separate rwlock for second level of immutable_page_map to avoid contention

Before this patch, `drop_buffers_for_immutable` would scan the entire
`slots` to proactively evict pages.

Some background on `drop_buffers_for_immutable`: it's really just a nice
courtesy to give back the slots that it used. Why is it just
nice-to-have? We're dropping the EphemeralFile, so, these slots are
never going to get accessed again.
The `find_victim` will eventually evict them.

.github/workflows/build_and_test.yml

@@ -737,34 +737,6 @@ jobs:
                            --destination neondatabase/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
                            --cleanup
 
-      # Due to a kaniko bug, we can't use cache for extensions image, thus it takes about the same amount of time as compute-node image to build (~10 min)
-      # During the transition period we need to have extensions in both places (in S3 and in compute-node image),
-      # so we won't build extension twice, but extract them from compute-node.
-      #
-      # For now we use extensions image only for new custom extensitons
-      - name: Kaniko build extensions only
-        run: |
-          # Kaniko is suposed to clean up after itself if --cleanup flag is set, but it doesn't.
-          # Despite some fixes were made in https://github.com/GoogleContainerTools/kaniko/pull/2504 (in kaniko v1.11.0),
-          # it still fails with error:
-          #   error building image: could not save file: copying file: symlink postgres /kaniko/1/usr/local/pgsql/bin/postmaster: file exists
-          #
-          # Ref https://github.com/GoogleContainerTools/kaniko/issues/1406
-          find /kaniko -maxdepth 1 -mindepth 1 -type d -regex "/kaniko/[0-9]*" -exec rm -rv {} \;
-
-          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true \
-                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache \
-                           --context . \
-                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }} \
-                           --build-arg PG_VERSION=${{ matrix.version }} \
-                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}} \
-                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com \
-                           --dockerfile Dockerfile.compute-node \
-                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
-                           --destination neondatabase/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
-                           --cleanup \
-                           --target postgres-extensions
-
       # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
       - name: Cleanup ECR folder
         run: rm -rf ~/.ecr
@@ -780,7 +752,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.15.4
+      VM_BUILDER_VERSION: v0.17.5
 
     steps:
       - name: Checkout
@@ -801,7 +773,11 @@ jobs:
 
       - name: Build vm image
         run: |
-          ./vm-builder -enable-file-cache -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}}
+          ./vm-builder \
+            -enable-file-cache \
+            -cgroup-uid=postgres \
+            -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}}
 
       - name: Pushing vm-compute-node image
         run: |
@@ -882,10 +858,8 @@ jobs:
           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:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{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/extensions-v15:${{needs.tag.outputs.build-tag}} latest
 
       - name: Push images to production ECR
         if: |
@@ -896,10 +870,8 @@ jobs:
           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:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-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/extensions-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:latest
 
       - name: Configure Docker Hub login
         run: |
@@ -921,65 +893,56 @@ jobs:
           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:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/extensions-v14:${{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/extensions-v15:${{needs.tag.outputs.build-tag}} latest
 
       - name: Cleanup ECR folder
         run: rm -rf ~/.ecr
 
-  upload-postgres-extensions-to-s3:
-    if: |
-      (github.ref_name == 'main' || github.ref_name == 'release') &&
-       github.event_name != 'workflow_dispatch'
-    runs-on: ${{ github.ref_name == 'release' && fromJSON('["self-hosted", "prod", "x64"]') || fromJSON('["self-hosted", "gen3", "small"]') }}
-    needs: [ tag, promote-images ]
-    strategy:
-      fail-fast: false
-      matrix:
-        version: [ v14, v15 ]
-
-    env:
-      EXTENSIONS_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
-      AWS_ACCESS_KEY_ID: ${{ github.ref_name == 'release' && secrets.AWS_ACCESS_KEY_PROD || secrets.AWS_ACCESS_KEY_DEV }}
-      AWS_SECRET_ACCESS_KEY: ${{ github.ref_name == 'release' && secrets.AWS_SECRET_KEY_PROD || secrets.AWS_SECRET_KEY_DEV }}
-      S3_BUCKETS: ${{ github.ref_name == 'release' && vars.S3_EXTENSIONS_BUCKETS_PROD || vars.S3_EXTENSIONS_BUCKETS_DEV }}
-
+  build-private-extensions:
+    runs-on: [ self-hosted, gen3, small ]
+    container:
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
+      options: --init
+    needs: [ tag ]
     steps:
-      - name: Pull postgres-extensions image
+      - name: Set PR's status to pending and request a remote CI test
         run: |
-          docker pull ${EXTENSIONS_IMAGE}
+          COMMIT_SHA=${{ github.event.pull_request.head.sha }}
+          COMMIT_SHA=${COMMIT_SHA:-${{ github.sha }}}
+          REMOTE_REPO="${{ github.repository_owner }}/build-custom-extensions"
 
-      - name: Create postgres-extensions container
-        id: create-container
-        run: |
-          EID=$(docker create ${EXTENSIONS_IMAGE} true)
-          echo "EID=${EID}" >> $GITHUB_OUTPUT
+          curl -f -X POST \
+          https://api.github.com/repos/${{ github.repository }}/statuses/$COMMIT_SHA \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"state\": \"pending\",
+              \"context\": \"build-and-upload-extensions\",
+              \"description\": \"[$REMOTE_REPO] Remote CI job is about to start\"
+            }"
 
-      - name: Extract postgres-extensions from container
-        run: |
-          rm -rf ./extensions-to-upload # Just in case
-          mkdir -p extensions-to-upload
-
-          docker cp ${{ steps.create-container.outputs.EID }}:/extensions/ ./extensions-to-upload/
-          docker cp ${{ steps.create-container.outputs.EID }}:/ext_index.json ./extensions-to-upload/
-
-      - name: Upload postgres-extensions to S3
-        run: |
-          for BUCKET in $(echo ${S3_BUCKETS:-[]} | jq --raw-output '.[]'); do
-            aws s3 cp --recursive --only-show-errors ./extensions-to-upload s3://${BUCKET}/${{ needs.tag.outputs.build-tag }}/${{ matrix.version }}
-          done
-
-      - name: Cleanup
-        if: ${{ always() && steps.create-container.outputs.EID }}
-        run: |
-          docker rm ${{ steps.create-container.outputs.EID }} || true
+          curl -f -X POST \
+          https://api.github.com/repos/$REMOTE_REPO/actions/workflows/build_and_upload_extensions.yml/dispatches \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"ref\": \"main\",
+              \"inputs\": {
+                \"ci_job_name\": \"build-and-upload-extensions\",
+                \"commit_hash\": \"$COMMIT_SHA\",
+                \"remote_repo\": \"${{ github.repository }}\",
+                \"compute_image_tag\": \"${{ needs.tag.outputs.build-tag }}\",
+                \"remote_branch_name\": \"${{ github.ref_name }}\"
+              }
+            }"
 
   deploy:
     runs-on: [ self-hosted, gen3, small ]
     container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/ansible:latest
-    needs: [ upload-postgres-extensions-to-s3, promote-images, tag, regress-tests ]
+    needs: [ promote-images, tag, regress-tests ]
     if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
     steps:
       - name: Fix git ownership

CODEOWNERS

@@ -1,11 +1,12 @@
-/compute_tools/ @neondatabase/control-plane
+/compute_tools/ @neondatabase/control-plane @neondatabase/compute
 /control_plane/ @neondatabase/compute @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/compute @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute 
-/libs/remote_storage/ @neondatabase/storage 
-/libs/safekeeper_api/ @neondatabase/safekeepers  
-/pageserver/ @neondatabase/compute @neondatabase/storage 
+/libs/postgres_ffi/ @neondatabase/compute
+/libs/remote_storage/ @neondatabase/storage
+/libs/safekeeper_api/ @neondatabase/safekeepers
+/libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
+/pageserver/ @neondatabase/compute @neondatabase/storage
 /pgxn/ @neondatabase/compute
-/proxy/ @neondatabase/control-plane 
+/proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
 /vendor/ @neondatabase/compute

Cargo.lock

@@ -190,7 +190,7 @@ checksum = "16e62a023e7c117e27523144c5d2459f4397fcc3cab0085af8e2224f643a0193"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -201,7 +201,7 @@ checksum = "b9ccdd8f2a161be9bd5c023df56f1b2a0bd1d83872ae53b71a84a12c9bf6e842"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -553,12 +553,13 @@ dependencies = [
 
 [[package]]
 name = "axum"
-version = "0.6.18"
+version = "0.6.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8175979259124331c1d7bf6586ee7e0da434155e4b2d48ec2c8386281d8df39"
+checksum = "3b829e4e32b91e643de6eafe82b1d90675f5874230191a4ffbc1b336dec4d6bf"
 dependencies = [
  "async-trait",
  "axum-core",
+ "base64 0.21.1",
  "bitflags",
  "bytes",
  "futures-util",
@@ -573,7 +574,13 @@ dependencies = [
  "pin-project-lite",
  "rustversion",
  "serde",
+ "serde_json",
+ "serde_path_to_error",
+ "serde_urlencoded",
+ "sha1",
  "sync_wrapper",
+ "tokio",
+ "tokio-tungstenite 0.20.0",
  "tower",
  "tower-layer",
  "tower-service",
@@ -673,7 +680,7 @@ dependencies = [
  "regex",
  "rustc-hash",
  "shlex",
- "syn 2.0.16",
+ "syn 2.0.28",
  "which",
 ]
 
@@ -765,6 +772,19 @@ version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
 
+[[package]]
+name = "cgroups-rs"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fb3af90c8d48ad5f432d8afb521b5b40c2a2fce46dd60e05912de51c47fba64"
+dependencies = [
+ "libc",
+ "log",
+ "nix 0.25.1",
+ "regex",
+ "thiserror",
+]
+
 [[package]]
 name = "chrono"
 version = "0.4.24"
@@ -849,7 +869,7 @@ dependencies = [
  "heck",
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -907,6 +927,7 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-compression",
+ "cfg-if",
  "chrono",
  "clap",
  "compute_api",
@@ -925,6 +946,7 @@ dependencies = [
  "tar",
  "tokio",
  "tokio-postgres",
+ "tokio-util",
  "toml_edit",
  "tracing",
  "tracing-opentelemetry",
@@ -932,6 +954,7 @@ dependencies = [
  "tracing-utils",
  "url",
  "utils",
+ "vm_monitor",
  "workspace_hack",
  "zstd",
 ]
@@ -978,7 +1001,7 @@ dependencies = [
  "comfy-table",
  "compute_api",
  "git-version",
- "nix",
+ "nix 0.26.2",
  "once_cell",
  "pageserver_api",
  "postgres",
@@ -1184,7 +1207,7 @@ dependencies = [
  "proc-macro2",
  "quote",
  "strsim",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -1195,7 +1218,7 @@ checksum = "29a358ff9f12ec09c3e61fef9b5a9902623a695a46a917b07f269bff1445611a"
 dependencies = [
  "darling_core",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -1260,7 +1283,7 @@ checksum = "487585f4d0c6655fe74905e2504d8ad6908e4db67f744eb140876906c2f3175d"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -1316,7 +1339,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -1512,7 +1535,7 @@ checksum = "89ca545a94061b6365f2c7355b4b32bd20df3ff95f02da9329b34ccc3bd6ee72"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -1863,8 +1886,8 @@ dependencies = [
  "hyper",
  "pin-project",
  "tokio",
- "tokio-tungstenite",
- "tungstenite",
+ "tokio-tungstenite 0.18.0",
+ "tungstenite 0.18.0",
 ]
 
 [[package]]
@@ -1928,6 +1951,19 @@ dependencies = [
  "libc",
 ]
 
+[[package]]
+name = "inotify"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fdd168d97690d0b8c412d6b6c10360277f4d7ee495c5d0d5d5fe0854923255cc"
+dependencies = [
+ "bitflags",
+ "futures-core",
+ "inotify-sys",
+ "libc",
+ "tokio",
+]
+
 [[package]]
 name = "inotify-sys"
 version = "0.1.5"
@@ -2251,6 +2287,18 @@ dependencies = [
  "tempfile",
 ]
 
+[[package]]
+name = "nix"
+version = "0.25.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f346ff70e7dbfd675fe90590b92d59ef2de15a8779ae305ebcbfd3f0caf59be4"
+dependencies = [
+ "autocfg",
+ "bitflags",
+ "cfg-if",
+ "libc",
+]
+
 [[package]]
 name = "nix"
 version = "0.26.2"
@@ -2285,7 +2333,7 @@ dependencies = [
  "crossbeam-channel",
  "filetime",
  "fsevent-sys",
- "inotify",
+ "inotify 0.9.6",
  "kqueue",
  "libc",
  "mio",
@@ -2293,6 +2341,15 @@ dependencies = [
  "windows-sys 0.45.0",
 ]
 
+[[package]]
+name = "ntapi"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e8a3895c6391c39d7fe7ebc444a87eb2991b2a0bc718fdabd071eec617fc68e4"
+dependencies = [
+ "winapi",
+]
+
 [[package]]
 name = "num-bigint"
 version = "0.4.3"
@@ -2386,7 +2443,7 @@ checksum = "a948666b637a0f465e8564c73e89d4dde00d72d4d473cc972f390fc3dcee7d9c"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -2573,7 +2630,7 @@ dependencies = [
  "hyper",
  "itertools",
  "metrics",
- "nix",
+ "nix 0.26.2",
  "num-traits",
  "num_cpus",
  "once_cell",
@@ -2596,6 +2653,7 @@ dependencies = [
  "serde_json",
  "serde_with",
  "signal-hook",
+ "smallvec",
  "storage_broker",
  "strum",
  "strum_macros",
@@ -2773,7 +2831,7 @@ checksum = "39407670928234ebc5e6e580247dd567ad73a3578460c5990f9503df207e8f07"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -2970,7 +3028,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3b69d39aab54d069e7f2fe8cb970493e7834601ca2d8c65fd7bbd183578080d1"
 dependencies = [
  "proc-macro2",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -2981,9 +3039,9 @@ checksum = "dc375e1527247fe1a97d8b7156678dfe7c1af2fc075c9a4db3690ecd2a148068"
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.64"
+version = "1.0.66"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "78803b62cbf1f46fde80d7c0e803111524b9877184cfe7c3033659490ac7a7da"
+checksum = "18fb31db3f9bddb2ea821cde30a9f70117e3f119938b5ee630b7403aa6e2ead9"
 dependencies = [
  "unicode-ident",
 ]
@@ -3145,9 +3203,9 @@ dependencies = [
 
 [[package]]
 name = "quote"
-version = "1.0.27"
+version = "1.0.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8f4f29d145265ec1c483c7c654450edde0bfe043d3938d6972630663356d9500"
+checksum = "50f3b39ccfb720540debaa0164757101c08ecb8d326b15358ce76a62c7e85965"
 dependencies = [
  "proc-macro2",
 ]
@@ -3569,9 +3627,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-webpki"
-version = "0.100.1"
+version = "0.100.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6207cd5ed3d8dca7816f8f3725513a34609c0c765bf652b8c3cb4cfd87db46b"
+checksum = "e98ff011474fa39949b7e5c0428f9b4937eda7da7848bbb947786b7be0b27dab"
 dependencies = [
  "ring",
  "untrusted",
@@ -3798,22 +3856,22 @@ dependencies = [
 
 [[package]]
 name = "serde"
-version = "1.0.163"
+version = "1.0.183"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2113ab51b87a539ae008b5c6c02dc020ffa39afd2d83cffcb3f4eb2722cebec2"
+checksum = "32ac8da02677876d532745a130fc9d8e6edfa81a269b107c5b00829b91d8eb3c"
 dependencies = [
  "serde_derive",
 ]
 
 [[package]]
 name = "serde_derive"
-version = "1.0.163"
+version = "1.0.183"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8c805777e3930c8883389c602315a24224bcc738b63905ef87cd1420353ea93e"
+checksum = "aafe972d60b0b9bee71a91b92fee2d4fb3c9d7e8f6b179aa99f27203d99a4816"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -3827,6 +3885,16 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "serde_path_to_error"
+version = "0.1.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4beec8bce849d58d06238cb50db2e1c417cfeafa4c63f692b15c82b7c80f8335"
+dependencies = [
+ "itoa",
+ "serde",
+]
+
 [[package]]
 name = "serde_spanned"
 version = "0.6.2"
@@ -3873,7 +3941,7 @@ dependencies = [
  "darling",
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -3972,9 +4040,9 @@ dependencies = [
 
 [[package]]
 name = "smallvec"
-version = "1.10.0"
+version = "1.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a507befe795404456341dfab10cef66ead4c041f62b8b11bbb92bffe5d0953e0"
+checksum = "62bb4feee49fdd9f707ef802e22365a35de4b7b299de4763d44bfea899442ff9"
 
 [[package]]
 name = "socket2"
@@ -4111,9 +4179,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.16"
+version = "2.0.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a6f671d4b5ffdb8eadec19c0ae67fe2639df8684bd7bc4b83d986b8db549cf01"
+checksum = "04361975b3f5e348b2189d8dc55bc942f278b2d482a6a0365de5bdd62d351567"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4139,14 +4207,29 @@ dependencies = [
 ]
 
 [[package]]
-name = "tar"
-version = "0.4.38"
+name = "sysinfo"
+version = "0.29.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4b55807c0344e1e6c04d7c965f5289c39a8d94ae23ed5c0b57aabac549f871c6"
+checksum = "165d6d8539689e3d3bc8b98ac59541e1f21c7de7c85d60dc80e43ae0ed2113db"
+dependencies = [
+ "cfg-if",
+ "core-foundation-sys",
+ "libc",
+ "ntapi",
+ "once_cell",
+ "rayon",
+ "winapi",
+]
+
+[[package]]
+name = "tar"
+version = "0.4.40"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b16afcea1f22891c49a00c751c7b63b2233284064f11a200fc624137c51e2ddb"
 dependencies = [
  "filetime",
  "libc",
- "xattr 0.2.3",
+ "xattr",
 ]
 
 [[package]]
@@ -4228,7 +4311,7 @@ checksum = "f9456a42c5b0d803c8cd86e73dd7cc9edd429499f37a3550d286d5e86720569f"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -4343,7 +4426,7 @@ checksum = "630bdcf245f78637c13ec01ffae6187cca34625e8c63150d424b59e55af2675e"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -4437,7 +4520,7 @@ dependencies = [
  "redox_syscall 0.3.5",
  "tokio",
  "tokio-stream",
- "xattr 1.0.0",
+ "xattr",
 ]
 
 [[package]]
@@ -4449,7 +4532,19 @@ dependencies = [
  "futures-util",
  "log",
  "tokio",
- "tungstenite",
+ "tungstenite 0.18.0",
+]
+
+[[package]]
+name = "tokio-tungstenite"
+version = "0.20.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b2dbec703c26b00d74844519606ef15d09a7d6857860f84ad223dec002ddea2"
+dependencies = [
+ "futures-util",
+ "log",
+ "tokio",
+ "tungstenite 0.20.0",
 ]
 
 [[package]]
@@ -4641,7 +4736,7 @@ checksum = "0f57e3ca2a01450b1a921183a9c9cbfda207fd822cef4ccb00a65402cbba7a74"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
 ]
 
 [[package]]
@@ -4770,6 +4865,25 @@ dependencies = [
  "utf-8",
 ]
 
+[[package]]
+name = "tungstenite"
+version = "0.20.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e862a1c4128df0112ab625f55cd5c934bcb4312ba80b39ae4b4835a3fd58e649"
+dependencies = [
+ "byteorder",
+ "bytes",
+ "data-encoding",
+ "http",
+ "httparse",
+ "log",
+ "rand",
+ "sha1",
+ "thiserror",
+ "url",
+ "utf-8",
+]
+
 [[package]]
 name = "typenum"
 version = "1.16.0"
@@ -4897,7 +5011,7 @@ dependencies = [
  "hyper",
  "jsonwebtoken",
  "metrics",
- "nix",
+ "nix 0.26.2",
  "once_cell",
  "pin-project-lite",
  "pq_proto",
@@ -4915,6 +5029,7 @@ dependencies = [
  "thiserror",
  "tokio",
  "tokio-stream",
+ "tokio-util",
  "tracing",
  "tracing-error",
  "tracing-subscriber",
@@ -4951,6 +5066,28 @@ version = "0.9.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "49874b5167b65d7193b8aba1567f5c7d93d001cafc34600cee003eda787e483f"
 
+[[package]]
+name = "vm_monitor"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "axum",
+ "cgroups-rs",
+ "clap",
+ "futures",
+ "inotify 0.10.2",
+ "serde",
+ "serde_json",
+ "sysinfo",
+ "tokio",
+ "tokio-postgres",
+ "tokio-stream",
+ "tokio-util",
+ "tracing",
+ "tracing-subscriber",
+ "workspace_hack",
+]
+
 [[package]]
 name = "vsimd"
 version = "0.8.0"
@@ -5021,7 +5158,7 @@ dependencies = [
  "once_cell",
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
  "wasm-bindgen-shared",
 ]
 
@@ -5055,7 +5192,7 @@ checksum = "e128beba882dd1eb6200e1dc92ae6c5dbaa4311aa7bb211ca035779e5efc39f8"
 dependencies = [
  "proc-macro2",
  "quote",
- "syn 2.0.16",
+ "syn 2.0.28",
  "wasm-bindgen-backend",
  "wasm-bindgen-shared",
 ]
@@ -5340,12 +5477,14 @@ name = "workspace_hack"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "axum",
  "bytes",
  "cc",
  "chrono",
  "clap",
  "clap_builder",
  "crossbeam-utils",
+ "digest",
  "either",
  "fail",
  "futures",
@@ -5354,6 +5493,7 @@ dependencies = [
  "futures-executor",
  "futures-sink",
  "futures-util",
+ "hyper",
  "itertools",
  "libc",
  "log",
@@ -5372,9 +5512,10 @@ dependencies = [
  "scopeguard",
  "serde",
  "serde_json",
+ "smallvec",
  "socket2 0.4.9",
  "syn 1.0.109",
- "syn 2.0.16",
+ "syn 2.0.28",
  "tokio",
  "tokio-rustls 0.23.4",
  "tokio-util",
@@ -5383,7 +5524,6 @@ dependencies = [
  "tower",
  "tracing",
  "tracing-core",
- "tracing-subscriber",
  "url",
 ]
 
@@ -5404,15 +5544,6 @@ dependencies = [
  "time",
 ]
 
-[[package]]
-name = "xattr"
-version = "0.2.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d1526bbe5aaeb5eb06885f4d987bcdfa5e23187055de9b83fe00156a821fabc"
-dependencies = [
- "libc",
-]
-
 [[package]]
 name = "xattr"
 version = "1.0.0"

Cargo.toml

@@ -23,6 +23,7 @@ members = [
     "libs/remote_storage",
     "libs/tracing-utils",
     "libs/postgres_ffi/wal_craft",
+    "libs/vm_monitor",
 ]
 
 [workspace.package]
@@ -41,12 +42,14 @@ aws-sdk-s3 = "0.27"
 aws-smithy-http = "0.55"
 aws-credential-types = "0.55"
 aws-types = "0.55"
+axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
 bindgen = "0.65"
 bstr = "1.0"
 byteorder = "1.4"
 bytes = "1.0"
+cfg-if = "1.0.0"
 chrono = { version = "0.4", default-features = false, features = ["clock"] }
 clap = { version = "4.0", features = ["derive"] }
 close_fds = "0.3.2"
@@ -74,6 +77,7 @@ humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
 hyper-tungstenite = "0.9"
+inotify = "0.10.2"
 itertools = "0.10"
 jsonwebtoken = "8"
 libc = "0.2"
@@ -105,12 +109,14 @@ rustls = "0.20"
 rustls-pemfile = "1"
 rustls-split = "0.3"
 scopeguard = "1.1"
+sysinfo = "0.29.2"
 sentry = { version = "0.30", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_with = "2.0"
 sha2 = "0.10.2"
 signal-hook = "0.3"
+smallvec = "1.11"
 socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
@@ -133,7 +139,7 @@ tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tracing = "0.1"
 tracing-error = "0.2.0"
 tracing-opentelemetry = "0.19.0"
-tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter"] }
+tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 url = "2.2"
 uuid = { version = "1.2", features = ["v4", "serde"] }
 walkdir = "2.3.2"
@@ -169,6 +175,7 @@ storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main br
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
+vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }

Dockerfile.compute-node

@@ -764,29 +764,6 @@ RUN rm -r /usr/local/pgsql/include
 # if they were to be used by other libraries.
 RUN rm /usr/local/pgsql/lib/lib*.a
 
-#########################################################################################
-#
-# Extenstion only
-#
-#########################################################################################
-FROM python:3.9-slim-bullseye AS generate-ext-index
-ARG PG_VERSION
-ARG BUILD_TAG
-RUN apt update && apt install -y zstd
-
-# copy the control files here
-COPY --from=kq-imcx-pg-build /extensions/ /extensions/
-COPY --from=pg-anon-pg-build /extensions/ /extensions/
-COPY --from=postgis-build /extensions/ /extensions/
-COPY scripts/combine_control_files.py ./combine_control_files.py
-RUN python3 ./combine_control_files.py ${PG_VERSION} ${BUILD_TAG} --public_extensions="anon,postgis"
-
-FROM scratch AS postgres-extensions
-# After the transition this layer will include all extensitons.
-# As for now, it's only a couple for testing purposses
-COPY --from=generate-ext-index /extensions/*.tar.zst /extensions/
-COPY --from=generate-ext-index /ext_index.json /ext_index.json
-
 #########################################################################################
 #
 # Final layer

compute_tools/Cargo.toml

@@ -8,6 +8,7 @@ license.workspace = true
 anyhow.workspace = true
 async-compression.workspace = true
 chrono.workspace = true
+cfg-if.workspace = true
 clap.workspace = true
 flate2.workspace = true
 futures.workspace = true
@@ -23,6 +24,7 @@ tar.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tokio-postgres.workspace = true
+tokio-util.workspace = true
 tracing.workspace = true
 tracing-opentelemetry.workspace = true
 tracing-subscriber.workspace = true
@@ -34,4 +36,5 @@ utils.workspace = true
 workspace_hack.workspace = true
 toml_edit.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
+vm_monitor = { version = "0.1", path = "../libs/vm_monitor/" }
 zstd = "0.12.4"

compute_tools/src/bin/compute_ctl.rs

@@ -35,7 +35,6 @@
 //!
 use std::collections::HashMap;
 use std::fs::File;
-use std::panic;
 use std::path::Path;
 use std::process::exit;
 use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
@@ -271,6 +270,55 @@ fn main() -> Result<()> {
         }
     };
 
+    // Start the vm-monitor if directed to. The vm-monitor only runs on linux
+    // because it requires cgroups.
+    cfg_if::cfg_if! {
+        if #[cfg(target_os = "linux")] {
+            use std::env;
+            use tokio_util::sync::CancellationToken;
+            use tracing::warn;
+            let vm_monitor_addr = matches.get_one::<String>("vm-monitor-addr");
+            let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
+            let cgroup = matches.get_one::<String>("cgroup");
+
+            // Only make a runtime if we need to.
+            // Note: it seems like you can make a runtime in an inner scope and
+            // if you start a task in it it won't be dropped. However, make it
+            // in the outermost scope just to be safe.
+            let rt = match (env::var_os("AUTOSCALING"), vm_monitor_addr) {
+                (None, None) => None,
+                (None, Some(_)) => {
+                    warn!("--vm-monitor-addr option set but AUTOSCALING env var not present");
+                    None
+                }
+                (Some(_), None) => {
+                    panic!("AUTOSCALING env var present but --vm-monitor-addr option not set")
+                }
+                (Some(_), Some(_)) => Some(
+                    tokio::runtime::Builder::new_multi_thread()
+                        .worker_threads(4)
+                        .enable_all()
+                        .build()
+                        .expect("failed to create tokio runtime for monitor"),
+                ),
+            };
+
+            // This token is used internally by the monitor to clean up all threads
+            let token = CancellationToken::new();
+
+            let vm_monitor = &rt.as_ref().map(|rt| {
+                rt.spawn(vm_monitor::start(
+                    Box::leak(Box::new(vm_monitor::Args {
+                        cgroup: cgroup.cloned(),
+                        pgconnstr: file_cache_connstr.cloned(),
+                        addr: vm_monitor_addr.cloned().unwrap(),
+                    })),
+                    token.clone(),
+                ))
+            });
+        }
+    }
+
     // Wait for the child Postgres process forever. In this state Ctrl+C will
     // propagate to Postgres and it will be shut down as well.
     if let Some(mut pg) = pg {
@@ -284,6 +332,24 @@ fn main() -> Result<()> {
         exit_code = ecode.code()
     }
 
+    // Terminate the vm_monitor so it releases the file watcher on
+    // /sys/fs/cgroup/neon-postgres.
+    // Note: the vm-monitor only runs on linux because it requires cgroups.
+    cfg_if::cfg_if! {
+        if #[cfg(target_os = "linux")] {
+            if let Some(handle) = vm_monitor {
+                // Kills all threads spawned by the monitor
+                token.cancel();
+                // Kills the actual task running the monitor
+                handle.abort();
+
+                // If handle is some, rt must have been used to produce it, and
+                // hence is also some
+                rt.unwrap().shutdown_timeout(Duration::from_secs(2));
+            }
+        }
+    }
+
     // Maybe sync safekeepers again, to speed up next startup
     let compute_state = compute.state.lock().unwrap().clone();
     let pspec = compute_state.pspec.as_ref().expect("spec must be set");
@@ -393,6 +459,29 @@ fn cli() -> clap::Command {
                 .long("remote-ext-config")
                 .value_name("REMOTE_EXT_CONFIG"),
         )
+        // TODO(fprasx): we currently have default arguments because the cloud PR
+        // to pass them in hasn't been merged yet. We should get rid of them once
+        // the PR is merged.
+        .arg(
+            Arg::new("vm-monitor-addr")
+                .long("vm-monitor-addr")
+                .default_value("0.0.0.0:10301")
+                .value_name("VM_MONITOR_ADDR"),
+        )
+        .arg(
+            Arg::new("cgroup")
+                .long("cgroup")
+                .default_value("neon-postgres")
+                .value_name("CGROUP"),
+        )
+        .arg(
+            Arg::new("filecache-connstr")
+                .long("filecache-connstr")
+                .default_value(
+                    "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable",
+                )
+                .value_name("FILECACHE_CONNSTR"),
+        )
 }
 
 #[test]

compute_tools/src/compute.rs

@@ -1,4 +1,5 @@
 use std::collections::HashMap;
+use std::env;
 use std::fs;
 use std::io::BufRead;
 use std::os::unix::fs::PermissionsExt;
@@ -175,6 +176,27 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
     }
 }
 
+/// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
+/// cgroup. Otherwise returns the default `Command::new(cmd)`
+///
+/// This function should be used to start postgres, as it will start it in the
+/// neon-postgres cgroup if we are a VM. This allows autoscaling to control
+/// postgres' resource usage. The cgroup will exist in VMs because vm-builder
+/// creates it during the sysinit phase of its inittab.
+fn maybe_cgexec(cmd: &str) -> Command {
+    // The cplane sets this env var for autoscaling computes.
+    // use `var_os` so we don't have to worry about the variable being valid
+    // unicode. Should never be an concern . . . but just in case
+    if env::var_os("AUTOSCALING").is_some() {
+        let mut command = Command::new("cgexec");
+        command.args(["-g", "memory:neon-postgres"]);
+        command.arg(cmd);
+        command
+    } else {
+        Command::new(cmd)
+    }
+}
+
 /// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
 /// that we give to customers
 fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
@@ -451,7 +473,7 @@ impl ComputeNode {
     pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
         let start_time = Utc::now();
 
-        let sync_handle = Command::new(&self.pgbin)
+        let sync_handle = maybe_cgexec(&self.pgbin)
             .args(["--sync-safekeepers"])
             .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
             .envs(if let Some(storage_auth_token) = &storage_auth_token {
@@ -586,7 +608,7 @@ impl ComputeNode {
 
         // Start postgres
         info!("starting postgres");
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
             .args(["-D", pgdata])
             .spawn()
             .expect("cannot start postgres process");
@@ -614,7 +636,7 @@ impl ComputeNode {
         let pgdata_path = Path::new(&self.pgdata);
 
         // Run postgres as a child process.
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
             .args(["-D", &self.pgdata])
             .envs(if let Some(storage_auth_token) = &storage_auth_token {
                 vec![("NEON_AUTH_TOKEN", storage_auth_token)]

deny.toml

@@ -4,7 +4,12 @@
 # to your expectations and requirements.
 
 # Root options
-targets = []
+targets = [
+    { triple = "x86_64-unknown-linux-gnu" },
+    { triple = "aarch64-unknown-linux-gnu" },
+    { triple = "aarch64-apple-darwin" },
+    { triple = "x86_64-apple-darwin" },
+]
 all-features = false
 no-default-features = false
 feature-depth = 1
@@ -18,7 +23,7 @@ vulnerability = "deny"
 unmaintained = "warn"
 yanked = "warn"
 notice = "warn"
-ignore = []
+ignore = ["RUSTSEC-2023-0052"]
 
 # This section is considered when running `cargo deny check licenses`
 # More documentation for the licenses section can be found here:

docs/rfcs/026-pageserver-s3-mvcc.md

@@ -0,0 +1,316 @@
+This is a copy from the [original Notion page](https://www.notion.so/neondatabase/Proposal-Pageserver-MVCC-S3-Storage-8a424c0c7ec5459e89d3e3f00e87657c?pvs=4), taken on 2023-08-16.
+
+This is for archival mostly.
+The RFC that we're likely to go with is https://github.com/neondatabase/neon/pull/4919.
+
+---
+
+# Proposal: Pageserver MVCC S3 Storage
+
+tl;dr: this proposal enables Control Plane to attach a tenant to a new pageserver without being 100% certain that it has been detached from the old pageserver. This enables us to automate failover if a pageserver dies (no human in the loop).
+
+# Problem Statement
+
+The current Neon architecture requires the Control Plane to guarantee that a tenant is only attached to one pageserver at a time. If a tenant is attached to multiple pageservers simultaneously, the pageservers will overwrite each other’s changes in S3 for that tenant, resulting in data loss for that tenant.
+
+The above imposes limitations on tenant relocation and future designs for high availability. For instance, Control Plane cannot relocate a tenant to another pageserver before it is 100% certain that the tenant is detached from the source pageserver. If the source pageserver is unresponsive, the tenant detach procedure cannot proceed, and Control Plane has no choice but to wait for either the source to become responsive again, or rely on a node failure detection mechanism to detect that the source pageserver is dead, and give permission to skip the detachment step. Either way, the tenant is unavailable for an extended period, and we have no means to improve it in the current architecture.
+
+Note that there is no 100% correct node failure detection mechanism, and even techniques to accelerate failure detection, such as ********************************shoot-the-other-node-in-the-head,******************************** have their limits. So, we currently rely on humans as node failure detectors: they get alerted via PagerDuty, assess the situation under high stress, and make the decision. If they make the wrong call, or the apparent dead pageserver somehow resurrects later, we’ll have data loss.
+
+Also, by relying on humans, we’re [incurring needless unscalable toil](https://sre.google/sre-book/eliminating-toil/): as Neon grows, pageserver failures will become more and more frequent because our fleet grows. Each instance will need quick response time to minimize downtime for the affected tenants, which implies higher toil, higher resulting attrition, and/or higher personnel cost.
+
+Lastly, there are foreseeable needs by operation and product such as zero-downtime relocation and automatic failover/HA. For such features, the ability to have a tenant purposefully or accidentally attached to more than one pageserver will greatly reduce risk of data loss, and improve availability.
+
+# High-Level Idea
+
+The core idea is to evolve the per-Tenant S3 state to an MVCC-like scheme, allowing multiple pageservers to operate on the same tenant S3 state without interference. To make changes to S3, pageservers acquire long-running transactions from Control Plane. After opening a transaction, Pageservers make PUTs directly against S3, but they keys include the transaction ID,  so overwrites never happen. Periodically, pageservers talk back to Control Plane to commit their transaction. This is where Control Plane enforces strict linearizability, favoring availability over work-conservation: commit is only granted if no transaction started after the one that’s requesting commit. Garbage collection is done through deadlists, and it’s simplified tremendously by above commit grant/reject policy.
+
+Minimal changes are required for safekeepers to allow WAL for a single timeline be consumed by more than one pageserver without premature truncation.
+
+**Above scheme makes it safe to attach tenants without a 100% correct node failure detection mechanism. Further, it makes it safe to interleave tenant-attachment to pageservers, unlocking new capabilities for (internal) product features:**
+
+- **Fast, Zero-Toil Failover on Network Partitions or Instance Failure**: if a pageserver is not reachable (network partition, hardware failure, overload) we want to spread its attached tenants to new pageservers to restore availability, within the range of *seconds*. We cannot afford gracious timeouts to maximize the probability that the unreachable pageserver has ceased writing to S3. This proposal enables us to attach the tenants to the replacement pageservers,  and redirect their computes, without having to wait for confirmation that the unreachable pageserver has ceased writing to S3.
+- **************************************Zero-Downtime Relocation:************************************** we want to be able to relocate tenants to different pageservers with minimized availability or a latency impact. This proposal enables us to attach the relocating Tenant to the destination Pageserver before detaching it from the source Pageserver. This can help minimize downtime because we can wait for the destination to catch up on WAL processing before redirecting Computes.
+
+# Design
+
+The core idea is to evolve the per-Tenant S3 state to a per-tenant MVCC-like scheme.
+
+To make S3 changes for a given tenant, Pageserver requests a transaction ID from control plane for that tenant. Without a transaction ID, Pageserver does not write to S3.
+
+Once Pageserver received a transaction ID it is allowed to produce new objects and overwrite objects created in this transaction. Pageserver is not allowed to delete any objects; instead, it marks the object as deleted by appending the key to the transaction’s deadlist for later deletion. Commits of transactions are serialized through Control Plane: when Pageserver wants to commit a transaction, it sends an RPC to Control Plane. Control Plane responds with a commit grant or commit reject message. Commit grant means that the transaction’s changes are now visible to subsequent transactions. Commit reject means that the transaction’s changes are not and never will be visible to another Pageserver instance, and the rejected Pageserver is to cease further activity on that tenant.
+
+## ****************************************************Commit grant/reject policy****************************************************
+
+For the purposes of Pageserver, we want **linearizability** of a tenant’s S3 state. Since our transactions are scoped per tenant, it is sufficient for linearizability to grant commit if and only if no other transaction has been started since the commit-requesting transaction started.
+
+For example, consider the case of a single tenant, attached to Pageserver A. Pageserver A has an open transaction but becomes unresponsive. Control Plane decides to relocate the tenant to another Pageserver B. It need *not* wait for A to be 100%-certainly down before B can start uploading to S3 for that tenant. Instead, B can start a new transaction right away, make progress, and get commit grants;  What about A? The transaction is RejectPending in Control Plane until A eventually becomes responsive again, tries to commit, gets a rejection, acknowledges it, and thus its transaction becomes RejectAcknowledge. If A is definitively dead, operator can also force-transition from state RejectPending to RejectAcknowledged. But critically, Control Plane doesn’t have for A’s transaction to become RejectAcknowledge before attaching the tenant to B.
+
+```mermaid
+sequenceDiagram
+
+   participant CP
+   participant A
+   participant S3
+   participant B
+
+	 CP -->> A: attach tenant
+   activate A
+	 A -->> CP: start txn
+	 CP -->> A: txn=23, last_committed_txn=22
+
+
+	 Note over CP,A: network partition
+	 CP --x A: heartbeat
+	 CP --x A: heartbeat
+
+	 Note over CP: relocate tenant to avoid downtime
+	 CP -->> B: attach tenant
+   activate B
+	 B -->> CP: start txn
+   Note over CP: mark A's txn 23 as RejectPending
+	 CP -->> B: txn=24, last-committed txn is 22
+	 B -->> S3: PUT X.layer.24<br>PUT index_part.json.24 referencing X.layer.24
+	 B -->> CP: request commit
+	 CP -->> B: granted
+   B -->> CP: start txn
+  CP -->> B: txn=25, last_committed_txn=22
+
+   A -->> S3: PUT Y.layer.23 <br> PUT index_part.json.23 referencing Y.layer.23
+  A --x CP: request commit
+	 A --x CP: request commit
+
+   Note over CP,A: partition is over
+
+   A -->> CP: request commit
+
+   Note over CP: most recently started txn is 25, not 23, reject
+
+   CP -->> A: reject
+   A -->> CP: acknowledge reject
+
+   Note over CP: mark A's txn 23 as RejectAcknowledged
+
+  deactivate A
+
+  B -->> S3: PUT 000-FFF_X-Y.layer.**************25**************<br>...
+
+  deactivate B
+
+
+```
+
+If a Pageserver gets a rejection to a commit request, it acknowledges rejection and cedes further S3 uploads for the tenant, until it receives a `/detach` request for the tenant (control plane has most likely attached the tenant to another pageserver in the meantime).
+
+In practice, Control Plane will probably extend the commit grant/reject schema above, taking into account the pageserver to which it last attached the tenant. In the above example, Control Plane could remember that the pageserver that is supposed to host the tenant is pageserver B, and reject start-txn and commit requests from pageserver A. It would also use such requests from A as a signal that A is reachable again, and retry the `/detach` .
+
+<aside>
+💡 A commit failure causes the tenant to become effectively `Broken`. Pageserver should persist this locally so it doesn’t bother ControlPlane for a new txn when Pageserver is restarted.
+
+</aside>
+
+## ********************Visibility********************
+
+We mentioned earlier that once a transaction commits, its changes are visible to subsequent transactions. But how does a given transaction know where to look for the data? There is no longer a single `index_part.json` per timeline, or a single `timelines/:timeline_id` prefix to look for; they’re all multi-versioned, suffixed by the txn number.
+The solution is: at transaction start, Pageserver receives the last-committed transaction ID from Control Plane (`last_committed_txn` in the diagram). last_commited_txn is the upper bound for what is visible for the current transaction. Control Plane keeps track of each open transaction’s last_committed_txn for purposes of garbage collection (see later paragraph).
+Equipped with last_committed_txn, Pageserver then discovers
+
+- the current index part of a timeline at `tenants/:tenant_id/timelines/:timeline_id/index_part.json.$last_committed_txn`. The `index_part.json.$last_committed_txn` has the exact same contents as the current architecture’s index_part.json, i.e. full list of layers.
+- the list of existent timelines as part of the `attach` RPC from CP;
+
+There is no other S3 state per tenant, so, that’s all the visibility required.
+An alternative to receiving the list of existent timelines from CP is to introduce a proper **********SetOfTimelines********** object in S3, and multi-version it just like above. For example, we could have a `tenants/:tenant_id/timelines.json.$txn` file that references `index_part.json.$last_committed_txn` . It can be added later if more separation between CP and PS is desired.
+
+So, the only MVCC’ed object types in this proposal are LayerFile and IndexPart (=individual timeline), but not the SetOfTimelines in a given tenant. Is this a problem? For example, the Pageserver’s garbage collection code needs to know the full set of timelines of a tenant. Otherwise it’ll make incorrect decisions. What if Pageserver A knows about timelines {R,S}, but another Pageserver B created an additional branch T, so, its set of timelines is {R,S,T}. Both pageservers will run GC code, and so, PS A may decide to delete a layer that’s still needed for branch T. Not a problem with this propsoal, because the effect of GC (i.e., layer deletion) is properly MVCC’ed.
+
+## Longevity Of Transactions & Availability
+
+Pageserver depends on Control Plane to start a new transaction. If ControlPlane is down, no new transactions can be started.
+
+Pageservers commit transactions based on a maximum amount of uncommitted changes that have accumulated in S3. A lower maximum increases dependence and load on ControlPlane which decreases availability. A higher maximum risks losing more work in the event of failover; the work will have to be re-done in a new transaction on the new node.
+
+Pageservers are persist the open txn id in local storage, so that they can resume the transaction after restart, without dependence on Control Plane.
+
+## **Operations**
+
+********PUTs:********
+
+- **layer files**
+    - current architecture: layer files are supposed to be write-once, but actually, there are edge-cases where we PUT the same layer file name twice; namely if we PUT the file to S3 but crash before uploading the index part that references it; then detach + attach, and re-run compaction, which is non-deterministic.
+    - this proposal: with transactions, we can now upload layers and index_part.json concurrently, just need to make sure layer file upload is done before we request txn commit.
+- **index part** upload: `index_part.json.$txn` may be created and subsequently overwritten multiple times in a transaction; it is an availability/work-loss trade-off how often to request a commit from CP.
+
+**************DELETEs**************: for deletion, we maintain a deadlist per transaction. It is located at `tenants/:tenant_id/deadlist/deadlist.json.$txn`. It is PUT once before the pageserver requests requests commit, and not changed after sending request to commit. An object created in the current txn need not (but can) be on the deadlist — it can be DELETEd immediately because it’s not visible to other transactions. An example use case would be an L0 layer that gets compacted within one transaction; or, if we ever start MVCC’ing the set of timelines of a tenant, a short-lived branch that is created & destroyed within one transaction.
+
+<aside>
+☝ **Deadlist Invariant:** if a an object is on a deadlist of transaction T, it is not referenced from anywhere else in the full state visible to T or any later started transaction > T.
+
+</aside>
+
+### Rationale For Deadlist.json
+
+Given that this proposal only MVCC’s layers and indexparts, one may ask why the deadlist isn’t part of indexpart. The reason is to not lose generality: the deadlist is just a list of keys; it is not necessary to understand the data format of the versioned object to process the deadlist. This is important for garbage collection / vacuuming, which we’ll come to in the next section.
+
+## Garbage Collection / Vacuuming
+
+After a transaction has reached reject-acknowledged state,  Control Plane initiates a garbage collection procedure for the aborted transaction.
+
+Control Plane is in the unique position about transaction states. Here is a sketch of the exact transaction states and what Control Plane keeps track of.
+
+```
+struct Tenant {
+  ...
+
+  txns: HashMap<TxnId, Transaction>,
+  // the most recently started txn's id; only most recently sarted can win
+  next_winner_txn: Option<TxnId>,
+}
+struct Transaction {
+  id: TxnId, // immutable
+  last_committed_txn: TxnId, // immutable; the most recent txn in state `Committed`
+                             // when self was started
+  pageserver_id: PageserverId,
+  state: enum {
+    Open,
+    Committed,
+    RejectPending,
+    RejectAcknowledged, // invariant: we know all S3 activity has ceded
+    GarbageCollected,
+  }
+}
+```
+
+Object creations & deletions by a rejected transaction have never been visible to other transactions. That is true for both RejectPending and RejectAcknowledged states. The difference is that, in RejectPending, the pageserver may still be uploading to S3, whereas in RejectAcknowledged, Control Plane can be certain that all S3 activity in the name of that transaction has ceded. So, once a transaction reaches state RejectAcknowledged state, it is safe to DELETE all objects created by that transaction, and discard the transaction’s deadlists.
+
+A transaction T in state Committed has subsequent transactions that may or may not reference the objects it created. None of the subsequent transaction can reference the objects on T’s deadlist, though, as per the Deadlist Invariant (see previous section).
+
+So, for garbage collection, we need to assess transactions in state Committed and RejectAcknowledged:
+
+- Commited: delete objects on the deadlist.
+    - We don’t need a LIST request here, the deadlist is sufficient. So, it’s really cheap.
+    - This is **not true MVCC garbage collection**; by deleting the objects on Committed transaction T ’s deadlist, we might delete data referenced by other transactions that were concurrent with T, i.e., they started while T was still open. However, the fact that T is committed means that the other transactions are RejectPending or RejectAcknowledged, so, they don’t matter. Pageservers executing these doomed RejectPending transactions must handle 404 for GETs gracefully, e.g., by trying to commit txn so they observe the rejection they’re destined to get anyways. 404’s for RejectAcknowledged is handled below.
+- RejectAcknowledged: delete all objects created in that txn, and discard deadlists.
+    - 404s / object-already-deleted type messages must be expected because of Committed garbage collection (see above)
+    - How to get this list of objects created in a txn? Open but solvable design question; Ideas:
+        - **Brute force**: within tenant prefix, search for all keys ending in `.$txn` and delete them.
+        - **WAL for PUTs**: before a txn PUTs an object, it logs to S3, or some other equivalently durable storage, that it’s going to do it. If we log to S3, this means we have to do an additional WAL PUT per “readl” PUT.
+        - ******************************LIST with reorg’ed S3 layout (preferred one right now):****************************** layout S3 key space such that `$txn` comes first, i.e., `tenants/:tenant_id/$txn/timelines/:timeline_id/*.json.$txn` . That way, when we need to GC a RejectAcknowledged txn, we just LIST the entire `tenants/:tenant_id/$txn` prefix and delete it. The cost of GC for RejectAcknowledged transactions is thus proportional to the number of objects created in that transaction.
+
+## Branches
+
+This proposal only MVCC’s layer files and and index_part.json, but leaves the tenant object not-MVCCed. We argued earlier that it’s fine to ignore this for now, because
+
+1. Control Plane can act as source-of-truth for the set of timelines, and
+2. The only operation that makes decision based on “set of timelines” is GC, which in turn only does layer deletions, and layer deletions ***are*** properly MVCC’ed.
+
+Now that we’ve introduced garbage collection, let’s elaborate a little more on (2). Recall our example from earlier: Pageserver A knows about timelines {R,S}, but another Pageserver B created an additional branch T, so, its set of timelines is {R,S,T}. Both pageservers will run GC code, and so, PS A may decide to delete a layer that’s still needed for branch T.
+
+How does the MVCC’ing of layer files protect us here? If A decides to delete that layer, it’s just on A’s transaction’s deadlist, but still present in S3 and usable by B. If A commits first, B won’t be able to commit and the layers in timeline T will be vacuumed. If B commits first, A’s deadlist is discarded and the layer continues to exist.
+
+## Safekeeper Changes
+
+We need to teach the safekeepers that there can be multiple pageservers requesting WAL for the same timeline, in order to prevent premature WAL truncation.
+
+In the current architecture, the Safekeeper service currently assumes only one Pageserver and is allowed to prune WAL older than that Pageserver’s `remote_consistent_lsn`. Safekeeper currently learns the `remote_consistent_lsn` through the walreceiver protocol.
+
+So, if we have a tenant attached to two pageservers at the same time, they will both try to stream WAL and the Safekeeper will get confused about which connection’s `remote_consistent_lsn` to use as a basis for WAL pruning.
+
+What do we need to change to make it work? We need to make sure that the Safekeepers only prune WAL up to the `remote_consistent_lsn` of the last-committed transaction.
+
+The straight-forward way to get it is to re-design WAL pruning as follows:
+
+1. Pageserver reports remote_consistent_lsn as part of transaction commit to Control Plane.
+2. Control Plane makes sure transaction state update is persisted.
+3. Control Plane (asynchronous to transaction commit) reconciles with Safekeepers to ensure WAL pruning happens.
+
+The above requires non-trivial changes, but, in the light of other planned projects such as restore-tenant-from-safekeeper-wal-backups, I think Control Plane will need to get involved in WAL pruning anyways.
+
+# How This Proposal Unlocks Future Features
+
+Let us revisit the example from the introduction where we were thinking about handling network partitions. Network partitions need to be solved first, because they’re unavoidable in distributed systems. We did that. Now let’s see how we can solve actual product problems:
+
+## **Fast, Zero-Toil Failover on Network Partitions or Instance Failure**
+
+The “Problem Statement” section outlined the current architecture’s problems with regards to network partitions or instance failure: it requires a 100% correct node-dead detector to make decisions, which doesn’t exist in reality. We rely instead on human toil: an oncall engineer has to inspect the situation and make a decision, which may be incorrect and in any case take time in the order of minutes, which means equivalent downtime for users.
+
+With this proposal, automatic failover for pageservers is trivial:
+
+If a pageserver is unresponsive from Control Plane’s / Compute’s perspective, Control Plane does the following:
+
+- attach all tenants of the unresponsive pageserver to new pageservers
+- switch over these tenants’ computes immediately;
+
+At this point, availability is restored and user pain relieved.
+
+What’s left is to somehow close the doomed transaction of the unresponsive pageserver, so that it beomes RejectAcknowledged, and GC can make progress. Since S3 is cheap, we can afford to wait a really long time here, especially if we put a soft bound on the amount of data a transaction may produce before it must commit. Procedure:
+
+1. Ensure the unresponsive pageserver is taken out of rotation for new attachments. That probably should happen as part of the routine above.
+2. Make a human operator investigate decide what to do (next morning, NO ONCALL ALERT):
+    1. Inspect the instance, investigate logs, understand root cause.
+    2. Try to re-establish connectivity between pageserver and Control Plane so that pageserver can retry commits, get rejected, ack rejection ⇒ enable GC.
+    3. Use below procedure to decomission pageserver.
+
+### Decomissioning A Pageserver (Dead or Alive-but-Unrespsonive)
+
+The solution, enabled by this proposal:
+
+1. Ensure that pageserver’s S3 credentials are revoked so that it cannot make new uploads, which wouldn’t be tracked anywhere.
+2. Let enough time pass for the S3 credential revocation to propagate. Amazon doesn’t give a guarantee here. As stated earlier, we can easily afford to wait here.
+3. Mark all Open and RejectPending transactions of that pageserver as RejectAcknowledge.
+
+Revocation of the S3 credentials is required so that, once we transition all the transactions of that pageserver to RejectAcknowledge, once garbage-collection pass is guaranteed to delete all objects that will ever exist for that pageserver. That way, we need not check *****GarbageCollected***** transactions every again.
+
+## Workflow: Zero-Downtime Relocation
+
+With zero-downtime relocation, the goal is to have the target pageserver warmed up, i.e., at the same `last_record_lsn` as the source pageserver, before switching over Computes from source to target pageserver.
+
+With this proposal, it works like so:
+
+1. Grant source pageserver its last open transaction. This one is doomed to be rejected later, unless the relocation fails.
+2. Grant target pageserver its first open transaction.
+3. Have target pageserver catch up on WAL, streaming from last-committed-txn’s remote_consistent_lsn onwards.
+4. Once target pageserver reports `last_record_lsn` close enough to source pageserver, target pageserver requests commit.
+5. Drain compute traffic from source to target pageserver. (Source can still answer requests until it tries to commit and gets reject, so, this will be quite smooth).
+
+Note that as soon as we complete step (4), the source pageserver’s transaction is doomed to be rejected later. Conversely, if the target can’t catch up fast enough, the source will make a transaction commit earlier. This will generally happen if there is a lot of write traffic coming in. The design space to make thing smooth here is large, but well explored in other areas of computing, e.g., VM live migration. We have all the important policy levers at hand, e.g.,
+
+- delaying source commits if we see target making progress
+- slowing down source consumption (need some signalling mechanism for it)
+- slowing down compute wal generation
+- …
+
+It doesn’t really matter, what’s important is that two pageservers can overlap.
+
+# Additional Trade-Offs / Remarks Brought Up During Peer Review
+
+This proposal was read by and discussed @Stas and @Dmitry Rodionov prior to publishing it with the broader team. (This does not mean they endorse this proposal!).
+
+Issues that we discussed:
+
+1. **Frequency of transactions:** If even idle tenants commit every 10min or so, that’s quite a lot of load on Control Plane. Can we minimize it by Equating Transaction Commit Period to Attachment Period? I.e. start txn on attach, commit on detach?
+    1. Would be nice, but, if a tenant is attached for 1 month, then PS dies, we lose 1 month of work.
+    2. ⇒ my solution to this problem: Adjusted this proposal to make transaction commit frequency proportional to amount of uncommitted data.
+        1. It’s ok to spend resources on active users, they pay us money to do it!
+        2. The amount of work per transaction is minimal.
+            1. In current Control Plane, it’s a small database transaction that is super unlikely to conflict with other transactions.
+            2. I have very little concerns about scalability of the commit workload on CP side because it's trivially horizontally scalable by sharding by tenant.
+        3. There's no super stringent availability requirement on control plane; if a txn can't commit because it can't reach the CP, PS can continue & retry in the background, speculating that it's CP downtime and not PS-partitioned-off scenario.
+        4. Without stringent availability requirement, there's flexibility for future changes to CP-side-implementation.
+2. ************************************************Does this proposal address mirroring / no-performance-degradation failover ?************************************************
+    1. No it doesn’t. It only provides the building block for attaching a tenant to a new pageserver without having to worry that the tenant is detached on the old pageserver.
+    2. A simple scheme to build no-performance-degradation failover on top of this proposal is to have an asynchronous read-only replica of a tenant on another pageserver in the same region.
+    3. Another more ambitious scheme to get no-performance-degradation would be [One-Pager: Layer File Spreading (Christian)](https://www.notion.so/One-Pager-Layer-File-Spreading-Christian-eb6b64182a214e11b3fceceee688d843?pvs=21); this proposal would be used in layer file spreading for risk-free automation of TenantLeader failover, which hasn’t been addressed Ithere.
+    4. In any way, failover would restart from an older S3 state, and need to re-ingest WAL before being able to server recently written pages.
+        1. Is that a show-stopper? I think not.
+        2. Is it suboptimal? Absolutely: if a pageserver instance fails, all its tenants will be distributed among the remaining pageservers (OK), and all these tenants will ask the safekeepers for WAL at the same time (BAD). So, pageserver instance failure will cause a load spike in safekeepers.
+            1. Personally I think that’s an OK trade-off to make.
+            2. There are countless options to avoid / mitigate the load spike. E.g., pro-actively streaming WAL to the standby read-only replica.
+
+3. ********************************************Does this proposal allow multiple writers for a tenant?********************************************
+    1. In abstract terms, this proposal provides a linearized history for a given S3 prefix.
+    2. In concrete terms, this proposal provides a linearized history per tenant.
+    3. There can be multiple writers at a given time, but only one of them will win to become part of the linearized history.
+4. ************************************************************************************Alternative ideas mentioned during meetings that should be turned into a written prospoal like this one:************************************************************************************
+    1. @Dmitry Rodionov : having linearized storage of index_part.json in some database that allows serializable transactions / atomic compare-and-swap PUT
+    2. @Dmitry Rodionov :
+    3. @Stas : something like this scheme, but somehow find a way to equate attachment duration with transaction duration, without losing work if pageserver dies months after attachment.

