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

Add safekeeper test truncating WAL.
We do it as a part of more complicated tests like test_compute_restarts, but let's have a simple test as well.
2026-01-24 13:50:37 +00:00 · 2024-05-28 16:29:05 +00:00 · 2024-05-28 11:08:29 +03:00 · 2024-05-27 15:57:57 +03:00 · 2024-05-25 06:06:32 +03:00 · 2024-05-25 06:06:32 +03:00
267 changed files with 12626 additions and 5418 deletions
--- a/.config/nextest.toml
+++ b/.config/nextest.toml
@@ -1,2 +1,2 @@
 [profile.default]
-slow-timeout = { period = "20s", terminate-after = 3 }
+slow-timeout = { period = "60s", terminate-after = 3 }
--- a/.dockerignore
+++ b/.dockerignore
@@ -17,6 +17,7 @@
 !libs/
 !neon_local/
 !pageserver/
+!patches/
 !pgxn/
 !proxy/
 !s3_scrubber/
--- a/.github/actionlint.yml
+++ b/.github/actionlint.yml
@@ -1,12 +1,11 @@
 self-hosted-runner:
  labels:
    - arm64
-    - dev
    - gen3
    - large
-    # Remove `macos-14` from the list after https://github.com/rhysd/actionlint/pull/392 is merged.
-    - macos-14
+    - large-arm64
    - small
+    - small-arm64
    - us-east-2
 config-variables:
  - REMOTE_STORAGE_AZURE_CONTAINER
--- a/.github/actions/neon-branch-create/action.yml
+++ b/.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:
--- a/.github/actions/neon-branch-delete/action.yml
+++ b/.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:
--- a/.github/actions/neon-project-create/action.yml
+++ b/.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:
--- a/.github/actions/neon-project-delete/action.yml
+++ b/.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:
--- a/.github/workflows/build-build-tools-image.yml
+++ b/.github/workflows/build-build-tools-image.yml
@@ -39,7 +39,7 @@ jobs:
      matrix:
        arch: [ x64, arm64 ]

-    runs-on: ${{ fromJson(format('["self-hosted", "dev", "{0}"]', matrix.arch)) }}
+    runs-on: ${{ fromJson(format('["self-hosted", "gen3", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}

    env:
      IMAGE_TAG: ${{ inputs.image-tag }}
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -341,6 +341,9 @@ jobs:
        env:
          NEXTEST_RETRIES: 3
        run: |
+          #nextest does not yet support running doctests
+          cargo test --doc $CARGO_FLAGS $CARGO_FEATURES
+
          for io_engine in std-fs tokio-epoll-uring ; do
            NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE=$io_engine ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES
          done
@@ -429,8 +432,9 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        build_type: [ debug, release ]
-        pg_version: [ v14, v15, v16 ]
+        build_type: [ debug ]
+        pg_version: [ v15 ]
+        pageserver_compaction_algorithm_kind: [ "legacy", "tiered" ]
    steps:
      - name: Checkout
        uses: actions/checkout@v4
@@ -458,6 +462,9 @@ jobs:
          PAGESERVER_GET_VECTORED_IMPL: vectored
          PAGESERVER_GET_IMPL: vectored
          PAGESERVER_VALIDATE_VEC_GET: true
+          PAGESERVER_DEFAULT_TENANT_CONFIG_COMPACTION_ALGORITHM: 'kind="${{ 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
@@ -543,9 +550,27 @@ jobs:
      # XXX: no coverage data handling here, since benchmarks are run on release builds,
      # while coverage is currently collected for the debug ones

+  report-benchmarks-failures:
+    needs: [ benchmarks, create-test-report ]
+    if: github.ref_name == 'main' && failure() && needs.benchmarks.result == 'failure'
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: slackapi/slack-github-action@v1
+      with:
+        channel-id: C060CNA47S9 # on-call-staging-storage-stream
+        slack-message: |
+          Benchmarks failed on main: ${{ github.event.head_commit.url }}
+
+          Allure report: ${{ needs.create-test-report.outputs.report-url }}
+      env:
+        SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
+
  create-test-report:
    needs: [ check-permissions, regress-tests, coverage-report, benchmarks, build-build-tools-image ]
    if: ${{ !cancelled() && contains(fromJSON('["skipped", "success"]'), needs.check-permissions.result) }}
+    outputs:
+      report-url: ${{ steps.create-allure-report.outputs.report-url }}

    runs-on: [ self-hosted, gen3, small ]
    container:
@@ -702,9 +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
@@ -726,12 +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: ${{ secrets.AWS_SECRET_KEY_DEV }}
-
      - uses: docker/build-push-action@v5
        with:
          context: .
@@ -743,25 +766,52 @@ jobs:
          push: true
          pull: true
          file: Dockerfile
-          cache-from: type=registry,ref=neondatabase/neon:cache
-          cache-to: type=registry,ref=neondatabase/neon:cache,mode=max
+          cache-from: type=registry,ref=neondatabase/neon:cache-${{ 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
@@ -808,15 +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: ${{ matrix.version == 'v16' }}
+        if: matrix.version == 'v16'
        uses: docker/build-push-action@v5
        with:
          target: compute-tools-image
@@ -830,14 +879,57 @@ jobs:
          pull: true
          file: Dockerfile.compute-node
          tags: |
-            369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{ 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 ]
@@ -845,11 +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
@@ -862,26 +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
@@ -899,7 +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"

@@ -925,78 +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 ]
--- a/.github/workflows/neon_extra_builds.yml
+++ b/.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, dev, arm64 ]
+    runs-on: [ self-hosted, small-arm64 ]

    env:
      # Use release build only, to have less debug info around
@@ -232,20 +232,20 @@ jobs:

      - name: Run cargo build
        run: |
-          mold -run cargo build --locked $CARGO_FLAGS $CARGO_FEATURES --bins --tests
+          mold -run cargo build --locked $CARGO_FLAGS $CARGO_FEATURES --bins --tests -j$(nproc)

      - name: Run cargo test
        env:
          NEXTEST_RETRIES: 3
        run: |
-          cargo nextest run $CARGO_FEATURES
+          cargo nextest run $CARGO_FEATURES -j$(nproc)

          # Run separate tests for real S3
          export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
          export REMOTE_STORAGE_S3_BUCKET=neon-github-ci-tests
          export REMOTE_STORAGE_S3_REGION=eu-central-1
          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          cargo nextest run --package remote_storage --test test_real_s3
+          cargo nextest run --package remote_storage --test test_real_s3 -j$(nproc)

          # Run separate tests for real Azure Blob Storage
          # XXX: replace region with `eu-central-1`-like region
@@ -255,12 +255,12 @@ jobs:
          export REMOTE_STORAGE_AZURE_CONTAINER="${{ vars.REMOTE_STORAGE_AZURE_CONTAINER }}"
          export REMOTE_STORAGE_AZURE_REGION="${{ vars.REMOTE_STORAGE_AZURE_REGION }}"
          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
-          cargo nextest run --package remote_storage --test test_real_azure
+          cargo nextest run --package remote_storage --test test_real_azure -j$(nproc)

  check-codestyle-rust-arm:
    needs: [ check-permissions, build-build-tools-image ]
    timeout-minutes: 90
-    runs-on: [ self-hosted, dev, arm64 ]
+    runs-on: [ self-hosted, small-arm64 ]

    container:
      image: ${{ needs.build-build-tools-image.outputs.image }}
@@ -269,6 +269,11 @@ jobs:
        password: ${{ secrets.NEON_DOCKERHUB_PASSWORD }}
      options: --init

+    strategy:
+      fail-fast: false
+      matrix:
+        build_type: [ debug, release ]
+
    steps:
      - name: Fix git ownership
        run: |
@@ -305,31 +310,35 @@ jobs:
            exit 1
          fi
          echo "CLIPPY_COMMON_ARGS=${CLIPPY_COMMON_ARGS}" >> $GITHUB_ENV
+
      - name: Run cargo clippy (debug)
+        if: matrix.build_type == 'debug'
        run: cargo hack --feature-powerset clippy $CLIPPY_COMMON_ARGS
      - name: Run cargo clippy (release)
+        if: matrix.build_type == 'release'
        run: cargo hack --feature-powerset clippy --release $CLIPPY_COMMON_ARGS

      - name: Check documentation generation
-        run: cargo doc --workspace --no-deps --document-private-items
+        if: matrix.build_type == 'release'
+        run: cargo doc --workspace --no-deps --document-private-items -j$(nproc)
        env:
            RUSTDOCFLAGS: "-Dwarnings -Arustdoc::private_intra_doc_links"

      # Use `${{ !cancelled() }}` to run quck tests after the longer clippy run
      - name: Check formatting
-        if: ${{ !cancelled() }}
+        if: ${{ !cancelled() && matrix.build_type == 'release' }}
        run: cargo fmt --all -- --check

      # https://github.com/facebookincubator/cargo-guppy/tree/bec4e0eb29dcd1faac70b1b5360267fc02bf830e/tools/cargo-hakari#2-keep-the-workspace-hack-up-to-date-in-ci
      - name: Check rust dependencies
-        if: ${{ !cancelled() }}
+        if: ${{ !cancelled() && matrix.build_type == 'release' }}
        run: |
          cargo hakari generate --diff  # workspace-hack Cargo.toml is up-to-date
          cargo hakari manage-deps --dry-run  # all workspace crates depend on workspace-hack

      # https://github.com/EmbarkStudios/cargo-deny
      - name: Check rust licenses/bans/advisories/sources
-        if: ${{ !cancelled() }}
+        if: ${{ !cancelled() && matrix.build_type == 'release' }}
        run: cargo deny check

  gather-rust-build-stats:
@@ -338,7 +347,7 @@ jobs:
      contains(github.event.pull_request.labels.*.name, 'run-extra-build-stats') ||
      contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
      github.ref_name == 'main'
-    runs-on: [ self-hosted, gen3, large ]
+    runs-on: [ self-hosted, large ]
    container:
      image: ${{ needs.build-build-tools-image.outputs.image }}
      credentials:
@@ -369,7 +378,7 @@ jobs:
        run: make walproposer-lib -j$(nproc)

      - name: Produce the build stats
-        run: cargo build --all --release --timings
+        run: cargo build --all --release --timings -j$(nproc)

      - name: Upload the build stats
        id: upload-stats
--- a/.github/workflows/release.yml
+++ b/.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
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -25,9 +25,9 @@ checksum = "f26201604c87b1e01bd3d98f8d5d9a8fcbb815e8cedb41ffccbeb4bf593a35fe"

 [[package]]
 name = "ahash"
-version = "0.8.9"
+version = "0.8.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d713b3834d76b85304d4d525563c1276e2e30dc97cc67bfb4585a4a29fc2c89f"
+checksum = "e89da841a80418a9b391ebaea17f5c112ffaaa96f621d2c285b5174da76b9011"
 dependencies = [
 "cfg-if",
 "const-random",
@@ -284,9 +284,9 @@ checksum = "d468802bab17cbc0cc575e9b053f41e72aa36bfa6b7f55e3529ffa43161b97fa"

 [[package]]
 name = "aws-config"
-version = "1.1.4"
+version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b30c39ebe61f75d1b3785362b1586b41991873c9ab3e317a9181c246fb71d82"
+checksum = "baaa0be6ee7d90b775ae6ccb6d2ba182b91219ec2001f92338773a094246af1d"
 dependencies = [
 "aws-credential-types",
 "aws-runtime",
@@ -309,14 +309,15 @@ dependencies = [
 "time",
 "tokio",
 "tracing",
+ "url",
 "zeroize",
 ]

 [[package]]
 name = "aws-credential-types"
-version = "1.1.8"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa8587ae17c8e967e4b05a62d495be2fb7701bec52a97f7acfe8a29f938384c8"
+checksum = "e16838e6c9e12125face1c1eff1343c75e3ff540de98ff7ebd61874a89bcfeb9"
 dependencies = [
 "aws-smithy-async",
 "aws-smithy-runtime-api",
@@ -326,9 +327,9 @@ dependencies = [

 [[package]]
 name = "aws-runtime"
-version = "1.1.8"
+version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b13dc54b4b49f8288532334bba8f87386a40571c47c37b1304979b556dc613c8"
+checksum = "785da4a15e7b166b505fd577e4560c7a7cd8fbdf842eb1336cbcbf8944ce56f1"
 dependencies = [
 "aws-credential-types",
 "aws-sigv4",
@@ -373,10 +374,11 @@ dependencies = [

 [[package]]
 name = "aws-sdk-s3"
-version = "1.14.0"
+version = "1.26.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "951f7730f51a2155c711c85c79f337fbc02a577fa99d2a0a8059acfce5392113"
+checksum = "7bc5ce518d4b8d16e0408de7bdf1b3097cec61a7daa979750a208f8d9934386d"
 dependencies = [
+ "ahash",
 "aws-credential-types",
 "aws-runtime",
 "aws-sigv4",
@@ -391,20 +393,25 @@ dependencies = [
 "aws-smithy-xml",
 "aws-types",
 "bytes",
+ "fastrand 2.0.0",
+ "hex",
+ "hmac",
 "http 0.2.9",
 "http-body 0.4.5",
+ "lru",
 "once_cell",
 "percent-encoding",
 "regex-lite",
+ "sha2",
 "tracing",
 "url",
 ]

 [[package]]
 name = "aws-sdk-sso"
-version = "1.12.0"
+version = "1.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f486420a66caad72635bc2ce0ff6581646e0d32df02aa39dc983bfe794955a5b"
+checksum = "ca3d6c4cba4e009391b72b0fcf12aff04ea3c9c3aa2ecaafa330326a8bd7e601"
 dependencies = [
 "aws-credential-types",
 "aws-runtime",
@@ -424,9 +431,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-ssooidc"
-version = "1.12.0"
+version = "1.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "39ddccf01d82fce9b4a15c8ae8608211ee7db8ed13a70b514bbfe41df3d24841"
+checksum = "73400dc239d14f63d932f4ca7b55af5e9ef1f857f7d70655249ccc287adb2570"
 dependencies = [
 "aws-credential-types",
 "aws-runtime",
@@ -446,9 +453,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-sts"
-version = "1.12.0"
+version = "1.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a591f8c7e6a621a501b2b5d2e88e1697fcb6274264523a6ad4d5959889a41ce"
+checksum = "10f8858308af76fba3e5ffcf1bb56af5471574d2bdfaf0159470c25bc2f760e5"
 dependencies = [
 "aws-credential-types",
 "aws-runtime",
@@ -469,9 +476,9 @@ dependencies = [

 [[package]]
 name = "aws-sigv4"
-version = "1.2.0"
+version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "11d6f29688a4be9895c0ba8bef861ad0c0dac5c15e9618b9b7a6c233990fc263"
+checksum = "58b56f1cbe6fd4d0c2573df72868f20ab1c125ca9c9dbce17927a463433a2e57"
 dependencies = [
 "aws-credential-types",
 "aws-smithy-eventstream",
@@ -498,9 +505,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-async"
-version = "1.1.8"
+version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d26ea8fa03025b2face2b3038a63525a10891e3d8829901d502e5384a0d8cd46"
+checksum = "62220bc6e97f946ddd51b5f1361f78996e704677afc518a4ff66b7a72ea1378c"
 dependencies = [
 "futures-util",
 "pin-project-lite",
@@ -509,9 +516,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-checksums"
-version = "0.60.4"
+version = "0.60.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "be2acd1b9c6ae5859999250ed5a62423aedc5cf69045b844432de15fa2f31f2b"
+checksum = "83fa43bc04a6b2441968faeab56e68da3812f978a670a5db32accbdcafddd12f"
 dependencies = [
 "aws-smithy-http",
 "aws-smithy-types",
@@ -541,9 +548,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-http"
-version = "0.60.7"
+version = "0.60.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f10fa66956f01540051b0aa7ad54574640f748f9839e843442d99b970d3aff9"
+checksum = "4a7de001a1b9a25601016d8057ea16e31a45fdca3751304c8edf4ad72e706c08"
 dependencies = [
 "aws-smithy-eventstream",
 "aws-smithy-runtime-api",
@@ -581,9 +588,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-runtime"
-version = "1.1.8"
+version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec81002d883e5a7fd2bb063d6fb51c4999eb55d404f4fff3dd878bf4733b9f01"
+checksum = "c9ac79e9f3a4d576f3cd4a470a0275b138d9e7b11b1cd514a6858ae0a79dd5bb"
 dependencies = [
 "aws-smithy-async",
 "aws-smithy-http",
@@ -594,6 +601,7 @@ dependencies = [
 "h2 0.3.26",
 "http 0.2.9",
 "http-body 0.4.5",
+ "http-body 1.0.0",
 "hyper 0.14.26",
 "hyper-rustls 0.24.0",
 "once_cell",
@@ -606,9 +614,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-runtime-api"
-version = "1.2.0"
+version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9acb931e0adaf5132de878f1398d83f8677f90ba70f01f65ff87f6d7244be1c5"
+checksum = "04ec42c2f5c0e7796a2848dde4d9f3bf8ce12ccbb3d5aa40c52fa0cdd61a1c47"
 dependencies = [
 "aws-smithy-async",
 "aws-smithy-types",
@@ -623,16 +631,19 @@ dependencies = [

 [[package]]
 name = "aws-smithy-types"
-version = "1.1.8"
+version = "1.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "abe14dceea1e70101d38fbf2a99e6a34159477c0fb95e68e05c66bd7ae4c3729"
+checksum = "baf98d97bba6ddaba180f1b1147e202d8fe04940403a95a3f826c790f931bbd1"
 dependencies = [
 "base64-simd",
 "bytes",
 "bytes-utils",
 "futures-core",
 "http 0.2.9",
+ "http 1.1.0",
 "http-body 0.4.5",
+ "http-body 1.0.0",
+ "http-body-util",
 "itoa",
 "num-integer",
 "pin-project-lite",
@@ -646,18 +657,18 @@ dependencies = [

 [[package]]
 name = "aws-smithy-xml"
-version = "0.60.7"
+version = "0.60.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "872c68cf019c0e4afc5de7753c4f7288ce4b71663212771bf5e4542eb9346ca9"
+checksum = "d123fbc2a4adc3c301652ba8e149bf4bc1d1725affb9784eb20c953ace06bf55"
 dependencies = [
 "xmlparser",
 ]

 [[package]]
 name = "aws-types"
-version = "1.1.8"
+version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0dbf2f3da841a8930f159163175cf6a3d16ddde517c1b0fba7aa776822800f40"
+checksum = "5a43b56df2c529fe44cb4d92bd64d0479883fb9608ff62daede4df5405381814"
 dependencies = [
 "aws-credential-types",
 "aws-smithy-async",
@@ -697,7 +708,7 @@ dependencies = [
 "sha1",
 "sync_wrapper",
 "tokio",
- "tokio-tungstenite 0.20.0",
+ "tokio-tungstenite",
 "tower",
 "tower-layer",
 "tower-service",
@@ -968,6 +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"
@@ -1055,9 +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",
@@ -1065,7 +1082,7 @@ dependencies = [
 "num-traits",
 "serde",
 "wasm-bindgen",
- "windows-targets 0.48.0",
+ "windows-targets 0.52.4",
 ]

 [[package]]
@@ -1092,7 +1109,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "defaa24ecc093c77630e6c15e17c51f5e187bf35ee514f4e2d67baaa96dae22b"
 dependencies = [
 "ciborium-io",
- "half",
+ "half 1.8.2",
 ]

 [[package]]
@@ -1222,8 +1239,10 @@ dependencies = [
 "serde_json",
 "signal-hook",
 "tar",
+ "thiserror",
 "tokio",
 "tokio-postgres",
+ "tokio-stream",
 "tokio-util",
 "toml_edit",
 "tracing",
@@ -1452,26 +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]]
@@ -1585,7 +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",
@@ -1986,6 +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"
@@ -2238,6 +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"
@@ -2264,9 +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",
@@ -2274,11 +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]]
@@ -2587,21 +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"
@@ -2679,7 +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]]
@@ -2935,6 +2966,15 @@ version = "0.4.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b5e6163cb8c49088c2c36f57875e58ccd8c87c7427f7fbd50ea6710b2f3f2e8f"

