Fix rust syntax

Add infrastructure to dynamically load postgres extensions and shared libraries from remote extension storage.

Before postgres start downloads list of available remote extensions and libraries, and also downloads 'shared_preload_libraries'. After postgres is running, 'compute_ctl' listens for HTTP requests to load files.

Postgres has new GUC 'extension_server_port' to specify port on which 'compute_ctl' listens for requests.

When PostgreSQL requests a file, 'compute_ctl' downloads it.

See more details about feature design and remote extension storage layout in docs/rfcs/024-extension-loading.md

---------

Co-authored-by: Anastasia Lubennikova <anastasia@neon.tech>
Co-authored-by: Alek Westover <alek.westover@gmail.com>

.cargo/config.toml

@@ -12,6 +12,11 @@ opt-level = 3
 # Turn on a small amount of optimization in Development mode.
 opt-level = 1
 
+[build]
+# This is only present for local builds, as it will be overridden
+# by the RUSTDOCFLAGS env var in CI.
+rustdocflags = ["-Arustdoc::private_intra_doc_links"]
+
 [alias]
 build_testing = ["build", "--features", "testing"]
 neon = ["run", "--bin", "neon_local"]

.dockerignore

@@ -21,4 +21,5 @@
 !workspace_hack/
 !neon_local/
 !scripts/ninstall.sh
+!scripts/combine_control_files.py
 !vm-cgconfig.conf

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

@@ -150,6 +150,14 @@ runs:
           EXTRA_PARAMS="--flaky-tests-json $TEST_OUTPUT/flaky.json $EXTRA_PARAMS"
         fi
 
+        # We use pytest-split plugin to run benchmarks in parallel on different CI runners
+        if [ "${TEST_SELECTION}" = "test_runner/performance" ] && [ "${{ inputs.build_type }}" != "remote" ]; then
+          mkdir -p $TEST_OUTPUT
+          poetry run ./scripts/benchmark_durations.py "${TEST_RESULT_CONNSTR}" --days 10 --output "$TEST_OUTPUT/benchmark_durations.json"
+
+          EXTRA_PARAMS="--durations-path $TEST_OUTPUT/benchmark_durations.json $EXTRA_PARAMS"
+        fi
+
         if [[ "${{ inputs.build_type }}" == "debug" ]]; then
           cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
         elif [[ "${{ inputs.build_type }}" == "release" ]]; then
@@ -201,4 +209,4 @@ runs:
       uses: ./.github/actions/allure-report-store
       with:
         report-dir: /tmp/test_output/allure/results
-        unique-key: ${{ inputs.build_type }}
+        unique-key: ${{ inputs.build_type }}-${{ inputs.pg_version }}

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

@@ -0,0 +1,55 @@
+name: Handle `approved-for-ci-run` label
+# This workflow helps to run CI pipeline for PRs made by external contributors (from forks).
+
+on:
+  pull_request:
+    types:
+      # Default types that triggers a workflow ([1]):
+      # - [1] https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#pull_request
+      - opened
+      - synchronize
+      - reopened
+      # Types that we wand to handle in addition to keep labels tidy:
+      - closed
+      # Actual magic happens here:
+      - labeled
+
+env:
+  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  PR_NUMBER: ${{ github.event.pull_request.number }}
+
+jobs:
+  remove-label:
+    # Remove `approved-for-ci-run` label if the workflow is triggered by changes in a PR.
+    # The PR should be reviewed and labelled manually again.
+
+    runs-on: [ ubuntu-latest ]
+
+    if: |
+      contains(fromJSON('["opened", "synchronize", "reopened", "closed"]'), github.event.action) &&
+      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
+
+    steps:
+      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"
+
+  create-branch:
+    # Create a local branch for an `approved-for-ci-run` labelled PR to run CI pipeline in it.
+
+    runs-on: [ ubuntu-latest ]
+
+    if: |
+      github.event.action == 'labeled' &&
+      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
+
+    steps:
+      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"
+
+      - uses: actions/checkout@v3
+        with:
+          ref: main
+
+      - run: gh pr checkout "${PR_NUMBER}"
+
+      - run: git checkout -b "ci-run/pr-${PR_NUMBER}"
+
+      - run: git push --force origin "ci-run/pr-${PR_NUMBER}"

.github/workflows/build_and_test.yml

@@ -5,6 +5,7 @@ on:
     branches:
       - main
       - release
+      - ci-run/pr-*
   pull_request:
 
 defaults:
@@ -127,6 +128,11 @@ jobs:
       - name: Run cargo clippy (release)
         run: cargo hack --feature-powerset clippy --release $CLIPPY_COMMON_ARGS
 
+      - name: Check documentation generation
+        run: cargo doc --workspace --no-deps --document-private-items
+        env:
+            RUSTDOCFLAGS: "-Dwarnings -Arustdoc::private_intra_doc_links"
+
       # Use `${{ !cancelled() }}` to run quck tests after the longer clippy run
       - name: Check formatting
         if: ${{ !cancelled() }}
@@ -390,13 +396,11 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
+        pytest_split_group: [ 1, 2, 3, 4 ]
         build_type: [ release ]
     steps:
       - name: Checkout
         uses: actions/checkout@v3
-        with:
-          submodules: true
-          fetch-depth: 1
 
       - name: Pytest benchmarks
         uses: ./.github/actions/run-python-test-set
@@ -405,9 +409,11 @@ jobs:
           test_selection: performance
           run_in_parallel: false
           save_perf_report: ${{ github.ref_name == 'main' }}
+          extra_params: --splits ${{ strategy.job-total }} --group ${{ matrix.pytest_split_group }}
         env:
           VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}"
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
@@ -788,7 +794,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.12.1
+      VM_BUILDER_VERSION: v0.13.1
 
     steps:
       - name: Checkout
@@ -949,22 +955,15 @@ jobs:
         version: [ v14, v15 ]
 
     env:
-      # While on transition period we extract public extensions from compute-node image and custom extensions from extensions image.
-      # Later all the extensions will be moved to extensions image.
-      EXTENSIONS_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:latest
-      COMPUTE_NODE_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:latest
+      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' &&
-          'neon-prod-extensions-ap-southeast-1 neon-prod-extensions-eu-central-1 neon-prod-extensions-us-east-1 neon-prod-extensions-us-east-2 neon-prod-extensions-us-west-2' ||
-          'neon-dev-extensions-eu-central-1 neon-dev-extensions-eu-west-1 neon-dev-extensions-us-east-2' }}
+      S3_BUCKETS: ${{ github.ref_name == 'release' && vars.S3_EXTENSIONS_BUCKETS_PROD || vars.S3_EXTENSIONS_BUCKETS_DEV }}
 
     steps:
       - name: Pull postgres-extensions image
         run: |
           docker pull ${EXTENSIONS_IMAGE}
-          docker pull ${COMPUTE_NODE_IMAGE}
 
       - name: Create postgres-extensions container
         id: create-container
@@ -972,44 +971,23 @@ jobs:
           EID=$(docker create ${EXTENSIONS_IMAGE} true)
           echo "EID=${EID}" >> $GITHUB_OUTPUT
 
-          CID=$(docker create ${COMPUTE_NODE_IMAGE} true)
-          echo "CID=${CID}" >> $GITHUB_OUTPUT
-
       - name: Extract postgres-extensions from container
         run: |
-          rm -rf ./extensions-to-upload ./custom-extensions # Just in case
+          rm -rf ./extensions-to-upload # Just in case
+          mkdir -p extensions-to-upload
 
-          # In compute image we have a bit different directory layout
-          mkdir -p extensions-to-upload/share
-          docker cp ${{ steps.create-container.outputs.CID }}:/usr/local/share/extension ./extensions-to-upload/share/extension
-          docker cp ${{ steps.create-container.outputs.CID }}:/usr/local/lib             ./extensions-to-upload/lib
-
-          # Delete Neon extensitons (they always present on compute-node image)
-          rm -rf ./extensions-to-upload/share/extension/neon*
-          rm -rf ./extensions-to-upload/lib/neon*
-
-          # Delete leftovers from the extension build step
-          rm -rf ./extensions-to-upload/lib/pgxs
-          rm -rf ./extensions-to-upload/lib/pkgconfig
-
-          docker cp ${{ steps.create-container.outputs.EID }}:/extensions ./custom-extensions
-          for EXT_NAME in $(ls ./custom-extensions); do
-            mkdir -p ./extensions-to-upload/${EXT_NAME}/share
-
-            mv ./custom-extensions/${EXT_NAME}/share/extension ./extensions-to-upload/${EXT_NAME}/share/extension
-            mv ./custom-extensions/${EXT_NAME}/lib             ./extensions-to-upload/${EXT_NAME}/lib
-          done
+          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}); do
+          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.CID || steps.create-container.outputs.EID) }}
+        if: ${{ always() && steps.create-container.outputs.EID }}
         run: |
-          docker rm ${{ steps.create-container.outputs.CID }} || true
           docker rm ${{ steps.create-container.outputs.EID }} || true
 
   deploy:

.github/workflows/neon_extra_builds.yml

@@ -3,7 +3,8 @@ name: Check neon with extra platform builds
 on:
   push:
     branches:
-    - main
+      - main
+      - ci-run/pr-*
   pull_request:
 
 defaults:

Cargo.lock

@@ -740,6 +740,9 @@ name = "cc"
 version = "1.0.79"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "50d30906286121d95be3d479533b458f87493b30a4b5f79a607db8f5d11aa91f"
+dependencies = [
+ "jobserver",
+]
 
 [[package]]
 name = "cexpr"
@@ -922,6 +925,7 @@ dependencies = [
  "url",
  "utils",
  "workspace_hack",
+ "zstd",
 ]
 
 [[package]]
@@ -1975,6 +1979,15 @@ version = "1.0.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "453ad9f582a441959e5f0d088b02ce04cfe8d51a8eaf077f12ac6d3e94164ca6"
 
+[[package]]
+name = "jobserver"
+version = "0.1.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "936cfd212a0155903bcbc060e316fb6cc7cbf2e1907329391ebadc1fe0ce77c2"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "js-sys"
 version = "0.3.63"
@@ -2382,9 +2395,9 @@ dependencies = [
 
 [[package]]
 name = "opentelemetry"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69d6c3d7288a106c0a363e4b0e8d308058d56902adefb16f4936f417ffef086e"
+checksum = "5f4b8347cc26099d3aeee044065ecc3ae11469796b4d65d065a23a584ed92a6f"
 dependencies = [
  "opentelemetry_api",
  "opentelemetry_sdk",
@@ -2392,9 +2405,9 @@ dependencies = [
 
 [[package]]
 name = "opentelemetry-http"
-version = "0.7.0"
+version = "0.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1edc79add46364183ece1a4542592ca593e6421c60807232f5b8f7a31703825d"
+checksum = "a819b71d6530c4297b49b3cae2939ab3a8cc1b9f382826a1bc29dd0ca3864906"
 dependencies = [
  "async-trait",
  "bytes",
@@ -2405,9 +2418,9 @@ dependencies = [
 
 [[package]]
 name = "opentelemetry-otlp"
-version = "0.11.0"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d1c928609d087790fc936a1067bdc310ae702bdf3b090c3f281b713622c8bbde"
+checksum = "8af72d59a4484654ea8eb183fea5ae4eb6a41d7ac3e3bae5f4d2a282a3a7d3ca"
 dependencies = [
  "async-trait",
  "futures",
@@ -2423,48 +2436,47 @@ dependencies = [
 
 [[package]]
 name = "opentelemetry-proto"
-version = "0.1.0"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d61a2f56df5574508dd86aaca016c917489e589ece4141df1b5e349af8d66c28"
+checksum = "045f8eea8c0fa19f7d48e7bc3128a39c2e5c533d5c61298c548dfefc1064474c"
 dependencies = [
  "futures",
  "futures-util",
  "opentelemetry",
  "prost",
  "tonic 0.8.3",
- "tonic-build 0.8.4",
 ]
 
 [[package]]
 name = "opentelemetry-semantic-conventions"
-version = "0.10.0"
+version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b02e0230abb0ab6636d18e2ba8fa02903ea63772281340ccac18e0af3ec9eeb"
+checksum = "24e33428e6bf08c6f7fcea4ddb8e358fab0fe48ab877a87c70c6ebe20f673ce5"
 dependencies = [
  "opentelemetry",
 ]
 
 [[package]]
 name = "opentelemetry_api"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c24f96e21e7acc813c7a8394ee94978929db2bcc46cf6b5014fc612bf7760c22"
+checksum = "ed41783a5bf567688eb38372f2b7a8530f5a607a4b49d38dd7573236c23ca7e2"
 dependencies = [
  "fnv",
  "futures-channel",
  "futures-util",
  "indexmap",
- "js-sys",
  "once_cell",
  "pin-project-lite",
  "thiserror",
+ "urlencoding",
 ]
 
 [[package]]
 name = "opentelemetry_sdk"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1ca41c4933371b61c2a2f214bf16931499af4ec90543604ec828f7a625c09113"
+checksum = "8b3a2a91fdbfdd4d212c0dcc2ab540de2c2bcbbd90be17de7a7daf8822d010c1"
 dependencies = [
  "async-trait",
  "crossbeam-channel",
@@ -2510,6 +2522,7 @@ dependencies = [
  "pageserver",
  "postgres_ffi",
  "svg_fmt",
+ "tokio",
  "utils",
  "workspace_hack",
 ]
@@ -2548,6 +2561,7 @@ dependencies = [
  "metrics",
  "nix",
  "num-traits",
+ "num_cpus",
  "once_cell",
  "pageserver_api",
  "pin-project-lite",
@@ -2784,7 +2798,7 @@ dependencies = [
 [[package]]
 name = "postgres"
 version = "0.19.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=9011f7110db12b5e15afaf98f8ac834501d50ddc#9011f7110db12b5e15afaf98f8ac834501d50ddc"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -2797,7 +2811,7 @@ dependencies = [
 [[package]]
 name = "postgres-native-tls"
 version = "0.5.0"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=9011f7110db12b5e15afaf98f8ac834501d50ddc#9011f7110db12b5e15afaf98f8ac834501d50ddc"
 dependencies = [
  "native-tls",
  "tokio",
@@ -2808,7 +2822,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=9011f7110db12b5e15afaf98f8ac834501d50ddc#9011f7110db12b5e15afaf98f8ac834501d50ddc"
 dependencies = [
  "base64 0.20.0",
  "byteorder",
@@ -2826,7 +2840,7 @@ dependencies = [
 [[package]]
 name = "postgres-types"
 version = "0.2.4"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=9011f7110db12b5e15afaf98f8ac834501d50ddc#9011f7110db12b5e15afaf98f8ac834501d50ddc"
 dependencies = [
  "bytes",
  "fallible-iterator",
@@ -2940,9 +2954,9 @@ checksum = "dc375e1527247fe1a97d8b7156678dfe7c1af2fc075c9a4db3690ecd2a148068"
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.58"
+version = "1.0.64"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa1fb82fc0c281dd9671101b66b771ebbe1eaf967b96ac8740dcba4b70005ca8"
+checksum = "78803b62cbf1f46fde80d7c0e803111524b9877184cfe7c3033659490ac7a7da"
 dependencies = [
  "unicode-ident",
 ]
@@ -3331,9 +3345,9 @@ dependencies = [
 
 [[package]]
 name = "reqwest-tracing"
-version = "0.4.4"
+version = "0.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "783e8130d2427ddd7897dd3f814d4a3aea31b05deb42a4fdf8c18258fe5aefd1"
+checksum = "1b97ad83c2fc18113346b7158d79732242002427c30f620fa817c1f32901e0a8"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -3858,7 +3872,8 @@ dependencies = [
 [[package]]
 name = "sharded-slab"
 version = "0.1.4"
-source = "git+https://github.com/neondatabase/sharded-slab.git?rev=98d16753ab01c61f0a028de44167307a00efea00#98d16753ab01c61f0a028de44167307a00efea00"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "900fba806f70c630b0a382d0d825e17a0f19fcd059a2ade1ff237bcddf446b31"
 dependencies = [
  "lazy_static",
 ]
@@ -4001,7 +4016,7 @@ dependencies = [
  "tokio",
  "tokio-stream",
  "tonic 0.9.2",
- "tonic-build 0.9.2",
+ "tonic-build",
  "tracing",
  "utils",
  "workspace_hack",
@@ -4102,7 +4117,7 @@ checksum = "4b55807c0344e1e6c04d7c965f5289c39a8d94ae23ed5c0b57aabac549f871c6"
 dependencies = [
  "filetime",
  "libc",
- "xattr",
+ "xattr 0.2.3",
 ]
 
 [[package]]
@@ -4315,7 +4330,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.7"
-source = "git+https://github.com/neondatabase/rust-postgres.git?rev=1aaedab101b23f7612042850d8f2036810fa7c7f#1aaedab101b23f7612042850d8f2036810fa7c7f"
+source = "git+https://github.com/neondatabase/rust-postgres.git?rev=9011f7110db12b5e15afaf98f8ac834501d50ddc#9011f7110db12b5e15afaf98f8ac834501d50ddc"
 dependencies = [
  "async-trait",
  "byteorder",
@@ -4383,16 +4398,17 @@ dependencies = [
 
 [[package]]
 name = "tokio-tar"
-version = "0.3.0"
-source = "git+https://github.com/neondatabase/tokio-tar.git?rev=404df61437de0feef49ba2ccdbdd94eb8ad6e142#404df61437de0feef49ba2ccdbdd94eb8ad6e142"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d5714c010ca3e5c27114c1cdeb9d14641ace49874aa5626d7149e47aedace75"
 dependencies = [
  "filetime",
  "futures-core",
  "libc",
- "redox_syscall 0.2.16",
+ "redox_syscall 0.3.5",
  "tokio",
  "tokio-stream",
- "xattr",
+ "xattr 1.0.0",
 ]
 
 [[package]]
@@ -4519,19 +4535,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "tonic-build"
-version = "0.8.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5bf5e9b9c0f7e0a7c027dcfaba7b2c60816c7049171f679d99ee2ff65d0de8c4"
-dependencies = [
- "prettyplease 0.1.25",
- "proc-macro2",
- "prost-build",
- "quote",
- "syn 1.0.109",
-]
-
 [[package]]
 name = "tonic-build"
 version = "0.9.2"
@@ -4655,9 +4658,9 @@ dependencies = [
 
 [[package]]
 name = "tracing-opentelemetry"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "21ebb87a95ea13271332df069020513ab70bdb5637ca42d6e492dc3bbbad48de"
+checksum = "00a39dcf9bfc1742fa4d6215253b33a6e474be78275884c216fc2a06267b3600"
 dependencies = [
  "once_cell",
  "opentelemetry",
@@ -4882,6 +4885,7 @@ dependencies = [
  "tempfile",
  "thiserror",
  "tokio",
+ "tokio-stream",
  "tracing",
  "tracing-error",
  "tracing-subscriber",
@@ -5308,6 +5312,7 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "bytes",
+ "cc",
  "chrono",
  "clap",
  "clap_builder",
@@ -5379,6 +5384,15 @@ dependencies = [
  "libc",
 ]
 
+[[package]]
+name = "xattr"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ea263437ca03c1522846a4ddafbca2542d0ad5ed9b784909d4b27b76f62bc34a"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "xmlparser"
 version = "0.13.5"
@@ -5399,3 +5413,33 @@ name = "zeroize"
 version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2a0956f1ba7c7909bfb66c2e9e4124ab6f6482560f6628b5aaeba39207c9aad9"
+
+[[package]]
+name = "zstd"
+version = "0.12.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1a27595e173641171fc74a1232b7b1c7a7cb6e18222c11e9dfb9888fa424c53c"
+dependencies = [
+ "zstd-safe",
+]
+
+[[package]]
+name = "zstd-safe"
+version = "6.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee98ffd0b48ee95e6c5168188e44a54550b1564d9d530ee21d5f0eaed1069581"
+dependencies = [
+ "libc",
+ "zstd-sys",
+]
+
+[[package]]
+name = "zstd-sys"
+version = "2.0.8+zstd.1.5.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5556e6ee25d32df2586c098bbfa278803692a20d0ab9565e049480d52707ec8c"
+dependencies = [
+ "cc",
+ "libc",
+ "pkg-config",
+]

Cargo.toml

@@ -84,9 +84,9 @@ notify = "5.0.0"
 num_cpus = "1.15"
 num-traits = "0.2.15"
 once_cell = "1.13"
-opentelemetry = "0.18.0"
-opentelemetry-otlp = { version = "0.11.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
-opentelemetry-semantic-conventions = "0.10.0"
+opentelemetry = "0.19.0"
+opentelemetry-otlp = { version = "0.12.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry-semantic-conventions = "0.11.0"
 parking_lot = "0.12"
 pbkdf2 = "0.12.1"
 pin-project-lite = "0.2"
@@ -95,7 +95,7 @@ prost = "0.11"
 rand = "0.8"
 regex = "1.4"
 reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_18"] }
+reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_19"] }
 reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
@@ -124,13 +124,14 @@ tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.9.0"
 tokio-rustls = "0.23"
 tokio-stream = "0.1"
+tokio-tar = "0.3"
 tokio-util = { version = "0.7", features = ["io"] }
 toml = "0.7"
 toml_edit = "0.19"
 tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tracing = "0.1"
 tracing-error = "0.2.0"
-tracing-opentelemetry = "0.18.0"
+tracing-opentelemetry = "0.19.0"
 tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter"] }
 url = "2.2"
 uuid = { version = "1.2", features = ["v4", "serde"] }
@@ -143,12 +144,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-tokio-tar = { git = "https://github.com/neondatabase/tokio-tar.git", rev="404df61437de0feef49ba2ccdbdd94eb8ad6e142" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -183,12 +183,7 @@ tonic-build = "0.9"
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-
-# Changes the MAX_THREADS limit from 4096 to 32768.
-# This is a temporary workaround for using tracing from many threads in safekeepers code,
-# until async safekeepers patch is merged to the main.
-sharded-slab = { git = "https://github.com/neondatabase/sharded-slab.git", rev="98d16753ab01c61f0a028de44167307a00efea00" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
 
 ################# Binary contents sections
 

Dockerfile.compute-node

@@ -13,7 +13,7 @@ FROM debian:bullseye-slim AS build-deps
 RUN apt update &&  \
     apt install -y git autoconf automake libtool build-essential bison flex libreadline-dev \
     zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget pkg-config libssl-dev \
-    libicu-dev libxslt1-dev liblz4-dev libzstd-dev
+    libicu-dev libxslt1-dev liblz4-dev libzstd-dev zstd
 
 #########################################################################################
 #
@@ -77,6 +77,7 @@ ENV PATH "/usr/local/pgsql/bin:$PATH"
 RUN wget https://download.osgeo.org/postgis/source/postgis-3.3.2.tar.gz -O postgis.tar.gz && \
     echo "9a2a219da005a1730a39d1959a1c7cec619b1efb009b65be80ffc25bad299068 postgis.tar.gz" | sha256sum --check && \
     mkdir postgis-src && cd postgis-src && tar xvzf ../postgis.tar.gz --strip-components=1 -C . && \
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /before.txt &&\
     ./autogen.sh && \
     ./configure --with-sfcgal=/usr/local/bin/sfcgal-config && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -89,17 +90,28 @@ RUN wget https://download.osgeo.org/postgis/source/postgis-3.3.2.tar.gz -O postg
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/postgis_tiger_geocoder.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/postgis_topology.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/address_standardizer.control && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/address_standardizer_data_us.control
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/address_standardizer_data_us.control && \
+    mkdir -p /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/postgis.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/postgis_raster.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/postgis_sfcgal.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/postgis_tiger_geocoder.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/postgis_topology.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/address_standardizer.control /extensions/postgis && \
+    cp /usr/local/pgsql/share/extension/address_standardizer_data_us.control /extensions/postgis
 
 RUN wget https://github.com/pgRouting/pgrouting/archive/v3.4.2.tar.gz -O pgrouting.tar.gz && \
     echo "cac297c07d34460887c4f3b522b35c470138760fe358e351ad1db4edb6ee306e pgrouting.tar.gz" | sha256sum --check && \
     mkdir pgrouting-src && cd pgrouting-src && tar xvzf ../pgrouting.tar.gz --strip-components=1 -C . && \
-    mkdir build && \
-    cd build && \
+    mkdir build && cd build && \
     cmake -DCMAKE_BUILD_TYPE=Release .. && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrouting.control
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrouting.control && \
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /after.txt &&\
+    cp /usr/local/pgsql/share/extension/pgrouting.control /extensions/postgis && \
+    sort -o /before.txt /before.txt && sort -o /after.txt /after.txt && \
+    comm -13 /before.txt /after.txt | tar --directory=/usr/local/pgsql --zstd -cf /extensions/postgis.tar.zst -T -
 
 #########################################################################################
 #
@@ -419,12 +431,16 @@ RUN apt-get update && \
     wget https://github.com/ketteq-neon/postgres-exts/archive/e0bd1a9d9313d7120c1b9c7bb15c48c0dede4c4e.tar.gz -O kq_imcx.tar.gz && \
     echo "dc93a97ff32d152d32737ba7e196d9687041cda15e58ab31344c2f2de8855336 kq_imcx.tar.gz" | sha256sum --check && \
     mkdir kq_imcx-src && cd kq_imcx-src && tar xvzf ../kq_imcx.tar.gz --strip-components=1 -C . && \
-    mkdir build && \
-    cd build && \
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /before.txt &&\
+    mkdir build && cd build && \
     cmake -DCMAKE_BUILD_TYPE=Release .. && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/kq_imcx.control
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/kq_imcx.control && \
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /after.txt &&\
+    mkdir -p /extensions/kq_imcx && cp /usr/local/pgsql/share/extension/kq_imcx.control /extensions/kq_imcx && \
+    sort -o /before.txt /before.txt && sort -o /after.txt /after.txt && \
+    comm -13 /before.txt /after.txt | tar --directory=/usr/local/pgsql --zstd -cf /extensions/kq_imcx.tar.zst -T -
 
 #########################################################################################
 #
@@ -535,10 +551,8 @@ FROM build-deps AS pg-embedding-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
 ENV PATH "/usr/local/pgsql/bin/:$PATH"
-# 2465f831ea1f8d49c1d74f8959adb7fc277d70cd made on 05/07/2023
-# There is no release tag yet
-RUN wget https://github.com/neondatabase/pg_embedding/archive/2465f831ea1f8d49c1d74f8959adb7fc277d70cd.tar.gz -O pg_embedding.tar.gz && \
-    echo "047af2b1f664a1e6e37867bd4eeaf5934fa27d6ba3d6c4461efa388ddf7cd1d5 pg_embedding.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/neondatabase/pg_embedding/archive/refs/tags/0.3.5.tar.gz -O pg_embedding.tar.gz && \
+    echo "0e95b27b8b6196e2cf0a0c9ec143fe2219b82e54c5bb4ee064e76398cbe69ae9 pg_embedding.tar.gz" | sha256sum --check && \
     mkdir pg_embedding-src && cd pg_embedding-src && tar xvzf ../pg_embedding.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -553,16 +567,17 @@ RUN wget https://github.com/neondatabase/pg_embedding/archive/2465f831ea1f8d49c1
 FROM build-deps AS pg-anon-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-# Kaniko doesn't allow to do `${from#/usr/local/pgsql/}`, so we use `${from:17}` instead
 ENV PATH "/usr/local/pgsql/bin/:$PATH"
 RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgresql_anonymizer-1.1.0.tar.gz -O pg_anon.tar.gz && \
     echo "08b09d2ff9b962f96c60db7e6f8e79cf7253eb8772516998fc35ece08633d3ad pg_anon.tar.gz" | sha256sum --check && \
     mkdir pg_anon-src && cd pg_anon-src && tar xvzf ../pg_anon.tar.gz --strip-components=1 -C . && \
-    find /usr/local/pgsql -type f | sort  > /before.txt && \
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /before.txt &&\
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/anon.control && \
-    find /usr/local/pgsql -type f | sort  > /after.txt && \
-    /bin/bash -c 'for from in $(comm -13 /before.txt /after.txt); do to=/extensions/anon/${from:17} && mkdir -p $(dirname ${to}) && cp -a ${from} ${to}; done'
+    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /after.txt &&\
+    mkdir -p /extensions/anon && cp /usr/local/pgsql/share/extension/anon.control /extensions/anon && \
+    sort -o /before.txt /before.txt && sort -o /after.txt /after.txt && \
+    comm -13 /before.txt /after.txt | tar --directory=/usr/local/pgsql --zstd -cf /extensions/anon.tar.zst -T -
 
 #########################################################################################
 #
@@ -754,16 +769,23 @@ 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 for new custom ones
-#
-# # Default extensions
-# COPY --from=postgres-cleanup-layer /usr/local/pgsql/share/extension /usr/local/pgsql/share/extension
-# COPY --from=postgres-cleanup-layer /usr/local/pgsql/lib             /usr/local/pgsql/lib
-# Custom extensions
-COPY --from=pg-anon-pg-build /extensions/anon/lib/ /extensions/anon/lib
-COPY --from=pg-anon-pg-build /extensions/anon/share/extension /extensions/anon/share/extension
+# 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
 
 #########################################################################################
 #

Makefile

@@ -108,6 +108,8 @@ postgres-%: postgres-configure-% \
 	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/contrib/pg_buffercache install
 	+@echo "Compiling pageinspect $*"
 	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/contrib/pageinspect install
+	+@echo "Compiling amcheck $*"
+	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/contrib/amcheck install
 
 .PHONY: postgres-clean-%
 postgres-clean-%:

compute_tools/Cargo.toml

@@ -34,3 +34,4 @@ utils.workspace = true
 workspace_hack.workspace = true
 toml_edit.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
+zstd = "0.12.4"

compute_tools/src/bin/compute_ctl.rs

@@ -30,7 +30,7 @@
 //!             -C 'postgresql://cloud_admin@localhost/postgres' \
 //!             -S /var/db/postgres/specs/current.json \
 //!             -b /usr/local/bin/postgres \
-//!             -r {"bucket": "my-bucket", "region": "eu-central-1", "endpoint": "http:://localhost:9000"} \
+//!             -r {"bucket": "neon-dev-extensions-eu-central-1", "region": "eu-central-1"}
 //! ```
 //!
 use std::collections::HashMap;
@@ -38,7 +38,7 @@ use std::fs::File;
 use std::panic;
 use std::path::Path;
 use std::process::exit;
-use std::sync::{mpsc, Arc, Condvar, Mutex, OnceLock};
+use std::sync::{mpsc, Arc, Condvar, Mutex, OnceLock, RwLock};
 use std::{thread, time::Duration};
 
 use anyhow::{Context, Result};
@@ -51,7 +51,6 @@ use compute_api::responses::ComputeStatus;
 
 use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec};
 use compute_tools::configurator::launch_configurator;
-use compute_tools::extension_server::launch_download_extensions;
 use compute_tools::extension_server::{get_pg_version, init_remote_storage};
 use compute_tools::http::api::launch_http_server;
 use compute_tools::logger::*;
@@ -59,43 +58,26 @@ use compute_tools::monitor::launch_monitor;
 use compute_tools::params::*;
 use compute_tools::spec::*;
 
-const BUILD_TAG_DEFAULT: &str = "local";
-const DEFAULT_REMOTE_EXT_CONFIG: &str = r#"{"bucket": "neon-dev-extensions", "region": "eu-central-1", "endpoint": null, "prefix": "5555"}"#;
+// this is an arbitrary build tag. Fine as a default / for testing purposes
+// in-case of not-set environment var
+const BUILD_TAG_DEFAULT: &str = "5670669815";
 
 fn main() -> Result<()> {
     init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
 
-    let build_tag = option_env!("BUILD_TAG").unwrap_or(BUILD_TAG_DEFAULT);
+    let build_tag = option_env!("BUILD_TAG")
+        .unwrap_or(BUILD_TAG_DEFAULT)
+        .to_string();
     info!("build_tag: {build_tag}");
 
     let matches = cli().get_matches();
     let pgbin_default = String::from("postgres");
     let pgbin = matches.get_one::<String>("pgbin").unwrap_or(&pgbin_default);
 
-    let remote_ext_config = matches
-        .get_one::<String>("remote-ext-config")
-        .map(|x| x.to_string());
-    // let remote_ext_config =
-    //     Some(remote_ext_config.unwrap_or(DEFAULT_REMOTE_EXT_CONFIG.to_string()));
-
+    let remote_ext_config = matches.get_one::<String>("remote-ext-config");
     let ext_remote_storage = remote_ext_config.map(|x| {
-        init_remote_storage(&x, build_tag)
-            .expect("cannot initialize remote extension storage from config")
+        init_remote_storage(x).expect("cannot initialize remote extension storage from config")
     });
-    // creds used to connect to remote extensions bucket
-    // let aws_creds = matches.get_one::<String>("awscreds");
-    // if let Some(aws_creds) = aws_creds {
-    //     // not sure if this is a bad idea?
-    //     let aws_creds_dict: serde_json::Value = serde_json::from_str(aws_creds)?;
-    //     std::env::set_var(
-    //         "AWS_ACCESS_KEY_ID",
-    //         aws_creds_dict["ID"].as_str().expect("config parse error"),
-    //     );
-    //     std::env::set_var(
-    //         "AWS_SECRET_ACCESS_KEY",
-    //         aws_creds_dict["key"].as_str().expect("config parse error"),
-    //     );
-    // }
 
     let http_port = *matches
         .get_one::<u16>("http-port")
@@ -214,7 +196,10 @@ fn main() -> Result<()> {
         state: Mutex::new(new_state),
         state_changed: Condvar::new(),
         ext_remote_storage,
-        available_extensions: OnceLock::new(),
+        ext_remote_paths: OnceLock::new(),
+        ext_download_progress: RwLock::new(HashMap::new()),
+        library_index: OnceLock::new(),
+        build_tag,
     };
     let compute = Arc::new(compute_node);
 
@@ -228,6 +213,13 @@ fn main() -> Result<()> {
     if !spec_set {
         // No spec provided, hang waiting for it.
         info!("no compute spec provided, waiting");
+
+        // TODO this can stall startups in the unlikely event that we bind
+        //      this compute node while it's busy prewarming. It's not too
+        //      bad because it's just 100ms and unlikely, but it's an
+        //      avoidable problem.
+        compute.prewarm_postgres()?;
+
         let mut state = compute.state.lock().unwrap();
         while state.status != ComputeStatus::ConfigurationPending {
             state = compute.state_changed.wait(state).unwrap();
@@ -258,12 +250,8 @@ fn main() -> Result<()> {
     drop(state);
 
     // Launch remaining service threads
-    let _monitor_handle = launch_monitor(&compute).expect("cannot launch compute monitor thread");
-    let _configurator_handle =
-        launch_configurator(&compute).expect("cannot launch configurator thread");
-
-    let _download_extensions_handle =
-        launch_download_extensions(&compute).expect("cannot launch download extensions thread");
+    let _monitor_handle = launch_monitor(&compute);
+    let _configurator_handle = launch_configurator(&compute);
 
     // Start Postgres
     let mut delay_exit = false;
@@ -403,12 +391,6 @@ fn cli() -> clap::Command {
                 .long("remote-ext-config")
                 .value_name("REMOTE_EXT_CONFIG"),
         )
-        .arg(
-            Arg::new("awscreds")
-                .short('k')
-                .long("awscreds")
-                .value_name("AWS_CREDENTIALS"),
-        )
 }
 
 #[test]

compute_tools/src/compute.rs

@@ -1,19 +1,22 @@
-use std::collections::HashSet;
+use std::collections::HashMap;
 use std::fs;
 use std::io::BufRead;
 use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
-use std::sync::{Condvar, Mutex, OnceLock};
+use std::sync::{Condvar, Mutex, OnceLock, RwLock};
 
 use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
 use futures::future::join_all;
+use futures::stream::FuturesUnordered;
+use futures::StreamExt;
 use postgres::{Client, NoTls};
+use regex::Regex;
 use tokio;
 use tokio_postgres;
-use tracing::{info, instrument, warn};
+use tracing::{error, info, instrument, warn};
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -21,10 +24,11 @@ use compute_api::responses::{ComputeMetrics, ComputeStatus};
 use compute_api::spec::{ComputeMode, ComputeSpec};
 use utils::measured_stream::MeasuredReader;
 
-use remote_storage::GenericRemoteStorage;
+use remote_storage::{GenericRemoteStorage, RemotePath};
 
 use crate::pg_helpers::*;
 use crate::spec::*;
+use crate::sync_sk::{check_if_synced, ping_safekeeper};
 use crate::{config, extension_server};
 
 /// Compute node info shared across several `compute_ctl` threads.
@@ -55,8 +59,22 @@ pub struct ComputeNode {
     pub state_changed: Condvar,
     ///  the S3 bucket that we search for extensions in
     pub ext_remote_storage: Option<GenericRemoteStorage>,
-    // cached lists of available extensions and libraries
-    pub available_extensions: OnceLock<HashSet<String>>,
+    // (key: extension name, value: path to extension archive in remote storage)
+    pub ext_remote_paths: OnceLock<HashMap<String, RemotePath>>,
+    // (key: library name, value: name of extension containing this library)
+    pub library_index: OnceLock<HashMap<String, String>>,
+    // key: ext_archive_name, value: started download time, download_completed?
+    pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
+    pub build_tag: String,
+}
+
+// store some metrics about download size that might impact startup time
+#[derive(Clone, Debug)]
+pub struct RemoteExtensionMetrics {
+    num_ext_downloaded: u64,
+    largest_ext_size: u64,
+    total_ext_download_size: u64,
+    prep_extensions_ms: u64,
 }
 
