Store read_only flag in pageserver timelines

## Problem
Names are not consistent between safekeeper migration RFC and the actual
implementation.

It's not used anywhere in production yet, so it's safe to rename. We
don't need to worry about backward compatibility.

- Follow up on https://github.com/neondatabase/neon/pull/12432

## Summary of changes
- rename term -> last_log_term in TimelineMembershipSwitchResponse 
- add missing fields to TimelineMembershipSwitchResponse in python

.dockerignore

@@ -4,6 +4,7 @@
 !Cargo.lock
 !Cargo.toml
 !Makefile
+!postgres.mk
 !rust-toolchain.toml
 !scripts/ninstall.sh
 !docker-compose/run-tests.sh

.github/actionlint.yml

@@ -7,6 +7,7 @@ self-hosted-runner:
     - small-metal
     - small-arm64
     - unit-perf
+    - unit-perf-aws-arm
     - us-east-2
 config-variables:
   - AWS_ECR_REGION

.github/workflows/build-macos.yml

@@ -32,161 +32,14 @@ permissions:
   contents: read
 
 jobs:
-  build-pgxn:
-    if: |
-      inputs.pg_versions != '[]' || inputs.rebuild_everything ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos')  ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
-      github.ref_name == 'main'
-    timeout-minutes: 30
-    runs-on: macos-15
-    strategy:
-      matrix:
-        postgres-version: ${{ inputs.rebuild_everything && fromJSON('["v14", "v15", "v16", "v17"]') || fromJSON(inputs.pg_versions) }}
-    env:
-      # Use release build only, to have less debug info around
-      # Hence keeping target/ (and general cache size) smaller
-      BUILD_TYPE: release
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - name: Checkout main repo
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set pg ${{ matrix.postgres-version }} for caching
-        id: pg_rev
-        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-${{ matrix.postgres-version }}) | tee -a "${GITHUB_OUTPUT}"
-
-      - name: Cache postgres ${{ matrix.postgres-version }} build
-        id: cache_pg
-        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
-        with:
-          path: pg_install/${{ matrix.postgres-version }}
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-${{ matrix.postgres-version }}-${{ steps.pg_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
-
-      - name: Checkout submodule vendor/postgres-${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          git submodule init vendor/postgres-${{ matrix.postgres-version }}
-          git submodule update --depth 1 --recursive
-
-      - name: Install build dependencies
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Build Postgres ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make postgres-${{ matrix.postgres-version }} -j$(sysctl -n hw.ncpu)
-
-      - name: Build Neon Pg Ext ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make "neon-pg-ext-${{ matrix.postgres-version }}" -j$(sysctl -n hw.ncpu)
-
-      - name: Get postgres headers ${{ matrix.postgres-version }}
-        if: steps.cache_pg.outputs.cache-hit != 'true'
-        run: |
-          make postgres-headers-${{ matrix.postgres-version }} -j$(sysctl -n hw.ncpu)
-
-      - name: Upload "pg_install/${{ matrix.postgres-version }}" artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: pg_install--${{ matrix.postgres-version }}
-          path: pg_install/${{ matrix.postgres-version }}
-          # The artifact is supposed to be used by the next job in the same workflow,
-          # so there’s no need to store it for too long.
-          retention-days: 1
-
-  build-walproposer-lib:
-    if: |
-      contains(inputs.pg_versions, 'v17') || inputs.rebuild_everything ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos')  ||
-      contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
-      github.ref_name == 'main'
-    timeout-minutes: 30
-    runs-on: macos-15
-    needs: [build-pgxn]
-    env:
-      # Use release build only, to have less debug info around
-      # Hence keeping target/ (and general cache size) smaller
-      BUILD_TYPE: release
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - name: Checkout main repo
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set pg v17 for caching
-        id: pg_rev
-        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v17) | tee -a "${GITHUB_OUTPUT}"
-
-      - name: Download "pg_install/v17" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v17
-          path: pg_install/v17
-
-      - name: Cache walproposer-lib
-        id: cache_walproposer_lib
-        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
-        with:
-          path: build/walproposer-lib
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-walproposer_lib-v17-${{ steps.pg_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
-
-      - name: Checkout submodule vendor/postgres-v17
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          git submodule init vendor/postgres-v17
-          git submodule update --depth 1 --recursive
-
-      - name: Install build dependencies
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Build walproposer-lib (only for v17)
-        if: steps.cache_walproposer_lib.outputs.cache-hit != 'true'
-        run:
-          make walproposer-lib -j$(sysctl -n hw.ncpu)
-
-      - name: Upload "build/walproposer-lib" artifact
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
-        with:
-          name: build--walproposer-lib
-          path: build/walproposer-lib
-          # The artifact is supposed to be used by the next job in the same workflow,
-          # so there’s no need to store it for too long.
-          retention-days: 1
-
-  cargo-build:
+  make-all:
     if: |
       inputs.pg_versions != '[]' || inputs.rebuild_rust_code || inputs.rebuild_everything ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-macos') ||
       contains(github.event.pull_request.labels.*.name, 'run-extra-build-*') ||
       github.ref_name == 'main'
-    timeout-minutes: 30
+    timeout-minutes: 60
     runs-on: macos-15
-    needs: [build-pgxn, build-walproposer-lib]
     env:
       # Use release build only, to have less debug info around
       # Hence keeping target/ (and general cache size) smaller
@@ -202,41 +55,53 @@ jobs:
         with:
           submodules: true
 