+[[package]]
+name = "lru"
+version = "0.12.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3262e75e648fce39813cb56ac41f3c3e3f65217ebf3844d818d1f9398cfb0dc"
+dependencies = [
+ "hashbrown 0.14.5",
+]
+
 [[package]]
 name = "match_cfg"
 version = "0.1.0"
@@ -2985,7 +3025,7 @@ checksum = "652bc741286361c06de8cb4d89b21a6437f120c508c51713663589eeb9928ac5"
 dependencies = [
 "bytes",
 "crossbeam-utils",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 "itoa",
 "lasso",
 "measured-derive",
@@ -3547,7 +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]]
@@ -3868,13 +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",
@@ -3882,12 +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",
@@ -3914,9 +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",
@@ -4339,6 +4381,7 @@ dependencies = [
 name = "proxy"
 version = "0.1.0"
 dependencies = [
+ "ahash",
 "anyhow",
 "async-compression",
 "async-trait",
@@ -4355,12 +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",
@@ -4370,7 +4415,6 @@ dependencies = [
 "humantime",
 "hyper 0.14.26",
 "hyper 1.2.0",
- "hyper-tungstenite",
 "hyper-util",
 "indexmap 2.0.1",
 "ipnet",
@@ -4415,7 +4459,6 @@ dependencies = [
 "smol_str",
 "socket2 0.5.5",
 "subtle",
- "sync_wrapper",
 "task-local-extensions",
 "thiserror",
 "tikv-jemalloc-ctl",
@@ -4424,6 +4467,7 @@ dependencies = [
 "tokio-postgres",
 "tokio-postgres-rustls",
 "tokio-rustls 0.25.0",
+ "tokio-tungstenite",
 "tokio-util",
 "tower-service",
 "tracing",
@@ -5932,7 +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"
@@ -6360,19 +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]]
@@ -6386,7 +6418,7 @@ dependencies = [
 "futures-io",
 "futures-sink",
 "futures-util",
- "hashbrown 0.14.0",
+ "hashbrown 0.14.5",
 "pin-project-lite",
 "tokio",
 "tracing",
@@ -6668,25 +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"
@@ -7457,6 +7470,7 @@ dependencies = [
 name = "workspace_hack"
 version = "0.1.0"
 dependencies = [
+ "ahash",
 "anyhow",
 "aws-config",
 "aws-runtime",
@@ -7482,7 +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",
--- a/Cargo.toml
+++ b/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"] }
@@ -52,14 +53,14 @@ azure_storage_blobs = "0.19"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
-aws-config = { version = "1.1.4", default-features = false, features=["rustls"] }
-aws-sdk-s3 = "1.14"
+aws-config = { version = "1.3", default-features = false, features=["rustls"] }
+aws-sdk-s3 = "1.26"
 aws-sdk-iam = "1.15.0"
-aws-smithy-async = { version = "1.1.4", default-features = false, features=["rt-tokio"] }
-aws-smithy-types = "1.1.4"
-aws-credential-types = "1.1.4"
-aws-sigv4 = { version = "1.2.0", features = ["sign-http"] }
-aws-types = "1.1.7"
+aws-smithy-async = { version = "1.2.1", default-features = false, features=["rt-tokio"] }
+aws-smithy-types = "1.1.9"
+aws-credential-types = "1.2.0"
+aws-sigv4 = { version = "1.2.1", features = ["sign-http"] }
+aws-types = "1.2.0"
 axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
@@ -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

--- a/Dockerfile.build-tools
+++ b/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 \
--- a/Dockerfile.compute-node
+++ b/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

 #########################################################################################
--- a/README.md
+++ b/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

--- a/compute_tools/Cargo.toml
+++ b/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
--- a/compute_tools/src/catalog.rs
+++ b/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()))))
+}
--- a/compute_tools/src/http/api.rs
+++ b/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();
--- a/compute_tools/src/http/openapi_spec.yaml
+++ b/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:
--- a/compute_tools/src/lib.rs
+++ b/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;
--- a/compute_tools/src/swap.rs
+++ b/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(|| {
--- a/control_plane/src/bin/neon_local.rs
+++ b/control_plane/src/bin/neon_local.rs
@@ -9,8 +9,11 @@ use anyhow::{anyhow, bail, Context, Result};
 use clap::{value_parser, Arg, ArgAction, ArgMatches, Command, ValueEnum};
 use compute_api::spec::ComputeMode;
 use control_plane::endpoint::ComputeControlPlane;
-use control_plane::local_env::{InitForceMode, LocalEnv};
-use control_plane::pageserver::{PageServerNode, PAGESERVER_REMOTE_STORAGE_DIR};
+use control_plane::local_env::{
+    InitForceMode, LocalEnv, NeonBroker, NeonLocalInitConf, NeonLocalInitPageserverConf,
+    SafekeeperConf,
+};
+use control_plane::pageserver::PageServerNode;
 use control_plane::safekeeper::SafekeeperNode;
 use control_plane::storage_controller::StorageController;
 use control_plane::{broker, local_env};
@@ -52,44 +55,6 @@ const DEFAULT_PG_VERSION: &str = "15";

 const DEFAULT_PAGESERVER_CONTROL_PLANE_API: &str = "http://127.0.0.1:1234/upcall/v1/";

-fn default_conf(num_pageservers: u16) -> String {
-    let mut template = format!(
-        r#"
-# Default built-in configuration, defined in main.rs
-control_plane_api = '{DEFAULT_PAGESERVER_CONTROL_PLANE_API}'
-
-[broker]
-listen_addr = '{DEFAULT_BROKER_ADDR}'
-
-[[safekeepers]]
-id = {DEFAULT_SAFEKEEPER_ID}
-pg_port = {DEFAULT_SAFEKEEPER_PG_PORT}
-http_port = {DEFAULT_SAFEKEEPER_HTTP_PORT}
-
-"#,
-    );
-
-    for i in 0..num_pageservers {
-        let pageserver_id = NodeId(DEFAULT_PAGESERVER_ID.0 + i as u64);
-        let pg_port = DEFAULT_PAGESERVER_PG_PORT + i;
-        let http_port = DEFAULT_PAGESERVER_HTTP_PORT + i;
-
-        template += &format!(
-            r#"
-[[pageservers]]
-id = {pageserver_id}
-listen_pg_addr = '127.0.0.1:{pg_port}'
-listen_http_addr = '127.0.0.1:{http_port}'
-pg_auth_type = '{trust_auth}'
-http_auth_type = '{trust_auth}'
-"#,
-            trust_auth = AuthType::Trust,
-        )
-    }
-
-    template
-}
-
 ///
 /// Timelines tree element used as a value in the HashMap.
 ///
@@ -152,7 +117,7 @@ fn main() -> Result<()> {
    };

    match subcommand_result {
-        Ok(Some(updated_env)) => updated_env.persist_config(&updated_env.base_data_dir)?,
+        Ok(Some(updated_env)) => updated_env.persist_config()?,
        Ok(None) => (),
        Err(e) => {
            eprintln!("command failed: {e:?}");
@@ -341,55 +306,65 @@ fn parse_timeline_id(sub_match: &ArgMatches) -> anyhow::Result<Option<TimelineId
 }

 fn handle_init(init_match: &ArgMatches) -> anyhow::Result<LocalEnv> {
-    let num_pageservers = init_match
-        .get_one::<u16>("num-pageservers")
-        .expect("num-pageservers arg has a default");
-    // Create config file
-    let toml_file: String = if let Some(config_path) = init_match.get_one::<PathBuf>("config") {
+    let num_pageservers = init_match.get_one::<u16>("num-pageservers");
+
+    let force = init_match.get_one("force").expect("we set a default value");
+
+    // Create the in-memory `LocalEnv` that we'd normally load from disk in `load_config`.
+    let init_conf: NeonLocalInitConf = if let Some(config_path) =
+        init_match.get_one::<PathBuf>("config")
+    {
+        // User (likely the Python test suite) provided a description of the environment.
+        if num_pageservers.is_some() {
+            bail!("Cannot specify both --num-pageservers and --config, use key `pageservers` in the --config file instead");
+        }
        // load and parse the file
-        std::fs::read_to_string(config_path).with_context(|| {
+        let contents = std::fs::read_to_string(config_path).with_context(|| {
            format!(
                "Could not read configuration file '{}'",
                config_path.display()
            )
-        })?
+        })?;
+        toml_edit::de::from_str(&contents)?
    } else {
-        // Built-in default config
-        default_conf(*num_pageservers)
+        // User (likely interactive) did not provide a description of the environment, give them the default
+        NeonLocalInitConf {
+            control_plane_api: Some(Some(DEFAULT_PAGESERVER_CONTROL_PLANE_API.parse().unwrap())),
+            broker: NeonBroker {
+                listen_addr: DEFAULT_BROKER_ADDR.parse().unwrap(),
+            },
+            safekeepers: vec![SafekeeperConf {
+                id: DEFAULT_SAFEKEEPER_ID,
+                pg_port: DEFAULT_SAFEKEEPER_PG_PORT,
+                http_port: DEFAULT_SAFEKEEPER_HTTP_PORT,
+                ..Default::default()
+            }],
+            pageservers: (0..num_pageservers.copied().unwrap_or(1))
+                .map(|i| {
+                    let pageserver_id = NodeId(DEFAULT_PAGESERVER_ID.0 + i as u64);
+                    let pg_port = DEFAULT_PAGESERVER_PG_PORT + i;
+                    let http_port = DEFAULT_PAGESERVER_HTTP_PORT + i;
+                    NeonLocalInitPageserverConf {
+                        id: pageserver_id,
+                        listen_pg_addr: format!("127.0.0.1:{pg_port}"),
+                        listen_http_addr: format!("127.0.0.1:{http_port}"),
+                        pg_auth_type: AuthType::Trust,
+                        http_auth_type: AuthType::Trust,
+                        other: Default::default(),
+                    }
+                })
+                .collect(),
+            pg_distrib_dir: None,
+            neon_distrib_dir: None,
+            default_tenant_id: TenantId::from_array(std::array::from_fn(|_| 0)),
+            storage_controller: None,
+            control_plane_compute_hook_api: None,
+        }
    };

-    let pageserver_config: toml_edit::Document =
-        if let Some(path) = init_match.get_one::<PathBuf>("pageserver-config") {
-            std::fs::read_to_string(path)?.parse()?
-        } else {
-            toml_edit::Document::new()
-        };
-
-    let pg_version = init_match
-        .get_one::<u32>("pg-version")
-        .copied()
-        .context("Failed to parse postgres version from the argument string")?;
-
-    let mut env =
-        LocalEnv::parse_config(&toml_file).context("Failed to create neon configuration")?;
-    let force = init_match.get_one("force").expect("we set a default value");
-    env.init(pg_version, force)
-        .context("Failed to initialize neon repository")?;
-
-    // Create remote storage location for default LocalFs remote storage
-    std::fs::create_dir_all(env.base_data_dir.join(PAGESERVER_REMOTE_STORAGE_DIR))?;
-
-    // Initialize pageserver, create initial tenant and timeline.
-    for ps_conf in &env.pageservers {
-        PageServerNode::from_env(&env, ps_conf)
-            .initialize(pageserver_config.clone())
-            .unwrap_or_else(|e| {
-                eprintln!("pageserver init failed: {e:?}");
-                exit(1);
-            });
-    }
-
-    Ok(env)
+    LocalEnv::init(init_conf, force)
+        .context("materialize initial neon_local environment on disk")?;
+    Ok(LocalEnv::load_config().expect("freshly written config should be loadable"))
 }

 /// The default pageserver is the one where CLI tenant/timeline operations are sent by default.
@@ -1418,9 +1393,7 @@ fn cli() -> Command {
    let num_pageservers_arg = Arg::new("num-pageservers")
        .value_parser(value_parser!(u16))
        .long("num-pageservers")
-        .help("How many pageservers to create (default 1)")
-        .required(false)
-        .default_value("1");
+        .help("How many pageservers to create (default 1)");

    let update_catalog = Arg::new("update-catalog")
        .value_parser(value_parser!(bool))
@@ -1454,14 +1427,6 @@ fn cli() -> Command {
                        .value_parser(value_parser!(PathBuf))
                        .value_name("config")
                )
-                .arg(
-                    Arg::new("pageserver-config")
-                        .long("pageserver-config")
-                        .required(false)
-                        .value_parser(value_parser!(PathBuf))
-                        .value_name("pageserver-config")
-                        .help("Merge the provided pageserver config into the one generated by neon_local."),
-                )
                .arg(pg_version_arg.clone())
                .arg(force_arg)
        )
--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -3,7 +3,7 @@
 //! Now it also provides init method which acts like a stub for proper installation
 //! script which will use local paths.

-use anyhow::{bail, ensure, Context};
+use anyhow::{bail, Context};

 use clap::ValueEnum;
 use postgres_backend::AuthType;
@@ -23,6 +23,8 @@ use utils::{
    id::{NodeId, TenantId, TenantTimelineId, TimelineId},
 };

+use crate::pageserver::PageServerNode;
+use crate::pageserver::PAGESERVER_REMOTE_STORAGE_DIR;
 use crate::safekeeper::SafekeeperNode;

 pub const DEFAULT_PG_VERSION: u32 = 15;
@@ -34,7 +36,7 @@ pub const DEFAULT_PG_VERSION: u32 = 15;
 // to 'neon_local init --config=<path>' option. See control_plane/simple.conf for
 // an example.
 //
-#[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
+#[derive(PartialEq, Eq, Clone, Debug)]
 pub struct LocalEnv {
    // Base directory for all the nodes (the pageserver, safekeepers and
    // compute endpoints).
@@ -42,59 +44,99 @@ pub struct LocalEnv {
    // This is not stored in the config file. Rather, this is the path where the
    // config file itself is. It is read from the NEON_REPO_DIR env variable or
    // '.neon' if not given.
-    #[serde(skip)]
    pub base_data_dir: PathBuf,

    // Path to postgres distribution. It's expected that "bin", "include",
    // "lib", "share" from postgres distribution are there. If at some point
    // in time we will be able to run against vanilla postgres we may split that
    // to four separate paths and match OS-specific installation layout.
-    #[serde(default)]
    pub pg_distrib_dir: PathBuf,

    // Path to pageserver binary.
-    #[serde(default)]
    pub neon_distrib_dir: PathBuf,

    // Default tenant ID to use with the 'neon_local' command line utility, when
    // --tenant_id is not explicitly specified.
-    #[serde(default)]
    pub default_tenant_id: Option<TenantId>,

    // used to issue tokens during e.g pg start
-    #[serde(default)]
    pub private_key_path: PathBuf,

    pub broker: NeonBroker,

    // Configuration for the storage controller (1 per neon_local environment)
-    #[serde(default)]
    pub storage_controller: NeonStorageControllerConf,

    /// This Vec must always contain at least one pageserver
+    /// Populdated by [`Self::load_config`] from the individual `pageserver.toml`s.
+    /// NB: not used anymore except for informing users that they need to change their `.neon/config`.
    pub pageservers: Vec<PageServerConf>,

-    #[serde(default)]
    pub safekeepers: Vec<SafekeeperConf>,

    // Control plane upcall API for pageserver: if None, we will not run storage_controller  If set, this will
    // be propagated into each pageserver's configuration.
-    #[serde(default)]
    pub control_plane_api: Option<Url>,

    // Control plane upcall API for storage controller.  If set, this will be propagated into the
    // storage controller's configuration.
-    #[serde(default)]
    pub control_plane_compute_hook_api: Option<Url>,

    /// Keep human-readable aliases in memory (and persist them to config), to hide ZId hex strings from the user.
-    #[serde(default)]
    // A `HashMap<String, HashMap<TenantId, TimelineId>>` would be more appropriate here,
    // but deserialization into a generic toml object as `toml::Value::try_from` fails with an error.
    // https://toml.io/en/v1.0.0 does not contain a concept of "a table inside another table".
+    pub branch_name_mappings: HashMap<String, Vec<(TenantId, TimelineId)>>,
+}
+
+/// On-disk state stored in `.neon/config`.
+#[derive(PartialEq, Eq, Clone, Debug, Default, Serialize, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+pub struct OnDiskConfig {
+    pub pg_distrib_dir: PathBuf,
+    pub neon_distrib_dir: PathBuf,
+    pub default_tenant_id: Option<TenantId>,
+    pub private_key_path: PathBuf,
+    pub broker: NeonBroker,
+    pub storage_controller: NeonStorageControllerConf,
+    #[serde(
+        skip_serializing,
+        deserialize_with = "fail_if_pageservers_field_specified"
+    )]
+    pub pageservers: Vec<PageServerConf>,
+    pub safekeepers: Vec<SafekeeperConf>,
+    pub control_plane_api: Option<Url>,
+    pub control_plane_compute_hook_api: Option<Url>,
    branch_name_mappings: HashMap<String, Vec<(TenantId, TimelineId)>>,
 }

+fn fail_if_pageservers_field_specified<'de, D>(_: D) -> Result<Vec<PageServerConf>, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    Err(serde::de::Error::custom(
+        "The 'pageservers' field is no longer used; pageserver.toml is now authoritative; \
+         Please remove the `pageservers` from your .neon/config.",
+    ))
+}
+
+/// The description of the neon_local env to be initialized by `neon_local init --config`.
+#[derive(Clone, Debug, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct NeonLocalInitConf {
+    // TODO: do we need this? Seems unused
+    pub pg_distrib_dir: Option<PathBuf>,
+    // TODO: do we need this? Seems unused
+    pub neon_distrib_dir: Option<PathBuf>,
+    pub default_tenant_id: TenantId,
+    pub broker: NeonBroker,
+    pub storage_controller: Option<NeonStorageControllerConf>,
+    pub pageservers: Vec<NeonLocalInitPageserverConf>,
+    pub safekeepers: Vec<SafekeeperConf>,
+    pub control_plane_api: Option<Option<Url>>,
+    pub control_plane_compute_hook_api: Option<Option<Url>>,
+}
+
 /// Broker config for cluster internal communication.
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 #[serde(default)]
@@ -110,6 +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 {
@@ -122,6 +167,7 @@ impl Default for NeonStorageControllerConf {
    fn default() -> Self {
        Self {
            max_unavailable: Self::DEFAULT_MAX_UNAVAILABLE_INTERVAL,
+            split_threshold: None,
        }
    }
 }
@@ -141,24 +187,18 @@ impl NeonBroker {
    }
 }

+// neon_local needs to know this subset of pageserver configuration.
+// For legacy reasons, this information is duplicated from `pageserver.toml` into `.neon/config`.
+// It can get stale if `pageserver.toml` is changed.
+// TODO(christian): don't store this at all in `.neon/config`, always load it from `pageserver.toml`
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 #[serde(default, deny_unknown_fields)]
 pub struct PageServerConf {
-    // node id
    pub id: NodeId,
-
-    // Pageserver connection settings
    pub listen_pg_addr: String,
    pub listen_http_addr: String,
-
-    // auth type used for the PG and HTTP ports
    pub pg_auth_type: AuthType,
    pub http_auth_type: AuthType,
-
-    pub(crate) virtual_file_io_engine: Option<String>,
-    pub(crate) get_vectored_impl: Option<String>,
-    pub(crate) get_impl: Option<String>,
-    pub(crate) validate_vectored_get: Option<bool>,
 }

 impl Default for PageServerConf {
@@ -169,10 +209,40 @@ impl Default for PageServerConf {
            listen_http_addr: String::new(),
            pg_auth_type: AuthType::Trust,
            http_auth_type: AuthType::Trust,
-            virtual_file_io_engine: None,
-            get_vectored_impl: None,
-            get_impl: None,
-            validate_vectored_get: None,
+        }
+    }
+}
+
+/// The toml that can be passed to `neon_local init --config`.
+/// This is a subset of the `pageserver.toml` configuration.
+// TODO(christian): use pageserver_api::config::ConfigToml (PR #7656)
+#[derive(Clone, Debug, serde::Deserialize, serde::Serialize)]
+pub struct NeonLocalInitPageserverConf {
+    pub id: NodeId,
+    pub listen_pg_addr: String,
+    pub listen_http_addr: String,
+    pub pg_auth_type: AuthType,
+    pub http_auth_type: AuthType,
+    #[serde(flatten)]
+    pub other: HashMap<String, toml::Value>,
+}
+
+impl From<&NeonLocalInitPageserverConf> for PageServerConf {
+    fn from(conf: &NeonLocalInitPageserverConf) -> Self {
+        let NeonLocalInitPageserverConf {
+            id,
+            listen_pg_addr,
+            listen_http_addr,
+            pg_auth_type,
+            http_auth_type,
+            other: _,
+        } = conf;
+        Self {
+            id: *id,
+            listen_pg_addr: listen_pg_addr.clone(),
+            listen_http_addr: listen_http_addr.clone(),
+            pg_auth_type: *pg_auth_type,
+            http_auth_type: *http_auth_type,
        }
    }
 }
@@ -360,44 +430,7 @@ impl LocalEnv {
            .collect()
    }