 #[derive(Clone, Debug)]
@@ -96,6 +114,7 @@ pub struct ParsedSpec {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
     pub pageserver_connstr: String,
+    pub safekeeper_connstrings: Vec<String>,
     pub storage_auth_token: Option<String>,
 }
 
@@ -113,6 +132,21 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
             .clone()
             .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
             .ok_or("pageserver connstr should be provided")?;
+        let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
+            if matches!(spec.mode, ComputeMode::Primary) {
+                spec.cluster
+                    .settings
+                    .find("neon.safekeepers")
+                    .ok_or("safekeeper connstrings should be provided")?
+                    .split(',')
+                    .map(|str| str.to_string())
+                    .collect()
+            } else {
+                vec![]
+            }
+        } else {
+            spec.safekeeper_connstrings.clone()
+        };
         let storage_auth_token = spec.storage_auth_token.clone();
         let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
             tenant_id
@@ -138,6 +172,7 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
         Ok(ParsedSpec {
             spec,
             pageserver_connstr,
+            safekeeper_connstrings,
             storage_auth_token,
             tenant_id,
             timeline_id,
@@ -319,6 +354,102 @@ impl ComputeNode {
         Ok(())
     }
 
+    pub async fn check_safekeepers_synced_async(
+        &self,
+        compute_state: &ComputeState,
+    ) -> Result<Option<Lsn>> {
+        // Construct a connection config for each safekeeper
+        let pspec: ParsedSpec = compute_state
+            .pspec
+            .as_ref()
+            .expect("spec must be set")
+            .clone();
+        let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
+        let sk_configs = sk_connstrs.into_iter().map(|connstr| {
+            // Format connstr
+            let id = connstr.clone();
+            let connstr = format!("postgresql://no_user@{}", connstr);
+            let options = format!(
+                "-c timeline_id={} tenant_id={}",
+                pspec.timeline_id, pspec.tenant_id
+            );
+
+            // Construct client
+            let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
+            config.options(&options);
+            if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
+                config.password(storage_auth_token);
+            }
+
+            (id, config)
+        });
+
+        // Create task set to query all safekeepers
+        let mut tasks = FuturesUnordered::new();
+        let quorum = sk_configs.len() / 2 + 1;
+        for (id, config) in sk_configs {
+            let timeout = tokio::time::Duration::from_millis(100);
+            let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
+            tasks.push(tokio::spawn(task));
+        }
+
+        // Get a quorum of responses or errors
+        let mut responses = Vec::new();
+        let mut join_errors = Vec::new();
+        let mut task_errors = Vec::new();
+        let mut timeout_errors = Vec::new();
+        while let Some(response) = tasks.next().await {
+            match response {
+                Ok(Ok(Ok(r))) => responses.push(r),
+                Ok(Ok(Err(e))) => task_errors.push(e),
+                Ok(Err(e)) => timeout_errors.push(e),
+                Err(e) => join_errors.push(e),
+            };
+            if responses.len() >= quorum {
+                break;
+            }
+            if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
+                break;
+            }
+        }
+
+        // In case of error, log and fail the check, but don't crash.
+        // We're playing it safe because these errors could be transient
+        // and we don't yet retry. Also being careful here allows us to
+        // be backwards compatible with safekeepers that don't have the
+        // TIMELINE_STATUS API yet.
+        if responses.len() < quorum {
+            error!(
+                "failed sync safekeepers check {:?} {:?} {:?}",
+                join_errors, task_errors, timeout_errors
+            );
+            return Ok(None);
+        }
+
+        Ok(check_if_synced(responses))
+    }
+
+    // Fast path for sync_safekeepers. If they're already synced we get the lsn
+    // in one roundtrip. If not, we should do a full sync_safekeepers.
+    pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
+        let start_time = Utc::now();
+
+        // Run actual work with new tokio runtime
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .expect("failed to create rt");
+        let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
+
+        // Record runtime
+        self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
+            .signed_duration_since(start_time)
+            .to_std()
+            .unwrap()
+            .as_millis() as u64;
+        result
+    }
+
     // Run `postgres` in a special mode with `--sync-safekeepers` argument
     // and return the reported LSN back to the caller.
     #[instrument(skip_all)]
@@ -389,10 +520,14 @@ impl ComputeNode {
         // cannot sync safekeepers.
         let lsn = match spec.mode {
             ComputeMode::Primary => {
-                info!("starting safekeepers syncing");
-                let lsn = self
-                    .sync_safekeepers(pspec.storage_auth_token.clone())
-                    .with_context(|| "failed to sync safekeepers")?;
+                info!("checking if safekeepers are synced");
+                let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
+                    lsn
+                } else {
+                    info!("starting safekeepers syncing");
+                    self.sync_safekeepers(pspec.storage_auth_token.clone())
+                        .with_context(|| "failed to sync safekeepers")?
+                };
                 info!("safekeepers synced at LSN {}", lsn);
                 lsn
             }
@@ -430,6 +565,50 @@ impl ComputeNode {
         Ok(())
     }
 
+    /// Start and stop a postgres process to warm up the VM for startup.
+    pub fn prewarm_postgres(&self) -> Result<()> {
+        info!("prewarming");
+
+        // Create pgdata
+        let pgdata = &format!("{}.warmup", self.pgdata);
+        create_pgdata(pgdata)?;
+
+        // Run initdb to completion
+        info!("running initdb");
+        let initdb_bin = Path::new(&self.pgbin).parent().unwrap().join("initdb");
+        Command::new(initdb_bin)
+            .args(["-D", pgdata])
+            .output()
+            .expect("cannot start initdb process");
+
+        // Write conf
+        use std::io::Write;
+        let conf_path = Path::new(pgdata).join("postgresql.conf");
+        let mut file = std::fs::File::create(conf_path)?;
+        writeln!(file, "shared_buffers=65536")?;
+        writeln!(file, "port=51055")?; // Nobody should be connecting
+        writeln!(file, "shared_preload_libraries = 'neon'")?;
+
+        // Start postgres
+        info!("starting postgres");
+        let mut pg = Command::new(&self.pgbin)
+            .args(["-D", pgdata])
+            .spawn()
+            .expect("cannot start postgres process");
+
+        // Stop it when it's ready
+        info!("waiting for postgres");
+        wait_for_postgres(&mut pg, Path::new(pgdata))?;
+        pg.kill()?;
+        info!("sent kill signal");
+        pg.wait()?;
+        info!("done prewarming");
+
+        // clean up
+        let _ok = fs::remove_dir_all(pgdata);
+        Ok(())
+    }
+
     /// Start Postgres as a child process and manage DBs/roles.
     /// After that this will hang waiting on the postmaster process to exit.
     #[instrument(skip_all)]
@@ -569,7 +748,7 @@ impl ComputeNode {
         // remote shared_preload_libraries before postgres start (if any)
         {
             let library_load_start_time = Utc::now();
-            self.prepare_preload_libraries(&compute_state)?;
+            let remote_ext_metrics = self.prepare_preload_libraries(&compute_state)?;
 
             let library_load_time = Utc::now()
                 .signed_duration_since(library_load_start_time)
@@ -577,11 +756,16 @@ impl ComputeNode {
                 .unwrap()
                 .as_millis() as u64;
             let mut state = self.state.lock().unwrap();
-            state.metrics.load_libraries_ms = library_load_time;
+            state.metrics.load_ext_ms = library_load_time;
+            state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
+            state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
+            state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
+            state.metrics.prep_extensions_ms = remote_ext_metrics.prep_extensions_ms;
             info!(
                 "Loading shared_preload_libraries took {:?}ms",
                 library_load_time
             );
+            info!("{:?}", remote_ext_metrics);
         }
 
         self.prepare_pgdata(&compute_state, extension_server_port)?;
@@ -735,46 +919,133 @@ LIMIT 100",
 
     // If remote extension storage is configured,
     // download extension control files
-    #[tokio::main]
     pub async fn prepare_external_extensions(&self, compute_state: &ComputeState) -> Result<()> {
         if let Some(ref ext_remote_storage) = self.ext_remote_storage {
             let pspec = compute_state.pspec.as_ref().expect("spec must be set");
             let spec = &pspec.spec;
-            let custom_ext_prefixes = spec.custom_extensions.clone().unwrap_or(Vec::new());
-            info!("custom_ext_prefixes: {:?}", &custom_ext_prefixes);
-            let available_extensions = extension_server::get_available_extensions(
+            let custom_ext = spec.custom_extensions.clone().unwrap_or(Vec::new());
+            info!("custom extensions: {:?}", &custom_ext);
+            let (ext_remote_paths, library_index) = extension_server::get_available_extensions(
                 ext_remote_storage,
                 &self.pgbin,
                 &self.pgversion,
-                &custom_ext_prefixes,
+                &custom_ext,
+                &self.build_tag,
             )
             .await?;
-            self.available_extensions
-                .set(available_extensions)
-                .expect("available_extensions.set error");
+            self.ext_remote_paths
+                .set(ext_remote_paths)
+                .expect("this is the only time we set ext_remote_paths");
+            self.library_index
+                .set(library_index)
+                .expect("this is the only time we set library_index");
         }
         Ok(())
     }
 
-    pub async fn download_extension(&self, ext_name: &str) -> Result<()> {
+    // download an archive, unzip and place files in correct locations
+    pub async fn download_extension(&self, ext_name: &str, is_library: bool) -> Result<u64> {
         match &self.ext_remote_storage {
             None => anyhow::bail!("No remote extension storage"),
             Some(remote_storage) => {
-                extension_server::download_extension(
-                    ext_name,
+                let mut real_ext_name = ext_name.to_string();
+                if is_library {
+                    // sometimes library names might have a suffix like
+                    // library.so or library.so.3. We strip this off
+                    // because library_index is based on the name without the file extension
+                    let strip_lib_suffix = Regex::new(r"\.so.*").unwrap();
+                    let lib_raw_name = strip_lib_suffix.replace(&real_ext_name, "").to_string();
+                    real_ext_name = self
+                        .library_index
+                        .get()
+                        .expect("must have already downloaded the library_index")[&lib_raw_name]
+                        .clone();
+                }
+
+                let ext_path = &self
+                    .ext_remote_paths
+                    .get()
+                    .expect("error accessing ext_remote_paths")[&real_ext_name];
+                let ext_archive_name = ext_path.object_name().expect("bad path");
+
+                let mut first_try = false;
+                if !self
+                    .ext_download_progress
+                    .read()
+                    .expect("lock err")
+                    .contains_key(ext_archive_name)
+                {
+                    self.ext_download_progress
+                        .write()
+                        .expect("lock err")
+                        .insert(ext_archive_name.to_string(), (Utc::now(), false));
+                    first_try = true;
+                }
+                let (download_start, download_completed) =
+                    self.ext_download_progress.read().expect("lock err")[ext_archive_name];
+                let start_time_delta = Utc::now()
+                    .signed_duration_since(download_start)
+                    .to_std()
+                    .unwrap()
+                    .as_millis() as u64;
+
+                // how long to wait for extension download if it was started by another process
+                const HANG_TIMEOUT: u64 = 3000; // milliseconds
+
+                if download_completed {
+                    info!("extension already downloaded, skipping re-download");
+                    return Ok(0);
+                } else if start_time_delta < HANG_TIMEOUT && !first_try {
+                    info!("download {ext_archive_name} already started by another process, hanging untill completion or timeout");
+                    let mut interval =
+                        tokio::time::interval(tokio::time::Duration::from_millis(500));
+                    loop {
+                        info!("waiting for download");
+                        interval.tick().await;
+                        let (_, download_completed_now) =
+                            self.ext_download_progress.read().expect("lock")[ext_archive_name];
+                        if download_completed_now {
+                            info!("download finished by whoever else downloaded it");
+                            return Ok(0);
+                        }
+                    }
+                    // NOTE: the above loop will get terminated
+                    // based on the timeout of the download function
+                }
+
+                // if extension hasn't been downloaded before or the previous
+                // attempt to download was at least HANG_TIMEOUT ms ago
+                // then we try to download it here
+                info!("downloading new extension {ext_archive_name}");
+
+                let download_size = extension_server::download_extension(
+                    &real_ext_name,
+                    ext_path,
                     remote_storage,
                     &self.pgbin,
-                    &self.pgversion,
                 )
-                .await
+                .await;
+                self.ext_download_progress
+                    .write()
+                    .expect("bad lock")
+                    .insert(ext_archive_name.to_string(), (download_start, true));
+                download_size
             }
         }
     }
 
     #[tokio::main]
-    pub async fn prepare_preload_libraries(&self, compute_state: &ComputeState) -> Result<()> {
+    pub async fn prepare_preload_libraries(
+        &self,
+        compute_state: &ComputeState,
+    ) -> Result<RemoteExtensionMetrics> {
         if self.ext_remote_storage.is_none() {
-            return Ok(());
+            return Ok(RemoteExtensionMetrics {
+                num_ext_downloaded: 0,
+                largest_ext_size: 0,
+                total_ext_download_size: 0,
+                prep_extensions_ms: 0,
+            });
         }
         let pspec = compute_state.pspec.as_ref().expect("spec must be set");
         let spec = &pspec.spec;
@@ -809,15 +1080,36 @@ LIMIT 100",
             libs_vec.extend(preload_libs_vec);
         }
 
+        info!("Download ext_index.json, find the extension paths");
+        let prep_ext_start_time = Utc::now();
+        self.prepare_external_extensions(compute_state).await?;
+        let prep_ext_time_delta = Utc::now()
+            .signed_duration_since(prep_ext_start_time)
+            .to_std()
+            .unwrap()
+            .as_millis() as u64;
+        info!("Prepare extensions took {prep_ext_time_delta}ms");
+
         info!("Downloading to shared preload libraries: {:?}", &libs_vec);
         let mut download_tasks = Vec::new();
         for library in &libs_vec {
-            download_tasks.push(self.download_extension(library));
+            download_tasks.push(self.download_extension(library, true));
         }
         let results = join_all(download_tasks).await;
+
+        let mut remote_ext_metrics = RemoteExtensionMetrics {
+            num_ext_downloaded: 0,
+            largest_ext_size: 0,
+            total_ext_download_size: 0,
+            prep_extensions_ms: prep_ext_time_delta,
+        };
         for result in results {
-            result?; // propogate any errors
+            let download_size = result?;
+            remote_ext_metrics.num_ext_downloaded += 1;
+            remote_ext_metrics.largest_ext_size =
+                std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
+            remote_ext_metrics.total_ext_download_size += download_size;
         }
-        Ok(())
+        Ok(remote_ext_metrics)
     }
 }

compute_tools/src/configurator.rs

@@ -1,7 +1,6 @@
 use std::sync::Arc;
 use std::thread;
 
-use anyhow::Result;
 use tracing::{error, info, instrument};
 
 use compute_api::responses::ComputeStatus;
@@ -42,9 +41,7 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
     }
 }
 
-pub fn launch_configurator(
-    compute: &Arc<ComputeNode>,
-) -> Result<thread::JoinHandle<()>, std::io::Error> {
+pub fn launch_configurator(compute: &Arc<ComputeNode>) -> thread::JoinHandle<()> {
     let compute = Arc::clone(compute);
 
     thread::Builder::new()
@@ -53,4 +50,5 @@ pub fn launch_configurator(
             configurator_main_loop(&compute);
             info!("configurator thread is exited");
         })
+        .expect("cannot launch configurator thread")
 }

compute_tools/src/extension_server.rs

@@ -1,35 +1,91 @@
 // Download extension files from the extension store
-// and put them in the right place in the postgres directory
+// and put them in the right place in the postgres directory (share / lib)
 /*
 The layout of the S3 bucket is as follows:
+5615610098 // this is an extension build number
+├── v14
+│   ├── extensions
+│   │   ├── anon.tar.zst
+│   │   └── embedding.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   ├── anon.tar.zst
+    │   └── embedding.tar.zst
+    └── ext_index.json
+5615261079
+├── v14
+│   ├── extensions
+│   │   └── anon.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   └── anon.tar.zst
+    └── ext_index.json
+5623261088
+├── v14
+│   ├── extensions
+│   │   └── embedding.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   └── embedding.tar.zst
+    └── ext_index.json
 
-v14/ext_index.json
-    -- this contains information necessary to create control files
-v14/extensions/test_ext1.tar.gz
-    -- this contains the library files and sql files necessary to create this extension
-v14/extensions/custom_ext1.tar.gz
+Note that build number cannot be part of prefix because we might need extensions
+from other build numbers.
 
-The difference between a private and public extensions is determined by who can
-load the extension this is specified in ext_index.json
+ext_index.json stores the control files and location of extension archives
+It also stores a list of public extensions and a library_index
 
-Speicially, ext_index.json has a list of public extensions, and a list of
-extensions enabled for specific tenant-ids.
+We don't need to duplicate extension.tar.zst files.
+We only need to upload a new one if it is updated.
+(Although currently we just upload every time anyways, hopefully will change
+this sometime)
+
+*access* is controlled by spec
+
+More specifically, here is an example ext_index.json
+{
+    "public_extensions": [
+        "anon",
+        "pg_buffercache"
+    ],
+    "library_index": {
+        "anon": "anon",
+        "pg_buffercache": "pg_buffercache"
+    },
+    "extension_data": {
+        "pg_buffercache": {
+            "control_data": {
+                "pg_buffercache.control": "# pg_buffercache extension \ncomment = 'examine the shared buffer cache' \ndefault_version = '1.3' \nmodule_pathname = '$libdir/pg_buffercache' \nrelocatable = true \ntrusted=true"
+            },
+            "archive_path": "5670669815/v14/extensions/pg_buffercache.tar.zst"
+        },
+        "anon": {
+            "control_data": {
+                "anon.control": "# PostgreSQL Anonymizer (anon) extension \ncomment = 'Data anonymization tools' \ndefault_version = '1.1.0' \ndirectory='extension/anon' \nrelocatable = false \nrequires = 'pgcrypto' \nsuperuser = false \nmodule_pathname = '$libdir/anon' \ntrusted = true \n"
+            },
+            "archive_path": "5670669815/v14/extensions/anon.tar.zst"
+        }
+    }
+}
 */
-use crate::compute::ComputeNode;
 use anyhow::Context;
 use anyhow::{self, Result};
-use flate2::read::GzDecoder;
+use futures::future::join_all;
 use remote_storage::*;
-use serde_json::{self, Value};
-use std::collections::HashSet;
+use serde_json;
+use std::collections::HashMap;
+use std::io::Read;
 use std::num::{NonZeroU32, NonZeroUsize};
 use std::path::Path;
 use std::str;
-use std::sync::Arc;
-use std::thread;
 use tar::Archive;
 use tokio::io::AsyncReadExt;
 use tracing::info;
+use tracing::log::warn;
+use zstd::stream::read::Decoder;
 
 fn get_pg_config(argument: &str, pgbin: &str) -> String {
     // gives the result of `pg_config [argument]`
@@ -61,154 +117,152 @@ pub fn get_pg_version(pgbin: &str) -> String {
     panic!("Unsuported postgres version {human_version}");
 }
 
-// download extension control files
-// if custom_ext_prefixes is provided - search also in custom extension paths
+// download control files for enabled_extensions
+// return Hashmaps converting library names to extension names (library_index)
+// and specifying the remote path to the archive for each extension name
 pub async fn get_available_extensions(
     remote_storage: &GenericRemoteStorage,
     pgbin: &str,
     pg_version: &str,
-    custom_ext_prefixes: &[String],
-) -> Result<HashSet<String>> {
+    custom_extensions: &[String],
+    build_tag: &str,
+) -> Result<(HashMap<String, RemotePath>, HashMap<String, String>)> {
     let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
-    let index_path = pg_version.to_owned() + "/ext_index.json";
+    let index_path = format!("{build_tag}/{pg_version}/ext_index.json");
     let index_path = RemotePath::new(Path::new(&index_path)).context("error forming path")?;
-    info!("download ext_index.json: {:?}", &index_path);
+    info!("download ext_index.json from: {:?}", &index_path);
 
-    // TODO: potential optimization: cache ext_index.json
     let mut download = remote_storage.download(&index_path).await?;
-    let mut write_data_buffer = Vec::new();
+    let mut ext_idx_buffer = Vec::new();
     download
         .download_stream
-        .read_to_end(&mut write_data_buffer)
+        .read_to_end(&mut ext_idx_buffer)
         .await?;
-    let ext_index_str = match str::from_utf8(&write_data_buffer) {
-        Ok(v) => v,
-        Err(e) => panic!("Invalid UTF-8 sequence: {}", e),
-    };
+    info!("ext_index downloaded");
 
-    let ext_index_full: Value = serde_json::from_str(ext_index_str)?;
-    let ext_index_full = ext_index_full.as_object().context("error parsing json")?;
-    let control_data = ext_index_full["control_data"]
-        .as_object()
-        .context("json parse error")?;
-    let enabled_extensions = ext_index_full["enabled_extensions"]
-        .as_object()
-        .context("json parse error")?;
-    info!("{:?}", control_data.clone());
-    info!("{:?}", enabled_extensions.clone());
+    #[derive(Debug, serde::Deserialize)]
+    struct Index {
+        public_extensions: Vec<String>,
+        library_index: HashMap<String, String>,
+        extension_data: HashMap<String, ExtensionData>,
+    }
 
-    let mut prefixes = vec!["public".to_string()];
-    prefixes.extend(custom_ext_prefixes.to_owned());
-    info!("{:?}", &prefixes);
-    let mut all_extensions = HashSet::new();
-    for prefix in prefixes {
-        let prefix_extensions = match enabled_extensions.get(&prefix) {
-            Some(Value::Array(ext_name)) => ext_name,
-            _ => {
-                info!("prefix {} has no extensions", prefix);
-                continue;
-            }
-        };
-        info!("{:?}", prefix_extensions);
-        for ext_name in prefix_extensions {
-            all_extensions.insert(ext_name.as_str().context("json parse error")?.to_string());
+    #[derive(Debug, serde::Deserialize)]
+    struct ExtensionData {
+        control_data: HashMap<String, String>,
+        archive_path: String,
+    }
+
+    let ext_index_full = serde_json::from_slice::<Index>(&ext_idx_buffer)?;
+    let mut enabled_extensions = ext_index_full.public_extensions;
+    enabled_extensions.extend_from_slice(custom_extensions);
+    let library_index = ext_index_full.library_index;
+    let all_extension_data = ext_index_full.extension_data;
+    info!("library_index: {:?}", library_index);
+
+    info!("enabled_extensions: {:?}", enabled_extensions);
+    let mut ext_remote_paths = HashMap::new();
+    let mut file_create_tasks = Vec::new();
+    for extension in enabled_extensions {
+        let ext_data = &all_extension_data[&extension];
+        for (control_file, control_contents) in &ext_data.control_data {
+            let extension_name = control_file
+                .strip_suffix(".control")
+                .expect("control files must end in .control");
+            ext_remote_paths.insert(
+                extension_name.to_string(),
+                RemotePath::from_string(&ext_data.archive_path)?,
+            );
+            let control_path = local_sharedir.join(control_file);
+            info!("writing file {:?}{:?}", control_path, control_contents);
+            file_create_tasks.push(tokio::fs::write(control_path, control_contents));
         }
     }
-
-    for prefix in &all_extensions {
-        let control_contents = control_data[prefix].as_str().context("json parse error")?;
-        let control_path = local_sharedir.join(prefix.to_owned() + ".control");
-
-        info!("WRITING FILE {:?}{:?}", control_path, control_contents);
-        std::fs::write(control_path, control_contents)?;
+    let results = join_all(file_create_tasks).await;
+    for result in results {
+        result?;
     }
-
-    Ok(all_extensions.into_iter().collect())
+    info!("ext_remote_paths {:?}", ext_remote_paths);
+    Ok((ext_remote_paths, library_index))
 }
 
-// download all sqlfiles (and possibly data files) for a given extension name
+// download the archive for a given extension,
+// unzip it, and place files in the appropriate locations (share/lib)
 pub async fn download_extension(
     ext_name: &str,
+    ext_path: &RemotePath,
     remote_storage: &GenericRemoteStorage,
     pgbin: &str,
-    pg_version: &str,
-) -> Result<()> {
-    // TODO: potential optimization: only download the extension if it doesn't exist
-    // problem: how would we tell if it exists?
-    let ext_name = ext_name.replace(".so", "");
-    let ext_name_targz = ext_name.to_owned() + ".tar.gz";
-    if Path::new(&ext_name_targz).exists() {
-        info!("extension {:?} already exists", ext_name_targz);
-        return Ok(());
-    }
-    let ext_path = RemotePath::new(
-        &Path::new(pg_version)
-            .join("extensions")
-            .join(ext_name_targz.clone()),
-    )?;
-    info!(
-        "Start downloading extension {:?} from {:?}",
-        ext_name, ext_path
-    );
-    let mut download = remote_storage.download(&ext_path).await?;
-    let mut write_data_buffer = Vec::new();
+) -> Result<u64> {
+    info!("Download extension {:?} from {:?}", ext_name, ext_path);
+    let mut download = remote_storage.download(ext_path).await?;
+    let mut download_buffer = Vec::new();
     download
         .download_stream
-        .read_to_end(&mut write_data_buffer)
+        .read_to_end(&mut download_buffer)
         .await?;
-    let unzip_dest = pgbin.strip_suffix("/bin/postgres").expect("bad pgbin");
-    let tar = GzDecoder::new(write_data_buffer.as_slice());
-    let mut archive = Archive::new(tar);
-    archive.unpack(unzip_dest)?;
+    let download_size = download_buffer.len() as u64;
+    // it's unclear whether it is more performant to decompress into memory or not
+    // TODO: decompressing into memory can be avoided
+    let mut decoder = Decoder::new(download_buffer.as_slice())?;
+    let mut decompress_buffer = Vec::new();
+    decoder.read_to_end(&mut decompress_buffer)?;
+    let mut archive = Archive::new(decompress_buffer.as_slice());
+    let unzip_dest = pgbin
+        .strip_suffix("/bin/postgres")
+        .expect("bad pgbin")
+        .to_string()
+        + "/download_extensions";
+    archive.unpack(&unzip_dest)?;
     info!("Download + unzip {:?} completed successfully", &ext_path);
 
-    let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
-    let zip_sharedir = format!("{unzip_dest}/extensions/{ext_name}/share/extension");
-    info!("mv {zip_sharedir:?}/* {local_sharedir:?}");
-    for file in std::fs::read_dir(zip_sharedir)? {
-        let old_file = file?.path();
-        let new_file =
-            Path::new(&local_sharedir).join(old_file.file_name().context("error parsing file")?);
-        std::fs::rename(old_file, new_file)?;
+    let sharedir_paths = (
+        unzip_dest.to_string() + "/share/extension",
+        Path::new(&get_pg_config("--sharedir", pgbin)).join("extension"),
+    );
+    let libdir_paths = (
+        unzip_dest.to_string() + "/lib",
+        Path::new(&get_pg_config("--libdir", pgbin)).join("postgresql"),
+    );
+    // move contents of the libdir / sharedir in unzipped archive to the correct local paths
+    for paths in [sharedir_paths, libdir_paths] {
+        let (zip_dir, real_dir) = paths;
+        info!("mv {zip_dir:?}/*  {real_dir:?}");
+        for file in std::fs::read_dir(zip_dir)? {
+            let old_file = file?.path();
+            let new_file =
+                Path::new(&real_dir).join(old_file.file_name().context("error parsing file")?);
+            info!("moving {old_file:?} to {new_file:?}");
+
+            // extension download failed: Directory not empty (os error 39)
+            match std::fs::rename(old_file, new_file) {
+                Ok(()) => info!("move succeeded"),
+                Err(e) => {
+                    warn!("move failed, probably because the extension already exists: {e}")
+                }
+            }
+        }
     }
-    let local_libdir = Path::new(&get_pg_config("--libdir", pgbin)).join("postgresql");
-    let zip_libdir = format!("{unzip_dest}/extensions/{ext_name}/lib");
-    info!("mv {zip_libdir:?}/* {local_libdir:?}");
-    for file in std::fs::read_dir(zip_libdir)? {
-        let old_file = file?.path();
-        let new_file =
-            Path::new(&local_libdir).join(old_file.file_name().context("error parsing file")?);
-        std::fs::rename(old_file, new_file)?;
-    }
-    Ok(())
+    info!("done moving extension {ext_name}");
+    Ok(download_size)
 }
 
-// This function initializes the necessary structs to use remmote storage (should be fairly cheap)
-pub fn init_remote_storage(
-    remote_ext_config: &str,
-    default_prefix: &str,
-) -> anyhow::Result<GenericRemoteStorage> {
-    let remote_ext_config: serde_json::Value = serde_json::from_str(remote_ext_config)?;
+// This function initializes the necessary structs to use remote storage
+pub fn init_remote_storage(remote_ext_config: &str) -> anyhow::Result<GenericRemoteStorage> {
+    #[derive(Debug, serde::Deserialize)]
+    struct RemoteExtJson {
+        bucket: String,
+        region: String,
+        endpoint: Option<String>,
+        prefix: Option<String>,
+    }
+    let remote_ext_json = serde_json::from_str::<RemoteExtJson>(remote_ext_config)?;
 
-    let remote_ext_bucket = remote_ext_config["bucket"]
-        .as_str()
-        .context("config parse error")?;
-    let remote_ext_region = remote_ext_config["region"]
-        .as_str()
-        .context("config parse error")?;
-    let remote_ext_endpoint = remote_ext_config["endpoint"].as_str();
-    let remote_ext_prefix = remote_ext_config["prefix"]
-        .as_str()
-        .unwrap_or(default_prefix)
-        .to_string();
-
-    // TODO: potentially allow modification of other parameters
-    // however, default values should be fine for now
     let config = S3Config {
-        bucket_name: remote_ext_bucket.to_string(),
-        bucket_region: remote_ext_region.to_string(),
-        prefix_in_bucket: Some(remote_ext_prefix),
-        endpoint: remote_ext_endpoint.map(|x| x.to_string()),
+        bucket_name: remote_ext_json.bucket,
+        bucket_region: remote_ext_json.region,
+        prefix_in_bucket: remote_ext_json.prefix,
+        endpoint: remote_ext_json.endpoint,
         concurrency_limit: NonZeroUsize::new(100).expect("100 != 0"),
         max_keys_per_list_response: None,
     };
@@ -219,19 +273,3 @@ pub fn init_remote_storage(
     };
     GenericRemoteStorage::from_config(&config)
 }
-
-pub fn launch_download_extensions(
-    compute: &Arc<ComputeNode>,
-) -> Result<thread::JoinHandle<()>, std::io::Error> {
-    let compute = Arc::clone(compute);
-    thread::Builder::new()
-        .name("download-extensions".into())
-        .spawn(move || {
-            info!("start download_extension_files");
-            let compute_state = compute.state.lock().expect("error unlocking compute.state");
-            compute
-                .prepare_external_extensions(&compute_state)
-                .expect("error preparing extensions");
-            info!("download_extension_files done, exiting thread");
-        })
-}

compute_tools/src/http/api.rs

@@ -125,13 +125,23 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
         (&Method::POST, route) if route.starts_with("/extension_server/") => {
             info!("serving {:?} POST request", route);
             info!("req.uri {:?}", req.uri());
-            let filename = route.split('/').last().unwrap().to_string();
-            info!(
-                "serving /extension_server POST request, filename: {:?}",
-                &filename
-            );
 
-            match compute.download_extension(&filename).await {
+            let mut is_library = false;
+            if let Some(params) = req.uri().query() {
+                info!("serving {:?} POST request with params: {}", route, params);
+                if params == "is_library=true" {
+                    is_library = true;
+                } else {
+                    let mut resp = Response::new(Body::from("Wrong request parameters"));
+                    *resp.status_mut() = StatusCode::BAD_REQUEST;
+                    return resp;
+                }
+            }
+
+            let filename = route.split('/').last().unwrap().to_string();
+            info!("serving /extension_server POST request, filename: {filename:?} is_library: {is_library}");
+
+            match compute.download_extension(&filename, is_library).await {
                 Ok(_) => Response::new(Body::from("OK")),
                 Err(e) => {
                     error!("extension download failed: {}", e);

compute_tools/src/lib.rs

@@ -14,3 +14,4 @@ pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
 pub mod spec;
+pub mod sync_sk;

compute_tools/src/monitor.rs

@@ -1,7 +1,6 @@
 use std::sync::Arc;
 use std::{thread, time};
 
-use anyhow::Result;
 use chrono::{DateTime, Utc};
 use postgres::{Client, NoTls};
 use tracing::{debug, info};
@@ -105,10 +104,11 @@ fn watch_compute_activity(compute: &ComputeNode) {
 }
 
 /// Launch a separate compute monitor thread and return its `JoinHandle`.
-pub fn launch_monitor(state: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>, std::io::Error> {
+pub fn launch_monitor(state: &Arc<ComputeNode>) -> thread::JoinHandle<()> {
     let state = Arc::clone(state);
 
     thread::Builder::new()
         .name("compute-monitor".into())
         .spawn(move || watch_compute_activity(&state))
+        .expect("cannot launch compute monitor thread")
 }

compute_tools/src/pg_helpers.rs

@@ -19,7 +19,7 @@ const POSTGRES_WAIT_TIMEOUT: Duration = Duration::from_millis(60 * 1000); // mil
 /// Escape a string for including it in a SQL literal. Wrapping the result
 /// with `E'{}'` or `'{}'` is not required, as it returns a ready-to-use
 /// SQL string literal, e.g. `'db'''` or `E'db\\'`.
-/// See https://github.com/postgres/postgres/blob/da98d005cdbcd45af563d0c4ac86d0e9772cd15f/src/backend/utils/adt/quote.c#L47
+/// See <https://github.com/postgres/postgres/blob/da98d005cdbcd45af563d0c4ac86d0e9772cd15f/src/backend/utils/adt/quote.c#L47>
 /// for the original implementation.
 pub fn escape_literal(s: &str) -> String {
     let res = s.replace('\'', "''").replace('\\', "\\\\");

compute_tools/src/sync_sk.rs

@@ -0,0 +1,98 @@
+// Utils for running sync_safekeepers
+use anyhow::Result;
+use tracing::info;
+use utils::lsn::Lsn;
+
+#[derive(Copy, Clone, Debug)]
+pub enum TimelineStatusResponse {
+    NotFound,
+    Ok(TimelineStatusOkResponse),
+}
+
+#[derive(Copy, Clone, Debug)]
+pub struct TimelineStatusOkResponse {
+    flush_lsn: Lsn,
+    commit_lsn: Lsn,
+}
+
+/// Get a safekeeper's metadata for our timeline. The id is only used for logging
+pub async fn ping_safekeeper(
+    id: String,
+    config: tokio_postgres::Config,
+) -> Result<TimelineStatusResponse> {
+    // TODO add retries
+
+    // Connect
+    info!("connecting to {}", id);
+    let (client, conn) = config.connect(tokio_postgres::NoTls).await?;
+    tokio::spawn(async move {
+        if let Err(e) = conn.await {
+            eprintln!("connection error: {}", e);
+        }
+    });
+
+    // Query
+    info!("querying {}", id);
+    let result = client.simple_query("TIMELINE_STATUS").await?;
+
+    // Parse result
+    info!("done with {}", id);
+    if let postgres::SimpleQueryMessage::Row(row) = &result[0] {
+        use std::str::FromStr;
+        let response = TimelineStatusResponse::Ok(TimelineStatusOkResponse {
+            flush_lsn: Lsn::from_str(row.get("flush_lsn").unwrap())?,
+            commit_lsn: Lsn::from_str(row.get("commit_lsn").unwrap())?,
+        });
+        Ok(response)
+    } else {
+        // Timeline doesn't exist
+        Ok(TimelineStatusResponse::NotFound)
+    }
+}
+
+/// Given a quorum of responses, check if safekeepers are synced at some Lsn
+pub fn check_if_synced(responses: Vec<TimelineStatusResponse>) -> Option<Lsn> {
+    // Check if all responses are ok
+    let ok_responses: Vec<TimelineStatusOkResponse> = responses
+        .iter()
+        .filter_map(|r| match r {
+            TimelineStatusResponse::Ok(ok_response) => Some(ok_response),
+            _ => None,
+        })
+        .cloned()
+        .collect();
+    if ok_responses.len() < responses.len() {
+        info!(
+            "not synced. Only {} out of {} know about this timeline",
+            ok_responses.len(),
+            responses.len()
+        );
+        return None;
+    }
+
+    // Get the min and the max of everything
+    let commit: Vec<Lsn> = ok_responses.iter().map(|r| r.commit_lsn).collect();
+    let flush: Vec<Lsn> = ok_responses.iter().map(|r| r.flush_lsn).collect();
+    let commit_max = commit.iter().max().unwrap();
+    let commit_min = commit.iter().min().unwrap();
+    let flush_max = flush.iter().max().unwrap();
+    let flush_min = flush.iter().min().unwrap();
+
+    // Check that all values are equal
+    if commit_min != commit_max {
+        info!("not synced. {:?} {:?}", commit_min, commit_max);
+        return None;
+    }
+    if flush_min != flush_max {
+        info!("not synced. {:?} {:?}", flush_min, flush_max);
+        return None;
+    }
+
+    // Check that commit == flush
+    if commit_max != flush_max {
+        info!("not synced. {:?} {:?}", commit_max, flush_max);
+        return None;
+    }
+
+    Some(*commit_max)
+}

control_plane/src/background_process.rs

@@ -10,7 +10,7 @@
 //! (non-Neon binaries don't necessarily follow our pidfile conventions).
 //! The pid stored in the file is later used to stop the service.
 //!
-//! See [`lock_file`] module for more info.
+//! See the [`lock_file`](utils::lock_file) module for more info.
 
