[DNM] matrix run test suite with legacy & tiered compaction

We do it as a part of more complicated tests like test_compute_restarts, but
let's have a simple test as well.

.dockerignore

@@ -17,6 +17,7 @@
 !libs/
 !neon_local/
 !pageserver/
+!patches/
 !pgxn/
 !proxy/
 !s3_scrubber/

.github/actionlint.yml

@@ -5,6 +5,7 @@ self-hosted-runner:
     - large
     - large-arm64
     - small
+    - small-arm64
     - us-east-2
 config-variables:
   - REMOTE_STORAGE_AZURE_CONTAINER

.github/actions/neon-branch-create/action.yml

@@ -3,13 +3,13 @@ description: 'Create Branch using API'
 
 inputs:
   api_key:
-    desctiption: 'Neon API key'
+    description: 'Neon API key'
     required: true
   project_id:
-    desctiption: 'ID of the Project to create Branch in'
+    description: 'ID of the Project to create Branch in'
     required: true
   api_host:
-    desctiption: 'Neon API host'
+    description: 'Neon API host'
     default: console-stage.neon.build
 outputs:
   dsn:

.github/actions/neon-branch-delete/action.yml

@@ -3,16 +3,16 @@ description: 'Delete Branch using API'
 
 inputs:
   api_key:
-    desctiption: 'Neon API key'
+    description: 'Neon API key'
     required: true
   project_id:
-    desctiption: 'ID of the Project which should be deleted'
+    description: 'ID of the Project which should be deleted'
     required: true
   branch_id:
-    desctiption: 'ID of the branch to delete'
+    description: 'ID of the branch to delete'
     required: true
   api_host:
-    desctiption: 'Neon API host'
+    description: 'Neon API host'
     default: console-stage.neon.build
 
 runs:

.github/actions/neon-project-create/action.yml

@@ -3,22 +3,22 @@ description: 'Create Neon Project using API'
 
 inputs:
   api_key:
-    desctiption: 'Neon API key'
+    description: 'Neon API key'
     required: true
   region_id:
-    desctiption: 'Region ID, if not set the project will be created in the default region'
+    description: 'Region ID, if not set the project will be created in the default region'
     default: aws-us-east-2
   postgres_version:
-    desctiption: 'Postgres version; default is 15'
-    default: 15
+    description: 'Postgres version; default is 15'
+    default: '15'
   api_host:
-    desctiption: 'Neon API host'
+    description: 'Neon API host'
     default: console-stage.neon.build
   provisioner:
-    desctiption: 'k8s-pod or k8s-neonvm'
+    description: 'k8s-pod or k8s-neonvm'
     default: 'k8s-pod'
   compute_units:
-    desctiption: '[Min, Max] compute units; Min and Max are used for k8s-neonvm with autoscaling, for k8s-pod values Min and Max should be equal'
+    description: '[Min, Max] compute units; Min and Max are used for k8s-neonvm with autoscaling, for k8s-pod values Min and Max should be equal'
     default: '[1, 1]'
 
 outputs:

.github/actions/neon-project-delete/action.yml

@@ -3,13 +3,13 @@ description: 'Delete Neon Project using API'
 
 inputs:
   api_key:
-    desctiption: 'Neon API key'
+    description: 'Neon API key'
     required: true
   project_id:
-    desctiption: 'ID of the Project to delete'
+    description: 'ID of the Project to delete'
     required: true
   api_host:
-    desctiption: 'Neon API host'
+    description: 'Neon API host'
     default: console-stage.neon.build
 
 runs:

.github/workflows/build_and_test.yml

@@ -432,8 +432,9 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        build_type: [ debug, release ]
-        pg_version: [ v14, v15, v16 ]
+        build_type: [ debug ]
+        pg_version: [ v15 ]
+        pageserver_compaction_algorithm_kind: [ "legacy", "tiered" ]
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -461,6 +462,9 @@ jobs:
           PAGESERVER_GET_VECTORED_IMPL: vectored
           PAGESERVER_GET_IMPL: vectored
           PAGESERVER_VALIDATE_VEC_GET: true
+          PAGESERVER_DEFAULT_TENANT_CONFIG_COMPACTION_ALGORITHM: 'kind="${{ matrix.pageserver_compaction_algorithm_kind }}"'
+          # catch the tests that override `tenant_config` as a whole without specifying the compaction algorithm `kind`
+          NEON_PAGESERVER_PANIC_ON_UNSPECIFIED_COMPACTION_ALGORITHM: true
 
       # Temporary disable this step until we figure out why it's so flaky
       # Ref https://github.com/neondatabase/neon/issues/4540
@@ -548,7 +552,7 @@ jobs:
 
   report-benchmarks-failures:
     needs: [ benchmarks, create-test-report ]
-    if: github.ref_name == 'main' && failure()
+    if: github.ref_name == 'main' && failure() && needs.benchmarks.result == 'failure'
     runs-on: ubuntu-latest
 
     steps:
@@ -723,9 +727,13 @@ jobs:
     uses: ./.github/workflows/trigger-e2e-tests.yml
     secrets: inherit
 
-  neon-image:
+  neon-image-arch:
     needs: [ check-permissions, build-build-tools-image, tag ]
-    runs-on: [ self-hosted, gen3, large ]
+    strategy:
+      matrix:
+        arch: [ x64, arm64 ]
+
+    runs-on: ${{ fromJson(format('["self-hosted", "gen3", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}
 
     steps:
       - name: Checkout
@@ -747,12 +755,6 @@ jobs:
           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: ${{ secrets.AWS_SECRET_KEY_DEV }}
-
       - uses: docker/build-push-action@v5
         with:
           context: .
@@ -764,25 +766,52 @@ jobs:
           push: true
           pull: true
           file: Dockerfile
-          cache-from: type=registry,ref=neondatabase/neon:cache
-          cache-to: type=registry,ref=neondatabase/neon:cache,mode=max
+          cache-from: type=registry,ref=neondatabase/neon:cache-${{ matrix.arch }}
+          cache-to: type=registry,ref=neondatabase/neon:cache-${{ matrix.arch }},mode=max
           tags: |
-            369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}}
-            neondatabase/neon:${{needs.tag.outputs.build-tag}}
+            neondatabase/neon:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
       - name: Remove custom docker config directory
         if: always()
         run: |
           rm -rf .docker-custom
 
-  compute-node-image:
-    needs: [ check-permissions, build-build-tools-image, tag ]
-    runs-on: [ self-hosted, gen3, large ]
+  neon-image:
+    needs: [ neon-image-arch, tag ]
+    runs-on: ubuntu-latest
 
+    steps:
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+
+      - name: Create multi-arch image
+        run: |
+          docker buildx imagetools create -t neondatabase/neon:${{ needs.tag.outputs.build-tag }} \
+                                             neondatabase/neon:${{ needs.tag.outputs.build-tag }}-x64 \
+                                             neondatabase/neon:${{ needs.tag.outputs.build-tag }}-arm64
+
+      - uses: docker/login-action@v3
+        with:
+          registry: 369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          username: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          password: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+      - name: Push multi-arch image to ECR
+        run: |
+          docker buildx imagetools create -t 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{ needs.tag.outputs.build-tag }} \
+                                                                                neondatabase/neon:${{ needs.tag.outputs.build-tag }}
+
+  compute-node-image-arch:
+    needs: [ check-permissions, build-build-tools-image, tag ]
     strategy:
       fail-fast: false
       matrix:
         version: [ v14, v15, v16 ]
+        arch: [ x64, arm64 ]
+
+    runs-on: ${{ fromJson(format('["self-hosted", "gen3", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}
 
     steps:
       - name: Checkout
@@ -829,15 +858,14 @@ jobs:
           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
+          cache-from: type=registry,ref=neondatabase/compute-node-${{ matrix.version }}:cache-${{ matrix.arch }}
+          cache-to: type=registry,ref=neondatabase/compute-node-${{ matrix.version }}:cache-${{ matrix.arch }},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}}
+            neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
       - name: Build compute-tools image
         # compute-tools are Postgres independent, so build it only once
-        if: ${{ matrix.version == 'v16' }}
+        if: matrix.version == 'v16'
         uses: docker/build-push-action@v5
         with:
           target: compute-tools-image
@@ -851,14 +879,57 @@ jobs:
           pull: true
           file: Dockerfile.compute-node
           tags: |
-            369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{ needs.tag.outputs.build-tag }}
-            neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}
+            neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
       - name: Remove custom docker config directory
         if: always()
         run: |
           rm -rf .docker-custom
 
+  compute-node-image:
+    needs: [ compute-node-image-arch, tag ]
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        version: [ v14, v15, v16 ]
+
+    steps:
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+
+      - name: Create multi-arch compute-node image
+        run: |
+          docker buildx imagetools create -t neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }} \
+                                             neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}-x64 \
+                                             neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}-arm64
+
+      - name: Create multi-arch compute-tools image
+        if: matrix.version == 'v16'
+        run: |
+          docker buildx imagetools create -t neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }} \
+                                             neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}-x64 \
+                                             neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}-arm64
+
+      - uses: docker/login-action@v3
+        with:
+          registry: 369495373322.dkr.ecr.eu-central-1.amazonaws.com
+          username: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+          password: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+      - name: Push multi-arch compute-node-${{ matrix.version }} image to ECR
+        run: |
+          docker buildx imagetools create -t 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: Push multi-arch compute-tools image to ECR
+        if: matrix.version == 'v16'
+        run: |
+          docker buildx imagetools create -t 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{ needs.tag.outputs.build-tag }} \
+                                                                                neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}
+
   vm-compute-node-image:
     needs: [ check-permissions, tag, compute-node-image ]
     runs-on: [ self-hosted, gen3, large ]
@@ -866,11 +937,8 @@ jobs:
       fail-fast: false
       matrix:
         version: [ v14, v15, v16 ]
-    defaults:
-      run:
-        shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.28.1
+      VM_BUILDER_VERSION: v0.29.3
 
     steps:
       - name: Checkout
@@ -883,26 +951,48 @@ jobs:
           curl -fL https://github.com/neondatabase/autoscaling/releases/download/$VM_BUILDER_VERSION/vm-builder -o vm-builder
           chmod +x vm-builder
 
+      # 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/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
+
       # Note: we need a separate pull step here because otherwise vm-builder will try to pull, and
       # it won't have the proper authentication (written at v0.6.0)
       - name: Pulling compute-node image
         run: |
-          docker pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+          docker pull neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
 
       - name: Build vm image
         run: |
           ./vm-builder \
             -spec=vm-image-spec.yaml \
-            -src=369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
-            -dst=369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+            -src=neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }} \
+            -dst=neondatabase/vm-compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
 
       - name: Pushing vm-compute-node image
         run: |
-          docker push 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+          docker push neondatabase/vm-compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
+
+      - name: Remove custom docker config directory
+        if: always()
+        run: |
+          rm -rf .docker-custom
 
   test-images:
     needs: [ check-permissions, tag, neon-image, compute-node-image ]
-    runs-on: [ self-hosted, gen3, small ]
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [ x64, arm64 ]
+
+    runs-on: ${{ fromJson(format('["self-hosted", "gen3", "{0}"]', matrix.arch == 'arm64' && 'small-arm64' || 'small')) }}
 
     steps:
       - name: Checkout
@@ -920,7 +1010,7 @@ jobs:
       - name: Verify image versions
         shell: bash # ensure no set -e for better error messages
         run: |
-          pageserver_version=$(docker run --rm 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}} "/bin/sh" "-c" "/usr/local/bin/pageserver --version")
+          pageserver_version=$(docker run --rm neondatabase/neon:${{ needs.tag.outputs.build-tag }} "/bin/sh" "-c" "/usr/local/bin/pageserver --version")
 
           echo "Pageserver version string: $pageserver_version"
 
@@ -946,78 +1036,48 @@ jobs:
 
   promote-images:
     needs: [ check-permissions, tag, test-images, vm-compute-node-image ]
-    runs-on: [ self-hosted, gen3, small ]
-    container: golang:1.19-bullseye
-    # Don't add if-condition here.
-    # The job should always be run because we have dependant other jobs that shouldn't be skipped
+    runs-on: ubuntu-latest
+
+    env:
+      VERSIONS: v14 v15 v16
 
     steps:
-      - name: Install Crane & ECR helper
-        run: |
-          go install github.com/google/go-containerregistry/cmd/crane@31786c6cbb82d6ec4fb8eb79cd9387905130534e # v0.11.0
-          go install github.com/awslabs/amazon-ecr-credential-helper/ecr-login/cli/docker-credential-ecr-login@69c85dc22db6511932bbf119e1a0cc5c90c69a7f # v0.6.0
+      - uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.NEON_DOCKERHUB_USERNAME }}
+          password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
 
-      - name: Configure ECR login
-        run: |
-          mkdir /github/home/.docker/
-          echo "{\"credsStore\":\"ecr-login\"}" > /github/home/.docker/config.json
+      - 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 }}
 
-      - name: Copy vm-compute-node images to Docker Hub
+      - name: Copy vm-compute-node images to ECR
         run: |
-          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} vm-compute-node-v14
-          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} vm-compute-node-v15
-          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} vm-compute-node-v16
+          for version in ${VERSIONS}; do
+            docker buildx imagetools create -t 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${version}:${{ needs.tag.outputs.build-tag }} \
+                                               neondatabase/vm-compute-node-${version}:${{ needs.tag.outputs.build-tag }}
+          done
 
       - name: Add latest tag to images
-        if: github.ref_name == 'main' || github.ref_name == 'release' || github.ref_name == 'release-proxy'
+        if: github.ref_name == 'main'
         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
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} latest
+          for repo in neondatabase 369495373322.dkr.ecr.eu-central-1.amazonaws.com; do
+            docker buildx imagetools create -t $repo/neon:latest \
+                                               $repo/neon:${{ needs.tag.outputs.build-tag }}
 
-      - name: Push images to production ECR
-        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
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:latest
+            docker buildx imagetools create -t $repo/compute-tools:latest \
+                                               $repo/compute-tools:${{ needs.tag.outputs.build-tag }}
 
-      - name: Configure Docker Hub login
-        run: |
-          # ECR Credential Helper & Docker Hub don't work together in config, hence reset
-          echo "" > /github/home/.docker/config.json
-          crane auth login -u ${{ secrets.NEON_DOCKERHUB_USERNAME }} -p ${{ secrets.NEON_DOCKERHUB_PASSWORD }} index.docker.io
+            for version in ${VERSIONS}; do
+              docker buildx imagetools create -t $repo/compute-node-${version}:latest \
+                                                 $repo/compute-node-${version}:${{ needs.tag.outputs.build-tag }}
 
-      - name: Push vm-compute-node to Docker Hub
-        run: |
-          crane push vm-compute-node-v14 neondatabase/vm-compute-node-v14:${{needs.tag.outputs.build-tag}}
-          crane push vm-compute-node-v15 neondatabase/vm-compute-node-v15:${{needs.tag.outputs.build-tag}}
-          crane push vm-compute-node-v16 neondatabase/vm-compute-node-v16:${{needs.tag.outputs.build-tag}}
-
-      - name: Push latest tags to Docker Hub
-        if: github.ref_name == 'main' || github.ref_name == 'release'|| github.ref_name == 'release-proxy'
-        run: |
-          crane tag neondatabase/neon:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/compute-tools:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/compute-node-v16:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} latest
-
-      - name: Cleanup ECR folder
-        run: rm -rf ~/.ecr
+              docker buildx imagetools create -t $repo/vm-compute-node-${version}:latest \
+                                                 $repo/vm-compute-node-${version}:${{ needs.tag.outputs.build-tag }}
+            done
+          done
 
   trigger-custom-extensions-build-and-wait:
     needs: [ check-permissions, tag ]

.github/workflows/neon_extra_builds.yml

@@ -136,7 +136,7 @@ jobs:
   check-linux-arm-build:
     needs: [ check-permissions, build-build-tools-image ]
     timeout-minutes: 90
-    runs-on: [ self-hosted, large-arm64 ]
+    runs-on: [ self-hosted, small-arm64 ]
 
     env:
       # Use release build only, to have less debug info around
@@ -260,7 +260,7 @@ jobs:
   check-codestyle-rust-arm:
     needs: [ check-permissions, build-build-tools-image ]
     timeout-minutes: 90
-    runs-on: [ self-hosted, large-arm64 ]
+    runs-on: [ self-hosted, small-arm64 ]
 
     container:
       image: ${{ needs.build-build-tools-image.outputs.image }}

.github/workflows/release.yml

@@ -53,7 +53,7 @@ jobs:
         GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
       run: |
         cat << EOF > body.md
-          ## Release ${RELEASE_DATE}
+          ## Storage & Compute release ${RELEASE_DATE}
 
           **Please merge this Pull Request using 'Create a merge commit' button**
         EOF

Cargo.lock

@@ -708,7 +708,7 @@ dependencies = [
  "sha1",
  "sync_wrapper",
  "tokio",
- "tokio-tungstenite 0.20.0",
+ "tokio-tungstenite",
  "tower",
  "tower-layer",
  "tower-service",
@@ -979,6 +979,12 @@ version = "3.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a3e2c3daef883ecc1b5d58c15adae93470a91d425f3532ba1695849656af3fc1"
 
+[[package]]
+name = "bytemuck"
+version = "1.16.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78834c15cb5d5efe3452d58b1e8ba890dd62d21907f867f383358198e56ebca5"
+
 [[package]]
 name = "byteorder"
 version = "1.4.3"
@@ -1066,9 +1072,9 @@ dependencies = [
 
 [[package]]
 name = "chrono"
-version = "0.4.31"
+version = "0.4.38"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f2c685bad3eb3d45a01354cedb7d5faa66194d1d58ba6e267a8de788f79db38"
+checksum = "a21f936df1771bf62b77f047b726c4625ff2e8aa607c01ec06e5a05bd8463401"
 dependencies = [
  "android-tzdata",
  "iana-time-zone",
@@ -1076,7 +1082,7 @@ dependencies = [
  "num-traits",
  "serde",
  "wasm-bindgen",
- "windows-targets 0.48.0",
+ "windows-targets 0.52.4",
 ]
 
 [[package]]
@@ -1103,7 +1109,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "defaa24ecc093c77630e6c15e17c51f5e187bf35ee514f4e2d67baaa96dae22b"
 dependencies = [
  "ciborium-io",
- "half",
+ "half 1.8.2",
 ]
 
 [[package]]
@@ -1233,8 +1239,10 @@ dependencies = [
  "serde_json",
  "signal-hook",
  "tar",
+ "thiserror",
  "tokio",
  "tokio-postgres",
+ "tokio-stream",
  "tokio-util",
  "toml_edit",
  "tracing",
@@ -1463,26 +1471,21 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-deque"
-version = "0.8.3"
+version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ce6fd6f855243022dcecf8702fef0c297d4338e226845fe067f6341ad9fa0cef"
+checksum = "613f8cc01fe9cf1a3eb3d7f488fd2fa8388403e97039e2f73692932e291a770d"
 dependencies = [
- "cfg-if",
  "crossbeam-epoch",
  "crossbeam-utils",
 ]
 
 [[package]]
 name = "crossbeam-epoch"
-version = "0.9.14"
+version = "0.9.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "46bd5f3f85273295a9d14aedfb86f6aadbff6d8f5295c4a9edb08e819dcf5695"
+checksum = "5b82ac4a3c2ca9c3460964f020e1402edd5753411d7737aa39c3714ad1b5420e"
 dependencies = [
- "autocfg",
- "cfg-if",
  "crossbeam-utils",
- "memoffset 0.8.0",
- "scopeguard",
 ]
 
 [[package]]
@@ -1596,7 +1599,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6943ae99c34386c84a470c499d3414f66502a41340aa895406e0d2e4a207b91d"
 dependencies = [
  "cfg-if",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
  "lock_api",
  "once_cell",
  "parking_lot_core 0.9.8",
@@ -1997,6 +2000,27 @@ dependencies = [
  "percent-encoding",
 ]
 
+[[package]]
+name = "framed-websockets"
+version = "0.1.0"
+source = "git+https://github.com/neondatabase/framed-websockets#34eff3d6f8cfccbc5f35e4f65314ff7328621127"
+dependencies = [
+ "base64 0.21.1",
+ "bytemuck",
+ "bytes",
+ "futures-core",
+ "futures-sink",
+ "http-body-util",
+ "hyper 1.2.0",
+ "hyper-util",
+ "pin-project",
+ "rand 0.8.5",
+ "sha1",
+ "thiserror",
+ "tokio",
+ "tokio-util",
+]
+
 [[package]]
 name = "fs2"
 version = "0.4.3"
@@ -2249,6 +2273,17 @@ version = "1.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eabb4a44450da02c90444cf74558da904edde8fb4e9035a9a6a4e15445af0bd7"
 
+[[package]]
+name = "half"
+version = "2.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6dd08c532ae367adf81c312a4580bc67f1d0fe8bc9c460520283f4c0ff277888"
+dependencies = [
+ "cfg-if",
+ "crunchy",
+ "num-traits",
+]
+
 [[package]]
 name = "hash32"
 version = "0.3.1"
@@ -2275,9 +2310,9 @@ dependencies = [
 
 [[package]]
 name = "hashbrown"
-version = "0.14.0"
+version = "0.14.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2c6201b9ff9fd90a5a3bac2e56a830d0caa509576f0e503818ee82c181b3437a"
+checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
 dependencies = [
  "ahash",
  "allocator-api2",
@@ -2285,11 +2320,11 @@ dependencies = [
 
 [[package]]
 name = "hashlink"
-version = "0.8.4"
+version = "0.9.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8094feaf31ff591f651a2664fb9cfd92bba7a60ce3197265e9482ebe753c8f7"
+checksum = "6ba4ff7128dee98c7dc9794b6a411377e1404dba1c97deb8d1a55297bd25d8af"
 dependencies = [
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 ]
 
 [[package]]
@@ -2598,21 +2633,6 @@ dependencies = [
  "tokio-native-tls",
 ]
 
-[[package]]
-name = "hyper-tungstenite"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a343d17fe7885302ed7252767dc7bb83609a874b6ff581142241ec4b73957ad"
-dependencies = [
- "http-body-util",
- "hyper 1.2.0",
- "hyper-util",
- "pin-project-lite",
- "tokio",
- "tokio-tungstenite 0.21.0",
- "tungstenite 0.21.0",
-]
-
 [[package]]
 name = "hyper-util"
 version = "0.1.3"
@@ -2690,7 +2710,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ad227c3af19d4914570ad36d30409928b75967c298feb9ea1969db3a610bb14e"
 dependencies = [
  "equivalent",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 ]
 
 [[package]]
@@ -2952,7 +2972,7 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d3262e75e648fce39813cb56ac41f3c3e3f65217ebf3844d818d1f9398cfb0dc"
 dependencies = [
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 ]
 
 [[package]]
@@ -3005,7 +3025,7 @@ checksum = "652bc741286361c06de8cb4d89b21a6437f120c508c51713663589eeb9928ac5"
 dependencies = [
  "bytes",
  "crossbeam-utils",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
  "itoa",
  "lasso",
  "measured-derive",
@@ -3567,7 +3587,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "49203cdcae0030493bad186b28da2fa25645fa276a51b6fec8010d281e02ef79"
 dependencies = [
  "dlv-list",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 ]
 
 [[package]]
@@ -3888,13 +3908,14 @@ dependencies = [
 
 [[package]]
 name = "parquet"
-version = "49.0.0"
-source = "git+https://github.com/neondatabase/arrow-rs?branch=neon-fix-bugs#8a0bc58aa67b98aabbd8eee7c6ca4281967ff9e9"
+version = "51.0.0"
+source = "git+https://github.com/apache/arrow-rs?branch=master#2534976a564be3d2d56312dc88fb1b6ed4cef829"
 dependencies = [
  "ahash",
  "bytes",
  "chrono",
- "hashbrown 0.14.0",
+ "half 2.4.1",
+ "hashbrown 0.14.5",
  "num",
  "num-bigint",
  "paste",
@@ -3902,12 +3923,13 @@ dependencies = [
  "thrift",
  "twox-hash",
  "zstd",
+ "zstd-sys",
 ]
 
 [[package]]
 name = "parquet_derive"
-version = "49.0.0"
-source = "git+https://github.com/neondatabase/arrow-rs?branch=neon-fix-bugs#8a0bc58aa67b98aabbd8eee7c6ca4281967ff9e9"
+version = "51.0.0"
+source = "git+https://github.com/apache/arrow-rs?branch=master#2534976a564be3d2d56312dc88fb1b6ed4cef829"
 dependencies = [
  "parquet",
  "proc-macro2",
@@ -3934,9 +3956,9 @@ checksum = "de3145af08024dea9fa9914f381a17b8fc6034dfb00f3a84013f7ff43f29ed4c"
 
 [[package]]
 name = "pbkdf2"
-version = "0.12.1"
+version = "0.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0ca0b5a68607598bf3bad68f32227a8164f6254833f84eafaac409cd6746c31"
+checksum = "f8ed6a7761f76e3b9f92dfb0a60a6a6477c61024b775147ff0973a02653abaf2"
 dependencies = [
  "digest",
  "hmac",
@@ -4359,6 +4381,7 @@ dependencies = [
 name = "proxy"
 version = "0.1.0"
 dependencies = [
+ "ahash",
  "anyhow",
  "async-compression",
  "async-trait",
@@ -4375,12 +4398,14 @@ dependencies = [
  "chrono",
  "clap",
  "consumption_metrics",
+ "crossbeam-deque",
  "dashmap",
  "env_logger",
  "fallible-iterator",
+ "framed-websockets",
  "futures",
  "git-version",
- "hashbrown 0.13.2",
+ "hashbrown 0.14.5",
  "hashlink",
  "hex",
  "hmac",
@@ -4390,7 +4415,6 @@ dependencies = [
  "humantime",
  "hyper 0.14.26",
  "hyper 1.2.0",
- "hyper-tungstenite",
  "hyper-util",
  "indexmap 2.0.1",
  "ipnet",
@@ -4435,7 +4459,6 @@ dependencies = [
  "smol_str",
  "socket2 0.5.5",
  "subtle",
- "sync_wrapper",
  "task-local-extensions",
  "thiserror",
  "tikv-jemalloc-ctl",
@@ -4444,6 +4467,7 @@ dependencies = [
  "tokio-postgres",
  "tokio-postgres-rustls",
  "tokio-rustls 0.25.0",
+ "tokio-tungstenite",
  "tokio-util",
  "tower-service",
  "tracing",
@@ -5952,7 +5976,7 @@ checksum = "81cdd64d312baedb58e21336b31bc043b77e01cc99033ce76ef539f78e965ebc"
 [[package]]
 name = "svg_fmt"
 version = "0.4.2"
-source = "git+https://github.com/neondatabase/fork--nical--rust_debug?branch=neon#c1820b28664b5df68de7f043fccf2ed5d67b6ae8"
+source = "git+https://github.com/nical/rust_debug?rev=28a7d96eecff2f28e75b1ea09f2d499a60d0e3b4#28a7d96eecff2f28e75b1ea09f2d499a60d0e3b4"
 
 [[package]]
 name = "syn"
@@ -6380,19 +6404,7 @@ dependencies = [
  "futures-util",
  "log",
  "tokio",
- "tungstenite 0.20.1",
-]
-
-[[package]]
-name = "tokio-tungstenite"
-version = "0.21.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c83b561d025642014097b66e6c1bb422783339e0909e4429cde4749d1990bc38"
-dependencies = [
- "futures-util",
- "log",
- "tokio",
- "tungstenite 0.21.0",
+ "tungstenite",
 ]
 
 [[package]]
@@ -6406,7 +6418,7 @@ dependencies = [
  "futures-io",
  "futures-sink",
  "futures-util",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
  "pin-project-lite",
  "tokio",
  "tracing",
@@ -6688,25 +6700,6 @@ dependencies = [
  "utf-8",
 ]
 
-[[package]]
-name = "tungstenite"
-version = "0.21.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9ef1a641ea34f399a848dea702823bbecfb4c486f911735368f1f137cb8257e1"
-dependencies = [
- "byteorder",
- "bytes",
- "data-encoding",
- "http 1.1.0",
- "httparse",
- "log",
- "rand 0.8.5",
- "sha1",
- "thiserror",
- "url",
- "utf-8",
-]
-
 [[package]]
 name = "twox-hash"
 version = "1.6.3"
@@ -7477,6 +7470,7 @@ dependencies = [
 name = "workspace_hack"
 version = "0.1.0"
 dependencies = [
+ "ahash",
  "anyhow",
  "aws-config",
  "aws-runtime",
@@ -7502,7 +7496,7 @@ dependencies = [
  "futures-sink",
  "futures-util",
  "getrandom 0.2.11",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
  "hex",
  "hmac",
  "hyper 0.14.26",

Cargo.toml

@@ -41,6 +41,7 @@ license = "Apache-2.0"
 
 ## All dependency versions, used in the project
 [workspace.dependencies]
+ahash = "0.8"
 anyhow = { version = "1.0", features = ["backtrace"] }
 arc-swap = "1.6"
 async-compression = { version = "0.4.0", features = ["tokio", "gzip", "zstd"] }
@@ -74,6 +75,7 @@ clap = { version = "4.0", features = ["derive"] }
 comfy-table = "6.1"
 const_format = "0.2"
 crc32c = "0.6"
+crossbeam-deque = "0.8.5"
 crossbeam-utils = "0.8.5"
 dashmap = { version = "5.5.0", features = ["raw-api"] }
 either = "1.8"
@@ -81,13 +83,14 @@ enum-map = "2.4.2"
 enumset = "1.0.12"
 fail = "0.5.0"
 fallible-iterator = "0.2"
+framed-websockets = { version = "0.1.0", git = "https://github.com/neondatabase/framed-websockets" }
 fs2 = "0.4.3"
 futures = "0.3"
 futures-core = "0.3"
 futures-util = "0.3"
 git-version = "0.3"
-hashbrown = "0.13"
-hashlink = "0.8.4"
+hashbrown = "0.14"
+hashlink = "0.9.1"
 hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
@@ -98,7 +101,7 @@ http-types = { version = "2", default-features = false }
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
-hyper-tungstenite = "0.13.0"
+tokio-tungstenite = "0.20.0"
 indexmap = "2"
 inotify = "0.10.2"
 ipnet = "2.9.0"
@@ -121,8 +124,8 @@ opentelemetry = "0.20.0"
 opentelemetry-otlp = { version = "0.13.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
 opentelemetry-semantic-conventions = "0.12.0"
 parking_lot = "0.12"
-parquet = { version = "49.0.0", default-features = false, features = ["zstd"] }
-parquet_derive = "49.0.0"
+parquet = { version = "51.0.0", default-features = false, features = ["zstd"] }
+parquet_derive = "51.0.0"
 pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pin-project-lite = "0.2"
 procfs = "0.14"
@@ -158,8 +161,8 @@ socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
 "subtle"  = "2.5.0"
-# https://github.com/nical/rust_debug/pull/4
-svg_fmt = { git = "https://github.com/neondatabase/fork--nical--rust_debug", branch = "neon" }
+# Our PR https://github.com/nical/rust_debug/pull/4 has been merged but no new version released yet
+svg_fmt = { git = "https://github.com/nical/rust_debug", rev = "28a7d96eecff2f28e75b1ea09f2d499a60d0e3b4" }
 sync_wrapper = "0.1.2"
 tar = "0.4"
 task-local-extensions = "0.1.4"
@@ -243,8 +246,8 @@ tonic-build = "0.9"
 tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch="neon" }
 
 # bug fixes for UUID
-parquet = { git = "https://github.com/neondatabase/arrow-rs", branch = "neon-fix-bugs" }
-parquet_derive = { git = "https://github.com/neondatabase/arrow-rs", branch = "neon-fix-bugs" }
+parquet = { git = "https://github.com/apache/arrow-rs", branch = "master" }
+parquet_derive = { git = "https://github.com/apache/arrow-rs", branch = "master" }
 
 ################# Binary contents sections
 

Dockerfile.build-tools

@@ -87,7 +87,7 @@ RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-$(uname -m).zip" -o "aws
     && rm awscliv2.zip
 
 # Mold: A Modern Linker
-ENV MOLD_VERSION v2.4.0
+ENV MOLD_VERSION v2.31.0
 RUN set -e \
     && git clone https://github.com/rui314/mold.git \
     && mkdir mold/build \

Dockerfile.compute-node

@@ -241,11 +241,17 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.7.tar.gz -
 FROM build-deps AS vector-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.1.tar.gz -O pgvector.tar.gz && \
-    echo "cc7a8e034a96e30a819911ac79d32f6bc47bdd1aa2de4d7d4904e26b83209dc8 pgvector.tar.gz" | sha256sum --check && \
+COPY patches/pgvector.patch /pgvector.patch
+
+# By default, pgvector Makefile uses `-march=native`. We don't want that, 
+# because we build the images on different machines than where we run them.
+# Pass OPTFLAGS="" to remove it.
+RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.7.0.tar.gz -O pgvector.tar.gz && \
+    echo "1b5503a35c265408b6eb282621c5e1e75f7801afc04eecb950796cfee2e3d1d8 pgvector.tar.gz" | sha256sum --check && \
     mkdir pgvector-src && cd pgvector-src && tar xvzf ../pgvector.tar.gz --strip-components=1 -C . && \
-    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
-    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
+    patch -p1 < /pgvector.patch && \
+    make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
+    make -j $(getconf _NPROCESSORS_ONLN) OPTFLAGS="" install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/vector.control
 
 #########################################################################################

README.md

@@ -1,4 +1,6 @@
-[![Neon](https://user-images.githubusercontent.com/13738772/236813940-dcfdcb5b-69d3-449b-a686-013febe834d4.png)](https://neon.tech)
+[![Neon](https://github.com/neondatabase/neon/assets/11527560/f15a17f0-836e-40c5-b35d-030606a6b660)](https://neon.tech)
+
+
 
 # Neon
 

compute_tools/Cargo.toml

@@ -27,10 +27,12 @@ reqwest = { workspace = true, features = ["json"] }
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tokio-postgres.workspace = true
 tokio-util.workspace = true
+tokio-stream.workspace = true
 tracing.workspace = true
 tracing-opentelemetry.workspace = true
 tracing-subscriber.workspace = true
 tracing-utils.workspace = true
+thiserror.workspace = true
 url.workspace = true
 
 compute_api.workspace = true

compute_tools/src/catalog.rs

@@ -0,0 +1,116 @@
+use compute_api::{
+    responses::CatalogObjects,
+    spec::{Database, Role},
+};
+use futures::Stream;
+use postgres::{Client, NoTls};
+use std::{path::Path, process::Stdio, result::Result, sync::Arc};
+use tokio::{
+    io::{AsyncBufReadExt, BufReader},
+    process::Command,
+    task,
+};
+use tokio_stream::{self as stream, StreamExt};
+use tokio_util::codec::{BytesCodec, FramedRead};
+use tracing::warn;
+
+use crate::{
+    compute::ComputeNode,
+    pg_helpers::{get_existing_dbs, get_existing_roles},
+};
+
+pub async fn get_dbs_and_roles(compute: &Arc<ComputeNode>) -> anyhow::Result<CatalogObjects> {
+    let connstr = compute.connstr.clone();
+    task::spawn_blocking(move || {
+        let mut client = Client::connect(connstr.as_str(), NoTls)?;
+        let roles: Vec<Role>;
+        {
+            let mut xact = client.transaction()?;
+            roles = get_existing_roles(&mut xact)?;
+        }
+        let databases: Vec<Database> = get_existing_dbs(&mut client)?.values().cloned().collect();
+
+        Ok(CatalogObjects { roles, databases })
+    })
+    .await?
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum SchemaDumpError {
+    #[error("Database does not exist.")]
+    DatabaseDoesNotExist,
+    #[error("Failed to execute pg_dump.")]
+    IO(#[from] std::io::Error),
+}
+
+// It uses the pg_dump utility to dump the schema of the specified database.
+// The output is streamed back to the caller and supposed to be streamed via HTTP.
+//
+// Before return the result with the output, it checks that pg_dump produced any output.
+// If not, it tries to parse the stderr output to determine if the database does not exist
+// and special error is returned.
+//
+// To make sure that the process is killed when the caller drops the stream, we use tokio kill_on_drop feature.
+pub async fn get_database_schema(
+    compute: &Arc<ComputeNode>,
+    dbname: &str,
+) -> Result<impl Stream<Item = Result<bytes::Bytes, std::io::Error>>, SchemaDumpError> {
+    let pgbin = &compute.pgbin;
+    let basepath = Path::new(pgbin).parent().unwrap();
+    let pgdump = basepath.join("pg_dump");
+    let mut connstr = compute.connstr.clone();
+    connstr.set_path(dbname);
+    let mut cmd = Command::new(pgdump)
+        .arg("--schema-only")
+        .arg(connstr.as_str())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped())
+        .kill_on_drop(true)
+        .spawn()?;
+
+    let stdout = cmd.stdout.take().ok_or_else(|| {
+        std::io::Error::new(std::io::ErrorKind::Other, "Failed to capture stdout.")
+    })?;
+
+    let stderr = cmd.stderr.take().ok_or_else(|| {
+        std::io::Error::new(std::io::ErrorKind::Other, "Failed to capture stderr.")
+    })?;
+
+    let mut stdout_reader = FramedRead::new(stdout, BytesCodec::new());
+    let stderr_reader = BufReader::new(stderr);
+
+    let first_chunk = match stdout_reader.next().await {
+        Some(Ok(bytes)) if !bytes.is_empty() => bytes,
+        Some(Err(e)) => {
+            return Err(SchemaDumpError::IO(e));
+        }
+        _ => {
+            let mut lines = stderr_reader.lines();
+            if let Some(line) = lines.next_line().await? {
+                if line.contains(&format!("FATAL:  database \"{}\" does not exist", dbname)) {
+                    return Err(SchemaDumpError::DatabaseDoesNotExist);
+                }
+                warn!("pg_dump stderr: {}", line)
+            }
+            tokio::spawn(async move {
+                while let Ok(Some(line)) = lines.next_line().await {
+                    warn!("pg_dump stderr: {}", line)
+                }
+            });
+
+            return Err(SchemaDumpError::IO(std::io::Error::new(
+                std::io::ErrorKind::Other,
+                "failed to start pg_dump",
+            )));
+        }
+    };
+    let initial_stream = stream::once(Ok(first_chunk.freeze()));
+    // Consume stderr and log warnings
+    tokio::spawn(async move {
+        let mut lines = stderr_reader.lines();
+        while let Ok(Some(line)) = lines.next_line().await {
+            warn!("pg_dump stderr: {}", line)
+        }
+    });
+    Ok(initial_stream.chain(stdout_reader.map(|res| res.map(|b| b.freeze()))))
+}

compute_tools/src/http/api.rs

@@ -5,17 +5,21 @@ use std::net::SocketAddr;
 use std::sync::Arc;
 use std::thread;
 
+use crate::catalog::SchemaDumpError;
+use crate::catalog::{get_database_schema, get_dbs_and_roles};
 use crate::compute::forward_termination_signal;
 use crate::compute::{ComputeNode, ComputeState, ParsedSpec};
 use compute_api::requests::ConfigurationRequest;
 use compute_api::responses::{ComputeStatus, ComputeStatusResponse, GenericAPIError};
 
 use anyhow::Result;
+use hyper::header::CONTENT_TYPE;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::{Body, Method, Request, Response, Server, StatusCode};
 use tokio::task;
 use tracing::{error, info, warn};
 use tracing_utils::http::OtelName;
+use utils::http::request::must_get_query_param;
 
 fn status_response_from_state(state: &ComputeState) -> ComputeStatusResponse {
     ComputeStatusResponse {
@@ -133,6 +137,34 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
             }
         }
 
+        (&Method::GET, "/dbs_and_roles") => {
+            info!("serving /dbs_and_roles GET request",);
+            match get_dbs_and_roles(compute).await {
+                Ok(res) => render_json(Body::from(serde_json::to_string(&res).unwrap())),
+                Err(_) => {
+                    render_json_error("can't get dbs and roles", StatusCode::INTERNAL_SERVER_ERROR)
+                }
+            }
+        }
+
+        (&Method::GET, "/database_schema") => {
+            let database = match must_get_query_param(&req, "database") {
+                Err(e) => return e.into_response(),
+                Ok(database) => database,
+            };
+            info!("serving /database_schema GET request with database: {database}",);
+            match get_database_schema(compute, &database).await {
+                Ok(res) => render_plain(Body::wrap_stream(res)),
+                Err(SchemaDumpError::DatabaseDoesNotExist) => {
+                    render_json_error("database does not exist", StatusCode::NOT_FOUND)
+                }
+                Err(e) => {
+                    error!("can't get schema dump: {}", e);
+                    render_json_error("can't get schema dump", StatusCode::INTERNAL_SERVER_ERROR)
+                }
+            }
+        }
+
         // download extension files from remote extension storage on demand
         (&Method::POST, route) if route.starts_with("/extension_server/") => {
             info!("serving {:?} POST request", route);
@@ -303,10 +335,25 @@ fn render_json_error(e: &str, status: StatusCode) -> Response<Body> {
     };
     Response::builder()
         .status(status)
+        .header(CONTENT_TYPE, "application/json")
         .body(Body::from(serde_json::to_string(&error).unwrap()))
         .unwrap()
 }
 
+fn render_json(body: Body) -> Response<Body> {
+    Response::builder()
+        .header(CONTENT_TYPE, "application/json")
+        .body(body)
+        .unwrap()
+}
+
+fn render_plain(body: Body) -> Response<Body> {
+    Response::builder()
+        .header(CONTENT_TYPE, "text/plain")
+        .body(body)
+        .unwrap()
+}
+
 async fn handle_terminate_request(compute: &Arc<ComputeNode>) -> Result<(), (String, StatusCode)> {
     {
         let mut state = compute.state.lock().unwrap();

compute_tools/src/http/openapi_spec.yaml

@@ -68,6 +68,51 @@ paths:
               schema:
                 $ref: "#/components/schemas/Info"
 
+  /dbs_and_roles:
+    get:
+      tags:
+        - Info
+      summary: Get databases and roles in the catalog.
+      description: ""
+      operationId: getDbsAndRoles
+      responses:
+        200:
+          description: Compute schema objects
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/DbsAndRoles"
+
+  /database_schema:
+    get:
+      tags:
+        - Info
+      summary: Get schema dump
+      parameters:
+        - name: database
+          in: query
+          description: Database name to dump.
+          required: true
+          schema:
+            type: string
+          example: "postgres"
+      description: Get schema dump in SQL format.
+      operationId: getDatabaseSchema
+      responses:
+        200:
+          description: Schema dump
+          content:
+            text/plain:
+              schema:
+                type: string
+                description: Schema dump in SQL format.
+        404:
+          description: Non existing database.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
+
   /check_writability:
     post:
       tags:
@@ -229,6 +274,73 @@ components:
         num_cpus:
           type: integer
 
+    DbsAndRoles:
+      type: object
+      description: Databases and Roles
+      required:
+        - roles
+        - databases
+      properties:
+        roles:
+          type: array
+          items:
+            $ref: "#/components/schemas/Role"
+        databases:
+          type: array
+          items:
+            $ref: "#/components/schemas/Database"
+
+    Database:
+      type: object
+      description: Database
+      required:
+        - name
+        - owner
+        - restrict_conn
+        - invalid
+      properties:
+        name:
+          type: string
+        owner:
+          type: string
+        options:
+          type: array
+          items:
+            $ref: "#/components/schemas/GenericOption"
+        restrict_conn:
+          type: boolean
+        invalid:
+          type: boolean
+
+    Role:
+      type: object
+      description: Role
+      required:
+        - name
+      properties:
+        name:
+          type: string
+        encrypted_password:
+          type: string
+        options:
+          type: array
+          items:
+            $ref: "#/components/schemas/GenericOption"
+
+    GenericOption:
+      type: object
+      description: Schema Generic option
+      required:
+        - name
+        - vartype
+      properties:
+        name:
+          type: string
+        value:
+          type: string
+        vartype:
+          type: string
+
     ComputeState:
       type: object
       required:

compute_tools/src/lib.rs

@@ -8,6 +8,7 @@ pub mod configurator;
 pub mod http;
 #[macro_use]
 pub mod logger;
+pub mod catalog;
 pub mod compute;
 pub mod extension_server;
 pub mod monitor;

compute_tools/src/swap.rs

@@ -1,3 +1,5 @@
+use std::path::Path;
+
 use anyhow::{anyhow, Context};
 use tracing::warn;
 
@@ -17,17 +19,24 @@ pub fn resize_swap(size_bytes: u64) -> anyhow::Result<()> {
         .arg(size_bytes.to_string())
         .spawn();
 
-    if matches!(&child_result, Err(e) if e.kind() == std::io::ErrorKind::NotFound) {
-        warn!("ignoring \"not found\" error from resize-swap to avoid swapoff while compute is running");
-        return Ok(());
-    }
-
     child_result
         .context("spawn() failed")
         .and_then(|mut child| child.wait().context("wait() failed"))
         .and_then(|status| match status.success() {
             true => Ok(()),
-            false => Err(anyhow!("process exited with {status}")),
+            false => {
+                // The command failed. Maybe it was because the resize-swap file doesn't exist?
+                // The --once flag causes it to delete itself on success so we don't disable swap
+                // while postgres is running; maybe this is fine.
+                match Path::new(RESIZE_SWAP_BIN).try_exists() {
+                    Err(_) | Ok(true) => Err(anyhow!("process exited with {status}")),
+                    // The path doesn't exist; we're actually ok 
+                    Ok(false) => {
+                        warn!("ignoring \"not found\" error from resize-swap to avoid swapoff while compute is running");
+                        Ok(())
+                    },
+                }
+            }
         })
         // wrap any prior error with the overall context that we couldn't run the command
         .with_context(|| {

control_plane/src/local_env.rs

@@ -152,6 +152,9 @@ pub struct NeonStorageControllerConf {
     /// Heartbeat timeout before marking a node offline
     #[serde(with = "humantime_serde")]
     pub max_unavailable: Duration,
+
+    /// Threshold for auto-splitting a tenant into shards
+    pub split_threshold: Option<u64>,
 }
 
 impl NeonStorageControllerConf {
@@ -164,6 +167,7 @@ impl Default for NeonStorageControllerConf {
     fn default() -> Self {
         Self {
             max_unavailable: Self::DEFAULT_MAX_UNAVAILABLE_INTERVAL,
+            split_threshold: None,
         }
     }
 }

control_plane/src/storage_controller.rs

@@ -243,9 +243,13 @@ impl StorageController {
                 anyhow::bail!("initdb failed with status {status}");
             }
 
+            // Write a minimal config file:
+            // - Specify the port, since this is chosen dynamically
+            // - Switch off fsync, since we're running on lightweight test environments and when e.g. scale testing
+            //   the storage controller we don't want a slow local disk to interfere with that.
             tokio::fs::write(
                 &pg_data_path.join("postgresql.conf"),
-                format!("port = {}", self.postgres_port),
+                format!("port = {}\nfsync=off\n", self.postgres_port),
             )
             .await?;
         };
@@ -305,6 +309,10 @@ impl StorageController {
             ));
         }
 
+        if let Some(split_threshold) = self.config.split_threshold.as_ref() {
+            args.push(format!("--split-threshold={split_threshold}"))
+        }
+
         background_process::start_process(
             COMMAND,
             &self.env.base_data_dir,

docker-compose/compute_wrapper/Dockerfile

@@ -1,4 +1,4 @@
-ARG REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
+ARG REPOSITORY=neondatabase
 ARG COMPUTE_IMAGE=compute-node-v14
 ARG TAG=latest
 

docker-compose/docker_compose_test.sh

@@ -8,8 +8,6 @@
 # Their defaults point at DockerHub `neondatabase/neon:latest` image.`,
 # to verify custom image builds (e.g pre-published ones).
 