-    /// Create a LocalEnv from a config file.
-    ///
-    /// Unlike 'load_config', this function fills in any defaults that are missing
-    /// from the config file.
-    pub fn parse_config(toml: &str) -> anyhow::Result<Self> {
-        let mut env: LocalEnv = toml::from_str(toml)?;
-
-        // Find postgres binaries.
-        // Follow POSTGRES_DISTRIB_DIR if set, otherwise look in "pg_install".
-        // Note that later in the code we assume, that distrib dirs follow the same pattern
-        // for all postgres versions.
-        if env.pg_distrib_dir == Path::new("") {
-            if let Some(postgres_bin) = env::var_os("POSTGRES_DISTRIB_DIR") {
-                env.pg_distrib_dir = postgres_bin.into();
-            } else {
-                let cwd = env::current_dir()?;
-                env.pg_distrib_dir = cwd.join("pg_install")
-            }
-        }
-
-        // Find neon binaries.
-        if env.neon_distrib_dir == Path::new("") {
-            env::current_exe()?
-                .parent()
-                .unwrap()
-                .clone_into(&mut env.neon_distrib_dir);
-        }
-
-        if env.pageservers.is_empty() {
-            anyhow::bail!("Configuration must contain at least one pageserver");
-        }
-
-        env.base_data_dir = base_path();
-
-        Ok(env)
-    }
-
-    /// Locate and load config
+    ///  Construct `Self` from on-disk state.
    pub fn load_config() -> anyhow::Result<Self> {
        let repopath = base_path();

@@ -411,38 +444,129 @@ impl LocalEnv {
        // TODO: check that it looks like a neon repository

        // load and parse file
-        let config = fs::read_to_string(repopath.join("config"))?;
-        let mut env: LocalEnv = toml::from_str(config.as_str())?;
+        let config_file_contents = fs::read_to_string(repopath.join("config"))?;
+        let on_disk_config: OnDiskConfig = toml::from_str(config_file_contents.as_str())?;
+        let mut env = {
+            let OnDiskConfig {
+                pg_distrib_dir,
+                neon_distrib_dir,
+                default_tenant_id,
+                private_key_path,
+                broker,
+                storage_controller,
+                pageservers,
+                safekeepers,
+                control_plane_api,
+                control_plane_compute_hook_api,
+                branch_name_mappings,
+            } = on_disk_config;
+            LocalEnv {
+                base_data_dir: repopath.clone(),
+                pg_distrib_dir,
+                neon_distrib_dir,
+                default_tenant_id,
+                private_key_path,
+                broker,
+                storage_controller,
+                pageservers,
+                safekeepers,
+                control_plane_api,
+                control_plane_compute_hook_api,
+                branch_name_mappings,
+            }
+        };

-        env.base_data_dir = repopath;
+        // The source of truth for pageserver configuration is the pageserver.toml.
+        assert!(
+            env.pageservers.is_empty(),
+            "we ensure this during deserialization"
+        );
+        env.pageservers = {
+            let iter = std::fs::read_dir(&repopath).context("open dir")?;
+            let mut pageservers = Vec::new();
+            for res in iter {
+                let dentry = res?;
+                const PREFIX: &str = "pageserver_";
+                let dentry_name = dentry
+                    .file_name()
+                    .into_string()
+                    .ok()
+                    .with_context(|| format!("non-utf8 dentry: {:?}", dentry.path()))
+                    .unwrap();
+                if !dentry_name.starts_with(PREFIX) {
+                    continue;
+                }
+                if !dentry.file_type().context("determine file type")?.is_dir() {
+                    anyhow::bail!("expected a directory, got {:?}", dentry.path());
+                }
+                let id = dentry_name[PREFIX.len()..]
+                    .parse::<NodeId>()
+                    .with_context(|| format!("parse id from {:?}", dentry.path()))?;
+                // TODO(christian): use pageserver_api::config::ConfigToml (PR #7656)
+                #[derive(serde::Serialize, serde::Deserialize)]
+                // (allow unknown fields, unlike PageServerConf)
+                struct PageserverConfigTomlSubset {
+                    id: NodeId,
+                    listen_pg_addr: String,
+                    listen_http_addr: String,
+                    pg_auth_type: AuthType,
+                    http_auth_type: AuthType,
+                }
+                let config_toml_path = dentry.path().join("pageserver.toml");
+                let config_toml: PageserverConfigTomlSubset = toml_edit::de::from_str(
+                    &std::fs::read_to_string(&config_toml_path)
+                        .with_context(|| format!("read {:?}", config_toml_path))?,
+                )
+                .context("parse pageserver.toml")?;
+                let PageserverConfigTomlSubset {
+                    id: config_toml_id,
+                    listen_pg_addr,
+                    listen_http_addr,
+                    pg_auth_type,
+                    http_auth_type,
+                } = config_toml;
+                let conf = PageServerConf {
+                    id: {
+                        anyhow::ensure!(
+                            config_toml_id == id,
+                            "id mismatch: config_toml.id={config_toml_id} id={id}",
+                        );
+                        id
+                    },
+                    listen_pg_addr,
+                    listen_http_addr,
+                    pg_auth_type,
+                    http_auth_type,
+                };
+                pageservers.push(conf);
+            }
+            pageservers
+        };

        Ok(env)
    }

-    pub fn persist_config(&self, base_path: &Path) -> anyhow::Result<()> {
-        // Currently, the user first passes a config file with 'neon_local init --config=<path>'
-        // We read that in, in `create_config`, and fill any missing defaults. Then it's saved
-        // to .neon/config. TODO: We lose any formatting and comments along the way, which is
-        // a bit sad.
-        let mut conf_content = r#"# This file describes a local deployment of the page server
-# and safekeeeper node. It is read by the 'neon_local' command-line
-# utility.
-"#
-        .to_string();
-
-        // Convert the LocalEnv to a toml file.
-        //
-        // This could be as simple as this:
-        //
-        // conf_content += &toml::to_string_pretty(env)?;
-        //
-        // But it results in a "values must be emitted before tables". I'm not sure
-        // why, AFAICS the table, i.e. 'safekeepers: Vec<SafekeeperConf>' is last.
-        // Maybe rust reorders the fields to squeeze avoid padding or something?
-        // In any case, converting to toml::Value first, and serializing that, works.
-        // See https://github.com/alexcrichton/toml-rs/issues/142
-        conf_content += &toml::to_string_pretty(&toml::Value::try_from(self)?)?;
+    pub fn persist_config(&self) -> anyhow::Result<()> {
+        Self::persist_config_impl(
+            &self.base_data_dir,
+            &OnDiskConfig {
+                pg_distrib_dir: self.pg_distrib_dir.clone(),
+                neon_distrib_dir: self.neon_distrib_dir.clone(),
+                default_tenant_id: self.default_tenant_id,
+                private_key_path: self.private_key_path.clone(),
+                broker: self.broker.clone(),
+                storage_controller: self.storage_controller.clone(),
+                pageservers: vec![], // it's skip_serializing anyway
+                safekeepers: self.safekeepers.clone(),
+                control_plane_api: self.control_plane_api.clone(),
+                control_plane_compute_hook_api: self.control_plane_compute_hook_api.clone(),
+                branch_name_mappings: self.branch_name_mappings.clone(),
+            },
+        )
+    }

+    pub fn persist_config_impl(base_path: &Path, config: &OnDiskConfig) -> anyhow::Result<()> {
+        let conf_content = &toml::to_string_pretty(config)?;
        let target_config_path = base_path.join("config");
        fs::write(&target_config_path, conf_content).with_context(|| {
            format!(
@@ -467,17 +591,13 @@ impl LocalEnv {
        }
    }

-    //
-    // Initialize a new Neon repository
-    //
-    pub fn init(&mut self, pg_version: u32, force: &InitForceMode) -> anyhow::Result<()> {
-        // check if config already exists
-        let base_path = &self.base_data_dir;
-        ensure!(
-            base_path != Path::new(""),
-            "repository base path is missing"
-        );
+    /// Materialize the [`NeonLocalInitConf`] to disk. Called during [`neon_local init`].
+    pub fn init(conf: NeonLocalInitConf, force: &InitForceMode) -> anyhow::Result<()> {
+        let base_path = base_path();
+        assert_ne!(base_path, Path::new(""));
+        let base_path = &base_path;

+        // create base_path dir
        if base_path.exists() {
            match force {
                InitForceMode::MustNotExist => {
@@ -509,74 +629,96 @@ impl LocalEnv {
                }
            }
        }
-
-        if !self.pg_bin_dir(pg_version)?.join("postgres").exists() {
-            bail!(
-                "Can't find postgres binary at {}",
-                self.pg_bin_dir(pg_version)?.display()
-            );
-        }
-        for binary in ["pageserver", "safekeeper"] {
-            if !self.neon_distrib_dir.join(binary).exists() {
-                bail!(
-                    "Can't find binary '{binary}' in neon distrib dir '{}'",
-                    self.neon_distrib_dir.display()
-                );
-            }
-        }
-
        if !base_path.exists() {
            fs::create_dir(base_path)?;
        }

+        let NeonLocalInitConf {
+            pg_distrib_dir,
+            neon_distrib_dir,
+            default_tenant_id,
+            broker,
+            storage_controller,
+            pageservers,
+            safekeepers,
+            control_plane_api,
+            control_plane_compute_hook_api,
+        } = conf;
+
+        // Find postgres binaries.
+        // Follow POSTGRES_DISTRIB_DIR if set, otherwise look in "pg_install".
+        // Note that later in the code we assume, that distrib dirs follow the same pattern
+        // for all postgres versions.
+        let pg_distrib_dir = pg_distrib_dir.unwrap_or_else(|| {
+            if let Some(postgres_bin) = env::var_os("POSTGRES_DISTRIB_DIR") {
+                postgres_bin.into()
+            } else {
+                let cwd = env::current_dir().unwrap();
+                cwd.join("pg_install")
+            }
+        });
+
+        // Find neon binaries.
+        let neon_distrib_dir = neon_distrib_dir
+            .unwrap_or_else(|| env::current_exe().unwrap().parent().unwrap().to_owned());
+
        // Generate keypair for JWT.
        //
        // The keypair is only needed if authentication is enabled in any of the
        // components. For convenience, we generate the keypair even if authentication
        // is not enabled, so that you can easily enable it after the initialization
-        // step. However, if the key generation fails, we treat it as non-fatal if
-        // authentication was not enabled.
-        if self.private_key_path == PathBuf::new() {
-            match generate_auth_keys(
-                base_path.join("auth_private_key.pem").as_path(),
-                base_path.join("auth_public_key.pem").as_path(),
-            ) {
-                Ok(()) => {
-                    self.private_key_path = PathBuf::from("auth_private_key.pem");
-                }
-                Err(e) => {
-                    if !self.auth_keys_needed() {
-                        eprintln!("Could not generate keypair for JWT authentication: {e}");
-                        eprintln!("Continuing anyway because authentication was not enabled");
-                        self.private_key_path = PathBuf::from("auth_private_key.pem");
-                    } else {
-                        return Err(e);
-                    }
-                }
-            }
+        // step.
+        generate_auth_keys(
+            base_path.join("auth_private_key.pem").as_path(),
+            base_path.join("auth_public_key.pem").as_path(),
+        )
+        .context("generate auth keys")?;
+        let private_key_path = PathBuf::from("auth_private_key.pem");
+
+        // create the runtime type because the remaining initialization code below needs
+        // a LocalEnv instance op operation
+        // TODO: refactor to avoid this, LocalEnv should only be constructed from on-disk state
+        let env = LocalEnv {
+            base_data_dir: base_path.clone(),
+            pg_distrib_dir,
+            neon_distrib_dir,
+            default_tenant_id: Some(default_tenant_id),
+            private_key_path,
+            broker,
+            storage_controller: storage_controller.unwrap_or_default(),
+            pageservers: pageservers.iter().map(Into::into).collect(),
+            safekeepers,
+            control_plane_api: control_plane_api.unwrap_or_default(),
+            control_plane_compute_hook_api: control_plane_compute_hook_api.unwrap_or_default(),
+            branch_name_mappings: Default::default(),
+        };
+
+        // create endpoints dir
+        fs::create_dir_all(env.endpoints_path())?;
+
+        // create safekeeper dirs
+        for safekeeper in &env.safekeepers {
+            fs::create_dir_all(SafekeeperNode::datadir_path_by_id(&env, safekeeper.id))?;
        }

-        fs::create_dir_all(self.endpoints_path())?;
-
-        for safekeeper in &self.safekeepers {
-            fs::create_dir_all(SafekeeperNode::datadir_path_by_id(self, safekeeper.id))?;
+        // initialize pageserver state
+        for (i, ps) in pageservers.into_iter().enumerate() {
+            let runtime_ps = &env.pageservers[i];
+            assert_eq!(&PageServerConf::from(&ps), runtime_ps);
+            fs::create_dir(env.pageserver_data_dir(ps.id))?;
+            PageServerNode::from_env(&env, runtime_ps)
+                .initialize(ps)
+                .context("pageserver init failed")?;
        }

-        for ps in &self.pageservers {
-            fs::create_dir(self.pageserver_data_dir(ps.id))?;
-        }
+        // setup remote remote location for default LocalFs remote storage
+        std::fs::create_dir_all(env.base_data_dir.join(PAGESERVER_REMOTE_STORAGE_DIR))?;

-        self.persist_config(base_path)
-    }
-
-    fn auth_keys_needed(&self) -> bool {
-        self.pageservers.iter().any(|ps| {
-            ps.pg_auth_type == AuthType::NeonJWT || ps.http_auth_type == AuthType::NeonJWT
-        }) || self.safekeepers.iter().any(|sk| sk.auth_enabled)
+        env.persist_config()
    }
 }

-fn base_path() -> PathBuf {
+pub fn base_path() -> PathBuf {
    match std::env::var_os("NEON_REPO_DIR") {
        Some(val) => PathBuf::from(val),
        None => PathBuf::from(".neon"),
@@ -619,31 +761,3 @@ fn generate_auth_keys(private_key_path: &Path, public_key_path: &Path) -> anyhow
    }
    Ok(())
 }
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn simple_conf_parsing() {
-        let simple_conf_toml = include_str!("../simple.conf");
-        let simple_conf_parse_result = LocalEnv::parse_config(simple_conf_toml);
-        assert!(
-            simple_conf_parse_result.is_ok(),
-            "failed to parse simple config {simple_conf_toml}, reason: {simple_conf_parse_result:?}"
-        );
-
-        let string_to_replace = "listen_addr = '127.0.0.1:50051'";
-        let spoiled_url_str = "listen_addr = '!@$XOXO%^&'";
-        let spoiled_url_toml = simple_conf_toml.replace(string_to_replace, spoiled_url_str);
-        assert!(
-            spoiled_url_toml.contains(spoiled_url_str),
-            "Failed to replace string {string_to_replace} in the toml file {simple_conf_toml}"
-        );
-        let spoiled_url_parse_result = LocalEnv::parse_config(&spoiled_url_toml);
-        assert!(
-            spoiled_url_parse_result.is_err(),
-            "expected toml with invalid Url {spoiled_url_toml} to fail the parsing, but got {spoiled_url_parse_result:?}"
-        );
-    }
-}
--- a/control_plane/src/pageserver.rs
+++ b/control_plane/src/pageserver.rs
@@ -30,7 +30,7 @@ use utils::{
    lsn::Lsn,
 };

-use crate::local_env::PageServerConf;
+use crate::local_env::{NeonLocalInitPageserverConf, PageServerConf};
 use crate::{background_process, local_env::LocalEnv};

 /// Directory within .neon which will be used by default for LocalFs remote storage.
@@ -76,9 +76,11 @@ impl PageServerNode {

    fn pageserver_init_make_toml(
        &self,
-        cli_overrides: toml_edit::Document,
+        conf: NeonLocalInitPageserverConf,
    ) -> anyhow::Result<toml_edit::Document> {
-        // TODO: this is a legacy code, it should be refactored to use toml_edit directly.
+        assert_eq!(&PageServerConf::from(&conf), &self.conf, "during neon_local init, we derive the runtime state of ps conf (self.conf) from the --config flag fully");
+
+        // TODO(christian): instead of what we do here, create a pageserver_api::config::ConfigToml (PR #7656)

        // FIXME: the paths should be shell-escaped to handle paths with spaces, quotas etc.
        let pg_distrib_dir_param = format!(
@@ -86,61 +88,9 @@ impl PageServerNode {
            self.env.pg_distrib_dir_raw().display()
        );

-        let PageServerConf {
-            id,
-            listen_pg_addr,
-            listen_http_addr,
-            pg_auth_type,
-            http_auth_type,
-            virtual_file_io_engine,
-            get_vectored_impl,
-            get_impl,
-            validate_vectored_get,
-        } = &self.conf;
-
-        let id = format!("id={}", id);
-
-        let http_auth_type_param = format!("http_auth_type='{}'", http_auth_type);
-        let listen_http_addr_param = format!("listen_http_addr='{}'", listen_http_addr);
-
-        let pg_auth_type_param = format!("pg_auth_type='{}'", pg_auth_type);
-        let listen_pg_addr_param = format!("listen_pg_addr='{}'", listen_pg_addr);
-        let virtual_file_io_engine = if let Some(virtual_file_io_engine) = virtual_file_io_engine {
-            format!("virtual_file_io_engine='{virtual_file_io_engine}'")
-        } else {
-            String::new()
-        };
-        let get_vectored_impl = if let Some(get_vectored_impl) = get_vectored_impl {
-            format!("get_vectored_impl='{get_vectored_impl}'")
-        } else {
-            String::new()
-        };
-        let get_impl = if let Some(get_impl) = get_impl {
-            format!("get_impl='{get_impl}'")
-        } else {
-            String::new()
-        };
-        let validate_vectored_get = if let Some(validate_vectored_get) = validate_vectored_get {
-            format!("validate_vectored_get={validate_vectored_get}")
-        } else {
-            String::new()
-        };
-
        let broker_endpoint_param = format!("broker_endpoint='{}'", self.env.broker.client_url());

-        let mut overrides = vec![
-            id,
-            pg_distrib_dir_param,
-            http_auth_type_param,
-            pg_auth_type_param,
-            listen_http_addr_param,
-            listen_pg_addr_param,
-            broker_endpoint_param,
-            virtual_file_io_engine,
-            get_vectored_impl,
-            get_impl,
-            validate_vectored_get,
-        ];
+        let mut overrides = vec![pg_distrib_dir_param, broker_endpoint_param];

        if let Some(control_plane_api) = &self.env.control_plane_api {
            overrides.push(format!(
@@ -150,7 +100,7 @@ impl PageServerNode {

            // Storage controller uses the same auth as pageserver: if JWT is enabled
            // for us, we will also need it to talk to them.
-            if matches!(http_auth_type, AuthType::NeonJWT) {
+            if matches!(conf.http_auth_type, AuthType::NeonJWT) {
                let jwt_token = self
                    .env
                    .generate_auth_token(&Claims::new(None, Scope::GenerationsApi))
@@ -159,20 +109,23 @@ impl PageServerNode {
            }
        }

-        if !cli_overrides.contains_key("remote_storage") {
+        if !conf.other.contains_key("remote_storage") {
            overrides.push(format!(
                "remote_storage={{local_path='../{PAGESERVER_REMOTE_STORAGE_DIR}'}}"
            ));
        }

-        if *http_auth_type != AuthType::Trust || *pg_auth_type != AuthType::Trust {
+        if conf.http_auth_type != AuthType::Trust || conf.pg_auth_type != AuthType::Trust {
            // Keys are generated in the toplevel repo dir, pageservers' workdirs
            // are one level below that, so refer to keys with ../
            overrides.push("auth_validation_public_key_path='../auth_public_key.pem'".to_owned());
        }

        // Apply the user-provided overrides
-        overrides.push(cli_overrides.to_string());
+        overrides.push(
+            toml_edit::ser::to_string_pretty(&conf)
+                .expect("we deserialized this from toml earlier"),
+        );

        // Turn `overrides` into a toml document.
        // TODO: above code is legacy code, it should be refactored to use toml_edit directly.
@@ -188,8 +141,8 @@ impl PageServerNode {
    }

    /// Initializes a pageserver node by creating its config with the overrides provided.
-    pub fn initialize(&self, config_overrides: toml_edit::Document) -> anyhow::Result<()> {
-        self.pageserver_init(config_overrides)
+    pub fn initialize(&self, conf: NeonLocalInitPageserverConf) -> anyhow::Result<()> {
+        self.pageserver_init(conf)
            .with_context(|| format!("Failed to run init for pageserver node {}", self.conf.id))
    }

