reset the map of aux files.
+                Err(PageReconstructError::Other(_) | PageReconstructError::WalRedo(_)) => {
+                    // Key is missing, we must insert an image as the basis for subsequent deltas.
+
+                    let mut dir = AuxFilesDirectory {
+                        files: HashMap::new(),
+                    };
+                    dir.upsert(file_path, content);
+                    self.put(
+                        AUX_FILES_KEY,
+                        Value::Image(Bytes::from(
+                            AuxFilesDirectory::ser(&dir).context("serialize")?,
+                        )),
+                    );
+                    dir
                 }
             }
         };
-        let path = path.to_string();
-        if content.is_empty() {
-            dir.files.remove(&path);
-        } else {
-            dir.files.insert(path, Bytes::copy_from_slice(content));
-        }
+
         self.pending_directory_entries
             .push((DirectoryKind::AuxFiles, dir.files.len()));
+        self.pending_aux_files = Some(dir);
 
-        self.put(
-            AUX_FILES_KEY,
-            Value::Image(Bytes::from(
-                AuxFilesDirectory::ser(&dir).context("serialize")?,
-            )),
-        );
         Ok(())
     }
 
@@ -1438,7 +1493,7 @@ impl<'a> DatadirModification<'a> {
             return Ok(());
         }
 
-        let writer = self.tline.writer().await;
+        let mut writer = self.tline.writer().await;
 
         // Flush relation and  SLRU data blocks, keep metadata.
         let mut retained_pending_updates = HashMap::<_, Vec<_>>::new();
@@ -1477,13 +1532,23 @@ impl<'a> DatadirModification<'a> {
     /// All the modifications in this atomic update are stamped by the specified LSN.
     ///
     pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
-        let writer = self.tline.writer().await;
+        let mut writer = self.tline.writer().await;
 
         let pending_nblocks = self.pending_nblocks;
         self.pending_nblocks = 0;
 
         if !self.pending_updates.is_empty() {
-            writer.put_batch(&self.pending_updates, ctx).await?;
+            let prev_pending_updates = std::mem::take(&mut self.pending_updates);
+
+            // The put_batch call below expects expects the inputs to be sorted by Lsn,
+            // so we do that first.
+            let lsn_ordered_batch: Vec<(Key, Lsn, Value)> = prev_pending_updates
+                .into_iter()
+                .map(|(key, vals)| vals.into_iter().map(move |(lsn, val)| (key, lsn, val)))
+                .kmerge_by(|lhs, rhs| lhs.1 .0 < rhs.1 .0)
+                .collect();
+
+            writer.put_batch(lsn_ordered_batch, ctx).await?;
             self.pending_updates.clear();
         }
 
@@ -1618,8 +1683,18 @@ struct RelDirectory {
 }
 
 #[derive(Debug, Serialize, Deserialize, Default)]
-struct AuxFilesDirectory {
-    files: HashMap<String, Bytes>,
+pub(crate) struct AuxFilesDirectory {
+    pub(crate) files: HashMap<String, Bytes>,
+}
+
+impl AuxFilesDirectory {
+    pub(crate) fn upsert(&mut self, key: String, value: Option<Bytes>) {
+        if let Some(value) = value {
+            self.files.insert(key, value);
+        } else {
+            self.files.remove(&key);
+        }
+    }
 }
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -1655,8 +1730,60 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 #[allow(clippy::bool_assert_comparison)]
 #[cfg(test)]
 mod tests {
-    //use super::repo_harness::*;
-    //use super::*;
+    use hex_literal::hex;
+    use utils::id::TimelineId;
+
+    use super::*;
+
+    use crate::{tenant::harness::TenantHarness, DEFAULT_PG_VERSION};
+
+    /// Test a round trip of aux file updates, from DatadirModification to reading back from the Timeline
+    #[tokio::test]
+    async fn aux_files_round_trip() -> anyhow::Result<()> {
+        let name = "aux_files_round_trip";
+        let harness = TenantHarness::create(name)?;
+
+        pub const TIMELINE_ID: TimelineId =
+            TimelineId::from_array(hex!("11223344556677881122334455667788"));
+
+        let (tenant, ctx) = harness.load().await;
+        let tline = tenant
+            .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        let tline = tline.raw_timeline().unwrap();
+
+        // First modification: insert two keys
+        let mut modification = tline.begin_modification(Lsn(0x1000));
+        modification.put_file("foo/bar1", b"content1", &ctx).await?;
+        modification.set_lsn(Lsn(0x1008))?;
+        modification.put_file("foo/bar2", b"content2", &ctx).await?;
+        modification.commit(&ctx).await?;
+        let expect_1008 = HashMap::from([
+            ("foo/bar1".to_string(), Bytes::from_static(b"content1")),
+            ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
+        ]);
+
+        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        assert_eq!(readback, expect_1008);
+
+        // Second modification: update one key, remove the other
+        let mut modification = tline.begin_modification(Lsn(0x2000));
+        modification.put_file("foo/bar1", b"content3", &ctx).await?;
+        modification.set_lsn(Lsn(0x2008))?;
+        modification.put_file("foo/bar2", b"", &ctx).await?;
+        modification.commit(&ctx).await?;
+        let expect_2008 =
+            HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
+
+        let readback = tline.list_aux_files(Lsn(0x2008), &ctx).await?;
+        assert_eq!(readback, expect_2008);
+
+        // Reading back in time works
+        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        assert_eq!(readback, expect_1008);
+
+        Ok(())
+    }
 
     /*
         fn assert_current_logical_size<R: Repository>(timeline: &DatadirTimeline<R>, lsn: Lsn) {

pageserver/src/task_mgr.rs

@@ -30,10 +30,6 @@
 //! only a single tenant or timeline.
 //!
 
-// Clippy 1.60 incorrectly complains about the tokio::task_local!() macro.
-// Silence it. See https://github.com/rust-lang/rust-clippy/issues/9224.
-#![allow(clippy::declare_interior_mutable_const)]
-
 use std::collections::HashMap;
 use std::fmt;
 use std::future::Future;
@@ -192,6 +188,7 @@ task_local! {
     serde::Serialize,
     serde::Deserialize,
     strum_macros::IntoStaticStr,
+    strum_macros::EnumString,
 )]
 pub enum TaskKind {
     // Pageserver startup, i.e., `main`
@@ -312,7 +309,6 @@ struct MutableTaskState {
 }
 
 struct PageServerTask {
-    #[allow(dead_code)] // unused currently
     task_id: PageserverTaskId,
 
     kind: TaskKind,

pageserver/src/tenant/config.rs

@@ -9,8 +9,8 @@
 //! may lead to a data loss.
 //!
 use anyhow::bail;
-use pageserver_api::models;
 use pageserver_api::models::EvictionPolicy;
+use pageserver_api::models::{self, ThrottleConfig};
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
 use serde::de::IntoDeserializer;
 use serde::{Deserialize, Serialize};
@@ -251,7 +251,7 @@ impl LocationConf {
         } else {
             ShardIdentity::new(
                 ShardNumber(conf.shard_number),
-                ShardCount(conf.shard_count),
+                ShardCount::new(conf.shard_count),
                 ShardStripeSize(conf.shard_stripe_size),
             )?
         };
@@ -285,7 +285,7 @@ impl Default for LocationConf {
 ///
 /// For storing and transmitting individual tenant's configuration, see
 /// TenantConfOpt.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct TenantConf {
     // Flush out an inmemory layer, if it's holding WAL older than this
     // This puts a backstop on how much WAL needs to be re-digested if the
@@ -348,11 +348,13 @@ pub struct TenantConf {
 
     /// If true then SLRU segments are dowloaded on demand, if false SLRU segments are included in basebackup
     pub lazy_slru_download: bool,
+
+    pub timeline_get_throttle: pageserver_api::models::ThrottleConfig,
 }
 
 /// Same as TenantConf, but this struct preserves the information about
 /// which parameters are set and which are not.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, Default)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
 pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
@@ -437,6 +439,9 @@ pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub lazy_slru_download: Option<bool>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub timeline_get_throttle: Option<pageserver_api::models::ThrottleConfig>,
 }
 
 impl TenantConfOpt {
@@ -485,6 +490,10 @@ impl TenantConfOpt {
             lazy_slru_download: self
                 .lazy_slru_download
                 .unwrap_or(global_conf.lazy_slru_download),
+            timeline_get_throttle: self
+                .timeline_get_throttle
+                .clone()
+                .unwrap_or(global_conf.timeline_get_throttle),
         }
     }
 }
@@ -524,6 +533,7 @@ impl Default for TenantConf {
             gc_feedback: false,
             heatmap_period: Duration::ZERO,
             lazy_slru_download: false,
+            timeline_get_throttle: crate::tenant::throttle::Config::disabled(),
         }
     }
 }
@@ -596,6 +606,7 @@ impl From<TenantConfOpt> for models::TenantConfig {
             gc_feedback: value.gc_feedback,
             heatmap_period: value.heatmap_period.map(humantime),
             lazy_slru_download: value.lazy_slru_download,
+            timeline_get_throttle: value.timeline_get_throttle.map(ThrottleConfig::from),
         }
     }
 }

pageserver/src/tenant/delete.rs

@@ -3,7 +3,7 @@ use std::sync::Arc;
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::{models::TenantState, shard::TenantShardId};
-use remote_storage::{GenericRemoteStorage, RemotePath};
+use remote_storage::{GenericRemoteStorage, RemotePath, TimeoutOrCancel};
 use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
 use tracing::{error, instrument, Instrument};
@@ -84,17 +84,17 @@ async fn create_remote_delete_mark(
             let data = bytes::Bytes::from_static(data);
             let stream = futures::stream::once(futures::future::ready(Ok(data)));
             remote_storage
-                .upload(stream, 0, &remote_mark_path, None)
+                .upload(stream, 0, &remote_mark_path, None, cancel)
                 .await
         },
-        |_e| false,
+        TimeoutOrCancel::caused_by_cancel,
         FAILED_UPLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         "mark_upload",
         cancel,
     )
     .await
-    .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+    .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
     .and_then(|x| x)
     .context("mark_upload")?;
 