libs/utils/Cargo.toml

@@ -26,6 +26,7 @@ serde_json.workspace = true
 signal-hook.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
+tokio-util.workspace = true
 tracing.workspace = true
 tracing-error.workspace = true
 tracing-subscriber = { workspace = true, features = ["json", "registry"] }

libs/utils/src/backoff.rs

@@ -1,18 +1,31 @@
 use std::fmt::{Debug, Display};
 
 use futures::Future;
+use tokio_util::sync::CancellationToken;
 
 pub const DEFAULT_BASE_BACKOFF_SECONDS: f64 = 0.1;
 pub const DEFAULT_MAX_BACKOFF_SECONDS: f64 = 3.0;
 
-pub async fn exponential_backoff(n: u32, base_increment: f64, max_seconds: f64) {
+pub async fn exponential_backoff(
+    n: u32,
+    base_increment: f64,
+    max_seconds: f64,
+    cancel: &CancellationToken,
+) {
     let backoff_duration_seconds =
         exponential_backoff_duration_seconds(n, base_increment, max_seconds);
     if backoff_duration_seconds > 0.0 {
         tracing::info!(
             "Backoff: waiting {backoff_duration_seconds} seconds before processing with the task",
         );
-        tokio::time::sleep(std::time::Duration::from_secs_f64(backoff_duration_seconds)).await;
+
+        drop(
+            tokio::time::timeout(
+                std::time::Duration::from_secs_f64(backoff_duration_seconds),
+                cancel.cancelled(),
+            )
+            .await,
+        )
     }
 }
 
@@ -24,28 +37,57 @@ pub fn exponential_backoff_duration_seconds(n: u32, base_increment: f64, max_sec
     }
 }
 
+/// Configure cancellation for a retried operation: when to cancel (the token), and
+/// what kind of error to return on cancellation
+pub struct Cancel<E, CF>
+where
+    E: Display + Debug + 'static,
+    CF: Fn() -> E,
+{
+    token: CancellationToken,
+    on_cancel: CF,
+}
+
+impl<E, CF> Cancel<E, CF>
+where
+    E: Display + Debug + 'static,
+    CF: Fn() -> E,
+{
+    pub fn new(token: CancellationToken, on_cancel: CF) -> Self {
+        Self { token, on_cancel }
+    }
+}
+
 /// retries passed operation until one of the following conditions are met:
 /// Encountered error is considered as permanent (non-retryable)
 /// Retries have been exhausted.
 /// `is_permanent` closure should be used to provide distinction between permanent/non-permanent errors
 /// When attempts cross `warn_threshold` function starts to emit log warnings.
 /// `description` argument is added to log messages. Its value should identify the `op` is doing
-pub async fn retry<T, O, F, E>(
+/// `cancel` argument is required: any time we are looping on retry, we should be using a CancellationToken
+/// to drop out promptly on shutdown.
+pub async fn retry<T, O, F, E, CF>(
     mut op: O,
     is_permanent: impl Fn(&E) -> bool,
     warn_threshold: u32,
     max_retries: u32,
     description: &str,
+    cancel: Cancel<E, CF>,
 ) -> Result<T, E>
 where
     // Not std::error::Error because anyhow::Error doesnt implement it.
     // For context see https://github.com/dtolnay/anyhow/issues/63
-    E: Display + Debug,
+    E: Display + Debug + 'static,
     O: FnMut() -> F,
     F: Future<Output = Result<T, E>>,
+    CF: Fn() -> E,
 {
     let mut attempts = 0;
     loop {
+        if cancel.token.is_cancelled() {
+            return Err((cancel.on_cancel)());
+        }
+
         let result = op().await;
         match result {
             Ok(_) => {
@@ -80,6 +122,7 @@ where
             attempts,
             DEFAULT_BASE_BACKOFF_SECONDS,
             DEFAULT_MAX_BACKOFF_SECONDS,
+            &cancel.token,
         )
         .await;
         attempts += 1;
@@ -132,6 +175,7 @@ mod tests {
             1,
             1,
             "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
         )
         .await;
 
@@ -157,6 +201,7 @@ mod tests {
             2,
             2,
             "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
         )
         .await
         .unwrap();
@@ -179,6 +224,7 @@ mod tests {
             2,
             2,
             "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
         )
         .await
         .unwrap_err();

libs/utils/src/lib.rs

@@ -68,44 +68,6 @@ pub mod completion;
 /// Reporting utilities
 pub mod error;
 
-mod failpoint_macro_helpers {
-
-    /// use with fail::cfg("$name", "return(2000)")
-    ///
-    /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the
-    /// specified time (in milliseconds). The main difference is that we use async
-    /// tokio sleep function. Another difference is that we print lines to the log,
-    /// which can be useful in tests to check that the failpoint was hit.
-    #[macro_export]
-    macro_rules! failpoint_sleep_millis_async {
-        ($name:literal) => {{
-            // If the failpoint is used with a "return" action, set should_sleep to the
-            // returned value (as string). Otherwise it's set to None.
-            let should_sleep = (|| {
-                ::fail::fail_point!($name, |x| x);
-                ::std::option::Option::None
-            })();
-
-            // Sleep if the action was a returned value
-            if let ::std::option::Option::Some(duration_str) = should_sleep {
-                $crate::failpoint_sleep_helper($name, duration_str).await
-            }
-        }};
-    }
-
-    // Helper function used by the macro. (A function has nicer scoping so we
-    // don't need to decorate everything with "::")
-    pub async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
-        let millis = duration_str.parse::<u64>().unwrap();
-        let d = std::time::Duration::from_millis(millis);
-
-        tracing::info!("failpoint {:?}: sleeping for {:?}", name, d);
-        tokio::time::sleep(d).await;
-        tracing::info!("failpoint {:?}: sleep done", name);
-    }
-}
-pub use failpoint_macro_helpers::failpoint_sleep_helper;
-
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:

libs/vm_monitor/Cargo.toml

@@ -0,0 +1,31 @@
+[package]
+name = "vm_monitor"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[[bin]]
+name = "vm-monitor"
+path = "./src/bin/monitor.rs"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow.workspace = true
+axum.workspace = true
+clap.workspace = true
+futures.workspace = true
+inotify.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+sysinfo.workspace = true
+tokio.workspace = true
+tokio-postgres.workspace = true
+tokio-stream.workspace = true
+tokio-util.workspace = true
+tracing.workspace = true
+tracing-subscriber.workspace = true
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+
+[target.'cfg(target_os = "linux")'.dependencies]
+cgroups-rs = "0.3.3"

libs/vm_monitor/README.md

@@ -0,0 +1,18 @@
+# `vm-monitor`
+
+The `vm-monitor` (or just monitor) is a core component of the autoscaling system,
+along with the `autoscale-scheduler` and the `autoscaler-agent`s. The monitor has
+two primary roles: 1) notifying agents when immediate upscaling is necessary due
+to memory conditions and 2) managing Postgres' file cache and a cgroup to carry
+out upscaling and downscaling decisions.
+
+## More on scaling
+
+We scale CPU and memory using NeonVM, our in-house QEMU tool for use with Kubernetes.
+To control thresholds for receiving memory usage notifications, we start Postgres
+in the `neon-postgres` cgroup and set its `memory.{max,high}`.
+
+* See also: [`neondatabase/autoscaling`](https://github.com/neondatabase/autoscaling/)
+* See also: [`neondatabase/vm-monitor`](https://github.com/neondatabase/vm-monitor/),
+where initial development of the monitor happened. The repository is no longer
+maintained but the commit history may be useful for debugging.

libs/vm_monitor/src/bin/monitor.rs

@@ -0,0 +1,33 @@
+// We expose a standalone binary _and_ start the monitor in `compute_ctl` so that
+// we can test the monitor as part of the entire autoscaling system in
+// neondatabase/autoscaling.
+//
+// The monitor was previously started by vm-builder, and for testing purposes,
+// we can mimic that setup with this binary.
+
+#[cfg(target_os = "linux")]
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    use clap::Parser;
+    use tokio_util::sync::CancellationToken;
+    use tracing_subscriber::EnvFilter;
+    use vm_monitor::Args;
+
+    let subscriber = tracing_subscriber::fmt::Subscriber::builder()
+        .json()
+        .with_file(true)
+        .with_line_number(true)
+        .with_span_list(true)
+        .with_env_filter(EnvFilter::from_default_env())
+        .finish();
+    tracing::subscriber::set_global_default(subscriber)?;
+
+    let args: &'static Args = Box::leak(Box::new(Args::parse()));
+    let token = CancellationToken::new();
+    vm_monitor::start(args, token).await
+}
+
+#[cfg(not(target_os = "linux"))]
+fn main() {
+    panic!("the monitor requires cgroups, which are only available on linux")
+}

libs/vm_monitor/src/cgroup.rs

@@ -0,0 +1,693 @@
+use std::{
+    fmt::{Debug, Display},
+    fs,
+    pin::pin,
+    sync::atomic::{AtomicU64, Ordering},
+};
+
+use anyhow::{anyhow, bail, Context};
+use cgroups_rs::{
+    freezer::FreezerController,
+    hierarchies::{self, is_cgroup2_unified_mode, UNIFIED_MOUNTPOINT},
+    memory::MemController,
+    MaxValue,
+    Subsystem::{Freezer, Mem},
+};
+use inotify::{EventStream, Inotify, WatchMask};
+use tokio::sync::mpsc::{self, error::TryRecvError};
+use tokio::time::{Duration, Instant};
+use tokio_stream::{Stream, StreamExt};
+use tracing::{info, warn};
+
+use crate::protocol::Resources;
+use crate::MiB;
+
+/// Monotonically increasing counter of the number of memory.high events
+/// the cgroup has experienced.
+///
+/// We use this to determine if a modification to the `memory.events` file actually
+/// changed the `high` field. If not, we don't care about the change. When we
+/// read the file, we check the `high` field in the file against `MEMORY_EVENT_COUNT`
+/// to see if it changed since last time.
+pub static MEMORY_EVENT_COUNT: AtomicU64 = AtomicU64::new(0);
+
+/// Monotonically increasing counter that gives each cgroup event a unique id.
+///
+/// This allows us to answer questions like "did this upscale arrive before this
+/// memory.high?". This static is also used by the `Sequenced` type to "tag" values
+/// with a sequence number. As such, prefer to used the `Sequenced` type rather
+/// than this static directly.
+static EVENT_SEQUENCE_NUMBER: AtomicU64 = AtomicU64::new(0);
+
+/// A memory event type reported in memory.events.
+#[derive(Debug, Eq, PartialEq, Copy, Clone)]
+pub enum MemoryEvent {
+    Low,
+    High,
+    Max,
+    Oom,
+    OomKill,
+    OomGroupKill,
+}
+
+impl MemoryEvent {
+    fn as_str(&self) -> &str {
+        match self {
+            MemoryEvent::Low => "low",
+            MemoryEvent::High => "high",
+            MemoryEvent::Max => "max",
+            MemoryEvent::Oom => "oom",
+            MemoryEvent::OomKill => "oom_kill",
+            MemoryEvent::OomGroupKill => "oom_group_kill",
+        }
+    }
+}
+
+impl Display for MemoryEvent {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.as_str())
+    }
+}
+
+/// Configuration for a `CgroupWatcher`
+#[derive(Debug, Clone)]
+pub struct Config {
+    // The target difference between the total memory reserved for the cgroup
+    // and the value of the cgroup's memory.high.
+    //
+    // In other words, memory.high + oom_buffer_bytes will equal the total memory that the cgroup may
+    // use (equal to system memory, minus whatever's taken out for the file cache).
+    oom_buffer_bytes: u64,
+
+    // The amount of memory, in bytes, below a proposed new value for
+    // memory.high that the cgroup's memory usage must be for us to downscale
+    //
+    // In other words, we can downscale only when:
+    //
+    //   memory.current + memory_high_buffer_bytes < (proposed) memory.high
+    //
+    // TODO: there's some minor issues with this approach -- in particular, that we might have
+    // memory in use by the kernel's page cache that we're actually ok with getting rid of.
+    pub(crate) memory_high_buffer_bytes: u64,
+
+    // The maximum duration, in milliseconds, that we're allowed to pause
+    // the cgroup for while waiting for the autoscaler-agent to upscale us
+    max_upscale_wait: Duration,
+
+    // The required minimum time, in milliseconds, that we must wait before re-freezing
+    // the cgroup while waiting for the autoscaler-agent to upscale us.
+    do_not_freeze_more_often_than: Duration,
+
+    // The amount of memory, in bytes, that we should periodically increase memory.high
+    // by while waiting for the autoscaler-agent to upscale us.
+    //
+    // This exists to avoid the excessive throttling that happens when a cgroup is above its
+    // memory.high for too long. See more here:
+    // https://github.com/neondatabase/autoscaling/issues/44#issuecomment-1522487217
+    memory_high_increase_by_bytes: u64,
+
+    // The period, in milliseconds, at which we should repeatedly increase the value
+    // of the cgroup's memory.high while we're waiting on upscaling and memory.high
+    // is still being hit.
+    //
+    // Technically speaking, this actually serves as a rate limit to moderate responding to
+    // memory.high events, but these are roughly equivalent if the process is still allocating
+    // memory.
+    memory_high_increase_every: Duration,
+}
+
+impl Config {
+    /// Calculate the new value for the cgroups memory.high based on system memory
+    pub fn calculate_memory_high_value(&self, total_system_mem: u64) -> u64 {
+        total_system_mem.saturating_sub(self.oom_buffer_bytes)
+    }
+}
+
+impl Default for Config {
+    fn default() -> Self {
+        Self {
+            oom_buffer_bytes: 100 * MiB,
+            memory_high_buffer_bytes: 100 * MiB,
+            // while waiting for upscale, don't freeze for more than 20ms every 1s
+            max_upscale_wait: Duration::from_millis(20),
+            do_not_freeze_more_often_than: Duration::from_millis(1000),
+            // while waiting for upscale, increase memory.high by 10MiB every 25ms
+            memory_high_increase_by_bytes: 10 * MiB,
+            memory_high_increase_every: Duration::from_millis(25),
+        }
+    }
+}
+
+/// Used to represent data that is associated with a certain point in time, such
+/// as an upscale request or memory.high event.
+///
+/// Internally, creating a `Sequenced` uses a static atomic counter to obtain
+/// a unique sequence number. Sequence numbers are monotonically increasing,
+/// allowing us to answer questions like "did this upscale happen after this
+/// memory.high event?" by comparing the sequence numbers of the two events.
+#[derive(Debug, Clone)]
+pub struct Sequenced<T> {
+    seqnum: u64,
+    data: T,
+}
+
+impl<T> Sequenced<T> {
+    pub fn new(data: T) -> Self {
+        Self {
+            seqnum: EVENT_SEQUENCE_NUMBER.fetch_add(1, Ordering::AcqRel),
+            data,
+        }
+    }
+}
+
+/// Responds to `MonitorEvents` to manage the cgroup: preventing it from being
+/// OOM killed or throttling.
+///
+/// The `CgroupWatcher` primarily achieves this by reading from a stream of
+/// `MonitorEvent`s. See `main_signals_loop` for details on how to keep the
+/// cgroup happy.
+#[derive(Debug)]
+pub struct CgroupWatcher {
+    pub config: Config,
+
+    /// The sequence number of the last upscale.
+    ///
+    /// If we receive a memory.high event that has a _lower_ sequence number than
+    /// `last_upscale_seqnum`, then we know it occured before the upscale, and we
+    /// can safely ignore it.
+    ///
+    /// Note: Like the `events` field, this doesn't _need_ interior mutability but we
+    /// use it anyways so that methods take `&self`, not `&mut self`.
+    last_upscale_seqnum: AtomicU64,
+
+    /// A channel on which we send messages to request upscale from the dispatcher.
+    upscale_requester: mpsc::Sender<()>,
+
+    /// The actual cgroup we are watching and managing.
+    cgroup: cgroups_rs::Cgroup,
+}
+
+/// Read memory.events for the desired event type.
+///
+/// `path` specifies the path to the desired `memory.events` file.
+/// For more info, see the `memory.events` section of the [kernel docs]
+/// <https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>
+fn get_event_count(path: &str, event: MemoryEvent) -> anyhow::Result<u64> {
+    let contents = fs::read_to_string(path)
+        .with_context(|| format!("failed to read memory.events from {path}"))?;
+
+    // Then contents of the file look like:
+    // low 42
+    // high 101
+    // ...
+    contents
+        .lines()
+        .filter_map(|s| s.split_once(' '))
+        .find(|(e, _)| *e == event.as_str())
+        .ok_or_else(|| anyhow!("failed to find entry for memory.{event} events in {path}"))
+        .and_then(|(_, count)| {
+            count
+                .parse::<u64>()
+                .with_context(|| format!("failed to parse memory.{event} as u64"))
+        })
+}
+
+/// Create an event stream that produces events whenever the file at the provided
+/// path is modified.
+fn create_file_watcher(path: &str) -> anyhow::Result<EventStream<[u8; 1024]>> {
+    info!("creating file watcher for {path}");
+    let inotify = Inotify::init().context("failed to initialize file watcher")?;
+    inotify
+        .watches()
+        .add(path, WatchMask::MODIFY)
+        .with_context(|| format!("failed to start watching {path}"))?;
+    inotify
+        // The inotify docs use [0u8; 1024] so we'll just copy them. We only need
+        // to store one event at a time - if the event gets written over, that's
+        // ok. We still see that there is an event. For more information, see:
+        // https://man7.org/linux/man-pages/man7/inotify.7.html
+        .into_event_stream([0u8; 1024])
+        .context("failed to start inotify event stream")
+}
+
+impl CgroupWatcher {
+    /// Create a new `CgroupWatcher`.
+    #[tracing::instrument(skip_all, fields(%name))]
+    pub fn new(
+        name: String,
+        // A channel on which to send upscale requests
+        upscale_requester: mpsc::Sender<()>,
+    ) -> anyhow::Result<(Self, impl Stream<Item = Sequenced<u64>>)> {
+        // TODO: clarify exactly why we need v2
+        // Make sure cgroups v2 (aka unified) are supported
+        if !is_cgroup2_unified_mode() {
+            anyhow::bail!("cgroups v2 not supported");
+        }
+        let cgroup = cgroups_rs::Cgroup::load(hierarchies::auto(), &name);
+
+        // Start monitoring the cgroup for memory events. In general, for
+        // cgroups v2 (aka unified), metrics are reported in files like
+        // > `/sys/fs/cgroup/{name}/{metric}`
+        // We are looking for `memory.high` events, which are stored in the
+        // file `memory.events`. For more info, see the `memory.events` section
+        // of https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
+        let path = format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name);
+        let memory_events = create_file_watcher(&path)
+            .with_context(|| format!("failed to create event watcher for {path}"))?
+            // This would be nice with with .inspect_err followed by .ok
+            .filter_map(move |_| match get_event_count(&path, MemoryEvent::High) {
+                Ok(high) => Some(high),
+                Err(error) => {
+                    // TODO: Might want to just panic here
+                    warn!(?error, "failed to read high events count from {}", &path);
+                    None
+                }
+            })
+            // Only report the event if the memory.high count increased
+            .filter_map(|high| {
+                if MEMORY_EVENT_COUNT.fetch_max(high, Ordering::AcqRel) < high {
+                    Some(high)
+                } else {
+                    None
+                }
+            })
+            .map(Sequenced::new);
+
+        let initial_count = get_event_count(
+            &format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name),
+            MemoryEvent::High,
+        )?;
+
+        info!(initial_count, "initial memory.high event count");
+
+        // Hard update `MEMORY_EVENT_COUNT` since there could have been processes
+        // running in the cgroup before that caused it to be non-zero.
+        MEMORY_EVENT_COUNT.fetch_max(initial_count, Ordering::AcqRel);
+
+        Ok((
+            Self {
+                cgroup,
+                upscale_requester,
+                last_upscale_seqnum: AtomicU64::new(0),
+                config: Default::default(),
+            },
+            memory_events,
+        ))
+    }
+
+    /// The entrypoint for the `CgroupWatcher`.
+    #[tracing::instrument(skip_all)]
+    pub async fn watch<E>(
+        &self,
+        // These are ~dependency injected~ (fancy, I know) because this function
+        // should never return.
+        // -> therefore: when we tokio::spawn it, we don't await the JoinHandle.
+        // -> therefore: if we want to stick it in an Arc so many threads can access
+        //    it, methods can never take mutable access.
+        //     - note: we use the Arc strategy so that a) we can call this function
+        //             right here and b) the runner can call the set/get_memory methods
+        // -> since calling recv() on a tokio::sync::mpsc::Receiver takes &mut self,
+        //    we just pass them in here instead of holding them in fields, as that
+        //    would require this method to take &mut self.
+        mut upscales: mpsc::Receiver<Sequenced<Resources>>,
+        events: E,
+    ) -> anyhow::Result<()>
+    where
+        E: Stream<Item = Sequenced<u64>>,
+    {
+        // There are several actions might do when receiving a `memory.high`,
+        // such as freezing the cgroup, or increasing its `memory.high`. We don't
+        // want to do these things too often (because postgres needs to run, and
+        // we only have so much memory). These timers serve as rate limits for this.
+        let mut wait_to_freeze = pin!(tokio::time::sleep(Duration::ZERO));
+        let mut wait_to_increase_memory_high = pin!(tokio::time::sleep(Duration::ZERO));
+        let mut events = pin!(events);
+
+        // Are we waiting to be upscaled? Could be true if we request upscale due
+        // to a memory.high event and it does not arrive in time.
+        let mut waiting_on_upscale = false;
+
+        loop {
+            tokio::select! {
+                upscale = upscales.recv() => {
+                    let Sequenced { seqnum, data } = upscale
+                        .context("failed to listen on upscale notification channel")?;
+                    self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+                    info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
+                }
+                event = events.next() => {
+                    let Some(Sequenced { seqnum, .. }) = event else {
+                        bail!("failed to listen for memory.high events")
+                    };
+                    // The memory.high came before our last upscale, so we consider
+                    // it resolved
+                    if self.last_upscale_seqnum.fetch_max(seqnum, Ordering::AcqRel) > seqnum {
+                        info!(
+                            "received memory.high event, but it came before our last upscale -> ignoring it"
+                        );
+                        continue;
+                    }
+
+                    // The memory.high came after our latest upscale. We don't
+                    // want to do anything yet, so peek the next event in hopes
+                    // that it's an upscale.
+                    if let Some(upscale_num) = self
+                        .upscaled(&mut upscales)
+                        .context("failed to check if we were upscaled")?
+                    {
+                        if upscale_num > seqnum {
+                            info!(
+                                "received memory.high event, but it came before our last upscale -> ignoring it"
+                            );
+                            continue;
+                        }
+                    }
+
+                    // If it's been long enough since we last froze, freeze the
+                    // cgroup and request upscale
+                    if wait_to_freeze.is_elapsed() {
+                        info!("received memory.high event -> requesting upscale");
+                        waiting_on_upscale = self
+                            .handle_memory_high_event(&mut upscales)
+                            .await
+                            .context("failed to handle upscale")?;
+                        wait_to_freeze
+                            .as_mut()
+                            .reset(Instant::now() + self.config.do_not_freeze_more_often_than);
+                        continue;
+                    }
+
+                    // Ok, we can't freeze, just request upscale
+                    if !waiting_on_upscale {
+                        info!("received memory.high event, but too soon to refreeze -> requesting upscale");
+
+                        // Make check to make sure we haven't been upscaled in the
+                        // meantine (can happen if the agent independently decides
+                        // to upscale us again)
+                        if self
+                            .upscaled(&mut upscales)
+                            .context("failed to check if we were upscaled")?
+                            .is_some()
+                        {
+                            info!("no need to request upscaling because we got upscaled");
+                            continue;
+                        }
+                        self.upscale_requester
+                            .send(())
+                            .await
+                            .context("failed to request upscale")?;
+                        continue;
+                    }
+
+                    // Shoot, we can't freeze or and we're still waiting on upscale,
+                    // increase memory.high to reduce throttling
+                    if wait_to_increase_memory_high.is_elapsed() {
+                        info!(
+                            "received memory.high event, \
+                            but too soon to refreeze and already requested upscale \
+                            -> increasing memory.high"
+                        );
+
+                        // Make check to make sure we haven't been upscaled in the
+                        // meantine (can happen if the agent independently decides
+                        // to upscale us again)
+                        if self
+                            .upscaled(&mut upscales)
+                            .context("failed to check if we were upscaled")?
+                            .is_some()
+                        {
+                            info!("no need to increase memory.high because got upscaled");
+                            continue;
+                        }
+
+                        // Request upscale anyways (the agent will handle deduplicating
+                        // requests)
+                        self.upscale_requester
+                            .send(())
+                            .await
+                            .context("failed to request upscale")?;
+
+                        let memory_high =
+                            self.get_high_bytes().context("failed to get memory.high")?;
+                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
+                        info!(
+                            current_high_bytes = memory_high,
+                            new_high_bytes = new_high,
+                            "updating memory.high"
+                        );
+                        self.set_high_bytes(new_high)
+                            .context("failed to set memory.high")?;
+                        wait_to_increase_memory_high
+                            .as_mut()
+                            .reset(Instant::now() + self.config.memory_high_increase_every)
+                    }
+
+                    // we can't do anything
+                }
+            };
+        }
+    }
+
+    /// Handle a `memory.high`, returning whether we are still waiting on upscale
+    /// by the time the function returns.
+    ///
+    /// The general plan for handling a `memory.high` event is as follows:
+    /// 1. Freeze the cgroup
+    /// 2. Start a timer for `self.config.max_upscale_wait`
+    /// 3. Request upscale
+    /// 4. After the timer elapses or we receive upscale, thaw the cgroup.
+    /// 5. Return whether or not we are still waiting for upscale. If we are,
+    ///    we'll increase the cgroups memory.high to avoid getting oom killed
+    #[tracing::instrument(skip_all)]
+    async fn handle_memory_high_event(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<bool> {
+        // Immediately freeze the cgroup before doing anything else.
+        info!("received memory.high event -> freezing cgroup");
+        self.freeze().context("failed to freeze cgroup")?;
+
+        // We'll use this for logging durations
+        let start_time = Instant::now();
+
+        // Await the upscale until we have to unfreeze
+        let timed =
+            tokio::time::timeout(self.config.max_upscale_wait, self.await_upscale(upscales));
+
+        // Request the upscale
+        info!(
+            wait = ?self.config.max_upscale_wait,
+            "sending request for immediate upscaling",
+        );
+        self.upscale_requester
+            .send(())
+            .await
+            .context("failed to request upscale")?;
+
+        let waiting_on_upscale = match timed.await {
+            Ok(Ok(())) => {
+                info!(elapsed = ?start_time.elapsed(), "received upscale in time");
+                false
+            }
+            // **important**: unfreeze the cgroup before ?-reporting the error
+            Ok(Err(e)) => {
+                info!("error waiting for upscale -> thawing cgroup");
+                self.thaw()
+                    .context("failed to thaw cgroup after errored waiting for upscale")?;
+                Err(e.context("failed to await upscale"))?
+            }
+            Err(_) => {
+                info!(elapsed = ?self.config.max_upscale_wait, "timed out waiting for upscale");
+                true
+            }
+        };
+
+        info!("thawing cgroup");
+        self.thaw().context("failed to thaw cgroup")?;
+
+        Ok(waiting_on_upscale)
+    }
+
+    /// Checks whether we were just upscaled, returning the upscale's sequence
+    /// number if so.
+    #[tracing::instrument(skip_all)]
+    fn upscaled(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<Option<u64>> {
+        let Sequenced { seqnum, data } = match upscales.try_recv() {
+            Ok(upscale) => upscale,
+            Err(TryRecvError::Empty) => return Ok(None),
+            Err(TryRecvError::Disconnected) => {
+                bail!("upscale notification channel was disconnected")
+            }
+        };
+
+        // Make sure to update the last upscale sequence number
+        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+        info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
+        Ok(Some(seqnum))
+    }
+
+    /// Await an upscale event, discarding any `memory.high` events received in
+    /// the process.
+    ///
+    /// This is used in `handle_memory_high_event`, where we need to listen
+    /// for upscales in particular so we know if we can thaw the cgroup early.
+    #[tracing::instrument(skip_all)]
+    async fn await_upscale(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<()> {
+        let Sequenced { seqnum, .. } = upscales
+            .recv()
+            .await
+            .context("error listening for upscales")?;
+
+        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+        Ok(())
+    }
+
+    /// Get the cgroup's name.
+    pub fn path(&self) -> &str {
+        self.cgroup.path()
+    }
+}
+
+/// Represents a set of limits we apply to a cgroup to control memory usage.
+///
+/// Setting these values also affects the thresholds for receiving usage alerts.
+#[derive(Debug)]
+pub struct MemoryLimits {
+    high: u64,
+    max: u64,
+}
+
+impl MemoryLimits {
+    pub fn new(high: u64, max: u64) -> Self {
+        Self { max, high }
+    }
+}
+
+// Methods for manipulating the actual cgroup
+impl CgroupWatcher {
+    /// Get a handle on the freezer subsystem.
+    fn freezer(&self) -> anyhow::Result<&FreezerController> {
+        if let Some(Freezer(freezer)) = self
+            .cgroup
+            .subsystems()
+            .iter()
+            .find(|sub| matches!(sub, Freezer(_)))
+        {
+            Ok(freezer)
+        } else {
+            anyhow::bail!("could not find freezer subsystem")
+        }
+    }
+
+    /// Attempt to freeze the cgroup.
+    pub fn freeze(&self) -> anyhow::Result<()> {
+        self.freezer()
+            .context("failed to get freezer subsystem")?
+            .freeze()
+            .context("failed to freeze")
+    }
+
+    /// Attempt to thaw the cgroup.
+    pub fn thaw(&self) -> anyhow::Result<()> {
+        self.freezer()
+            .context("failed to get freezer subsystem")?
+            .thaw()
+            .context("failed to thaw")
+    }
+
+    /// Get a handle on the memory subsystem.
+    ///
+    /// Note: this method does not require `self.memory_update_lock` because
+    /// getting a handle to the subsystem does not access any of the files we
+    /// care about, such as memory.high and memory.events
+    fn memory(&self) -> anyhow::Result<&MemController> {
+        if let Some(Mem(memory)) = self
+            .cgroup
+            .subsystems()
+            .iter()
+            .find(|sub| matches!(sub, Mem(_)))
+        {
+            Ok(memory)
+        } else {
+            anyhow::bail!("could not find memory subsystem")
+        }
+    }
+
+    /// Get cgroup current memory usage.
+    pub fn current_memory_usage(&self) -> anyhow::Result<u64> {
+        Ok(self
+            .memory()
+            .context("failed to get memory subsystem")?
+            .memory_stat()
+            .usage_in_bytes)
+    }
+
+    /// Set cgroup memory.high threshold.
+    pub fn set_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
+        self.memory()
+            .context("failed to get memory subsystem")?
+            .set_mem(cgroups_rs::memory::SetMemory {
+                low: None,
+                high: Some(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64)),
+                min: None,
+                max: None,
+            })
+            .context("failed to set memory.high")
+    }
+
+    /// Set cgroup memory.high and memory.max.
+    pub fn set_limits(&self, limits: &MemoryLimits) -> anyhow::Result<()> {
+        info!(
+            limits.high,
+            limits.max,
+            path = self.path(),
+            "writing new memory limits",
+        );
+        self.memory()
+            .context("failed to get memory subsystem while setting memory limits")?
+            .set_mem(cgroups_rs::memory::SetMemory {
+                min: None,
+                low: None,
+                high: Some(MaxValue::Value(
+                    u64::min(limits.high, i64::MAX as u64) as i64
+                )),
+                max: Some(MaxValue::Value(u64::min(limits.max, i64::MAX as u64) as i64)),
+            })
+            .context("failed to set memory limits")
+    }
+
+    /// Given some amount of available memory, set the desired cgroup memory limits
+    pub fn set_memory_limits(&mut self, available_memory: u64) -> anyhow::Result<()> {
+        let new_high = self.config.calculate_memory_high_value(available_memory);
+        let limits = MemoryLimits::new(new_high, available_memory);
+        info!(
+            path = self.path(),
+            memory = ?limits,
+            "setting cgroup memory",
+        );
+        self.set_limits(&limits)
+            .context("failed to set cgroup memory limits")?;
+        Ok(())
+    }
+
+    /// Get memory.high threshold.
+    pub fn get_high_bytes(&self) -> anyhow::Result<u64> {
+        let high = self
+            .memory()
+            .context("failed to get memory subsystem while getting memory statistics")?
+            .get_mem()
+            .map(|mem| mem.high)
+            .context("failed to get memory statistics from subsystem")?;
+        match high {
+            Some(MaxValue::Max) => Ok(i64::MAX as u64),
+            Some(MaxValue::Value(high)) => Ok(high as u64),
+            None => anyhow::bail!("failed to read memory.high from memory subsystem"),
+        }
+    }
+}

libs/vm_monitor/src/dispatcher.rs

@@ -0,0 +1,155 @@
+//! Managing the websocket connection and other signals in the monitor.
+//!
+//! Contains types that manage the interaction (not data interchange, see `protocol`)
+//! between informant and monitor, allowing us to to process and send messages in a
+//! straightforward way. The dispatcher also manages that signals that come from
+//! the cgroup (requesting upscale), and the signals that go to the cgroup
+//! (notifying it of upscale).
+
+use anyhow::{bail, Context};
+use axum::extract::ws::{Message, WebSocket};
+use futures::{
+    stream::{SplitSink, SplitStream},
+    SinkExt, StreamExt,
+};
+use tokio::sync::mpsc;
+use tracing::info;
+
+use crate::cgroup::Sequenced;
+use crate::protocol::{
+    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, Resources, PROTOCOL_MAX_VERSION,
+    PROTOCOL_MIN_VERSION,
+};
+
+/// The central handler for all communications in the monitor.
+///
+/// The dispatcher has two purposes:
+/// 1. Manage the connection to the informant, sending and receiving messages.
+/// 2. Communicate with the cgroup manager, notifying it when upscale is received,
+///    and sending a message to the informant when the cgroup manager requests
+///    upscale.
+#[derive(Debug)]
+pub struct Dispatcher {
+    /// We read informant messages of of `source`
+    pub(crate) source: SplitStream<WebSocket>,
+
+    /// We send messages to the informant through `sink`
+    sink: SplitSink<WebSocket, Message>,
+
+    /// Used to notify the cgroup when we are upscaled.
+    pub(crate) notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
+
+    /// When the cgroup requests upscale it will send on this channel. In response
+    /// we send an `UpscaleRequst` to the agent.
+    pub(crate) request_upscale_events: mpsc::Receiver<()>,
+
+    /// The protocol version we have agreed to use with the informant. This is negotiated
+    /// during the creation of the dispatcher, and should be the highest shared protocol
+    /// version.
+    ///
+    // NOTE: currently unused, but will almost certainly be used in the futures
+    // as the protocol changes
+    #[allow(unused)]
+    pub(crate) proto_version: ProtocolVersion,
+}
+
+impl Dispatcher {
+    /// Creates a new dispatcher using the passed-in connection.
+    ///
+    /// Performs a negotiation with the informant to determine the highest protocol
+    /// version that both support. This consists of two steps:
+    /// 1. Wait for the informant to sent the range of protocols it supports.
+    /// 2. Send a protocol version that works for us as well, or an error if there
+    ///    is no compatible version.
+    pub async fn new(
+        stream: WebSocket,
+        notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
+        request_upscale_events: mpsc::Receiver<()>,
+    ) -> anyhow::Result<Self> {
+        let (mut sink, mut source) = stream.split();
+
+        // Figure out the highest protocol version we both support
+        info!("waiting for informant to send protocol version range");
+        let Some(message) = source.next().await else {
+            bail!("websocket connection closed while performing protocol handshake")
+        };
+
+        let message = message.context("failed to read protocol version range off connection")?;
+
+        let Message::Text(message_text) = message else {
+            // All messages should be in text form, since we don't do any
+            // pinging/ponging. See nhooyr/websocket's implementation and the
+            // informant/agent for more info
+            bail!("received non-text message during proocol handshake: {message:?}")
+        };
+
+        let monitor_range = ProtocolRange {
+            min: PROTOCOL_MIN_VERSION,
+            max: PROTOCOL_MAX_VERSION,
+        };
+
+        let informant_range: ProtocolRange = serde_json::from_str(&message_text)
+            .context("failed to deserialize protocol version range")?;
+
+        info!(range = ?informant_range, "received protocol version range");
+
+        let highest_shared_version = match monitor_range.highest_shared_version(&informant_range) {
+            Ok(version) => {
+                sink.send(Message::Text(
+                    serde_json::to_string(&ProtocolResponse::Version(version)).unwrap(),
+                ))
+                .await
+                .context("failed to notify informant of negotiated protocol version")?;
+                version
+            }
+            Err(e) => {
+                sink.send(Message::Text(
+                    serde_json::to_string(&ProtocolResponse::Error(format!(
+                        "Received protocol version range {} which does not overlap with {}",
+                        informant_range, monitor_range
+                    )))
+                    .unwrap(),
+                ))
+                .await
+                .context(
+                    "failed to notify informant of no overlap between protocol version ranges",
+                )?;
+                Err(e).context("error determining suitable protocol version range")?
+            }
+        };
+
+        Ok(Self {
+            sink,
+            source,
+            notify_upscale_events,
+            request_upscale_events,
+            proto_version: highest_shared_version,
+        })
+    }
+
+    /// Notify the cgroup manager that we have received upscale and wait for
+    /// the acknowledgement.
+    #[tracing::instrument(skip_all, fields(?resources))]
+    pub async fn notify_upscale(&self, resources: Sequenced<Resources>) -> anyhow::Result<()> {
+        self.notify_upscale_events
+            .send(resources)
+            .await
+            .context("failed to send resources and oneshot sender across channel")
+    }
+
+    /// Send a message to the informant.
+    ///
+    /// Although this function is small, it has one major benefit: it is the only
+    /// way to send data accross the connection, and you can only pass in a proper
+    /// `MonitorMessage`. Without safeguards like this, it's easy to accidentally
+    /// serialize the wrong thing and send it, since `self.sink.send` will take
+    /// any string.
+    pub async fn send(&mut self, message: OutboundMsg) -> anyhow::Result<()> {
+        info!(?message, "sending message");
+        let json = serde_json::to_string(&message).context("failed to serialize message")?;
+        self.sink
+            .send(Message::Text(json))
+            .await
+            .context("stream error sending message")
+    }
+}