-# XXX: Current does not work on M1 macs due to x86_64 Docker images compiled only, and no seccomp support in M1 Docker emulation layer.
-
 set -eux -o pipefail
 
 SCRIPT_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"

libs/compute_api/src/responses.rs

@@ -3,7 +3,7 @@
 use chrono::{DateTime, Utc};
 use serde::{Deserialize, Serialize, Serializer};
 
-use crate::spec::ComputeSpec;
+use crate::spec::{ComputeSpec, Database, Role};
 
 #[derive(Serialize, Debug, Deserialize)]
 pub struct GenericAPIError {
@@ -113,6 +113,12 @@ pub struct ComputeMetrics {
     pub total_ext_download_size: u64,
 }
 
+#[derive(Clone, Debug, Default, Serialize)]
+pub struct CatalogObjects {
+    pub roles: Vec<Role>,
+    pub databases: Vec<Database>,
+}
+
 /// Response of the `/computes/{compute_id}/spec` control-plane API.
 /// This is not actually a compute API response, so consider moving
 /// to a different place.

libs/pageserver_api/src/keyspace.rs

@@ -307,7 +307,7 @@ impl KeySpace {
     }
 
     /// Merge another keyspace into the current one.
-    /// Note: the keyspaces must not ovelap (enforced via assertions)
+    /// Note: the keyspaces must not overlap (enforced via assertions). To merge overlapping key ranges, use `KeySpaceRandomAccum`.
     pub fn merge(&mut self, other: &KeySpace) {
         let all_ranges = self
             .ranges

libs/pageserver_api/src/models.rs

@@ -9,7 +9,7 @@ use std::{
     collections::HashMap,
     io::{BufRead, Read},
     num::{NonZeroU64, NonZeroUsize},
-    str::FromStr,
+    sync::atomic::AtomicUsize,
     time::{Duration, SystemTime},
 };
 
@@ -161,6 +161,22 @@ impl std::fmt::Debug for TenantState {
     }
 }
 
+/// A temporary lease to a specific lsn inside a timeline.
+/// Access to the lsn is guaranteed by the pageserver until the expiration indicated by `valid_until`.
+#[serde_as]
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+pub struct LsnLease {
+    #[serde_as(as = "SystemTimeAsRfc3339Millis")]
+    pub valid_until: SystemTime,
+}
+
+serde_with::serde_conv!(
+    SystemTimeAsRfc3339Millis,
+    SystemTime,
+    |time: &SystemTime| humantime::format_rfc3339_millis(*time).to_string(),
+    |value: String| -> Result<_, humantime::TimestampError> { humantime::parse_rfc3339(&value) }
+);
+
 /// The only [`TenantState`] variants we could be `TenantState::Activating` from.
 #[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum ActivatingFrom {
@@ -289,7 +305,7 @@ pub struct TenantConfig {
     pub compaction_period: Option<String>,
     pub compaction_threshold: Option<usize>,
     // defer parsing compaction_algorithm, like eviction_policy
-    pub compaction_algorithm: Option<CompactionAlgorithm>,
+    pub compaction_algorithm: Option<CompactionAlgorithmSettings>,
     pub gc_horizon: Option<u64>,
     pub gc_period: Option<String>,
     pub image_creation_threshold: Option<usize>,
@@ -308,28 +324,100 @@ pub struct TenantConfig {
     pub switch_aux_file_policy: Option<AuxFilePolicy>,
 }
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+/// The policy for the aux file storage. It can be switched through `switch_aux_file_policy`
+/// tenant config. When the first aux file written, the policy will be persisted in the
+/// `index_part.json` file and has a limited migration path.
+///
+/// Currently, we only allow the following migration path:
+///
+/// Unset -> V1
+///       -> V2
+///       -> CrossValidation -> V2
+#[derive(
+    Eq,
+    PartialEq,
+    Debug,
+    Copy,
+    Clone,
+    strum_macros::EnumString,
+    strum_macros::Display,
+    serde_with::DeserializeFromStr,
+    serde_with::SerializeDisplay,
+)]
+#[strum(serialize_all = "kebab-case")]
 pub enum AuxFilePolicy {
+    /// V1 aux file policy: store everything in AUX_FILE_KEY
+    #[strum(ascii_case_insensitive)]
     V1,
+    /// V2 aux file policy: store in the AUX_FILE keyspace
+    #[strum(ascii_case_insensitive)]
     V2,
+    /// Cross validation runs both formats on the write path and does validation
+    /// on the read path.
+    #[strum(ascii_case_insensitive)]
     CrossValidation,
 }
 
-impl FromStr for AuxFilePolicy {
-    type Err = anyhow::Error;
+impl AuxFilePolicy {
+    pub fn is_valid_migration_path(from: Option<Self>, to: Self) -> bool {
+        matches!(
+            (from, to),
+            (None, _) | (Some(AuxFilePolicy::CrossValidation), AuxFilePolicy::V2)
+        )
+    }
 
-    fn from_str(s: &str) -> Result<Self, Self::Err> {
-        let s = s.to_lowercase();
-        if s == "v1" {
-            Ok(Self::V1)
-        } else if s == "v2" {
-            Ok(Self::V2)
-        } else if s == "crossvalidation" || s == "cross_validation" {
-            Ok(Self::CrossValidation)
-        } else {
-            anyhow::bail!("cannot parse {} to aux file policy", s)
+    /// If a tenant writes aux files without setting `switch_aux_policy`, this value will be used.
+    pub fn default_tenant_config() -> Self {
+        Self::V1
+    }
+}
+
+/// The aux file policy memory flag. Users can store `Option<AuxFilePolicy>` into this atomic flag. 0 == unspecified.
+pub struct AtomicAuxFilePolicy(AtomicUsize);
+
+impl AtomicAuxFilePolicy {
+    pub fn new(policy: Option<AuxFilePolicy>) -> Self {
+        Self(AtomicUsize::new(
+            policy.map(AuxFilePolicy::to_usize).unwrap_or_default(),
+        ))
+    }
+
+    pub fn load(&self) -> Option<AuxFilePolicy> {
+        match self.0.load(std::sync::atomic::Ordering::Acquire) {
+            0 => None,
+            other => Some(AuxFilePolicy::from_usize(other)),
         }
     }
+
+    pub fn store(&self, policy: Option<AuxFilePolicy>) {
+        self.0.store(
+            policy.map(AuxFilePolicy::to_usize).unwrap_or_default(),
+            std::sync::atomic::Ordering::Release,
+        );
+    }
+}
+
+impl AuxFilePolicy {
+    pub fn to_usize(self) -> usize {
+        match self {
+            Self::V1 => 1,
+            Self::CrossValidation => 2,
+            Self::V2 => 3,
+        }
+    }
+
+    pub fn try_from_usize(this: usize) -> Option<Self> {
+        match this {
+            1 => Some(Self::V1),
+            2 => Some(Self::CrossValidation),
+            3 => Some(Self::V2),
+            _ => None,
+        }
+    }
+
+    pub fn from_usize(this: usize) -> Self {
+        Self::try_from_usize(this).unwrap()
+    }
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -350,13 +438,30 @@ impl EvictionPolicy {
     }
 }
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(tag = "kind")]
+#[derive(
+    Eq,
+    PartialEq,
+    Debug,
+    Copy,
+    Clone,
+    strum_macros::EnumString,
+    strum_macros::Display,
+    serde_with::DeserializeFromStr,
+    serde_with::SerializeDisplay,
+)]
+#[strum(serialize_all = "kebab-case")]
 pub enum CompactionAlgorithm {
+    #[strum(disabled)]
+    NotSpecified,
     Legacy,
     Tiered,
 }
 
+#[derive(Eq, PartialEq, Debug, Clone, Serialize, Deserialize)]
+pub struct CompactionAlgorithmSettings {
+    pub kind: CompactionAlgorithm,
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 pub struct EvictionPolicyLayerAccessThreshold {
     #[serde(with = "humantime_serde")]
@@ -604,6 +709,9 @@ pub struct TimelineInfo {
     pub state: TimelineState,
 
     pub walreceiver_status: String,
+
+    /// The last aux file policy being used on this timeline
+    pub last_aux_file_policy: Option<AuxFilePolicy>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -710,6 +818,8 @@ pub enum HistoricLayerInfo {
         lsn_end: Lsn,
         remote: bool,
         access_stats: LayerAccessStats,
+
+        l0: bool,
     },
     Image {
         layer_file_name: String,
@@ -745,6 +855,16 @@ impl HistoricLayerInfo {
         };
         *field = value;
     }
+    pub fn layer_file_size(&self) -> u64 {
+        match self {
+            HistoricLayerInfo::Delta {
+                layer_file_size, ..
+            } => *layer_file_size,
+            HistoricLayerInfo::Image {
+                layer_file_size, ..
+            } => *layer_file_size,
+        }
+    }
 }
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -752,6 +872,16 @@ pub struct DownloadRemoteLayersTaskSpawnRequest {
     pub max_concurrent_downloads: NonZeroUsize,
 }
 
+#[derive(Debug, Serialize, Deserialize)]
+pub struct IngestAuxFilesRequest {
+    pub aux_files: HashMap<String, String>,
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+pub struct ListAuxFilesRequest {
+    pub lsn: Lsn,
+}
+
 #[derive(Debug, Serialize, Deserialize, Clone)]
 pub struct DownloadRemoteLayersTaskInfo {
     pub task_id: String,
@@ -776,9 +906,6 @@ pub struct TimelineGcRequest {
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct WalRedoManagerProcessStatus {
     pub pid: u32,
-    /// The strum-generated `into::<&'static str>()` for `pageserver::walredo::ProcessKind`.
-    /// `ProcessKind` are a transitory thing, so, they have no enum representation in `pageserver_api`.
-    pub kind: Cow<'static, str>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -817,6 +944,55 @@ pub struct TenantScanRemoteStorageResponse {
     pub shards: Vec<TenantScanRemoteStorageShard>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone)]
+#[serde(rename_all = "snake_case")]
+pub enum TenantSorting {
+    ResidentSize,
+    MaxLogicalSize,
+}
+
+impl Default for TenantSorting {
+    fn default() -> Self {
+        Self::ResidentSize
+    }
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct TopTenantShardsRequest {
+    // How would you like to sort the tenants?
+    pub order_by: TenantSorting,
+
+    // How many results?
+    pub limit: usize,
+
+    // Omit tenants with more than this many shards (e.g. if this is the max number of shards
+    // that the caller would ever split to)
+    pub where_shards_lt: Option<ShardCount>,
+
+    // Omit tenants where the ordering metric is less than this (this is an optimization to
+    // let us quickly exclude numerous tiny shards)
+    pub where_gt: Option<u64>,
+}
+
+#[derive(Serialize, Deserialize, Debug, PartialEq, Eq)]
+pub struct TopTenantShardItem {
+    pub id: TenantShardId,
+
+    /// Total size of layers on local disk for all timelines in this tenant
+    pub resident_size: u64,
+
+    /// Total size of layers in remote storage for all timelines in this tenant
+    pub physical_size: u64,
+
+    /// The largest logical size of a timeline within this tenant
+    pub max_logical_size: u64,
+}
+
+#[derive(Serialize, Deserialize, Debug, Default)]
+pub struct TopTenantShardsResponse {
+    pub shards: Vec<TopTenantShardItem>,
+}
+
 pub mod virtual_file {
     #[derive(
         Copy,
@@ -1242,6 +1418,7 @@ impl PagestreamBeMessage {
 #[cfg(test)]
 mod tests {
     use serde_json::json;
+    use std::str::FromStr;
 
     use super::*;
 
@@ -1449,4 +1626,69 @@ mod tests {
             assert_eq!(actual, expected, "example on {line}");
         }
     }
+
+    #[test]
+    fn test_aux_file_migration_path() {
+        assert!(AuxFilePolicy::is_valid_migration_path(
+            None,
+            AuxFilePolicy::V1
+        ));
+        assert!(AuxFilePolicy::is_valid_migration_path(
+            None,
+            AuxFilePolicy::V2
+        ));
+        assert!(AuxFilePolicy::is_valid_migration_path(
+            None,
+            AuxFilePolicy::CrossValidation
+        ));
+        // Self-migration is not a valid migration path, and the caller should handle it by itself.
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V1),
+            AuxFilePolicy::V1
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V2),
+            AuxFilePolicy::V2
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::CrossValidation),
+            AuxFilePolicy::CrossValidation
+        ));
+        // Migrations not allowed
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::CrossValidation),
+            AuxFilePolicy::V1
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V1),
+            AuxFilePolicy::V2
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V2),
+            AuxFilePolicy::V1
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V2),
+            AuxFilePolicy::CrossValidation
+        ));
+        assert!(!AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::V1),
+            AuxFilePolicy::CrossValidation
+        ));
+        // Migrations allowed
+        assert!(AuxFilePolicy::is_valid_migration_path(
+            Some(AuxFilePolicy::CrossValidation),
+            AuxFilePolicy::V2
+        ));
+    }
+
+    #[test]
+    fn test_aux_parse() {
+        assert_eq!(AuxFilePolicy::from_str("V2").unwrap(), AuxFilePolicy::V2);
+        assert_eq!(AuxFilePolicy::from_str("v2").unwrap(), AuxFilePolicy::V2);
+        assert_eq!(
+            AuxFilePolicy::from_str("cross-validation").unwrap(),
+            AuxFilePolicy::CrossValidation
+        );
+    }
 }

libs/pageserver_api/src/shard.rs

@@ -125,7 +125,7 @@ impl ShardCount {
 
     /// `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 {
+    pub const fn new(val: u8) -> Self {
         Self(val)
     }
 }
@@ -559,6 +559,14 @@ impl ShardIdentity {
         }
     }
 
+    /// Obtains the shard number and count combined into a `ShardIndex`.
+    pub fn shard_index(&self) -> ShardIndex {
+        ShardIndex {
+            shard_count: self.count,
+            shard_number: self.number,
+        }
+    }
+
     pub fn shard_slug(&self) -> String {
         if self.count > ShardCount(0) {
             format!("-{:02x}{:02x}", self.number.0, self.count.0)

libs/postgres_backend/src/lib.rs

@@ -820,10 +820,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         Ok(ProcessMsgResult::Continue)
     }
 
-    /// Log as info/error result of handling COPY stream and send back
-    /// ErrorResponse if that makes sense. Shutdown the stream if we got
-    /// Terminate. TODO: transition into waiting for Sync msg if we initiate the
-    /// close.
+    /// - Log as info/error result of handling COPY stream and send back
+    ///   ErrorResponse if that makes sense.
+    /// - Shutdown the stream if we got Terminate.
+    /// - Then close the connection because we don't handle exiting from COPY
+    ///   stream normally.
     pub async fn handle_copy_stream_end(&mut self, end: CopyStreamHandlerEnd) {
         use CopyStreamHandlerEnd::*;
 
@@ -849,10 +850,6 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             }
         }
 
-        if let Terminate = &end {
-            self.state = ProtoState::Closed;
-        }
-
         let err_to_send_and_errcode = match &end {
             ServerInitiated(_) => Some((end.to_string(), SQLSTATE_SUCCESSFUL_COMPLETION)),
             Other(_) => Some((format!("{end:#}"), SQLSTATE_INTERNAL_ERROR)),
@@ -882,6 +879,12 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                 error!("failed to send ErrorResponse: {}", ee);
             }
         }
+
+        // Proper COPY stream finishing to continue using the connection is not
+        // implemented at the server side (we don't need it so far). To prevent
+        // further usages of the connection, close it.
+        self.framed.shutdown().await.ok();
+        self.state = ProtoState::Closed;
     }
 }
 

libs/postgres_connection/src/lib.rs

@@ -178,6 +178,13 @@ impl PgConnectionConfig {
     }
 }
 
+impl fmt::Display for PgConnectionConfig {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        // The password is intentionally hidden and not part of this display string.
+        write!(f, "postgresql://{}:{}", self.host, self.port)
+    }
+}
+
 impl fmt::Debug for PgConnectionConfig {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         // We want `password: Some(REDACTED-STRING)`, not `password: Some("REDACTED-STRING")`

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -359,7 +359,7 @@ impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
             // Is there enough space on the page for another logical message and an
             // XLOG_SWITCH? If not, start over.
             let page_remain = XLOG_BLCKSZ as u64 - u64::from(after_lsn) % XLOG_BLCKSZ as u64;
-            if page_remain < base_size - XLOG_SIZE_OF_XLOG_RECORD as u64 {
+            if page_remain < base_size + XLOG_SIZE_OF_XLOG_RECORD as u64 {
                 continue;
             }
 
@@ -373,31 +373,29 @@ impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
                 "SELECT pg_logical_emit_message(false, 'swch', REPEAT('a', $1))",
                 &[&(repeats as i32)],
             )?;
-            break;
-        }
-        info!(
-            "current_wal_insert_lsn={}, XLOG_SIZE_OF_XLOG_RECORD={}",
-            client.pg_current_wal_insert_lsn()?,
-            XLOG_SIZE_OF_XLOG_RECORD
-        );
+            info!(
+                "current_wal_insert_lsn={}, XLOG_SIZE_OF_XLOG_RECORD={}",
+                client.pg_current_wal_insert_lsn()?,
+                XLOG_SIZE_OF_XLOG_RECORD
+            );
 
-        // Emit the XLOG_SWITCH
-        let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
-        let xlog_switch_record_end: PgLsn = client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
-        let next_segment = PgLsn::from(0x0200_0000);
-        ensure!(
-            xlog_switch_record_end < next_segment,
-            "XLOG_SWITCH record ended on or after the expected segment boundary: {} > {}",
-            xlog_switch_record_end,
-            next_segment
-        );
-        ensure!(
-            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
-            "XLOG_SWITCH message ended not on page boundary: {}, offset = {}",
-            xlog_switch_record_end,
-            u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ
-        );
-        Ok(vec![before_xlog_switch, xlog_switch_record_end])
+            // Emit the XLOG_SWITCH
+            let before_xlog_switch = client.pg_current_wal_insert_lsn()?;
+            let xlog_switch_record_end: PgLsn =
+                client.query_one("SELECT pg_switch_wal()", &[])?.get(0);
+
+            if u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ
+                != XLOG_SIZE_OF_XLOG_SHORT_PHD
+            {
+                warn!(
+                    "XLOG_SWITCH message ended not on page boundary: {}, offset = {}, repeating",
+                    xlog_switch_record_end,
+                    u64::from(xlog_switch_record_end) as usize % XLOG_BLCKSZ
+                );
+                continue;
+            }
+            return Ok(vec![before_xlog_switch, xlog_switch_record_end]);
+        }
     }
 }
 

libs/safekeeper_api/src/models.rs

@@ -50,6 +50,9 @@ pub struct SkTimelineInfo {
     pub safekeeper_connstr: Option<String>,
     #[serde(default)]
     pub http_connstr: Option<String>,
+    // Minimum of all active RO replicas flush LSN
+    #[serde(default = "lsn_invalid")]
+    pub standby_horizon: Lsn,
 }
 
 #[derive(Debug, Clone, Deserialize, Serialize)]

libs/utils/src/failpoint_support.rs

@@ -9,6 +9,33 @@ use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
 
+/// Declare a failpoint that can use the `pause` failpoint action.
+/// We don't want to block the executor thread, hence, spawn_blocking + await.
+#[macro_export]
+macro_rules! pausable_failpoint {
+    ($name:literal) => {
+        if cfg!(feature = "testing") {
+            tokio::task::spawn_blocking({
+                let current = tracing::Span::current();
+                move || {
+                    let _entered = current.entered();
+                    tracing::info!("at failpoint {}", $name);
+                    fail::fail_point!($name);
+                }
+            })
+            .await
+            .expect("spawn_blocking");
+        }
+    };
+    ($name:literal, $cond:expr) => {
+        if cfg!(feature = "testing") {
+            if $cond {
+                pausable_failpoint!($name)
+            }
+        }
+    };
+}
+
 /// use with fail::cfg("$name", "return(2000)")
 ///
 /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the

libs/utils/src/sync/gate.rs

@@ -135,7 +135,8 @@ impl Gate {
         let started_at = std::time::Instant::now();
         let mut do_close = std::pin::pin!(self.do_close());
 
-        let nag_after = Duration::from_secs(1);
+        // with 1s we rarely saw anything, let's try if we get more gate closing reasons with 100ms
+        let nag_after = Duration::from_millis(100);
 
         let Err(_timeout) = tokio::time::timeout(nag_after, &mut do_close).await else {
             return;

libs/walproposer/src/walproposer.rs

@@ -496,9 +496,9 @@ mod tests {
                 // TODO: When updating Postgres versions, this test will cause
                 // problems. Postgres version in message needs updating.
                 //
-                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160002, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
+                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160003, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
                 vec![
-                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 2, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 3, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 158, 76, 143, 54, 6, 60, 108, 110,
                     147, 188, 32, 214, 90, 130, 15, 61, 158, 76, 143, 54, 6, 60, 108, 110, 147,
                     188, 32, 214, 90, 130, 15, 61, 1, 0, 0, 0, 0, 0, 0, 1,

pageserver/benches/bench_walredo.rs

@@ -30,47 +30,27 @@
 //! 2024-04-15 on i3en.3xlarge
 //!
 //! ```text
-//! async-short/1           time:   [24.584 µs 24.737 µs 24.922 µs]
-//! async-short/2           time:   [33.479 µs 33.660 µs 33.888 µs]
-//! async-short/4           time:   [42.713 µs 43.046 µs 43.440 µs]
-//! async-short/8           time:   [71.814 µs 72.478 µs 73.240 µs]
-//! async-short/16          time:   [132.73 µs 134.45 µs 136.22 µs]
-//! async-short/32          time:   [258.31 µs 260.73 µs 263.27 µs]
-//! async-short/64          time:   [511.61 µs 514.44 µs 517.51 µs]
-//! async-short/128         time:   [992.64 µs 998.23 µs 1.0042 ms]
-//! async-medium/1          time:   [110.11 µs 110.50 µs 110.96 µs]
-//! async-medium/2          time:   [153.06 µs 153.85 µs 154.99 µs]
-//! async-medium/4          time:   [317.51 µs 319.92 µs 322.85 µs]
-//! async-medium/8          time:   [638.30 µs 644.68 µs 652.12 µs]
-//! async-medium/16         time:   [1.2651 ms 1.2773 ms 1.2914 ms]
-//! async-medium/32         time:   [2.5117 ms 2.5410 ms 2.5720 ms]
-//! async-medium/64         time:   [4.8088 ms 4.8555 ms 4.9047 ms]
-//! async-medium/128        time:   [8.8311 ms 8.9849 ms 9.1263 ms]
-//! sync-short/1            time:   [25.503 µs 25.626 µs 25.771 µs]
-//! sync-short/2            time:   [30.850 µs 31.013 µs 31.208 µs]
-//! sync-short/4            time:   [45.543 µs 45.856 µs 46.193 µs]
-//! sync-short/8            time:   [84.114 µs 84.639 µs 85.220 µs]
-//! sync-short/16           time:   [185.22 µs 186.15 µs 187.13 µs]
-//! sync-short/32           time:   [377.43 µs 378.87 µs 380.46 µs]
-//! sync-short/64           time:   [756.49 µs 759.04 µs 761.70 µs]
-//! sync-short/128          time:   [1.4825 ms 1.4874 ms 1.4923 ms]
-//! sync-medium/1           time:   [105.66 µs 106.01 µs 106.43 µs]
-//! sync-medium/2           time:   [153.10 µs 153.84 µs 154.72 µs]
-//! sync-medium/4           time:   [327.13 µs 329.44 µs 332.27 µs]
-//! sync-medium/8           time:   [654.26 µs 658.73 µs 663.63 µs]
-//! sync-medium/16          time:   [1.2682 ms 1.2748 ms 1.2816 ms]
-//! sync-medium/32          time:   [2.4456 ms 2.4595 ms 2.4731 ms]
-//! sync-medium/64          time:   [4.6523 ms 4.6890 ms 4.7256 ms]
-//! sync-medium/128         time:   [8.7215 ms 8.8323 ms 8.9344 ms]
+//! short/1           time:   [24.584 µs 24.737 µs 24.922 µs]
+//! short/2           time:   [33.479 µs 33.660 µs 33.888 µs]
+//! short/4           time:   [42.713 µs 43.046 µs 43.440 µs]
+//! short/8           time:   [71.814 µs 72.478 µs 73.240 µs]
+//! short/16          time:   [132.73 µs 134.45 µs 136.22 µs]
+//! short/32          time:   [258.31 µs 260.73 µs 263.27 µs]
+//! short/64          time:   [511.61 µs 514.44 µs 517.51 µs]
+//! short/128         time:   [992.64 µs 998.23 µs 1.0042 ms]
+//! medium/1          time:   [110.11 µs 110.50 µs 110.96 µs]
+//! medium/2          time:   [153.06 µs 153.85 µs 154.99 µs]
+//! medium/4          time:   [317.51 µs 319.92 µs 322.85 µs]
+//! medium/8          time:   [638.30 µs 644.68 µs 652.12 µs]
+//! medium/16         time:   [1.2651 ms 1.2773 ms 1.2914 ms]
+//! medium/32         time:   [2.5117 ms 2.5410 ms 2.5720 ms]
+//! medium/64         time:   [4.8088 ms 4.8555 ms 4.9047 ms]
+//! medium/128        time:   [8.8311 ms 8.9849 ms 9.1263 ms]
 //! ```
 
 use bytes::{Buf, Bytes};
 use criterion::{BenchmarkId, Criterion};
-use pageserver::{
-    config::PageServerConf,
-    walrecord::NeonWalRecord,
-    walredo::{PostgresRedoManager, ProcessKind},
-};
+use pageserver::{config::PageServerConf, walrecord::NeonWalRecord, walredo::PostgresRedoManager};
 use pageserver_api::{key::Key, shard::TenantShardId};
 use std::{
     sync::Arc,
@@ -80,39 +60,32 @@ use tokio::{sync::Barrier, task::JoinSet};
 use utils::{id::TenantId, lsn::Lsn};
 
 fn bench(c: &mut Criterion) {
-    for process_kind in &[ProcessKind::Async, ProcessKind::Sync] {
-        {
-            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-            for nclients in nclients {
-                let mut group = c.benchmark_group(format!("{process_kind}-short"));
-                group.bench_with_input(
-                    BenchmarkId::from_parameter(nclients),
-                    &nclients,
-                    |b, nclients| {
-                        let redo_work = Arc::new(Request::short_input());
-                        b.iter_custom(|iters| {
-                            bench_impl(*process_kind, Arc::clone(&redo_work), iters, *nclients)
-                        });
-                    },
-                );
-            }
+    {
+        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+        for nclients in nclients {
+            let mut group = c.benchmark_group("short");
+            group.bench_with_input(
+                BenchmarkId::from_parameter(nclients),
+                &nclients,
+                |b, nclients| {
+                    let redo_work = Arc::new(Request::short_input());
+                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
+                },
+            );
         }
-
-        {
-            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-            for nclients in nclients {
-                let mut group = c.benchmark_group(format!("{process_kind}-medium"));
-                group.bench_with_input(
-                    BenchmarkId::from_parameter(nclients),
-                    &nclients,
-                    |b, nclients| {
-                        let redo_work = Arc::new(Request::medium_input());
-                        b.iter_custom(|iters| {
-                            bench_impl(*process_kind, Arc::clone(&redo_work), iters, *nclients)
-                        });
-                    },
-                );
-            }
+    }
+    {
+        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+        for nclients in nclients {
+            let mut group = c.benchmark_group("medium");
+            group.bench_with_input(
+                BenchmarkId::from_parameter(nclients),
+                &nclients,
+                |b, nclients| {
+                    let redo_work = Arc::new(Request::medium_input());
+                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
+                },
+            );
         }
     }
 }
@@ -120,16 +93,10 @@ criterion::criterion_group!(benches, bench);
 criterion::criterion_main!(benches);
 
 // Returns the sum of each client's wall-clock time spent executing their share of the n_redos.
-fn bench_impl(
-    process_kind: ProcessKind,
-    redo_work: Arc<Request>,
-    n_redos: u64,
-    nclients: u64,
-) -> Duration {
+fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration {
     let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
-    let mut conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
-    conf.walredo_process_kind = process_kind;
+    let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
     let conf = Box::leak(Box::new(conf));
     let tenant_shard_id = TenantShardId::unsharded(TenantId::generate());
 
@@ -158,27 +125,13 @@ fn bench_impl(
         });
     }
 
-    let elapsed = rt.block_on(async move {
+    rt.block_on(async move {
         let mut total_wallclock_time = Duration::ZERO;
         while let Some(res) = tasks.join_next().await {
             total_wallclock_time += res.unwrap();
         }
         total_wallclock_time
-    });
-
-    // consistency check to ensure process kind setting worked
-    if nredos_per_client > 0 {
-        assert_eq!(
-            manager
-                .status()
-                .process
-                .map(|p| p.kind)
-                .expect("the benchmark work causes a walredo process to be spawned"),
-            std::borrow::Cow::Borrowed(process_kind.into())
-        );
-    }
-
-    elapsed
+    })
 }
 
 async fn client(

pageserver/client/src/mgmt_api.rs

@@ -1,8 +1,12 @@
+use std::collections::HashMap;
+
+use bytes::Bytes;
 use pageserver_api::{models::*, shard::TenantShardId};
 use reqwest::{IntoUrl, Method, StatusCode};
 use utils::{
     http::error::HttpErrorBody,
     id::{TenantId, TimelineId},
+    lsn::Lsn,
 };
 
 pub mod util;
@@ -486,6 +490,18 @@ impl Client {
             .map_err(Error::ReceiveBody)
     }
 
+    pub async fn top_tenant_shards(
+        &self,
+        request: TopTenantShardsRequest,
+    ) -> Result<TopTenantShardsResponse> {
+        let uri = format!("{}/v1/top_tenants", self.mgmt_api_endpoint);
+        self.request(Method::POST, uri, request)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
     pub async fn layer_map_info(
         &self,
         tenant_shard_id: TenantShardId,
@@ -549,4 +565,57 @@ impl Client {
             }),
         }
     }
+
+    pub async fn ingest_aux_files(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        aux_files: HashMap<String, String>,
+    ) -> Result<bool> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/ingest_aux_files",
+            self.mgmt_api_endpoint, tenant_shard_id, timeline_id
+        );
+        let resp = self
+            .request_noerror(Method::POST, &uri, IngestAuxFilesRequest { aux_files })
+            .await?;
+        match resp.status() {
+            StatusCode::OK => Ok(true),
+            status => Err(match resp.json::<HttpErrorBody>().await {
+                Ok(HttpErrorBody { msg }) => Error::ApiError(status, msg),
+                Err(_) => {
+                    Error::ReceiveErrorBody(format!("Http error ({}) at {}.", status.as_u16(), uri))
+                }
+            }),
+        }
+    }
+
+    pub async fn list_aux_files(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        lsn: Lsn,
+    ) -> Result<HashMap<String, Bytes>> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/list_aux_files",
+            self.mgmt_api_endpoint, tenant_shard_id, timeline_id
+        );
+        let resp = self
+            .request_noerror(Method::POST, &uri, ListAuxFilesRequest { lsn })
+            .await?;
+        match resp.status() {
+            StatusCode::OK => {
+                let resp: HashMap<String, Bytes> = resp.json().await.map_err(|e| {
+                    Error::ApiError(StatusCode::INTERNAL_SERVER_ERROR, format!("{e}"))
+                })?;
+                Ok(resp)
+            }
+            status => Err(match resp.json::<HttpErrorBody>().await {
+                Ok(HttpErrorBody { msg }) => Error::ApiError(status, msg),
+                Err(_) => {
+                    Error::ReceiveErrorBody(format!("Http error ({}) at {}.", status.as_u16(), uri))
+                }
+            }),
+        }
+    }
 }

pageserver/compaction/src/bin/compaction-simulator.rs

@@ -1,4 +1,5 @@
 use clap::{Parser, Subcommand};
+use pageserver_compaction::helpers::PAGE_SZ;
 use pageserver_compaction::simulator::MockTimeline;
 use rand::Rng;
 use std::io::Write;
@@ -51,7 +52,7 @@ async fn simulate(cmd: &SimulateCmd, results_path: &Path) -> anyhow::Result<()>
     let mut executor = MockTimeline::new();
 
     // Convert the logical size in MB into a key range.
-    let key_range = 0..((cmd.logical_size * 1024 * 1024) / 8192);
+    let key_range = 0..((cmd.logical_size * 1024 * 1024) / PAGE_SZ);
     //let key_range = u64::MIN..u64::MAX;
     println!(
         "starting simulation with key range {:016X}-{:016X}",

pageserver/compaction/src/compact_tiered.rs

@@ -25,7 +25,7 @@ use std::collections::{HashSet, VecDeque};
 use std::ops::Range;
 
 use crate::helpers::{
-    accum_key_values, keyspace_total_size, merge_delta_keys_buffered, overlaps_with,
+    accum_key_values, keyspace_total_size, merge_delta_keys_buffered, overlaps_with, PAGE_SZ,
 };
 use crate::interface::*;
 use utils::lsn::Lsn;
@@ -379,7 +379,7 @@ where
                 .get_keyspace(&job.key_range, job.lsn_range.end, ctx)
                 .await?,
             &self.shard_identity,
-        ) * 8192;
+        ) * PAGE_SZ;
 
         let wal_size = job
             .input_layers
@@ -441,7 +441,7 @@ where
         let mut window = KeyspaceWindow::new(
             E::Key::MIN..E::Key::MAX,
             keyspace,
-            self.target_file_size / 8192,
+            self.target_file_size / PAGE_SZ,
         );
         while let Some(key_range) = window.choose_next_image(&self.shard_identity) {
             new_jobs.push(CompactionJob::<E> {
@@ -530,8 +530,6 @@ where
         // If we have accumulated only a narrow band of keyspace, create an
         // image layer. Otherwise write a delta layer.
 
-        // FIXME: deal with the case of lots of values for same key
-
         // FIXME: we are ignoring images here. Did we already divide the work
         // so that we won't encounter them here?
 
@@ -550,39 +548,94 @@ where
         let mut new_jobs = Vec::new();
 
         // Slide a window through the keyspace
-        let mut key_accum = std::pin::pin!(accum_key_values(key_value_stream));
+        let mut key_accum =
+            std::pin::pin!(accum_key_values(key_value_stream, self.target_file_size));
         let mut all_in_window: bool = false;
         let mut window = Window::new();
+
+        // Helper function to create a job for a new delta layer with given key-lsn
+        // rectangle.
+        let create_delta_job = |key_range, lsn_range: &Range<Lsn>, new_jobs: &mut Vec<_>| {
+            // The inputs for the job are all the input layers of the original job that
+            // overlap with the rectangle.
+            let batch_layers: Vec<LayerId> = job
+                .input_layers
+                .iter()
+                .filter(|layer_id| {
+                    overlaps_with(self.layers[layer_id.0].layer.key_range(), &key_range)
+                })
+                .cloned()
+                .collect();
+            assert!(!batch_layers.is_empty());
+            new_jobs.push(CompactionJob {
+                key_range,
+                lsn_range: lsn_range.clone(),
+                strategy: CompactionStrategy::CreateDelta,
+                input_layers: batch_layers,
+                completed: false,
+            });
+        };
+
         loop {
-            if all_in_window && window.elems.is_empty() {
+            if all_in_window && window.is_empty() {
                 // All done!
                 break;
             }
+
+            // If we now have enough keyspace for next delta layer in the window, create a
+            // new delta layer
             if let Some(key_range) = window.choose_next_delta(self.target_file_size, !all_in_window)
             {
-                let batch_layers: Vec<LayerId> = job
-                    .input_layers
-                    .iter()
-                    .filter(|layer_id| {
-                        overlaps_with(self.layers[layer_id.0].layer.key_range(), &key_range)
-                    })
-                    .cloned()
-                    .collect();
-                assert!(!batch_layers.is_empty());
-                new_jobs.push(CompactionJob {
-                    key_range,
-                    lsn_range: job.lsn_range.clone(),
-                    strategy: CompactionStrategy::CreateDelta,
-                    input_layers: batch_layers,
-                    completed: false,
-                });
-            } else {
-                assert!(!all_in_window);
-                if let Some(next_key) = key_accum.next().await.transpose()? {
-                    window.feed(next_key.key, next_key.size);
-                } else {
+                create_delta_job(key_range, &job.lsn_range, &mut new_jobs);
+                continue;
+            }
+            assert!(!all_in_window);
+
+            // Process next key in the key space
+            match key_accum.next().await.transpose()? {
+                None => {
                     all_in_window = true;
                 }
+                Some(next_key) if next_key.partition_lsns.is_empty() => {
+                    // Normal case: extend the window by the key
+                    window.feed(next_key.key, next_key.size);
+                }
+                Some(next_key) => {
+                    // A key with too large size impact for a single delta layer. This
+                    // case occurs if you make a huge number of updates for a single key.
+                    //
+                    // Drain the window with has_more = false to make a clean cut before
+                    // the key, and then make dedicated delta layers for the single key.
+                    //
+                    // We cannot cluster the key with the others, because we don't want
+                    // layer files to overlap with each other in the lsn,key space (no
+                    // overlaps for the rectangles).
+                    let key = next_key.key;
+                    debug!("key {key} with size impact larger than the layer size");
+                    while !window.is_empty() {
+                        let has_more = false;
+                        let key_range = window.choose_next_delta(self.target_file_size, has_more)
+                            .expect("with has_more==false, choose_next_delta always returns something for a non-empty Window");
+                        create_delta_job(key_range, &job.lsn_range, &mut new_jobs);
+                    }
+
+                    // Not really required: but here for future resilience:
+                    // We make a "gap" here, so any structure the window holds should
+                    // probably be reset.
+                    window = Window::new();
+
+                    let mut prior_lsn = job.lsn_range.start;
+                    let mut lsn_ranges = Vec::new();
+                    for (lsn, _size) in next_key.partition_lsns.iter() {
+                        lsn_ranges.push(prior_lsn..*lsn);
+                        prior_lsn = *lsn;
+                    }
+                    lsn_ranges.push(prior_lsn..job.lsn_range.end);
+                    for lsn_range in lsn_ranges {
+                        let key_range = key..key.next();
+                        create_delta_job(key_range, &lsn_range, &mut new_jobs);
+                    }
+                }
             }
         }
 
@@ -610,8 +663,8 @@ where
     }
 }
 
-// Sliding window through keyspace and values
-// This is used by over_with_images to decide on good split points
+/// Sliding window through keyspace and values for image layer
+/// This is used by [`LevelCompactionState::cover_with_images`] to decide on good split points
 struct KeyspaceWindow<K> {
     head: KeyspaceWindowHead<K>,
 
@@ -751,9 +804,9 @@ struct WindowElement<K> {
     accum_size: u64,
 }
 
-// Sliding window through keyspace and values
-//
-// This is used to decide what layer to write next, from the beginning of the window.
+/// Sliding window through keyspace and values for delta layer tiling
+///
+/// This is used to decide which delta layer to write next.
 struct Window<K> {
     elems: VecDeque<WindowElement<K>>,
 
@@ -777,11 +830,13 @@ where
     fn feed(&mut self, key: K, size: u64) {
         let last_size;
         if let Some(last) = self.elems.back_mut() {
-            assert!(last.last_key <= key);
-            if key == last.last_key {
-                last.accum_size += size;
-                return;
-            }
+            // We require the keys to be strictly increasing for the window.
+            // Keys should already have been deduplicated by `accum_key_values`
+            assert!(
+                last.last_key < key,
+                "last_key(={}) >= key(={key})",
+                last.last_key
+            );
             last_size = last.accum_size;
         } else {
             last_size = 0;
@@ -803,6 +858,10 @@ where
         self.elems.front().unwrap().accum_size - self.splitoff_size
     }
 
+    fn is_empty(&self) -> bool {
+        self.elems.is_empty()
+    }
+
     fn commit_upto(&mut self, mut upto: usize) {
         while upto > 1 {
             let popped = self.elems.pop_front().unwrap();
@@ -865,7 +924,7 @@ where
         // If we're willing to stretch it up to 1.25 target size, could we
         // gobble up the rest of the work? This avoids creating very small
         // "tail" layers at the end of the keyspace
-        if !has_more && self.remain_size() < target_size * 5 / 3 {
+        if !has_more && self.remain_size() < target_size * 5 / 4 {
             self.commit_upto(self.elems.len());
         } else {
             let delta_split_at = self.find_size_split(target_size);

pageserver/compaction/src/helpers.rs

@@ -16,6 +16,8 @@ use std::pin::Pin;
 use std::task::{ready, Poll};
 use utils::lsn::Lsn;
 
+pub const PAGE_SZ: u64 = 8192;
+
 pub fn keyspace_total_size<K>(
     keyspace: &CompactionKeySpace<K>,
     shard_identity: &ShardIdentity,
@@ -235,9 +237,14 @@ pub struct KeySize<K> {
     pub key: K,
     pub num_values: u64,
     pub size: u64,
+    /// The lsns to partition at (if empty then no per-lsn partitioning)
+    pub partition_lsns: Vec<(Lsn, u64)>,
 }
 
-pub fn accum_key_values<'a, I, K, D, E>(input: I) -> impl Stream<Item = Result<KeySize<K>, E>>
+pub fn accum_key_values<'a, I, K, D, E>(
+    input: I,
+    target_size: u64,
+) -> impl Stream<Item = Result<KeySize<K>, E>>
 where
     K: Eq + PartialOrd + Display + Copy,
     I: Stream<Item = Result<D, E>>,
@@ -249,25 +256,35 @@ where
 
         if let Some(first) = input.next().await {
             let first = first?;
+            let mut part_size = first.size();
             let mut accum: KeySize<K> = KeySize {
                 key: first.key(),
                 num_values: 1,
-                size: first.size(),
+                size: part_size,
+                partition_lsns: Vec::new(),
             };
             let mut last_key = accum.key;
             while let Some(this) = input.next().await {
                 let this = this?;
                 if this.key() == accum.key {
-                    accum.size += this.size();
+                    let add_size = this.size();
+                    if part_size + add_size > target_size {
+                        accum.partition_lsns.push((this.lsn(), part_size));
+                        part_size = 0;
+                    }
+                    part_size += add_size;
+                    accum.size += add_size;
                     accum.num_values += 1;
                 } else {
                     assert!(last_key <= accum.key, "last_key={last_key} <= accum.key={}", accum.key);
                     last_key = accum.key;
                     yield accum;
+                    part_size = this.size();
                     accum = KeySize {
                         key: this.key(),
                         num_values: 1,
-                        size: this.size(),
+                        size: part_size,
+                        partition_lsns: Vec::new(),
                     };
                 }
             }

pageserver/compaction/src/identify_levels.rs

@@ -184,6 +184,12 @@ impl<L> Level<L> {
         }
         let mut events: Vec<Event<K>> = Vec::new();
         for (idx, l) in self.layers.iter().enumerate() {
+            let key_range = l.key_range();
+            if key_range.end == key_range.start.next() && l.is_delta() {
+                // Ignore single-key delta layers as they can be stacked on top of each other
+                // as that is the only way to cut further.
+                continue;
+            }
             events.push(Event {
                 key: l.key_range().start,
                 layer_idx: idx,

pageserver/compaction/src/simulator.rs

@@ -14,6 +14,7 @@ use std::ops::Range;
 use std::sync::Arc;
 use std::sync::Mutex;
 
+use crate::helpers::PAGE_SZ;
 use crate::helpers::{merge_delta_keys, overlaps_with};
 
 use crate::interface;
@@ -379,8 +380,8 @@ impl interface::CompactionLayer<Key> for MockLayer {
     }
     fn file_size(&self) -> u64 {
         match self {
-            MockLayer::Delta(this) => this.file_size(),
-            MockLayer::Image(this) => this.file_size(),
+            MockLayer::Delta(this) => this.file_size,
+            MockLayer::Image(this) => this.file_size,
         }
     }
     fn short_id(&self) -> String {
@@ -509,7 +510,7 @@ impl interface::CompactionJobExecutor for MockTimeline {
         let new_layer = Arc::new(MockImageLayer {
             key_range: key_range.clone(),
             lsn_range: lsn..lsn,
-            file_size: accum_size * 8192,
+            file_size: accum_size * PAGE_SZ,
             deleted: Mutex::new(false),
         });
         info!(

pageserver/compaction/tests/tests.rs

@@ -20,10 +20,6 @@ pub(crate) fn setup_logging() {
 /// even if we produce an extremely narrow delta layer, spanning just that one
 /// key, we still too many records to fit in the target file size. We need to
 /// split in the LSN dimension too in that case.
-///
-/// TODO: The code to avoid this problem has not been implemented yet! So the
-/// assertion currently fails, but we need to make it not fail.
-#[ignore]
 #[tokio::test]
 async fn test_many_updates_for_single_key() {
     setup_logging();
@@ -43,9 +39,9 @@ async fn test_many_updates_for_single_key() {
     }
     for l in executor.live_layers.iter() {
         assert!(l.file_size() < executor.target_file_size * 2);
-        // sanity check that none of the delta layers are stupidly small either
+        // Sanity check that none of the delta layers are empty either.
         if l.is_delta() {
-            assert!(l.file_size() > executor.target_file_size / 2);
+            assert!(l.file_size() > 0);
         }
     }
 }

pageserver/ctl/src/draw_timeline_dir.rs

@@ -52,7 +52,6 @@
 
 use anyhow::{Context, Result};
 use pageserver::repository::Key;
-use pageserver::METADATA_FILE_NAME;
 use std::cmp::Ordering;
 use std::io::{self, BufRead};
 use std::path::PathBuf;
@@ -83,6 +82,11 @@ fn parse_filename(name: &str) -> (Range<Key>, Range<Lsn>) {
     let split: Vec<&str> = name.split("__").collect();
     let keys: Vec<&str> = split[0].split('-').collect();
     let mut lsns: Vec<&str> = split[1].split('-').collect();
+
+    if lsns.last().expect("should").len() == 8 {
+        lsns.pop();
+    }
+
     if lsns.len() == 1 {
         lsns.push(lsns[0]);
     }
@@ -154,10 +158,6 @@ pub fn main() -> Result<()> {
         let line = PathBuf::from_str(&line).unwrap();
         let filename = line.file_name().unwrap();
         let filename = filename.to_str().unwrap();
-        if filename == METADATA_FILE_NAME {
-            // Don't try and parse "metadata" like a key-lsn range
-            continue;
-        }
         let (key_range, lsn_range) = parse_filename(filename);
         files.push(Layer {
             filename: filename.to_owned(),

pageserver/ctl/src/index_part.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 
 use anyhow::Context;
 use camino::Utf8PathBuf;
-use pageserver::tenant::remote_timeline_client::index::IndexLayerMetadata;
+use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata;
 use pageserver::tenant::storage_layer::LayerName;
 use pageserver::tenant::{metadata::TimelineMetadata, IndexPart};
 use utils::lsn::Lsn;
@@ -19,7 +19,7 @@ pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
             let des: IndexPart = IndexPart::from_s3_bytes(&bytes).context("deserialize")?;
             #[derive(serde::Serialize)]
             struct Output<'a> {
-                layer_metadata: &'a HashMap<LayerName, IndexLayerMetadata>,
+                layer_metadata: &'a HashMap<LayerName, LayerFileMetadata>,
                 disk_consistent_lsn: Lsn,
                 timeline_metadata: &'a TimelineMetadata,
             }

pageserver/ctl/src/main.rs

@@ -219,6 +219,7 @@ fn handle_metadata(
     let mut meta = TimelineMetadata::from_bytes(&metadata_bytes)?;
     println!("Current metadata:\n{meta:?}");
     let mut update_meta = false;
+    // TODO: simplify this part
     if let Some(disk_consistent_lsn) = disk_consistent_lsn {
         meta = TimelineMetadata::new(
             *disk_consistent_lsn,

pageserver/pagebench/src/cmd/aux_files.rs

@@ -0,0 +1,98 @@
+use pageserver_api::models::{AuxFilePolicy, TenantConfig, TenantConfigRequest};
+use pageserver_api::shard::TenantShardId;
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+/// Ingest aux files into the pageserver.
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    #[clap(long, default_value = "http://localhost:9898")]
+    mgmt_api_endpoint: String,
+    #[clap(long, default_value = "postgres://postgres@localhost:64000")]
+    page_service_connstring: String,
+    #[clap(long)]
+    pageserver_jwt: Option<String>,
+
+    targets: Option<Vec<TenantTimelineId>>,
+}
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()
+        .unwrap();
+
+    let main_task = rt.spawn(main_impl(args));
+    rt.block_on(main_task).unwrap()
+}
+
+async fn main_impl(args: Args) -> anyhow::Result<()> {
+    let args: &'static Args = Box::leak(Box::new(args));
+
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        args.mgmt_api_endpoint.clone(),
+        args.pageserver_jwt.as_deref(),
+    ));
+
+    // discover targets
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: None,
+            targets: {
+                if let Some(targets) = &args.targets {
+                    if targets.len() != 1 {
+                        anyhow::bail!("must specify exactly one target");
+                    }
+                    Some(targets.clone())
+                } else {
+                    None
+                }
+            },
+        },
+    )
+    .await?;
+
+    let timeline = timelines[0];
+    let tenant_shard_id = TenantShardId::unsharded(timeline.tenant_id);
+    let timeline_id = timeline.timeline_id;
+
+    println!("operating on timeline {}", timeline);
+
+    mgmt_api_client
+        .tenant_config(&TenantConfigRequest {
+            tenant_id: timeline.tenant_id,
+            config: TenantConfig {
+                switch_aux_file_policy: Some(AuxFilePolicy::V2),
+                ..Default::default()
+            },
+        })
+        .await?;
+
+    for batch in 0..100 {
+        let items = (0..100)
+            .map(|id| {
+                (
+                    format!("pg_logical/mappings/{:03}.{:03}", batch, id),
+                    format!("{:08}", id),
+                )
+            })
+            .collect::<HashMap<_, _>>();
+        let file_cnt = items.len();
+        mgmt_api_client
+            .ingest_aux_files(tenant_shard_id, timeline_id, items)
+            .await?;
+        println!("ingested {file_cnt} files");
+    }
+
+    let files = mgmt_api_client
+        .list_aux_files(tenant_shard_id, timeline_id, Lsn(Lsn::MAX.0 - 1))
+        .await?;
+
+    println!("{} files found", files.len());
+
+    anyhow::Ok(())
+}

pageserver/pagebench/src/cmd/ondemand_download_churn.rs

@@ -2,9 +2,11 @@ use pageserver_api::{models::HistoricLayerInfo, shard::TenantShardId};
 
 use pageserver_client::mgmt_api;
 use rand::seq::SliceRandom;
+use tokio_util::sync::CancellationToken;
 use tracing::{debug, info};
 use utils::id::{TenantTimelineId, TimelineId};
 
+use std::{f64, sync::Arc};
 use tokio::{
     sync::{mpsc, OwnedSemaphorePermit},
     task::JoinSet,
@@ -12,10 +14,7 @@ use tokio::{
 
 use std::{
     num::NonZeroUsize,
-    sync::{
-        atomic::{AtomicU64, Ordering},
-        Arc,
-    },
+    sync::atomic::{AtomicU64, Ordering},
     time::{Duration, Instant},
 };
 
@@ -51,19 +50,31 @@ pub(crate) fn main(args: Args) -> anyhow::Result<()> {
     Ok(())
 }
 
+#[derive(serde::Serialize)]
+struct Output {
+    downloads_count: u64,
+    downloads_bytes: u64,
+    evictions_count: u64,
+    timeline_restarts: u64,
+    #[serde(with = "humantime_serde")]
+    runtime: Duration,
+}
+
 #[derive(Debug, Default)]
 struct LiveStats {
-    evictions: AtomicU64,
-    downloads: AtomicU64,
+    evictions_count: AtomicU64,
+    downloads_count: AtomicU64,
+    downloads_bytes: AtomicU64,
     timeline_restarts: AtomicU64,
 }
 
 impl LiveStats {
     fn eviction_done(&self) {
-        self.evictions.fetch_add(1, Ordering::Relaxed);
+        self.evictions_count.fetch_add(1, Ordering::Relaxed);
     }
-    fn download_done(&self) {
-        self.downloads.fetch_add(1, Ordering::Relaxed);
+    fn download_done(&self, size: u64) {
+        self.downloads_count.fetch_add(1, Ordering::Relaxed);
+        self.downloads_bytes.fetch_add(size, Ordering::Relaxed);
     }
     fn timeline_restart_done(&self) {
         self.timeline_restarts.fetch_add(1, Ordering::Relaxed);
@@ -92,28 +103,49 @@ async fn main_impl(args: Args) -> anyhow::Result<()> {
     )
     .await?;
 
+    let token = CancellationToken::new();
     let mut tasks = JoinSet::new();
 
-    let live_stats = Arc::new(LiveStats::default());
+    let periodic_stats = Arc::new(LiveStats::default());
+    let total_stats = Arc::new(LiveStats::default());
+
+    let start = Instant::now();
     tasks.spawn({
-        let live_stats = Arc::clone(&live_stats);
+        let periodic_stats = Arc::clone(&periodic_stats);
+        let total_stats = Arc::clone(&total_stats);
+        let cloned_token = token.clone();
         async move {
             let mut last_at = Instant::now();
             loop {
+                if cloned_token.is_cancelled() {
+                    return;
+                }
                 tokio::time::sleep_until((last_at + Duration::from_secs(1)).into()).await;
                 let now = Instant::now();
                 let delta: Duration = now - last_at;
                 last_at = now;
 
                 let LiveStats {
-                    evictions,
-                    downloads,
+                    evictions_count,
+                    downloads_count,
+                    downloads_bytes,
                     timeline_restarts,
-                } = &*live_stats;
-                let evictions = evictions.swap(0, Ordering::Relaxed) as f64 / delta.as_secs_f64();
-                let downloads = downloads.swap(0, Ordering::Relaxed) as f64 / delta.as_secs_f64();
+                } = &*periodic_stats;
+                let evictions_count = evictions_count.swap(0, Ordering::Relaxed);
+                let downloads_count = downloads_count.swap(0, Ordering::Relaxed);
+                let downloads_bytes = downloads_bytes.swap(0, Ordering::Relaxed);
                 let timeline_restarts = timeline_restarts.swap(0, Ordering::Relaxed);
-                info!("evictions={evictions:.2}/s downloads={downloads:.2}/s timeline_restarts={timeline_restarts}");
+
+                total_stats.evictions_count.fetch_add(evictions_count, Ordering::Relaxed);
+                total_stats.downloads_count.fetch_add(downloads_count, Ordering::Relaxed);
+                total_stats.downloads_bytes.fetch_add(downloads_bytes, Ordering::Relaxed);
+                total_stats.timeline_restarts.fetch_add(timeline_restarts, Ordering::Relaxed);
+
+                let evictions_per_s = evictions_count as f64 / delta.as_secs_f64();
+                let downloads_per_s = downloads_count as f64 / delta.as_secs_f64();
+                let downloads_mibs_per_s = downloads_bytes as f64 / delta.as_secs_f64() / ((1 << 20) as f64);
+
+                info!("evictions={evictions_per_s:.2}/s downloads={downloads_per_s:.2}/s download_bytes={downloads_mibs_per_s:.2}MiB/s timeline_restarts={timeline_restarts}");
             }
         }
     });
@@ -124,14 +156,42 @@ async fn main_impl(args: Args) -> anyhow::Result<()> {
                 args,
                 Arc::clone(&mgmt_api_client),
                 tl,
-                Arc::clone(&live_stats),
+                Arc::clone(&periodic_stats),
+                token.clone(),
             ));
         }
     }
+    if let Some(runtime) = args.runtime {
+        tokio::spawn(async move {
+            tokio::time::sleep(runtime.into()).await;
+            token.cancel();
+        });
+    }
 
     while let Some(res) = tasks.join_next().await {
         res.unwrap();
     }
+    let end = Instant::now();
+    let duration: Duration = end - start;
+
+    let output = {
+        let LiveStats {
+            evictions_count,
+            downloads_count,
+            downloads_bytes,
+            timeline_restarts,
+        } = &*total_stats;
+        Output {
+            downloads_count: downloads_count.load(Ordering::Relaxed),
+            downloads_bytes: downloads_bytes.load(Ordering::Relaxed),
+            evictions_count: evictions_count.load(Ordering::Relaxed),
+            timeline_restarts: timeline_restarts.load(Ordering::Relaxed),
+            runtime: duration,
+        }
+    };
+    let output = serde_json::to_string_pretty(&output).unwrap();
+    println!("{output}");
+
     Ok(())
 }
 
@@ -140,6 +200,7 @@ async fn timeline_actor(
     mgmt_api_client: Arc<pageserver_client::mgmt_api::Client>,
     timeline: TenantTimelineId,
     live_stats: Arc<LiveStats>,
+    token: CancellationToken,
 ) {
     // TODO: support sharding
     let tenant_shard_id = TenantShardId::unsharded(timeline.tenant_id);
@@ -149,7 +210,7 @@ async fn timeline_actor(
         layers: Vec<mpsc::Sender<OwnedSemaphorePermit>>,
         concurrency: Arc<tokio::sync::Semaphore>,
     }
-    loop {
+    while !token.is_cancelled() {
         debug!("restarting timeline");
         let layer_map_info = mgmt_api_client
             .layer_map_info(tenant_shard_id, timeline.timeline_id)
@@ -185,7 +246,7 @@ async fn timeline_actor(
 
         live_stats.timeline_restart_done();
 
-        loop {
+        while !token.is_cancelled() {
             assert!(!timeline.joinset.is_empty());
             if let Some(res) = timeline.joinset.try_join_next() {
                 debug!(?res, "a layer actor exited, should not happen");
@@ -255,7 +316,7 @@ async fn layer_actor(
                     .layer_ondemand_download(tenant_shard_id, timeline_id, layer.layer_file_name())
                     .await
                     .unwrap();
-                live_stats.download_done();
+                live_stats.download_done(layer.layer_file_size());
                 did_it
             }
         };

pageserver/pagebench/src/main.rs

@@ -14,6 +14,7 @@ mod util {
 
 /// The pagebench CLI sub-commands, dispatched in [`main`] below.
 mod cmd {
+    pub(super) mod aux_files;
     pub(super) mod basebackup;
     pub(super) mod getpage_latest_lsn;
     pub(super) mod ondemand_download_churn;
@@ -27,6 +28,7 @@ enum Args {
     GetPageLatestLsn(cmd::getpage_latest_lsn::Args),
     TriggerInitialSizeCalculation(cmd::trigger_initial_size_calculation::Args),
     OndemandDownloadChurn(cmd::ondemand_download_churn::Args),
+    AuxFiles(cmd::aux_files::Args),
 }
 
 fn main() {
@@ -46,6 +48,7 @@ fn main() {
             cmd::trigger_initial_size_calculation::main(args)
         }
         Args::OndemandDownloadChurn(args) => cmd::ondemand_download_churn::main(args),
+        Args::AuxFiles(args) => cmd::aux_files::main(args),
     }
     .unwrap()
 }

pageserver/src/aux_file.rs

@@ -1,15 +1,39 @@
+use std::sync::Arc;
+
+use ::metrics::IntGauge;
 use bytes::{Buf, BufMut, Bytes};
 use pageserver_api::key::{Key, AUX_KEY_PREFIX, METADATA_KEY_SIZE};
 use tracing::warn;
 
-/// Create a metadata key from a hash, encoded as [AUX_KEY_PREFIX, 2B directory prefix, first 13B of 128b xxhash].
+// BEGIN Copyright (c) 2017 Servo Contributors
+
+/// Const version of FNV hash.
+#[inline]
+#[must_use]
+pub const fn fnv_hash(bytes: &[u8]) -> u128 {
+    const INITIAL_STATE: u128 = 0x6c62272e07bb014262b821756295c58d;
+    const PRIME: u128 = 0x0000000001000000000000000000013B;
+
+    let mut hash = INITIAL_STATE;
+    let mut i = 0;
+    while i < bytes.len() {
+        hash ^= bytes[i] as u128;
+        hash = hash.wrapping_mul(PRIME);
+        i += 1;
+    }
+    hash
+}
+
+// END Copyright (c) 2017 Servo Contributors
+
+/// Create a metadata key from a hash, encoded as [AUX_KEY_PREFIX, 2B directory prefix, least significant 13B of FNV hash].
 fn aux_hash_to_metadata_key(dir_level1: u8, dir_level2: u8, data: &[u8]) -> Key {
-    let mut key = [0; METADATA_KEY_SIZE];
-    let hash = twox_hash::xxh3::hash128(data).to_be_bytes();
+    let mut key: [u8; 16] = [0; METADATA_KEY_SIZE];
+    let hash = fnv_hash(data).to_be_bytes();
     key[0] = AUX_KEY_PREFIX;
     key[1] = dir_level1;
     key[2] = dir_level2;
-    key[3..16].copy_from_slice(&hash[0..13]);
+    key[3..16].copy_from_slice(&hash[3..16]);
     Key::from_metadata_key_fixed_size(&key)
 }
 
@@ -140,6 +164,55 @@ pub fn encode_file_value(files: &[(&str, &[u8])]) -> anyhow::Result<Vec<u8>> {
     Ok(encoded)
 }
 