@@ -184,15 +184,15 @@ async fn remove_tenant_remote_delete_mark(
     if let Some(remote_storage) = remote_storage {
         let path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
         backoff::retry(
-            || async { remote_storage.delete(&path).await },
-            |_e| false,
+            || async { remote_storage.delete(&path, cancel).await },
+            TimeoutOrCancel::caused_by_cancel,
             FAILED_UPLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "remove_tenant_remote_delete_mark",
             cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)
         .context("remove_tenant_remote_delete_mark")?;
     }
@@ -246,6 +246,8 @@ async fn cleanup_remaining_fs_traces(
 
     rm(conf.tenant_deleted_mark_file_path(tenant_shard_id), false).await?;
 
+    rm(conf.tenant_heatmap_path(tenant_shard_id), false).await?;
+
     fail::fail_point!("tenant-delete-before-remove-tenant-dir", |_| {
         Err(anyhow::anyhow!(
             "failpoint: tenant-delete-before-remove-tenant-dir"

pageserver/src/tenant/disk_btree.rs

@@ -36,7 +36,6 @@ use crate::{
 pub const VALUE_SZ: usize = 5;
 pub const MAX_VALUE: u64 = 0x007f_ffff_ffff;
 
-#[allow(dead_code)]
 pub const PAGE_SZ: usize = 8192;
 
 #[derive(Clone, Copy, Debug)]

pageserver/src/tenant/ephemeral_file.rs

@@ -6,6 +6,7 @@ use crate::context::RequestContext;
 use crate::page_cache::{self, PAGE_SZ};
 use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReader};
 use crate::virtual_file::{self, VirtualFile};
+use bytes::BytesMut;
 use camino::Utf8PathBuf;
 use pageserver_api::shard::TenantShardId;
 use std::cmp::min;
@@ -26,7 +27,10 @@ pub struct EphemeralFile {
     /// An ephemeral file is append-only.
     /// We keep the last page, which can still be modified, in [`Self::mutable_tail`].
     /// The other pages, which can no longer be modified, are accessed through the page cache.
-    mutable_tail: [u8; PAGE_SZ],
+    ///
+    /// None <=> IO is ongoing.
+    /// Size is fixed to PAGE_SZ at creation time and must not be changed.
+    mutable_tail: Option<BytesMut>,
 }
 
 impl EphemeralFile {
@@ -60,7 +64,7 @@ impl EphemeralFile {
             _timeline_id: timeline_id,
             file,
             len: 0,
-            mutable_tail: [0u8; PAGE_SZ],
+            mutable_tail: Some(BytesMut::zeroed(PAGE_SZ)),
         })
     }
 
@@ -103,7 +107,13 @@ impl EphemeralFile {
             };
         } else {
             debug_assert_eq!(blknum as u64, self.len / PAGE_SZ as u64);
-            Ok(BlockLease::EphemeralFileMutableTail(&self.mutable_tail))
+            Ok(BlockLease::EphemeralFileMutableTail(
+                self.mutable_tail
+                    .as_deref()
+                    .expect("we're not doing IO, it must be Some()")
+                    .try_into()
+                    .expect("we ensure that it's always PAGE_SZ"),
+            ))
         }
     }
 
@@ -135,21 +145,27 @@ impl EphemeralFile {
             ) -> Result<(), io::Error> {
                 let mut src_remaining = src;
                 while !src_remaining.is_empty() {
-                    let dst_remaining = &mut self.ephemeral_file.mutable_tail[self.off..];
+                    let dst_remaining = &mut self
+                        .ephemeral_file
+                        .mutable_tail
+                        .as_deref_mut()
+                        .expect("IO is not yet ongoing")[self.off..];
                     let n = min(dst_remaining.len(), src_remaining.len());
                     dst_remaining[..n].copy_from_slice(&src_remaining[..n]);
                     self.off += n;
                     src_remaining = &src_remaining[n..];
                     if self.off == PAGE_SZ {
-                        match self
+                        let mutable_tail = std::mem::take(&mut self.ephemeral_file.mutable_tail)
+                            .expect("IO is not yet ongoing");
+                        let (mutable_tail, res) = self
                             .ephemeral_file
                             .file
-                            .write_all_at(
-                                &self.ephemeral_file.mutable_tail,
-                                self.blknum as u64 * PAGE_SZ as u64,
-                            )
-                            .await
-                        {
+                            .write_all_at(mutable_tail, self.blknum as u64 * PAGE_SZ as u64)
+                            .await;
+                        // TODO: If we panic before we can put the mutable_tail back, subsequent calls will fail.
+                        // I.e., the IO isn't retryable if we panic.
+                        self.ephemeral_file.mutable_tail = Some(mutable_tail);
+                        match res {
                             Ok(_) => {
                                 // Pre-warm the page cache with what we just wrote.
                                 // This isn't necessary for coherency/correctness, but it's how we've always done it.
@@ -169,7 +185,12 @@ impl EphemeralFile {
                                     Ok(page_cache::ReadBufResult::NotFound(mut write_guard)) => {
                                         let buf: &mut [u8] = write_guard.deref_mut();
                                         debug_assert_eq!(buf.len(), PAGE_SZ);
-                                        buf.copy_from_slice(&self.ephemeral_file.mutable_tail);
+                                        buf.copy_from_slice(
+                                            self.ephemeral_file
+                                                .mutable_tail
+                                                .as_deref()
+                                                .expect("IO is not ongoing"),
+                                        );
                                         let _ = write_guard.mark_valid();
                                         // pre-warm successful
                                     }
@@ -181,7 +202,11 @@ impl EphemeralFile {
                                 // Zero the buffer for re-use.
                                 // Zeroing is critical for correcntess because the write_blob code below
                                 // and similarly read_blk expect zeroed pages.
-                                self.ephemeral_file.mutable_tail.fill(0);
+                                self.ephemeral_file
+                                    .mutable_tail
+                                    .as_deref_mut()
+                                    .expect("IO is not ongoing")
+                                    .fill(0);
                                 // This block is done, move to next one.
                                 self.blknum += 1;
                                 self.off = 0;

pageserver/src/tenant/layer_map.rs

@@ -52,8 +52,7 @@ use crate::repository::Key;
 use crate::tenant::storage_layer::InMemoryLayer;
 use anyhow::Result;
 use pageserver_api::keyspace::KeySpaceAccum;
-use std::cmp::Ordering;
-use std::collections::{BTreeMap, VecDeque};
+use std::collections::{HashMap, VecDeque};
 use std::iter::Peekable;
 use std::ops::Range;
 use std::sync::Arc;
@@ -147,43 +146,28 @@ impl Drop for BatchedUpdates<'_> {
 }
 
 /// Return value of LayerMap::search
-#[derive(Eq, PartialEq, Debug)]
+#[derive(Eq, PartialEq, Debug, Hash)]
 pub struct SearchResult {
     pub layer: Arc<PersistentLayerDesc>,
     pub lsn_floor: Lsn,
 }
 
-pub struct OrderedSearchResult(SearchResult);
-
-impl Ord for OrderedSearchResult {
-    fn cmp(&self, other: &Self) -> Ordering {
-        self.0.lsn_floor.cmp(&other.0.lsn_floor)
-    }
-}
-
-impl PartialOrd for OrderedSearchResult {
-    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-impl PartialEq for OrderedSearchResult {
-    fn eq(&self, other: &Self) -> bool {
-        self.0.lsn_floor == other.0.lsn_floor
-    }
-}
-
-impl Eq for OrderedSearchResult {}
-
+/// Return value of [`LayerMap::range_search`]
+///
+/// Contains a mapping from a layer description to a keyspace
+/// accumulator that contains all the keys which intersect the layer
+/// from the original search space. Keys that were not found are accumulated
+/// in a separate key space accumulator.
+#[derive(Debug)]
 pub struct RangeSearchResult {
-    pub found: BTreeMap<OrderedSearchResult, KeySpaceAccum>,
+    pub found: HashMap<SearchResult, KeySpaceAccum>,
     pub not_found: KeySpaceAccum,
 }
 
 impl RangeSearchResult {
     fn new() -> Self {
         Self {
-            found: BTreeMap::new(),
+            found: HashMap::new(),
             not_found: KeySpaceAccum::new(),
         }
     }
@@ -314,7 +298,7 @@ where
             Some(search_result) => self
                 .result
                 .found
-                .entry(OrderedSearchResult(search_result))
+                .entry(search_result)
                 .or_default()
                 .add_range(covered_range),
             None => self.pad_range(covered_range),
@@ -362,6 +346,35 @@ where
     }
 }
 
+#[derive(PartialEq, Eq, Hash, Debug, Clone)]
+pub enum InMemoryLayerHandle {
+    Open {
+        lsn_floor: Lsn,
+        end_lsn: Lsn,
+    },
+    Frozen {
+        idx: usize,
+        lsn_floor: Lsn,
+        end_lsn: Lsn,
+    },
+}
+
+impl InMemoryLayerHandle {
+    pub fn get_lsn_floor(&self) -> Lsn {
+        match self {
+            InMemoryLayerHandle::Open { lsn_floor, .. } => *lsn_floor,
+            InMemoryLayerHandle::Frozen { lsn_floor, .. } => *lsn_floor,
+        }
+    }
+
+    pub fn get_end_lsn(&self) -> Lsn {
+        match self {
+            InMemoryLayerHandle::Open { end_lsn, .. } => *end_lsn,
+            InMemoryLayerHandle::Frozen { end_lsn, .. } => *end_lsn,
+        }
+    }
+}
+
 impl LayerMap {
     ///
     /// Find the latest layer (by lsn.end) that covers the given
@@ -556,6 +569,43 @@ impl LayerMap {
         self.historic.iter()
     }
 