-      - name: Download "pg_install/v14" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v14
-          path: pg_install/v14
-
-      - name: Download "pg_install/v15" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v15
-          path: pg_install/v15
-
-      - name: Download "pg_install/v16" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v16
-          path: pg_install/v16
-
-      - name: Download "pg_install/v17" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: pg_install--v17
-          path: pg_install/v17
-
-      - name: Download "build/walproposer-lib" artifact
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
-        with:
-          name: build--walproposer-lib
-          path: build/walproposer-lib
-
-      # `actions/download-artifact` doesn't preserve permissions:
-      # https://github.com/actions/download-artifact?tab=readme-ov-file#permission-loss
-      - name: Make pg_install/v*/bin/* executable
+      - name: Install build dependencies
         run: |
-          chmod +x pg_install/v*/bin/*
+          brew install flex bison openssl protobuf icu4c
+
+      - name: Set extra env for macOS
+        run: |
+          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
+          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
+
+      - name: Restore "pg_install/" cache
+        id: cache_pg
+        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
+        with:
+          path: pg_install
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ env.BUILD_TYPE }}-pg-install-v14-${{ hashFiles('Makefile', 'postgres.mk', 'vendor/revisions.json') }}
+
+      - name: Checkout vendor/postgres submodules
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          git submodule init
+          git submodule update --depth 1 --recursive
+
+      - name: Build Postgres
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          make postgres -j$(sysctl -n hw.ncpu)
+
+      # This isn't strictly necessary, but it makes the cached and non-cached builds more similar,
+      # When pg_install is restored from cache, there is no 'build/' directory. By removing it
+      # in a non-cached build too, we enforce that the rest of the steps don't depend on it,
+      # so that we notice any build caching bugs earlier.
+      - name: Remove build artifacts
+        if: steps.cache_pg.outputs.cache-hit != 'true'
+        run: |
+          rm -rf build
+
+      # Explicitly update the rust toolchain before running 'make'. The parallel make build can
+      # invoke 'cargo build' more than once in parallel, for different crates.  That's OK, 'cargo'
+      # does its own locking to prevent concurrent builds from stepping on each other's
+      # toes. However, it will first try to update the toolchain, and that step is not locked the
+      # same way. To avoid two toolchain updates running in parallel and stepping on each other's
+      # toes, ensure that the toolchain is up-to-date beforehand.
+      - name: Update rust toolchain
+        run: |
+          rustup --version &&
+          rustup update &&
+          rustup show
 
       - name: Cache cargo deps
         uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
@@ -248,17 +113,12 @@ jobs:
             target
           key: v1-${{ runner.os }}-${{ runner.arch }}-cargo-${{ hashFiles('./Cargo.lock') }}-${{ hashFiles('./rust-toolchain.toml') }}-rust
 
-      - name: Install build dependencies
-        run: |
-          brew install flex bison openssl protobuf icu4c
-
-      - name: Set extra env for macOS
-        run: |
-          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
-          echo 'CPPFLAGS=-I/usr/local/opt/openssl@3/include' >> $GITHUB_ENV
-
-      - name: Run cargo build
-        run: cargo build --all --release -j$(sysctl -n hw.ncpu)
+      # Build the neon-specific postgres extensions, and all the Rust bits.
+      #
+      # Pass PG_INSTALL_CACHED=1 because PostgreSQL was already built and cached
+      # separately.
+      - name: Build all
+        run: PG_INSTALL_CACHED=1 BUILD_TYPE=release make -j$(sysctl -n hw.ncpu) all
 
       - name: Check that no warnings are produced
         run: ./run_clippy.sh

.github/workflows/build_and_test.yml

@@ -69,7 +69,7 @@ jobs:
           submodules: true
 
       - name: Check for file changes
-        uses: step-security/paths-filter@v3
+        uses: dorny/paths-filter@de90cc6fb38fc0963ad72b210f1f284cd68cea36  # v3.0.2
         id: files-changed
         with:
           token: ${{ secrets.GITHUB_TOKEN }}
@@ -306,14 +306,14 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [ self-hosted, unit-perf ]
+    runs-on: [ self-hosted, unit-perf-aws-arm ]
     container:
       image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
       credentials:
         username: ${{ github.actor }}
         password: ${{ secrets.GITHUB_TOKEN }}
       # for changed limits, see comments on `options:` earlier in this file
-      options: --init --shm-size=512mb --ulimit memlock=67108864:67108864
+      options: --init --shm-size=512mb --ulimit memlock=67108864:67108864 --ulimit nofile=65536:65536 --security-opt seccomp=unconfined
     strategy:
       fail-fast: false
       matrix:

.github/workflows/large_oltp_benchmark.yml

@@ -153,7 +153,7 @@ jobs:
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
 
     - name: Benchmark database maintenance
-      if: ${{ matrix.test_maintenance == 'true' }}
+      if: ${{ matrix.test_maintenance }}
       uses: ./.github/actions/run-python-test-set
       with:
         build_type: ${{ env.BUILD_TYPE }}