 use std::ffi::OsStr;
 use std::io::Write;

control_plane/src/broker.rs

@@ -2,8 +2,9 @@
 //!
 //! In the local test environment, the data for each safekeeper is stored in
 //!
+//! ```text
 //!   .neon/safekeepers/<safekeeper id>
-//!
+//! ```
 use anyhow::Context;
 
 use std::path::PathBuf;

control_plane/src/endpoint.rs

@@ -2,7 +2,9 @@
 //!
 //! In the local test environment, the data for each endpoint is stored in
 //!
+//! ```text
 //!   .neon/endpoints/<endpoint id>
+//! ```
 //!
 //! Some basic information about the endpoint, like the tenant and timeline IDs,
 //! are stored in the `endpoint.json` file. The `endpoint.json` file is created
@@ -22,7 +24,7 @@
 //!
 //! Directory contents:
 //!
-//! ```ignore
+//! ```text
 //! .neon/endpoints/main/
 //!     compute.log               - log output of `compute_ctl` and `postgres`
 //!     endpoint.json             - serialized `EndpointConf` struct
@@ -287,7 +289,7 @@ impl Endpoint {
                         .env
                         .safekeepers
                         .iter()
-                        .map(|sk| format!("localhost:{}", sk.pg_port))
+                        .map(|sk| format!("localhost:{}", sk.get_compute_port()))
                         .collect::<Vec<String>>()
                         .join(",");
                     conf.append("neon.safekeepers", &safekeepers);
@@ -316,7 +318,7 @@ impl Endpoint {
                     .env
                     .safekeepers
                     .iter()
-                    .map(|x| x.pg_port.to_string())
+                    .map(|x| x.get_compute_port().to_string())
                     .collect::<Vec<_>>()
                     .join(",");
                 let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
@@ -466,7 +468,7 @@ impl Endpoint {
                     .iter()
                     .find(|node| node.id == sk_id)
                     .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
-                safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.pg_port));
+                safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
             }
         }
 
@@ -491,15 +493,7 @@ impl Endpoint {
             pageserver_connstring: Some(pageserver_connstring),
             safekeeper_connstrings,
             storage_auth_token: auth_token.clone(),
-            // TODO FIXME: This is a hack to test custom extensions locally.
-            // In test_download_extensions, we assume that the custom extension
-            // prefix is the tenant ID. So we set it here.
-            //
-            // The proper way to implement this is to pass the custom extension
-            // in spec, but we don't have a way to do that yet in the python tests.
-            // NEW HACK: we enable the anon custom extension for everyone! this is of course just for testing
-            // how will we do it for real?
-            custom_extensions: Some(vec!["123454321".to_string(), self.tenant_id.to_string()]),
+            custom_extensions: Some(vec![]),
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -581,9 +575,7 @@ impl Endpoint {
                 }
                 Err(e) => {
                     if attempt == MAX_ATTEMPTS {
-                        return Err(e).context(
-                            "timed out waiting to connect to compute_ctl HTTP; last error: {e}",
-                        );
+                        return Err(e).context("timed out waiting to connect to compute_ctl HTTP");
                     }
                 }
             }

control_plane/src/local_env.rs

@@ -137,6 +137,7 @@ impl Default for PageServerConf {
 pub struct SafekeeperConf {
     pub id: NodeId,
     pub pg_port: u16,
+    pub pg_tenant_only_port: Option<u16>,
     pub http_port: u16,
     pub sync: bool,
     pub remote_storage: Option<String>,
@@ -149,6 +150,7 @@ impl Default for SafekeeperConf {
         Self {
             id: NodeId(0),
             pg_port: 0,
+            pg_tenant_only_port: None,
             http_port: 0,
             sync: true,
             remote_storage: None,
@@ -158,6 +160,14 @@ impl Default for SafekeeperConf {
     }
 }
 
+impl SafekeeperConf {
+    /// Compute is served by port on which only tenant scoped tokens allowed, if
+    /// it is configured.
+    pub fn get_compute_port(&self) -> u16 {
+        self.pg_tenant_only_port.unwrap_or(self.pg_port)
+    }
+}
+
 impl LocalEnv {
     pub fn pg_distrib_dir_raw(&self) -> PathBuf {
         self.pg_distrib_dir.clone()

control_plane/src/safekeeper.rs

@@ -2,8 +2,9 @@
 //!
 //! In the local test environment, the data for each safekeeper is stored in
 //!
+//! ```text
 //!   .neon/safekeepers/<safekeeper id>
-//!
+//! ```
 use std::io::Write;
 use std::path::PathBuf;
 use std::process::Child;
@@ -119,45 +120,55 @@ impl SafekeeperNode {
         let availability_zone = format!("sk-{}", id_string);
 
         let mut args = vec![
-            "-D",
-            datadir.to_str().with_context(|| {
-                format!("Datadir path {datadir:?} cannot be represented as a unicode string")
-            })?,
-            "--id",
-            &id_string,
-            "--listen-pg",
-            &listen_pg,
-            "--listen-http",
-            &listen_http,
-            "--availability-zone",
-            &availability_zone,
+            "-D".to_owned(),
+            datadir
+                .to_str()
+                .with_context(|| {
+                    format!("Datadir path {datadir:?} cannot be represented as a unicode string")
+                })?
+                .to_owned(),
+            "--id".to_owned(),
+            id_string,
+            "--listen-pg".to_owned(),
+            listen_pg,
+            "--listen-http".to_owned(),
+            listen_http,
+            "--availability-zone".to_owned(),
+            availability_zone,
         ];
+        if let Some(pg_tenant_only_port) = self.conf.pg_tenant_only_port {
+            let listen_pg_tenant_only = format!("127.0.0.1:{}", pg_tenant_only_port);
+            args.extend(["--listen-pg-tenant-only".to_owned(), listen_pg_tenant_only]);
+        }
         if !self.conf.sync {
-            args.push("--no-sync");
+            args.push("--no-sync".to_owned());
         }
 
         let broker_endpoint = format!("{}", self.env.broker.client_url());
-        args.extend(["--broker-endpoint", &broker_endpoint]);
+        args.extend(["--broker-endpoint".to_owned(), broker_endpoint]);
 
         let mut backup_threads = String::new();
         if let Some(threads) = self.conf.backup_threads {
             backup_threads = threads.to_string();
-            args.extend(["--backup-threads", &backup_threads]);
+            args.extend(["--backup-threads".to_owned(), backup_threads]);
         } else {
             drop(backup_threads);
         }
 
         if let Some(ref remote_storage) = self.conf.remote_storage {
-            args.extend(["--remote-storage", remote_storage]);
+            args.extend(["--remote-storage".to_owned(), remote_storage.clone()]);
         }
 
         let key_path = self.env.base_data_dir.join("auth_public_key.pem");
         if self.conf.auth_enabled {
             args.extend([
-                "--auth-validation-public-key-path",
-                key_path.to_str().with_context(|| {
-                    format!("Key path {key_path:?} cannot be represented as a unicode string")
-                })?,
+                "--auth-validation-public-key-path".to_owned(),
+                key_path
+                    .to_str()
+                    .with_context(|| {
+                        format!("Key path {key_path:?} cannot be represented as a unicode string")
+                    })?
+                    .to_owned(),
             ]);
         }
 

docs/pageserver-thread-mgmt.md

@@ -30,8 +30,8 @@ or similar, to wake up on shutdown.
 
 In async Rust, futures can be "cancelled" at any await point, by
 dropping the Future. For example, `tokio::select!` returns as soon as
-one of the Futures returns, and drops the others. `tokio::timeout!` is
-another example. In the Rust ecosystem, some functions are
+one of the Futures returns, and drops the others. `tokio::time::timeout`
+is another example. In the Rust ecosystem, some functions are
 cancellation-safe, meaning they can be safely dropped without
 side-effects, while others are not. See documentation of
 `tokio::select!` for examples.
@@ -42,9 +42,9 @@ function that you call cannot be assumed to be async
 cancellation-safe, and must be polled to completion.
 
 The downside of non-cancellation safe code is that you have to be very
-careful when using `tokio::select!`, `tokio::timeout!`, and other such
-functions that can cause a Future to be dropped. They can only be used
-with functions that are explicitly documented to be cancellation-safe,
+careful when using `tokio::select!`, `tokio::time::timeout`, and other
+such functions that can cause a Future to be dropped. They can only be
+used with functions that are explicitly documented to be cancellation-safe,
 or you need to spawn a separate task to shield from the cancellation.
 
 At the entry points to the code, we also take care to poll futures to

docs/rfcs/024-extension-loading.md

@@ -141,37 +141,86 @@ popular extensions.
 
 The layout of the S3 bucket is as follows:
 ```
-v14/ext_index.json
-    -- this contains information necessary to create control files
-v14/extensions/test_ext1.tar.gz
-    -- this contains the library files and sql files necessary to create this extension
-v14/extensions/custom_ext1.tar.gz
+5615610098 // this is an extension build number
+├── v14
+│   ├── extensions
+│   │   ├── anon.tar.zst
+│   │   └── embedding.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   ├── anon.tar.zst
+    │   └── embedding.tar.zst
+    └── ext_index.json
+5615261079
+├── v14
+│   ├── extensions
+│   │   └── anon.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   └── anon.tar.zst
+    └── ext_index.json
+5623261088
+├── v14
+│   ├── extensions
+│   │   └── embedding.tar.zst
+│   └── ext_index.json
+└── v15
+    ├── extensions
+    │   └── embedding.tar.zst
+    └── ext_index.json
 ```
-The difference between private and public extensions is determined by who can
-load the extension. This is specified in `ext_index.json`.
-Speicially, `ext_index.json` has a list of public extensions, and a list of
-extensions enabled for specific tenant-ids. Here is an example `ext_index.json`:
+
+Note that build number cannot be part of prefix because we might need extensions
+from other build numbers.
+
+`ext_index.json` stores the control files and location of extension archives. 
+It also stores a list of public extensions and a library_index
+
+We don't need to duplicate `extension.tar.zst`` files.
+We only need to upload a new one if it is updated.
+(Although currently we just upload every time anyways, hopefully will change
+this sometime)
+
+*access* is controlled by spec
+
+More specifically, here is an example ext_index.json
 ```
 {
-  "enabled_extensions": {
-    "123454321": [
-      "anon"
+    "public_extensions": [
+        "anon",
+        "pg_buffercache"
     ],
-    "public": [
-      "embedding"
-    ]
-  },
-  "control_data": {
-    "embedding": "comment = 'hnsw index' \ndefault_version = '0.1.0' \nmodule_pathname = '$libdir/embedding' \nrelocatable = true \ntrusted = true",
-    "anon": "# PostgreSQL Anonymizer (anon) extension \ncomment = 'Data anonymization tools' \ndefault_version = '1.1.0' \ndirectory='extension/anon' \nrelocatable = false \nrequires = 'pgcrypto' \nsuperuser = false \nmodule_pathname = '$libdir/anon' \ntrusted = true \n"
-  }
+    "library_index": {
+        "anon": "anon",
+        "pg_buffercache": "pg_buffercache"
+        // for more complex extensions like postgis
+        // we might have something like:
+        // address_standardizer: postgis
+        // postgis_tiger: postgis
+    },
+    "extension_data": {
+        "pg_buffercache": {
+            "control_data": {
+                "pg_buffercache.control": "# pg_buffercache extension \ncomment = 'examine the shared buffer cache' \ndefault_version = '1.3' \nmodule_pathname = '$libdir/pg_buffercache' \nrelocatable = true \ntrusted=true"
+            },
+            "archive_path": "5670669815/v14/extensions/pg_buffercache.tar.zst"
+        },
+        "anon": {
+            "control_data": {
+                "anon.control": "# PostgreSQL Anonymizer (anon) extension \ncomment = 'Data anonymization tools' \ndefault_version = '1.1.0' \ndirectory='extension/anon' \nrelocatable = false \nrequires = 'pgcrypto' \nsuperuser = false \nmodule_pathname = '$libdir/anon' \ntrusted = true \n"
+            },
+            "archive_path": "5670669815/v14/extensions/anon.tar.zst"
+        }
+    }
 }
 ```
 
 ### How to add new extension to the Extension Storage?
 
 Simply upload build artifacts to the S3 bucket.
-Implement a CI step for that. Splitting it from ompute-node-image build.
+Implement a CI step for that. Splitting it from compute-node-image build.
 
 ### How do we deal with extension versions and updates?
 

docs/tools.md

@@ -0,0 +1,22 @@
+# Useful development tools
+
+This readme contains some hints on how to set up some optional development tools.
+
+## ccls
+
+[ccls](https://github.com/MaskRay/ccls) is a c/c++ language server. It requires some setup
+to work well. There are different ways to do it but here's what works for me:
+1. Make a common parent directory for all your common neon projects. (for example, `~/src/neondatabase/`)
+2. Go to `vendor/postgres-v15`
+3. Run `make clean && ./configure`
+4. Install [bear](https://github.com/rizsotto/Bear), and run `bear -- make -j4`
+5. Copy the generated `compile_commands.json` to `~/src/neondatabase` (or equivalent)
+6. Run `touch ~/src/neondatabase/.ccls-root` this will make the `compile_commands.json` file discoverable in all subdirectories
+
+With this setup you will get decent lsp mileage inside the postgres repo, and also any postgres extensions that you put in `~/src/neondatabase/`, like `pg_embedding`, or inside `~/src/neondatabase/neon/pgxn` as well.
+
+Some additional tips for various IDEs:
+
+### Emacs
+
+To improve performance: `(setq lsp-lens-enable nil)`

libs/compute_api/src/responses.rs

@@ -70,12 +70,17 @@ where
 pub struct ComputeMetrics {
     pub wait_for_spec_ms: u64,
     pub sync_safekeepers_ms: u64,
+    pub sync_sk_check_ms: u64,
     pub basebackup_ms: u64,
     pub basebackup_bytes: u64,
     pub start_postgres_ms: u64,
     pub config_ms: u64,
     pub total_startup_ms: u64,
-    pub load_libraries_ms: u64,
+    pub load_ext_ms: u64,
+    pub num_ext_downloaded: u64,
+    pub largest_ext_size: u64, // these are measured in bytes
+    pub total_ext_download_size: u64,
+    pub prep_extensions_ms: u64,
 }
 
 /// Response of the `/computes/{compute_id}/spec` control-plane API.

libs/consumption_metrics/src/lib.rs

@@ -5,7 +5,7 @@ use chrono::{DateTime, Utc};
 use rand::Rng;
 use serde::Serialize;
 
-#[derive(Serialize, Debug, Clone, Eq, PartialEq, Ord, PartialOrd)]
+#[derive(Serialize, Debug, Clone, Copy, Eq, PartialEq, Ord, PartialOrd)]
 #[serde(tag = "type")]
 pub enum EventType {
     #[serde(rename = "absolute")]
@@ -17,6 +17,32 @@ pub enum EventType {
     },
 }
 
+impl EventType {
+    pub fn absolute_time(&self) -> Option<&DateTime<Utc>> {
+        use EventType::*;
+        match self {
+            Absolute { time } => Some(time),
+            _ => None,
+        }
+    }
+
+    pub fn incremental_timerange(&self) -> Option<std::ops::Range<&DateTime<Utc>>> {
+        // these can most likely be thought of as Range or RangeFull
+        use EventType::*;
+        match self {
+            Incremental {
+                start_time,
+                stop_time,
+            } => Some(start_time..stop_time),
+            _ => None,
+        }
+    }
+
+    pub fn is_incremental(&self) -> bool {
+        matches!(self, EventType::Incremental { .. })
+    }
+}
+
 #[derive(Serialize, Debug, Clone, Eq, PartialEq, Ord, PartialOrd)]
 pub struct Event<Extra> {
     #[serde(flatten)]
@@ -31,7 +57,7 @@ pub struct Event<Extra> {
     pub extra: Extra,
 }
 
-pub fn idempotency_key(node_id: String) -> String {
+pub fn idempotency_key(node_id: &str) -> String {
     format!(
         "{}-{}-{:04}",
         Utc::now(),
@@ -45,6 +71,6 @@ pub const CHUNK_SIZE: usize = 1000;
 // Just a wrapper around a slice of events
 // to serialize it as `{"events" : [ ] }
 #[derive(serde::Serialize)]
-pub struct EventChunk<'a, T> {
-    pub events: &'a [T],
+pub struct EventChunk<'a, T: Clone> {
+    pub events: std::borrow::Cow<'a, [T]>,
 }

libs/metrics/src/lib.rs

@@ -6,6 +6,7 @@ use once_cell::sync::Lazy;
 use prometheus::core::{AtomicU64, Collector, GenericGauge, GenericGaugeVec};
 pub use prometheus::opts;
 pub use prometheus::register;
+pub use prometheus::Error;
 pub use prometheus::{core, default_registry, proto};
 pub use prometheus::{exponential_buckets, linear_buckets};
 pub use prometheus::{register_counter_vec, Counter, CounterVec};

libs/metrics/src/metric_vec_duration.rs

@@ -1,4 +1,4 @@
-//! Helpers for observing duration on HistogramVec / CounterVec / GaugeVec / MetricVec<T>.
+//! Helpers for observing duration on `HistogramVec` / `CounterVec` / `GaugeVec` / `MetricVec<T>`.
 
 use std::{future::Future, time::Instant};
 

libs/pageserver_api/src/models.rs

@@ -9,6 +9,7 @@ use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 use strum_macros;
 use utils::{
+    completion,
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
@@ -76,7 +77,12 @@ pub enum TenantState {
     /// system is being shut down.
     ///
     /// Transitions out of this state are possible through `set_broken()`.
-    Stopping,
+    Stopping {
+        // Because of https://github.com/serde-rs/serde/issues/2105 this has to be a named field,
+        // otherwise it will not be skipped during deserialization
+        #[serde(skip)]
+        progress: completion::Barrier,
+    },
     /// The tenant is recognized by the pageserver, but can no longer be used for
     /// any operations.
     ///
@@ -118,7 +124,7 @@ impl TenantState {
             // Why is Stopping a Maybe case? Because, during pageserver shutdown,
             // we set the Stopping state irrespective of whether the tenant
             // has finished attaching or not.
-            Self::Stopping => Maybe,
+            Self::Stopping { .. } => Maybe,
         }
     }
 
@@ -411,12 +417,16 @@ pub struct LayerResidenceEvent {
     pub reason: LayerResidenceEventReason,
 }
 
-/// The reason for recording a given [`ResidenceEvent`].
+/// The reason for recording a given [`LayerResidenceEvent`].
 #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
 pub enum LayerResidenceEventReason {
     /// The layer map is being populated, e.g. during timeline load or attach.
     /// This includes [`RemoteLayer`] objects created in [`reconcile_with_remote`].
     /// We need to record such events because there is no persistent storage for the events.
+    ///
+    // https://github.com/rust-lang/rust/issues/74481
+    /// [`RemoteLayer`]: ../../tenant/storage_layer/struct.RemoteLayer.html
+    /// [`reconcile_with_remote`]: ../../tenant/struct.Timeline.html#method.reconcile_with_remote
     LayerLoad,
     /// We just created the layer (e.g., freeze_and_flush or compaction).
     /// Such layers are always [`LayerResidenceStatus::Resident`].
@@ -924,7 +934,13 @@ mod tests {
                 "Activating",
             ),
             (line!(), TenantState::Active, "Active"),
-            (line!(), TenantState::Stopping, "Stopping"),
+            (
+                line!(),
+                TenantState::Stopping {
+                    progress: utils::completion::Barrier::default(),
+                },
+                "Stopping",
+            ),
             (
                 line!(),
                 TenantState::Broken {

libs/pageserver_api/src/reltag.rs

@@ -60,8 +60,9 @@ impl Ord for RelTag {
 
 /// Display RelTag in the same format that's used in most PostgreSQL debug messages:
 ///
+/// ```text
 /// <spcnode>/<dbnode>/<relnode>[_fsm|_vm|_init]
-///
+/// ```
 impl fmt::Display for RelTag {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         if let Some(forkname) = forknumber_to_name(self.forknum) {

libs/postgres_ffi/src/nonrelfile_utils.rs

@@ -57,9 +57,9 @@ pub fn slru_may_delete_clogsegment(segpage: u32, cutoff_page: u32) -> bool {
 // Multixact utils
 
 pub fn mx_offset_to_flags_offset(xid: MultiXactId) -> usize {
-    ((xid / pg_constants::MULTIXACT_MEMBERS_PER_MEMBERGROUP as u32) as u16
-        % pg_constants::MULTIXACT_MEMBERGROUPS_PER_PAGE
-        * pg_constants::MULTIXACT_MEMBERGROUP_SIZE) as usize
+    ((xid / pg_constants::MULTIXACT_MEMBERS_PER_MEMBERGROUP as u32)
+        % pg_constants::MULTIXACT_MEMBERGROUPS_PER_PAGE as u32
+        * pg_constants::MULTIXACT_MEMBERGROUP_SIZE as u32) as usize
 }
 
 pub fn mx_offset_to_flags_bitshift(xid: MultiXactId) -> u16 {
@@ -81,3 +81,41 @@ fn mx_offset_to_member_page(xid: u32) -> u32 {
 pub fn mx_offset_to_member_segment(xid: u32) -> i32 {
     (mx_offset_to_member_page(xid) / pg_constants::SLRU_PAGES_PER_SEGMENT) as i32
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_multixid_calc() {
+        // Check that the mx_offset_* functions produce the same values as the
+        // corresponding PostgreSQL C macros (MXOffsetTo*). These test values
+        // were generated by calling the PostgreSQL macros with a little C
+        // program.
+        assert_eq!(mx_offset_to_member_segment(0), 0);
+        assert_eq!(mx_offset_to_member_page(0), 0);
+        assert_eq!(mx_offset_to_flags_offset(0), 0);
+        assert_eq!(mx_offset_to_flags_bitshift(0), 0);
+        assert_eq!(mx_offset_to_member_offset(0), 4);
+        assert_eq!(mx_offset_to_member_segment(1), 0);
+        assert_eq!(mx_offset_to_member_page(1), 0);
+        assert_eq!(mx_offset_to_flags_offset(1), 0);
+        assert_eq!(mx_offset_to_flags_bitshift(1), 8);
+        assert_eq!(mx_offset_to_member_offset(1), 8);
+        assert_eq!(mx_offset_to_member_segment(123456789), 2358);
+        assert_eq!(mx_offset_to_member_page(123456789), 75462);
+        assert_eq!(mx_offset_to_flags_offset(123456789), 4780);
+        assert_eq!(mx_offset_to_flags_bitshift(123456789), 8);
+        assert_eq!(mx_offset_to_member_offset(123456789), 4788);
+        assert_eq!(mx_offset_to_member_segment(u32::MAX - 1), 82040);
+        assert_eq!(mx_offset_to_member_page(u32::MAX - 1), 2625285);
+        assert_eq!(mx_offset_to_flags_offset(u32::MAX - 1), 5160);
+        assert_eq!(mx_offset_to_flags_bitshift(u32::MAX - 1), 16);
+        assert_eq!(mx_offset_to_member_offset(u32::MAX - 1), 5172);
+        assert_eq!(mx_offset_to_member_segment(u32::MAX), 82040);
+        assert_eq!(mx_offset_to_member_page(u32::MAX), 2625285);
+        assert_eq!(mx_offset_to_flags_offset(u32::MAX), 5160);
+        assert_eq!(mx_offset_to_flags_bitshift(u32::MAX), 24);
+        assert_eq!(mx_offset_to_member_offset(u32::MAX), 5176);
+    }
+}

libs/postgres_ffi/src/relfile_utils.rs

@@ -49,14 +49,16 @@ pub fn forknumber_to_name(forknum: u8) -> Option<&'static str> {
     }
 }
 
-///
 /// Parse a filename of a relation file. Returns (relfilenode, forknum, segno) tuple.
 ///
 /// Formats:
+///
+/// ```text
 /// <oid>
 /// <oid>_<fork name>
 /// <oid>.<segment number>
 /// <oid>_<fork name>.<segment number>
+/// ```
 ///
 /// See functions relpath() and _mdfd_segpath() in PostgreSQL sources.
 ///

libs/pq_proto/src/framed.rs

@@ -5,11 +5,11 @@
 //! It is similar to what tokio_util::codec::Framed with appropriate codec
 //! provides, but `FramedReader` and `FramedWriter` read/write parts can be used
 //! separately without using split from futures::stream::StreamExt (which
-//! allocates box[1] in polling internally). tokio::io::split is used for splitting
+//! allocates a [Box] in polling internally). tokio::io::split is used for splitting
 //! instead. Plus we customize error messages more than a single type for all io
 //! calls.
 //!
-//! [1] https://docs.rs/futures-util/0.3.26/src/futures_util/lock/bilock.rs.html#107
+//! [Box]: https://docs.rs/futures-util/0.3.26/src/futures_util/lock/bilock.rs.html#107
 use bytes::{Buf, BytesMut};
 use std::{
     future::Future,
@@ -117,7 +117,7 @@ impl<S: AsyncWrite + Unpin> Framed<S> {
 impl<S: AsyncRead + AsyncWrite + Unpin> Framed<S> {
     /// Split into owned read and write parts. Beware of potential issues with
     /// using halves in different tasks on TLS stream:
-    /// https://github.com/tokio-rs/tls/issues/40
+    /// <https://github.com/tokio-rs/tls/issues/40>
     pub fn split(self) -> (FramedReader<S>, FramedWriter<S>) {
         let (read_half, write_half) = tokio::io::split(self.stream);
         let reader = FramedReader {

libs/pq_proto/src/lib.rs

@@ -179,7 +179,7 @@ pub struct FeExecuteMessage {
 #[derive(Debug)]
 pub struct FeCloseMessage;
 
-/// An error occured while parsing or serializing raw stream into Postgres
+/// An error occurred while parsing or serializing raw stream into Postgres
 /// messages.
 #[derive(thiserror::Error, Debug)]
 pub enum ProtocolError {
@@ -934,6 +934,15 @@ impl<'a> BeMessage<'a> {
     }
 }
 
+fn terminate_code(code: &[u8; 5]) -> [u8; 6] {
+    let mut terminated = [0; 6];
+    for (i, &elem) in code.iter().enumerate() {
+        terminated[i] = elem;
+    }
+
+    terminated
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -965,12 +974,3 @@ mod tests {
         assert_eq!(split_options(&params), ["foo bar", " \\", "baz ", "lol"]);
     }
 }
-
-fn terminate_code(code: &[u8; 5]) -> [u8; 6] {
-    let mut terminated = [0; 6];
-    for (i, &elem) in code.iter().enumerate() {
-        terminated[i] = elem;
-    }
-
-    terminated
-}

libs/remote_storage/src/lib.rs

@@ -34,12 +34,12 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNCS: usize = 50;
 pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// Currently, sync happens with AWS S3, that has two limits on requests per second:
 /// ~200 RPS for IAM services
-/// https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html
+/// <https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html>
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
-/// https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/
+/// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
-/// https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax
+/// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
@@ -50,6 +50,12 @@ const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
 pub struct RemotePath(PathBuf);
 
+impl std::fmt::Display for RemotePath {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0.display())
+    }
+}
+
 impl RemotePath {
     pub fn new(relative_path: &Path) -> anyhow::Result<Self> {
         anyhow::ensure!(
@@ -59,6 +65,10 @@ impl RemotePath {
         Ok(Self(relative_path.to_path_buf()))
     }
 
+    pub fn from_string(relative_path: &str) -> anyhow::Result<Self> {
+        Self::new(Path::new(relative_path))
+    }
+
     pub fn with_base(&self, base_path: &Path) -> PathBuf {
         base_path.join(&self.0)
     }

libs/remote_storage/src/local_fs.rs

@@ -151,10 +151,7 @@ impl RemoteStorage for LocalFs {
         let mut files = vec![];
         let mut directory_queue = vec![full_path.clone()];
 
-        while !directory_queue.is_empty() {
-            let cur_folder = directory_queue
-                .pop()
-                .expect("queue cannot be empty: we just checked");
+        while let Some(cur_folder) = directory_queue.pop() {
             let mut entries = fs::read_dir(cur_folder.clone()).await?;
             while let Some(entry) = entries.next_entry().await? {
                 let file_name: PathBuf = entry.file_name().into();

libs/remote_storage/src/s3_bucket.rs

@@ -200,13 +200,17 @@ impl S3Bucket {
         )
     }
 
-    fn relative_path_to_s3_object(&self, path: &RemotePath) -> String {
-        let mut full_path = self.prefix_in_bucket.clone().unwrap_or_default();
-        for segment in path.0.iter() {
-            full_path.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
-            full_path.push_str(segment.to_str().unwrap_or_default());
+    pub fn relative_path_to_s3_object(&self, path: &RemotePath) -> String {
+        assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path
+            .get_path()
+            .to_string_lossy()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR)
+            .to_string();
+        match &self.prefix_in_bucket {
+            Some(prefix) => prefix.clone() + "/" + &path_string,
+            None => path_string,
         }
-        full_path
     }
 
     async fn download_object(&self, request: GetObjectRequest) -> Result<Download, DownloadError> {
@@ -349,17 +353,10 @@ impl RemoteStorage for S3Bucket {
 
     /// See the doc for `RemoteStorage::list_files`
     async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
-        let mut folder_name = folder
+        let folder_name = folder
             .map(|p| self.relative_path_to_s3_object(p))
             .or_else(|| self.prefix_in_bucket.clone());
 
-        // remove leading "/" if one exists
-        if let Some(folder_name_slash) = folder_name.clone() {
-            if folder_name_slash.starts_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
-                folder_name = Some(folder_name_slash[1..].to_string());
-            }
-        }
-
         // AWS may need to break the response into several parts
         let mut continuation_token = None;
         let mut all_files = vec![];
@@ -434,10 +431,12 @@ impl RemoteStorage for S3Bucket {
     }
 
     async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+        // if prefix is not none then download file `prefix/from`
+        // if prefix is none then download file `from`
         self.download_object(GetObjectRequest {
             bucket: self.bucket_name.clone(),
             key: self.relative_path_to_s3_object(from),
-            ..GetObjectRequest::default()
+            range: None,
         })
         .await
     }
@@ -530,3 +529,63 @@ impl RemoteStorage for S3Bucket {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::num::NonZeroUsize;
+    use std::path::Path;
+
+    use crate::{RemotePath, S3Bucket, S3Config};
+
+    #[test]
+    fn relative_path() {
+        let all_paths = vec!["", "some/path", "some/path/"];
+        let all_paths: Vec<RemotePath> = all_paths
+            .iter()
+            .map(|x| RemotePath::new(Path::new(x)).expect("bad path"))
+            .collect();
+        let prefixes = [
+            None,
+            Some(""),
+            Some("test/prefix"),
+            Some("test/prefix/"),
+            Some("/test/prefix/"),
+        ];
+        let expected_outputs = vec![
+            vec!["", "some/path", "some/path"],
+            vec!["/", "/some/path", "/some/path"],
+            vec![
+                "test/prefix/",
+                "test/prefix/some/path",
+                "test/prefix/some/path",
+            ],
+            vec![
+                "test/prefix/",
+                "test/prefix/some/path",
+                "test/prefix/some/path",
+            ],
+            vec![
+                "test/prefix/",
+                "test/prefix/some/path",
+                "test/prefix/some/path",
+            ],
+        ];
+
+        for (prefix_idx, prefix) in prefixes.iter().enumerate() {
+            let config = S3Config {
+                bucket_name: "bucket".to_owned(),
+                bucket_region: "region".to_owned(),
+                prefix_in_bucket: prefix.map(str::to_string),
+                endpoint: None,
+                concurrency_limit: NonZeroUsize::new(100).unwrap(),
+                max_keys_per_list_response: Some(5),
+            };
+            let storage = S3Bucket::new(&config).expect("remote storage init");
+            for (test_path_idx, test_path) in all_paths.iter().enumerate() {
+                let result = storage.relative_path_to_s3_object(test_path);
+                let expected = expected_outputs[prefix_idx][test_path_idx];
+                assert_eq!(result, expected);
+            }
+        }
+    }
+}

libs/remote_storage/tests/test_real_s3.rs

@@ -19,7 +19,7 @@ static LOGGING_DONE: OnceCell<()> = OnceCell::new();
 
 const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
 
-const BASE_PREFIX: &str = "test/";
+const BASE_PREFIX: &str = "test";
 
 /// Tests that S3 client can list all prefixes, even if the response come paginated and requires multiple S3 queries.
 /// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified.

libs/tenant_size_model/src/calculation.rs

@@ -21,7 +21,7 @@ use crate::{SegmentMethod, SegmentSizeResult, SizeResult, StorageModel};
 // 2. D+C+a+b
 // 3. D+A+B
 
-/// [`Segment`] which has had it's size calculated.
+/// `Segment` which has had its size calculated.
 #[derive(Clone, Debug)]
 struct SegmentSize {
     method: SegmentMethod,

libs/tracing-utils/src/http.rs

@@ -33,7 +33,7 @@ pub enum OtelName<'a> {
 /// directly into HTTP servers. However, I couldn't find one for Hyper,
 /// so I had to write our own. OpenTelemetry website has a registry of
 /// instrumentation libraries at:
-/// https://opentelemetry.io/registry/?language=rust&component=instrumentation
+/// <https://opentelemetry.io/registry/?language=rust&component=instrumentation>
 /// If a Hyper crate appears, consider switching to that.
 pub async fn tracing_handler<F, R>(
     req: Request<Body>,

libs/utils/Cargo.toml

@@ -42,6 +42,10 @@ workspace_hack.workspace = true
 
 const_format.workspace = true
 
+# to use tokio channels as streams, this is faster to compile than async_stream
+# why is it only here? no other crate should use it, streams are rarely needed.
+tokio-stream = { version = "0.1.14" }
+
 [dev-dependencies]
 byteorder.workspace = true
 bytes.workspace = true

libs/utils/src/auth.rs

@@ -16,7 +16,7 @@ use crate::id::TenantId;
 /// Algorithm to use. We require EdDSA.
 const STORAGE_TOKEN_ALGORITHM: Algorithm = Algorithm::EdDSA;
 
-#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
+#[derive(Debug, Serialize, Deserialize, Clone, Copy, PartialEq)]
 #[serde(rename_all = "lowercase")]
 pub enum Scope {
     // Provides access to all data for a specific tenant (specified in `struct Claims` below)

libs/utils/src/completion.rs

@@ -12,6 +12,13 @@ pub struct Completion(mpsc::Sender<()>);
 #[derive(Clone)]
 pub struct Barrier(Arc<Mutex<mpsc::Receiver<()>>>);
 
+impl Default for Barrier {
+    fn default() -> Self {
+        let (_, rx) = channel();
+        rx
+    }
+}
+
 impl Barrier {
     pub async fn wait(self) {
         self.0.lock().await.recv().await;
@@ -24,6 +31,15 @@ impl Barrier {
     }
 }
 
+impl PartialEq for Barrier {
+    fn eq(&self, other: &Self) -> bool {
+        // we don't use dyn so this is good
+        Arc::ptr_eq(&self.0, &other.0)
+    }
+}
+
+impl Eq for Barrier {}
+
 /// Create new Guard and Barrier pair.
 pub fn channel() -> (Completion, Barrier) {
     let (tx, rx) = mpsc::channel::<()>(1);

libs/utils/src/error.rs

@@ -0,0 +1,111 @@
+/// Create a reporter for an error that outputs similar to [`anyhow::Error`] with Display with alternative setting.
+///
+/// It can be used with `anyhow::Error` as well.
+///
+/// Why would one use this instead of converting to `anyhow::Error` on the spot? Because
+/// anyhow::Error would also capture a stacktrace on the spot, which you would later discard after
+/// formatting.
+///
+/// ## Usage
+///
+/// ```rust
+/// #[derive(Debug, thiserror::Error)]
+/// enum MyCoolError {
+///   #[error("should never happen")]
+///   Bad(#[source] std::io::Error),
+/// }
+///
+/// # fn failing_call() -> Result<(), MyCoolError> { Err(MyCoolError::Bad(std::io::ErrorKind::PermissionDenied.into())) }
+///
+/// # fn main() {
+/// use utils::error::report_compact_sources;
+///
+/// if let Err(e) = failing_call() {
+///     let e = report_compact_sources(&e);
+///     assert_eq!(format!("{e}"), "should never happen: permission denied");
+/// }
+/// # }
+/// ```
+///
+/// ## TODO
+///
+/// When we are able to describe return position impl trait in traits, this should of course be an
+/// extension trait. Until then avoid boxing with this more ackward interface.
+pub fn report_compact_sources<E: std::error::Error>(e: &E) -> impl std::fmt::Display + '_ {
+    struct AnyhowDisplayAlternateAlike<'a, E>(&'a E);
+
+    impl<E: std::error::Error> std::fmt::Display for AnyhowDisplayAlternateAlike<'_, E> {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            write!(f, "{}", self.0)?;
+
+            // why is E a generic parameter here? hope that rustc will see through a default
+            // Error::source implementation and leave the following out if there cannot be any
+            // sources:
+            Sources(self.0.source()).try_for_each(|src| write!(f, ": {}", src))
+        }
+    }
+
+    struct Sources<'a>(Option<&'a (dyn std::error::Error + 'static)>);
+
+    impl<'a> Iterator for Sources<'a> {
+        type Item = &'a (dyn std::error::Error + 'static);
+
+        fn next(&mut self) -> Option<Self::Item> {
+            let rem = self.0;
+
+            let next = self.0.and_then(|x| x.source());
+            self.0 = next;
+            rem
+        }
+    }
+
+    AnyhowDisplayAlternateAlike(e)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::report_compact_sources;
+
+    #[test]
+    fn report_compact_sources_examples() {
+        use std::fmt::Write;
+
+        #[derive(Debug, thiserror::Error)]
+        enum EvictionError {
+            #[error("cannot evict a remote layer")]
+            CannotEvictRemoteLayer,
+            #[error("stat failed")]
+            StatFailed(#[source] std::io::Error),
+            #[error("layer was no longer part of LayerMap")]
+            LayerNotFound(#[source] anyhow::Error),
+        }
+
+        let examples = [
+            (
+                line!(),
+                EvictionError::CannotEvictRemoteLayer,
+                "cannot evict a remote layer",
+            ),
+            (
+                line!(),
+                EvictionError::StatFailed(std::io::ErrorKind::PermissionDenied.into()),
+                "stat failed: permission denied",
+            ),
+            (
+                line!(),
+                EvictionError::LayerNotFound(anyhow::anyhow!("foobar")),
+                "layer was no longer part of LayerMap: foobar",
+            ),
+        ];
+
+        let mut s = String::new();
+
+        for (line, example, expected) in examples {
+            s.clear();
+
+            write!(s, "{}", report_compact_sources(&example)).expect("string grows");
+
+            assert_eq!(s, expected, "example on line {line}");
+        }
+    }
+}

libs/utils/src/fs_ext.rs

@@ -24,12 +24,29 @@ pub async fn is_directory_empty(path: impl AsRef<Path>) -> anyhow::Result<bool>
     Ok(dir.next_entry().await?.is_none())
 }
 