libs/vm_monitor/src/filecache.rs

@@ -0,0 +1,306 @@
+//! Logic for configuring and scaling the Postgres file cache.
+
+use std::num::NonZeroU64;
+
+use crate::MiB;
+use anyhow::{anyhow, Context};
+use tokio_postgres::{types::ToSql, Client, NoTls, Row};
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info};
+
+/// Manages Postgres' file cache by keeping a connection open.
+#[derive(Debug)]
+pub struct FileCacheState {
+    client: Client,
+    conn_str: String,
+    pub(crate) config: FileCacheConfig,
+
+    /// A token for cancelling spawned threads during shutdown.
+    token: CancellationToken,
+}
+
+#[derive(Debug)]
+pub struct FileCacheConfig {
+    /// Whether the file cache is *actually* stored in memory (e.g. by writing to
+    /// a tmpfs or shmem file). If true, the size of the file cache will be counted against the
+    /// memory available for the cgroup.
+    pub(crate) in_memory: bool,
+
+    /// The size of the file cache, in terms of the size of the resource it consumes
+    /// (currently: only memory)
+    ///
+    /// For example, setting `resource_multipler = 0.75` gives the cache a target size of 75% of total
+    /// resources.
+    ///
+    /// This value must be strictly between 0 and 1.
+    resource_multiplier: f64,
+
+    /// The required minimum amount of memory, in bytes, that must remain available
+    /// after subtracting the file cache.
+    ///
+    /// This value must be non-zero.
+    min_remaining_after_cache: NonZeroU64,
+
+    /// Controls the rate of increase in the file cache's size as it grows from zero
+    /// (when total resources equals min_remaining_after_cache) to the desired size based on
+    /// `resource_multiplier`.
+    ///
+    /// A `spread_factor` of zero means that all additional resources will go to the cache until it
+    /// reaches the desired size. Setting `spread_factor` to N roughly means "for every 1 byte added to
+    /// the cache's size, N bytes are reserved for the rest of the system, until the cache gets to
+    /// its desired size".
+    ///
+    /// This value must be >= 0, and must retain an increase that is more than what would be given by
+    /// `resource_multiplier`. For example, setting `resource_multiplier` = 0.75 but `spread_factor` = 1
+    /// would be invalid, because `spread_factor` would induce only 50% usage - never reaching the 75%
+    /// as desired by `resource_multiplier`.
+    ///
+    /// `spread_factor` is too large if `(spread_factor + 1) * resource_multiplier >= 1`.
+    spread_factor: f64,
+}
+
+impl Default for FileCacheConfig {
+    fn default() -> Self {
+        Self {
+            in_memory: true,
+            // 75 %
+            resource_multiplier: 0.75,
+            // 640 MiB; (512 + 128)
+            min_remaining_after_cache: NonZeroU64::new(640 * MiB).unwrap(),
+            // ensure any increase in file cache size is split 90-10 with 10% to other memory
+            spread_factor: 0.1,
+        }
+    }
+}
+
+impl FileCacheConfig {
+    /// Make sure fields of the config are consistent.
+    pub fn validate(&self) -> anyhow::Result<()> {
+        // Single field validity
+        anyhow::ensure!(
+            0.0 < self.resource_multiplier && self.resource_multiplier < 1.0,
+            "resource_multiplier must be between 0.0 and 1.0 exclusive, got {}",
+            self.resource_multiplier
+        );
+        anyhow::ensure!(
+            self.spread_factor >= 0.0,
+            "spread_factor must be >= 0, got {}",
+            self.spread_factor
+        );
+
+        // Check that `resource_multiplier` and `spread_factor` are valid w.r.t. each other.
+        //
+        // As shown in `calculate_cache_size`, we have two lines resulting from `resource_multiplier` and
+        // `spread_factor`, respectively. They are:
+        //
+        //                 `total`           `min_remaining_after_cache`
+        //   size = ————————————————————— - —————————————————————————————
+        //           `spread_factor` + 1         `spread_factor` + 1
+        //
+        // and
+        //
+        //   size = `resource_multiplier` × total
+        //
+        // .. where `total` is the total resources. These are isomorphic to the typical 'y = mx + b'
+        // form, with y = "size" and x = "total".
+        //
+        // These lines intersect at:
+        //
+        //               `min_remaining_after_cache`
+        //   ———————————————————————————————————————————————————
+        //    1 - `resource_multiplier` × (`spread_factor` + 1)
+        //
+        // We want to ensure that this value (a) exists, and (b) is >= `min_remaining_after_cache`. This is
+        // guaranteed when '`resource_multiplier` × (`spread_factor` + 1)' is less than 1.
+        // (We also need it to be >= 0, but that's already guaranteed.)
+
+        let intersect_factor = self.resource_multiplier * (self.spread_factor + 1.0);
+        anyhow::ensure!(
+            intersect_factor < 1.0,
+            "incompatible resource_multipler and spread_factor"
+        );
+        Ok(())
+    }
+
+    /// Calculate the desired size of the cache, given the total memory
+    pub fn calculate_cache_size(&self, total: u64) -> u64 {
+        // *Note*: all units are in bytes, until the very last line.
+        let available = total.saturating_sub(self.min_remaining_after_cache.get());
+        if available == 0 {
+            return 0;
+        }
+
+        // Conversions to ensure we don't overflow from floating-point ops
+        let size_from_spread =
+            i64::max(0, (available as f64 / (1.0 + self.spread_factor)) as i64) as u64;
+
+        let size_from_normal = (total as f64 * self.resource_multiplier) as u64;
+
+        let byte_size = u64::min(size_from_spread, size_from_normal);
+
+        // The file cache operates in units of mebibytes, so the sizes we produce should
+        // be rounded to a mebibyte. We round down to be conservative.
+        byte_size / MiB * MiB
+    }
+}
+
+impl FileCacheState {
+    /// Connect to the file cache.
+    #[tracing::instrument(skip_all, fields(%conn_str, ?config))]
+    pub async fn new(
+        conn_str: &str,
+        config: FileCacheConfig,
+        token: CancellationToken,
+    ) -> anyhow::Result<Self> {
+        config.validate().context("file cache config is invalid")?;
+
+        info!(conn_str, "connecting to Postgres file cache");
+        let client = FileCacheState::connect(conn_str, token.clone())
+            .await
+            .context("failed to connect to postgres file cache")?;
+
+        let conn_str = conn_str.to_string();
+        Ok(Self {
+            client,
+            config,
+            conn_str,
+            token,
+        })
+    }
+
+    /// Connect to Postgres.
+    ///
+    /// Aborts the spawned thread if the kill signal is received. This is not
+    /// a method as it is called in [`FileCacheState::new`].
+    #[tracing::instrument(skip_all, fields(%conn_str))]
+    async fn connect(conn_str: &str, token: CancellationToken) -> anyhow::Result<Client> {
+        let (client, conn) = tokio_postgres::connect(conn_str, NoTls)
+            .await
+            .context("failed to connect to pg client")?;
+
+        // The connection object performs the actual communication with the database,
+        // so spawn it off to run on its own. See tokio-postgres docs.
+        crate::spawn_with_cancel(
+            token,
+            |res| {
+                if let Err(error) = res {
+                    error!(%error, "postgres error")
+                }
+            },
+            conn,
+        );
+
+        Ok(client)
+    }
+
+    /// Execute a query with a retry if necessary.
+    ///
+    /// If the initial query fails, we restart the database connection and attempt
+    /// if again.
+    #[tracing::instrument(skip_all, fields(%statement))]
+    pub async fn query_with_retry(
+        &mut self,
+        statement: &str,
+        params: &[&(dyn ToSql + Sync)],
+    ) -> anyhow::Result<Vec<Row>> {
+        match self
+            .client
+            .query(statement, params)
+            .await
+            .context("failed to execute query")
+        {
+            Ok(rows) => Ok(rows),
+            Err(e) => {
+                error!(error = ?e, "postgres error: {e} -> retrying");
+
+                let client = FileCacheState::connect(&self.conn_str, self.token.clone())
+                    .await
+                    .context("failed to connect to postgres file cache")?;
+                info!("successfully reconnected to postgres client");
+
+                // Replace the old client and attempt the query with the new one
+                self.client = client;
+                self.client
+                    .query(statement, params)
+                    .await
+                    .context("failed to execute query a second time")
+            }
+        }
+    }
+
+    /// Get the current size of the file cache.
+    #[tracing::instrument(skip_all)]
+    pub async fn get_file_cache_size(&mut self) -> anyhow::Result<u64> {
+        self.query_with_retry(
+            // The file cache GUC variable is in MiB, but the conversion with
+            // pg_size_bytes means that the end result we get is in bytes.
+            "SELECT pg_size_bytes(current_setting('neon.file_cache_size_limit'));",
+            &[],
+        )
+        .await
+        .context("failed to query pg for file cache size")?
+        .first()
+        .ok_or_else(|| anyhow!("file cache size query returned no rows"))?
+        // pg_size_bytes returns a bigint which is the same as an i64.
+        .try_get::<_, i64>(0)
+        // Since the size of the table is not negative, the cast is sound.
+        .map(|bytes| bytes as u64)
+        .context("failed to extract file cache size from query result")
+    }
+
+    /// Attempt to set the file cache size, returning the size it was actually
+    /// set to.
+    #[tracing::instrument(skip_all, fields(%num_bytes))]
+    pub async fn set_file_cache_size(&mut self, num_bytes: u64) -> anyhow::Result<u64> {
+        let max_bytes = self
+            // The file cache GUC variable is in MiB, but the conversion with pg_size_bytes
+            // means that the end result we get is in bytes.
+            .query_with_retry(
+                "SELECT pg_size_bytes(current_setting('neon.max_file_cache_size'));",
+                &[],
+            )
+            .await
+            .context("failed to query pg for max file cache size")?
+            .first()
+            .ok_or_else(|| anyhow!("max file cache size query returned no rows"))?
+            .try_get::<_, i64>(0)
+            .map(|bytes| bytes as u64)
+            .context("failed to extract max file cache size from query result")?;
+
+        let max_mb = max_bytes / MiB;
+        let num_mb = u64::min(num_bytes, max_bytes) / MiB;
+
+        let capped = if num_bytes > max_bytes {
+            " (capped by maximum size)"
+        } else {
+            ""
+        };
+
+        info!(
+            size = num_mb,
+            max = max_mb,
+            "updating file cache size {capped}",
+        );
+
+        // note: even though the normal ways to get the cache size produce values with trailing "MB"
+        // (hence why we call pg_size_bytes in `get_file_cache_size`'s query), the format
+        // it expects to set the value is "integer number of MB" without trailing units.
+        // For some reason, this *really* wasn't working with normal arguments, so that's
+        // why we're constructing the query here.
+        self.client
+            .query(
+                &format!("ALTER SYSTEM SET neon.file_cache_size_limit = {};", num_mb),
+                &[],
+            )
+            .await
+            .context("failed to change file cache size limit")?;
+
+        // must use pg_reload_conf to have the settings change take effect
+        self.client
+            .execute("SELECT pg_reload_conf();", &[])
+            .await
+            .context("failed to reload config")?;
+
+        Ok(num_mb * MiB)
+    }
+}

libs/vm_monitor/src/lib.rs

@@ -0,0 +1,205 @@
+#![cfg(target_os = "linux")]
+
+use anyhow::Context;
+use axum::{
+    extract::{ws::WebSocket, State, WebSocketUpgrade},
+    response::Response,
+};
+use axum::{routing::get, Router, Server};
+use clap::Parser;
+use futures::Future;
+use std::{fmt::Debug, time::Duration};
+use sysinfo::{RefreshKind, System, SystemExt};
+use tokio::{sync::broadcast, task::JoinHandle};
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info};
+
+use runner::Runner;
+
+// Code that interfaces with agent
+pub mod dispatcher;
+pub mod protocol;
+
+pub mod cgroup;
+pub mod filecache;
+pub mod runner;
+
+/// The vm-monitor is an autoscaling component started by compute_ctl.
+///
+/// It carries out autoscaling decisions (upscaling/downscaling) and responds to
+/// memory pressure by making requests to the autoscaler-agent.
+#[derive(Debug, Parser)]
+pub struct Args {
+    /// The name of the cgroup we should monitor for memory.high events. This
+    /// is the cgroup that postgres should be running in.
+    #[arg(short, long)]
+    pub cgroup: Option<String>,
+
+    /// The connection string for the Postgres file cache we should manage.
+    #[arg(short, long)]
+    pub pgconnstr: Option<String>,
+
+    /// The address we should listen on for connection requests. For the
+    /// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
+    #[arg(short, long)]
+    pub addr: String,
+}
+
+impl Args {
+    pub fn addr(&self) -> &str {
+        &self.addr
+    }
+}
+
+/// The number of bytes in one mebibyte.
+#[allow(non_upper_case_globals)]
+const MiB: u64 = 1 << 20;
+
+/// Convert a quantity in bytes to a quantity in mebibytes, generally for display
+/// purposes. (Most calculations in this crate use bytes directly)
+pub fn bytes_to_mebibytes(bytes: u64) -> f32 {
+    (bytes as f32) / (MiB as f32)
+}
+
+pub fn get_total_system_memory() -> u64 {
+    System::new_with_specifics(RefreshKind::new().with_memory()).total_memory()
+}
+
+/// Global app state for the Axum server
+#[derive(Debug, Clone)]
+pub struct ServerState {
+    /// Used to close old connections.
+    ///
+    /// When a new connection is made, we send a message signalling to the old
+    /// connection to close.
+    pub sender: broadcast::Sender<()>,
+
+    /// Used to cancel all spawned threads in the monitor.
+    pub token: CancellationToken,
+
+    // The CLI args
+    pub args: &'static Args,
+}
+
+/// Spawn a thread that may get cancelled by the provided [`CancellationToken`].
+///
+/// This is mainly meant to be called with futures that will be pending for a very
+/// long time, or are not mean to return. If it is not desirable for the future to
+/// ever resolve, such as in the case of [`cgroup::CgroupWatcher::watch`], the error can
+/// be logged with `f`.
+pub fn spawn_with_cancel<T, F>(
+    token: CancellationToken,
+    f: F,
+    future: T,
+) -> JoinHandle<Option<T::Output>>
+where
+    T: Future + Send + 'static,
+    T::Output: Send + 'static,
+    F: FnOnce(&T::Output) + Send + 'static,
+{
+    tokio::spawn(async move {
+        tokio::select! {
+            _ = token.cancelled() => {
+                info!("received global kill signal");
+                None
+            }
+            res = future => {
+                f(&res);
+                Some(res)
+            }
+        }
+    })
+}
+
+/// The entrypoint to the binary.
+///
+/// Set up tracing, parse arguments, and start an http server.
+pub async fn start(args: &'static Args, token: CancellationToken) -> anyhow::Result<()> {
+    // This channel is used to close old connections. When a new connection is
+    // made, we send a message signalling to the old connection to close.
+    let (sender, _) = tokio::sync::broadcast::channel::<()>(1);
+
+    let app = Router::new()
+        // This route gets upgraded to a websocket connection. We only support
+        // one connection at a time, which we enforce by killing old connections
+        // when we receive a new one.
+        .route("/monitor", get(ws_handler))
+        .with_state(ServerState {
+            sender,
+            token,
+            args,
+        });
+
+    let addr = args.addr();
+    let bound = Server::try_bind(&addr.parse().expect("parsing address should not fail"))
+        .with_context(|| format!("failed to bind to {addr}"))?;
+
+    info!(addr, "server bound");
+
+    bound
+        .serve(app.into_make_service())
+        .await
+        .context("server exited")?;
+
+    Ok(())
+}
+
+/// Handles incoming websocket connections.
+///
+/// If we are already to connected to an informant, we kill that old connection
+/// and accept the new one.
+#[tracing::instrument(name = "/monitor", skip_all, fields(?args))]
+pub async fn ws_handler(
+    ws: WebSocketUpgrade,
+    State(ServerState {
+        sender,
+        token,
+        args,
+    }): State<ServerState>,
+) -> Response {
+    // Kill the old monitor
+    info!("closing old connection if there is one");
+    let _ = sender.send(());
+
+    // Start the new one. Wow, the cycle of death and rebirth
+    let closer = sender.subscribe();
+    ws.on_upgrade(|ws| start_monitor(ws, args, closer, token))
+}
+
+/// Starts the monitor. If startup fails or the monitor exits, an error will
+/// be logged and our internal state will be reset to allow for new connections.
+#[tracing::instrument(skip_all, fields(?args))]
+async fn start_monitor(
+    ws: WebSocket,
+    args: &Args,
+    kill: broadcast::Receiver<()>,
+    token: CancellationToken,
+) {
+    info!("accepted new websocket connection -> starting monitor");
+    let timeout = Duration::from_secs(4);
+    let monitor = tokio::time::timeout(
+        timeout,
+        Runner::new(Default::default(), args, ws, kill, token),
+    )
+    .await;
+    let mut monitor = match monitor {
+        Ok(Ok(monitor)) => monitor,
+        Ok(Err(error)) => {
+            error!(?error, "failed to create monitor");
+            return;
+        }
+        Err(_) => {
+            error!(
+                ?timeout,
+                "creating monitor timed out (probably waiting to receive protocol range)"
+            );
+            return;
+        }
+    };
+    info!("connected to informant");
+
+    match monitor.run().await {
+        Ok(()) => info!("monitor was killed due to new connection"),
+        Err(e) => error!(error = ?e, "monitor terminated unexpectedly"),
+    }
+}

libs/vm_monitor/src/protocol.rs

@@ -0,0 +1,241 @@
+//! Types representing protocols and actual informant-monitor messages.
+//!
+//! The pervasive use of serde modifiers throughout this module is to ease
+//! serialization on the go side. Because go does not have enums (which model
+//! messages well), it is harder to model messages, and we accomodate that with
+//! serde.
+//!
+//! *Note*: the informant sends and receives messages in different ways.
+//!
+//! The informant serializes messages in the form and then sends them. The use
+//! of `#[serde(tag = "type", content = "content")]` allows us to use `Type`
+//! to determine how to deserialize `Content`.
+//! ```ignore
+//! struct {
+//!     Content any
+//!     Type    string
+//!     Id      uint64
+//! }
+//! ```
+//! and receives messages in the form:
+//! ```ignore
+//! struct {
+//!     {fields embedded}
+//!     Type string
+//!     Id   uint64
+//! }
+//! ```
+//! After reading the type field, the informant will decode the entire message
+//! again, this time into the correct type using the embedded fields.
+//! Because the informant cannot just extract the json contained in a certain field
+//! (it initially deserializes to `map[string]interface{}`), we keep the fields
+//! at the top level, so the entire piece of json can be deserialized into a struct,
+//! such as a `DownscaleResult`, with the `Type` and `Id` fields ignored.
+
+use core::fmt;
+use std::cmp;
+
+use serde::{de::Error, Deserialize, Serialize};
+
+/// A Message we send to the informant.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct OutboundMsg {
+    #[serde(flatten)]
+    pub(crate) inner: OutboundMsgKind,
+    pub(crate) id: usize,
+}
+
+impl OutboundMsg {
+    pub fn new(inner: OutboundMsgKind, id: usize) -> Self {
+        Self { inner, id }
+    }
+}
+
+/// The different underlying message types we can send to the informant.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+#[serde(tag = "type")]
+pub enum OutboundMsgKind {
+    /// Indicates that the informant sent an invalid message, i.e, we couldn't
+    /// properly deserialize it.
+    InvalidMessage { error: String },
+    /// Indicates that we experienced an internal error while processing a message.
+    /// For example, if a cgroup operation fails while trying to handle an upscale,
+    /// we return `InternalError`.
+    InternalError { error: String },
+    /// Returned to the informant once we have finished handling an upscale. If the
+    /// handling was unsuccessful, an `InternalError` will get returned instead.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    UpscaleConfirmation {},
+    /// Indicates to the monitor that we are urgently requesting resources.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    UpscaleRequest {},
+    /// Returned to the informant once we have finished attempting to downscale. If
+    /// an error occured trying to do so, an `InternalError` will get returned instead.
+    /// However, if we are simply unsuccessful (for example, do to needing the resources),
+    /// that gets included in the `DownscaleResult`.
+    DownscaleResult {
+        // FIXME for the future (once the informant is deprecated)
+        // As of the time of writing, the informant/agent version of this struct is
+        // called api.DownscaleResult. This struct has uppercase fields which are
+        // serialized as such. Thus, we serialize using uppercase names so we don't
+        // have to make a breaking change to the agent<->informant protocol. Once
+        // the informant has been superseded by the monitor, we can add the correct
+        // struct tags to api.DownscaleResult without causing a breaking change,
+        // since we don't need to support the agent<->informant protocol anymore.
+        #[serde(rename = "Ok")]
+        ok: bool,
+        #[serde(rename = "Status")]
+        status: String,
+    },
+    /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
+    /// informant.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    HealthCheck {},
+}
+
+/// A message received form the informant.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct InboundMsg {
+    #[serde(flatten)]
+    pub(crate) inner: InboundMsgKind,
+    pub(crate) id: usize,
+}
+
+/// The different underlying message types we can receive from the informant.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+#[serde(tag = "type", content = "content")]
+pub enum InboundMsgKind {
+    /// Indicates that the we sent an invalid message, i.e, we couldn't
+    /// properly deserialize it.
+    InvalidMessage { error: String },
+    /// Indicates that the informan experienced an internal error while processing
+    /// a message. For example, if it failed to request upsacle from the agent, it
+    /// would return an `InternalError`.
+    InternalError { error: String },
+    /// Indicates to us that we have been granted more resources. We should respond
+    /// with an `UpscaleConfirmation` when done handling the resources (increasins
+    /// file cache size, cgorup memory limits).
+    UpscaleNotification { granted: Resources },
+    /// A request to reduce resource usage. We should response with a `DownscaleResult`,
+    /// when done.
+    DownscaleRequest { target: Resources },
+    /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
+    /// informant.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    HealthCheck {},
+}
+
+/// Represents the resources granted to a VM.
+#[derive(Serialize, Deserialize, Debug, Clone, Copy)]
+// Renamed because the agent/informant has multiple resources types:
+// `Resources` (milliCPU/memory slots)
+// `Allocation` (vCPU/bytes) <- what we correspond to
+#[serde(rename(serialize = "Allocation", deserialize = "Allocation"))]
+pub struct Resources {
+    /// Number of vCPUs
+    pub(crate) cpu: f64,
+    /// Bytes of memory
+    pub(crate) mem: u64,
+}
+
+impl Resources {
+    pub fn new(cpu: f64, mem: u64) -> Self {
+        Self { cpu, mem }
+    }
+}
+
+pub const PROTOCOL_MIN_VERSION: ProtocolVersion = ProtocolVersion::V1_0;
+pub const PROTOCOL_MAX_VERSION: ProtocolVersion = ProtocolVersion::V1_0;
+
+#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Ord, Eq, Serialize, Deserialize)]
+pub struct ProtocolVersion(u8);
+
+impl ProtocolVersion {
+    /// Represents v1.0 of the informant<-> monitor protocol - the initial version
+    ///
+    /// Currently the latest version.
+    const V1_0: ProtocolVersion = ProtocolVersion(1);
+}
+
+impl fmt::Display for ProtocolVersion {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match *self {
+            ProtocolVersion(0) => f.write_str("<invalid: zero>"),
+            ProtocolVersion::V1_0 => f.write_str("v1.0"),
+            other => write!(f, "<unknown: {other}>"),
+        }
+    }
+}
+
+/// A set of protocol bounds that determines what we are speaking.
+///
+/// These bounds are inclusive.
+#[derive(Debug)]
+pub struct ProtocolRange {
+    pub min: ProtocolVersion,
+    pub max: ProtocolVersion,
+}
+
+// Use a custom deserialize impl to ensure that `self.min <= self.max`
+impl<'de> Deserialize<'de> for ProtocolRange {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        #[derive(Deserialize)]
+        struct InnerProtocolRange {
+            min: ProtocolVersion,
+            max: ProtocolVersion,
+        }
+        let InnerProtocolRange { min, max } = InnerProtocolRange::deserialize(deserializer)?;
+        if min > max {
+            Err(D::Error::custom(format!(
+                "min version = {min} is greater than max version = {max}",
+            )))
+        } else {
+            Ok(ProtocolRange { min, max })
+        }
+    }
+}
+
+impl fmt::Display for ProtocolRange {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        if self.min == self.max {
+            f.write_fmt(format_args!("{}", self.max))
+        } else {
+            f.write_fmt(format_args!("{} to {}", self.min, self.max))
+        }
+    }
+}
+
+impl ProtocolRange {
+    /// Find the highest shared version between two `ProtocolRange`'s
+    pub fn highest_shared_version(&self, other: &Self) -> anyhow::Result<ProtocolVersion> {
+        // We first have to make sure the ranges are overlapping. Once we know
+        // this, we can merge the ranges by taking the max of the mins and the
+        // mins of the maxes.
+        if self.min > other.max {
+            anyhow::bail!(
+                "Non-overlapping bounds: other.max = {} was less than self.min = {}",
+                other.max,
+                self.min,
+            )
+        } else if self.max < other.min {
+            anyhow::bail!(
+                "Non-overlappinng bounds: self.max = {} was less than other.min = {}",
+                self.max,
+                other.min
+            )
+        } else {
+            Ok(cmp::min(self.max, other.max))
+        }
+    }
+}
+
+/// We send this to the monitor after negotiating which protocol to use
+#[derive(Serialize, Debug)]
+#[serde(rename_all = "camelCase")]
+pub enum ProtocolResponse {
+    Error(String),
+    Version(ProtocolVersion),
+}

libs/vm_monitor/src/runner.rs

@@ -0,0 +1,460 @@
+//! Exposes the `Runner`, which handles messages received from informant and
+//! sends upscale requests.
+//!
+//! This is the "Monitor" part of the monitor binary and is the main entrypoint for
+//! all functionality.
+
+use std::sync::Arc;
+use std::{fmt::Debug, mem};
+
+use anyhow::{bail, Context};
+use axum::extract::ws::{Message, WebSocket};
+use futures::StreamExt;
+use tokio::sync::broadcast;
+use tokio::sync::mpsc;
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info, warn};
+
+use crate::cgroup::{CgroupWatcher, MemoryLimits, Sequenced};
+use crate::dispatcher::Dispatcher;
+use crate::filecache::{FileCacheConfig, FileCacheState};
+use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
+use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args, MiB};
+
+/// Central struct that interacts with informant, dispatcher, and cgroup to handle
+/// signals from the informant.
+#[derive(Debug)]
+pub struct Runner {
+    config: Config,
+    filecache: Option<FileCacheState>,
+    cgroup: Option<Arc<CgroupWatcher>>,
+    dispatcher: Dispatcher,
+
+    /// We "mint" new message ids by incrementing this counter and taking the value.
+    ///
+    /// **Note**: This counter is always odd, so that we avoid collisions between the IDs generated
+    /// by us vs the autoscaler-agent.
+    counter: usize,
+
+    /// A signal to kill the main thread produced by `self.run()`. This is triggered
+    /// when the server receives a new connection. When the thread receives the
+    /// signal off this channel, it will gracefully shutdown.
+    kill: broadcast::Receiver<()>,
+}
+
+/// Configuration for a `Runner`
+#[derive(Debug)]
+pub struct Config {
+    /// `sys_buffer_bytes` gives the estimated amount of memory, in bytes, that the kernel uses before
+    /// handing out the rest to userspace. This value is the estimated difference between the
+    /// *actual* physical memory and the amount reported by `grep MemTotal /proc/meminfo`.
+    ///
+    /// For more information, refer to `man 5 proc`, which defines MemTotal as "Total usable RAM
+    /// (i.e., physical RAM minus a few reserved bits and the kernel binary code)".
+    ///
+    /// We only use `sys_buffer_bytes` when calculating the system memory from the *external* memory
+    /// size, rather than the self-reported memory size, according to the kernel.
+    ///
+    /// TODO: this field is only necessary while we still have to trust the autoscaler-agent's
+    /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
+    /// should be removed once we have a better solution there.
+    sys_buffer_bytes: u64,
+}
+
+impl Default for Config {
+    fn default() -> Self {
+        Self {
+            sys_buffer_bytes: 100 * MiB,
+        }
+    }
+}
+
+impl Runner {
+    /// Create a new monitor.
+    #[tracing::instrument(skip_all, fields(?config, ?args))]
+    pub async fn new(
+        config: Config,
+        args: &Args,
+        ws: WebSocket,
+        kill: broadcast::Receiver<()>,
+        token: CancellationToken,
+    ) -> anyhow::Result<Runner> {
+        anyhow::ensure!(
+            config.sys_buffer_bytes != 0,
+            "invalid monitor Config: sys_buffer_bytes cannot be 0"
+        );
+
+        // *NOTE*: the dispatcher and cgroup manager talk through these channels
+        // so make sure they each get the correct half, nothing is droppped, etc.
+        let (notified_send, notified_recv) = mpsc::channel(1);
+        let (requesting_send, requesting_recv) = mpsc::channel(1);
+
+        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+            .await
+            .context("error creating new dispatcher")?;
+
+        let mut state = Runner {
+            config,
+            filecache: None,
+            cgroup: None,
+            dispatcher,
+            counter: 1, // NB: must be odd, see the comment about the field for more.
+            kill,
+        };
+
+        let mut file_cache_reserved_bytes = 0;
+        let mem = get_total_system_memory();
+
+        // We need to process file cache initialization before cgroup initialization, so that the memory
+        // allocated to the file cache is appropriately taken into account when we decide the cgroup's
+        // memory limits.
+        if let Some(connstr) = &args.pgconnstr {
+            info!("initializing file cache");
+            let config: FileCacheConfig = Default::default();
+            if !config.in_memory {
+                panic!("file cache not in-memory implemented")
+            }
+
+            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
+                .await
+                .context("failed to create file cache")?;
+
+            let size = file_cache
+                .get_file_cache_size()
+                .await
+                .context("error getting file cache size")?;
+
+            let new_size = file_cache.config.calculate_cache_size(mem);
+            info!(
+                initial = bytes_to_mebibytes(size),
+                new = bytes_to_mebibytes(new_size),
+                "setting initial file cache size",
+            );
+
+            // note: even if size == new_size, we want to explicitly set it, just
+            // to make sure that we have the permissions to do so
+            let actual_size = file_cache
+                .set_file_cache_size(new_size)
+                .await
+                .context("failed to set file cache size, possibly due to inadequate permissions")?;
+            if actual_size != new_size {
+                info!("file cache size actually got set to {actual_size}")
+            }
+            file_cache_reserved_bytes = actual_size;
+
+            state.filecache = Some(file_cache);
+        }
+
+        if let Some(name) = &args.cgroup {
+            let (mut cgroup, cgroup_event_stream) =
+                CgroupWatcher::new(name.clone(), requesting_send)
+                    .context("failed to create cgroup manager")?;
+
+            let available = mem - file_cache_reserved_bytes;
+
+            cgroup
+                .set_memory_limits(available)
+                .context("failed to set cgroup memory limits")?;
+
+            let cgroup = Arc::new(cgroup);
+
+            // Some might call this . . . cgroup v2
+            let cgroup_clone = Arc::clone(&cgroup);
+
+            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
+                cgroup_clone.watch(notified_recv, cgroup_event_stream).await
+            });
+
+            state.cgroup = Some(cgroup);
+        } else {
+            // *NOTE*: We need to forget the sender so that its drop impl does not get ran.
+            // This allows us to poll it in `Monitor::run` regardless of whether we
+            // are managing a cgroup or not. If we don't forget it, all receives will
+            // immediately return an error because the sender is droped and it will
+            // claim all select! statements, effectively turning `Monitor::run` into
+            // `loop { fail to receive }`.
+            mem::forget(requesting_send);
+        }
+
+        Ok(state)
+    }
+
+    /// Attempt to downscale filecache + cgroup
+    #[tracing::instrument(skip_all, fields(?target))]
+    pub async fn try_downscale(&mut self, target: Resources) -> anyhow::Result<(bool, String)> {
+        // Nothing to adjust
+        if self.cgroup.is_none() && self.filecache.is_none() {
+            info!("no action needed for downscale (no cgroup or file cache enabled)");
+            return Ok((
+                true,
+                "monitor is not managing cgroup or file cache".to_string(),
+            ));
+        }
+
+        let requested_mem = target.mem;
+        let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
+        let expected_file_cache_mem_usage = self
+            .filecache
+            .as_ref()
+            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
+            .unwrap_or(0);
+        let mut new_cgroup_mem_high = 0;
+        if let Some(cgroup) = &self.cgroup {
+            new_cgroup_mem_high = cgroup
+                .config
+                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+
+            let current = cgroup
+                .current_memory_usage()
+                .context("failed to fetch cgroup memory")?;
+
+            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+                let status = format!(
+                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
+                    "calculated memory.high too low",
+                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    bytes_to_mebibytes(current),
+                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                );
+
+                info!(status, "discontinuing downscale");
+
+                return Ok((false, status));
+            }
+        }
+
+        // The downscaling has been approved. Downscale the file cache, then the cgroup.
+        let mut status = vec![];
+        let mut file_cache_mem_usage = 0;
+        if let Some(file_cache) = &mut self.filecache {
+            if !file_cache.config.in_memory {
+                panic!("file cache not in-memory unimplemented")
+            }
+
+            let actual_usage = file_cache
+                .set_file_cache_size(expected_file_cache_mem_usage)
+                .await
+                .context("failed to set file cache size")?;
+            file_cache_mem_usage = actual_usage;
+            let message = format!(
+                "set file cache size to {} MiB",
+                bytes_to_mebibytes(actual_usage)
+            );
+            info!("downscale: {message}");
+            status.push(message);
+        }
+
+        if let Some(cgroup) = &self.cgroup {
+            let available_memory = usable_system_memory - file_cache_mem_usage;
+
+            if file_cache_mem_usage != expected_file_cache_mem_usage {
+                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+            }
+
+            let limits = MemoryLimits::new(
+                // new_cgroup_mem_high is initialized to 0 but it is guarancontextd to not be here
+                // since it is properly initialized in the previous cgroup if let block
+                new_cgroup_mem_high,
+                available_memory,
+            );
+            cgroup
+                .set_limits(&limits)
+                .context("failed to set cgroup memory limits")?;
+
+            let message = format!(
+                "set cgroup memory.high to {} MiB, of new max {} MiB",
+                bytes_to_mebibytes(new_cgroup_mem_high),
+                bytes_to_mebibytes(available_memory)
+            );
+            info!("downscale: {message}");
+            status.push(message);
+        }
+
+        // TODO: make this status thing less jank
+        let status = status.join("; ");
+        Ok((true, status))
+    }
+
+    /// Handle new resources
+    #[tracing::instrument(skip_all, fields(?resources))]
+    pub async fn handle_upscale(&mut self, resources: Resources) -> anyhow::Result<()> {
+        if self.filecache.is_none() && self.cgroup.is_none() {
+            info!("no action needed for upscale (no cgroup or file cache enabled)");
+            return Ok(());
+        }
+
+        let new_mem = resources.mem;
+        let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
+
+        // Get the file cache's expected contribution to the memory usage
+        let mut file_cache_mem_usage = 0;
+        if let Some(file_cache) = &mut self.filecache {
+            if !file_cache.config.in_memory {
+                panic!("file cache not in-memory unimplemented");
+            }
+
+            let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
+            info!(
+                target = bytes_to_mebibytes(expected_usage),
+                total = bytes_to_mebibytes(new_mem),
+                "updating file cache size",
+            );
+
+            let actual_usage = file_cache
+                .set_file_cache_size(expected_usage)
+                .await
+                .context("failed to set file cache size")?;
+
+            if actual_usage != expected_usage {
+                warn!(
+                    "file cache was set to a different size that we wanted: target = {} Mib, actual= {} Mib",
+                    bytes_to_mebibytes(expected_usage),
+                    bytes_to_mebibytes(actual_usage)
+                )
+            }
+            file_cache_mem_usage = actual_usage;
+        }
+
+        if let Some(cgroup) = &self.cgroup {
+            let available_memory = usable_system_memory - file_cache_mem_usage;
+            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+            info!(
+                target = bytes_to_mebibytes(new_cgroup_mem_high),
+                total = bytes_to_mebibytes(new_mem),
+                name = cgroup.path(),
+                "updating cgroup memory.high",
+            );
+            let limits = MemoryLimits::new(new_cgroup_mem_high, available_memory);
+            cgroup
+                .set_limits(&limits)
+                .context("failed to set file cache size")?;
+        }
+
+        Ok(())
+    }
+
+    /// Take in a message and perform some action, such as downscaling or upscaling,
+    /// and return a message to be send back.
+    #[tracing::instrument(skip_all, fields(%id, message = ?inner))]
+    pub async fn process_message(
+        &mut self,
+        InboundMsg { inner, id }: InboundMsg,
+    ) -> anyhow::Result<Option<OutboundMsg>> {
+        match inner {
+            InboundMsgKind::UpscaleNotification { granted } => {
+                self.handle_upscale(granted)
+                    .await
+                    .context("failed to handle upscale")?;
+                self.dispatcher
+                    .notify_upscale(Sequenced::new(granted))
+                    .await
+                    .context("failed to notify notify cgroup of upscale")?;
+                Ok(Some(OutboundMsg::new(
+                    OutboundMsgKind::UpscaleConfirmation {},
+                    id,
+                )))
+            }
+            InboundMsgKind::DownscaleRequest { target } => self
+                .try_downscale(target)
+                .await
+                .context("failed to downscale")
+                .map(|(ok, status)| {
+                    Some(OutboundMsg::new(
+                        OutboundMsgKind::DownscaleResult { ok, status },
+                        id,
+                    ))
+                }),
+            InboundMsgKind::InvalidMessage { error } => {
+                warn!(
+                    %error, id, "received notification of an invalid message we sent"
+                );
+                Ok(None)
+            }
+            InboundMsgKind::InternalError { error } => {
+                warn!(error, id, "informant experienced an internal error");
+                Ok(None)
+            }
+            InboundMsgKind::HealthCheck {} => {
+                Ok(Some(OutboundMsg::new(OutboundMsgKind::HealthCheck {}, id)))
+            }
+        }
+    }
+
+    // TODO: don't propagate errors, probably just warn!?
+    #[tracing::instrument(skip_all)]
+    pub async fn run(&mut self) -> anyhow::Result<()> {
+        info!("starting dispatcher");
+        loop {
+            tokio::select! {
+                signal = self.kill.recv() => {
+                    match signal {
+                        Ok(()) => return Ok(()),
+                        Err(e) => bail!("failed to receive kill signal: {e}")
+                    }
+                }
+                // we need to propagate an upscale request
+                request = self.dispatcher.request_upscale_events.recv() => {
+                    if request.is_none() {
+                        bail!("failed to listen for upscale event from cgroup")
+                    }
+                    info!("cgroup asking for upscale; forwarding request");
+                    self.counter += 2; // Increment, preserving parity (i.e. keep the
+                                       // counter odd). See the field comment for more.
+                    self.dispatcher
+                        .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
+                        .await
+                        .context("failed to send message")?;
+                }
+                // there is a message from the informant
+                msg = self.dispatcher.source.next() => {
+                    if let Some(msg) = msg {
+                        // Don't use 'message' as a key as the string also uses
+                        // that for its key
+                        info!(?msg, "received message");
+                        match msg {
+                            Ok(msg) => {
+                                let message: InboundMsg = match msg {
+                                    Message::Text(text) => {
+                                        serde_json::from_str(&text).context("failed to deserialize text message")?
+                                    }
+                                    other => {
+                                        warn!(
+                                            // Don't use 'message' as a key as the
+                                            // string also uses that for its key
+                                            msg = ?other,
+                                            "informant should only send text messages but received different type"
+                                        );
+                                        continue
+                                    },
+                                };
+
+                                let out = match self.process_message(message.clone()).await {
+                                    Ok(Some(out)) => out,
+                                    Ok(None) => continue,
+                                    Err(e) => {
+                                        let error = e.to_string();
+                                        warn!(?error, "error handling message");
+                                        OutboundMsg::new(
+                                            OutboundMsgKind::InternalError {
+                                                error
+                                            },
+                                            message.id
+                                        )
+                                    }
+                                };
+
+                                self.dispatcher
+                                    .send(out)
+                                    .await
+                                    .context("failed to send message")?;
+                            }
+                            Err(e) => warn!("{e}"),
+                        }
+                    } else {
+                        anyhow::bail!("dispatcher connection closed")
+                    }
+                }
+            }
+        }
+    }
+}

pageserver/Cargo.toml

@@ -51,6 +51,7 @@ serde.workspace = true
 serde_json = { workspace = true, features = ["raw_value"] }
 serde_with.workspace = true
 signal-hook.workspace = true
+smallvec = { workspace = true, features = ["write"] }
 svg_fmt.workspace = true
 sync_wrapper.workspace = true
 tokio-tar.workspace = true

pageserver/benches/bench_layer_map.rs

@@ -215,7 +215,6 @@ fn bench_sequential(c: &mut Criterion) {
             TimelineId::generate(),
             zero.add(10 * i32)..zero.add(10 * i32 + 1),
             Lsn(i),
-            false,
             0,
         );
         updates.insert_historic(layer);

pageserver/ctl/src/layer_map_analyzer.rs

@@ -10,7 +10,7 @@ use std::{fs, path::Path, str};
 
 use pageserver::page_cache::PAGE_SZ;
 use pageserver::repository::{Key, KEY_SIZE};
-use pageserver::tenant::block_io::{BlockReader, FileBlockReader};
+use pageserver::tenant::block_io::FileBlockReader;
 use pageserver::tenant::disk_btree::{DiskBtreeReader, VisitDirection};
 use pageserver::tenant::storage_layer::delta_layer::{Summary, DELTA_KEY_SIZE};
 use pageserver::tenant::storage_layer::range_overlaps;
@@ -99,6 +99,7 @@ async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
     let file = FileBlockReader::new(VirtualFile::open(path)?);
     let summary_blk = file.read_blk(0)?;
     let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+    drop(summary_blk); // so we don't borrow `file` for too long
     let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
         actual_summary.index_start_blk,
         actual_summary.index_root_blk,

pageserver/ctl/src/layers.rs

@@ -44,8 +44,6 @@ pub(crate) enum LayerCmd {
 }
 
 async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
-    use pageserver::tenant::block_io::BlockReader;
-
     let path = path.as_ref();
     virtual_file::init(10);
     page_cache::init(100);
@@ -70,7 +68,7 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
             },
         )
         .await?;
-    let cursor = BlockCursor::new(&file);
+    let cursor = BlockCursor::new_fileblockreader_virtual(&file);
     for (k, v) in all {
         let value = cursor.read_blob(v.pos()).await?;
         println!("key:{} value_len:{}", k, value.len());

pageserver/src/bin/pageserver.rs

@@ -6,11 +6,12 @@ use std::{env, ops::ControlFlow, path::Path, str::FromStr};
 
 use anyhow::{anyhow, Context};
 use clap::{Arg, ArgAction, Command};
-use fail::FailScenario;
+
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::WALRECEIVER_RUNTIME;
+use pageserver::tenant::TenantSharedResources;
 use remote_storage::GenericRemoteStorage;
 use tokio::time::Instant;
 use tracing::*;
@@ -121,7 +122,7 @@ fn main() -> anyhow::Result<()> {
     }
 
     // Initialize up failpoints support
-    let scenario = FailScenario::setup();
+    let scenario = pageserver::failpoint_support::init();
 
     // Basic initialization of things that don't change after startup
     virtual_file::init(conf.max_file_descriptors);
@@ -382,8 +383,10 @@ fn start_pageserver(
 
     BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
         conf,
-        broker_client.clone(),
-        remote_storage.clone(),
+        TenantSharedResources {
+            broker_client: broker_client.clone(),
+            remote_storage: remote_storage.clone(),
+        },
         order,
     ))?;
 

pageserver/src/failpoint_support.rs

@@ -0,0 +1,86 @@
+/// use with fail::cfg("$name", "return(2000)")
+///
+/// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the
+/// specified time (in milliseconds). The main difference is that we use async
+/// tokio sleep function. Another difference is that we print lines to the log,
+/// which can be useful in tests to check that the failpoint was hit.
+#[macro_export]
+macro_rules! __failpoint_sleep_millis_async {
+    ($name:literal) => {{
+        // If the failpoint is used with a "return" action, set should_sleep to the
+        // returned value (as string). Otherwise it's set to None.
+        let should_sleep = (|| {
+            ::fail::fail_point!($name, |x| x);
+            ::std::option::Option::None
+        })();
+
+        // Sleep if the action was a returned value
+        if let ::std::option::Option::Some(duration_str) = should_sleep {
+            $crate::failpoint_support::failpoint_sleep_helper($name, duration_str).await
+        }
+    }};
+}
+pub use __failpoint_sleep_millis_async as sleep_millis_async;
+
+// Helper function used by the macro. (A function has nicer scoping so we
+// don't need to decorate everything with "::")
+#[doc(hidden)]
+pub(crate) async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
+    let millis = duration_str.parse::<u64>().unwrap();
+    let d = std::time::Duration::from_millis(millis);
+
+    tracing::info!("failpoint {:?}: sleeping for {:?}", name, d);
+    tokio::time::sleep(d).await;
+    tracing::info!("failpoint {:?}: sleep done", name);
+}
+
+pub fn init() -> fail::FailScenario<'static> {
+    // The failpoints lib provides support for parsing the `FAILPOINTS` env var.
+    // We want non-default behavior for `exit`, though, so, we handle it separately.
+    //
+    // Format for FAILPOINTS is "name=actions" separated by ";".
+    let actions = std::env::var("FAILPOINTS");
+    if actions.is_ok() {
+        std::env::remove_var("FAILPOINTS");
+    } else {
+        // let the library handle non-utf8, or nothing for not present
+    }
+
+    let scenario = fail::FailScenario::setup();
+
+    if let Ok(val) = actions {
+        val.split(';')
+            .enumerate()
+            .map(|(i, s)| s.split_once('=').ok_or((i, s)))
+            .for_each(|res| {
+                let (name, actions) = match res {
+                    Ok(t) => t,
+                    Err((i, s)) => {
+                        panic!(
+                            "startup failpoints: missing action on the {}th failpoint; try `{s}=return`",
+                            i + 1,
+                        );
+                    }
+                };
+                if let Err(e) = apply_failpoint(name, actions) {
+                    panic!("startup failpoints: failed to apply failpoint {name}={actions}: {e}");
+                }
+            });
+    }
+
+    scenario
+}
+
+pub(crate) fn apply_failpoint(name: &str, actions: &str) -> Result<(), String> {
+    if actions == "exit" {
+        fail::cfg_callback(name, exit_failpoint)
+    } else {
+        fail::cfg(name, actions)
+    }
+}
+
+#[inline(never)]
+fn exit_failpoint() {
+    tracing::info!("Exit requested by failpoint");
+    std::process::exit(1);
+}

pageserver/src/http/routes.rs

@@ -517,7 +517,6 @@ async fn timeline_delete_handler(
         .instrument(info_span!("timeline_delete", %tenant_id, %timeline_id))
         .await?;
 
-    // FIXME: needs to be an error for console to retry it. Ideally Accepted should be used and retried until 404.
     json_response(StatusCode::ACCEPTED, ())
 }
 
@@ -980,14 +979,7 @@ async fn failpoints_handler(
 
         // We recognize one extra "action" that's not natively recognized
         // by the failpoints crate: exit, to immediately kill the process
-        let cfg_result = if fp.actions == "exit" {
-            fail::cfg_callback(fp.name, || {
-                info!("Exit requested by failpoint");
-                std::process::exit(1);
-            })
-        } else {
-            fail::cfg(fp.name, &fp.actions)
-        };
+        let cfg_result = crate::failpoint_support::apply_failpoint(&fp.name, &fp.actions);
 
         if let Err(err_msg) = cfg_result {
             return Err(ApiError::BadRequest(anyhow!(

pageserver/src/lib.rs

@@ -21,6 +21,8 @@ pub mod walingest;
 pub mod walrecord;
 pub mod walredo;
 
+pub mod failpoint_support;
+
 use std::path::Path;
 
 use crate::task_mgr::TaskKind;

pageserver/src/metrics.rs

@@ -6,7 +6,7 @@ use metrics::{
     HistogramVec, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
-use strum::VariantNames;
+use strum::{EnumCount, IntoEnumIterator, VariantNames};
 use strum_macros::{EnumVariantNames, IntoStaticStr};
 use utils::id::{TenantId, TimelineId};
 
@@ -570,23 +570,160 @@ pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-const SMGR_QUERY_TIME_OPERATIONS: &[&str] = &[
-    "get_rel_exists",
-    "get_rel_size",
-    "get_page_at_lsn",
-    "get_db_size",
-];
+#[derive(Debug)]
+struct GlobalAndPerTimelineHistogram {
+    global: Histogram,
+    per_tenant_timeline: Histogram,
+}
 
-pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+impl GlobalAndPerTimelineHistogram {
+    fn observe(&self, value: f64) {
+        self.global.observe(value);
+        self.per_tenant_timeline.observe(value);
+    }
+}
+
+struct GlobalAndPerTimelineHistogramTimer<'a> {
+    h: &'a GlobalAndPerTimelineHistogram,
+    start: std::time::Instant,
+}
+
+impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
+    fn drop(&mut self) {
+        let elapsed = self.start.elapsed();
+        self.h.observe(elapsed.as_secs_f64());
+    }
+}
+
+#[derive(
+    Debug,
+    Clone,
+    Copy,
+    IntoStaticStr,
+    strum_macros::EnumCount,
+    strum_macros::EnumIter,
+    strum_macros::FromRepr,
+)]
+#[strum(serialize_all = "snake_case")]
+pub enum SmgrQueryType {
+    GetRelExists,
+    GetRelSize,
+    GetPageAtLsn,
+    GetDbSize,
+}
+
+#[derive(Debug)]
+pub struct SmgrQueryTimePerTimeline {
+    metrics: [GlobalAndPerTimelineHistogram; SmgrQueryType::COUNT],
+}
+
+static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_smgr_query_seconds",
-        "Time spent on smgr query handling",
+        "Time spent on smgr query handling, aggegated by query type and tenant/timeline.",
         &["smgr_query_type", "tenant_id", "timeline_id"],
         CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
 
+static SMGR_QUERY_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
+    register_histogram_vec!(
+        "pageserver_smgr_query_seconds_global",
+        "Time spent on smgr query handling, aggregated by query type.",
+        &["smgr_query_type"],
+        CRITICAL_OP_BUCKETS.into(),
+    )
+    .expect("failed to define a metric")
+});
+
+impl SmgrQueryTimePerTimeline {
+    pub(crate) fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
+        let tenant_id = tenant_id.to_string();
+        let timeline_id = timeline_id.to_string();
+        let metrics = std::array::from_fn(|i| {
+            let op = SmgrQueryType::from_repr(i).unwrap();
+            let global = SMGR_QUERY_TIME_GLOBAL
+                .get_metric_with_label_values(&[op.into()])
+                .unwrap();
+            let per_tenant_timeline = SMGR_QUERY_TIME_PER_TENANT_TIMELINE
+                .get_metric_with_label_values(&[op.into(), &tenant_id, &timeline_id])
+                .unwrap();
+            GlobalAndPerTimelineHistogram {
+                global,
+                per_tenant_timeline,
+            }
+        });
+        Self { metrics }
+    }
+    pub(crate) fn start_timer(&self, op: SmgrQueryType) -> impl Drop + '_ {
+        let metric = &self.metrics[op as usize];
+        GlobalAndPerTimelineHistogramTimer {
+            h: metric,
+            start: std::time::Instant::now(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod smgr_query_time_tests {
+    use strum::IntoEnumIterator;
+    use utils::id::{TenantId, TimelineId};
+
+    // Regression test, we used hard-coded string constants before using an enum.
+    #[test]
+    fn op_label_name() {
+        use super::SmgrQueryType::*;
+        let expect: [(super::SmgrQueryType, &'static str); 4] = [
+            (GetRelExists, "get_rel_exists"),
+            (GetRelSize, "get_rel_size"),
+            (GetPageAtLsn, "get_page_at_lsn"),
+            (GetDbSize, "get_db_size"),
+        ];
+        for (op, expect) in expect {
+            let actual: &'static str = op.into();
+            assert_eq!(actual, expect);
+        }
+    }
+
+    #[test]
+    fn basic() {
+        let ops: Vec<_> = super::SmgrQueryType::iter().collect();
+
+        for op in &ops {
+            let tenant_id = TenantId::generate();
+            let timeline_id = TimelineId::generate();
+            let metrics = super::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);
+
+            let get_counts = || {
+                let global: u64 = ops
+                    .iter()
+                    .map(|op| metrics.metrics[*op as usize].global.get_sample_count())
+                    .sum();
+                let per_tenant_timeline: u64 = ops
+                    .iter()
+                    .map(|op| {
+                        metrics.metrics[*op as usize]
+                            .per_tenant_timeline
+                            .get_sample_count()
+                    })
+                    .sum();
+                (global, per_tenant_timeline)
+            };
+
+            let (pre_global, pre_per_tenant_timeline) = get_counts();
+            assert_eq!(pre_per_tenant_timeline, 0);
+
+            let timer = metrics.start_timer(*op);
+            drop(timer);
+
+            let (post_global, post_per_tenant_timeline) = get_counts();
+            assert_eq!(post_per_tenant_timeline, 1);
+            assert!(post_global > pre_global);
+        }
+    }
+}
+
 // keep in sync with control plane Go code so that we can validate
 // compute's basebackup_ms metric with our perspective in the context of SLI/SLO.
 static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
@@ -1045,6 +1182,12 @@ impl Drop for TimelineMetrics {
             .write()
             .unwrap()
             .remove(tenant_id, timeline_id);
+
+        // The following metrics are born outside of the TimelineMetrics lifecycle but still
+        // removed at the end of it. The idea is to have the metrics outlive the
+        // entity during which they're observed, e.g., the smgr metrics shall
+        // outlive an individual smgr connection, but not the timeline.
+
         for op in StorageTimeOperation::VARIANTS {
             let _ =
                 STORAGE_TIME_SUM_PER_TIMELINE.remove_label_values(&[op, tenant_id, timeline_id]);
@@ -1056,8 +1199,12 @@ impl Drop for TimelineMetrics {
             let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, timeline_id]);
         }
 