+    /// Get a handle for the first in memory layer that matches the provided predicate.
+    /// The handle should be used with [`Self::get_in_memory_layer`] to retrieve the actual layer.
+    ///
+    /// Note: [`Self::find_in_memory_layer`] and [`Self::get_in_memory_layer`] should be called during
+    /// the same exclusive region established by holding the layer manager lock.
+    pub fn find_in_memory_layer<Pred>(&self, mut pred: Pred) -> Option<InMemoryLayerHandle>
+    where
+        Pred: FnMut(&Arc<InMemoryLayer>) -> bool,
+    {
+        if let Some(open) = &self.open_layer {
+            if pred(open) {
+                return Some(InMemoryLayerHandle::Open {
+                    lsn_floor: open.get_lsn_range().start,
+                    end_lsn: open.get_lsn_range().end,
+                });
+            }
+        }
+
+        let pos = self.frozen_layers.iter().rev().position(pred);
+        pos.map(|rev_idx| {
+            let idx = self.frozen_layers.len() - 1 - rev_idx;
+            InMemoryLayerHandle::Frozen {
+                idx,
+                lsn_floor: self.frozen_layers[idx].get_lsn_range().start,
+                end_lsn: self.frozen_layers[idx].get_lsn_range().end,
+            }
+        })
+    }
+
+    /// Get the layer pointed to by the provided handle.
+    pub fn get_in_memory_layer(&self, handle: &InMemoryLayerHandle) -> Option<Arc<InMemoryLayer>> {
+        match handle {
+            InMemoryLayerHandle::Open { .. } => self.open_layer.clone(),
+            InMemoryLayerHandle::Frozen { idx, .. } => self.frozen_layers.get(*idx).cloned(),
+        }
+    }
+
     ///
     /// Divide the whole given range of keys into sub-ranges based on the latest
     /// image layer that covers each range at the specified lsn (inclusive).
@@ -869,6 +919,8 @@ impl LayerMap {
 
 #[cfg(test)]
 mod tests {
+    use pageserver_api::keyspace::KeySpace;
+
     use super::*;
 
     #[derive(Clone)]
@@ -895,15 +947,15 @@ mod tests {
 
     fn assert_range_search_result_eq(lhs: RangeSearchResult, rhs: RangeSearchResult) {
         assert_eq!(lhs.not_found.to_keyspace(), rhs.not_found.to_keyspace());
-        let lhs: Vec<_> = lhs
+        let lhs: HashMap<SearchResult, KeySpace> = lhs
             .found
             .into_iter()
-            .map(|(search_result, accum)| (search_result.0, accum.to_keyspace()))
+            .map(|(search_result, accum)| (search_result, accum.to_keyspace()))
             .collect();
-        let rhs: Vec<_> = rhs
+        let rhs: HashMap<SearchResult, KeySpace> = rhs
             .found
             .into_iter()
-            .map(|(search_result, accum)| (search_result.0, accum.to_keyspace()))
+            .map(|(search_result, accum)| (search_result, accum.to_keyspace()))
             .collect();
 
         assert_eq!(lhs, rhs);
@@ -923,7 +975,7 @@ mod tests {
                 Some(res) => {
                     range_search_result
                         .found
-                        .entry(OrderedSearchResult(res))
+                        .entry(res)
                         .or_default()
                         .add_key(key);
                 }

pageserver/src/tenant/metadata.rs

@@ -8,20 +8,11 @@
 //!
 //! [`remote_timeline_client`]: super::remote_timeline_client
 
-use std::io::{self};
-
-use anyhow::{ensure, Context};
-use pageserver_api::shard::TenantShardId;
+use anyhow::ensure;
 use serde::{de::Error, Deserialize, Serialize, Serializer};
-use thiserror::Error;
 use utils::bin_ser::SerializeError;
-use utils::crashsafe::path_with_suffix_extension;
 use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn};
 
-use crate::config::PageServerConf;
-use crate::virtual_file::VirtualFile;
-use crate::TEMP_FILE_SUFFIX;
-
 /// Use special format number to enable backward compatibility.
 const METADATA_FORMAT_VERSION: u16 = 4;
 
@@ -268,43 +259,6 @@ impl Serialize for TimelineMetadata {
     }
 }
 
-/// Save timeline metadata to file
-#[tracing::instrument(skip_all, fields(%tenant_id=tenant_shard_id.tenant_id, %shard_id=tenant_shard_id.shard_slug(), %timeline_id))]
-pub async fn save_metadata(
-    conf: &'static PageServerConf,
-    tenant_shard_id: &TenantShardId,
-    timeline_id: &TimelineId,
-    data: &TimelineMetadata,
-) -> anyhow::Result<()> {
-    let path = conf.metadata_path(tenant_shard_id, timeline_id);
-    let temp_path = path_with_suffix_extension(&path, TEMP_FILE_SUFFIX);
-    let metadata_bytes = data.to_bytes().context("serialize metadata")?;
-    VirtualFile::crashsafe_overwrite(&path, &temp_path, metadata_bytes)
-        .await
-        .context("write metadata")?;
-    Ok(())
-}
-
-#[derive(Error, Debug)]
-pub enum LoadMetadataError {
-    #[error(transparent)]
-    Read(#[from] io::Error),
-
-    #[error(transparent)]
-    Decode(#[from] anyhow::Error),
-}
-
-pub fn load_metadata(
-    conf: &'static PageServerConf,
-    tenant_shard_id: &TenantShardId,
-    timeline_id: &TimelineId,
-) -> Result<TimelineMetadata, LoadMetadataError> {
-    let metadata_path = conf.metadata_path(tenant_shard_id, timeline_id);
-    let metadata_bytes = std::fs::read(metadata_path)?;
-
-    Ok(TimelineMetadata::from_bytes(&metadata_bytes)?)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;

pageserver/src/tenant/mgr.rs

@@ -2,6 +2,7 @@
 //! page server.
 
 use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
+use futures::stream::StreamExt;
 use itertools::Itertools;
 use pageserver_api::key::Key;
 use pageserver_api::models::ShardParameters;
@@ -31,6 +32,7 @@ use crate::control_plane_client::{
     ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
 };
 use crate::deletion_queue::DeletionQueueClient;
+use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
 use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
 use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::{
@@ -40,7 +42,7 @@ use crate::tenant::config::{
 use crate::tenant::delete::DeleteTenantFlow;
 use crate::tenant::span::debug_assert_current_span_has_tenant_id;
 use crate::tenant::{AttachedTenantConf, SpawnMode, Tenant, TenantState};
-use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, TEMP_FILE_SUFFIX};
+use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TEMP_FILE_SUFFIX};
 
 use utils::crashsafe::path_with_suffix_extension;
 use utils::fs_ext::PathExt;
@@ -357,12 +359,6 @@ fn load_tenant_config(
         return Ok(None);
     }
 
-    let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
-    if tenant_ignore_mark_file.exists() {
-        info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
-        return Ok(None);
-    }
-
     let tenant_shard_id = match tenant_dir_path
         .file_name()
         .unwrap_or_default()
@@ -375,6 +371,59 @@ fn load_tenant_config(
         }
     };
 
+    // Clean up legacy `metadata` files.
+    // Doing it here because every single tenant directory is visited here.
+    // In any later code, there's different treatment of tenant dirs
+    // ... depending on whether the tenant is in re-attach response or not
+    // ... epending on whether the tenant is ignored or not
+    assert_eq!(
+        &conf.tenant_path(&tenant_shard_id),
+        &tenant_dir_path,
+        "later use of conf....path() methods would be dubious"
+    );
+    let timelines: Vec<TimelineId> = match conf.timelines_path(&tenant_shard_id).read_dir_utf8() {
+        Ok(iter) => {
+            let mut timelines = Vec::new();
+            for res in iter {
+                let p = res?;
+                let Some(timeline_id) = p.file_name().parse::<TimelineId>().ok() else {
+                    // skip any entries that aren't TimelineId, such as
+                    // - *.___temp dirs
+                    // - unfinished initdb uploads (test_non_uploaded_root_timeline_is_deleted_after_restart)
+                    continue;
+                };
+                timelines.push(timeline_id);
+            }
+            timelines
+        }
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => vec![],
+        Err(e) => return Err(anyhow::anyhow!(e)),
+    };
+    for timeline_id in timelines {
+        let timeline_path = &conf.timeline_path(&tenant_shard_id, &timeline_id);
+        let metadata_path = timeline_path.join(METADATA_FILE_NAME);
+        match std::fs::remove_file(&metadata_path) {
+            Ok(()) => {
+                crashsafe::fsync(timeline_path)
+                    .context("fsync timeline dir after removing legacy metadata file")?;
+                info!("removed legacy metadata file at {metadata_path}");
+            }
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
+                // something removed the file earlier, or it was never there
+                // We don't care, this software version doesn't write it again, so, we're good.
+            }
+            Err(e) => {
+                anyhow::bail!("remove legacy metadata file: {e}: {metadata_path}");
+            }
+        }
+    }
+
+    let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
+    if tenant_ignore_mark_file.exists() {
+        info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
+        return Ok(None);
+    }
+
     Ok(Some((
         tenant_shard_id,
         Tenant::load_tenant_config(conf, &tenant_shard_id),
@@ -483,7 +532,7 @@ pub async fn init_tenant_mgr(
                             TenantSlot::Secondary(SecondaryTenant::new(
                                 tenant_shard_id,
                                 location_conf.shard,
-                                location_conf.tenant_conf,
+                                location_conf.tenant_conf.clone(),
                                 &SecondaryLocationConfig { warm: false },
                             )),
                         );
@@ -793,7 +842,7 @@ pub(crate) async fn set_new_tenant_config(
     info!("configuring tenant {tenant_id}");
     let tenant = get_tenant(tenant_shard_id, true)?;
 