@@ -209,7 +162,7 @@ impl PageServerNode {
        self.start_node().await
    }

-    fn pageserver_init(&self, cli_overrides: toml_edit::Document) -> anyhow::Result<()> {
+    fn pageserver_init(&self, conf: NeonLocalInitPageserverConf) -> anyhow::Result<()> {
        let datadir = self.repo_path();
        let node_id = self.conf.id;
        println!(
@@ -221,7 +174,7 @@ impl PageServerNode {
        io::stdout().flush()?;

        let config = self
-            .pageserver_init_make_toml(cli_overrides)
+            .pageserver_init_make_toml(conf)
            .context("make pageserver toml")?;
        let config_file_path = datadir.join("pageserver.toml");
        let mut config_file = std::fs::OpenOptions::new()
--- a/control_plane/src/storage_controller.rs
+++ b/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,
--- a/docker-compose/compute_wrapper/Dockerfile
+++ b/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

--- a/docker-compose/docker_compose_test.sh
+++ b/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 )"
--- a/libs/compute_api/src/responses.rs
+++ b/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.
--- a/libs/pageserver_api/src/keyspace.rs
+++ b/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
--- a/libs/pageserver_api/src/models.rs
+++ b/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
+        );
+    }
 }
--- a/libs/pageserver_api/src/shard.rs
+++ b/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)
--- a/libs/postgres_backend/src/lib.rs
+++ b/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;
    }
 }

--- a/libs/postgres_connection/src/lib.rs
+++ b/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")`
--- a/libs/postgres_ffi/wal_craft/src/lib.rs
+++ b/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]);
+        }
    }
 }

--- a/libs/remote_storage/src/s3_bucket.rs
+++ b/libs/remote_storage/src/s3_bucket.rs
@@ -27,7 +27,7 @@ use aws_config::{
 };
 use aws_credential_types::provider::SharedCredentialsProvider;
 use aws_sdk_s3::{
-    config::{AsyncSleep, Builder, IdentityCache, Region, SharedAsyncSleep},
+    config::{AsyncSleep, IdentityCache, Region, SharedAsyncSleep},
    error::SdkError,
    operation::get_object::GetObjectError,
    types::{Delete, DeleteMarkerEntry, ObjectIdentifier, ObjectVersion, StorageClass},
@@ -75,13 +75,13 @@ struct GetObjectRequest {
 }
 impl S3Bucket {
    /// Creates the S3 storage, errors if incorrect AWS S3 configuration provided.
-    pub fn new(aws_config: &S3Config, timeout: Duration) -> anyhow::Result<Self> {
+    pub fn new(remote_storage_config: &S3Config, timeout: Duration) -> anyhow::Result<Self> {
        tracing::debug!(
            "Creating s3 remote storage for S3 bucket {}",
-            aws_config.bucket_name
+            remote_storage_config.bucket_name
        );

-        let region = Some(Region::new(aws_config.bucket_region.clone()));
+        let region = Some(Region::new(remote_storage_config.bucket_region.clone()));

        let provider_conf = ProviderConfig::without_region().with_region(region.clone());

@@ -113,6 +113,38 @@ impl S3Bucket {
        // AWS SDK requires us to specify how the RetryConfig should sleep when it wants to back off
        let sleep_impl: Arc<dyn AsyncSleep> = Arc::new(TokioSleep::new());

+        let sdk_config_loader: aws_config::ConfigLoader = aws_config::defaults(
+            #[allow(deprecated)] /* TODO: https://github.com/neondatabase/neon/issues/7665 */
+            BehaviorVersion::v2023_11_09(),
+        )
+        .region(region)
+        .identity_cache(IdentityCache::lazy().build())
+        .credentials_provider(SharedCredentialsProvider::new(credentials_provider))
+        .sleep_impl(SharedAsyncSleep::from(sleep_impl));
+
+        let sdk_config: aws_config::SdkConfig = std::thread::scope(|s| {
+            s.spawn(|| {
+                // TODO: make this function async.
+                tokio::runtime::Builder::new_current_thread()
+                    .enable_all()
+                    .build()
+                    .unwrap()
+                    .block_on(sdk_config_loader.load())
+            })
+            .join()
+            .unwrap()
+        });
+
+        let mut s3_config_builder = aws_sdk_s3::config::Builder::from(&sdk_config);
+
+        // Technically, the `remote_storage_config.endpoint` field only applies to S3 interactions.
+        // (In case we ever re-use the `sdk_config` for more than just the S3 client in the future)
+        if let Some(custom_endpoint) = remote_storage_config.endpoint.clone() {
+            s3_config_builder = s3_config_builder
+                .endpoint_url(custom_endpoint)
+                .force_path_style(true);
+        }
+
        // We do our own retries (see [`backoff::retry`]).  However, for the AWS SDK to enable rate limiting in response to throttling
        // responses (e.g. 429 on too many ListObjectsv2 requests), we must provide a retry config.  We set it to use at most one
        // attempt, and enable 'Adaptive' mode, which causes rate limiting to be enabled.
@@ -120,42 +152,36 @@ impl S3Bucket {
        retry_config
            .set_max_attempts(Some(1))
            .set_mode(Some(RetryMode::Adaptive));
+        s3_config_builder = s3_config_builder.retry_config(retry_config.build());

-        let mut config_builder = Builder::default()
-            .behavior_version(BehaviorVersion::v2023_11_09())
-            .region(region)
-            .identity_cache(IdentityCache::lazy().build())
-            .credentials_provider(SharedCredentialsProvider::new(credentials_provider))
-            .retry_config(retry_config.build())
-            .sleep_impl(SharedAsyncSleep::from(sleep_impl));
+        let s3_config = s3_config_builder.build();
+        let client = aws_sdk_s3::Client::from_conf(s3_config);

-        if let Some(custom_endpoint) = aws_config.endpoint.clone() {
-            config_builder = config_builder
-                .endpoint_url(custom_endpoint)
-                .force_path_style(true);
-        }
+        let prefix_in_bucket = remote_storage_config
+            .prefix_in_bucket
+            .as_deref()
+            .map(|prefix| {
+                let mut prefix = prefix;
+                while prefix.starts_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix = &prefix[1..]
+                }

-        let client = Client::from_conf(config_builder.build());
+                let mut prefix = prefix.to_string();
+                while prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix.pop();
+                }
+                prefix
+            });

-        let prefix_in_bucket = aws_config.prefix_in_bucket.as_deref().map(|prefix| {
-            let mut prefix = prefix;
-            while prefix.starts_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
-                prefix = &prefix[1..]
-            }
-
-            let mut prefix = prefix.to_string();
-            while prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
-                prefix.pop();
-            }
-            prefix
-        });
        Ok(Self {
            client,
-            bucket_name: aws_config.bucket_name.clone(),
-            max_keys_per_list_response: aws_config.max_keys_per_list_response,
+            bucket_name: remote_storage_config.bucket_name.clone(),
+            max_keys_per_list_response: remote_storage_config.max_keys_per_list_response,
            prefix_in_bucket,
-            concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
-            upload_storage_class: aws_config.upload_storage_class.clone(),
+            concurrency_limiter: ConcurrencyLimiter::new(
+                remote_storage_config.concurrency_limit.get(),
+            ),
+            upload_storage_class: remote_storage_config.upload_storage_class.clone(),
            timeout,
        })
    }
--- a/libs/safekeeper_api/src/models.rs
+++ b/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)]
--- a/libs/utils/src/failpoint_support.rs
+++ b/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
--- a/libs/utils/src/poison.rs
+++ b/libs/utils/src/poison.rs
@@ -3,7 +3,7 @@
 //!  # Example
 //!
 //!  ```
-//!  # tokio_test::block_on(async {
+//!  # tokio::runtime::Builder::new_current_thread().enable_all().build().unwrap().block_on(async {
 //!  use utils::poison::Poison;
 //!  use std::time::Duration;
 //!
--- a/libs/utils/src/sync/gate.rs
+++ b/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;
--- a/libs/walproposer/src/walproposer.rs
+++ b/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,
--- a/pageserver/benches/bench_layer_map.rs
+++ b/pageserver/benches/bench_layer_map.rs
@@ -1,7 +1,7 @@
 use pageserver::keyspace::{KeyPartitioning, KeySpace};
 use pageserver::repository::Key;
 use pageserver::tenant::layer_map::LayerMap;
-use pageserver::tenant::storage_layer::LayerFileName;
+use pageserver::tenant::storage_layer::LayerName;
 use pageserver::tenant::storage_layer::PersistentLayerDesc;
 use pageserver_api::shard::TenantShardId;
 use rand::prelude::{SeedableRng, SliceRandom, StdRng};
@@ -28,7 +28,7 @@ fn build_layer_map(filename_dump: PathBuf) -> LayerMap {
    let mut updates = layer_map.batch_update();
    for fname in filenames {
        let fname = fname.unwrap();
-        let fname = LayerFileName::from_str(&fname).unwrap();
+        let fname = LayerName::from_str(&fname).unwrap();
        let layer = PersistentLayerDesc::from(fname);

        let lsn_range = layer.get_lsn_range();
--- a/pageserver/benches/bench_walredo.rs
+++ b/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(
--- a/pageserver/client/src/mgmt_api.rs
+++ b/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))
+                }
+            }),
+        }
+    }
 }
--- a/pageserver/compaction/src/bin/compaction-simulator.rs
+++ b/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}",
--- a/pageserver/compaction/src/compact_tiered.rs
+++ b/pageserver/compaction/src/compact_tiered.rs
@@ -24,7 +24,9 @@ use tracing::{debug, info};
 use std::collections::{HashSet, VecDeque};
 use std::ops::Range;

-use crate::helpers::{accum_key_values, keyspace_total_size, merge_delta_keys, overlaps_with};
+use crate::helpers::{
+    accum_key_values, keyspace_total_size, merge_delta_keys_buffered, overlaps_with, PAGE_SZ,
+};
 use crate::interface::*;
 use utils::lsn::Lsn;

@@ -104,7 +106,13 @@ pub async fn compact_tiered<E: CompactionJobExecutor>(
            ctx,
        )
        .await?;
-        if target_file_size == u64::MAX {
+        if current_level_target_height == u64::MAX {
+            // our target height includes all possible lsns
+            info!(
+                level = current_level_no,
+                depth = depth,
+                "compaction loop reached max current_level_target_height"
+            );
            break;
        }
        current_level_no += 1;
@@ -371,7 +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
@@ -433,7 +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> {
@@ -522,8 +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?

@@ -535,43 +541,101 @@ where
            }
        }
        // Open stream
-        let key_value_stream = std::pin::pin!(merge_delta_keys::<E>(deltas.as_slice(), ctx));
+        let key_value_stream =
+            std::pin::pin!(merge_delta_keys_buffered::<E>(deltas.as_slice(), ctx)
+                .await?
+                .map(Result::<_, anyhow::Error>::Ok));
        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);
+                    }
+                }
            }
        }

@@ -599,8 +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>,

@@ -740,9 +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>>,

@@ -766,11 +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;
@@ -792,6 +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();
@@ -854,7 +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);
--- a/pageserver/compaction/src/helpers.rs
+++ b/pageserver/compaction/src/helpers.rs
@@ -9,10 +9,14 @@ use pageserver_api::shard::ShardIdentity;
 use pin_project_lite::pin_project;
 use std::collections::BinaryHeap;
 use std::collections::VecDeque;
+use std::fmt::Display;
 use std::future::Future;
 use std::ops::{DerefMut, Range};
 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>,
@@ -108,17 +112,40 @@ pub fn merge_delta_keys<'a, E: CompactionJobExecutor>(
    }
 }

+pub async fn merge_delta_keys_buffered<'a, E: CompactionJobExecutor + 'a>(
+    layers: &'a [E::DeltaLayer],
+    ctx: &'a E::RequestContext,
+) -> anyhow::Result<impl Stream<Item = <E::DeltaLayer as CompactionDeltaLayer<E>>::DeltaEntry<'a>>>
+{
+    let mut keys = Vec::new();
+    for l in layers {
+        // Boxing and casting to LoadFuture is required to obtain the right Sync bound.
+        // If we do l.load_keys(ctx).await? directly, there is a compilation error.
+        let load_future: LoadFuture<'a, _> = Box::pin(l.load_keys(ctx));
+        keys.extend(load_future.await?.into_iter());
+    }
+    keys.sort_by_key(|k| (k.key(), k.lsn()));
+    let stream = futures::stream::iter(keys.into_iter());
+    Ok(stream)
+}
+
 enum LazyLoadLayer<'a, E: CompactionJobExecutor> {
    Loaded(VecDeque<<E::DeltaLayer as CompactionDeltaLayer<E>>::DeltaEntry<'a>>),
    Unloaded(&'a E::DeltaLayer),
 }
 impl<'a, E: CompactionJobExecutor> LazyLoadLayer<'a, E> {
-    fn key(&self) -> E::Key {
+    fn min_key(&self) -> E::Key {
        match self {
            Self::Loaded(entries) => entries.front().unwrap().key(),
            Self::Unloaded(dl) => dl.key_range().start,
        }
    }
+    fn min_lsn(&self) -> Lsn {
+        match self {
+            Self::Loaded(entries) => entries.front().unwrap().lsn(),
+            Self::Unloaded(dl) => dl.lsn_range().start,
+        }
+    }
 }
 impl<'a, E: CompactionJobExecutor> PartialOrd for LazyLoadLayer<'a, E> {
    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
@@ -128,12 +155,12 @@ impl<'a, E: CompactionJobExecutor> PartialOrd for LazyLoadLayer<'a, E> {
 impl<'a, E: CompactionJobExecutor> Ord for LazyLoadLayer<'a, E> {
    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
        // reverse order so that we get a min-heap
-        other.key().cmp(&self.key())
+        (other.min_key(), other.min_lsn()).cmp(&(self.min_key(), self.min_lsn()))
    }
 }
 impl<'a, E: CompactionJobExecutor> PartialEq for LazyLoadLayer<'a, E> {
    fn eq(&self, other: &Self) -> bool {
-        self.key().eq(&other.key())
+        self.cmp(other) == std::cmp::Ordering::Equal
    }
 }
 impl<'a, E: CompactionJobExecutor> Eq for LazyLoadLayer<'a, E> {}
@@ -210,11 +237,16 @@ 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,
+    K: Eq + PartialOrd + Display + Copy,
    I: Stream<Item = Result<D, E>>,
    D: CompactionDeltaEntry<'a, K>,
 {
@@ -224,25 +256,39 @@ 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(),
                    };
                }
            }
+            assert!(last_key <= accum.key, "last_key={last_key} <= accum.key={}", accum.key);
            yield accum;
        }
    }
--- a/pageserver/compaction/src/identify_levels.rs
+++ b/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,
--- a/pageserver/compaction/src/simulator.rs
+++ b/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!(
--- a/pageserver/compaction/tests/tests.rs
+++ b/pageserver/compaction/tests/tests.rs
@@ -1,23 +1,35 @@
+use once_cell::sync::OnceCell;
 use pageserver_compaction::interface::CompactionLayer;
 use pageserver_compaction::simulator::MockTimeline;
+use utils::logging;
+
+static LOG_HANDLE: OnceCell<()> = OnceCell::new();
+
+pub(crate) fn setup_logging() {
+    LOG_HANDLE.get_or_init(|| {
+        logging::init(
+            logging::LogFormat::Test,
+            logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
+            logging::Output::Stdout,
+        )
+        .expect("Failed to init test logging")
+    });
+}

 /// Test the extreme case that there are so many updates for a single key that
 /// 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();
    let mut executor = MockTimeline::new();
-    executor.target_file_size = 10_000_000; // 10 MB
+    executor.target_file_size = 1_000_000; // 1 MB

-    // Ingest 100 MB of updates to a single key.
+    // Ingest 10 MB of updates to a single key.
    for _ in 1..1000 {
        executor.ingest_uniform(100, 10, &(0..100_000)).unwrap();
-        executor.ingest_uniform(10_000, 10, &(0..1)).unwrap();
+        executor.ingest_uniform(1000, 10, &(0..1)).unwrap();
        executor.compact().await.unwrap();
    }

@@ -27,9 +39,32 @@ 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);
        }
    }
 }
+
+#[tokio::test]
+async fn test_simple_updates() {
+    setup_logging();
+    let mut executor = MockTimeline::new();
+    executor.target_file_size = 500_000; // 500 KB
+
+    // Ingest some traffic.
+    for _ in 1..400 {
+        executor.ingest_uniform(100, 500, &(0..100_000)).unwrap();
+    }
+
+    for l in executor.live_layers.iter() {
+        println!("layer {}: {}", l.short_id(), l.file_size());
+    }
+
+    println!("Running compaction...");
+    executor.compact().await.unwrap();
+
+    for l in executor.live_layers.iter() {
+        println!("layer {}: {}", l.short_id(), l.file_size());
+    }
+}
--- a/pageserver/ctl/src/draw_timeline_dir.rs
+++ b/pageserver/ctl/src/draw_timeline_dir.rs
@@ -28,6 +28,8 @@
 //! # From an `index_part.json` in S3
 //! (jq -r '.layer_metadata | keys[]' | cargo  run -p pagectl draw-timeline ) < index_part.json-00000016 > out.svg
 //!
+//! # enrich with lines for gc_cutoff and a child branch point
+//! cat <(jq -r '.historic_layers[] | .layer_file_name' < layers.json) <(echo -e 'gc_cutoff:0000001CE3FE32C9\nbranch:0000001DE3FE32C9') | cargo run --bin pagectl draw-timeline >| out.svg
 //! ```
 //!
 //! ## Viewing
@@ -48,9 +50,8 @@
 //! ```
 //!