-        for op in SMGR_QUERY_TIME_OPERATIONS {
-            let _ = SMGR_QUERY_TIME.remove_label_values(&[op, tenant_id, timeline_id]);
+        for op in SmgrQueryType::iter() {
+            let _ = SMGR_QUERY_TIME_PER_TENANT_TIMELINE.remove_label_values(&[
+                op.into(),
+                tenant_id,
+                timeline_id,
+            ]);
         }
     }
 }

pageserver/src/page_cache.rs

@@ -10,6 +10,42 @@
 //! PostgreSQL buffer size, and a Slot struct for each buffer to contain
 //! information about what's stored in the buffer.
 //!
+//! # Types Of Pages
+//!
+//! [`PageCache`] only supports immutable pages.
+//! Hence there is no need to worry about coherency.
+//!
+//! Two types of pages are supported:
+//!
+//! * **Materialized pages**, filled & used by page reconstruction
+//! * **Immutable File pages**, filled & used by [`crate::tenant::block_io`] and [`crate::tenant::ephemeral_file`].
+//!
+//! Note that [`crate::tenant::ephemeral_file::EphemeralFile`] is generally mutable, but, it's append-only.
+//! It uses the page cache only for the blocks that are already fully written and immutable.
+//!
+//! # Filling The Page Cache
+//!
+//! Page cache maps from a cache key to a buffer slot.
+//! The cache key uniquely identifies the piece of data that is being cached.
+//!
+//! The cache key for **materialized pages** is  [`TenantId`], [`TimelineId`], [`Key`], and [`Lsn`].
+//! Use [`PageCache::memorize_materialized_page`] and [`PageCache::lookup_materialized_page`] for fill & access.
+//!
+//! The cache key for **immutable file** pages is [`FileId`] and a block number.
+//! Users of page cache that wish to page-cache an arbitrary (immutable!) on-disk file do the following:
+//! * Have a mechanism to deterministically associate the on-disk file with a [`FileId`].
+//! * Get a [`FileId`] using [`next_file_id`].
+//! * Use the mechanism to associate the on-disk file with the returned [`FileId`].
+//! * Use [`PageCache::read_immutable_buf`] to get a [`ReadBufResult`].
+//! * If the page was already cached, it'll be the [`ReadBufResult::Found`] variant that contains
+//!   a read guard for the page. Just use it.
+//! * If the page was not cached, it'll be the [`ReadBufResult::NotFound`] variant that contains
+//!   a write guard for the page. Fill the page with the contents of the on-disk file.
+//!   Then call [`PageWriteGuard::mark_valid`] to mark the page as valid.
+//!   Then try again to [`PageCache::read_immutable_buf`].
+//!   Unless there's high cache pressure, the page should now be cached.
+//!   (TODO: allow downgrading the write guard to a read guard to ensure forward progress.)
+//!
 //! # Locking
 //!
 //! There are two levels of locking involved: There's one lock for the "mapping"
@@ -40,20 +76,18 @@ use std::{
     collections::{hash_map::Entry, HashMap},
     convert::TryInto,
     sync::{
-        atomic::{AtomicU8, AtomicUsize, Ordering},
-        RwLock, RwLockReadGuard, RwLockWriteGuard, TryLockError,
+        atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
+        RwLock, RwLockReadGuard, RwLockWriteGuard, TryLockError, Arc, TryLockResult,
     },
 };
 
 use anyhow::Context;
 use once_cell::sync::OnceCell;
-use tracing::error;
 use utils::{
     id::{TenantId, TimelineId},
     lsn::Lsn,
 };
 
-use crate::tenant::{block_io, ephemeral_file, writeback_ephemeral_file};
 use crate::{metrics::PageCacheSizeMetrics, repository::Key};
 
 static PAGE_CACHE: OnceCell<PageCache> = OnceCell::new();
@@ -87,6 +121,17 @@ pub fn get() -> &'static PageCache {
 pub const PAGE_SZ: usize = postgres_ffi::BLCKSZ as usize;
 const MAX_USAGE_COUNT: u8 = 5;
 
+/// See module-level comment.
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
+pub struct FileId(u64);
+
+static NEXT_ID: AtomicU64 = AtomicU64::new(1);
+
+/// See module-level comment.
+pub fn next_file_id() -> FileId {
+    FileId(NEXT_ID.fetch_add(1, Ordering::Relaxed))
+}
+
 ///
 /// CacheKey uniquely identifies a "thing" to cache in the page cache.
 ///
@@ -97,12 +142,8 @@ enum CacheKey {
         hash_key: MaterializedPageHashKey,
         lsn: Lsn,
     },
-    EphemeralPage {
-        file_id: ephemeral_file::FileId,
-        blkno: u32,
-    },
     ImmutableFilePage {
-        file_id: block_io::FileId,
+        file_id: FileId,
         blkno: u32,
     },
 }
@@ -128,7 +169,6 @@ struct Slot {
 struct SlotInner {
     key: Option<CacheKey>,
     buf: &'static mut [u8; PAGE_SZ],
-    dirty: bool,
 }
 
 impl Slot {
@@ -165,6 +205,8 @@ impl Slot {
     }
 }
 
+type ImmutablePageMap = HashMap<FileId, Arc<RwLock<HashMap<u32, usize>>>>;
+
 pub struct PageCache {
     /// This contains the mapping from the cache key to buffer slot that currently
     /// contains the page, if any.
@@ -177,9 +219,7 @@ pub struct PageCache {
     /// can have a separate mapping map, next to this field.
     materialized_page_map: RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,
 
-    ephemeral_page_map: RwLock<HashMap<(ephemeral_file::FileId, u32), usize>>,
-
-    immutable_page_map: RwLock<HashMap<(block_io::FileId, u32), usize>>,
+    immutable_page_map: RwLock<ImmutablePageMap>,
 
     /// The actual buffers with their metadata.
     slots: Box<[Slot]>,
@@ -258,14 +298,6 @@ impl PageWriteGuard<'_> {
         );
         self.valid = true;
     }
-    pub fn mark_dirty(&mut self) {
-        // only ephemeral pages can be dirty ATM.
-        assert!(matches!(
-            self.inner.key,
-            Some(CacheKey::EphemeralPage { .. })
-        ));
-        self.inner.dirty = true;
-    }
 }
 
 impl Drop for PageWriteGuard<'_> {
@@ -280,7 +312,6 @@ impl Drop for PageWriteGuard<'_> {
             let self_key = self.inner.key.as_ref().unwrap();
             PAGE_CACHE.get().unwrap().remove_mapping(self_key);
             self.inner.key = None;
-            self.inner.dirty = false;
         }
     }
 }
@@ -388,79 +419,65 @@ impl PageCache {
         Ok(())
     }
 
-    // Section 1.2: Public interface functions for working with Ephemeral pages.
+    // Section 1.2: Public interface functions for working with immutable file pages.
 
-    pub fn read_ephemeral_buf(
-        &self,
-        file_id: ephemeral_file::FileId,
-        blkno: u32,
-    ) -> anyhow::Result<ReadBufResult> {
-        let mut cache_key = CacheKey::EphemeralPage { file_id, blkno };
-
-        self.lock_for_read(&mut cache_key)
-    }
-
-    pub fn write_ephemeral_buf(
-        &self,
-        file_id: ephemeral_file::FileId,
-        blkno: u32,
-    ) -> anyhow::Result<WriteBufResult> {
-        let cache_key = CacheKey::EphemeralPage { file_id, blkno };
-
-        self.lock_for_write(&cache_key)
-    }
-
-    /// Immediately drop all buffers belonging to given file, without writeback
-    pub fn drop_buffers_for_ephemeral(&self, drop_file_id: ephemeral_file::FileId) {
-        for slot_idx in 0..self.slots.len() {
-            let slot = &self.slots[slot_idx];
-
-            let mut inner = slot.inner.write().unwrap();
-            if let Some(key) = &inner.key {
-                match key {
-                    CacheKey::EphemeralPage { file_id, blkno: _ } if *file_id == drop_file_id => {
-                        // remove mapping for old buffer
-                        self.remove_mapping(key);
-                        inner.key = None;
-                        inner.dirty = false;
-                    }
-                    _ => {}
-                }
-            }
-        }
-    }
-
-    // Section 1.3: Public interface functions for working with immutable file pages.
-
-    pub fn read_immutable_buf(
-        &self,
-        file_id: block_io::FileId,
-        blkno: u32,
-    ) -> anyhow::Result<ReadBufResult> {
+    pub fn read_immutable_buf(&self, file_id: FileId, blkno: u32) -> anyhow::Result<ReadBufResult> {
         let mut cache_key = CacheKey::ImmutableFilePage { file_id, blkno };
 
         self.lock_for_read(&mut cache_key)
     }
 
-    /// Immediately drop all buffers belonging to given file, without writeback
-    pub fn drop_buffers_for_immutable(&self, drop_file_id: block_io::FileId) {
-        for slot_idx in 0..self.slots.len() {
-            let slot = &self.slots[slot_idx];
+    /// Immediately drop all buffers belonging to given file
+    pub fn drop_buffers_for_immutable(&self, drop_file_id: FileId) {
+        let map = self.immutable_page_map.read().unwrap();
+        let Some(block_nos_arc_rwl) = map.get(&drop_file_id).map(Arc::clone) else {
+            return
+        };
+        drop(map); // avoid contention on immutable_page_map
 
-            let mut inner = slot.inner.write().unwrap();
-            if let Some(key) = &inner.key {
-                match key {
-                    CacheKey::ImmutableFilePage { file_id, blkno: _ }
-                        if *file_id == drop_file_id =>
-                    {
-                        // remove mapping for old buffer
-                        self.remove_mapping(key);
+        let mut block_nos_guard = block_nos_arc_rwl.write().unwrap();
+        block_nos_guard.retain(|block_no, slot_idx| {
+            let expect_cache_key = CacheKey::ImmutableFilePage {
+                file_id: drop_file_id,
+                blkno: *block_no,
+            };
+            let slot = &self.slots[*slot_idx];
+            match slot.inner.try_write() {
+                TryLockResult::Ok(mut inner) => {
+                    // check again, could have been taken since added to immutable_page_map
+                    if inner.key == Some(expect_cache_key) {
+                        // immutable_page_map was still in sync with reality
+                        // TODO: find a way to share code with `remove_mapping`
+                        self.size_metrics.current_bytes_immutable.sub_page_sz(1);
                         inner.key = None;
-                        inner.dirty = false;
                     }
-                    _ => {}
+                    false
+                }
+                TryLockResult::Err(e @ TryLockError::Poisoned(_)) => {
+                    panic!("slot lock {slot_idx} poisoned: {e:?}")
+                }
+                TryLockResult::Err(TryLockError::WouldBlock) => {
+                    // This function is only called from the `Drop` impl of EphemeralFile.
+                    // So, there shouldn't be any page cache users that are reading from
+                    // that EphemeralFile anymore, because,
+                    // 1. EphemeralFile doesn't hand out page cache read guards that outlive EphemeralFile and
+                    // 2. there can't be any concurrent `EphemeralFile::reads`s because it's being dropped.
+                    //
+                    // So, the only reason why the page is locked right now is `find_victim`
+                    // trying to claim this page.
+                    // So, let `find_victim` do the work.
+                    // Also, it will call `remove_mapping()` to remove the (file_id, block_no)
+                    // from the immutable_page_map; so, keep  the entry in the map here, otherwise
+                    // the `remove_mapping()` call will panic.
+                    true
                 }
             }
+        });
+        if block_nos_guard.is_empty() {
+            drop(block_nos_guard); // remove_mapping_immutable_file_common re-aquires the lock
+            drop(block_nos_arc_rwl);
+            let mut map = self.immutable_page_map.write().unwrap();
+            self.remove_mapping_immutable_file_common(&mut map, drop_file_id)
         }
     }
 
@@ -534,10 +551,6 @@ impl PageCache {
             CacheKey::MaterializedPage { .. } => {
                 unreachable!("Materialized pages use lookup_materialized_page")
             }
-            CacheKey::EphemeralPage { .. } => (
-                &crate::metrics::PAGE_CACHE.read_accesses_ephemeral,
-                &crate::metrics::PAGE_CACHE.read_hits_ephemeral,
-            ),
             CacheKey::ImmutableFilePage { .. } => (
                 &crate::metrics::PAGE_CACHE.read_accesses_immutable,
                 &crate::metrics::PAGE_CACHE.read_hits_immutable,
@@ -578,7 +591,6 @@ impl PageCache {
             // Make the slot ready
             let slot = &self.slots[slot_idx];
             inner.key = Some(cache_key.clone());
-            inner.dirty = false;
             slot.usage_count.store(1, Ordering::Relaxed);
 
             return Ok(ReadBufResult::NotFound(PageWriteGuard {
@@ -640,7 +652,6 @@ impl PageCache {
             // Make the slot ready
             let slot = &self.slots[slot_idx];
             inner.key = Some(cache_key.clone());
-            inner.dirty = false;
             slot.usage_count.store(1, Ordering::Relaxed);
 
             return Ok(WriteBufResult::NotFound(PageWriteGuard {
@@ -679,13 +690,11 @@ impl PageCache {
                 *lsn = version.lsn;
                 Some(version.slot_idx)
             }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let map = self.ephemeral_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
-            }
             CacheKey::ImmutableFilePage { file_id, blkno } => {
                 let map = self.immutable_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
+                let block_nos = map.get(file_id)?;
+                let block_nos = block_nos.read().unwrap();
+                block_nos.get(blkno).copied()
             }
         }
     }
@@ -706,13 +715,11 @@ impl PageCache {
                     None
                 }
             }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let map = self.ephemeral_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
-            }
             CacheKey::ImmutableFilePage { file_id, blkno } => {
                 let map = self.immutable_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
+                let block_nos = map.get(file_id)?;
+                let block_nos = block_nos.read().unwrap();
+                block_nos.get(blkno).copied()
             }
         }
     }
@@ -743,21 +750,40 @@ impl PageCache {
                     panic!("could not find old key in mapping")
                 }
             }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let mut map = self.ephemeral_page_map.write().unwrap();
-                map.remove(&(*file_id, *blkno))
-                    .expect("could not find old key in mapping");
-                self.size_metrics.current_bytes_ephemeral.sub_page_sz(1);
-            }
             CacheKey::ImmutableFilePage { file_id, blkno } => {
-                let mut map = self.immutable_page_map.write().unwrap();
-                map.remove(&(*file_id, *blkno))
-                    .expect("could not find old key in mapping");
+                let map = self.immutable_page_map.read().unwrap();
+                let block_nos = map.get(file_id).expect("could not find file_id in mapping");
+                let mut block_nos = block_nos.write().unwrap();
+                block_nos
+                    .remove(blkno)
+                    .expect("could not find blkno in mapping");
                 self.size_metrics.current_bytes_immutable.sub_page_sz(1);
+                if block_nos.is_empty() {
+                    drop(block_nos);
+                    // re-lock map in write mode
+                    drop(map);
+                    let mut map = self.immutable_page_map.write().unwrap();
+                    self.remove_mapping_immutable_file_common(&mut map, *file_id)
+                }
             }
         }
     }
 
+    fn remove_mapping_immutable_file_common(
+        &self,
+        map: &mut RwLockWriteGuard<ImmutablePageMap>,
+        file_id: FileId,
+    ) {
+        let Some(block_nos_rwl) = map.get(&file_id) else {
+            return;
+        };
+        let block_nos = block_nos_rwl.read().unwrap();
+        if block_nos.is_empty() {
+            drop(block_nos);
+            map.remove(&file_id);
+        }
+    }
+
     ///
     /// Insert mapping for given key.
     ///
@@ -788,20 +814,12 @@ impl PageCache {
                     }
                 }
             }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let mut map = self.ephemeral_page_map.write().unwrap();
-                match map.entry((*file_id, *blkno)) {
-                    Entry::Occupied(entry) => Some(*entry.get()),
-                    Entry::Vacant(entry) => {
-                        entry.insert(slot_idx);
-                        self.size_metrics.current_bytes_ephemeral.add_page_sz(1);
-                        None
-                    }
-                }
-            }
+
             CacheKey::ImmutableFilePage { file_id, blkno } => {
                 let mut map = self.immutable_page_map.write().unwrap();
-                match map.entry((*file_id, *blkno)) {
+                let block_nos = map.entry(*file_id).or_default();
+                let mut block_nos = block_nos.write().unwrap();
+                match block_nos.entry(*blkno) {
                     Entry::Occupied(entry) => Some(*entry.get()),
                     Entry::Vacant(entry) => {
                         entry.insert(slot_idx);
@@ -849,25 +867,8 @@ impl PageCache {
                     }
                 };
                 if let Some(old_key) = &inner.key {
-                    if inner.dirty {
-                        if let Err(err) = Self::writeback(old_key, inner.buf) {
-                            // Writing the page to disk failed.
-                            //
-                            // FIXME: What to do here, when? We could propagate the error to the
-                            // caller, but victim buffer is generally unrelated to the original
-                            // call. It can even belong to a different tenant. Currently, we
-                            // report the error to the log and continue the clock sweep to find
-                            // a different victim. But if the problem persists, the page cache
-                            // could fill up with dirty pages that we cannot evict, and we will
-                            // loop retrying the writebacks indefinitely.
-                            error!("writeback of buffer {:?} failed: {}", old_key, err);
-                            continue;
-                        }
-                    }
-
                     // remove mapping for old buffer
                     self.remove_mapping(old_key);
-                    inner.dirty = false;
                     inner.key = None;
                 }
                 return Ok((slot_idx, inner));
@@ -875,28 +876,6 @@ impl PageCache {
         }
     }
 
-    fn writeback(cache_key: &CacheKey, buf: &[u8]) -> Result<(), std::io::Error> {
-        match cache_key {
-            CacheKey::MaterializedPage {
-                hash_key: _,
-                lsn: _,
-            } => Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
-                "unexpected dirty materialized page",
-            )),
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                writeback_ephemeral_file(*file_id, *blkno, buf)
-            }
-            CacheKey::ImmutableFilePage {
-                file_id: _,
-                blkno: _,
-            } => Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
-                "unexpected dirty immutable page",
-            )),
-        }
-    }
-
     /// Initialize a new page cache
     ///
     /// This should be called only once at page server startup.
@@ -907,7 +886,6 @@ impl PageCache {
 
         let size_metrics = &crate::metrics::PAGE_CACHE_SIZE;
         size_metrics.max_bytes.set_page_sz(num_pages);
-        size_metrics.current_bytes_ephemeral.set_page_sz(0);
         size_metrics.current_bytes_immutable.set_page_sz(0);
         size_metrics.current_bytes_materialized_page.set_page_sz(0);
 
@@ -917,11 +895,7 @@ impl PageCache {
                 let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();
 
                 Slot {
-                    inner: RwLock::new(SlotInner {
-                        key: None,
-                        buf,
-                        dirty: false,
-                    }),
+                    inner: RwLock::new(SlotInner { key: None, buf }),
                     usage_count: AtomicU8::new(0),
                 }
             })
@@ -929,7 +903,6 @@ impl PageCache {
 
         Self {
             materialized_page_map: Default::default(),
-            ephemeral_page_map: Default::default(),
             immutable_page_map: Default::default(),
             slots,
             next_evict_slot: AtomicUsize::new(0),

pageserver/src/page_service.rs

@@ -50,7 +50,8 @@ use crate::basebackup;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir::import_wal_from_tar;
-use crate::metrics::{LIVE_CONNECTIONS_COUNT, SMGR_QUERY_TIME};
+use crate::metrics;
+use crate::metrics::LIVE_CONNECTIONS_COUNT;
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant;
@@ -306,39 +307,6 @@ async fn page_service_conn_main(
     }
 }
 
-struct PageRequestMetrics {
-    get_rel_exists: metrics::Histogram,
-    get_rel_size: metrics::Histogram,
-    get_page_at_lsn: metrics::Histogram,
-    get_db_size: metrics::Histogram,
-}
-
-impl PageRequestMetrics {
-    fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
-        let tenant_id = tenant_id.to_string();
-        let timeline_id = timeline_id.to_string();
-
-        let get_rel_exists =
-            SMGR_QUERY_TIME.with_label_values(&["get_rel_exists", &tenant_id, &timeline_id]);
-
-        let get_rel_size =
-            SMGR_QUERY_TIME.with_label_values(&["get_rel_size", &tenant_id, &timeline_id]);
-
-        let get_page_at_lsn =
-            SMGR_QUERY_TIME.with_label_values(&["get_page_at_lsn", &tenant_id, &timeline_id]);
-
-        let get_db_size =
-            SMGR_QUERY_TIME.with_label_values(&["get_db_size", &tenant_id, &timeline_id]);
-
-        Self {
-            get_rel_exists,
-            get_rel_size,
-            get_page_at_lsn,
-            get_db_size,
-        }
-    }
-}
-
 struct PageServerHandler {
     _conf: &'static PageServerConf,
     broker_client: storage_broker::BrokerClientChannel,
@@ -406,7 +374,7 @@ impl PageServerHandler {
         pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
         pgb.flush().await?;
 
-        let metrics = PageRequestMetrics::new(&tenant_id, &timeline_id);
+        let metrics = metrics::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);
 
         loop {
             let msg = tokio::select! {
@@ -446,21 +414,21 @@ impl PageServerHandler {
 
             let response = match neon_fe_msg {
                 PagestreamFeMessage::Exists(req) => {
-                    let _timer = metrics.get_rel_exists.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetRelExists);
                     self.handle_get_rel_exists_request(&timeline, &req, &ctx)
                         .await
                 }
                 PagestreamFeMessage::Nblocks(req) => {
-                    let _timer = metrics.get_rel_size.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetRelSize);
                     self.handle_get_nblocks_request(&timeline, &req, &ctx).await
                 }
                 PagestreamFeMessage::GetPage(req) => {
-                    let _timer = metrics.get_page_at_lsn.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetPageAtLsn);
                     self.handle_get_page_at_lsn_request(&timeline, &req, &ctx)
                         .await
                 }
                 PagestreamFeMessage::DbSize(req) => {
-                    let _timer = metrics.get_db_size.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetDbSize);
                     self.handle_db_size_request(&timeline, &req, &ctx).await
                 }
             };
@@ -984,8 +952,8 @@ where
                 false
             };
 
-            metrics::metric_vec_duration::observe_async_block_duration_by_result(
-                &*crate::metrics::BASEBACKUP_QUERY_TIME,
+            ::metrics::metric_vec_duration::observe_async_block_duration_by_result(
+                &*metrics::BASEBACKUP_QUERY_TIME,
                 async move {
                     self.handle_basebackup_request(
                         pgb,

pageserver/src/tenant.rs

@@ -29,6 +29,7 @@ use std::collections::hash_map::Entry;
 use std::collections::BTreeSet;
 use std::collections::HashMap;
 use std::fmt::Debug;
+use std::fmt::Display;
 use std::fs;
 use std::fs::File;
 use std::fs::OpenOptions;
@@ -55,6 +56,7 @@ use self::remote_timeline_client::RemoteTimelineClient;
 use self::timeline::uninit::TimelineUninitMark;
 use self::timeline::uninit::UninitializedTimeline;
 use self::timeline::EvictionTaskTenantState;
+use self::timeline::TimelineResources;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir;
@@ -70,7 +72,6 @@ use crate::tenant::remote_timeline_client::index::IndexPart;
 use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
 use crate::tenant::storage_layer::DeltaLayer;
 use crate::tenant::storage_layer::ImageLayer;
-use crate::tenant::storage_layer::Layer;
 use crate::InitializationOrder;
 
 use crate::tenant::timeline::delete::DeleteTimelineFlow;
@@ -136,9 +137,6 @@ pub use timeline::{
     LocalLayerInfoForDiskUsageEviction, LogicalSizeCalculationCause, PageReconstructError, Timeline,
 };
 
-// re-export this function so that page_cache.rs can use it.
-pub use crate::tenant::ephemeral_file::writeback as writeback_ephemeral_file;
-
 // re-export for use in remote_timeline_client.rs
 pub use crate::tenant::metadata::save_metadata;
 
@@ -152,6 +150,14 @@ pub const TENANT_ATTACHING_MARKER_FILENAME: &str = "attaching";
 
 pub const TENANT_DELETED_MARKER_FILE_NAME: &str = "deleted";
 
+/// References to shared objects that are passed into each tenant, such
+/// as the shared remote storage client and process initialization state.
+#[derive(Clone)]
+pub struct TenantSharedResources {
+    pub broker_client: storage_broker::BrokerClientChannel,
+    pub remote_storage: Option<GenericRemoteStorage>,
+}
+
 ///
 /// Tenant consists of multiple timelines. Keep them in a hash table.
 ///
@@ -391,7 +397,7 @@ impl Tenant {
     async fn timeline_init_and_sync(
         &self,
         timeline_id: TimelineId,
-        remote_client: Option<RemoteTimelineClient>,
+        resources: TimelineResources,
         remote_startup_data: Option<RemoteStartupData>,
         local_metadata: Option<TimelineMetadata>,
         ancestor: Option<Arc<Timeline>>,
@@ -412,17 +418,57 @@ impl Tenant {
             timeline_id,
             up_to_date_metadata,
             ancestor.clone(),
-            remote_client,
+            resources,
             init_order,
             CreateTimelineCause::Load,
         )?;
-        let new_disk_consistent_lsn = timeline.get_disk_consistent_lsn();
+        let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
         anyhow::ensure!(
-            new_disk_consistent_lsn.is_valid(),
+            disk_consistent_lsn.is_valid(),
             "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
         );
+        assert_eq!(
+            disk_consistent_lsn,
+            up_to_date_metadata.disk_consistent_lsn(),
+            "these are used interchangeably"
+        );
+
+        // Save the metadata file to local disk.
+        if !picked_local {
+            save_metadata(
+                self.conf,
+                &tenant_id,
+                &timeline_id,
+                up_to_date_metadata,
+                first_save,
+            )
+            .context("save_metadata")?;
+        }
+
+        let index_part = remote_startup_data.as_ref().map(|x| &x.index_part);
+
+        if let Some(index_part) = index_part {
+            timeline
+                .remote_client
+                .as_ref()
+                .unwrap()
+                .init_upload_queue(index_part)?;
+        } else if self.remote_storage.is_some() {
+            // No data on the remote storage, but we have local metadata file. We can end up
+            // here with timeline_create being interrupted before finishing index part upload.
+            // By doing what we do here, the index part upload is retried.
+            // If control plane retries timeline creation in the meantime, the mgmt API handler
+            // for timeline creation will coalesce on the upload we queue here.
+            let rtc = timeline.remote_client.as_ref().unwrap();
+            rtc.init_upload_queue_for_empty_remote(up_to_date_metadata)?;
+            rtc.schedule_index_upload_for_metadata_update(up_to_date_metadata)?;
+        }
+
         timeline
-            .load_layer_map(new_disk_consistent_lsn)
+            .load_layer_map(
+                disk_consistent_lsn,
+                remote_startup_data.map(|x| x.index_part),
+            )
             .await
             .with_context(|| {
                 format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
@@ -446,19 +492,6 @@ impl Tenant {
             }
         };
 
-        if self.remote_storage.is_some() {
-            // Reconcile local state with remote storage, downloading anything that's
-            // missing locally, and scheduling uploads for anything that's missing
-            // in remote storage.
-            timeline
-                .reconcile_with_remote(
-                    up_to_date_metadata,
-                    remote_startup_data.as_ref().map(|r| &r.index_part),
-                )
-                .await
-                .context("failed to reconcile with remote")?
-        }
-
         // Sanity check: a timeline should have some content.
         anyhow::ensure!(
             ancestor.is_some()
@@ -473,18 +506,6 @@ impl Tenant {
             "Timeline has no ancestor and no layer files"
         );
 
-        // Save the metadata file to local disk.
-        if !picked_local {
-            save_metadata(
-                self.conf,
-                &tenant_id,
-                &timeline_id,
-                up_to_date_metadata,
-                first_save,
-            )
-            .context("save_metadata")?;
-        }
-
         Ok(())
     }
 
@@ -502,6 +523,7 @@ impl Tenant {
         conf: &'static PageServerConf,
         tenant_id: TenantId,
         broker_client: storage_broker::BrokerClientChannel,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
         remote_storage: GenericRemoteStorage,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Tenant>> {
@@ -516,7 +538,7 @@ impl Tenant {
             tenant_conf,
             wal_redo_manager,
             tenant_id,
-            Some(remote_storage),
+            Some(remote_storage.clone()),
         ));
 
         // Do all the hard work in the background
@@ -531,17 +553,61 @@ impl Tenant {
             "attach tenant",
             false,
             async move {
+                // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
+                let make_broken = |t: &Tenant, err: anyhow::Error| {
+                    error!("attach failed, setting tenant state to Broken: {err:?}");
+                    t.state.send_modify(|state| {
+                        assert_eq!(
+                            *state,
+                            TenantState::Attaching,
+                            "the attach task owns the tenant state until activation is complete"
+                        );
+                        *state = TenantState::broken_from_reason(err.to_string());
+                    });
+                };
+
+                let pending_deletion = {
+                    match DeleteTenantFlow::should_resume_deletion(
+                        conf,
+                        Some(&remote_storage),
+                        &tenant_clone,
+                    )
+                    .await
+                    {
+                        Ok(should_resume_deletion) => should_resume_deletion,
+                        Err(err) => {
+                            make_broken(&tenant_clone, anyhow::anyhow!(err));
+                            return Ok(());
+                        }
+                    }
+                };
+
+                info!("pending_deletion {}", pending_deletion.is_some());
+
+                if let Some(deletion) = pending_deletion {
+                    match DeleteTenantFlow::resume_from_attach(
+                        deletion,
+                        &tenant_clone,
+                        tenants,
+                        &ctx,
+                    )
+                    .await
+                    {
+                        Err(err) => {
+                            make_broken(&tenant_clone, anyhow::anyhow!(err));
+                            return Ok(());
+                        }
+                        Ok(()) => return Ok(()),
+                    }
+                }
+
                 match tenant_clone.attach(&ctx).await {
                     Ok(()) => {
                         info!("attach finished, activating");
                         tenant_clone.activate(broker_client, None, &ctx);
                     }
                     Err(e) => {
-                        error!("attach failed, setting tenant state to Broken: {:?}", e);
-                        tenant_clone.state.send_modify(|state| {
-                            assert_eq!(*state, TenantState::Attaching, "the attach task owns the tenant state until activation is complete");
-                            *state = TenantState::broken_from_reason(e.to_string());
-                        });
+                        make_broken(&tenant_clone, anyhow::anyhow!(e));
                     }
                 }
                 Ok(())
@@ -619,6 +685,9 @@ impl Tenant {
                 .instrument(info_span!("download_index_part", %timeline_id)),
             );
         }
+
+        let mut timelines_to_resume_deletions = vec![];
+
         // Wait for all the download tasks to complete & collect results.
         let mut remote_index_and_client = HashMap::new();
         let mut timeline_ancestors = HashMap::new();
@@ -629,15 +698,15 @@ impl Tenant {
             debug!("successfully downloaded index part for timeline {timeline_id}");
             match index_part {
                 MaybeDeletedIndexPart::IndexPart(index_part) => {
-                    timeline_ancestors.insert(
-                        timeline_id,
-                        index_part.parse_metadata().context("parse_metadata")?,
-                    );
+                    timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
                     remote_index_and_client.insert(timeline_id, (index_part, client));
                 }
-                MaybeDeletedIndexPart::Deleted(_) => {
-                    info!("timeline {} is deleted, skipping", timeline_id);
-                    continue;
+                MaybeDeletedIndexPart::Deleted(index_part) => {
+                    info!(
+                        "timeline {} is deleted, picking to resume deletion",
+                        timeline_id
+                    );
+                    timelines_to_resume_deletions.push((timeline_id, index_part, client));
                 }
             }
         }
@@ -652,14 +721,41 @@ impl Tenant {
                 .expect("just put it in above");
 
             // TODO again handle early failure
-            self.load_remote_timeline(timeline_id, index_part, remote_metadata, remote_client, ctx)
-                .await
-                .with_context(|| {
-                    format!(
-                        "failed to load remote timeline {} for tenant {}",
-                        timeline_id, self.tenant_id
-                    )
-                })?;
+            self.load_remote_timeline(
+                timeline_id,
+                index_part,
+                remote_metadata,
+                TimelineResources {
+                    remote_client: Some(remote_client),
+                },
+                ctx,
+            )
+            .await
+            .with_context(|| {
+                format!(
+                    "failed to load remote timeline {} for tenant {}",
+                    timeline_id, self.tenant_id
+                )
+            })?;
+        }
+
+        // Walk through deleted timelines, resume deletion
+        for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
+            remote_timeline_client
+                .init_upload_queue_stopped_to_continue_deletion(&index_part)
+                .context("init queue stopped")
+                .map_err(LoadLocalTimelineError::ResumeDeletion)?;
+
+            DeleteTimelineFlow::resume_deletion(
+                Arc::clone(self),
+                timeline_id,
+                &index_part.metadata,
+                Some(remote_timeline_client),
+                None,
+            )
+            .await
+            .context("resume_deletion")
+            .map_err(LoadLocalTimelineError::ResumeDeletion)?;
         }
 
         std::fs::remove_file(&marker_file)
@@ -667,7 +763,7 @@ impl Tenant {
         crashsafe::fsync(marker_file.parent().expect("marker file has parent dir"))
             .context("fsync tenant directory after unlinking attach marker file")?;
 
-        utils::failpoint_sleep_millis_async!("attach-before-activate");
+        crate::failpoint_support::sleep_millis_async!("attach-before-activate");
 
         info!("Done");
 
@@ -695,7 +791,7 @@ impl Tenant {
         timeline_id: TimelineId,
         index_part: IndexPart,
         remote_metadata: TimelineMetadata,
-        remote_client: RemoteTimelineClient,
+        resources: TimelineResources,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         span::debug_assert_current_span_has_tenant_id();
@@ -725,7 +821,7 @@ impl Tenant {
 
         self.timeline_init_and_sync(
             timeline_id,
-            Some(remote_client),
+            resources,
             Some(RemoteStartupData {
                 index_part,
                 remote_metadata,
@@ -772,8 +868,7 @@ impl Tenant {
     pub(crate) fn spawn_load(
         conf: &'static PageServerConf,
         tenant_id: TenantId,
-        broker_client: storage_broker::BrokerClientChannel,
-        remote_storage: Option<GenericRemoteStorage>,
+        resources: TenantSharedResources,
         init_order: Option<InitializationOrder>,
         tenants: &'static tokio::sync::RwLock<TenantsMap>,
         ctx: &RequestContext,
@@ -788,6 +883,9 @@ impl Tenant {
             }
         };
 
+        let broker_client = resources.broker_client;
+        let remote_storage = resources.remote_storage;
+
         let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
         let tenant = Tenant::new(
             TenantState::Loading,
@@ -811,6 +909,7 @@ impl Tenant {
             "initial tenant load",
             false,
             async move {
+                // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
                 let make_broken = |t: &Tenant, err: anyhow::Error| {
                     error!("load failed, setting tenant state to Broken: {err:?}");
                     t.state.send_modify(|state| {
@@ -858,7 +957,7 @@ impl Tenant {
                         .as_mut()
                         .and_then(|x| x.initial_logical_size_attempt.take());
 
-                    match DeleteTenantFlow::resume(
+                    match DeleteTenantFlow::resume_from_load(
                         deletion,
                         &tenant_clone,
                         init_order.as_ref(),
@@ -880,7 +979,7 @@ impl Tenant {
 
                 match tenant_clone.load(init_order.as_ref(), &ctx).await {
                     Ok(()) => {
-                        debug!("load finished",);
+                        debug!("load finished");
 
                         tenant_clone.activate(broker_client, background_jobs_can_start, &ctx);
                     }
@@ -1076,7 +1175,7 @@ impl Tenant {
 
         debug!("loading tenant task");
 
-        utils::failpoint_sleep_millis_async!("before-loading-tenant");
+        crate::failpoint_support::sleep_millis_async!("before-loading-tenant");
 
         // Load in-memory state to reflect the local files on disk
         //
@@ -1172,16 +1271,9 @@ impl Tenant {
     ) -> Result<(), LoadLocalTimelineError> {
         span::debug_assert_current_span_has_tenant_id();
 
-        let remote_client = self.remote_storage.as_ref().map(|remote_storage| {
-            RemoteTimelineClient::new(
-                remote_storage.clone(),
-                self.conf,
-                self.tenant_id,
-                timeline_id,
-            )
-        });
+        let mut resources = self.build_timeline_resources(timeline_id);
 
-        let (remote_startup_data, remote_client) = match remote_client {
+        let (remote_startup_data, remote_client) = match resources.remote_client {
             Some(remote_client) => match remote_client.download_index_file().await {
                 Ok(index_part) => {
                     let index_part = match index_part {
@@ -1219,10 +1311,7 @@ impl Tenant {
                         }
                     };
 
-                    let remote_metadata = index_part
-                        .parse_metadata()
-                        .context("parse_metadata")
-                        .map_err(LoadLocalTimelineError::Load)?;
+                    let remote_metadata = index_part.metadata.clone();
                     (
                         Some(RemoteStartupData {
                             index_part,
@@ -1269,9 +1358,10 @@ impl Tenant {
                     return Ok(());
                 }
 
-                (None, remote_client)
+                (None, resources.remote_client)
             }
         };
+        resources.remote_client = remote_client;
 
         let ancestor = if let Some(ancestor_timeline_id) = local_metadata.ancestor_timeline() {
             let ancestor_timeline = self.get_timeline(ancestor_timeline_id, false)
@@ -1284,7 +1374,7 @@ impl Tenant {
 
         self.timeline_init_and_sync(
             timeline_id,
-            remote_client,
+            resources,
             remote_startup_data,
             Some(local_metadata),
             ancestor,
@@ -1773,7 +1863,7 @@ impl Tenant {
         // It's mesed up.
         // we just ignore the failure to stop
 
-        match self.set_stopping(shutdown_progress, false).await {
+        match self.set_stopping(shutdown_progress, false, false).await {
             Ok(()) => {}
             Err(SetStoppingError::Broken) => {
                 // assume that this is acceptable
@@ -1815,15 +1905,18 @@ impl Tenant {
     /// This function is not cancel-safe!
     ///
     /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
+    /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
     async fn set_stopping(
         &self,
         progress: completion::Barrier,
         allow_transition_from_loading: bool,
+        allow_transition_from_attaching: bool,
     ) -> Result<(), SetStoppingError> {
         let mut rx = self.state.subscribe();
 
         // cannot stop before we're done activating, so wait out until we're done activating
         rx.wait_for(|state| match state {
+            TenantState::Attaching if allow_transition_from_attaching => true,
             TenantState::Activating(_) | TenantState::Attaching => {
                 info!(
                     "waiting for {} to turn Active|Broken|Stopping",
@@ -1840,12 +1933,19 @@ impl Tenant {
         // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
         let mut err = None;
         let stopping = self.state.send_if_modified(|current_state| match current_state {
-            TenantState::Activating(_) | TenantState::Attaching => {
-                unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
+            TenantState::Activating(_) => {
+                unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
+            }
+            TenantState::Attaching => {
+                if !allow_transition_from_attaching {
+                    unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
+                };
+                *current_state = TenantState::Stopping { progress };
+                true
             }
             TenantState::Loading => {
                 if !allow_transition_from_loading {
-                    unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
+                    unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
                 };
                 *current_state = TenantState::Stopping { progress };
                 true
@@ -1921,7 +2021,8 @@ impl Tenant {
         self.set_broken_no_wait(reason)
     }
 
-    pub(crate) fn set_broken_no_wait(&self, reason: String) {
+    pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
+        let reason = reason.to_string();
         self.state.send_modify(|current_state| {
             match *current_state {
                 TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
@@ -2145,7 +2246,7 @@ impl Tenant {
         new_timeline_id: TimelineId,
         new_metadata: &TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
-        remote_client: Option<RemoteTimelineClient>,
+        resources: TimelineResources,
         init_order: Option<&InitializationOrder>,
         cause: CreateTimelineCause,
     ) -> anyhow::Result<Arc<Timeline>> {
@@ -2174,7 +2275,7 @@ impl Tenant {
             new_timeline_id,
             self.tenant_id,
             Arc::clone(&self.walredo_mgr),
-            remote_client,
+            resources,
             pg_version,
             initial_logical_size_can_start.cloned(),
             initial_logical_size_attempt.cloned().flatten(),
@@ -2416,7 +2517,9 @@ impl Tenant {
             .refresh_gc_info_internal(target_timeline_id, horizon, pitr, ctx)
             .await?;
 
-        utils::failpoint_sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
+        crate::failpoint_support::sleep_millis_async!(
+            "gc_iteration_internal_after_getting_gc_timelines"
+        );
 
         // If there is nothing to GC, we don't want any messages in the INFO log.
         if !gc_timelines.is_empty() {
@@ -2820,6 +2923,23 @@ impl Tenant {
         Ok(timeline)
     }
 
+    /// Call this before constructing a timeline, to build its required structures
+    fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
+        let remote_client = if let Some(remote_storage) = self.remote_storage.as_ref() {
+            let remote_client = RemoteTimelineClient::new(
+                remote_storage.clone(),
+                self.conf,
+                self.tenant_id,
+                timeline_id,
+            );
+            Some(remote_client)
+        } else {
+            None
+        };
+
+        TimelineResources { remote_client }
+    }
+
     /// Creates intermediate timeline structure and its files.
     ///
     /// An empty layer map is initialized, and new data and WAL can be imported starting
@@ -2836,25 +2956,17 @@ impl Tenant {
     ) -> anyhow::Result<UninitializedTimeline> {
         let tenant_id = self.tenant_id;
 
-        let remote_client = if let Some(remote_storage) = self.remote_storage.as_ref() {
-            let remote_client = RemoteTimelineClient::new(
-                remote_storage.clone(),
-                self.conf,
-                tenant_id,
-                new_timeline_id,
-            );
+        let resources = self.build_timeline_resources(new_timeline_id);
+        if let Some(remote_client) = &resources.remote_client {
             remote_client.init_upload_queue_for_empty_remote(new_metadata)?;
-            Some(remote_client)
-        } else {
-            None
-        };
+        }
 
         let timeline_struct = self
             .create_timeline_struct(
                 new_timeline_id,
                 new_metadata,
                 ancestor,
-                remote_client,
+                resources,
                 None,
                 CreateTimelineCause::Load,
             )
@@ -3927,6 +4039,31 @@ mod tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn delta_layer_dumping() -> anyhow::Result<()> {
+        let (tenant, ctx) = TenantHarness::create("test_layer_dumping")?.load().await;
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        make_some_layers(tline.as_ref(), Lsn(0x20)).await?;
+
+        let layer_map = tline.layers.read().await;
+        let level0_deltas = layer_map.layer_map().get_level0_deltas()?;
+
+        assert!(!level0_deltas.is_empty());
+
+        for delta in level0_deltas {
+            let delta = layer_map.get_from_desc(&delta);
+            // Ensure we are dumping a delta layer here
+            let delta = delta.downcast_delta_layer().unwrap();
+
+            delta.dump(false, &ctx).await.unwrap();
+            delta.dump(true, &ctx).await.unwrap();
+        }
+
+        Ok(())
+    }
+
     #[tokio::test]
     async fn corrupt_metadata() -> anyhow::Result<()> {
         const TEST_NAME: &str = "corrupt_metadata";
@@ -3964,7 +4101,7 @@ mod tests {
         let mut found_error_message = false;
         let mut err_source = err.source();
         while let Some(source) = err_source {
-            if source.to_string() == "metadata checksum mismatch" {
+            if source.to_string().contains("metadata checksum mismatch") {
                 found_error_message = true;
                 break;
             }

pageserver/src/tenant/blob_io.rs

@@ -12,14 +12,11 @@
 //! len >= 128: 1XXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
 use crate::page_cache::PAGE_SZ;
-use crate::tenant::block_io::{BlockCursor, BlockReader};
+use crate::tenant::block_io::BlockCursor;
 use std::cmp::min;
 use std::io::{Error, ErrorKind};
 
-impl<R> BlockCursor<R>
-where
-    R: BlockReader,
-{
+impl<'a> BlockCursor<'a> {
     /// Read a blob into a new buffer.
     pub async fn read_blob(&self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
         let mut buf = Vec::new();

pageserver/src/tenant/block_io.rs

@@ -2,11 +2,14 @@
 //! Low-level Block-oriented I/O functions
 //!
 