-    if tenant.tenant_shard_id().shard_count > ShardCount(0) {
+    if !tenant.tenant_shard_id().shard_count.is_unsharded() {
         // Note that we use ShardParameters::default below.
         return Err(SetNewTenantConfigError::Other(anyhow::anyhow!(
             "This API may only be used on single-sharded tenants, use the /location_config API for sharded tenants"
@@ -804,7 +853,7 @@ pub(crate) async fn set_new_tenant_config(
     // API to use is the location_config/ endpoint, which lets the caller provide
     // the full LocationConf.
     let location_conf = LocationConf::attached_single(
-        new_tenant_conf,
+        new_tenant_conf.clone(),
         tenant.generation,
         &ShardParameters::default(),
     );
@@ -1375,7 +1424,7 @@ impl TenantManager {
         result
     }
 
-    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), new_shard_count=%new_shard_count.0))]
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), new_shard_count=%new_shard_count.literal()))]
     pub(crate) async fn shard_split(
         &self,
         tenant_shard_id: TenantShardId,
@@ -1385,11 +1434,10 @@ impl TenantManager {
         let tenant = get_tenant(tenant_shard_id, true)?;
 
         // Plan: identify what the new child shards will be
-        let effective_old_shard_count = std::cmp::max(tenant_shard_id.shard_count.0, 1);
-        if new_shard_count <= ShardCount(effective_old_shard_count) {
+        if new_shard_count.count() <= tenant_shard_id.shard_count.count() {
             anyhow::bail!("Requested shard count is not an increase");
         }
-        let expansion_factor = new_shard_count.0 / effective_old_shard_count;
+        let expansion_factor = new_shard_count.count() / tenant_shard_id.shard_count.count();
         if !expansion_factor.is_power_of_two() {
             anyhow::bail!("Requested split is not a power of two");
         }
@@ -1439,8 +1487,10 @@ impl TenantManager {
             }
         };
 
-        // TODO: hardlink layers from the parent into the child shard directories so that they don't immediately re-download
-        // TODO: erase the dentries from the parent
+        // Optimization: hardlink layers from the parent into the children, so that they don't have to
+        // re-download & duplicate the data referenced in their initial IndexPart
+        self.shard_split_hardlink(parent, child_shards.clone())
+            .await?;
 
         // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
         // child shards to reach this point.
@@ -1464,7 +1514,7 @@ impl TenantManager {
                     attach_mode: AttachmentMode::Single,
                 }),
                 shard: child_shard_identity,
-                tenant_conf: parent_tenant_conf,
+                tenant_conf: parent_tenant_conf.clone(),
             };
 
             self.upsert_location(
@@ -1479,13 +1529,24 @@ impl TenantManager {
 
         // Phase 4: wait for child chards WAL ingest to catch up to target LSN
         for child_shard_id in &child_shards {
+            let child_shard_id = *child_shard_id;
             let child_shard = {
                 let locked = TENANTS.read().unwrap();
                 let peek_slot =
-                    tenant_map_peek_slot(&locked, child_shard_id, TenantSlotPeekMode::Read)?;
+                    tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
                 peek_slot.and_then(|s| s.get_attached()).cloned()
             };
             if let Some(t) = child_shard {
+                // Wait for the child shard to become active: this should be very quick because it only
+                // has to download the index_part that we just uploaded when creating it.
+                if let Err(e) = t.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await {
+                    // This is not fatal: we have durably created the child shard.  It just makes the
+                    // split operation less seamless for clients, as we will may detach the parent
+                    // shard before the child shards are fully ready to serve requests.
+                    tracing::warn!("Failed to wait for shard {child_shard_id} to activate: {e}");
+                    continue;
+                }
+
                 let timelines = t.timelines.lock().unwrap().clone();
                 for timeline in timelines.values() {
                     let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
@@ -1517,7 +1578,7 @@ impl TenantManager {
             }
         }
 
-        // Phase 5: Shut down the parent shard.
+        // Phase 5: Shut down the parent shard, and erase it from disk
         let (_guard, progress) = completion::channel();
         match parent.shutdown(progress, false).await {
             Ok(()) => {}
@@ -1525,6 +1586,24 @@ impl TenantManager {
                 other.wait().await;
             }
         }
+        let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
+        let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
+            .await
+            .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
+        task_mgr::spawn(
+            task_mgr::BACKGROUND_RUNTIME.handle(),
+            TaskKind::MgmtRequest,
+            None,
+            None,
+            "tenant_files_delete",
+            false,
+            async move {
+                fs::remove_dir_all(tmp_path.as_path())
+                    .await
+                    .with_context(|| format!("tenant directory {:?} deletion", tmp_path))
+            },
+        );
+
         parent_slot_guard.drop_old_value()?;
 
         // Phase 6: Release the InProgress on the parent shard
@@ -1532,6 +1611,130 @@ impl TenantManager {
 
         Ok(child_shards)
     }
+
+    /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
+    /// to avoid the children downloading them again.
+    ///
+    /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
+    async fn shard_split_hardlink(
+        &self,
+        parent_shard: &Tenant,
+        child_shards: Vec<TenantShardId>,
+    ) -> anyhow::Result<()> {
+        debug_assert_current_span_has_tenant_id();
+
+        let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
+        let (parent_timelines, parent_layers) = {
+            let mut parent_layers = Vec::new();
+            let timelines = parent_shard.timelines.lock().unwrap().clone();
+            let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
+            for timeline in timelines.values() {
+                let timeline_layers = timeline
+                    .layers
+                    .read()
+                    .await
+                    .resident_layers()
+                    .collect::<Vec<_>>()
+                    .await;
+                for layer in timeline_layers {
+                    let relative_path = layer
+                        .local_path()
+                        .strip_prefix(&parent_path)
+                        .context("Removing prefix from parent layer path")?;
+                    parent_layers.push(relative_path.to_owned());
+                }
+            }
+            debug_assert!(
+                !parent_layers.is_empty(),
+                "shutdown cannot empty the layermap"
+            );
+            (parent_timelines, parent_layers)
+        };
+
+        let mut child_prefixes = Vec::new();
+        let mut create_dirs = Vec::new();
+
+        for child in child_shards {
+            let child_prefix = self.conf.tenant_path(&child);
+            create_dirs.push(child_prefix.clone());
+            create_dirs.extend(
+                parent_timelines
+                    .iter()
+                    .map(|t| self.conf.timeline_path(&child, t)),
+            );
+
+            child_prefixes.push(child_prefix);
+        }
+
+        // Since we will do a large number of small filesystem metadata operations, batch them into
+        // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
+        let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
+            for dir in &create_dirs {
+                if let Err(e) = std::fs::create_dir_all(dir) {
+                    // Ignore AlreadyExists errors, drop out on all other errors
+                    match e.kind() {
+                        std::io::ErrorKind::AlreadyExists => {}
+                        _ => {
+                            return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
+                        }
+                    }
+                }
+            }
+
+            for child_prefix in child_prefixes {
+                for relative_layer in &parent_layers {
+                    let parent_path = parent_path.join(relative_layer);
+                    let child_path = child_prefix.join(relative_layer);
+                    if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
+                        match e.kind() {
+                            std::io::ErrorKind::AlreadyExists => {}
+                            std::io::ErrorKind::NotFound => {
+                                tracing::info!(
+                                    "Layer {} not found during hard-linking, evicted during split?",
+                                    relative_layer
+                                );
+                            }
+                            _ => {
+                                return Err(anyhow::anyhow!(e).context(format!(
+                                    "Hard linking {relative_layer} into {child_prefix}"
+                                )))
+                            }
+                        }
+                    }
+                }
+            }
+
+            // Durability is not required for correctness, but if we crashed during split and
+            // then came restarted with empty timeline dirs, it would be very inefficient to
+            // re-populate from remote storage.
+            for dir in create_dirs {
+                if let Err(e) = crashsafe::fsync(&dir) {
+                    // Something removed a newly created timeline dir out from underneath us?  Extremely
+                    // unexpected, but not worth panic'ing over as this whole function is just an
+                    // optimization.
+                    tracing::warn!("Failed to fsync directory {dir}: {e}")
+                }
+            }
+
+            Ok(parent_layers.len())
+        });
+
+        match jh.await {
+            Ok(Ok(layer_count)) => {
+                tracing::info!(count = layer_count, "Hard linked layers into child shards");
+            }
+            Ok(Err(e)) => {
+                // This is an optimization, so we tolerate failure.
+                tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
+            }
+            Err(e) => {
+                // This is something totally unexpected like a panic, so bail out.
+                anyhow::bail!("Error joining hard linking task: {e}");
+            }
+        }
+
+        Ok(())
+    }
 }
 
 #[derive(Debug, thiserror::Error)]

pageserver/src/tenant/remote_timeline_client.rs

@@ -196,14 +196,12 @@ pub(crate) use upload::upload_initdb_dir;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
-use utils::timeout::{timeout_cancellable, TimeoutCancellableError};
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
-use std::time::Duration;
 