+pub fn ignore_not_found(e: io::Error) -> io::Result<()> {
+    if e.kind() == io::ErrorKind::NotFound {
+        Ok(())
+    } else {
+        Err(e)
+    }
+}
+
+pub fn ignore_absent_files<F>(fs_operation: F) -> io::Result<()>
+where
+    F: Fn() -> io::Result<()>,
+{
+    fs_operation().or_else(ignore_not_found)
+}
+
 #[cfg(test)]
 mod test {
     use std::path::PathBuf;
 
     use crate::fs_ext::is_directory_empty;
 
+    use super::ignore_absent_files;
+
     #[test]
     fn is_empty_dir() {
         use super::PathExt;
@@ -75,4 +92,21 @@ mod test {
         std::fs::remove_file(&file_path).unwrap();
         assert!(is_directory_empty(file_path).await.is_err());
     }
+
+    #[test]
+    fn ignore_absent_files_works() {
+        let dir = tempfile::tempdir().unwrap();
+        let dir_path = dir.path();
+
+        let file_path: PathBuf = dir_path.join("testfile");
+
+        ignore_absent_files(|| std::fs::remove_file(&file_path)).expect("should execute normally");
+
+        let f = std::fs::File::create(&file_path).unwrap();
+        drop(f);
+
+        ignore_absent_files(|| std::fs::remove_file(&file_path)).expect("should execute normally");
+
+        assert!(!file_path.exists());
+    }
 }

libs/utils/src/http/endpoint.rs

@@ -9,7 +9,6 @@ use metrics::{register_int_counter, Encoder, IntCounter, TextEncoder};
 use once_cell::sync::Lazy;
 use routerify::ext::RequestExt;
 use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
-use tokio::task::JoinError;
 use tracing::{self, debug, info, info_span, warn, Instrument};
 
 use std::future::Future;
@@ -148,26 +147,140 @@ impl Drop for RequestCancelled {
 }
 
 async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    use bytes::{Bytes, BytesMut};
+    use std::io::Write as _;
+    use tokio::sync::mpsc;
+    use tokio_stream::wrappers::ReceiverStream;
+
     SERVE_METRICS_COUNT.inc();
 
-    let mut buffer = vec![];
-    let encoder = TextEncoder::new();
+    /// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
+    struct ChannelWriter {
+        buffer: BytesMut,
+        tx: mpsc::Sender<std::io::Result<Bytes>>,
+        written: usize,
+    }
 
-    let metrics = tokio::task::spawn_blocking(move || {
-        // Currently we take a lot of mutexes while collecting metrics, so it's
-        // better to spawn a blocking task to avoid blocking the event loop.
-        metrics::gather()
-    })
-    .await
-    .map_err(|e: JoinError| ApiError::InternalServerError(e.into()))?;
-    encoder.encode(&metrics, &mut buffer).unwrap();
+    impl ChannelWriter {
+        fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
+            assert_ne!(buf_len, 0);
+            ChannelWriter {
+                // split about half off the buffer from the start, because we flush depending on
+                // capacity. first flush will come sooner than without this, but now resizes will
+                // have better chance of picking up the "other" half. not guaranteed of course.
+                buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
+                tx,
+                written: 0,
+            }
+        }
+
+        fn flush0(&mut self) -> std::io::Result<usize> {
+            let n = self.buffer.len();
+            if n == 0 {
+                return Ok(0);
+            }
+
+            tracing::trace!(n, "flushing");
+            let ready = self.buffer.split().freeze();
+
+            // not ideal to call from blocking code to block_on, but we are sure that this
+            // operation does not spawn_blocking other tasks
+            let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
+                self.tx.send(Ok(ready)).await.map_err(|_| ())?;
+
+                // throttle sending to allow reuse of our buffer in `write`.
+                self.tx.reserve().await.map_err(|_| ())?;
+
+                // now the response task has picked up the buffer and hopefully started
+                // sending it to the client.
+                Ok(())
+            });
+            if res.is_err() {
+                return Err(std::io::ErrorKind::BrokenPipe.into());
+            }
+            self.written += n;
+            Ok(n)
+        }
+
+        fn flushed_bytes(&self) -> usize {
+            self.written
+        }
+    }
+
+    impl std::io::Write for ChannelWriter {
+        fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
+            let remaining = self.buffer.capacity() - self.buffer.len();
+
+            let out_of_space = remaining < buf.len();
+
+            let original_len = buf.len();
+
+            if out_of_space {
+                let can_still_fit = buf.len() - remaining;
+                self.buffer.extend_from_slice(&buf[..can_still_fit]);
+                buf = &buf[can_still_fit..];
+                self.flush0()?;
+            }
+
+            // assume that this will often under normal operation just move the pointer back to the
+            // beginning of allocation, because previous split off parts are already sent and
+            // dropped.
+            self.buffer.extend_from_slice(buf);
+            Ok(original_len)
+        }
+
+        fn flush(&mut self) -> std::io::Result<()> {
+            self.flush0().map(|_| ())
+        }
+    }
+
+    let started_at = std::time::Instant::now();
+
+    let (tx, rx) = mpsc::channel(1);
+
+    let body = Body::wrap_stream(ReceiverStream::new(rx));
+
+    let mut writer = ChannelWriter::new(128 * 1024, tx);
+
+    let encoder = TextEncoder::new();
 
     let response = Response::builder()
         .status(200)
         .header(CONTENT_TYPE, encoder.format_type())
-        .body(Body::from(buffer))
+        .body(body)
         .unwrap();
 
+    let span = info_span!("blocking");
+    tokio::task::spawn_blocking(move || {
+        let _span = span.entered();
+        let metrics = metrics::gather();
+        let res = encoder
+            .encode(&metrics, &mut writer)
+            .and_then(|_| writer.flush().map_err(|e| e.into()));
+
+        match res {
+            Ok(()) => {
+                tracing::info!(
+                    bytes = writer.flushed_bytes(),
+                    elapsed_ms = started_at.elapsed().as_millis(),
+                    "responded /metrics"
+                );
+            }
+            Err(e) => {
+                tracing::warn!("failed to write out /metrics response: {e:#}");
+                // semantics of this error are quite... unclear. we want to error the stream out to
+                // abort the response to somehow notify the client that we failed.
+                //
+                // though, most likely the reason for failure is that the receiver is already gone.
+                drop(
+                    writer
+                        .tx
+                        .blocking_send(Err(std::io::ErrorKind::BrokenPipe.into())),
+                );
+            }
+        }
+    });
+
     Ok(response)
 }
 

libs/utils/src/http/json.rs

@@ -14,7 +14,7 @@ pub async fn json_request<T: for<'de> Deserialize<'de>>(
         .map_err(ApiError::BadRequest)
 }
 
-/// Will be removed as part of https://github.com/neondatabase/neon/issues/4282
+/// Will be removed as part of <https://github.com/neondatabase/neon/issues/4282>
 pub async fn json_request_or_empty_body<T: for<'de> Deserialize<'de>>(
     request: &mut Request<Body>,
 ) -> Result<Option<T>, ApiError> {

libs/utils/src/id.rs

@@ -1,5 +1,7 @@
+use std::ffi::OsStr;
 use std::{fmt, str::FromStr};
 
+use anyhow::Context;
 use hex::FromHex;
 use rand::Rng;
 use serde::{Deserialize, Serialize};
@@ -213,6 +215,18 @@ pub struct TimelineId(Id);
 
 id_newtype!(TimelineId);
 
+impl TryFrom<Option<&OsStr>> for TimelineId {
+    type Error = anyhow::Error;
+
+    fn try_from(value: Option<&OsStr>) -> Result<Self, Self::Error> {
+        value
+            .and_then(OsStr::to_str)
+            .unwrap_or_default()
+            .parse::<TimelineId>()
+            .with_context(|| format!("Could not parse timeline id from {:?}", value))
+    }
+}
+
 /// Neon Tenant Id represents identifiar of a particular tenant.
 /// Is used for distinguishing requests and data belonging to different users.
 ///

libs/utils/src/lib.rs

@@ -63,6 +63,9 @@ pub mod rate_limit;
 /// Simple once-barrier and a guard which keeps barrier awaiting.
 pub mod completion;
 
+/// Reporting utilities
+pub mod error;
+
 mod failpoint_macro_helpers {
 
     /// use with fail::cfg("$name", "return(2000)")
@@ -130,8 +133,8 @@ pub use failpoint_macro_helpers::failpoint_sleep_helper;
 /// Note that with git_version prefix is `git:` and in case of git version from env its `git-env:`.
 ///
 /// #############################################################################################
-/// TODO this macro is not the way the library is intended to be used, see https://github.com/neondatabase/neon/issues/1565 for details.
-/// We use `cachepot` to reduce our current CI build times: https://github.com/neondatabase/cloud/pull/1033#issuecomment-1100935036
+/// TODO this macro is not the way the library is intended to be used, see <https://github.com/neondatabase/neon/issues/1565> for details.
+/// We use `cachepot` to reduce our current CI build times: <https://github.com/neondatabase/cloud/pull/1033#issuecomment-1100935036>
 /// Yet, it seems to ignore the GIT_VERSION env variable, passed to Docker build, even with build.rs that contains
 /// `println!("cargo:rerun-if-env-changed=GIT_VERSION");` code for cachepot cache invalidation.
 /// The problem needs further investigation and regular `const` declaration instead of a macro.

libs/utils/src/lock_file.rs

@@ -1,9 +1,10 @@
 //! A module to create and read lock files.
 //!
 //! File locking is done using [`fcntl::flock`] exclusive locks.
-//! The only consumer of this module is currently [`pid_file`].
-//! See the module-level comment there for potential pitfalls
-//! with lock files that are used to store PIDs (pidfiles).
+//! The only consumer of this module is currently
+//! [`pid_file`](crate::pid_file). See the module-level comment
+//! there for potential pitfalls with lock files that are used
+//! to store PIDs (pidfiles).
 
 use std::{
     fs,
@@ -81,7 +82,7 @@ pub fn create_exclusive(lock_file_path: &Path) -> anyhow::Result<UnwrittenLockFi
 }
 
 /// Returned by [`read_and_hold_lock_file`].
-/// Check out the [`pid_file`] module for what the variants mean
+/// Check out the [`pid_file`](crate::pid_file) module for what the variants mean
 /// and potential caveats if the lock files that are used to store PIDs.
 pub enum LockFileRead {
     /// No file exists at the given path.

libs/utils/src/logging.rs

@@ -112,7 +112,7 @@ pub fn init(
 ///
 /// When the return value is dropped, the hook is reverted to std default hook (prints to stderr).
 /// If the assumptions about the initialization order are not held, use
-/// [`TracingPanicHookGuard::disarm`] but keep in mind, if tracing is stopped, then panics will be
+/// [`TracingPanicHookGuard::forget`] but keep in mind, if tracing is stopped, then panics will be
 /// lost.
 #[must_use]
 pub fn replace_panic_hook_with_tracing_panic_hook() -> TracingPanicHookGuard {

libs/utils/src/seqwait.rs

@@ -23,9 +23,9 @@ pub enum SeqWaitError {
 
 /// Monotonically increasing value
 ///
-/// It is handy to store some other fields under the same mutex in SeqWait<S>
+/// It is handy to store some other fields under the same mutex in `SeqWait<S>`
 /// (e.g. store prev_record_lsn). So we allow SeqWait to be parametrized with
-/// any type that can expose counter. <V> is the type of exposed counter.
+/// any type that can expose counter. `V` is the type of exposed counter.
 pub trait MonotonicCounter<V> {
     /// Bump counter value and check that it goes forward
     /// N.B.: new_val is an actual new value, not a difference.
@@ -90,7 +90,7 @@ impl<T: Ord> Eq for Waiter<T> {}
 /// [`wait_for`]: SeqWait::wait_for
 /// [`advance`]: SeqWait::advance
 ///
-/// <S> means Storage, <V> is type of counter that this storage exposes.
+/// `S` means Storage, `V` is type of counter that this storage exposes.
 ///
 pub struct SeqWait<S, V>
 where

libs/utils/src/tracing_span_assert.rs

@@ -1,8 +1,15 @@
 //! Assert that the current [`tracing::Span`] has a given set of fields.
 //!
+//! Can only produce meaningful positive results when tracing has been configured as in example.
+//! Absence of `tracing_error::ErrorLayer` is not detected yet.
+//!
+//! `#[cfg(test)]` code will get a pass when using the `check_fields_present` macro in case tracing
+//! is completly unconfigured.
+//!
 //! # Usage
 //!
-//! ```
+//! ```rust
+//! # fn main() {
 //! use tracing_subscriber::prelude::*;
 //! let registry = tracing_subscriber::registry()
 //!    .with(tracing_error::ErrorLayer::default());
@@ -20,23 +27,18 @@
 //!
 //! use utils::tracing_span_assert::{check_fields_present, MultiNameExtractor};
 //! let extractor = MultiNameExtractor::new("TestExtractor", ["test", "test_id"]);
-//! match check_fields_present([&extractor]) {
-//!    Ok(()) => {},
-//!    Err(missing) => {
-//!        panic!("Missing fields: {:?}", missing.into_iter().map(|f| f.name() ).collect::<Vec<_>>());
-//!    }
+//! if let Err(missing) = check_fields_present!([&extractor]) {
+//!    // if you copypaste this to a custom assert method, remember to add #[track_caller]
+//!    // to get the "user" code location for the panic.
+//!    panic!("Missing fields: {missing:?}");
 //! }
+//! # }
 //! ```
 //!
-//! Recommended reading: https://docs.rs/tracing-subscriber/0.3.16/tracing_subscriber/layer/index.html#per-layer-filtering
+//! Recommended reading: <https://docs.rs/tracing-subscriber/0.3.16/tracing_subscriber/layer/index.html#per-layer-filtering>
 //!
 
-use std::{
-    collections::HashSet,
-    fmt::{self},
-    hash::{Hash, Hasher},
-};
-
+#[derive(Debug)]
 pub enum ExtractionResult {
     Present,
     Absent,
@@ -71,51 +73,103 @@ impl<const L: usize> Extractor for MultiNameExtractor<L> {
     }
 }
 
-struct MemoryIdentity<'a>(&'a dyn Extractor);
-
-impl<'a> MemoryIdentity<'a> {
-    fn as_ptr(&self) -> *const () {
-        self.0 as *const _ as *const ()
-    }
-}
-impl<'a> PartialEq for MemoryIdentity<'a> {
-    fn eq(&self, other: &Self) -> bool {
-        self.as_ptr() == other.as_ptr()
-    }
-}
-impl<'a> Eq for MemoryIdentity<'a> {}
-impl<'a> Hash for MemoryIdentity<'a> {
-    fn hash<H: Hasher>(&self, state: &mut H) {
-        self.as_ptr().hash(state);
-    }
-}
-impl<'a> fmt::Debug for MemoryIdentity<'a> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{:p}: {}", self.as_ptr(), self.0.name())
-    }
-}
-
-/// The extractor names passed as keys to [`new`].
-pub fn check_fields_present<const L: usize>(
+/// Checks that the given extractors are satisfied with the current span hierarchy.
+///
+/// This should not be called directly, but used through [`check_fields_present`] which allows
+/// `Summary::Unconfigured` only when the calling crate is being `#[cfg(test)]` as a conservative default.
+#[doc(hidden)]
+pub fn check_fields_present0<const L: usize>(
     must_be_present: [&dyn Extractor; L],
-) -> Result<(), Vec<&dyn Extractor>> {
-    let mut missing: HashSet<MemoryIdentity> =
-        HashSet::from_iter(must_be_present.into_iter().map(|r| MemoryIdentity(r)));
+) -> Result<Summary, Vec<&dyn Extractor>> {
+    let mut missing = must_be_present.into_iter().collect::<Vec<_>>();
     let trace = tracing_error::SpanTrace::capture();
     trace.with_spans(|md, _formatted_fields| {
-        missing.retain(|extractor| match extractor.0.extract(md.fields()) {
+        // when trying to understand the inner workings of how does the matching work, note that
+        // this closure might be called zero times if the span is disabled. normally it is called
+        // once per span hierarchy level.
+        missing.retain(|extractor| match extractor.extract(md.fields()) {
             ExtractionResult::Present => false,
             ExtractionResult::Absent => true,
         });
-        !missing.is_empty() // continue walking up until we've found all missing
+
+        // continue walking up until we've found all missing
+        !missing.is_empty()
     });
     if missing.is_empty() {
-        Ok(())
+        Ok(Summary::FoundEverything)
+    } else if !tracing_subscriber_configured() {
+        Ok(Summary::Unconfigured)
     } else {
-        Err(missing.into_iter().map(|mi| mi.0).collect())
+        // we can still hit here if a tracing subscriber has been configured but the ErrorLayer is
+        // missing, which can be annoying. for this case, we could probably use
+        // SpanTrace::status().
+        //
+        // another way to end up here is with RUST_LOG=pageserver=off while configuring the
+        // logging, though I guess in that case the SpanTrace::status() == EMPTY would be valid.
+        // this case is covered by test `not_found_if_tracing_error_subscriber_has_wrong_filter`.
+        Err(missing)
     }
 }
 
+/// Checks that the given extractors are satisfied with the current span hierarchy.
+///
+/// The macro is the preferred way of checking if fields exist while passing checks if a test does
+/// not have tracing configured.
+///
+/// Why mangled name? Because #[macro_export] will expose it at utils::__check_fields_present.
+/// However we can game a module namespaced macro for `use` purposes by re-exporting the
+/// #[macro_export] exported name with an alias (below).
+#[doc(hidden)]
+#[macro_export]
+macro_rules! __check_fields_present {
+    ($extractors:expr) => {{
+        {
+            use $crate::tracing_span_assert::{check_fields_present0, Summary::*, Extractor};
+
+            match check_fields_present0($extractors) {
+                Ok(FoundEverything) => Ok(()),
+                Ok(Unconfigured) if cfg!(test) => {
+                    // allow unconfigured in tests
+                    Ok(())
+                },
+                Ok(Unconfigured) => {
+                    panic!("utils::tracing_span_assert: outside of #[cfg(test)] expected tracing to be configured with tracing_error::ErrorLayer")
+                },
+                Err(missing) => Err(missing)
+            }
+        }
+    }}
+}
+
+pub use crate::__check_fields_present as check_fields_present;
+
+/// Explanation for why the check was deemed ok.
+///
+/// Mainly useful for testing, or configuring per-crate behaviour as in with
+/// [`check_fields_present`].
+#[derive(Debug)]
+pub enum Summary {
+    /// All extractors were found.
+    ///
+    /// Should only happen when tracing is properly configured.
+    FoundEverything,
+
+    /// Tracing has not been configured at all. This is ok for tests running without tracing set
+    /// up.
+    Unconfigured,
+}
+
+fn tracing_subscriber_configured() -> bool {
+    let mut noop_configured = false;
+    tracing::dispatcher::get_default(|d| {
+        // it is possible that this closure will not be invoked, but the current implementation
+        // always invokes it
+        noop_configured = d.is::<tracing::subscriber::NoSubscriber>();
+    });
+
+    !noop_configured
+}
+
 #[cfg(test)]
 mod tests {
 
@@ -123,6 +177,36 @@ mod tests {
 
     use super::*;
 
+    use std::{
+        collections::HashSet,
+        fmt::{self},
+        hash::{Hash, Hasher},
+    };
+
+    struct MemoryIdentity<'a>(&'a dyn Extractor);
+
+    impl<'a> MemoryIdentity<'a> {
+        fn as_ptr(&self) -> *const () {
+            self.0 as *const _ as *const ()
+        }
+    }
+    impl<'a> PartialEq for MemoryIdentity<'a> {
+        fn eq(&self, other: &Self) -> bool {
+            self.as_ptr() == other.as_ptr()
+        }
+    }
+    impl<'a> Eq for MemoryIdentity<'a> {}
+    impl<'a> Hash for MemoryIdentity<'a> {
+        fn hash<H: Hasher>(&self, state: &mut H) {
+            self.as_ptr().hash(state);
+        }
+    }
+    impl<'a> fmt::Debug for MemoryIdentity<'a> {
+        fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::fmt::Result {
+            write!(f, "{:p}: {}", self.as_ptr(), self.0.name())
+        }
+    }
+
     struct Setup {
         _current_thread_subscriber_guard: tracing::subscriber::DefaultGuard,
         tenant_extractor: MultiNameExtractor<2>,
@@ -159,7 +243,8 @@ mod tests {
         let setup = setup_current_thread();
         let span = tracing::info_span!("root", tenant_id = "tenant-1", timeline_id = "timeline-1");
         let _guard = span.enter();
-        check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
     }
 
     #[test]
@@ -167,8 +252,8 @@ mod tests {
         let setup = setup_current_thread();
         let span = tracing::info_span!("root", timeline_id = "timeline-1");
         let _guard = span.enter();
-        let missing =
-            check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor])
+            .unwrap_err();
         assert_missing(missing, vec![&setup.tenant_extractor]);
     }
 
@@ -185,7 +270,8 @@ mod tests {
         let span = tracing::info_span!("grandchild", timeline_id = "timeline-1");
         let _guard = span.enter();
 
-        check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
     }
 
     #[test]
@@ -198,7 +284,7 @@ mod tests {
         let span = tracing::info_span!("child", timeline_id = "timeline-1");
         let _guard = span.enter();
 
-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
         assert_missing(missing, vec![&setup.tenant_extractor]);
     }
 
@@ -207,7 +293,8 @@ mod tests {
         let setup = setup_current_thread();
         let span = tracing::info_span!("root", tenant_id = "tenant-1", timeline_id = "timeline-1");
         let _guard = span.enter();
-        check_fields_present([&setup.tenant_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
     }
 
     #[test]
@@ -223,7 +310,8 @@ mod tests {
         let span = tracing::info_span!("grandchild", timeline_id = "timeline-1");
         let _guard = span.enter();
 
-        check_fields_present([&setup.tenant_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
     }
 
     #[test]
@@ -231,7 +319,7 @@ mod tests {
         let setup = setup_current_thread();
         let span = tracing::info_span!("root", timeline_id = "timeline-1");
         let _guard = span.enter();
-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
         assert_missing(missing, vec![&setup.tenant_extractor]);
     }
 
@@ -245,43 +333,107 @@ mod tests {
         let span = tracing::info_span!("child", timeline_id = "timeline-1");
         let _guard = span.enter();
 
-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
         assert_missing(missing, vec![&setup.tenant_extractor]);
     }
 
     #[test]
-    fn tracing_error_subscriber_not_set_up() {
+    fn tracing_error_subscriber_not_set_up_straight_line() {
         // no setup
-
         let span = tracing::info_span!("foo", e = "some value");
         let _guard = span.enter();
 
         let extractor = MultiNameExtractor::new("E", ["e"]);
-        let missing = check_fields_present([&extractor]).unwrap_err();
-        assert_missing(missing, vec![&extractor]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+
+        // similarly for a not found key
+        let extractor = MultiNameExtractor::new("F", ["foobar"]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
     }
 
     #[test]
-    #[should_panic]
-    fn panics_if_tracing_error_subscriber_has_wrong_filter() {
+    fn tracing_error_subscriber_not_set_up_with_instrument() {
+        // no setup
+
+        // demo a case where span entering is used to establish a parent child connection, but
+        // when we re-enter the subspan SpanTrace::with_spans iterates over nothing.
+        let span = tracing::info_span!("foo", e = "some value");
+        let _guard = span.enter();
+
+        let subspan = tracing::info_span!("bar", f = "foobar");
+        drop(_guard);
+
+        // normally this would work, but without any tracing-subscriber configured, both
+        // check_field_present find nothing
+        let _guard = subspan.enter();
+        let extractors: [&dyn Extractor; 2] = [
+            &MultiNameExtractor::new("E", ["e"]),
+            &MultiNameExtractor::new("F", ["f"]),
+        ];
+
+        let res = check_fields_present0(extractors);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+
+        // similarly for a not found key
+        let extractor = MultiNameExtractor::new("G", ["g"]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+    }
+
+    #[test]
+    fn tracing_subscriber_configured() {
+        // this will fail if any utils::logging::init callers appear, but let's hope they do not
+        // appear.
+        assert!(!super::tracing_subscriber_configured());
+
+        let _g = setup_current_thread();
+
+        assert!(super::tracing_subscriber_configured());
+    }
+
+    #[test]
+    fn not_found_when_disabled_by_filter() {
         let r = tracing_subscriber::registry().with({
-            tracing_error::ErrorLayer::default().with_filter(
-                tracing_subscriber::filter::dynamic_filter_fn(|md, _| {
-                    if md.is_span() && *md.level() == tracing::Level::INFO {
-                        return false;
-                    }
-                    true
-                }),
-            )
+            tracing_error::ErrorLayer::default().with_filter(tracing_subscriber::filter::filter_fn(
+                |md| !(md.is_span() && *md.level() == tracing::Level::INFO),
+            ))
         });
 
         let _guard = tracing::subscriber::set_default(r);
 
+        // this test is a rather tricky one, it has a number of possible outcomes depending on the
+        // execution order when executed with other tests even if no test sets the global default
+        // subscriber.
+
         let span = tracing::info_span!("foo", e = "some value");
         let _guard = span.enter();
 
-        let extractor = MultiNameExtractor::new("E", ["e"]);
-        let missing = check_fields_present([&extractor]).unwrap_err();
-        assert_missing(missing, vec![&extractor]);
+        let extractors: [&dyn Extractor; 1] = [&MultiNameExtractor::new("E", ["e"])];
+
+        if span.is_disabled() {
+            // the tests are running single threaded, or we got lucky and no other tests subscriber
+            // was got to register their per-CALLSITE::META interest between `set_default` and
+            // creation of the span, thus the filter got to apply and registered interest of Never,
+            // so the span was never created.
+            //
+            // as the span is disabled, no keys were recorded to it, leading check_fields_present0
+            // to find an error.
+
+            let missing = check_fields_present0(extractors).unwrap_err();
+            assert_missing(missing, vec![extractors[0]]);
+        } else {
+            // when the span is enabled, it is because some other test is running at the same time,
+            // and that tests registry has filters which are interested in our above span.
+            //
+            // because the span is now enabled, all keys will be found for it. the
+            // tracing_error::SpanTrace does not consider layer filters during the span hierarchy
+            // walk (SpanTrace::with_spans), nor is the SpanTrace::status a reliable indicator in
+            // this test-induced issue.
+
+            let res = check_fields_present0(extractors);
+            assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
+        }
     }
 }

pageserver/Cargo.toml

@@ -35,6 +35,8 @@ humantime-serde.workspace = true
 hyper.workspace = true
 itertools.workspace = true
 nix.workspace = true
+# hack to get the number of worker threads tokio uses
+num_cpus = { version = "1.15" }
 num-traits.workspace = true
 once_cell.workspace = true
 pin-project-lite.workspace = true
@@ -82,6 +84,7 @@ strum_macros.workspace = true
 criterion.workspace = true
 hex-literal.workspace = true
 tempfile.workspace = true
+tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time", "test-util"] }
 
 [[bench]]
 name = "bench_layer_map"

pageserver/ctl/Cargo.toml

@@ -13,6 +13,7 @@ clap = { workspace = true, features = ["string"] }
 git-version.workspace = true
 pageserver = { path = ".." }
 postgres_ffi.workspace = true
+tokio.workspace = true
 utils.workspace = true
 svg_fmt.workspace = true
 workspace_hack.workspace = true

pageserver/ctl/src/draw_timeline_dir.rs

@@ -7,10 +7,10 @@
 //! - The y axis represents LSN, growing upwards.
 //!
 //! Coordinates in both axis are compressed for better readability.
-//! (see https://medium.com/algorithms-digest/coordinate-compression-2fff95326fb)
+//! (see <https://medium.com/algorithms-digest/coordinate-compression-2fff95326fb>)
 //!
 //! Example use:
-//! ```
+//! ```bash
 //! $ ls test_output/test_pgbench\[neon-45-684\]/repo/tenants/$TENANT/timelines/$TIMELINE | \
 //! $   grep "__" | cargo run --release --bin pagectl draw-timeline-dir > out.svg
 //! $ firefox out.svg
@@ -20,7 +20,7 @@
 //! or from pageserver log files.
 //!
 //! TODO Consider shipping this as a grafana panel plugin:
-//!      https://grafana.com/tutorials/build-a-panel-plugin/
+//!      <https://grafana.com/tutorials/build-a-panel-plugin/>
 use anyhow::Result;
 use pageserver::repository::Key;
 use std::cmp::Ordering;

pageserver/ctl/src/layer_map_analyzer.rs

@@ -95,7 +95,7 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {
 }
 
 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
-fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
+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())?;
@@ -129,7 +129,7 @@ fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
     Ok(holes)
 }
 
-pub(crate) fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
+pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
     let storage_path = &cmd.path;
     let max_holes = cmd.max_holes.unwrap_or(DEFAULT_MAX_HOLES);
 
@@ -160,7 +160,7 @@ pub(crate) fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
                     parse_filename(&layer.file_name().into_string().unwrap())
                 {
                     if layer_file.is_delta {
-                        layer_file.holes = get_holes(&layer.path(), max_holes)?;
+                        layer_file.holes = get_holes(&layer.path(), max_holes).await?;
                         n_deltas += 1;
                     }
                     layers.push(layer_file);

pageserver/ctl/src/layers.rs

@@ -43,8 +43,7 @@ pub(crate) enum LayerCmd {
     },
 }
 
-fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
-    use pageserver::tenant::blob_io::BlobCursor;
+async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
     use pageserver::tenant::block_io::BlockReader;
 
     let path = path.as_ref();
@@ -69,7 +68,7 @@ fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
             true
         },
     )?;
-    let mut cursor = BlockCursor::new(&file);
+    let cursor = BlockCursor::new(&file);
     for (k, v) in all {
         let value = cursor.read_blob(v.pos())?;
         println!("key:{} value_len:{}", k, value.len());
@@ -78,7 +77,7 @@ fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
     Ok(())
 }
 
-pub(crate) fn main(cmd: &LayerCmd) -> Result<()> {
+pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
     match cmd {
         LayerCmd::List { path } => {
             for tenant in fs::read_dir(path.join("tenants"))? {
@@ -153,7 +152,7 @@ pub(crate) fn main(cmd: &LayerCmd) -> Result<()> {
                         );
 
                         if layer_file.is_delta {
-                            read_delta_file(layer.path())?;
+                            read_delta_file(layer.path()).await?;
                         } else {
                             anyhow::bail!("not supported yet :(");
                         }

pageserver/ctl/src/main.rs

@@ -72,12 +72,13 @@ struct AnalyzeLayerMapCmd {
     max_holes: Option<usize>,
 }
 
-fn main() -> anyhow::Result<()> {
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
     let cli = CliOpts::parse();
 
     match cli.command {
         Commands::Layer(cmd) => {
-            layers::main(&cmd)?;
+            layers::main(&cmd).await?;
         }
         Commands::Metadata(cmd) => {
             handle_metadata(&cmd)?;
@@ -86,7 +87,7 @@ fn main() -> anyhow::Result<()> {
             draw_timeline_dir::main()?;
         }
         Commands::AnalyzeLayerMap(cmd) => {
-            layer_map_analyzer::main(&cmd)?;
+            layer_map_analyzer::main(&cmd).await?;
         }
         Commands::PrintLayerFile(cmd) => {
             if let Err(e) = read_pg_control_file(&cmd.path) {
@@ -94,7 +95,7 @@ fn main() -> anyhow::Result<()> {
                     "Failed to read input file as a pg control one: {e:#}\n\
                     Attempting to read it as layer file"
                 );
-                print_layerfile(&cmd.path)?;
+                print_layerfile(&cmd.path).await?;
             }
         }
     };
@@ -113,12 +114,12 @@ fn read_pg_control_file(control_file_path: &Path) -> anyhow::Result<()> {
     Ok(())
 }
 
-fn print_layerfile(path: &Path) -> anyhow::Result<()> {
+async fn print_layerfile(path: &Path) -> anyhow::Result<()> {
     // Basic initialization of things that don't change after startup
     virtual_file::init(10);
     page_cache::init(100);
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
-    dump_layerfile_from_path(path, true, &ctx)
+    dump_layerfile_from_path(path, true, &ctx).await
 }
 
 fn handle_metadata(

pageserver/src/basebackup.rs

@@ -19,12 +19,6 @@ use tokio::io;
 use tokio::io::AsyncWrite;
 use tracing::*;
 
-/// NB: This relies on a modified version of tokio_tar that does *not* write the
-/// end-of-archive marker (1024 zero bytes), when the Builder struct is dropped
-/// without explicitly calling 'finish' or 'into_inner'!
-///
-/// See https://github.com/neondatabase/tokio-tar/pull/1
-///
 use tokio_tar::{Builder, EntryType, Header};
 
 use crate::context::RequestContext;

pageserver/src/bin/pageserver.rs

@@ -396,8 +396,8 @@ fn start_pageserver(
 
             let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
 
-            let init_sizes_done = tokio::select! {
-                _ = &mut init_sizes_done => {
+            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
+                Ok(_) => {
                     let now = std::time::Instant::now();
                     tracing::info!(
                         from_init_done_millis = (now - init_done).as_millis(),
@@ -406,7 +406,7 @@ fn start_pageserver(
                     );
                     None
                 }
-                _ = tokio::time::sleep(timeout) => {
+                Err(_) => {
                     tracing::info!(
                         timeout_millis = timeout.as_millis(),
                         "Initial logical size timeout elapsed; starting background jobs"

pageserver/src/config.rs

@@ -33,7 +33,8 @@ use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{TENANT_ATTACHING_MARKER_FILENAME, TIMELINES_SEGMENT_NAME};
 use crate::{
-    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TIMELINE_UNINIT_MARK_SUFFIX,
+    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TIMELINE_DELETE_MARK_SUFFIX,
+    TIMELINE_UNINIT_MARK_SUFFIX,
 };
 
 pub mod defaults {
@@ -171,11 +172,13 @@ pub struct PageServerConf {
 
     pub log_format: LogFormat,
 
-    /// Number of concurrent [`Tenant::gather_size_inputs`] allowed.
+    /// Number of concurrent [`Tenant::gather_size_inputs`](crate::tenant::Tenant::gather_size_inputs) allowed.
     pub concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore,
     /// Limit of concurrent [`Tenant::gather_size_inputs`] issued by module `eviction_task`.
     /// The number of permits is the same as `concurrent_tenant_size_logical_size_queries`.
     /// See the comment in `eviction_task` for details.
+    ///
+    /// [`Tenant::gather_size_inputs`]: crate::tenant::Tenant::gather_size_inputs
     pub eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore,
 
     // How often to collect metrics and send them to the metrics endpoint.
@@ -599,6 +602,17 @@ impl PageServerConf {
         )
     }
 
+    pub fn timeline_delete_mark_file_path(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> PathBuf {
+        path_with_suffix_extension(
+            self.timeline_path(&tenant_id, &timeline_id),
+            TIMELINE_DELETE_MARK_SUFFIX,
+        )
+    }
+
     pub fn traces_path(&self) -> PathBuf {
         self.workdir.join("traces")
     }
@@ -993,6 +1007,8 @@ impl ConfigurableSemaphore {
     /// Require a non-zero initial permits, because using permits == 0 is a crude way to disable a
     /// feature such as [`Tenant::gather_size_inputs`]. Otherwise any semaphore using future will
     /// behave like [`futures::future::pending`], just waiting until new permits are added.
+    ///
+    /// [`Tenant::gather_size_inputs`]: crate::tenant::Tenant::gather_size_inputs
     pub fn new(initial_permits: NonZeroUsize) -> Self {
         ConfigurableSemaphore {
             initial_permits,

pageserver/src/consumption_metrics.rs

@@ -7,27 +7,23 @@ use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
 use crate::tenant::{mgr, LogicalSizeCalculationCause};
 use anyhow;
-use chrono::Utc;
+use chrono::{DateTime, Utc};
 use consumption_metrics::{idempotency_key, Event, EventChunk, EventType, CHUNK_SIZE};
 use pageserver_api::models::TenantState;
 use reqwest::Url;
 use serde::Serialize;
 use serde_with::{serde_as, DisplayFromStr};
 use std::collections::HashMap;
-use std::time::Duration;
+use std::sync::Arc;
+use std::time::{Duration, SystemTime};
 use tracing::*;
 use utils::id::{NodeId, TenantId, TimelineId};
-
-const WRITTEN_SIZE: &str = "written_size";
-const SYNTHETIC_STORAGE_SIZE: &str = "synthetic_storage_size";
-const RESIDENT_SIZE: &str = "resident_size";
-const REMOTE_STORAGE_SIZE: &str = "remote_storage_size";
-const TIMELINE_LOGICAL_SIZE: &str = "timeline_logical_size";
+use utils::lsn::Lsn;
 
 const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
 
 #[serde_as]
-#[derive(Serialize, Debug)]
+#[derive(Serialize, Debug, Clone, Copy)]
 struct Ids {
     #[serde_as(as = "DisplayFromStr")]
     tenant_id: TenantId,
@@ -38,10 +34,142 @@ struct Ids {
 