+/// An estimation of the size of aux files.
+pub struct AuxFileSizeEstimator {
+    aux_file_size_gauge: IntGauge,
+    size: Arc<std::sync::Mutex<Option<isize>>>,
+}
+
+impl AuxFileSizeEstimator {
+    pub fn new(aux_file_size_gauge: IntGauge) -> Self {
+        Self {
+            aux_file_size_gauge,
+            size: Arc::new(std::sync::Mutex::new(None)),
+        }
+    }
+
+    pub fn on_base_backup(&self, new_size: usize) {
+        let mut guard = self.size.lock().unwrap();
+        *guard = Some(new_size as isize);
+        self.report(new_size as isize);
+    }
+
+    pub fn on_add(&self, file_size: usize) {
+        let mut guard = self.size.lock().unwrap();
+        if let Some(size) = &mut *guard {
+            *size += file_size as isize;
+            self.report(*size);
+        }
+    }
+
+    pub fn on_remove(&self, file_size: usize) {
+        let mut guard = self.size.lock().unwrap();
+        if let Some(size) = &mut *guard {
+            *size -= file_size as isize;
+            self.report(*size);
+        }
+    }
+
+    pub fn on_update(&self, old_size: usize, new_size: usize) {
+        let mut guard = self.size.lock().unwrap();
+        if let Some(size) = &mut *guard {
+            *size += new_size as isize - old_size as isize;
+            self.report(*size);
+        }
+    }
+
+    pub fn report(&self, size: isize) {
+        self.aux_file_size_gauge.set(size as i64);
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -148,15 +221,19 @@ mod tests {
     fn test_hash_portable() {
         // AUX file encoding requires the hash to be portable across all platforms. This test case checks
         // if the algorithm produces the same hash across different environments.
+
         assert_eq!(
-            305317690835051308206966631765527126151,
-            twox_hash::xxh3::hash128("test1".as_bytes())
+            265160408618497461376862998434862070044,
+            super::fnv_hash("test1".as_bytes())
         );
         assert_eq!(
-            85104974691013376326742244813280798847,
-            twox_hash::xxh3::hash128("test/test2".as_bytes())
+            295486155126299629456360817749600553988,
+            super::fnv_hash("test/test2".as_bytes())
+        );
+        assert_eq!(
+            144066263297769815596495629667062367629,
+            super::fnv_hash("".as_bytes())
         );
-        assert_eq!(0, twox_hash::xxh3::hash128("".as_bytes()));
     }
 
     #[test]
@@ -164,28 +241,28 @@ mod tests {
         // To correct retrieve AUX files, the generated keys for the same file must be the same for all versions
         // of the page server.
         assert_eq!(
-            "6200000101E5B20C5F8DD5AA3289D6D9EAFA",
-            encode_aux_file_key("pg_logical/mappings/test1").to_string()
+            "62000001017F8B83D94F7081693471ABF91C",
+            encode_aux_file_key("pg_logical/mappings/test1").to_string(),
         );
         assert_eq!(
-            "620000010239AAC544893139B26F501B97E6",
-            encode_aux_file_key("pg_logical/snapshots/test2").to_string()
+            "62000001027F8E83D94F7081693471ABFCCD",
+            encode_aux_file_key("pg_logical/snapshots/test2").to_string(),
         );
         assert_eq!(
-            "620000010300000000000000000000000000",
-            encode_aux_file_key("pg_logical/replorigin_checkpoint").to_string()
+            "62000001032E07BB014262B821756295C58D",
+            encode_aux_file_key("pg_logical/replorigin_checkpoint").to_string(),
         );
         assert_eq!(
-            "62000001FF8635AF2134B7266EC5B4189FD6",
-            encode_aux_file_key("pg_logical/unsupported").to_string()
+            "62000001FF4F38E1C74754E7D03C1A660178",
+            encode_aux_file_key("pg_logical/unsupported").to_string(),
         );
         assert_eq!(
-            "6200000201772D0E5D71DE14DA86142A1619",
+            "62000002017F8D83D94F7081693471ABFB92",
             encode_aux_file_key("pg_replslot/test3").to_string()
         );
         assert_eq!(
-            "620000FFFF1866EBEB53B807B26A2416F317",
-            encode_aux_file_key("other_file_not_supported").to_string()
+            "620000FFFF2B6ECC8AEF93F643DC44F15E03",
+            encode_aux_file_key("other_file_not_supported").to_string(),
         );
     }
 

pageserver/src/bin/pageserver.rs

@@ -284,7 +284,6 @@ fn start_pageserver(
     ))
     .unwrap();
     pageserver::preinitialize_metrics();
-    pageserver::metrics::wal_redo::set_process_kind_metric(conf.walredo_process_kind);
 
     // If any failpoints were set from FAILPOINTS environment variable,
     // print them to the log for debugging purposes
@@ -383,7 +382,7 @@ fn start_pageserver(
     let shutdown_pageserver = tokio_util::sync::CancellationToken::new();
 
     // Set up remote storage client
-    let remote_storage = Some(create_remote_storage_client(conf)?);
+    let remote_storage = create_remote_storage_client(conf)?;
 
     // Set up deletion queue
     let (deletion_queue, deletion_workers) = DeletionQueue::new(
@@ -516,16 +515,12 @@ fn start_pageserver(
         }
     });
 
-    let secondary_controller = if let Some(remote_storage) = &remote_storage {
-        secondary::spawn_tasks(
-            tenant_manager.clone(),
-            remote_storage.clone(),
-            background_jobs_barrier.clone(),
-            shutdown_pageserver.clone(),
-        )
-    } else {
-        secondary::null_controller()
-    };
+    let secondary_controller = secondary::spawn_tasks(
+        tenant_manager.clone(),
+        remote_storage.clone(),
+        background_jobs_barrier.clone(),
+        shutdown_pageserver.clone(),
+    );
 
     // shared state between the disk-usage backed eviction background task and the http endpoint
     // that allows triggering disk-usage based eviction manually. note that the http endpoint
@@ -533,15 +528,13 @@ fn start_pageserver(
     // been configured.
     let disk_usage_eviction_state: Arc<disk_usage_eviction_task::State> = Arc::default();
 
-    if let Some(remote_storage) = &remote_storage {
-        launch_disk_usage_global_eviction_task(
-            conf,
-            remote_storage.clone(),
-            disk_usage_eviction_state.clone(),
-            tenant_manager.clone(),
-            background_jobs_barrier.clone(),
-        )?;
-    }
+    launch_disk_usage_global_eviction_task(
+        conf,
+        remote_storage.clone(),
+        disk_usage_eviction_state.clone(),
+        tenant_manager.clone(),
+        background_jobs_barrier.clone(),
+    )?;
 
     // Start up the service to handle HTTP mgmt API request. We created the
     // listener earlier already.
@@ -654,17 +647,20 @@ fn start_pageserver(
             None,
             "libpq endpoint listener",
             true,
-            async move {
-                page_service::libpq_listener_main(
-                    conf,
-                    broker_client,
-                    pg_auth,
-                    pageserver_listener,
-                    conf.pg_auth_type,
-                    libpq_ctx,
-                    task_mgr::shutdown_token(),
-                )
-                .await
+            {
+                let tenant_manager = tenant_manager.clone();
+                async move {
+                    page_service::libpq_listener_main(
+                        tenant_manager,
+                        broker_client,
+                        pg_auth,
+                        pageserver_listener,
+                        conf.pg_auth_type,
+                        libpq_ctx,
+                        task_mgr::shutdown_token(),
+                    )
+                    .await
+                }
             },
         );
     }
@@ -693,14 +689,7 @@ fn start_pageserver(
             // Right now that tree doesn't reach very far, and `task_mgr` is used instead.
             // The plan is to change that over time.
             shutdown_pageserver.take();
-            let bg_remote_storage = remote_storage.clone();
-            let bg_deletion_queue = deletion_queue.clone();
-            pageserver::shutdown_pageserver(
-                &tenant_manager,
-                bg_remote_storage.map(|_| bg_deletion_queue),
-                0,
-            )
-            .await;
+            pageserver::shutdown_pageserver(&tenant_manager, deletion_queue.clone(), 0).await;
             unreachable!()
         })
     }

pageserver/src/config.rs

@@ -5,7 +5,7 @@
 //! See also `settings.md` for better description on every parameter.
 
 use anyhow::{anyhow, bail, ensure, Context, Result};
-use pageserver_api::shard::TenantShardId;
+use pageserver_api::{models::CompactionAlgorithm, shard::TenantShardId};
 use remote_storage::{RemotePath, RemoteStorageConfig};
 use serde;
 use serde::de::IntoDeserializer;
@@ -15,7 +15,7 @@ use utils::crashsafe::path_with_suffix_extension;
 use utils::id::ConnectionId;
 use utils::logging::SecretString;
 
-use once_cell::sync::OnceCell;
+use once_cell::sync::{Lazy, OnceCell};
 use reqwest::Url;
 use std::num::NonZeroUsize;
 use std::str::FromStr;
@@ -99,7 +99,7 @@ pub mod defaults {
 
     pub const DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB: usize = 0;
 
-    pub const DEFAULT_WALREDO_PROCESS_KIND: &str = "sync";
+    pub const DEFAULT_WALREDO_PROCESS_KIND: &str = "async";
 
     ///
     /// Default built-in configuration file.
@@ -1067,6 +1067,19 @@ impl PageServerConf {
 
         conf.default_tenant_conf = t_conf.merge(TenantConf::default());
 
+        {
+            const VAR_NAME: &str = "NEON_PAGESERVER_PANIC_ON_UNSPECIFIED_COMPACTION_ALGORITHM";
+            static VAR: Lazy<Option<bool>> = Lazy::new(|| utils::env::var(VAR_NAME));
+            if VAR.unwrap_or(false)
+                && conf.default_tenant_conf.compaction_algorithm.kind
+                    == CompactionAlgorithm::NotSpecified
+            {
+                panic!(
+                        "Unspecified compaction algorithm in default tenant configuration. \
+                        Set the algorithm explicitly in the pageserver.toml's `tenant_config` field or unset the environment variable {VAR_NAME}");
+            }
+        }
+
         Ok(conf)
     }
 

pageserver/src/deletion_queue.rs

@@ -632,7 +632,7 @@ impl DeletionQueue {
     ///
     /// If remote_storage is None, then the returned workers will also be None.
     pub fn new<C>(
-        remote_storage: Option<GenericRemoteStorage>,
+        remote_storage: GenericRemoteStorage,
         control_plane_client: Option<C>,
         conf: &'static PageServerConf,
     ) -> (Self, Option<DeletionQueueWorkers<C>>)
@@ -658,23 +658,6 @@ impl DeletionQueue {
         // longer to flush after Tenants have all been torn down.
         let cancel = CancellationToken::new();
 
-        let remote_storage = match remote_storage {
-            None => {
-                return (
-                    Self {
-                        client: DeletionQueueClient {
-                            tx,
-                            executor_tx,
-                            lsn_table: lsn_table.clone(),
-                        },
-                        cancel,
-                    },
-                    None,
-                )
-            }
-            Some(r) => r,
-        };
-
         (
             Self {
                 client: DeletionQueueClient {
@@ -765,7 +748,7 @@ mod test {
         /// Simulate a pageserver restart by destroying and recreating the deletion queue
         async fn restart(&mut self) {
             let (deletion_queue, workers) = DeletionQueue::new(
-                Some(self.storage.clone()),
+                self.storage.clone(),
                 Some(self.mock_control_plane.clone()),
                 self.harness.conf,
             );
@@ -875,7 +858,7 @@ mod test {
         let mock_control_plane = MockControlPlane::new();
 
         let (deletion_queue, worker) = DeletionQueue::new(
-            Some(storage.clone()),
+            storage.clone(),
             Some(mock_control_plane.clone()),
             harness.conf,
         );

pageserver/src/disk_usage_eviction_task.rs

@@ -534,18 +534,12 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                     });
                 }
                 EvictionLayer::Secondary(layer) => {
-                    let file_size = layer.metadata.file_size();
-                    let tenant_manager = tenant_manager.clone();
+                    let file_size = layer.metadata.file_size;
 
                     js.spawn(async move {
                         layer
                             .secondary_tenant
-                            .evict_layer(
-                                tenant_manager.get_conf(),
-                                layer.timeline_id,
-                                layer.name,
-                                layer.metadata,
-                            )
+                            .evict_layer(layer.timeline_id, layer.name)
                             .await;
                         Ok(file_size)
                     });
@@ -647,7 +641,7 @@ impl EvictionLayer {
     pub(crate) fn get_file_size(&self) -> u64 {
         match self {
             Self::Attached(l) => l.layer_desc().file_size,
-            Self::Secondary(sl) => sl.metadata.file_size(),
+            Self::Secondary(sl) => sl.metadata.file_size,
         }
     }
 }

pageserver/src/http/openapi_spec.yml

@@ -257,6 +257,37 @@ paths:
               schema:
                 $ref: "#/components/schemas/LsnByTimestampResponse"
 
+  /v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/lsn_lease:
+    parameters:
+      - name: tenant_shard_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    post:
+      description: Obtain lease for the given LSN
+      parameters:
+        - name: lsn
+          in: query
+          required: true
+          schema:
+            type: string
+            format: hex
+          description: A LSN to obtain the lease for
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/LsnLease"
+
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/do_gc:
     parameters:
       - name: tenant_id
@@ -581,6 +612,80 @@ paths:
               schema:
                 $ref: "#/components/schemas/Error"
 
+  /v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/detach_ancestor:
+    parameters:
+      - name: tenant_shard_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        ŕequired: true
+        schema:
+          type: string
+
+    put:
+      description: |
+        Detach a timeline from its ancestor and reparent all ancestors timelines with lower `ancestor_lsn`.
+        Current implementation might not be retryable across failure cases, but will be enhanced in future.
+        Detaching should be expected to be expensive operation. Timeouts should be retried.
+      responses:
+        "200":
+          description: |
+            The timeline has been detached from it's ancestor (now or earlier), and at least the returned timelines have been reparented.
+            If any timelines were deleted after reparenting, they might not be on this list.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/AncestorDetached"
+
+        "400":
+          description: |
+            Number of early checks meaning the timeline cannot be detached now:
+              - the ancestor of timeline has an ancestor: not supported, see RFC
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+
+        "404":
+          description: Tenant or timeline not found.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
+
+        "409":
+          description: |
+            The timeline can never be detached:
+              - timeline has no ancestor, implying that the timeline has never had an ancestor
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ConflictError"
+
+        "500":
+          description: |
+            Transient error, for example, pageserver shutdown happened while
+            processing the request but we were unable to distinguish that. Must
+            be retried.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+
+        "503":
+          description: |
+            Temporarily unavailable, please retry. Possible reasons:
+              - another timeline detach for the same tenant is underway, please retry later
+              - detected shutdown error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ServiceUnavailableError"
+
+
   /v1/tenant/:
     get:
       description: Get tenants list
@@ -980,6 +1085,15 @@ components:
           type: string
           enum: [past, present, future, nodata]
 
+    LsnLease:
+      type: object
+      required:
+        - valid_until
+      properties:
+        valid_until:
+          type: string
+          format: date-time
+
     PageserverUtilization:
       type: object
       required:
@@ -1037,6 +1151,19 @@ components:
           format: int64
           description: How many bytes of layer content were in the latest layer heatmap
 
+    AncestorDetached:
+      type: object
+      required:
+        - reparented_timelines
+      properties:
+        reparented_timelines:
+          type: array
+          description: Set of reparented timeline ids
+          properties:
+            type: string
+            format: hex
+            description: TimelineId
+
 
     Error:
       type: object

pageserver/src/http/routes.rs

@@ -1,6 +1,8 @@
 //!
 //! Management HTTP API
 //!
+use std::cmp::Reverse;
+use std::collections::BinaryHeap;
 use std::collections::HashMap;
 use std::str::FromStr;
 use std::sync::Arc;
@@ -14,6 +16,9 @@ use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
+use pageserver_api::models::AuxFilePolicy;
+use pageserver_api::models::IngestAuxFilesRequest;
+use pageserver_api::models::ListAuxFilesRequest;
 use pageserver_api::models::LocationConfig;
 use pageserver_api::models::LocationConfigListResponse;
 use pageserver_api::models::ShardParameters;
@@ -24,7 +29,11 @@ use pageserver_api::models::TenantScanRemoteStorageShard;
 use pageserver_api::models::TenantShardLocation;
 use pageserver_api::models::TenantShardSplitRequest;
 use pageserver_api::models::TenantShardSplitResponse;
+use pageserver_api::models::TenantSorting;
 use pageserver_api::models::TenantState;
+use pageserver_api::models::TopTenantShardItem;
+use pageserver_api::models::TopTenantShardsRequest;
+use pageserver_api::models::TopTenantShardsResponse;
 use pageserver_api::models::{
     DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
     TenantLoadRequest, TenantLocationConfigRequest,
@@ -66,6 +75,7 @@ use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::storage_layer::LayerName;
 use crate::tenant::timeline::CompactFlags;
 use crate::tenant::timeline::Timeline;
+use crate::tenant::GetTimelineError;
 use crate::tenant::SpawnMode;
 use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
 use crate::{config::PageServerConf, tenant::mgr};
@@ -104,7 +114,7 @@ pub struct State {
     tenant_manager: Arc<TenantManager>,
     auth: Option<Arc<SwappableJwtAuth>>,
     allowlist_routes: Vec<Uri>,
-    remote_storage: Option<GenericRemoteStorage>,
+    remote_storage: GenericRemoteStorage,
     broker_client: storage_broker::BrokerClientChannel,
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     deletion_queue_client: DeletionQueueClient,
@@ -118,7 +128,7 @@ impl State {
         conf: &'static PageServerConf,
         tenant_manager: Arc<TenantManager>,
         auth: Option<Arc<SwappableJwtAuth>>,
-        remote_storage: Option<GenericRemoteStorage>,
+        remote_storage: GenericRemoteStorage,
         broker_client: storage_broker::BrokerClientChannel,
         disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
         deletion_queue_client: DeletionQueueClient,
@@ -271,6 +281,13 @@ impl From<GetTenantError> for ApiError {
     }
 }
 
+impl From<GetTimelineError> for ApiError {
+    fn from(gte: GetTimelineError) -> Self {
+        // Rationale: tenant is activated only after eligble timelines activate
+        ApiError::NotFound(gte.into())
+    }
+}
+
 impl From<GetActiveTenantError> for ApiError {
     fn from(e: GetActiveTenantError) -> ApiError {
         match e {
@@ -378,7 +395,7 @@ async fn build_timeline_info_common(
         let guard = timeline.last_received_wal.lock().unwrap();
         if let Some(info) = guard.as_ref() {
             (
-                Some(format!("{:?}", info.wal_source_connconf)), // Password is hidden, but it's for statistics only.
+                Some(format!("{}", info.wal_source_connconf)), // Password is hidden, but it's for statistics only.
                 Some(info.last_received_msg_lsn),
                 Some(info.last_received_msg_ts),
             )
@@ -433,6 +450,8 @@ async fn build_timeline_info_common(
         state,
 
         walreceiver_status,
+
+        last_aux_file_policy: timeline.last_aux_file_policy.load(),
     };
     Ok(info)
 }
@@ -633,9 +652,7 @@ async fn timeline_preserve_initdb_handler(
             .tenant_manager
             .get_attached_tenant_shard(tenant_shard_id)?;
 
-        let timeline = tenant
-            .get_timeline(timeline_id, false)
-            .map_err(|e| ApiError::NotFound(e.into()))?;
+        let timeline = tenant.get_timeline(timeline_id, false)?;
 
         timeline
             .preserve_initdb_archive()
@@ -677,9 +694,7 @@ async fn timeline_detail_handler(
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
-        let timeline = tenant
-            .get_timeline(timeline_id, false)
-            .map_err(|e| ApiError::NotFound(e.into()))?;
+        let timeline = tenant.get_timeline(timeline_id, false)?;
 
         let timeline_info = build_timeline_info(
             &timeline,
@@ -813,12 +828,6 @@ async fn tenant_attach_handler(
 
     let generation = get_request_generation(state, maybe_body.as_ref().and_then(|r| r.generation))?;
 
-    if state.remote_storage.is_none() {
-        return Err(ApiError::BadRequest(anyhow!(
-            "attach_tenant is not possible because pageserver was configured without remote storage"
-        )));
-    }
-
     let tenant_shard_id = TenantShardId::unsharded(tenant_id);
     let shard_params = ShardParameters::default();
     let location_conf = LocationConf::attached_single(tenant_conf, generation, &shard_params);
@@ -1643,12 +1652,6 @@ async fn tenant_time_travel_remote_storage_handler(
         )));
     }
 
-    let Some(storage) = state.remote_storage.as_ref() else {
-        return Err(ApiError::InternalServerError(anyhow::anyhow!(
-            "remote storage not configured, cannot run time travel"
-        )));
-    };
-
     if timestamp > done_if_after {
         return Err(ApiError::BadRequest(anyhow!(
             "The done_if_after timestamp comes before the timestamp to recover to"
@@ -1658,7 +1661,7 @@ async fn tenant_time_travel_remote_storage_handler(
     tracing::info!("Issuing time travel request internally. timestamp={timestamp_raw}, done_if_after={done_if_after_raw}");
 
     remote_timeline_client::upload::time_travel_recover_tenant(
-        storage,
+        &state.remote_storage,
         &tenant_shard_id,
         timestamp,
         done_if_after,
@@ -1703,6 +1706,32 @@ async fn handle_tenant_break(
     json_response(StatusCode::OK, ())
 }
 
+// Obtains an lsn lease on the given timeline.
+async fn lsn_lease_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let lsn: Lsn = parse_query_param(&request, "lsn")?
+        .ok_or_else(|| ApiError::BadRequest(anyhow!("missing 'lsn' query parameter")))?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+
+    let state = get_state(&request);
+
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+    let result = timeline
+        .make_lsn_lease(lsn, &ctx)
+        .map_err(|e| ApiError::InternalServerError(e.context("lsn lease http handler")))?;
+
+    json_response(StatusCode::OK, result)
+}
+
 // Run GC immediately on given timeline.
 async fn timeline_gc_handler(
     mut request: Request<Body>,
@@ -1715,12 +1744,7 @@ async fn timeline_gc_handler(
     let gc_req: TimelineGcRequest = json_request(&mut request).await?;
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let wait_task_done = mgr::immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx)?;
-    let gc_result = wait_task_done
-        .await
-        .context("wait for gc task")
-        .map_err(ApiError::InternalServerError)?
-        .map_err(ApiError::InternalServerError)?;
+    let gc_result = mgr::immediate_gc(tenant_shard_id, timeline_id, gc_req, cancel, &ctx).await?;
 
     json_response(StatusCode::OK, gc_result)
 }
@@ -1743,6 +1767,8 @@ async fn timeline_compact_handler(
     if Some(true) == parse_query_param::<_, bool>(&request, "force_image_layer_creation")? {
         flags |= CompactFlags::ForceImageLayerCreation;
     }
+    let wait_until_uploaded =
+        parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
 
     async {
         let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
@@ -1751,6 +1777,9 @@ async fn timeline_compact_handler(
             .compact(&cancel, flags, &ctx)
             .await
             .map_err(|e| ApiError::InternalServerError(e.into()))?;
+        if wait_until_uploaded {
+            timeline.remote_client.wait_completion().await.map_err(ApiError::InternalServerError)?;
+        }
         json_response(StatusCode::OK, ())
     }
     .instrument(info_span!("manual_compaction", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
@@ -1775,6 +1804,8 @@ async fn timeline_checkpoint_handler(
     if Some(true) == parse_query_param::<_, bool>(&request, "force_image_layer_creation")? {
         flags |= CompactFlags::ForceImageLayerCreation;
     }
+    let wait_until_uploaded =
+        parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
 
     async {
         let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
@@ -1788,6 +1819,10 @@ async fn timeline_checkpoint_handler(
             .await
             .map_err(|e| ApiError::InternalServerError(e.into()))?;
 
+        if wait_until_uploaded {
+            timeline.remote_client.wait_completion().await.map_err(ApiError::InternalServerError)?;
+        }
+
         json_response(StatusCode::OK, ())
     }
     .instrument(info_span!("manual_checkpoint", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
@@ -1871,14 +1906,11 @@ async fn timeline_detach_ancestor_handler(
         let ctx = RequestContext::new(TaskKind::DetachAncestor, DownloadBehavior::Download);
         let ctx = &ctx;
 
-        let timeline = tenant
-            .get_timeline(timeline_id, true)
-            .map_err(|e| ApiError::NotFound(e.into()))?;
+        let timeline = tenant.get_timeline(timeline_id, true)?;
 
         let (_guard, prepared) = timeline
             .prepare_to_detach_from_ancestor(&tenant, options, ctx)
-            .await
-            .map_err(|e| ApiError::InternalServerError(e.into()))?;
+            .await?;
 
         let res = state
             .tenant_manager
@@ -1908,11 +1940,6 @@ async fn deletion_queue_flush(
 ) -> Result<Response<Body>, ApiError> {
     let state = get_state(&r);
 
-    if state.remote_storage.is_none() {
-        // Nothing to do if remote storage is disabled.
-        return json_response(StatusCode::OK, ());
-    }
-
     let execute = parse_query_param(&r, "execute")?.unwrap_or(false);
 
     let flush = async {
@@ -2017,9 +2044,7 @@ async fn active_timeline_of_active_tenant(
 
     tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
-    tenant
-        .get_timeline(timeline_id, true)
-        .map_err(|e| ApiError::NotFound(e.into()))
+    Ok(tenant.get_timeline(timeline_id, true)?)
 }
 
 async fn always_panic_handler(
@@ -2077,18 +2102,11 @@ async fn disk_usage_eviction_run(
     };
 
     let state = get_state(&r);
-
-    let Some(storage) = state.remote_storage.as_ref() else {
-        return Err(ApiError::InternalServerError(anyhow::anyhow!(
-            "remote storage not configured, cannot run eviction iteration"
-        )));
-    };
-
     let eviction_state = state.disk_usage_eviction_state.clone();
 
     let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
         &eviction_state,
-        storage,
+        &state.remote_storage,
         usage,
         &state.tenant_manager,
         config.eviction_order,
@@ -2125,29 +2143,23 @@ async fn tenant_scan_remote_handler(
     let state = get_state(&request);
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
 
-    let Some(remote_storage) = state.remote_storage.as_ref() else {
-        return Err(ApiError::BadRequest(anyhow::anyhow!(
-            "Remote storage not configured"
-        )));
-    };
-
     let mut response = TenantScanRemoteStorageResponse::default();
 
     let (shards, _other_keys) =
-        list_remote_tenant_shards(remote_storage, tenant_id, cancel.clone())
+        list_remote_tenant_shards(&state.remote_storage, tenant_id, cancel.clone())
             .await
             .map_err(|e| ApiError::InternalServerError(anyhow::anyhow!(e)))?;
 
     for tenant_shard_id in shards {
         let (timeline_ids, _other_keys) =
-            list_remote_timelines(remote_storage, tenant_shard_id, cancel.clone())
+            list_remote_timelines(&state.remote_storage, tenant_shard_id, cancel.clone())
                 .await
                 .map_err(|e| ApiError::InternalServerError(anyhow::anyhow!(e)))?;
 
         let mut generation = Generation::none();
         for timeline_id in timeline_ids {
             match download_index_part(
-                remote_storage,
+                &state.remote_storage,
                 &tenant_shard_id,
                 &timeline_id,
                 Generation::MAX,
@@ -2296,6 +2308,31 @@ async fn post_tracing_event_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn force_aux_policy_switch_handler(
+    mut r: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
+    let tenant_shard_id: TenantShardId = parse_request_param(&r, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&r, "timeline_id")?;
+    let policy: AuxFilePolicy = json_request(&mut r).await?;
+
+    let state = get_state(&r);
+
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?;
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+    timeline
+        .do_switch_aux_policy(policy)
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, ())
+}
+
 async fn put_io_engine_handler(
     mut r: Request<Body>,
     _cancel: CancellationToken,
@@ -2358,6 +2395,150 @@ async fn get_utilization(
         .map_err(ApiError::InternalServerError)
 }
 
+async fn list_aux_files(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    let body: ListAuxFilesRequest = json_request(&mut request).await?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let state = get_state(&request);
+
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let files = timeline.list_aux_files(body.lsn, &ctx).await?;
+    json_response(StatusCode::OK, files)
+}
+
+async fn ingest_aux_files(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    let body: IngestAuxFilesRequest = json_request(&mut request).await?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let state = get_state(&request);
+
+    let timeline =
+        active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
+            .await?;
+
+    let mut modification = timeline.begin_modification(
+        Lsn(timeline.get_last_record_lsn().0 + 8), /* advance LSN by 8 */
+    );
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    for (fname, content) in body.aux_files {
+        modification
+            .put_file(&fname, content.as_bytes(), &ctx)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+    }
+    modification
+        .commit(&ctx)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, ())
+}
+
+/// Report on the largest tenants on this pageserver, for the storage controller to identify
+/// candidates for splitting
+async fn post_top_tenants(
+    mut r: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
+    let request: TopTenantShardsRequest = json_request(&mut r).await?;
+    let state = get_state(&r);
+
+    fn get_size_metric(sizes: &TopTenantShardItem, order_by: &TenantSorting) -> u64 {
+        match order_by {
+            TenantSorting::ResidentSize => sizes.resident_size,
+            TenantSorting::MaxLogicalSize => sizes.max_logical_size,
+        }
+    }
+
+    #[derive(Eq, PartialEq)]
+    struct HeapItem {
+        metric: u64,
+        sizes: TopTenantShardItem,
+    }
+
+    impl PartialOrd for HeapItem {
+        fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+            Some(self.cmp(other))
+        }
+    }
+
+    /// Heap items have reverse ordering on their metric: this enables using BinaryHeap, which
+    /// supports popping the greatest item but not the smallest.
+    impl Ord for HeapItem {
+        fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+            Reverse(self.metric).cmp(&Reverse(other.metric))
+        }
+    }
+
+    let mut top_n: BinaryHeap<HeapItem> = BinaryHeap::with_capacity(request.limit);
+
+    // FIXME: this is a lot of clones to take this tenant list
+    for (tenant_shard_id, tenant_slot) in state.tenant_manager.list() {
+        if let Some(shards_lt) = request.where_shards_lt {
+            // Ignore tenants which already have >= this many shards
+            if tenant_shard_id.shard_count >= shards_lt {
+                continue;
+            }
+        }
+
+        let sizes = match tenant_slot {
+            TenantSlot::Attached(tenant) => tenant.get_sizes(),
+            TenantSlot::Secondary(_) | TenantSlot::InProgress(_) => {
+                continue;
+            }
+        };
+        let metric = get_size_metric(&sizes, &request.order_by);
+
+        if let Some(gt) = request.where_gt {
+            // Ignore tenants whose metric is <= the lower size threshold, to do less sorting work
+            if metric <= gt {
+                continue;
+            }
+        };
+
+        match top_n.peek() {
+            None => {
+                // Top N list is empty: candidate becomes first member
+                top_n.push(HeapItem { metric, sizes });
+            }
+            Some(i) if i.metric > metric && top_n.len() < request.limit => {
+                // Lowest item in list is greater than our candidate, but we aren't at limit yet: push to end
+                top_n.push(HeapItem { metric, sizes });
+            }
+            Some(i) if i.metric > metric => {
+                // List is at limit and lowest value is greater than our candidate, drop it.
+            }
+            Some(_) => top_n.push(HeapItem { metric, sizes }),
+        }
+
+        while top_n.len() > request.limit {
+            top_n.pop();
+        }
+    }
+
+    json_response(
+        StatusCode::OK,
+        TopTenantShardsResponse {
+            shards: top_n.into_iter().map(|i| i.sizes).collect(),
+        },
+    )
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -2570,6 +2751,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/get_timestamp_of_lsn",
             |r| api_handler(r, get_timestamp_of_lsn_handler),
         )
+        .post(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/lsn_lease",
+            |r| api_handler(r, lsn_lease_handler),
+        )
         .put(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/do_gc",
             |r| api_handler(r, timeline_gc_handler),
@@ -2643,6 +2828,19 @@ pub fn make_router(
             |r| api_handler(r, timeline_collect_keyspace),
         )
         .put("/v1/io_engine", |r| api_handler(r, put_io_engine_handler))
+        .put(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/force_aux_policy_switch",
+            |r| api_handler(r, force_aux_policy_switch_handler),
+        )
         .get("/v1/utilization", |r| api_handler(r, get_utilization))
+        .post(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/ingest_aux_files",
+            |r| testing_api_handler("ingest_aux_files", r, ingest_aux_files),
+        )
+        .post(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/list_aux_files",
+            |r| testing_api_handler("list_aux_files", r, list_aux_files),
+        )
+        .post("/v1/top_tenants", |r| api_handler(r, post_top_tenants))
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -57,7 +57,7 @@ pub use crate::metrics::preinitialize_metrics;
 #[tracing::instrument(skip_all, fields(%exit_code))]
 pub async fn shutdown_pageserver(
     tenant_manager: &TenantManager,
-    deletion_queue: Option<DeletionQueue>,
+    mut deletion_queue: DeletionQueue,
     exit_code: i32,
 ) {
     use std::time::Duration;
@@ -89,9 +89,7 @@ pub async fn shutdown_pageserver(
     .await;
 
     // Best effort to persist any outstanding deletions, to avoid leaking objects
-    if let Some(mut deletion_queue) = deletion_queue {
-        deletion_queue.shutdown(Duration::from_secs(5)).await;
-    }
+    deletion_queue.shutdown(Duration::from_secs(5)).await;
 
     // Shut down the HTTP endpoint last, so that you can still check the server's
     // status while it's shutting down.
@@ -114,10 +112,6 @@ pub async fn shutdown_pageserver(
     std::process::exit(exit_code);
 }
 
-/// The name of the metadata file pageserver creates per timeline.
-/// Full path: `tenants/<tenant_id>/timelines/<timeline_id>/metadata`.
-pub const METADATA_FILE_NAME: &str = "metadata";
-
 /// Per-tenant configuration file.
 /// Full path: `tenants/<tenant_id>/config`.
 pub(crate) const TENANT_CONFIG_NAME: &str = "config";

pageserver/src/metrics.rs

@@ -525,6 +525,15 @@ static LAST_RECORD_LSN: Lazy<IntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static STANDBY_HORIZON: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "pageserver_standby_horizon",
+        "Standby apply LSN for which GC is hold off, by timeline.",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
 static RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_resident_physical_size",
@@ -585,6 +594,15 @@ static CURRENT_LOGICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define current logical size metric")
 });
 
+static AUX_FILE_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "pageserver_aux_file_estimated_size",
+        "The size of all aux files for a timeline in aux file v2 store.",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) mod initial_logical_size {
     use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
     use once_cell::sync::Lazy;
@@ -1849,7 +1867,6 @@ pub(crate) struct WalIngestMetrics {
     pub(crate) records_received: IntCounter,
     pub(crate) records_committed: IntCounter,
     pub(crate) records_filtered: IntCounter,
-    pub(crate) time_spent_on_ingest: Histogram,
 }
 
 pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| WalIngestMetrics {
@@ -1873,12 +1890,6 @@ pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| WalIngestMet
         "Number of WAL records filtered out due to sharding"
     )
     .expect("failed to define a metric"),
-    time_spent_on_ingest: register_histogram!(
-        "pageserver_wal_ingest_put_value_seconds",
-        "Actual time spent on ingesting a record",
-        redo_histogram_time_buckets!(),
-    )
-    .expect("failed to define a metric"),
 });
 
 pub(crate) static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
@@ -1990,29 +2001,6 @@ impl Default for WalRedoProcessCounters {
 pub(crate) static WAL_REDO_PROCESS_COUNTERS: Lazy<WalRedoProcessCounters> =
     Lazy::new(WalRedoProcessCounters::default);
 
-#[cfg(not(test))]
-pub mod wal_redo {
-    use super::*;
-
-    static PROCESS_KIND: Lazy<std::sync::Mutex<UIntGaugeVec>> = Lazy::new(|| {
-        std::sync::Mutex::new(
-            register_uint_gauge_vec!(
-                "pageserver_wal_redo_process_kind",
-                "The configured process kind for walredo",
-                &["kind"],
-            )
-            .unwrap(),
-        )
-    });
-
-    pub fn set_process_kind_metric(kind: crate::walredo::ProcessKind) {
-        // use guard to avoid races around the next two steps
-        let guard = PROCESS_KIND.lock().unwrap();
-        guard.reset();
-        guard.with_label_values(&[&format!("{kind}")]).set(1);
-    }
-}
-
 /// Similar to `prometheus::HistogramTimer` but does not record on drop.
 pub(crate) struct StorageTimeMetricsTimer {
     metrics: StorageTimeMetrics,
@@ -2112,9 +2100,11 @@ pub(crate) struct TimelineMetrics {
     pub garbage_collect_histo: StorageTimeMetrics,
     pub find_gc_cutoffs_histo: StorageTimeMetrics,
     pub last_record_gauge: IntGauge,
-    resident_physical_size_gauge: UIntGauge,
+    pub standby_horizon_gauge: IntGauge,
+    pub resident_physical_size_gauge: UIntGauge,
     /// copy of LayeredTimeline.current_logical_size
     pub current_logical_size_gauge: UIntGauge,
+    pub aux_file_size_gauge: IntGauge,
     pub directory_entries_count_gauge: Lazy<UIntGauge, Box<dyn Send + Fn() -> UIntGauge>>,
     pub evictions: IntCounter,
     pub evictions_with_low_residence_duration: std::sync::RwLock<EvictionsWithLowResidenceDuration>,
@@ -2180,6 +2170,9 @@ impl TimelineMetrics {
         let last_record_gauge = LAST_RECORD_LSN
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
+        let standby_horizon_gauge = STANDBY_HORIZON
+            .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
+            .unwrap();
         let resident_physical_size_gauge = RESIDENT_PHYSICAL_SIZE
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
@@ -2187,6 +2180,9 @@ impl TimelineMetrics {
         let current_logical_size_gauge = CURRENT_LOGICAL_SIZE
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
+        let aux_file_size_gauge = AUX_FILE_SIZE
+            .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
+            .unwrap();
         // TODO use impl Trait syntax here once we have ability to use it: https://github.com/rust-lang/rust/issues/63065
         let directory_entries_count_gauge_closure = {
             let tenant_shard_id = *tenant_shard_id;
@@ -2222,8 +2218,10 @@ impl TimelineMetrics {
             find_gc_cutoffs_histo,
             load_layer_map_histo,
             last_record_gauge,
+            standby_horizon_gauge,
             resident_physical_size_gauge,
             current_logical_size_gauge,
+            aux_file_size_gauge,
             directory_entries_count_gauge,
             evictions,
             evictions_with_low_residence_duration: std::sync::RwLock::new(
@@ -2255,6 +2253,7 @@ impl TimelineMetrics {
         let timeline_id = &self.timeline_id;
         let shard_id = &self.shard_id;
         let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, shard_id, timeline_id]);
+        let _ = STANDBY_HORIZON.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         {
             RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
             let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
@@ -2264,6 +2263,7 @@ impl TimelineMetrics {
             let _ = metric.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         }
         let _ = EVICTIONS.remove_label_values(&[tenant_id, shard_id, timeline_id]);
+        let _ = AUX_FILE_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
 
         self.evictions_with_low_residence_duration
             .write()
@@ -2320,6 +2320,7 @@ use pin_project_lite::pin_project;
 use std::collections::HashMap;
 use std::num::NonZeroUsize;
 use std::pin::Pin;
+use std::sync::atomic::AtomicU64;
 use std::sync::{Arc, Mutex};
 use std::task::{Context, Poll};
 use std::time::{Duration, Instant};
@@ -2329,35 +2330,35 @@ use crate::task_mgr::TaskKind;
 use crate::tenant::mgr::TenantSlot;
 
 /// Maintain a per timeline gauge in addition to the global gauge.
-struct PerTimelineRemotePhysicalSizeGauge {
-    last_set: u64,
+pub(crate) struct PerTimelineRemotePhysicalSizeGauge {
+    last_set: AtomicU64,
     gauge: UIntGauge,
 }
 
 impl PerTimelineRemotePhysicalSizeGauge {
     fn new(per_timeline_gauge: UIntGauge) -> Self {
         Self {
-            last_set: per_timeline_gauge.get(),
+            last_set: AtomicU64::new(0),
             gauge: per_timeline_gauge,
         }
     }
-    fn set(&mut self, sz: u64) {
+    pub(crate) fn set(&self, sz: u64) {
         self.gauge.set(sz);
-        if sz < self.last_set {
-            REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set - sz);
+        let prev = self.last_set.swap(sz, std::sync::atomic::Ordering::Relaxed);
+        if sz < prev {
+            REMOTE_PHYSICAL_SIZE_GLOBAL.sub(prev - sz);
         } else {
-            REMOTE_PHYSICAL_SIZE_GLOBAL.add(sz - self.last_set);
+            REMOTE_PHYSICAL_SIZE_GLOBAL.add(sz - prev);
         };
-        self.last_set = sz;
     }
-    fn get(&self) -> u64 {
+    pub(crate) fn get(&self) -> u64 {
         self.gauge.get()
     }
 }
 
 impl Drop for PerTimelineRemotePhysicalSizeGauge {
     fn drop(&mut self) {
-        REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set);
+        REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set.load(std::sync::atomic::Ordering::Relaxed));
     }
 }
 
@@ -2365,7 +2366,7 @@ pub(crate) struct RemoteTimelineClientMetrics {
     tenant_id: String,
     shard_id: String,
     timeline_id: String,
-    remote_physical_size_gauge: Mutex<Option<PerTimelineRemotePhysicalSizeGauge>>,
+    pub(crate) remote_physical_size_gauge: PerTimelineRemotePhysicalSizeGauge,
     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>>,
@@ -2373,38 +2374,27 @@ pub(crate) struct RemoteTimelineClientMetrics {
 
 impl RemoteTimelineClientMetrics {
     pub fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self {
+        let tenant_id_str = tenant_shard_id.tenant_id.to_string();
+        let shard_id_str = format!("{}", tenant_shard_id.shard_slug());
+        let timeline_id_str = timeline_id.to_string();
+
+        let remote_physical_size_gauge = PerTimelineRemotePhysicalSizeGauge::new(
+            REMOTE_PHYSICAL_SIZE
+                .get_metric_with_label_values(&[&tenant_id_str, &shard_id_str, &timeline_id_str])
+                .unwrap(),
+        );
+
         RemoteTimelineClientMetrics {
-            tenant_id: tenant_shard_id.tenant_id.to_string(),
-            shard_id: format!("{}", tenant_shard_id.shard_slug()),
-            timeline_id: timeline_id.to_string(),
+            tenant_id: tenant_id_str,
+            shard_id: shard_id_str,
+            timeline_id: timeline_id_str,
             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),
+            remote_physical_size_gauge,
         }
     }
 
-    pub(crate) fn remote_physical_size_set(&self, sz: u64) {
-        let mut guard = self.remote_physical_size_gauge.lock().unwrap();
-        let gauge = guard.get_or_insert_with(|| {
-            PerTimelineRemotePhysicalSizeGauge::new(
-                REMOTE_PHYSICAL_SIZE
-                    .get_metric_with_label_values(&[
-                        &self.tenant_id,
-                        &self.shard_id,
-                        &self.timeline_id,
-                    ])
-                    .unwrap(),
-            )
-        });
-        gauge.set(sz);
-    }
-
-    pub(crate) fn remote_physical_size_get(&self) -> u64 {
-        let guard = self.remote_physical_size_gauge.lock().unwrap();
-        guard.as_ref().map(|gauge| gauge.get()).unwrap_or(0)
-    }
-
     pub fn remote_operation_time(
         &self,
         file_kind: &RemoteOpFileKind,

pageserver/src/page_service.rs

@@ -19,6 +19,7 @@ use pageserver_api::models::{
 };
 use pageserver_api::shard::ShardIndex;
 use pageserver_api::shard::ShardNumber;
+use pageserver_api::shard::TenantShardId;
 use postgres_backend::{is_expected_io_error, AuthType, PostgresBackend, QueryError};
 use pq_proto::framed::ConnectionError;
 use pq_proto::FeStartupPacket;
@@ -32,6 +33,8 @@ use std::str;
 use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
+use std::time::Instant;
+use std::time::SystemTime;
 use tokio::io::AsyncWriteExt;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::io::StreamReader;
@@ -49,7 +52,6 @@ use utils::{
 use crate::auth::check_permission;
 use crate::basebackup;
 use crate::basebackup::BasebackupError;
-use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir::import_wal_from_tar;
 use crate::metrics;
@@ -59,13 +61,15 @@ use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id;
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
-use crate::tenant::mgr;
-use crate::tenant::mgr::get_active_tenant_with_timeout;
 use crate::tenant::mgr::GetActiveTenantError;
+use crate::tenant::mgr::GetTenantError;
+use crate::tenant::mgr::ShardResolveResult;
 use crate::tenant::mgr::ShardSelector;
+use crate::tenant::mgr::TenantManager;
 use crate::tenant::timeline::WaitLsnError;
 use crate::tenant::GetTimelineError;
 use crate::tenant::PageReconstructError;
+use crate::tenant::Tenant;
 use crate::tenant::Timeline;
 use crate::trace::Tracer;
 use pageserver_api::key::rel_block_to_key;
@@ -135,7 +139,7 @@ async fn read_tar_eof(mut reader: (impl AsyncRead + Unpin)) -> anyhow::Result<()
 /// Listens for connections, and launches a new handler task for each.
 ///
 pub async fn libpq_listener_main(
-    conf: &'static PageServerConf,
+    tenant_manager: Arc<TenantManager>,
     broker_client: storage_broker::BrokerClientChannel,
     auth: Option<Arc<SwappableJwtAuth>>,
     listener: TcpListener,
@@ -180,7 +184,7 @@ pub async fn libpq_listener_main(
                     "serving compute connection task",
                     false,
                     page_service_conn_main(
-                        conf,
+                        tenant_manager.clone(),
                         broker_client.clone(),
                         local_auth,
                         socket,
@@ -203,7 +207,7 @@ pub async fn libpq_listener_main(
 
 #[instrument(skip_all, fields(peer_addr))]
 async fn page_service_conn_main(
-    conf: &'static PageServerConf,
+    tenant_manager: Arc<TenantManager>,
     broker_client: storage_broker::BrokerClientChannel,
     auth: Option<Arc<SwappableJwtAuth>>,
     socket: tokio::net::TcpStream,
@@ -256,11 +260,14 @@ async fn page_service_conn_main(
     socket.set_timeout(Some(std::time::Duration::from_millis(socket_timeout_ms)));
     let socket = std::pin::pin!(socket);
 
+    fail::fail_point!("ps::connection-start::pre-login");
+
     // XXX: pgbackend.run() should take the connection_ctx,
     // and create a child per-query context when it invokes process_query.
     // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
     // and create the per-query context in process_query ourselves.
-    let mut conn_handler = PageServerHandler::new(conf, broker_client, auth, connection_ctx);
+    let mut conn_handler =
+        PageServerHandler::new(tenant_manager, broker_client, auth, connection_ctx);
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
 
     match pgbackend
@@ -291,11 +298,12 @@ struct HandlerTimeline {
 }
 
 struct PageServerHandler {
-    _conf: &'static PageServerConf,
     broker_client: storage_broker::BrokerClientChannel,
     auth: Option<Arc<SwappableJwtAuth>>,
     claims: Option<Claims>,
 
+    tenant_manager: Arc<TenantManager>,
+
     /// The context created for the lifetime of the connection
     /// services by this PageServerHandler.
     /// For each query received over the connection,
@@ -381,13 +389,13 @@ impl From<WaitLsnError> for QueryError {
 
 impl PageServerHandler {
     pub fn new(
-        conf: &'static PageServerConf,
+        tenant_manager: Arc<TenantManager>,
         broker_client: storage_broker::BrokerClientChannel,
         auth: Option<Arc<SwappableJwtAuth>>,
         connection_ctx: RequestContext,
     ) -> Self {
         PageServerHandler {
-            _conf: conf,
+            tenant_manager,
             broker_client,
             auth,
             claims: None,
@@ -552,13 +560,9 @@ impl PageServerHandler {
     {
         debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
 
-        let tenant = mgr::get_active_tenant_with_timeout(
-            tenant_id,
-            ShardSelector::First,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await?;
+        let tenant = self
+            .get_active_tenant_with_timeout(tenant_id, ShardSelector::First, ACTIVE_TENANT_TIMEOUT)
+            .await?;
 
         // Make request tracer if needed
         let mut tracer = if tenant.get_trace_read_requests() {
@@ -601,6 +605,7 @@ impl PageServerHandler {
             };
 
             trace!("query: {copy_data_bytes:?}");
+            fail::fail_point!("ps::handle-pagerequest-message");
 
             // Trace request if needed
             if let Some(t) = tracer.as_mut() {
@@ -615,6 +620,7 @@ impl PageServerHandler {
 
             let (response, span) = match neon_fe_msg {
                 PagestreamFeMessage::Exists(req) => {
+                    fail::fail_point!("ps::handle-pagerequest-message::exists");
                     let span = tracing::info_span!("handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.request_lsn);
                     (
                         self.handle_get_rel_exists_request(tenant_id, timeline_id, &req, &ctx)
@@ -624,6 +630,7 @@ impl PageServerHandler {
                     )
                 }
                 PagestreamFeMessage::Nblocks(req) => {
+                    fail::fail_point!("ps::handle-pagerequest-message::nblocks");
                     let span = tracing::info_span!("handle_get_nblocks_request", rel = %req.rel, req_lsn = %req.request_lsn);
                     (
                         self.handle_get_nblocks_request(tenant_id, timeline_id, &req, &ctx)
@@ -633,6 +640,7 @@ impl PageServerHandler {
                     )
                 }
                 PagestreamFeMessage::GetPage(req) => {
+                    fail::fail_point!("ps::handle-pagerequest-message::getpage");
                     // shard_id is filled in by the handler
                     let span = tracing::info_span!("handle_get_page_at_lsn_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.request_lsn);
                     (
@@ -643,6 +651,7 @@ impl PageServerHandler {
                     )
                 }
                 PagestreamFeMessage::DbSize(req) => {
+                    fail::fail_point!("ps::handle-pagerequest-message::dbsize");
                     let span = tracing::info_span!("handle_db_size_request", dbnode = %req.dbnode, req_lsn = %req.request_lsn);
                     (
                         self.handle_db_size_request(tenant_id, timeline_id, &req, &ctx)
@@ -652,6 +661,7 @@ impl PageServerHandler {
                     )
                 }
                 PagestreamFeMessage::GetSlruSegment(req) => {
+                    fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
                     let span = tracing::info_span!("handle_get_slru_segment_request", kind = %req.kind, segno = %req.segno, req_lsn = %req.request_lsn);
                     (
                         self.handle_get_slru_segment_request(tenant_id, timeline_id, &req, &ctx)
@@ -726,13 +736,9 @@ impl PageServerHandler {
 
         // Create empty timeline
         info!("creating new timeline");
-        let tenant = get_active_tenant_with_timeout(
-            tenant_id,
-            ShardSelector::Zero,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await?;
+        let tenant = self
+            .get_active_tenant_with_timeout(tenant_id, ShardSelector::Zero, ACTIVE_TENANT_TIMEOUT)
+            .await?;
         let timeline = tenant
             .create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)
             .await?;
@@ -909,6 +915,39 @@ impl PageServerHandler {
         }
     }
 
+    #[instrument(skip_all, fields(shard_id, %lsn))]
+    async fn handle_make_lsn_lease<IO>(
+        &self,
+        pgb: &mut PostgresBackend<IO>,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<(), QueryError>
+    where
+        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
+    {
+        let shard_selector = ShardSelector::Known(tenant_shard_id.to_index());
+        let timeline = self
+            .get_active_tenant_timeline(tenant_shard_id.tenant_id, timeline_id, shard_selector)
+            .await?;
+        let lease = timeline.make_lsn_lease(lsn, ctx)?;
+        let valid_until = lease
+            .valid_until
+            .duration_since(SystemTime::UNIX_EPOCH)
+            .map_err(|e| QueryError::Other(e.into()))?;
+
+        pgb.write_message_noflush(&BeMessage::RowDescription(&[RowDescriptor::text_col(
+            b"valid_until",
+        )]))?
+        .write_message_noflush(&BeMessage::DataRow(&[Some(
+            &valid_until.as_millis().to_be_bytes(),
+        )]))?
+        .write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
+
+        Ok(())
+    }
+
     #[instrument(skip_all, fields(shard_id))]
     async fn handle_get_rel_exists_request(
         &mut self,
@@ -1370,18 +1409,69 @@ impl PageServerHandler {
         timeline_id: TimelineId,
         selector: ShardSelector,
     ) -> Result<Arc<Timeline>, GetActiveTimelineError> {
-        let tenant = get_active_tenant_with_timeout(
-            tenant_id,
-            selector,
-            ACTIVE_TENANT_TIMEOUT,
-            &task_mgr::shutdown_token(),
-        )
-        .await
-        .map_err(GetActiveTimelineError::Tenant)?;
+        let tenant = self
+            .get_active_tenant_with_timeout(tenant_id, selector, ACTIVE_TENANT_TIMEOUT)
+            .await
+            .map_err(GetActiveTimelineError::Tenant)?;
         let timeline = tenant.get_timeline(timeline_id, true)?;
         set_tracing_field_shard_id(&timeline);
         Ok(timeline)
     }
+
+    /// Get a shard's [`Tenant`] in its active state, if present.  If we don't find the shard and some
+    /// slots for this tenant are `InProgress` then we will wait.
+    /// If we find the [`Tenant`] and it's not yet in state [`TenantState::Active`], we will wait.
+    ///
+    /// `timeout` is used as a total timeout for the whole wait operation.
+    async fn get_active_tenant_with_timeout(
+        &self,
+        tenant_id: TenantId,
+        shard_selector: ShardSelector,
+        timeout: Duration,
+    ) -> Result<Arc<Tenant>, GetActiveTenantError> {
+        let wait_start = Instant::now();
+        let deadline = wait_start + timeout;
+
+        // Resolve TenantId to TenantShardId.  This is usually a quick one-shot thing, the loop is
+        // for handling the rare case that the slot we're accessing is InProgress.
+        let tenant_shard = loop {
+            let resolved = self
+                .tenant_manager
+                .resolve_attached_shard(&tenant_id, shard_selector);
+            match resolved {
+                ShardResolveResult::Found(tenant_shard) => break tenant_shard,
+                ShardResolveResult::NotFound => {
+                    return Err(GetActiveTenantError::NotFound(GetTenantError::NotFound(
+                        tenant_id,
+                    )));
+                }
+                ShardResolveResult::InProgress(barrier) => {
+                    // We can't authoritatively answer right now: wait for InProgress state
+                    // to end, then try again
+                    tokio::select! {
+                        _ = self.await_connection_cancelled() => {
+                            return Err(GetActiveTenantError::Cancelled)
+                        },
+                        _  = barrier.wait() => {
+                            // The barrier completed: proceed around the loop to try looking up again
+                        },
+                        _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
+                            return Err(GetActiveTenantError::WaitForActiveTimeout {
+                                latest_state: None,
+                                wait_time: timeout,
+                            });
+                        }
+                    }
+                }
+            };
+        };
+
+        tracing::debug!("Waiting for tenant to enter active state...");
+        tenant_shard
+            .wait_to_become_active(deadline.duration_since(Instant::now()))
+            .await?;
+        Ok(tenant_shard)
+    }
 }
 