-use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
+use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath, TimeoutOrCancel};
 use std::ops::DerefMut;
 use tracing::{debug, error, info, instrument, warn};
 use tracing::{info_span, Instrument};
@@ -263,11 +261,6 @@ pub(crate) const INITDB_PRESERVED_PATH: &str = "initdb-preserved.tar.zst";
 /// Default buffer size when interfacing with [`tokio::fs::File`].
 pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
 
-/// This timeout is intended to deal with hangs in lower layers, e.g. stuck TCP flows.  It is not
-/// intended to be snappy enough for prompt shutdown, as we have a CancellationToken for that.
-pub(crate) const UPLOAD_TIMEOUT: Duration = Duration::from_secs(120);
-pub(crate) const DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(120);
-
 pub enum MaybeDeletedIndexPart {
     IndexPart(IndexPart),
     Deleted(IndexPart),
@@ -331,40 +324,6 @@ pub struct RemoteTimelineClient {
     cancel: CancellationToken,
 }
 
-/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to anyhow.
-///
-/// This is a convenience for the various upload functions.  In future
-/// the anyhow::Error result should be replaced with a more structured type that
-/// enables callers to avoid handling shutdown as an error.
-async fn upload_cancellable<F>(cancel: &CancellationToken, future: F) -> anyhow::Result<()>
-where
-    F: std::future::Future<Output = anyhow::Result<()>>,
-{
-    match timeout_cancellable(UPLOAD_TIMEOUT, cancel, future).await {
-        Ok(Ok(())) => Ok(()),
-        Ok(Err(e)) => Err(e),
-        Err(TimeoutCancellableError::Timeout) => Err(anyhow::anyhow!("Timeout")),
-        Err(TimeoutCancellableError::Cancelled) => Err(anyhow::anyhow!("Shutting down")),
-    }
-}
-/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to DownloaDError.
-async fn download_cancellable<F, R>(
-    cancel: &CancellationToken,
-    future: F,
-) -> Result<R, DownloadError>
-where
-    F: std::future::Future<Output = Result<R, DownloadError>>,
-{
-    match timeout_cancellable(DOWNLOAD_TIMEOUT, cancel, future).await {
-        Ok(Ok(r)) => Ok(r),
-        Ok(Err(e)) => Err(e),
-        Err(TimeoutCancellableError::Timeout) => {
-            Err(DownloadError::Other(anyhow::anyhow!("Timed out")))
-        }
-        Err(TimeoutCancellableError::Cancelled) => Err(DownloadError::Cancelled),
-    }
-}
-
 impl RemoteTimelineClient {
     ///
     /// Create a remote storage client for given timeline
@@ -655,7 +614,7 @@ impl RemoteTimelineClient {
             metadata,
         );
         let op = UploadOp::UploadMetadata(index_part, disk_consistent_lsn);
-        self.calls_unfinished_metric_begin(&op);
+        self.metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
         upload_queue.latest_files_changes_since_metadata_upload_scheduled = 0;
 
@@ -695,7 +654,7 @@ impl RemoteTimelineClient {
             metadata.generation, metadata.shard
         );
         let op = UploadOp::UploadLayer(layer, metadata);
-        self.calls_unfinished_metric_begin(&op);
+        self.metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
     }
 
@@ -864,10 +823,14 @@ impl RemoteTimelineClient {
         }
 
         // schedule the actual deletions
+        if with_metadata.is_empty() {
+            // avoid scheduling the op & bumping the metric
+            return;
+        }
         let op = UploadOp::Delete(Delete {
             layers: with_metadata,
         });
-        self.calls_unfinished_metric_begin(&op);
+        self.metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
     }
 
@@ -1050,7 +1013,7 @@ impl RemoteTimelineClient {
             &self.cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)?;
 
         // all good, disarm the guard and mark as success
@@ -1082,14 +1045,14 @@ impl RemoteTimelineClient {
                 upload::preserve_initdb_archive(&self.storage_impl, tenant_id, timeline_id, cancel)
                     .await
             },
-            |_e| false,
+            TimeoutOrCancel::caused_by_cancel,
             FAILED_DOWNLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "preserve_initdb_tar_zst",
             &cancel.clone(),
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancellled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)
         .context("backing up initdb archive")?;
         Ok(())
@@ -1151,7 +1114,7 @@ impl RemoteTimelineClient {
         let remaining = download_retry(
             || async {
                 self.storage_impl
-                    .list_files(Some(&timeline_storage_path), None)
+                    .list_files(Some(&timeline_storage_path), None, &cancel)
                     .await
             },
             "list remaining files",
@@ -1445,6 +1408,10 @@ impl RemoteTimelineClient {
                 Ok(()) => {
                     break;
                 }
+                Err(e) if TimeoutOrCancel::caused_by_cancel(&e) => {
+                    // loop around to do the proper stopping
+                    continue;
+                }
                 Err(e) => {
                     let retries = task.retries.fetch_add(1, Ordering::SeqCst);
 
@@ -1553,10 +1520,10 @@ impl RemoteTimelineClient {
                 .await;
         }
 
-        self.calls_unfinished_metric_end(&task.op);
+        self.metric_end(&task.op);
     }
 
-    fn calls_unfinished_metric_impl(
+    fn metric_impl(
         &self,
         op: &UploadOp,
     ) -> Option<(
@@ -1593,17 +1560,17 @@ impl RemoteTimelineClient {
         Some(res)
     }
 
-    fn calls_unfinished_metric_begin(&self, op: &UploadOp) {
-        let (file_kind, op_kind, track_bytes) = match self.calls_unfinished_metric_impl(op) {
+    fn metric_begin(&self, op: &UploadOp) {
+        let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
             Some(x) => x,
             None => return,
         };
         let guard = self.metrics.call_begin(&file_kind, &op_kind, track_bytes);
-        guard.will_decrement_manually(); // in unfinished_ops_metric_end()
+        guard.will_decrement_manually(); // in metric_end(), see right below
     }
 
-    fn calls_unfinished_metric_end(&self, op: &UploadOp) {
-        let (file_kind, op_kind, track_bytes) = match self.calls_unfinished_metric_impl(op) {
+    fn metric_end(&self, op: &UploadOp) {
+        let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
             Some(x) => x,
             None => return,
         };
@@ -1688,7 +1655,7 @@ impl RemoteTimelineClient {
 
                 // Tear down queued ops
                 for op in qi.queued_operations.into_iter() {
-                    self.calls_unfinished_metric_end(&op);
+                    self.metric_end(&op);
                     // Dropping UploadOp::Barrier() here will make wait_completion() return with an Err()
                     // which is exactly what we want to happen.
                     drop(op);

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -11,16 +11,14 @@ use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::shard::TenantShardId;
 use tokio::fs::{self, File, OpenOptions};
 use tokio::io::{AsyncSeekExt, AsyncWriteExt};
+use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
-use utils::timeout::timeout_cancellable;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::remote_timeline_client::{
-    download_cancellable, remote_layer_path, remote_timelines_path, DOWNLOAD_TIMEOUT,
-};
+use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::Generation;
 use crate::virtual_file::on_fatal_io_error;
@@ -83,60 +81,28 @@ pub async fn download_layer_file<'a>(
                 .with_context(|| format!("create a destination file for layer '{temp_file_path}'"))
                 .map_err(DownloadError::Other)?;
 
-            // Cancellation safety: it is safe to cancel this future, because it isn't writing to a local
-            // file: the write to local file doesn't start until after the request header is returned
-            // and we start draining the body stream below
-            let download = download_cancellable(cancel, storage.download(&remote_path))
-                .await
-                .with_context(|| {
-                    format!(
-                    "open a download stream for layer with remote storage path '{remote_path:?}'"
-                )
-                })
-                .map_err(DownloadError::Other)?;
+            let download = storage.download(&remote_path, cancel).await?;
 
             let mut destination_file =
                 tokio::io::BufWriter::with_capacity(super::BUFFER_SIZE, destination_file);
 
             let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
 
-            // Cancellation safety: it is safe to cancel this future because it is writing into a temporary file,
-            // and we will unlink the temporary file if there is an error.  This unlink is important because we
-            // are in a retry loop, and we wouldn't want to leave behind a rogue write I/O to a file that
-            // we will imminiently try and write to again.
-            let bytes_amount: u64 = match timeout_cancellable(
-                DOWNLOAD_TIMEOUT,
-                cancel,
-                tokio::io::copy_buf(&mut reader, &mut destination_file),
-            )
-            .await
-            .with_context(|| {
-                format!(
-                    "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
-                )
-            })
-            .map_err(DownloadError::Other)?
-            {
-                Ok(b) => Ok(b),
+            let bytes_amount = tokio::io::copy_buf(&mut reader, &mut destination_file).await;
+
+            match bytes_amount {
+                Ok(bytes_amount) => {
+                    let destination_file = destination_file.into_inner();
+                    Ok((destination_file, bytes_amount))
+                }
                 Err(e) => {
-                    // Remove incomplete files: on restart Timeline would do this anyway, but we must
-                    // do it here for the retry case.
                     if let Err(e) = tokio::fs::remove_file(&temp_file_path).await {
                         on_fatal_io_error(&e, &format!("Removing temporary file {temp_file_path}"));
                     }
-                    Err(e)
+
+                    Err(e.into())
                 }
             }
-            .with_context(|| {
-                format!(
-                    "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
-                )
-            })
-            .map_err(DownloadError::Other)?;
-
-            let destination_file = destination_file.into_inner();
-
-            Ok((destination_file, bytes_amount))
         },
         &format!("download {remote_path:?}"),
         cancel,
@@ -218,9 +184,11 @@ pub async fn list_remote_timelines(
 
     let listing = download_retry_forever(
         || {
-            download_cancellable(
+            storage.list(
+                Some(&remote_path),
+                ListingMode::WithDelimiter,
+                None,
                 &cancel,
-                storage.list(Some(&remote_path), ListingMode::WithDelimiter, None),
             )
         },
         &format!("list timelines for {tenant_shard_id}"),
@@ -259,26 +227,20 @@ async fn do_download_index_part(
     index_generation: Generation,
     cancel: &CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
-    use futures::stream::StreamExt;
-
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
 
     let index_part_bytes = download_retry_forever(
         || async {
-            // Cancellation: if is safe to cancel this future because we're just downloading into
-            // a memory buffer, not touching local disk.
-            let index_part_download =
-                download_cancellable(cancel, storage.download(&remote_path)).await?;
+            let download = storage.download(&remote_path, cancel).await?;
 