 /// Key that uniquely identifies the object, this metric describes.
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct PageserverConsumptionMetricsKey {
-    pub tenant_id: TenantId,
-    pub timeline_id: Option<TimelineId>,
-    pub metric: &'static str,
+struct MetricsKey {
+    tenant_id: TenantId,
+    timeline_id: Option<TimelineId>,
+    metric: &'static str,
+}
+
+impl MetricsKey {
+    const fn absolute_values(self) -> AbsoluteValueFactory {
+        AbsoluteValueFactory(self)
+    }
+    const fn incremental_values(self) -> IncrementalValueFactory {
+        IncrementalValueFactory(self)
+    }
+}
+
+/// Helper type which each individual metric kind can return to produce only absolute values.
+struct AbsoluteValueFactory(MetricsKey);
+
+impl AbsoluteValueFactory {
+    fn at(self, time: DateTime<Utc>, val: u64) -> (MetricsKey, (EventType, u64)) {
+        let key = self.0;
+        (key, (EventType::Absolute { time }, val))
+    }
+}
+
+/// Helper type which each individual metric kind can return to produce only incremental values.
+struct IncrementalValueFactory(MetricsKey);
+
+impl IncrementalValueFactory {
+    #[allow(clippy::wrong_self_convention)]
+    fn from_previous_up_to(
+        self,
+        prev_end: DateTime<Utc>,
+        up_to: DateTime<Utc>,
+        val: u64,
+    ) -> (MetricsKey, (EventType, u64)) {
+        let key = self.0;
+        // cannot assert prev_end < up_to because these are realtime clock based
+        (
+            key,
+            (
+                EventType::Incremental {
+                    start_time: prev_end,
+                    stop_time: up_to,
+                },
+                val,
+            ),
+        )
+    }
+
+    fn key(&self) -> &MetricsKey {
+        &self.0
+    }
+}
+
+// the static part of a MetricsKey
+impl MetricsKey {
+    /// Absolute value of [`Timeline::get_last_record_lsn`].
+    ///
+    /// [`Timeline::get_last_record_lsn`]: crate::tenant::Timeline::get_last_record_lsn
+    const fn written_size(tenant_id: TenantId, timeline_id: TimelineId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            metric: "written_size",
+        }
+        .absolute_values()
+    }
+
+    /// Values will be the difference of the latest [`MetricsKey::written_size`] to what we
+    /// previously sent, starting from the previously sent incremental time range ending at the
+    /// latest absolute measurement.
+    const fn written_size_delta(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> IncrementalValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            // the name here is correctly about data not size, because that is what is wanted by
+            // downstream pipeline
+            metric: "written_data_bytes_delta",
+        }
+        .incremental_values()
+    }
+
+    /// Exact [`Timeline::get_current_logical_size`].
+    ///
+    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
+    const fn timeline_logical_size(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: Some(timeline_id),
+            metric: "timeline_logical_size",
+        }
+        .absolute_values()
+    }
+
+    /// [`Tenant::remote_size`]
+    ///
+    /// [`Tenant::remote_size`]: crate::tenant::Tenant::remote_size
+    const fn remote_storage_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: "remote_storage_size",
+        }
+        .absolute_values()
+    }
+
+    /// Sum of [`Timeline::resident_physical_size`] for each `Tenant`.
+    ///
+    /// [`Timeline::resident_physical_size`]: crate::tenant::Timeline::resident_physical_size
+    const fn resident_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: "resident_size",
+        }
+        .absolute_values()
+    }
+
+    /// [`Tenant::cached_synthetic_size`] as refreshed by [`calculate_synthetic_size_worker`].
+    ///
+    /// [`Tenant::cached_synthetic_size`]: crate::tenant::Tenant::cached_synthetic_size
+    const fn synthetic_size(tenant_id: TenantId) -> AbsoluteValueFactory {
+        MetricsKey {
+            tenant_id,
+            timeline_id: None,
+            metric: "synthetic_storage_size",
+        }
+        .absolute_values()
+    }
 }
 
 /// Main thread that serves metrics collection
@@ -79,7 +207,7 @@ pub async fn collect_metrics(
         .timeout(DEFAULT_HTTP_REPORTING_TIMEOUT)
         .build()
         .expect("Failed to create http client with timeout");
-    let mut cached_metrics: HashMap<PageserverConsumptionMetricsKey, u64> = HashMap::new();
+    let mut cached_metrics = HashMap::new();
     let mut prev_iteration_time: std::time::Instant = std::time::Instant::now();
 
     loop {
@@ -119,15 +247,15 @@ pub async fn collect_metrics(
 ///
 /// TODO
 /// - refactor this function (chunking+sending part) to reuse it in proxy module;
-pub async fn collect_metrics_iteration(
+async fn collect_metrics_iteration(
     client: &reqwest::Client,
-    cached_metrics: &mut HashMap<PageserverConsumptionMetricsKey, u64>,
+    cached_metrics: &mut HashMap<MetricsKey, (EventType, u64)>,
     metric_collection_endpoint: &reqwest::Url,
     node_id: NodeId,
     ctx: &RequestContext,
     send_cached: bool,
 ) {
-    let mut current_metrics: Vec<(PageserverConsumptionMetricsKey, u64)> = Vec::new();
+    let mut current_metrics: Vec<(MetricsKey, (EventType, u64))> = Vec::new();
     trace!(
         "starting collect_metrics_iteration. metric_collection_endpoint: {}",
         metric_collection_endpoint
@@ -161,99 +289,65 @@ pub async fn collect_metrics_iteration(
         let mut tenant_resident_size = 0;
 
         // iterate through list of timelines in tenant
-        for timeline in tenant.list_timelines().iter() {
+        for timeline in tenant.list_timelines() {
             // collect per-timeline metrics only for active timelines
-            if timeline.is_active() {
-                let timeline_written_size = u64::from(timeline.get_last_record_lsn());
 
-                current_metrics.push((
-                    PageserverConsumptionMetricsKey {
+            let timeline_id = timeline.timeline_id;
+
+            match TimelineSnapshot::collect(&timeline, ctx) {
+                Ok(Some(snap)) => {
+                    snap.to_metrics(
                         tenant_id,
-                        timeline_id: Some(timeline.timeline_id),
-                        metric: WRITTEN_SIZE,
-                    },
-                    timeline_written_size,
-                ));
-
-                let span = info_span!("collect_metrics_iteration", tenant_id = %timeline.tenant_id, timeline_id = %timeline.timeline_id);
-                match span.in_scope(|| timeline.get_current_logical_size(ctx)) {
-                    // Only send timeline logical size when it is fully calculated.
-                    Ok((size, is_exact)) if is_exact => {
-                        current_metrics.push((
-                            PageserverConsumptionMetricsKey {
-                                tenant_id,
-                                timeline_id: Some(timeline.timeline_id),
-                                metric: TIMELINE_LOGICAL_SIZE,
-                            },
-                            size,
-                        ));
-                    }
-                    Ok((_, _)) => {}
-                    Err(err) => {
-                        error!(
-                            "failed to get current logical size for timeline {}: {err:?}",
-                            timeline.timeline_id
-                        );
-                        continue;
-                    }
-                };
+                        timeline_id,
+                        Utc::now(),
+                        &mut current_metrics,
+                        cached_metrics,
+                    );
+                }
+                Ok(None) => {}
+                Err(e) => {
+                    error!(
+                        "failed to get metrics values for tenant {tenant_id} timeline {}: {e:#?}",
+                        timeline.timeline_id
+                    );
+                    continue;
+                }
             }
 
-            let timeline_resident_size = timeline.get_resident_physical_size();
-            tenant_resident_size += timeline_resident_size;
+            tenant_resident_size += timeline.resident_physical_size();
         }
 
-        match tenant.get_remote_size().await {
-            Ok(tenant_remote_size) => {
-                current_metrics.push((
-                    PageserverConsumptionMetricsKey {
-                        tenant_id,
-                        timeline_id: None,
-                        metric: REMOTE_STORAGE_SIZE,
-                    },
-                    tenant_remote_size,
-                ));
-            }
-            Err(err) => {
-                error!(
-                    "failed to get remote size for tenant {}: {err:?}",
-                    tenant_id
-                );
-            }
-        }
+        current_metrics
+            .push(MetricsKey::remote_storage_size(tenant_id).at(Utc::now(), tenant.remote_size()));
 
-        current_metrics.push((
-            PageserverConsumptionMetricsKey {
-                tenant_id,
-                timeline_id: None,
-                metric: RESIDENT_SIZE,
-            },
-            tenant_resident_size,
-        ));
+        current_metrics
+            .push(MetricsKey::resident_size(tenant_id).at(Utc::now(), tenant_resident_size));
 
         // Note that this metric is calculated in a separate bgworker
         // Here we only use cached value, which may lag behind the real latest one
-        let tenant_synthetic_size = tenant.get_cached_synthetic_size();
+        let synthetic_size = tenant.cached_synthetic_size();
 
-        if tenant_synthetic_size != 0 {
+        if synthetic_size != 0 {
             // only send non-zeroes because otherwise these show up as errors in logs
-            current_metrics.push((
-                PageserverConsumptionMetricsKey {
-                    tenant_id,
-                    timeline_id: None,
-                    metric: SYNTHETIC_STORAGE_SIZE,
-                },
-                tenant_synthetic_size,
-            ));
+            current_metrics
+                .push(MetricsKey::synthetic_size(tenant_id).at(Utc::now(), synthetic_size));
         }
     }
 
     // Filter metrics, unless we want to send all metrics, including cached ones.
     // See: https://github.com/neondatabase/neon/issues/3485
     if !send_cached {
-        current_metrics.retain(|(curr_key, curr_val)| match cached_metrics.get(curr_key) {
-            Some(val) => val != curr_val,
-            None => true,
+        current_metrics.retain(|(curr_key, (kind, curr_val))| {
+            if kind.is_incremental() {
+                // incremental values (currently only written_size_delta) should not get any cache
+                // deduplication because they will be used by upstream for "is still alive."
+                true
+            } else {
+                match cached_metrics.get(curr_key) {
+                    Some((_, val)) => val != curr_val,
+                    None => true,
+                }
+            }
         });
     }
 
@@ -268,14 +362,16 @@ pub async fn collect_metrics_iteration(
 
     let mut chunk_to_send: Vec<Event<Ids>> = Vec::with_capacity(CHUNK_SIZE);
 
+    let node_id = node_id.to_string();
+
     for chunk in chunks {
         chunk_to_send.clear();
 
         // enrich metrics with type,timestamp and idempotency key before sending
-        chunk_to_send.extend(chunk.iter().map(|(curr_key, curr_val)| Event {
-            kind: EventType::Absolute { time: Utc::now() },
+        chunk_to_send.extend(chunk.iter().map(|(curr_key, (when, curr_val))| Event {
+            kind: *when,
             metric: curr_key.metric,
-            idempotency_key: idempotency_key(node_id.to_string()),
+            idempotency_key: idempotency_key(&node_id),
             value: *curr_val,
             extra: Ids {
                 tenant_id: curr_key.tenant_id,
@@ -283,17 +379,14 @@ pub async fn collect_metrics_iteration(
             },
         }));
 
-        let chunk_json = serde_json::value::to_raw_value(&EventChunk {
-            events: &chunk_to_send,
-        })
-        .expect("PageserverConsumptionMetric should not fail serialization");
-
         const MAX_RETRIES: u32 = 3;
 
         for attempt in 0..MAX_RETRIES {
             let res = client
                 .post(metric_collection_endpoint.clone())
-                .json(&chunk_json)
+                .json(&EventChunk {
+                    events: (&chunk_to_send).into(),
+                })
                 .send()
                 .await;
 
@@ -329,6 +422,130 @@ pub async fn collect_metrics_iteration(
     }
 }
 
+/// Internal type to make timeline metric production testable.
+///
+/// As this value type contains all of the information needed from a timeline to produce the
+/// metrics, it can easily be created with different values in test.
+struct TimelineSnapshot {
+    loaded_at: (Lsn, SystemTime),
+    last_record_lsn: Lsn,
+    current_exact_logical_size: Option<u64>,
+}
+
+impl TimelineSnapshot {
+    /// Collect the metrics from an actual timeline.
+    ///
+    /// Fails currently only when [`Timeline::get_current_logical_size`] fails.
+    ///
+    /// [`Timeline::get_current_logical_size`]: crate::tenant::Timeline::get_current_logical_size
+    fn collect(
+        t: &Arc<crate::tenant::Timeline>,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Option<Self>> {
+        use anyhow::Context;
+
+        if !t.is_active() {
+            // no collection for broken or stopping needed, we will still keep the cached values
+            // though at the caller.
+            Ok(None)
+        } else {
+            let loaded_at = t.loaded_at;
+            let last_record_lsn = t.get_last_record_lsn();
+
+            let current_exact_logical_size = {
+                let span = info_span!("collect_metrics_iteration", tenant_id = %t.tenant_id, timeline_id = %t.timeline_id);
+                let res = span
+                    .in_scope(|| t.get_current_logical_size(ctx))
+                    .context("get_current_logical_size");
+                match res? {
+                    // Only send timeline logical size when it is fully calculated.
+                    (size, is_exact) if is_exact => Some(size),
+                    (_, _) => None,
+                }
+            };
+
+            Ok(Some(TimelineSnapshot {
+                loaded_at,
+                last_record_lsn,
+                current_exact_logical_size,
+            }))
+        }
+    }
+
+    /// Produce the timeline consumption metrics into the `metrics` argument.
+    fn to_metrics(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        now: DateTime<Utc>,
+        metrics: &mut Vec<(MetricsKey, (EventType, u64))>,
+        cache: &HashMap<MetricsKey, (EventType, u64)>,
+    ) {
+        let timeline_written_size = u64::from(self.last_record_lsn);
+
+        let (key, written_size_now) =
+            MetricsKey::written_size(tenant_id, timeline_id).at(now, timeline_written_size);
+
+        // last_record_lsn can only go up, right now at least, TODO: #2592 or related
+        // features might change this.
+
+        let written_size_delta_key = MetricsKey::written_size_delta(tenant_id, timeline_id);
+
+        // use this when available, because in a stream of incremental values, it will be
+        // accurate where as when last_record_lsn stops moving, we will only cache the last
+        // one of those.
+        let last_stop_time = cache
+            .get(written_size_delta_key.key())
+            .map(|(until, _val)| {
+                until
+                    .incremental_timerange()
+                    .expect("never create EventType::Absolute for written_size_delta")
+                    .end
+            });
+
+        // by default, use the last sent written_size as the basis for
+        // calculating the delta. if we don't yet have one, use the load time value.
+        let prev = cache
+            .get(&key)
+            .map(|(prev_at, prev)| {
+                // use the prev time from our last incremental update, or default to latest
+                // absolute update on the first round.
+                let prev_at = prev_at
+                    .absolute_time()
+                    .expect("never create EventType::Incremental for written_size");
+                let prev_at = last_stop_time.unwrap_or(prev_at);
+                (*prev_at, *prev)
+            })
+            .unwrap_or_else(|| {
+                // if we don't have a previous point of comparison, compare to the load time
+                // lsn.
+                let (disk_consistent_lsn, loaded_at) = &self.loaded_at;
+                (DateTime::from(*loaded_at), disk_consistent_lsn.0)
+            });
+
+        // written_size_bytes_delta
+        metrics.extend(
+            if let Some(delta) = written_size_now.1.checked_sub(prev.1) {
+                let up_to = written_size_now
+                    .0
+                    .absolute_time()
+                    .expect("never create EventType::Incremental for written_size");
+                let key_value = written_size_delta_key.from_previous_up_to(prev.0, *up_to, delta);
+                Some(key_value)
+            } else {
+                None
+            },
+        );
+
+        // written_size
+        metrics.push((key, written_size_now));
+
+        if let Some(size) = self.current_exact_logical_size {
+            metrics.push(MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, size));
+        }
+    }
+}
+
 /// Caclculate synthetic size for each active tenant
 pub async fn calculate_synthetic_size_worker(
     synthetic_size_calculation_interval: Duration,
@@ -343,7 +560,7 @@ pub async fn calculate_synthetic_size_worker(
             _ = task_mgr::shutdown_watcher() => {
                 return Ok(());
             },
-        tick_at = ticker.tick() => {
+            tick_at = ticker.tick() => {
 
                 let tenants = match mgr::list_tenants().await {
                     Ok(tenants) => tenants,
@@ -379,3 +596,149 @@ pub async fn calculate_synthetic_size_worker(
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use std::time::SystemTime;
+    use utils::{
+        id::{TenantId, TimelineId},
+        lsn::Lsn,
+    };
+
+    use crate::consumption_metrics::MetricsKey;
+
+    use super::TimelineSnapshot;
+    use chrono::{DateTime, Utc};
+
+    #[test]
+    fn startup_collected_timeline_metrics_before_advancing() {
+        let tenant_id = TenantId::generate();
+        let timeline_id = TimelineId::generate();
+
+        let mut metrics = Vec::new();
+        let cache = HashMap::new();
+
+        let initdb_lsn = Lsn(0x10000);
+        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+        let snap = TimelineSnapshot {
+            loaded_at: (disk_consistent_lsn, SystemTime::now()),
+            last_record_lsn: disk_consistent_lsn,
+            current_exact_logical_size: Some(0x42000),
+        };
+
+        let now = DateTime::<Utc>::from(SystemTime::now());
+
+        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+        assert_eq!(
+            metrics,
+            &[
+                MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
+                    snap.loaded_at.1.into(),
+                    now,
+                    0
+                ),
+                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+            ]
+        );
+    }
+
+    #[test]
+    fn startup_collected_timeline_metrics_second_round() {
+        let tenant_id = TenantId::generate();
+        let timeline_id = TimelineId::generate();
+
+        let [now, before, init] = time_backwards();
+
+        let now = DateTime::<Utc>::from(now);
+        let before = DateTime::<Utc>::from(before);
+
+        let initdb_lsn = Lsn(0x10000);
+        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+        let mut metrics = Vec::new();
+        let cache = HashMap::from([
+            MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0)
+        ]);
+
+        let snap = TimelineSnapshot {
+            loaded_at: (disk_consistent_lsn, init),
+            last_record_lsn: disk_consistent_lsn,
+            current_exact_logical_size: Some(0x42000),
+        };
+
+        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+        assert_eq!(
+            metrics,
+            &[
+                MetricsKey::written_size_delta(tenant_id, timeline_id)
+                    .from_previous_up_to(before, now, 0),
+                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+            ]
+        );
+    }
+
+    #[test]
+    fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
+        let tenant_id = TenantId::generate();
+        let timeline_id = TimelineId::generate();
+
+        let [now, just_before, before, init] = time_backwards();
+
+        let now = DateTime::<Utc>::from(now);
+        let just_before = DateTime::<Utc>::from(just_before);
+        let before = DateTime::<Utc>::from(before);
+
+        let initdb_lsn = Lsn(0x10000);
+        let disk_consistent_lsn = Lsn(initdb_lsn.0 * 2);
+
+        let mut metrics = Vec::new();
+        let cache = HashMap::from([
+            // at t=before was the last time the last_record_lsn changed
+            MetricsKey::written_size(tenant_id, timeline_id).at(before, disk_consistent_lsn.0),
+            // end time of this event is used for the next ones
+            MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
+                before,
+                just_before,
+                0,
+            ),
+        ]);
+
+        let snap = TimelineSnapshot {
+            loaded_at: (disk_consistent_lsn, init),
+            last_record_lsn: disk_consistent_lsn,
+            current_exact_logical_size: Some(0x42000),
+        };
+
+        snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
+
+        assert_eq!(
+            metrics,
+            &[
+                MetricsKey::written_size_delta(tenant_id, timeline_id).from_previous_up_to(
+                    just_before,
+                    now,
+                    0
+                ),
+                MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
+                MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
+            ]
+        );
+    }
+
+    fn time_backwards<const N: usize>() -> [std::time::SystemTime; N] {
+        let mut times = [std::time::SystemTime::UNIX_EPOCH; N];
+        times[0] = std::time::SystemTime::now();
+        for behind in 1..N {
+            times[behind] = times[0] - std::time::Duration::from_secs(behind as u64);
+        }
+
+        times
+    }
+}

pageserver/src/context.rs

@@ -179,6 +179,9 @@ impl RequestContext {
     /// a context and you are unwilling to change all callers to provide one.
     ///
     /// Before we add cancellation, we should get rid of this method.
+    ///
+    /// [`attached_child`]: Self::attached_child
+    /// [`detached_child`]: Self::detached_child
     pub fn todo_child(task_kind: TaskKind, download_behavior: DownloadBehavior) -> Self {
         Self::new(task_kind, download_behavior)
     }

pageserver/src/disk_usage_eviction_task.rs

@@ -60,7 +60,7 @@ use utils::serde_percent::Percent;
 use crate::{
     config::PageServerConf,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
-    tenant::{self, storage_layer::PersistentLayer, Timeline},
+    tenant::{self, storage_layer::PersistentLayer, timeline::EvictionError, Timeline},
 };
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -166,11 +166,11 @@ async fn disk_usage_eviction_task(
         .await;
 
         let sleep_until = start + task_config.period;
-        tokio::select! {
-            _ = tokio::time::sleep_until(sleep_until) => {},
-            _ = cancel.cancelled() => {
-                break
-            }
+        if tokio::time::timeout_at(sleep_until, cancel.cancelled())
+            .await
+            .is_ok()
+        {
+            break;
         }
     }
 }
@@ -390,13 +390,22 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                     assert_eq!(results.len(), batch.len());
                     for (result, layer) in results.into_iter().zip(batch.iter()) {
                         match result {
-                            Some(Ok(true)) => {
+                            Some(Ok(())) => {
                                 usage_assumed.add_available_bytes(layer.file_size());
                             }
-                            Some(Ok(false)) => {
-                                // this is:
-                                // - Replacement::{NotFound, Unexpected}
-                                // - it cannot be is_remote_layer, filtered already
+                            Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
+                                unreachable!("get_local_layers_for_disk_usage_eviction finds only local layers")
+                            }
+                            Some(Err(EvictionError::FileNotFound)) => {
+                                evictions_failed.file_sizes += layer.file_size();
+                                evictions_failed.count += 1;
+                            }
+                            Some(Err(
+                                e @ EvictionError::LayerNotFound(_)
+                                | e @ EvictionError::StatFailed(_),
+                            )) => {
+                                let e = utils::error::report_compact_sources(&e);
+                                warn!(%layer, "failed to evict layer: {e}");
                                 evictions_failed.file_sizes += layer.file_size();
                                 evictions_failed.count += 1;
                             }
@@ -404,10 +413,6 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                                 assert!(cancel.is_cancelled());
                                 return;
                             }
-                            Some(Err(e)) => {
-                                // we really shouldn't be getting this, precondition failure
-                                error!("failed to evict layer: {:#}", e);
-                            }
                         }
                     }
                 }
@@ -540,12 +545,12 @@ async fn collect_eviction_candidates(
         // We could be better here, e.g., sum of all L0 layers + most recent L1 layer.
         // That's what's typically used by the various background loops.
         //
-        // The default can be overriden with a fixed value in the tenant conf.
+        // The default can be overridden with a fixed value in the tenant conf.
         // A default override can be put in the default tenant conf in the pageserver.toml.
         let min_resident_size = if let Some(s) = tenant.get_min_resident_size_override() {
             debug!(
                 tenant_id=%tenant.tenant_id(),
-                overriden_size=s,
+                overridden_size=s,
                 "using overridden min resident size for tenant"
             );
             s

pageserver/src/http/routes.rs

@@ -994,31 +994,29 @@ async fn timeline_gc_handler(
 // Run compaction immediately on given timeline.
 async fn timeline_compact_handler(
     request: Request<Body>,
-    _cancel: CancellationToken,
+    cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_id))?;
 
-    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let result_receiver = mgr::immediate_compact(tenant_id, timeline_id, &ctx)
-        .await
-        .context("spawn compaction task")
-        .map_err(ApiError::InternalServerError)?;
-
-    let result: anyhow::Result<()> = result_receiver
-        .await
-        .context("receive compaction result")
-        .map_err(ApiError::InternalServerError)?;
-    result.map_err(ApiError::InternalServerError)?;
-
-    json_response(StatusCode::OK, ())
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+        timeline
+            .compact(&cancel, &ctx)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+        json_response(StatusCode::OK, ())
+    }
+    .instrument(info_span!("manual_compaction", %tenant_id, %timeline_id))
+    .await
 }
 
 // Run checkpoint immediately on given timeline.
 async fn timeline_checkpoint_handler(
     request: Request<Body>,
-    _cancel: CancellationToken,
+    cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
@@ -1031,13 +1029,13 @@ async fn timeline_checkpoint_handler(
             .await
             .map_err(ApiError::InternalServerError)?;
         timeline
-            .compact(&ctx)
+            .compact(&cancel, &ctx)
             .await
             .map_err(ApiError::InternalServerError)?;
 
         json_response(StatusCode::OK, ())
     }
-    .instrument(info_span!("manual_checkpoint", tenant_id = %tenant_id, timeline_id = %timeline_id))
+    .instrument(info_span!("manual_checkpoint", %tenant_id, %timeline_id))
     .await
 }
 
@@ -1143,7 +1141,7 @@ async fn disk_usage_eviction_run(
     let Some(storage) = state.remote_storage.clone() else {
         return Err(ApiError::InternalServerError(anyhow::anyhow!(
             "remote storage not configured, cannot run eviction iteration"
-        )))
+        )));
     };
 
     let state = state.disk_usage_eviction_state.clone();

pageserver/src/lib.rs

@@ -109,6 +109,8 @@ pub const TEMP_FILE_SUFFIX: &str = "___temp";
 /// Full path: `tenants/<tenant_id>/timelines/<timeline_id>___uninit`.
 pub const TIMELINE_UNINIT_MARK_SUFFIX: &str = "___uninit";
 
+pub const TIMELINE_DELETE_MARK_SUFFIX: &str = "___delete";
+
 /// A marker file to prevent pageserver from loading a certain tenant on restart.
 /// Different from [`TIMELINE_UNINIT_MARK_SUFFIX`] due to semantics of the corresponding
 /// `ignore` management API command, that expects the ignored tenant to be properly loaded
@@ -123,15 +125,30 @@ pub fn is_temporary(path: &Path) -> bool {
     }
 }
 
-pub fn is_uninit_mark(path: &Path) -> bool {
+fn ends_with_suffix(path: &Path, suffix: &str) -> bool {
     match path.file_name() {
-        Some(name) => name
-            .to_string_lossy()
-            .ends_with(TIMELINE_UNINIT_MARK_SUFFIX),
+        Some(name) => name.to_string_lossy().ends_with(suffix),
         None => false,
     }
 }
 
+pub fn is_uninit_mark(path: &Path) -> bool {
+    ends_with_suffix(path, TIMELINE_UNINIT_MARK_SUFFIX)
+}
+
+pub fn is_delete_mark(path: &Path) -> bool {
+    ends_with_suffix(path, TIMELINE_DELETE_MARK_SUFFIX)
+}
+
+fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
+    if let Some(e) = e.io_error() {
+        if e.kind() == std::io::ErrorKind::NotFound {
+            return true;
+        }
+    }
+    false
+}
+
 /// During pageserver startup, we need to order operations not to exhaust tokio worker threads by
 /// blocking.
 ///

pageserver/src/metrics.rs

@@ -6,7 +6,6 @@ use metrics::{
     IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
-use pageserver_api::models::TenantState;
 use strum::VariantNames;
 use strum_macros::{EnumVariantNames, IntoStaticStr};
 use utils::id::{TenantId, TimelineId};
@@ -74,7 +73,7 @@ pub static STORAGE_TIME_COUNT_PER_TIMELINE: Lazy<IntCounterVec> = Lazy::new(|| {
 // Buckets for background operations like compaction, GC, size calculation
 const STORAGE_OP_BUCKETS: &[f64] = &[0.010, 0.100, 1.0, 10.0, 100.0, 1000.0];
 
-pub static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
+pub(crate) static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_storage_operations_seconds_global",
         "Time spent on storage operations",
@@ -84,18 +83,17 @@ pub static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-static READ_NUM_FS_LAYERS: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static READ_NUM_FS_LAYERS: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
         "pageserver_read_num_fs_layers",
         "Number of persistent layers accessed for processing a read request, including those in the cache",
-        &["tenant_id", "timeline_id"],
         vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 10.0, 20.0, 50.0, 100.0],
     )
     .expect("failed to define a metric")
 });
 
 // Metrics collected on operations on the storage repository.
-pub static RECONSTRUCT_TIME: Lazy<Histogram> = Lazy::new(|| {
+pub(crate) static RECONSTRUCT_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_getpage_reconstruct_seconds",
         "Time spent in reconstruct_value (reconstruct a page from deltas)",
@@ -104,7 +102,7 @@ pub static RECONSTRUCT_TIME: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_materialized_cache_hits_direct_total",
         "Number of cache hits from materialized page cache without redo",
@@ -112,17 +110,16 @@ pub static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-static GET_RECONSTRUCT_DATA_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static GET_RECONSTRUCT_DATA_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
         "pageserver_getpage_get_reconstruct_data_seconds",
         "Time spent in get_reconstruct_value_data",
-        &["tenant_id", "timeline_id"],
         CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
 
-pub static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static MATERIALIZED_PAGE_CACHE_HIT: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_materialized_cache_hits_total",
         "Number of cache hits from materialized page cache",
@@ -246,11 +243,10 @@ pub static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> = Lazy::new(|| PageCacheS
     },
 });
 
-static WAIT_LSN_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static WAIT_LSN_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
         "pageserver_wait_lsn_seconds",
         "Time spent waiting for WAL to arrive",
-        &["tenant_id", "timeline_id"],
         CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
@@ -284,7 +280,7 @@ static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_remote_ondemand_downloaded_layers_total",
         "Total on-demand downloaded layers"
@@ -292,7 +288,7 @@ pub static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
     .unwrap()
 });
 
-pub static REMOTE_ONDEMAND_DOWNLOADED_BYTES: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_BYTES: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_remote_ondemand_downloaded_bytes_total",
         "Total bytes of layers on-demand downloaded",
@@ -309,16 +305,29 @@ static CURRENT_LOGICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define current logical size metric")
 });
 
-pub static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
+pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_tenant_states_count",
         "Count of tenants per state",
-        &["tenant_id", "state"]
+        &["state"]
     )
     .expect("Failed to register pageserver_tenant_states_count metric")
 });
 
-pub static TENANT_SYNTHETIC_SIZE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
+/// A set of broken tenants.
+///
+/// These are expected to be so rare that a set is fine. Set as in a new timeseries per each broken
+/// tenant.
+pub(crate) static BROKEN_TENANTS_SET: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_broken_tenants_count",
+        "Set of broken tenants",
+        &["tenant_id"]
+    )
+    .expect("Failed to register pageserver_tenant_states_count metric")
+});
+
+pub(crate) static TENANT_SYNTHETIC_SIZE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_tenant_synthetic_cached_size_bytes",
         "Synthetic size of each tenant in bytes",
@@ -376,7 +385,7 @@ static EVICTIONS_WITH_LOW_RESIDENCE_DURATION: Lazy<IntCounterVec> = Lazy::new(||
     .expect("failed to define a metric")
 });
 
-pub static UNEXPECTED_ONDEMAND_DOWNLOADS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static UNEXPECTED_ONDEMAND_DOWNLOADS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_unexpected_ondemand_downloads_count",
         "Number of unexpected on-demand downloads. \
@@ -385,7 +394,7 @@ pub static UNEXPECTED_ONDEMAND_DOWNLOADS: Lazy<IntCounter> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-/// Each [`Timeline`]'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
+/// Each `Timeline`'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
 #[derive(Debug)]
 pub struct EvictionsWithLowResidenceDuration {
     data_source: &'static str,
@@ -499,23 +508,31 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
     30.000,   // 30000 ms
 ];
 
-const STORAGE_IO_TIME_OPERATIONS: &[&str] = &[
-    "open", "close", "read", "write", "seek", "fsync", "gc", "metadata",
-];
-
-const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];
-
-pub static STORAGE_IO_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+/// Tracks time taken by fs operations near VirtualFile.
+///
+/// Operations:
+/// - open ([`std::fs::OpenOptions::open`])
+/// - close (dropping [`std::fs::File`])
+/// - close-by-replace (close by replacement algorithm)
+/// - read (`read_at`)
+/// - write (`write_at`)
+/// - seek (modify internal position or file length query)
+/// - fsync ([`std::fs::File::sync_all`])
+/// - metadata ([`std::fs::File::metadata`])
+pub(crate) static STORAGE_IO_TIME: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_io_operations_seconds",
         "Time spent in IO operations",
-        &["operation", "tenant_id", "timeline_id"],
+        &["operation"],
         STORAGE_IO_TIME_BUCKETS.into()
     )
     .expect("failed to define a metric")
 });
 
-pub static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
+const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];
+
+// Needed for the https://neonprod.grafana.net/d/5uK9tHL4k/picking-tenant-for-relocation?orgId=1
+pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
     register_int_gauge_vec!(
         "pageserver_io_operations_bytes_total",
         "Total amount of bytes read/written in IO operations",
@@ -541,6 +558,17 @@ pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+// 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(|| {
+    // Go code uses milliseconds. Variable is called `computeStartupBuckets`
+    [
+        5, 10, 20, 30, 50, 70, 100, 120, 150, 200, 250, 300, 350, 400, 450, 500, 600, 800, 1000,
+        1500, 2000, 2500, 3000, 5000, 10000, 20000, 40000, 60000,
+    ]
+    .map(|ms| (ms as f64) / 1000.0)
+});
+
 pub struct BasebackupQueryTime(HistogramVec);
 pub static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
     BasebackupQueryTime({
@@ -548,7 +576,7 @@ pub static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
             "pageserver_basebackup_query_seconds",
             "Histogram of basebackup queries durations, by result type",
             &["result"],
-            CRITICAL_OP_BUCKETS.into(),
+            COMPUTE_STARTUP_BUCKETS.to_vec(),
         )
         .expect("failed to define a metric")
     })
@@ -594,7 +622,7 @@ static REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST: Lazy<HistogramVec> = Lazy::new
          at a given instant. It gives you a better idea of the queue depth \
          than plotting the gauge directly, since operations may complete faster \
          than the sampling interval.",
-        &["tenant_id", "timeline_id", "file_kind", "op_kind"],
+        &["file_kind", "op_kind"],
         // The calls_unfinished gauge is an integer gauge, hence we have integer buckets.
         vec![0.0, 1.0, 2.0, 4.0, 6.0, 8.0, 10.0, 15.0, 20.0, 40.0, 60.0, 80.0, 100.0, 500.0],
     )
@@ -651,18 +679,18 @@ impl RemoteOpFileKind {
     }
 }
 
-pub static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+pub(crate) static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_remote_operation_seconds",
         "Time spent on remote storage operations. \
         Grouped by tenant, timeline, operation_kind and status. \
         Does not account for time spent waiting in remote timeline client's queues.",
-        &["tenant_id", "timeline_id", "file_kind", "op_kind", "status"]
+        &["file_kind", "op_kind", "status"]
     )
     .expect("failed to define a metric")
 });
 
-pub static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
+pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_tenant_task_events",
         "Number of task start/stop/fail events.",
@@ -671,7 +699,7 @@ pub static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("Failed to register tenant_task_events metric")
 });
 
-pub static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
+pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_background_loop_period_overrun_count",
         "Incremented whenever warn_when_period_overrun() logs a warning.",
@@ -682,7 +710,7 @@ pub static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new
 
 // walreceiver metrics
 
-pub static WALRECEIVER_STARTED_CONNECTIONS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static WALRECEIVER_STARTED_CONNECTIONS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_walreceiver_started_connections_total",
         "Number of started walreceiver connections"
@@ -690,7 +718,7 @@ pub static WALRECEIVER_STARTED_CONNECTIONS: Lazy<IntCounter> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WALRECEIVER_ACTIVE_MANAGERS: Lazy<IntGauge> = Lazy::new(|| {
+pub(crate) static WALRECEIVER_ACTIVE_MANAGERS: Lazy<IntGauge> = Lazy::new(|| {
     register_int_gauge!(
         "pageserver_walreceiver_active_managers",
         "Number of active walreceiver managers"
@@ -698,7 +726,7 @@ pub static WALRECEIVER_ACTIVE_MANAGERS: Lazy<IntGauge> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WALRECEIVER_SWITCHES: Lazy<IntCounterVec> = Lazy::new(|| {
+pub(crate) static WALRECEIVER_SWITCHES: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_walreceiver_switches_total",
         "Number of walreceiver manager change_connection calls",
@@ -707,7 +735,7 @@ pub static WALRECEIVER_SWITCHES: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WALRECEIVER_BROKER_UPDATES: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static WALRECEIVER_BROKER_UPDATES: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_walreceiver_broker_updates_total",
         "Number of received broker updates in walreceiver"
@@ -715,7 +743,7 @@ pub static WALRECEIVER_BROKER_UPDATES: Lazy<IntCounter> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WALRECEIVER_CANDIDATES_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
+pub(crate) static WALRECEIVER_CANDIDATES_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_walreceiver_candidates_events_total",
         "Number of walreceiver candidate events",
@@ -724,10 +752,10 @@ pub static WALRECEIVER_CANDIDATES_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WALRECEIVER_CANDIDATES_ADDED: Lazy<IntCounter> =
+pub(crate) static WALRECEIVER_CANDIDATES_ADDED: Lazy<IntCounter> =
     Lazy::new(|| WALRECEIVER_CANDIDATES_EVENTS.with_label_values(&["add"]));
 