 #[async_trait::async_trait]
@@ -1423,6 +1513,7 @@ where
         _pgb: &mut PostgresBackend<IO>,
         _sm: &FeStartupPacket,
     ) -> Result<(), QueryError> {
+        fail::fail_point!("ps::connection-start::startup-packet");
         Ok(())
     }
 
@@ -1437,11 +1528,12 @@ where
             Err(QueryError::SimulatedConnectionError)
         });
 
+        fail::fail_point!("ps::connection-start::process-query");
+
         let ctx = self.connection_ctx.attached_child();
         debug!("process query {query_string:?}");
-        if query_string.starts_with("pagestream_v2 ") {
-            let (_, params_raw) = query_string.split_at("pagestream_v2 ".len());
-            let params = params_raw.split(' ').collect::<Vec<_>>();
+        let parts = query_string.split_whitespace().collect::<Vec<_>>();
+        if let Some(params) = parts.strip_prefix(&["pagestream_v2"]) {
             if params.len() != 2 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for pagestream command"
@@ -1466,9 +1558,7 @@ where
                 ctx,
             )
             .await?;
-        } else if query_string.starts_with("pagestream ") {
-            let (_, params_raw) = query_string.split_at("pagestream ".len());
-            let params = params_raw.split(' ').collect::<Vec<_>>();
+        } else if let Some(params) = parts.strip_prefix(&["pagestream"]) {
             if params.len() != 2 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for pagestream command"
@@ -1493,10 +1583,7 @@ where
                 ctx,
             )
             .await?;
-        } else if query_string.starts_with("basebackup ") {
-            let (_, params_raw) = query_string.split_at("basebackup ".len());
-            let params = params_raw.split_whitespace().collect::<Vec<_>>();
-
+        } else if let Some(params) = parts.strip_prefix(&["basebackup"]) {
             if params.len() < 2 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for basebackup command"
@@ -1514,26 +1601,23 @@ where
 
             self.check_permission(Some(tenant_id))?;
 
-            let lsn = if params.len() >= 3 {
+            let lsn = if let Some(lsn_str) = params.get(2) {
                 Some(
-                    Lsn::from_str(params[2])
-                        .with_context(|| format!("Failed to parse Lsn from {}", params[2]))?,
+                    Lsn::from_str(lsn_str)
+                        .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?,
                 )
             } else {
                 None
             };
 
-            let gzip = if params.len() >= 4 {
-                if params[3] == "--gzip" {
-                    true
-                } else {
+            let gzip = match params.get(3) {
+                Some(&"--gzip") => true,
+                None => false,
+                Some(third_param) => {
                     return Err(QueryError::Other(anyhow::anyhow!(
-                        "Parameter in position 3 unknown {}",
-                        params[3],
-                    )));
+                        "Parameter in position 3 unknown {third_param}",
+                    )))
                 }
-            } else {
-                false
             };
 
             let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording(&ctx);
@@ -1557,10 +1641,7 @@ where
             res?;
         }
         // return pair of prev_lsn and last_lsn
-        else if query_string.starts_with("get_last_record_rlsn ") {
-            let (_, params_raw) = query_string.split_at("get_last_record_rlsn ".len());
-            let params = params_raw.split_whitespace().collect::<Vec<_>>();
-
+        else if let Some(params) = parts.strip_prefix(&["get_last_record_rlsn"]) {
             if params.len() != 2 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for get_last_record_rlsn command"
@@ -1602,10 +1683,7 @@ where
             .await?;
         }
         // same as basebackup, but result includes relational data as well
-        else if query_string.starts_with("fullbackup ") {
-            let (_, params_raw) = query_string.split_at("fullbackup ".len());
-            let params = params_raw.split_whitespace().collect::<Vec<_>>();
-
+        else if let Some(params) = parts.strip_prefix(&["fullbackup"]) {
             if params.len() < 2 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for fullbackup command"
@@ -1622,18 +1700,18 @@ where
                 .record("timeline_id", field::display(timeline_id));
 
             // The caller is responsible for providing correct lsn and prev_lsn.
-            let lsn = if params.len() > 2 {
+            let lsn = if let Some(lsn_str) = params.get(2) {
                 Some(
-                    Lsn::from_str(params[2])
-                        .with_context(|| format!("Failed to parse Lsn from {}", params[2]))?,
+                    Lsn::from_str(lsn_str)
+                        .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?,
                 )
             } else {
                 None
             };
-            let prev_lsn = if params.len() > 3 {
+            let prev_lsn = if let Some(prev_lsn_str) = params.get(3) {
                 Some(
-                    Lsn::from_str(params[3])
-                        .with_context(|| format!("Failed to parse Lsn from {}", params[3]))?,
+                    Lsn::from_str(prev_lsn_str)
+                        .with_context(|| format!("Failed to parse Lsn from {prev_lsn_str}"))?,
                 )
             } else {
                 None
@@ -1666,8 +1744,7 @@ where
             // 2. Run:
             // cat my_backup/base.tar | psql -h $PAGESERVER \
             //     -c "import basebackup $TENANT $TIMELINE $START_LSN $END_LSN $PG_VERSION"
-            let (_, params_raw) = query_string.split_at("import basebackup ".len());
-            let params = params_raw.split_whitespace().collect::<Vec<_>>();
+            let params = &parts[2..];
             if params.len() != 5 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for import basebackup command"
@@ -1716,8 +1793,7 @@ where
             //
             // Files are scheduled to be persisted to remote storage, and the
             // caller should poll the http api to check when that is done.
-            let (_, params_raw) = query_string.split_at("import wal ".len());
-            let params = params_raw.split_whitespace().collect::<Vec<_>>();
+            let params = &parts[2..];
             if params.len() != 4 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for import wal command"
@@ -1755,10 +1831,45 @@ where
             // important because psycopg2 executes "SET datestyle TO 'ISO'"
             // on connect
             pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
-        } else if query_string.starts_with("show ") {
+        } else if query_string.starts_with("lease lsn ") {
+            let params = &parts[2..];
+            if params.len() != 3 {
+                return Err(QueryError::Other(anyhow::anyhow!(
+                    "invalid param number {} for lease lsn command",
+                    params.len()
+                )));
+            }
+
+            let tenant_shard_id = TenantShardId::from_str(params[0])
+                .with_context(|| format!("Failed to parse tenant id from {}", params[0]))?;
+            let timeline_id = TimelineId::from_str(params[1])
+                .with_context(|| format!("Failed to parse timeline id from {}", params[1]))?;
+
+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_shard_id))
+                .record("timeline_id", field::display(timeline_id));
+
+            self.check_permission(Some(tenant_shard_id.tenant_id))?;
+
+            // The caller is responsible for providing correct lsn.
+            let lsn = Lsn::from_str(params[2])
+                .with_context(|| format!("Failed to parse Lsn from {}", params[2]))?;
+
+            match self
+                .handle_make_lsn_lease(pgb, tenant_shard_id, timeline_id, lsn, &ctx)
+                .await
+            {
+                Ok(()) => pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?,
+                Err(e) => {
+                    error!("error obtaining lsn lease for {lsn}: {e:?}");
+                    pgb.write_message_noflush(&BeMessage::ErrorResponse(
+                        &e.to_string(),
+                        Some(e.pg_error_code()),
+                    ))?
+                }
+            };
+        } else if let Some(params) = parts.strip_prefix(&["show"]) {
             // show <tenant_id>
-            let (_, params_raw) = query_string.split_at("show ".len());
-            let params = params_raw.split(' ').collect::<Vec<_>>();
             if params.len() != 1 {
                 return Err(QueryError::Other(anyhow::anyhow!(
                     "invalid param number for config command"
@@ -1771,13 +1882,13 @@ where
 
             self.check_permission(Some(tenant_id))?;
 
-            let tenant = get_active_tenant_with_timeout(
-                tenant_id,
-                ShardSelector::Zero,
-                ACTIVE_TENANT_TIMEOUT,
-                &task_mgr::shutdown_token(),
-            )
-            .await?;
+            let tenant = self
+                .get_active_tenant_with_timeout(
+                    tenant_id,
+                    ShardSelector::Zero,
+                    ACTIVE_TENANT_TIMEOUT,
+                )
+                .await?;
             pgb.write_message_noflush(&BeMessage::RowDescription(&[
                 RowDescriptor::int8_col(b"checkpoint_distance"),
                 RowDescriptor::int8_col(b"checkpoint_timeout"),

pageserver/src/pgdatadir_mapping.rs

@@ -9,7 +9,6 @@
 use super::tenant::{PageReconstructError, Timeline};
 use crate::context::RequestContext;
 use crate::keyspace::{KeySpace, KeySpaceAccum};
-use crate::metrics::WAL_INGEST;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id;
 use crate::walrecord::NeonWalRecord;
 use crate::{aux_file, repository::*};
@@ -35,12 +34,16 @@ use std::ops::ControlFlow;
 use std::ops::Range;
 use strum::IntoEnumIterator;
 use tokio_util::sync::CancellationToken;
-use tracing::{debug, trace, warn};
+use tracing::{debug, info, trace, warn};
 use utils::bin_ser::DeserializeError;
 use utils::vec_map::{VecMap, VecMapOrdering};
 use utils::{bin_ser::BeSer, lsn::Lsn};
 
-const MAX_AUX_FILE_DELTAS: usize = 1024;
+/// Max delta records appended to the AUX_FILES_KEY (for aux v1). The write path will write a full image once this threshold is reached.
+pub const MAX_AUX_FILE_DELTAS: usize = 1024;
+
+/// Max number of aux-file-related delta layers. The compaction will create a new image layer once this threshold is reached.
+pub const MAX_AUX_FILE_V2_DELTAS: usize = 64;
 
 #[derive(Debug)]
 pub enum LsnForTimestamp {
@@ -699,13 +702,17 @@ impl Timeline {
             .await
             .context("scan")?;
         let mut result = HashMap::new();
+        let mut sz = 0;
         for (_, v) in kv {
             let v = v.context("get value")?;
             let v = aux_file::decode_file_value_bytes(&v).context("value decode")?;
             for (fname, content) in v {
+                sz += fname.len();
+                sz += content.len();
                 result.insert(fname, content);
             }
         }
+        self.aux_file_size_estimator.on_base_backup(sz);
         Ok(result)
     }
 
@@ -714,10 +721,11 @@ impl Timeline {
         lsn: Lsn,
         ctx: &RequestContext,
     ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
-        match self.get_switch_aux_file_policy() {
-            AuxFilePolicy::V1 => self.list_aux_files_v1(lsn, ctx).await,
-            AuxFilePolicy::V2 => self.list_aux_files_v2(lsn, ctx).await,
-            AuxFilePolicy::CrossValidation => {
+        let current_policy = self.last_aux_file_policy.load();
+        match current_policy {
+            Some(AuxFilePolicy::V1) | None => self.list_aux_files_v1(lsn, ctx).await,
+            Some(AuxFilePolicy::V2) => self.list_aux_files_v2(lsn, ctx).await,
+            Some(AuxFilePolicy::CrossValidation) => {
                 let v1_result = self.list_aux_files_v1(lsn, ctx).await;
                 let v2_result = self.list_aux_files_v2(lsn, ctx).await;
                 match (v1_result, v2_result) {
@@ -1465,7 +1473,40 @@ impl<'a> DatadirModification<'a> {
         content: &[u8],
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        let policy = self.tline.get_switch_aux_file_policy();
+        let switch_policy = self.tline.get_switch_aux_file_policy();
+
+        let policy = {
+            let current_policy = self.tline.last_aux_file_policy.load();
+            // Allowed switch path:
+            // * no aux files -> v1/v2/cross-validation
+            // * cross-validation->v2
+
+            let current_policy = if current_policy.is_none() {
+                // This path will only be hit once per tenant: we will decide the final policy in this code block.
+                // The next call to `put_file` will always have `last_aux_file_policy != None`.
+                let lsn = Lsn::max(self.tline.get_last_record_lsn(), self.lsn);
+                let aux_files_key_v1 = self.tline.list_aux_files_v1(lsn, ctx).await?;
+                if aux_files_key_v1.is_empty() {
+                    None
+                } else {
+                    self.tline.do_switch_aux_policy(AuxFilePolicy::V1)?;
+                    Some(AuxFilePolicy::V1)
+                }
+            } else {
+                current_policy
+            };
+
+            if AuxFilePolicy::is_valid_migration_path(current_policy, switch_policy) {
+                self.tline.do_switch_aux_policy(switch_policy)?;
+                info!(current=?current_policy, next=?switch_policy, "switching aux file policy");
+                switch_policy
+            } else {
+                // This branch handles non-valid migration path, and the case that switch_policy == current_policy.
+                // And actually, because the migration path always allow unspecified -> *, this unwrap_or will never be hit.
+                current_policy.unwrap_or(AuxFilePolicy::default_tenant_config())
+            }
+        };
+
         if let AuxFilePolicy::V2 | AuxFilePolicy::CrossValidation = policy {
             let key = aux_file::encode_aux_file_key(path);
             // retrieve the key from the engine
@@ -1474,23 +1515,45 @@ impl<'a> DatadirModification<'a> {
                 Err(PageReconstructError::MissingKey(_)) => None,
                 Err(e) => return Err(e.into()),
             };
-            let files = if let Some(ref old_val) = old_val {
+            let files: Vec<(&str, &[u8])> = if let Some(ref old_val) = old_val {
                 aux_file::decode_file_value(old_val)?
             } else {
                 Vec::new()
             };
-            let new_files = if content.is_empty() {
-                files
-                    .into_iter()
-                    .filter(|(p, _)| &path != p)
-                    .collect::<Vec<_>>()
-            } else {
-                files
-                    .into_iter()
-                    .filter(|(p, _)| &path != p)
-                    .chain(std::iter::once((path, content)))
-                    .collect::<Vec<_>>()
-            };
+            let mut other_files = Vec::with_capacity(files.len());
+            let mut modifying_file = None;
+            for file @ (p, content) in files {
+                if path == p {
+                    assert!(
+                        modifying_file.is_none(),
+                        "duplicated entries found for {}",
+                        path
+                    );
+                    modifying_file = Some(content);
+                } else {
+                    other_files.push(file);
+                }
+            }
+            let mut new_files = other_files;
+            match (modifying_file, content.is_empty()) {
+                (Some(old_content), false) => {
+                    self.tline
+                        .aux_file_size_estimator
+                        .on_update(old_content.len(), content.len());
+                    new_files.push((path, content));
+                }
+                (Some(old_content), true) => {
+                    self.tline
+                        .aux_file_size_estimator
+                        .on_remove(old_content.len());
+                    // not adding the file key to the final `new_files` vec.
+                }
+                (None, false) => {
+                    self.tline.aux_file_size_estimator.on_add(content.len());
+                    new_files.push((path, content));
+                }
+                (None, true) => anyhow::bail!("removing non-existing aux file: {}", path),
+            }
             let new_val = aux_file::encode_file_value(&new_files)?;
             self.put(key, Value::Image(new_val.into()));
         }
@@ -1651,8 +1714,6 @@ impl<'a> DatadirModification<'a> {
     pub async fn commit(&mut self, ctx: &RequestContext) -> anyhow::Result<()> {
         let mut writer = self.tline.writer().await;
 
-        let timer = WAL_INGEST.time_spent_on_ingest.start_timer();
-
         let pending_nblocks = self.pending_nblocks;
         self.pending_nblocks = 0;
 
@@ -1692,8 +1753,6 @@ impl<'a> DatadirModification<'a> {
             writer.update_directory_entries_count(kind, count as u64);
         }
 
-        timer.observe_duration();
-
         Ok(())
     }
 
@@ -1729,6 +1788,12 @@ impl<'a> DatadirModification<'a> {
         self.tline.get(key, lsn, ctx).await
     }
 
+    /// Only used during unit tests, force putting a key into the modification.
+    #[cfg(test)]
+    pub(crate) fn put_for_test(&mut self, key: Key, val: Value) {
+        self.put(key, val);
+    }
+
     fn put(&mut self, key: Key, val: Value) {
         let values = self.pending_updates.entry(key).or_default();
         // Replace the previous value if it exists at the same lsn

pageserver/src/tenant/blob_io.rs

@@ -238,10 +238,13 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
                         io_buf,
                         Err(Error::new(
                             ErrorKind::Other,
-                            format!("blob too large ({} bytes)", len),
+                            format!("blob too large ({len} bytes)"),
                         )),
                     );
                 }
+                if len > 0x0fff_ffff {
+                    tracing::warn!("writing blob above future limit ({len} bytes)");
+                }
                 let mut len_buf = (len as u32).to_be_bytes();
                 len_buf[0] |= 0x80;
                 io_buf.extend_from_slice(&len_buf[..]);

pageserver/src/tenant/config.rs

@@ -11,6 +11,7 @@
 use anyhow::bail;
 use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::models::CompactionAlgorithm;
+use pageserver_api::models::CompactionAlgorithmSettings;
 use pageserver_api::models::EvictionPolicy;
 use pageserver_api::models::{self, ThrottleConfig};
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
@@ -39,8 +40,6 @@ pub mod defaults {
 
     pub const DEFAULT_COMPACTION_PERIOD: &str = "20 s";
     pub const DEFAULT_COMPACTION_THRESHOLD: usize = 10;
-    pub const DEFAULT_COMPACTION_ALGORITHM: super::CompactionAlgorithm =
-        super::CompactionAlgorithm::Legacy;
 
     pub const DEFAULT_GC_HORIZON: u64 = 64 * 1024 * 1024;
 
@@ -320,7 +319,7 @@ pub struct TenantConf {
     pub compaction_period: Duration,
     // Level0 delta layer threshold for compaction.
     pub compaction_threshold: usize,
-    pub compaction_algorithm: CompactionAlgorithm,
+    pub compaction_algorithm: CompactionAlgorithmSettings,
     // Determines how much history is retained, to allow
     // branching and read replicas at an older point in time.
     // The unit is #of bytes of WAL.
@@ -373,6 +372,8 @@ pub struct TenantConf {
 
     /// Switch to a new aux file policy. Switching this flag requires the user has not written any aux file into
     /// the storage before, and this flag cannot be switched back. Otherwise there will be data corruptions.
+    /// There is a `last_aux_file_policy` flag which gets persisted in `index_part.json` once the first aux
+    /// file is written.
     pub switch_aux_file_policy: AuxFilePolicy,
 }
 
@@ -404,7 +405,7 @@ pub struct TenantConfOpt {
 
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
-    pub compaction_algorithm: Option<CompactionAlgorithm>,
+    pub compaction_algorithm: Option<CompactionAlgorithmSettings>,
 
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
@@ -495,7 +496,9 @@ impl TenantConfOpt {
                 .unwrap_or(global_conf.compaction_threshold),
             compaction_algorithm: self
                 .compaction_algorithm
-                .unwrap_or(global_conf.compaction_algorithm),
+                .as_ref()
+                .unwrap_or(&global_conf.compaction_algorithm)
+                .clone(),
             gc_horizon: self.gc_horizon.unwrap_or(global_conf.gc_horizon),
             gc_period: self.gc_period.unwrap_or(global_conf.gc_period),
             image_creation_threshold: self
@@ -548,7 +551,15 @@ impl Default for TenantConf {
             compaction_period: humantime::parse_duration(DEFAULT_COMPACTION_PERIOD)
                 .expect("cannot parse default compaction period"),
             compaction_threshold: DEFAULT_COMPACTION_THRESHOLD,
-            compaction_algorithm: DEFAULT_COMPACTION_ALGORITHM,
+            compaction_algorithm: CompactionAlgorithmSettings {
+                kind: if cfg!(test) {
+                    // Rust tests rely on a valid implicit default (TODO: fix this)
+                    CompactionAlgorithm::Legacy
+                } else {
+                    // Python tests are subject to NotSpecified handling
+                    CompactionAlgorithm::NotSpecified
+                },
+            },
             gc_horizon: DEFAULT_GC_HORIZON,
             gc_period: humantime::parse_duration(DEFAULT_GC_PERIOD)
                 .expect("cannot parse default gc period"),
@@ -574,7 +585,7 @@ impl Default for TenantConf {
             lazy_slru_download: false,
             timeline_get_throttle: crate::tenant::throttle::Config::disabled(),
             image_layer_creation_check_threshold: DEFAULT_IMAGE_LAYER_CREATION_CHECK_THRESHOLD,
-            switch_aux_file_policy: AuxFilePolicy::V1,
+            switch_aux_file_policy: AuxFilePolicy::default_tenant_config(),
         }
     }
 }

pageserver/src/tenant/delete.rs

@@ -8,7 +8,7 @@ use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
 use tracing::{error, instrument, Instrument};
 
-use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId};
+use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
 
 use crate::{
     config::PageServerConf,
@@ -181,25 +181,23 @@ async fn ensure_timelines_dir_empty(timelines_path: &Utf8Path) -> Result<(), Del
 
 async fn remove_tenant_remote_delete_mark(
     conf: &PageServerConf,
-    remote_storage: Option<&GenericRemoteStorage>,
+    remote_storage: &GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
     cancel: &CancellationToken,
 ) -> Result<(), DeleteTenantError> {
-    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, 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::Error::new(TimeoutOrCancel::Cancel))
-        .and_then(|x| x)
-        .context("remove_tenant_remote_delete_mark")?;
-    }
+    let path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
+    backoff::retry(
+        || 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::Error::new(TimeoutOrCancel::Cancel))
+    .and_then(|x| x)
+    .context("remove_tenant_remote_delete_mark")?;
     Ok(())
 }
 
@@ -297,7 +295,7 @@ impl DeleteTenantFlow {
     #[instrument(skip_all)]
     pub(crate) async fn run(
         conf: &'static PageServerConf,
-        remote_storage: Option<GenericRemoteStorage>,
+        remote_storage: GenericRemoteStorage,
         tenants: &'static std::sync::RwLock<TenantsMap>,
         tenant: Arc<Tenant>,
         cancel: &CancellationToken,
@@ -308,9 +306,7 @@ impl DeleteTenantFlow {
 
         let mut guard = Self::prepare(&tenant).await?;
 
-        if let Err(e) =
-            Self::run_inner(&mut guard, conf, remote_storage.as_ref(), &tenant, cancel).await
-        {
+        if let Err(e) = Self::run_inner(&mut guard, conf, &remote_storage, &tenant, cancel).await {
             tenant.set_broken(format!("{e:#}")).await;
             return Err(e);
         }
@@ -327,7 +323,7 @@ impl DeleteTenantFlow {
     async fn run_inner(
         guard: &mut OwnedMutexGuard<Self>,
         conf: &'static PageServerConf,
-        remote_storage: Option<&GenericRemoteStorage>,
+        remote_storage: &GenericRemoteStorage,
         tenant: &Tenant,
         cancel: &CancellationToken,
     ) -> Result<(), DeleteTenantError> {
@@ -339,14 +335,9 @@ impl DeleteTenantFlow {
             ))?
         });
 
-        // IDEA: implement detach as delete without remote storage. Then they would use the same lock (deletion_progress) so wont contend.
-        // Though sounds scary, different mark name?
-        // Detach currently uses remove_dir_all so in case of a crash we can end up in a weird state.
-        if let Some(remote_storage) = &remote_storage {
-            create_remote_delete_mark(conf, remote_storage, &tenant.tenant_shard_id, cancel)
-                .await
-                .context("remote_mark")?
-        }
+        create_remote_delete_mark(conf, remote_storage, &tenant.tenant_shard_id, cancel)
+            .await
+            .context("remote_mark")?;
 
         fail::fail_point!("tenant-delete-before-create-local-mark", |_| {
             Err(anyhow::anyhow!(
@@ -483,7 +474,7 @@ impl DeleteTenantFlow {
     fn schedule_background(
         guard: OwnedMutexGuard<Self>,
         conf: &'static PageServerConf,
-        remote_storage: Option<GenericRemoteStorage>,
+        remote_storage: GenericRemoteStorage,
         tenants: &'static std::sync::RwLock<TenantsMap>,
         tenant: Arc<Tenant>,
     ) {
@@ -512,7 +503,7 @@ impl DeleteTenantFlow {
     async fn background(
         mut guard: OwnedMutexGuard<Self>,
         conf: &PageServerConf,
-        remote_storage: Option<GenericRemoteStorage>,
+        remote_storage: GenericRemoteStorage,
         tenants: &'static std::sync::RwLock<TenantsMap>,
         tenant: &Arc<Tenant>,
     ) -> Result<(), DeleteTenantError> {
@@ -551,7 +542,7 @@ impl DeleteTenantFlow {
 
         remove_tenant_remote_delete_mark(
             conf,
-            remote_storage.as_ref(),
+            &remote_storage,
             &tenant.tenant_shard_id,
             &task_mgr::shutdown_token(),
         )

pageserver/src/tenant/mgr.rs

@@ -7,7 +7,7 @@ use itertools::Itertools;
 use pageserver_api::key::Key;
 use pageserver_api::models::LocationConfigMode;
 use pageserver_api::shard::{
-    ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId,
+    ShardCount, ShardIdentity, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId,
 };
 use pageserver_api::upcall_api::ReAttachResponseTenant;
 use rand::{distributions::Alphanumeric, Rng};
@@ -16,10 +16,9 @@ use std::cmp::Ordering;
 use std::collections::{BTreeMap, HashMap};
 use std::ops::Deref;
 use std::sync::Arc;
-use std::time::{Duration, Instant};
+use std::time::Duration;
 use sysinfo::SystemExt;
 use tokio::fs;
-use utils::timeout::{timeout_cancellable, TimeoutCancellableError};
 
 use anyhow::Context;
 use once_cell::sync::Lazy;
@@ -47,7 +46,7 @@ use crate::tenant::span::debug_assert_current_span_has_tenant_id;
 use crate::tenant::storage_layer::inmemory_layer;
 use crate::tenant::timeline::ShutdownMode;
 use crate::tenant::{AttachedTenantConf, SpawnMode, Tenant, TenantState};
-use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TEMP_FILE_SUFFIX};
+use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, TEMP_FILE_SUFFIX};
 
 use utils::crashsafe::path_with_suffix_extension;
 use utils::fs_ext::PathExt;
@@ -119,6 +118,7 @@ pub(crate) enum TenantsMapRemoveResult {
 
 /// When resolving a TenantId to a shard, we may be looking for the 0th
 /// shard, or we might be looking for whichever shard holds a particular page.
+#[derive(Copy, Clone)]
 pub(crate) enum ShardSelector {
     /// Only return the 0th shard, if it is present.  If a non-0th shard is present,
     /// ignore it.
@@ -127,6 +127,8 @@ pub(crate) enum ShardSelector {
     First,
     /// Pick the shard that holds this key
     Page(Key),
+    /// The shard ID is known: pick the given shard
+    Known(ShardIndex),
 }
 
 /// A convenience for use with the re_attach ControlPlaneClient function: rather
@@ -169,6 +171,14 @@ impl TenantStartupMode {
     }
 }
 
+/// Result type for looking up a TenantId to a specific shard
+pub(crate) enum ShardResolveResult {
+    NotFound,
+    Found(Arc<Tenant>),
+    // Wait for this barrrier, then query again
+    InProgress(utils::completion::Barrier),
+}
+
 impl TenantsMap {
     /// Convenience function for typical usage, where we want to get a `Tenant` object, for
     /// working with attached tenants.  If the TenantId is in the map but in Secondary state,
@@ -182,51 +192,6 @@ impl TenantsMap {
         }
     }
 
-    /// A page service client sends a TenantId, and to look up the correct Tenant we must
-    /// resolve this to a fully qualified TenantShardId.
-    fn resolve_attached_shard(
-        &self,
-        tenant_id: &TenantId,
-        selector: ShardSelector,
-    ) -> Option<TenantShardId> {
-        let mut want_shard = None;
-        match self {
-            TenantsMap::Initializing => None,
-            TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
-                for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
-                    // Ignore all slots that don't contain an attached tenant
-                    let tenant = match &slot.1 {
-                        TenantSlot::Attached(t) => t,
-                        _ => continue,
-                    };
-
-                    match selector {
-                        ShardSelector::First => return Some(*slot.0),
-                        ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
-                            return Some(*slot.0)
-                        }
-                        ShardSelector::Page(key) => {
-                            // First slot we see for this tenant, calculate the expected shard number
-                            // for the key: we will use this for checking if this and subsequent
-                            // slots contain the key, rather than recalculating the hash each time.
-                            if want_shard.is_none() {
-                                want_shard = Some(tenant.shard_identity.get_shard_number(&key));
-                            }
-
-                            if Some(tenant.shard_identity.number) == want_shard {
-                                return Some(*slot.0);
-                            }
-                        }
-                        _ => continue,
-                    }
-                }
-
-                // Fall through: we didn't find an acceptable shard
-                None
-            }
-        }
-    }
-
     /// Only for use from DeleteTenantFlow.  This method directly removes a TenantSlot from the map.
     ///
     /// The normal way to remove a tenant is using a SlotGuard, which will gracefully remove the guarded
@@ -391,22 +356,17 @@ async fn init_load_generations(
     // deletion list entries may still be valid.  We provide that by pushing a recovery operation into
     // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
     // are processed, even though we don't block on recovery completing here.
-    //
-    // Must only do this if remote storage is enabled, otherwise deletion queue
-    // is not running and channel push will fail.
-    if resources.remote_storage.is_some() {
-        let attached_tenants = generations
-            .iter()
-            .flat_map(|(id, start_mode)| {
-                match start_mode {
-                    TenantStartupMode::Attached((_mode, generation)) => Some(generation),
-                    TenantStartupMode::Secondary => None,
-                }
-                .map(|gen| (*id, *gen))
-            })
-            .collect();
-        resources.deletion_queue_client.recover(attached_tenants)?;
-    }
+    let attached_tenants = generations
+        .iter()
+        .flat_map(|(id, start_mode)| {
+            match start_mode {
+                TenantStartupMode::Attached((_mode, generation)) => Some(generation),
+                TenantStartupMode::Secondary => None,
+            }
+            .map(|gen| (*id, *gen))
+        })
+        .collect();
+    resources.deletion_queue_client.recover(attached_tenants)?;
 
     Ok(Some(generations))
 }
@@ -460,53 +420,6 @@ 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");
@@ -611,6 +524,7 @@ pub async fn init_tenant_mgr(
                     TenantSlot::Attached(Tenant::create_broken_tenant(
                         conf,
                         tenant_shard_id,
+                        resources.remote_storage.clone(),
                         format!("{}", e),
                     )),
                 );
@@ -803,6 +717,7 @@ fn tenant_spawn(
         "Cannot load tenant, ignore mark found at {tenant_ignore_mark:?}"
     );
 
+    let remote_storage = resources.remote_storage.clone();
     let tenant = match Tenant::spawn(
         conf,
         tenant_shard_id,
@@ -817,7 +732,7 @@ fn tenant_spawn(
         Ok(tenant) => tenant,
         Err(e) => {
             error!("Failed to spawn tenant {tenant_shard_id}, reason: {e:#}");
-            Tenant::create_broken_tenant(conf, tenant_shard_id, format!("{e:#}"))
+            Tenant::create_broken_tenant(conf, tenant_shard_id, remote_storage, format!("{e:#}"))
         }
     };
 
@@ -2103,6 +2018,77 @@ impl TenantManager {
 
         Ok(reparented)
     }
+
+    /// A page service client sends a TenantId, and to look up the correct Tenant we must
+    /// resolve this to a fully qualified TenantShardId.
+    ///
+    /// During shard splits: we shall see parent shards in InProgress state and skip them, and
+    /// instead match on child shards which should appear in Attached state.  Very early in a shard
+    /// split, or in other cases where a shard is InProgress, we will return our own InProgress result
+    /// to instruct the caller to wait for that to finish before querying again.
+    pub(crate) fn resolve_attached_shard(
+        &self,
+        tenant_id: &TenantId,
+        selector: ShardSelector,
+    ) -> ShardResolveResult {
+        let tenants = self.tenants.read().unwrap();
+        let mut want_shard = None;
+        let mut any_in_progress = None;
+
+        match &*tenants {
+            TenantsMap::Initializing => ShardResolveResult::NotFound,
+            TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
+                for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
+                    // Ignore all slots that don't contain an attached tenant
+                    let tenant = match &slot.1 {
+                        TenantSlot::Attached(t) => t,
+                        TenantSlot::InProgress(barrier) => {
+                            // We might still find a usable shard, but in case we don't, remember that
+                            // we saw at least one InProgress slot, so that we can distinguish this case
+                            // from a simple NotFound in our return value.
+                            any_in_progress = Some(barrier.clone());
+                            continue;
+                        }
+                        _ => continue,
+                    };
+
+                    match selector {
+                        ShardSelector::First => return ShardResolveResult::Found(tenant.clone()),
+                        ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
+                            return ShardResolveResult::Found(tenant.clone())
+                        }
+                        ShardSelector::Page(key) => {
+                            // First slot we see for this tenant, calculate the expected shard number
+                            // for the key: we will use this for checking if this and subsequent
+                            // slots contain the key, rather than recalculating the hash each time.
+                            if want_shard.is_none() {
+                                want_shard = Some(tenant.shard_identity.get_shard_number(&key));
+                            }
+
+                            if Some(tenant.shard_identity.number) == want_shard {
+                                return ShardResolveResult::Found(tenant.clone());
+                            }
+                        }
+                        ShardSelector::Known(shard)
+                            if tenant.shard_identity.shard_index() == shard =>
+                        {
+                            return ShardResolveResult::Found(tenant.clone());
+                        }
+                        _ => continue,
+                    }
+                }
+
+                // Fall through: we didn't find a slot that was in Attached state & matched our selector.  If
+                // we found one or more InProgress slot, indicate to caller that they should retry later.  Otherwise
+                // this requested shard simply isn't found.
+                if let Some(barrier) = any_in_progress {
+                    ShardResolveResult::InProgress(barrier)
+                } else {
+                    ShardResolveResult::NotFound
+                }
+            }
+        }
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -2151,105 +2137,6 @@ pub(crate) enum GetActiveTenantError {
     Broken(String),
 }
 
-/// Get a [`Tenant`] in its active state. If the tenant_id is currently in [`TenantSlot::InProgress`]
-/// state, then wait for up to `timeout`.  If the [`Tenant`] is not currently in [`TenantState::Active`],
-/// then wait for up to `timeout` (minus however long we waited for the slot).
-pub(crate) async fn get_active_tenant_with_timeout(
-    tenant_id: TenantId,
-    shard_selector: ShardSelector,
-    timeout: Duration,
-    cancel: &CancellationToken,
-) -> Result<Arc<Tenant>, GetActiveTenantError> {
-    enum WaitFor {
-        Barrier(utils::completion::Barrier),
-        Tenant(Arc<Tenant>),
-    }
-
-    let wait_start = Instant::now();
-    let deadline = wait_start + timeout;
-
-    let (wait_for, tenant_shard_id) = {
-        let locked = TENANTS.read().unwrap();
-
-        // Resolve TenantId to TenantShardId
-        let tenant_shard_id = locked
-            .resolve_attached_shard(&tenant_id, shard_selector)
-            .ok_or(GetActiveTenantError::NotFound(GetTenantError::NotFound(
-                tenant_id,
-            )))?;
-
-        let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
-            .map_err(GetTenantError::MapState)?;
-        match peek_slot {
-            Some(TenantSlot::Attached(tenant)) => {
-                match tenant.current_state() {
-                    TenantState::Active => {
-                        // Fast path: we don't need to do any async waiting.
-                        return Ok(tenant.clone());
-                    }
-                    _ => {
-                        tenant.activate_now();
-                        (WaitFor::Tenant(tenant.clone()), tenant_shard_id)
-                    }
-                }
-            }
-            Some(TenantSlot::Secondary(_)) => {
-                return Err(GetActiveTenantError::NotFound(GetTenantError::NotActive(
-                    tenant_shard_id,
-                )))
-            }
-            Some(TenantSlot::InProgress(barrier)) => {
-                (WaitFor::Barrier(barrier.clone()), tenant_shard_id)
-            }
-            None => {
-                return Err(GetActiveTenantError::NotFound(GetTenantError::NotFound(
-                    tenant_id,
-                )))
-            }
-        }
-    };
-
-    let tenant = match wait_for {
-        WaitFor::Barrier(barrier) => {
-            tracing::debug!("Waiting for tenant InProgress state to pass...");
-            timeout_cancellable(
-                deadline.duration_since(Instant::now()),
-                cancel,
-                barrier.wait(),
-            )
-            .await
-            .map_err(|e| match e {
-                TimeoutCancellableError::Timeout => GetActiveTenantError::WaitForActiveTimeout {
-                    latest_state: None,
-                    wait_time: wait_start.elapsed(),
-                },
-                TimeoutCancellableError::Cancelled => GetActiveTenantError::Cancelled,
-            })?;
-            {
-                let locked = TENANTS.read().unwrap();
-                let peek_slot =
-                    tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
-                        .map_err(GetTenantError::MapState)?;
-                match peek_slot {
-                    Some(TenantSlot::Attached(tenant)) => tenant.clone(),
-                    _ => {
-                        return Err(GetActiveTenantError::NotFound(GetTenantError::NotActive(
-                            tenant_shard_id,
-                        )))
-                    }
-                }
-            }
-        }
-        WaitFor::Tenant(tenant) => tenant,
-    };
-
-    tracing::debug!("Waiting for tenant to enter active state...");
-    tenant
-        .wait_to_become_active(deadline.duration_since(Instant::now()))
-        .await?;
-    Ok(tenant)
-}
-
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum DeleteTimelineError {
     #[error("Tenant {0}")]
@@ -2276,7 +2163,7 @@ pub(crate) async fn load_tenant(
     tenant_id: TenantId,
     generation: Generation,
     broker_client: storage_broker::BrokerClientChannel,
-    remote_storage: Option<GenericRemoteStorage>,
+    remote_storage: GenericRemoteStorage,
     deletion_queue_client: DeletionQueueClient,
     ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
@@ -2880,86 +2767,73 @@ use {
     utils::http::error::ApiError,
 };
 
-pub(crate) fn immediate_gc(
+#[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))]
+pub(crate) async fn immediate_gc(
     tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
     gc_req: TimelineGcRequest,
     cancel: CancellationToken,
     ctx: &RequestContext,
-) -> Result<tokio::sync::oneshot::Receiver<Result<GcResult, anyhow::Error>>, ApiError> {
-    let guard = TENANTS.read().unwrap();
-
-    let tenant = guard
-        .get(&tenant_shard_id)
-        .cloned()
-        .with_context(|| format!("tenant {tenant_shard_id}"))
-        .map_err(|e| ApiError::NotFound(e.into()))?;
+) -> Result<GcResult, ApiError> {
+    let tenant = {
+        let guard = TENANTS.read().unwrap();
+        guard
+            .get(&tenant_shard_id)
+            .cloned()
+            .with_context(|| format!("tenant {tenant_shard_id}"))
+            .map_err(|e| ApiError::NotFound(e.into()))?
+    };
 
     let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
     // Use tenant's pitr setting
     let pitr = tenant.get_pitr_interval();
 
+    tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
+
     // Run in task_mgr to avoid race with tenant_detach operation
-    let ctx = ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
-    let (task_done, wait_task_done) = tokio::sync::oneshot::channel();
-    let span = info_span!("manual_gc", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id);
+    let ctx: RequestContext =
+        ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
 
-    // TODO: spawning is redundant now, need to hold the gate
-    task_mgr::spawn(
-        &tokio::runtime::Handle::current(),
-        TaskKind::GarbageCollector,
-        Some(tenant_shard_id),
-        Some(timeline_id),
-        &format!("timeline_gc_handler garbage collection run for tenant {tenant_shard_id} timeline {timeline_id}"),
-        false,
-        async move {
-            fail::fail_point!("immediate_gc_task_pre");
+    let _gate_guard = tenant.gate.enter().map_err(|_| ApiError::ShuttingDown)?;
 
-            #[allow(unused_mut)]
-            let mut result = tenant
-                .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
-                .await;
-                // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
-                // better once the types support it.
+    fail::fail_point!("immediate_gc_task_pre");
 
-            #[cfg(feature = "testing")]
-            {
-                // we need to synchronize with drop completion for python tests without polling for
-                // log messages
-                if let Ok(result) = result.as_mut() {
-                    let mut js = tokio::task::JoinSet::new();
-                    for layer in std::mem::take(&mut result.doomed_layers) {
-                        js.spawn(layer.wait_drop());
-                    }
-                    tracing::info!(total = js.len(), "starting to wait for the gc'd layers to be dropped");
-                    while let Some(res) = js.join_next().await {
-                        res.expect("wait_drop should not panic");
-                    }
-                }
+    #[allow(unused_mut)]
+    let mut result = tenant
+        .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
+        .await;
+    // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
+    // better once the types support it.
 
-                let timeline = tenant.get_timeline(timeline_id, false).ok();
-                let rtc = timeline.as_ref().and_then(|x| x.remote_client.as_ref());
-
-                if let Some(rtc) = rtc {
-                    // layer drops schedule actions on remote timeline client to actually do the
-                    // deletions; don't care about the shutdown error, just exit fast
-                    drop(rtc.wait_completion().await);
-                }
+    #[cfg(feature = "testing")]
+    {
+        // we need to synchronize with drop completion for python tests without polling for
+        // log messages
+        if let Ok(result) = result.as_mut() {
+            let mut js = tokio::task::JoinSet::new();
+            for layer in std::mem::take(&mut result.doomed_layers) {
+                js.spawn(layer.wait_drop());
             }
-
-            match task_done.send(result) {
-                Ok(_) => (),
-                Err(result) => error!("failed to send gc result: {result:?}"),
+            tracing::info!(
+                total = js.len(),
+                "starting to wait for the gc'd layers to be dropped"
+            );
+            while let Some(res) = js.join_next().await {
+                res.expect("wait_drop should not panic");
             }
-            Ok(())
         }
-        .instrument(span)
-    );
 
-    // drop the guard until after we've spawned the task so that timeline shutdown will wait for the task
-    drop(guard);
+        let timeline = tenant.get_timeline(timeline_id, false).ok();
+        let rtc = timeline.as_ref().map(|x| &x.remote_client);
 
-    Ok(wait_task_done)
+        if let Some(rtc) = rtc {
+            // layer drops schedule actions on remote timeline client to actually do the
+            // deletions; don't care about the shutdown error, just exit fast
+            drop(rtc.wait_completion().await);
+        }
+    }
+
+    result.map_err(ApiError::InternalServerError)
 }
 
 #[cfg(test)]

pageserver/src/tenant/remote_timeline_client.rs

@@ -189,6 +189,7 @@ use camino::Utf8Path;
 use chrono::{NaiveDateTime, Utc};
 
 pub(crate) use download::download_initdb_tar_zst;
+use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::shard::{ShardIndex, TenantShardId};
 use scopeguard::ScopeGuard;
 use tokio_util::sync::CancellationToken;
@@ -196,6 +197,7 @@ pub(crate) use upload::upload_initdb_dir;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
+use utils::pausable_failpoint;
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
@@ -317,7 +319,7 @@ pub struct RemoteTimelineClient {
 
     upload_queue: Mutex<UploadQueue>,
 
-    metrics: Arc<RemoteTimelineClientMetrics>,
+    pub(crate) metrics: Arc<RemoteTimelineClientMetrics>,
 
     storage_impl: GenericRemoteStorage,
 
@@ -461,11 +463,11 @@ impl RemoteTimelineClient {
         } else {
             0
         };
-        self.metrics.remote_physical_size_set(size);
+        self.metrics.remote_physical_size_gauge.set(size);
     }
 
     pub fn get_remote_physical_size(&self) -> u64 {
-        self.metrics.remote_physical_size_get()
+        self.metrics.remote_physical_size_gauge.get()
     }
 
     //
@@ -518,6 +520,7 @@ impl RemoteTimelineClient {
         &self,
         layer_file_name: &LayerName,
         layer_metadata: &LayerFileMetadata,
+        local_path: &Utf8Path,
         cancel: &CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<u64> {
@@ -536,6 +539,7 @@ impl RemoteTimelineClient {
                 self.timeline_id,
                 layer_file_name,
                 layer_metadata,
+                local_path,
                 cancel,
                 ctx,
             )
@@ -609,6 +613,17 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    /// Launch an index-file upload operation in the background, with only aux_file_policy flag updated.
+    pub(crate) fn schedule_index_upload_for_aux_file_policy_update(
+        self: &Arc<Self>,
+        last_aux_file_policy: Option<AuxFilePolicy>,
+    ) -> anyhow::Result<()> {
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+        upload_queue.last_aux_file_policy = last_aux_file_policy;
+        self.schedule_index_upload(upload_queue);
+        Ok(())
+    }
     ///
     /// Launch an index-file upload operation in the background, if necessary.
     ///
@@ -1127,6 +1142,11 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    pub(crate) fn is_deleting(&self) -> bool {
+        let mut locked = self.upload_queue.lock().unwrap();
+        locked.stopped_mut().is_ok()
+    }
+
     pub(crate) async fn preserve_initdb_archive(
         self: &Arc<Self>,
         tenant_id: &TenantId,
@@ -1173,7 +1193,7 @@ impl RemoteTimelineClient {
                     &self.storage_impl,
                     uploaded.local_path(),
                     &remote_path,
-                    uploaded.metadata().file_size(),
+                    uploaded.metadata().file_size,
                     cancel,
                 )
                 .await
@@ -1554,7 +1574,7 @@ impl RemoteTimelineClient {
                         &self.storage_impl,
                         local_path,
                         &remote_path,
-                        layer_metadata.file_size(),
+                        layer_metadata.file_size,
                         &self.cancel,
                     )
                     .measure_remote_op(
@@ -1749,7 +1769,7 @@ impl RemoteTimelineClient {
             UploadOp::UploadLayer(_, m) => (
                 RemoteOpFileKind::Layer,
                 RemoteOpKind::Upload,
-                RemoteTimelineClientMetricsCallTrackSize::Bytes(m.file_size()),
+                RemoteTimelineClientMetricsCallTrackSize::Bytes(m.file_size),
             ),
             UploadOp::UploadMetadata(_, _) => (
                 RemoteOpFileKind::Index,
@@ -1844,6 +1864,7 @@ impl RemoteTimelineClient {
                         dangling_files: HashMap::default(),
                         shutting_down: false,
                         shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
+                        last_aux_file_policy: initialized.last_aux_file_policy,
                     };
 
                     let upload_queue = std::mem::replace(
@@ -2132,7 +2153,7 @@ mod tests {
             tenant_ctx: _tenant_ctx,
         } = test_setup;
 
-        let client = timeline.remote_client.as_ref().unwrap();
+        let client = &timeline.remote_client;
 
         // Download back the index.json, and check that the list of files is correct
         let initial_index_part = match client
@@ -2323,7 +2344,7 @@ mod tests {
             timeline,
             ..
         } = TestSetup::new("metrics").await.unwrap();
-        let client = timeline.remote_client.as_ref().unwrap();
+        let client = &timeline.remote_client;
 
         let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
         let local_path = local_layer_path(

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -21,7 +21,6 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
-use crate::tenant::storage_layer::layer::local_layer_path;
 use crate::tenant::storage_layer::LayerName;
 use crate::tenant::Generation;
 use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
@@ -50,19 +49,13 @@ pub async fn download_layer_file<'a>(
     timeline_id: TimelineId,
     layer_file_name: &'a LayerName,
     layer_metadata: &'a LayerFileMetadata,
+    local_path: &Utf8Path,
     cancel: &CancellationToken,
     ctx: &RequestContext,
 ) -> Result<u64, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
     let timeline_path = conf.timeline_path(&tenant_shard_id, &timeline_id);
-    let local_path = local_layer_path(
-        conf,
-        &tenant_shard_id,
-        &timeline_id,
-        layer_file_name,
-        &layer_metadata.generation,
-    );
 
     let remote_path = remote_layer_path(
         &tenant_shard_id.tenant_id,
@@ -82,7 +75,7 @@ pub async fn download_layer_file<'a>(
     // For more context about durable_rename check this email from postgres mailing list:
     // https://www.postgresql.org/message-id/56583BDD.9060302@2ndquadrant.com
     // If pageserver crashes the temp file will be deleted on startup and re-downloaded.
-    let temp_file_path = path_with_suffix_extension(&local_path, TEMP_DOWNLOAD_EXTENSION);
+    let temp_file_path = path_with_suffix_extension(local_path, TEMP_DOWNLOAD_EXTENSION);
 
     let bytes_amount = download_retry(
         || async { download_object(storage, &remote_path, &temp_file_path, cancel, ctx).await },
@@ -91,7 +84,7 @@ pub async fn download_layer_file<'a>(
     )
     .await?;
 
-    let expected = layer_metadata.file_size();
+    let expected = layer_metadata.file_size;
     if expected != bytes_amount {
         return Err(DownloadError::Other(anyhow!(
             "According to layer file metadata should have downloaded {expected} bytes but downloaded {bytes_amount} bytes into file {temp_file_path:?}",

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -5,6 +5,7 @@
 use std::collections::HashMap;
 
 use chrono::NaiveDateTime;
+use pageserver_api::models::AuxFilePolicy;
 use serde::{Deserialize, Serialize};
 use utils::id::TimelineId;
 
@@ -16,46 +17,6 @@ use pageserver_api::shard::ShardIndex;
 
 use utils::lsn::Lsn;
 
-/// Metadata gathered for each of the layer files.
-///
-/// Fields have to be `Option`s because remote [`IndexPart`]'s can be from different version, which
-/// might have less or more metadata depending if upgrading or rolling back an upgrade.
-#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
-//#[cfg_attr(test, derive(Default))]
-pub struct LayerFileMetadata {
-    file_size: u64,
-
-    pub(crate) generation: Generation,
-
-    pub(crate) shard: ShardIndex,
-}
-
-impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
-    fn from(other: &IndexLayerMetadata) -> Self {
-        LayerFileMetadata {
-            file_size: other.file_size,
-            generation: other.generation,
-            shard: other.shard,
-        }
-    }
-}
-
-impl LayerFileMetadata {
-    pub fn new(file_size: u64, generation: Generation, shard: ShardIndex) -> Self {
-        LayerFileMetadata {
-            file_size,
-            generation,
-            shard,
-        }
-    }
-
-    pub fn file_size(&self) -> u64 {
-        self.file_size
-    }
-}
-
-// TODO seems like another part of the remote storage file format
-// compatibility issue, see https://github.com/neondatabase/neon/issues/3072
 /// In-memory representation of an `index_part.json` file
 ///
 /// Contains the data about all files in the timeline, present remotely and its metadata.
@@ -76,7 +37,7 @@ pub struct IndexPart {
     ///
     /// Older versions of `IndexPart` will not have this property or have only a part of metadata
     /// that latest version stores.
-    pub layer_metadata: HashMap<LayerName, IndexLayerMetadata>,
+    pub layer_metadata: HashMap<LayerName, LayerFileMetadata>,
 
     // 'disk_consistent_lsn' is a copy of the 'disk_consistent_lsn' in the metadata.
     // It's duplicated for convenience when reading the serialized structure, but is
@@ -88,6 +49,16 @@ pub struct IndexPart {
 
     #[serde(default)]
     pub(crate) lineage: Lineage,
+
+    /// Describes the kind of aux files stored in the timeline.
+    ///
+    /// The value is modified during file ingestion when the latest wanted value communicated via tenant config is applied if it is acceptable.
+    /// A V1 setting after V2 files have been committed is not accepted.
+    ///
+    /// None means no aux files have been written to the storage before the point
+    /// when this flag is introduced.
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub(crate) last_aux_file_policy: Option<AuxFilePolicy>,
 }
 
 impl IndexPart {
@@ -101,10 +72,11 @@ impl IndexPart {
     ///      is always generated from the keys of `layer_metadata`)
     /// - 4: timeline_layers is fully removed.
     /// - 5: lineage was added
-    const LATEST_VERSION: usize = 5;
+    /// - 6: last_aux_file_policy is added.
+    const LATEST_VERSION: usize = 6;
 
     // Versions we may see when reading from a bucket.
-    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5];
+    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5, 6];
 
     pub const FILE_NAME: &'static str = "index_part.json";
 
@@ -113,11 +85,9 @@ impl IndexPart {
         disk_consistent_lsn: Lsn,
         metadata: TimelineMetadata,
         lineage: Lineage,
+        last_aux_file_policy: Option<AuxFilePolicy>,
     ) -> Self {
-        let layer_metadata = layers_and_metadata
-            .iter()
-            .map(|(k, v)| (k.to_owned(), IndexLayerMetadata::from(v)))
-            .collect();
+        let layer_metadata = layers_and_metadata.clone();
 
         Self {
             version: Self::LATEST_VERSION,
@@ -126,6 +96,7 @@ impl IndexPart {
             metadata,
             deleted_at: None,
             lineage,
+            last_aux_file_policy,
         }
     }
 
@@ -155,8 +126,13 @@ impl IndexPart {
             example_metadata.disk_consistent_lsn(),
             example_metadata,
             Default::default(),
+            Some(AuxFilePolicy::V1),
         )
     }
+
+    pub(crate) fn last_aux_file_policy(&self) -> Option<AuxFilePolicy> {
+        self.last_aux_file_policy
+    }
 }
 
 impl From<&UploadQueueInitialized> for IndexPart {
@@ -165,13 +141,22 @@ impl From<&UploadQueueInitialized> for IndexPart {
         let metadata = uq.latest_metadata.clone();
         let lineage = uq.latest_lineage.clone();
 
-        Self::new(&uq.latest_files, disk_consistent_lsn, metadata, lineage)
+        Self::new(
+            &uq.latest_files,
+            disk_consistent_lsn,
+            metadata,
+            lineage,
+            uq.last_aux_file_policy,
+        )
     }
 }
 