.github/workflows/neon_extra_builds.yml

@@ -53,7 +53,7 @@ jobs:
           submodules: true
 
       - name: Check for Postgres changes
-        uses: step-security/paths-filter@v3
+        uses: dorny/paths-filter@1441771bbfdd59dcd748680ee64ebd8faab1a242  #v3
         id: files_changed
         with:
           token: ${{ github.token }}

.github/workflows/periodic_pagebench.yml

@@ -1,4 +1,4 @@
-name: Periodic pagebench performance test on unit-perf hetzner runner
+name: Periodic pagebench performance test on unit-perf-aws-arm runners
 
 on:
   schedule:
@@ -40,7 +40,7 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [ self-hosted, unit-perf ]
+    runs-on: [ self-hosted, unit-perf-aws-arm ]
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm
       credentials:

.github/workflows/pre-merge-checks.yml

@@ -34,7 +34,7 @@ jobs:
 
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
-      - uses: step-security/changed-files@3dbe17c78367e7d60f00d78ae6781a35be47b4a1 # v45.0.1
+      - uses: tj-actions/changed-files@ed68ef82c095e0d48ec87eccea555d944a631a4c # v46.0.5
         id: python-src
         with:
           files: |
@@ -45,7 +45,7 @@ jobs:
             poetry.lock
             pyproject.toml
 
-      - uses: step-security/changed-files@3dbe17c78367e7d60f00d78ae6781a35be47b4a1 # v45.0.1
+      - uses: tj-actions/changed-files@ed68ef82c095e0d48ec87eccea555d944a631a4c # v46.0.5
         id: rust-src
         with:
           files: |

.github/workflows/proxy-benchmark.yml

@@ -1,4 +1,4 @@
-name: Periodic proxy performance test on unit-perf hetzner runner
+name: Periodic proxy performance test on unit-perf-aws-arm runners
 
 on:
   push: # TODO: remove after testing
@@ -32,7 +32,7 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [self-hosted, unit-perf]
+    runs-on: [self-hosted, unit-perf-aws-arm]
     timeout-minutes: 60  # 1h timeout
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm
@@ -60,22 +60,23 @@ jobs:
         } >> "$GITHUB_ENV"
 
     - name: Run proxy-bench
-      run: ./${PROXY_BENCH_PATH}/run.sh
+      run: ${PROXY_BENCH_PATH}/run.sh
 
     - name: Ingest Bench Results # neon repo script
-      if: success()
+      if: always()
       run: |
         mkdir -p $TEST_OUTPUT
         python $NEON_DIR/scripts/proxy_bench_results_ingest.py --out $TEST_OUTPUT
 
     - name: Push Metrics to Proxy perf database
-      if: success()
+      if: always()
       env:
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PROXY_TEST_RESULT_CONNSTR }}"
         REPORT_FROM: $TEST_OUTPUT
       run: $NEON_DIR/scripts/generate_and_push_perf_report.sh
 
     - name: Docker cleanup
+      if: always()
       run: docker compose down
 
     - name: Notify Failure

.gitignore

@@ -6,6 +6,7 @@
 /tmp_check_cli
 __pycache__/
 test_output/
+neon_previous/
 .vscode
 .idea
 *.swp

Cargo.lock

@@ -1083,6 +1083,25 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
+[[package]]
+name = "cbindgen"
+version = "0.29.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "975982cdb7ad6a142be15bdf84aea7ec6a9e5d4d797c004d43185b24cfe4e684"
+dependencies = [
+ "clap",
+ "heck",
+ "indexmap 2.9.0",
+ "log",
+ "proc-macro2",
+ "quote",
+ "serde",
+ "serde_json",
+ "syn 2.0.100",
+ "tempfile",
+ "toml",
+]
+
 [[package]]
 name = "cc"
 version = "1.2.16"
@@ -1267,6 +1286,15 @@ dependencies = [
  "unicode-width",
 ]
 
+[[package]]
+name = "communicator"
+version = "0.1.0"
+dependencies = [
+ "cbindgen",
+ "neon-shmem",
+ "workspace_hack",
+]
+
 [[package]]
 name = "compute_api"
 version = "0.1.0"
@@ -1279,6 +1307,7 @@ dependencies = [
  "remote_storage",
  "serde",
  "serde_json",
+ "url",
  "utils",
 ]
 
@@ -1304,6 +1333,7 @@ dependencies = [
  "fail",
  "flate2",
  "futures",
+ "hostname-validator",
  "http 1.1.0",
  "indexmap 2.9.0",
  "itertools 0.10.5",
@@ -1316,6 +1346,7 @@ dependencies = [
  "opentelemetry",
  "opentelemetry_sdk",
  "p256 0.13.2",
+ "pageserver_page_api",
  "postgres",
  "postgres_initdb",
  "postgres_versioninfo",
@@ -1335,6 +1366,7 @@ dependencies = [
  "tokio-postgres",
  "tokio-stream",
  "tokio-util",
+ "tonic 0.13.1",
  "tower 0.5.2",
  "tower-http",
  "tower-otel",
@@ -2768,6 +2800,12 @@ dependencies = [
  "windows",
 ]
 