-            let mut index_part_bytes = Vec::new();
-            let mut stream = std::pin::pin!(index_part_download.download_stream);
-            while let Some(chunk) = stream.next().await {
-                let chunk = chunk
-                    .with_context(|| format!("download index part at {remote_path:?}"))
-                    .map_err(DownloadError::Other)?;
-                index_part_bytes.extend_from_slice(&chunk[..]);
-            }
-            Ok(index_part_bytes)
+            let mut bytes = Vec::new();
+
+            let stream = download.download_stream;
+            let mut stream = StreamReader::new(stream);
+
+            tokio::io::copy_buf(&mut stream, &mut bytes).await?;
+
+            Ok(bytes)
         },
         &format!("download {remote_path:?}"),
         cancel,
@@ -373,7 +335,7 @@ pub(super) async fn download_index_part(
     let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
 
     let indices = download_retry(
-        || async { storage.list_files(Some(&index_prefix), None).await },
+        || async { storage.list_files(Some(&index_prefix), None, cancel).await },
         "list index_part files",
         cancel,
     )
@@ -446,24 +408,17 @@ pub(crate) async fn download_initdb_tar_zst(
                 .with_context(|| format!("tempfile creation {temp_path}"))
                 .map_err(DownloadError::Other)?;
 
-            let download = match download_cancellable(cancel, storage.download(&remote_path)).await
-            {
+            let download = match storage.download(&remote_path, cancel).await {
                 Ok(dl) => dl,
                 Err(DownloadError::NotFound) => {
-                    download_cancellable(cancel, storage.download(&remote_preserved_path)).await?
+                    storage.download(&remote_preserved_path, cancel).await?
                 }
                 Err(other) => Err(other)?,
             };
             let mut download = tokio_util::io::StreamReader::new(download.download_stream);
             let mut writer = tokio::io::BufWriter::with_capacity(super::BUFFER_SIZE, file);
 
-            // TODO: this consumption of the response body should be subject to timeout + cancellation, but
-            // not without thinking carefully about how to recover safely from cancelling a write to
-            // local storage (e.g. by writing into a temp file as we do in download_layer)
-            tokio::io::copy_buf(&mut download, &mut writer)
-                .await
-                .with_context(|| format!("download initdb.tar.zst at {remote_path:?}"))
-                .map_err(DownloadError::Other)?;
+            tokio::io::copy_buf(&mut download, &mut writer).await?;
 
             let mut file = writer.into_inner();
 

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -16,7 +16,7 @@ use crate::{
     config::PageServerConf,
     tenant::remote_timeline_client::{
         index::IndexPart, remote_index_path, remote_initdb_archive_path,
-        remote_initdb_preserved_archive_path, remote_path, upload_cancellable,
+        remote_initdb_preserved_archive_path, remote_path,
     },
 };
 use remote_storage::{GenericRemoteStorage, TimeTravelError};
@@ -49,16 +49,15 @@ pub(crate) async fn upload_index_part<'a>(
     let index_part_bytes = bytes::Bytes::from(index_part_bytes);
 
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, generation);
-    upload_cancellable(
-        cancel,
-        storage.upload_storage_object(
+    storage
+        .upload_storage_object(
             futures::stream::once(futures::future::ready(Ok(index_part_bytes))),
             index_part_size,
             &remote_path,
-        ),
-    )
-    .await
-    .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
+            cancel,
+        )
+        .await
+        .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
 }
 
 /// Attempts to upload given layer files.
@@ -115,11 +114,10 @@ pub(super) async fn upload_timeline_layer<'a>(
 
     let reader = tokio_util::io::ReaderStream::with_capacity(source_file, super::BUFFER_SIZE);
 
-    upload_cancellable(cancel, storage.upload(reader, fs_size, &storage_path, None))
+    storage
+        .upload(reader, fs_size, &storage_path, None, cancel)
         .await
-        .with_context(|| format!("upload layer from local path '{source_path}'"))?;
-
-    Ok(())
+        .with_context(|| format!("upload layer from local path '{source_path}'"))
 }
 
 /// Uploads the given `initdb` data to the remote storage.
@@ -139,12 +137,10 @@ pub(crate) async fn upload_initdb_dir(
     let file = tokio_util::io::ReaderStream::with_capacity(initdb_tar_zst, super::BUFFER_SIZE);
 
     let remote_path = remote_initdb_archive_path(tenant_id, timeline_id);
-    upload_cancellable(
-        cancel,
-        storage.upload_storage_object(file, size as usize, &remote_path),
-    )
-    .await
-    .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
+    storage
+        .upload_storage_object(file, size as usize, &remote_path, cancel)
+        .await
+        .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
 }
 
 pub(crate) async fn preserve_initdb_archive(
@@ -155,7 +151,8 @@ pub(crate) async fn preserve_initdb_archive(
 ) -> anyhow::Result<()> {
     let source_path = remote_initdb_archive_path(tenant_id, timeline_id);
     let dest_path = remote_initdb_preserved_archive_path(tenant_id, timeline_id);
-    upload_cancellable(cancel, storage.copy_object(&source_path, &dest_path))
+    storage
+        .copy_object(&source_path, &dest_path, cancel)
         .await
         .with_context(|| format!("backing up initdb archive for '{tenant_id} / {timeline_id}'"))
 }

pageserver/src/tenant/secondary.rs

@@ -133,7 +133,7 @@ impl SecondaryTenant {
     }
 
     pub(crate) fn set_tenant_conf(&self, config: &TenantConfOpt) {
-        *(self.tenant_conf.lock().unwrap()) = *config;
+        *(self.tenant_conf.lock().unwrap()) = config.clone();
     }
 
     /// For API access: generate a LocationConfig equivalent to the one that would be used to
@@ -144,13 +144,13 @@ impl SecondaryTenant {
 
         let conf = models::LocationConfigSecondary { warm: conf.warm };
 
-        let tenant_conf = *self.tenant_conf.lock().unwrap();
+        let tenant_conf = self.tenant_conf.lock().unwrap().clone();
         models::LocationConfig {
             mode: models::LocationConfigMode::Secondary,
             generation: None,
             secondary_conf: Some(conf),
             shard_number: self.tenant_shard_id.shard_number.0,
-            shard_count: self.tenant_shard_id.shard_count.0,
+            shard_count: self.tenant_shard_id.shard_count.literal(),
             shard_stripe_size: self.shard_identity.stripe_size.0,
             tenant_conf: tenant_conf.into(),
         }

pageserver/src/tenant/secondary/downloader.rs

@@ -37,6 +37,7 @@ use crate::tenant::{
     remote_timeline_client::{download::download_layer_file, remote_heatmap_path},
 };
 
+use camino::Utf8PathBuf;
 use chrono::format::{DelayedFormat, StrftimeItems};
 use futures::Future;
 use pageserver_api::shard::TenantShardId;
@@ -44,7 +45,7 @@ use rand::Rng;
 use remote_storage::{DownloadError, GenericRemoteStorage};
 
 use tokio_util::sync::CancellationToken;
-use tracing::{info_span, instrument, Instrument};
+use tracing::{info_span, instrument, warn, Instrument};
 use utils::{
     backoff, completion::Barrier, crashsafe::path_with_suffix_extension, fs_ext, id::TimelineId,
 };
@@ -438,8 +439,14 @@ impl From<std::io::Error> for UpdateError {
     fn from(value: std::io::Error) -> Self {
         if let Some(nix::errno::Errno::ENOSPC) = value.raw_os_error().map(nix::errno::from_i32) {
             UpdateError::NoSpace
+        } else if value
+            .get_ref()
+            .and_then(|x| x.downcast_ref::<DownloadError>())
+            .is_some()
+        {
+            UpdateError::from(DownloadError::from(value))
         } else {
-            // An I/O error from e.g. tokio::io::copy is most likely a remote storage issue
+            // An I/O error from e.g. tokio::io::copy_buf is most likely a remote storage issue
             UpdateError::Other(anyhow::anyhow!(value))
         }
     }
@@ -523,12 +530,13 @@ impl<'a> TenantDownloader<'a> {
         tracing::debug!("Downloading heatmap for secondary tenant",);
 
         let heatmap_path = remote_heatmap_path(tenant_shard_id);
+        let cancel = &self.secondary_state.cancel;
 