-/// Serialized form of [`LayerFileMetadata`].
-#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub struct IndexLayerMetadata {
+/// Metadata gathered for each of the layer files.
+///
+/// Fields have to be `Option`s because remote [`IndexPart`]'s can be from different version, which
+/// might have less or more metadata depending if upgrading or rolling back an upgrade.
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
+pub struct LayerFileMetadata {
     pub file_size: u64,
 
     #[serde(default = "Generation::none")]
@@ -183,12 +168,12 @@ pub struct IndexLayerMetadata {
     pub shard: ShardIndex,
 }
 
-impl From<&LayerFileMetadata> for IndexLayerMetadata {
-    fn from(other: &LayerFileMetadata) -> Self {
-        IndexLayerMetadata {
-            file_size: other.file_size,
-            generation: other.generation,
-            shard: other.shard,
+impl LayerFileMetadata {
+    pub fn new(file_size: u64, generation: Generation, shard: ShardIndex) -> Self {
+        LayerFileMetadata {
+            file_size,
+            generation,
+            shard,
         }
     }
 }
@@ -282,12 +267,12 @@ mod tests {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
             layer_metadata: HashMap::from([
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
                     file_size: 25600000,
                     generation: Generation::none(),
                     shard: ShardIndex::unsharded()
                 }),
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
@@ -299,6 +284,7 @@ mod tests {
             metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
             deleted_at: None,
             lineage: Lineage::default(),
+            last_aux_file_policy: None,
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -323,12 +309,12 @@ mod tests {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
             layer_metadata: HashMap::from([
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
                     file_size: 25600000,
                     generation: Generation::none(),
                     shard: ShardIndex::unsharded()
                 }),
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
@@ -340,6 +326,7 @@ mod tests {
             metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
             deleted_at: None,
             lineage: Lineage::default(),
+            last_aux_file_policy: None,
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -365,12 +352,12 @@ mod tests {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 2,
             layer_metadata: HashMap::from([
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
                     file_size: 25600000,
                     generation: Generation::none(),
                     shard: ShardIndex::unsharded()
                 }),
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
@@ -383,6 +370,7 @@ mod tests {
             deleted_at: Some(chrono::NaiveDateTime::parse_from_str(
                 "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap()),
             lineage: Lineage::default(),
+            last_aux_file_policy: None,
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -428,6 +416,7 @@ mod tests {
             .unwrap(),
             deleted_at: None,
             lineage: Lineage::default(),
+            last_aux_file_policy: None,
         };
 
         let empty_layers_parsed = IndexPart::from_s3_bytes(empty_layers_json.as_bytes()).unwrap();
@@ -451,12 +440,12 @@ mod tests {
         let expected = IndexPart {
             version: 4,
             layer_metadata: HashMap::from([
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
                     file_size: 25600000,
                     generation: Generation::none(),
                     shard: ShardIndex::unsharded()
                 }),
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
@@ -468,6 +457,7 @@ mod tests {
             metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
             deleted_at: Some(parse_naive_datetime("2023-07-31T09:00:00.123000000")),
             lineage: Lineage::default(),
+            last_aux_file_policy: None,
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -492,12 +482,12 @@ mod tests {
         let expected = IndexPart {
             version: 5,
             layer_metadata: HashMap::from([
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF420-00000000014EF499".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF420-00000000014EF499".parse().unwrap(), LayerFileMetadata {
                     file_size: 23289856,
                     generation: Generation::new(1),
                     shard: ShardIndex::unsharded(),
                 }),
-                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF499-00000000015A7619".parse().unwrap(), IndexLayerMetadata {
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF499-00000000015A7619".parse().unwrap(), LayerFileMetadata {
                     file_size: 1015808,
                     generation: Generation::new(1),
                     shard: ShardIndex::unsharded(),
@@ -511,6 +501,57 @@ mod tests {
                 reparenting_history: vec![TimelineId::from_str("e1bfd8c633d713d279e6fcd2bcc15b6d").unwrap()],
                 original_ancestor: Some((TimelineId::from_str("e2bfd8c633d713d279e6fcd2bcc15b6d").unwrap(), Lsn::from_str("0/15A7618").unwrap(), parse_naive_datetime("2024-05-07T18:52:36.322426563"))),
             },
+            last_aux_file_policy: None,
+        };
+
+        let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v6_indexpart_is_parsed() {
+        let example = r#"{
+            "version":6,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata_bytes":[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+            "deleted_at": "2023-07-31T09:00:00.123",
+            "lineage":{
+                "original_ancestor":["e2bfd8c633d713d279e6fcd2bcc15b6d","0/15A7618","2024-05-07T18:52:36.322426563"],
+                "reparenting_history":["e1bfd8c633d713d279e6fcd2bcc15b6d"]
+            },
+            "last_aux_file_policy": "V2"
+        }"#;
+
+        let expected = IndexPart {
+            version: 6,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
+                    // serde_json should always parse this but this might be a double with jq for
+                    // example.
+                    file_size: 9007199254741001,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
+            deleted_at: Some(chrono::NaiveDateTime::parse_from_str(
+                "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap()),
+            lineage: Lineage {
+                reparenting_history_truncated: false,
+                reparenting_history: vec![TimelineId::from_str("e1bfd8c633d713d279e6fcd2bcc15b6d").unwrap()],
+                original_ancestor: Some((TimelineId::from_str("e2bfd8c633d713d279e6fcd2bcc15b6d").unwrap(), Lsn::from_str("0/15A7618").unwrap(), parse_naive_datetime("2024-05-07T18:52:36.322426563"))),
+            },
+            last_aux_file_policy: Some(AuxFilePolicy::V2),
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -9,7 +9,7 @@ use std::time::SystemTime;
 use tokio::fs::{self, File};
 use tokio::io::AsyncSeekExt;
 use tokio_util::sync::CancellationToken;
-use utils::backoff;
+use utils::{backoff, pausable_failpoint};
 
 use super::Generation;
 use crate::tenant::remote_timeline_client::{

pageserver/src/tenant/secondary.rs

@@ -6,11 +6,9 @@ mod scheduler;
 use std::{sync::Arc, time::SystemTime};
 
 use crate::{
-    config::PageServerConf,
     context::RequestContext,
     disk_usage_eviction_task::DiskUsageEvictionInfo,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
-    virtual_file::MaybeFatalIo,
 };
 
 use self::{
@@ -21,9 +19,8 @@ use self::{
 use super::{
     config::{SecondaryLocationConfig, TenantConfOpt},
     mgr::TenantManager,
-    remote_timeline_client::LayerFileMetadata,
     span::debug_assert_current_span_has_tenant_id,
-    storage_layer::{layer::local_layer_path, LayerName},
+    storage_layer::LayerName,
 };
 
 use pageserver_api::{
@@ -178,13 +175,7 @@ impl SecondaryTenant {
 
     /// Cancellation safe, but on cancellation the eviction will go through
     #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline_id, name=%name))]
-    pub(crate) async fn evict_layer(
-        self: &Arc<Self>,
-        conf: &PageServerConf,
-        timeline_id: TimelineId,
-        name: LayerName,
-        metadata: LayerFileMetadata,
-    ) {
+    pub(crate) async fn evict_layer(self: &Arc<Self>, timeline_id: TimelineId, name: LayerName) {
         debug_assert_current_span_has_tenant_id();
 
         let guard = match self.gate.enter() {
@@ -197,41 +188,11 @@ impl SecondaryTenant {
 
         let now = SystemTime::now();
 
-        let local_path = local_layer_path(
-            conf,
-            &self.tenant_shard_id,
-            &timeline_id,
-            &name,
-            &metadata.generation,
-        );
-
         let this = self.clone();
 
         // spawn it to be cancellation safe
         tokio::task::spawn_blocking(move || {
             let _guard = guard;
-            // We tolerate ENOENT, because between planning eviction and executing
-            // it, the secondary downloader could have seen an updated heatmap that
-            // resulted in a layer being deleted.
-            // Other local I/O errors are process-fatal: these should never happen.
-            let deleted = std::fs::remove_file(local_path);
-
-            let not_found = deleted
-                .as_ref()
-                .is_err_and(|x| x.kind() == std::io::ErrorKind::NotFound);
-
-            let deleted = if not_found {
-                false
-            } else {
-                deleted
-                    .map(|()| true)
-                    .fatal_err("Deleting layer during eviction")
-            };
-
-            if !deleted {
-                // skip updating accounting and putting perhaps later timestamp
-                return;
-            }
 
             // Update the timeline's state.  This does not have to be synchronized with
             // the download process, because:
@@ -250,8 +211,15 @@ impl SecondaryTenant {
             // of the cache.
             let mut detail = this.detail.lock().unwrap();
             if let Some(timeline_detail) = detail.timelines.get_mut(&timeline_id) {
-                timeline_detail.on_disk_layers.remove(&name);
-                timeline_detail.evicted_at.insert(name, now);
+                let removed = timeline_detail.on_disk_layers.remove(&name);
+
+                // We might race with removal of the same layer during downloads, if it was removed
+                // from the heatmap.  If we see that the OnDiskState is gone, then no need to
+                // do a physical deletion or store in evicted_at.
+                if let Some(removed) = removed {
+                    removed.remove_blocking();
+                    timeline_detail.evicted_at.insert(name, now);
+                }
             }
         })
         .await

pageserver/src/tenant/secondary/downloader.rs

@@ -26,7 +26,7 @@ use crate::{
         tasks::{warn_when_period_overrun, BackgroundLoopKind},
     },
     virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile},
-    METADATA_FILE_NAME, TEMP_FILE_SUFFIX,
+    TEMP_FILE_SUFFIX,
 };
 
 use super::{
@@ -62,14 +62,10 @@ use super::{
     CommandRequest, DownloadCommand,
 };
 
-/// For each tenant, how long must have passed since the last download_tenant call before
-/// calling it again.  This is approximately the time by which local data is allowed
-/// to fall behind remote data.
-///
-/// TODO: this should just be a default, and the actual period should be controlled
-/// via the heatmap itself
-/// `<ttps://github.com/neondatabase/neon/issues/6200>`
-const DOWNLOAD_FRESHEN_INTERVAL: Duration = Duration::from_millis(60000);
+/// For each tenant, default period for how long must have passed since the last download_tenant call before
+/// calling it again.  This default is replaced with the value of [`HeatMapTenant::upload_period_ms`] after first
+/// download, if the uploader populated it.
+const DEFAULT_DOWNLOAD_INTERVAL: Duration = Duration::from_millis(60000);
 
 pub(super) async fn downloader_task(
     tenant_manager: Arc<TenantManager>,
@@ -90,7 +86,7 @@ pub(super) async fn downloader_task(
 
     scheduler
         .run(command_queue, background_jobs_can_start, cancel)
-        .instrument(info_span!("secondary_downloads"))
+        .instrument(info_span!("secondary_download_scheduler"))
         .await
 }
 
@@ -104,6 +100,7 @@ struct SecondaryDownloader {
 pub(super) struct OnDiskState {
     metadata: LayerFileMetadata,
     access_time: SystemTime,
+    local_path: Utf8PathBuf,
 }
 
 impl OnDiskState {
@@ -114,12 +111,26 @@ impl OnDiskState {
         _ame: LayerName,
         metadata: LayerFileMetadata,
         access_time: SystemTime,
+        local_path: Utf8PathBuf,
     ) -> Self {
         Self {
             metadata,
             access_time,
+            local_path,
         }
     }
+
+    // This is infallible, because all errors are either acceptable (ENOENT), or totally
+    // unexpected (fatal).
+    pub(super) fn remove_blocking(&self) {
+        // We tolerate ENOENT, because between planning eviction and executing
+        // it, the secondary downloader could have seen an updated heatmap that
+        // resulted in a layer being deleted.
+        // Other local I/O errors are process-fatal: these should never happen.
+        std::fs::remove_file(&self.local_path)
+            .or_else(fs_ext::ignore_not_found)
+            .fatal_err("Deleting secondary layer")
+    }
 }
 
 #[derive(Debug, Clone, Default)]
@@ -130,14 +141,22 @@ pub(super) struct SecondaryDetailTimeline {
     pub(super) evicted_at: HashMap<LayerName, SystemTime>,
 }
 
+// Aspects of a heatmap that we remember after downloading it
+#[derive(Clone, Debug)]
+struct DownloadSummary {
+    etag: Etag,
+    #[allow(unused)]
+    mtime: SystemTime,
+    upload_period: Duration,
+}
+
 /// This state is written by the secondary downloader, it is opaque
 /// to TenantManager
 #[derive(Debug)]
 pub(super) struct SecondaryDetail {
     pub(super) config: SecondaryLocationConfig,
 
-    last_download: Option<Instant>,
-    last_etag: Option<Etag>,
+    last_download: Option<DownloadSummary>,
     next_download: Option<Instant>,
     pub(super) timelines: HashMap<TimelineId, SecondaryDetailTimeline>,
 }
@@ -167,7 +186,6 @@ impl SecondaryDetail {
         Self {
             config,
             last_download: None,
-            last_etag: None,
             next_download: None,
             timelines: HashMap::new(),
         }
@@ -221,9 +239,8 @@ impl SecondaryDetail {
 
 struct PendingDownload {
     secondary_state: Arc<SecondaryTenant>,
-    last_download: Option<Instant>,
+    last_download: Option<DownloadSummary>,
     target_time: Option<Instant>,
-    period: Option<Duration>,
 }
 
 impl scheduler::PendingJob for PendingDownload {
@@ -273,10 +290,17 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
 
         tracing::debug!("Secondary tenant download completed");
 
-        // Update freshened_at even if there was an error: we don't want errored tenants to implicitly
-        // take priority to run again.
         let mut detail = secondary_state.detail.lock().unwrap();
-        detail.next_download = Some(Instant::now() + period_jitter(DOWNLOAD_FRESHEN_INTERVAL, 5));
+
+        let period = detail
+            .last_download
+            .as_ref()
+            .map(|d| d.upload_period)
+            .unwrap_or(DEFAULT_DOWNLOAD_INTERVAL);
+
+        // We advance next_download irrespective of errors: we don't want error cases to result in
+        // expensive busy-polling.
+        detail.next_download = Some(Instant::now() + period_jitter(period, 5));
     }
 
     async fn schedule(&mut self) -> SchedulingResult<PendingDownload> {
@@ -309,11 +333,11 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
                     if detail.next_download.is_none() {
                         // Initialize randomly in the range from 0 to our interval: this uniformly spreads the start times.  Subsequent
                         // rounds will use a smaller jitter to avoid accidentally synchronizing later.
-                        detail.next_download = Some(now.checked_add(period_warmup(DOWNLOAD_FRESHEN_INTERVAL)).expect(
+                        detail.next_download = Some(now.checked_add(period_warmup(DEFAULT_DOWNLOAD_INTERVAL)).expect(
                         "Using our constant, which is known to be small compared with clock range",
                     ));
                     }
-                    (detail.last_download, detail.next_download.unwrap())
+                    (detail.last_download.clone(), detail.next_download.unwrap())
                 };
 
                 if now > next_download {
@@ -321,7 +345,6 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
                         secondary_state: secondary_tenant,
                         last_download,
                         target_time: Some(next_download),
-                        period: Some(DOWNLOAD_FRESHEN_INTERVAL),
                     })
                 } else {
                     None
@@ -347,7 +370,6 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
 
         Ok(PendingDownload {
             target_time: None,
-            period: None,
             last_download: None,
             secondary_state: tenant,
         })
@@ -364,7 +386,6 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
             secondary_state,
             last_download,
             target_time,
-            period,
         } = job;
 
         let (completion, barrier) = utils::completion::channel();
@@ -386,7 +407,7 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
                     tracing::warn!("Insufficient space while downloading.  Will retry later.");
                 }
                 Err(UpdateError::Cancelled) => {
-                    tracing::debug!("Shut down while downloading");
+                    tracing::info!("Shut down while downloading");
                 },
                 Err(UpdateError::Deserialize(e)) => {
                     tracing::error!("Corrupt content while downloading tenant: {e}");
@@ -401,20 +422,15 @@ impl JobGenerator<PendingDownload, RunningDownload, CompleteDownload, DownloadCo
 
             // If the job had a target execution time, we may check our final execution
             // time against that for observability purposes.
-            if let (Some(target_time), Some(period)) = (target_time, period) {
-                // Only track execution lag if this isn't our first download: otherwise, it is expected
-                // that execution will have taken longer than our configured interval, for example
-                // when starting up a pageserver and
-                if last_download.is_some() {
-                    // Elapsed time includes any scheduling lag as well as the execution of the job
-                    let elapsed = Instant::now().duration_since(target_time);
+            if let (Some(target_time), Some(last_download)) = (target_time, last_download) {
+                // Elapsed time includes any scheduling lag as well as the execution of the job
+                let elapsed = Instant::now().duration_since(target_time);
 
-                    warn_when_period_overrun(
-                        elapsed,
-                        period,
-                        BackgroundLoopKind::SecondaryDownload,
-                    );
-                }
+                warn_when_period_overrun(
+                    elapsed,
+                    last_download.upload_period,
+                    BackgroundLoopKind::SecondaryDownload,
+                );
             }
 
             CompleteDownload {
@@ -503,12 +519,12 @@ impl<'a> TenantDownloader<'a> {
         let tenant_shard_id = self.secondary_state.get_tenant_shard_id();
 
         // We will use the etag from last successful download to make the download conditional on changes
-        let last_etag = self
+        let last_download = self
             .secondary_state
             .detail
             .lock()
             .unwrap()
-            .last_etag
+            .last_download
             .clone();
 
         // Download the tenant's heatmap
@@ -517,7 +533,7 @@ impl<'a> TenantDownloader<'a> {
             etag: heatmap_etag,
             bytes: heatmap_bytes,
         } = match tokio::select!(
-            bytes = self.download_heatmap(last_etag.as_ref()) => {bytes?},
+            bytes = self.download_heatmap(last_download.as_ref().map(|d| &d.etag)) => {bytes?},
             _ = self.secondary_state.cancel.cancelled() => return Ok(())
         ) {
             HeatMapDownload::Unmodified => {
@@ -546,6 +562,39 @@ impl<'a> TenantDownloader<'a> {
             heatmap.timelines.len()
         );
 
+        // Get or initialize the local disk state for the timelines we will update
+        let mut timeline_states = HashMap::new();
+        for timeline in &heatmap.timelines {
+            let timeline_state = self
+                .secondary_state
+                .detail
+                .lock()
+                .unwrap()
+                .timelines
+                .get(&timeline.timeline_id)
+                .cloned();
+
+            let timeline_state = match timeline_state {
+                Some(t) => t,
+                None => {
+                    // We have no existing state: need to scan local disk for layers first.
+                    let timeline_state =
+                        init_timeline_state(self.conf, tenant_shard_id, timeline).await;
+
+                    // Re-acquire detail lock now that we're done with async load from local FS
+                    self.secondary_state
+                        .detail
+                        .lock()
+                        .unwrap()
+                        .timelines
+                        .insert(timeline.timeline_id, timeline_state.clone());
+                    timeline_state
+                }
+            };
+
+            timeline_states.insert(timeline.timeline_id, timeline_state);
+        }
+
         // Clean up any local layers that aren't in the heatmap.  We do this first for all timelines, on the general
         // principle that deletions should be done before writes wherever possible, and so that we can use this
         // phase to initialize our SecondaryProgress.
@@ -556,6 +605,10 @@ impl<'a> TenantDownloader<'a> {
 
         // Download the layers in the heatmap
         for timeline in heatmap.timelines {
+            let timeline_state = timeline_states
+                .remove(&timeline.timeline_id)
+                .expect("Just populated above");
+
             if self.secondary_state.cancel.is_cancelled() {
                 tracing::debug!(
                     "Cancelled before downloading timeline {}",
@@ -565,7 +618,7 @@ impl<'a> TenantDownloader<'a> {
             }
 
             let timeline_id = timeline.timeline_id;
-            self.download_timeline(timeline, ctx)
+            self.download_timeline(timeline, timeline_state, ctx)
                 .instrument(tracing::info_span!(
                     "secondary_download_timeline",
                     tenant_id=%tenant_shard_id.tenant_id,
@@ -577,7 +630,30 @@ impl<'a> TenantDownloader<'a> {
 
         // Only update last_etag after a full successful download: this way will not skip
         // the next download, even if the heatmap's actual etag is unchanged.
-        self.secondary_state.detail.lock().unwrap().last_etag = Some(heatmap_etag);
+        self.secondary_state.detail.lock().unwrap().last_download = Some(DownloadSummary {
+            etag: heatmap_etag,
+            mtime: heatmap_mtime,
+            upload_period: heatmap
+                .upload_period_ms
+                .map(|ms| Duration::from_millis(ms as u64))
+                .unwrap_or(DEFAULT_DOWNLOAD_INTERVAL),
+        });
+
+        // Robustness: we should have updated progress properly, but in case we didn't, make sure
+        // we don't leave the tenant in a state where we claim to have successfully downloaded
+        // everything, but our progress is incomplete.  The invariant here should be that if
+        // we have set `last_download` to this heatmap's etag, then the next time we see that
+        // etag we can safely do no work (i.e. we must be complete).
+        let mut progress = self.secondary_state.progress.lock().unwrap();
+        debug_assert!(progress.layers_downloaded == progress.layers_total);
+        debug_assert!(progress.bytes_downloaded == progress.bytes_total);
+        if progress.layers_downloaded != progress.layers_total
+            || progress.bytes_downloaded != progress.bytes_total
+        {
+            tracing::warn!("Correcting drift in progress stats ({progress:?})");
+            progress.layers_downloaded = progress.layers_total;
+            progress.bytes_downloaded = progress.bytes_total;
+        }
 
         Ok(())
     }
@@ -633,7 +709,7 @@ impl<'a> TenantDownloader<'a> {
                 let mut layer_byte_count: u64 = timeline_state
                     .on_disk_layers
                     .values()
-                    .map(|l| l.metadata.file_size())
+                    .map(|l| l.metadata.file_size)
                     .sum();
 
                 // Remove on-disk layers that are no longer present in heatmap
@@ -644,7 +720,7 @@ impl<'a> TenantDownloader<'a> {
                         .get(layer_file_name)
                         .unwrap()
                         .metadata
-                        .file_size();
+                        .file_size;
 
                     let local_path = local_layer_path(
                         self.conf,
@@ -754,6 +830,7 @@ impl<'a> TenantDownloader<'a> {
     async fn download_timeline(
         &self,
         timeline: HeatMapTimeline,
+        timeline_state: SecondaryDetailTimeline,
         ctx: &RequestContext,
     ) -> Result<(), UpdateError> {
         debug_assert_current_span_has_tenant_and_timeline_id();
@@ -762,34 +839,6 @@ impl<'a> TenantDownloader<'a> {
         // Accumulate updates to the state
         let mut touched = Vec::new();
 
-        // Clone a view of what layers already exist on disk
-        let timeline_state = self
-            .secondary_state
-            .detail
-            .lock()
-            .unwrap()
-            .timelines
-            .get(&timeline.timeline_id)
-            .cloned();
-
-        let timeline_state = match timeline_state {
-            Some(t) => t,
-            None => {
-                // We have no existing state: need to scan local disk for layers first.
-                let timeline_state =
-                    init_timeline_state(self.conf, tenant_shard_id, &timeline).await;
-
-                // Re-acquire detail lock now that we're done with async load from local FS
-                self.secondary_state
-                    .detail
-                    .lock()
-                    .unwrap()
-                    .timelines
-                    .insert(timeline.timeline_id, timeline_state.clone());
-                timeline_state
-            }
-        };
-
         tracing::debug!(timeline_id=%timeline.timeline_id, "Downloading layers, {} in heatmap", timeline.layers.len());
 
         // Download heatmap layers that are not present on local disk, or update their
@@ -807,20 +856,12 @@ impl<'a> TenantDownloader<'a> {
                 if cfg!(debug_assertions) {
                     // Debug for https://github.com/neondatabase/neon/issues/6966: check that the files we think
                     // are already present on disk are really there.
-                    let local_path = local_layer_path(
-                        self.conf,
-                        tenant_shard_id,
-                        &timeline.timeline_id,
-                        &layer.name,
-                        &layer.metadata.generation,
-                    );
-
-                    match tokio::fs::metadata(&local_path).await {
+                    match tokio::fs::metadata(&on_disk.local_path).await {
                         Ok(meta) => {
                             tracing::debug!(
                                 "Layer {} present at {}, size {}",
                                 layer.name,
-                                local_path,
+                                on_disk.local_path,
                                 meta.len(),
                             );
                         }
@@ -828,7 +869,7 @@ impl<'a> TenantDownloader<'a> {
                             tracing::warn!(
                                 "Layer {} not found at {} ({})",
                                 layer.name,
-                                local_path,
+                                on_disk.local_path,
                                 e
                             );
                             debug_assert!(false);
@@ -836,9 +877,7 @@ impl<'a> TenantDownloader<'a> {
                     }
                 }
 
-                if on_disk.metadata != LayerFileMetadata::from(&layer.metadata)
-                    || on_disk.access_time != layer.access_time
-                {
+                if on_disk.metadata != layer.metadata || on_disk.access_time != layer.access_time {
                     // We already have this layer on disk.  Update its access time.
                     tracing::debug!(
                         "Access time updated for layer {}: {} -> {}",
@@ -874,67 +913,16 @@ impl<'a> TenantDownloader<'a> {
                 }
             }
 
-            // Failpoint for simulating slow remote storage
-            failpoint_support::sleep_millis_async!(
-                "secondary-layer-download-sleep",
-                &self.secondary_state.cancel
-            );
-
-            // Note: no backoff::retry wrapper here because download_layer_file does its own retries internally
-            let downloaded_bytes = match download_layer_file(
-                self.conf,
-                self.remote_storage,
-                *tenant_shard_id,
-                timeline.timeline_id,
-                &layer.name,
-                &LayerFileMetadata::from(&layer.metadata),
-                &self.secondary_state.cancel,
-                ctx,
-            )
-            .await
+            match self
+                .download_layer(tenant_shard_id, &timeline.timeline_id, layer, ctx)
+                .await?
             {
-                Ok(bytes) => bytes,
-                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;
+                Some(layer) => touched.push(layer),
+                None => {
+                    // Not an error but we didn't download it: remote layer is missing.  Don't add it to the list of
+                    // things to consider touched.
                 }
-                Err(e) => return Err(e.into()),
-            };
-
-            if downloaded_bytes != layer.metadata.file_size {
-                let local_path = local_layer_path(
-                    self.conf,
-                    tenant_shard_id,
-                    &timeline.timeline_id,
-                    &layer.name,
-                    &layer.metadata.generation,
-                );
-
-                tracing::warn!(
-                    "Downloaded layer {} with unexpected size {} != {}.  Removing download.",
-                    layer.name,
-                    downloaded_bytes,
-                    layer.metadata.file_size
-                );
-
-                tokio::fs::remove_file(&local_path)
-                    .await
-                    .or_else(fs_ext::ignore_not_found)?;
-            } else {
-                tracing::info!("Downloaded layer {}, size {}", layer.name, downloaded_bytes);
-                let mut progress = self.secondary_state.progress.lock().unwrap();
-                progress.bytes_downloaded += downloaded_bytes;
-                progress.layers_downloaded += 1;
             }
-
-            SECONDARY_MODE.download_layer.inc();
-            touched.push(layer)
         }
 
         // Write updates to state to record layers we just downloaded or touched.
@@ -951,13 +939,21 @@ impl<'a> TenantDownloader<'a> {
                         v.get_mut().access_time = t.access_time;
                     }
                     Entry::Vacant(e) => {
+                        let local_path = local_layer_path(
+                            self.conf,
+                            tenant_shard_id,
+                            &timeline.timeline_id,
+                            &t.name,
+                            &t.metadata.generation,
+                        );
                         e.insert(OnDiskState::new(
                             self.conf,
                             tenant_shard_id,
                             &timeline.timeline_id,
                             t.name,
-                            LayerFileMetadata::from(&t.metadata),
+                            t.metadata.clone(),
                             t.access_time,
+                            local_path,
                         ));
                     }
                 }
@@ -966,6 +962,99 @@ impl<'a> TenantDownloader<'a> {
 
         Ok(())
     }
+
+    async fn download_layer(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+        layer: HeatMapLayer,
+        ctx: &RequestContext,
+    ) -> Result<Option<HeatMapLayer>, UpdateError> {
+        // Failpoint for simulating slow remote storage
+        failpoint_support::sleep_millis_async!(
+            "secondary-layer-download-sleep",
+            &self.secondary_state.cancel
+        );
+
+        let local_path = local_layer_path(
+            self.conf,
+            tenant_shard_id,
+            timeline_id,
+            &layer.name,
+            &layer.metadata.generation,
+        );
+
+        // Note: no backoff::retry wrapper here because download_layer_file does its own retries internally
+        tracing::info!(
+            "Starting download of layer {}, size {}",
+            layer.name,
+            layer.metadata.file_size
+        );
+        let downloaded_bytes = match download_layer_file(
+            self.conf,
+            self.remote_storage,
+            *tenant_shard_id,
+            *timeline_id,
+            &layer.name,
+            &layer.metadata,
+            &local_path,
+            &self.secondary_state.cancel,
+            ctx,
+        )
+        .await
+        {
+            Ok(bytes) => bytes,
+            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
+                );
+
+                // If the layer is 404, adjust the progress statistics to reflect that we will not download it.
+                let mut progress = self.secondary_state.progress.lock().unwrap();
+                progress.layers_total = progress.layers_total.saturating_sub(1);
+                progress.bytes_total = progress
+                    .bytes_total
+                    .saturating_sub(layer.metadata.file_size);
+
+                return Ok(None);
+            }
+            Err(e) => return Err(e.into()),
+        };
+
+        if downloaded_bytes != layer.metadata.file_size {
+            let local_path = local_layer_path(
+                self.conf,
+                tenant_shard_id,
+                timeline_id,
+                &layer.name,
+                &layer.metadata.generation,
+            );
+
+            tracing::warn!(
+                "Downloaded layer {} with unexpected size {} != {}.  Removing download.",
+                layer.name,
+                downloaded_bytes,
+                layer.metadata.file_size
+            );
+
+            tokio::fs::remove_file(&local_path)
+                .await
+                .or_else(fs_ext::ignore_not_found)?;
+        } else {
+            tracing::info!("Downloaded layer {}, size {}", layer.name, downloaded_bytes);
+            let mut progress = self.secondary_state.progress.lock().unwrap();
+            progress.bytes_downloaded += downloaded_bytes;
+            progress.layers_downloaded += 1;
+        }
+
+        SECONDARY_MODE.download_layer.inc();
+
+        Ok(Some(layer))
+    }
 }
 
 /// Scan local storage and build up Layer objects based on the metadata in a HeatMapTimeline
@@ -1015,11 +1104,7 @@ async fn init_timeline_state(
             .fatal_err(&format!("Read metadata on {}", file_path));
 
         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)
+        if crate::is_temporary(&file_path)
             || is_temp_download_file(&file_path)
             || is_ephemeral_file(file_name)
         {
@@ -1059,8 +1144,9 @@ async fn init_timeline_state(
                                     tenant_shard_id,
                                     &heatmap.timeline_id,
                                     name,
-                                    LayerFileMetadata::from(&remote_meta.metadata),
+                                    remote_meta.metadata.clone(),
                                     remote_meta.access_time,
+                                    file_path,
                                 ),
                             );
                         }

pageserver/src/tenant/secondary/heatmap.rs

@@ -1,6 +1,6 @@
 use std::time::SystemTime;
 
-use crate::tenant::{remote_timeline_client::index::IndexLayerMetadata, storage_layer::LayerName};
+use crate::tenant::{remote_timeline_client::index::LayerFileMetadata, storage_layer::LayerName};
 
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr, TimestampSeconds};
@@ -15,6 +15,14 @@ pub(super) struct HeatMapTenant {
     pub(super) generation: Generation,
 
     pub(super) timelines: Vec<HeatMapTimeline>,
+
+    /// Uploaders provide their own upload period in the heatmap, as a hint to downloaders
+    /// of how frequently it is worthwhile to check for updates.
+    ///
+    /// This is optional for backward compat, and because we sometimes might upload
+    /// a heatmap explicitly via API for a tenant that has no periodic upload configured.
+    #[serde(default)]
+    pub(super) upload_period_ms: Option<u128>,
 }
 
 #[serde_as]
@@ -30,7 +38,7 @@ pub(crate) struct HeatMapTimeline {
 #[derive(Serialize, Deserialize)]
 pub(crate) struct HeatMapLayer {
     pub(super) name: LayerName,
-    pub(super) metadata: IndexLayerMetadata,
+    pub(super) metadata: LayerFileMetadata,
 
     #[serde_as(as = "TimestampSeconds<i64>")]
     pub(super) access_time: SystemTime,
@@ -41,7 +49,7 @@ pub(crate) struct HeatMapLayer {
 impl HeatMapLayer {
     pub(crate) fn new(
         name: LayerName,
-        metadata: IndexLayerMetadata,
+        metadata: LayerFileMetadata,
         access_time: SystemTime,
     ) -> Self {
         Self {
@@ -81,4 +89,21 @@ impl HeatMapTenant {
 
         stats
     }
+
+    pub(crate) fn strip_atimes(self) -> Self {
+        Self {
+            timelines: self
+                .timelines
+                .into_iter()
+                .map(|mut tl| {
+                    for layer in &mut tl.layers {
+                        layer.access_time = SystemTime::UNIX_EPOCH;
+                    }
+                    tl
+                })
+                .collect(),
+            generation: self.generation,
+            upload_period_ms: self.upload_period_ms,
+        }
+    }
 }

pageserver/src/tenant/secondary/heatmap_uploader.rs

@@ -53,7 +53,7 @@ pub(super) async fn heatmap_uploader_task(
 
     scheduler
         .run(command_queue, background_jobs_can_start, cancel)
-        .instrument(info_span!("heatmap_uploader"))
+        .instrument(info_span!("heatmap_upload_scheduler"))
         .await
 }
 
@@ -80,7 +80,7 @@ impl RunningJob for WriteInProgress {
 
 struct UploadPending {
     tenant: Arc<Tenant>,
-    last_digest: Option<md5::Digest>,
+    last_upload: Option<LastUploadState>,
     target_time: Option<Instant>,
     period: Option<Duration>,
 }
@@ -94,7 +94,7 @@ impl scheduler::PendingJob for UploadPending {
 struct WriteComplete {
     tenant_shard_id: TenantShardId,
     completed_at: Instant,
-    digest: Option<md5::Digest>,
+    uploaded: Option<LastUploadState>,
     next_upload: Option<Instant>,
 }
 
@@ -115,10 +115,7 @@ struct UploaderTenantState {
     tenant: Weak<Tenant>,
 
     /// Digest of the serialized heatmap that we last successfully uploaded
-    ///
-    /// md5 is generally a bad hash.  We use it because it's convenient for interop with AWS S3's ETag,
-    /// which is also an md5sum.
-    last_digest: Option<md5::Digest>,
+    last_upload_state: Option<LastUploadState>,
 
     /// When the last upload attempt completed (may have been successful or failed)
     last_upload: Option<Instant>,
@@ -187,7 +184,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
                     tenant: Arc::downgrade(&tenant),
                     last_upload: None,
                     next_upload: Some(now.checked_add(period_warmup(period)).unwrap_or(now)),
-                    last_digest: None,
+                    last_upload_state: None,
                 });
 
             // Decline to do the upload if insufficient time has passed
@@ -195,10 +192,10 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
                 return;
             }
 
-            let last_digest = state.last_digest;
+            let last_upload = state.last_upload_state.clone();
             result.jobs.push(UploadPending {
                 tenant,
-                last_digest,
+                last_upload,
                 target_time: state.next_upload,
                 period: Some(period),
             });
@@ -218,7 +215,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
     ) {
         let UploadPending {
             tenant,
-            last_digest,
+            last_upload,
             target_time,
             period,
         } = job;
@@ -231,16 +228,16 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
             let _completion = completion;
 
             let started_at = Instant::now();
-            let digest = match upload_tenant_heatmap(remote_storage, &tenant, last_digest).await {
-                Ok(UploadHeatmapOutcome::Uploaded(digest)) => {
+            let uploaded = match upload_tenant_heatmap(remote_storage, &tenant, last_upload.clone()).await {
+                Ok(UploadHeatmapOutcome::Uploaded(uploaded)) => {
                     let duration = Instant::now().duration_since(started_at);
                     SECONDARY_MODE
                         .upload_heatmap_duration
                         .observe(duration.as_secs_f64());
                     SECONDARY_MODE.upload_heatmap.inc();
-                    Some(digest)
+                    Some(uploaded)
                 }
-                Ok(UploadHeatmapOutcome::NoChange | UploadHeatmapOutcome::Skipped) => last_digest,
+                Ok(UploadHeatmapOutcome::NoChange | UploadHeatmapOutcome::Skipped) => last_upload,
                 Err(UploadHeatmapError::Upload(e)) => {
                     tracing::warn!(
                         "Failed to upload heatmap for tenant {}: {e:#}",
@@ -251,11 +248,11 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
                         .upload_heatmap_duration
                         .observe(duration.as_secs_f64());
                     SECONDARY_MODE.upload_heatmap_errors.inc();
-                    last_digest
+                    last_upload
                 }
                 Err(UploadHeatmapError::Cancelled) => {
                     tracing::info!("Cancelled heatmap upload, shutting down");
-                    last_digest
+                    last_upload
                 }
             };
 
@@ -277,7 +274,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
             WriteComplete {
                     tenant_shard_id: *tenant.get_tenant_shard_id(),
                     completed_at: now,
-                    digest,
+                    uploaded,
                     next_upload,
                 }
         }.instrument(info_span!(parent: None, "heatmap_upload", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))))
@@ -299,7 +296,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
 
         Ok(UploadPending {
             // Ignore our state for last digest: this forces an upload even if nothing has changed
-            last_digest: None,
+            last_upload: None,
             tenant,
             target_time: None,
             period: None,
@@ -312,7 +309,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
         let WriteComplete {
             tenant_shard_id,
             completed_at,
-            digest,
+            uploaded,
             next_upload,
         } = completion;
         use std::collections::hash_map::Entry;
@@ -322,7 +319,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
             }
             Entry::Occupied(mut entry) => {
                 entry.get_mut().last_upload = Some(completed_at);
-                entry.get_mut().last_digest = digest;
+                entry.get_mut().last_upload_state = uploaded;
                 entry.get_mut().next_upload = next_upload
             }
         }
@@ -331,7 +328,7 @@ impl JobGenerator<UploadPending, WriteInProgress, WriteComplete, UploadCommand>
 
 enum UploadHeatmapOutcome {
     /// We successfully wrote to remote storage, with this digest.
-    Uploaded(md5::Digest),
+    Uploaded(LastUploadState),
     /// We did not upload because the heatmap digest was unchanged since the last upload
     NoChange,
     /// We skipped the upload for some reason, such as tenant/timeline not ready
@@ -347,12 +344,25 @@ enum UploadHeatmapError {
     Upload(#[from] anyhow::Error),
 }
 
+/// Digests describing the heatmap we most recently uploaded successfully.
+///
+/// md5 is generally a bad hash.  We use it because it's convenient for interop with AWS S3's ETag,
+/// which is also an md5sum.
+#[derive(Clone)]
+struct LastUploadState {
+    // Digest of json-encoded HeatMapTenant
+    uploaded_digest: md5::Digest,
+
+    // Digest without atimes set.
+    layers_only_digest: md5::Digest,
+}
+
 /// The inner upload operation.  This will skip if `last_digest` is Some and matches the digest
 /// of the object we would have uploaded.
 async fn upload_tenant_heatmap(
     remote_storage: GenericRemoteStorage,
     tenant: &Arc<Tenant>,
-    last_digest: Option<md5::Digest>,
+    last_upload: Option<LastUploadState>,
 ) -> Result<UploadHeatmapOutcome, UploadHeatmapError> {
     debug_assert_current_span_has_tenant_id();
 
@@ -368,6 +378,7 @@ async fn upload_tenant_heatmap(
     let mut heatmap = HeatMapTenant {
         timelines: Vec::new(),
         generation,
+        upload_period_ms: tenant.get_heatmap_period().map(|p| p.as_millis()),
     };
     let timelines = tenant.timelines.lock().unwrap().clone();
 
@@ -396,15 +407,31 @@ async fn upload_tenant_heatmap(
 
     // Serialize the heatmap
     let bytes = serde_json::to_vec(&heatmap).map_err(|e| anyhow::anyhow!(e))?;
-    let bytes = bytes::Bytes::from(bytes);
-    let size = bytes.len();
 
     // Drop out early if nothing changed since our last upload
     let digest = md5::compute(&bytes);
-    if Some(digest) == last_digest {
+    if Some(&digest) == last_upload.as_ref().map(|d| &d.uploaded_digest) {
         return Ok(UploadHeatmapOutcome::NoChange);
     }
 
+    // Calculate a digest that omits atimes, so that we can distinguish actual changes in
+    // layers from changes only in atimes.
+    let heatmap_size_bytes = heatmap.get_stats().bytes;
+    let layers_only_bytes =
+        serde_json::to_vec(&heatmap.strip_atimes()).map_err(|e| anyhow::anyhow!(e))?;
+    let layers_only_digest = md5::compute(&layers_only_bytes);
+    if heatmap_size_bytes < tenant.get_checkpoint_distance() {
+        // For small tenants, skip upload if only atimes changed. This avoids doing frequent
+        // uploads from long-idle tenants whose atimes are just incremented by periodic
+        // size calculations.
+        if Some(&layers_only_digest) == last_upload.as_ref().map(|d| &d.layers_only_digest) {
+            return Ok(UploadHeatmapOutcome::NoChange);
+        }
+    }
+
+    let bytes = bytes::Bytes::from(bytes);
+    let size = bytes.len();
+
     let path = remote_heatmap_path(tenant.get_tenant_shard_id());
 
     let cancel = &tenant.cancel;
@@ -436,5 +463,8 @@ async fn upload_tenant_heatmap(
 
     tracing::info!("Successfully uploaded {size} byte heatmap to {path}");
 
-    Ok(UploadHeatmapOutcome::Uploaded(digest))
+    Ok(UploadHeatmapOutcome::Uploaded(LastUploadState {
+        uploaded_digest: digest,
+        layers_only_digest,
+    }))
 }

pageserver/src/tenant/secondary/scheduler.rs

@@ -179,6 +179,13 @@ where
             // Schedule some work, if concurrency limit permits it
             self.spawn_pending();
 
+            // This message is printed every scheduling iteration as proof of liveness when looking at logs
+            tracing::info!(
+                "Status: {} tasks running, {} pending",
+                self.running.len(),
+                self.pending.len()
+            );
+
             // Between scheduling iterations, we will:
             //  - Drain any complete tasks and spawn pending tasks
             //  - Handle incoming administrative commands
@@ -258,7 +265,11 @@ where
 
         self.tasks.spawn(fut);
 
-        self.running.insert(tenant_shard_id, in_progress);
+        let replaced = self.running.insert(tenant_shard_id, in_progress);
+        debug_assert!(replaced.is_none());
+        if replaced.is_some() {
+            tracing::warn!(%tenant_shard_id, "Unexpectedly spawned a task when one was already running")
+        }
     }
 
     /// For all pending tenants that are elegible for execution, spawn their task.
@@ -268,7 +279,9 @@ where
         while !self.pending.is_empty() && self.running.len() < self.concurrency {
             // unwrap: loop condition includes !is_empty()
             let pending = self.pending.pop_front().unwrap();
-            self.do_spawn(pending);
+            if !self.running.contains_key(pending.get_tenant_shard_id()) {
+                self.do_spawn(pending);
+            }
         }
     }
 
@@ -321,7 +334,8 @@ where
 
         let tenant_shard_id = job.get_tenant_shard_id();
         let barrier = if let Some(barrier) = self.get_running(tenant_shard_id) {
-            tracing::info!("Command already running, waiting for it");
+            tracing::info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
+                           "Command already running, waiting for it");
             barrier
         } else {
             let running = self.spawn_now(job);

pageserver/src/tenant/storage_layer.rs

@@ -113,12 +113,20 @@ impl From<VectoredValueReconstructState> for ValueReconstructState {
     }
 }
 
-/// Bag of data accumulated during a vectored get
+/// Bag of data accumulated during a vectored get..
 pub(crate) struct ValuesReconstructState {
+    /// The keys will be removed after `get_vectored` completes. The caller outside `Timeline`
+    /// should not expect to get anything from this hashmap.
     pub(crate) keys: HashMap<Key, Result<VectoredValueReconstructState, PageReconstructError>>,
-
+    /// The keys which are already retrieved
     keys_done: KeySpaceRandomAccum,
+
+    /// The keys covered by the image layers
+    keys_with_image_coverage: Option<Range<Key>>,
+
+    // Statistics that are still accessible as a caller of `get_vectored_impl`.
     layers_visited: u32,
+    delta_layers_visited: u32,
 }
 
 impl ValuesReconstructState {
@@ -126,7 +134,9 @@ impl ValuesReconstructState {
         Self {
             keys: HashMap::new(),
             keys_done: KeySpaceRandomAccum::new(),
+            keys_with_image_coverage: None,
             layers_visited: 0,
+            delta_layers_visited: 0,
         }
     }
 
@@ -140,8 +150,17 @@ impl ValuesReconstructState {
         }
     }
 
-    pub(crate) fn on_layer_visited(&mut self) {
+    pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
         self.layers_visited += 1;
+        if let ReadableLayer::PersistentLayer(layer) = layer {
+            if layer.layer_desc().is_delta() {
+                self.delta_layers_visited += 1;
+            }
+        }
+    }
+
+    pub(crate) fn get_delta_layers_visited(&self) -> u32 {
+        self.delta_layers_visited
     }
 
     pub(crate) fn get_layers_visited(&self) -> u32 {
@@ -171,6 +190,16 @@ impl ValuesReconstructState {
         }
     }
 
+    /// On hitting image layer, we can mark all keys in this range as done, because
+    /// if the image layer does not contain a key, it is deleted/never added.
+    pub(crate) fn on_image_layer_visited(&mut self, key_range: &Range<Key>) {
+        let prev_val = self.keys_with_image_coverage.replace(key_range.clone());
+        assert_eq!(
+            prev_val, None,
+            "should consume the keyspace before the next iteration"
+        );
+    }
+
     /// Update the state collected for a given key.
     /// Returns true if this was the last value needed for the key and false otherwise.
     ///
@@ -233,8 +262,12 @@ impl ValuesReconstructState {
 
     /// 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()
+    /// Returns individual keys done, and the image layer coverage.
+    pub(crate) fn consume_done_keys(&mut self) -> (KeySpace, Option<Range<Key>>) {
+        (
+            self.keys_done.consume_keyspace(),
+            self.keys_with_image_coverage.take(),
+        )
     }
 }
 

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -47,7 +47,7 @@ use hex;
 use itertools::Itertools;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::LayerAccessKind;
-use pageserver_api::shard::TenantShardId;
+use pageserver_api::shard::{ShardIdentity, TenantShardId};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs::File;
@@ -158,6 +158,7 @@ pub struct ImageLayerInner {
     index_start_blk: u32,
     index_root_blk: u32,
 
+    key_range: Range<Key>,
     lsn: Lsn,
 
     file: VirtualFile,
@@ -419,6 +420,7 @@ impl ImageLayerInner {
             file,
             file_id,
             max_vectored_read_bytes,
+            key_range: actual_summary.key_range,
         }))
     }
 
@@ -471,19 +473,27 @@ impl ImageLayerInner {
         ctx: &RequestContext,
     ) -> Result<(), GetVectoredError> {
         let reads = self
-            .plan_reads(keyspace, ctx)
+            .plan_reads(keyspace, None, ctx)
             .await
             .map_err(GetVectoredError::Other)?;
 
         self.do_reads_and_update_state(reads, reconstruct_state, ctx)
             .await;
 
+        reconstruct_state.on_image_layer_visited(&self.key_range);
+
         Ok(())
     }
 
+    /// Traverse the layer's index to build read operations on the overlap of the input keyspace
+    /// and the keys in this layer.
+    ///
+    /// If shard_identity is provided, it will be used to filter keys down to those stored on
+    /// this shard.
     async fn plan_reads(
         &self,
         keyspace: KeySpace,
+        shard_identity: Option<&ShardIdentity>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Vec<VectoredRead>> {
         let mut planner = VectoredReadPlanner::new(
@@ -503,7 +513,6 @@ impl ImageLayerInner {
 
         for range in keyspace.ranges.iter() {
             let mut range_end_handled = false;
-
             let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
             range.start.write_to_byte_slice(&mut search_key);
 
@@ -516,12 +525,22 @@ impl ImageLayerInner {
                 let key = Key::from_slice(&raw_key[..KEY_SIZE]);
                 assert!(key >= range.start);
 
+                let flag = if let Some(shard_identity) = shard_identity {
+                    if shard_identity.is_key_disposable(&key) {
+                        BlobFlag::Ignore
+                    } else {
+                        BlobFlag::None
+                    }
+                } else {
+                    BlobFlag::None
+                };
+
                 if key >= range.end {
                     planner.handle_range_end(offset);
                     range_end_handled = true;
                     break;
                 } else {
-                    planner.handle(key, self.lsn, offset, BlobFlag::None);
+                    planner.handle(key, self.lsn, offset, flag);
                 }
             }
 
@@ -534,6 +553,50 @@ impl ImageLayerInner {
         Ok(planner.finish())
     }
 
+    /// Given a key range, select the parts of that range that should be retained by the ShardIdentity,
+    /// then execute vectored GET operations, passing the results of all read keys into the writer.
+    pub(super) async fn filter(
+        &self,
+        shard_identity: &ShardIdentity,
+        writer: &mut ImageLayerWriter,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<usize> {
+        // Fragment the range into the regions owned by this ShardIdentity
+        let plan = self
+            .plan_reads(
+                KeySpace {
+                    // If asked for the total key space, plan_reads will give us all the keys in the layer
+                    ranges: vec![Key::MIN..Key::MAX],
+                },
+                Some(shard_identity),
+                ctx,
+            )
+            .await?;
+
+        let vectored_blob_reader = VectoredBlobReader::new(&self.file);
+        let mut key_count = 0;
+        for read in plan.into_iter() {
+            let buf_size = read.size();
+
+            let buf = BytesMut::with_capacity(buf_size);
+            let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
+
+            let frozen_buf = blobs_buf.buf.freeze();
+
+            for meta in blobs_buf.blobs.iter() {
+                let img_buf = frozen_buf.slice(meta.start..meta.end);
+
+                key_count += 1;
+                writer
+                    .put_image(meta.meta.key, img_buf, ctx)
+                    .await
+                    .context(format!("Storing key {}", meta.meta.key))?;
+            }
+        }
+
+        Ok(key_count)
+    }
+
     async fn do_reads_and_update_state(
         &self,
         reads: Vec<VectoredRead>,
@@ -646,7 +709,7 @@ impl ImageLayerWriterInner {
                 lsn,
             },
         );
-        info!("new image layer {path}");
+        trace!("creating image layer {}", path);
         let mut file = {
             VirtualFile::open_with_options(
                 &path,
@@ -766,7 +829,7 @@ impl ImageLayerWriterInner {
         // FIXME: why not carry the virtualfile here, it supports renaming?
         let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
 
-        trace!("created image layer {}", layer.local_path());
+        info!("created image layer {}", layer.local_path());
 
         Ok(layer)
     }
@@ -851,3 +914,136 @@ impl Drop for ImageLayerWriter {
         }
     }
 }
+
+#[cfg(test)]
+mod test {
+    use bytes::Bytes;
+    use pageserver_api::{
+        key::Key,
+        shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize},
+    };
+    use utils::{id::TimelineId, lsn::Lsn};
+
+    use crate::{tenant::harness::TenantHarness, DEFAULT_PG_VERSION};
+
+    use super::ImageLayerWriter;
+
+    #[tokio::test]
+    async fn image_layer_rewrite() {
+        let harness = TenantHarness::create("test_image_layer_rewrite").unwrap();
+        let (tenant, ctx) = harness.load().await;
+
+        // The LSN at which we will create an image layer to filter
+        let lsn = Lsn(0xdeadbeef0000);
+
+        let timeline_id = TimelineId::generate();
+        let timeline = tenant
+            .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
+            .await
+            .unwrap();
+
+        // This key range contains several 0x8000 page stripes, only one of which belongs to shard zero
+        let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
+        let range = input_start..input_end;
+
+        // Build an image layer to filter
+        let resident = {
+            let mut writer = ImageLayerWriter::new(
+                harness.conf,
+                timeline_id,
+                harness.tenant_shard_id,
+                &range,
+                lsn,
+                &ctx,
+            )
+            .await
+            .unwrap();
+
+            let foo_img = Bytes::from_static(&[1, 2, 3, 4]);
+            let mut key = range.start;
+            while key < range.end {
+                writer.put_image(key, foo_img.clone(), &ctx).await.unwrap();
+
+                key = key.next();
+            }
+            writer.finish(&timeline, &ctx).await.unwrap()
+        };
+        let original_size = resident.metadata().file_size;
+
+        // Filter for various shards: this exercises cases like values at start of key range, end of key
+        // range, middle of key range.
+        for shard_number in 0..4 {
+            let mut filtered_writer = ImageLayerWriter::new(
+                harness.conf,
+                timeline_id,
+                harness.tenant_shard_id,
+                &range,
+                lsn,
+                &ctx,
+            )
+            .await
+            .unwrap();
+
+            // TenantHarness gave us an unsharded tenant, but we'll use a sharded ShardIdentity
+            // to exercise filter()
+            let shard_identity = ShardIdentity::new(
+                ShardNumber(shard_number),
+                ShardCount::new(4),
+                ShardStripeSize(0x8000),
+            )
+            .unwrap();
+
+            let wrote_keys = resident
+                .filter(&shard_identity, &mut filtered_writer, &ctx)
+                .await
+                .unwrap();
+            let replacement = if wrote_keys > 0 {
+                Some(filtered_writer.finish(&timeline, &ctx).await.unwrap())
+            } else {
+                None
+            };
+
+            // This exact size and those below will need updating as/when the layer encoding changes, but
+            // should be deterministic for a given version of the format, as we used no randomness generating the input.
+            assert_eq!(original_size, 1597440);
+
+            match shard_number {
+                0 => {
+                    // We should have written out just one stripe for our shard identity
+                    assert_eq!(wrote_keys, 0x8000);
+                    let replacement = replacement.unwrap();
+
+                    // We should have dropped some of the data
+                    assert!(replacement.metadata().file_size < original_size);
+                    assert!(replacement.metadata().file_size > 0);
+
+                    // Assert that we dropped ~3/4 of the data.
+                    assert_eq!(replacement.metadata().file_size, 417792);
+                }
+                1 => {
+                    // Shard 1 has no keys in our input range
+                    assert_eq!(wrote_keys, 0x0);
+                    assert!(replacement.is_none());
+                }
+                2 => {
+                    // Shard 2 has one stripes in the input range
+                    assert_eq!(wrote_keys, 0x8000);
+                    let replacement = replacement.unwrap();
+                    assert!(replacement.metadata().file_size < original_size);
+                    assert!(replacement.metadata().file_size > 0);
+                    assert_eq!(replacement.metadata().file_size, 417792);
+                }
+                3 => {
+                    // Shard 3 has two stripes in the input range
+                    assert_eq!(wrote_keys, 0x10000);
+                    let replacement = replacement.unwrap();
+                    assert!(replacement.metadata().file_size < original_size);
+                    assert!(replacement.metadata().file_size > 0);
+                    assert_eq!(replacement.metadata().file_size, 811008);
+                }
+                _ => unreachable!(),
+            }
+        }
+    }
+}

pageserver/src/tenant/storage_layer/layer.rs

@@ -4,7 +4,7 @@ use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::{
     HistoricLayerInfo, LayerAccessKind, LayerResidenceEventReason, LayerResidenceStatus,
 };
-use pageserver_api::shard::{ShardIndex, TenantShardId};
+use pageserver_api::shard::{ShardIdentity, ShardIndex, TenantShardId};
 use std::ops::Range;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use std::sync::{Arc, Weak};
@@ -12,7 +12,7 @@ use std::time::{Duration, SystemTime};
 use tracing::Instrument;
 use utils::id::TimelineId;
 use utils::lsn::Lsn;
-use utils::sync::heavier_once_cell;
+use utils::sync::{gate, heavier_once_cell};
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
@@ -23,10 +23,10 @@ use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::{remote_timeline_client::LayerFileMetadata, Timeline};
 
 use super::delta_layer::{self, DeltaEntry};
-use super::image_layer;
+use super::image_layer::{self};
 use super::{
-    AsLayerDesc, LayerAccessStats, LayerAccessStatsReset, LayerName, PersistentLayerDesc,
-    ValueReconstructResult, ValueReconstructState, ValuesReconstructState,
+    AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
+    PersistentLayerDesc, ValueReconstructResult, ValueReconstructState, ValuesReconstructState,
 };
 
 use utils::generation::Generation;
@@ -129,19 +129,16 @@ pub(crate) fn local_layer_path(
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     layer_file_name: &LayerName,
-    _generation: &Generation,
+    generation: &Generation,
 ) -> Utf8PathBuf {
     let timeline_path = conf.timeline_path(tenant_shard_id, timeline_id);
 