-use anyhow::Result;
+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;
@@ -81,6 +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]);
    }
@@ -90,6 +96,33 @@ fn parse_filename(name: &str) -> (Range<Key>, Range<Lsn>) {
    (keys, lsns)
 }

+#[derive(Clone, Copy)]
+enum LineKind {
+    GcCutoff,
+    Branch,
+}
+
+impl From<LineKind> for Fill {
+    fn from(value: LineKind) -> Self {
+        match value {
+            LineKind::GcCutoff => Fill::Color(rgb(255, 0, 0)),
+            LineKind::Branch => Fill::Color(rgb(0, 255, 0)),
+        }
+    }
+}
+
+impl FromStr for LineKind {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> std::prelude::v1::Result<Self, Self::Err> {
+        Ok(match s {
+            "gc_cutoff" => LineKind::GcCutoff,
+            "branch" => LineKind::Branch,
+            _ => anyhow::bail!("unsupported linekind: {s}"),
+        })
+    }
+}
+
 pub fn main() -> Result<()> {
    // Parse layer filenames from stdin
    struct Layer {
@@ -99,15 +132,32 @@ pub fn main() -> Result<()> {
    }
    let mut files: Vec<Layer> = vec![];
    let stdin = io::stdin();
-    for line in stdin.lock().lines() {
+
+    let mut lines: Vec<(Lsn, LineKind)> = vec![];
+
+    for (lineno, line) in stdin.lock().lines().enumerate() {
+        let lineno = lineno + 1;
+
        let line = line.unwrap();
+        if let Some((kind, lsn)) = line.split_once(':') {
+            let (kind, lsn) = LineKind::from_str(kind)
+                .context("parse kind")
+                .and_then(|kind| {
+                    if lsn.contains('/') {
+                        Lsn::from_str(lsn)
+                    } else {
+                        Lsn::from_hex(lsn)
+                    }
+                    .map(|lsn| (kind, lsn))
+                    .context("parse lsn")
+                })
+                .with_context(|| format!("parse {line:?} on {lineno}"))?;
+            lines.push((lsn, kind));
+            continue;
+        }
        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(),
@@ -117,8 +167,9 @@ pub fn main() -> Result<()> {
    }

    // Collect all coordinates
-    let mut keys: Vec<Key> = vec![];
-    let mut lsns: Vec<Lsn> = vec![];
+    let mut keys: Vec<Key> = Vec::with_capacity(files.len());
+    let mut lsns: Vec<Lsn> = Vec::with_capacity(files.len() + lines.len());
+
    for Layer {
        key_range: keyr,
        lsn_range: lsnr,
@@ -131,6 +182,8 @@ pub fn main() -> Result<()> {
        lsns.push(lsnr.end);
    }

+    lsns.extend(lines.iter().map(|(lsn, _)| *lsn));
+
    // Analyze
    let key_map = build_coordinate_compression_map(keys);
    let lsn_map = build_coordinate_compression_map(lsns);
@@ -144,10 +197,13 @@ pub fn main() -> Result<()> {
    println!(
        "{}",
        BeginSvg {
-            w: key_map.len() as f32,
+            w: (key_map.len() + 10) as f32,
            h: stretch * lsn_map.len() as f32
        }
    );
+
+    let xmargin = 0.05; // Height-dependent margin to disambiguate overlapping deltas
+
    for Layer {
        filename,
        key_range: keyr,
@@ -169,7 +225,6 @@ pub fn main() -> Result<()> {
        let mut lsn_diff = (lsn_end - lsn_start) as f32;
        let mut fill = Fill::None;
        let mut ymargin = 0.05 * lsn_diff; // Height-dependent margin to disambiguate overlapping deltas
-        let xmargin = 0.05; // Height-dependent margin to disambiguate overlapping deltas
        let mut lsn_offset = 0.0;

        // Fill in and thicken rectangle if it's an
@@ -189,7 +244,7 @@ pub fn main() -> Result<()> {
        println!(
            "    {}",
            rectangle(
-                key_start as f32 + stretch * xmargin,
+                5.0 + key_start as f32 + stretch * xmargin,
                stretch * (lsn_max as f32 - (lsn_end as f32 - ymargin - lsn_offset)),
                key_diff as f32 - stretch * 2.0 * xmargin,
                stretch * (lsn_diff - 2.0 * ymargin)
@@ -200,6 +255,26 @@ pub fn main() -> Result<()> {
            .comment(filename)
        );
    }
+
+    for (lsn, kind) in lines {
+        let lsn_start = *lsn_map.get(&lsn).unwrap();
+        let lsn_end = lsn_start;
+        let stretch = 2.0;
+        let lsn_diff = 0.3;
+        let lsn_offset = -lsn_diff / 2.0;
+        let ymargin = 0.05;
+        println!(
+            "{}",
+            rectangle(
+                0.0f32 + stretch * xmargin,
+                stretch * (lsn_map.len() as f32 - (lsn_end as f32 - ymargin - lsn_offset)),
+                (key_map.len() + 10) as f32,
+                stretch * (lsn_diff - 2.0 * ymargin)
+            )
+            .fill(kind)
+        );
+    }
+
    println!("{}", EndSvg);

    eprintln!("num_images: {}", num_images);
--- a/pageserver/ctl/src/index_part.rs
+++ b/pageserver/ctl/src/index_part.rs
@@ -2,8 +2,8 @@ use std::collections::HashMap;

 use anyhow::Context;
 use camino::Utf8PathBuf;
-use pageserver::tenant::remote_timeline_client::index::IndexLayerMetadata;
-use pageserver::tenant::storage_layer::LayerFileName;
+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<LayerFileName, IndexLayerMetadata>,
+                layer_metadata: &'a HashMap<LayerName, LayerFileMetadata>,
                disk_consistent_lsn: Lsn,
                timeline_metadata: &'a TimelineMetadata,
            }
--- a/pageserver/ctl/src/layer_map_analyzer.rs
+++ b/pageserver/ctl/src/layer_map_analyzer.rs
@@ -100,7 +100,7 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {

 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
 async fn get_holes(path: &Utf8Path, max_holes: usize, ctx: &RequestContext) -> Result<Vec<Hole>> {
-    let file = VirtualFile::open(path).await?;
+    let file = VirtualFile::open(path, ctx).await?;
    let file_id = page_cache::next_file_id();
    let block_reader = FileBlockReader::new(&file, file_id);
    let summary_blk = block_reader.read_blk(0, ctx).await?;
--- a/pageserver/ctl/src/layers.rs
+++ b/pageserver/ctl/src/layers.rs
@@ -61,7 +61,7 @@ async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result
    let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
    page_cache::init(100);
-    let file = VirtualFile::open(path).await?;
+    let file = VirtualFile::open(path, ctx).await?;
    let file_id = page_cache::next_file_id();
    let block_reader = FileBlockReader::new(&file, file_id);
    let summary_blk = block_reader.read_blk(0, ctx).await?;
--- a/pageserver/ctl/src/main.rs
+++ b/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,
--- a/pageserver/pagebench/src/cmd/aux_files.rs
+++ b/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(())
+}
--- a/pageserver/pagebench/src/cmd/ondemand_download_churn.rs
+++ b/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
            }
        };
--- a/pageserver/pagebench/src/main.rs
+++ b/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()
 }
--- a/pageserver/src/aux_file.rs
+++ b/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(),
        );
    }

--- a/pageserver/src/basebackup.rs
+++ b/pageserver/src/basebackup.rs
@@ -601,7 +601,7 @@ where
        // add zenith.signal file
        let mut zenith_signal = String::new();
        if self.prev_record_lsn == Lsn(0) {
-            if self.lsn == self.timeline.get_ancestor_lsn() {
+            if self.timeline.is_ancestor_lsn(self.lsn) {
                write!(zenith_signal, "PREV LSN: none")
                    .map_err(|e| BasebackupError::Server(e.into()))?;
            } else {
--- a/pageserver/src/bin/pageserver.rs
+++ b/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
@@ -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!()
        })
    }
@@ -708,12 +697,11 @@ fn start_pageserver(

 fn create_remote_storage_client(
    conf: &'static PageServerConf,
-) -> anyhow::Result<Option<GenericRemoteStorage>> {
+) -> anyhow::Result<GenericRemoteStorage> {
    let config = if let Some(config) = &conf.remote_storage_config {
        config
    } else {
-        tracing::warn!("no remote storage configured, this is a deprecated configuration");
-        return Ok(None);
+        anyhow::bail!("no remote storage configured, this is a deprecated configuration");
    };

    // Create the client
@@ -733,7 +721,7 @@ fn create_remote_storage_client(
            GenericRemoteStorage::unreliable_wrapper(remote_storage, conf.test_remote_failures);
    }

-    Ok(Some(remote_storage))
+    Ok(remote_storage)
 }

 fn cli() -> Command {
--- a/pageserver/src/config.rs
+++ b/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)
    }

--- a/pageserver/src/deletion_queue.rs
+++ b/pageserver/src/deletion_queue.rs
@@ -38,7 +38,7 @@ use deleter::DeleterMessage;
 use list_writer::ListWriterQueueMessage;
 use validator::ValidatorQueueMessage;

-use crate::{config::PageServerConf, tenant::storage_layer::LayerFileName};
+use crate::{config::PageServerConf, tenant::storage_layer::LayerName};

 // TODO: configurable for how long to wait before executing deletions

@@ -479,7 +479,7 @@ impl DeletionQueueClient {
        tenant_shard_id: TenantShardId,
        timeline_id: TimelineId,
        current_generation: Generation,
-        layers: Vec<(LayerFileName, LayerFileMetadata)>,
+        layers: Vec<(LayerName, LayerFileMetadata)>,
    ) -> Result<(), DeletionQueueError> {
        if current_generation.is_none() {
            debug!("Enqueuing deletions in legacy mode, skipping queue");
@@ -511,7 +511,7 @@ impl DeletionQueueClient {
        tenant_shard_id: TenantShardId,
        timeline_id: TimelineId,
        current_generation: Generation,
-        layers: Vec<(LayerFileName, LayerFileMetadata)>,
+        layers: Vec<(LayerName, LayerFileMetadata)>,
    ) -> Result<(), DeletionQueueError> {
        metrics::DELETION_QUEUE
            .keys_submitted
@@ -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 {
@@ -734,20 +717,20 @@ mod test {
    use crate::{
        control_plane_client::RetryForeverError,
        repository::Key,
-        tenant::{harness::TenantHarness, storage_layer::DeltaFileName},
+        tenant::{harness::TenantHarness, storage_layer::DeltaLayerName},
    };

    use super::*;
    pub const TIMELINE_ID: TimelineId =
        TimelineId::from_array(hex!("11223344556677881122334455667788"));

-    pub const EXAMPLE_LAYER_NAME: LayerFileName = LayerFileName::Delta(DeltaFileName {
+    pub const EXAMPLE_LAYER_NAME: LayerName = LayerName::Delta(DeltaLayerName {
        key_range: Key::from_i128(0x0)..Key::from_i128(0xFFFFFFFFFFFFFFFF),
        lsn_range: Lsn(0x00000000016B59D8)..Lsn(0x00000000016B5A51),
    });

    // When you need a second layer in a test.
-    pub const EXAMPLE_LAYER_NAME_ALT: LayerFileName = LayerFileName::Delta(DeltaFileName {
+    pub const EXAMPLE_LAYER_NAME_ALT: LayerName = LayerName::Delta(DeltaLayerName {
        key_range: Key::from_i128(0x0)..Key::from_i128(0xFFFFFFFFFFFFFFFF),
        lsn_range: Lsn(0x00000000016B5A51)..Lsn(0x00000000016B5A61),
    });
@@ -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,
            );
@@ -797,7 +780,7 @@ mod test {
        /// Returns remote layer file name, suitable for use in assert_remote_files
        fn write_remote_layer(
            &self,
-            file_name: LayerFileName,
+            file_name: LayerName,
            gen: Generation,
        ) -> anyhow::Result<String> {
            let tenant_shard_id = self.harness.tenant_shard_id;
@@ -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,
        );
@@ -952,7 +935,7 @@ mod test {
        let client = ctx.deletion_queue.new_client();
        client.recover(HashMap::new())?;

-        let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
+        let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
        let tenant_shard_id = ctx.harness.tenant_shard_id;

        let content: Vec<u8> = "victim1 contents".into();
--- a/pageserver/src/deletion_queue/list_writer.rs
+++ b/pageserver/src/deletion_queue/list_writer.rs
@@ -34,7 +34,7 @@ use crate::deletion_queue::TEMP_SUFFIX;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::remote_timeline_client::LayerFileMetadata;
-use crate::tenant::storage_layer::LayerFileName;
+use crate::tenant::storage_layer::LayerName;
 use crate::virtual_file::on_fatal_io_error;
 use crate::virtual_file::MaybeFatalIo;

@@ -59,7 +59,7 @@ pub(super) struct DeletionOp {
    // `layers` and `objects` are both just lists of objects.  `layers` is used if you do not
    // have a config object handy to project it to a remote key, and need the consuming worker
    // to do it for you.
-    pub(super) layers: Vec<(LayerFileName, LayerFileMetadata)>,
+    pub(super) layers: Vec<(LayerName, LayerFileMetadata)>,
    pub(super) objects: Vec<RemotePath>,

    /// The _current_ generation of the Tenant shard attachment in which we are enqueuing
--- a/pageserver/src/disk_usage_eviction_task.rs
+++ b/pageserver/src/disk_usage_eviction_task.rs
@@ -64,7 +64,7 @@ use crate::{
        mgr::TenantManager,
        remote_timeline_client::LayerFileMetadata,
        secondary::SecondaryTenant,
-        storage_layer::{AsLayerDesc, EvictionError, Layer, LayerFileName},
+        storage_layer::{AsLayerDesc, EvictionError, Layer, LayerName},
    },
 };

@@ -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)
                    });
@@ -604,7 +598,7 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 pub(crate) struct EvictionSecondaryLayer {
    pub(crate) secondary_tenant: Arc<SecondaryTenant>,
    pub(crate) timeline_id: TimelineId,
-    pub(crate) name: LayerFileName,
+    pub(crate) name: LayerName,
    pub(crate) metadata: LayerFileMetadata,
 }

@@ -637,9 +631,9 @@ impl EvictionLayer {
        }
    }

-    pub(crate) fn get_name(&self) -> LayerFileName {
+    pub(crate) fn get_name(&self) -> LayerName {
        match self {
-            Self::Attached(l) => l.layer_desc().filename(),
+            Self::Attached(l) => l.layer_desc().layer_name(),
            Self::Secondary(sl) => sl.name.clone(),
        }
    }
@@ -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,
        }
    }
 }
--- a/pageserver/src/http/openapi_spec.yml
+++ b/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
@@ -420,25 +451,6 @@ paths:
          description: Tenant scheduled to load successfully

  /v1/tenant/{tenant_id}/synthetic_size:
-    parameters:
-      - name: tenant_id
-        in: path
-        required: true
-        schema:
-          type: string
-    get:
-      description: |
-        Calculate tenant's synthetic size
-      responses:
-        "200":
-          description: Tenant's synthetic size
-          content:
-            application/json:
-              schema:
-                $ref: "#/components/schemas/SyntheticSizeResponse"
-
-  # This route has no handler. TODO: remove?
-  /v1/tenant/{tenant_id}/size:
    parameters:
      - name: tenant_id
        in: path
@@ -468,19 +480,9 @@ paths:
          content:
            application/json:
              schema:
-                type: object
-                required:
-                  - id
-                  - size
-                properties:
-                  id:
-                    type: string
-                    format: hex
-                  size:
-                    type: integer
-                    nullable: true
-                    description: |
-                      Size metric in bytes or null if inputs_only=true was given.
+                $ref: "#/components/schemas/SyntheticSizeResponse"
+            text/html:
+              description: SVG representation of the tenant and it's timelines.
        "401":
          description: Unauthorized Error
          content:
@@ -610,6 +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
@@ -929,6 +1005,9 @@ components:
          format: hex
        size:
          type: integer
+          nullable: true
+          description: |
+            Size metric in bytes or null if inputs_only=true was given.
        segment_sizes:
          type: array
          items:
@@ -1006,6 +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:
@@ -1063,6 +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
--- a/pageserver/src/http/routes.rs
+++ b/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,
@@ -63,9 +72,10 @@ use crate::tenant::remote_timeline_client::list_remote_timelines;
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
-use crate::tenant::storage_layer::LayerFileName;
+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);
@@ -1229,7 +1238,7 @@ async fn layer_download_handler(
    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
    let layer_file_name = get_request_param(&request, "layer_file_name")?;
    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
-    let layer_name = LayerFileName::from_str(layer_file_name)
+    let layer_name = LayerName::from_str(layer_file_name)
        .map_err(|s| ApiError::BadRequest(anyhow::anyhow!(s)))?;
    let state = get_state(&request);

@@ -1261,7 +1270,7 @@ async fn evict_timeline_layer_handler(
    let layer_file_name = get_request_param(&request, "layer_file_name")?;
    let state = get_state(&request);

-    let layer_name = LayerFileName::from_str(layer_file_name)
+    let layer_name = LayerName::from_str(layer_file_name)
        .map_err(|s| ApiError::BadRequest(anyhow::anyhow!(s)))?;

    let timeline =
@@ -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))
 }
--- a/pageserver/src/lib.rs
+++ b/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";
--- a/pageserver/src/metrics.rs
+++ b/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,
--- a/pageserver/src/page_service.rs
+++ b/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"),
--- a/pageserver/src/pgdatadir_mapping.rs
+++ b/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;

@@ -1671,7 +1732,7 @@ impl<'a> DatadirModification<'a> {
        }

        if !self.pending_deletions.is_empty() {
-            writer.delete_batch(&self.pending_deletions).await?;
+            writer.delete_batch(&self.pending_deletions, ctx).await?;
            self.pending_deletions.clear();
        }

@@ -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
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
--- a/pageserver/src/tenant/blob_io.rs
+++ b/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[..]);
@@ -299,7 +302,7 @@ mod tests {
        // Write part (in block to drop the file)
        let mut offsets = Vec::new();
        {
-            let file = VirtualFile::create(pathbuf.as_path()).await?;
+            let file = VirtualFile::create(pathbuf.as_path(), &ctx).await?;
            let mut wtr = BlobWriter::<BUFFERED>::new(file, 0);
            for blob in blobs.iter() {
                let (_, res) = wtr.write_blob(blob.clone(), &ctx).await;
@@ -314,7 +317,7 @@ mod tests {
            wtr.flush_buffer(&ctx).await?;
        }

-        let file = VirtualFile::open(pathbuf.as_path()).await?;
+        let file = VirtualFile::open(pathbuf.as_path(), &ctx).await?;
        let rdr = BlockReaderRef::VirtualFile(&file);
        let rdr = BlockCursor::new(rdr);
        for (idx, (blob, offset)) in blobs.iter().zip(offsets.iter()).enumerate() {
--- a/pageserver/src/tenant/block_io.rs
+++ b/pageserver/src/tenant/block_io.rs
@@ -102,7 +102,7 @@ impl<'a> BlockReaderRef<'a> {
            #[cfg(test)]
            TestDisk(r) => r.read_blk(blknum),
            #[cfg(test)]
-            VirtualFile(r) => r.read_blk(blknum).await,
+            VirtualFile(r) => r.read_blk(blknum, ctx).await,
        }
    }
 }
@@ -177,10 +177,11 @@ impl<'a> FileBlockReader<'a> {
        &self,
        buf: PageWriteGuard<'static>,
        blkno: u32,
+        ctx: &RequestContext,
    ) -> Result<PageWriteGuard<'static>, std::io::Error> {
        assert!(buf.len() == PAGE_SZ);
        self.file
-            .read_exact_at_page(buf, blkno as u64 * PAGE_SZ as u64)
+            .read_exact_at_page(buf, blkno as u64 * PAGE_SZ as u64, ctx)
            .await
    }
    /// Read a block.
@@ -206,7 +207,7 @@ impl<'a> FileBlockReader<'a> {
            ReadBufResult::Found(guard) => Ok(guard.into()),
            ReadBufResult::NotFound(write_guard) => {
                // Read the page from disk into the buffer
-                let write_guard = self.fill_buffer(write_guard, blknum).await?;
+                let write_guard = self.fill_buffer(write_guard, blknum, ctx).await?;
                Ok(write_guard.mark_valid().into())
            }
        }
--- a/pageserver/src/tenant/config.rs
+++ b/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(),
        }
    }
 }
--- a/pageserver/src/tenant/delete.rs
+++ b/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(),
        )
--- a/pageserver/src/tenant/ephemeral_file.rs
+++ b/pageserver/src/tenant/ephemeral_file.rs
@@ -28,6 +28,7 @@ impl EphemeralFile {
        conf: &PageServerConf,
        tenant_shard_id: TenantShardId,
        timeline_id: TimelineId,
+        ctx: &RequestContext,
    ) -> Result<EphemeralFile, io::Error> {
        static NEXT_FILENAME: AtomicU64 = AtomicU64::new(1);
        let filename_disambiguator =
@@ -45,6 +46,7 @@ impl EphemeralFile {
                .read(true)
                .write(true)
                .create(true),
+            ctx,
        )
        .await?;

@@ -153,7 +155,7 @@ mod tests {
    async fn test_ephemeral_blobs() -> Result<(), io::Error> {
        let (conf, tenant_id, timeline_id, ctx) = harness("ephemeral_blobs")?;

-        let mut file = EphemeralFile::create(conf, tenant_id, timeline_id).await?;
+        let mut file = EphemeralFile::create(conf, tenant_id, timeline_id, &ctx).await?;

        let pos_foo = file.write_blob(b"foo", &ctx).await?;
        assert_eq!(
--- a/pageserver/src/tenant/ephemeral_file/page_caching.rs
+++ b/pageserver/src/tenant/ephemeral_file/page_caching.rs
@@ -78,7 +78,7 @@ impl RW {
                    page_cache::ReadBufResult::NotFound(write_guard) => {
                        let write_guard = writer
                            .file
-                            .read_exact_at_page(write_guard, blknum as u64 * PAGE_SZ as u64)
+                            .read_exact_at_page(write_guard, blknum as u64 * PAGE_SZ as u64, ctx)
                            .await?;
                        let read_guard = write_guard.mark_valid();
                        return Ok(BlockLease::PageReadGuard(read_guard));
--- a/pageserver/src/tenant/metadata.rs
+++ b/pageserver/src/tenant/metadata.rs
@@ -214,12 +214,12 @@ impl TimelineMetadata {
        self.body.ancestor_timeline = Some(*timeline);
    }

-    pub fn detach_from_ancestor(&mut self, timeline: &TimelineId, ancestor_lsn: &Lsn) {
+    pub fn detach_from_ancestor(&mut self, branchpoint: &(TimelineId, Lsn)) {
        if let Some(ancestor) = self.body.ancestor_timeline {
-            assert_eq!(ancestor, *timeline);
+            assert_eq!(ancestor, branchpoint.0);
        }
        if self.body.ancestor_lsn != Lsn(0) {
-            assert_eq!(self.body.ancestor_lsn, *ancestor_lsn);
+            assert_eq!(self.body.ancestor_lsn, branchpoint.1);
        }
        self.body.ancestor_timeline = None;
        self.body.ancestor_lsn = Lsn(0);
--- a/pageserver/src/tenant/mgr.rs
+++ b/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)]
--- a/pageserver/src/tenant/remote_timeline_client.rs
+++ b/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};
@@ -240,7 +242,7 @@ use utils::id::{TenantId, TimelineId};
 use self::index::IndexPart;

 use super::metadata::MetadataUpdate;
-use super::storage_layer::{Layer, LayerFileName, ResidentLayer};
+use super::storage_layer::{Layer, LayerName, ResidentLayer};
 use super::upload_queue::SetDeletedFlagProgress;
 use super::Generation;