         let heatmap_bytes = backoff::retry(
             || async {
                 let download = self
                     .remote_storage
-                    .download(&heatmap_path)
+                    .download(&heatmap_path, cancel)
                     .await
                     .map_err(UpdateError::from)?;
                 let mut heatmap_bytes = Vec::new();
@@ -540,7 +548,7 @@ impl<'a> TenantDownloader<'a> {
             FAILED_DOWNLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "download heatmap",
-            &self.secondary_state.cancel,
+            cancel,
         )
         .await
         .ok_or_else(|| UpdateError::Cancelled)
@@ -671,20 +679,17 @@ impl<'a> TenantDownloader<'a> {
             .await
             {
                 Ok(bytes) => bytes,
-                Err(e) => {
-                    if let DownloadError::NotFound = e {
-                        // A heatmap might be out of date and refer to a layer that doesn't exist any more.
-                        // This is harmless: continue to download the next layer. It is expected during compaction
-                        // GC.
-                        tracing::debug!(
-                            "Skipped downloading missing layer {}, raced with compaction/gc?",
-                            layer.name
-                        );
-                        continue;
-                    } else {
-                        return Err(e.into());
-                    }
+                Err(DownloadError::NotFound) => {
+                    // A heatmap might be out of date and refer to a layer that doesn't exist any more.
+                    // This is harmless: continue to download the next layer. It is expected during compaction
+                    // GC.
+                    tracing::debug!(
+                        "Skipped downloading missing layer {}, raced with compaction/gc?",
+                        layer.name
+                    );
+                    continue;
                 }
+                Err(e) => return Err(e.into()),
             };
 
             if downloaded_bytes != layer.metadata.file_size {
@@ -774,19 +779,33 @@ async fn init_timeline_state(
         .await
         .fatal_err(&format!("Listing {timeline_path}"))
     {
-        let dentry_file_name = dentry.file_name();
-        let file_name = dentry_file_name.to_string_lossy();
-        let local_meta = dentry.metadata().await.fatal_err(&format!(
-            "Read metadata on {}",
-            dentry.path().to_string_lossy()
-        ));
+        let Ok(file_path) = Utf8PathBuf::from_path_buf(dentry.path()) else {
+            tracing::warn!("Malformed filename at {}", dentry.path().to_string_lossy());
+            continue;
+        };
+        let local_meta = dentry
+            .metadata()
+            .await
+            .fatal_err(&format!("Read metadata on {}", file_path));
 
-        // Secondary mode doesn't use local metadata files, but they might have been left behind by an attached tenant.
+        let file_name = file_path.file_name().expect("created it from the dentry");
         if file_name == METADATA_FILE_NAME {
+            // Secondary mode doesn't use local metadata files, but they might have been left behind by an attached tenant.
+            warn!(path=?dentry.path(), "found legacy metadata file, these should have been removed in load_tenant_config");
+            continue;
+        } else if crate::is_temporary(&file_path) {
+            // Temporary files are frequently left behind from restarting during downloads
+            tracing::info!("Cleaning up temporary file {file_path}");
+            if let Err(e) = tokio::fs::remove_file(&file_path)
+                .await
+                .or_else(fs_ext::ignore_not_found)
+            {
+                tracing::error!("Failed to remove temporary file {file_path}: {e}");
+            }
             continue;
         }
 
-        match LayerFileName::from_str(&file_name) {
+        match LayerFileName::from_str(file_name) {
             Ok(name) => {
                 let remote_meta = heatmap_metadata.get(&name);
                 match remote_meta {

pageserver/src/tenant/secondary/heatmap_uploader.rs

@@ -21,18 +21,17 @@ use futures::Future;
 use md5;
 use pageserver_api::shard::TenantShardId;
 use rand::Rng;
-use remote_storage::GenericRemoteStorage;
+use remote_storage::{GenericRemoteStorage, TimeoutOrCancel};
 
 use super::{
+    heatmap::HeatMapTenant,
     scheduler::{self, JobGenerator, RunningJob, SchedulingResult, TenantBackgroundJobs},
-    CommandRequest,
+    CommandRequest, UploadCommand,
 };
 use tokio_util::sync::CancellationToken;
 use tracing::{info_span, instrument, Instrument};
 use utils::{backoff, completion::Barrier, yielding_loop::yielding_loop};
 
-use super::{heatmap::HeatMapTenant, UploadCommand};
-
 pub(super) async fn heatmap_uploader_task(
     tenant_manager: Arc<TenantManager>,
     remote_storage: GenericRemoteStorage,
@@ -417,10 +416,10 @@ async fn upload_tenant_heatmap(
         || async {
             let bytes = futures::stream::once(futures::future::ready(Ok(bytes.clone())));
             remote_storage
-                .upload_storage_object(bytes, size, &path)
+                .upload_storage_object(bytes, size, &path, cancel)
                 .await
         },
-        |_| false,
+        TimeoutOrCancel::caused_by_cancel,
         3,
         u32::MAX,
         "Uploading heatmap",

pageserver/src/tenant/storage_layer.rs

@@ -8,15 +8,21 @@ pub(crate) mod layer;
 mod layer_desc;
 
 use crate::context::{AccessStatsBehavior, RequestContext};
+use crate::repository::Value;
 use crate::task_mgr::TaskKind;
 use crate::walrecord::NeonWalRecord;
 use bytes::Bytes;
 use enum_map::EnumMap;
 use enumset::EnumSet;
 use once_cell::sync::Lazy;
+use pageserver_api::key::Key;
+use pageserver_api::keyspace::{KeySpace, KeySpaceRandomAccum};
 use pageserver_api::models::{
     LayerAccessKind, LayerResidenceEvent, LayerResidenceEventReason, LayerResidenceStatus,
 };
+use std::cmp::{Ordering, Reverse};
+use std::collections::hash_map::Entry;
+use std::collections::{BinaryHeap, HashMap};
 use std::ops::Range;
 use std::sync::Mutex;
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
@@ -34,6 +40,11 @@ pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
 
 pub(crate) use layer::{EvictionError, Layer, ResidentLayer};
 
+use super::layer_map::InMemoryLayerHandle;
+use super::timeline::layer_manager::LayerManager;
+use super::timeline::GetVectoredError;
+use super::PageReconstructError;
+
 pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
 where
     T: PartialOrd<T>,
@@ -67,6 +78,277 @@ pub struct ValueReconstructState {
     pub img: Option<(Lsn, Bytes)>,
 }
 