+[[package]]
+name = "hostname-validator"
+version = "1.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f558a64ac9af88b5ba400d99b579451af0d39c6d360980045b91aac966d705e2"
+
 [[package]]
 name = "http"
 version = "0.2.9"
@@ -4408,6 +4446,7 @@ dependencies = [
  "postgres_backend",
  "postgres_ffi_types",
  "postgres_versioninfo",
+ "posthog_client_lite",
  "rand 0.8.5",
  "remote_storage",
  "reqwest",
@@ -4418,6 +4457,7 @@ dependencies = [
  "strum",
  "strum_macros",
  "thiserror 1.0.69",
+ "tracing",
  "tracing-utils",
  "utils",
 ]
@@ -4474,12 +4514,14 @@ dependencies = [
  "bytes",
  "futures",
  "pageserver_api",
- "postgres_ffi",
+ "postgres_ffi_types",
  "prost 0.13.5",
+ "prost-types 0.13.5",
  "strum",
  "strum_macros",
  "thiserror 1.0.69",
  "tokio",
+ "tokio-util",
  "tonic 0.13.1",
  "tonic-build",
  "utils",
@@ -5152,7 +5194,7 @@ dependencies = [
  "petgraph",
  "prettyplease",
  "prost 0.13.5",
- "prost-types 0.13.3",
+ "prost-types 0.13.5",
  "regex",
  "syn 2.0.100",
  "tempfile",
@@ -5195,9 +5237,9 @@ dependencies = [
 
 [[package]]
 name = "prost-types"
-version = "0.13.3"
+version = "0.13.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4759aa0d3a6232fb8dbdb97b61de2c20047c68aca932c7ed76da9d788508d670"
+checksum = "52c2c1bf36ddb1a1c396b3601a3cec27c2462e45f07c386894ec3ccf5332bd16"
 dependencies = [
  "prost 0.13.5",
 ]
@@ -6804,6 +6846,7 @@ dependencies = [
  "chrono",
  "clap",
  "clashmap",
+ "compute_api",
  "control_plane",
  "cron",
  "diesel",
@@ -6815,6 +6858,7 @@ dependencies = [
  "hex",
  "http-utils",
  "humantime",
+ "humantime-serde",
  "hyper 0.14.30",
  "itertools 0.10.5",
  "json-structural-diff",
@@ -6825,6 +6869,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "postgres_connection",
+ "posthog_client_lite",
  "rand 0.8.5",
  "regex",
  "reqwest",
@@ -7635,7 +7680,7 @@ dependencies = [
  "prettyplease",
  "proc-macro2",
  "prost-build 0.13.3",
- "prost-types 0.13.3",
+ "prost-types 0.13.5",
  "quote",
  "syn 2.0.100",
 ]
@@ -7647,7 +7692,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f9687bd5bfeafebdded2356950f278bba8226f0b32109537c4253406e09aafe1"
 dependencies = [
  "prost 0.13.5",
- "prost-types 0.13.3",
+ "prost-types 0.13.5",
  "tokio",
  "tokio-stream",
  "tonic 0.13.1",

Cargo.toml

@@ -46,6 +46,7 @@ members = [
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
     "endpoint_storage",
+    "pgxn/neon/communicator",
 ]
 
 [workspace.package]
@@ -152,6 +153,7 @@ pprof = { version = "0.14", features = ["criterion", "flamegraph", "frame-pointe
 procfs = "0.16"
 prometheus = {version = "0.13", default-features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.13.5"
+prost-types = "0.13.5"
 rand = "0.8"
 redis = { version = "0.29.2", features = ["tokio-rustls-comp", "keep-alive"] }
 regex = "1.10.2"
@@ -199,7 +201,7 @@ tokio-postgres-rustls = "0.12.0"
 tokio-rustls = { version = "0.26.0", default-features = false, features = ["tls12", "ring"]}
 tokio-stream = "0.1"
 tokio-tar = "0.3"
-tokio-util = { version = "0.7.10", features = ["io", "rt"] }
+tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
 toml_edit = "0.22"
 tonic = { version = "0.13.1", default-features = false, features = ["channel", "codegen", "gzip", "prost", "router", "server", "tls-ring", "tls-native-roots", "zstd"] }
@@ -254,6 +256,7 @@ desim = { version = "0.1", path = "./libs/desim" }
 endpoint_storage = { version = "0.0.1", path = "./endpoint_storage/" }
 http-utils = { version = "0.1", path = "./libs/http-utils/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
+neon-shmem = { version = "0.1", path = "./libs/neon-shmem/" }
 pageserver = { path = "./pageserver" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
 pageserver_client = { path = "./pageserver/client" }
@@ -283,6 +286,7 @@ walproposer = { version = "0.1", path = "./libs/walproposer/" }
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
 
 ## Build dependencies
+cbindgen = "0.29.0"
 criterion = "0.5.1"
 rcgen = "0.13"
 rstest = "0.18"

Dockerfile

@@ -30,7 +30,18 @@ ARG BASE_IMAGE_SHA=debian:${DEBIAN_FLAVOR}
 ARG BASE_IMAGE_SHA=${BASE_IMAGE_SHA/debian:bookworm-slim/debian@$BOOKWORM_SLIM_SHA}
 ARG BASE_IMAGE_SHA=${BASE_IMAGE_SHA/debian:bullseye-slim/debian@$BULLSEYE_SLIM_SHA}
 