-    timeline_path.join(layer_file_name.to_string())
-
-    // TODO: switch to enabling new-style layer paths after next release
-    // if generation.is_none() {
-    //     // Without a generation, we may only use legacy path style
-    //     timeline_path.join(layer_file_name.to_string())
-    // } else {
-    //     timeline_path.join(format!("{}-v1{}", layer_file_name, generation.get_suffix()))
-    // }
+    if generation.is_none() {
+        // Without a generation, we may only use legacy path style
+        timeline_path.join(layer_file_name.to_string())
+    } else {
+        timeline_path.join(format!("{}-v1{}", layer_file_name, generation.get_suffix()))
+    }
 }
 
 impl Layer {
@@ -164,7 +161,7 @@ impl Layer {
             timeline.tenant_shard_id,
             timeline.timeline_id,
             file_name,
-            metadata.file_size(),
+            metadata.file_size,
         );
 
         let access_stats = LayerAccessStats::for_loading_layer(LayerResidenceStatus::Evicted);
@@ -197,7 +194,7 @@ impl Layer {
             timeline.tenant_shard_id,
             timeline.timeline_id,
             file_name,
-            metadata.file_size(),
+            metadata.file_size,
         );
 
         let access_stats = LayerAccessStats::for_loading_layer(LayerResidenceStatus::Resident);
@@ -230,7 +227,7 @@ impl Layer {
 
         timeline
             .metrics
-            .resident_physical_size_add(metadata.file_size());
+            .resident_physical_size_add(metadata.file_size);
 
         ResidentLayer { downloaded, owner }
     }
@@ -588,9 +585,6 @@ struct LayerInner {
     /// [`Timeline::gate`] at the same time.
     timeline: Weak<Timeline>,
 
-    /// Cached knowledge of [`Timeline::remote_client`] being `Some`.
-    have_remote_client: bool,
-
     access_stats: LayerAccessStats,
 
     /// This custom OnceCell is backed by std mutex, but only held for short time periods.
@@ -735,23 +729,23 @@ impl Drop for LayerInner {
             if removed {
                 timeline.metrics.resident_physical_size_sub(file_size);
             }
-            if let Some(remote_client) = timeline.remote_client.as_ref() {
-                let res = remote_client.schedule_deletion_of_unlinked(vec![(file_name, meta)]);
+            let res = timeline
+                .remote_client
+                .schedule_deletion_of_unlinked(vec![(file_name, meta)]);
 
-                if let Err(e) = res {
-                    // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
-                    // demonstrating this deadlock (without spawn_blocking): stop will drop
-                    // queued items, which will have ResidentLayer's, and those drops would try
-                    // to re-entrantly lock the RemoteTimelineClient inner state.
-                    if !timeline.is_active() {
-                        tracing::info!("scheduling deletion on drop failed: {e:#}");
-                    } else {
-                        tracing::warn!("scheduling deletion on drop failed: {e:#}");
-                    }
-                    LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
+            if let Err(e) = res {
+                // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
+                // demonstrating this deadlock (without spawn_blocking): stop will drop
+                // queued items, which will have ResidentLayer's, and those drops would try
+                // to re-entrantly lock the RemoteTimelineClient inner state.
+                if !timeline.is_active() {
+                    tracing::info!("scheduling deletion on drop failed: {e:#}");
                 } else {
-                    LAYER_IMPL_METRICS.inc_completed_deletes();
+                    tracing::warn!("scheduling deletion on drop failed: {e:#}");
                 }
+                LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
+            } else {
+                LAYER_IMPL_METRICS.inc_completed_deletes();
             }
         });
     }
@@ -789,7 +783,6 @@ impl LayerInner {
             path: local_path,
             desc,
             timeline: Arc::downgrade(timeline),
-            have_remote_client: timeline.remote_client.is_some(),
             access_stats,
             wanted_deleted: AtomicBool::new(false),
             inner,
@@ -818,8 +811,6 @@ impl LayerInner {
     /// in a new attempt to evict OR join the previously started attempt.
     #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, ret, err(level = tracing::Level::DEBUG), fields(layer=%self))]
     pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
-        assert!(self.have_remote_client);
-
         let mut rx = self.status.as_ref().unwrap().subscribe();
 
         {
@@ -976,10 +967,6 @@ impl LayerInner {
             return Err(DownloadError::NotFile(ft));
         }
 
-        if timeline.remote_client.as_ref().is_none() {
-            return Err(DownloadError::NoRemoteStorage);
-        }
-
         if let Some(ctx) = ctx {
             self.check_expected_download(ctx)?;
         }
@@ -1116,15 +1103,12 @@ impl LayerInner {
         permit: heavier_once_cell::InitPermit,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<DownloadedLayer>> {
-        let client = timeline
+        let result = timeline
             .remote_client
-            .as_ref()
-            .expect("checked before download_init_and_wait");
-
-        let result = client
             .download_layer_file(
                 &self.desc.layer_name(),
                 &self.metadata(),
+                &self.path,
                 &timeline.cancel,
                 ctx,
             )
@@ -1280,6 +1264,7 @@ impl LayerInner {
                 lsn_end: lsn_range.end,
                 remote: !resident,
                 access_stats,
+                l0: crate::tenant::layer_map::LayerMap::is_l0(self.layer_desc()),
             }
         } else {
             let lsn = self.desc.image_layer_lsn();
@@ -1296,20 +1281,10 @@ impl LayerInner {
 
     /// `DownloadedLayer` is being dropped, so it calls this method.
     fn on_downloaded_layer_drop(self: Arc<LayerInner>, only_version: usize) {
-        let can_evict = self.have_remote_client;
-
         // we cannot know without inspecting LayerInner::inner if we should evict or not, even
         // though here it is very likely
         let span = tracing::info_span!(parent: None, "layer_evict", tenant_id = %self.desc.tenant_shard_id.tenant_id, shard_id = %self.desc.tenant_shard_id.shard_slug(), timeline_id = %self.desc.timeline_id, layer=%self, version=%only_version);
 
-        if !can_evict {
-            // it would be nice to assert this case out, but we are in drop
-            span.in_scope(|| {
-                tracing::error!("bug in struct Layer: ResidentOrWantedEvicted has been downgraded while we have no remote storage");
-            });
-            return;
-        }
-
         // NOTE: this scope *must* never call `self.inner.get` because evict_and_wait might
         // drop while the `self.inner` is being locked, leading to a deadlock.
 
@@ -1358,7 +1333,7 @@ impl LayerInner {
 
         is_good_to_continue(&rx.borrow_and_update())?;
 
-        let Ok(_gate) = timeline.gate.enter() else {
+        let Ok(gate) = timeline.gate.enter() else {
             return Err(EvictionCancelled::TimelineGone);
         };
 
@@ -1446,7 +1421,7 @@ impl LayerInner {
         Self::spawn_blocking(move || {
             let _span = span.entered();
 
-            let res = self.evict_blocking(&timeline, &permit);
+            let res = self.evict_blocking(&timeline, &gate, &permit);
 
             let waiters = self.inner.initializer_count();
 
@@ -1472,6 +1447,7 @@ impl LayerInner {
     fn evict_blocking(
         &self,
         timeline: &Timeline,
+        _gate: &gate::GateGuard,
         _permit: &heavier_once_cell::InitPermit,
     ) -> Result<(), EvictionCancelled> {
         // now accesses to `self.inner.get_or_init*` wait on the semaphore or the `_permit`
@@ -1581,8 +1557,6 @@ pub(crate) enum EvictionError {
 pub(crate) enum DownloadError {
     #[error("timeline has already shutdown")]
     TimelineShutdown,
-    #[error("no remote storage configured")]
-    NoRemoteStorage,
     #[error("context denies downloading")]
     ContextAndConfigReallyDeniesDownloads,
     #[error("downloading is really required but not allowed by this method")]
@@ -1828,16 +1802,15 @@ impl ResidentLayer {
         use LayerKind::*;
 
         let owner = &self.owner.0;
-
         match self.downloaded.get(owner, ctx).await? {
             Delta(ref d) => {
+                // this is valid because the DownloadedLayer::kind is a OnceCell, not a
+                // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
+                // while it's being held.
                 owner
                     .access_stats
                     .record_access(LayerAccessKind::KeyIter, ctx);
 
-                // this is valid because the DownloadedLayer::kind is a OnceCell, not a
-                // Mutex<OnceCell>, so we cannot go and deinitialize the value with OnceCell::take
-                // while it's being held.
                 delta_layer::DeltaLayerInner::load_keys(d, ctx)
                     .await
                     .with_context(|| format!("Layer index is corrupted for {self}"))
@@ -1846,6 +1819,23 @@ impl ResidentLayer {
         }
     }
 
+    /// Read all they keys in this layer which match the ShardIdentity, and write them all to
+    /// the provided writer.  Return the number of keys written.
+    #[tracing::instrument(level = tracing::Level::DEBUG, skip_all, fields(layer=%self))]
+    pub(crate) async fn filter<'a>(
+        &'a self,
+        shard_identity: &ShardIdentity,
+        writer: &mut ImageLayerWriter,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<usize> {
+        use LayerKind::*;
+
+        match self.downloaded.get(&self.owner.0, ctx).await? {
+            Delta(_) => anyhow::bail!(format!("cannot filter() on a delta layer {self}")),
+            Image(i) => i.filter(shard_identity, writer, ctx).await,
+        }
+    }
+
     /// Returns the amount of keys and values written to the writer.
     pub(crate) async fn copy_delta_prefix(
         &self,

pageserver/src/tenant/storage_layer/layer/tests.rs

@@ -145,7 +145,7 @@ async fn smoke_test() {
         .await
         .expect("the local layer file still exists");
 
-    let rtc = timeline.remote_client.as_ref().unwrap();
+    let rtc = &timeline.remote_client;
 
     {
         let layers = &[layer];
@@ -761,13 +761,7 @@ async fn eviction_cancellation_on_drop() {
     timeline.freeze_and_flush().await.unwrap();
 
     // wait for the upload to complete so our Arc::strong_count assertion holds
-    timeline
-        .remote_client
-        .as_ref()
-        .unwrap()
-        .wait_completion()
-        .await
-        .unwrap();
+    timeline.remote_client.wait_completion().await.unwrap();
 
     let (evicted_layer, not_evicted) = {
         let mut layers = {

pageserver/src/tenant/storage_layer/layer_name.rs

@@ -347,37 +347,33 @@ impl<'de> serde::de::Visitor<'de> for LayerNameVisitor {
 mod test {
     use super::*;
     #[test]
-    fn image_layer_parse() -> anyhow::Result<()> {
+    fn image_layer_parse() {
         let expected = LayerName::Image(ImageLayerName {
             key_range: Key::from_i128(0)
                 ..Key::from_hex("000000067F00000001000004DF0000000006").unwrap(),
             lsn: Lsn::from_hex("00000000014FED58").unwrap(),
         });
-        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-v1-00000001").map_err(|s| anyhow::anyhow!(s))?;
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-v1-00000001").unwrap();
         assert_eq!(parsed, expected,);
 
         // Omitting generation suffix is valid
-        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58").map_err(|s| anyhow::anyhow!(s))?;
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58").unwrap();
         assert_eq!(parsed, expected,);
-
-        Ok(())
     }
 
     #[test]
-    fn delta_layer_parse() -> anyhow::Result<()> {
+    fn delta_layer_parse() {
         let expected = LayerName::Delta(DeltaLayerName {
             key_range: Key::from_i128(0)
                 ..Key::from_hex("000000067F00000001000004DF0000000006").unwrap(),
             lsn_range: Lsn::from_hex("00000000014FED58").unwrap()
                 ..Lsn::from_hex("000000000154C481").unwrap(),
         });
-        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481-v1-00000001").map_err(|s| anyhow::anyhow!(s))?;
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481-v1-00000001").unwrap();
         assert_eq!(parsed, expected);
 
         // Omitting generation suffix is valid
-        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481").map_err(|s| anyhow::anyhow!(s))?;
+        let parsed = LayerName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481").unwrap();
         assert_eq!(parsed, expected);
-
-        Ok(())
     }
 }

pageserver/src/tenant/tasks.rs

@@ -17,7 +17,7 @@ use crate::tenant::{Tenant, TenantState};
 use rand::Rng;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::{backoff, completion};
+use utils::{backoff, completion, pausable_failpoint};
 
 static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
     once_cell::sync::Lazy::new(|| {

pageserver/src/tenant/timeline/compaction.rs

@@ -9,7 +9,10 @@ use std::ops::{Deref, Range};
 use std::sync::Arc;
 
 use super::layer_manager::LayerManager;
-use super::{CompactFlags, DurationRecorder, ImageLayerCreationMode, RecordedDuration, Timeline};
+use super::{
+    CompactFlags, CreateImageLayersError, DurationRecorder, ImageLayerCreationMode,
+    RecordedDuration, Timeline,
+};
 
 use anyhow::{anyhow, Context};
 use enumset::EnumSet;
@@ -22,14 +25,13 @@ use tracing::{debug, info, info_span, trace, warn, Instrument};
 use utils::id::TimelineId;
 
 use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
+use crate::page_cache;
 use crate::tenant::storage_layer::{AsLayerDesc, PersistentLayerDesc};
-use crate::tenant::timeline::{drop_rlock, is_rel_fsm_block_key, is_rel_vm_block_key, Hole};
+use crate::tenant::timeline::{drop_rlock, Hole, ImageLayerCreationOutcome};
 use crate::tenant::timeline::{DeltaLayerWriter, ImageLayerWriter};
 use crate::tenant::timeline::{Layer, ResidentLayer};
 use crate::tenant::DeltaLayer;
-use crate::tenant::PageReconstructError;
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
-use crate::{page_cache, ZERO_PAGE};
 
 use crate::keyspace::KeySpace;
 use crate::repository::Key;
@@ -116,9 +118,13 @@ impl Timeline {
 
                 // 3. Create new image layers for partitions that have been modified
                 // "enough".
-                let dense_layers = self
+                let mut partitioning = dense_partitioning;
+                partitioning
+                    .parts
+                    .extend(sparse_partitioning.into_dense().parts);
+                let image_layers = self
                     .create_image_layers(
-                        &dense_partitioning,
+                        &partitioning,
                         lsn,
                         if flags.contains(CompactFlags::ForceImageLayerCreation) {
                             ImageLayerCreationMode::Force
@@ -130,24 +136,8 @@ impl Timeline {
                     .await
                     .map_err(anyhow::Error::from)?;
 
-                // For now, nothing will be produced...
-                let sparse_layers = self
-                    .create_image_layers(
-                        &sparse_partitioning.clone().into_dense(),
-                        lsn,
-                        if flags.contains(CompactFlags::ForceImageLayerCreation) {
-                            ImageLayerCreationMode::Force
-                        } else {
-                            ImageLayerCreationMode::Try
-                        },
-                        &image_ctx,
-                    )
-                    .await
-                    .map_err(anyhow::Error::from)?;
-                assert!(sparse_layers.is_empty());
-
-                self.upload_new_image_layers(dense_layers)?;
-                dense_partitioning.parts.len()
+                self.upload_new_image_layers(image_layers)?;
+                partitioning.parts.len()
             }
             Err(err) => {
                 // no partitioning? This is normal, if the timeline was just created
@@ -186,13 +176,24 @@ impl Timeline {
     async fn compact_shard_ancestors(
         self: &Arc<Self>,
         rewrite_max: usize,
-        _ctx: &RequestContext,
+        ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         let mut drop_layers = Vec::new();
-        let layers_to_rewrite: Vec<Layer> = Vec::new();
+        let mut layers_to_rewrite: Vec<Layer> = Vec::new();
 
-        // We will use the PITR cutoff as a condition for rewriting layers.
-        let pitr_cutoff = self.gc_info.read().unwrap().cutoffs.pitr;
+        // We will use the Lsn cutoff of the last GC as a threshold for rewriting layers: if a
+        // layer is behind this Lsn, it indicates that the layer is being retained beyond the
+        // pitr_interval, for example because a branchpoint references it.
+        //
+        // Holding this read guard also blocks [`Self::gc_timeline`] from entering while we
+        // are rewriting layers.
+        let latest_gc_cutoff = self.get_latest_gc_cutoff_lsn();
+
+        tracing::info!(
+            "latest_gc_cutoff: {}, pitr cutoff {}",
+            *latest_gc_cutoff,
+            self.gc_info.read().unwrap().cutoffs.pitr
+        );
 
         let layers = self.layers.read().await;
         for layer_desc in layers.layer_map().iter_historic_layers() {
@@ -251,9 +252,9 @@ impl Timeline {
 
             // Don't bother re-writing a layer if it is within the PITR window: it will age-out eventually
             // without incurring the I/O cost of a rewrite.
-            if layer_desc.get_lsn_range().end >= pitr_cutoff {
-                debug!(%layer, "Skipping rewrite of layer still in PITR window ({} >= {})",
-                    layer_desc.get_lsn_range().end, pitr_cutoff);
+            if layer_desc.get_lsn_range().end >= *latest_gc_cutoff {
+                debug!(%layer, "Skipping rewrite of layer still in GC window ({} >= {})",
+                    layer_desc.get_lsn_range().end, *latest_gc_cutoff);
                 continue;
             }
 
@@ -263,13 +264,10 @@ impl Timeline {
                 continue;
             }
 
-            // Only rewrite layers if they would have different remote paths: either they belong to this
-            // shard but an old generation, or they belonged to another shard.  This also implicitly
-            // guarantees that the layer is persistent in remote storage (as only remote persistent
-            // layers are carried across shard splits, any local-only layer would be in the current generation)
-            if layer.metadata().generation == self.generation
-                && layer.metadata().shard.shard_count == self.shard_identity.count
-            {
+            // Only rewrite layers if their generations differ.  This guarantees:
+            //  - that local rewrite is safe, as local layer paths will differ between existing layer and rewritten one
+            //  - that the layer is persistent in remote storage, as we only see old-generation'd layer via loading from remote storage
+            if layer.metadata().generation == self.generation {
                 debug!(%layer, "Skipping rewrite, is not from old generation");
                 continue;
             }
@@ -282,26 +280,77 @@ impl Timeline {
             }
 
             // Fall through: all our conditions for doing a rewrite passed.
-            // TODO: implement rewriting
-            tracing::debug!(%layer, "Would rewrite layer");
+            layers_to_rewrite.push(layer);
         }
 
-        // Drop the layers read lock: we will acquire it for write in [`Self::rewrite_layers`]
+        // Drop read lock on layer map before we start doing time-consuming I/O
         drop(layers);
 
-        // TODO: collect layers to rewrite
-        let replace_layers = Vec::new();
+        let mut replace_image_layers = Vec::new();
+
+        for layer in layers_to_rewrite {
+            tracing::info!(layer=%layer, "Rewriting layer after shard split...");
+            let mut image_layer_writer = ImageLayerWriter::new(
+                self.conf,
+                self.timeline_id,
+                self.tenant_shard_id,
+                &layer.layer_desc().key_range,
+                layer.layer_desc().image_layer_lsn(),
+                ctx,
+            )
+            .await?;
+
+            // Safety of layer rewrites:
+            // - We are writing to a different local file path than we are reading from, so the old Layer
+            //   cannot interfere with the new one.
+            // - In the page cache, contents for a particular VirtualFile are stored with a file_id that
+            //   is different for two layers with the same name (in `ImageLayerInner::new` we always
+            //   acquire a fresh id from [`crate::page_cache::next_file_id`].  So readers do not risk
+            //   reading the index from one layer file, and then data blocks from the rewritten layer file.
+            // - Any readers that have a reference to the old layer will keep it alive until they are done
+            //   with it. If they are trying to promote from remote storage, that will fail, but this is the same
+            //   as for compaction generally: compaction is allowed to delete layers that readers might be trying to use.
+            // - We do not run concurrently with other kinds of compaction, so the only layer map writes we race with are:
+            //    - GC, which at worst witnesses us "undelete" a layer that they just deleted.
+            //    - ingestion, which only inserts layers, therefore cannot collide with us.
+            let resident = layer.download_and_keep_resident().await?;
+
+            let keys_written = resident
+                .filter(&self.shard_identity, &mut image_layer_writer, ctx)
+                .await?;
+
+            if keys_written > 0 {
+                let new_layer = image_layer_writer.finish(self, ctx).await?;
+                tracing::info!(layer=%new_layer, "Rewrote layer, {} -> {} bytes",
+                    layer.metadata().file_size,
+                    new_layer.metadata().file_size);
+
+                replace_image_layers.push((layer, new_layer));
+            } else {
+                // Drop the old layer.  Usually for this case we would already have noticed that
+                // the layer has no data for us with the ShardedRange check above, but
+                drop_layers.push(layer);
+            }
+        }
+
+        // At this point, we have replaced local layer files with their rewritten form, but not yet uploaded
+        // metadata to reflect that. If we restart here, the replaced layer files will look invalid (size mismatch
+        // to remote index) and be removed. This is inefficient but safe.
+        fail::fail_point!("compact-shard-ancestors-localonly");
 
         // Update the LayerMap so that readers will use the new layers, and enqueue it for writing to remote storage
-        self.rewrite_layers(replace_layers, drop_layers).await?;
+        self.rewrite_layers(replace_image_layers, drop_layers)
+            .await?;
 
-        if let Some(remote_client) = self.remote_client.as_ref() {
-            // We wait for all uploads to complete before finishing this compaction stage.  This is not
-            // necessary for correctness, but it simplifies testing, and avoids proceeding with another
-            // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
-            // load.
-            remote_client.wait_completion().await?;
-        }
+        fail::fail_point!("compact-shard-ancestors-enqueued");
+
+        // We wait for all uploads to complete before finishing this compaction stage.  This is not
+        // necessary for correctness, but it simplifies testing, and avoids proceeding with another
+        // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
+        // load.
+        self.remote_client.wait_completion().await?;
+
+        fail::fail_point!("compact-shard-ancestors-persistent");
 
         Ok(())
     }
@@ -501,8 +550,11 @@ impl Timeline {
 
         for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
             if let Some(prev_key) = prev {
-                // just first fast filter
-                if next_key.to_i128() - prev_key.to_i128() >= min_hole_range {
+                // just first fast filter, do not create hole entries for metadata keys. The last hole in the
+                // compaction is the gap between data key and metadata keys.
+                if next_key.to_i128() - prev_key.to_i128() >= min_hole_range
+                    && !Key::is_metadata_key(&prev_key)
+                {
                     let key_range = prev_key..next_key;
                     // Measuring hole by just subtraction of i128 representation of key range boundaries
                     // has not so much sense, because largest holes will corresponds field1/field2 changes.
@@ -1161,10 +1213,10 @@ impl TimelineAdaptor {
         lsn: Lsn,
         key_range: &Range<Key>,
         ctx: &RequestContext,
-    ) -> Result<(), PageReconstructError> {
+    ) -> Result<(), CreateImageLayersError> {
         let timer = self.timeline.metrics.create_images_time_histo.start_timer();
 
-        let mut image_layer_writer = ImageLayerWriter::new(
+        let image_layer_writer = ImageLayerWriter::new(
             self.timeline.conf,
             self.timeline.timeline_id,
             self.timeline.tenant_shard_id,
@@ -1175,47 +1227,34 @@ impl TimelineAdaptor {
         .await?;
 
         fail_point!("image-layer-writer-fail-before-finish", |_| {
-            Err(PageReconstructError::Other(anyhow::anyhow!(
+            Err(CreateImageLayersError::Other(anyhow::anyhow!(
                 "failpoint image-layer-writer-fail-before-finish"
             )))
         });
-        let keyspace_ranges = self.get_keyspace(key_range, lsn, ctx).await?;
-        for range in &keyspace_ranges {
-            let mut key = range.start;
-            while key < range.end {
-                let img = match self.timeline.get(key, lsn, ctx).await {
-                    Ok(img) => img,
-                    Err(err) => {
-                        // If we fail to reconstruct a VM or FSM page, we can zero the
-                        // page without losing any actual user data. That seems better
-                        // than failing repeatedly and getting stuck.
-                        //
-                        // We had a bug at one point, where we truncated the FSM and VM
-                        // in the pageserver, but the Postgres didn't know about that
-                        // and continued to generate incremental WAL records for pages
-                        // that didn't exist in the pageserver. Trying to replay those
-                        // WAL records failed to find the previous image of the page.
-                        // This special case allows us to recover from that situation.
-                        // See https://github.com/neondatabase/neon/issues/2601.
-                        //
-                        // Unfortunately we cannot do this for the main fork, or for
-                        // any metadata keys, keys, as that would lead to actual data
-                        // loss.
-                        if is_rel_fsm_block_key(key) || is_rel_vm_block_key(key) {
-                            warn!("could not reconstruct FSM or VM key {key}, filling with zeros: {err:?}");
-                            ZERO_PAGE.clone()
-                        } else {
-                            return Err(err);
-                        }
-                    }
-                };
-                image_layer_writer.put_image(key, img, ctx).await?;
-                key = key.next();
-            }
-        }
-        let image_layer = image_layer_writer.finish(&self.timeline, ctx).await?;
 
-        self.new_images.push(image_layer);
+        let keyspace = KeySpace {
+            ranges: self.get_keyspace(key_range, lsn, ctx).await?,
+        };
+        // TODO set proper (stateful) start. The create_image_layer_for_rel_blocks function mostly
+        let start = Key::MIN;
+        let ImageLayerCreationOutcome {
+            image,
+            next_start_key: _,
+        } = self
+            .timeline
+            .create_image_layer_for_rel_blocks(
+                &keyspace,
+                image_layer_writer,
+                lsn,
+                ctx,
+                key_range.clone(),
+                start,
+            )
+            .await?;
+
+        if let Some(image_layer) = image {
+            self.new_images.push(image_layer);
+        }
 
         timer.stop_and_record();
 

pageserver/src/tenant/timeline/delete.rs

@@ -7,7 +7,7 @@ use anyhow::Context;
 use pageserver_api::{models::TimelineState, shard::TenantShardId};
 use tokio::sync::OwnedMutexGuard;
 use tracing::{error, info, instrument, Instrument};
-use utils::{crashsafe, fs_ext, id::TimelineId};
+use utils::{crashsafe, fs_ext, id::TimelineId, pausable_failpoint};
 
 use crate::{
     config::PageServerConf,
@@ -26,19 +26,21 @@ use super::{Timeline, TimelineResources};
 /// during attach or pageserver restart.
 /// See comment in persist_index_part_with_deleted_flag.
 async fn set_deleted_in_remote_index(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
-    if let Some(remote_client) = timeline.remote_client.as_ref() {
-        match remote_client.persist_index_part_with_deleted_flag().await {
-            // If we (now, or already) marked it successfully as deleted, we can proceed
-            Ok(()) | Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(_)) => (),
-            // Bail out otherwise
-            //
-            // AlreadyInProgress shouldn't happen, because the 'delete_lock' prevents
-            // two tasks from performing the deletion at the same time. The first task
-            // that starts deletion should run it to completion.
-            Err(e @ PersistIndexPartWithDeletedFlagError::AlreadyInProgress(_))
-            | Err(e @ PersistIndexPartWithDeletedFlagError::Other(_)) => {
-                return Err(DeleteTimelineError::Other(anyhow::anyhow!(e)));
-            }
+    match timeline
+        .remote_client
+        .persist_index_part_with_deleted_flag()
+        .await
+    {
+        // If we (now, or already) marked it successfully as deleted, we can proceed
+        Ok(()) | Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(_)) => (),
+        // Bail out otherwise
+        //
+        // AlreadyInProgress shouldn't happen, because the 'delete_lock' prevents
+        // two tasks from performing the deletion at the same time. The first task
+        // that starts deletion should run it to completion.
+        Err(e @ PersistIndexPartWithDeletedFlagError::AlreadyInProgress(_))
+        | Err(e @ PersistIndexPartWithDeletedFlagError::Other(_)) => {
+            return Err(DeleteTimelineError::Other(anyhow::anyhow!(e)));
         }
     }
     Ok(())
@@ -117,11 +119,11 @@ pub(super) async fn delete_local_timeline_directory(
 
 /// Removes remote layers and an index file after them.
 async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<()> {
-    if let Some(remote_client) = &timeline.remote_client {
-        remote_client.delete_all().await.context("delete_all")?
-    };
-
-    Ok(())
+    timeline
+        .remote_client
+        .delete_all()
+        .await
+        .context("delete_all")
 }
 
 // This function removs remaining traces of a timeline on disk.
@@ -260,7 +262,7 @@ impl DeleteTimelineFlow {
         tenant: Arc<Tenant>,
         timeline_id: TimelineId,
         local_metadata: &TimelineMetadata,
-        remote_client: Option<RemoteTimelineClient>,
+        remote_client: RemoteTimelineClient,
         deletion_queue_client: DeletionQueueClient,
     ) -> anyhow::Result<()> {
         // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
@@ -278,6 +280,8 @@ impl DeleteTimelineFlow {
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.
                 CreateTimelineCause::Delete,
+                // Aux file policy is not needed for deletion, assuming deletion does not read aux keyspace
+                None,
             )
             .context("create_timeline_struct")?;
 

pageserver/src/tenant/timeline/detach_ancestor.rs

@@ -12,7 +12,7 @@ use crate::{
 };
 use tokio_util::sync::CancellationToken;
 use tracing::Instrument;
-use utils::{completion, generation::Generation, id::TimelineId, lsn::Lsn};
+use utils::{completion, generation::Generation, http::error::ApiError, id::TimelineId, lsn::Lsn};
 
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum Error {
@@ -41,6 +41,27 @@ pub(crate) enum Error {
     Unexpected(#[source] anyhow::Error),
 }
 
+impl From<Error> for ApiError {
+    fn from(value: Error) -> Self {
+        match value {
+            e @ Error::NoAncestor => ApiError::Conflict(e.to_string()),
+            // TODO: ApiError converts the anyhow using debug formatting ... just stop using ApiError?
+            e @ Error::TooManyAncestors => ApiError::BadRequest(anyhow::anyhow!("{}", e)),
+            Error::ShuttingDown => ApiError::ShuttingDown,
+            Error::OtherTimelineDetachOngoing(_) => {
+                ApiError::ResourceUnavailable("other timeline detach is already ongoing".into())
+            }
+            // All of these contain shutdown errors, in fact, it's the most common
+            e @ Error::FlushAncestor(_)
+            | e @ Error::RewrittenDeltaDownloadFailed(_)
+            | e @ Error::CopyDeltaPrefix(_)
+            | e @ Error::UploadRewritten(_)
+            | e @ Error::CopyFailed(_)
+            | e @ Error::Unexpected(_) => ApiError::InternalServerError(e.into()),
+        }
+    }
+}
+
 pub(crate) struct PreparedTimelineDetach {
     layers: Vec<Layer>,
 }
@@ -56,7 +77,7 @@ impl Default for Options {
     fn default() -> Self {
         Self {
             rewrite_concurrency: std::num::NonZeroUsize::new(2).unwrap(),
-            copy_concurrency: std::num::NonZeroUsize::new(10).unwrap(),
+            copy_concurrency: std::num::NonZeroUsize::new(100).unwrap(),
         }
     }
 }
@@ -70,15 +91,16 @@ pub(super) async fn prepare(
 ) -> Result<(completion::Completion, PreparedTimelineDetach), Error> {
     use Error::*;
 
-    if detached.remote_client.as_ref().is_none() {
-        unimplemented!("no new code for running without remote storage");
-    }
-
     let Some((ancestor, ancestor_lsn)) = detached
         .ancestor_timeline
         .as_ref()
         .map(|tl| (tl.clone(), detached.ancestor_lsn))
     else {
+        // TODO: check if we have already been detached; for this we need to read the stored data
+        // on remote client, for that we need a follow-up which makes uploads cheaper and maintains
+        // a projection of the commited data.
+        //
+        // the error is wrong per openapi
         return Err(NoAncestor);
     };
 
@@ -88,7 +110,7 @@ pub(super) async fn prepare(
 
     if ancestor.ancestor_timeline.is_some() {
         // non-technical requirement; we could flatten N ancestors just as easily but we chose
-        // not to
+        // not to, at least initially
         return Err(TooManyAncestors);
     }
 
@@ -315,8 +337,6 @@ async fn upload_rewritten_layer(
     // FIXME: better shuttingdown error
     target
         .remote_client
-        .as_ref()
-        .unwrap()
         .upload_layer_file(&copied, cancel)
         .await
         .map_err(UploadRewritten)?;
@@ -406,8 +426,6 @@ async fn remote_copy(
     // FIXME: better shuttingdown error
     adoptee
         .remote_client
-        .as_ref()
-        .unwrap()
         .copy_timeline_layer(adopted, &owned, cancel)
         .await
         .map(move |()| owned)
@@ -421,11 +439,6 @@ pub(super) async fn complete(
     prepared: PreparedTimelineDetach,
     _ctx: &RequestContext,
 ) -> Result<Vec<TimelineId>, anyhow::Error> {
-    let rtc = detached
-        .remote_client
-        .as_ref()
-        .expect("has to have a remote timeline client for timeline ancestor detach");
-
     let PreparedTimelineDetach { layers } = prepared;
 
     let ancestor = detached
@@ -442,11 +455,13 @@ pub(super) async fn complete(
     //
     // this is not perfect, but it avoids us a retry happening after a compaction or gc on restart
     // which could give us a completely wrong layer combination.
-    rtc.schedule_adding_existing_layers_to_index_detach_and_wait(
-        &layers,
-        (ancestor.timeline_id, ancestor_lsn),
-    )
-    .await?;
+    detached
+        .remote_client
+        .schedule_adding_existing_layers_to_index_detach_and_wait(
+            &layers,
+            (ancestor.timeline_id, ancestor_lsn),
+        )
+        .await?;
 
     let mut tasks = tokio::task::JoinSet::new();
 
@@ -491,8 +506,6 @@ pub(super) async fn complete(
                 async move {
                     let res = timeline
                         .remote_client
-                        .as_ref()
-                        .expect("reparented has to have remote client because detached has one")
                         .schedule_reparenting_and_wait(&new_parent)
                         .await;
 

pageserver/src/tenant/timeline/eviction_task.rs

@@ -23,7 +23,7 @@ use std::{
 use pageserver_api::models::{EvictionPolicy, EvictionPolicyLayerAccessThreshold};
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
-use tracing::{debug, error, info, info_span, instrument, warn, Instrument};
+use tracing::{debug, info, info_span, instrument, warn, Instrument};
 
 use crate::{
     context::{DownloadBehavior, RequestContext},
@@ -211,11 +211,6 @@ impl Timeline {
 
         // So, we just need to deal with this.
 
-        if self.remote_client.is_none() {
-            error!("no remote storage configured, cannot evict layers");
-            return ControlFlow::Continue(());
-        }
-
         let mut js = tokio::task::JoinSet::new();
         {
             let guard = self.layers.read().await;

pageserver/src/tenant/timeline/init.rs

@@ -7,30 +7,28 @@ use crate::{
             index::{IndexPart, LayerFileMetadata},
         },
         storage_layer::LayerName,
-        Generation,
     },
-    METADATA_FILE_NAME,
 };
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
-use pageserver_api::shard::ShardIndex;
-use std::{collections::HashMap, str::FromStr};
+use std::{
+    collections::{hash_map, HashMap},
+    str::FromStr,
+};
 use utils::lsn::Lsn;
 
 /// Identified files in the timeline directory.
 pub(super) enum Discovered {
     /// The only one we care about
-    Layer(LayerName, Utf8PathBuf, u64),
+    Layer(LayerName, LocalLayerFileMetadata),
     /// Old ephmeral files from previous launches, should be removed
     Ephemeral(String),
     /// Old temporary timeline files, unsure what these really are, should be removed
     Temporary(String),
     /// Temporary on-demand download files, should be removed
     TemporaryDownload(String),
-    /// "metadata" file we persist locally and include in `index_part.json`
-    Metadata,
     /// Backup file from previously future layers
-    IgnoredBackup,
+    IgnoredBackup(Utf8PathBuf),
     /// Unrecognized, warn about these
     Unknown(String),
 }
@@ -46,14 +44,15 @@ pub(super) fn scan_timeline_dir(path: &Utf8Path) -> anyhow::Result<Vec<Discovere
         let discovered = match LayerName::from_str(&file_name) {
             Ok(file_name) => {
                 let file_size = direntry.metadata()?.len();
-                Discovered::Layer(file_name, direntry.path().to_owned(), file_size)
+                Discovered::Layer(
+                    file_name,
+                    LocalLayerFileMetadata::new(direntry.path().to_owned(), file_size),
+                )
             }
             Err(_) => {
-                if file_name == METADATA_FILE_NAME {
-                    Discovered::Metadata
-                } else if file_name.ends_with(".old") {
+                if file_name.ends_with(".old") {
                     // ignore these
-                    Discovered::IgnoredBackup
+                    Discovered::IgnoredBackup(direntry.path().to_owned())
                 } else if remote_timeline_client::is_temp_download_file(direntry.path()) {
                     Discovered::TemporaryDownload(file_name)
                 } else if is_ephemeral_file(&file_name) {
@@ -76,37 +75,32 @@ pub(super) fn scan_timeline_dir(path: &Utf8Path) -> anyhow::Result<Vec<Discovere
 /// this structure extends it with metadata describing the layer's presence in local storage.
 #[derive(Clone, Debug)]
 pub(super) struct LocalLayerFileMetadata {
-    pub(super) metadata: LayerFileMetadata,
+    pub(super) file_size: u64,
     pub(super) local_path: Utf8PathBuf,
 }
 
 impl LocalLayerFileMetadata {
-    pub fn new(
-        local_path: Utf8PathBuf,
-        file_size: u64,
-        generation: Generation,
-        shard: ShardIndex,
-    ) -> Self {
+    pub fn new(local_path: Utf8PathBuf, file_size: u64) -> Self {
         Self {
             local_path,
-            metadata: LayerFileMetadata::new(file_size, generation, shard),
+            file_size,
         }
     }
 }
 
-/// Decision on what to do with a layer file after considering its local and remote metadata.
+/// For a layer that is present in remote metadata, this type describes how to handle
+/// it during startup: it is either Resident (and we have some metadata about a local file),
+/// or it is Evicted (and we only have remote metadata).
 #[derive(Clone, Debug)]
 pub(super) enum Decision {
     /// The layer is not present locally.
     Evicted(LayerFileMetadata),
-    /// The layer is present locally, but local metadata does not match remote; we must
-    /// delete it and treat it as evicted.
-    UseRemote {
+    /// The layer is present locally, and metadata matches: we may hook up this layer to the
+    /// existing file in local storage.
+    Resident {
         local: LocalLayerFileMetadata,
         remote: LayerFileMetadata,
     },
-    /// The layer is present locally, and metadata matches.
-    UseLocal(LocalLayerFileMetadata),
 }
 
 /// A layer needs to be left out of the layer map.
@@ -122,77 +116,81 @@ pub(super) enum DismissedLayer {
     /// In order to make crash safe updates to layer map, we must dismiss layers which are only
     /// found locally or not yet included in the remote `index_part.json`.
     LocalOnly(LocalLayerFileMetadata),
+
+    /// The layer exists in remote storage but the local layer's metadata (e.g. file size)
+    /// does not match it
+    BadMetadata(LocalLayerFileMetadata),
 }
 
 /// Merges local discoveries and remote [`IndexPart`] to a collection of decisions.
 pub(super) fn reconcile(
-    discovered: Vec<(LayerName, Utf8PathBuf, u64)>,
+    local_layers: Vec<(LayerName, LocalLayerFileMetadata)>,
     index_part: Option<&IndexPart>,
     disk_consistent_lsn: Lsn,
-    generation: Generation,
-    shard: ShardIndex,
 ) -> Vec<(LayerName, Result<Decision, DismissedLayer>)> {
-    use Decision::*;
+    let Some(index_part) = index_part else {
+        // If we have no remote metadata, no local layer files are considered valid to load
+        return local_layers
+            .into_iter()
+            .map(|(layer_name, local_metadata)| {
+                (layer_name, Err(DismissedLayer::LocalOnly(local_metadata)))
+            })
+            .collect();
+    };
 
-    // name => (local_metadata, remote_metadata)
-    type Collected =
-        HashMap<LayerName, (Option<LocalLayerFileMetadata>, Option<LayerFileMetadata>)>;
+    let mut result = Vec::new();
 
-    let mut discovered = discovered
-        .into_iter()
-        .map(|(layer_name, local_path, file_size)| {
-            (
-                layer_name,
-                // The generation and shard here will be corrected to match IndexPart in the merge below, unless
-                // it is not in IndexPart, in which case using our current generation makes sense
-                // because it will be uploaded in this generation.
-                (
-                    Some(LocalLayerFileMetadata::new(
-                        local_path, file_size, generation, shard,
-                    )),
-                    None,
-                ),
-            )
-        })
-        .collect::<Collected>();
+    let mut remote_layers = HashMap::new();
 
-    // merge any index_part information, when available
+    // Construct Decisions for layers that are found locally, if they're in remote metadata.  Otherwise
+    // construct DismissedLayers to get rid of them.
+    for (layer_name, local_metadata) in local_layers {
+        let Some(remote_metadata) = index_part.layer_metadata.get(&layer_name) else {
+            result.push((layer_name, Err(DismissedLayer::LocalOnly(local_metadata))));
+            continue;
+        };
+
+        if remote_metadata.file_size != local_metadata.file_size {
+            result.push((layer_name, Err(DismissedLayer::BadMetadata(local_metadata))));
+            continue;
+        }
+
+        remote_layers.insert(
+            layer_name,
+            Decision::Resident {
+                local: local_metadata,
+                remote: remote_metadata.clone(),
+            },
+        );
+    }
+
+    // Construct Decision for layers that were not found locally
     index_part
-        .as_ref()
-        .map(|ip| ip.layer_metadata.iter())
-        .into_iter()
-        .flatten()
-        .map(|(name, metadata)| (name, LayerFileMetadata::from(metadata)))
+        .layer_metadata
+        .iter()
         .for_each(|(name, metadata)| {
-            if let Some(existing) = discovered.get_mut(name) {
-                existing.1 = Some(metadata);
-            } else {
-                discovered.insert(name.to_owned(), (None, Some(metadata)));
+            if let hash_map::Entry::Vacant(entry) = remote_layers.entry(name.clone()) {
+                entry.insert(Decision::Evicted(metadata.clone()));
             }
         });
 
-    discovered
-        .into_iter()
-        .map(|(name, (local, remote))| {
-            let decision = if name.is_in_future(disk_consistent_lsn) {
-                Err(DismissedLayer::Future { local })
-            } else {
-                match (local, remote) {
-                    (Some(local), Some(remote)) if local.metadata != remote => {
-                        Ok(UseRemote { local, remote })
-                    }
-                    (Some(x), Some(_)) => Ok(UseLocal(x)),
-                    (None, Some(x)) => Ok(Evicted(x)),
-                    (Some(x), None) => Err(DismissedLayer::LocalOnly(x)),
-                    (None, None) => {
-                        unreachable!("there must not be any non-local non-remote files")
-                    }
-                }
-            };
+    // For layers that were found in authoritative remote metadata, apply a final check that they are within
+    // the disk_consistent_lsn.
+    result.extend(remote_layers.into_iter().map(|(name, decision)| {
+        if name.is_in_future(disk_consistent_lsn) {
+            match decision {
+                Decision::Evicted(_remote) => (name, Err(DismissedLayer::Future { local: None })),
+                Decision::Resident {
+                    local,
+                    remote: _remote,
+                } => (name, Err(DismissedLayer::Future { local: Some(local) })),
+            }
+        } else {
+            (name, Ok(decision))
+        }
+    }));
 
-            (name, decision)
-        })
-        .collect::<Vec<_>>()
+    result
 }
 
 pub(super) fn cleanup(path: &Utf8Path, kind: &str) -> anyhow::Result<()> {
@@ -201,25 +199,15 @@ pub(super) fn cleanup(path: &Utf8Path, kind: &str) -> anyhow::Result<()> {
     std::fs::remove_file(path).with_context(|| format!("failed to remove {kind} at {path}"))
 }
 
-pub(super) fn cleanup_local_file_for_remote(
-    local: &LocalLayerFileMetadata,
-    remote: &LayerFileMetadata,
-) -> anyhow::Result<()> {
-    let local_size = local.metadata.file_size();
-    let remote_size = remote.file_size();
+pub(super) fn cleanup_local_file_for_remote(local: &LocalLayerFileMetadata) -> anyhow::Result<()> {
+    let local_size = local.file_size;
     let path = &local.local_path;
-
     let file_name = path.file_name().expect("must be file path");
-    tracing::warn!("removing local file {file_name:?} because it has unexpected length {local_size}; length in remote index is {remote_size}");
-    if let Err(err) = crate::tenant::timeline::rename_to_backup(path) {
-        assert!(
-            path.exists(),
-            "we would leave the local_layer without a file if this does not hold: {path}",
-        );
-        Err(err)
-    } else {
-        Ok(())
-    }
+    tracing::warn!(
+        "removing local file {file_name:?} because it has unexpected length {local_size};"
+    );
+
+    std::fs::remove_file(path).with_context(|| format!("failed to remove layer at {path}"))
 }
 
 pub(super) fn cleanup_future_layer(
@@ -241,8 +229,8 @@ pub(super) fn cleanup_local_only_file(
 ) -> anyhow::Result<()> {
     let kind = name.kind();
     tracing::info!(
-        "found local-only {kind} layer {name}, metadata {:?}",
-        local.metadata
+        "found local-only {kind} layer {name} size {}",
+        local.file_size
     );
     std::fs::remove_file(&local.local_path)?;
     Ok(())

pageserver/src/tenant/timeline/layer_manager.rs

@@ -212,13 +212,34 @@ impl LayerManager {
         &mut self,
         rewrite_layers: &[(Layer, ResidentLayer)],
         drop_layers: &[Layer],
-        _metrics: &TimelineMetrics,
+        metrics: &TimelineMetrics,
     ) {
         let mut updates = self.layer_map.batch_update();
+        for (old_layer, new_layer) in rewrite_layers {
+            debug_assert_eq!(
+                old_layer.layer_desc().key_range,
+                new_layer.layer_desc().key_range
+            );
+            debug_assert_eq!(
+                old_layer.layer_desc().lsn_range,
+                new_layer.layer_desc().lsn_range
+            );
 
-        // TODO: implement rewrites (currently this code path only used for drops)
-        assert!(rewrite_layers.is_empty());
+            // Safety: we may never rewrite the same file in-place.  Callers are responsible
+            // for ensuring that they only rewrite layers after something changes the path,
+            // such as an increment in the generation number.
+            assert_ne!(old_layer.local_path(), new_layer.local_path());
 
+            Self::delete_historic_layer(old_layer, &mut updates, &mut self.layer_fmgr);
+
+            Self::insert_historic_layer(
+                new_layer.as_ref().clone(),
+                &mut updates,
+                &mut self.layer_fmgr,
+            );
+
+            metrics.record_new_file_metrics(new_layer.layer_desc().file_size);
+        }
         for l in drop_layers {
             Self::delete_historic_layer(l, &mut updates, &mut self.layer_fmgr);
         }

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

@@ -705,6 +705,7 @@ impl ConnectionManagerState {
                     commit_lsn: info.commit_lsn,
                     safekeeper_connstr: info.safekeeper_connstr,
                     availability_zone: info.availability_zone,
+                    standby_horizon: info.standby_horizon,
                 }
             }
             MessageType::SafekeeperDiscoveryResponse => {
@@ -725,6 +726,21 @@ impl ConnectionManagerState {
 
         WALRECEIVER_BROKER_UPDATES.inc();
 
+        trace!(
+            "safekeeper info update: standby_horizon(cutoff)={}",
+            timeline_update.standby_horizon
+        );
+        if timeline_update.standby_horizon != 0 {
+            // ignore reports from safekeepers not connected to replicas
+            self.timeline
+                .standby_horizon
+                .store(Lsn(timeline_update.standby_horizon));
+            self.timeline
+                .metrics
+                .standby_horizon_gauge
+                .set(timeline_update.standby_horizon as i64);
+        }
+
         let new_safekeeper_id = NodeId(timeline_update.safekeeper_id);
         let old_entry = self.wal_stream_candidates.insert(
             new_safekeeper_id,
@@ -1094,6 +1110,7 @@ mod tests {
                 commit_lsn,
                 safekeeper_connstr: safekeeper_connstr.to_owned(),
                 availability_zone: None,
+                standby_horizon: 0,
             },
             latest_update,
         }

pageserver/src/tenant/upload_queue.rs

@@ -8,6 +8,7 @@ use std::collections::{HashMap, VecDeque};
 use std::fmt::Debug;
 
 use chrono::NaiveDateTime;
+use pageserver_api::models::AuxFilePolicy;
 use std::sync::Arc;
 use tracing::info;
 use utils::lsn::AtomicLsn;
@@ -60,6 +61,9 @@ pub(crate) struct UploadQueueInitialized {
     /// Part of the flattened "next" `index_part.json`.
     pub(crate) latest_lineage: Lineage,
 
+    /// The last aux file policy used on this timeline.
+    pub(crate) last_aux_file_policy: Option<AuxFilePolicy>,
+
     /// `disk_consistent_lsn` from the last metadata file that was successfully
     /// uploaded. `Lsn(0)` if nothing was uploaded yet.
     /// Unlike `latest_files` or `latest_metadata`, this value is never ahead.
@@ -189,6 +193,7 @@ impl UploadQueue {
             dangling_files: HashMap::new(),
             shutting_down: false,
             shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
+            last_aux_file_policy: Default::default(),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -208,10 +213,7 @@ impl UploadQueue {
 
         let mut files = HashMap::with_capacity(index_part.layer_metadata.len());
         for (layer_name, layer_metadata) in &index_part.layer_metadata {
-            files.insert(
-                layer_name.to_owned(),
-                LayerFileMetadata::from(layer_metadata),
-            );
+            files.insert(layer_name.to_owned(), layer_metadata.clone());
         }
 
         info!(
@@ -239,6 +241,7 @@ impl UploadQueue {
             dangling_files: HashMap::new(),
             shutting_down: false,
             shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
+            last_aux_file_policy: index_part.last_aux_file_policy(),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -316,9 +319,7 @@ impl std::fmt::Display for UploadOp {
                 write!(
                     f,
                     "UploadLayer({}, size={:?}, gen={:?})",
-                    layer,
-                    metadata.file_size(),
-                    metadata.generation
+                    layer, metadata.file_size, metadata.generation
                 )
             }
             UploadOp::UploadMetadata(_, lsn) => {

pageserver/src/walredo.rs

@@ -153,10 +153,7 @@ impl PostgresRedoManager {
             process: self
                 .redo_process
                 .get()
-                .map(|p| WalRedoManagerProcessStatus {
-                    pid: p.id(),
-                    kind: std::borrow::Cow::Borrowed(p.kind().into()),
-                }),
+                .map(|p| WalRedoManagerProcessStatus { pid: p.id() }),
         }
     }
 }

pageserver/src/walredo/process.rs

@@ -1,7 +1,10 @@
+/// Layer of indirection previously used to support multiple implementations.
+/// Subject to removal: <https://github.com/neondatabase/neon/issues/7753>
 use std::time::Duration;
 
 use bytes::Bytes;
 use pageserver_api::{reltag::RelTag, shard::TenantShardId};
+use tracing::warn;
 use utils::lsn::Lsn;
 
 use crate::{config::PageServerConf, walrecord::NeonWalRecord};
@@ -12,7 +15,6 @@ mod protocol;
 
 mod process_impl {
     pub(super) mod process_async;
-    pub(super) mod process_std;
 }
 