@@ -317,7 +319,7 @@ pub struct RemoteTimelineClient {

    upload_queue: Mutex<UploadQueue>,

-    metrics: Arc<RemoteTimelineClientMetrics>,
+    pub(crate) metrics: Arc<RemoteTimelineClientMetrics>,

    storage_impl: GenericRemoteStorage,

@@ -437,6 +439,19 @@ impl RemoteTimelineClient {
        }
    }

+    /// Returns true if this timeline was previously detached at this Lsn and the remote timeline
+    /// client is currently initialized.
+    pub(crate) fn is_previous_ancestor_lsn(&self, lsn: Lsn) -> bool {
+        // technically this is a dirty read, but given how timeline detach ancestor is implemented
+        // via tenant restart, the lineage has always been uploaded.
+        self.upload_queue
+            .lock()
+            .unwrap()
+            .initialized_mut()
+            .map(|uq| uq.latest_lineage.is_previous_ancestor_lsn(lsn))
+            .unwrap_or(false)
+    }
+
    fn update_remote_physical_size_gauge(&self, current_remote_index_part: Option<&IndexPart>) {
        let size: u64 = if let Some(current_remote_index_part) = current_remote_index_part {
            current_remote_index_part
@@ -448,11 +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()
    }

    //
@@ -503,8 +518,9 @@ impl RemoteTimelineClient {
    /// On success, returns the size of the downloaded file.
    pub async fn download_layer_file(
        &self,
-        layer_file_name: &LayerFileName,
+        layer_file_name: &LayerName,
        layer_metadata: &LayerFileMetadata,
+        local_path: &Utf8Path,
        cancel: &CancellationToken,
        ctx: &RequestContext,
    ) -> anyhow::Result<u64> {
@@ -523,6 +539,7 @@ impl RemoteTimelineClient {
                self.timeline_id,
                layer_file_name,
                layer_metadata,
+                local_path,
                cancel,
                ctx,
            )
@@ -596,6 +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.
    ///
@@ -628,7 +656,7 @@ impl RemoteTimelineClient {
        );

        let index_part = IndexPart::from(&*upload_queue);
-        let op = UploadOp::UploadMetadata(index_part, disk_consistent_lsn);
+        let op = UploadOp::UploadMetadata(Box::new(index_part), disk_consistent_lsn);
        self.metric_begin(&op);
        upload_queue.queued_operations.push_back(op);
        upload_queue.latest_files_changes_since_metadata_upload_scheduled = 0;
@@ -647,7 +675,14 @@ impl RemoteTimelineClient {
            let mut guard = self.upload_queue.lock().unwrap();
            let upload_queue = guard.initialized_mut()?;

+            let Some(prev) = upload_queue.latest_metadata.ancestor_timeline() else {
+                return Err(anyhow::anyhow!(
+                    "cannot reparent without a current ancestor"
+                ));
+            };
+
            upload_queue.latest_metadata.reparent(new_parent);
+            upload_queue.latest_lineage.record_previous_ancestor(&prev);

            self.schedule_index_upload(upload_queue);

@@ -670,14 +705,13 @@ impl RemoteTimelineClient {
            let mut guard = self.upload_queue.lock().unwrap();
            let upload_queue = guard.initialized_mut()?;

-            upload_queue
-                .latest_metadata
-                .detach_from_ancestor(&adopted.0, &adopted.1);
+            upload_queue.latest_metadata.detach_from_ancestor(&adopted);
+            upload_queue.latest_lineage.record_detaching(&adopted);

            for layer in layers {
                upload_queue
                    .latest_files
-                    .insert(layer.layer_desc().filename(), layer.metadata());
+                    .insert(layer.layer_desc().layer_name(), layer.metadata());
            }

            self.schedule_index_upload(upload_queue);
@@ -713,7 +747,7 @@ impl RemoteTimelineClient {

        upload_queue
            .latest_files
-            .insert(layer.layer_desc().filename(), metadata.clone());
+            .insert(layer.layer_desc().layer_name(), metadata.clone());
        upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;

        info!(
@@ -737,7 +771,7 @@ impl RemoteTimelineClient {
    /// successfully.
    pub fn schedule_layer_file_deletion(
        self: &Arc<Self>,
-        names: &[LayerFileName],
+        names: &[LayerName],
    ) -> anyhow::Result<()> {
        let mut guard = self.upload_queue.lock().unwrap();
        let upload_queue = guard.initialized_mut()?;
@@ -765,7 +799,7 @@ impl RemoteTimelineClient {
        // the layer files as "dangling". this is fine, at worst case we create work for the
        // scrubber.

-        let names = gc_layers.iter().map(|x| x.layer_desc().filename());
+        let names = gc_layers.iter().map(|x| x.layer_desc().layer_name());

        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);

@@ -780,9 +814,9 @@ impl RemoteTimelineClient {
        self: &Arc<Self>,
        upload_queue: &mut UploadQueueInitialized,
        names: I,
-    ) -> Vec<(LayerFileName, LayerFileMetadata)>
+    ) -> Vec<(LayerName, LayerFileMetadata)>
    where
-        I: IntoIterator<Item = LayerFileName>,
+        I: IntoIterator<Item = LayerName>,
    {
        // Decorate our list of names with each name's metadata, dropping
        // names that are unexpectedly missing from our metadata.  This metadata
@@ -832,7 +866,7 @@ impl RemoteTimelineClient {
    /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
    pub(crate) fn schedule_deletion_of_unlinked(
        self: &Arc<Self>,
-        layers: Vec<(LayerFileName, LayerFileMetadata)>,
+        layers: Vec<(LayerName, LayerFileMetadata)>,
    ) -> anyhow::Result<()> {
        let mut guard = self.upload_queue.lock().unwrap();
        let upload_queue = guard.initialized_mut()?;
@@ -845,7 +879,7 @@ impl RemoteTimelineClient {
    fn schedule_deletion_of_unlinked0(
        self: &Arc<Self>,
        upload_queue: &mut UploadQueueInitialized,
-        mut with_metadata: Vec<(LayerFileName, LayerFileMetadata)>,
+        mut with_metadata: Vec<(LayerName, LayerFileMetadata)>,
    ) {
        // Filter out any layers which were not created by this tenant shard.  These are
        // layers that originate from some ancestor shard after a split, and may still
@@ -914,7 +948,7 @@ impl RemoteTimelineClient {
            self.schedule_layer_file_upload0(upload_queue, layer.clone());
        }

-        let names = compacted_from.iter().map(|x| x.layer_desc().filename());
+        let names = compacted_from.iter().map(|x| x.layer_desc().layer_name());

        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
        self.launch_queued_tasks(upload_queue);
@@ -1108,6 +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,
@@ -1144,7 +1183,7 @@ impl RemoteTimelineClient {
            &self.tenant_shard_id.tenant_id,
            &self.timeline_id,
            self.tenant_shard_id.to_index(),
-            &uploaded.layer_desc().filename(),
+            &uploaded.layer_desc().layer_name(),
            uploaded.metadata().generation,
        );

@@ -1154,7 +1193,7 @@ impl RemoteTimelineClient {
                    &self.storage_impl,
                    uploaded.local_path(),
                    &remote_path,
-                    uploaded.metadata().file_size(),
+                    uploaded.metadata().file_size,
                    cancel,
                )
                .await
@@ -1185,7 +1224,7 @@ impl RemoteTimelineClient {
                .get_timeline_id()
                .expect("Source timeline should be alive"),
            self.tenant_shard_id.to_index(),
-            &adopted.layer_desc().filename(),
+            &adopted.layer_desc().layer_name(),
            adopted.metadata().generation,
        );

@@ -1193,7 +1232,7 @@ impl RemoteTimelineClient {
            &self.tenant_shard_id.tenant_id,
            &self.timeline_id,
            self.tenant_shard_id.to_index(),
-            &adopted_as.layer_desc().filename(),
+            &adopted_as.layer_desc().layer_name(),
            adopted_as.metadata().generation,
        );

@@ -1527,7 +1566,7 @@ impl RemoteTimelineClient {
                        &self.tenant_shard_id.tenant_id,
                        &self.timeline_id,
                        layer_metadata.shard,
-                        &layer.layer_desc().filename(),
+                        &layer.layer_desc().layer_name(),
                        layer_metadata.generation,
                    );

@@ -1535,7 +1574,7 @@ impl RemoteTimelineClient {
                        &self.storage_impl,
                        local_path,
                        &remote_path,
-                        layer_metadata.file_size(),
+                        layer_metadata.file_size,
                        &self.cancel,
                    )
                    .measure_remote_op(
@@ -1730,7 +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,
@@ -1811,6 +1850,7 @@ impl RemoteTimelineClient {
                        latest_files: initialized.latest_files.clone(),
                        latest_files_changes_since_metadata_upload_scheduled: 0,
                        latest_metadata: initialized.latest_metadata.clone(),
+                        latest_lineage: initialized.latest_lineage.clone(),
                        projected_remote_consistent_lsn: None,
                        visible_remote_consistent_lsn: initialized
                            .visible_remote_consistent_lsn
@@ -1824,6 +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(
@@ -1896,14 +1937,14 @@ pub fn remote_layer_path(
    tenant_id: &TenantId,
    timeline_id: &TimelineId,
    shard: ShardIndex,
-    layer_file_name: &LayerFileName,
+    layer_file_name: &LayerName,
    generation: Generation,
 ) -> RemotePath {
    // Generation-aware key format
    let path = format!(
        "tenants/{tenant_id}{0}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{1}{2}",
        shard.get_suffix(),
-        layer_file_name.file_name(),
+        layer_file_name,
        generation.get_suffix()
    );

@@ -2000,8 +2041,8 @@ mod tests {
        TimelineMetadata::from_bytes(&metadata.to_bytes().unwrap()).unwrap()
    }

-    fn assert_file_list(a: &HashSet<LayerFileName>, b: &[&str]) {
-        let mut avec: Vec<String> = a.iter().map(|x| x.file_name()).collect();
+    fn assert_file_list(a: &HashSet<LayerName>, b: &[&str]) {
+        let mut avec: Vec<String> = a.iter().map(|x| x.to_string()).collect();
        avec.sort();

        let mut bvec = b.to_vec();
@@ -2112,7 +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
@@ -2127,7 +2168,7 @@ mod tests {
            .layer_metadata
            .keys()
            .map(|f| f.to_owned())
-            .collect::<HashSet<LayerFileName>>();
+            .collect::<HashSet<LayerName>>();
        let initial_layer = {
            assert!(initial_layers.len() == 1);
            initial_layers.into_iter().next().unwrap()
@@ -2153,7 +2194,7 @@ mod tests {
            ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap(), dummy_contents("baz"))
        ]
        .into_iter()
-        .map(|(name, contents): (LayerFileName, Vec<u8>)| {
+        .map(|(name, contents): (LayerName, Vec<u8>)| {

            let local_path = local_layer_path(
                harness.conf,
@@ -2234,9 +2275,9 @@ mod tests {
                .map(|f| f.to_owned())
                .collect(),
            &[
-                &initial_layer.file_name(),
-                &layers[0].layer_desc().filename().file_name(),
-                &layers[1].layer_desc().filename().file_name(),
+                &initial_layer.to_string(),
+                &layers[0].layer_desc().layer_name().to_string(),
+                &layers[1].layer_desc().layer_name().to_string(),
            ],
        );
        assert_eq!(index_part.metadata, metadata);
@@ -2250,7 +2291,7 @@ mod tests {
        // keep using schedule_layer_file_deletion because we don't have a way to wait for the
        // spawn_blocking started by the drop.
        client
-            .schedule_layer_file_deletion(&[layers[0].layer_desc().filename()])
+            .schedule_layer_file_deletion(&[layers[0].layer_desc().layer_name()])
            .unwrap();
        {
            let mut guard = client.upload_queue.lock().unwrap();
@@ -2268,9 +2309,9 @@ mod tests {
        }
        assert_remote_files(
            &[
-                &initial_layer.file_name(),
-                &layers[0].layer_desc().filename().file_name(),
-                &layers[1].layer_desc().filename().file_name(),
+                &initial_layer.to_string(),
+                &layers[0].layer_desc().layer_name().to_string(),
+                &layers[1].layer_desc().layer_name().to_string(),
                "index_part.json",
            ],
            &remote_timeline_dir,
@@ -2283,9 +2324,9 @@ mod tests {

        assert_remote_files(
            &[
-                &initial_layer.file_name(),
-                &layers[1].layer_desc().filename().file_name(),
-                &layers[2].layer_desc().filename().file_name(),
+                &initial_layer.to_string(),
+                &layers[1].layer_desc().layer_name().to_string(),
+                &layers[2].layer_desc().layer_name().to_string(),
                "index_part.json",
            ],
            &remote_timeline_dir,
@@ -2303,9 +2344,9 @@ 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: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
+        let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
        let local_path = local_layer_path(
            harness.conf,
            &timeline.tenant_shard_id,
--- a/pageserver/src/tenant/remote_timeline_client/download.rs
+++ b/pageserver/src/tenant/remote_timeline_client/download.rs
@@ -21,8 +21,7 @@ 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::LayerFileName;
+use crate::tenant::storage_layer::LayerName;
 use crate::tenant::Generation;
 use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
 use crate::TEMP_FILE_SUFFIX;
@@ -48,21 +47,15 @@ pub async fn download_layer_file<'a>(
    storage: &'a GenericRemoteStorage,
    tenant_shard_id: TenantShardId,
    timeline_id: TimelineId,
-    layer_file_name: &'a LayerFileName,
+    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:?}",
@@ -112,14 +105,17 @@ pub async fn download_layer_file<'a>(
    // We use fatal_err() below because the after the rename above,
    // the in-memory state of the filesystem already has the layer file in its final place,
    // and subsequent pageserver code could think it's durable while it really isn't.
-    let work = async move {
-        let timeline_dir = VirtualFile::open(&timeline_path)
-            .await
-            .fatal_err("VirtualFile::open for timeline dir fsync");
-        timeline_dir
-            .sync_all()
-            .await
-            .fatal_err("VirtualFile::sync_all timeline dir");
+    let work = {
+        let ctx = ctx.detached_child(ctx.task_kind(), ctx.download_behavior());
+        async move {
+            let timeline_dir = VirtualFile::open(&timeline_path, &ctx)
+                .await
+                .fatal_err("VirtualFile::open for timeline dir fsync");
+            timeline_dir
+                .sync_all()
+                .await
+                .fatal_err("VirtualFile::sync_all timeline dir");
+        }
    };
    crate::virtual_file::io_engine::get()
        .spawn_blocking_and_block_on_if_std(work)
@@ -196,7 +192,7 @@ async fn download_object<'a>(
            use crate::virtual_file::owned_buffers_io::{self, util::size_tracking_writer};
            use bytes::BytesMut;
            async {
-                let destination_file = VirtualFile::create(dst_path)
+                let destination_file = VirtualFile::create(dst_path, ctx)
                    .await
                    .with_context(|| format!("create a destination file for layer '{dst_path}'"))
                    .map_err(DownloadError::Other)?;
--- a/pageserver/src/tenant/remote_timeline_client/index.rs
+++ b/pageserver/src/tenant/remote_timeline_client/index.rs
@@ -5,56 +5,18 @@
 use std::collections::HashMap;

 use chrono::NaiveDateTime;
+use pageserver_api::models::AuxFilePolicy;
 use serde::{Deserialize, Serialize};
+use utils::id::TimelineId;

 use crate::tenant::metadata::TimelineMetadata;
-use crate::tenant::storage_layer::LayerFileName;
+use crate::tenant::storage_layer::LayerName;
 use crate::tenant::upload_queue::UploadQueueInitialized;
 use crate::tenant::Generation;
 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.
@@ -75,7 +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<LayerFileName, 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
@@ -84,6 +46,19 @@ pub struct IndexPart {

    #[serde(rename = "metadata_bytes")]
    pub metadata: TimelineMetadata,
+
+    #[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 {
@@ -96,22 +71,23 @@ impl IndexPart {
    /// - 3: no longer deserialize `timeline_layers` (serialized format is the same, but timeline_layers
    ///      is always generated from the keys of `layer_metadata`)
    /// - 4: timeline_layers is fully removed.
-    const LATEST_VERSION: usize = 4;
+    /// - 5: lineage was added
+    /// - 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];
+    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5, 6];

    pub const FILE_NAME: &'static str = "index_part.json";

    fn new(
-        layers_and_metadata: &HashMap<LayerFileName, LayerFileMetadata>,
+        layers_and_metadata: &HashMap<LayerName, LayerFileMetadata>,
        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,
@@ -119,6 +95,8 @@ impl IndexPart {
            disk_consistent_lsn,
            metadata,
            deleted_at: None,
+            lineage,
+            last_aux_file_policy,
        }
    }

@@ -147,22 +125,38 @@ impl IndexPart {
            &HashMap::new(),
            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 {
    fn from(uq: &UploadQueueInitialized) -> Self {
        let disk_consistent_lsn = uq.latest_metadata.disk_consistent_lsn();
        let metadata = uq.latest_metadata.clone();
+        let lineage = uq.latest_lineage.clone();

-        Self::new(&uq.latest_files, disk_consistent_lsn, metadata)
+        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")]
@@ -174,18 +168,86 @@ 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,
        }
    }
 }

+/// Limited history of earlier ancestors.
+///
+/// A timeline can have more than 1 earlier ancestor, in the rare case that it was repeatedly
+/// reparented by having an later timeline be detached from it's ancestor.
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Default)]
+pub(crate) struct Lineage {
+    /// Has the `reparenting_history` been truncated to [`Lineage::REMEMBER_AT_MOST`].
+    #[serde(skip_serializing_if = "is_false", default)]
+    reparenting_history_truncated: bool,
+
+    /// Earlier ancestors, truncated when [`Self::reparenting_history_truncated`]
+    ///
+    /// These are stored in case we want to support WAL based DR on the timeline. There can be many
+    /// of these and at most one [`Self::original_ancestor`]. There cannot be more reparentings
+    /// after [`Self::original_ancestor`] has been set.
+    #[serde(skip_serializing_if = "Vec::is_empty", default)]
+    reparenting_history: Vec<TimelineId>,
+
+    /// The ancestor from which this timeline has been detached from and when.
+    ///
+    /// If you are adding support for detaching from a hierarchy, consider changing the ancestry
+    /// into a `Vec<(TimelineId, Lsn)>` to be a path instead.
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    original_ancestor: Option<(TimelineId, Lsn, NaiveDateTime)>,
+}
+
+fn is_false(b: &bool) -> bool {
+    !b
+}
+
+impl Lineage {
+    const REMEMBER_AT_MOST: usize = 100;
+
+    pub(crate) fn record_previous_ancestor(&mut self, old_ancestor: &TimelineId) {
+        if self.reparenting_history.last() == Some(old_ancestor) {
+            // do not re-record it
+            return;
+        }
+
+        let drop_oldest = self.reparenting_history.len() + 1 >= Self::REMEMBER_AT_MOST;
+
+        self.reparenting_history_truncated |= drop_oldest;
+        if drop_oldest {
+            self.reparenting_history.remove(0);
+        }
+        self.reparenting_history.push(*old_ancestor);
+    }
+
+    pub(crate) fn record_detaching(&mut self, branchpoint: &(TimelineId, Lsn)) {
+        assert!(self.original_ancestor.is_none());
+
+        self.original_ancestor =
+            Some((branchpoint.0, branchpoint.1, chrono::Utc::now().naive_utc()));
+    }
+
+    /// The queried lsn is most likely the basebackup lsn, and this answers question "is it allowed
+    /// to start a read/write primary at this lsn".
+    ///
+    /// Returns true if the Lsn was previously a branch point.
+    pub(crate) fn is_previous_ancestor_lsn(&self, lsn: Lsn) -> bool {
+        self.original_ancestor
+            .as_ref()
+            .is_some_and(|(_, ancestor_lsn, _)| lsn == *ancestor_lsn)
+    }
+}
+
 #[cfg(test)]
 mod tests {
+    use std::str::FromStr;
+
    use super::*;