+#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
+pub(crate) enum ValueReconstructSituation {
+    Complete,
+    #[default]
+    Continue,
+}
+
+/// Reconstruct data accumulated for a single key during a vectored get
+#[derive(Debug, Default, Clone)]
+pub(crate) struct VectoredValueReconstructState {
+    pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
+    pub(crate) img: Option<(Lsn, Bytes)>,
+
+    situation: ValueReconstructSituation,
+}
+
+impl VectoredValueReconstructState {
+    fn get_cached_lsn(&self) -> Option<Lsn> {
+        self.img.as_ref().map(|img| img.0)
+    }
+}
+
+impl From<VectoredValueReconstructState> for ValueReconstructState {
+    fn from(mut state: VectoredValueReconstructState) -> Self {
+        // walredo expects the records to be descending in terms of Lsn
+        state.records.sort_by_key(|(lsn, _)| Reverse(*lsn));
+
+        ValueReconstructState {
+            records: state.records,
+            img: state.img,
+        }
+    }
+}
+
+/// Bag of data accumulated during a vectored get
+pub(crate) struct ValuesReconstructState {
+    pub(crate) keys: HashMap<Key, Result<VectoredValueReconstructState, PageReconstructError>>,
+
+    keys_done: KeySpaceRandomAccum,
+}
+
+impl ValuesReconstructState {
+    pub(crate) fn new() -> Self {
+        Self {
+            keys: HashMap::new(),
+            keys_done: KeySpaceRandomAccum::new(),
+        }
+    }
+
+    /// Associate a key with the error which it encountered and mark it as done
+    pub(crate) fn on_key_error(&mut self, key: Key, err: PageReconstructError) {
+        let previous = self.keys.insert(key, Err(err));
+        if let Some(Ok(state)) = previous {
+            if state.situation == ValueReconstructSituation::Continue {
+                self.keys_done.add_key(key);
+            }
+        }
+    }
+
+    /// Update the state collected for a given key.
+    /// Returns true if this was the last value needed for the key and false otherwise.
+    ///
+    /// If the key is done after the update, mark it as such.
+    pub(crate) fn update_key(
+        &mut self,
+        key: &Key,
+        lsn: Lsn,
+        value: Value,
+    ) -> ValueReconstructSituation {
+        let state = self
+            .keys
+            .entry(*key)
+            .or_insert(Ok(VectoredValueReconstructState::default()));
+
+        if let Ok(state) = state {
+            let key_done = match state.situation {
+                ValueReconstructSituation::Complete => unreachable!(),
+                ValueReconstructSituation::Continue => match value {
+                    Value::Image(img) => {
+                        state.img = Some((lsn, img));
+                        true
+                    }
+                    Value::WalRecord(rec) => {
+                        let reached_cache =
+                            state.get_cached_lsn().map(|clsn| clsn + 1) == Some(lsn);
+                        let will_init = rec.will_init();
+                        state.records.push((lsn, rec));
+                        will_init || reached_cache
+                    }
+                },
+            };
+
+            if key_done && state.situation == ValueReconstructSituation::Continue {
+                state.situation = ValueReconstructSituation::Complete;
+                self.keys_done.add_key(*key);
+            }
+
+            state.situation
+        } else {
+            ValueReconstructSituation::Complete
+        }
+    }
+
+    /// Returns the Lsn at which this key is cached if one exists.
+    /// The read path should go no further than this Lsn for the given key.
+    pub(crate) fn get_cached_lsn(&self, key: &Key) -> Option<Lsn> {
+        self.keys
+            .get(key)
+            .and_then(|k| k.as_ref().ok())
+            .and_then(|state| state.get_cached_lsn())
+    }
+
+    /// Returns the key space describing the keys that have
+    /// been marked as completed since the last call to this function.
+    pub(crate) fn consume_done_keys(&mut self) -> KeySpace {
+        self.keys_done.consume_keyspace()
+    }
+}
+
+impl Default for ValuesReconstructState {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Description of layer to be read - the layer map can turn
+/// this description into the actual layer.
+#[derive(PartialEq, Eq, Hash, Debug, Clone)]
+pub(crate) enum ReadableLayerDesc {
+    Persistent {
+        desc: PersistentLayerDesc,
+        lsn_floor: Lsn,
+        lsn_ceil: Lsn,
+    },
+    InMemory {
+        handle: InMemoryLayerHandle,
+        lsn_ceil: Lsn,
+    },
+}
+
+/// Wraper for 'ReadableLayerDesc' sorted by Lsn
+#[derive(Debug)]
+struct ReadableLayerDescOrdered(ReadableLayerDesc);
+
+/// Data structure which maintains a fringe of layers for the
+/// read path. The fringe is the set of layers which intersects
+/// the current keyspace that the search is descending on.
+/// Each layer tracks the keyspace that intersects it.
+///
+/// The fringe must appear sorted by Lsn. Hence, it uses
+/// a two layer indexing scheme.
+#[derive(Debug)]
+pub(crate) struct LayerFringe {
+    layers_by_lsn: BinaryHeap<ReadableLayerDescOrdered>,
+    layers: HashMap<ReadableLayerDesc, KeySpace>,
+}
+
+impl LayerFringe {
+    pub(crate) fn new() -> Self {
+        LayerFringe {
+            layers_by_lsn: BinaryHeap::new(),
+            layers: HashMap::new(),
+        }
+    }
+
+    pub(crate) fn next_layer(&mut self) -> Option<(ReadableLayerDesc, KeySpace)> {
+        let handle = match self.layers_by_lsn.pop() {
+            Some(h) => h,
+            None => return None,
+        };
+
+        let removed = self.layers.remove_entry(&handle.0);
+        match removed {
+            Some((layer, keyspace)) => Some((layer, keyspace)),
+            None => unreachable!("fringe internals are always consistent"),
+        }
+    }
+
+    pub(crate) fn update(&mut self, layer: ReadableLayerDesc, keyspace: KeySpace) {
+        let entry = self.layers.entry(layer.clone());
+        match entry {
+            Entry::Occupied(mut entry) => {
+                entry.get_mut().merge(&keyspace);
+            }
+            Entry::Vacant(entry) => {
+                self.layers_by_lsn
+                    .push(ReadableLayerDescOrdered(entry.key().clone()));
+                entry.insert(keyspace);
+            }
+        }
+    }
+}
+
+impl Default for LayerFringe {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Ord for ReadableLayerDescOrdered {
+    fn cmp(&self, other: &Self) -> Ordering {
+        let ord = self.0.get_lsn_ceil().cmp(&other.0.get_lsn_ceil());
+        if ord == std::cmp::Ordering::Equal {
+            self.0
+                .get_lsn_floor()
+                .cmp(&other.0.get_lsn_floor())
+                .reverse()
+        } else {
+            ord
+        }
+    }
+}
+
+impl PartialOrd for ReadableLayerDescOrdered {
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl PartialEq for ReadableLayerDescOrdered {
+    fn eq(&self, other: &Self) -> bool {
+        self.0.get_lsn_floor() == other.0.get_lsn_floor()
+            && self.0.get_lsn_ceil() == other.0.get_lsn_ceil()
+    }
+}
+
+impl Eq for ReadableLayerDescOrdered {}
+
+impl ReadableLayerDesc {
+    pub(crate) fn get_lsn_floor(&self) -> Lsn {
+        match self {
+            ReadableLayerDesc::Persistent { lsn_floor, .. } => *lsn_floor,
+            ReadableLayerDesc::InMemory { handle, .. } => handle.get_lsn_floor(),
+        }
+    }
+
+    pub(crate) fn get_lsn_ceil(&self) -> Lsn {
+        match self {
+            ReadableLayerDesc::Persistent { lsn_ceil, .. } => *lsn_ceil,
+            ReadableLayerDesc::InMemory { lsn_ceil, .. } => *lsn_ceil,
+        }
+    }
+
+    pub(crate) async fn get_values_reconstruct_data(
+        &self,
+        layer_manager: &LayerManager,
+        keyspace: KeySpace,
+        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<(), GetVectoredError> {
+        match self {
+            ReadableLayerDesc::Persistent { desc, lsn_ceil, .. } => {
+                let layer = layer_manager.get_from_desc(desc);
+                layer
+                    .get_values_reconstruct_data(keyspace, *lsn_ceil, reconstruct_state, ctx)
+                    .await
+            }
+            ReadableLayerDesc::InMemory { handle, lsn_ceil } => {
+                let layer = layer_manager
+                    .layer_map()
+                    .get_in_memory_layer(handle)
+                    .unwrap();
+
+                layer
+                    .get_values_reconstruct_data(keyspace, *lsn_ceil, reconstruct_state, ctx)
+                    .await
+            }
+        }
+    }
+}
+
 /// Return value from [`Layer::get_value_reconstruct_data`]
 #[derive(Clone, Copy, Debug)]
 pub enum ValueReconstructResult {

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -35,16 +35,19 @@ use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
-use crate::tenant::Timeline;
+use crate::tenant::timeline::GetVectoredError;
+use crate::tenant::{PageReconstructError, Timeline};
 use crate::virtual_file::{self, VirtualFile};
 use crate::{walrecord, TEMP_FILE_SUFFIX};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
-use anyhow::{bail, ensure, Context, Result};
+use anyhow::{anyhow, bail, ensure, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
+use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::LayerAccessKind;
 use pageserver_api::shard::TenantShardId;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
+use std::collections::BTreeMap;
 use std::fs::File;
 use std::io::SeekFrom;
 use std::ops::Range;
@@ -59,7 +62,10 @@ use utils::{
     lsn::Lsn,
 };
 
-use super::{AsLayerDesc, LayerAccessStats, PersistentLayerDesc, ResidentLayer};
+use super::{
+    AsLayerDesc, LayerAccessStats, PersistentLayerDesc, ResidentLayer, ValueReconstructSituation,
+    ValuesReconstructState,
+};
 
 ///
 /// Header stored in the beginning of the file
@@ -818,6 +824,133 @@ impl DeltaLayerInner {
         }
     }
 
+    // Look up the keys in the provided keyspace and update
+    // the reconstruct state with whatever is found.
+    //
+    // If the key is cached, go no further than the cached Lsn.
+    //
+    // Currently, the index is visited for each range, but this
+    // can be further optimised to visit the index only once.
+    pub(super) async fn get_values_reconstruct_data(
+        &self,
+        keyspace: KeySpace,
+        end_lsn: Lsn,
+        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<(), GetVectoredError> {
+        let file = &self.file;
+        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
+            self.index_start_blk,
+            self.index_root_blk,
+            file,
+        );
+
+        let mut offsets: BTreeMap<Key, Vec<(Lsn, u64)>> = BTreeMap::new();
+
+        for range in keyspace.ranges.iter() {
+            let mut ignore_key = None;
+
+            // Scan the page versions backwards, starting from the last key in the range.
+            // to collect all the offsets at which need to be read.
+            let end_key = DeltaKey::from_key_lsn(&range.end, Lsn(end_lsn.0 - 1));
+            tree_reader
+                .visit(
+                    &end_key.0,
+                    VisitDirection::Backwards,
+                    |raw_key, value| {
+                        let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                        let entry_lsn = DeltaKey::extract_lsn_from_buf(raw_key);
+
+                        if entry_lsn >= end_lsn {
+                            return true;
+                        }
+
+                        if key < range.start {
+                            return false;
+                        }
+
+                        if key >= range.end {
+                            return true;
+                        }
+
+                        if Some(key) == ignore_key {
+                            return true;
+                        }
+
+                        if let Some(cached_lsn) = reconstruct_state.get_cached_lsn(&key) {
+                            if entry_lsn <= cached_lsn {
+                                return key != range.start;
+                            }
+                        }
+
+                        let blob_ref = BlobRef(value);
+                        let lsns_at = offsets.entry(key).or_default();
+                        lsns_at.push((entry_lsn, blob_ref.pos()));
+
+                        if blob_ref.will_init() {
+                            if key == range.start {
+                                return false;
+                            } else {
+                                ignore_key = Some(key);
+                                return true;
+                            }
+                        }
+
+                        true
+                    },
+                    &RequestContextBuilder::extend(ctx)
+                        .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
+                        .build(),
+                )
+                .await
+                .map_err(|err| GetVectoredError::Other(anyhow!(err)))?;
+        }
+
+        let ctx = &RequestContextBuilder::extend(ctx)
+            .page_content_kind(PageContentKind::DeltaLayerValue)
+            .build();
+
+        let cursor = file.block_cursor();
+        let mut buf = Vec::new();
+        for (key, lsns_at) in offsets {
+            for (lsn, block_offset) in lsns_at {
+                let res = cursor.read_blob_into_buf(block_offset, &mut buf, ctx).await;
+
+                if let Err(e) = res {
+                    reconstruct_state.on_key_error(
+                        key,
+                        PageReconstructError::from(anyhow!(e).context(format!(
+                            "Failed to read blob from virtual file {}",
+                            file.file.path
+                        ))),
+                    );
+
+                    break;
+                }
+
+                let value = Value::des(&buf);
+                if let Err(e) = value {
+                    reconstruct_state.on_key_error(
+                        key,
+                        PageReconstructError::from(anyhow!(e).context(format!(
+                            "Failed to deserialize file blob from virtual file {}",
+                            file.file.path
+                        ))),
+                    );
+
+                    break;
+                }
+
+                let key_situation = reconstruct_state.update_key(&key, lsn, value.unwrap());
+                if key_situation == ValueReconstructSituation::Complete {
+                    break;
+                }
+            }
+        }
+
+        Ok(())
+    }
+
     pub(super) async fn load_keys<'a>(
         &'a self,
         ctx: &RequestContext,