+use super::ephemeral_file::EphemeralFile;
+use super::storage_layer::delta_layer::{Adapter, DeltaLayerInner};
 use crate::page_cache::{self, PageReadGuard, ReadBufResult, PAGE_SZ};
+use crate::virtual_file::VirtualFile;
 use bytes::Bytes;
+use std::fs::File;
 use std::ops::{Deref, DerefMut};
 use std::os::unix::fs::FileExt;
-use std::sync::atomic::AtomicU64;
 
 /// This is implemented by anything that can read 8 kB (PAGE_SZ)
 /// blocks, using the page cache
@@ -14,72 +17,87 @@ use std::sync::atomic::AtomicU64;
 /// There are currently two implementations: EphemeralFile, and FileBlockReader
 /// below.
 pub trait BlockReader {
-    ///
-    /// Read a block. Returns a "lease" object that can be used to
-    /// access to the contents of the page. (For the page cache, the
-    /// lease object represents a lock on the buffer.)
-    ///
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error>;
-
     ///
     /// Create a new "cursor" for reading from this reader.
     ///
     /// A cursor caches the last accessed page, allowing for faster
     /// access if the same block is accessed repeatedly.
-    fn block_cursor(&self) -> BlockCursor<&Self>
-    where
-        Self: Sized,
-    {
-        BlockCursor::new(self)
-    }
+    fn block_cursor(&self) -> BlockCursor<'_>;
 }
 
 impl<B> BlockReader for &B
 where
     B: BlockReader,
 {
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        (*self).read_blk(blknum)
+    fn block_cursor(&self) -> BlockCursor<'_> {
+        (*self).block_cursor()
     }
 }
 
-/// A block accessible for reading
-///
-/// During builds with `#[cfg(test)]`, this is a proper enum
-/// with two variants to support testing code. During normal
-/// builds, it just has one variant and is thus a cheap newtype
-/// wrapper of [`PageReadGuard`]
-pub enum BlockLease {
-    PageReadGuard(PageReadGuard<'static>),
+/// Reference to an in-memory copy of an immutable on-disk block.
+pub enum BlockLease<'a> {
+    /// [crate::page_cache::PageCache::drop_buffers_for_immutable] relies on the read guard
+    /// not outiving the EphemeralFile. See the comment in there for details.
+    PageReadGuard(PageReadGuard<'a>),
+    EphemeralFileMutableTail(&'a [u8; PAGE_SZ]),
     #[cfg(test)]
     Rc(std::rc::Rc<[u8; PAGE_SZ]>),
 }
 
-impl From<PageReadGuard<'static>> for BlockLease {
-    fn from(value: PageReadGuard<'static>) -> Self {
+impl From<PageReadGuard<'static>> for BlockLease<'static> {
+    fn from(value: PageReadGuard<'static>) -> BlockLease<'static> {
         BlockLease::PageReadGuard(value)
     }
 }
 
 #[cfg(test)]
-impl From<std::rc::Rc<[u8; PAGE_SZ]>> for BlockLease {
+impl<'a> From<std::rc::Rc<[u8; PAGE_SZ]>> for BlockLease<'a> {
     fn from(value: std::rc::Rc<[u8; PAGE_SZ]>) -> Self {
         BlockLease::Rc(value)
     }
 }
 
-impl Deref for BlockLease {
+impl<'a> Deref for BlockLease<'a> {
     type Target = [u8; PAGE_SZ];
 
     fn deref(&self) -> &Self::Target {
         match self {
             BlockLease::PageReadGuard(v) => v.deref(),
+            BlockLease::EphemeralFileMutableTail(v) => v,
             #[cfg(test)]
             BlockLease::Rc(v) => v.deref(),
         }
     }
 }
 
+/// Provides the ability to read blocks from different sources,
+/// similar to using traits for this purpose.
+///
+/// Unlike traits, we also support the read function to be async though.
+pub(crate) enum BlockReaderRef<'a> {
+    FileBlockReaderVirtual(&'a FileBlockReader<VirtualFile>),
+    FileBlockReaderFile(&'a FileBlockReader<std::fs::File>),
+    EphemeralFile(&'a EphemeralFile),
+    Adapter(Adapter<&'a DeltaLayerInner>),
+    #[cfg(test)]
+    TestDisk(&'a super::disk_btree::tests::TestDisk),
+}
+
+impl<'a> BlockReaderRef<'a> {
+    #[inline(always)]
+    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+        use BlockReaderRef::*;
+        match self {
+            FileBlockReaderVirtual(r) => r.read_blk(blknum),
+            FileBlockReaderFile(r) => r.read_blk(blknum),
+            EphemeralFile(r) => r.read_blk(blknum),
+            Adapter(r) => r.read_blk(blknum),
+            #[cfg(test)]
+            TestDisk(r) => r.read_blk(blknum),
+        }
+    }
+}
+
 ///
 /// A "cursor" for efficiently reading multiple pages from a BlockReader
 ///
@@ -97,32 +115,31 @@ impl Deref for BlockLease {
 /// // do stuff with 'buf'
 /// ```
 ///
-pub struct BlockCursor<R>
-where
-    R: BlockReader,
-{
-    reader: R,
+pub struct BlockCursor<'a> {
+    reader: BlockReaderRef<'a>,
 }
 
-impl<R> BlockCursor<R>
-where
-    R: BlockReader,
-{
-    pub fn new(reader: R) -> Self {
+impl<'a> BlockCursor<'a> {
+    pub(crate) fn new(reader: BlockReaderRef<'a>) -> Self {
         BlockCursor { reader }
     }
+    // Needed by cli
+    pub fn new_fileblockreader_virtual(reader: &'a FileBlockReader<VirtualFile>) -> Self {
+        BlockCursor {
+            reader: BlockReaderRef::FileBlockReaderVirtual(reader),
+        }
+    }
 
+    /// Read a block.
+    ///
+    /// Returns a "lease" object that can be used to
+    /// access to the contents of the page. (For the page cache, the
+    /// lease object represents a lock on the buffer.)
+    #[inline(always)]
     pub fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
         self.reader.read_blk(blknum)
     }
 }
-static NEXT_ID: AtomicU64 = AtomicU64::new(1);
-#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
-pub struct FileId(u64);
-
-fn next_file_id() -> FileId {
-    FileId(NEXT_ID.fetch_add(1, std::sync::atomic::Ordering::Relaxed))
-}
 
 /// An adapter for reading a (virtual) file using the page cache.
 ///
@@ -132,7 +149,7 @@ pub struct FileBlockReader<F> {
     pub file: F,
 
     /// Unique ID of this file, used as key in the page cache.
-    file_id: FileId,
+    file_id: page_cache::FileId,
 }
 
 impl<F> FileBlockReader<F>
@@ -140,7 +157,7 @@ where
     F: FileExt,
 {
     pub fn new(file: F) -> Self {
-        let file_id = next_file_id();
+        let file_id = page_cache::next_file_id();
 
         FileBlockReader { file_id, file }
     }
@@ -150,14 +167,12 @@ where
         assert!(buf.len() == PAGE_SZ);
         self.file.read_exact_at(buf, blkno as u64 * PAGE_SZ as u64)
     }
-}
-
-impl<F> BlockReader for FileBlockReader<F>
-where
-    F: FileExt,
-{
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        // Look up the right page
+    /// Read a block.
+    ///
+    /// Returns a "lease" object that can be used to
+    /// access to the contents of the page. (For the page cache, the
+    /// lease object represents a lock on the buffer.)
+    pub fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
         let cache = page_cache::get();
         loop {
             match cache
@@ -182,6 +197,18 @@ where
     }
 }
 
+impl BlockReader for FileBlockReader<File> {
+    fn block_cursor(&self) -> BlockCursor<'_> {
+        BlockCursor::new(BlockReaderRef::FileBlockReaderFile(self))
+    }
+}
+
+impl BlockReader for FileBlockReader<VirtualFile> {
+    fn block_cursor(&self) -> BlockCursor<'_> {
+        BlockCursor::new(BlockReaderRef::FileBlockReaderVirtual(self))
+    }
+}
+
 ///
 /// Trait for block-oriented output
 ///

pageserver/src/tenant/delete.rs

@@ -7,6 +7,7 @@ use anyhow::Context;
 use pageserver_api::models::TenantState;
 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 use tokio::sync::OwnedMutexGuard;
+use tokio_util::sync::CancellationToken;
 use tracing::{error, info, instrument, warn, Instrument, Span};
 
 use utils::{
@@ -82,6 +83,8 @@ async fn create_remote_delete_mark(
         FAILED_UPLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         "mark_upload",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
     )
     .await
     .context("mark_upload")?;
@@ -171,6 +174,8 @@ async fn remove_tenant_remote_delete_mark(
             FAILED_UPLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "remove_tenant_remote_delete_mark",
+            // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+            backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
         )
         .await
         .context("remove_tenant_remote_delete_mark")?;
@@ -252,6 +257,8 @@ pub(crate) async fn remote_delete_mark_exists(
         SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
         SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
         "fetch_tenant_deletion_mark",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
     )
     .await;
 
@@ -275,8 +282,9 @@ pub(crate) async fn remote_delete_mark_exists(
 /// It is resumable from any step in case a crash/restart occurs.
 /// There are three entrypoints to the process:
 /// 1. [`DeleteTenantFlow::run`] this is the main one called by a management api handler.
-/// 2. [`DeleteTenantFlow::resume`] is called during restarts when local or remote deletion marks are still there.
-/// Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
+/// 2. [`DeleteTenantFlow::resume_from_load`] is called during restarts when local or remote deletion marks are still there.
+/// 3. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
+///  Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
 #[derive(Default)]
 pub enum DeleteTenantFlow {
     #[default]
@@ -403,7 +411,7 @@ impl DeleteTenantFlow {
         }
     }
 
-    pub(crate) async fn resume(
+    pub(crate) async fn resume_from_load(
         guard: DeletionGuard,
         tenant: &Arc<Tenant>,
         init_order: Option<&InitializationOrder>,
@@ -413,7 +421,7 @@ impl DeleteTenantFlow {
         let (_, progress) = completion::channel();
 
         tenant
-            .set_stopping(progress, true)
+            .set_stopping(progress, true, false)
             .await
             .expect("cant be stopping or broken");
 
@@ -441,6 +449,31 @@ impl DeleteTenantFlow {
         .await
     }
 
+    pub(crate) async fn resume_from_attach(
+        guard: DeletionGuard,
+        tenant: &Arc<Tenant>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        ctx: &RequestContext,
+    ) -> Result<(), DeleteTenantError> {
+        let (_, progress) = completion::channel();
+
+        tenant
+            .set_stopping(progress, false, true)
+            .await
+            .expect("cant be stopping or broken");
+
+        tenant.attach(ctx).await.context("attach")?;
+
+        Self::background(
+            guard,
+            tenant.conf,
+            tenant.remote_storage.clone(),
+            tenants,
+            tenant,
+        )
+        .await
+    }
+
     async fn prepare(
         tenants: &tokio::sync::RwLock<TenantsMap>,
         tenant_id: TenantId,

pageserver/src/tenant/disk_btree.rs

@@ -259,9 +259,10 @@ where
     {
         let mut stack = Vec::new();
         stack.push((self.root_blk, None));
+        let block_cursor = self.reader.block_cursor();
         while let Some((node_blknum, opt_iter)) = stack.pop() {
             // Locate the node.
-            let node_buf = self.reader.read_blk(self.start_blk + node_blknum)?;
+            let node_buf = block_cursor.read_blk(self.start_blk + node_blknum)?;
 
             let node = OnDiskNode::deparse(node_buf.as_ref())?;
             let prefix_len = node.prefix_len as usize;
@@ -353,8 +354,10 @@ where
 
         stack.push((self.root_blk, String::new(), 0, 0, 0));
 
+        let block_cursor = self.reader.block_cursor();
+
         while let Some((blknum, path, depth, child_idx, key_off)) = stack.pop() {
-            let blk = self.reader.read_blk(self.start_blk + blknum)?;
+            let blk = block_cursor.read_blk(self.start_blk + blknum)?;
             let buf: &[u8] = blk.as_ref();
             let node = OnDiskNode::<L>::deparse(buf)?;
 
@@ -683,29 +686,32 @@ impl<const L: usize> BuildNode<L> {
 }
 
 #[cfg(test)]
-mod tests {
+pub(crate) mod tests {
     use super::*;
-    use crate::tenant::block_io::BlockLease;
+    use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReaderRef};
     use rand::Rng;
     use std::collections::BTreeMap;
     use std::sync::atomic::{AtomicUsize, Ordering};
 
     #[derive(Clone, Default)]
-    struct TestDisk {
+    pub(crate) struct TestDisk {
         blocks: Vec<Bytes>,
     }
     impl TestDisk {
         fn new() -> Self {
             Self::default()
         }
-    }
-    impl BlockReader for TestDisk {
-        fn read_blk(&self, blknum: u32) -> io::Result<BlockLease> {
+        pub(crate) fn read_blk(&self, blknum: u32) -> io::Result<BlockLease> {
             let mut buf = [0u8; PAGE_SZ];
             buf.copy_from_slice(&self.blocks[blknum as usize]);
             Ok(std::rc::Rc::new(buf).into())
         }
     }
+    impl BlockReader for TestDisk {
+        fn block_cursor(&self) -> BlockCursor<'_> {
+            BlockCursor::new(BlockReaderRef::TestDisk(self))
+        }
+    }
     impl BlockWriter for &mut TestDisk {
         fn write_blk(&mut self, buf: Bytes) -> io::Result<u32> {
             let blknum = self.blocks.len();

pageserver/src/tenant/ephemeral_file.rs

@@ -2,54 +2,30 @@
 //! used to keep in-memory layers spilled on disk.
 
 use crate::config::PageServerConf;
-use crate::page_cache::{self, ReadBufResult, WriteBufResult, PAGE_SZ};
-use crate::tenant::blob_io::BlobWriter;
-use crate::tenant::block_io::{BlockLease, BlockReader};
+use crate::page_cache::{self, PAGE_SZ};
+use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReader};
 use crate::virtual_file::VirtualFile;
-use once_cell::sync::Lazy;
 use std::cmp::min;
-use std::collections::HashMap;
 use std::fs::OpenOptions;
 use std::io::{self, ErrorKind};
 use std::ops::DerefMut;
 use std::os::unix::prelude::FileExt;
 use std::path::PathBuf;
-use std::sync::{Arc, RwLock};
+use std::sync::atomic::AtomicU64;
 use tracing::*;
 use utils::id::{TenantId, TimelineId};
 
-///
-/// This is the global cache of file descriptors (File objects).
-///
-static EPHEMERAL_FILES: Lazy<RwLock<EphemeralFiles>> = Lazy::new(|| {
-    RwLock::new(EphemeralFiles {
-        next_file_id: FileId(1),
-        files: HashMap::new(),
-    })
-});
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-pub struct FileId(u64);
-
-impl std::fmt::Display for FileId {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.0)
-    }
-}
-
-pub struct EphemeralFiles {
-    next_file_id: FileId,
-
-    files: HashMap<FileId, Arc<VirtualFile>>,
-}
-
 pub struct EphemeralFile {
-    file_id: FileId,
+    page_cache_file_id: page_cache::FileId,
+
     _tenant_id: TenantId,
     _timeline_id: TimelineId,
-    file: Arc<VirtualFile>,
-
-    pub size: u64,
+    file: VirtualFile,
+    len: u64,
+    /// An ephemeral file is append-only.
+    /// We keep the last page, which can still be modified, in [`Self::mutable_tail`].
+    /// The other pages, which can no longer be modified, are accessed through the page cache.
+    mutable_tail: [u8; PAGE_SZ],
 }
 
 impl EphemeralFile {
@@ -58,74 +34,179 @@ impl EphemeralFile {
         tenant_id: TenantId,
         timeline_id: TimelineId,
     ) -> Result<EphemeralFile, io::Error> {
-        let mut l = EPHEMERAL_FILES.write().unwrap();
-        let file_id = l.next_file_id;
-        l.next_file_id = FileId(l.next_file_id.0 + 1);
+        static NEXT_FILENAME: AtomicU64 = AtomicU64::new(1);
+        let filename_disambiguator =
+            NEXT_FILENAME.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
 
         let filename = conf
             .timeline_path(&tenant_id, &timeline_id)
-            .join(PathBuf::from(format!("ephemeral-{}", file_id)));
+            .join(PathBuf::from(format!("ephemeral-{filename_disambiguator}")));
 
         let file = VirtualFile::open_with_options(
             &filename,
             OpenOptions::new().read(true).write(true).create(true),
         )?;
-        let file_rc = Arc::new(file);
-        l.files.insert(file_id, file_rc.clone());
 
         Ok(EphemeralFile {
-            file_id,
+            page_cache_file_id: page_cache::next_file_id(),
             _tenant_id: tenant_id,
             _timeline_id: timeline_id,
-            file: file_rc,
-            size: 0,
+            file,
+            len: 0,
+            mutable_tail: [0u8; PAGE_SZ],
         })
     }
 
-    fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), io::Error> {
-        let mut off = 0;
-        while off < PAGE_SZ {
-            let n = self
-                .file
-                .read_at(&mut buf[off..], blkno as u64 * PAGE_SZ as u64 + off as u64)?;
-
-            if n == 0 {
-                // Reached EOF. Fill the rest of the buffer with zeros.
-                const ZERO_BUF: [u8; PAGE_SZ] = [0u8; PAGE_SZ];
-
-                buf[off..].copy_from_slice(&ZERO_BUF[off..]);
-                break;
-            }
-
-            off += n;
-        }
-        Ok(())
+    pub(crate) fn len(&self) -> u64 {
+        self.len
     }
 
-    fn get_buf_for_write(
-        &self,
-        blkno: u32,
-    ) -> Result<page_cache::PageWriteGuard<'static>, io::Error> {
-        // Look up the right page
-        let cache = page_cache::get();
-        let mut write_guard = match cache
-            .write_ephemeral_buf(self.file_id, blkno)
-            .map_err(|e| to_io_error(e, "Failed to write ephemeral buf"))?
-        {
-            WriteBufResult::Found(guard) => guard,
-            WriteBufResult::NotFound(mut guard) => {
-                // Read the page from disk into the buffer
-                // TODO: if we're overwriting the whole page, no need to read it in first
-                self.fill_buffer(guard.deref_mut(), blkno)?;
-                guard.mark_valid();
+    pub(crate) fn read_blk(&self, blknum: u32) -> Result<BlockLease, io::Error> {
+        let flushed_blknums = 0..self.len / PAGE_SZ as u64;
+        if flushed_blknums.contains(&(blknum as u64)) {
+            let cache = page_cache::get();
+            loop {
+                match cache
+                    .read_immutable_buf(self.page_cache_file_id, blknum)
+                    .map_err(|e| {
+                        std::io::Error::new(
+                            std::io::ErrorKind::Other,
+                            // order path before error because error is anyhow::Error => might have many contexts
+                            format!(
+                                "ephemeral file: read immutable page #{}: {}: {:#}",
+                                blknum,
+                                self.file.path.display(),
+                                e,
+                            ),
+                        )
+                    })? {
+                    page_cache::ReadBufResult::Found(guard) => {
+                        return Ok(BlockLease::PageReadGuard(guard))
+                    }
+                    page_cache::ReadBufResult::NotFound(mut write_guard) => {
+                        let buf: &mut [u8] = write_guard.deref_mut();
+                        debug_assert_eq!(buf.len(), PAGE_SZ);
+                        self.file
+                            .read_exact_at(&mut buf[..], blknum as u64 * PAGE_SZ as u64)?;
+                        write_guard.mark_valid();
 
-                // And then fall through to modify it.
-                guard
+                        // Swap for read lock
+                        continue;
+                    }
+                };
             }
-        };
-        write_guard.mark_dirty();
+        } else {
+            debug_assert_eq!(blknum as u64, self.len / PAGE_SZ as u64);
+            Ok(BlockLease::EphemeralFileMutableTail(&self.mutable_tail))
+        }
+    }
 
-        Ok(write_guard)
+    pub(crate) async fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, io::Error> {
+        struct Writer<'a> {
+            ephemeral_file: &'a mut EphemeralFile,
+            /// The block to which the next [`push_bytes`] will write.
+            blknum: u32,
+            /// The offset inside the block identified by [`blknum`] to which [`push_bytes`] will write.
+            off: usize,
+        }
+        impl<'a> Writer<'a> {
+            fn new(ephemeral_file: &'a mut EphemeralFile) -> io::Result<Writer<'a>> {
+                Ok(Writer {
+                    blknum: (ephemeral_file.len / PAGE_SZ as u64) as u32,
+                    off: (ephemeral_file.len % PAGE_SZ as u64) as usize,
+                    ephemeral_file,
+                })
+            }
+            #[inline(always)]
+            async fn push_bytes(&mut self, src: &[u8]) -> Result<(), io::Error> {
+                let mut src_remaining = src;
+                while !src_remaining.is_empty() {
+                    let dst_remaining = &mut self.ephemeral_file.mutable_tail[self.off..];
+                    let n = min(dst_remaining.len(), src_remaining.len());
+                    dst_remaining[..n].copy_from_slice(&src_remaining[..n]);
+                    self.off += n;
+                    src_remaining = &src_remaining[n..];
+                    if self.off == PAGE_SZ {
+                        match self.ephemeral_file.file.write_all_at(
+                            &self.ephemeral_file.mutable_tail,
+                            self.blknum as u64 * PAGE_SZ as u64,
+                        ) {
+                            Ok(_) => {
+                                // Pre-warm the page cache with what we just wrote.
+                                // This isn't necessary for coherency/correctness, but it's how we've always done it.
+                                let cache = page_cache::get();
+                                match cache.read_immutable_buf(
+                                    self.ephemeral_file.page_cache_file_id,
+                                    self.blknum,
+                                ) {
+                                    Ok(page_cache::ReadBufResult::Found(_guard)) => {
+                                        // This function takes &mut self, so, it shouldn't be possible to reach this point.
+                                        unreachable!("we just wrote blknum {} and this function takes &mut self, so, no concurrent read_blk is possible", self.blknum);
+                                    }
+                                    Ok(page_cache::ReadBufResult::NotFound(mut write_guard)) => {
+                                        let buf: &mut [u8] = write_guard.deref_mut();
+                                        debug_assert_eq!(buf.len(), PAGE_SZ);
+                                        buf.copy_from_slice(&self.ephemeral_file.mutable_tail);
+                                        write_guard.mark_valid();
+                                        // pre-warm successful
+                                    }
+                                    Err(e) => {
+                                        error!("ephemeral_file write_blob failed to get immutable buf to pre-warm page cache: {e:?}");
+                                        // fail gracefully, it's not the end of the world if we can't pre-warm the cache here
+                                    }
+                                }
+                                // Zero the buffer for re-use.
+                                // Zeroing is critical for correcntess because the write_blob code below
+                                // and similarly read_blk expect zeroed pages.
+                                self.ephemeral_file.mutable_tail.fill(0);
+                                // This block is done, move to next one.
+                                self.blknum += 1;
+                                self.off = 0;
+                            }
+                            Err(e) => {
+                                return Err(std::io::Error::new(
+                                    ErrorKind::Other,
+                                    // order error before path because path is long and error is short
+                                    format!(
+                                        "ephemeral_file: write_blob: write-back full tail blk #{}: {:#}: {}",
+                                        self.blknum,
+                                        e,
+                                        self.ephemeral_file.file.path.display(),
+                                    ),
+                                ));
+                            }
+                        }
+                    }
+                }
+                Ok(())
+            }
+        }
+
+        let pos = self.len;
+        let mut writer = Writer::new(self)?;
+
+        // Write the length field
+        if srcbuf.len() < 0x80 {
+            // short one-byte length header
+            let len_buf = [srcbuf.len() as u8];
+            writer.push_bytes(&len_buf).await?;
+        } else {
+            let mut len_buf = u32::to_be_bytes(srcbuf.len() as u32);
+            len_buf[0] |= 0x80;
+            writer.push_bytes(&len_buf).await?;
+        }
+
+        // Write the payload
+        writer.push_bytes(srcbuf).await?;
+
+        if srcbuf.len() < 0x80 {
+            self.len += 1;
+        } else {
+            self.len += 4;
+        }
+        self.len += srcbuf.len() as u64;
+
+        Ok(pos)
     }
 }
 
@@ -138,99 +219,11 @@ pub fn is_ephemeral_file(filename: &str) -> bool {
     }
 }
 
-impl BlobWriter for EphemeralFile {
-    fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, io::Error> {
-        struct Writer<'a> {
-            ephemeral_file: &'a mut EphemeralFile,
-            /// The block to which the next [`push_bytes`] will write.
-            blknum: u32,
-            /// The offset inside the block identified by [`blknum`] to which [`push_bytes`] will write.
-            off: usize,
-            /// Used by [`push_bytes`] to memoize the page cache write guard across calls to it.
-            memo_page_guard: MemoizedPageWriteGuard,
-        }
-        struct MemoizedPageWriteGuard {
-            guard: page_cache::PageWriteGuard<'static>,
-            /// The block number of the page in `guard`.
-            blknum: u32,
-        }
-        impl<'a> Writer<'a> {
-            fn new(ephemeral_file: &'a mut EphemeralFile) -> io::Result<Writer<'a>> {
-                let blknum = (ephemeral_file.size / PAGE_SZ as u64) as u32;
-                Ok(Writer {
-                    blknum,
-                    off: (ephemeral_file.size % PAGE_SZ as u64) as usize,
-                    memo_page_guard: MemoizedPageWriteGuard {
-                        guard: ephemeral_file.get_buf_for_write(blknum)?,
-                        blknum,
-                    },
-                    ephemeral_file,
-                })
-            }
-            #[inline(always)]
-            fn push_bytes(&mut self, src: &[u8]) -> Result<(), io::Error> {
-                // `src_remaining` is the remaining bytes to be written
-                let mut src_remaining = src;
-                while !src_remaining.is_empty() {
-                    let page = if self.memo_page_guard.blknum == self.blknum {
-                        &mut self.memo_page_guard.guard
-                    } else {
-                        self.memo_page_guard.guard =
-                            self.ephemeral_file.get_buf_for_write(self.blknum)?;
-                        self.memo_page_guard.blknum = self.blknum;
-                        &mut self.memo_page_guard.guard
-                    };
-                    let dst_remaining = &mut page[self.off..];
-                    let n = min(dst_remaining.len(), src_remaining.len());
-                    dst_remaining[..n].copy_from_slice(&src_remaining[..n]);
-                    self.off += n;
-                    src_remaining = &src_remaining[n..];
-                    if self.off == PAGE_SZ {
-                        // This block is done, move to next one.
-                        self.blknum += 1;
-                        self.off = 0;
-                    }
-                }
-                Ok(())
-            }
-        }
-
-        let pos = self.size;
-        let mut writer = Writer::new(self)?;
-
-        // Write the length field
-        if srcbuf.len() < 0x80 {
-            // short one-byte length header
-            let len_buf = [srcbuf.len() as u8];
-            writer.push_bytes(&len_buf)?;
-        } else {
-            let mut len_buf = u32::to_be_bytes(srcbuf.len() as u32);
-            len_buf[0] |= 0x80;
-            writer.push_bytes(&len_buf)?;
-        }
-
-        // Write the payload
-        writer.push_bytes(srcbuf)?;
-
-        if srcbuf.len() < 0x80 {
-            self.size += 1;
-        } else {
-            self.size += 4;
-        }
-        self.size += srcbuf.len() as u64;
-
-        Ok(pos)
-    }
-}
-
 impl Drop for EphemeralFile {
     fn drop(&mut self) {
         // drop all pages from page cache
         let cache = page_cache::get();
-        cache.drop_buffers_for_ephemeral(self.file_id);
-
-        // remove entry from the hash map
-        EPHEMERAL_FILES.write().unwrap().files.remove(&self.file_id);
+        cache.drop_buffers_for_immutable(self.page_cache_file_id);
 
         // unlink the file
         let res = std::fs::remove_file(&self.file.path);
@@ -250,59 +243,16 @@ impl Drop for EphemeralFile {
     }
 }
 
-pub fn writeback(file_id: FileId, blkno: u32, buf: &[u8]) -> Result<(), io::Error> {
-    if let Some(file) = EPHEMERAL_FILES.read().unwrap().files.get(&file_id) {
-        match file.write_all_at(buf, blkno as u64 * PAGE_SZ as u64) {
-            Ok(_) => Ok(()),
-            Err(e) => Err(io::Error::new(
-                ErrorKind::Other,
-                format!(
-                    "failed to write back to ephemeral file at {} error: {}",
-                    file.path.display(),
-                    e
-                ),
-            )),
-        }
-    } else {
-        Err(io::Error::new(
-            ErrorKind::Other,
-            "could not write back page, not found in ephemeral files hash",
-        ))
-    }
-}
-
 impl BlockReader for EphemeralFile {
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, io::Error> {
-        // Look up the right page
-        let cache = page_cache::get();
-        loop {
-            match cache
-                .read_ephemeral_buf(self.file_id, blknum)
-                .map_err(|e| to_io_error(e, "Failed to read ephemeral buf"))?
-            {
-                ReadBufResult::Found(guard) => return Ok(guard.into()),
-                ReadBufResult::NotFound(mut write_guard) => {
-                    // Read the page from disk into the buffer
-                    self.fill_buffer(write_guard.deref_mut(), blknum)?;
-                    write_guard.mark_valid();
-
-                    // Swap for read lock
-                    continue;
-                }
-            };
-        }
+    fn block_cursor(&self) -> super::block_io::BlockCursor<'_> {
+        BlockCursor::new(super::block_io::BlockReaderRef::EphemeralFile(self))
     }
 }
 
-fn to_io_error(e: anyhow::Error, context: &str) -> io::Error {
-    io::Error::new(ErrorKind::Other, format!("{context}: {e:#}"))
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::tenant::blob_io::BlobWriter;
-    use crate::tenant::block_io::BlockCursor;
+    use crate::tenant::block_io::{BlockCursor, BlockReaderRef};
     use rand::{thread_rng, RngCore};
     use std::fs;
     use std::str::FromStr;
@@ -330,12 +280,12 @@ mod tests {
 
         let mut file = EphemeralFile::create(conf, tenant_id, timeline_id)?;
 
-        let pos_foo = file.write_blob(b"foo")?;
+        let pos_foo = file.write_blob(b"foo").await?;
         assert_eq!(
             b"foo",
             file.block_cursor().read_blob(pos_foo).await?.as_slice()
         );
-        let pos_bar = file.write_blob(b"bar")?;
+        let pos_bar = file.write_blob(b"bar").await?;
         assert_eq!(
             b"foo",
             file.block_cursor().read_blob(pos_foo).await?.as_slice()
@@ -348,17 +298,17 @@ mod tests {
         let mut blobs = Vec::new();
         for i in 0..10000 {
             let data = Vec::from(format!("blob{}", i).as_bytes());
-            let pos = file.write_blob(&data)?;
+            let pos = file.write_blob(&data).await?;
             blobs.push((pos, data));
         }
         // also test with a large blobs
         for i in 0..100 {
             let data = format!("blob{}", i).as_bytes().repeat(100);
-            let pos = file.write_blob(&data)?;
+            let pos = file.write_blob(&data).await?;
             blobs.push((pos, data));
         }
 
-        let cursor = BlockCursor::new(&file);
+        let cursor = BlockCursor::new(BlockReaderRef::EphemeralFile(&file));
         for (pos, expected) in blobs {
             let actual = cursor.read_blob(pos).await?;
             assert_eq!(actual, expected);
@@ -368,7 +318,7 @@ mod tests {
         let mut large_data = Vec::new();
         large_data.resize(20000, 0);
         thread_rng().fill_bytes(&mut large_data);
-        let pos_large = file.write_blob(&large_data)?;
+        let pos_large = file.write_blob(&large_data).await?;
         let result = file.block_cursor().read_blob(pos_large).await?;
         assert_eq!(result, large_data);
 

pageserver/src/tenant/layer_map.rs

@@ -50,7 +50,6 @@ use crate::context::RequestContext;
 use crate::keyspace::KeyPartitioning;
 use crate::repository::Key;
 use crate::tenant::storage_layer::InMemoryLayer;
-use crate::tenant::storage_layer::Layer;
 use anyhow::Result;
 use std::collections::VecDeque;
 use std::ops::Range;

pageserver/src/tenant/metadata.rs

@@ -12,7 +12,7 @@ use std::fs::{File, OpenOptions};
 use std::io::{self, Write};
 
 use anyhow::{bail, ensure, Context};
-use serde::{Deserialize, Serialize};
+use serde::{de::Error, Deserialize, Serialize, Serializer};
 use thiserror::Error;
 use tracing::info_span;
 use utils::bin_ser::SerializeError;
@@ -232,6 +232,28 @@ impl TimelineMetadata {
     }
 }
 
+impl<'de> Deserialize<'de> for TimelineMetadata {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let bytes = Vec::<u8>::deserialize(deserializer)?;
+        Self::from_bytes(bytes.as_slice()).map_err(|e| D::Error::custom(format!("{e}")))
+    }
+}
+
+impl Serialize for TimelineMetadata {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        let bytes = self
+            .to_bytes()
+            .map_err(|e| serde::ser::Error::custom(format!("{e}")))?;
+        bytes.serialize(serializer)
+    }
+}
+
 /// Save timeline metadata to file
 pub fn save_metadata(
     conf: &'static PageServerConf,

pageserver/src/tenant/mgr.rs

@@ -27,8 +27,9 @@ use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME};
 use utils::fs_ext::PathExt;
 use utils::id::{TenantId, TimelineId};
 
-use super::delete::{remote_delete_mark_exists, DeleteTenantError};
+use super::delete::DeleteTenantError;
 use super::timeline::delete::DeleteTimelineFlow;
+use super::TenantSharedResources;
 
 /// The tenants known to the pageserver.
 /// The enum variants are used to distinguish the different states that the pageserver can be in.
@@ -66,8 +67,7 @@ static TENANTS: Lazy<RwLock<TenantsMap>> = Lazy::new(|| RwLock::new(TenantsMap::
 #[instrument(skip_all)]
 pub async fn init_tenant_mgr(
     conf: &'static PageServerConf,
-    broker_client: storage_broker::BrokerClientChannel,
-    remote_storage: Option<GenericRemoteStorage>,
+    resources: TenantSharedResources,
     init_order: InitializationOrder,
 ) -> anyhow::Result<()> {
     // Scan local filesystem for attached tenants
@@ -125,8 +125,7 @@ pub async fn init_tenant_mgr(
                     match schedule_local_tenant_processing(
                         conf,
                         &tenant_dir_path,
-                        broker_client.clone(),
-                        remote_storage.clone(),
+                        resources.clone(),
                         Some(init_order.clone()),
                         &TENANTS,
                         &ctx,
@@ -162,8 +161,7 @@ pub async fn init_tenant_mgr(
 pub(crate) fn schedule_local_tenant_processing(
     conf: &'static PageServerConf,
     tenant_path: &Path,
-    broker_client: storage_broker::BrokerClientChannel,
-    remote_storage: Option<GenericRemoteStorage>,
+    resources: TenantSharedResources,
     init_order: Option<InitializationOrder>,
     tenants: &'static tokio::sync::RwLock<TenantsMap>,
     ctx: &RequestContext,
@@ -200,8 +198,15 @@ pub(crate) fn schedule_local_tenant_processing(
 
     let tenant = if conf.tenant_attaching_mark_file_path(&tenant_id).exists() {
         info!("tenant {tenant_id} has attaching mark file, resuming its attach operation");
-        if let Some(remote_storage) = remote_storage {
-            match Tenant::spawn_attach(conf, tenant_id, broker_client, remote_storage, ctx) {
+        if let Some(remote_storage) = resources.remote_storage {
+            match Tenant::spawn_attach(
+                conf,
+                tenant_id,
+                resources.broker_client,
+                tenants,
+                remote_storage,
+                ctx,
+            ) {
                 Ok(tenant) => tenant,
                 Err(e) => {
                     error!("Failed to spawn_attach tenant {tenant_id}, reason: {e:#}");
@@ -219,15 +224,7 @@ pub(crate) fn schedule_local_tenant_processing(
     } else {
         info!("tenant {tenant_id} is assumed to be loadable, starting load operation");
         // Start loading the tenant into memory. It will initially be in Loading state.
-        Tenant::spawn_load(
-            conf,
-            tenant_id,
-            broker_client,
-            remote_storage,
-            init_order,
-            tenants,
-            ctx,
-        )
+        Tenant::spawn_load(conf, tenant_id, resources, init_order, tenants, ctx)
     };
     Ok(tenant)
 }
@@ -362,8 +359,12 @@ pub async fn create_tenant(
         // TODO: tenant directory remains on disk if we bail out from here on.
         //       See https://github.com/neondatabase/neon/issues/4233
 
+        let tenant_resources = TenantSharedResources {
+            broker_client,
+            remote_storage,
+        };
         let created_tenant =
-            schedule_local_tenant_processing(conf, &tenant_directory, broker_client, remote_storage, None, &TENANTS, ctx)?;
+            schedule_local_tenant_processing(conf, &tenant_directory, tenant_resources, None, &TENANTS, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 
@@ -522,7 +523,11 @@ pub async fn load_tenant(
                 .with_context(|| format!("Failed to remove tenant ignore mark {tenant_ignore_mark:?} during tenant loading"))?;
         }
 
-        let new_tenant = schedule_local_tenant_processing(conf, &tenant_path, broker_client, remote_storage, None, &TENANTS, ctx)
+        let resources = TenantSharedResources {
+            broker_client,
+            remote_storage,
+        };
+        let new_tenant = schedule_local_tenant_processing(conf, &tenant_path,  resources, None,  &TENANTS, ctx)
             .with_context(|| {
                 format!("Failed to schedule tenant processing in path {tenant_path:?}")
             })?;
@@ -591,12 +596,6 @@ pub async fn attach_tenant(
     remote_storage: GenericRemoteStorage,
     ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
-    // Temporary solution, proper one would be to resume deletion, but that needs more plumbing around Tenant::load/Tenant::attach
-    // Corresponding issue https://github.com/neondatabase/neon/issues/5006
-    if remote_delete_mark_exists(conf, &tenant_id, &remote_storage).await? {
-        return Err(anyhow::anyhow!("Tenant is marked as deleted on remote storage").into());
-    }
-
     tenant_map_insert(tenant_id, || {
         let tenant_dir = create_tenant_files(conf, tenant_conf, &tenant_id, CreateTenantFilesMode::Attach)?;
         // TODO: tenant directory remains on disk if we bail out from here on.
@@ -609,7 +608,11 @@ pub async fn attach_tenant(
             .context("check for attach marker file existence")?;
         anyhow::ensure!(marker_file_exists, "create_tenant_files should have created the attach marker file");
 
-        let attached_tenant = schedule_local_tenant_processing(conf, &tenant_dir, broker_client, Some(remote_storage), None, &TENANTS, ctx)?;
+        let resources = TenantSharedResources {
+            broker_client,
+            remote_storage: Some(remote_storage),
+        };
+        let attached_tenant = schedule_local_tenant_processing(conf, &tenant_dir, resources, None, &TENANTS, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -135,7 +135,7 @@
 //! - Initiate upload queue with that [`IndexPart`].
 //! - Reschedule all lost operations by comparing the local filesystem state
 //!   and remote state as per [`IndexPart`]. This is done in
-//!   [`Tenant::timeline_init_and_sync`] and [`Timeline::reconcile_with_remote`].
+//!   [`Tenant::timeline_init_and_sync`].
 //!
 //! Note that if we crash during file deletion between the index update
 //! that removes the file from the list of files, and deleting the remote file,
@@ -172,7 +172,6 @@
 //!   transitioning it from `TenantState::Attaching` to `TenantState::Active` state.
 //!   This starts the timelines' WAL-receivers and the tenant's GC & Compaction loops.
 //!
-//! Most of the above steps happen in [`Timeline::reconcile_with_remote`] or its callers.
 //! We keep track of the fact that a client is in `Attaching` state in a marker
 //! file on the local disk. This is critical because, when we restart the pageserver,
 //! we do not want to do the `List timelines` step for each tenant that has already
@@ -192,14 +191,14 @@
 //! not created and the uploads are skipped.
 //! Theoretically, it should be ok to remove and re-add remote storage configuration to
 //! the pageserver config at any time, since it doesn't make a difference to
-//! `reconcile_with_remote`.
+//! [`Timeline::load_layer_map`].
 //! Of course, the remote timeline dir must not change while we have de-configured
 //! remote storage, i.e., the pageserver must remain the owner of the given prefix
 //! in remote storage.
 //! But note that we don't test any of this right now.
 //!
 //! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
-//! [`Timeline::reconcile_with_remote`]: super::Timeline::reconcile_with_remote
+//! [`Timeline::load_layer_map`]: super::Timeline::load_layer_map
 
 mod delete;
 mod download;
@@ -211,6 +210,7 @@ use chrono::{NaiveDateTime, Utc};
 // re-export these
 pub use download::{is_temp_download_file, list_remote_timelines};
 use scopeguard::ScopeGuard;
+use tokio_util::sync::CancellationToken;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
@@ -231,6 +231,7 @@ use crate::metrics::{
     RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
     REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
 };
+use crate::task_mgr::shutdown_token;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
 use crate::tenant::upload_queue::Delete;
@@ -353,6 +354,10 @@ impl RemoteTimelineClient {
         let mut upload_queue = self.upload_queue.lock().unwrap();
         upload_queue.initialize_with_current_remote_index_part(index_part)?;
         self.update_remote_physical_size_gauge(Some(index_part));
+        info!(
+            "initialized upload queue from remote index with {} layer files",
+            index_part.layer_metadata.len()
+        );
         Ok(())
     }
 
@@ -365,6 +370,7 @@ impl RemoteTimelineClient {
         let mut upload_queue = self.upload_queue.lock().unwrap();
         upload_queue.initialize_empty_remote(local_metadata)?;
         self.update_remote_physical_size_gauge(None);
+        info!("initialized upload queue as empty");
         Ok(())
     }
 
@@ -535,8 +541,7 @@ impl RemoteTimelineClient {
         // ahead of what's _actually_ on the remote during index upload.
         upload_queue.latest_metadata = metadata.clone();
 
-        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
-        self.schedule_index_upload(upload_queue, metadata_bytes);
+        self.schedule_index_upload(upload_queue, upload_queue.latest_metadata.clone());
 
         Ok(())
     }
@@ -556,8 +561,7 @@ impl RemoteTimelineClient {
         let upload_queue = guard.initialized_mut()?;
 
         if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
-            let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
-            self.schedule_index_upload(upload_queue, metadata_bytes);
+            self.schedule_index_upload(upload_queue, upload_queue.latest_metadata.clone());
         }
 
         Ok(())
@@ -567,7 +571,7 @@ impl RemoteTimelineClient {
     fn schedule_index_upload(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        metadata_bytes: Vec<u8>,
+        metadata: TimelineMetadata,
     ) {
         info!(
             "scheduling metadata upload with {} files ({} changed)",
@@ -580,7 +584,7 @@ impl RemoteTimelineClient {
         let index_part = IndexPart::new(
             upload_queue.latest_files.clone(),
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
         );
         let op = UploadOp::UploadMetadata(index_part, disk_consistent_lsn);
         self.calls_unfinished_metric_begin(&op);
@@ -636,7 +640,7 @@ impl RemoteTimelineClient {
 
         // Deleting layers doesn't affect the values stored in TimelineMetadata,
         // so we don't need update it. Just serialize it.
-        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
+        let metadata = upload_queue.latest_metadata.clone();
 
         // Update the remote index file, removing the to-be-deleted files from the index,
         // before deleting the actual files.
@@ -647,12 +651,13 @@ impl RemoteTimelineClient {
         // to syntactically forbid ? or bail! calls here.
         let no_bail_here = || {
             for name in names {
-                upload_queue.latest_files.remove(name);
-                upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
+                if upload_queue.latest_files.remove(name).is_some() {
+                    upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
+                }
             }
 
             if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
-                self.schedule_index_upload(upload_queue, metadata_bytes);
+                self.schedule_index_upload(upload_queue, metadata);
             }
 
             // schedule the actual deletions
@@ -754,7 +759,7 @@ impl RemoteTimelineClient {
         pausable_failpoint!("persist_deleted_index_part");
 
         backoff::retry(
-            || async {
+            || {
                 upload::upload_index_part(
                     self.conf,
                     &self.storage_impl,
@@ -762,7 +767,6 @@ impl RemoteTimelineClient {
                     &self.timeline_id,
                     &index_part_with_deleted_at,
                 )
-                .await
             },
             |_e| false,
             1,
@@ -771,6 +775,8 @@ impl RemoteTimelineClient {
             // when executed as part of tenant deletion this happens in the background
             2,
             "persist_index_part_with_deleted_flag",
+            // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+            backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
         )
         .await?;
 
@@ -857,6 +863,7 @@ impl RemoteTimelineClient {
             FAILED_DOWNLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "list_prefixes",
+            backoff::Cancel::new(shutdown_token(), || anyhow::anyhow!("Cancelled!")),
         )
         .await
         .context("list prefixes")?;
@@ -880,6 +887,7 @@ impl RemoteTimelineClient {
                 FAILED_UPLOAD_WARN_THRESHOLD,
                 FAILED_REMOTE_OP_RETRIES,
                 "delete_objects",
+                backoff::Cancel::new(shutdown_token(), || anyhow::anyhow!("Cancelled!")),
             )
             .await
             .context("delete_objects")?;
@@ -901,6 +909,7 @@ impl RemoteTimelineClient {
             FAILED_UPLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "delete_index",
+            backoff::Cancel::new(shutdown_token(), || anyhow::anyhow!("Cancelled")),
         )
         .await
         .context("delete_index")?;
@@ -1066,6 +1075,15 @@ impl RemoteTimelineClient {
                     .await
                 }
                 UploadOp::UploadMetadata(ref index_part, _lsn) => {
+                    let mention_having_future_layers = if cfg!(feature = "testing") {
+                        index_part
+                            .layer_metadata
+                            .keys()
+                            .any(|x| x.is_in_future(*_lsn))
+                    } else {
+                        false
+                    };
+
                     let res = upload::upload_index_part(
                         self.conf,
                         &self.storage_impl,
@@ -1083,6 +1101,10 @@ impl RemoteTimelineClient {
                     .await;
                     if res.is_ok() {
                         self.update_remote_physical_size_gauge(Some(index_part));
+                        if mention_having_future_layers {
+                            // find rationale near crate::tenant::timeline::init::cleanup_future_layer
+                            tracing::info!(disk_consistent_lsn=%_lsn, "uploaded an index_part.json with future layers -- this is ok! if shutdown now, expect future layer cleanup");
+                        }
                     }
                     res
                 }
@@ -1134,14 +1156,13 @@ impl RemoteTimelineClient {
                     }
 
                     // sleep until it's time to retry, or we're cancelled
-                    tokio::select! {
-                        _ = task_mgr::shutdown_watcher() => { },
-                        _ = exponential_backoff(
-                            retries,
-                            DEFAULT_BASE_BACKOFF_SECONDS,
-                            DEFAULT_MAX_BACKOFF_SECONDS,
-                        ) => { },
-                    };
+                    exponential_backoff(
+                        retries,
+                        DEFAULT_BASE_BACKOFF_SECONDS,
+                        DEFAULT_MAX_BACKOFF_SECONDS,
+                        &shutdown_token(),
+                    )
+                    .await;
                 }
             }
         }
@@ -1578,14 +1599,17 @@ mod tests {
         };
 
         assert_file_list(
-            &index_part.timeline_layers,
+            &index_part
+                .layer_metadata
+                .keys()
+                .map(|f| f.to_owned())
+                .collect(),
             &[
                 &layer_file_name_1.file_name(),
                 &layer_file_name_2.file_name(),
             ],
         );
-        let downloaded_metadata = index_part.parse_metadata().unwrap();
-        assert_eq!(downloaded_metadata, metadata);
+        assert_eq!(index_part.metadata, metadata);
 
         // Schedule upload and then a deletion. Check that the deletion is queued
         client

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -11,6 +11,7 @@ use std::time::Duration;
 use anyhow::{anyhow, Context};
 use tokio::fs;
 use tokio::io::AsyncWriteExt;
+use tokio_util::sync::CancellationToken;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
@@ -280,6 +281,10 @@ where
         FAILED_DOWNLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         description,
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || -> DownloadError {
+            unreachable!()
+        }),
     )
     .await
 }

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -62,10 +62,9 @@ pub struct IndexPart {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub deleted_at: Option<NaiveDateTime>,
 
-    /// Layer names, which are stored on the remote storage.
-    ///
-    /// Additional metadata can might exist in `layer_metadata`.
-    pub timeline_layers: HashSet<LayerFileName>,
+    /// Legacy field: equal to the keys of `layer_metadata`, only written out for forward compat
+    #[serde(default, skip_deserializing)]
+    timeline_layers: HashSet<LayerFileName>,
 
     /// Per layer file name metadata, which can be present for a present or missing layer file.
     ///
@@ -74,10 +73,13 @@ pub struct IndexPart {
     pub layer_metadata: HashMap<LayerFileName, IndexLayerMetadata>,
 
     // 'disk_consistent_lsn' is a copy of the 'disk_consistent_lsn' in the metadata.
-    // It's duplicated here for convenience.
+    // It's duplicated for convenience when reading the serialized structure, but is
+    // private because internally we would read from metadata instead.
     #[serde_as(as = "DisplayFromStr")]
-    pub disk_consistent_lsn: Lsn,
-    metadata_bytes: Vec<u8>,
+    disk_consistent_lsn: Lsn,
+
+    #[serde(rename = "metadata_bytes")]
+    pub metadata: TimelineMetadata,
 }
 
 impl IndexPart {
@@ -85,13 +87,17 @@ impl IndexPart {
     /// used to understand later versions.
     ///
     /// Version is currently informative only.
-    const LATEST_VERSION: usize = 2;
+    /// Version history
+    /// - 2: added `deleted_at`
+    /// - 3: no longer deserialize `timeline_layers` (serialized format is the same, but timeline_layers
+    ///      is always generated from the keys of `layer_metadata`)
+    const LATEST_VERSION: usize = 3;
     pub const FILE_NAME: &'static str = "index_part.json";
 
     pub fn new(
         layers_and_metadata: HashMap<LayerFileName, LayerFileMetadata>,
         disk_consistent_lsn: Lsn,
-        metadata_bytes: Vec<u8>,
+        metadata: TimelineMetadata,
     ) -> Self {
         let mut timeline_layers = HashSet::with_capacity(layers_and_metadata.len());
         let mut layer_metadata = HashMap::with_capacity(layers_and_metadata.len());
@@ -107,14 +113,10 @@ impl IndexPart {
             timeline_layers,
             layer_metadata,
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
             deleted_at: None,
         }
     }
-
-    pub fn parse_metadata(&self) -> anyhow::Result<TimelineMetadata> {
-        TimelineMetadata::from_bytes(&self.metadata_bytes)
-    }
 }
 
 impl TryFrom<&UploadQueueInitialized> for IndexPart {
@@ -122,12 +124,12 @@ impl TryFrom<&UploadQueueInitialized> for IndexPart {
 
     fn try_from(upload_queue: &UploadQueueInitialized) -> Result<Self, Self::Error> {
         let disk_consistent_lsn = upload_queue.latest_metadata.disk_consistent_lsn();
-        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
+        let metadata = upload_queue.latest_metadata.clone();
 
         Ok(Self::new(
             upload_queue.latest_files.clone(),
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
         ))
     }
 }
@@ -166,7 +168,7 @@ mod tests {
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
-            timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
+            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
@@ -178,7 +180,7 @@ mod tests {
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
-            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].to_vec(),
+            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,
         };
 
@@ -197,13 +199,13 @@ mod tests {
                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
             },
             "disk_consistent_lsn":"0/16960E8",
-            "metadata_bytes":[112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
+            "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]
         }"#;
 
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
-            timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
+            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
@@ -215,7 +217,7 @@ mod tests {
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
-            metadata_bytes: [112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
+            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,
         };
 
@@ -234,14 +236,14 @@ mod tests {
                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
             },
             "disk_consistent_lsn":"0/16960E8",
-            "metadata_bytes":[112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+            "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"
         }"#;
 
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 2,
-            timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
+            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
@@ -253,7 +255,7 @@ mod tests {
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
-            metadata_bytes: [112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
+            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())
         };
@@ -277,7 +279,7 @@ mod tests {
             timeline_layers: HashSet::new(),
             layer_metadata: HashMap::new(),
             disk_consistent_lsn: "0/2532648".parse::<Lsn>().unwrap(),
-            metadata_bytes: [
+            metadata: TimelineMetadata::from_bytes(&[
                 136, 151, 49, 208, 0, 70, 0, 4, 0, 0, 0, 0, 2, 83, 38, 72, 1, 0, 0, 0, 0, 2, 83,
                 38, 32, 1, 87, 198, 240, 135, 97, 119, 45, 125, 38, 29, 155, 161, 140, 141, 255,
                 210, 0, 0, 0, 0, 2, 83, 38, 72, 0, 0, 0, 0, 1, 73, 240, 192, 0, 0, 0, 0, 1, 73,
@@ -298,8 +300,8 @@ mod tests {
                 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                 0, 0,
-            ]
-            .to_vec(),
+            ])
+            .unwrap(),
             deleted_at: None,
         };
 

pageserver/src/tenant/storage_layer.rs

@@ -41,8 +41,6 @@ pub use inmemory_layer::InMemoryLayer;
 pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
 pub use remote_layer::RemoteLayer;
 