    #[test]
@@ -205,12 +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,
@@ -221,6 +283,8 @@ mod tests {
            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: None,
+            lineage: Lineage::default(),
+            last_aux_file_policy: None,
        };

        let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -245,12 +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,
@@ -261,6 +325,8 @@ mod tests {
            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: None,
+            lineage: Lineage::default(),
+            last_aux_file_policy: None,
        };

        let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();
@@ -286,12 +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,
@@ -302,7 +368,9 @@ mod tests {
            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())
+                "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();
@@ -347,6 +415,8 @@ 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();
@@ -370,12 +440,101 @@ 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,
+                    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(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();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v5_indexpart_is_parsed() {
+        let example = r#"{
+            "version":5,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF420-00000000014EF499":{"file_size":23289856,"generation":1},
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF499-00000000015A7619":{"file_size":1015808,"generation":1}},
+                "disk_consistent_lsn":"0/15A7618",
+                "metadata_bytes":[226,88,25,241,0,46,0,4,0,0,0,0,1,90,118,24,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,78,244,32,0,0,0,0,1,78,244,32,0,0,0,16,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+                "lineage":{
+                    "original_ancestor":["e2bfd8c633d713d279e6fcd2bcc15b6d","0/15A7618","2024-05-07T18:52:36.322426563"],
+                    "reparenting_history":["e1bfd8c633d713d279e6fcd2bcc15b6d"]
+                }
+        }"#;
+
+        let expected = IndexPart {
+            version: 5,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF420-00000000014EF499".parse().unwrap(), LayerFileMetadata {
+                    file_size: 23289856,
+                    generation: Generation::new(1),
+                    shard: ShardIndex::unsharded(),
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000014EF499-00000000015A7619".parse().unwrap(), LayerFileMetadata {
+                    file_size: 1015808,
+                    generation: Generation::new(1),
+                    shard: ShardIndex::unsharded(),
+                })
+            ]),
+            disk_consistent_lsn: Lsn::from_str("0/15A7618").unwrap(),
+            metadata: TimelineMetadata::from_bytes(&[226,88,25,241,0,46,0,4,0,0,0,0,1,90,118,24,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,78,244,32,0,0,0,0,1,78,244,32,0,0,0,16,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
+            deleted_at: None,
+            lineage: Lineage {
+                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: 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,
@@ -387,9 +546,19 @@ 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(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();
        assert_eq!(part, expected);
    }
+
+    fn parse_naive_datetime(s: &str) -> NaiveDateTime {
+        chrono::NaiveDateTime::parse_from_str(s, "%Y-%m-%dT%H:%M:%S.%f").unwrap()
+    }
 }
--- a/pageserver/src/tenant/remote_timeline_client/upload.rs
+++ b/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::{
--- a/pageserver/src/tenant/secondary.rs
+++ b/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, LayerFileName},
+    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: LayerFileName,
-        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
--- a/pageserver/src/tenant/secondary/downloader.rs
+++ b/pageserver/src/tenant/secondary/downloader.rs
@@ -22,11 +22,11 @@ use crate::{
            FAILED_REMOTE_OP_RETRIES,
        },
        span::debug_assert_current_span_has_tenant_id,
-        storage_layer::{layer::local_layer_path, LayerFileName},
+        storage_layer::{layer::local_layer_path, LayerName},
        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 {
@@ -111,23 +108,46 @@ impl OnDiskState {
        _conf: &'static PageServerConf,
        _tenant_shard_id: &TenantShardId,
        _imeline_id: &TimelineId,
-        _ame: LayerFileName,
+        _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)]
 pub(super) struct SecondaryDetailTimeline {
-    pub(super) on_disk_layers: HashMap<LayerFileName, OnDiskState>,
+    pub(super) on_disk_layers: HashMap<LayerName, OnDiskState>,

    /// We remember when layers were evicted, to prevent re-downloading them.
-    pub(super) evicted_at: HashMap<LayerFileName, SystemTime>,
+    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
@@ -136,8 +156,7 @@ pub(super) struct SecondaryDetailTimeline {
 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
@@ -997,7 +1086,7 @@ async fn init_timeline_state(

    // As we iterate through layers found on disk, we will look up their metadata from this map.
    // Layers not present in metadata will be discarded.
-    let heatmap_metadata: HashMap<&LayerFileName, &HeatMapLayer> =
+    let heatmap_metadata: HashMap<&LayerName, &HeatMapLayer> =
        heatmap.layers.iter().map(|l| (&l.name, l)).collect();

    while let Some(dentry) = dir
@@ -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)
        {
@@ -1034,7 +1119,7 @@ async fn init_timeline_state(
            continue;
        }

-        match LayerFileName::from_str(file_name) {
+        match LayerName::from_str(file_name) {
            Ok(name) => {
                let remote_meta = heatmap_metadata.get(&name);
                match remote_meta {
@@ -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,
                                ),
                            );
                        }
--- a/pageserver/src/tenant/secondary/heatmap.rs
+++ b/pageserver/src/tenant/secondary/heatmap.rs
@@ -1,8 +1,6 @@
 use std::time::SystemTime;

-use crate::tenant::{
-    remote_timeline_client::index::IndexLayerMetadata, storage_layer::LayerFileName,
-};
+use crate::tenant::{remote_timeline_client::index::LayerFileMetadata, storage_layer::LayerName};

 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr, TimestampSeconds};
@@ -17,6 +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]
@@ -31,8 +37,8 @@ pub(crate) struct HeatMapTimeline {
 #[serde_as]
 #[derive(Serialize, Deserialize)]
 pub(crate) struct HeatMapLayer {
-    pub(super) name: LayerFileName,
-    pub(super) metadata: IndexLayerMetadata,
+    pub(super) name: LayerName,
+    pub(super) metadata: LayerFileMetadata,

    #[serde_as(as = "TimestampSeconds<i64>")]
    pub(super) access_time: SystemTime,
@@ -42,8 +48,8 @@ pub(crate) struct HeatMapLayer {

 impl HeatMapLayer {
    pub(crate) fn new(
-        name: LayerFileName,
-        metadata: IndexLayerMetadata,
+        name: LayerName,
+        metadata: LayerFileMetadata,
        access_time: SystemTime,
    ) -> Self {
        Self {
@@ -83,4 +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,
+        }
+    }
 }
--- a/pageserver/src/tenant/secondary/heatmap_uploader.rs
+++ b/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,
+    }))
 }
--- a/pageserver/src/tenant/secondary/scheduler.rs
+++ b/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);
--- a/pageserver/src/tenant/storage_layer.rs
+++ b/pageserver/src/tenant/storage_layer.rs
@@ -1,11 +1,11 @@
 //! Common traits and structs for layers

 pub mod delta_layer;
-mod filename;
 pub mod image_layer;
 pub(crate) mod inmemory_layer;
 pub(crate) mod layer;
 mod layer_desc;
+mod layer_name;

 use crate::context::{AccessStatsBehavior, RequestContext};
 use crate::repository::Value;
@@ -34,10 +34,10 @@ use utils::rate_limit::RateLimit;
 use utils::{id::TimelineId, lsn::Lsn};

 pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
-pub use filename::{DeltaFileName, ImageFileName, LayerFileName};
 pub use image_layer::{ImageLayer, ImageLayerWriter};
 pub use inmemory_layer::InMemoryLayer;
 pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
+pub use layer_name::{DeltaLayerName, ImageLayerName, LayerName};

 pub(crate) use layer::{EvictionError, Layer, ResidentLayer};

@@ -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(),
+        )
    }
 }

@@ -646,8 +679,8 @@ pub mod tests {

    use super::*;

-    impl From<DeltaFileName> for PersistentLayerDesc {
-        fn from(value: DeltaFileName) -> Self {
+    impl From<DeltaLayerName> for PersistentLayerDesc {
+        fn from(value: DeltaLayerName) -> Self {
            PersistentLayerDesc::new_delta(
                TenantShardId::from([0; 18]),
                TimelineId::from_array([0; 16]),
@@ -658,8 +691,8 @@ pub mod tests {
        }
    }

-    impl From<ImageFileName> for PersistentLayerDesc {
-        fn from(value: ImageFileName) -> Self {
+    impl From<ImageLayerName> for PersistentLayerDesc {
+        fn from(value: ImageLayerName) -> Self {
            PersistentLayerDesc::new_img(
                TenantShardId::from([0; 18]),
                TimelineId::from_array([0; 16]),
@@ -670,11 +703,11 @@ pub mod tests {
        }
    }

-    impl From<LayerFileName> for PersistentLayerDesc {
-        fn from(value: LayerFileName) -> Self {
+    impl From<LayerName> for PersistentLayerDesc {
+        fn from(value: LayerName) -> Self {
            match value {
-                LayerFileName::Delta(d) => Self::from(d),
-                LayerFileName::Image(i) => Self::from(i),
+                LayerName::Delta(d) => Self::from(d),
+                LayerName::Image(i) => Self::from(i),
            }
        }
    }
--- a/pageserver/src/tenant/storage_layer/delta_layer.rs
+++ b/pageserver/src/tenant/storage_layer/delta_layer.rs
@@ -57,6 +57,7 @@ use std::fs::File;
 use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
+use std::str::FromStr;
 use std::sync::Arc;
 use tokio::sync::OnceCell;
 use tracing::*;
@@ -68,7 +69,8 @@ use utils::{
 };

 use super::{
-    AsLayerDesc, LayerAccessStats, PersistentLayerDesc, ResidentLayer, ValuesReconstructState,
+    AsLayerDesc, LayerAccessStats, LayerName, PersistentLayerDesc, ResidentLayer,
+    ValuesReconstructState,
 };

 ///
@@ -309,13 +311,13 @@ impl DeltaLayer {
            .and_then(|res| res)?;

        // not production code
-        let actual_filename = path.file_name().unwrap().to_owned();
-        let expected_filename = self.layer_desc().filename().file_name();
+        let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
+        let expected_layer_name = self.layer_desc().layer_name();

-        if actual_filename != expected_filename {
+        if actual_layer_name != expected_layer_name {
            println!("warning: filename does not match what is expected from in-file summary");
-            println!("actual: {:?}", actual_filename);
-            println!("expected: {:?}", expected_filename);
+            println!("actual: {:?}", actual_layer_name.to_string());
+            println!("expected: {:?}", expected_layer_name.to_string());
        }

        Ok(Arc::new(loaded))
@@ -392,6 +394,7 @@ impl DeltaLayerWriterInner {
        tenant_shard_id: TenantShardId,
        key_start: Key,
        lsn_range: Range<Lsn>,
+        ctx: &RequestContext,
    ) -> anyhow::Result<Self> {
        // Create the file initially with a temporary filename. We don't know
        // the end key yet, so we cannot form the final filename yet. We will
@@ -402,7 +405,7 @@ impl DeltaLayerWriterInner {
        let path =
            DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);

-        let mut file = VirtualFile::create(&path).await?;
+        let mut file = VirtualFile::create(&path, ctx).await?;
        // make room for the header block
        file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
        let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
@@ -584,6 +587,7 @@ impl DeltaLayerWriter {
        tenant_shard_id: TenantShardId,
        key_start: Key,
        lsn_range: Range<Lsn>,
+        ctx: &RequestContext,
    ) -> anyhow::Result<Self> {
        Ok(Self {
            inner: Some(
@@ -593,6 +597,7 @@ impl DeltaLayerWriter {
                    tenant_shard_id,
                    key_start,
                    lsn_range,
+                    ctx,
                )
                .await?,
            ),
@@ -699,6 +704,7 @@ impl DeltaLayer {
        let mut file = VirtualFile::open_with_options(
            path,
            virtual_file::OpenOptions::new().read(true).write(true),
+            ctx,
        )
        .await
        .with_context(|| format!("Failed to open file '{}'", path))?;
@@ -732,7 +738,7 @@ impl DeltaLayerInner {
        max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
        ctx: &RequestContext,
    ) -> Result<Result<Self, anyhow::Error>, anyhow::Error> {
-        let file = match VirtualFile::open(path).await {
+        let file = match VirtualFile::open(path, ctx).await {
            Ok(file) => file,
            Err(e) => return Ok(Err(anyhow::Error::new(e).context("open layer file"))),
        };
@@ -906,7 +912,7 @@ impl DeltaLayerInner {
        .await
        .map_err(GetVectoredError::Other)?;

-        self.do_reads_and_update_state(reads, reconstruct_state)
+        self.do_reads_and_update_state(reads, reconstruct_state, ctx)
            .await;

        reconstruct_state.on_lsn_advanced(&keyspace, self.lsn_range.start);
@@ -1010,6 +1016,7 @@ impl DeltaLayerInner {
        &self,
        reads: Vec<VectoredRead>,
        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
    ) {
        let vectored_blob_reader = VectoredBlobReader::new(&self.file);
        let mut ignore_key_with_err = None;
@@ -1027,7 +1034,7 @@ impl DeltaLayerInner {
        // track when a key is done.
        for read in reads.into_iter().rev() {
            let res = vectored_blob_reader
-                .read_blobs(&read, buf.take().expect("Should have a buffer"))
+                .read_blobs(&read, buf.take().expect("Should have a buffer"), ctx)
                .await;

            let blobs_buf = match res {
@@ -1272,7 +1279,7 @@ impl DeltaLayerInner {

                buf.clear();
                buf.reserve(read.size());
-                let res = reader.read_blobs(&read, buf).await?;
+                let res = reader.read_blobs(&read, buf, ctx).await?;

                for blob in res.blobs {
                    let key = blob.meta.key;
@@ -1789,6 +1796,7 @@ mod test {
            harness.tenant_shard_id,
            entries_meta.key_range.start,
            entries_meta.lsn_range.clone(),
+            &ctx,
        )
        .await?;

@@ -1846,7 +1854,7 @@ mod test {

            for read in vectored_reads {
                let blobs_buf = vectored_blob_reader
-                    .read_blobs(&read, buf.take().expect("Should have a buffer"))
+                    .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
                    .await?;
                for meta in blobs_buf.blobs.iter() {
                    let value = &blobs_buf.buf[meta.start..meta.end];
@@ -1976,6 +1984,7 @@ mod test {
                tenant.tenant_shard_id,
                Key::MIN,
                Lsn(0x11)..truncate_at,
+                ctx,
            )
            .await
            .unwrap();
--- a/pageserver/src/tenant/storage_layer/image_layer.rs
+++ b/pageserver/src/tenant/storage_layer/image_layer.rs
@@ -47,13 +47,14 @@ 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;
 use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
+use std::str::FromStr;
 use std::sync::Arc;
 use tokio::sync::OnceCell;
 use tokio_stream::StreamExt;
@@ -65,8 +66,10 @@ use utils::{
    lsn::Lsn,
 };

-use super::filename::ImageFileName;
-use super::{AsLayerDesc, Layer, PersistentLayerDesc, ResidentLayer, ValuesReconstructState};
+use super::layer_name::ImageLayerName;
+use super::{
+    AsLayerDesc, Layer, LayerName, PersistentLayerDesc, ResidentLayer, ValuesReconstructState,
+};

 ///
 /// Header stored in the beginning of the file
@@ -155,6 +158,7 @@ pub struct ImageLayerInner {
    index_start_blk: u32,
    index_root_blk: u32,

+    key_range: Range<Key>,
    lsn: Lsn,

    file: VirtualFile,
@@ -231,7 +235,7 @@ impl ImageLayer {
        conf: &PageServerConf,
        timeline_id: TimelineId,
        tenant_shard_id: TenantShardId,
-        fname: &ImageFileName,
+        fname: &ImageLayerName,
    ) -> Utf8PathBuf {
        let rand_string: String = rand::thread_rng()
            .sample_iter(&Alphanumeric)
@@ -267,13 +271,13 @@ impl ImageLayer {
            .and_then(|res| res)?;

        // not production code
-        let actual_filename = path.file_name().unwrap().to_owned();
-        let expected_filename = self.layer_desc().filename().file_name();
+        let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
+        let expected_layer_name = self.layer_desc().layer_name();

-        if actual_filename != expected_filename {
+        if actual_layer_name != expected_layer_name {
            println!("warning: filename does not match what is expected from in-file summary");
-            println!("actual: {:?}", actual_filename);
-            println!("expected: {:?}", expected_filename);
+            println!("actual: {:?}", actual_layer_name.to_string());
+            println!("expected: {:?}", expected_layer_name.to_string());
        }

        Ok(loaded)
@@ -340,6 +344,7 @@ impl ImageLayer {
        let mut file = VirtualFile::open_with_options(
            path,
            virtual_file::OpenOptions::new().read(true).write(true),
+            ctx,
        )
        .await
        .with_context(|| format!("Failed to open file '{}'", path))?;
@@ -374,7 +379,7 @@ impl ImageLayerInner {
        max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
        ctx: &RequestContext,
    ) -> Result<Result<Self, anyhow::Error>, anyhow::Error> {
-        let file = match VirtualFile::open(path).await {
+        let file = match VirtualFile::open(path, ctx).await {
            Ok(file) => file,
            Err(e) => return Ok(Err(anyhow::Error::new(e).context("open layer file"))),
        };
@@ -415,6 +420,7 @@ impl ImageLayerInner {
            file,
            file_id,
            max_vectored_read_bytes,
+            key_range: actual_summary.key_range,
        }))
    }

@@ -467,19 +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)
+        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(
@@ -499,7 +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);

@@ -512,12 +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);
                }
            }

@@ -530,10 +553,55 @@ 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>,
        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
    ) {
        let max_vectored_read_bytes = self
            .max_vectored_read_bytes
@@ -562,7 +630,7 @@ impl ImageLayerInner {
            }

            let buf = BytesMut::with_capacity(buf_size);
-            let res = vectored_blob_reader.read_blobs(&read, buf).await;
+            let res = vectored_blob_reader.read_blobs(&read, buf, ctx).await;

            match res {
                Ok(blobs_buf) => {
@@ -628,6 +696,7 @@ impl ImageLayerWriterInner {
        tenant_shard_id: TenantShardId,
        key_range: &Range<Key>,
        lsn: Lsn,
+        ctx: &RequestContext,
    ) -> anyhow::Result<Self> {
        // Create the file initially with a temporary filename.
        // We'll atomically rename it to the final name when we're done.
@@ -635,18 +704,19 @@ impl ImageLayerWriterInner {
            conf,
            timeline_id,
            tenant_shard_id,
-            &ImageFileName {
+            &ImageLayerName {
                key_range: key_range.clone(),
                lsn,
            },
        );
-        info!("new image layer {path}");
+        trace!("creating image layer {}", path);
        let mut file = {
            VirtualFile::open_with_options(
                &path,
                virtual_file::OpenOptions::new()
                    .write(true)
                    .create_new(true),
+                ctx,
            )
            .await?
        };
@@ -759,7 +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)
    }
@@ -801,10 +871,11 @@ impl ImageLayerWriter {
        tenant_shard_id: TenantShardId,
        key_range: &Range<Key>,
        lsn: Lsn,
+        ctx: &RequestContext,
    ) -> anyhow::Result<ImageLayerWriter> {
        Ok(Self {
            inner: Some(
-                ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn)
+                ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn, ctx)
                    .await?,
            ),
        })
@@ -843,3 +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!(),
+            }
+        }
+    }
+}
--- a/pageserver/src/tenant/storage_layer/inmemory_layer.rs
+++ b/pageserver/src/tenant/storage_layer/inmemory_layer.rs
@@ -473,10 +473,11 @@ impl InMemoryLayer {
        timeline_id: TimelineId,
        tenant_shard_id: TenantShardId,
        start_lsn: Lsn,
+        ctx: &RequestContext,
    ) -> Result<InMemoryLayer> {
        trace!("initializing new empty InMemoryLayer for writing on timeline {timeline_id} at {start_lsn}");

-        let file = EphemeralFile::create(conf, tenant_shard_id, timeline_id).await?;
+        let file = EphemeralFile::create(conf, tenant_shard_id, timeline_id, ctx).await?;
        let key = InMemoryLayerFileId(file.page_cache_file_id());

        Ok(InMemoryLayer {
@@ -642,6 +643,7 @@ impl InMemoryLayer {
            self.tenant_shard_id,
            Key::MIN,
            self.start_lsn..end_lsn,
+            ctx,
        )
        .await?;