-pub static WALRECEIVER_CANDIDATES_REMOVED: Lazy<IntCounter> =
+pub(crate) static WALRECEIVER_CANDIDATES_REMOVED: Lazy<IntCounter> =
     Lazy::new(|| WALRECEIVER_CANDIDATES_EVENTS.with_label_values(&["remove"]));
 
 // Metrics collected on WAL redo operations
@@ -774,7 +802,7 @@ macro_rules! redo_bytes_histogram_count_buckets {
     };
 }
 
-pub static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
+pub(crate) static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wal_redo_seconds",
         "Time spent on WAL redo",
@@ -783,7 +811,7 @@ pub static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WAL_REDO_WAIT_TIME: Lazy<Histogram> = Lazy::new(|| {
+pub(crate) static WAL_REDO_WAIT_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wal_redo_wait_seconds",
         "Time spent waiting for access to the Postgres WAL redo process",
@@ -792,7 +820,7 @@ pub static WAL_REDO_WAIT_TIME: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
+pub(crate) static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wal_redo_records_histogram",
         "Histogram of number of records replayed per redo in the Postgres WAL redo process",
@@ -801,7 +829,7 @@ pub static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WAL_REDO_BYTES_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
+pub(crate) static WAL_REDO_BYTES_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wal_redo_bytes_histogram",
         "Histogram of number of records replayed per redo sent to Postgres",
@@ -810,7 +838,8 @@ pub static WAL_REDO_BYTES_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
-pub static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
+// FIXME: isn't this already included by WAL_REDO_RECORDS_HISTOGRAM which has _count?
+pub(crate) static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "pageserver_replayed_wal_records_total",
         "Number of WAL records replayed in WAL redo process"
@@ -818,7 +847,7 @@ pub static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
     .unwrap()
 });
 
-/// Similar to [`prometheus::HistogramTimer`] but does not record on drop.
+/// Similar to `prometheus::HistogramTimer` but does not record on drop.
 pub struct StorageTimeMetricsTimer {
     metrics: StorageTimeMetrics,
     start: Instant,
@@ -876,7 +905,7 @@ impl StorageTimeMetrics {
 
     /// Starts timing a new operation.
     ///
-    /// Note: unlike [`prometheus::HistogramTimer`] the returned timer does not record on drop.
+    /// Note: unlike `prometheus::HistogramTimer` the returned timer does not record on drop.
     pub fn start_timer(&self) -> StorageTimeMetricsTimer {
         StorageTimeMetricsTimer::new(self.clone())
     }
@@ -886,7 +915,6 @@ impl StorageTimeMetrics {
 pub struct TimelineMetrics {
     tenant_id: String,
     timeline_id: String,
-    pub get_reconstruct_data_time_histo: Histogram,
     pub flush_time_histo: StorageTimeMetrics,
     pub compact_time_histo: StorageTimeMetrics,
     pub create_images_time_histo: StorageTimeMetrics,
@@ -895,9 +923,7 @@ pub struct TimelineMetrics {
     pub load_layer_map_histo: StorageTimeMetrics,
     pub garbage_collect_histo: StorageTimeMetrics,
     pub last_record_gauge: IntGauge,
-    pub wait_lsn_time_histo: Histogram,
     pub resident_physical_size_gauge: UIntGauge,
-    pub read_num_fs_layers: Histogram,
     /// copy of LayeredTimeline.current_logical_size
     pub current_logical_size_gauge: UIntGauge,
     pub num_persistent_files_created: IntCounter,
@@ -914,9 +940,6 @@ impl TimelineMetrics {
     ) -> Self {
         let tenant_id = tenant_id.to_string();
         let timeline_id = timeline_id.to_string();
-        let get_reconstruct_data_time_histo = GET_RECONSTRUCT_DATA_TIME
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
         let flush_time_histo =
             StorageTimeMetrics::new(StorageTimeOperation::LayerFlush, &tenant_id, &timeline_id);
         let compact_time_histo =
@@ -937,9 +960,6 @@ impl TimelineMetrics {
         let last_record_gauge = LAST_RECORD_LSN
             .get_metric_with_label_values(&[&tenant_id, &timeline_id])
             .unwrap();
-        let wait_lsn_time_histo = WAIT_LSN_TIME
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
         let resident_physical_size_gauge = RESIDENT_PHYSICAL_SIZE
             .get_metric_with_label_values(&[&tenant_id, &timeline_id])
             .unwrap();
@@ -955,16 +975,12 @@ impl TimelineMetrics {
         let evictions = EVICTIONS
             .get_metric_with_label_values(&[&tenant_id, &timeline_id])
             .unwrap();
-        let read_num_fs_layers = READ_NUM_FS_LAYERS
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
         let evictions_with_low_residence_duration =
             evictions_with_low_residence_duration_builder.build(&tenant_id, &timeline_id);
 
         TimelineMetrics {
             tenant_id,
             timeline_id,
-            get_reconstruct_data_time_histo,
             flush_time_histo,
             compact_time_histo,
             create_images_time_histo,
@@ -973,7 +989,6 @@ impl TimelineMetrics {
             garbage_collect_histo,
             load_layer_map_histo,
             last_record_gauge,
-            wait_lsn_time_histo,
             resident_physical_size_gauge,
             current_logical_size_gauge,
             num_persistent_files_created,
@@ -982,7 +997,6 @@ impl TimelineMetrics {
             evictions_with_low_residence_duration: std::sync::RwLock::new(
                 evictions_with_low_residence_duration,
             ),
-            read_num_fs_layers,
         }
     }
 }
@@ -991,15 +1005,12 @@ impl Drop for TimelineMetrics {
     fn drop(&mut self) {
         let tenant_id = &self.tenant_id;
         let timeline_id = &self.timeline_id;
-        let _ = GET_RECONSTRUCT_DATA_TIME.remove_label_values(&[tenant_id, timeline_id]);
         let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = WAIT_LSN_TIME.remove_label_values(&[tenant_id, timeline_id]);
         let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
         let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
         let _ = NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, timeline_id]);
         let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, timeline_id]);
         let _ = EVICTIONS.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = READ_NUM_FS_LAYERS.remove_label_values(&[tenant_id, timeline_id]);
 
         self.evictions_with_low_residence_duration
             .write()
@@ -1011,9 +1022,6 @@ impl Drop for TimelineMetrics {
             let _ =
                 STORAGE_TIME_COUNT_PER_TIMELINE.remove_label_values(&[op, tenant_id, timeline_id]);
         }
-        for op in STORAGE_IO_TIME_OPERATIONS {
-            let _ = STORAGE_IO_TIME.remove_label_values(&[op, tenant_id, timeline_id]);
-        }
 
         for op in STORAGE_IO_SIZE_OPERATIONS {
             let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, timeline_id]);
@@ -1028,9 +1036,7 @@ impl Drop for TimelineMetrics {
 pub fn remove_tenant_metrics(tenant_id: &TenantId) {
     let tid = tenant_id.to_string();
     let _ = TENANT_SYNTHETIC_SIZE_METRIC.remove_label_values(&[&tid]);
-    for state in TenantState::VARIANTS {
-        let _ = TENANT_STATE_METRIC.remove_label_values(&[&tid, state]);
-    }
+    // we leave the BROKEN_TENANTS_SET entry if any
 }
 
 use futures::Future;
@@ -1045,9 +1051,7 @@ pub struct RemoteTimelineClientMetrics {
     tenant_id: String,
     timeline_id: String,
     remote_physical_size_gauge: Mutex<Option<UIntGauge>>,
-    remote_operation_time: Mutex<HashMap<(&'static str, &'static str, &'static str), Histogram>>,
     calls_unfinished_gauge: Mutex<HashMap<(&'static str, &'static str), IntGauge>>,
-    calls_started_hist: Mutex<HashMap<(&'static str, &'static str), Histogram>>,
     bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
     bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
 }
@@ -1057,14 +1061,13 @@ impl RemoteTimelineClientMetrics {
         RemoteTimelineClientMetrics {
             tenant_id: tenant_id.to_string(),
             timeline_id: timeline_id.to_string(),
-            remote_operation_time: Mutex::new(HashMap::default()),
             calls_unfinished_gauge: Mutex::new(HashMap::default()),
-            calls_started_hist: Mutex::new(HashMap::default()),
             bytes_started_counter: Mutex::new(HashMap::default()),
             bytes_finished_counter: Mutex::new(HashMap::default()),
             remote_physical_size_gauge: Mutex::new(None),
         }
     }
+
     pub fn remote_physical_size_gauge(&self) -> UIntGauge {
         let mut guard = self.remote_physical_size_gauge.lock().unwrap();
         guard
@@ -1078,26 +1081,17 @@ impl RemoteTimelineClientMetrics {
             })
             .clone()
     }
+
     pub fn remote_operation_time(
         &self,
         file_kind: &RemoteOpFileKind,
         op_kind: &RemoteOpKind,
         status: &'static str,
     ) -> Histogram {
-        let mut guard = self.remote_operation_time.lock().unwrap();
         let key = (file_kind.as_str(), op_kind.as_str(), status);
-        let metric = guard.entry(key).or_insert_with(move || {
-            REMOTE_OPERATION_TIME
-                .get_metric_with_label_values(&[
-                    &self.tenant_id.to_string(),
-                    &self.timeline_id.to_string(),
-                    key.0,
-                    key.1,
-                    key.2,
-                ])
-                .unwrap()
-        });
-        metric.clone()
+        REMOTE_OPERATION_TIME
+            .get_metric_with_label_values(&[key.0, key.1, key.2])
+            .unwrap()
     }
 
     fn calls_unfinished_gauge(
@@ -1125,19 +1119,10 @@ impl RemoteTimelineClientMetrics {
         file_kind: &RemoteOpFileKind,
         op_kind: &RemoteOpKind,
     ) -> Histogram {
-        let mut guard = self.calls_started_hist.lock().unwrap();
         let key = (file_kind.as_str(), op_kind.as_str());
-        let metric = guard.entry(key).or_insert_with(move || {
-            REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST
-                .get_metric_with_label_values(&[
-                    &self.tenant_id.to_string(),
-                    &self.timeline_id.to_string(),
-                    key.0,
-                    key.1,
-                ])
-                .unwrap()
-        });
-        metric.clone()
+        REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST
+            .get_metric_with_label_values(&[key.0, key.1])
+            .unwrap()
     }
 
     fn bytes_started_counter(
@@ -1256,7 +1241,7 @@ impl RemoteTimelineClientMetrics {
     /// Update the metrics that change when a call to the remote timeline client instance starts.
     ///
     /// Drop the returned guard object once the operation is finished to updates corresponding metrics that track completions.
-    /// Or, use [`RemoteTimelineClientCallMetricGuard::will_decrement_manually`] and [`call_end`] if that
+    /// Or, use [`RemoteTimelineClientCallMetricGuard::will_decrement_manually`] and [`call_end`](Self::call_end) if that
     /// is more suitable.
     /// Never do both.
     pub(crate) fn call_begin(
@@ -1289,7 +1274,7 @@ impl RemoteTimelineClientMetrics {
 
     /// Manually udpate the metrics that track completions, instead of using the guard object.
     /// Using the guard object is generally preferable.
-    /// See [`call_begin`] for more context.
+    /// See [`call_begin`](Self::call_begin) for more context.
     pub(crate) fn call_end(
         &self,
         file_kind: &RemoteOpFileKind,
@@ -1317,15 +1302,10 @@ impl Drop for RemoteTimelineClientMetrics {
             tenant_id,
             timeline_id,
             remote_physical_size_gauge,
-            remote_operation_time,
             calls_unfinished_gauge,
-            calls_started_hist,
             bytes_started_counter,
             bytes_finished_counter,
         } = self;
-        for ((a, b, c), _) in remote_operation_time.get_mut().unwrap().drain() {
-            let _ = REMOTE_OPERATION_TIME.remove_label_values(&[tenant_id, timeline_id, a, b, c]);
-        }
         for ((a, b), _) in calls_unfinished_gauge.get_mut().unwrap().drain() {
             let _ = REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE.remove_label_values(&[
                 tenant_id,
@@ -1334,14 +1314,6 @@ impl Drop for RemoteTimelineClientMetrics {
                 b,
             ]);
         }
-        for ((a, b), _) in calls_started_hist.get_mut().unwrap().drain() {
-            let _ = REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST.remove_label_values(&[
-                tenant_id,
-                timeline_id,
-                a,
-                b,
-            ]);
-        }
         for ((a, b), _) in bytes_started_counter.get_mut().unwrap().drain() {
             let _ = REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER.remove_label_values(&[
                 tenant_id,
@@ -1423,15 +1395,51 @@ impl<F: Future<Output = Result<O, E>>, O, E> Future for MeasuredRemoteOp<F> {
 }
 
 pub fn preinitialize_metrics() {
-    // We want to alert on this metric increasing.
-    // Initialize it eagerly, so that our alert rule can distinguish absence of the metric from metric value 0.
-    assert_eq!(UNEXPECTED_ONDEMAND_DOWNLOADS.get(), 0);
-    UNEXPECTED_ONDEMAND_DOWNLOADS.reset();
+    // Python tests need these and on some we do alerting.
+    //
+    // FIXME(4813): make it so that we have no top level metrics as this fn will easily fall out of
+    // order:
+    // - global metrics reside in a Lazy<PageserverMetrics>
+    //   - access via crate::metrics::PS_METRICS.materialized_page_cache_hit.inc()
+    // - could move the statics into TimelineMetrics::new()?
 
-    // Same as above for this metric, but, it's a Vec-type metric for which we don't know all the labels.
-    BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT.reset();
+    // counters
+    [
+        &MATERIALIZED_PAGE_CACHE_HIT,
+        &MATERIALIZED_PAGE_CACHE_HIT_DIRECT,
+        &UNEXPECTED_ONDEMAND_DOWNLOADS,
+        &WALRECEIVER_STARTED_CONNECTIONS,
+        &WALRECEIVER_BROKER_UPDATES,
+        &WALRECEIVER_CANDIDATES_ADDED,
+        &WALRECEIVER_CANDIDATES_REMOVED,
+    ]
+    .into_iter()
+    .for_each(|c| {
+        Lazy::force(c);
+    });
 
-    // Python tests need these.
-    MATERIALIZED_PAGE_CACHE_HIT_DIRECT.get();
-    MATERIALIZED_PAGE_CACHE_HIT.get();
+    // countervecs
+    [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
+        .into_iter()
+        .for_each(|c| {
+            Lazy::force(c);
+        });
+
+    // gauges
+    WALRECEIVER_ACTIVE_MANAGERS.get();
+
+    // histograms
+    [
+        &READ_NUM_FS_LAYERS,
+        &RECONSTRUCT_TIME,
+        &WAIT_LSN_TIME,
+        &WAL_REDO_TIME,
+        &WAL_REDO_WAIT_TIME,
+        &WAL_REDO_RECORDS_HISTOGRAM,
+        &WAL_REDO_BYTES_HISTOGRAM,
+    ]
+    .into_iter()
+    .for_each(|h| {
+        Lazy::force(h);
+    });
 }

pageserver/src/pgdatadir_mapping.rs

@@ -1131,7 +1131,7 @@ impl<'a> DatadirModification<'a> {
     /// context, breaking the atomicity is OK. If the import is interrupted, the
     /// whole import fails and the timeline will be deleted anyway.
     /// (Or to be precise, it will be left behind for debugging purposes and
-    /// ignored, see https://github.com/neondatabase/neon/pull/1809)
+    /// ignored, see <https://github.com/neondatabase/neon/pull/1809>)
     ///
     /// Note: A consequence of flushing the pending operations is that they
     /// won't be visible to subsequent operations until `commit`. The function

pageserver/src/task_mgr.rs

@@ -130,11 +130,25 @@ pub static WALRECEIVER_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
 pub static BACKGROUND_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
     tokio::runtime::Builder::new_multi_thread()
         .thread_name("background op worker")
+        // if you change the number of worker threads please change the constant below
         .enable_all()
         .build()
         .expect("Failed to create background op runtime")
 });
 
+pub(crate) static BACKGROUND_RUNTIME_WORKER_THREADS: Lazy<usize> = Lazy::new(|| {
+    // force init and thus panics
+    let _ = BACKGROUND_RUNTIME.handle();
+    // replicates tokio-1.28.1::loom::sys::num_cpus which is not available publicly
+    // tokio would had already panicked for parsing errors or NotUnicode
+    //
+    // this will be wrong if any of the runtimes gets their worker threads configured to something
+    // else, but that has not been needed in a long time.
+    std::env::var("TOKIO_WORKER_THREADS")
+        .map(|s| s.parse::<usize>().unwrap())
+        .unwrap_or_else(|_e| usize::max(1, num_cpus::get()))
+});
+
 #[derive(Debug, Clone, Copy)]
 pub struct PageserverTaskId(u64);
 
@@ -205,7 +219,7 @@ pub enum TaskKind {
     ///
     /// Walreceiver uses its own abstraction called `TaskHandle` to represent the activity of establishing and handling a connection.
     /// That abstraction doesn't use `task_mgr`.
-    /// The [`WalReceiverManager`] task ensures that this `TaskHandle` task does not outlive the [`WalReceiverManager`] task.
+    /// The `WalReceiverManager` task ensures that this `TaskHandle` task does not outlive the `WalReceiverManager` task.
     /// For the `RequestContext` that we hand to the TaskHandle, we use the [`WalReceiverConnectionHandler`] task kind.
     ///
     /// Once the connection is established, the `TaskHandle` task creates a
@@ -213,16 +227,21 @@ pub enum TaskKind {
     /// the `Connection` object.
     /// A `CancellationToken` created by the `TaskHandle` task ensures
     /// that the [`WalReceiverConnectionPoller`] task will cancel soon after as the `TaskHandle` is dropped.
+    ///
+    /// [`WalReceiverConnectionHandler`]: Self::WalReceiverConnectionHandler
+    /// [`WalReceiverConnectionPoller`]: Self::WalReceiverConnectionPoller
     WalReceiverManager,
 
-    /// The `TaskHandle` task that executes [`walreceiver_connection::handle_walreceiver_connection`].
+    /// The `TaskHandle` task that executes `handle_walreceiver_connection`.
     /// Not a `task_mgr` task, but we use this `TaskKind` for its `RequestContext`.
     /// See the comment on [`WalReceiverManager`].
+    ///
+    /// [`WalReceiverManager`]: Self::WalReceiverManager
     WalReceiverConnectionHandler,
 
     /// The task that polls the `tokio-postgres::Connection` object.
-    /// Spawned by task [`WalReceiverConnectionHandler`].
-    /// See the comment on [`WalReceiverManager`].
+    /// Spawned by task [`WalReceiverConnectionHandler`](Self::WalReceiverConnectionHandler).
+    /// See the comment on [`WalReceiverManager`](Self::WalReceiverManager).
     WalReceiverConnectionPoller,
 
     // Garbage collection worker. One per tenant
@@ -506,17 +525,13 @@ pub async fn shutdown_tasks(
                     warn!(name = task.name, tenant_id = ?tenant_id, timeline_id = ?timeline_id, kind = ?task_kind, "stopping left-over");
                 }
             }
-            let join_handle = tokio::select! {
-                biased;
-                _ = &mut join_handle => { None },
-                _ = tokio::time::sleep(std::time::Duration::from_secs(1)) => {
-                    // allow some time to elapse before logging to cut down the number of log
-                    // lines.
-                    info!("waiting for {} to shut down", task.name);
-                    Some(join_handle)
-                }
-            };
-            if let Some(join_handle) = join_handle {
+            if tokio::time::timeout(std::time::Duration::from_secs(1), &mut join_handle)
+                .await
+                .is_err()
+            {
+                // allow some time to elapse before logging to cut down the number of log
+                // lines.
+                info!("waiting for {} to shut down", task.name);
                 // we never handled this return value, but:
                 // - we don't deschedule which would lead to is_cancelled
                 // - panics are already logged (is_panicked)
@@ -544,7 +559,7 @@ pub fn current_task_id() -> Option<PageserverTaskId> {
 pub async fn shutdown_watcher() {
     let token = SHUTDOWN_TOKEN
         .try_with(|t| t.clone())
-        .expect("shutdown_requested() called in an unexpected task or thread");
+        .expect("shutdown_watcher() called in an unexpected task or thread");
 
     token.cancelled().await;
 }

pageserver/src/tenant/blob_io.rs

@@ -16,30 +16,20 @@ use crate::tenant::block_io::{BlockCursor, BlockReader};
 use std::cmp::min;
 use std::io::{Error, ErrorKind};
 
-/// For reading
-pub trait BlobCursor {
+impl<R> BlockCursor<R>
+where
+    R: BlockReader,
+{
     /// Read a blob into a new buffer.
-    fn read_blob(&mut self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
+    pub fn read_blob(&self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
         let mut buf = Vec::new();
         self.read_blob_into_buf(offset, &mut buf)?;
         Ok(buf)
     }
-
     /// Read blob into the given buffer. Any previous contents in the buffer
     /// are overwritten.
-    fn read_blob_into_buf(
-        &mut self,
-        offset: u64,
-        dstbuf: &mut Vec<u8>,
-    ) -> Result<(), std::io::Error>;
-}
-
-impl<R> BlobCursor for BlockCursor<R>
-where
-    R: BlockReader,
-{
-    fn read_blob_into_buf(
-        &mut self,
+    pub fn read_blob_into_buf(
+        &self,
         offset: u64,
         dstbuf: &mut Vec<u8>,
     ) -> Result<(), std::io::Error> {

pageserver/src/tenant/block_io.rs

@@ -80,7 +80,7 @@ where
         BlockCursor { reader }
     }
 
-    pub fn read_blk(&mut self, blknum: u32) -> Result<R::BlockLease, std::io::Error> {
+    pub fn read_blk(&self, blknum: u32) -> Result<R::BlockLease, std::io::Error> {
         self.reader.read_blk(blknum)
     }
 }

pageserver/src/tenant/disk_btree.rs

@@ -390,39 +390,42 @@ where
     }
 
     #[allow(dead_code)]
-    pub fn dump(&self) -> Result<()> {
-        self.dump_recurse(self.root_blk, &[], 0)
-    }
+    pub async fn dump(&self) -> Result<()> {
+        let mut stack = Vec::new();
 
-    fn dump_recurse(&self, blknum: u32, path: &[u8], depth: usize) -> Result<()> {
-        let blk = self.reader.read_blk(self.start_blk + blknum)?;
-        let buf: &[u8] = blk.as_ref();
+        stack.push((self.root_blk, String::new(), 0, 0, 0));
 
-        let node = OnDiskNode::<L>::deparse(buf)?;
+        while let Some((blknum, path, depth, child_idx, key_off)) = stack.pop() {
+            let blk = self.reader.read_blk(self.start_blk + blknum)?;
+            let buf: &[u8] = blk.as_ref();
+            let node = OnDiskNode::<L>::deparse(buf)?;
 
-        print!("{:indent$}", "", indent = depth * 2);
-        println!(
-            "blk #{}: path {}: prefix {}, suffix_len {}",
-            blknum,
-            hex::encode(path),
-            hex::encode(node.prefix),
-            node.suffix_len
-        );
+            if child_idx == 0 {
+                print!("{:indent$}", "", indent = depth * 2);
+                let path_prefix = stack
+                    .iter()
+                    .map(|(_blknum, path, ..)| path.as_str())
+                    .collect::<String>();
+                println!(
+                    "blk #{blknum}: path {path_prefix}{path}: prefix {}, suffix_len {}",
+                    hex::encode(node.prefix),
+                    node.suffix_len
+                );
+            }
 
-        let mut idx = 0;
-        let mut key_off = 0;
-        while idx < node.num_children {
+            if child_idx + 1 < node.num_children {
+                let key_off = key_off + node.suffix_len as usize;
+                stack.push((blknum, path.clone(), depth, child_idx + 1, key_off));
+            }
             let key = &node.keys[key_off..key_off + node.suffix_len as usize];
-            let val = node.value(idx as usize);
+            let val = node.value(child_idx as usize);
+
             print!("{:indent$}", "", indent = depth * 2 + 2);
             println!("{}: {}", hex::encode(key), hex::encode(val.0));
 
             if node.level > 0 {
-                let child_path = [path, node.prefix].concat();
-                self.dump_recurse(val.to_blknum(), &child_path, depth + 1)?;
+                stack.push((val.to_blknum(), hex::encode(node.prefix), depth + 1, 0, 0));
             }
-            idx += 1;
-            key_off += node.suffix_len as usize;
         }
         Ok(())
     }
@@ -442,7 +445,7 @@ where
     writer: W,
 
     ///
-    /// stack[0] is the current root page, stack.last() is the leaf.
+    /// `stack[0]` is the current root page, `stack.last()` is the leaf.
     ///
     /// We maintain the length of the stack to be always greater than zero.
     /// Two exceptions are:
@@ -754,8 +757,8 @@ mod tests {
         }
     }
 
-    #[test]
-    fn basic() -> Result<()> {
+    #[tokio::test]
+    async fn basic() -> Result<()> {
         let mut disk = TestDisk::new();
         let mut writer = DiskBtreeBuilder::<_, 6>::new(&mut disk);
 
@@ -775,7 +778,7 @@ mod tests {
 
         let reader = DiskBtreeReader::new(0, root_offset, disk);
 
-        reader.dump()?;
+        reader.dump().await?;
 
         // Test the `get` function on all the keys.
         for (key, val) in all_data.iter() {
@@ -835,8 +838,8 @@ mod tests {
         Ok(())
     }
 
-    #[test]
-    fn lots_of_keys() -> Result<()> {
+    #[tokio::test]
+    async fn lots_of_keys() -> Result<()> {
         let mut disk = TestDisk::new();
         let mut writer = DiskBtreeBuilder::<_, 8>::new(&mut disk);
 
@@ -856,7 +859,7 @@ mod tests {
 
         let reader = DiskBtreeReader::new(0, root_offset, disk);
 
-        reader.dump()?;
+        reader.dump().await?;
 
         use std::sync::Mutex;
 
@@ -994,8 +997,8 @@ mod tests {
     ///
     /// This test contains a particular data set, see disk_btree_test_data.rs
     ///
-    #[test]
-    fn particular_data() -> Result<()> {
+    #[tokio::test]
+    async fn particular_data() -> Result<()> {
         // Build a tree from it
         let mut disk = TestDisk::new();
         let mut writer = DiskBtreeBuilder::<_, 26>::new(&mut disk);
@@ -1022,7 +1025,7 @@ mod tests {
         })?;
         assert_eq!(count, disk_btree_test_data::TEST_DATA.len());
 
-        reader.dump()?;
+        reader.dump().await?;
 
         Ok(())
     }

pageserver/src/tenant/ephemeral_file.rs

@@ -328,7 +328,7 @@ fn to_io_error(e: anyhow::Error, context: &str) -> io::Error {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::tenant::blob_io::{BlobCursor, BlobWriter};
+    use crate::tenant::blob_io::BlobWriter;
     use crate::tenant::block_io::BlockCursor;
     use rand::{seq::SliceRandom, thread_rng, RngCore};
     use std::fs;
@@ -420,7 +420,7 @@ mod tests {
             blobs.push((pos, data));
         }
 
-        let mut cursor = BlockCursor::new(&file);
+        let cursor = BlockCursor::new(&file);
         for (pos, expected) in blobs {
             let actual = cursor.read_blob(pos)?;
             assert_eq!(actual, expected);

pageserver/src/tenant/layer_map.rs

@@ -16,7 +16,7 @@
 //! Other read methods are less critical but still impact performance of background tasks.
 //!
 //! This data structure relies on a persistent/immutable binary search tree. See the
-//! following lecture for an introduction https://www.youtube.com/watch?v=WqCWghETNDc&t=581s
+//! following lecture for an introduction <https://www.youtube.com/watch?v=WqCWghETNDc&t=581s>
 //! Summary: A persistent/immutable BST (and persistent data structures in general) allows
 //! you to modify the tree in such a way that each modification creates a new "version"
 //! of the tree. When you modify it, you get a new version, but all previous versions are
@@ -40,7 +40,7 @@
 //! afterwards. We can add layers as long as they have larger LSNs than any previous layer in
 //! the map, but if we need to remove a layer, or insert anything with an older LSN, we need
 //! to throw away most of the persistent BST and build a new one, starting from the oldest
-//! LSN. See `LayerMap::flush_updates()`.
+//! LSN. See [`LayerMap::flush_updates()`].
 //!
 
 mod historic_layer_coverage;
@@ -626,17 +626,17 @@ impl LayerMap {
 
     /// debugging function to print out the contents of the layer map
     #[allow(unused)]
-    pub fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         println!("Begin dump LayerMap");
 
         println!("open_layer:");
         if let Some(open_layer) = &self.open_layer {
-            open_layer.dump(verbose, ctx)?;
+            open_layer.dump(verbose, ctx).await?;
         }
 
         println!("frozen_layers:");
         for frozen_layer in self.frozen_layers.iter() {
-            frozen_layer.dump(verbose, ctx)?;
+            frozen_layer.dump(verbose, ctx).await?;
         }
 
         println!("historic_layers:");

pageserver/src/tenant/layer_map/historic_layer_coverage.rs

@@ -122,8 +122,7 @@ impl<Value: Clone> HistoricLayerCoverage<Value> {
         self.head = self
             .historic
             .iter()
-            .rev()
-            .next()
+            .next_back()
             .map(|(_, v)| v.clone())
             .unwrap_or_default();
     }
@@ -412,7 +411,7 @@ fn test_persistent_overlapping() {
 /// still be more critical.
 ///
 /// See this for more on persistent and retroactive techniques:
-/// https://www.youtube.com/watch?v=WqCWghETNDc&t=581s
+/// <https://www.youtube.com/watch?v=WqCWghETNDc&t=581s>
 pub struct BufferedHistoricLayerCoverage<Value> {
     /// A persistent layer map that we rebuild when we need to retroactively update
     historic_coverage: HistoricLayerCoverage<Value>,

pageserver/src/tenant/layer_map/layer_coverage.rs

@@ -2,7 +2,7 @@ use std::ops::Range;
 
 // NOTE the `im` crate has 20x more downloads and also has
 // persistent/immutable BTree. But it's bugged so rpds is a
-// better choice https://github.com/neondatabase/neon/issues/3395
+// better choice <https://github.com/neondatabase/neon/issues/3395>
 use rpds::RedBlackTreeMapSync;
 
 /// Data structure that can efficiently:
@@ -11,7 +11,7 @@ use rpds::RedBlackTreeMapSync;
 /// - insert layers in non-decreasing lsn.start order
 ///
 /// For a detailed explanation and justification of this approach, see:
-/// https://neon.tech/blog/persistent-structures-in-neons-wal-indexing
+/// <https://neon.tech/blog/persistent-structures-in-neons-wal-indexing>
 ///
 /// NOTE The struct is parameterized over Value for easier
 ///      testing, but in practice it's some sort of layer.
@@ -113,8 +113,7 @@ impl<Value: Clone> LayerCoverage<Value> {
     pub fn query(&self, key: i128) -> Option<Value> {
         self.nodes
             .range(..=key)
-            .rev()
-            .next()?
+            .next_back()?
             .1
             .as_ref()
             .map(|(_, v)| v.clone())

pageserver/src/tenant/manifest.rs

@@ -24,7 +24,7 @@
 //! Currently, this is not used in the system. Future refactors will ensure
 //! the storage state will be recorded in this file, and the system can be
 //! recovered from this file. This is tracked in
-//! https://github.com/neondatabase/neon/issues/4418
+//! <https://github.com/neondatabase/neon/issues/4418>
 
 use std::io::{self, Read, Write};
 

pageserver/src/tenant/metadata.rs

@@ -1,16 +1,19 @@
 //! Every image of a certain timeline from [`crate::tenant::Tenant`]
 //! has a metadata that needs to be stored persistently.
 //!
-//! Later, the file gets is used in [`crate::remote_storage::storage_sync`] as a part of
+//! Later, the file gets used in [`remote_timeline_client`] as a part of
 //! external storage import and export operations.
 //!
 //! The module contains all structs and related helper methods related to timeline metadata.
+//!
+//! [`remote_timeline_client`]: super::remote_timeline_client
 
 use std::fs::{File, OpenOptions};
-use std::io::Write;
+use std::io::{self, Write};
 
 use anyhow::{bail, ensure, Context};
 use serde::{Deserialize, Serialize};
+use thiserror::Error;
 use tracing::info_span;
 use utils::bin_ser::SerializeError;
 use utils::{
@@ -265,24 +268,24 @@ pub fn save_metadata(
     Ok(())
 }
 
+#[derive(Error, Debug)]
+pub enum LoadMetadataError {
+    #[error(transparent)]
+    Read(#[from] io::Error),
+
+    #[error(transparent)]
+    Decode(#[from] anyhow::Error),
+}
+
 pub fn load_metadata(
     conf: &'static PageServerConf,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
-) -> anyhow::Result<TimelineMetadata> {
+) -> Result<TimelineMetadata, LoadMetadataError> {
     let metadata_path = conf.metadata_path(tenant_id, timeline_id);
-    let metadata_bytes = std::fs::read(&metadata_path).with_context(|| {
-        format!(
-            "Failed to read metadata bytes from path {}",
-            metadata_path.display()
-        )
-    })?;
-    TimelineMetadata::from_bytes(&metadata_bytes).with_context(|| {
-        format!(
-            "Failed to parse metadata bytes from path {}",
-            metadata_path.display()
-        )
-    })
+    let metadata_bytes = std::fs::read(metadata_path)?;
+
+    Ok(TimelineMetadata::from_bytes(&metadata_bytes)?)
 }
 
 #[cfg(test)]

pageserver/src/tenant/mgr.rs

@@ -26,6 +26,8 @@ use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME};
 use utils::fs_ext::PathExt;
 use utils::id::{TenantId, TimelineId};
 
+use super::timeline::delete::DeleteTimelineFlow;
+
 /// The tenants known to the pageserver.
 /// The enum variants are used to distinguish the different states that the pageserver can be in.
 enum TenantsMap {
@@ -233,11 +235,17 @@ pub fn schedule_local_tenant_processing(
 /// That could be easily misinterpreted by control plane, the consumer of the
 /// management API. For example, it could attach the tenant on a different pageserver.
 /// We would then be in split-brain once this pageserver restarts.
-#[instrument]
+#[instrument(skip_all)]
 pub async fn shutdown_all_tenants() {
+    shutdown_all_tenants0(&TENANTS).await
+}
+
+async fn shutdown_all_tenants0(tenants: &tokio::sync::RwLock<TenantsMap>) {
+    use utils::completion;
+
     // Prevent new tenants from being created.
     let tenants_to_shut_down = {
-        let mut m = TENANTS.write().await;
+        let mut m = tenants.write().await;
         match &mut *m {
             TenantsMap::Initializing => {
                 *m = TenantsMap::ShuttingDown(HashMap::default());
@@ -262,14 +270,41 @@ pub async fn shutdown_all_tenants() {
     for (tenant_id, tenant) in tenants_to_shut_down {
         join_set.spawn(
             async move {
-                let freeze_and_flush = true;
+                // ordering shouldn't matter for this, either we store true right away or never
+                let ordering = std::sync::atomic::Ordering::Relaxed;
+                let joined_other = std::sync::atomic::AtomicBool::new(false);
 
-                match tenant.shutdown(freeze_and_flush).await {
-                    Ok(()) => debug!("tenant successfully stopped"),
-                    Err(super::ShutdownError::AlreadyStopping) => {
-                        warn!("tenant was already shutting down")
+                let mut shutdown = std::pin::pin!(async {
+                    let freeze_and_flush = true;
+
+                    let res = {
+                        let (_guard, shutdown_progress) = completion::channel();
+                        tenant.shutdown(shutdown_progress, freeze_and_flush).await
+                    };
+
+                    if let Err(other_progress) = res {
+                        // join the another shutdown in progress
+                        joined_other.store(true, ordering);
+                        other_progress.wait().await;
                     }
-                }
+                });
+
+                // in practice we might not have a lot time to go, since systemd is going to
+                // SIGKILL us at 10s, but we can try. delete tenant might take a while, so put out
+                // a warning.
+                let warning = std::time::Duration::from_secs(5);
+                let mut warning = std::pin::pin!(tokio::time::sleep(warning));
+
+                tokio::select! {
+                    _ = &mut shutdown => {},
+                    _ = &mut warning => {
+                        let joined_other = joined_other.load(ordering);
+                        warn!(%joined_other, "waiting for the shutdown to complete");
+                        shutdown.await;
+                    }
+                };
+
+                debug!("tenant successfully stopped");
             }
             .instrument(info_span!("shutdown", %tenant_id)),
         );
@@ -388,12 +423,10 @@ pub enum DeleteTimelineError {
 pub async fn delete_timeline(
     tenant_id: TenantId,
     timeline_id: TimelineId,
-    ctx: &RequestContext,
+    _ctx: &RequestContext,
 ) -> Result<(), DeleteTimelineError> {
     let tenant = get_tenant(tenant_id, true).await?;
-    tenant
-        .prepare_and_schedule_delete_timeline(timeline_id, ctx)
-        .await?;
+    DeleteTimelineFlow::run(&tenant, timeline_id).await?;
     Ok(())
 }
 
@@ -413,6 +446,15 @@ pub async fn detach_tenant(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
     detach_ignored: bool,
+) -> Result<(), TenantStateError> {
+    detach_tenant0(conf, &TENANTS, tenant_id, detach_ignored).await
+}
+
+async fn detach_tenant0(
+    conf: &'static PageServerConf,
+    tenants: &tokio::sync::RwLock<TenantsMap>,
+    tenant_id: TenantId,
+    detach_ignored: bool,
 ) -> Result<(), TenantStateError> {
     let local_files_cleanup_operation = |tenant_id_to_clean| async move {
         let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
@@ -425,7 +467,8 @@ pub async fn detach_tenant(
     };
 
     let removal_result =
-        remove_tenant_from_memory(tenant_id, local_files_cleanup_operation(tenant_id)).await;
+        remove_tenant_from_memory(tenants, tenant_id, local_files_cleanup_operation(tenant_id))
+            .await;
 
     // Ignored tenants are not present in memory and will bail the removal from memory operation.
     // Before returning the error, check for ignored tenant removal case — we only need to clean its local files then.
@@ -472,7 +515,15 @@ pub async fn ignore_tenant(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
 ) -> Result<(), TenantStateError> {
-    remove_tenant_from_memory(tenant_id, async {
+    ignore_tenant0(conf, &TENANTS, tenant_id).await
+}
+
+async fn ignore_tenant0(
+    conf: &'static PageServerConf,
+    tenants: &tokio::sync::RwLock<TenantsMap>,
+    tenant_id: TenantId,
+) -> Result<(), TenantStateError> {
+    remove_tenant_from_memory(tenants, tenant_id, async {
         let ignore_mark_file = conf.tenant_ignore_mark_file_path(&tenant_id);
         fs::File::create(&ignore_mark_file)
             .await
@@ -597,18 +648,21 @@ where
 /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
 /// operation would be needed to remove it.
 async fn remove_tenant_from_memory<V, F>(
+    tenants: &tokio::sync::RwLock<TenantsMap>,
     tenant_id: TenantId,
     tenant_cleanup: F,
 ) -> Result<V, TenantStateError>
 where
     F: std::future::Future<Output = anyhow::Result<V>>,
 {
+    use utils::completion;
+
     // It's important to keep the tenant in memory after the final cleanup, to avoid cleanup races.
     // The exclusive lock here ensures we don't miss the tenant state updates before trying another removal.
     // tenant-wde cleanup operations may take some time (removing the entire tenant directory), we want to
     // avoid holding the lock for the entire process.
     let tenant = {
-        TENANTS
+        tenants
             .write()
             .await
             .get(&tenant_id)
@@ -616,14 +670,20 @@ where
             .ok_or(TenantStateError::NotFound(tenant_id))?
     };
 
+    // allow pageserver shutdown to await for our completion
+    let (_guard, progress) = completion::channel();
+
+    // whenever we remove a tenant from memory, we don't want to flush and wait for upload
     let freeze_and_flush = false;
 
     // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
     // that we can continue safely to cleanup.
-    match tenant.shutdown(freeze_and_flush).await {
+    match tenant.shutdown(progress, freeze_and_flush).await {
         Ok(()) => {}
-        Err(super::ShutdownError::AlreadyStopping) => {
-            return Err(TenantStateError::IsStopping(tenant_id))
+        Err(_other) => {
+            // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
+            // wait for it but return an error right away because these are distinct requests.
+            return Err(TenantStateError::IsStopping(tenant_id));
         }
     }
 
@@ -632,14 +692,14 @@ where
         .with_context(|| format!("Failed to run cleanup for tenant {tenant_id}"))
     {
         Ok(hook_value) => {
-            let mut tenants_accessor = TENANTS.write().await;
+            let mut tenants_accessor = tenants.write().await;
             if tenants_accessor.remove(&tenant_id).is_none() {
                 warn!("Tenant {tenant_id} got removed from memory before operation finished");
             }
             Ok(hook_value)
         }
         Err(e) => {
-            let tenants_accessor = TENANTS.read().await;
+            let tenants_accessor = tenants.read().await;
             match tenants_accessor.get(&tenant_id) {
                 Some(tenant) => {
                     tenant.set_broken(e.to_string()).await;
@@ -708,51 +768,108 @@ pub async fn immediate_gc(
     Ok(wait_task_done)
 }
 
-pub async fn immediate_compact(
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
-    ctx: &RequestContext,
-) -> Result<tokio::sync::oneshot::Receiver<anyhow::Result<()>>, ApiError> {
-    let guard = TENANTS.read().await;
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Arc;
+    use tracing::{info_span, Instrument};
 
-    let tenant = guard
-        .get(&tenant_id)
-        .map(Arc::clone)
-        .with_context(|| format!("tenant {tenant_id}"))
-        .map_err(|e| ApiError::NotFound(e.into()))?;
+    use super::{super::harness::TenantHarness, TenantsMap};
 
-    let timeline = tenant
-        .get_timeline(timeline_id, true)
-        .map_err(|e| ApiError::NotFound(e.into()))?;
+    #[tokio::test(start_paused = true)]
+    async fn shutdown_joins_remove_tenant_from_memory() {
+        // the test is a bit ugly with the lockstep together with spawned tasks. the aim is to make
+        // sure `shutdown_all_tenants0` per-tenant processing joins in any active
+        // remove_tenant_from_memory calls, which is enforced by making the operation last until
+        // we've ran `shutdown_all_tenants0` for a long time.
 