-use super::timeline::layer_manager::LayerManager;
-
 pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
 where
     T: PartialOrd<T>,
@@ -175,16 +173,9 @@ impl LayerAccessStats {
     ///
     /// [`LayerLoad`]: LayerResidenceEventReason::LayerLoad
     /// [`record_residence_event`]: Self::record_residence_event
-    pub(crate) fn for_loading_layer(
-        layer_map_lock_held_witness: &LayerManager,
-        status: LayerResidenceStatus,
-    ) -> Self {
+    pub(crate) fn for_loading_layer(status: LayerResidenceStatus) -> Self {
         let new = LayerAccessStats(Mutex::new(LayerAccessStatsLocked::default()));
-        new.record_residence_event(
-            layer_map_lock_held_witness,
-            status,
-            LayerResidenceEventReason::LayerLoad,
-        );
+        new.record_residence_event(status, LayerResidenceEventReason::LayerLoad);
         new
     }
 
@@ -197,7 +188,6 @@ impl LayerAccessStats {
     /// [`record_residence_event`]: Self::record_residence_event
     pub(crate) fn clone_for_residence_change(
         &self,
-        layer_map_lock_held_witness: &LayerManager,
         new_status: LayerResidenceStatus,
     ) -> LayerAccessStats {
         let clone = {
@@ -205,11 +195,7 @@ impl LayerAccessStats {
             inner.clone()
         };
         let new = LayerAccessStats(Mutex::new(clone));
-        new.record_residence_event(
-            layer_map_lock_held_witness,
-            new_status,
-            LayerResidenceEventReason::ResidenceChange,
-        );
+        new.record_residence_event(new_status, LayerResidenceEventReason::ResidenceChange);
         new
     }
 
@@ -229,7 +215,6 @@ impl LayerAccessStats {
     ///
     pub(crate) fn record_residence_event(
         &self,
-        _layer_map_lock_held_witness: &LayerManager,
         status: LayerResidenceStatus,
         reason: LayerResidenceEventReason,
     ) {
@@ -344,23 +329,6 @@ impl LayerAccessStats {
 /// are used in (timeline).
 #[async_trait::async_trait]
 pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync + 'static {
-    /// Range of keys that this layer covers
-    fn get_key_range(&self) -> Range<Key>;
-
-    /// Inclusive start bound of the LSN range that this layer holds
-    /// Exclusive end bound of the LSN range that this layer holds.
-    ///
-    /// - For an open in-memory layer, this is MAX_LSN.
-    /// - For a frozen in-memory layer or a delta layer, this is a valid end bound.
-    /// - An image layer represents snapshot at one LSN, so end_lsn is always the snapshot LSN + 1
-    fn get_lsn_range(&self) -> Range<Lsn>;
-
-    /// Does this layer only contain some data for the key-range (incremental),
-    /// or does it contain a version of every page? This is important to know
-    /// for garbage collecting old layers: an incremental layer depends on
-    /// the previous non-incremental layer.
-    fn is_incremental(&self) -> bool;
-
     ///
     /// Return data needed to reconstruct given page at LSN.
     ///
@@ -380,9 +348,6 @@ pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync + 'static {
         reconstruct_data: &mut ValueReconstructState,
         ctx: &RequestContext,
     ) -> Result<ValueReconstructResult>;
-
-    /// Dump summary of the contents of the layer to stdout
-    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()>;
 }
 
 /// Get a layer descriptor from a layer.
@@ -467,7 +432,6 @@ pub mod tests {
                 TimelineId::from_array([0; 16]),
                 value.key_range,
                 value.lsn,
-                false,
                 233,
             )
         }

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -51,7 +51,6 @@ use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
-use tokio::runtime::Handle;
 use tokio::sync::OnceCell;
 use tracing::*;
 
@@ -177,10 +176,6 @@ impl DeltaKey {
         Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
     }
 
-    fn extract_key_from_buf(buf: &[u8]) -> Key {
-        Key::from_slice(&buf[..KEY_SIZE])
-    }
-
     fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
         let mut lsn_buf = [0u8; 8];
         lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
@@ -244,8 +239,54 @@ impl std::fmt::Debug for DeltaLayerInner {
 
 #[async_trait::async_trait]
 impl Layer for DeltaLayer {
-    /// debugging function to print out the contents of the layer
-    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_state, ctx)
+            .await
+    }
+}
+/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
+impl std::fmt::Display for DeltaLayer {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.layer_desc().short_id())
+    }
+}
+
+impl AsLayerDesc for DeltaLayer {
+    fn layer_desc(&self) -> &PersistentLayerDesc {
+        &self.desc
+    }
+}
+
+impl PersistentLayer for DeltaLayer {
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        Some(self)
+    }
+
+    fn local_path(&self) -> Option<PathBuf> {
+        self.local_path()
+    }
+
+    fn delete_resident_layer_file(&self) -> Result<()> {
+        self.delete_resident_layer_file()
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        self.info(reset)
+    }
+
+    fn access_stats(&self) -> &LayerAccessStats {
+        self.access_stats()
+    }
+}
+
+impl DeltaLayer {
+    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         println!(
             "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} size {} ----",
             self.desc.tenant_id,
@@ -277,12 +318,11 @@ impl Layer for DeltaLayer {
 
         tree_reader.dump().await?;
 
-        let cursor = file.block_cursor();
+        let keys = DeltaLayerInner::load_keys(&inner).await?;
 
         // A subroutine to dump a single blob
-        let dump_blob = |blob_ref: BlobRef| -> anyhow::Result<String> {
-            // TODO this is not ideal, but on the other hand we are in dumping code...
-            let buf = Handle::current().block_on(cursor.read_blob(blob_ref.pos()))?;
+        async fn dump_blob(val: ValueRef<'_>) -> Result<String> {
+            let buf = val.reader.read_blob(val.blob_ref.pos()).await?;
             let val = Value::des(&buf)?;
             let desc = match val {
                 Value::Image(img) => {
@@ -299,31 +339,24 @@ impl Layer for DeltaLayer {
                 }
             };
             Ok(desc)
-        };
+        }
 
-        tree_reader
-            .visit(
-                &[0u8; DELTA_KEY_SIZE],
-                VisitDirection::Forwards,
-                |delta_key, val| {
-                    let blob_ref = BlobRef(val);
-                    let key = DeltaKey::extract_key_from_buf(delta_key);
-                    let lsn = DeltaKey::extract_lsn_from_buf(delta_key);
-
-                    let desc = match dump_blob(blob_ref) {
-                        Ok(desc) => desc,
-                        Err(err) => format!("ERROR: {}", err),
-                    };
-                    println!("  key {} at {}: {}", key, lsn, desc);
-                    true
-                },
-            )
-            .await?;
+        for entry in keys {
+            let DeltaEntry { key, lsn, val, .. } = entry;
+            let desc = match dump_blob(val).await {
+                Ok(desc) => desc,
+                Err(err) => {
+                    let err: anyhow::Error = err;
+                    format!("ERROR: {err}")
+                }
+            };
+            println!("  key {key} at {lsn}: {desc}");
+        }
 
         Ok(())
     }
 
-    async fn get_value_reconstruct_data(
+    pub(crate) async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -342,52 +375,19 @@ impl Layer for DeltaLayer {
             .await
     }
 
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_key_range(&self) -> Range<Key> {
-        self.layer_desc().key_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_lsn_range(&self) -> Range<Lsn> {
-        self.layer_desc().lsn_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn is_incremental(&self) -> bool {
-        self.layer_desc().is_incremental
-    }
-}
-/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-impl std::fmt::Display for DeltaLayer {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.layer_desc().short_id())
-    }
-}
-
-impl AsLayerDesc for DeltaLayer {
-    fn layer_desc(&self) -> &PersistentLayerDesc {
-        &self.desc
-    }
-}
-
-impl PersistentLayer for DeltaLayer {
-    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
-        Some(self)
-    }
-
-    fn local_path(&self) -> Option<PathBuf> {
+    pub(crate) fn local_path(&self) -> Option<PathBuf> {
         Some(self.path())
     }
 
-    fn delete_resident_layer_file(&self) -> Result<()> {
+    pub(crate) fn delete_resident_layer_file(&self) -> Result<()> {
         // delete underlying file
         fs::remove_file(self.path())?;
         Ok(())
     }
 
-    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
-        let layer_file_name = self.filename().file_name();
-        let lsn_range = self.get_lsn_range();
+    pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_range = self.layer_desc().lsn_range.clone();
 
         let access_stats = self.access_stats.as_api_model(reset);
 
@@ -401,12 +401,10 @@ impl PersistentLayer for DeltaLayer {
         }
     }
 
-    fn access_stats(&self) -> &LayerAccessStats {
+    pub(crate) fn access_stats(&self) -> &LayerAccessStats {
         &self.access_stats
     }
-}
 
-impl DeltaLayer {
     fn path_for(
         path_or_conf: &PathOrConf,
         tenant_id: &TenantId,
@@ -552,17 +550,12 @@ impl DeltaLayer {
     /// Loads all keys stored in the layer. Returns key, lsn, value size and value reference.
     ///
     /// The value can be obtained via the [`ValueRef::load`] function.
-    pub(crate) async fn load_keys(
-        &self,
-        ctx: &RequestContext,
-    ) -> Result<Vec<DeltaEntry<Ref<&'_ DeltaLayerInner>>>> {
+    pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> Result<Vec<DeltaEntry<'_>>> {
         let inner = self
             .load(LayerAccessKind::KeyIter, ctx)
             .await
             .context("load delta layer keys")?;
-
-        let inner = Ref(&**inner);
-        DeltaLayerInner::load_keys(&inner)
+        DeltaLayerInner::load_keys(inner)
             .await
             .context("Layer index is corrupted")
     }
@@ -855,6 +848,7 @@ impl DeltaLayerInner {
 
         let summary_blk = file.read_blk(0)?;
         let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        drop(summary_blk); // so we don't borrow `file` for too long
 
         if let Some(mut expected_summary) = summary {
             // production code path
@@ -958,14 +952,14 @@ impl DeltaLayerInner {
 
     pub(super) async fn load_keys<T: AsRef<DeltaLayerInner> + Clone>(
         this: &T,
-    ) -> Result<Vec<DeltaEntry<T>>> {
+    ) -> Result<Vec<DeltaEntry<'_>>> {
         let dl = this.as_ref();
         let file = &dl.file;
 
         let tree_reader =
             DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(dl.index_start_blk, dl.index_root_blk, file);
 
-        let mut all_keys: Vec<DeltaEntry<T>> = Vec::new();
+        let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
 
         tree_reader
             .visit(
@@ -975,7 +969,9 @@ impl DeltaLayerInner {
                     let delta_key = DeltaKey::from_slice(key);
                     let val_ref = ValueRef {
                         blob_ref: BlobRef(value),
-                        reader: BlockCursor::new(Adapter(this.clone())),
+                        reader: BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(
+                            Adapter(dl),
+                        )),
                     };
                     let pos = BlobRef(value).pos();
                     if let Some(last) = all_keys.last_mut() {
@@ -1004,43 +1000,23 @@ impl DeltaLayerInner {
     }
 }
 
-/// Cloneable borrow wrapper to make borrows behave like smart pointers.
-///
-/// Shared references are trivially copyable. This wrapper avoids (confusion) to otherwise attempt
-/// cloning DeltaLayerInner.
-pub(crate) struct Ref<T>(T);
-
-impl<'a, T> AsRef<T> for Ref<&'a T> {
-    fn as_ref(&self) -> &T {
-        self.0
-    }
-}
-
-impl<'a, T> Clone for Ref<&'a T> {
-    fn clone(&self) -> Self {
-        *self
-    }
-}
-
-impl<'a, T> Copy for Ref<&'a T> {}
-
 /// A set of data associated with a delta layer key and its value
-pub struct DeltaEntry<T: AsRef<DeltaLayerInner>> {
+pub struct DeltaEntry<'a> {
     pub key: Key,
     pub lsn: Lsn,
     /// Size of the stored value
     pub size: u64,
     /// Reference to the on-disk value
-    pub val: ValueRef<T>,
+    pub val: ValueRef<'a>,
 }
 
 /// Reference to an on-disk value
-pub struct ValueRef<T: AsRef<DeltaLayerInner>> {
+pub struct ValueRef<'a> {
     blob_ref: BlobRef,
-    reader: BlockCursor<Adapter<T>>,
+    reader: BlockCursor<'a>,
 }
 
-impl<T: AsRef<DeltaLayerInner>> ValueRef<T> {
+impl<'a> ValueRef<'a> {
     /// Loads the value from disk
     pub async fn load(&self) -> Result<Value> {
         // theoretically we *could* record an access time for each, but it does not really matter
@@ -1050,10 +1026,10 @@ impl<T: AsRef<DeltaLayerInner>> ValueRef<T> {
     }
 }
 
-struct Adapter<T: AsRef<DeltaLayerInner>>(T);
+pub(crate) struct Adapter<T>(T);
 
-impl<T: AsRef<DeltaLayerInner>> BlockReader for Adapter<T> {
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
+    pub(crate) fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
         self.0.as_ref().file.read_blk(blknum)
     }
 }

pageserver/src/tenant/storage_layer/filename.rs

@@ -212,9 +212,20 @@ pub enum LayerFileName {
 }
 
 impl LayerFileName {
-    pub fn file_name(&self) -> String {
+    pub(crate) fn file_name(&self) -> String {
         self.to_string()
     }
+
+    /// 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::*;
+        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,
+            _ => false,
+        }
+    }
 }
 
 impl fmt::Display for LayerFileName {
@@ -263,8 +274,8 @@ impl serde::Serialize for LayerFileName {
         S: serde::Serializer,
     {
         match self {
-            Self::Image(fname) => serializer.serialize_str(&fname.to_string()),
-            Self::Delta(fname) => serializer.serialize_str(&fname.to_string()),
+            Self::Image(fname) => serializer.collect_str(fname),
+            Self::Delta(fname) => serializer.collect_str(fname),
         }
     }
 }

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -169,8 +169,52 @@ impl std::fmt::Debug for ImageLayerInner {
 
 #[async_trait::async_trait]
 impl Layer for ImageLayer {
-    /// debugging function to print out the contents of the layer
-    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    /// Look up given page in the file
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_state: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_state, ctx)
+            .await
+    }
+}
+
+/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
+impl std::fmt::Display for ImageLayer {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.layer_desc().short_id())
+    }
+}
+
+impl AsLayerDesc for ImageLayer {
+    fn layer_desc(&self) -> &PersistentLayerDesc {
+        &self.desc
+    }
+}
+
+impl PersistentLayer for ImageLayer {
+    fn local_path(&self) -> Option<PathBuf> {
+        self.local_path()
+    }
+
+    fn delete_resident_layer_file(&self) -> Result<()> {
+        self.delete_resident_layer_file()
+    }
+
+    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        self.info(reset)
+    }
+
+    fn access_stats(&self) -> &LayerAccessStats {
+        self.access_stats()
+    }
+}
+
+impl ImageLayer {
+    pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         println!(
             "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
             self.desc.tenant_id,
@@ -178,7 +222,7 @@ impl Layer for ImageLayer {
             self.desc.key_range.start,
             self.desc.key_range.end,
             self.lsn,
-            self.desc.is_incremental,
+            self.desc.is_incremental(),
             self.desc.file_size
         );
 
@@ -203,8 +247,7 @@ impl Layer for ImageLayer {
         Ok(())
     }
 
-    /// Look up given page in the file
-    async fn get_value_reconstruct_data(
+    pub(crate) async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -225,65 +268,33 @@ impl Layer for ImageLayer {
             .with_context(|| format!("read {}", self.path().display()))
     }
 
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_key_range(&self) -> Range<Key> {
-        self.layer_desc().key_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_lsn_range(&self) -> Range<Lsn> {
-        self.layer_desc().lsn_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn is_incremental(&self) -> bool {
-        self.layer_desc().is_incremental
-    }
-}
-
-/// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-impl std::fmt::Display for ImageLayer {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.layer_desc().short_id())
-    }
-}
-
-impl AsLayerDesc for ImageLayer {
-    fn layer_desc(&self) -> &PersistentLayerDesc {
-        &self.desc
-    }
-}
-
-impl PersistentLayer for ImageLayer {
-    fn local_path(&self) -> Option<PathBuf> {
+    pub(crate) fn local_path(&self) -> Option<PathBuf> {
         Some(self.path())
     }
 
-    fn delete_resident_layer_file(&self) -> Result<()> {
+    pub(crate) fn delete_resident_layer_file(&self) -> Result<()> {
         // delete underlying file
         fs::remove_file(self.path())?;
         Ok(())
     }
 
-    fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
-        let layer_file_name = self.filename().file_name();
-        let lsn_range = self.get_lsn_range();
+    pub(crate) fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_start = self.layer_desc().image_layer_lsn();
 
         HistoricLayerInfo::Image {
             layer_file_name,
             layer_file_size: self.desc.file_size,
-            lsn_start: lsn_range.start,
+            lsn_start,
             remote: false,
             access_stats: self.access_stats.as_api_model(reset),
         }
     }
 
-    fn access_stats(&self) -> &LayerAccessStats {
+    pub(crate) fn access_stats(&self) -> &LayerAccessStats {
         &self.access_stats
     }
-}
 
-impl ImageLayer {
     fn path_for(
         path_or_conf: &PathOrConf,
         timeline_id: TimelineId,
@@ -371,7 +382,6 @@ impl ImageLayer {
                 timeline_id,
                 filename.key_range.clone(),
                 filename.lsn,
-                false,
                 file_size,
             ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
             lsn: filename.lsn,
@@ -398,7 +408,6 @@ impl ImageLayer {
                 summary.timeline_id,
                 summary.key_range,
                 summary.lsn,
-                false,
                 metadata.len(),
             ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
             lsn: summary.lsn,
@@ -433,6 +442,7 @@ impl ImageLayerInner {
         let file = FileBlockReader::new(file);
         let summary_blk = file.read_blk(0)?;
         let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
+        drop(summary_blk); // so we don't borrow `file` for too long
 
         if let Some(mut expected_summary) = summary {
             // production code path
@@ -500,7 +510,6 @@ struct ImageLayerWriterInner {
     tenant_id: TenantId,
     key_range: Range<Key>,
     lsn: Lsn,
-    is_incremental: bool,
 
     blob_writer: WriteBlobWriter<VirtualFile>,
     tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
@@ -516,7 +525,6 @@ impl ImageLayerWriterInner {
         tenant_id: TenantId,
         key_range: &Range<Key>,
         lsn: Lsn,
-        is_incremental: bool,
     ) -> anyhow::Result<Self> {
         // Create the file initially with a temporary filename.
         // We'll atomically rename it to the final name when we're done.
@@ -551,7 +559,6 @@ impl ImageLayerWriterInner {
             lsn,
             tree: tree_builder,
             blob_writer,
-            is_incremental,
         };
 
         Ok(writer)
@@ -612,7 +619,6 @@ impl ImageLayerWriterInner {
             self.timeline_id,
             self.key_range.clone(),
             self.lsn,
-            self.is_incremental, // for now, image layer ALWAYS covers the full range
             metadata.len(),
         );
 
@@ -687,7 +693,6 @@ impl ImageLayerWriter {
         tenant_id: TenantId,
         key_range: &Range<Key>,
         lsn: Lsn,
-        is_incremental: bool,
     ) -> anyhow::Result<ImageLayerWriter> {
         Ok(Self {
             inner: Some(ImageLayerWriterInner::new(
@@ -696,7 +701,6 @@ impl ImageLayerWriter {
                 tenant_id,
                 key_range,
                 lsn,
-                is_incremental,
             )?),
         })
     }

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -7,14 +7,12 @@
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::repository::{Key, Value};
-use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::BlockReader;
 use crate::tenant::ephemeral_file::EphemeralFile;
 use crate::tenant::storage_layer::{ValueReconstructResult, ValueReconstructState};
 use crate::walrecord;
 use anyhow::{ensure, Result};
 use pageserver_api::models::InMemoryLayerInfo;
-use std::cell::RefCell;
 use std::collections::HashMap;
 use std::sync::OnceLock;
 use tracing::*;
@@ -32,12 +30,6 @@ use tokio::sync::RwLock;
 
 use super::{DeltaLayer, DeltaLayerWriter, Layer};
 
-thread_local! {
-    /// A buffer for serializing object during [`InMemoryLayer::put_value`].
-    /// This buffer is reused for each serialization to avoid additional malloc calls.
-    static SER_BUFFER: RefCell<Vec<u8>> = RefCell::new(Vec::new());
-}
-
 pub struct InMemoryLayer {
     conf: &'static PageServerConf,
     tenant_id: TenantId,
@@ -85,11 +77,11 @@ impl std::fmt::Debug for InMemoryLayerInner {
 }
 
 impl InMemoryLayer {
-    pub fn get_timeline_id(&self) -> TimelineId {
+    pub(crate) fn get_timeline_id(&self) -> TimelineId {
         self.timeline_id
     }
 
-    pub fn info(&self) -> InMemoryLayerInfo {
+    pub(crate) fn info(&self) -> InMemoryLayerInfo {
         let lsn_start = self.start_lsn;
 
         if let Some(&lsn_end) = self.end_lsn.get() {
@@ -99,32 +91,22 @@ impl InMemoryLayer {
         }
     }
 
-    fn assert_writable(&self) {
+    pub(crate) fn assert_writable(&self) {
         assert!(self.end_lsn.get().is_none());
     }
 
-    fn end_lsn_or_max(&self) -> Lsn {
+    pub(crate) fn end_lsn_or_max(&self) -> Lsn {
         self.end_lsn.get().copied().unwrap_or(Lsn::MAX)
     }
-}
 
-#[async_trait::async_trait]
-impl Layer for InMemoryLayer {
-    fn get_key_range(&self) -> Range<Key> {
-        Key::MIN..Key::MAX
-    }
-
-    fn get_lsn_range(&self) -> Range<Lsn> {
+    pub(crate) fn get_lsn_range(&self) -> Range<Lsn> {
         self.start_lsn..self.end_lsn_or_max()
     }
 
-    fn is_incremental(&self) -> bool {
-        // in-memory layer is always considered incremental.
-        true
-    }
-
     /// debugging function to print out the contents of the layer
-    async fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
+    ///
+    /// this is likely completly unused
+    pub async fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
         let inner = self.inner.read().await;
 
         let end_str = self.end_lsn_or_max();
@@ -171,7 +153,7 @@ impl Layer for InMemoryLayer {
     }
 
     /// Look up given value in the layer.
-    async fn get_value_reconstruct_data(
+    pub(crate) async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -221,6 +203,20 @@ impl Layer for InMemoryLayer {
     }
 }
 
+#[async_trait::async_trait]
+impl Layer for InMemoryLayer {
+    async fn get_value_reconstruct_data(
+        &self,
+        key: Key,
+        lsn_range: Range<Lsn>,
+        reconstruct_data: &mut ValueReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<ValueReconstructResult> {
+        self.get_value_reconstruct_data(key, lsn_range, reconstruct_data, ctx)
+            .await
+    }
+}
+
 impl std::fmt::Display for InMemoryLayer {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         let end_lsn = self.end_lsn_or_max();
@@ -230,11 +226,11 @@ impl std::fmt::Display for InMemoryLayer {
 
 impl InMemoryLayer {
     ///
-    /// Get layer size on the disk
+    /// Get layer size.
     ///
     pub async fn size(&self) -> Result<u64> {
         let inner = self.inner.read().await;
-        Ok(inner.file.size)
+        Ok(inner.file.len())
     }
 
     ///
@@ -269,17 +265,17 @@ impl InMemoryLayer {
     /// Adds the page version to the in-memory tree
     pub async fn put_value(&self, key: Key, lsn: Lsn, val: &Value) -> Result<()> {
         trace!("put_value key {} at {}/{}", key, self.timeline_id, lsn);
-        let mut inner = self.inner.write().await;
+        let inner: &mut _ = &mut *self.inner.write().await;
         self.assert_writable();
 
         let off = {
-            SER_BUFFER.with(|x| -> Result<_> {
-                let mut buf = x.borrow_mut();
-                buf.clear();
-                val.ser_into(&mut (*buf))?;
-                let off = inner.file.write_blob(&buf)?;
-                Ok(off)
-            })?
+            // Avoid doing allocations for "small" values.
+            // In the regression test suite, the limit of 256 avoided allocations in 95% of cases:
+            // https://github.com/neondatabase/neon/pull/5056#discussion_r1301975061
+            let mut buf = smallvec::SmallVec::<[u8; 256]>::new();
+            buf.clear();
+            val.ser_into(&mut buf)?;
+            inner.file.write_blob(&buf).await?
         };
 
         let vec_map = inner.index.entry(key).or_default();
@@ -317,7 +313,7 @@ impl InMemoryLayer {
     /// Write this frozen in-memory layer to disk.
     ///
     /// Returns a new delta layer with all the same data as this in-memory layer
-    pub async fn write_to_disk(&self) -> Result<DeltaLayer> {
+    pub(crate) async fn write_to_disk(&self) -> Result<DeltaLayer> {
         // Grab the lock in read-mode. We hold it over the I/O, but because this
         // layer is not writeable anymore, no one should be trying to acquire the
         // write lock on it, so we shouldn't block anyone. There's one exception

pageserver/src/tenant/storage_layer/layer_desc.rs

@@ -19,16 +19,17 @@ use serde::{Deserialize, Serialize};
 pub struct PersistentLayerDesc {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
+    /// Range of keys that this layer covers
     pub key_range: Range<Key>,
-    /// For image layer, this is `[lsn, lsn+1)`.
+    /// Inclusive start, exclusive end of the LSN range that this layer holds.
+    ///
+    /// - For an open in-memory layer, the end bound is MAX_LSN
+    /// - For a frozen in-memory layer or a delta layer, the end bound is a valid lsn after the
+    /// range start
+    /// - An image layer represents snapshot at one LSN, so end_lsn is always the snapshot LSN + 1
     pub lsn_range: Range<Lsn>,
-    /// Whether this is a delta layer.
+    /// Whether this is a delta layer, and also, is this incremental.
     pub is_delta: bool,
-    /// Whether this layer only contains page images for part of the keys in the range. In the current implementation, this should
-    /// always be equal to `is_delta`. If we land the partial image layer PR someday, image layer could also be
-    /// incremental.
-    pub is_incremental: bool,
-    /// File size
     pub file_size: u64,
 }
 
@@ -61,7 +62,6 @@ impl PersistentLayerDesc {
             key_range,
             lsn_range: Lsn(0)..Lsn(1),
             is_delta: false,
-            is_incremental: false,
             file_size: 0,
         }
     }
@@ -71,7 +71,6 @@ impl PersistentLayerDesc {
         timeline_id: TimelineId,
         key_range: Range<Key>,
         lsn: Lsn,
-        is_incremental: bool,
         file_size: u64,
     ) -> Self {
         Self {
@@ -80,7 +79,6 @@ impl PersistentLayerDesc {
             key_range,
             lsn_range: Self::image_layer_lsn_range(lsn),
             is_delta: false,
-            is_incremental,
             file_size,
         }
     }
@@ -98,7 +96,6 @@ impl PersistentLayerDesc {
             key_range,
             lsn_range,
             is_delta: true,
-            is_incremental: true,
             file_size,
         }
     }
@@ -164,8 +161,12 @@ impl PersistentLayerDesc {
         self.tenant_id
     }
 
+    /// Does this layer only contain some data for the key-range (incremental),
+    /// or does it contain a version of every page? This is important to know
+    /// for garbage collecting old layers: an incremental layer depends on
+    /// the previous non-incremental layer.
     pub fn is_incremental(&self) -> bool {
-        self.is_incremental
+        self.is_delta
     }
 
     pub fn is_delta(&self) -> bool {
@@ -182,7 +183,7 @@ impl PersistentLayerDesc {
             self.lsn_range.start,
             self.lsn_range.end,
             self.is_delta,
-            self.is_incremental,
+            self.is_incremental(),
             self.file_size,
         );
 

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -60,7 +60,7 @@ impl std::fmt::Debug for RemoteLayer {
         f.debug_struct("RemoteLayer")
             .field("file_name", &self.desc.filename())
             .field("layer_metadata", &self.layer_metadata)
-            .field("is_incremental", &self.desc.is_incremental)
+            .field("is_incremental", &self.desc.is_incremental())
             .finish()
     }
 }
@@ -76,39 +76,6 @@ impl Layer for RemoteLayer {
     ) -> Result<ValueReconstructResult> {
         bail!("layer {self} needs to be downloaded");
     }
-
-    /// debugging function to print out the contents of the layer
-    async fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
-        println!(
-            "----- remote layer for ten {} tli {} keys {}-{} lsn {}-{} is_delta {} is_incremental {} size {} ----",
-            self.desc.tenant_id,
-            self.desc.timeline_id,
-            self.desc.key_range.start,
-            self.desc.key_range.end,
-            self.desc.lsn_range.start,
-            self.desc.lsn_range.end,
-            self.desc.is_delta,
-            self.desc.is_incremental,
-            self.desc.file_size,
-        );
-
-        Ok(())
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_key_range(&self) -> Range<Key> {
-        self.layer_desc().key_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn get_lsn_range(&self) -> Range<Lsn> {
-        self.layer_desc().lsn_range.clone()
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    fn is_incremental(&self) -> bool {
-        self.layer_desc().is_incremental
-    }
 }
 
 /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
@@ -142,8 +109,8 @@ impl PersistentLayer for RemoteLayer {
     }
 
     fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
-        let layer_file_name = self.filename().file_name();
-        let lsn_range = self.get_lsn_range();
+        let layer_file_name = self.layer_desc().filename().file_name();
+        let lsn_range = self.layer_desc().lsn_range.clone();
 
         if self.desc.is_delta {
             HistoricLayerInfo::Delta {
@@ -184,7 +151,6 @@ impl RemoteLayer {
                 timelineid,
                 fname.key_range.clone(),
                 fname.lsn,
-                false,
                 layer_metadata.file_size(),
             ),
             layer_metadata: layer_metadata.clone(),
@@ -217,9 +183,9 @@ impl RemoteLayer {
     }
 
     /// Create a Layer struct representing this layer, after it has been downloaded.
-    pub fn create_downloaded_layer(
+    pub(crate) fn create_downloaded_layer(
         &self,
-        layer_map_lock_held_witness: &LayerManager,
+        _layer_map_lock_held_witness: &LayerManager,
         conf: &'static PageServerConf,
         file_size: u64,
     ) -> Arc<dyn PersistentLayer> {
@@ -231,10 +197,8 @@ impl RemoteLayer {
                 self.desc.tenant_id,
                 &fname,
                 file_size,
-                self.access_stats.clone_for_residence_change(
-                    layer_map_lock_held_witness,
-                    LayerResidenceStatus::Resident,
-                ),
+                self.access_stats
+                    .clone_for_residence_change(LayerResidenceStatus::Resident),
             ))
         } else {
             let fname = self.desc.image_file_name();
@@ -244,10 +208,8 @@ impl RemoteLayer {
                 self.desc.tenant_id,
                 &fname,
                 file_size,
-                self.access_stats.clone_for_residence_change(
-                    layer_map_lock_held_witness,
-                    LayerResidenceStatus::Resident,
-                ),
+                self.access_stats
+                    .clone_for_residence_change(LayerResidenceStatus::Resident),
             ))
         }
     }

pageserver/src/tenant/timeline.rs

@@ -1,5 +1,6 @@
 pub mod delete;
 mod eviction_task;
+mod init;
 pub mod layer_manager;
 mod logical_size;
 pub mod span;
@@ -27,7 +28,6 @@ use utils::id::TenantTimelineId;
 
 use std::cmp::{max, min, Ordering};
 use std::collections::{BinaryHeap, HashMap, HashSet};
-use std::fs;
 use std::ops::{Deref, Range};
 use std::path::{Path, PathBuf};
 use std::pin::pin;
@@ -38,15 +38,13 @@ use std::time::{Duration, Instant, SystemTime};
 use crate::context::{
     AccessStatsBehavior, DownloadBehavior, RequestContext, RequestContextBuilder,
 };
-use crate::tenant::remote_timeline_client::{self, index::LayerFileMetadata};
+use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
 use crate::tenant::storage_layer::delta_layer::DeltaEntry;
 use crate::tenant::storage_layer::{
-    DeltaFileName, DeltaLayerWriter, ImageFileName, ImageLayerWriter, InMemoryLayer,
-    LayerAccessStats, LayerFileName, RemoteLayer,
+    DeltaLayerWriter, ImageLayerWriter, InMemoryLayer, LayerAccessStats, LayerFileName, RemoteLayer,
 };
 use crate::tenant::timeline::logical_size::CurrentLogicalSize;
 use crate::tenant::{
-    ephemeral_file::is_ephemeral_file,
     layer_map::{LayerMap, SearchResult},
     metadata::{save_metadata, TimelineMetadata},
     par_fsync,
@@ -78,11 +76,10 @@ use utils::{
 use crate::page_cache;
 use crate::repository::GcResult;
 use crate::repository::{Key, Value};
+use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::walredo::WalRedoManager;
-use crate::METADATA_FILE_NAME;
 use crate::ZERO_PAGE;
-use crate::{is_temporary, task_mgr};
 
 use self::delete::DeleteTimelineFlow;
 pub(super) use self::eviction_task::EvictionTaskTenantState;
@@ -95,7 +92,7 @@ use super::config::TenantConf;
 use super::remote_timeline_client::index::IndexPart;
 use super::remote_timeline_client::RemoteTimelineClient;
 use super::storage_layer::{
-    AsLayerDesc, DeltaLayer, ImageLayer, Layer, LayerAccessStatsReset, PersistentLayerDesc,
+    AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStatsReset, PersistentLayerDesc,
 };
 
 #[derive(Debug, PartialEq, Eq, Clone, Copy)]
@@ -140,6 +137,12 @@ fn drop_rlock<T>(rlock: tokio::sync::OwnedRwLockReadGuard<T>) {
 fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
     drop(rlock)
 }
+
+/// The outward-facing resources required to build a Timeline
+pub struct TimelineResources {
+    pub remote_client: Option<RemoteTimelineClient>,
+}
+
 pub struct Timeline {
     conf: &'static PageServerConf,
     tenant_conf: Arc<RwLock<TenantConfOpt>>,
@@ -1205,7 +1208,7 @@ impl Timeline {
                 &layer_metadata,
                 local_layer
                     .access_stats()
-                    .clone_for_residence_change(layer_mgr, LayerResidenceStatus::Evicted),
+                    .clone_for_residence_change(LayerResidenceStatus::Evicted),
             ),
             LayerFileName::Delta(delta_name) => RemoteLayer::new_delta(
                 self.tenant_id,
@@ -1214,7 +1217,7 @@ impl Timeline {
                 &layer_metadata,
                 local_layer
                     .access_stats()
-                    .clone_for_residence_change(layer_mgr, LayerResidenceStatus::Evicted),
+                    .clone_for_residence_change(LayerResidenceStatus::Evicted),
             ),
         });
 
@@ -1374,7 +1377,7 @@ impl Timeline {
         timeline_id: TimelineId,
         tenant_id: TenantId,
         walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
-        remote_client: Option<RemoteTimelineClient>,
+        resources: TimelineResources,
         pg_version: u32,
         initial_logical_size_can_start: Option<completion::Barrier>,
         initial_logical_size_attempt: Option<completion::Completion>,
@@ -1409,7 +1412,7 @@ impl Timeline {
                 walredo_mgr,
                 walreceiver: Mutex::new(None),
 
-                remote_client: remote_client.map(Arc::new),
+                remote_client: resources.remote_client.map(Arc::new),
 
                 // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
                 last_record_lsn: SeqWait::new(RecordLsn {
@@ -1512,7 +1515,7 @@ impl Timeline {
         let layer_flush_start_rx = self.layer_flush_start_tx.subscribe();
         let self_clone = Arc::clone(self);
 
-        info!("spawning flush loop");
+        debug!("spawning flush loop");
         *flush_loop_state = FlushLoopState::Running {
             #[cfg(test)]
             expect_initdb_optimization: false,
@@ -1583,9 +1586,7 @@ impl Timeline {
         ));
     }
 
-    ///
     /// Initialize with an empty layer map. Used when creating a new timeline.
-    ///
     pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
         let mut layers = self.layers.try_write().expect(
             "in the context where we call this function, no other task has access to the object",
@@ -1593,10 +1594,16 @@ impl Timeline {
         layers.initialize_empty(Lsn(start_lsn.0));
     }
 
-    ///
-    /// Scan the timeline directory to populate the layer map.
-    ///
-    pub(super) async fn load_layer_map(&self, disk_consistent_lsn: Lsn) -> anyhow::Result<()> {
+    /// Scan the timeline directory, cleanup, populate the layer map, and schedule uploads for local-only
+    /// files.
+    pub(super) async fn load_layer_map(
+        &self,
+        disk_consistent_lsn: Lsn,
+        index_part: Option<IndexPart>,
+    ) -> anyhow::Result<()> {
+        use init::{Decision::*, Discovered, FutureLayer};
+        use LayerFileName::*;
+
         let mut guard = self.layers.write().await;
 
         let timer = self.metrics.load_layer_map_histo.start_timer();
@@ -1604,102 +1611,153 @@ impl Timeline {
         // Scan timeline directory and create ImageFileName and DeltaFilename
         // structs representing all files on disk
         let timeline_path = self.conf.timeline_path(&self.tenant_id, &self.timeline_id);
-        // total size of layer files in the current timeline directory
-        let mut total_physical_size = 0;
+        let (conf, tenant_id, timeline_id) = (self.conf, self.tenant_id, self.timeline_id);
+        let span = tracing::Span::current();
 
-        let mut loaded_layers = Vec::<Arc<dyn PersistentLayer>>::new();
+        let (loaded_layers, to_sync, total_physical_size) = tokio::task::spawn_blocking({
+            move || {
+                let _g = span.entered();
+                let discovered = init::scan_timeline_dir(&timeline_path)?;
+                let mut discovered_layers = Vec::with_capacity(discovered.len());
+                let mut unrecognized_files = Vec::new();
 
-        for direntry in fs::read_dir(timeline_path)? {
-            let direntry = direntry?;
-            let direntry_path = direntry.path();
-            let fname = direntry.file_name();
-            let fname = fname.to_string_lossy();
+                let mut path = timeline_path;
 
-            if let Some(filename) = ImageFileName::parse_str(&fname) {
-                // create an ImageLayer struct for each image file.
-                if filename.lsn > disk_consistent_lsn {
-                    info!(
-                        "found future image layer {} on timeline {} disk_consistent_lsn is {}",
-                        filename, self.timeline_id, disk_consistent_lsn
-                    );
-
-                    rename_to_backup(&direntry_path)?;
-                    continue;
+                for discovered in discovered {
+                    let (name, kind) = match discovered {
+                        Discovered::Layer(file_name, file_size) => {
+                            discovered_layers.push((file_name, file_size));
+                            continue;
+                        }
+                        Discovered::Metadata | Discovered::IgnoredBackup => {
+                            continue;
+                        }
+                        Discovered::Unknown(file_name) => {
+                            // we will later error if there are any
+                            unrecognized_files.push(file_name);
+                            continue;
+                        }
+                        Discovered::Ephemeral(name) => (name, "old ephemeral file"),
+                        Discovered::Temporary(name) => (name, "temporary timeline file"),
+                        Discovered::TemporaryDownload(name) => (name, "temporary download"),
+                    };
+                    path.push(name);
+                    init::cleanup(&path, kind)?;
+                    path.pop();
                 }
 
-                let file_size = direntry_path.metadata()?.len();
-                let stats =
-                    LayerAccessStats::for_loading_layer(&guard, LayerResidenceStatus::Resident);
-
-                let layer = ImageLayer::new(
-                    self.conf,
-                    self.timeline_id,
-                    self.tenant_id,
-                    &filename,
-                    file_size,
-                    stats,
-                );
-
-                total_physical_size += file_size;
-                loaded_layers.push(Arc::new(layer));
-            } else if let Some(filename) = DeltaFileName::parse_str(&fname) {
-                // Create a DeltaLayer struct for each delta file.
-                // The end-LSN is exclusive, while disk_consistent_lsn is
-                // inclusive. For example, if disk_consistent_lsn is 100, it is
-                // OK for a delta layer to have end LSN 101, but if the end LSN
-                // is 102, then it might not have been fully flushed to disk
-                // before crash.
-                if filename.lsn_range.end > disk_consistent_lsn + 1 {
-                    info!(
-                        "found future delta layer {} on timeline {} disk_consistent_lsn is {}",
-                        filename, self.timeline_id, disk_consistent_lsn
+                if !unrecognized_files.is_empty() {
+                    // assume that if there are any there are many many.
+                    let n = unrecognized_files.len();
+                    let first = &unrecognized_files[..n.min(10)];
+                    anyhow::bail!(
+                        "unrecognized files in timeline dir (total {n}), first 10: {first:?}"
                     );
-
-                    rename_to_backup(&direntry_path)?;
-                    continue;
                 }
 
-                let file_size = direntry_path.metadata()?.len();
-                let stats =
-                    LayerAccessStats::for_loading_layer(&guard, LayerResidenceStatus::Resident);
+                let decided =
+                    init::reconcile(discovered_layers, index_part.as_ref(), disk_consistent_lsn);
 
-                let layer = DeltaLayer::new(
-                    self.conf,
-                    self.timeline_id,
-                    self.tenant_id,
-                    &filename,
-                    file_size,
-                    stats,
-                );
+                let mut loaded_layers = Vec::new();
+                let mut needs_upload = Vec::new();
+                let mut needs_cleanup = Vec::new();
+                let mut total_physical_size = 0;
 
-                total_physical_size += file_size;
-                loaded_layers.push(Arc::new(layer));
-            } else if fname == METADATA_FILE_NAME || fname.ends_with(".old") {
-                // ignore these
-            } else if remote_timeline_client::is_temp_download_file(&direntry_path) {
-                info!(
-                    "skipping temp download file, reconcile_with_remote will resume / clean up: {}",
-                    fname
-                );
-            } else if is_ephemeral_file(&fname) {
-                // Delete any old ephemeral files
-                trace!("deleting old ephemeral file in timeline dir: {}", fname);
-                fs::remove_file(&direntry_path)?;
-            } else if is_temporary(&direntry_path) {
-                info!("removing temp timeline file at {}", direntry_path.display());
-                fs::remove_file(&direntry_path).with_context(|| {
-                    format!(
-                        "failed to remove temp download file at {}",
-                        direntry_path.display()
-                    )
-                })?;
-            } else {
-                warn!("unrecognized filename in timeline dir: {}", fname);
+                for (name, decision) in decided {
+                    let decision = match decision {
+                        Ok(UseRemote { local, remote }) => {
+                            path.push(name.file_name());
+                            init::cleanup_local_file_for_remote(&path, &local, &remote)?;
+                            path.pop();
+
+                            UseRemote { local, remote }
+                        }
+                        Ok(decision) => decision,
+                        Err(FutureLayer { local }) => {
+                            if local.is_some() {
+                                path.push(name.file_name());
+                                init::cleanup_future_layer(&path, &name, disk_consistent_lsn)?;
+                                path.pop();
+                            }
+                            needs_cleanup.push(name);
+                            continue;
+                        }
+                    };
+
+                    match &name {
+                        Delta(d) => assert!(d.lsn_range.end <= disk_consistent_lsn + 1),
+                        Image(i) => assert!(i.lsn <= disk_consistent_lsn),
+                    }
+
+                    let status = match &decision {
+                        UseLocal(_) | NeedsUpload(_) => LayerResidenceStatus::Resident,
+                        Evicted(_) | UseRemote { .. } => LayerResidenceStatus::Evicted,
+                    };
+
+                    let stats = LayerAccessStats::for_loading_layer(status);
+
+                    let layer: Arc<dyn PersistentLayer> = match (name, &decision) {
+                        (Delta(d), UseLocal(m) | NeedsUpload(m)) => {
+                            total_physical_size += m.file_size();
+                            Arc::new(DeltaLayer::new(
+                                conf,
+                                timeline_id,
+                                tenant_id,
+                                &d,
+                                m.file_size(),
+                                stats,
+                            ))
+                        }
+                        (Image(i), UseLocal(m) | NeedsUpload(m)) => {
+                            total_physical_size += m.file_size();
+                            Arc::new(ImageLayer::new(
+                                conf,
+                                timeline_id,
+                                tenant_id,
+                                &i,
+                                m.file_size(),
+                                stats,
+                            ))
+                        }
+                        (Delta(d), Evicted(remote) | UseRemote { remote, .. }) => Arc::new(
+                            RemoteLayer::new_delta(tenant_id, timeline_id, &d, remote, stats),
+                        ),
+                        (Image(i), Evicted(remote) | UseRemote { remote, .. }) => Arc::new(
+                            RemoteLayer::new_img(tenant_id, timeline_id, &i, remote, stats),
+                        ),
+                    };
+
+                    if let NeedsUpload(m) = decision {
+                        needs_upload.push((layer.clone(), m));
+                    }
+
+                    loaded_layers.push(layer);
+                }
+                Ok((
+                    loaded_layers,
+                    (needs_upload, needs_cleanup),
+                    total_physical_size,
+                ))
             }
-        }
+        })
+        .await
+        .map_err(anyhow::Error::new)
+        .and_then(|x| x)?;
 
         let num_layers = loaded_layers.len();
-        guard.initialize_local_layers(loaded_layers, Lsn(disk_consistent_lsn.0) + 1);
+
+        guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
+
+        if let Some(rtc) = self.remote_client.as_ref() {
+            let (needs_upload, needs_cleanup) = to_sync;
+            for (layer, m) in needs_upload {
+                rtc.schedule_layer_file_upload(&layer.layer_desc().filename(), &m)?;
+            }
+            rtc.schedule_layer_file_deletion(&needs_cleanup)?;
+            rtc.schedule_index_upload_for_file_changes()?;
+            // Tenant::create_timeline will wait for these uploads to happen before returning, or
+            // on retry.
+        }
 
         info!(
             "loaded layer map with {} layers at {}, total physical size: {}",
@@ -1710,236 +1768,6 @@ impl Timeline {
             .set(total_physical_size);
 
         timer.stop_and_record();
-
-        Ok(())
-    }
-
-    async fn create_remote_layers(
-        &self,
-        index_part: &IndexPart,
-        local_layers: HashMap<LayerFileName, Arc<dyn PersistentLayer>>,
-        up_to_date_disk_consistent_lsn: Lsn,
-    ) -> anyhow::Result<HashMap<LayerFileName, Arc<dyn PersistentLayer>>> {
-        // Are we missing some files that are present in remote storage?
-        // Create RemoteLayer instances for them.
-        let mut local_only_layers = local_layers;
-
-        // We're holding a layer map lock for a while but this
-        // method is only called during init so it's fine.
-        let mut guard = self.layers.write().await;
-
-        let mut corrupted_local_layers = Vec::new();
-        let mut added_remote_layers = Vec::new();
-        for remote_layer_name in &index_part.timeline_layers {
-            let local_layer = local_only_layers.remove(remote_layer_name);
-
-            let remote_layer_metadata = index_part
-                .layer_metadata
-                .get(remote_layer_name)
-                .map(LayerFileMetadata::from)
-                .with_context(|| {
-                    format!(
-                        "No remote layer metadata found for layer {}",
-                        remote_layer_name.file_name()
-                    )
-                })?;
-
-            // Is the local layer's size different from the size stored in the
-            // remote index file?
-            // If so, rename_to_backup those files & replace their local layer with
-            // a RemoteLayer in the layer map so that we re-download them on-demand.
-            if let Some(local_layer) = local_layer {
-                let local_layer_path = local_layer
-                    .local_path()
-                    .expect("caller must ensure that local_layers only contains local layers");
-                ensure!(
-                    local_layer_path.exists(),
-                    "every layer from local_layers must exist on disk: {}",
-                    local_layer_path.display()
-                );
-
-                let remote_size = remote_layer_metadata.file_size();
-                let metadata = local_layer_path.metadata().with_context(|| {
-                    format!(
-                        "get file size of local layer {}",
-                        local_layer_path.display()
-                    )
-                })?;
-                let local_size = metadata.len();
-                if local_size != remote_size {
-                    warn!("removing local file {local_layer_path:?} because it has unexpected length {local_size}; length in remote index is {remote_size}");
-                    if let Err(err) = rename_to_backup(&local_layer_path) {
-                        assert!(local_layer_path.exists(), "we would leave the local_layer without a file if this does not hold: {}", local_layer_path.display());
-                        anyhow::bail!("could not rename file {local_layer_path:?}: {err:?}");
-                    } else {
-                        self.metrics.resident_physical_size_gauge.sub(local_size);
-                        corrupted_local_layers.push(local_layer);
-                        // fall-through to adding the remote layer
-                    }
-                } else {
-                    debug!(
-                        "layer is present locally and file size matches remote, using it: {}",
-                        local_layer_path.display()
-                    );
-                    continue;
-                }
-            }
-
-            info!(
-                "remote layer does not exist locally, creating remote layer: {}",
-                remote_layer_name.file_name()
-            );
-
-            match remote_layer_name {
-                LayerFileName::Image(imgfilename) => {
-                    if imgfilename.lsn > up_to_date_disk_consistent_lsn {
-                        info!(
-                        "found future image layer {} on timeline {} remote_consistent_lsn is {}",
-                        imgfilename, self.timeline_id, up_to_date_disk_consistent_lsn
-                    );
-                        continue;
-                    }
-                    let stats =
-                        LayerAccessStats::for_loading_layer(&guard, LayerResidenceStatus::Evicted);
-
-                    let remote_layer = RemoteLayer::new_img(
-                        self.tenant_id,
-                        self.timeline_id,
-                        imgfilename,
-                        &remote_layer_metadata,
-                        stats,
-                    );
-                    let remote_layer = Arc::new(remote_layer);
-                    added_remote_layers.push(remote_layer);
-                }
-                LayerFileName::Delta(deltafilename) => {
-                    // Create a RemoteLayer for the delta file.
-                    // The end-LSN is exclusive, while disk_consistent_lsn is
-                    // inclusive. For example, if disk_consistent_lsn is 100, it is
-                    // OK for a delta layer to have end LSN 101, but if the end LSN
-                    // is 102, then it might not have been fully flushed to disk
-                    // before crash.
-                    if deltafilename.lsn_range.end > up_to_date_disk_consistent_lsn + 1 {
-                        info!(
-                            "found future delta layer {} on timeline {} remote_consistent_lsn is {}",
-                            deltafilename, self.timeline_id, up_to_date_disk_consistent_lsn
-                        );
-                        continue;
-                    }
-                    let stats =
-                        LayerAccessStats::for_loading_layer(&guard, LayerResidenceStatus::Evicted);
-
-                    let remote_layer = RemoteLayer::new_delta(
-                        self.tenant_id,
-                        self.timeline_id,
-                        deltafilename,
-                        &remote_layer_metadata,
-                        stats,
-                    );
-                    let remote_layer = Arc::new(remote_layer);
-                    added_remote_layers.push(remote_layer);
-                }
-            }
-        }
-        guard.initialize_remote_layers(corrupted_local_layers, added_remote_layers);
-        Ok(local_only_layers)
-    }
-
-    /// This function will synchronize local state with what we have in remote storage.
-    ///
-    /// Steps taken:
-    /// 1. Initialize upload queue based on `index_part`.
-    /// 2. Create `RemoteLayer` instances for layers that exist only on the remote.
-    ///    The list of layers on the remote comes from `index_part`.
-    ///    The list of local layers is given by the layer map's `iter_historic_layers()`.
-    ///    So, the layer map must have been loaded already.
-    /// 3. Schedule upload of local-only layer files (which will then also update the remote
-    ///    IndexPart to include the new layer files).
-    ///
-    /// Refer to the [`remote_timeline_client`] module comment for more context.
-    ///
-    /// # TODO
-    /// May be a bit cleaner to do things based on populated remote client,
-    /// and then do things based on its upload_queue.latest_files.
-    #[instrument(skip(self, index_part, up_to_date_metadata))]
-    pub async fn reconcile_with_remote(
-        &self,
-        up_to_date_metadata: &TimelineMetadata,
-        index_part: Option<&IndexPart>,
-    ) -> anyhow::Result<()> {
-        info!("starting");
-        let remote_client = self
-            .remote_client
-            .as_ref()
-            .ok_or_else(|| anyhow!("cannot download without remote storage"))?;
-
-        let disk_consistent_lsn = up_to_date_metadata.disk_consistent_lsn();
-
-        let local_layers = {
-            let guard = self.layers.read().await;
-            let layers = guard.layer_map();
-            layers
-                .iter_historic_layers()
-                .map(|l| (l.filename(), guard.get_from_desc(&l)))
-                .collect::<HashMap<_, _>>()
-        };
-
-        // If no writes happen, new branches do not have any layers, only the metadata file.
-        let has_local_layers = !local_layers.is_empty();
-        let local_only_layers = match index_part {
-            Some(index_part) => {
-                info!(
-                    "initializing upload queue from remote index with {} layer files",
-                    index_part.timeline_layers.len()
-                );
-                remote_client.init_upload_queue(index_part)?;
-                self.create_remote_layers(index_part, local_layers, disk_consistent_lsn)
-                    .await?
-            }
-            None => {
-                info!("initializing upload queue as empty");
-                remote_client.init_upload_queue_for_empty_remote(up_to_date_metadata)?;
-                local_layers
-            }
-        };
-
-        if has_local_layers {
-            // Are there local files that don't exist remotely? Schedule uploads for them.
-            // Local timeline metadata will get uploaded to remove along witht he layers.
-            for (layer_name, layer) in &local_only_layers {
-                // XXX solve this in the type system
-                let layer_path = layer
-                    .local_path()
-                    .expect("local_only_layers only contains local layers");
-                let layer_size = layer_path
-                    .metadata()
-                    .with_context(|| format!("failed to get file {layer_path:?} metadata"))?
-                    .len();
-                info!("scheduling {layer_path:?} for upload");
-                remote_client
-                    .schedule_layer_file_upload(layer_name, &LayerFileMetadata::new(layer_size))?;
-            }
-            remote_client.schedule_index_upload_for_file_changes()?;
-        } else if index_part.is_none() {
-            // No data on the remote storage, no local layers, local metadata file.
-            //
-            // TODO https://github.com/neondatabase/neon/issues/3865
-            // Currently, console does not wait for the timeline data upload to the remote storage
-            // and considers the timeline created, expecting other pageserver nodes to work with it.
-            // Branch metadata upload could get interrupted (e.g pageserver got killed),
-            // hence any locally existing branch metadata with no remote counterpart should be uploaded,
-            // otherwise any other pageserver won't see the branch on `attach`.
-            //
-            // After the issue gets implemented, pageserver should rather remove the branch,
-            // since absence on S3 means we did not acknowledge the branch creation and console will have to retry,
-            // no need to keep the old files.
-            remote_client.schedule_index_upload_for_metadata_update(up_to_date_metadata)?;
-        } else {
-            // Local timeline has a metadata file, remote one too, both have no layers to sync.
-        }
-
-        info!("Done");
-
         Ok(())
     }
 