--- a/pageserver/src/tenant/storage_layer/layer.rs
+++ b/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, LayerFileName, PersistentLayerDesc,
-    ValueReconstructResult, ValueReconstructState, ValuesReconstructState,
+    AsLayerDesc, ImageLayerWriter, LayerAccessStats, LayerAccessStatsReset, LayerName,
+    PersistentLayerDesc, ValueReconstructResult, ValueReconstructState, ValuesReconstructState,
 };

 use utils::generation::Generation;
@@ -128,19 +128,17 @@ pub(crate) fn local_layer_path(
    conf: &PageServerConf,
    tenant_shard_id: &TenantShardId,
    timeline_id: &TimelineId,
-    layer_file_name: &LayerFileName,
-    _generation: &Generation,
+    layer_file_name: &LayerName,
+    generation: &Generation,
 ) -> Utf8PathBuf {
    let timeline_path = conf.timeline_path(tenant_shard_id, timeline_id);

-    timeline_path.join(layer_file_name.file_name())
-
-    // TOOD: include generation in the name in now+1 releases.
-    // timeline_path.join(format!(
-    //     "{}{}",
-    //     layer_file_name.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 {
@@ -148,7 +146,7 @@ impl Layer {
    pub(crate) fn for_evicted(
        conf: &'static PageServerConf,
        timeline: &Arc<Timeline>,
-        file_name: LayerFileName,
+        file_name: LayerName,
        metadata: LayerFileMetadata,
    ) -> Self {
        let local_path = local_layer_path(
@@ -163,7 +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);
@@ -189,14 +187,14 @@ impl Layer {
        conf: &'static PageServerConf,
        timeline: &Arc<Timeline>,
        local_path: Utf8PathBuf,
-        file_name: LayerFileName,
+        file_name: LayerName,
        metadata: LayerFileMetadata,
    ) -> ResidentLayer {
        let desc = PersistentLayerDesc::from_filename(
            timeline.tenant_shard_id,
            timeline.timeline_id,
            file_name,
-            metadata.file_size(),
+            metadata.file_size,
        );

        let access_stats = LayerAccessStats::for_loading_layer(LayerResidenceStatus::Resident);
@@ -229,7 +227,7 @@ impl Layer {

        timeline
            .metrics
-            .resident_physical_size_add(metadata.file_size());
+            .resident_physical_size_add(metadata.file_size);

        ResidentLayer { downloaded, owner }
    }
@@ -261,7 +259,7 @@ impl Layer {
                conf,
                &timeline.tenant_shard_id,
                &timeline.timeline_id,
-                &desc.filename(),
+                &desc.layer_name(),
                &timeline.generation,
            );

@@ -587,9 +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.
@@ -689,7 +684,7 @@ impl Drop for LayerInner {
        let span = tracing::info_span!(parent: None, "layer_delete", tenant_id = %self.layer_desc().tenant_shard_id.tenant_id, shard_id=%self.layer_desc().tenant_shard_id.shard_slug(), timeline_id = %self.layer_desc().timeline_id);

        let path = std::mem::take(&mut self.path);
-        let file_name = self.layer_desc().filename();
+        let file_name = self.layer_desc().layer_name();
        let file_size = self.layer_desc().file_size;
        let timeline = self.timeline.clone();
        let meta = self.metadata();
@@ -734,23 +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();
            }
        });
    }
@@ -782,11 +777,12 @@ impl LayerInner {

        LayerInner {
            conf,
-            debug_str: { format!("timelines/{}/{}", timeline.timeline_id, desc.filename()).into() },
+            debug_str: {
+                format!("timelines/{}/{}", timeline.timeline_id, desc.layer_name()).into()
+            },
            path: local_path,
            desc,
            timeline: Arc::downgrade(timeline),
-            have_remote_client: timeline.remote_client.is_some(),
            access_stats,
            wanted_deleted: AtomicBool::new(false),
            inner,
@@ -815,8 +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();

        {
@@ -973,10 +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)?;
        }
@@ -1113,15 +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.filename(),
+                &self.desc.layer_name(),
                &self.metadata(),
+                &self.path,
                &timeline.cancel,
                ctx,
            )
@@ -1257,7 +1244,7 @@ impl LayerInner {
    }

    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
-        let layer_file_name = self.desc.filename().file_name();
+        let layer_name = self.desc.layer_name().to_string();

        let resident = self
            .inner
@@ -1271,18 +1258,19 @@ impl LayerInner {
            let lsn_range = &self.desc.lsn_range;

            HistoricLayerInfo::Delta {
-                layer_file_name,
+                layer_file_name: layer_name,
                layer_file_size: self.desc.file_size,
                lsn_start: lsn_range.start,
                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();

            HistoricLayerInfo::Image {
-                layer_file_name,
+                layer_file_name: layer_name,
                layer_file_size: self.desc.file_size,
                lsn_start: lsn,
                remote: !resident,
@@ -1293,20 +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.

@@ -1355,7 +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);
        };

@@ -1443,7 +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();

@@ -1469,6 +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`
@@ -1578,8 +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")]
@@ -1825,16 +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}"))
@@ -1843,6 +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,
--- a/pageserver/src/tenant/storage_layer/layer/tests.rs
+++ b/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 = {
--- a/pageserver/src/tenant/storage_layer/layer_desc.rs
+++ b/pageserver/src/tenant/storage_layer/layer_desc.rs
@@ -5,7 +5,7 @@ use utils::{id::TimelineId, lsn::Lsn};

 use crate::repository::Key;

-use super::{DeltaFileName, ImageFileName, LayerFileName};
+use super::{DeltaLayerName, ImageLayerName, LayerName};

 use serde::{Deserialize, Serialize};

@@ -51,7 +51,7 @@ impl PersistentLayerDesc {
    }

    pub fn short_id(&self) -> impl Display {
-        self.filename()
+        self.layer_name()
    }

    #[cfg(test)]
@@ -103,14 +103,14 @@ impl PersistentLayerDesc {
    pub fn from_filename(
        tenant_shard_id: TenantShardId,
        timeline_id: TimelineId,
-        filename: LayerFileName,
+        filename: LayerName,
        file_size: u64,
    ) -> Self {
        match filename {
-            LayerFileName::Image(i) => {
+            LayerName::Image(i) => {
                Self::new_img(tenant_shard_id, timeline_id, i.key_range, i.lsn, file_size)
            }
-            LayerFileName::Delta(d) => Self::new_delta(
+            LayerName::Delta(d) => Self::new_delta(
                tenant_shard_id,
                timeline_id,
                d.key_range,
@@ -132,34 +132,34 @@ impl PersistentLayerDesc {
        lsn..(lsn + 1)
    }

-    /// Get a delta file name for this layer.
+    /// Get a delta layer name for this layer.
    ///
    /// Panic: if this is not a delta layer.
-    pub fn delta_file_name(&self) -> DeltaFileName {
+    pub fn delta_layer_name(&self) -> DeltaLayerName {
        assert!(self.is_delta);
-        DeltaFileName {
+        DeltaLayerName {
            key_range: self.key_range.clone(),
            lsn_range: self.lsn_range.clone(),
        }
    }

-    /// Get a delta file name for this layer.
+    /// Get a image layer name for this layer.
    ///
    /// Panic: if this is not an image layer, or the lsn range is invalid
-    pub fn image_file_name(&self) -> ImageFileName {
+    pub fn image_layer_name(&self) -> ImageLayerName {
        assert!(!self.is_delta);
        assert!(self.lsn_range.start + 1 == self.lsn_range.end);
-        ImageFileName {
+        ImageLayerName {
            key_range: self.key_range.clone(),
            lsn: self.lsn_range.start,
        }
    }

-    pub fn filename(&self) -> LayerFileName {
+    pub fn layer_name(&self) -> LayerName {
        if self.is_delta {
-            self.delta_file_name().into()
+            self.delta_layer_name().into()
        } else {
-            self.image_file_name().into()
+            self.image_layer_name().into()
        }
    }

--- a/pageserver/src/tenant/storage_layer/layer_name.rs
+++ b/pageserver/src/tenant/storage_layer/layer_name.rs
@@ -15,29 +15,29 @@ use super::PersistentLayerDesc;

 // Note: Timeline::load_layer_map() relies on this sort order
 #[derive(PartialEq, Eq, Clone, Hash)]
-pub struct DeltaFileName {
+pub struct DeltaLayerName {
    pub key_range: Range<Key>,
    pub lsn_range: Range<Lsn>,
 }

-impl std::fmt::Debug for DeltaFileName {
+impl std::fmt::Debug for DeltaLayerName {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        use super::RangeDisplayDebug;

-        f.debug_struct("DeltaFileName")
+        f.debug_struct("DeltaLayerName")
            .field("key_range", &RangeDisplayDebug(&self.key_range))
            .field("lsn_range", &self.lsn_range)
            .finish()
    }
 }

-impl PartialOrd for DeltaFileName {
+impl PartialOrd for DeltaLayerName {
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
 }

-impl Ord for DeltaFileName {
+impl Ord for DeltaLayerName {
    fn cmp(&self, other: &Self) -> Ordering {
        let mut cmp = self.key_range.start.cmp(&other.key_range.start);
        if cmp != Ordering::Equal {
@@ -57,16 +57,14 @@ impl Ord for DeltaFileName {
    }
 }

-/// Represents the filename of a DeltaLayer
+/// Represents the region of the LSN-Key space covered by a DeltaLayer
 ///
 /// ```text
 ///    <key start>-<key end>__<LSN start>-<LSN end>
 /// ```
-impl DeltaFileName {
-    ///
-    /// Parse a string as a delta file name. Returns None if the filename does not
-    /// match the expected pattern.
-    ///
+impl DeltaLayerName {
+    /// Parse the part of a delta layer's file name that represents the LayerName. Returns None
+    /// if the filename does not match the expected pattern.
    pub fn parse_str(fname: &str) -> Option<Self> {
        let mut parts = fname.split("__");
        let mut key_parts = parts.next()?.split('-');
@@ -105,14 +103,14 @@ impl DeltaFileName {
            // or panic?
        }

-        Some(DeltaFileName {
+        Some(DeltaLayerName {
            key_range: key_start..key_end,
            lsn_range: start_lsn..end_lsn,
        })
    }
 }

-impl fmt::Display for DeltaFileName {
+impl fmt::Display for DeltaLayerName {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(
            f,
@@ -126,29 +124,29 @@ impl fmt::Display for DeltaFileName {
 }

 #[derive(PartialEq, Eq, Clone, Hash)]
-pub struct ImageFileName {
+pub struct ImageLayerName {
    pub key_range: Range<Key>,
    pub lsn: Lsn,
 }

-impl std::fmt::Debug for ImageFileName {
+impl std::fmt::Debug for ImageLayerName {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        use super::RangeDisplayDebug;

-        f.debug_struct("ImageFileName")
+        f.debug_struct("ImageLayerName")
            .field("key_range", &RangeDisplayDebug(&self.key_range))
            .field("lsn", &self.lsn)
            .finish()
    }
 }

-impl PartialOrd for ImageFileName {
+impl PartialOrd for ImageLayerName {
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
 }

-impl Ord for ImageFileName {
+impl Ord for ImageLayerName {
    fn cmp(&self, other: &Self) -> Ordering {
        let mut cmp = self.key_range.start.cmp(&other.key_range.start);
        if cmp != Ordering::Equal {
@@ -164,7 +162,7 @@ impl Ord for ImageFileName {
    }
 }

-impl ImageFileName {
+impl ImageLayerName {
    pub fn lsn_as_range(&self) -> Range<Lsn> {
        // Saves from having to copypaste this all over
        PersistentLayerDesc::image_layer_lsn_range(self.lsn)
@@ -172,16 +170,14 @@ impl ImageFileName {
 }

 ///
-/// Represents the filename of an ImageLayer
+/// Represents the part of the Key-LSN space covered by an ImageLayer
 ///
 /// ```text
 ///    <key start>-<key end>__<LSN>
 /// ```
-impl ImageFileName {
-    ///
-    /// Parse a string as an image file name. Returns None if the filename does not
-    /// match the expected pattern.
-    ///
+impl ImageLayerName {
+    /// Parse a string as then LayerName part of an image layer file name. Returns None if the
+    /// filename does not match the expected pattern.
    pub fn parse_str(fname: &str) -> Option<Self> {
        let mut parts = fname.split("__");
        let mut key_parts = parts.next()?.split('-');
@@ -202,14 +198,14 @@ impl ImageFileName {

        let lsn = Lsn::from_hex(lsn_str).ok()?;

-        Some(ImageFileName {
+        Some(ImageLayerName {
            key_range: key_start..key_end,
            lsn,
        })
    }
 }

-impl fmt::Display for ImageFileName {
+impl fmt::Display for ImageLayerName {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(
            f,
@@ -220,21 +216,24 @@ impl fmt::Display for ImageFileName {
        )
    }
 }
+
+/// LayerName is the logical identity of a layer within a LayerMap at a moment in time.  The
+/// LayerName is not a unique filename, as the same LayerName may have multiple physical incarnations
+/// over time (e.g. across shard splits or compression). The physical filenames of layers in local
+/// storage and object names in remote storage consist of the LayerName plus some extra qualifiers
+/// that uniquely identify the physical incarnation of a layer (see [crate::tenant::remote_timeline_client::remote_layer_path])
+/// and [`crate::tenant::storage_layer::layer::local_layer_path`])
 #[derive(Debug, PartialEq, Eq, Hash, Clone)]
-pub enum LayerFileName {
-    Image(ImageFileName),
-    Delta(DeltaFileName),
+pub enum LayerName {
+    Image(ImageLayerName),
+    Delta(DeltaLayerName),
 }

-impl LayerFileName {
-    pub fn file_name(&self) -> String {
-        self.to_string()
-    }
-
+impl LayerName {
    /// Determines if this layer file is considered to be in future meaning we will discard these
    /// layers during timeline initialization from the given disk_consistent_lsn.
    pub(crate) fn is_in_future(&self, disk_consistent_lsn: Lsn) -> bool {
-        use LayerFileName::*;
+        use LayerName::*;
        match self {
            Image(file_name) if file_name.lsn > disk_consistent_lsn => true,
            Delta(file_name) if file_name.lsn_range.end > disk_consistent_lsn + 1 => true,
@@ -243,7 +242,7 @@ impl LayerFileName {
    }

    pub(crate) fn kind(&self) -> &'static str {
-        use LayerFileName::*;
+        use LayerName::*;
        match self {
            Delta(_) => "delta",
            Image(_) => "image",
@@ -251,7 +250,7 @@ impl LayerFileName {
    }
 }

-impl fmt::Display for LayerFileName {
+impl fmt::Display for LayerName {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::Image(fname) => write!(f, "{fname}"),
@@ -260,25 +259,25 @@ impl fmt::Display for LayerFileName {
    }
 }

-impl From<ImageFileName> for LayerFileName {
-    fn from(fname: ImageFileName) -> Self {
+impl From<ImageLayerName> for LayerName {
+    fn from(fname: ImageLayerName) -> Self {
        Self::Image(fname)
    }
 }
-impl From<DeltaFileName> for LayerFileName {
-    fn from(fname: DeltaFileName) -> Self {
+impl From<DeltaLayerName> for LayerName {
+    fn from(fname: DeltaLayerName) -> Self {
        Self::Delta(fname)
    }
 }

-impl FromStr for LayerFileName {
+impl FromStr for LayerName {
    type Err = String;

    /// Conversion from either a physical layer filename, or the string-ization of
    /// Self. When loading a physical layer filename, we drop any extra information
    /// not needed to build Self.
    fn from_str(value: &str) -> Result<Self, Self::Err> {
-        let gen_suffix_regex = Regex::new("^(?<base>.+)-(?<gen>[0-9a-f]{8})$").unwrap();
+        let gen_suffix_regex = Regex::new("^(?<base>.+)(?<gen>-v1-[0-9a-f]{8})$").unwrap();
        let file_name: Cow<str> = match gen_suffix_regex.captures(value) {
            Some(captures) => captures
                .name("base")
@@ -288,8 +287,8 @@ impl FromStr for LayerFileName {
            None => value.into(),
        };

-        let delta = DeltaFileName::parse_str(&file_name);
-        let image = ImageFileName::parse_str(&file_name);
+        let delta = DeltaLayerName::parse_str(&file_name);
+        let image = ImageLayerName::parse_str(&file_name);
        let ok = match (delta, image) {
            (None, None) => {
                return Err(format!(
@@ -304,7 +303,7 @@ impl FromStr for LayerFileName {
    }
 }

-impl serde::Serialize for LayerFileName {
+impl serde::Serialize for LayerName {
    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
    where
        S: serde::Serializer,
@@ -316,19 +315,19 @@ impl serde::Serialize for LayerFileName {
    }
 }

-impl<'de> serde::Deserialize<'de> for LayerFileName {
+impl<'de> serde::Deserialize<'de> for LayerName {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: serde::Deserializer<'de>,
    {
-        deserializer.deserialize_string(LayerFileNameVisitor)
+        deserializer.deserialize_string(LayerNameVisitor)
    }
 }

-struct LayerFileNameVisitor;
+struct LayerNameVisitor;

-impl<'de> serde::de::Visitor<'de> for LayerFileNameVisitor {
-    type Value = LayerFileName;
+impl<'de> serde::de::Visitor<'de> for LayerNameVisitor {
+    type Value = LayerName;

    fn expecting(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
        write!(
@@ -348,37 +347,33 @@ impl<'de> serde::de::Visitor<'de> for LayerFileNameVisitor {
 mod test {
    use super::*;
    #[test]
-    fn image_layer_parse() -> anyhow::Result<()> {
-        let expected = LayerFileName::Image(ImageFileName {
+    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 = LayerFileName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-00000001").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 = LayerFileName::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<()> {
-        let expected = LayerFileName::Delta(DeltaFileName {
+    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 = LayerFileName::from_str("000000000000000000000000000000000000-000000067F00000001000004DF0000000006__00000000014FED58-000000000154C481-00000001").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 = LayerFileName::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(())
    }
 }
--- a/pageserver/src/tenant/tasks.rs
+++ b/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(|| {
@@ -41,7 +41,7 @@ static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore
        tokio::sync::Semaphore::new(permits)
    });

-#[derive(Debug, PartialEq, Eq, Clone, Copy, strum_macros::IntoStaticStr)]
+#[derive(Debug, PartialEq, Eq, Clone, Copy, strum_macros::IntoStaticStr, enum_map::Enum)]
 #[strum(serialize_all = "snake_case")]
 pub(crate) enum BackgroundLoopKind {
    Compaction,
@@ -57,19 +57,25 @@ pub(crate) enum BackgroundLoopKind {

 impl BackgroundLoopKind {
    fn as_static_str(&self) -> &'static str {
-        let s: &'static str = self.into();
-        s
+        self.into()
    }
 }

+static PERMIT_GAUGES: once_cell::sync::Lazy<
+    enum_map::EnumMap<BackgroundLoopKind, metrics::IntCounterPair>,
+> = once_cell::sync::Lazy::new(|| {
+    enum_map::EnumMap::from_array(std::array::from_fn(|i| {
+        let kind = <BackgroundLoopKind as enum_map::Enum>::from_usize(i);
+        crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_GAUGE.with_label_values(&[kind.into()])
+    }))
+});
+
 /// Cancellation safe.
 pub(crate) async fn concurrent_background_tasks_rate_limit_permit(
    loop_kind: BackgroundLoopKind,
    _ctx: &RequestContext,
 ) -> tokio::sync::SemaphorePermit<'static> {
-    let _guard = crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_GAUGE
-        .with_label_values(&[loop_kind.as_static_str()])
-        .guard();
+    let _guard = PERMIT_GAUGES[loop_kind].guard();

    pausable_failpoint!(
        "initial-size-calculation-permit-pause",
--- a/pageserver/src/tenant/timeline.rs
+++ b/pageserver/src/tenant/timeline.rs
--- a/pageserver/src/tenant/timeline/compaction.rs
+++ b/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.
@@ -700,6 +752,7 @@ impl Timeline {
                                debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
                                lsn_range.clone()
                            },
+                            ctx,
                        )
                        .await?,
                    );
@@ -755,6 +808,7 @@ impl Timeline {
                &self
                    .conf
                    .timeline_path(&self.tenant_shard_id, &self.timeline_id),
+                ctx,
            )
            .await
            .fatal_err("VirtualFile::open for timeline dir fsync");
@@ -1093,6 +1147,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
            self.timeline.tenant_shard_id,
            key_range.start,
            lsn_range.clone(),
+            ctx,
        )
        .await?;

@@ -1158,60 +1213,48 @@ 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,
            key_range,
            lsn,
+            ctx,
        )
        .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();

--- a/Show More
+++ b/Show More