-    // Run in task_mgr to avoid race with tenant_detach operation
-    let ctx = ctx.detached_child(TaskKind::Compaction, DownloadBehavior::Download);
-    let (task_done, wait_task_done) = tokio::sync::oneshot::channel();
-    task_mgr::spawn(
-        &tokio::runtime::Handle::current(),
-        TaskKind::Compaction,
-        Some(tenant_id),
-        Some(timeline_id),
-        &format!(
-            "timeline_compact_handler compaction run for tenant {tenant_id} timeline {timeline_id}"
-        ),
-        false,
-        async move {
-            let result = timeline
-                .compact(&ctx)
-                .instrument(info_span!("manual_compact", %tenant_id, %timeline_id))
-                .await;
+        let (t, _ctx) = TenantHarness::create("shutdown_joins_detach")
+            .unwrap()
+            .load()
+            .await;
 
-            match task_done.send(result) {
-                Ok(_) => (),
-                Err(result) => error!("failed to send compaction result: {result:?}"),
-            }
-            Ok(())
-        },
-    );
+        // harness loads it to active, which is forced and nothing is running on the tenant
 
-    // drop the guard until after we've spawned the task so that timeline shutdown will wait for the task
-    drop(guard);
+        let id = t.tenant_id();
 
-    Ok(wait_task_done)
+        // tenant harness configures the logging and we cannot escape it
+        let _e = info_span!("testing", tenant_id = %id).entered();
+
+        let tenants = HashMap::from([(id, t.clone())]);
+        let tenants = Arc::new(tokio::sync::RwLock::new(TenantsMap::Open(tenants)));
+
+        let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
+        let (until_cleanup_started, cleanup_started) = utils::completion::channel();
+
+        // start a "detaching operation", which will take a while, until can_complete_cleanup
+        let cleanup_task = {
+            let jh = tokio::spawn({
+                let tenants = tenants.clone();
+                async move {
+                    let cleanup = async move {
+                        drop(until_cleanup_started);
+                        can_complete_cleanup.wait().await;
+                        anyhow::Ok(())
+                    };
+                    super::remove_tenant_from_memory(&tenants, id, cleanup).await
+                }
+                .instrument(info_span!("foobar", tenant_id = %id))
+            });
+
+            // now the long cleanup should be in place, with the stopping state
+            cleanup_started.wait().await;
+            jh
+        };
+
+        let mut cleanup_progress = std::pin::pin!(t
+            .shutdown(utils::completion::Barrier::default(), false)
+            .await
+            .unwrap_err()
+            .wait());
+
+        let mut shutdown_task = {
+            let (until_shutdown_started, shutdown_started) = utils::completion::channel();
+
+            let shutdown_task = tokio::spawn(async move {
+                drop(until_shutdown_started);
+                super::shutdown_all_tenants0(&tenants).await;
+            });
+
+            shutdown_started.wait().await;
+            shutdown_task
+        };
+
+        // if the joining in is removed from shutdown_all_tenants0, the shutdown_task should always
+        // get to complete within timeout and fail the test. it is expected to continue awaiting
+        // until completion or SIGKILL during normal shutdown.
+        //
+        // the timeout is long to cover anything that shutdown_task could be doing, but it is
+        // handled instantly because we use tokio's time pausing in this test. 100s is much more than
+        // what we get from systemd on shutdown (10s).
+        let long_time = std::time::Duration::from_secs(100);
+        tokio::select! {
+            _ = &mut shutdown_task => unreachable!("shutdown must continue, until_cleanup_completed is not dropped"),
+            _ = &mut cleanup_progress => unreachable!("cleanup progress must continue, until_cleanup_completed is not dropped"),
+            _ = tokio::time::sleep(long_time) => {},
+        }
+
+        // allow the remove_tenant_from_memory and thus eventually the shutdown to continue
+        drop(until_cleanup_completed);
+
+        let (je, ()) = tokio::join!(shutdown_task, cleanup_progress);
+        je.expect("Tenant::shutdown shutdown not have panicked");
+        cleanup_task
+            .await
+            .expect("no panicking")
+            .expect("remove_tenant_from_memory failed");
+
+        futures::future::poll_immediate(
+            t.shutdown(utils::completion::Barrier::default(), false)
+                .await
+                .unwrap_err()
+                .wait(),
+        )
+        .await
+        .expect("the stopping progress must still be complete");
+    }
 }

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
-//!   [`Timeline::timeline_init_and_sync`] and [`Timeline::reconcile_with_remote`].
+//!   [`Tenant::timeline_init_and_sync`] and [`Timeline::reconcile_with_remote`].
 //!
 //! 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,
@@ -163,8 +163,8 @@
 //!   - download their remote [`IndexPart`]s
 //!   - create `Timeline` struct and a `RemoteTimelineClient`
 //!   - initialize the client's upload queue with its `IndexPart`
-//!   - create [`RemoteLayer`] instances for layers that are referenced by `IndexPart`
-//!     but not present locally
+//!   - create [`RemoteLayer`](super::storage_layer::RemoteLayer) instances
+//!     for layers that are referenced by `IndexPart` but not present locally
 //!   - schedule uploads for layers that are only present locally.
 //!   - if the remote `IndexPart`'s metadata was newer than the metadata in
 //!     the local filesystem, write the remote metadata to the local filesystem
@@ -198,6 +198,8 @@
 //! 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
 
 mod delete;
 mod download;
@@ -512,7 +514,7 @@ impl RemoteTimelineClient {
     /// updated metadata.
     ///
     /// The upload will be added to the queue immediately, but it
-    /// won't be performed until all previosuly scheduled layer file
+    /// won't be performed until all previously scheduled layer file
     /// upload operations have completed successfully.  This is to
     /// ensure that when the index file claims that layers X, Y and Z
     /// exist in remote storage, they really do. To wait for the upload
@@ -623,7 +625,7 @@ impl RemoteTimelineClient {
     /// Note: This schedules an index file upload before the deletions.  The
     /// deletion won't actually be performed, until any previously scheduled
     /// upload operations, and the index file upload, have completed
-    /// succesfully.
+    /// successfully.
     pub fn schedule_layer_file_deletion(
         self: &Arc<Self>,
         names: &[LayerFileName],
@@ -825,7 +827,7 @@ impl RemoteTimelineClient {
             )
         };
 
-        receiver.changed().await?;
+        receiver.changed().await.context("upload queue shut down")?;
 
         // Do not delete index part yet, it is needed for possible retry. If we remove it first
         // and retry will arrive to different pageserver there wont be any traces of it on remote storage
@@ -840,23 +842,37 @@ impl RemoteTimelineClient {
         let remaining: Vec<RemotePath> = remaining
             .into_iter()
             .filter(|p| p.object_name() != Some(IndexPart::FILE_NAME))
+            .inspect(|path| {
+                if let Some(name) = path.object_name() {
+                    info!(%name, "deleting a file not referenced from index_part.json");
+                } else {
+                    warn!(%path, "deleting a nameless or non-utf8 object not referenced from index_part.json");
+                }
+            })
             .collect();
 
         if !remaining.is_empty() {
-            warn!(
-                "Found {} files not bound to index_file.json, proceeding with their deletion",
-                remaining.len()
-            );
-            warn!("About to remove {} files", remaining.len());
             self.storage_impl.delete_objects(&remaining).await?;
         }
 
+        fail::fail_point!("timeline-delete-before-index-delete", |_| {
+            Err(anyhow::anyhow!(
+                "failpoint: timeline-delete-before-index-delete"
+            ))?
+        });
+
         let index_file_path = timeline_storage_path.join(Path::new(IndexPart::FILE_NAME));
 
         debug!("deleting index part");
         self.storage_impl.delete(&index_file_path).await?;
 
-        info!(deletions_queued, "done deleting, including index_part.json");
+        fail::fail_point!("timeline-delete-after-index-delete", |_| {
+            Err(anyhow::anyhow!(
+                "failpoint: timeline-delete-after-index-delete"
+            ))?
+        });
+
+        info!(prefix=%timeline_storage_path, referenced=deletions_queued, not_referenced=%remaining.len(), "done deleting in timeline prefix, including index_part.json");
 
         Ok(())
     }
@@ -1101,7 +1117,7 @@ impl RemoteTimelineClient {
             debug!("remote task {} completed successfully", task.op);
         }
 
-        // The task has completed succesfully. Remove it from the in-progress list.
+        // The task has completed successfully. Remove it from the in-progress list.
         {
             let mut upload_queue_guard = self.upload_queue.lock().unwrap();
             let upload_queue = match upload_queue_guard.deref_mut() {

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -223,6 +223,45 @@ mod tests {
         assert_eq!(part, expected);
     }
 
+    #[test]
+    fn v2_indexpart_is_parsed_with_deleted_at() {
+        let example = r#"{
+            "version":2,
+            "timeline_layers":["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9"],
+            "missing_layers":["This shouldn't fail deserialization"],
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "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],
+            "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()]),
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                    file_size: 25600000,
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                    // serde_json should always parse this but this might be a double with jq for
+                    // example.
+                    file_size: 9007199254741001,
+                })
+            ]),
+            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(),
+            deleted_at: Some(chrono::NaiveDateTime::parse_from_str(
+                "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap())
+        };
+
+        let part = serde_json::from_str::<IndexPart>(example).unwrap();
+        assert_eq!(part, expected);
+    }
+
     #[test]
     fn empty_layers_are_parsed() {
         let empty_layers_json = r#"{

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -62,12 +62,11 @@ pub(super) async fn upload_timeline_layer<'a>(
     let source_file = match source_file_res {
         Ok(source_file) => source_file,
         Err(e) if e.kind() == ErrorKind::NotFound => {
-            // In some situations we might run into the underlying file being deleted by
-            // e.g. compaction before the uploader gets to it. In that instance, we don't
-            // want to retry the error: a deleted file won't come back. In theory, the
-            // file might not have been written in the first place, which also indicates
-            // a bug. Still log the situation so that we can keep an eye on it.
-            // See https://github.com/neondatabase/neon/issues/4526
+            // If we encounter this arm, it wasn't intended, but it's also not
+            // a big problem, if it's because the file was deleted before an
+            // upload. However, a nonexistent file can also be indicative of
+            // something worse, like when a file is scheduled for upload before
+            // it has been written to disk yet.
             info!(path = %source_path.display(), "File to upload doesn't exist. Likely the file has been deleted and an upload is not required any more.");
             return Ok(());
         }

pageserver/src/tenant/size.rs

@@ -110,11 +110,11 @@ pub struct TimelineInputs {
 ///
 /// Tenant size does not consider the latest state, but only the state until next_gc_cutoff, which
 /// is updated on-demand, during the start of this calculation and separate from the
-/// [`Timeline::latest_gc_cutoff`].
+/// [`TimelineInputs::latest_gc_cutoff`].
 ///
 /// For timelines in general:
 ///
-/// ```ignore
+/// ```text
 /// 0-----|---------|----|------------| · · · · · |·> lsn
 ///   initdb_lsn  branchpoints*  next_gc_cutoff  latest
 /// ```

pageserver/src/tenant/span.rs

@@ -11,10 +11,7 @@ pub(crate) static TENANT_ID_EXTRACTOR: once_cell::sync::Lazy<MultiNameExtractor<
 #[cfg(debug_assertions)]
 #[track_caller]
 pub(crate) fn debug_assert_current_span_has_tenant_id() {
-    if let Err(missing) = check_fields_present([&*TENANT_ID_EXTRACTOR]) {
-        panic!(
-            "missing extractors: {:?}",
-            missing.into_iter().map(|e| e.name()).collect::<Vec<_>>()
-        )
+    if let Err(missing) = check_fields_present!([&*TENANT_ID_EXTRACTOR]) {
+        panic!("missing extractors: {missing:?}")
     }
 }

pageserver/src/tenant/storage_layer.rs

@@ -9,7 +9,7 @@ mod remote_layer;
 
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
-use crate::repository::{Key, Value};
+use crate::repository::Key;
 use crate::task_mgr::TaskKind;
 use crate::walrecord::NeonWalRecord;
 use anyhow::Result;
@@ -34,7 +34,7 @@ use utils::{
     lsn::Lsn,
 };
 
-pub use delta_layer::{DeltaLayer, DeltaLayerWriter};
+pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
 pub use filename::{DeltaFileName, ImageFileName, LayerFileName};
 pub use image_layer::{ImageLayer, ImageLayerWriter};
 pub use inmemory_layer::InMemoryLayer;
@@ -162,6 +162,9 @@ impl LayerAccessStats {
     /// The caller is responsible for recording a residence event
     /// using [`record_residence_event`] before calling `latest_activity`.
     /// If they don't, [`latest_activity`] will return `None`.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
+    /// [`latest_activity`]: Self::latest_activity
     pub(crate) fn empty_will_record_residence_event_later() -> Self {
         LayerAccessStats(Mutex::default())
     }
@@ -169,6 +172,9 @@ impl LayerAccessStats {
     /// Create an empty stats object and record a [`LayerLoad`] event with the given residence status.
     ///
     /// See [`record_residence_event`] for why you need to do this while holding the layer map lock.
+    ///
+    /// [`LayerLoad`]: LayerResidenceEventReason::LayerLoad
+    /// [`record_residence_event`]: Self::record_residence_event
     pub(crate) fn for_loading_layer(
         layer_map_lock_held_witness: &LayerManager,
         status: LayerResidenceStatus,
@@ -187,6 +193,8 @@ impl LayerAccessStats {
     /// The `new_status` is not recorded in `self`.
     ///
     /// See [`record_residence_event`] for why you need to do this while holding the layer map lock.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
     pub(crate) fn clone_for_residence_change(
         &self,
         layer_map_lock_held_witness: &LayerManager,
@@ -294,11 +302,13 @@ impl LayerAccessStats {
     /// implementation error. This function logs a rate-limited warning in that case.
     ///
     /// TODO: use type system to avoid the need for `fallback`.
-    /// The approach in https://github.com/neondatabase/neon/pull/3775
+    /// The approach in <https://github.com/neondatabase/neon/pull/3775>
     /// could be used to enforce that a residence event is recorded
     /// before a layer is added to the layer map. We could also have
     /// a layer wrapper type that holds the LayerAccessStats, and ensure
     /// that that type can only be produced by inserting into the layer map.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
     pub(crate) fn latest_activity(&self) -> Option<SystemTime> {
         let locked = self.0.lock().unwrap();
         let inner = &locked.for_eviction_policy;
@@ -323,12 +333,13 @@ impl LayerAccessStats {
 }
 
 /// Supertrait of the [`Layer`] trait that captures the bare minimum interface
-/// required by [`LayerMap`].
+/// required by [`LayerMap`](super::layer_map::LayerMap).
 ///
 /// All layers should implement a minimal `std::fmt::Debug` without tenant or
 /// timeline names, because those are known in the context of which the layers
 /// are used in (timeline).
-pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
+#[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>;
 
@@ -358,7 +369,7 @@ pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
     /// is available. If this returns ValueReconstructResult::Continue, look up
     /// the predecessor layer and call again with the same 'reconstruct_data' to
     /// collect more data.
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -367,15 +378,9 @@ pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
     ) -> Result<ValueReconstructResult>;
 
     /// Dump summary of the contents of the layer to stdout
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()>;
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()>;
 }
 
-/// Returned by [`Layer::iter`]
-pub type LayerIter<'i> = Box<dyn Iterator<Item = Result<(Key, Lsn, Value)>> + 'i + Send>;
-
-/// Returned by [`Layer::key_iter`]
-pub type LayerKeyIter<'i> = Box<dyn Iterator<Item = (Key, Lsn, u64)> + 'i + Send>;
-
 /// Get a layer descriptor from a layer.
 pub trait AsLayerDesc {
     /// Get the layer descriptor.
@@ -416,15 +421,6 @@ pub trait PersistentLayer: Layer + AsLayerDesc {
     // `None` for `RemoteLayer`.
     fn local_path(&self) -> Option<PathBuf>;
 
-    /// Iterate through all keys and values stored in the layer
-    fn iter(&self, ctx: &RequestContext) -> Result<LayerIter<'_>>;
-
-    /// Iterate through all keys stored in the layer. Returns key, lsn and value size
-    /// It is used only for compaction and so is currently implemented only for DeltaLayer
-    fn key_iter(&self, _ctx: &RequestContext) -> Result<LayerKeyIter<'_>> {
-        panic!("Not implemented")
-    }
-
     /// Permanently remove this layer from disk.
     fn delete_resident_layer_file(&self) -> Result<()>;
 
@@ -432,6 +428,10 @@ pub trait PersistentLayer: Layer + AsLayerDesc {
         None
     }
 
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        None
+    }
+
     fn is_remote_layer(&self) -> bool {
         false
     }

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -7,14 +7,18 @@
 //! must be page images or WAL records with the 'will_init' flag set, so that
 //! they can be replayed without referring to an older page version.
 //!
-//! The delta files are stored in timelines/<timeline_id> directory.  Currently,
+//! The delta files are stored in `timelines/<timeline_id>` directory.  Currently,
 //! there are no subdirectories, and each delta file is named like this:
 //!
+//! ```text
 //!    <key start>-<key end>__<start LSN>-<end LSN>
+//! ```
 //!
 //! For example:
 //!
+//! ```text
 //!    000000067F000032BE0000400000000020B6-000000067F000032BE0000400000000030B6__000000578C6B29-0000000057A50051
+//! ```
 //!
 //! Every delta file consists of three parts: "summary", "index", and
 //! "values". The summary is a fixed size header at the beginning of the file,
@@ -27,7 +31,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::{PageReadGuard, PAGE_SZ};
 use crate::repository::{Key, Value, KEY_SIZE};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -47,6 +51,7 @@ use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
+use std::sync::Arc;
 use tracing::*;
 
 use utils::{
@@ -56,8 +61,8 @@ use utils::{
 };
 
 use super::{
-    AsLayerDesc, DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, LayerIter,
-    LayerKeyIter, PathOrConf, PersistentLayerDesc,
+    AsLayerDesc, DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, PathOrConf,
+    PersistentLayerDesc,
 };
 
 ///
@@ -184,7 +189,7 @@ pub struct DeltaLayer {
 
     access_stats: LayerAccessStats,
 
-    inner: OnceCell<DeltaLayerInner>,
+    inner: OnceCell<Arc<DeltaLayerInner>>,
 }
 
 impl std::fmt::Debug for DeltaLayer {
@@ -218,9 +223,10 @@ impl std::fmt::Debug for DeltaLayerInner {
     }
 }
 
+#[async_trait::async_trait]
 impl Layer for DeltaLayer {
     /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         println!(
             "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} size {} ----",
             self.desc.tenant_id,
@@ -250,12 +256,12 @@ impl Layer for DeltaLayer {
             file,
         );
 
-        tree_reader.dump()?;
+        tree_reader.dump().await?;
 
-        let mut cursor = file.block_cursor();
+        let cursor = file.block_cursor();
 
         // A subroutine to dump a single blob
-        let mut dump_blob = |blob_ref: BlobRef| -> anyhow::Result<String> {
+        let dump_blob = |blob_ref: BlobRef| -> anyhow::Result<String> {
             let buf = cursor.read_blob(blob_ref.pos())?;
             let val = Value::des(&buf)?;
             let desc = match val {
@@ -295,7 +301,7 @@ impl Layer for DeltaLayer {
         Ok(())
     }
 
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -337,7 +343,7 @@ impl Layer for DeltaLayer {
             })?;
 
             // Ok, 'offsets' now contains the offsets of all the entries we need to read
-            let mut cursor = file.block_cursor();
+            let cursor = file.block_cursor();
             let mut buf = Vec::new();
             for (entry_lsn, pos) in offsets {
                 cursor.read_blob_into_buf(pos, &mut buf).with_context(|| {
@@ -410,27 +416,14 @@ impl AsLayerDesc for DeltaLayer {
 }
 
 impl PersistentLayer for DeltaLayer {
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        Some(self)
+    }
+
     fn local_path(&self) -> Option<PathBuf> {
         Some(self.path())
     }
 
-    fn iter(&self, ctx: &RequestContext) -> Result<LayerIter<'_>> {
-        let inner = self
-            .load(LayerAccessKind::KeyIter, ctx)
-            .context("load delta layer")?;
-        Ok(match DeltaValueIter::new(inner) {
-            Ok(iter) => Box::new(iter),
-            Err(err) => Box::new(std::iter::once(Err(err))),
-        })
-    }
-
-    fn key_iter(&self, ctx: &RequestContext) -> Result<LayerKeyIter<'_>> {
-        let inner = self.load(LayerAccessKind::KeyIter, ctx)?;
-        Ok(Box::new(
-            DeltaKeyIter::new(inner).context("Layer index is corrupted")?,
-        ))
-    }
-
     fn delete_resident_layer_file(&self) -> Result<()> {
         // delete underlying file
         fs::remove_file(self.path())?;
@@ -500,7 +493,11 @@ impl DeltaLayer {
     /// Open the underlying file and read the metadata into memory, if it's
     /// not loaded already.
     ///
-    fn load(&self, access_kind: LayerAccessKind, ctx: &RequestContext) -> Result<&DeltaLayerInner> {
+    fn load(
+        &self,
+        access_kind: LayerAccessKind,
+        ctx: &RequestContext,
+    ) -> Result<&Arc<DeltaLayerInner>> {
         self.access_stats
             .record_access(access_kind, ctx.task_kind());
         // Quick exit if already loaded
@@ -509,7 +506,7 @@ impl DeltaLayer {
             .with_context(|| format!("Failed to load delta layer {}", self.path().display()))
     }
 
-    fn load_inner(&self) -> Result<DeltaLayerInner> {
+    fn load_inner(&self) -> Result<Arc<DeltaLayerInner>> {
         let path = self.path();
 
         let file = VirtualFile::open(&path)
@@ -544,11 +541,11 @@ impl DeltaLayer {
 
         debug!("loaded from {}", &path.display());
 
-        Ok(DeltaLayerInner {
+        Ok(Arc::new(DeltaLayerInner {
             file,
             index_start_blk: actual_summary.index_start_blk,
             index_root_blk: actual_summary.index_root_blk,
-        })
+        }))
     }
 
     /// Create a DeltaLayer struct representing an existing file on disk.
@@ -613,6 +610,24 @@ impl DeltaLayer {
             &self.layer_name(),
         )
     }
+
+    /// Obtains all keys and value references stored in the layer
+    ///
+    /// The value can be obtained via the [`ValueRef::load`] function.
+    pub fn load_val_refs(&self, ctx: &RequestContext) -> Result<Vec<(Key, Lsn, ValueRef)>> {
+        let inner = self
+            .load(LayerAccessKind::KeyIter, ctx)
+            .context("load delta layer")?;
+        DeltaLayerInner::load_val_refs(inner).context("Layer index is corrupted")
+    }
+
+    /// Loads all keys stored in the layer. Returns key, lsn and value size.
+    pub fn load_keys(&self, ctx: &RequestContext) -> Result<Vec<(Key, Lsn, u64)>> {
+        let inner = self
+            .load(LayerAccessKind::KeyIter, ctx)
+            .context("load delta layer keys")?;
+        inner.load_keys().context("Layer index is corrupted")
+    }
 }
 
 /// A builder object for constructing a new delta layer.
@@ -800,7 +815,7 @@ impl DeltaLayerWriterInner {
 ///
 /// # Note
 ///
-/// As described in https://github.com/neondatabase/neon/issues/2650, it's
+/// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
 /// possible for the writer to drop before `finish` is actually called. So this
 /// could lead to odd temporary files in the directory, exhausting file system.
 /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
@@ -883,121 +898,41 @@ impl Drop for DeltaLayerWriter {
     }
 }
 
-///
-/// Iterator over all key-value pairse stored in a delta layer
-///
-/// FIXME: This creates a Vector to hold the offsets of all key value pairs.
-/// That takes up quite a lot of memory. Should do this in a more streaming
-/// fashion.
-///
-struct DeltaValueIter<'a> {
-    all_offsets: Vec<(DeltaKey, BlobRef)>,
-    next_idx: usize,
-    reader: BlockCursor<Adapter<'a>>,
-}
-
-struct Adapter<'a>(&'a DeltaLayerInner);
-
-impl<'a> BlockReader for Adapter<'a> {
-    type BlockLease = PageReadGuard<'static>;
-
-    fn read_blk(&self, blknum: u32) -> Result<Self::BlockLease, std::io::Error> {
-        self.0.file.read_blk(blknum)
-    }
-}
-
-impl<'a> Iterator for DeltaValueIter<'a> {
-    type Item = Result<(Key, Lsn, Value)>;
-
-    fn next(&mut self) -> Option<Self::Item> {
-        self.next_res().transpose()
-    }
-}
-
-impl<'a> DeltaValueIter<'a> {
-    fn new(inner: &'a DeltaLayerInner) -> Result<Self> {
-        let file = &inner.file;
+impl DeltaLayerInner {
+    fn load_val_refs(this: &Arc<DeltaLayerInner>) -> Result<Vec<(Key, Lsn, ValueRef)>> {
+        let file = &this.file;
         let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            inner.index_start_blk,
-            inner.index_root_blk,
+            this.index_start_blk,
+            this.index_root_blk,
             file,
         );
 
-        let mut all_offsets: Vec<(DeltaKey, BlobRef)> = Vec::new();
+        let mut all_offsets = Vec::<(Key, Lsn, ValueRef)>::new();
         tree_reader.visit(
             &[0u8; DELTA_KEY_SIZE],
             VisitDirection::Forwards,
             |key, value| {
-                all_offsets.push((DeltaKey::from_slice(key), BlobRef(value)));
+                let delta_key = DeltaKey::from_slice(key);
+                let val_ref = ValueRef {
+                    blob_ref: BlobRef(value),
+                    reader: BlockCursor::new(Adapter(this.clone())),
+                };
+                all_offsets.push((delta_key.key(), delta_key.lsn(), val_ref));
                 true
             },
         )?;
 
-        let iter = DeltaValueIter {
-            all_offsets,
-            next_idx: 0,
-            reader: BlockCursor::new(Adapter(inner)),
-        };
-
-        Ok(iter)
+        Ok(all_offsets)
     }
-
-    fn next_res(&mut self) -> Result<Option<(Key, Lsn, Value)>> {
-        if self.next_idx < self.all_offsets.len() {
-            let (delta_key, blob_ref) = &self.all_offsets[self.next_idx];
-
-            let key = delta_key.key();
-            let lsn = delta_key.lsn();
-
-            let buf = self.reader.read_blob(blob_ref.pos())?;
-            let val = Value::des(&buf)?;
-            self.next_idx += 1;
-            Ok(Some((key, lsn, val)))
-        } else {
-            Ok(None)
-        }
-    }
-}
-///
-/// Iterator over all keys stored in a delta layer
-///
-/// FIXME: This creates a Vector to hold all keys.
-/// That takes up quite a lot of memory. Should do this in a more streaming
-/// fashion.
-///
-struct DeltaKeyIter {
-    all_keys: Vec<(DeltaKey, u64)>,
-    next_idx: usize,
-}
-
-impl Iterator for DeltaKeyIter {
-    type Item = (Key, Lsn, u64);
-
-    fn next(&mut self) -> Option<Self::Item> {
-        if self.next_idx < self.all_keys.len() {
-            let (delta_key, size) = &self.all_keys[self.next_idx];
-
-            let key = delta_key.key();
-            let lsn = delta_key.lsn();
-
-            self.next_idx += 1;
-            Some((key, lsn, *size))
-        } else {
-            None
-        }
-    }
-}
-
-impl<'a> DeltaKeyIter {
-    fn new(inner: &'a DeltaLayerInner) -> Result<Self> {
-        let file = &inner.file;
+    fn load_keys(&self) -> Result<Vec<(Key, Lsn, u64)>> {
+        let file = &self.file;
         let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            inner.index_start_blk,
-            inner.index_root_blk,
+            self.index_start_blk,
+            self.index_root_blk,
             file,
         );
 
-        let mut all_keys: Vec<(DeltaKey, u64)> = Vec::new();
+        let mut all_keys: Vec<(Key, Lsn, u64)> = Vec::new();
         tree_reader.visit(
             &[0u8; DELTA_KEY_SIZE],
             VisitDirection::Forwards,
@@ -1005,46 +940,48 @@ impl<'a> DeltaKeyIter {
                 let delta_key = DeltaKey::from_slice(key);
                 let pos = BlobRef(value).pos();
                 if let Some(last) = all_keys.last_mut() {
-                    if last.0.key() == delta_key.key() {
+                    if last.0 == delta_key.key() {
                         return true;
                     } else {
                         // subtract offset of new key BLOB and first blob of this key
                         // to get total size if values associated with this key
-                        let first_pos = last.1;
-                        last.1 = pos - first_pos;
+                        let first_pos = last.2;
+                        last.2 = pos - first_pos;
                     }
                 }
-                all_keys.push((delta_key, pos));
+                all_keys.push((delta_key.key(), delta_key.lsn(), pos));
                 true
             },
         )?;
         if let Some(last) = all_keys.last_mut() {
             // Last key occupies all space till end of layer
-            last.1 = std::fs::metadata(&file.file.path)?.len() - last.1;
+            last.2 = std::fs::metadata(&file.file.path)?.len() - last.2;
         }
-        let iter = DeltaKeyIter {
-            all_keys,
-            next_idx: 0,
-        };
-
-        Ok(iter)
+        Ok(all_keys)
     }
 }
 
-#[cfg(test)]
-mod test {
-    use super::DeltaKeyIter;
-    use super::DeltaLayer;
-    use super::DeltaValueIter;
+/// Reference to an on-disk value
+pub struct ValueRef {
+    blob_ref: BlobRef,
+    reader: BlockCursor<Adapter>,
+}
 
-    // We will soon need the iters to be send in the compaction code.
-    // Cf https://github.com/neondatabase/neon/pull/4462#issuecomment-1587398883
-    // Cf https://github.com/neondatabase/neon/issues/4471
-    #[test]
-    fn is_send() {
-        fn assert_send<T: Send>() {}
-        assert_send::<DeltaLayer>();
-        assert_send::<DeltaValueIter>();
-        assert_send::<DeltaKeyIter>();
+impl ValueRef {
+    /// Loads the value from disk
+    pub fn load(&self) -> Result<Value> {
+        let buf = self.reader.read_blob(self.blob_ref.pos())?;
+        let val = Value::des(&buf)?;
+        Ok(val)
+    }
+}
+
+struct Adapter(Arc<DeltaLayerInner>);
+
+impl BlockReader for Adapter {
+    type BlockLease = PageReadGuard<'static>;
+
+    fn read_blk(&self, blknum: u32) -> Result<Self::BlockLease, std::io::Error> {
+        self.0.file.read_blk(blknum)
     }
 }

pageserver/src/tenant/storage_layer/filename.rs

@@ -57,8 +57,9 @@ impl Ord for DeltaFileName {
 
 /// Represents the filename of a DeltaLayer
 ///
+/// ```text
 ///    <key start>-<key end>__<LSN start>-<LSN end>
-///
+/// ```
 impl DeltaFileName {
     ///
     /// Parse a string as a delta file name. Returns None if the filename does not
@@ -162,7 +163,9 @@ impl ImageFileName {
 ///
 /// Represents the filename of an ImageLayer
 ///
+/// ```text
 ///    <key start>-<key end>__<LSN>
+/// ```
 impl ImageFileName {
     ///
     /// Parse a string as an image file name. Returns None if the filename does not

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -7,11 +7,15 @@
 //! timelines/<timeline_id> directory.  Currently, there are no
 //! subdirectories, and each image layer file is named like this:
 //!
+//! ```text
 //!    <key start>-<key end>__<LSN>
+//! ```
 //!
 //! For example:
 //!
+//! ```text
 //!    000000067F000032BE0000400000000070B6-000000067F000032BE0000400000000080B6__00000000346BC568
+//! ```
 //!
 //! Every image layer file consists of three parts: "summary",
 //! "index", and "values".  The summary is a fixed size header at the
@@ -23,7 +27,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::repository::{Key, KEY_SIZE};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
 use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -34,6 +38,7 @@ use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 use anyhow::{bail, ensure, Context, Result};
 use bytes::Bytes;
 use hex;
+use once_cell::sync::OnceCell;
 use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
@@ -43,7 +48,6 @@ use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::path::{Path, PathBuf};
-use std::sync::{RwLock, RwLockReadGuard};
 use tracing::*;
 
 use utils::{
@@ -53,9 +57,7 @@ use utils::{
 };
 
 use super::filename::ImageFileName;
-use super::{
-    AsLayerDesc, Layer, LayerAccessStatsReset, LayerIter, PathOrConf, PersistentLayerDesc,
-};
+use super::{AsLayerDesc, Layer, LayerAccessStatsReset, PathOrConf, PersistentLayerDesc};
 