-# Build Postgres
+# Naive way:
+#
+# 1. COPY . .
+# 1. make neon-pg-ext
+# 2. cargo build <storage binaries>
+#
+# But to enable docker to cache intermediate layers, we perform a few preparatory steps:
+#
+# - Build all postgres versions, depending on just the contents of vendor/
+# - Use cargo chef to build all rust dependencies
+
+# 1. Build all postgres versions
 FROM $REPOSITORY/$IMAGE:$TAG AS pg-build
 WORKDIR /home/nonroot
 
@@ -38,16 +49,15 @@ COPY --chown=nonroot vendor/postgres-v14 vendor/postgres-v14
 COPY --chown=nonroot vendor/postgres-v15 vendor/postgres-v15
 COPY --chown=nonroot vendor/postgres-v16 vendor/postgres-v16
 COPY --chown=nonroot vendor/postgres-v17 vendor/postgres-v17
-COPY --chown=nonroot pgxn pgxn
 COPY --chown=nonroot Makefile Makefile
+COPY --chown=nonroot postgres.mk postgres.mk
 COPY --chown=nonroot scripts/ninstall.sh scripts/ninstall.sh
 
 ENV BUILD_TYPE=release
 RUN set -e \
-    && mold -run make -j $(nproc) -s neon-pg-ext \
-    && tar -C pg_install -czf /home/nonroot/postgres_install.tar.gz .
+    && mold -run make -j $(nproc) -s postgres
 
-# Prepare cargo-chef recipe
+# 2. Prepare cargo-chef recipe
 FROM $REPOSITORY/$IMAGE:$TAG AS plan
 WORKDIR /home/nonroot
 
@@ -55,23 +65,22 @@ COPY --chown=nonroot . .
 
 RUN cargo chef prepare --recipe-path recipe.json
 
-# Build neon binaries
+# Main build image
 FROM $REPOSITORY/$IMAGE:$TAG AS build
 WORKDIR /home/nonroot
 ARG GIT_VERSION=local
 ARG BUILD_TAG
-
-COPY --from=pg-build /home/nonroot/pg_install/v14/include/postgresql/server pg_install/v14/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v15/include/postgresql/server pg_install/v15/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v16/include/postgresql/server pg_install/v16/include/postgresql/server
-COPY --from=pg-build /home/nonroot/pg_install/v17/include/postgresql/server pg_install/v17/include/postgresql/server
-COPY --from=plan     /home/nonroot/recipe.json                              recipe.json
-
 ARG ADDITIONAL_RUSTFLAGS=""
 
+# 3. Build cargo dependencies. Note that this step doesn't depend on anything else than
+# `recipe.json`, so the layer can be reused as long as none of the dependencies change.
+COPY --from=plan     /home/nonroot/recipe.json                              recipe.json
 RUN set -e \
     && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo chef cook --locked --release --recipe-path recipe.json
 
+# Perform the main build. We reuse the Postgres build artifacts from the intermediate 'pg-build'
+# layer, and the cargo dependencies built in the previous step.
+COPY --chown=nonroot --from=pg-build /home/nonroot/pg_install/ pg_install
 COPY --chown=nonroot . .
 
 RUN set -e \
@@ -86,10 +95,10 @@ RUN set -e \
       --bin endpoint_storage \
       --bin neon_local \
       --bin storage_scrubber \
-      --locked --release
+      --locked --release \
+    && mold -run make -j $(nproc) -s neon-pg-ext
 
-# Build final image
-#
+# Assemble the final image
 FROM $BASE_IMAGE_SHA
 WORKDIR /data
 
@@ -129,12 +138,15 @@ COPY --from=build --chown=neon:neon /home/nonroot/target/release/proxy
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/endpoint_storage    /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/neon_local          /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_scrubber    /usr/local/bin
+COPY --from=build /home/nonroot/pg_install/v14 /usr/local/v14/
+COPY --from=build /home/nonroot/pg_install/v15 /usr/local/v15/
+COPY --from=build /home/nonroot/pg_install/v16 /usr/local/v16/
+COPY --from=build /home/nonroot/pg_install/v17 /usr/local/v17/
 
-COPY --from=pg-build /home/nonroot/pg_install/v14 /usr/local/v14/
-COPY --from=pg-build /home/nonroot/pg_install/v15 /usr/local/v15/
-COPY --from=pg-build /home/nonroot/pg_install/v16 /usr/local/v16/
-COPY --from=pg-build /home/nonroot/pg_install/v17 /usr/local/v17/
-COPY --from=pg-build /home/nonroot/postgres_install.tar.gz /data/
+# Deprecated: Old deployment scripts use this tarball which contains all the Postgres binaries.
+# That's obsolete, since all the same files are also present under /usr/local/v*. But to keep the
+# old scripts working for now, create the tarball.
+RUN tar -C /usr/local -cvzf /data/postgres_install.tar.gz v14 v15 v16 v17
 
 # By default, pageserver uses `.neon/` working directory in WORKDIR, so create one and fill it with the dummy config.
 # Now, when `docker run ... pageserver` is run, it can start without errors, yet will have some default dummy values.