 #[derive(
@@ -34,10 +36,7 @@ pub enum Kind {
     Async,
 }
 
-pub(crate) enum Process {
-    Sync(process_impl::process_std::WalRedoProcess),
-    Async(process_impl::process_async::WalRedoProcess),
-}
+pub(crate) struct Process(process_impl::process_async::WalRedoProcess);
 
 impl Process {
     #[inline(always)]
@@ -46,18 +45,17 @@ impl Process {
         tenant_shard_id: TenantShardId,
         pg_version: u32,
     ) -> anyhow::Result<Self> {
-        Ok(match conf.walredo_process_kind {
-            Kind::Sync => Self::Sync(process_impl::process_std::WalRedoProcess::launch(
-                conf,
-                tenant_shard_id,
-                pg_version,
-            )?),
-            Kind::Async => Self::Async(process_impl::process_async::WalRedoProcess::launch(
-                conf,
-                tenant_shard_id,
-                pg_version,
-            )?),
-        })
+        if conf.walredo_process_kind != Kind::Async {
+            warn!(
+                configured = %conf.walredo_process_kind,
+                "the walredo_process_kind setting has been turned into a no-op, using async implementation"
+            );
+        }
+        Ok(Self(process_impl::process_async::WalRedoProcess::launch(
+            conf,
+            tenant_shard_id,
+            pg_version,
+        )?))
     }
 
     #[inline(always)]
@@ -69,29 +67,12 @@ impl Process {
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
     ) -> anyhow::Result<Bytes> {
-        match self {
-            Process::Sync(p) => {
-                p.apply_wal_records(rel, blknum, base_img, records, wal_redo_timeout)
-                    .await
-            }
-            Process::Async(p) => {
-                p.apply_wal_records(rel, blknum, base_img, records, wal_redo_timeout)
-                    .await
-            }
-        }
+        self.0
+            .apply_wal_records(rel, blknum, base_img, records, wal_redo_timeout)
+            .await
     }
 
     pub(crate) fn id(&self) -> u32 {
-        match self {
-            Process::Sync(p) => p.id(),
-            Process::Async(p) => p.id(),
-        }
-    }
-
-    pub(crate) fn kind(&self) -> Kind {
-        match self {
-            Process::Sync(_) => Kind::Sync,
-            Process::Async(_) => Kind::Async,
-        }
+        self.0.id()
     }
 }

pageserver/src/walredo/process/process_impl/process_std.rs

@@ -1,405 +0,0 @@
-use self::no_leak_child::NoLeakChild;
-use crate::{
-    config::PageServerConf,
-    metrics::{WalRedoKillCause, WAL_REDO_PROCESS_COUNTERS, WAL_REDO_RECORD_COUNTER},
-    walrecord::NeonWalRecord,
-    walredo::process::{no_leak_child, protocol},
-};
-use anyhow::Context;
-use bytes::Bytes;
-use nix::poll::{PollFd, PollFlags};
-use pageserver_api::{reltag::RelTag, shard::TenantShardId};
-use postgres_ffi::BLCKSZ;
-use std::os::fd::AsRawFd;
-#[cfg(feature = "testing")]
-use std::sync::atomic::AtomicUsize;
-use std::{
-    collections::VecDeque,
-    io::{Read, Write},
-    process::{ChildStdin, ChildStdout, Command, Stdio},
-    sync::{Mutex, MutexGuard},
-    time::Duration,
-};
-use tracing::{debug, error, instrument, Instrument};
-use utils::{lsn::Lsn, nonblock::set_nonblock};
-
-pub struct WalRedoProcess {
-    #[allow(dead_code)]
-    conf: &'static PageServerConf,
-    tenant_shard_id: TenantShardId,
-    // Some() on construction, only becomes None on Drop.
-    child: Option<NoLeakChild>,
-    stdout: Mutex<ProcessOutput>,
-    stdin: Mutex<ProcessInput>,
-    /// Counter to separate same sized walredo inputs failing at the same millisecond.
-    #[cfg(feature = "testing")]
-    dump_sequence: AtomicUsize,
-}
-
-struct ProcessInput {
-    stdin: ChildStdin,
-    n_requests: usize,
-}
-
-struct ProcessOutput {
-    stdout: ChildStdout,
-    pending_responses: VecDeque<Option<Bytes>>,
-    n_processed_responses: usize,
-}
-
-impl WalRedoProcess {
-    //
-    // Start postgres binary in special WAL redo mode.
-    //
-    #[instrument(skip_all,fields(pg_version=pg_version))]
-    pub(crate) fn launch(
-        conf: &'static PageServerConf,
-        tenant_shard_id: TenantShardId,
-        pg_version: u32,
-    ) -> anyhow::Result<Self> {
-        crate::span::debug_assert_current_span_has_tenant_id();
-
-        let pg_bin_dir_path = conf.pg_bin_dir(pg_version).context("pg_bin_dir")?; // TODO these should be infallible.
-        let pg_lib_dir_path = conf.pg_lib_dir(pg_version).context("pg_lib_dir")?;
-
-        use no_leak_child::NoLeakChildCommandExt;
-        // Start postgres itself
-        let child = Command::new(pg_bin_dir_path.join("postgres"))
-            // the first arg must be --wal-redo so the child process enters into walredo mode
-            .arg("--wal-redo")
-            // the child doesn't process this arg, but, having it in the argv helps indentify the
-            // walredo process for a particular tenant when debugging a pagserver
-            .args(["--tenant-shard-id", &format!("{tenant_shard_id}")])
-            .stdin(Stdio::piped())
-            .stderr(Stdio::piped())
-            .stdout(Stdio::piped())
-            .env_clear()
-            .env("LD_LIBRARY_PATH", &pg_lib_dir_path)
-            .env("DYLD_LIBRARY_PATH", &pg_lib_dir_path)
-            // NB: The redo process is not trusted after we sent it the first
-            // walredo work. Before that, it is trusted. Specifically, we trust
-            // it to
-            // 1. close all file descriptors except stdin, stdout, stderr because
-            //    pageserver might not be 100% diligent in setting FD_CLOEXEC on all
-            //    the files it opens, and
-            // 2. to use seccomp to sandbox itself before processing the first
-            //    walredo request.
-            .spawn_no_leak_child(tenant_shard_id)
-            .context("spawn process")?;
-        WAL_REDO_PROCESS_COUNTERS.started.inc();
-        let mut child = scopeguard::guard(child, |child| {
-            error!("killing wal-redo-postgres process due to a problem during launch");
-            child.kill_and_wait(WalRedoKillCause::Startup);
-        });
-
-        let stdin = child.stdin.take().unwrap();
-        let stdout = child.stdout.take().unwrap();
-        let stderr = child.stderr.take().unwrap();
-        let stderr = tokio::process::ChildStderr::from_std(stderr)
-            .context("convert to tokio::ChildStderr")?;
-        macro_rules! set_nonblock_or_log_err {
-        ($file:ident) => {{
-            let res = set_nonblock($file.as_raw_fd());
-            if let Err(e) = &res {
-                error!(error = %e, file = stringify!($file), pid = child.id(), "set_nonblock failed");
-            }
-            res
-        }};
-    }
-        set_nonblock_or_log_err!(stdin)?;
-        set_nonblock_or_log_err!(stdout)?;
-
-        // all fallible operations post-spawn are complete, so get rid of the guard
-        let child = scopeguard::ScopeGuard::into_inner(child);
-
-        tokio::spawn(
-            async move {
-                scopeguard::defer! {
-                    debug!("wal-redo-postgres stderr_logger_task finished");
-                    crate::metrics::WAL_REDO_PROCESS_COUNTERS.active_stderr_logger_tasks_finished.inc();
-                }
-                debug!("wal-redo-postgres stderr_logger_task started");
-                crate::metrics::WAL_REDO_PROCESS_COUNTERS.active_stderr_logger_tasks_started.inc();
-
-                use tokio::io::AsyncBufReadExt;
-                let mut stderr_lines = tokio::io::BufReader::new(stderr);
-                let mut buf = Vec::new();
-                let res = loop {
-                    buf.clear();
-                    // TODO we don't trust the process to cap its stderr length.
-                    // Currently it can do unbounded Vec allocation.
-                    match stderr_lines.read_until(b'\n', &mut buf).await {
-                        Ok(0) => break Ok(()), // eof
-                        Ok(num_bytes) => {
-                            let output = String::from_utf8_lossy(&buf[..num_bytes]);
-                            error!(%output, "received output");
-                        }
-                        Err(e) => {
-                            break Err(e);
-                        }
-                    }
-                };
-                match res {
-                    Ok(()) => (),
-                    Err(e) => {
-                        error!(error=?e, "failed to read from walredo stderr");
-                    }
-                }
-            }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %pg_version))
-        );
-
-        Ok(Self {
-            conf,
-            tenant_shard_id,
-            child: Some(child),
-            stdin: Mutex::new(ProcessInput {
-                stdin,
-                n_requests: 0,
-            }),
-            stdout: Mutex::new(ProcessOutput {
-                stdout,
-                pending_responses: VecDeque::new(),
-                n_processed_responses: 0,
-            }),
-            #[cfg(feature = "testing")]
-            dump_sequence: AtomicUsize::default(),
-        })
-    }
-
-    pub(crate) fn id(&self) -> u32 {
-        self.child
-            .as_ref()
-            .expect("must not call this during Drop")
-            .id()
-    }
-
-    // Apply given WAL records ('records') over an old page image. Returns
-    // new page image.
-    //
-    #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), pid=%self.id()))]
-    pub(crate) async fn apply_wal_records(
-        &self,
-        rel: RelTag,
-        blknum: u32,
-        base_img: &Option<Bytes>,
-        records: &[(Lsn, NeonWalRecord)],
-        wal_redo_timeout: Duration,
-    ) -> anyhow::Result<Bytes> {
-        let tag = protocol::BufferTag { rel, blknum };
-        let input = self.stdin.lock().unwrap();
-
-        // Serialize all the messages to send the WAL redo process first.
-        //
-        // This could be problematic if there are millions of records to replay,
-        // but in practice the number of records is usually so small that it doesn't
-        // matter, and it's better to keep this code simple.
-        //
-        // Most requests start with a before-image with BLCKSZ bytes, followed by
-        // by some other WAL records. Start with a buffer that can hold that
-        // comfortably.
-        let mut writebuf: Vec<u8> = Vec::with_capacity((BLCKSZ as usize) * 3);
-        protocol::build_begin_redo_for_block_msg(tag, &mut writebuf);
-        if let Some(img) = base_img {
-            protocol::build_push_page_msg(tag, img, &mut writebuf);
-        }
-        for (lsn, rec) in records.iter() {
-            if let NeonWalRecord::Postgres {
-                will_init: _,
-                rec: postgres_rec,
-            } = rec
-            {
-                protocol::build_apply_record_msg(*lsn, postgres_rec, &mut writebuf);
-            } else {
-                anyhow::bail!("tried to pass neon wal record to postgres WAL redo");
-            }
-        }
-        protocol::build_get_page_msg(tag, &mut writebuf);
-        WAL_REDO_RECORD_COUNTER.inc_by(records.len() as u64);
-
-        let res = self.apply_wal_records0(&writebuf, input, wal_redo_timeout);
-
-        if res.is_err() {
-            // not all of these can be caused by this particular input, however these are so rare
-            // in tests so capture all.
-            self.record_and_log(&writebuf);
-        }
-
-        res
-    }
-
-    fn apply_wal_records0(
-        &self,
-        writebuf: &[u8],
-        input: MutexGuard<ProcessInput>,
-        wal_redo_timeout: Duration,
-    ) -> anyhow::Result<Bytes> {
-        let mut proc = { input }; // TODO: remove this legacy rename, but this keep the patch small.
-        let mut nwrite = 0usize;
-
-        while nwrite < writebuf.len() {
-            let mut stdin_pollfds = [PollFd::new(&proc.stdin, PollFlags::POLLOUT)];
-            let n = loop {
-                match nix::poll::poll(&mut stdin_pollfds[..], wal_redo_timeout.as_millis() as i32) {
-                    Err(nix::errno::Errno::EINTR) => continue,
-                    res => break res,
-                }
-            }?;
-
-            if n == 0 {
-                anyhow::bail!("WAL redo timed out");
-            }
-
-            // If 'stdin' is writeable, do write.
-            let in_revents = stdin_pollfds[0].revents().unwrap();
-            if in_revents & (PollFlags::POLLERR | PollFlags::POLLOUT) != PollFlags::empty() {
-                nwrite += proc.stdin.write(&writebuf[nwrite..])?;
-            }
-            if in_revents.contains(PollFlags::POLLHUP) {
-                // We still have more data to write, but the process closed the pipe.
-                anyhow::bail!("WAL redo process closed its stdin unexpectedly");
-            }
-        }
-        let request_no = proc.n_requests;
-        proc.n_requests += 1;
-        drop(proc);
-
-        // To improve walredo performance we separate sending requests and receiving
-        // responses. Them are protected by different mutexes (output and input).
-        // If thread T1, T2, T3 send requests D1, D2, D3 to walredo process
-        // then there is not warranty that T1 will first granted output mutex lock.
-        // To address this issue we maintain number of sent requests, number of processed
-        // responses and ring buffer with pending responses. After sending response
-        // (under input mutex), threads remembers request number. Then it releases
-        // input mutex, locks output mutex and fetch in ring buffer all responses until
-        // its stored request number. The it takes correspondent element from
-        // pending responses ring buffer and truncate all empty elements from the front,
-        // advancing processed responses number.
-
-        let mut output = self.stdout.lock().unwrap();
-        let n_processed_responses = output.n_processed_responses;
-        while n_processed_responses + output.pending_responses.len() <= request_no {
-            // We expect the WAL redo process to respond with an 8k page image. We read it
-            // into this buffer.
-            let mut resultbuf = vec![0; BLCKSZ.into()];
-            let mut nresult: usize = 0; // # of bytes read into 'resultbuf' so far
-            while nresult < BLCKSZ.into() {
-                let mut stdout_pollfds = [PollFd::new(&output.stdout, PollFlags::POLLIN)];
-                // We do two things simultaneously: reading response from stdout
-                // and forward any logging information that the child writes to its stderr to the page server's log.
-                let n = loop {
-                    match nix::poll::poll(
-                        &mut stdout_pollfds[..],
-                        wal_redo_timeout.as_millis() as i32,
-                    ) {
-                        Err(nix::errno::Errno::EINTR) => continue,
-                        res => break res,
-                    }
-                }?;
-
-                if n == 0 {
-                    anyhow::bail!("WAL redo timed out");
-                }
-
-                // If we have some data in stdout, read it to the result buffer.
-                let out_revents = stdout_pollfds[0].revents().unwrap();
-                if out_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
-                    nresult += output.stdout.read(&mut resultbuf[nresult..])?;
-                }
-                if out_revents.contains(PollFlags::POLLHUP) {
-                    anyhow::bail!("WAL redo process closed its stdout unexpectedly");
-                }
-            }
-            output
-                .pending_responses
-                .push_back(Some(Bytes::from(resultbuf)));
-        }
-        // Replace our request's response with None in `pending_responses`.
-        // Then make space in the ring buffer by clearing out any seqence of contiguous
-        // `None`'s from the front of `pending_responses`.
-        // NB: We can't pop_front() because other requests' responses because another
-        // requester might have grabbed the output mutex before us:
-        // T1: grab input mutex
-        // T1: send request_no 23
-        // T1: release input mutex
-        // T2: grab input mutex
-        // T2: send request_no 24
-        // T2: release input mutex
-        // T2: grab output mutex
-        // T2: n_processed_responses + output.pending_responses.len() <= request_no
-        //            23                                0                   24
-        // T2: enters poll loop that reads stdout
-        // T2: put response for 23 into pending_responses
-        // T2: put response for 24 into pending_resposnes
-        // pending_responses now looks like this: Front Some(response_23) Some(response_24) Back
-        // T2: takes its response_24
-        // pending_responses now looks like this: Front Some(response_23) None Back
-        // T2: does the while loop below
-        // pending_responses now looks like this: Front Some(response_23) None Back
-        // T2: releases output mutex
-        // T1: grabs output mutex
-        // T1: n_processed_responses + output.pending_responses.len() > request_no
-        //            23                                2                   23
-        // T1: skips poll loop that reads stdout
-        // T1: takes its response_23
-        // pending_responses now looks like this: Front None None Back
-        // T2: does the while loop below
-        // pending_responses now looks like this: Front Back
-        // n_processed_responses now has value 25
-        let res = output.pending_responses[request_no - n_processed_responses]
-            .take()
-            .expect("we own this request_no, nobody else is supposed to take it");
-        while let Some(front) = output.pending_responses.front() {
-            if front.is_none() {
-                output.pending_responses.pop_front();
-                output.n_processed_responses += 1;
-            } else {
-                break;
-            }
-        }
-        Ok(res)
-    }
-
-    #[cfg(feature = "testing")]
-    fn record_and_log(&self, writebuf: &[u8]) {
-        use std::sync::atomic::Ordering;
-
-        let millis = std::time::SystemTime::now()
-            .duration_since(std::time::SystemTime::UNIX_EPOCH)
-            .unwrap()
-            .as_millis();
-
-        let seq = self.dump_sequence.fetch_add(1, Ordering::Relaxed);
-
-        // these files will be collected to an allure report
-        let filename = format!("walredo-{millis}-{}-{seq}.walredo", writebuf.len());
-
-        let path = self.conf.tenant_path(&self.tenant_shard_id).join(&filename);
-
-        let res = std::fs::OpenOptions::new()
-            .write(true)
-            .create_new(true)
-            .read(true)
-            .open(path)
-            .and_then(|mut f| f.write_all(writebuf));
-
-        // trip up allowed_errors
-        if let Err(e) = res {
-            tracing::error!(target=%filename, length=writebuf.len(), "failed to write out the walredo errored input: {e}");
-        } else {
-            tracing::error!(filename, "erroring walredo input saved");
-        }
-    }
-
-    #[cfg(not(feature = "testing"))]
-    fn record_and_log(&self, _: &[u8]) {}
-}
-
-impl Drop for WalRedoProcess {
-    fn drop(&mut self) {
-        self.child
-            .take()
-            .expect("we only do this once")
-            .kill_and_wait(WalRedoKillCause::WalRedoProcessDrop);
-        // no way to wait for stderr_logger_task from Drop because that is async only
-    }
-}

patches/pgvector.patch

@@ -0,0 +1,78 @@
+From 0b0194a57bd0f3598bd57dbedd0df3932330169d Mon Sep 17 00:00:00 2001
+From: Heikki Linnakangas <heikki.linnakangas@iki.fi>
+Date: Fri, 2 Feb 2024 22:26:45 +0200
+Subject: [PATCH 1/1] Make v0.6.0 work with Neon
+
+Now that the WAL-logging happens as a separate step at the end of the
+build, we need a few neon-specific hints to make it work.
+---
+ src/hnswbuild.c | 36 ++++++++++++++++++++++++++++++++++++
+ 1 file changed, 36 insertions(+)
+
+diff --git a/src/hnswbuild.c b/src/hnswbuild.c
+index 680789b..ec54dea 100644
+--- a/src/hnswbuild.c
++++ b/src/hnswbuild.c
+@@ -840,9 +840,17 @@ HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc)
+ 
+ 	hnswarea = shm_toc_lookup(toc, PARALLEL_KEY_HNSW_AREA, false);
+ 
++#ifdef NEON_SMGR
++	smgr_start_unlogged_build(RelationGetSmgr(indexRel));
++#endif
++
+ 	/* Perform inserts */
+ 	HnswParallelScanAndInsert(heapRel, indexRel, hnswshared, hnswarea, false);
+ 
++#ifdef NEON_SMGR
++	smgr_finish_unlogged_build_phase_1(RelationGetSmgr(indexRel));
++#endif
++
+ 	/* Close relations within worker */
+ 	index_close(indexRel, indexLockmode);
+ 	table_close(heapRel, heapLockmode);
+@@ -1089,13 +1097,41 @@ BuildIndex(Relation heap, Relation index, IndexInfo *indexInfo,
+ 	SeedRandom(42);
+ #endif
+ 
++#ifdef NEON_SMGR
++	smgr_start_unlogged_build(RelationGetSmgr(index));
++#endif
++
+ 	InitBuildState(buildstate, heap, index, indexInfo, forkNum);
+ 
+ 	BuildGraph(buildstate, forkNum);
+ 
++#ifdef NEON_SMGR
++	smgr_finish_unlogged_build_phase_1(RelationGetSmgr(index));
++#endif
++
+ 	if (RelationNeedsWAL(index))
++	{
+ 		log_newpage_range(index, forkNum, 0, RelationGetNumberOfBlocks(index), true);
+ 
++#ifdef NEON_SMGR
++		{
++#if PG_VERSION_NUM >= 160000
++			RelFileLocator rlocator = RelationGetSmgr(index)->smgr_rlocator.locator;
++#else
++			RelFileNode rlocator = RelationGetSmgr(index)->smgr_rnode.node;
++#endif
++
++			SetLastWrittenLSNForBlockRange(XactLastRecEnd, rlocator,
++										   MAIN_FORKNUM, 0, RelationGetNumberOfBlocks(index));
++			SetLastWrittenLSNForRelation(XactLastRecEnd, rlocator, MAIN_FORKNUM);
++		}
++#endif
++	}
++
++#ifdef NEON_SMGR
++	smgr_end_unlogged_build(RelationGetSmgr(index));
++#endif
++
+ 	FreeBuildState(buildstate);
+ }
+ 
+-- 
+2.39.2
+

pgxn/neon/libpagestore.c

@@ -49,9 +49,8 @@ char	   *neon_auth_token;
 int			readahead_buffer_size = 128;
 int			flush_every_n_requests = 8;
 
-int         neon_protocol_version = 1;
+int         neon_protocol_version = 2;
 
-static int	n_reconnect_attempts = 0;
 static int	max_reconnect_attempts = 60;
 static int	stripe_size;
 
@@ -95,18 +94,37 @@ static shmem_startup_hook_type prev_shmem_startup_hook;
 static PagestoreShmemState *pagestore_shared;
 static uint64 pagestore_local_counter = 0;
 
+typedef enum PSConnectionState {
+	PS_Disconnected,			/* no connection yet */
+	PS_Connecting_Startup,		/* connection starting up */
+	PS_Connecting_PageStream,	/* negotiating pagestream */ 
+	PS_Connected,				/* connected, pagestream established */
+} PSConnectionState;
+
 /* This backend's per-shard connections */
 typedef struct
 {
-	PGconn	   *conn;
+	TimestampTz		last_connect_time; /* read-only debug value */
+	TimestampTz		last_reconnect_time;
+	uint32			delay_us;
+	int				n_reconnect_attempts;
 
 	/*---
-	 * WaitEventSet containing:
-	 * - WL_SOCKET_READABLE on 'conn'
-	 * - WL_LATCH_SET on MyLatch, and
-	 * - WL_EXIT_ON_PM_DEATH.
+	 * Pageserver connection state, i.e.
+	 *	disconnected: conn == NULL, wes == NULL;
+	 *	conn_startup: connection initiated, waiting for connection establishing
+	 *	conn_ps:      PageStream query sent, waiting for confirmation
+	 *	connected:    PageStream established
 	 */
-	WaitEventSet *wes;
+	PSConnectionState state;
+	PGconn		   *conn;
+	/*---
+	 * WaitEventSet containing:
+	 *	- WL_SOCKET_READABLE on 'conn'
+	 *	- WL_LATCH_SET on MyLatch, and
+	 *	- WL_EXIT_ON_PM_DEATH.
+	 */
+	WaitEventSet   *wes_read;
 } PageServer;
 
 static PageServer page_servers[MAX_SHARDS];
@@ -303,119 +321,277 @@ get_shard_number(BufferTag *tag)
 	return hash % n_shards;
 }
 
+static inline void
+CLEANUP_AND_DISCONNECT(PageServer *shard) 
+{
+	if (shard->wes_read)
+	{
+		FreeWaitEventSet(shard->wes_read);
+		shard->wes_read = NULL;
+	}
+	if (shard->conn)
+	{
+		PQfinish(shard->conn);
+		shard->conn = NULL;
+	}
+
+	shard->state = PS_Disconnected;
+}
+
+/*
+ * Connect to a pageserver, or continue to try to connect if we're yet to
+ * complete the connection (e.g. due to receiving an earlier cancellation
+ * during connection start).
+ * Returns true if successfully connected; false if the connection failed.
+ * 
+ * Throws errors in unrecoverable situations, or when this backend's query
+ * is canceled.
+ */
 static bool
 pageserver_connect(shardno_t shard_no, int elevel)
 {
-	char	   *query;
-	int			ret;
-	const char *keywords[3];
-	const char *values[3];
-	int			n;
-	PGconn	   *conn;
-	WaitEventSet *wes;
+	PageServer *shard = &page_servers[shard_no];
 	char		connstr[MAX_PAGESERVER_CONNSTRING_SIZE];
 
-	static TimestampTz last_connect_time = 0;
-	static uint64_t delay_us = MIN_RECONNECT_INTERVAL_USEC;
-	TimestampTz now;
-	uint64_t	us_since_last_connect;
-	bool	broke_from_loop = false;
-
-	Assert(page_servers[shard_no].conn == NULL);
-
 	/*
 	 * Get the connection string for this shard. If the shard map has been
 	 * updated since we last looked, this will also disconnect any existing
 	 * pageserver connections as a side effect.
+	 * Note that connstr is used both during connection start, and when we
+	 * log the successful connection.
 	 */
 	load_shard_map(shard_no, connstr, NULL);
 
-	now = GetCurrentTimestamp();
-	us_since_last_connect = now - last_connect_time;
-	if (us_since_last_connect < MAX_RECONNECT_INTERVAL_USEC)
+	switch (shard->state)
 	{
-		pg_usleep(delay_us);
-		delay_us *= 2;
-	}
-	else
+	case PS_Disconnected:
 	{
-		delay_us = MIN_RECONNECT_INTERVAL_USEC;
-	}
+		const char *keywords[3];
+		const char *values[3];
+		int			n_pgsql_params;
+		TimestampTz	now;
+		int64		us_since_last_attempt;
 
-	/*
-	 * Connect using the connection string we got from the
-	 * neon.pageserver_connstring GUC. If the NEON_AUTH_TOKEN environment
-	 * variable was set, use that as the password.
-	 *
-	 * The connection options are parsed in the order they're given, so when
-	 * we set the password before the connection string, the connection string
-	 * can override the password from the env variable. Seems useful, although
-	 * we don't currently use that capability anywhere.
-	 */
-	n = 0;
-	if (neon_auth_token)
-	{
-		keywords[n] = "password";
-		values[n] = neon_auth_token;
-		n++;
+		/* Make sure we start with a clean slate */
+		CLEANUP_AND_DISCONNECT(shard);
+
+		neon_shard_log(shard_no, DEBUG5, "Connection state: Disconnected");
+
+		now = GetCurrentTimestamp();
+		us_since_last_attempt = (int64) (now - shard->last_reconnect_time);
+		shard->last_reconnect_time = now;
+
+		/*
+		 * If we did other tasks between reconnect attempts, then we won't
+		 * need to wait as long as a full delay.
+		 */
+		if (us_since_last_attempt < shard->delay_us)
+		{
+			pg_usleep(shard->delay_us - us_since_last_attempt);
+		}
+
+		/* update the delay metric */
+		shard->delay_us = Min(shard->delay_us * 2, MAX_RECONNECT_INTERVAL_USEC);
+
+		/*
+		 * Connect using the connection string we got from the
+		 * neon.pageserver_connstring GUC. If the NEON_AUTH_TOKEN environment
+		 * variable was set, use that as the password.
+		 *
+		 * The connection options are parsed in the order they're given, so when
+		 * we set the password before the connection string, the connection string
+		 * can override the password from the env variable. Seems useful, although
+		 * we don't currently use that capability anywhere.
+		 */
+		keywords[0] = "dbname";
+		values[0] = connstr;
+		n_pgsql_params = 1;
+
+		if (neon_auth_token)
+		{
+			keywords[1] = "password";
+			values[1] = neon_auth_token;
+			n_pgsql_params++;
+		}
+
+		keywords[n_pgsql_params] = NULL;
+		values[n_pgsql_params] = NULL;
+
+		shard->conn = PQconnectStartParams(keywords, values, 1);
+		if (!shard->conn)
+		{
+			neon_shard_log(shard_no, elevel, "Failed to connect to pageserver: out of memory");
+			return false;
+		}
+
+		shard->state = PS_Connecting_Startup;
+		/* fallthrough */
 	}
-	keywords[n] = "dbname";
-	values[n] = connstr;
-	n++;
-	keywords[n] = NULL;
-	values[n] = NULL;
-	n++;
-	conn = PQconnectdbParams(keywords, values, 1);
-	last_connect_time = GetCurrentTimestamp();
-
-	if (PQstatus(conn) == CONNECTION_BAD)
+	case PS_Connecting_Startup:
 	{
-		char	   *msg = pchomp(PQerrorMessage(conn));
+		char	   *pagestream_query;
+		int			ps_send_query_ret;
+		bool		connected = false;
+		int poll_result = PGRES_POLLING_WRITING;
+		neon_shard_log(shard_no, DEBUG5, "Connection state: Connecting_Startup");
 
-		PQfinish(conn);
+		do
+		{
+			WaitEvent	event;
 
-		ereport(elevel,
-				(errcode(ERRCODE_SQLCLIENT_UNABLE_TO_ESTABLISH_SQLCONNECTION),
-				 errmsg(NEON_TAG "[shard %d] could not establish connection to pageserver", shard_no),
-				 errdetail_internal("%s", msg)));
-		pfree(msg);
-		return false;
-	}
-	switch (neon_protocol_version)
-	{
+			switch (poll_result)
+			{
+			default: /* unknown/unused states are handled as a failed connection */
+			case PGRES_POLLING_FAILED:
+				{
+					char	   *pqerr = PQerrorMessage(shard->conn);
+					char	   *msg = NULL;
+					neon_shard_log(shard_no, DEBUG5, "POLLING_FAILED");
+
+					if (pqerr)
+						msg = pchomp(pqerr);
+
+					CLEANUP_AND_DISCONNECT(shard);
+
+					if (msg)
+					{
+						neon_shard_log(shard_no, elevel,
+									   "could not connect to pageserver: %s",
+									   msg);
+						pfree(msg);
+					}
+					else
+						neon_shard_log(shard_no, elevel,
+									   "could not connect to pageserver");
+
+					return false;
+				}
+			case PGRES_POLLING_READING:
+				/* Sleep until there's something to do */
+				while (true)
+				{
+					int rc = WaitLatchOrSocket(MyLatch,
+											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_READABLE,
+											   PQsocket(shard->conn),
+											   0,
+											   PG_WAIT_EXTENSION);
+					elog(DEBUG5, "PGRES_POLLING_READING=>%d", rc);
+					if (rc & WL_LATCH_SET)
+					{
+						ResetLatch(MyLatch);
+						/* query cancellation, backend shutdown */
+						CHECK_FOR_INTERRUPTS();
+					}
+					if (rc & WL_SOCKET_READABLE)
+						break;
+				}
+				/* PQconnectPoll() handles the socket polling state updates */
+
+				break;
+			case PGRES_POLLING_WRITING:
+				/* Sleep until there's something to do */
+				while (true)
+				{
+					int rc = WaitLatchOrSocket(MyLatch,
+											   WL_EXIT_ON_PM_DEATH | WL_LATCH_SET | WL_SOCKET_WRITEABLE,
+											   PQsocket(shard->conn),
+											   0,
+											   PG_WAIT_EXTENSION);
+					elog(DEBUG5, "PGRES_POLLING_WRITING=>%d", rc);
+					if (rc & WL_LATCH_SET)
+					{
+						ResetLatch(MyLatch);
+						/* query cancellation, backend shutdown */
+						CHECK_FOR_INTERRUPTS();
+					}
+					if (rc & WL_SOCKET_WRITEABLE)
+						break;
+				}
+				/* PQconnectPoll() handles the socket polling state updates */
+
+				break;
+			case PGRES_POLLING_OK:
+				neon_shard_log(shard_no, DEBUG5, "POLLING_OK");
+				connected = true;
+				break;
+			}
+			poll_result = PQconnectPoll(shard->conn);
+			elog(DEBUG5, "PQconnectPoll=>%d", poll_result);
+		}
+		while (!connected);
+
+		/* No more polling needed; connection succeeded */
+		shard->last_connect_time = GetCurrentTimestamp();
+
+		shard->wes_read = CreateWaitEventSet(TopMemoryContext, 3);
+		AddWaitEventToSet(shard->wes_read, WL_LATCH_SET, PGINVALID_SOCKET,
+						  MyLatch, NULL);
+		AddWaitEventToSet(shard->wes_read, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
+						  NULL, NULL);
+		AddWaitEventToSet(shard->wes_read, WL_SOCKET_READABLE, PQsocket(shard->conn), NULL, NULL);
+
+
+		switch (neon_protocol_version)
+		{
 		case 2:
-			query = psprintf("pagestream_v2 %s %s", neon_tenant, neon_timeline);
+			pagestream_query = psprintf("pagestream_v2 %s %s", neon_tenant, neon_timeline);
 			break;
 		case 1:
-			query = psprintf("pagestream %s %s", neon_tenant, neon_timeline);
+			pagestream_query = psprintf("pagestream %s %s", neon_tenant, neon_timeline);
 			break;
 		default:
 			elog(ERROR, "unexpected neon_protocol_version %d", neon_protocol_version);
-	}
-	ret = PQsendQuery(conn, query);
-	pfree(query);
-	if (ret != 1)
-	{
-		PQfinish(conn);
-		neon_shard_log(shard_no, elevel, "could not send pagestream command to pageserver");
-		return false;
-	}
+		}
 
-	wes = CreateWaitEventSet(TopMemoryContext, 3);
-	AddWaitEventToSet(wes, WL_LATCH_SET, PGINVALID_SOCKET,
-					  MyLatch, NULL);
-	AddWaitEventToSet(wes, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
-					  NULL, NULL);
-	AddWaitEventToSet(wes, WL_SOCKET_READABLE, PQsocket(conn), NULL, NULL);
+		if (PQstatus(shard->conn) == CONNECTION_BAD)
+		{
+			char	   *msg = pchomp(PQerrorMessage(shard->conn));
 
-	PG_TRY();
+			CLEANUP_AND_DISCONNECT(shard);
+
+			ereport(elevel,
+					(errcode(ERRCODE_SQLCLIENT_UNABLE_TO_ESTABLISH_SQLCONNECTION),
+						errmsg(NEON_TAG "[shard %d] could not establish connection to pageserver", shard_no),
+						errdetail_internal("%s", msg)));
+			pfree(msg);
+			return false;
+		}
+
+		ps_send_query_ret = PQsendQuery(shard->conn, pagestream_query);
+		pfree(pagestream_query);
+		if (ps_send_query_ret != 1)
+		{
+			CLEANUP_AND_DISCONNECT(shard);
+
+			neon_shard_log(shard_no, elevel, "could not send pagestream command to pageserver");
+			return false;
+		}
+
+		shard->state = PS_Connecting_PageStream;
+		/* fallthrough */
+	}
+	case PS_Connecting_PageStream:
 	{
-		while (PQisBusy(conn))
+		neon_shard_log(shard_no, DEBUG5, "Connection state: Connecting_PageStream");
+
+		if (PQstatus(shard->conn) == CONNECTION_BAD)
+		{
+			char	   *msg = pchomp(PQerrorMessage(shard->conn));
+			CLEANUP_AND_DISCONNECT(shard);
+			ereport(elevel,
+					(errcode(ERRCODE_SQLCLIENT_UNABLE_TO_ESTABLISH_SQLCONNECTION),
+						errmsg(NEON_TAG "[shard %d] could not establish connection to pageserver", shard_no),
+						errdetail_internal("%s", msg)));
+			pfree(msg);
+			return false;
+		}
+
+		while (PQisBusy(shard->conn))
 		{
 			WaitEvent	event;
 
 			/* Sleep until there's something to do */
-			(void) WaitEventSetWait(wes, -1L, &event, 1, PG_WAIT_EXTENSION);
+			(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1, PG_WAIT_EXTENSION);
 			ResetLatch(MyLatch);
 
 			CHECK_FOR_INTERRUPTS();
@@ -423,40 +599,37 @@ pageserver_connect(shardno_t shard_no, int elevel)
 			/* Data available in socket? */
 			if (event.events & WL_SOCKET_READABLE)
 			{
-				if (!PQconsumeInput(conn))
+				if (!PQconsumeInput(shard->conn))
 				{
-					char	   *msg = pchomp(PQerrorMessage(conn));
-
-					PQfinish(conn);
-					FreeWaitEventSet(wes);
+					char	   *msg = pchomp(PQerrorMessage(shard->conn));
 
+					CLEANUP_AND_DISCONNECT(shard);
 					neon_shard_log(shard_no, elevel, "could not complete handshake with pageserver: %s",
 								   msg);
-					/* Returning from inside PG_TRY is bad, so we break/return later */
-					broke_from_loop = true;
-					break;
+					pfree(msg);
+					return false;
 				}
 			}
 		}
-	}
-	PG_CATCH();
-	{
-		PQfinish(conn);
-		FreeWaitEventSet(wes);
-		PG_RE_THROW();
-	}
-	PG_END_TRY();
 
-	if (broke_from_loop)
-	{
-		return false;
+		shard->state = PS_Connected;
+		/* fallthrough */
 	}
+	case PS_Connected:
+		/*
+		 * We successfully connected. Future connections to this PageServer
+		 * will do fast retries again, with exponential backoff.
+		 */
+		shard->delay_us = MIN_RECONNECT_INTERVAL_USEC;
 
-	neon_shard_log(shard_no, LOG, "libpagestore: connected to '%s' with protocol version %d", connstr, neon_protocol_version);
-	page_servers[shard_no].conn = conn;
-	page_servers[shard_no].wes = wes;
-
-	return true;
+		neon_shard_log(shard_no, DEBUG5, "Connection state: Connected");
+		neon_shard_log(shard_no, LOG, "libpagestore: connected to '%s' with protocol version %d", connstr, neon_protocol_version);
+		return true;
+	default:
+		neon_shard_log(shard_no, ERROR, "libpagestore: invalid connection state %d", shard->state);
+	}
+	/* This shouldn't be hit */
+	Assert(false);
 }
 
 /*
@@ -476,7 +649,7 @@ retry:
 		WaitEvent	event;
 
 		/* Sleep until there's something to do */
-		(void) WaitEventSetWait(page_servers[shard_no].wes, -1L, &event, 1, PG_WAIT_EXTENSION);
+		(void) WaitEventSetWait(page_servers[shard_no].wes_read, -1L, &event, 1, PG_WAIT_EXTENSION);
 		ResetLatch(MyLatch);
 
 		CHECK_FOR_INTERRUPTS();
@@ -502,7 +675,8 @@ retry:
 
 /*
  * Reset prefetch and drop connection to the shard.
- * It also drops connection to all other shards involved in prefetch.
+ * It also drops connection to all other shards involved in prefetch, through
+ * prefetch_on_ps_disconnect().
  */
 static void
 pageserver_disconnect(shardno_t shard_no)
@@ -512,9 +686,6 @@ pageserver_disconnect(shardno_t shard_no)
 	 * whole prefetch queue, even for other pageservers. It should not
 	 * cause big problems, because connection loss is supposed to be a
 	 * rare event.
-	 *
-	 * Prefetch state should be reset even if page_servers[shard_no].conn == NULL,
-	 * because prefetch request may be registered before connection is established.
 	 */
 	prefetch_on_ps_disconnect();
 
@@ -527,37 +698,36 @@ pageserver_disconnect(shardno_t shard_no)
 static void
 pageserver_disconnect_shard(shardno_t shard_no)
 {
+	PageServer *shard = &page_servers[shard_no];
 	/*
 	 * If anything goes wrong while we were sending a request, it's not clear
 	 * what state the connection is in. For example, if we sent the request
 	 * but didn't receive a response yet, we might receive the response some
 	 * time later after we have already sent a new unrelated request. Close
 	 * the connection to avoid getting confused.
+	 * Similarly, even when we're in PS_DISCONNECTED, we may have junk to
+	 * clean up: It is possible that we encountered an error allocating any
+	 * of the wait event sets or the psql connection, or failed when we tried
+	 * to attach wait events to the WaitEventSets.
 	 */
-	if (page_servers[shard_no].conn)
-	{
-		neon_shard_log(shard_no, LOG, "dropping connection to page server due to error");
-		PQfinish(page_servers[shard_no].conn);
-		page_servers[shard_no].conn = NULL;
-	}
-	if (page_servers[shard_no].wes != NULL)
-	{
-		FreeWaitEventSet(page_servers[shard_no].wes);
-		page_servers[shard_no].wes = NULL;
-	}
+	CLEANUP_AND_DISCONNECT(shard);
+
+	shard->state = PS_Disconnected;
 }
 
 static bool
 pageserver_send(shardno_t shard_no, NeonRequest *request)
 {
 	StringInfoData req_buff;
-	PGconn	   *pageserver_conn = page_servers[shard_no].conn;
+	PageServer *shard = &page_servers[shard_no];
+	PGconn	   *pageserver_conn;
 
 	/* If the connection was lost for some reason, reconnect */
-	if (pageserver_conn && PQstatus(pageserver_conn) == CONNECTION_BAD)
+	if (shard->state == PS_Connected && PQstatus(shard->conn) == CONNECTION_BAD)
 	{
 		neon_shard_log(shard_no, LOG, "pageserver_send disconnect bad connection");
 		pageserver_disconnect(shard_no);
+		pageserver_conn = NULL;
 	}
 
 	req_buff = nm_pack_request(request);
@@ -571,17 +741,19 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 	 * https://github.com/neondatabase/neon/issues/1138 So try to reestablish
 	 * connection in case of failure.
 	 */
-	if (!page_servers[shard_no].conn)
+	if (shard->state != PS_Connected)
 	{
-		while (!pageserver_connect(shard_no, n_reconnect_attempts < max_reconnect_attempts ? LOG : ERROR))
+		while (!pageserver_connect(shard_no, shard->n_reconnect_attempts < max_reconnect_attempts ? LOG : ERROR))
 		{
 			HandleMainLoopInterrupts();
-			n_reconnect_attempts += 1;
+			shard->n_reconnect_attempts += 1;
 		}
-		n_reconnect_attempts = 0;
+		shard->n_reconnect_attempts = 0;
+	} else {
+		Assert(shard->conn != NULL);
 	}
 
-	pageserver_conn = page_servers[shard_no].conn;
+	pageserver_conn = shard->conn;
 
 	/*
 	 * Send request.
@@ -590,13 +762,17 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 	 * should use async mode and check for interrupts while waiting. In
 	 * practice, our requests are small enough to always fit in the output and
 	 * TCP buffer.
+	 *
+	 * Note that this also will fail when the connection is in the
+	 * PGRES_POLLING_WRITING state. It's kinda dirty to disconnect at this
+	 * point, but on the grand scheme of things it's only a small issue.
 	 */
 	if (PQputCopyData(pageserver_conn, req_buff.data, req_buff.len) <= 0)
 	{
 		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
 
 		pageserver_disconnect(shard_no);
-		neon_shard_log(shard_no, LOG, "pageserver_send disconnect because failed to send page request (try to reconnect): %s", msg);
+		neon_shard_log(shard_no, LOG, "pageserver_send disconnected: failed to send page request (try to reconnect): %s", msg);
 		pfree(msg);
 		pfree(req_buff.data);
 		return false;
@@ -611,6 +787,7 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 		neon_shard_log(shard_no, PageStoreTrace, "sent request: %s", msg);
 		pfree(msg);
 	}
+
 	return true;
 }
 
@@ -619,58 +796,68 @@ pageserver_receive(shardno_t shard_no)
 {
 	StringInfoData resp_buff;
 	NeonResponse *resp;
-	PGconn	   *pageserver_conn = page_servers[shard_no].conn;
+	PageServer *shard = &page_servers[shard_no];
+	PGconn	   *pageserver_conn = shard->conn;
+	/* read response */
+	int			rc;
 
-	if (!pageserver_conn)
-		return NULL;
-
-	PG_TRY();
+	if (shard->state != PS_Connected)
 	{
-		/* read response */
-		int			rc;
+		neon_shard_log(shard_no, LOG,
+					   "pageserver_receive: returning NULL for non-connected pageserver connection: 0x%02x",
+					   shard->state);
+		return NULL;
+	}
 
-		rc = call_PQgetCopyData(shard_no, &resp_buff.data);
-		if (rc >= 0)
+	Assert(pageserver_conn);
+
+	rc = call_PQgetCopyData(shard_no, &resp_buff.data);
+	if (rc >= 0)
+	{
+		/* call_PQgetCopyData handles rc == 0 */
+		Assert(rc > 0);
+
+		PG_TRY();
 		{
 			resp_buff.len = rc;
 			resp_buff.cursor = 0;
 			resp = nm_unpack_response(&resp_buff);
 			PQfreemem(resp_buff.data);
-
-			if (message_level_is_interesting(PageStoreTrace))
-			{
-				char	   *msg = nm_to_string((NeonMessage *) resp);
-
-				neon_shard_log(shard_no, PageStoreTrace, "got response: %s", msg);
-				pfree(msg);
-			}
 		}
-		else if (rc == -1)
+		PG_CATCH();
 		{
-			neon_shard_log(shard_no, LOG, "pageserver_receive disconnect because call_PQgetCopyData returns -1: %s", pchomp(PQerrorMessage(pageserver_conn)));
+			neon_shard_log(shard_no, LOG, "pageserver_receive: disconnect due malformatted response");
 			pageserver_disconnect(shard_no);
-			resp = NULL;
+			PG_RE_THROW();
 		}
-		else if (rc == -2)
-		{
-			char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
+		PG_END_TRY();
 
-			pageserver_disconnect(shard_no);
-			neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect because could not read COPY data: %s", msg);
-		}
-		else
+		if (message_level_is_interesting(PageStoreTrace))
 		{
-			pageserver_disconnect(shard_no);
-			neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect because unexpected PQgetCopyData return value: %d", rc);
+			char	   *msg = nm_to_string((NeonMessage *) resp);
+
+			neon_shard_log(shard_no, PageStoreTrace, "got response: %s", msg);
+			pfree(msg);
 		}
 	}
-	PG_CATCH();
+	else if (rc == -1)
 	{
-		neon_shard_log(shard_no, LOG, "pageserver_receive disconnect due to caught exception");
+		neon_shard_log(shard_no, LOG, "pageserver_receive disconnect: psql end of copy data: %s", pchomp(PQerrorMessage(pageserver_conn)));
 		pageserver_disconnect(shard_no);
-		PG_RE_THROW();
+		resp = NULL;
+	}
+	else if (rc == -2)
+	{
+		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
+
+		pageserver_disconnect(shard_no);
+		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: could not read COPY data: %s", msg);
+	}
+	else
+	{
+		pageserver_disconnect(shard_no);
+		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc);
 	}
-	PG_END_TRY();
 
 	return (NeonResponse *) resp;
 }
@@ -681,7 +868,7 @@ pageserver_flush(shardno_t shard_no)
 {
 	PGconn	   *pageserver_conn = page_servers[shard_no].conn;
 
-	if (!pageserver_conn)
+	if (page_servers[shard_no].state != PS_Connected)
 	{
 		neon_shard_log(shard_no, WARNING, "Tried to flush while disconnected");
 	}
@@ -697,6 +884,7 @@ pageserver_flush(shardno_t shard_no)
 			return false;
 		}
 	}
+
 	return true;
 }
 
@@ -860,7 +1048,7 @@ pg_init_libpagestore(void)
 							"Version of compute<->page server protocol",
 							NULL,
 							&neon_protocol_version,
-							1, /* default to old protocol for now */
+							2, /* use protocol version 2 */
 							1, /* min */
 							2, /* max */
 							PGC_SU_BACKEND,
@@ -891,5 +1079,7 @@ pg_init_libpagestore(void)
 		dbsize_hook = neon_dbsize;
 	}
 
+	memset(page_servers, 0, sizeof(page_servers));
+
 	lfc_init();
 }

pgxn/neon/pagestore_client.h

@@ -237,18 +237,50 @@ extern void neon_zeroextend(SMgrRelation reln, ForkNumber forknum,
 extern bool neon_prefetch(SMgrRelation reln, ForkNumber forknum,
 						  BlockNumber blocknum);
 
+/*
+ * LSN values associated with each request to the pageserver
+ */
+typedef struct
+{
+	/*
+	 * 'request_lsn' is the main value that determines which page version to
+	 * fetch.
+	 */
+	XLogRecPtr request_lsn;
+
+	/*
+	 * A hint to the pageserver that the requested page hasn't been modified
+	 * between this LSN and 'request_lsn'. That allows the pageserver to
+	 * return the page faster, without waiting for 'request_lsn' to arrive in
+	 * the pageserver, as long as 'not_modified_since' has arrived.
+	 */
+	XLogRecPtr not_modified_since;
+
+	/*
+	 * 'effective_request_lsn' is not included in the request that's sent to
+	 * the pageserver, but is used to keep track of the latest LSN of when the
+	 * request was made. In a standby server, this is always the same as the
+	 * 'request_lsn', but in the primary we use UINT64_MAX as the
+	 * 'request_lsn' to request the latest page version, so we need this
+	 * separate field to remember that latest LSN was when the request was
+	 * made. It's needed to manage prefetch request, to verify if the response
+	 * to a prefetched request is still valid.
+	 */
+	XLogRecPtr effective_request_lsn;
+} neon_request_lsns;
+
 #if PG_MAJORVERSION_NUM < 16
 extern void neon_read(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 					  char *buffer);
 extern PGDLLEXPORT void neon_read_at_lsn(NRelFileInfo rnode, ForkNumber forkNum, BlockNumber blkno,
-										 XLogRecPtr request_lsn, XLogRecPtr not_modified_since, char *buffer);
+										 neon_request_lsns request_lsns, char *buffer);
 extern void neon_write(SMgrRelation reln, ForkNumber forknum,
 					   BlockNumber blocknum, char *buffer, bool skipFsync);
 #else
 extern void neon_read(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 					  void *buffer);
 extern PGDLLEXPORT void neon_read_at_lsn(NRelFileInfo rnode, ForkNumber forkNum, BlockNumber blkno,
-										 XLogRecPtr request_lsn, XLogRecPtr not_modified_since, void *buffer);
+										 neon_request_lsns request_lsns, void *buffer);
 extern void neon_write(SMgrRelation reln, ForkNumber forknum,
 					   BlockNumber blocknum, const void *buffer, bool skipFsync);
 #endif

pgxn/neon/pagestore_smgr.c

@@ -45,6 +45,7 @@
  */
 #include "postgres.h"
 
+#include "access/parallel.h"
 #include "access/xact.h"
 #include "access/xlog.h"
 #include "access/xlogdefs.h"
@@ -93,6 +94,10 @@ static char *hexdump_page(char *page);
 
 const int	SmgrTrace = DEBUG5;
 
+#define NEON_PANIC_CONNECTION_STATE(shard_no, elvl, message, ...) \
+	neon_shard_log(shard_no, elvl, "Broken connection state: " message, \
+				   ##__VA_ARGS__)
+
 page_server_api *page_server;
 
 /* unlogged relation build states */
@@ -168,8 +173,7 @@ typedef enum PrefetchStatus
 typedef struct PrefetchRequest
 {
 	BufferTag	buftag;			/* must be first entry in the struct */
-	XLogRecPtr	request_lsn;
-	XLogRecPtr	not_modified_since;
+	neon_request_lsns request_lsns;
 	NeonResponse *response;		/* may be null */
 	PrefetchStatus status;
 	shardno_t   shard_no;
@@ -271,16 +275,15 @@ static PrefetchState *MyPState;
 
 static bool compact_prefetch_buffers(void);
 static void consume_prefetch_responses(void);
-static uint64 prefetch_register_buffer(BufferTag tag, XLogRecPtr *force_request_lsn, XLogRecPtr *force_not_modified_since);
+static uint64 prefetch_register_buffer(BufferTag tag, neon_request_lsns *force_request_lsns);
 static bool prefetch_read(PrefetchRequest *slot);
-static void prefetch_do_request(PrefetchRequest *slot, XLogRecPtr *force_request_lsn, XLogRecPtr *force_not_modified_since);
+static void prefetch_do_request(PrefetchRequest *slot, neon_request_lsns *force_request_lsns);
 static bool prefetch_wait_for(uint64 ring_index);
 static void prefetch_cleanup_trailing_unused(void);
 static inline void prefetch_set_unused(uint64 ring_index);
 
-static void neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
-								 XLogRecPtr *request_lsn, XLogRecPtr *not_modified_since);
-static bool neon_prefetch_response_usable(XLogRecPtr request_lsn, XLogRecPtr not_modified_since,
+static neon_request_lsns neon_get_request_lsns(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno);
+static bool neon_prefetch_response_usable(neon_request_lsns request_lsns,
 										  PrefetchRequest *slot);
 
 static bool
@@ -338,8 +341,7 @@ compact_prefetch_buffers(void)
 		target_slot->shard_no = source_slot->shard_no;
 		target_slot->status = source_slot->status;
 		target_slot->response = source_slot->response;
-		target_slot->request_lsn = source_slot->request_lsn;
-		target_slot->not_modified_since = source_slot->not_modified_since;
+		target_slot->request_lsns = source_slot->request_lsns;
 		target_slot->my_ring_index = empty_ring_index;
 
 		prfh_delete(MyPState->prf_hash, source_slot);
@@ -358,8 +360,9 @@ compact_prefetch_buffers(void)
 		};
 		source_slot->response = NULL;
 		source_slot->my_ring_index = 0;
-		source_slot->request_lsn = InvalidXLogRecPtr;
-		source_slot->not_modified_since = InvalidXLogRecPtr;
+		source_slot->request_lsns = (neon_request_lsns) {
+			InvalidXLogRecPtr, InvalidXLogRecPtr, InvalidXLogRecPtr
+		};
 
 		/* update bookkeeping */
 		n_moved++;
@@ -527,6 +530,8 @@ prefetch_flush_requests(void)
  *
  * NOTE: this function may indirectly update MyPState->pfs_hash; which
  * invalidates any active pointers into the hash table.
+ * NOTE: callers should make sure they can handle query cancellations in this
+ * function's call path.
  */
 static bool
 prefetch_wait_for(uint64 ring_index)
@@ -562,6 +567,8 @@ prefetch_wait_for(uint64 ring_index)
  *
  * NOTE: this function may indirectly update MyPState->pfs_hash; which
  * invalidates any active pointers into the hash table.
+ *
+ * NOTE: this does IO, and can get canceled out-of-line.
  */
 static bool
 prefetch_read(PrefetchRequest *slot)
@@ -573,6 +580,14 @@ prefetch_read(PrefetchRequest *slot)
 	Assert(slot->response == NULL);
 	Assert(slot->my_ring_index == MyPState->ring_receive);
 