@@ -2846,7 +2674,6 @@ impl Timeline {
             if let Some(ref l) = delta_layer_to_add {
                 // TODO: move access stats, metrics update, etc. into layer manager.
                 l.access_stats().record_residence_event(
-                    &guard,
                     LayerResidenceStatus::Resident,
                     LayerResidenceEventReason::LayerCreate,
                 );
@@ -3143,7 +2970,6 @@ impl Timeline {
                     self.tenant_id,
                     &img_range,
                     lsn,
-                    false, // image layer always covers the full range
                 )?;
 
                 fail_point!("image-layer-writer-fail-before-finish", |_| {
@@ -3236,7 +3062,6 @@ impl Timeline {
                 .add(metadata.len());
             let l = Arc::new(l);
             l.access_stats().record_residence_event(
-                &guard,
                 LayerResidenceStatus::Resident,
                 LayerResidenceEventReason::LayerCreate,
             );
@@ -3404,8 +3229,8 @@ impl Timeline {
     /// start of level0 files compaction, the on-demand download should be revisited as well.
     ///
     /// [`compact_inner`]: Self::compact_inner
-    fn compact_level0_phase1(
-        self: Arc<Self>,
+    async fn compact_level0_phase1(
+        self: &Arc<Self>,
         _layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
         guard: tokio::sync::OwnedRwLockReadGuard<LayerManager>,
         mut stats: CompactLevel0Phase1StatsBuilder,
@@ -3556,7 +3381,7 @@ impl Timeline {
             .collect();
         for dl in downcast_deltas.iter() {
             // TODO: replace this with an await once we fully go async
-            all_keys.extend(Handle::current().block_on(DeltaLayer::load_keys(dl, ctx))?);
+            all_keys.extend(DeltaLayer::load_keys(dl, ctx).await?);
         }
 
         // The current stdlib sorting implementation is designed in a way where it is
@@ -3670,107 +3495,103 @@ impl Timeline {
         let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
         let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
 
-        // TODO remove this block_on wrapper once we fully go async
-        Handle::current().block_on(async {
-            for &DeltaEntry {
-                key, lsn, ref val, ..
-            } in all_values_iter
-            {
-                let value = val.load().await?;
-                let same_key = prev_key.map_or(false, |prev_key| prev_key == key);
-                // We need to check key boundaries once we reach next key or end of layer with the same key
-                if !same_key || lsn == dup_end_lsn {
-                    let mut next_key_size = 0u64;
-                    let is_dup_layer = dup_end_lsn.is_valid();
-                    dup_start_lsn = Lsn::INVALID;
-                    if !same_key {
-                        dup_end_lsn = Lsn::INVALID;
-                    }
-                    // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
-                    for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
-                        next_key_size = next_size;
-                        if key != next_key {
-                            if dup_end_lsn.is_valid() {
-                                // We are writting segment with duplicates:
-                                // place all remaining values of this key in separate segment
-                                dup_start_lsn = dup_end_lsn; // new segments starts where old stops
-                                dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
-                            }
-                            break;
-                        }
-                        key_values_total_size += next_size;
-                        // Check if it is time to split segment: if total keys size is larger than target file size.
-                        // We need to avoid generation of empty segments if next_size > target_file_size.
-                        if key_values_total_size > target_file_size && lsn != next_lsn {
-                            // Split key between multiple layers: such layer can contain only single key
-                            dup_start_lsn = if dup_end_lsn.is_valid() {
-                                dup_end_lsn // new segment with duplicates starts where old one stops
-                            } else {
-                                lsn // start with the first LSN for this key
-                            };
-                            dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
-                            break;
-                        }
-                    }
-                    // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
-                    if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
-                        dup_start_lsn = dup_end_lsn;
-                        dup_end_lsn = lsn_range.end;
-                    }
-                    if writer.is_some() {
-                        let written_size = writer.as_mut().unwrap().size();
-                        let contains_hole =
-                            next_hole < holes.len() && key >= holes[next_hole].key_range.end;
-                        // check if key cause layer overflow or contains hole...
-                        if is_dup_layer
-                            || dup_end_lsn.is_valid()
-                            || written_size + key_values_total_size > target_file_size
-                            || contains_hole
-                        {
-                            // ... if so, flush previous layer and prepare to write new one
-                            new_layers.push(Arc::new(
-                                writer.take().unwrap().finish(prev_key.unwrap().next())?,
-                            ));
-                            writer = None;
-
-                            if contains_hole {
-                                // skip hole
-                                next_hole += 1;
-                            }
-                        }
-                    }
-                    // Remember size of key value because at next iteration we will access next item
-                    key_values_total_size = next_key_size;
+        for &DeltaEntry {
+            key, lsn, ref val, ..
+        } in all_values_iter
+        {
+            let value = val.load().await?;
+            let same_key = prev_key.map_or(false, |prev_key| prev_key == key);
+            // We need to check key boundaries once we reach next key or end of layer with the same key
+            if !same_key || lsn == dup_end_lsn {
+                let mut next_key_size = 0u64;
+                let is_dup_layer = dup_end_lsn.is_valid();
+                dup_start_lsn = Lsn::INVALID;
+                if !same_key {
+                    dup_end_lsn = Lsn::INVALID;
                 }
-                if writer.is_none() {
-                    // Create writer if not initiaized yet
-                    writer = Some(DeltaLayerWriter::new(
-                        self.conf,
-                        self.timeline_id,
-                        self.tenant_id,
-                        key,
+                // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
+                for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
+                    next_key_size = next_size;
+                    if key != next_key {
                         if dup_end_lsn.is_valid() {
-                            // this is a layer containing slice of values of the same key
-                            debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
-                            dup_start_lsn..dup_end_lsn
+                            // We are writting segment with duplicates:
+                            // place all remaining values of this key in separate segment
+                            dup_start_lsn = dup_end_lsn; // new segments starts where old stops
+                            dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
+                        }
+                        break;
+                    }
+                    key_values_total_size += next_size;
+                    // Check if it is time to split segment: if total keys size is larger than target file size.
+                    // We need to avoid generation of empty segments if next_size > target_file_size.
+                    if key_values_total_size > target_file_size && lsn != next_lsn {
+                        // Split key between multiple layers: such layer can contain only single key
+                        dup_start_lsn = if dup_end_lsn.is_valid() {
+                            dup_end_lsn // new segment with duplicates starts where old one stops
                         } else {
-                            debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
-                            lsn_range.clone()
-                        },
-                    )?);
+                            lsn // start with the first LSN for this key
+                        };
+                        dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
+                        break;
+                    }
                 }
+                // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
+                if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
+                    dup_start_lsn = dup_end_lsn;
+                    dup_end_lsn = lsn_range.end;
+                }
+                if writer.is_some() {
+                    let written_size = writer.as_mut().unwrap().size();
+                    let contains_hole =
+                        next_hole < holes.len() && key >= holes[next_hole].key_range.end;
+                    // check if key cause layer overflow or contains hole...
+                    if is_dup_layer
+                        || dup_end_lsn.is_valid()
+                        || written_size + key_values_total_size > target_file_size
+                        || contains_hole
+                    {
+                        // ... if so, flush previous layer and prepare to write new one
+                        new_layers.push(Arc::new(
+                            writer.take().unwrap().finish(prev_key.unwrap().next())?,
+                        ));
+                        writer = None;
 
-                fail_point!("delta-layer-writer-fail-before-finish", |_| {
-                    Result::<_>::Err(anyhow::anyhow!(
-                        "failpoint delta-layer-writer-fail-before-finish"
-                    ))
-                });
-
-                writer.as_mut().unwrap().put_value(key, lsn, value)?;
-                prev_key = Some(key);
+                        if contains_hole {
+                            // skip hole
+                            next_hole += 1;
+                        }
+                    }
+                }
+                // Remember size of key value because at next iteration we will access next item
+                key_values_total_size = next_key_size;
             }
-            Ok(())
-        })?;
+            if writer.is_none() {
+                // Create writer if not initiaized yet
+                writer = Some(DeltaLayerWriter::new(
+                    self.conf,
+                    self.timeline_id,
+                    self.tenant_id,
+                    key,
+                    if dup_end_lsn.is_valid() {
+                        // this is a layer containing slice of values of the same key
+                        debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
+                        dup_start_lsn..dup_end_lsn
+                    } else {
+                        debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
+                        lsn_range.clone()
+                    },
+                )?);
+            }
+
+            fail_point!("delta-layer-writer-fail-before-finish", |_| {
+                Err(CompactionError::Other(anyhow::anyhow!(
+                    "failpoint delta-layer-writer-fail-before-finish"
+                )))
+            });
+
+            writer.as_mut().unwrap().put_value(key, lsn, value)?;
+            prev_key = Some(key);
+        }
         if let Some(writer) = writer {
             new_layers.push(Arc::new(writer.finish(prev_key.unwrap().next())?));
         }
@@ -3783,10 +3604,10 @@ impl Timeline {
             // we still might easily hit the limit otherwise.
             let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
             for layer in new_layers.iter() {
-                if layer.desc.file_size > warn_limit {
+                if layer.layer_desc().file_size > warn_limit {
                     warn!(
                         %layer,
-                        "created delta file of size {} larger than double of target of {target_file_size}", layer.desc.file_size
+                        "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
                     );
                 }
             }
@@ -3804,7 +3625,7 @@ impl Timeline {
 
         stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
         stats.new_deltas_count = Some(new_layers.len());
-        stats.new_deltas_size = Some(new_layers.iter().map(|l| l.desc.file_size).sum());
+        stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
 
         match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
             .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
@@ -3844,8 +3665,7 @@ impl Timeline {
             deltas_to_compact,
         } = {
             let phase1_span = info_span!("compact_level0_phase1");
-            let myself = Arc::clone(self);
-            let ctx = ctx.attached_child(); // technically, the spawn_blocking can outlive this future
+            let ctx = ctx.attached_child();
             let mut stats = CompactLevel0Phase1StatsBuilder {
                 version: Some(2),
                 tenant_id: Some(self.tenant_id),
@@ -3859,18 +3679,15 @@ impl Timeline {
             stats.read_lock_acquisition_micros =
                 DurationRecorder::Recorded(RecordedDuration(now - begin), now);
             let layer_removal_cs = layer_removal_cs.clone();
-            tokio::task::spawn_blocking(move || {
-                let _entered = phase1_span.enter();
-                myself.compact_level0_phase1(
-                    layer_removal_cs,
-                    phase1_layers_locked,
-                    stats,
-                    target_file_size,
-                    &ctx,
-                )
-            })
-            .await
-            .context("spawn_blocking")??
+            self.compact_level0_phase1(
+                layer_removal_cs,
+                phase1_layers_locked,
+                stats,
+                target_file_size,
+                &ctx,
+            )
+            .instrument(phase1_span)
+            .await?
         };
 
         if new_layers.is_empty() && deltas_to_compact.is_empty() {
@@ -3924,7 +3741,6 @@ impl Timeline {
 
             new_layer_paths.insert(new_delta_path, LayerFileMetadata::new(metadata.len()));
             l.access_stats().record_residence_event(
-                &guard,
                 LayerResidenceStatus::Resident,
                 LayerResidenceEventReason::LayerCreate,
             );
@@ -4843,7 +4659,8 @@ fn rename_to_backup(path: &Path) -> anyhow::Result<()> {
     for i in 0u32.. {
         new_path.set_file_name(format!("{filename}.{i}.old"));
         if !new_path.exists() {
-            std::fs::rename(path, &new_path)?;
+            std::fs::rename(path, &new_path)
+                .with_context(|| format!("rename {path:?} to {new_path:?}"))?;
             return Ok(());
         }
     }

pageserver/src/tenant/timeline/delete.rs

@@ -25,7 +25,7 @@ use crate::{
     InitializationOrder,
 };
 
-use super::Timeline;
+use super::{Timeline, TimelineResources};
 
 /// Now that the Timeline is in Stopping state, request all the related tasks to shut down.
 async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
@@ -416,7 +416,7 @@ impl DeleteTimelineFlow {
                 timeline_id,
                 local_metadata,
                 None, // Ancestor is not needed for deletion.
-                remote_client,
+                TimelineResources { remote_client },
                 init_order,
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.

pageserver/src/tenant/timeline/init.rs

@@ -0,0 +1,199 @@
+use crate::{
+    is_temporary,
+    tenant::{
+        ephemeral_file::is_ephemeral_file,
+        remote_timeline_client::{
+            self,
+            index::{IndexPart, LayerFileMetadata},
+        },
+        storage_layer::LayerFileName,
+    },
+    METADATA_FILE_NAME,
+};
+use anyhow::Context;
+use std::{collections::HashMap, ffi::OsString, path::Path, str::FromStr};
+use utils::lsn::Lsn;
+
+/// Identified files in the timeline directory.
+pub(super) enum Discovered {
+    /// The only one we care about
+    Layer(LayerFileName, u64),
+    /// Old ephmeral files from previous launches, should be removed
+    Ephemeral(OsString),
+    /// Old temporary timeline files, unsure what these really are, should be removed
+    Temporary(OsString),
+    /// Temporary on-demand download files, should be removed
+    TemporaryDownload(OsString),
+    /// "metadata" file we persist locally and include in `index_part.json`
+    Metadata,
+    /// Backup file from previously future layers
+    IgnoredBackup,
+    /// Unrecognized, warn about these
+    Unknown(OsString),
+}
+
+/// Scans the timeline directory for interesting files.
+pub(super) fn scan_timeline_dir(path: &Path) -> anyhow::Result<Vec<Discovered>> {
+    let mut ret = Vec::new();
+
+    for direntry in std::fs::read_dir(path)? {
+        let direntry = direntry?;
+        let direntry_path = direntry.path();
+        let file_name = direntry.file_name();
+
+        let fname = file_name.to_string_lossy();
+
+        let discovered = match LayerFileName::from_str(&fname) {
+            Ok(file_name) => {
+                let file_size = direntry.metadata()?.len();
+                Discovered::Layer(file_name, file_size)
+            }
+            Err(_) => {
+                if fname == METADATA_FILE_NAME {
+                    Discovered::Metadata
+                } else if fname.ends_with(".old") {
+                    // ignore these
+                    Discovered::IgnoredBackup
+                } else if remote_timeline_client::is_temp_download_file(&direntry_path) {
+                    Discovered::TemporaryDownload(file_name)
+                } else if is_ephemeral_file(&fname) {
+                    Discovered::Ephemeral(file_name)
+                } else if is_temporary(&direntry_path) {
+                    Discovered::Temporary(file_name)
+                } else {
+                    Discovered::Unknown(file_name)
+                }
+            }
+        };
+
+        ret.push(discovered);
+    }
+
+    Ok(ret)
+}
+
+/// Decision on what to do with a layer file after considering its local and remote metadata.
+#[derive(Clone)]
+pub(super) enum Decision {
+    /// The layer is not present locally.
+    Evicted(LayerFileMetadata),
+    /// The layer is present locally, but local metadata does not match remote; we must
+    /// delete it and treat it as evicted.
+    UseRemote {
+        local: LayerFileMetadata,
+        remote: LayerFileMetadata,
+    },
+    /// The layer is present locally, and metadata matches.
+    UseLocal(LayerFileMetadata),
+    /// The layer is only known locally, it needs to be uploaded.
+    NeedsUpload(LayerFileMetadata),
+}
+
+/// The related layer is is in future compared to disk_consistent_lsn, it must not be loaded.
+#[derive(Debug)]
+pub(super) struct FutureLayer {
+    /// The local metadata. `None` if the layer is only known through [`IndexPart`].
+    pub(super) local: Option<LayerFileMetadata>,
+}
+
+/// Merges local discoveries and remote [`IndexPart`] to a collection of decisions.
+///
+/// This function should not gain additional reasons to fail than [`FutureLayer`], consider adding
+/// the checks earlier to [`scan_timeline_dir`].
+pub(super) fn reconcile(
+    discovered: Vec<(LayerFileName, u64)>,
+    index_part: Option<&IndexPart>,
+    disk_consistent_lsn: Lsn,
+) -> Vec<(LayerFileName, Result<Decision, FutureLayer>)> {
+    use Decision::*;
+
+    // name => (local, remote)
+    type Collected = HashMap<LayerFileName, (Option<LayerFileMetadata>, Option<LayerFileMetadata>)>;
+
+    let mut discovered = discovered
+        .into_iter()
+        .map(|(name, file_size)| (name, (Some(LayerFileMetadata::new(file_size)), None)))
+        .collect::<Collected>();
+
+    // merge any index_part information, when available
+    index_part
+        .as_ref()
+        .map(|ip| ip.layer_metadata.iter())
+        .into_iter()
+        .flatten()
+        .map(|(name, metadata)| (name, LayerFileMetadata::from(metadata)))
+        .for_each(|(name, metadata)| {
+            if let Some(existing) = discovered.get_mut(name) {
+                existing.1 = Some(metadata);
+            } else {
+                discovered.insert(name.to_owned(), (None, Some(metadata)));
+            }
+        });
+
+    discovered
+        .into_iter()
+        .map(|(name, (local, remote))| {
+            let decision = if name.is_in_future(disk_consistent_lsn) {
+                Err(FutureLayer { local })
+            } else {
+                Ok(match (local, remote) {
+                    (Some(local), Some(remote)) if local != remote => UseRemote { local, remote },
+                    (Some(x), Some(_)) => UseLocal(x),
+                    (None, Some(x)) => Evicted(x),
+                    (Some(x), None) => NeedsUpload(x),
+                    (None, None) => {
+                        unreachable!("there must not be any non-local non-remote files")
+                    }
+                })
+            };
+
+            (name, decision)
+        })
+        .collect::<Vec<_>>()
+}
+
+pub(super) fn cleanup(path: &Path, kind: &str) -> anyhow::Result<()> {
+    let file_name = path.file_name().expect("must be file path");
+    tracing::debug!(kind, ?file_name, "cleaning up");
+    std::fs::remove_file(path)
+        .with_context(|| format!("failed to remove {kind} at {}", path.display()))
+}
+
+pub(super) fn cleanup_local_file_for_remote(
+    path: &Path,
+    local: &LayerFileMetadata,
+    remote: &LayerFileMetadata,
+) -> anyhow::Result<()> {
+    let local_size = local.file_size();
+    let remote_size = remote.file_size();
+
+    let file_name = path.file_name().expect("must be file path");
+    tracing::warn!("removing local file {file_name:?} because it has unexpected length {local_size}; length in remote index is {remote_size}");
+    if let Err(err) = crate::tenant::timeline::rename_to_backup(path) {
+        assert!(
+            path.exists(),
+            "we would leave the local_layer without a file if this does not hold: {}",
+            path.display()
+        );
+        Err(err)
+    } else {
+        Ok(())
+    }
+}
+
+pub(super) fn cleanup_future_layer(
+    path: &Path,
+    name: &LayerFileName,
+    disk_consistent_lsn: Lsn,
+) -> anyhow::Result<()> {
+    use LayerFileName::*;
+    let kind = match name {
+        Delta(_) => "delta",
+        Image(_) => "image",
+    };
+    // future image layers are allowed to be produced always for not yet flushed to disk
+    // lsns stored in InMemoryLayer.
+    tracing::info!("found future {kind} layer {name} disk_consistent_lsn is {disk_consistent_lsn}");
+    crate::tenant::timeline::rename_to_backup(path)?;
+    Ok(())
+}

pageserver/src/tenant/timeline/layer_manager.rs

@@ -12,38 +12,38 @@ use crate::{
     tenant::{
         layer_map::{BatchedUpdates, LayerMap},
         storage_layer::{
-            AsLayerDesc, DeltaLayer, ImageLayer, InMemoryLayer, Layer, PersistentLayer,
-            PersistentLayerDesc, PersistentLayerKey, RemoteLayer,
+            AsLayerDesc, DeltaLayer, ImageLayer, InMemoryLayer, PersistentLayer,
+            PersistentLayerDesc, PersistentLayerKey,
         },
         timeline::compare_arced_layers,
     },
 };
 
 /// Provides semantic APIs to manipulate the layer map.
-pub struct LayerManager {
+pub(crate) struct LayerManager {
     layer_map: LayerMap,
     layer_fmgr: LayerFileManager,
 }
 
 /// After GC, the layer map changes will not be applied immediately. Users should manually apply the changes after
 /// scheduling deletes in remote client.
-pub struct ApplyGcResultGuard<'a>(BatchedUpdates<'a>);
+pub(crate) struct ApplyGcResultGuard<'a>(BatchedUpdates<'a>);
 
 impl ApplyGcResultGuard<'_> {
-    pub fn flush(self) {
+    pub(crate) fn flush(self) {
         self.0.flush();
     }
 }
 
 impl LayerManager {
-    pub fn create() -> Self {
+    pub(crate) fn create() -> Self {
         Self {
             layer_map: LayerMap::default(),
             layer_fmgr: LayerFileManager::new(),
         }
     }
 
-    pub fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<dyn PersistentLayer> {
+    pub(crate) fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<dyn PersistentLayer> {
         self.layer_fmgr.get_from_desc(desc)
     }
 
@@ -51,18 +51,12 @@ impl LayerManager {
     ///
     /// We expect users only to be able to get an immutable layer map. If users want to make modifications,
     /// they should use the below semantic APIs. This design makes us step closer to immutable storage state.
-    pub fn layer_map(&self) -> &LayerMap {
+    pub(crate) fn layer_map(&self) -> &LayerMap {
         &self.layer_map
     }
 
-    /// Get a mutable reference to the layer map. This function will be removed once `flush_frozen_layer`
-    /// gets a refactor.
-    pub fn layer_map_mut(&mut self) -> &mut LayerMap {
-        &mut self.layer_map
-    }
-
     /// Replace layers in the layer file manager, used in evictions and layer downloads.
-    pub fn replace_and_verify(
+    pub(crate) fn replace_and_verify(
         &mut self,
         expected: Arc<dyn PersistentLayer>,
         new: Arc<dyn PersistentLayer>,
@@ -73,7 +67,7 @@ impl LayerManager {
     /// Called from `load_layer_map`. Initialize the layer manager with:
     /// 1. all on-disk layers
     /// 2. next open layer (with disk disk_consistent_lsn LSN)
-    pub fn initialize_local_layers(
+    pub(crate) fn initialize_local_layers(
         &mut self,
         on_disk_layers: Vec<Arc<dyn PersistentLayer>>,
         next_open_layer_at: Lsn,
@@ -87,28 +81,13 @@ impl LayerManager {
     }
 
     /// Initialize when creating a new timeline, called in `init_empty_layer_map`.
-    pub fn initialize_empty(&mut self, next_open_layer_at: Lsn) {
+    pub(crate) fn initialize_empty(&mut self, next_open_layer_at: Lsn) {
         self.layer_map.next_open_layer_at = Some(next_open_layer_at);
     }
 
-    pub fn initialize_remote_layers(
-        &mut self,
-        corrupted_local_layers: Vec<Arc<dyn PersistentLayer>>,
-        remote_layers: Vec<Arc<RemoteLayer>>,
-    ) {
-        let mut updates = self.layer_map.batch_update();
-        for layer in corrupted_local_layers {
-            Self::remove_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
-        }
-        for layer in remote_layers {
-            Self::insert_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
-        }
-        updates.flush();
-    }
-
     /// Open a new writable layer to append data if there is no open layer, otherwise return the current open layer,
     /// called within `get_layer_for_write`.
-    pub fn get_layer_for_write(
+    pub(crate) fn get_layer_for_write(
         &mut self,
         lsn: Lsn,
         last_record_lsn: Lsn,
@@ -163,7 +142,7 @@ impl LayerManager {
     }
 
     /// Called from `freeze_inmem_layer`, returns true if successfully frozen.
-    pub async fn try_freeze_in_memory_layer(
+    pub(crate) async fn try_freeze_in_memory_layer(
         &mut self,
         Lsn(last_record_lsn): Lsn,
         last_freeze_at: &AtomicLsn,
@@ -185,7 +164,7 @@ impl LayerManager {
     }
 
     /// Add image layers to the layer map, called from `create_image_layers`.
-    pub fn track_new_image_layers(&mut self, image_layers: Vec<ImageLayer>) {
+    pub(crate) fn track_new_image_layers(&mut self, image_layers: Vec<ImageLayer>) {
         let mut updates = self.layer_map.batch_update();
         for layer in image_layers {
             Self::insert_historic_layer(Arc::new(layer), &mut updates, &mut self.layer_fmgr);
@@ -194,7 +173,7 @@ impl LayerManager {
     }
 
     /// Flush a frozen layer and add the written delta layer to the layer map.
-    pub fn finish_flush_l0_layer(
+    pub(crate) fn finish_flush_l0_layer(
         &mut self,
         delta_layer: Option<DeltaLayer>,
         frozen_layer_for_check: &Arc<InMemoryLayer>,
@@ -214,7 +193,7 @@ impl LayerManager {
     }
 
     /// Called when compaction is completed.
-    pub fn finish_compact_l0(
+    pub(crate) fn finish_compact_l0(
         &mut self,
         layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
         compact_from: Vec<Arc<dyn PersistentLayer>>,
@@ -242,7 +221,7 @@ impl LayerManager {
     }
 
     /// Called when garbage collect the timeline. Returns a guard that will apply the updates to the layer map.
-    pub fn finish_gc_timeline(
+    pub(crate) fn finish_gc_timeline(
         &mut self,
         layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
         gc_layers: Vec<Arc<dyn PersistentLayer>>,
@@ -271,16 +250,6 @@ impl LayerManager {
         mapping.insert(layer);
     }
 
-    /// Helper function to remove a layer into the layer map and file manager
-    fn remove_historic_layer(
-        layer: Arc<dyn PersistentLayer>,
-        updates: &mut BatchedUpdates<'_>,
-        mapping: &mut LayerFileManager,
-    ) {
-        updates.remove_historic(layer.layer_desc());
-        mapping.remove(layer);
-    }
-
     /// Removes the layer from local FS (if present) and from memory.
     /// Remote storage is not affected by this operation.
     fn delete_historic_layer(
@@ -313,7 +282,7 @@ impl LayerManager {
     }
 }
 
-pub struct LayerFileManager<T: AsLayerDesc + ?Sized = dyn PersistentLayer>(
+pub(crate) struct LayerFileManager<T: AsLayerDesc + ?Sized = dyn PersistentLayer>(
     HashMap<PersistentLayerKey, Arc<T>>,
 );
 

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

@@ -17,7 +17,7 @@ use crate::metrics::{
     WALRECEIVER_ACTIVE_MANAGERS, WALRECEIVER_BROKER_UPDATES, WALRECEIVER_CANDIDATES_ADDED,
     WALRECEIVER_CANDIDATES_REMOVED, WALRECEIVER_SWITCHES,
 };
-use crate::task_mgr::TaskKind;
+use crate::task_mgr::{shutdown_token, TaskKind};
 use crate::tenant::{debug_assert_current_span_has_tenant_and_timeline_id, Timeline};
 use anyhow::Context;
 use chrono::{NaiveDateTime, Utc};
@@ -211,11 +211,14 @@ async fn subscribe_for_timeline_updates(
     id: TenantTimelineId,
 ) -> Streaming<SafekeeperTimelineInfo> {
     let mut attempt = 0;
+    let cancel = shutdown_token();
+
     loop {
         exponential_backoff(
             attempt,
             DEFAULT_BASE_BACKOFF_SECONDS,
             DEFAULT_MAX_BACKOFF_SECONDS,
+            &cancel,
         )
         .await;
         attempt += 1;

pageserver/src/tenant/upload_queue.rs

@@ -140,36 +140,24 @@ impl UploadQueue {
             }
         }
 
-        let mut files = HashMap::with_capacity(index_part.timeline_layers.len());
-        for layer_name in &index_part.timeline_layers {
-            match index_part
-                .layer_metadata
-                .get(layer_name)
-                .map(LayerFileMetadata::from)
-            {
-                Some(layer_metadata) => {
-                    files.insert(layer_name.to_owned(), layer_metadata);
-                }
-                None => {
-                    anyhow::bail!(
-                        "No remote layer metadata found for layer {}",
-                        layer_name.file_name()
-                    );
-                }
-            }
+        let mut files = HashMap::with_capacity(index_part.layer_metadata.len());
+        for (layer_name, layer_metadata) in &index_part.layer_metadata {
+            files.insert(
+                layer_name.to_owned(),
+                LayerFileMetadata::from(layer_metadata),
+            );
         }
 
-        let index_part_metadata = index_part.parse_metadata()?;
         info!(
             "initializing upload queue with remote index_part.disk_consistent_lsn: {}",
-            index_part_metadata.disk_consistent_lsn()
+            index_part.metadata.disk_consistent_lsn()
         );
 
         let state = UploadQueueInitialized {
             latest_files: files,
             latest_files_changes_since_metadata_upload_scheduled: 0,
-            latest_metadata: index_part_metadata.clone(),
-            last_uploaded_consistent_lsn: index_part_metadata.disk_consistent_lsn(),
+            latest_metadata: index_part.metadata.clone(),
+            last_uploaded_consistent_lsn: index_part.metadata.disk_consistent_lsn(),
             // what follows are boring default initializations
             task_counter: 0,
             num_inprogress_layer_uploads: 0,

pageserver/src/walingest.rs

@@ -312,7 +312,7 @@ impl<'a> WalIngest<'a> {
                 // particular point in the WAL. For more fine-grained control,
                 // we could peek into the message and only pause if it contains
                 // a particular string, for example, but this is enough for now.
-                utils::failpoint_sleep_millis_async!("wal-ingest-logical-message-sleep");
+                crate::failpoint_support::sleep_millis_async!("wal-ingest-logical-message-sleep");
             }
         }
 

proxy/src/http/conn_pool.rs

@@ -64,13 +64,13 @@ pub struct EndpointConnPool {
     total_conns: usize,
 }
 
-/// This is cheap and not hugely secure.
-/// But probably good enough for in memory only hashes.
+/// 4096 is the number of rounds that SCRAM-SHA-256 recommends.
+/// It's not the 600,000 that OWASP recommends... but our passwords are high entropy anyway.
 ///
-/// Still takes 3.5ms to hash on my hardware.
+/// Still takes 1.4ms to hash on my hardware.
 /// We don't want to ruin the latency improvements of using the pool by making password verification take too long
 const PARAMS: Params = Params {
-    rounds: 10_000,
+    rounds: 4096,
     output_length: 32,
 };
 
@@ -99,6 +99,10 @@ pub struct GlobalConnPool {
     max_conns_per_endpoint: usize,
 
     proxy_config: &'static crate::config::ProxyConfig,
+
+    // Using a lock to remove any race conditions.
+    // Eg cleaning up connections while a new connection is returned
+    closed: RwLock<bool>,
 }
 
 impl GlobalConnPool {
@@ -108,9 +112,24 @@ impl GlobalConnPool {
             global_pool_size: AtomicUsize::new(0),
             max_conns_per_endpoint: MAX_CONNS_PER_ENDPOINT,
             proxy_config: config,
+            closed: RwLock::new(false),
         })
     }
 
+    pub fn shutdown(&self) {
+        *self.closed.write() = true;
+
+        self.global_pool.retain(|_, endpoint_pool| {
+            let mut pool = endpoint_pool.write();
+            // by clearing this hashmap, we remove the slots that a connection can be returned to.
+            // when returning, it drops the connection if the slot doesn't exist
+            pool.pools.clear();
+            pool.total_conns = 0;
+
+            false
+        });
+    }
+
     pub async fn get(
         &self,
         conn_info: &ConnInfo,
@@ -208,7 +227,20 @@ impl GlobalConnPool {
         new_client
     }
 
-    pub async fn put(&self, conn_info: &ConnInfo, client: Client) -> anyhow::Result<()> {
+    pub fn put(&self, conn_info: &ConnInfo, client: Client) -> anyhow::Result<()> {
+        // We want to hold this open while we return. This ensures that the pool can't close
+        // while we are in the middle of returning the connection.
+        let closed = self.closed.read();
+        if *closed {
+            info!("pool: throwing away connection '{conn_info}' because pool is closed");
+            return Ok(());
+        }
+
+        if client.inner.is_closed() {
+            info!("pool: throwing away connection '{conn_info}' because connection is closed");
+            return Ok(());
+        }
+
         let pool = self.get_or_create_endpoint_pool(&conn_info.hostname);
 
         // return connection to the pool
@@ -376,9 +408,9 @@ async fn connect_to_compute_once(
     let (tx, mut rx) = tokio::sync::watch::channel(session);
 
     let conn_id = uuid::Uuid::new_v4();
-    let span = info_span!(parent: None, "connection", %conn_info, %conn_id);
+    let span = info_span!(parent: None, "connection", %conn_id);
     span.in_scope(|| {
-        info!(%session, "new connection");
+        info!(%conn_info, %session, "new connection");
     });
 
     tokio::spawn(
@@ -388,26 +420,28 @@ async fn connect_to_compute_once(
                 info!(%session, "changed session");
             }
 
-            let message = ready!(connection.poll_message(cx));
+            loop {
+                let message = ready!(connection.poll_message(cx));
 
-            match message {
-                Some(Ok(AsyncMessage::Notice(notice))) => {
-                    info!(%session, "notice: {}", notice);
-                    Poll::Pending
+                match message {
+                    Some(Ok(AsyncMessage::Notice(notice))) => {
+                        info!(%session, "notice: {}", notice);
+                    }
+                    Some(Ok(AsyncMessage::Notification(notif))) => {
+                        warn!(%session, pid = notif.process_id(), channel = notif.channel(), "notification received");
+                    }
+                    Some(Ok(_)) => {
+                        warn!(%session, "unknown message");
+                    }
+                    Some(Err(e)) => {
+                        error!(%session, "connection error: {}", e);
+                        return Poll::Ready(())
+                    }
+                    None => {
+                        info!("connection closed");
+                        return Poll::Ready(())
+                    }
                 }
-                Some(Ok(AsyncMessage::Notification(notif))) => {
-                    warn!(%session, pid = notif.process_id(), channel = notif.channel(), "notification received");
-                    Poll::Pending
-                }
-                Some(Ok(_)) => {
-                    warn!(%session, "unknown message");
-                    Poll::Pending
-                }
-                Some(Err(e)) => {
-                    error!(%session, "connection error: {}", e);
-                    Poll::Ready(())
-                }
-                None => Poll::Ready(()),
             }
         })
         .instrument(span)

proxy/src/http/sql_over_http.rs

@@ -16,7 +16,6 @@ use tokio_postgres::types::Type;
 use tokio_postgres::GenericClient;
 use tokio_postgres::IsolationLevel;
 use tokio_postgres::Row;
-use tracing::Instrument;
 use url::Url;
 
 use super::conn_pool::ConnInfo;
@@ -286,13 +285,12 @@ pub async fn handle(
     };
 
     if allow_pool {
+        let current_span = tracing::Span::current();
         // return connection to the pool
-        tokio::task::spawn(
-            async move {
-                let _ = conn_pool.put(&conn_info, client).await;
-            }
-            .in_current_span(),
-        );
+        tokio::task::spawn_blocking(move || {
+            let _span = current_span.enter();
+            let _ = conn_pool.put(&conn_info, client);
+        });
     }
 
     result

proxy/src/http/websocket.rs

@@ -269,6 +269,18 @@ pub async fn task_main(
 
     let conn_pool: Arc<GlobalConnPool> = GlobalConnPool::new(config);
 
+    // shutdown the connection pool
+    tokio::spawn({
+        let cancellation_token = cancellation_token.clone();
+        let conn_pool = conn_pool.clone();
+        async move {
+            cancellation_token.cancelled().await;
+            tokio::task::spawn_blocking(move || conn_pool.shutdown())
+                .await
+                .unwrap();
+        }
+    });
+
     let tls_config = config.tls_config.as_ref().map(|cfg| cfg.to_server_config());
     let tls_acceptor: tokio_rustls::TlsAcceptor = match tls_config {
         Some(config) => config.into(),

safekeeper/src/http/routes.rs

@@ -15,8 +15,10 @@ use tokio::fs::File;
 use tokio::io::AsyncReadExt;
 use utils::http::endpoint::request_span;
 
+use crate::receive_wal::WalReceiverState;
 use crate::safekeeper::ServerInfo;
 use crate::safekeeper::Term;
+use crate::send_wal::WalSenderState;
 use crate::{debug_dump, pull_timeline};
 
 use crate::timelines_global_map::TimelineDeleteForceResult;
@@ -99,6 +101,8 @@ pub struct TimelineStatus {
     pub peer_horizon_lsn: Lsn,
     #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn: Lsn,
+    pub walsenders: Vec<WalSenderState>,
+    pub walreceivers: Vec<WalReceiverState>,
 }
 
 fn check_permission(request: &Request<Body>, tenant_id: Option<TenantId>) -> Result<(), ApiError> {
@@ -149,6 +153,8 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
         backup_lsn: inmem.backup_lsn,
         peer_horizon_lsn: inmem.peer_horizon_lsn,
         remote_consistent_lsn: tli.get_walsenders().get_remote_consistent_lsn(),
+        walsenders: tli.get_walsenders().get_all(),
+        walreceivers: tli.get_walreceivers().get_all(),
     };
     json_response(StatusCode::OK, status)
 }

safekeeper/src/receive_wal.rs

@@ -11,11 +11,16 @@ use crate::wal_service::ConnectionId;
 use crate::GlobalTimelines;
 use anyhow::{anyhow, Context};
 use bytes::BytesMut;
+use parking_lot::MappedMutexGuard;
+use parking_lot::Mutex;
+use parking_lot::MutexGuard;
 use postgres_backend::CopyStreamHandlerEnd;
 use postgres_backend::PostgresBackend;
 use postgres_backend::PostgresBackendReader;
 use postgres_backend::QueryError;
 use pq_proto::BeMessage;
+use serde::Deserialize;
+use serde::Serialize;
 use std::net::SocketAddr;
 use std::sync::Arc;
 use tokio::io::AsyncRead;
@@ -32,6 +37,105 @@ use tracing::*;
 use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
 
+/// Registry of WalReceivers (compute connections). Timeline holds it (wrapped
+/// in Arc).
+pub struct WalReceivers {
+    mutex: Mutex<WalReceiversShared>,
+}
+
+/// Id under which walreceiver is registered in shmem.
+type WalReceiverId = usize;
+
+impl WalReceivers {
+    pub fn new() -> Arc<WalReceivers> {
+        Arc::new(WalReceivers {
+            mutex: Mutex::new(WalReceiversShared { slots: Vec::new() }),
+        })
+    }
+
+    /// Register new walreceiver. Returned guard provides access to the slot and
+    /// automatically deregisters in Drop.
+    pub fn register(self: &Arc<WalReceivers>) -> WalReceiverGuard {
+        let slots = &mut self.mutex.lock().slots;
+        let walreceiver = WalReceiverState::Voting;
+        // find empty slot or create new one
+        let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
+            slots[pos] = Some(walreceiver);
+            pos
+        } else {
+            let pos = slots.len();
+            slots.push(Some(walreceiver));
+            pos
+        };
+        WalReceiverGuard {
+            id: pos,
+            walreceivers: self.clone(),
+        }
+    }
+
+    /// Get reference to locked slot contents. Slot must exist (registered
+    /// earlier).
+    fn get_slot<'a>(
+        self: &'a Arc<WalReceivers>,
+        id: WalReceiverId,
+    ) -> MappedMutexGuard<'a, WalReceiverState> {
+        MutexGuard::map(self.mutex.lock(), |locked| {
+            locked.slots[id]
+                .as_mut()
+                .expect("walreceiver doesn't exist")
+        })
+    }
+
+    /// Get number of walreceivers (compute connections).
+    pub fn get_num(self: &Arc<WalReceivers>) -> usize {
+        self.mutex.lock().slots.iter().flatten().count()
+    }
+
+    /// Get state of all walreceivers.
+    pub fn get_all(self: &Arc<WalReceivers>) -> Vec<WalReceiverState> {
+        self.mutex.lock().slots.iter().flatten().cloned().collect()
+    }
+
+    /// Unregister walsender.
+    fn unregister(self: &Arc<WalReceivers>, id: WalReceiverId) {
+        let mut shared = self.mutex.lock();
+        shared.slots[id] = None;
+    }
+}
+
+/// Only a few connections are expected (normally one), so store in Vec.
+struct WalReceiversShared {
+    slots: Vec<Option<WalReceiverState>>,
+}
+
+/// Walreceiver status. Currently only whether it passed voting stage and
+/// started receiving the stream, but it is easy to add more if needed.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub enum WalReceiverState {
+    Voting,
+    Streaming,
+}
+
+/// Scope guard to access slot in WalSenders registry and unregister from it in
+/// Drop.
+pub struct WalReceiverGuard {
+    id: WalReceiverId,
+    walreceivers: Arc<WalReceivers>,
+}
+
+impl WalReceiverGuard {
+    /// Get reference to locked shared state contents.
+    fn get(&self) -> MappedMutexGuard<WalReceiverState> {
+        self.walreceivers.get_slot(self.id)
+    }
+}
+
+impl Drop for WalReceiverGuard {
+    fn drop(&mut self) {
+        self.walreceivers.unregister(self.id);
+    }
+}
+
 const MSG_QUEUE_SIZE: usize = 256;
 const REPLY_QUEUE_SIZE: usize = 16;
 
@@ -246,10 +350,13 @@ impl WalAcceptor {
     /// it must mean that network thread terminated.
     async fn run(&mut self) -> anyhow::Result<()> {
         // Register the connection and defer unregister.
-        self.tli.on_compute_connect().await?;
-        let _guard = ComputeConnectionGuard {
+        // Order of the next two lines is important: we want first to remove our entry and then
+        // update status which depends on registered connections.
+        let _compute_conn_guard = ComputeConnectionGuard {
             timeline: Arc::clone(&self.tli),
         };
+        let walreceiver_guard = self.tli.get_walreceivers().register();
+        self.tli.update_status_notify().await?;
 
         // After this timestamp we will stop processing AppendRequests and send a response
         // to the walproposer. walproposer sends at least one AppendRequest per second,
@@ -263,6 +370,11 @@ impl WalAcceptor {
             }
             let mut next_msg = opt_msg.unwrap();
 
+            // Update walreceiver state in shmem for reporting.
+            if let ProposerAcceptorMessage::Elected(_) = &next_msg {
+                *walreceiver_guard.get() = WalReceiverState::Streaming;
+            }
+
             let reply_msg = if matches!(next_msg, ProposerAcceptorMessage::AppendRequest(_)) {
                 // loop through AppendRequest's while it's readily available to
                 // write as many WAL as possible without fsyncing
@@ -311,6 +423,7 @@ impl WalAcceptor {
     }
 }
 
+/// Calls update_status_notify in drop to update timeline status.
 struct ComputeConnectionGuard {
     timeline: Arc<Timeline>,
 }
@@ -318,11 +431,9 @@ struct ComputeConnectionGuard {
 impl Drop for ComputeConnectionGuard {
     fn drop(&mut self) {
         let tli = self.timeline.clone();
-        // tokio forbids to call blocking_send inside the runtime, and see
-        // comments in on_compute_disconnect why we call blocking_send.
         tokio::spawn(async move {
-            if let Err(e) = tli.on_compute_disconnect().await {
-                error!("failed to unregister compute connection: {}", e);
+            if let Err(e) = tli.update_status_notify().await {
+                error!("failed to update timeline status: {}", e);
             }
         });
     }

safekeeper/src/timeline.rs

@@ -23,6 +23,7 @@ use utils::{
 use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
 
+use crate::receive_wal::WalReceivers;
 use crate::safekeeper::{
     AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, SafeKeeperState,
     SafekeeperMemState, ServerInfo, Term,
@@ -164,8 +165,8 @@ impl SharedState {
         })
     }
 