Makefile

@@ -4,11 +4,14 @@ ROOT_PROJECT_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 # managers.
 POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install/
 
+# Supported PostgreSQL versions
+POSTGRES_VERSIONS = v17 v16 v15 v14
+
 # CARGO_BUILD_FLAGS: Extra flags to pass to `cargo build`. `--locked`
 # and `--features testing` are popular examples.
 #
-# CARGO_PROFILE: You can also set to override the cargo profile to
-# use. By default, it is derived from BUILD_TYPE.
+# CARGO_PROFILE: Set to override the cargo profile to use. By default,
+# it is derived from BUILD_TYPE.
 
 # All intermediate build artifacts are stored here.
 BUILD_DIR := build
@@ -27,11 +30,18 @@ ifeq ($(BUILD_TYPE),release)
 	PG_CFLAGS += -O2 -g3 $(CFLAGS)
 	PG_LDFLAGS = $(LDFLAGS)
 	CARGO_PROFILE ?= --profile=release
+	# NEON_CARGO_ARTIFACT_TARGET_DIR is the directory where `cargo build` places
+	# the final build artifacts. There is unfortunately no easy way of changing
+	# it to a fully predictable path, nor to extract the path with a simple
+	# command. See https://github.com/rust-lang/cargo/issues/9661 and
+	# https://github.com/rust-lang/cargo/issues/6790.
+	NEON_CARGO_ARTIFACT_TARGET_DIR = $(ROOT_PROJECT_DIR)/target/release
 else ifeq ($(BUILD_TYPE),debug)
 	PG_CONFIGURE_OPTS = --enable-debug --with-openssl --enable-cassert --enable-depend
 	PG_CFLAGS += -O0 -g3 $(CFLAGS)
 	PG_LDFLAGS = $(LDFLAGS)
 	CARGO_PROFILE ?= --profile=dev
+	NEON_CARGO_ARTIFACT_TARGET_DIR = $(ROOT_PROJECT_DIR)/target/debug
 else
 	$(error Bad build type '$(BUILD_TYPE)', see Makefile for options)
 endif
@@ -95,94 +105,30 @@ CACHEDIR_TAG_CONTENTS := "Signature: 8a477f597d28d172789f06886806bc55"
 # Top level Makefile to build Neon and PostgreSQL
 #
 .PHONY: all
-all: neon postgres neon-pg-ext
+all: neon postgres-install neon-pg-ext
 
 ### Neon Rust bits
 #
-# The 'postgres_ffi' depends on the Postgres headers.
+# The 'postgres_ffi' crate depends on the Postgres headers.
 .PHONY: neon
-neon: postgres-headers walproposer-lib cargo-target-dir
+neon: postgres-headers-install walproposer-lib cargo-target-dir
 	+@echo "Compiling Neon"
 	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS) $(CARGO_PROFILE)
+
 .PHONY: cargo-target-dir
 cargo-target-dir:
 	# https://github.com/rust-lang/cargo/issues/14281
 	mkdir -p target
 	test -e target/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > target/CACHEDIR.TAG
 
-### PostgreSQL parts
-# Some rules are duplicated for Postgres v14 and 15. We may want to refactor
-# to avoid the duplication in the future, but it's tolerable for now.
-#
-$(BUILD_DIR)/%/config.status:
-	mkdir -p $(BUILD_DIR)
-	test -e $(BUILD_DIR)/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > $(BUILD_DIR)/CACHEDIR.TAG
-
-	+@echo "Configuring Postgres $* build"
-	@test -s $(ROOT_PROJECT_DIR)/vendor/postgres-$*/configure || { \
-		echo "\nPostgres submodule not found in $(ROOT_PROJECT_DIR)/vendor/postgres-$*/, execute "; \
-		echo "'git submodule update --init --recursive --depth 2 --progress .' in project root.\n"; \
-		exit 1; }
-	mkdir -p $(BUILD_DIR)/$*
-
-	VERSION=$*; \
-	EXTRA_VERSION=$$(cd $(ROOT_PROJECT_DIR)/vendor/postgres-$$VERSION && git rev-parse HEAD); \
-	(cd $(BUILD_DIR)/$$VERSION && \
-	env PATH="$(EXTRA_PATH_OVERRIDES):$$PATH" $(ROOT_PROJECT_DIR)/vendor/postgres-$$VERSION/configure \
-		CFLAGS='$(PG_CFLAGS)' LDFLAGS='$(PG_LDFLAGS)' \
-		$(PG_CONFIGURE_OPTS) --with-extra-version=" ($$EXTRA_VERSION)" \
-		--prefix=$(abspath $(POSTGRES_INSTALL_DIR))/$$VERSION > configure.log)
-
-# nicer alias to run 'configure'
-# Note: I've been unable to use templates for this part of our configuration.
-# I'm not sure why it wouldn't work, but this is the only place (apart from
-# the "build-all-versions" entry points) where direct mention of PostgreSQL
-# versions is used.
-.PHONY: postgres-configure-v17
-postgres-configure-v17: $(BUILD_DIR)/v17/config.status
-.PHONY: postgres-configure-v16
-postgres-configure-v16: $(BUILD_DIR)/v16/config.status
-.PHONY: postgres-configure-v15
-postgres-configure-v15: $(BUILD_DIR)/v15/config.status
-.PHONY: postgres-configure-v14
-postgres-configure-v14: $(BUILD_DIR)/v14/config.status
-
-# Install the PostgreSQL header files into $(POSTGRES_INSTALL_DIR)/<version>/include
-.PHONY: postgres-headers-%
-postgres-headers-%: postgres-configure-%
-	+@echo "Installing PostgreSQL $* headers"
-	$(MAKE) -C $(BUILD_DIR)/$*/src/include MAKELEVEL=0 install
-
-# Compile and install PostgreSQL
-.PHONY: postgres-%
-postgres-%: postgres-configure-% \
-		  postgres-headers-% # to prevent `make install` conflicts with neon's `postgres-headers`
-	+@echo "Compiling PostgreSQL $*"
-	$(MAKE) -C $(BUILD_DIR)/$* MAKELEVEL=0 install
-	+@echo "Compiling pg_prewarm $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_prewarm install
-	+@echo "Compiling pg_buffercache $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_buffercache install
-	+@echo "Compiling pg_visibility $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_visibility install
-	+@echo "Compiling pageinspect $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pageinspect install
-	+@echo "Compiling pg_trgm $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/pg_trgm install
-	+@echo "Compiling amcheck $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/amcheck install
-	+@echo "Compiling test_decoding $*"
-	$(MAKE) -C $(BUILD_DIR)/$*/contrib/test_decoding install
-
-.PHONY: postgres-check-%
-postgres-check-%: postgres-%
-	$(MAKE) -C $(BUILD_DIR)/$* MAKELEVEL=0 check
-
 .PHONY: neon-pg-ext-%
-neon-pg-ext-%: postgres-%
+neon-pg-ext-%: postgres-install-% cargo-target-dir
 	+@echo "Compiling neon-specific Postgres extensions for $*"
 	mkdir -p $(BUILD_DIR)/pgxn-$*
-	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config COPT='$(COPT)' \
+	$(MAKE) PG_CONFIG="$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config" COPT='$(COPT)' \
+		NEON_CARGO_ARTIFACT_TARGET_DIR="$(NEON_CARGO_ARTIFACT_TARGET_DIR)" \
+		CARGO_BUILD_FLAGS="$(CARGO_BUILD_FLAGS)" \
+		CARGO_PROFILE="$(CARGO_PROFILE)" \
 		-C $(BUILD_DIR)/pgxn-$*\
 		-f $(ROOT_PROJECT_DIR)/pgxn/Makefile  install
 