+	if (slot->status != PRFS_REQUESTED ||
+		slot->response != NULL ||
+		slot->my_ring_index != MyPState->ring_receive)
+		neon_shard_log(slot->shard_no, ERROR,
+					   "Incorrect prefetch read: status=%d response=%p my=%lu receive=%lu",
+					   slot->status, slot->response,
+					   (long)slot->my_ring_index, (long)MyPState->ring_receive);
+
 	old = MemoryContextSwitchTo(MyPState->errctx);
 	response = (NeonResponse *) page_server->receive(slot->shard_no);
 	MemoryContextSwitchTo(old);
@@ -590,6 +605,11 @@ prefetch_read(PrefetchRequest *slot)
 	}
 	else
 	{
+		neon_shard_log(slot->shard_no, WARNING,
+					   "No response from reading prefetch entry %lu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
+					   (long)slot->my_ring_index,
+					   RelFileInfoFmt(BufTagGetNRelFileInfo(slot->buftag)),
+					   slot->buftag.forkNum, slot->buftag.blockNum);
 		return false;
 	}
 }
@@ -604,6 +624,7 @@ void
 prefetch_on_ps_disconnect(void)
 {
 	MyPState->ring_flush = MyPState->ring_unused;
+
 	while (MyPState->ring_receive < MyPState->ring_unused)
 	{
 		PrefetchRequest *slot;
@@ -626,6 +647,7 @@ prefetch_on_ps_disconnect(void)
 		slot->status = PRFS_TAG_REMAINS;
 		MyPState->n_requests_inflight -= 1;
 		MyPState->ring_receive += 1;
+
 		prefetch_set_unused(ring_index);
 	}
 }
@@ -689,9 +711,11 @@ prefetch_set_unused(uint64 ring_index)
  * prefetch_wait_for().
  */
 static void
-prefetch_do_request(PrefetchRequest *slot, XLogRecPtr *force_request_lsn, XLogRecPtr *force_not_modified_since)
+prefetch_do_request(PrefetchRequest *slot, neon_request_lsns *force_request_lsns)
 {
 	bool		found;
+	uint64		mySlotNo = slot->my_ring_index;
+
 	NeonGetPageRequest request = {
 		.req.tag = T_NeonGetPageRequest,
 		/* lsn and not_modified_since are filled in below */
@@ -700,28 +724,25 @@ prefetch_do_request(PrefetchRequest *slot, XLogRecPtr *force_request_lsn, XLogRe
 		.blkno = slot->buftag.blockNum,
 	};
 
-	Assert(((force_request_lsn != NULL) == (force_not_modified_since != NULL)));
+	Assert(mySlotNo == MyPState->ring_unused);
 
-	if (force_request_lsn)
-	{
-		request.req.lsn = *force_request_lsn;
-		request.req.not_modified_since = *force_not_modified_since;
-	}
+	if (force_request_lsns)
+		slot->request_lsns = *force_request_lsns;
 	else
-	{
-		neon_get_request_lsn(BufTagGetNRelFileInfo(slot->buftag),
-							 slot->buftag.forkNum,
-							 slot->buftag.blockNum,
-							 &request.req.lsn,
-							 &request.req.not_modified_since);
-	}
-	slot->request_lsn = request.req.lsn;
-	slot->not_modified_since = request.req.not_modified_since;
+		slot->request_lsns = neon_get_request_lsns(BufTagGetNRelFileInfo(slot->buftag),
+												   slot->buftag.forkNum,
+												   slot->buftag.blockNum);
+	request.req.lsn = slot->request_lsns.request_lsn;
+	request.req.not_modified_since = slot->request_lsns.not_modified_since;
 
 	Assert(slot->response == NULL);
 	Assert(slot->my_ring_index == MyPState->ring_unused);
 
-	while (!page_server->send(slot->shard_no, (NeonRequest *) &request));
+	while (!page_server->send(slot->shard_no, (NeonRequest *) &request))
+	{
+		Assert(mySlotNo == MyPState->ring_unused);
+		/* loop */
+	}
 
 	/* update prefetch state */
 	MyPState->n_requests_inflight += 1;
@@ -732,7 +753,6 @@ prefetch_do_request(PrefetchRequest *slot, XLogRecPtr *force_request_lsn, XLogRe
 
 	/* update slot state */
 	slot->status = PRFS_REQUESTED;
-
 	prfh_insert(MyPState->prf_hash, slot, &found);
 	Assert(!found);
 }
@@ -742,25 +762,22 @@ prefetch_do_request(PrefetchRequest *slot, XLogRecPtr *force_request_lsn, XLogRe
  *
  * Register that we may want the contents of BufferTag in the near future.
  *
- * If force_request_lsn and force_not_modified_since are not NULL, those
- * values are sent to the pageserver. If they are NULL, we utilize the
- * lastWrittenLsn -infrastructure to fill them in.
+ * If force_request_lsns is not NULL, those values are sent to the
+ * pageserver. If NULL, we utilize the lastWrittenLsn -infrastructure
+ * to calculate the LSNs to send.
  *
  * NOTE: this function may indirectly update MyPState->pfs_hash; which
  * invalidates any active pointers into the hash table.
  */
 
 static uint64
-prefetch_register_buffer(BufferTag tag, XLogRecPtr *force_request_lsn,
-						 XLogRecPtr *force_not_modified_since)
+prefetch_register_buffer(BufferTag tag, neon_request_lsns *force_request_lsns)
 {
 	uint64		ring_index;
 	PrefetchRequest req;
 	PrefetchRequest *slot;
 	PrfHashEntry *entry;
 
-	Assert(((force_request_lsn != NULL) == (force_not_modified_since != NULL)));
-
 	/* use an intermediate PrefetchRequest struct to ensure correct alignment */
 	req.buftag = tag;
 Retry:
@@ -781,10 +798,9 @@ Retry:
 		 * If the caller specified a request LSN to use, only accept prefetch
 		 * responses that satisfy that request.
 		 */
-		if (force_request_lsn)
+		if (force_request_lsns)
 		{
-			if (!neon_prefetch_response_usable(*force_request_lsn,
-											   *force_not_modified_since, slot))
+			if (!neon_prefetch_response_usable(*force_request_lsns, slot))
 			{
 				/* Wait for the old request to finish and discard it */
 				if (!prefetch_wait_for(ring_index))
@@ -886,7 +902,7 @@ Retry:
 	slot->shard_no = get_shard_number(&tag);
 	slot->my_ring_index = ring_index;
 
-	prefetch_do_request(slot, force_request_lsn, force_not_modified_since);
+	prefetch_do_request(slot, force_request_lsns);
 	Assert(slot->status == PRFS_REQUESTED);
 	Assert(MyPState->ring_last <= ring_index &&
 		   ring_index < MyPState->ring_unused);
@@ -908,6 +924,10 @@ Retry:
 	return ring_index;
 }
 
+/*
+ * Note: this function can get canceled and use a long jump to the next catch
+ * context. Take care.
+ */
 static NeonResponse *
 page_server_request(void const *req)
 {
@@ -939,19 +959,38 @@ page_server_request(void const *req)
 	 * Current sharding model assumes that all metadata is present only at shard 0.
 	 * We still need to call get_shard_no() to check if shard map is up-to-date.
 	 */
-	if (((NeonRequest *) req)->tag != T_NeonGetPageRequest || ((NeonGetPageRequest *) req)->forknum != MAIN_FORKNUM)
+	if (((NeonRequest *) req)->tag != T_NeonGetPageRequest ||
+		((NeonGetPageRequest *) req)->forknum != MAIN_FORKNUM)
 	{
 		shard_no = 0;
 	}
 
 	do
 	{
-		while (!page_server->send(shard_no, (NeonRequest *) req) || !page_server->flush(shard_no));
-		consume_prefetch_responses();
-		resp = page_server->receive(shard_no);
-	} while (resp == NULL);
-	return resp;
+		PG_TRY();
+		{
+			while (!page_server->send(shard_no, (NeonRequest *) req)
+				   || !page_server->flush(shard_no))
+			{
+				/* do nothing */
+			}
+			consume_prefetch_responses();
+			resp = page_server->receive(shard_no);
+		}
+		PG_CATCH();
+		{
+			/*
+			 * Cancellation in this code needs to be handled better at some
+			 * point, but this currently seems fine for now.
+			 */
+			page_server->disconnect(shard_no);
+			PG_RE_THROW();
+		}
+		PG_END_TRY();
 
+	} while (resp == NULL);
+
+	return resp;
 }
 
 
@@ -1363,6 +1402,10 @@ PageIsEmptyHeapPage(char *buffer)
 	return memcmp(buffer, empty_page.data, BLCKSZ) == 0;
 }
 
+/*
+ * A page is being evicted from the shared buffer cache. Update the
+ * last-written LSN of the page, and WAL-log it if needed.
+ */
 static void
 #if PG_MAJORVERSION_NUM < 16
 neon_wallog_page(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, char *buffer, bool force)
@@ -1371,12 +1414,7 @@ neon_wallog_page(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, co
 #endif
 {
 	XLogRecPtr	lsn = PageGetLSN((Page) buffer);
-
-	if (ShutdownRequestPending)
-		return;
-	/* Don't log any pages if we're not allowed to do so. */
-	if (!XLogInsertAllowed())
-		return;
+	bool		log_page;
 
 	/*
 	 * Whenever a VM or FSM page is evicted, WAL-log it. FSM and (some) VM
@@ -1385,9 +1423,21 @@ neon_wallog_page(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, co
 	 * correctness, the non-logged updates are not critical. But we want to
 	 * have a reasonably up-to-date VM and FSM in the page server.
 	 */
-	if ((force || forknum == FSM_FORKNUM || forknum == VISIBILITYMAP_FORKNUM) && !RecoveryInProgress())
+	log_page = false;
+	if (force)
+	{
+		Assert(XLogInsertAllowed());
+		log_page = true;
+	}
+	else if (XLogInsertAllowed() &&
+			 !ShutdownRequestPending &&
+			 (forknum == FSM_FORKNUM || forknum == VISIBILITYMAP_FORKNUM))
+	{
+		log_page = true;
+	}
+
+	if (log_page)
 	{
-		/* FSM is never WAL-logged and we don't care. */
 		XLogRecPtr	recptr;
 
 		recptr = log_newpage_copy(&InfoFromSMgrRel(reln), forknum, blocknum,
@@ -1400,7 +1450,8 @@ neon_wallog_page(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, co
 						RelFileInfoFmt(InfoFromSMgrRel(reln)),
 						forknum, LSN_FORMAT_ARGS(lsn))));
 	}
-	else if (lsn == InvalidXLogRecPtr)
+
+	if (lsn == InvalidXLogRecPtr)
 	{
 		/*
 		 * When PostgreSQL extends a relation, it calls smgrextend() with an
@@ -1436,19 +1487,31 @@ neon_wallog_page(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, co
 							RelFileInfoFmt(InfoFromSMgrRel(reln)),
 							forknum)));
 		}
-		else
+		else if (forknum != FSM_FORKNUM && forknum != VISIBILITYMAP_FORKNUM)
 		{
-			ereport(PANIC,
+			/*
+			 * Its a bad sign if there is a page with zero LSN in the buffer
+			 * cache in a standby, too. However, PANICing seems like a cure
+			 * worse than the disease, as the damage has likely already been
+			 * done in the primary. So in a standby, make this an assertion,
+			 * and in a release build just LOG the error and soldier on. We
+			 * update the last-written LSN of the page with a conservative
+			 * value in that case, which is the last replayed LSN.
+			 */
+			ereport(RecoveryInProgress() ? LOG : PANIC,
 					(errmsg(NEON_TAG "Page %u of relation %u/%u/%u.%u is evicted with zero LSN",
 							blocknum,
 							RelFileInfoFmt(InfoFromSMgrRel(reln)),
 							forknum)));
+			Assert(false);
+
+			lsn = GetXLogReplayRecPtr(NULL); /* in standby mode, soldier on */
 		}
 	}
 	else
 	{
 		ereport(SmgrTrace,
-				(errmsg(NEON_TAG "Page %u of relation %u/%u/%u.%u is already wal logged at lsn=%X/%X",
+				(errmsg(NEON_TAG "Evicting page %u of relation %u/%u/%u.%u with lsn=%X/%X",
 						blocknum,
 						RelFileInfoFmt(InfoFromSMgrRel(reln)),
 						forknum, LSN_FORMAT_ARGS(lsn))));
@@ -1529,11 +1592,11 @@ nm_adjust_lsn(XLogRecPtr lsn)
 /*
  * Return LSN for requesting pages and number of blocks from page server
  */
-static void
-neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
-					 XLogRecPtr *request_lsn, XLogRecPtr *not_modified_since)
+static neon_request_lsns
+neon_get_request_lsns(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno)
 {
 	XLogRecPtr	last_written_lsn;
+	neon_request_lsns result;
 
 	last_written_lsn = GetLastWrittenLSN(rinfo, forknum, blkno);
 	last_written_lsn = nm_adjust_lsn(last_written_lsn);
@@ -1541,13 +1604,98 @@ neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
 
 	if (RecoveryInProgress())
 	{
-		/* Request the page at the last replayed LSN. */
-		*request_lsn = GetXLogReplayRecPtr(NULL);
-		*not_modified_since = last_written_lsn;
-		Assert(last_written_lsn <= *request_lsn);
+		/*---
+		 * In broad strokes, a replica always requests the page at the current
+		 * replay LSN. But looking closer, what exactly is the replay LSN? Is
+		 * it the last replayed record, or the record being replayed? And does
+		 * the startup process performing the replay need to do something
+		 * differently than backends running queries? Let's take a closer look
+		 * at the different scenarios:
+		 *
+		 * 1. Startup process reads a page, last_written_lsn is old.
+		 *
+		 * Read the old version of the page. We will apply the WAL record on
+		 * it to bring it up-to-date.
+		 *
+		 * We could read the new version, with the changes from this WAL
+		 * record already applied, to offload the work of replaying the record
+		 * to the pageserver. The pageserver might not have received the WAL
+		 * record yet, though, so a read of the old page version and applying
+		 * the record ourselves is likely faster. Also, the redo function
+		 * might be surprised if the changes have already applied. That's
+		 * normal during crash recovery, but not in hot standby.
+		 *
+		 * 2. Startup process reads a page, last_written_lsn == record we're
+		 *    replaying.
+		 *
+		 * Can this happen? There are a few theoretical cases when it might:
+		 *
+		 * A) The redo function reads the same page twice. We had already read
+		 *    and applied the changes once, and now we're reading it for the
+		 *    second time.  That would be a rather silly thing for a redo
+		 *    function to do, and I'm not aware of any that would do it.
+		 *
+		 * B) The redo function modifies multiple pages, and it already
+		 *    applied the changes to one of the pages, released the lock on
+		 *    it, and is now reading a second page.  Furthermore, the first
+		 *    page was already evicted from the buffer cache, and also from
+		 *    the last-written LSN cache, so that the per-relation or global
+		 *    last-written LSN was already updated. All the WAL redo functions
+		 *    hold the locks on pages that they modify, until all the changes
+		 *    have been modified (?), which would make that impossible.
+		 *    However, we skip the locking, if the page isn't currently in the
+		 *    page cache (see neon_redo_read_buffer_filter below).
+		 *
+		 * Even if the one of the above cases were possible in theory, they
+		 * would also require the pages being modified by the redo function to
+		 * be immediately evicted from the page cache.
+		 *
+		 * So this probably does not happen in practice. But if it does, we
+		 * request the new version, including the changes from the record
+		 * being replayed. That seems like the correct behavior in any case.
+		 *
+		 * 3. Backend process reads a page with old last-written LSN
+		 *
+		 * Nothing special here. Read the old version.
+		 *
+		 * 4. Backend process reads a page with last_written_lsn == record being replayed
+		 *
+		 * This can happen, if the redo function has started to run, and saw
+		 * that the page isn't present in the page cache (see
+		 * neon_redo_read_buffer_filter below).  Normally, in a normal
+		 * Postgres server, the redo function would hold a lock on the page,
+		 * so we would get blocked waiting the redo function to release the
+		 * lock. To emulate that, wait for the WAL replay of the record to
+		 * finish.
+		 */
+		/* Request the page at the end of the last fully replayed LSN. */
+		XLogRecPtr replay_lsn = GetXLogReplayRecPtr(NULL);
 
-		neon_log(DEBUG1, "neon_get_request_lsn request lsn %X/%X, not_modified_since %X/%X",
-				 LSN_FORMAT_ARGS(*request_lsn), LSN_FORMAT_ARGS(*not_modified_since));
+		if (last_written_lsn > replay_lsn)
+		{
+			/* GetCurrentReplayRecPtr was introduced in v15 */
+#if PG_VERSION_NUM >= 150000
+			Assert(last_written_lsn == GetCurrentReplayRecPtr(NULL));
+#endif
+
+			/*
+			 * Cases 2 and 4. If this is a backend (case 4), the
+			 * neon_read_at_lsn() call later will wait for the WAL record to be
+			 * fully replayed.
+			 */
+			result.request_lsn = last_written_lsn;
+		}
+		else
+		{
+			/* cases 1 and 3 */
+			result.request_lsn = replay_lsn;
+		}
+		result.not_modified_since = last_written_lsn;
+		result.effective_request_lsn = result.request_lsn;
+		Assert(last_written_lsn <= result.request_lsn);
+
+		neon_log(DEBUG1, "neon_get_request_lsns request lsn %X/%X, not_modified_since %X/%X",
+				 LSN_FORMAT_ARGS(result.request_lsn), LSN_FORMAT_ARGS(result.not_modified_since));
 	}
 	else
 	{
@@ -1559,7 +1707,7 @@ neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
 		 * must still in the buffer cache, so our request cannot concern
 		 * those.
 		 */
-		neon_log(DEBUG1, "neon_get_request_lsn GetLastWrittenLSN lsn %X/%X ",
+		neon_log(DEBUG1, "neon_get_request_lsns GetLastWrittenLSN lsn %X/%X",
 				 LSN_FORMAT_ARGS(last_written_lsn));
 
 		/*
@@ -1585,16 +1733,33 @@ neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
 		}
 
 		/*
-		 * Request the latest version of the page. The most up-to-date request
-		 * LSN we could use would be the current insert LSN, but to avoid the
-		 * overhead of looking it up, use 'flushlsn' instead. This relies on
-		 * the assumption that if the page was modified since the last WAL
-		 * flush, it should still be in the buffer cache, and we wouldn't be
-		 * requesting it.
+		 * Request the very latest version of the page. In principle we
+		 * want to read the page at the current insert LSN, and we could
+		 * use that value in the request. However, there's a corner case
+		 * with pageserver's garbage collection. If the GC horizon is
+		 * set to a very small value, it's possible that by the time
+		 * that the pageserver processes our request, the GC horizon has
+		 * already moved past the LSN we calculate here. Standby servers
+		 * always have that problem as the can always lag behind the
+		 * primary, but for the primary we can avoid it by always
+		 * requesting the latest page, by setting request LSN to
+		 * UINT64_MAX.
+		 *
+		 * Remember the current LSN, however, so that we can later
+		 * correctly determine if the response to the request is still
+		 * valid. The most up-to-date LSN we could use for that purpose
+		 * would be the current insert LSN, but to avoid the overhead of
+		 * looking it up, use 'flushlsn' instead. This relies on the
+		 * assumption that if the page was modified since the last WAL
+		 * flush, it should still be in the buffer cache, and we
+		 * wouldn't be requesting it.
 		 */
-		*request_lsn = flushlsn;
-		*not_modified_since = last_written_lsn;
+		result.request_lsn = UINT64_MAX;
+		result.not_modified_since = last_written_lsn;
+		result.effective_request_lsn = flushlsn;
 	}
+
+	return result;
 }
 
 /*
@@ -1604,12 +1769,16 @@ neon_get_request_lsn(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
  * satisfy a page read now.
  */
 static bool
-neon_prefetch_response_usable(XLogRecPtr request_lsn, XLogRecPtr not_modified_since,
+neon_prefetch_response_usable(neon_request_lsns request_lsns,
 							  PrefetchRequest *slot)
 {
 	/* sanity check the LSN's on the old and the new request */
-	Assert(request_lsn >= not_modified_since);
-	Assert(slot->request_lsn >= slot->not_modified_since);
+	Assert(request_lsns.request_lsn >= request_lsns.not_modified_since);
+	Assert(request_lsns.effective_request_lsn >= request_lsns.not_modified_since);
+	Assert(request_lsns.effective_request_lsn <= request_lsns.request_lsn);
+	Assert(slot->request_lsns.request_lsn >= slot->request_lsns.not_modified_since);
+	Assert(slot->request_lsns.effective_request_lsn >= slot->request_lsns.not_modified_since);
+	Assert(slot->request_lsns.effective_request_lsn <= slot->request_lsns.request_lsn);
 	Assert(slot->status != PRFS_UNUSED);
 
 	/*
@@ -1627,26 +1796,40 @@ neon_prefetch_response_usable(XLogRecPtr request_lsn, XLogRecPtr not_modified_si
 	 * calculate LSNs "out of order" with each other, but the prefetch queue
 	 * is backend-private at the moment.)
 	 */
-	if (request_lsn < slot->request_lsn || not_modified_since < slot->not_modified_since)
+	if (request_lsns.effective_request_lsn < slot->request_lsns.effective_request_lsn ||
+		request_lsns.not_modified_since < slot->request_lsns.not_modified_since)
 	{
 		ereport(LOG,
 				(errcode(ERRCODE_IO_ERROR),
 				 errmsg(NEON_TAG "request with unexpected LSN after prefetch"),
 				 errdetail("Request %X/%X not_modified_since %X/%X, prefetch %X/%X not_modified_since %X/%X)",
-						   LSN_FORMAT_ARGS(request_lsn), LSN_FORMAT_ARGS(not_modified_since),
-						   LSN_FORMAT_ARGS(slot->request_lsn), LSN_FORMAT_ARGS(slot->not_modified_since))));
+						   LSN_FORMAT_ARGS(request_lsns.effective_request_lsn),
+						   LSN_FORMAT_ARGS(request_lsns.not_modified_since),
+						   LSN_FORMAT_ARGS(slot->request_lsns.effective_request_lsn),
+						   LSN_FORMAT_ARGS(slot->request_lsns.not_modified_since))));
 		return false;
 	}
 
 	/*---
-	 * Each request to the pageserver carries two LSN values:
-	 * `not_modified_since` and `request_lsn`. The (not_modified_since,
-	 * request_lsn] range of each request is effectively a claim that the page
-	 * has not been modified between those LSNs.  If the range of the old
-	 * request in the queue overlaps with the new request, we know that the
-	 * page hasn't been modified in the union of the ranges. We can use the
-	 * response to old request to satisfy the new request in that case. For
-	 * example:
+	 * Each request to the pageserver has three LSN values associated with it:
+	 * `not_modified_since`, `request_lsn`, and 'effective_request_lsn'.
+	 * `not_modified_since` and `request_lsn` are sent to the pageserver, but
+	 * in the primary node, we always use UINT64_MAX as the `request_lsn`, so
+	 * we remember `effective_request_lsn` separately. In a primary,
+	 * `effective_request_lsn` is the last flush WAL position when the request
+	 * was sent to the pageserver. That's logically the LSN that we are
+	 * requesting the page at, but we send UINT64_MAX to the pageserver so
+	 * that if the GC horizon advances past that position, we still get a
+	 * valid response instead of an error.
+	 *
+	 * To determine whether a response to a GetPage request issued earlier is
+	 * still valid to satisfy a new page read, we look at the
+	 * (not_modified_since, effective_request_lsn] range of the request. It is
+	 * effectively a claim that the page has not been modified between those
+	 * LSNs.  If the range of the old request in the queue overlaps with the
+	 * new request, we know that the page hasn't been modified in the union of
+	 * the ranges. We can use the response to old request to satisfy the new
+	 * request in that case. For example:
 	 *
 	 *              100      500
 	 * Old request:  +--------+
@@ -1675,9 +1858,9 @@ neon_prefetch_response_usable(XLogRecPtr request_lsn, XLogRecPtr not_modified_si
 	 */
 
 	/* this follows from the checks above */
-	Assert(request_lsn >= slot->not_modified_since);
+	Assert(request_lsns.effective_request_lsn >= slot->request_lsns.not_modified_since);
 
-	return not_modified_since <= slot->request_lsn;
+	return request_lsns.not_modified_since <= slot->request_lsns.effective_request_lsn;
 }
 
 /*
@@ -1689,8 +1872,7 @@ neon_exists(SMgrRelation reln, ForkNumber forkNum)
 	bool		exists;
 	NeonResponse *resp;
 	BlockNumber n_blocks;
-	XLogRecPtr	request_lsn;
-	XLogRecPtr	not_modified_since;
+	neon_request_lsns request_lsns;
 
 	switch (reln->smgr_relpersistence)
 	{
@@ -1745,15 +1927,15 @@ neon_exists(SMgrRelation reln, ForkNumber forkNum)
 		return false;
 	}
 
-	neon_get_request_lsn(InfoFromSMgrRel(reln), forkNum, REL_METADATA_PSEUDO_BLOCKNO,
-						 &request_lsn, &not_modified_since);
+	request_lsns = neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, REL_METADATA_PSEUDO_BLOCKNO);
 	{
 		NeonExistsRequest request = {
 			.req.tag = T_NeonExistsRequest,
-			.req.lsn = request_lsn,
-			.req.not_modified_since = not_modified_since,
+			.req.lsn = request_lsns.request_lsn,
+			.req.not_modified_since = request_lsns.not_modified_since,
 			.rinfo = InfoFromSMgrRel(reln),
-		.forknum = forkNum};
+			.forknum = forkNum
+		};
 
 		resp = page_server_request(&request);
 	}
@@ -1770,13 +1952,15 @@ neon_exists(SMgrRelation reln, ForkNumber forkNum)
 					 errmsg(NEON_TAG "could not read relation existence of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 							RelFileInfoFmt(InfoFromSMgrRel(reln)),
 							forkNum,
-							(uint32) (request_lsn >> 32), (uint32) request_lsn),
+							LSN_FORMAT_ARGS(request_lsns.effective_request_lsn)),
 					 errdetail("page server returned error: %s",
 							   ((NeonErrorResponse *) resp)->message)));
 			break;
 
 		default:
-			neon_log(ERROR, "unexpected response from page server with tag 0x%02x in neon_exists", resp->tag);
+			NEON_PANIC_CONNECTION_STATE(-1, PANIC,
+										"Expected Exists (0x%02x) or Error (0x%02x) response to ExistsRequest, but got 0x%02x",
+										T_NeonExistsResponse, T_NeonErrorResponse, resp->tag);
 	}
 	pfree(resp);
 	return exists;
@@ -2135,7 +2319,7 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
 
 	CopyNRelFileInfoToBufTag(tag, InfoFromSMgrRel(reln));
 
-	ring_index = prefetch_register_buffer(tag, NULL, NULL);
+	ring_index = prefetch_register_buffer(tag, NULL);
 
 	Assert(ring_index < MyPState->ring_unused &&
 		   MyPState->ring_last <= ring_index);
@@ -2188,10 +2372,10 @@ neon_writeback(SMgrRelation reln, ForkNumber forknum,
 void
 #if PG_MAJORVERSION_NUM < 16
 neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
-				 XLogRecPtr request_lsn, XLogRecPtr not_modified_since, char *buffer)
+				 neon_request_lsns request_lsns, char *buffer)
 #else
 neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
-				 XLogRecPtr request_lsn, XLogRecPtr not_modified_since, void *buffer)
+				 neon_request_lsns request_lsns, void *buffer)
 #endif
 {
 	NeonResponse *resp;
@@ -2223,18 +2407,18 @@ neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 	 * value of the LwLsn cache when the entry is not found.
 	 */
 	if (RecoveryInProgress() && !(MyBackendType == B_STARTUP))
-		XLogWaitForReplayOf(request_lsn);
+		XLogWaitForReplayOf(request_lsns.request_lsn);
 
 	/*
 	 * Try to find prefetched page in the list of received pages.
 	 */
-  Retry:
+Retry:
 	entry = prfh_lookup(MyPState->prf_hash, (PrefetchRequest *) &buftag);
 
 	if (entry != NULL)
 	{
 		slot = entry->slot;
-		if (neon_prefetch_response_usable(request_lsn, not_modified_since, slot))
+		if (neon_prefetch_response_usable(request_lsns, slot))
 		{
 			ring_index = slot->my_ring_index;
 			pgBufferUsage.prefetch.hits += 1;
@@ -2268,8 +2452,7 @@ neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		{
 			pgBufferUsage.prefetch.misses += 1;
 
-			ring_index = prefetch_register_buffer(buftag, &request_lsn,
-												  &not_modified_since);
+			ring_index = prefetch_register_buffer(buftag, &request_lsns);
 			slot = GetPrfSlot(ring_index);
 		}
 		else
@@ -2310,12 +2493,14 @@ neon_read_at_lsn(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 							slot->shard_no, blkno,
 							RelFileInfoFmt(rinfo),
 							forkNum,
-							(uint32) (request_lsn >> 32), (uint32) request_lsn),
+							LSN_FORMAT_ARGS(request_lsns.effective_request_lsn)),
 					 errdetail("page server returned error: %s",
 							   ((NeonErrorResponse *) resp)->message)));
 			break;
 		default:
-			neon_log(ERROR, "unexpected response from page server with tag 0x%02x in neon_read_at_lsn", resp->tag);
+			NEON_PANIC_CONNECTION_STATE(slot->shard_no, PANIC,
+										"Expected GetPage (0x%02x) or Error (0x%02x) response to GetPageRequest, but got 0x%02x",
+										T_NeonGetPageResponse, T_NeonErrorResponse, resp->tag);
 	}
 
 	/* buffer was used, clean up for later reuse */
@@ -2333,8 +2518,7 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, char *buffer
 neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer)
 #endif
 {
-	XLogRecPtr	request_lsn;
-	XLogRecPtr	not_modified_since;
+	neon_request_lsns request_lsns;
 
 	switch (reln->smgr_relpersistence)
 	{
@@ -2359,9 +2543,8 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 		return;
 	}
 
-	neon_get_request_lsn(InfoFromSMgrRel(reln), forkNum, blkno,
-						 &request_lsn, &not_modified_since);
-	neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsn, not_modified_since, buffer);
+	request_lsns = neon_get_request_lsns(InfoFromSMgrRel(reln), forkNum, blkno);
+	neon_read_at_lsn(InfoFromSMgrRel(reln), forkNum, blkno, request_lsns, buffer);
 
 #ifdef DEBUG_COMPARE_LOCAL
 	if (forkNum == MAIN_FORKNUM && IS_LOCAL_REL(reln))
@@ -2530,8 +2713,7 @@ neon_nblocks(SMgrRelation reln, ForkNumber forknum)
 {
 	NeonResponse *resp;
 	BlockNumber n_blocks;
-	XLogRecPtr	request_lsn;
-	XLogRecPtr	not_modified_since;
+	neon_request_lsns request_lsns;
 
 	switch (reln->smgr_relpersistence)
 	{
@@ -2558,13 +2740,12 @@ neon_nblocks(SMgrRelation reln, ForkNumber forknum)
 		return n_blocks;
 	}
 
-	neon_get_request_lsn(InfoFromSMgrRel(reln), forknum, REL_METADATA_PSEUDO_BLOCKNO,
-						 &request_lsn, &not_modified_since);
+	request_lsns = neon_get_request_lsns(InfoFromSMgrRel(reln), forknum, REL_METADATA_PSEUDO_BLOCKNO);
 	{
 		NeonNblocksRequest request = {
 			.req.tag = T_NeonNblocksRequest,
-			.req.lsn = request_lsn,
-			.req.not_modified_since = not_modified_since,
+			.req.lsn = request_lsns.request_lsn,
+			.req.not_modified_since = request_lsns.not_modified_since,
 			.rinfo = InfoFromSMgrRel(reln),
 			.forknum = forknum,
 		};
@@ -2584,21 +2765,23 @@ neon_nblocks(SMgrRelation reln, ForkNumber forknum)
 					 errmsg(NEON_TAG "could not read relation size of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 							RelFileInfoFmt(InfoFromSMgrRel(reln)),
 							forknum,
-							(uint32) (request_lsn >> 32), (uint32) request_lsn),
+							LSN_FORMAT_ARGS(request_lsns.effective_request_lsn)),
 					 errdetail("page server returned error: %s",
 							   ((NeonErrorResponse *) resp)->message)));
 			break;
 
 		default:
-			neon_log(ERROR, "unexpected response from page server with tag 0x%02x in neon_nblocks", resp->tag);
+			NEON_PANIC_CONNECTION_STATE(-1, PANIC,
+										"Expected Nblocks (0x%02x) or Error (0x%02x) response to NblocksRequest, but got 0x%02x",
+										T_NeonNblocksResponse, T_NeonErrorResponse, resp->tag);
 	}
 	update_cached_relsize(InfoFromSMgrRel(reln), forknum, n_blocks);
 
 	neon_log(SmgrTrace, "neon_nblocks: rel %u/%u/%u fork %u (request LSN %X/%08X): %u blocks",
-		 RelFileInfoFmt(InfoFromSMgrRel(reln)),
-		 forknum,
-		 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-		 n_blocks);
+			 RelFileInfoFmt(InfoFromSMgrRel(reln)),
+			 forknum,
+			 LSN_FORMAT_ARGS(request_lsns.effective_request_lsn),
+			 n_blocks);
 
 	pfree(resp);
 	return n_blocks;
@@ -2612,17 +2795,15 @@ neon_dbsize(Oid dbNode)
 {
 	NeonResponse *resp;
 	int64		db_size;
-	XLogRecPtr	request_lsn,
-				not_modified_since;
+	neon_request_lsns request_lsns;
 	NRelFileInfo dummy_node = {0};
 
-	neon_get_request_lsn(dummy_node, MAIN_FORKNUM, REL_METADATA_PSEUDO_BLOCKNO,
-						 &request_lsn, &not_modified_since);
+	request_lsns = neon_get_request_lsns(dummy_node, MAIN_FORKNUM, REL_METADATA_PSEUDO_BLOCKNO);
 	{
 		NeonDbSizeRequest request = {
 			.req.tag = T_NeonDbSizeRequest,
-			.req.lsn = request_lsn,
-			.req.not_modified_since = not_modified_since,
+			.req.lsn = request_lsns.request_lsn,
+			.req.not_modified_since = request_lsns.not_modified_since,
 			.dbNode = dbNode,
 		};
 
@@ -2639,20 +2820,19 @@ neon_dbsize(Oid dbNode)
 			ereport(ERROR,
 					(errcode(ERRCODE_IO_ERROR),
 					 errmsg(NEON_TAG "could not read db size of db %u from page server at lsn %X/%08X",
-							dbNode,
-							(uint32) (request_lsn >> 32), (uint32) request_lsn),
+							dbNode, LSN_FORMAT_ARGS(request_lsns.effective_request_lsn)),
 					 errdetail("page server returned error: %s",
 							   ((NeonErrorResponse *) resp)->message)));
 			break;
 
 		default:
-			neon_log(ERROR, "unexpected response from page server with tag 0x%02x in neon_dbsize", resp->tag);
+			NEON_PANIC_CONNECTION_STATE(-1, PANIC,
+										"Expected DbSize (0x%02x) or Error (0x%02x) response to DbSizeRequest, but got 0x%02x",
+										T_NeonDbSizeResponse, T_NeonErrorResponse, resp->tag);
 	}
 
 	neon_log(SmgrTrace, "neon_dbsize: db %u (request LSN %X/%08X): %ld bytes",
-		 dbNode,
-		 (uint32) (request_lsn >> 32), (uint32) request_lsn,
-		 db_size);
+			 dbNode, LSN_FORMAT_ARGS(request_lsns.effective_request_lsn), db_size);
 
 	pfree(resp);
 	return db_size;
@@ -2812,10 +2992,14 @@ neon_start_unlogged_build(SMgrRelation reln)
 	reln->smgr_relpersistence = RELPERSISTENCE_UNLOGGED;
 
 	/*
+	 * Create the local file. In a parallel build, the leader is expected to
+	 * call this first and do it.
+	 *
 	 * FIXME: should we pass isRedo true to create the tablespace dir if it
 	 * doesn't exist? Is it needed?
 	 */
-	mdcreate(reln, MAIN_FORKNUM, false);
+	if (!IsParallelWorker())
+		mdcreate(reln, MAIN_FORKNUM, false);
 }
 
 /*
@@ -2839,7 +3023,17 @@ neon_finish_unlogged_build_phase_1(SMgrRelation reln)
 	Assert(unlogged_build_phase == UNLOGGED_BUILD_PHASE_1);
 	Assert(reln->smgr_relpersistence == RELPERSISTENCE_UNLOGGED);
 
-	unlogged_build_phase = UNLOGGED_BUILD_PHASE_2;
+	/*
+	 * In a parallel build, (only) the leader process performs the 2nd
+	 * phase.
+	 */
+	if (IsParallelWorker())
+	{
+		unlogged_build_rel = NULL;
+		unlogged_build_phase = UNLOGGED_BUILD_NOT_IN_PROGRESS;
+	}
+	else
+		unlogged_build_phase = UNLOGGED_BUILD_PHASE_2;
 }
 
 /*
@@ -2897,6 +3091,10 @@ neon_read_slru_segment(SMgrRelation reln, const char* path, int segno, void* buf
 	XLogRecPtr request_lsn,
 		not_modified_since;
 
+	/*
+	 * Compute a request LSN to use, similar to neon_get_request_lsns() but the
+	 * logic is a bit simpler.
+	 */
 	if (RecoveryInProgress())
 	{
 		request_lsn = GetXLogReplayRecPtr(NULL);
@@ -2908,10 +3106,10 @@ neon_read_slru_segment(SMgrRelation reln, const char* path, int segno, void* buf
 			 */
 			request_lsn = GetRedoStartLsn();
 		}
+		request_lsn = nm_adjust_lsn(request_lsn);
 	}
 	else
-		request_lsn = GetXLogInsertRecPtr();
-	request_lsn = nm_adjust_lsn(request_lsn);
+		request_lsn = UINT64_MAX;
 
 	/*
 	 * GetRedoStartLsn() returns LSN of basebackup. We know that the SLRU
@@ -2969,7 +3167,9 @@ neon_read_slru_segment(SMgrRelation reln, const char* path, int segno, void* buf
 			break;
 
 		default:
-			neon_log(ERROR, "unexpected response from page server with tag 0x%02x in neon_read_slru_segment", resp->tag);
+			NEON_PANIC_CONNECTION_STATE(-1, PANIC,
+										"Expected GetSlruSegment (0x%02x) or Error (0x%02x) response to GetSlruSegmentRequest, but got 0x%02x",
+										T_NeonGetSlruSegmentResponse, T_NeonErrorResponse, resp->tag);
 	}
 	pfree(resp);
 
@@ -3187,7 +3387,7 @@ neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)
 	BufferTag	tag;
 	uint32		hash;
 	LWLock	   *partitionLock;
-	Buffer		buffer;
+	int			buf_id;
 	bool		no_redo_needed;
 
 	if (old_redo_read_buffer_filter && old_redo_read_buffer_filter(record, block_id))
@@ -3225,20 +3425,20 @@ neon_redo_read_buffer_filter(XLogReaderState *record, uint8 block_id)
 	else
 	{
 		/* Try to find the relevant buffer */
-		buffer = BufTableLookup(&tag, hash);
+		buf_id = BufTableLookup(&tag, hash);
 
-		no_redo_needed = buffer < 0;
+		no_redo_needed = buf_id < 0;
 	}
-	/* In both cases st lwlsn past this WAL record */
-	SetLastWrittenLSNForBlock(end_recptr, rinfo, forknum, blkno);
 
 	/*
 	 * we don't have the buffer in memory, update lwLsn past this record, also
 	 * evict page from file cache
 	 */
 	if (no_redo_needed)
+	{
+		SetLastWrittenLSNForBlock(end_recptr, rinfo, forknum, blkno);
 		lfc_evict(rinfo, forknum, blkno);
-
+	}
 
 	LWLockRelease(partitionLock);
 

pgxn/neon/walproposer_pg.c

@@ -1852,34 +1852,30 @@ static void
 CombineHotStanbyFeedbacks(HotStandbyFeedback *hs, WalProposer *wp)
 {
 	hs->ts = 0;
-	hs->xmin.value = ~0;		/* largest unsigned value */
-	hs->catalog_xmin.value = ~0;	/* largest unsigned value */
+	hs->xmin = InvalidFullTransactionId;
+	hs->catalog_xmin = InvalidFullTransactionId;
 
 	for (int i = 0; i < wp->n_safekeepers; i++)
 	{
-		if (wp->safekeeper[i].appendResponse.hs.ts != 0)
+
+		if (wp->safekeeper[i].state == SS_ACTIVE)
 		{
 			HotStandbyFeedback *skhs = &wp->safekeeper[i].appendResponse.hs;
 
 			if (FullTransactionIdIsNormal(skhs->xmin)
-				&& FullTransactionIdPrecedes(skhs->xmin, hs->xmin))
+				&& (!FullTransactionIdIsValid(hs->xmin) || FullTransactionIdPrecedes(skhs->xmin, hs->xmin)))
 			{
 				hs->xmin = skhs->xmin;
 				hs->ts = skhs->ts;
 			}
 			if (FullTransactionIdIsNormal(skhs->catalog_xmin)
-				&& FullTransactionIdPrecedes(skhs->catalog_xmin, hs->xmin))
+				&& (!FullTransactionIdIsValid(hs->catalog_xmin) || FullTransactionIdPrecedes(skhs->catalog_xmin, hs->catalog_xmin)))
 			{
 				hs->catalog_xmin = skhs->catalog_xmin;
 				hs->ts = skhs->ts;
 			}
 		}
 	}
-
-	if (hs->xmin.value == ~0)
-		hs->xmin = InvalidFullTransactionId;
-	if (hs->catalog_xmin.value == ~0)
-		hs->catalog_xmin = InvalidFullTransactionId;
 }
 
 /*
@@ -1946,14 +1942,28 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, Safekeeper *sk)
 	}
 
 	CombineHotStanbyFeedbacks(&hsFeedback, wp);
-	if (hsFeedback.ts != 0 && memcmp(&hsFeedback, &agg_hs_feedback, sizeof hsFeedback) != 0)
+	if (memcmp(&hsFeedback, &agg_hs_feedback, sizeof hsFeedback) != 0)
 	{
+		FullTransactionId xmin = hsFeedback.xmin;
+		FullTransactionId catalog_xmin = hsFeedback.catalog_xmin;
+		FullTransactionId next_xid = ReadNextFullTransactionId();
+		/*
+		 * Page server is updating nextXid in checkpoint each 1024 transactions,
+		 * so feedback xmin can be actually larger then nextXid and
+		 * function TransactionIdInRecentPast return false in this case,
+		 * preventing update of slot's xmin.
+		 */
+		if (FullTransactionIdPrecedes(next_xid, xmin))
+			xmin = next_xid;
+		if (FullTransactionIdPrecedes(next_xid, catalog_xmin))
+			catalog_xmin = next_xid;
 		agg_hs_feedback = hsFeedback;
+		elog(DEBUG2, "ProcessStandbyHSFeedback(xmin=%d, catalog_xmin=%d", XidFromFullTransactionId(hsFeedback.xmin), XidFromFullTransactionId(hsFeedback.catalog_xmin));
 		ProcessStandbyHSFeedback(hsFeedback.ts,
-								 XidFromFullTransactionId(hsFeedback.xmin),
-								 EpochFromFullTransactionId(hsFeedback.xmin),
-								 XidFromFullTransactionId(hsFeedback.catalog_xmin),
-								 EpochFromFullTransactionId(hsFeedback.catalog_xmin));
+								 XidFromFullTransactionId(xmin),
+								 EpochFromFullTransactionId(xmin),
+								 XidFromFullTransactionId(catalog_xmin),
+								 EpochFromFullTransactionId(catalog_xmin));
 	}
 
 	CheckGracefulShutdown(wp);

pgxn/neon_test_utils/neontest.c

@@ -48,10 +48,10 @@ PG_FUNCTION_INFO_V1(neon_xlogflush);
  */
 #if PG_MAJORVERSION_NUM < 16
 typedef void (*neon_read_at_lsn_type) (NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
-									   XLogRecPtr request_lsn, XLogRecPtr not_modified_since, char *buffer);
+									   neon_request_lsns request_lsns, char *buffer);
 #else
 typedef void (*neon_read_at_lsn_type) (NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
-									   XLogRecPtr request_lsn, XLogRecPtr not_modified_since, void *buffer);
+									   neon_request_lsns request_lsns, void *buffer);
 #endif
 
 static neon_read_at_lsn_type neon_read_at_lsn_ptr;
@@ -298,9 +298,7 @@ get_raw_page_at_lsn(PG_FUNCTION_ARGS)
 	text	   *relname;
 	text	   *forkname;
 	uint32		blkno;
-
-	XLogRecPtr	request_lsn;
-	XLogRecPtr	not_modified_since;
+	neon_request_lsns	request_lsns;
 
 	if (PG_NARGS() != 5)
 		elog(ERROR, "unexpected number of arguments in SQL function signature");
@@ -312,8 +310,15 @@ get_raw_page_at_lsn(PG_FUNCTION_ARGS)
 	forkname = PG_GETARG_TEXT_PP(1);
 	blkno = PG_GETARG_UINT32(2);
 
-	request_lsn = PG_ARGISNULL(3) ? GetXLogInsertRecPtr() : PG_GETARG_LSN(3);
-	not_modified_since = PG_ARGISNULL(4) ? request_lsn : PG_GETARG_LSN(4);
+	request_lsns.request_lsn = PG_ARGISNULL(3) ? GetXLogInsertRecPtr() : PG_GETARG_LSN(3);
+	request_lsns.not_modified_since = PG_ARGISNULL(4) ? request_lsns.request_lsn : PG_GETARG_LSN(4);
+	/*
+	 * For the time being, use the same LSN for request and
+	 * effective request LSN. If any test needed to use UINT64_MAX
+	 * as the request LSN, we'd need to add effective_request_lsn
+	 * as a new argument.
+	 */
+	request_lsns.effective_request_lsn = request_lsns.request_lsn;
 
 	if (!superuser())
 		ereport(ERROR,
@@ -367,7 +372,8 @@ get_raw_page_at_lsn(PG_FUNCTION_ARGS)
 	SET_VARSIZE(raw_page, BLCKSZ + VARHDRSZ);
 	raw_page_data = VARDATA(raw_page);
 
-	neon_read_at_lsn(InfoFromRelation(rel), forknum, blkno, request_lsn, not_modified_since, raw_page_data);
+	neon_read_at_lsn(InfoFromRelation(rel), forknum, blkno, request_lsns,
+					 raw_page_data);
 
 	relation_close(rel, AccessShareLock);
 
@@ -413,19 +419,25 @@ get_raw_page_at_lsn_ex(PG_FUNCTION_ARGS)
 
 		ForkNumber	forknum = PG_GETARG_UINT32(3);
 		uint32		blkno = PG_GETARG_UINT32(4);
-		XLogRecPtr	request_lsn;
-		XLogRecPtr	not_modified_since;
+		neon_request_lsns	request_lsns;
 
 		/* Initialize buffer to copy to */
 		bytea	   *raw_page = (bytea *) palloc(BLCKSZ + VARHDRSZ);
 
-		request_lsn = PG_ARGISNULL(5) ? GetXLogInsertRecPtr() : PG_GETARG_LSN(5);
-		not_modified_since = PG_ARGISNULL(6) ? request_lsn : PG_GETARG_LSN(6);
+		request_lsns.request_lsn = PG_ARGISNULL(5) ? GetXLogInsertRecPtr() : PG_GETARG_LSN(5);
+		request_lsns.not_modified_since = PG_ARGISNULL(6) ? request_lsns.request_lsn : PG_GETARG_LSN(6);
+		/*
+		 * For the time being, use the same LSN for request
+		 * and effective request LSN. If any test needed to
+		 * use UINT64_MAX as the request LSN, we'd need to add
+		 * effective_request_lsn as a new argument.
+		 */
+		request_lsns.effective_request_lsn = request_lsns.request_lsn;
 
 		SET_VARSIZE(raw_page, BLCKSZ + VARHDRSZ);
 		raw_page_data = VARDATA(raw_page);
 
-		neon_read_at_lsn(rinfo, forknum, blkno, request_lsn, not_modified_since, raw_page_data);
+		neon_read_at_lsn(rinfo, forknum, blkno, request_lsns, raw_page_data);
 		PG_RETURN_BYTEA_P(raw_page);
 	}
 }

poetry.lock

@@ -2405,6 +2405,7 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
+    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -2529,13 +2530,13 @@ files = [
 
 [[package]]
 name = "requests"
-version = "2.31.0"
+version = "2.32.0"
 description = "Python HTTP for Humans."
 optional = false
-python-versions = ">=3.7"
+python-versions = ">=3.8"
 files = [
-    {file = "requests-2.31.0-py3-none-any.whl", hash = "sha256:58cd2187c01e70e6e26505bca751777aa9f2ee0b7f4300988b709f44e013003f"},
-    {file = "requests-2.31.0.tar.gz", hash = "sha256:942c5a758f98d790eaed1a29cb6eefc7ffb0d1cf7af05c3d2791656dbd6ad1e1"},
+    {file = "requests-2.32.0-py3-none-any.whl", hash = "sha256:f2c3881dddb70d056c5bd7600a4fae312b2a300e39be6a118d30b90bd27262b5"},
+    {file = "requests-2.32.0.tar.gz", hash = "sha256:fa5490319474c82ef1d2c9bc459d3652e3ae4ef4c4ebdd18a21145a47ca4b6b8"},
 ]
 
 [package.dependencies]
@@ -2959,6 +2960,16 @@ files = [
     {file = "wrapt-1.14.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:8ad85f7f4e20964db4daadcab70b47ab05c7c1cf2a7c1e51087bfaa83831854c"},
     {file = "wrapt-1.14.1-cp310-cp310-win32.whl", hash = "sha256:a9a52172be0b5aae932bef82a79ec0a0ce87288c7d132946d645eba03f0ad8a8"},
     {file = "wrapt-1.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:6d323e1554b3d22cfc03cd3243b5bb815a51f5249fdcbb86fda4bf62bab9e164"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ecee4132c6cd2ce5308e21672015ddfed1ff975ad0ac8d27168ea82e71413f55"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2020f391008ef874c6d9e208b24f28e31bcb85ccff4f335f15a3251d222b92d9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2feecf86e1f7a86517cab34ae6c2f081fd2d0dac860cb0c0ded96d799d20b335"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:240b1686f38ae665d1b15475966fe0472f78e71b1b4903c143a842659c8e4cb9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9008dad07d71f68487c91e96579c8567c98ca4c3881b9b113bc7b33e9fd78b8"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:6447e9f3ba72f8e2b985a1da758767698efa72723d5b59accefd716e9e8272bf"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:acae32e13a4153809db37405f5eba5bac5fbe2e2ba61ab227926a22901051c0a"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:49ef582b7a1152ae2766557f0550a9fcbf7bbd76f43fbdc94dd3bf07cc7168be"},
+    {file = "wrapt-1.14.1-cp311-cp311-win32.whl", hash = "sha256:358fe87cc899c6bb0ddc185bf3dbfa4ba646f05b1b0b9b5a27c2cb92c2cea204"},
+    {file = "wrapt-1.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:26046cd03936ae745a502abf44dac702a5e6880b2b01c29aea8ddf3353b68224"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_i686.whl", hash = "sha256:43ca3bbbe97af00f49efb06e352eae40434ca9d915906f77def219b88e85d907"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_x86_64.whl", hash = "sha256:6b1a564e6cb69922c7fe3a678b9f9a3c54e72b469875aa8018f18b4d1dd1adf3"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux2010_i686.whl", hash = "sha256:00b6d4ea20a906c0ca56d84f93065b398ab74b927a7a3dbd470f6fc503f95dc3"},
@@ -3196,4 +3207,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.9"
-content-hash = "dcde14c58a32bda5f123319a069352c458b3719f3c62977991eebb9803a46a9e"
+content-hash = "16ebd6a46768be7f67dbdb4ee5903b167d94edc9965f29252f038c67e9e907b0"

proxy/Cargo.toml

@@ -9,6 +9,7 @@ default = []
 testing = []
 
 [dependencies]
+ahash.workspace = true
 anyhow.workspace = true
 async-compression.workspace = true
 async-trait.workspace = true
@@ -24,8 +25,10 @@ camino.workspace = true
 chrono.workspace = true
 clap.workspace = true
 consumption_metrics.workspace = true
+crossbeam-deque.workspace = true
 dashmap.workspace = true
 env_logger.workspace = true
+framed-websockets.workspace = true
 futures.workspace = true
 git-version.workspace = true
 hashbrown.workspace = true
@@ -35,7 +38,6 @@ hmac.workspace = true
 hostname.workspace = true
 http.workspace = true
 humantime.workspace = true
-hyper-tungstenite.workspace = true
 hyper.workspace = true
 hyper1 = { package = "hyper", version = "1.2", features = ["server"] }
 hyper-util = { version = "0.1", features = ["server", "http1", "http2", "tokio"] }
@@ -52,7 +54,6 @@ opentelemetry.workspace = true
 parking_lot.workspace = true
 parquet.workspace = true
 parquet_derive.workspace = true
-pbkdf2 = { workspace = true, features = ["simple", "std"] }
 pin-project-lite.workspace = true
 postgres_backend.workspace = true
 pq_proto.workspace = true
@@ -76,7 +77,6 @@ smol_str.workspace = true
 smallvec.workspace = true
 socket2.workspace = true
 subtle.workspace = true
-sync_wrapper.workspace = true
 task-local-extensions.workspace = true
 thiserror.workspace = true
 tikv-jemallocator.workspace = true
@@ -106,6 +106,8 @@ workspace_hack.workspace = true
 [dev-dependencies]
 camino-tempfile.workspace = true
 fallible-iterator.workspace = true
+tokio-tungstenite.workspace = true
+pbkdf2 = { workspace = true, features = ["simple", "std"] }
 rcgen.workspace = true
 rstest.workspace = true
 tokio-postgres-rustls.workspace = true