-    fn is_active(&self, remote_consistent_lsn: Lsn) -> bool {
-        self.is_wal_backup_required()
+    fn is_active(&self, num_computes: usize, remote_consistent_lsn: Lsn) -> bool {
+        self.is_wal_backup_required(num_computes)
             // FIXME: add tracking of relevant pageservers and check them here individually,
             // otherwise migration won't work (we suspend too early).
             || remote_consistent_lsn < self.sk.inmem.commit_lsn
@@ -173,29 +174,34 @@ impl SharedState {
 
     /// Mark timeline active/inactive and return whether s3 offloading requires
     /// start/stop action.
-    fn update_status(&mut self, remote_consistent_lsn: Lsn, ttid: TenantTimelineId) -> bool {
-        let is_active = self.is_active(remote_consistent_lsn);
+    fn update_status(
+        &mut self,
+        num_computes: usize,
+        remote_consistent_lsn: Lsn,
+        ttid: TenantTimelineId,
+    ) -> bool {
+        let is_active = self.is_active(num_computes, remote_consistent_lsn);
         if self.active != is_active {
             info!("timeline {} active={} now", ttid, is_active);
         }
         self.active = is_active;
-        self.is_wal_backup_action_pending()
+        self.is_wal_backup_action_pending(num_computes)
     }
 
     /// Should we run s3 offloading in current state?
-    fn is_wal_backup_required(&self) -> bool {
+    fn is_wal_backup_required(&self, num_computes: usize) -> bool {
         let seg_size = self.get_wal_seg_size();
-        self.num_computes > 0 ||
+        num_computes > 0 ||
         // Currently only the whole segment is offloaded, so compare segment numbers.
-               (self.sk.inmem.commit_lsn.segment_number(seg_size) >
-                self.sk.inmem.backup_lsn.segment_number(seg_size))
+            (self.sk.inmem.commit_lsn.segment_number(seg_size) >
+             self.sk.inmem.backup_lsn.segment_number(seg_size))
     }
 
     /// Is current state of s3 offloading is not what it ought to be?
-    fn is_wal_backup_action_pending(&self) -> bool {
-        let res = self.wal_backup_active != self.is_wal_backup_required();
+    fn is_wal_backup_action_pending(&self, num_computes: usize) -> bool {
+        let res = self.wal_backup_active != self.is_wal_backup_required(num_computes);
         if res {
-            let action_pending = if self.is_wal_backup_required() {
+            let action_pending = if self.is_wal_backup_required(num_computes) {
                 "start"
             } else {
                 "stop"
@@ -210,8 +216,8 @@ impl SharedState {
 
     /// Returns whether s3 offloading is required and sets current status as
     /// matching.
-    fn wal_backup_attend(&mut self) -> bool {
-        self.wal_backup_active = self.is_wal_backup_required();
+    fn wal_backup_attend(&mut self, num_computes: usize) -> bool {
+        self.wal_backup_active = self.is_wal_backup_required(num_computes);
         self.wal_backup_active
     }
 
@@ -295,6 +301,7 @@ pub struct Timeline {
     /// while holding it, ensuring that consensus checks are in order.
     mutex: Mutex<SharedState>,
     walsenders: Arc<WalSenders>,
+    walreceivers: Arc<WalReceivers>,
 
     /// Cancellation channel. Delete/cancel will send `true` here as a cancellation signal.
     cancellation_tx: watch::Sender<bool>,
@@ -329,6 +336,7 @@ impl Timeline {
             commit_lsn_watch_rx,
             mutex: Mutex::new(shared_state),
             walsenders: WalSenders::new(rcl),
+            walreceivers: WalReceivers::new(),
             cancellation_rx,
             cancellation_tx,
             timeline_dir: conf.timeline_dir(&ttid),
@@ -355,6 +363,7 @@ impl Timeline {
             commit_lsn_watch_rx,
             mutex: Mutex::new(SharedState::create_new(&conf, &ttid, state)?),
             walsenders: WalSenders::new(Lsn(0)),
+            walreceivers: WalReceivers::new(),
             cancellation_rx,
             cancellation_tx,
             timeline_dir: conf.timeline_dir(&ttid),
@@ -441,40 +450,22 @@ impl Timeline {
     }
 
     fn update_status(&self, shared_state: &mut SharedState) -> bool {
-        shared_state.update_status(self.get_walsenders().get_remote_consistent_lsn(), self.ttid)
+        shared_state.update_status(
+            self.walreceivers.get_num(),
+            self.get_walsenders().get_remote_consistent_lsn(),
+            self.ttid,
+        )
     }
 
-    /// Register compute connection, starting timeline-related activity if it is
-    /// not running yet.
-    pub async fn on_compute_connect(&self) -> Result<()> {
+    /// Update timeline status and kick wal backup launcher to stop/start offloading if needed.
+    pub async fn update_status_notify(&self) -> Result<()> {
         if self.is_cancelled() {
             bail!(TimelineError::Cancelled(self.ttid));
         }
-
-        let is_wal_backup_action_pending: bool;
-        {
+        let is_wal_backup_action_pending: bool = {
             let mut shared_state = self.write_shared_state().await;
-            shared_state.num_computes += 1;
-            is_wal_backup_action_pending = self.update_status(&mut shared_state);
-        }
-        // Wake up wal backup launcher, if offloading not started yet.
-        if is_wal_backup_action_pending {
-            // Can fail only if channel to a static thread got closed, which is not normal at all.
-            self.wal_backup_launcher_tx.send(self.ttid).await?;
-        }
-        Ok(())
-    }
-
-    /// De-register compute connection, shutting down timeline activity if
-    /// pageserver doesn't need catchup.
-    pub async fn on_compute_disconnect(&self) -> Result<()> {
-        let is_wal_backup_action_pending: bool;
-        {
-            let mut shared_state = self.write_shared_state().await;
-            shared_state.num_computes -= 1;
-            is_wal_backup_action_pending = self.update_status(&mut shared_state);
-        }
-        // Wake up wal backup launcher, if it is time to stop the offloading.
+            self.update_status(&mut shared_state)
+        };
         if is_wal_backup_action_pending {
             // Can fail only if channel to a static thread got closed, which is not normal at all.
             self.wal_backup_launcher_tx.send(self.ttid).await?;
@@ -519,7 +510,9 @@ impl Timeline {
             return false;
         }
 
-        self.write_shared_state().await.wal_backup_attend()
+        self.write_shared_state()
+            .await
+            .wal_backup_attend(self.walreceivers.get_num())
     }
 
     /// Returns commit_lsn watch channel.
@@ -650,6 +643,10 @@ impl Timeline {
         &self.walsenders
     }
 
+    pub fn get_walreceivers(&self) -> &Arc<WalReceivers> {
+        &self.walreceivers
+    }
+
     /// Returns flush_lsn.
     pub async fn get_flush_lsn(&self) -> Lsn {
         self.write_shared_state().await.sk.wal_store.flush_lsn()

scripts/combine_control_files.py

@@ -1,76 +0,0 @@
-#! /usr/bin/env python3
-# Script to generate ext_index.json metadata file
-# that stores content of the control files and location of extension archives
-# for all extensions in extensions subdir.
-import argparse
-import json
-import subprocess
-from pathlib import Path
-
-"""
-# ext_index.json example:
-{
-    "public_extensions": [
-        "anon"
-    ],
-    "library_index": {
-        "anon": "anon",
-        // for more complex extensions like postgis
-        // we might have something like:
-        // address_standardizer: postgis
-        // postgis_tiger: postgis
-    },
-    "extension_data": {
-        "anon": {
-            "control_data": {
-                "anon.control": "# PostgreSQL Anonymizer (anon) extension \ncomment = 'Data anonymization tools' \ndefault_version = '1.1.0' \ndirectory='extension/anon' \nrelocatable = false \nrequires = 'pgcrypto' \nsuperuser = false \nmodule_pathname = '$libdir/anon' \ntrusted = true \n"
-            },
-            "archive_path": "5648391853/v15/extensions/anon.tar.zst"
-        }
-    }
-}
-"""
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="generate ext_index.json")
-    parser.add_argument("pg_version", type=str, choices=["v14", "v15"], help="pg_version")
-    parser.add_argument("BUILD_TAG", type=str, help="BUILD_TAG for this compute image")
-    parser.add_argument("--public_extensions", type=str, help="list of public extensions")
-    args = parser.parse_args()
-    pg_version = args.pg_version
-    BUILD_TAG = args.BUILD_TAG
-    public_ext_list = args.public_extensions.split(",")
-
-    ext_index = {}
-    library_index = {}
-    EXT_PATH = Path("extensions")
-    for extension in EXT_PATH.iterdir():
-        if extension.is_dir():
-            control_data = {}
-            for control_file in extension.glob("*.control"):
-                if control_file.suffix != ".control":
-                    continue
-                with open(control_file, "r") as f:
-                    control_data[control_file.name] = f.read()
-            ext_index[extension.name] = {
-                "control_data": control_data,
-                "archive_path": f"{BUILD_TAG}/{pg_version}/extensions/{extension.name}.tar.zst",
-            }
-        elif extension.suffix == ".zst":
-            file_list = (
-                str(subprocess.check_output(["tar", "tf", str(extension)]), "utf-8")
-                .strip()
-                .split("\n")
-            )
-            for file in file_list:
-                if file.endswith(".so") and file.startswith("lib/"):
-                    lib_name = file[4:-3]
-                    library_index[lib_name] = extension.name.replace(".tar.zst", "")
-
-    all_data = {
-        "public_extensions": public_ext_list,
-        "library_index": library_index,
-        "extension_data": ext_index,
-    }
-    with open("ext_index.json", "w") as f:
-        json.dump(all_data, f)

test_runner/fixtures/metrics.py

@@ -70,6 +70,7 @@ PAGESERVER_GLOBAL_METRICS: Tuple[str, ...] = (
     "pageserver_getpage_reconstruct_seconds_count",
     "pageserver_getpage_reconstruct_seconds_sum",
     *[f"pageserver_basebackup_query_seconds_{x}" for x in ["bucket", "count", "sum"]],
+    *histogram("pageserver_smgr_query_seconds_global"),
     *histogram("pageserver_read_num_fs_layers"),
     *histogram("pageserver_getpage_get_reconstruct_data_seconds"),
     *histogram("pageserver_wait_lsn_seconds"),

test_runner/fixtures/neon_fixtures.py

@@ -427,6 +427,7 @@ class NeonEnvBuilder:
         default_branch_name: str = DEFAULT_BRANCH_NAME,
         preserve_database_files: bool = False,
         initial_tenant: Optional[TenantId] = None,
+        initial_timeline: Optional[TimelineId] = None,
     ):
         self.repo_dir = repo_dir
         self.rust_log_override = rust_log_override
@@ -452,6 +453,7 @@ class NeonEnvBuilder:
         self.pg_version = pg_version
         self.preserve_database_files = preserve_database_files
         self.initial_tenant = initial_tenant or TenantId.generate()
+        self.initial_timeline = initial_timeline or TimelineId.generate()
 
     def init_configs(self) -> NeonEnv:
         # Cannot create more than one environment from one builder
@@ -473,9 +475,10 @@ class NeonEnvBuilder:
             f"Services started, creating initial tenant {env.initial_tenant} and its initial timeline"
         )
         initial_tenant, initial_timeline = env.neon_cli.create_tenant(
-            tenant_id=env.initial_tenant, conf=initial_tenant_conf
+            tenant_id=env.initial_tenant, conf=initial_tenant_conf, timeline_id=env.initial_timeline
         )
-        env.initial_timeline = initial_timeline
+        assert env.initial_tenant == initial_tenant
+        assert env.initial_timeline == initial_timeline
         log.info(f"Initial timeline {initial_tenant}/{initial_timeline} created successfully")
 
         return env
@@ -784,7 +787,7 @@ class NeonEnv:
         # generate initial tenant ID here instead of letting 'neon init' generate it,
         # so that we don't need to dig it out of the config file afterwards.
         self.initial_tenant = config.initial_tenant
-        self.initial_timeline: Optional[TimelineId] = None
+        self.initial_timeline = config.initial_timeline
 
         # Create a config file corresponding to the options
         toml = textwrap.dedent(

test_runner/fixtures/pageserver/utils.py

@@ -315,4 +315,4 @@ MANY_SMALL_LAYERS_TENANT_CONFIG = {
 
 
 def poll_for_remote_storage_iterations(remote_storage_kind: RemoteStorageKind) -> int:
-    return 40 if remote_storage_kind is RemoteStorageKind.REAL_S3 else 10
+    return 40 if remote_storage_kind is RemoteStorageKind.REAL_S3 else 15

test_runner/regress/test_disk_usage_eviction.py

@@ -1,4 +1,3 @@
-import shutil
 import time
 from dataclasses import dataclass
 from typing import Dict, Tuple
@@ -14,7 +13,7 @@ from fixtures.neon_fixtures import (
 )
 from fixtures.pageserver.http import PageserverHttpClient
 from fixtures.pageserver.utils import wait_for_upload_queue_empty
-from fixtures.remote_storage import LocalFsStorage, RemoteStorageKind
+from fixtures.remote_storage import RemoteStorageKind
 from fixtures.types import Lsn, TenantId, TimelineId
 from fixtures.utils import wait_until
 
@@ -138,22 +137,14 @@ def eviction_env(request, neon_env_builder: NeonEnvBuilder, pg_bin: PgBin) -> Ev
 
     neon_env_builder.enable_remote_storage(RemoteStorageKind.LOCAL_FS, f"{request.node.name}")
 
-    env = neon_env_builder.init_start()
+    # initial tenant will not be present on this pageserver
+    env = neon_env_builder.init_configs()
+    env.start()
     pageserver_http = env.pageserver.http_client()
 
     # allow because we are invoking this manually; we always warn on executing disk based eviction
     env.pageserver.allowed_errors.append(r".* running disk usage based eviction due to pressure.*")
 
-    # remove the initial tenant
-    assert env.initial_timeline
-    pageserver_http.tenant_detach(env.initial_tenant)
-    assert isinstance(env.remote_storage, LocalFsStorage)
-    tenant_remote_storage = env.remote_storage.root / "tenants" / str(env.initial_tenant)
-    assert tenant_remote_storage.is_dir()
-    shutil.rmtree(tenant_remote_storage)
-    env.initial_tenant = TenantId("0" * 32)
-    env.initial_timeline = None
-
     # Choose small layer_size so that we can use low pgbench_scales and still get a large count of layers.
     # Large count of layers and small layer size is good for testing because it makes evictions predictable.
     # Predictable in the sense that many layer evictions will be required to reach the eviction target, because

test_runner/regress/test_gc_aggressive.py

@@ -11,8 +11,7 @@ from fixtures.neon_fixtures import (
     wait_for_last_flush_lsn,
 )
 from fixtures.remote_storage import RemoteStorageKind
-from fixtures.types import TenantId, TimelineId
-from fixtures.utils import query_scalar
+from fixtures.types import TimelineId
 
 # Test configuration
 #
@@ -71,13 +70,11 @@ def test_gc_aggressive(neon_env_builder: NeonEnvBuilder):
     # Disable pitr, because here we want to test branch creation after GC
     neon_env_builder.pageserver_config_override = "tenant_config={pitr_interval = '0 sec'}"
     env = neon_env_builder.init_start()
-    env.neon_cli.create_branch("test_gc_aggressive", "main")
+    timeline = env.neon_cli.create_branch("test_gc_aggressive", "main")
     endpoint = env.endpoints.create_start("test_gc_aggressive")
     log.info("postgres is running on test_gc_aggressive branch")
 
     with endpoint.cursor() as cur:
-        timeline = TimelineId(query_scalar(cur, "SHOW neon.timeline_id"))
-
         # Create table, and insert the first 100 rows
         cur.execute("CREATE TABLE foo (id int, counter int, t text)")
         cur.execute(
@@ -109,7 +106,8 @@ def test_gc_index_upload(neon_env_builder: NeonEnvBuilder, remote_storage_kind:
     )
 
     env = neon_env_builder.init_start()
-    env.neon_cli.create_branch("test_gc_index_upload", "main")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_gc_index_upload", "main")
     endpoint = env.endpoints.create_start("test_gc_index_upload")
 
     pageserver_http = env.pageserver.http_client()
@@ -117,9 +115,6 @@ def test_gc_index_upload(neon_env_builder: NeonEnvBuilder, remote_storage_kind:
     pg_conn = endpoint.connect()
     cur = pg_conn.cursor()
 
-    tenant_id = TenantId(query_scalar(cur, "SHOW neon.tenant_id"))
-    timeline_id = TimelineId(query_scalar(cur, "SHOW neon.timeline_id"))
-
     cur.execute("CREATE TABLE foo (id int, counter int, t text)")
     cur.execute(
         """

test_runner/regress/test_gc_cutoff.py

@@ -12,13 +12,8 @@ from fixtures.neon_fixtures import NeonEnvBuilder, PgBin
 # test anyway, so it doesn't need any special attention here.
 @pytest.mark.timeout(600)
 def test_gc_cutoff(neon_env_builder: NeonEnvBuilder, pg_bin: PgBin):
-    env = neon_env_builder.init_start()
-
-    pageserver_http = env.pageserver.http_client()
-
-    # Use aggressive GC and checkpoint settings, so that we also exercise GC during the test
-    tenant_id, _ = env.neon_cli.create_tenant(
-        conf={
+    env = neon_env_builder.init_start(
+        initial_tenant_conf={
             "gc_period": "10 s",
             "gc_horizon": f"{1024 ** 2}",
             "checkpoint_distance": f"{1024 ** 2}",
@@ -29,6 +24,11 @@ def test_gc_cutoff(neon_env_builder: NeonEnvBuilder, pg_bin: PgBin):
             "image_creation_threshold": "2",
         }
     )
+
+    pageserver_http = env.pageserver.http_client()
+
+    # Use aggressive GC and checkpoint settings, so that we also exercise GC during the test
+    tenant_id = env.initial_tenant
     endpoint = env.endpoints.create_start("main", tenant_id=tenant_id)
     connstr = endpoint.connstr(options="-csynchronous_commit=off")
     pg_bin.run_capture(["pgbench", "-i", "-s10", connstr])
@@ -39,5 +39,4 @@ def test_gc_cutoff(neon_env_builder: NeonEnvBuilder, pg_bin: PgBin):
         with pytest.raises(subprocess.SubprocessError):
             pg_bin.run_capture(["pgbench", "-P1", "-N", "-c5", "-T500", "-Mprepared", connstr])
         env.pageserver.stop()
-        env.pageserver.start()
-        pageserver_http.configure_failpoints(("after-timeline-gc-removed-layers", "exit"))
+        env.pageserver.start(extra_env_vars={"FAILPOINTS": "after-timeline-gc-removed-layers=exit"})

test_runner/regress/test_large_schema.py

@@ -74,9 +74,9 @@ def test_large_schema(neon_env_builder: NeonEnvBuilder):
     cur.execute("select * from pg_depend order by refclassid, refobjid, refobjsubid")
 
     # Check layer file sizes
-    tenant_id = endpoint.safe_psql("show neon.tenant_id")[0][0]
-    timeline_id = endpoint.safe_psql("show neon.timeline_id")[0][0]
-    timeline_path = "{}/tenants/{}/timelines/{}/".format(env.repo_dir, tenant_id, timeline_id)
+    timeline_path = "{}/tenants/{}/timelines/{}/".format(
+        env.repo_dir, env.initial_tenant, env.initial_timeline
+    )
     for filename in os.listdir(timeline_path):
         if filename.startswith("00000"):
             log.info(f"layer {filename} size is {os.path.getsize(timeline_path + filename)}")

test_runner/regress/test_layer_eviction.py

@@ -8,7 +8,7 @@ from fixtures.neon_fixtures import (
 )
 from fixtures.pageserver.utils import wait_for_last_record_lsn, wait_for_upload
 from fixtures.remote_storage import RemoteStorageKind
-from fixtures.types import Lsn, TenantId, TimelineId
+from fixtures.types import Lsn
 from fixtures.utils import query_scalar
 
 
@@ -34,8 +34,8 @@ def test_basic_eviction(
     client = env.pageserver.http_client()
     endpoint = env.endpoints.create_start("main")
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Create a number of layers in the tenant
     with endpoint.cursor() as cur:

test_runner/regress/test_metric_collection.py

@@ -18,8 +18,7 @@ from fixtures.neon_fixtures import (
 )
 from fixtures.port_distributor import PortDistributor
 from fixtures.remote_storage import RemoteStorageKind
-from fixtures.types import TenantId, TimelineId
-from fixtures.utils import query_scalar
+from fixtures.types import TenantId
 from pytest_httpserver import HTTPServer
 from werkzeug.wrappers.request import Request
 from werkzeug.wrappers.response import Response
@@ -115,15 +114,13 @@ def test_metric_collection(
     # Order of fixtures shutdown is not specified, and if http server gets down
     # before pageserver, pageserver log might contain such errors in the end.
     env.pageserver.allowed_errors.append(".*metrics endpoint refused the sent metrics*")
-    env.neon_cli.create_branch("test_metric_collection")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_metric_collection")
     endpoint = env.endpoints.create_start("test_metric_collection")
 
     pg_conn = endpoint.connect()
     cur = pg_conn.cursor()
 
-    tenant_id = TenantId(query_scalar(cur, "SHOW neon.tenant_id"))
-    timeline_id = TimelineId(query_scalar(cur, "SHOW neon.timeline_id"))
-
     cur.execute("CREATE TABLE foo (id int, counter int, t text)")
     cur.execute(
         """

test_runner/regress/test_ondemand_download.py

@@ -78,8 +78,8 @@ def test_ondemand_download_large_rel(
 
     client = env.pageserver.http_client()
 
-    tenant_id = endpoint.safe_psql("show neon.tenant_id")[0][0]
-    timeline_id = endpoint.safe_psql("show neon.timeline_id")[0][0]
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # We want to make sure that the data is large enough that the keyspace is partitioned.
     num_rows = 1000000
@@ -183,8 +183,8 @@ def test_ondemand_download_timetravel(
 
     client = env.pageserver.http_client()
 
-    tenant_id = endpoint.safe_psql("show neon.tenant_id")[0][0]
-    timeline_id = endpoint.safe_psql("show neon.timeline_id")[0][0]
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     lsns = []
 
@@ -342,8 +342,8 @@ def test_download_remote_layers_api(
 
     client = env.pageserver.http_client()
 
-    tenant_id = endpoint.safe_psql("show neon.tenant_id")[0][0]
-    timeline_id = endpoint.safe_psql("show neon.timeline_id")[0][0]
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     table_len = 10000
     with endpoint.cursor() as cur:
@@ -369,7 +369,7 @@ def test_download_remote_layers_api(
     filled_current_physical = get_api_current_physical_size()
     log.info(filled_current_physical)
     filled_size = get_resident_physical_size()
-    log.info(filled_size)
+    log.info(f"filled_size: {filled_size}")
     assert filled_current_physical == filled_size, "we don't yet do layer eviction"
 
     env.pageserver.stop()
@@ -377,7 +377,7 @@ def test_download_remote_layers_api(
     # remove all the layer files
     # XXX only delete some of the layer files, to show that it really just downloads all the layers
     for layer in (Path(env.repo_dir) / "tenants").glob("*/timelines/*/*-*_*"):
-        log.info(f"unlinking layer {layer}")
+        log.info(f"unlinking layer {layer.name}")
         layer.unlink()
 
     # Shut down safekeepers before starting the pageserver.
@@ -403,7 +403,7 @@ def test_download_remote_layers_api(
         filled_current_physical == get_api_current_physical_size()
     ), "current_physical_size is sum of loaded layer sizes, independent of whether local or remote"
     post_unlink_size = get_resident_physical_size()
-    log.info(post_unlink_size)
+    log.info(f"post_unlink_size: {post_unlink_size}")
     assert (
         post_unlink_size < filled_size
     ), "we just deleted layers and didn't cause anything to re-download them yet"
@@ -516,7 +516,6 @@ def test_compaction_downloads_on_demand_without_image_creation(
 
     tenant_id = env.initial_tenant
     timeline_id = env.initial_timeline
-    assert timeline_id is not None
 
     with env.endpoints.create_start("main") as endpoint:
         # no particular reason to create the layers like this, but we are sure
@@ -590,7 +589,6 @@ def test_compaction_downloads_on_demand_with_image_creation(
     env = neon_env_builder.init_start(initial_tenant_conf=stringify(conf))
     tenant_id = env.initial_tenant
     timeline_id = env.initial_timeline
-    assert timeline_id is not None
 
     pageserver_http = env.pageserver.http_client()
 

test_runner/regress/test_read_trace.py

@@ -2,7 +2,7 @@ from contextlib import closing
 
 from fixtures.neon_fixtures import NeonEnvBuilder
 from fixtures.pageserver.utils import wait_for_last_record_lsn
-from fixtures.types import Lsn, TenantId, TimelineId
+from fixtures.types import Lsn
 from fixtures.utils import query_scalar
 
 
@@ -12,24 +12,21 @@ from fixtures.utils import query_scalar
 # Additionally, tests that pageserver is able to create tenants with custom configs.
 def test_read_request_tracing(neon_env_builder: NeonEnvBuilder):
     neon_env_builder.num_safekeepers = 1
-    env = neon_env_builder.init_start()
-
-    tenant, _ = env.neon_cli.create_tenant(
-        conf={
+    env = neon_env_builder.init_start(
+        initial_tenant_conf={
             "trace_read_requests": "true",
         }
     )
 
-    timeline = env.neon_cli.create_timeline("test_trace_replay", tenant_id=tenant)
-    endpoint = env.endpoints.create_start("test_trace_replay", "main", tenant)
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+    endpoint = env.endpoints.create_start("main")
 
     with closing(endpoint.connect()) as conn:
         with conn.cursor() as cur:
             cur.execute("create table t (i integer);")
             cur.execute(f"insert into t values (generate_series(1,{10000}));")
             cur.execute("select count(*) from t;")
-            tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-            timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
             current_lsn = Lsn(query_scalar(cur, "SELECT pg_current_wal_flush_lsn()"))
     # wait until pageserver receives that data
     pageserver_http = env.pageserver.http_client()
@@ -38,5 +35,5 @@ def test_read_request_tracing(neon_env_builder: NeonEnvBuilder):
     # Stop postgres so we drop the connection and flush the traces
     endpoint.stop()
 
-    trace_path = env.repo_dir / "traces" / str(tenant) / str(timeline)
+    trace_path = env.repo_dir / "traces" / str(tenant_id) / str(timeline_id)
     assert trace_path.exists()

test_runner/regress/test_remote_storage.py

@@ -95,12 +95,12 @@ def test_remote_storage_backup_and_restore(
 
     client = env.pageserver.http_client()
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Thats because of UnreliableWrapper's injected failures
     env.pageserver.allowed_errors.append(
-        f".*failed to fetch tenant deletion mark at tenants/({tenant_id}|{env.initial_tenant})/deleted attempt 1.*"
+        f".*failed to fetch tenant deletion mark at tenants/{tenant_id}/deleted attempt 1.*"
     )
 
     checkpoint_numbers = range(1, 3)
@@ -403,8 +403,7 @@ def test_remote_timeline_client_calls_started_metric(
     )
 
     tenant_id = env.initial_tenant
-    assert env.initial_timeline is not None
-    timeline_id: TimelineId = env.initial_timeline
+    timeline_id = env.initial_timeline
 
     client = env.pageserver.http_client()
 
@@ -542,8 +541,7 @@ def test_timeline_deletion_with_files_stuck_in_upload_queue(
         }
     )
     tenant_id = env.initial_tenant
-    assert env.initial_timeline is not None
-    timeline_id: TimelineId = env.initial_timeline
+    timeline_id = env.initial_timeline
 
     timeline_path = env.timeline_dir(tenant_id, timeline_id)
 
@@ -808,8 +806,7 @@ def test_compaction_delete_before_upload(
     )
 
     tenant_id = env.initial_tenant
-    assert env.initial_timeline is not None
-    timeline_id: TimelineId = env.initial_timeline
+    timeline_id = env.initial_timeline
 
     client = env.pageserver.http_client()
 

test_runner/regress/test_tenant_delete.py

@@ -48,6 +48,11 @@ def test_tenant_delete_smoke(
 
     env = neon_env_builder.init_start()
 
+    # lucky race with stopping from flushing a layer we fail to schedule any uploads
+    env.pageserver.allowed_errors.append(
+        ".*layer flush task.+: could not flush frozen layer: update_metadata_file"
+    )
+
     ps_http = env.pageserver.http_client()
 
     # first try to delete non existing tenant
@@ -287,9 +292,8 @@ def test_delete_tenant_exercise_crash_safety_failpoints(
         )
 
 
-# TODO resume deletion (https://github.com/neondatabase/neon/issues/5006)
 @pytest.mark.parametrize("remote_storage_kind", available_remote_storages())
-def test_deleted_tenant_ignored_on_attach(
+def test_tenant_delete_is_resumed_on_attach(
     neon_env_builder: NeonEnvBuilder,
     remote_storage_kind: RemoteStorageKind,
     pg_bin: PgBin,
@@ -331,6 +335,8 @@ def test_deleted_tenant_ignored_on_attach(
         (
             # allow errors caused by failpoints
             f".*failpoint: {failpoint}",
+            # From deletion polling
+            f".*NotFound: tenant {env.initial_tenant}.*",
             # It appears when we stopped flush loop during deletion (attempt) and then pageserver is stopped
             ".*freeze_and_flush_on_shutdown.*failed to freeze and flush: cannot flush frozen layers when flush_loop is not running, state is Exited",
             # error from http response is also logged
@@ -376,20 +382,17 @@ def test_deleted_tenant_ignored_on_attach(
     env.pageserver.start()
 
     # now we call attach
-    with pytest.raises(
-        PageserverApiException, match="Tenant is marked as deleted on remote storage"
-    ):
-        ps_http.tenant_attach(tenant_id=tenant_id)
+    ps_http.tenant_attach(tenant_id=tenant_id)
 
-    # delete should be resumed (not yet)
-    # wait_tenant_status_404(ps_http, tenant_id, iterations)
+    # delete should be resumed
+    wait_tenant_status_404(ps_http, tenant_id, iterations)
 
     # we shouldn've created tenant dir on disk
     tenant_path = env.tenant_dir(tenant_id=tenant_id)
     assert not tenant_path.exists()
 
     if remote_storage_kind in available_s3_storages():
-        assert_prefix_not_empty(
+        assert_prefix_empty(
             neon_env_builder,
             prefix="/".join(
                 (

test_runner/regress/test_tenant_detach.py

@@ -463,8 +463,8 @@ def test_detach_while_attaching(
 
     client = env.pageserver.http_client()
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Attempts to connect from compute to pageserver while the tenant is
     # temporarily detached produces these errors in the pageserver log.
@@ -615,8 +615,8 @@ def test_ignored_tenant_download_missing_layers(
     pageserver_http = env.pageserver.http_client()
     endpoint = env.endpoints.create_start("main")
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Attempts to connect from compute to pageserver while the tenant is
     # temporarily detached produces these errors in the pageserver log.
@@ -679,10 +679,10 @@ def test_ignored_tenant_stays_broken_without_metadata(
     )
     env = neon_env_builder.init_start()
     pageserver_http = env.pageserver.http_client()
-    endpoint = env.endpoints.create_start("main")
+    env.endpoints.create_start("main")
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Attempts to connect from compute to pageserver while the tenant is
     # temporarily detached produces these errors in the pageserver log.
@@ -723,9 +723,9 @@ def test_load_attach_negatives(
     )
     env = neon_env_builder.init_start()
     pageserver_http = env.pageserver.http_client()
-    endpoint = env.endpoints.create_start("main")
+    env.endpoints.create_start("main")
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
+    tenant_id = env.initial_tenant
 
     # Attempts to connect from compute to pageserver while the tenant is
     # temporarily detached produces these errors in the pageserver log.
@@ -773,8 +773,8 @@ def test_ignore_while_attaching(
 
     pageserver_http = env.pageserver.http_client()
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Attempts to connect from compute to pageserver while the tenant is
     # temporarily detached produces these errors in the pageserver log.

test_runner/regress/test_tenants_with_remote_storage.py

@@ -142,8 +142,8 @@ def test_tenants_attached_after_download(
 
     client = env.pageserver.http_client()
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     # Thats because of UnreliableWrapper's injected failures
     env.pageserver.allowed_errors.append(
@@ -252,8 +252,8 @@ def test_tenant_redownloads_truncated_file_on_startup(
     pageserver_http = env.pageserver.http_client()
     endpoint = env.endpoints.create_start("main")
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
     with endpoint.cursor() as cur:
         cur.execute("CREATE TABLE t1 AS VALUES (123, 'foobar');")

test_runner/regress/test_threshold_based_eviction.py

@@ -10,7 +10,6 @@ from fixtures.neon_fixtures import (
 )
 from fixtures.pageserver.http import LayerMapInfo
 from fixtures.remote_storage import RemoteStorageKind
-from fixtures.types import TimelineId
 from pytest_httpserver import HTTPServer
 
 # NB: basic config change tests are in test_tenant_conf.py
@@ -45,7 +44,6 @@ def test_threshold_based_eviction(
     )
 
     tenant_id, timeline_id = env.initial_tenant, env.initial_timeline
-    assert isinstance(timeline_id, TimelineId)
 
     ps_http = env.pageserver.http_client()
     assert ps_http.tenant_config(tenant_id).effective_config["eviction_policy"] == {

test_runner/regress/test_timeline_delete.py

@@ -17,6 +17,7 @@ from fixtures.neon_fixtures import (
 )
 from fixtures.pageserver.http import PageserverApiException
 from fixtures.pageserver.utils import (
+    MANY_SMALL_LAYERS_TENANT_CONFIG,
     assert_prefix_empty,
     assert_prefix_not_empty,
     poll_for_remote_storage_iterations,
@@ -34,7 +35,7 @@ from fixtures.remote_storage import (
     available_s3_storages,
 )
 from fixtures.types import Lsn, TenantId, TimelineId
-from fixtures.utils import query_scalar, wait_until
+from fixtures.utils import query_scalar, run_pg_bench_small, wait_until
 
 
 def test_timeline_delete(neon_simple_env: NeonEnv):
@@ -208,7 +209,7 @@ def test_delete_timeline_exercise_crash_safety_failpoints(
     timeline_id = env.neon_cli.create_timeline("delete")
     with env.endpoints.create_start("delete") as endpoint:
         # generate enough layers
-        pg_bin.run(["pgbench", "-i", "-I dtGvp", "-s1", endpoint.connstr()])
+        run_pg_bench_small(pg_bin, endpoint.connstr())
         if remote_storage_kind is RemoteStorageKind.NOOP:
             wait_for_last_flush_lsn(env, endpoint, env.initial_tenant, timeline_id)
         else:
@@ -358,8 +359,8 @@ def test_timeline_resurrection_on_attach(
     ps_http = env.pageserver.http_client()
     pg = env.endpoints.create_start("main")
 
-    tenant_id = TenantId(pg.safe_psql("show neon.tenant_id")[0][0])
-    main_timeline_id = TimelineId(pg.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    main_timeline_id = env.initial_timeline
 
     with pg.cursor() as cur:
         cur.execute("CREATE TABLE f (i integer);")
@@ -487,15 +488,7 @@ def test_timeline_delete_fail_before_local_delete(neon_env_builder: NeonEnvBuild
     # Wait for tenant to finish loading.
     wait_until_tenant_active(ps_http, tenant_id=env.initial_tenant, iterations=10, period=1)
 
-    try:
-        data = ps_http.timeline_detail(env.initial_tenant, leaf_timeline_id)
-        log.debug(f"detail {data}")
-    except PageserverApiException as e:
-        log.debug(e)
-        if e.status_code != 404:
-            raise
-    else:
-        raise Exception("detail succeeded (it should return 404)")
+    wait_timeline_detail_404(ps_http, env.initial_tenant, leaf_timeline_id, iterations=4)
 
     assert (
         not leaf_timeline_path.exists()
@@ -519,8 +512,6 @@ def test_timeline_delete_fail_before_local_delete(neon_env_builder: NeonEnvBuild
         ),
     )
 
-    assert env.initial_timeline is not None
-
     for timeline_id in (intermediate_timeline_id, env.initial_timeline):
         timeline_delete_wait_completed(
             ps_http, tenant_id=env.initial_tenant, timeline_id=timeline_id
@@ -723,13 +714,9 @@ def test_timeline_delete_works_for_remote_smoke(
     ps_http = env.pageserver.http_client()
     pg = env.endpoints.create_start("main")
 
-    tenant_id = TenantId(pg.safe_psql("show neon.tenant_id")[0][0])
-    main_timeline_id = TimelineId(pg.safe_psql("show neon.timeline_id")[0][0])
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
 
-    assert tenant_id == env.initial_tenant
-    assert main_timeline_id == env.initial_timeline
-
-    assert env.initial_timeline is not None
     timeline_ids = [env.initial_timeline]
     for i in range(2):
         branch_timeline_id = env.neon_cli.create_branch(f"new{i}", "main")
@@ -750,9 +737,8 @@ def test_timeline_delete_works_for_remote_smoke(
             log.info("waiting for checkpoint upload")
             wait_for_upload(ps_http, tenant_id, branch_timeline_id, current_lsn)
             log.info("upload of checkpoint is done")
-            timeline_id = TimelineId(pg.safe_psql("show neon.timeline_id")[0][0])
 
-        timeline_ids.append(timeline_id)
+        timeline_ids.append(branch_timeline_id)
 
     for timeline_id in timeline_ids:
         assert_prefix_not_empty(
@@ -812,7 +798,7 @@ def test_delete_orphaned_objects(
     timeline_id = env.neon_cli.create_timeline("delete")
     with env.endpoints.create_start("delete") as endpoint:
         # generate enough layers
-        pg_bin.run(["pgbench", "-i", "-I dtGvp", "-s1", endpoint.connstr()])
+        run_pg_bench_small(pg_bin, endpoint.connstr())
         last_flush_lsn_upload(env, endpoint, env.initial_tenant, timeline_id)
 
     # write orphaned file that is missing from the index
@@ -848,3 +834,121 @@ def test_delete_orphaned_objects(
         )
 
     assert env.remote_storage.index_path(env.initial_tenant, timeline_id).exists()
+
+
+@pytest.mark.parametrize("remote_storage_kind", available_remote_storages())
+def test_timeline_delete_resumed_on_attach(
+    neon_env_builder: NeonEnvBuilder,
+    remote_storage_kind: RemoteStorageKind,
+    pg_bin: PgBin,
+):
+    neon_env_builder.enable_remote_storage(
+        remote_storage_kind=remote_storage_kind,
+        test_name="test_deleted_tenant_ignored_on_attach",
+    )
+
+    env = neon_env_builder.init_start(initial_tenant_conf=MANY_SMALL_LAYERS_TENANT_CONFIG)
+
+    tenant_id = env.initial_tenant
+
+    ps_http = env.pageserver.http_client()
+
+    timeline_id = env.neon_cli.create_timeline("delete")
+    with env.endpoints.create_start("delete") as endpoint:
+        # generate enough layers
+        run_pg_bench_small(pg_bin, endpoint.connstr())
+        last_flush_lsn_upload(env, endpoint, env.initial_tenant, timeline_id)
+
+        if remote_storage_kind in available_s3_storages():
+            assert_prefix_not_empty(
+                neon_env_builder,
+                prefix="/".join(
+                    (
+                        "tenants",
+                        str(env.initial_tenant),
+                        "timelines",
+                        str(timeline_id),
+                    )
+                ),
+            )
+
+    # failpoint before we remove index_part from s3
+    failpoint = "timeline-delete-during-rm"
+    ps_http.configure_failpoints((failpoint, "return"))
+
+    env.pageserver.allowed_errors.extend(
+        (
+            # allow errors caused by failpoints
+            f".*failpoint: {failpoint}",
+            # It appears when we stopped flush loop during deletion (attempt) and then pageserver is stopped
+            ".*freeze_and_flush_on_shutdown.*failed to freeze and flush: cannot flush frozen layers when flush_loop is not running, state is Exited",
+            # error from http response is also logged
+            ".*InternalServerError\\(Tenant is marked as deleted on remote storage.*",
+            # Polling after attach may fail with this
+            f".*InternalServerError\\(Tenant {tenant_id} is not active.*",
+            '.*shutdown_pageserver{exit_code=0}: stopping left-over name="remote upload".*',
+        )
+    )
+
+    iterations = poll_for_remote_storage_iterations(remote_storage_kind)
+
+    ps_http.timeline_delete(tenant_id, timeline_id)
+
+    timeline_info = wait_until_timeline_state(
+        pageserver_http=ps_http,
+        tenant_id=env.initial_tenant,
+        timeline_id=timeline_id,
+        expected_state="Broken",
+        iterations=iterations,
+    )
+
+    reason = timeline_info["state"]["Broken"]["reason"]
+    log.info(f"timeline broken: {reason}")
+
+    # failpoint may not be the only error in the stack
+    assert reason.endswith(f"failpoint: {failpoint}"), reason
+
+    if remote_storage_kind in available_s3_storages():
+        assert_prefix_not_empty(
+            neon_env_builder,
+            prefix="/".join(
+                (
+                    "tenants",
+                    str(tenant_id),
+                    "timelines",
+                    str(timeline_id),
+                )
+            ),
+        )
+
+    # now we stop pageserver and remove local tenant state
+    env.endpoints.stop_all()
+    env.pageserver.stop()
+
+    dir_to_clear = Path(env.repo_dir) / "tenants"
+    shutil.rmtree(dir_to_clear)
+    os.mkdir(dir_to_clear)
+
+    env.pageserver.start()
+
+    # now we call attach
+    ps_http.tenant_attach(tenant_id=tenant_id)
+
+    # delete should be resumed
+    wait_timeline_detail_404(ps_http, env.initial_tenant, timeline_id, iterations=iterations)
+
+    tenant_path = env.timeline_dir(tenant_id=tenant_id, timeline_id=timeline_id)
+    assert not tenant_path.exists()
+
+    if remote_storage_kind in available_s3_storages():
+        assert_prefix_empty(
+            neon_env_builder,
+            prefix="/".join(
+                (
+                    "tenants",
+                    str(timeline_id),
+                    "timelines",
+                    str(timeline_id),
+                )
+            ),
+        )

test_runner/regress/test_wal_acceptor.py

@@ -270,7 +270,8 @@ def test_broker(neon_env_builder: NeonEnvBuilder):
     neon_env_builder.enable_local_fs_remote_storage()
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_broker", "main")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_broker", "main")
 
     # FIXME: Is this expected?
     env.pageserver.allowed_errors.append(
@@ -280,10 +281,6 @@ def test_broker(neon_env_builder: NeonEnvBuilder):
     endpoint = env.endpoints.create_start("test_broker")
     endpoint.safe_psql("CREATE TABLE t(key int primary key, value text)")
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     # wait until remote_consistent_lsn gets advanced on all safekeepers
     clients = [sk.http_client() for sk in env.safekeepers]
     stat_before = [cli.timeline_status(tenant_id, timeline_id) for cli in clients]
@@ -325,7 +322,8 @@ def test_wal_removal(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
         ".*init_tenant_mgr: marking .* as locally complete, while it doesnt exist in remote index.*"
     )
 
-    env.neon_cli.create_branch("test_safekeepers_wal_removal")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_safekeepers_wal_removal")
     endpoint = env.endpoints.create_start("test_safekeepers_wal_removal")
 
     # Note: it is important to insert at least two segments, as currently
@@ -338,9 +336,6 @@ def test_wal_removal(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
         ]
     )
 
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     # force checkpoint to advance remote_consistent_lsn
     pageserver_conn_options = {}
     if auth_enabled:
@@ -451,13 +446,10 @@ def test_wal_backup(neon_env_builder: NeonEnvBuilder, remote_storage_kind: Remot
 
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_safekeepers_wal_backup")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_safekeepers_wal_backup")
     endpoint = env.endpoints.create_start("test_safekeepers_wal_backup")
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     pg_conn = endpoint.connect()
     cur = pg_conn.cursor()
     cur.execute("create table t(key int, value text)")
@@ -505,14 +497,11 @@ def test_s3_wal_replay(neon_env_builder: NeonEnvBuilder, remote_storage_kind: Re
     neon_env_builder.remote_storage_users = RemoteStorageUsers.SAFEKEEPER
 
     env = neon_env_builder.init_start()
-    env.neon_cli.create_branch("test_s3_wal_replay")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_s3_wal_replay")
 
     endpoint = env.endpoints.create_start("test_s3_wal_replay")
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     expected_sum = 0
 
     with closing(endpoint.connect()) as conn:
@@ -796,15 +785,12 @@ def test_timeline_status(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
     neon_env_builder.auth_enabled = auth_enabled
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_timeline_status")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_timeline_status")
     endpoint = env.endpoints.create_start("test_timeline_status")
 
     wa = env.safekeepers[0]
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     if not auth_enabled:
         wa_http_cli = wa.http_client()
         wa_http_cli.check_status()
@@ -887,15 +873,12 @@ def test_start_replication_term(neon_env_builder: NeonEnvBuilder):
 
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_start_replication_term")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_start_replication_term")
     endpoint = env.endpoints.create_start("test_start_replication_term")
 
     endpoint.safe_psql("CREATE TABLE t(key int primary key, value text)")
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     sk = env.safekeepers[0]
     sk_http_cli = sk.http_client()
     tli_status = sk_http_cli.timeline_status(tenant_id, timeline_id)
@@ -922,15 +905,12 @@ def test_sk_auth(neon_env_builder: NeonEnvBuilder):
     neon_env_builder.auth_enabled = True
     env = neon_env_builder.init_start()
 
-    env.neon_cli.create_branch("test_sk_auth")
-    endpoint = env.endpoints.create_start("test_sk_auth")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_sk_auth")
+    env.endpoints.create_start("test_sk_auth")
 
     sk = env.safekeepers[0]
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     tenant_token = env.auth_keys.generate_tenant_token(tenant_id)
     full_token = env.auth_keys.generate_safekeeper_token()
 
@@ -1185,7 +1165,8 @@ def test_replace_safekeeper(neon_env_builder: NeonEnvBuilder):
 
     neon_env_builder.num_safekeepers = 4
     env = neon_env_builder.init_start()
-    env.neon_cli.create_branch("test_replace_safekeeper")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_replace_safekeeper")
 
     log.info("Use only first 3 safekeepers")
     env.safekeepers[3].stop()
@@ -1193,10 +1174,6 @@ def test_replace_safekeeper(neon_env_builder: NeonEnvBuilder):
     endpoint.active_safekeepers = [1, 2, 3]
     endpoint.start()
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     execute_payload(endpoint)
     show_statuses(env.safekeepers, tenant_id, timeline_id)
 
@@ -1448,7 +1425,8 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
 
     neon_env_builder.num_safekeepers = 4
     env = neon_env_builder.init_start()
-    env.neon_cli.create_branch("test_pull_timeline")
+    tenant_id = env.initial_tenant
+    timeline_id = env.neon_cli.create_branch("test_pull_timeline")
 
     log.info("Use only first 3 safekeepers")
     env.safekeepers[3].stop()
@@ -1456,10 +1434,6 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
     endpoint.active_safekeepers = [1, 2, 3]
     endpoint.start()
 
-    # learn neon timeline from compute
-    tenant_id = TenantId(endpoint.safe_psql("show neon.tenant_id")[0][0])
-    timeline_id = TimelineId(endpoint.safe_psql("show neon.timeline_id")[0][0])
-
     execute_payload(endpoint)
     show_statuses(env.safekeepers, tenant_id, timeline_id)
 

workspace_hack/Cargo.toml

@@ -14,11 +14,13 @@ publish = false
 ### BEGIN HAKARI SECTION
 [dependencies]
 anyhow = { version = "1", features = ["backtrace"] }
+axum = { version = "0.6", features = ["ws"] }
 bytes = { version = "1", features = ["serde"] }
 chrono = { version = "0.4", default-features = false, features = ["clock", "serde"] }
 clap = { version = "4", features = ["derive", "string"] }
 clap_builder = { version = "4", default-features = false, features = ["color", "help", "std", "string", "suggestions", "usage"] }
 crossbeam-utils = { version = "0.8" }
+digest = { version = "0.10", features = ["mac", "std"] }
 either = { version = "1" }
 fail = { version = "0.5", default-features = false, features = ["failpoints"] }
 futures = { version = "0.3" }
@@ -27,6 +29,7 @@ futures-core = { version = "0.3" }
 futures-executor = { version = "0.3" }
 futures-sink = { version = "0.3" }
 futures-util = { version = "0.3", features = ["channel", "io", "sink"] }
+hyper = { version = "0.14", features = ["full"] }
 itertools = { version = "0.10" }
 libc = { version = "0.2", features = ["extra_traits"] }
 log = { version = "0.4", default-features = false, features = ["std"] }
@@ -45,6 +48,7 @@ rustls = { version = "0.20", features = ["dangerous_configuration"] }
 scopeguard = { version = "1" }
 serde = { version = "1", features = ["alloc", "derive"] }
 serde_json = { version = "1", features = ["raw_value"] }
+smallvec = { version = "1", default-features = false, features = ["write"] }
 socket2 = { version = "0.4", default-features = false, features = ["all"] }
 tokio = { version = "1", features = ["fs", "io-std", "io-util", "macros", "net", "process", "rt-multi-thread", "signal", "test-util"] }
 tokio-rustls = { version = "0.23" }
@@ -54,7 +58,6 @@ toml_edit = { version = "0.19", features = ["serde"] }
 tower = { version = "0.4", features = ["balance", "buffer", "limit", "retry", "timeout", "util"] }
 tracing = { version = "0.1", features = ["log"] }
 tracing-core = { version = "0.1" }
-tracing-subscriber = { version = "0.3", default-features = false, features = ["env-filter", "fmt", "json", "smallvec", "tracing-log"] }
 url = { version = "2", features = ["serde"] }
 
 [build-dependencies]