 ///
 /// Header stored in the beginning of the file
@@ -113,7 +115,7 @@ pub struct ImageLayer {
 
     access_stats: LayerAccessStats,
 
-    inner: RwLock<ImageLayerInner>,
+    inner: OnceCell<ImageLayerInner>,
 }
 
 impl std::fmt::Debug for ImageLayer {
@@ -130,30 +132,27 @@ impl std::fmt::Debug for ImageLayer {
 }
 
 pub struct ImageLayerInner {
-    /// If false, the 'index' has not been loaded into memory yet.
-    loaded: bool,
-
     // values copied from summary
     index_start_blk: u32,
     index_root_blk: u32,
 
-    /// Reader object for reading blocks from the file. (None if not loaded yet)
-    file: Option<FileBlockReader<VirtualFile>>,
+    /// Reader object for reading blocks from the file.
+    file: FileBlockReader<VirtualFile>,
 }
 
 impl std::fmt::Debug for ImageLayerInner {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("ImageLayerInner")
-            .field("loaded", &self.loaded)
             .field("index_start_blk", &self.index_start_blk)
             .field("index_root_blk", &self.index_root_blk)
             .finish()
     }
 }
 
+#[async_trait::async_trait]
 impl Layer for ImageLayer {
     /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
         println!(
             "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
             self.desc.tenant_id,
@@ -170,11 +169,11 @@ impl Layer for ImageLayer {
         }
 
         let inner = self.load(LayerAccessKind::Dump, ctx)?;
-        let file = inner.file.as_ref().unwrap();
+        let file = &inner.file;
         let tree_reader =
             DiskBtreeReader::<_, KEY_SIZE>::new(inner.index_start_blk, inner.index_root_blk, file);
 
-        tree_reader.dump()?;
+        tree_reader.dump().await?;
 
         tree_reader.visit(&[0u8; KEY_SIZE], VisitDirection::Forwards, |key, value| {
             println!("key: {} offset {}", hex::encode(key), value);
@@ -185,7 +184,7 @@ impl Layer for ImageLayer {
     }
 
     /// Look up given page in the file
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -198,7 +197,7 @@ impl Layer for ImageLayer {
 
         let inner = self.load(LayerAccessKind::GetValueReconstructData, ctx)?;
 
-        let file = inner.file.as_ref().unwrap();
+        let file = &inner.file;
         let tree_reader = DiskBtreeReader::new(inner.index_start_blk, inner.index_root_blk, file);
 
         let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
@@ -254,10 +253,6 @@ impl PersistentLayer for ImageLayer {
         Some(self.path())
     }
 
-    fn iter(&self, _ctx: &RequestContext) -> Result<LayerIter<'_>> {
-        unimplemented!();
-    }
-
     fn delete_resident_layer_file(&self) -> Result<()> {
         // delete underlying file
         fs::remove_file(self.path())?;
@@ -317,52 +312,26 @@ impl ImageLayer {
     /// Open the underlying file and read the metadata into memory, if it's
     /// not loaded already.
     ///
-    fn load(
-        &self,
-        access_kind: LayerAccessKind,
-        ctx: &RequestContext,
-    ) -> Result<RwLockReadGuard<ImageLayerInner>> {
+    fn load(&self, access_kind: LayerAccessKind, ctx: &RequestContext) -> Result<&ImageLayerInner> {
         self.access_stats
             .record_access(access_kind, ctx.task_kind());
         loop {
-            // Quick exit if already loaded
-            let inner = self.inner.read().unwrap();
-            if inner.loaded {
+            if let Some(inner) = self.inner.get() {
                 return Ok(inner);
             }
-
-            // Need to open the file and load the metadata. Upgrade our lock to
-            // a write lock. (Or rather, release and re-lock in write mode.)
-            drop(inner);
-            let mut inner = self.inner.write().unwrap();
-            if !inner.loaded {
-                self.load_inner(&mut inner).with_context(|| {
-                    format!("Failed to load image layer {}", self.path().display())
-                })?
-            } else {
-                // Another thread loaded it while we were not holding the lock.
-            }
-
-            // We now have the file open and loaded. There's no function to do
-            // that in the std library RwLock, so we have to release and re-lock
-            // in read mode. (To be precise, the lock guard was moved in the
-            // above call to `load_inner`, so it's already been released). And
-            // while we do that, another thread could unload again, so we have
-            // to re-check and retry if that happens.
-            drop(inner);
+            self.inner
+                .get_or_try_init(|| self.load_inner())
+                .with_context(|| format!("Failed to load image layer {}", self.path().display()))?;
         }
     }
 
-    fn load_inner(&self, inner: &mut ImageLayerInner) -> Result<()> {
+    fn load_inner(&self) -> Result<ImageLayerInner> {
         let path = self.path();
 
         // Open the file if it's not open already.
-        if inner.file.is_none() {
-            let file = VirtualFile::open(&path)
-                .with_context(|| format!("Failed to open file '{}'", path.display()))?;
-            inner.file = Some(FileBlockReader::new(file));
-        }
-        let file = inner.file.as_mut().unwrap();
+        let file = VirtualFile::open(&path)
+            .with_context(|| format!("Failed to open file '{}'", path.display()))?;
+        let file = FileBlockReader::new(file);
         let summary_blk = file.read_blk(0)?;
         let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
 
@@ -390,10 +359,11 @@ impl ImageLayer {
             }
         }
 
-        inner.index_start_blk = actual_summary.index_start_blk;
-        inner.index_root_blk = actual_summary.index_root_blk;
-        inner.loaded = true;
-        Ok(())
+        Ok(ImageLayerInner {
+            index_start_blk: actual_summary.index_start_blk,
+            index_root_blk: actual_summary.index_root_blk,
+            file,
+        })
     }
 
     /// Create an ImageLayer struct representing an existing file on disk
@@ -417,12 +387,7 @@ impl ImageLayer {
             ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
             lsn: filename.lsn,
             access_stats,
-            inner: RwLock::new(ImageLayerInner {
-                loaded: false,
-                file: None,
-                index_start_blk: 0,
-                index_root_blk: 0,
-            }),
+            inner: OnceCell::new(),
         }
     }
 
@@ -449,12 +414,7 @@ impl ImageLayer {
             ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
             lsn: summary.lsn,
             access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
-            inner: RwLock::new(ImageLayerInner {
-                file: None,
-                loaded: false,
-                index_start_blk: 0,
-                index_root_blk: 0,
-            }),
+            inner: OnceCell::new(),
         })
     }
 
@@ -615,12 +575,7 @@ impl ImageLayerWriterInner {
             desc,
             lsn: self.lsn,
             access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
-            inner: RwLock::new(ImageLayerInner {
-                loaded: false,
-                file: None,
-                index_start_blk,
-                index_root_blk,
-            }),
+            inner: OnceCell::new(),
         };
 
         // fsync the file
@@ -660,7 +615,7 @@ impl ImageLayerWriterInner {
 ///
 /// # Note
 ///
-/// As described in https://github.com/neondatabase/neon/issues/2650, it's
+/// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
 /// possible for the writer to drop before `finish` is actually called. So this
 /// could lead to odd temporary files in the directory, exhausting file system.
 /// This structure wraps `ImageLayerWriterInner` and also contains `Drop`

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -7,7 +7,7 @@
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::repository::{Key, Value};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter};
+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};
@@ -110,6 +110,7 @@ impl InMemoryLayer {
     }
 }
 
+#[async_trait::async_trait]
 impl Layer for InMemoryLayer {
     fn get_key_range(&self) -> Range<Key> {
         Key::MIN..Key::MAX
@@ -132,7 +133,7 @@ impl Layer for InMemoryLayer {
     }
 
     /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
         let inner = self.inner.read().unwrap();
 
         let end_str = inner
@@ -150,7 +151,7 @@ impl Layer for InMemoryLayer {
             return Ok(());
         }
 
-        let mut cursor = inner.file.block_cursor();
+        let cursor = inner.file.block_cursor();
         let mut buf = Vec::new();
         for (key, vec_map) in inner.index.iter() {
             for (lsn, pos) in vec_map.as_slice() {
@@ -183,7 +184,7 @@ impl Layer for InMemoryLayer {
     }
 
     /// Look up given value in the layer.
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
         &self,
         key: Key,
         lsn_range: Range<Lsn>,
@@ -195,7 +196,7 @@ impl Layer for InMemoryLayer {
 
         let inner = self.inner.read().unwrap();
 
-        let mut reader = inner.file.block_cursor();
+        let reader = inner.file.block_cursor();
 
         // Scan the page versions backwards, starting from `lsn`.
         if let Some(vec_map) = inner.index.get(&key) {
@@ -353,7 +354,7 @@ impl InMemoryLayer {
 
         let mut buf = Vec::new();
 
-        let mut cursor = inner.file.block_cursor();
+        let cursor = inner.file.block_cursor();
 
         let mut keys: Vec<(&Key, &VecMap<Lsn, u64>)> = inner.index.iter().collect();
         keys.sort_by_key(|k| k.0);

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -20,12 +20,12 @@ use utils::{
 
 use super::filename::{DeltaFileName, ImageFileName};
 use super::{
-    AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset, LayerIter,
-    LayerKeyIter, LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
+    AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset,
+    LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
 };
 
 /// RemoteLayer is a not yet downloaded [`ImageLayer`] or
-/// [`crate::storage_layer::DeltaLayer`].
+/// [`DeltaLayer`](super::DeltaLayer).
 ///
 /// RemoteLayer might be downloaded on-demand during operations which are
 /// allowed download remote layers and during which, it gets replaced with a
@@ -50,6 +50,8 @@ pub struct RemoteLayer {
     /// It is very unlikely to accumulate these in the Timeline's LayerMap, but having this avoids
     /// a possible fast loop between `Timeline::get_reconstruct_data` and
     /// `Timeline::download_remote_layer`, which also logs.
+    ///
+    /// [`ongoing_download`]: Self::ongoing_download
     pub(crate) download_replacement_failure: std::sync::atomic::AtomicBool,
 }
 
@@ -63,8 +65,9 @@ impl std::fmt::Debug for RemoteLayer {
     }
 }
 
+#[async_trait::async_trait]
 impl Layer for RemoteLayer {
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
         &self,
         _key: Key,
         _lsn_range: Range<Lsn>,
@@ -75,7 +78,7 @@ impl Layer for RemoteLayer {
     }
 
     /// debugging function to print out the contents of the layer
-    fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
+    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,
@@ -126,14 +129,6 @@ impl PersistentLayer for RemoteLayer {
         None
     }
 
-    fn iter(&self, _ctx: &RequestContext) -> Result<LayerIter<'_>> {
-        bail!("cannot iterate a remote layer");
-    }
-
-    fn key_iter(&self, _ctx: &RequestContext) -> Result<LayerKeyIter<'_>> {
-        bail!("cannot iterate a remote layer");
-    }
-
     fn delete_resident_layer_file(&self) -> Result<()> {
         bail!("remote layer has no layer file");
     }

pageserver/src/tenant/tasks.rs

@@ -111,7 +111,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 Duration::from_secs(10)
             } else {
                 // Run compaction
-                if let Err(e) = tenant.compaction_iteration(&ctx).await {
+                if let Err(e) = tenant.compaction_iteration(&cancel, &ctx).await {
                     error!("Compaction failed, retrying in {:?}: {e:?}", wait_duration);
                     wait_duration
                 } else {
@@ -122,12 +122,12 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             warn_when_period_overrun(started_at.elapsed(), period, "compaction");
 
             // Sleep
-            tokio::select! {
-                _ = cancel.cancelled() => {
-                    info!("received cancellation request during idling");
-                    break;
-                },
-                _ = tokio::time::sleep(sleep_duration) => {},
+            if tokio::time::timeout(sleep_duration, cancel.cancelled())
+                .await
+                .is_ok()
+            {
+                info!("received cancellation request during idling");
+                break;
             }
         }
     }
@@ -196,12 +196,12 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
             warn_when_period_overrun(started_at.elapsed(), period, "gc");
 
             // Sleep
-            tokio::select! {
-                _ = cancel.cancelled() => {
-                    info!("received cancellation request during idling");
-                    break;
-                },
-                _ = tokio::time::sleep(sleep_duration) => {},
+            if tokio::time::timeout(sleep_duration, cancel.cancelled())
+                .await
+                .is_ok()
+            {
+                info!("received cancellation request during idling");
+                break;
             }
         }
     }
@@ -263,9 +263,9 @@ pub(crate) async fn random_init_delay(
         rng.gen_range(Duration::ZERO..=period)
     };
 
-    tokio::select! {
-        _ = cancel.cancelled() => Err(Cancelled),
-        _ = tokio::time::sleep(d) => Ok(()),
+    match tokio::time::timeout(d, cancel.cancelled()).await {
+        Ok(_) => Err(Cancelled),
+        Err(_) => Ok(()),
     }
 }
 

pageserver/src/tenant/timeline/delete.rs

@@ -0,0 +1,576 @@
+use std::{
+    ops::{Deref, DerefMut},
+    sync::Arc,
+};
+
+use anyhow::Context;
+use pageserver_api::models::TimelineState;
+use tokio::sync::OwnedMutexGuard;
+use tracing::{debug, error, info, instrument, warn, Instrument, Span};
+use utils::{
+    crashsafe, fs_ext,
+    id::{TenantId, TimelineId},
+};
+
+use crate::{
+    config::PageServerConf,
+    task_mgr::{self, TaskKind},
+    tenant::{
+        metadata::TimelineMetadata,
+        remote_timeline_client::{
+            self, PersistIndexPartWithDeletedFlagError, RemoteTimelineClient,
+        },
+        CreateTimelineCause, DeleteTimelineError, Tenant,
+    },
+    InitializationOrder,
+};
+
+use super::Timeline;
+
+/// Now that the Timeline is in Stopping state, request all the related tasks to shut down.
+async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
+    // Stop the walreceiver first.
+    debug!("waiting for wal receiver to shutdown");
+    let maybe_started_walreceiver = { timeline.walreceiver.lock().unwrap().take() };
+    if let Some(walreceiver) = maybe_started_walreceiver {
+        walreceiver.stop().await;
+    }
+    debug!("wal receiver shutdown confirmed");
+
+    // Prevent new uploads from starting.
+    if let Some(remote_client) = timeline.remote_client.as_ref() {
+        let res = remote_client.stop();
+        match res {
+            Ok(()) => {}
+            Err(e) => match e {
+                remote_timeline_client::StopError::QueueUninitialized => {
+                    // This case shouldn't happen currently because the
+                    // load and attach code bails out if _any_ of the timeline fails to fetch its IndexPart.
+                    // That is, before we declare the Tenant as Active.
+                    // But we only allow calls to delete_timeline on Active tenants.
+                    return Err(DeleteTimelineError::Other(anyhow::anyhow!("upload queue is uninitialized, likely the timeline was in Broken state prior to this call because it failed to fetch IndexPart during load or attach, check the logs")));
+                }
+            },
+        }
+    }
+
+    // Stop & wait for the remaining timeline tasks, including upload tasks.
+    // NB: This and other delete_timeline calls do not run as a task_mgr task,
+    //     so, they are not affected by this shutdown_tasks() call.
+    info!("waiting for timeline tasks to shutdown");
+    task_mgr::shutdown_tasks(None, Some(timeline.tenant_id), Some(timeline.timeline_id)).await;
+
+    fail::fail_point!("timeline-delete-before-index-deleted-at", |_| {
+        Err(anyhow::anyhow!(
+            "failpoint: timeline-delete-before-index-deleted-at"
+        ))?
+    });
+    Ok(())
+}
+
+/// Mark timeline as deleted in S3 so we won't pick it up next time
+/// during attach or pageserver restart.
+/// See comment in persist_index_part_with_deleted_flag.
+async fn set_deleted_in_remote_index(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
+    if let Some(remote_client) = timeline.remote_client.as_ref() {
+        match remote_client.persist_index_part_with_deleted_flag().await {
+            // If we (now, or already) marked it successfully as deleted, we can proceed
+            Ok(()) | Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(_)) => (),
+            // Bail out otherwise
+            //
+            // AlreadyInProgress shouldn't happen, because the 'delete_lock' prevents
+            // two tasks from performing the deletion at the same time. The first task
+            // that starts deletion should run it to completion.
+            Err(e @ PersistIndexPartWithDeletedFlagError::AlreadyInProgress(_))
+            | Err(e @ PersistIndexPartWithDeletedFlagError::Other(_)) => {
+                return Err(DeleteTimelineError::Other(anyhow::anyhow!(e)));
+            }
+        }
+    }
+    Ok(())
+}
+
+// We delete local files first, so if pageserver restarts after local files deletion then remote deletion is not continued.
+// This can be solved with inversion of these steps. But even if these steps are inverted then, when index_part.json
+// gets deleted there is no way to distinguish between "this timeline is good, we just didnt upload it to remote"
+// and "this timeline is deleted we should continue with removal of local state". So to avoid the ambiguity we use a mark file.
+// After index part is deleted presence of this mark file indentifies that it was a deletion intention.
+// So we can just remove the mark file.
+async fn create_delete_mark(
+    conf: &PageServerConf,
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+) -> Result<(), DeleteTimelineError> {
+    fail::fail_point!("timeline-delete-before-delete-mark", |_| {
+        Err(anyhow::anyhow!(
+            "failpoint: timeline-delete-before-delete-mark"
+        ))?
+    });
+    let marker_path = conf.timeline_delete_mark_file_path(tenant_id, timeline_id);
+
+    // Note: we're ok to replace existing file.
+    let _ = std::fs::OpenOptions::new()
+        .write(true)
+        .create(true)
+        .open(&marker_path)
+        .with_context(|| format!("could not create delete marker file {marker_path:?}"))?;
+
+    crashsafe::fsync_file_and_parent(&marker_path).context("sync_mark")?;
+    Ok(())
+}
+
+/// Grab the layer_removal_cs lock, and actually perform the deletion.
+///
+/// This lock prevents prevents GC or compaction from running at the same time.
+/// The GC task doesn't register itself with the timeline it's operating on,
+/// so it might still be running even though we called `shutdown_tasks`.
+///
+/// Note that there are still other race conditions between
+/// GC, compaction and timeline deletion. See
+/// <https://github.com/neondatabase/neon/issues/2671>
+///
+/// No timeout here, GC & Compaction should be responsive to the
+/// `TimelineState::Stopping` change.
+async fn delete_local_layer_files(
+    conf: &PageServerConf,
+    tenant_id: TenantId,
+    timeline: &Timeline,
+) -> anyhow::Result<()> {
+    info!("waiting for layer_removal_cs.lock()");
+    let layer_removal_guard = timeline.layer_removal_cs.lock().await;
+    info!("got layer_removal_cs.lock(), deleting layer files");
+
+    // NB: storage_sync upload tasks that reference these layers have been cancelled
+    //     by the caller.
+
+    let local_timeline_directory = conf.timeline_path(&tenant_id, &timeline.timeline_id);
+
+    fail::fail_point!("timeline-delete-before-rm", |_| {
+        Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
+    });
+
+    // NB: This need not be atomic because the deleted flag in the IndexPart
+    // will be observed during tenant/timeline load. The deletion will be resumed there.
+    //
+    // For configurations without remote storage, we guarantee crash-safety by persising delete mark file.
+    //
+    // Note that here we do not bail out on std::io::ErrorKind::NotFound.
+    // This can happen if we're called a second time, e.g.,
+    // because of a previous failure/cancellation at/after
+    // failpoint timeline-delete-after-rm.
+    //
+    // It can also happen if we race with tenant detach, because,
+    // it doesn't grab the layer_removal_cs lock.
+    //
+    // For now, log and continue.
+    // warn! level is technically not appropriate for the
+    // first case because we should expect retries to happen.
+    // But the error is so rare, it seems better to get attention if it happens.
+    //
+    // Note that metadata removal is skipped, this is not technically needed,
+    // but allows to reuse timeline loading code during resumed deletion.
+    // (we always expect that metadata is in place when timeline is being loaded)
+
+    #[cfg(feature = "testing")]
+    let mut counter = 0;
+
+    // Timeline directory may not exist if we failed to delete mark file and request was retried.
+    if !local_timeline_directory.exists() {
+        return Ok(());
+    }
+
+    let metadata_path = conf.metadata_path(&tenant_id, &timeline.timeline_id);
+
+    for entry in walkdir::WalkDir::new(&local_timeline_directory).contents_first(true) {
+        #[cfg(feature = "testing")]
+        {
+            counter += 1;
+            if counter == 2 {
+                fail::fail_point!("timeline-delete-during-rm", |_| {
+                    Err(anyhow::anyhow!("failpoint: timeline-delete-during-rm"))?
+                });
+            }
+        }
+
+        let entry = entry?;
+        if entry.path() == metadata_path {
+            debug!("found metadata, skipping");
+            continue;
+        }
+
+        if entry.path() == local_timeline_directory {
+            // Keeping directory because metedata file is still there
+            debug!("found timeline dir itself, skipping");
+            continue;
+        }
+
+        let metadata = match entry.metadata() {
+            Ok(metadata) => metadata,
+            Err(e) => {
+                if crate::is_walkdir_io_not_found(&e) {
+                    warn!(
+                        timeline_dir=?local_timeline_directory,
+                        path=?entry.path().display(),
+                        "got not found err while removing timeline dir, proceeding anyway"
+                    );
+                    continue;
+                }
+                anyhow::bail!(e);
+            }
+        };
+
+        let r = if metadata.is_dir() {
+            // There shouldnt be any directories inside timeline dir as of current layout.
+            tokio::fs::remove_dir(entry.path()).await
+        } else {
+            tokio::fs::remove_file(entry.path()).await
+        };
+
+        if let Err(e) = r {
+            if e.kind() == std::io::ErrorKind::NotFound {
+                warn!(
+                    timeline_dir=?local_timeline_directory,
+                    path=?entry.path().display(),
+                    "got not found err while removing timeline dir, proceeding anyway"
+                );
+                continue;
+            }
+            anyhow::bail!(anyhow::anyhow!(
+                "Failed to remove: {}. Error: {e}",
+                entry.path().display()
+            ));
+        }
+    }
+
+    info!("finished deleting layer files, releasing layer_removal_cs.lock()");
+    drop(layer_removal_guard);
+
+    fail::fail_point!("timeline-delete-after-rm", |_| {
+        Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm"))?
+    });
+
+    Ok(())
+}
+
+/// Removes remote layers and an index file after them.
+async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<()> {
+    if let Some(remote_client) = &timeline.remote_client {
+        remote_client.delete_all().await.context("delete_all")?
+    };
+
+    Ok(())
+}
+
+// This function removs remaining traces of a timeline on disk.
+// Namely: metadata file, timeline directory, delete mark.
+// Note: io::ErrorKind::NotFound are ignored for metadata and timeline dir.
+// delete mark should be present because it is the last step during deletion.
+// (nothing can fail after its deletion)
+async fn cleanup_remaining_timeline_fs_traces(
+    conf: &PageServerConf,
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+) -> anyhow::Result<()> {
+    // Remove local metadata
+    tokio::fs::remove_file(conf.metadata_path(&tenant_id, &timeline_id))
+        .await
+        .or_else(fs_ext::ignore_not_found)
+        .context("remove metadata")?;
+
+    fail::fail_point!("timeline-delete-after-rm-metadata", |_| {
+        Err(anyhow::anyhow!(
+            "failpoint: timeline-delete-after-rm-metadata"
+        ))?
+    });
+
+    // Remove timeline dir
+    tokio::fs::remove_dir(conf.timeline_path(&tenant_id, &timeline_id))
+        .await
+        .or_else(fs_ext::ignore_not_found)
+        .context("timeline dir")?;
+
+    fail::fail_point!("timeline-delete-after-rm-dir", |_| {
+        Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm-dir"))?
+    });
+
+    // Remove delete mark
+    tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_id, timeline_id))
+        .await
+        .context("remove delete mark")
+}
+
+/// It is important that this gets called when DeletionGuard is being held.
+/// For more context see comments in [`DeleteTimelineFlow::prepare`]
+async fn remove_timeline_from_tenant(
+    tenant: &Tenant,
+    timeline_id: TimelineId,
+    _: &DeletionGuard, // using it as a witness
+) -> anyhow::Result<()> {
+    // Remove the timeline from the map.
+    let mut timelines = tenant.timelines.lock().unwrap();
+    let children_exist = timelines
+        .iter()
+        .any(|(_, entry)| entry.get_ancestor_timeline_id() == Some(timeline_id));
+    // XXX this can happen because `branch_timeline` doesn't check `TimelineState::Stopping`.
+    // We already deleted the layer files, so it's probably best to panic.
+    // (Ideally, above remove_dir_all is atomic so we don't see this timeline after a restart)
+    if children_exist {
+        panic!("Timeline grew children while we removed layer files");
+    }
+
+    timelines
+        .remove(&timeline_id)
+        .expect("timeline that we were deleting was concurrently removed from 'timelines' map");
+
+    drop(timelines);
+
+    Ok(())
+}
+
+/// Orchestrates timeline shut down of all timeline tasks, removes its in-memory structures,
+/// and deletes its data from both disk and s3.
+/// The sequence of steps:
+/// 1. Set deleted_at in remote index part.
+/// 2. Create local mark file.
+/// 3. Delete local files except metadata (it is simpler this way, to be able to reuse timeline initialization code that expects metadata)
+/// 4. Delete remote layers
+/// 5. Delete index part
+/// 6. Delete meta, timeline directory
+/// 7. Delete mark file
+/// It is resumable from any step in case a crash/restart occurs.
+/// There are three entrypoints to the process:
+/// 1. [`DeleteTimelineFlow::run`] this is the main one called by a management api handler.
+/// 2. [`DeleteTimelineFlow::resume_deletion`] is called during restarts when local metadata is still present
+/// and we possibly neeed to continue deletion of remote files.
+/// 3. [`DeleteTimelineFlow::cleanup_remaining_timeline_fs_traces`] is used when we deleted remote
+/// index but still have local metadata, timeline directory and delete mark.
+/// Note the only other place that messes around timeline delete mark is the logic that scans directory with timelines during tenant load.
+#[derive(Default)]
+pub enum DeleteTimelineFlow {
+    #[default]
+    NotStarted,
+    InProgress,
+    Finished,
+}
+
+impl DeleteTimelineFlow {
+    // These steps are run in the context of management api request handler.
+    // Long running steps are continued to run in the background.
+    // NB: If this fails half-way through, and is retried, the retry will go through
+    // all the same steps again. Make sure the code here is idempotent, and don't
+    // error out if some of the shutdown tasks have already been completed!
+    #[instrument(skip_all, fields(tenant_id=%tenant.tenant_id, %timeline_id))]
+    pub async fn run(
+        tenant: &Arc<Tenant>,
+        timeline_id: TimelineId,
+    ) -> Result<(), DeleteTimelineError> {
+        let (timeline, mut guard) = Self::prepare(tenant, timeline_id)?;
+
+        guard.mark_in_progress()?;
+
+        stop_tasks(&timeline).await?;
+
+        set_deleted_in_remote_index(&timeline).await?;
+
+        create_delete_mark(tenant.conf, timeline.tenant_id, timeline.timeline_id).await?;
+
+        fail::fail_point!("timeline-delete-before-schedule", |_| {
+            Err(anyhow::anyhow!(
+                "failpoint: timeline-delete-before-schedule"
+            ))?
+        });
+
+        Self::schedule_background(guard, tenant.conf, Arc::clone(tenant), timeline);
+
+        Ok(())
+    }
+
+    fn mark_in_progress(&mut self) -> anyhow::Result<()> {
+        match self {
+            Self::Finished => anyhow::bail!("Bug. Is in finished state"),
+            Self::InProgress { .. } => { /* We're in a retry */ }
+            Self::NotStarted => { /* Fresh start */ }
+        }
+
+        *self = Self::InProgress;
+
+        Ok(())
+    }
+
+    /// Shortcut to create Timeline in stopping state and spawn deletion task.
+    pub async fn resume_deletion(
+        tenant: Arc<Tenant>,
+        timeline_id: TimelineId,
+        local_metadata: &TimelineMetadata,
+        remote_client: Option<RemoteTimelineClient>,
+        init_order: Option<&InitializationOrder>,
+    ) -> anyhow::Result<()> {
+        // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
+        // RemoteTimelineClient is the only functioning part.
+        let timeline = tenant
+            .create_timeline_struct(
+                timeline_id,
+                local_metadata,
+                None, // Ancestor is not needed for deletion.
+                remote_client,
+                init_order,
+                // Important. We dont pass ancestor above because it can be missing.
+                // Thus we need to skip the validation here.
+                CreateTimelineCause::Delete,
+            )
+            .context("create_timeline_struct")?;
+
+        let mut guard = DeletionGuard(
+            Arc::clone(&timeline.delete_progress)
+                .try_lock_owned()
+                .expect("cannot happen because we're the only owner"),
+        );
+
+        // We meed to do this because when console retries delete request we shouldnt answer with 404
+        // because 404 means successful deletion.
+        {
+            let mut locked = tenant.timelines.lock().unwrap();
+            locked.insert(timeline_id, Arc::clone(&timeline));
+        }
+
+        guard.mark_in_progress()?;
+
+        // Note that delete mark can be missing on resume
+        // because we create delete mark after we set deleted_at in the index part.
+        create_delete_mark(tenant.conf, tenant.tenant_id, timeline_id).await?;
+
+        Self::schedule_background(guard, tenant.conf, tenant, timeline);
+
+        Ok(())
+    }
+
+    pub async fn cleanup_remaining_timeline_fs_traces(
+        tenant: &Tenant,
+        timeline_id: TimelineId,
+    ) -> anyhow::Result<()> {
+        cleanup_remaining_timeline_fs_traces(tenant.conf, tenant.tenant_id, timeline_id).await
+    }
+
+    fn prepare(
+        tenant: &Tenant,
+        timeline_id: TimelineId,
+    ) -> Result<(Arc<Timeline>, DeletionGuard), DeleteTimelineError> {
+        // Note the interaction between this guard and deletion guard.
+        // Here we attempt to lock deletion guard when we're holding a lock on timelines.
+        // This is important because when you take into account `remove_timeline_from_tenant`
+        // we remove timeline from memory when we still hold the deletion guard.
+        // So here when timeline deletion is finished timeline wont be present in timelines map at all
+        // which makes the following sequence impossible:
+        // T1: get preempted right before the try_lock on `Timeline::delete_progress`
+        // T2: do a full deletion, acquire and drop `Timeline::delete_progress`
+        // T1: acquire deletion lock, do another `DeleteTimelineFlow::run`
+        // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
+        let timelines = tenant.timelines.lock().unwrap();
+
+        let timeline = match timelines.get(&timeline_id) {
+            Some(t) => t,
+            None => return Err(DeleteTimelineError::NotFound),
+        };
+
+        // Ensure that there are no child timelines **attached to that pageserver**,
+        // because detach removes files, which will break child branches
+        let children: Vec<TimelineId> = timelines
+            .iter()
+            .filter_map(|(id, entry)| {
+                if entry.get_ancestor_timeline_id() == Some(timeline_id) {
+                    Some(*id)
+                } else {
+                    None
+                }
+            })
+            .collect();
+
+        if !children.is_empty() {
+            return Err(DeleteTimelineError::HasChildren(children));
+        }
+
+        // Note that using try_lock here is important to avoid a deadlock.
+        // Here we take lock on timelines and then the deletion guard.
+        // At the end of the operation we're holding the guard and need to lock timelines map
+        // to remove the timeline from it.
+        // Always if you have two locks that are taken in different order this can result in a deadlock.
+        let delete_lock_guard = DeletionGuard(
+            Arc::clone(&timeline.delete_progress)
+                .try_lock_owned()
+                .map_err(|_| DeleteTimelineError::AlreadyInProgress)?,
+        );
+
+        timeline.set_state(TimelineState::Stopping);
+
+        Ok((Arc::clone(timeline), delete_lock_guard))
+    }
+
+    fn schedule_background(
+        guard: DeletionGuard,
+        conf: &'static PageServerConf,
+        tenant: Arc<Tenant>,
+        timeline: Arc<Timeline>,
+    ) {
+        let tenant_id = timeline.tenant_id;
+        let timeline_id = timeline.timeline_id;
+
+        task_mgr::spawn(
+            task_mgr::BACKGROUND_RUNTIME.handle(),
+            TaskKind::TimelineDeletionWorker,
+            Some(tenant_id),
+            Some(timeline_id),
+            "timeline_delete",
+            false,
+            async move {
+                if let Err(err) = Self::background(guard, conf, &tenant, &timeline).await {
+                    error!("Error: {err:#}");
+                    timeline.set_broken(format!("{err:#}"))
+                };
+                Ok(())
+            }
+            .instrument({
+                let span =
+                    tracing::info_span!(parent: None, "delete_timeline", tenant_id=%tenant_id, timeline_id=%timeline_id);
+                span.follows_from(Span::current());
+                span
+            }),
+        );
+    }
+
+    async fn background(
+        mut guard: DeletionGuard,
+        conf: &PageServerConf,
+        tenant: &Tenant,
+        timeline: &Timeline,
+    ) -> Result<(), DeleteTimelineError> {
+        delete_local_layer_files(conf, tenant.tenant_id, timeline).await?;
+
+        delete_remote_layers_and_index(timeline).await?;
+
+        pausable_failpoint!("in_progress_delete");
+
+        cleanup_remaining_timeline_fs_traces(conf, tenant.tenant_id, timeline.timeline_id).await?;
+
+        remove_timeline_from_tenant(tenant, timeline.timeline_id, &guard).await?;
+
+        *guard.0 = Self::Finished;
+
+        Ok(())
+    }
+}
+
+struct DeletionGuard(OwnedMutexGuard<DeleteTimelineFlow>);
+
+impl Deref for DeletionGuard {
+    type Target = DeleteTimelineFlow;
+
+    fn deref(&self) -> &Self::Target {
+        &self.0
+    }
+}
+
+impl DerefMut for DeletionGuard {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.0
+    }
+}

pageserver/src/tenant/timeline/eviction_task.rs

@@ -30,6 +30,7 @@ use crate::{
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
         storage_layer::PersistentLayer,
+        timeline::EvictionError,
         LogicalSizeCalculationCause, Tenant,
     },
 };
@@ -100,11 +101,11 @@ impl Timeline {
             match cf {
                 ControlFlow::Break(()) => break,
                 ControlFlow::Continue(sleep_until) => {
-                    tokio::select! {
-                        _ = cancel.cancelled() => {
-                            break;
-                        }
-                        _ = tokio::time::sleep_until(sleep_until) => { }
+                    if tokio::time::timeout_at(sleep_until, cancel.cancelled())
+                        .await
+                        .is_ok()
+                    {
+                        break;
                     }
                 }
             }
@@ -270,20 +271,22 @@ impl Timeline {
                 None => {
                     stats.skipped_for_shutdown += 1;
                 }
-                Some(Ok(true)) => {
-                    debug!("evicted layer {l:?}");
+                Some(Ok(())) => {
                     stats.evicted += 1;
                 }
-                Some(Ok(false)) => {
-                    debug!("layer is not evictable: {l:?}");
+                Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
                     stats.not_evictable += 1;
                 }
-                Some(Err(e)) => {
-                    // This variant is the case where an unexpected error happened during eviction.
-                    // Expected errors that result in non-eviction are `Some(Ok(false))`.
-                    // So, dump Debug here to gather as much info as possible in this rare case.
-                    warn!("failed to evict layer {l:?}: {e:?}");
-                    stats.errors += 1;
+                Some(Err(EvictionError::FileNotFound)) => {
+                    // compaction/gc removed the file while we were waiting on layer_removal_cs
+                    stats.not_evictable += 1;
+                }
+                Some(Err(
+                    e @ EvictionError::LayerNotFound(_) | e @ EvictionError::StatFailed(_),
+                )) => {
+                    let e = utils::error::report_compact_sources(&e);
+                    warn!(layer = %l, "failed to evict layer: {e}");
+                    stats.not_evictable += 1;
                 }
             }
         }
@@ -305,8 +308,13 @@ impl Timeline {
         ctx: &RequestContext,
     ) -> ControlFlow<()> {
         let mut state = self.eviction_task_timeline_state.lock().await;
+
+        // Only do the imitate_layer accesses approximately as often as the threshold.  A little
+        // more frequently, to avoid this period racing with the threshold/period-th eviction iteration.
+        let inter_imitate_period = p.threshold.checked_sub(p.period).unwrap_or(p.threshold);
+
         match state.last_layer_access_imitation {
-            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
+            Some(ts) if ts.elapsed() < inter_imitate_period => { /* no need to run */ }
             _ => {
                 self.imitate_timeline_cached_layer_accesses(cancel, ctx)
                     .await;
@@ -329,7 +337,7 @@ impl Timeline {
         };
         let mut state = tenant.eviction_task_tenant_state.lock().await;
         match state.last_layer_access_imitation {
-            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
+            Some(ts) if ts.elapsed() < inter_imitate_period => { /* no need to run */ }
             _ => {
                 self.imitate_synthetic_size_calculation_worker(&tenant, ctx, cancel)
                     .await;