@@ -218,39 +164,14 @@ ifeq ($(UNAME_S),Linux)
 		pg_crc32c.o
 endif
 
+# Shorthand to call neon-pg-ext-% target for all Postgres versions
 .PHONY: neon-pg-ext
-neon-pg-ext: \
-	neon-pg-ext-v14 \
-	neon-pg-ext-v15 \
-	neon-pg-ext-v16 \
-	neon-pg-ext-v17
-
-# shorthand to build all Postgres versions
-.PHONY: postgres
-postgres: \
-	postgres-v14 \
-	postgres-v15 \
-	postgres-v16 \
-	postgres-v17
-
-.PHONY: postgres-headers
-postgres-headers: \
-	postgres-headers-v14 \
-	postgres-headers-v15 \
-	postgres-headers-v16 \
-	postgres-headers-v17
-
-.PHONY: postgres-check
-postgres-check: \
-	postgres-check-v14 \
-	postgres-check-v15 \
-	postgres-check-v16 \
-	postgres-check-v17
+neon-pg-ext: $(foreach pg_version,$(POSTGRES_VERSIONS),neon-pg-ext-$(pg_version))
 
 # This removes everything
 .PHONY: distclean
 distclean:
-	$(RM) -r $(POSTGRES_INSTALL_DIR)
+	$(RM) -r $(POSTGRES_INSTALL_DIR) $(BUILD_DIR)
 	$(CARGO_CMD_PREFIX) cargo clean
 
 .PHONY: fmt
@@ -298,3 +219,19 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 .PHONY: setup-pre-commit-hook
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit
+
+# Targets for building PostgreSQL are defined in postgres.mk.
+#
+# But if the caller has indicated that PostgreSQL is already
+# installed, by setting the PG_INSTALL_CACHED variable, skip it.
+ifdef PG_INSTALL_CACHED
+postgres-install: skip-install
+$(foreach pg_version,$(POSTGRES_VERSIONS),postgres-install-$(pg_version)): skip-install
+postgres-headers-install:
+	+@echo "Skipping installation of PostgreSQL headers because PG_INSTALL_CACHED is set"
+skip-install:
+	+@echo "Skipping PostgreSQL installation because PG_INSTALL_CACHED is set"
+
+else
+include postgres.mk
+endif

build-tools.Dockerfile

@@ -165,6 +165,7 @@ RUN curl -fsSL \
     && rm sql_exporter.tar.gz
 
 # protobuf-compiler (protoc)
+# Keep the version the same as in compute/compute-node.Dockerfile
 ENV PROTOC_VERSION=25.1
 RUN curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
     && unzip -q protoc.zip -d protoc \
@@ -179,7 +180,7 @@ RUN curl -sL "https://github.com/peak/s5cmd/releases/download/v${S5CMD_VERSION}/
     && mv s5cmd /usr/local/bin/s5cmd
 
 # LLVM
-ENV LLVM_VERSION=19
+ENV LLVM_VERSION=20
 RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && echo "deb http://apt.llvm.org/${DEBIAN_VERSION}/ llvm-toolchain-${DEBIAN_VERSION}-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
     && apt update \
@@ -292,7 +293,7 @@ WORKDIR /home/nonroot
 
 # Rust
 # Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
-ENV RUSTC_VERSION=1.87.0
+ENV RUSTC_VERSION=1.88.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
 ARG RUSTFILT_VERSION=0.2.1

compute/compute-node.Dockerfile

@@ -115,6 +115,9 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION
 
+# Keep in sync with build-tools.Dockerfile
+ENV PROTOC_VERSION=25.1
+
 # Use strict mode for bash to catch errors early
 SHELL ["/bin/bash", "-euo", "pipefail", "-c"]
 
@@ -149,8 +152,14 @@ RUN case $DEBIAN_VERSION in \
     libclang-dev \
     jsonnet \
     $VERSION_INSTALLS \
-    && apt clean && rm -rf /var/lib/apt/lists/* && \
-    useradd -ms /bin/bash nonroot -b /home
+    && apt clean && rm -rf /var/lib/apt/lists/* \
+    && useradd -ms /bin/bash nonroot -b /home \
+    # Install protoc from binary release, since Debian's versions are too old.
+    && curl -fsSL "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-$(uname -m | sed 's/aarch64/aarch_64/g').zip" -o "protoc.zip" \
+    && unzip -q protoc.zip -d protoc \
+    && mv protoc/bin/protoc /usr/local/bin/protoc \
+    && mv protoc/include/google /usr/local/include/google \
+    && rm -rf protoc.zip protoc
 
 #########################################################################################
 #
@@ -1170,7 +1179,7 @@ COPY --from=pgrag-src /ext-src/ /ext-src/
 # Install it using virtual environment, because Python 3.11 (the default version on Debian 12 (Bookworm)) complains otherwise
 WORKDIR /ext-src/onnxruntime-src
 RUN apt update && apt install --no-install-recommends --no-install-suggests -y \
-    python3 python3-pip python3-venv protobuf-compiler && \
+    python3 python3-pip python3-venv && \
     apt clean && rm -rf /var/lib/apt/lists/* && \
     python3 -m venv venv && \
     . venv/bin/activate && \
@@ -1563,6 +1572,7 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) && \
 FROM build-deps AS pgaudit-src
 ARG PG_VERSION
 WORKDIR /ext-src
+COPY "compute/patches/pgaudit-parallel_workers-${PG_VERSION}.patch" .
 RUN case "${PG_VERSION}" in \
     "v14") \
     export PGAUDIT_VERSION=1.6.3 \
@@ -1585,7 +1595,8 @@ RUN case "${PG_VERSION}" in \
     esac && \
     wget https://github.com/pgaudit/pgaudit/archive/refs/tags/${PGAUDIT_VERSION}.tar.gz -O pgaudit.tar.gz && \
     echo "${PGAUDIT_CHECKSUM} pgaudit.tar.gz" | sha256sum --check && \
-    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C .
+    mkdir pgaudit-src && cd pgaudit-src && tar xzf ../pgaudit.tar.gz --strip-components=1 -C . && \
+    patch -p1 < "/ext-src/pgaudit-parallel_workers-${PG_VERSION}.patch"
 
 FROM pg-build AS pgaudit-build
 COPY --from=pgaudit-src /ext-src/ /ext-src/
@@ -1625,11 +1636,14 @@ RUN make install USE_PGXS=1 -j $(getconf _NPROCESSORS_ONLN)
 # compile neon extensions
 #
 #########################################################################################
-FROM pg-build AS neon-ext-build
+FROM pg-build-with-cargo AS neon-ext-build
 ARG PG_VERSION
 
-COPY pgxn/ pgxn/
-RUN make -j $(getconf _NPROCESSORS_ONLN) -C pgxn -s install-compute
+USER root
+COPY . .
+
+RUN make -j $(getconf _NPROCESSORS_ONLN) -C pgxn -s install-compute \
+      BUILD_TYPE=release CARGO_BUILD_FLAGS="--locked --release" NEON_CARGO_ARTIFACT_TARGET_DIR="$(pwd)/target/release"
 
 #########################################################################################
 #
@@ -1974,7 +1988,7 @@ RUN apt update && \
         locales \
         lsof \
         procps \
-        rsyslog \
+        rsyslog-gnutls \
         screen \
         tcpdump \
         $VERSION_INSTALLS && \

compute/etc/neon_collector.jsonnet

@@ -8,6 +8,8 @@
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
     import 'sql_exporter/compute_logical_snapshots_bytes.libsonnet',
     import 'sql_exporter/compute_max_connections.libsonnet',
+    import 'sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet',
+    import 'sql_exporter/compute_pg_oldest_mxid_age.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',
     import 'sql_exporter/compute_subscriptions_count.libsonnet',
     import 'sql_exporter/connection_counts.libsonnet',

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.libsonnet

@@ -0,0 +1,13 @@
+{
+  metric_name: 'compute_pg_oldest_frozen_xid_age',
+  type: 'gauge',
+  help: 'Age of oldest XIDs that have not been frozen by VACUUM. An indicator of how long it has been since VACUUM last ran.',
+  key_labels: [
+    'database_name',
+  ],
+  value_label: 'metric',
+  values: [
+    'frozen_xid_age',
+  ],
+  query: importstr 'sql_exporter/compute_pg_oldest_frozen_xid_age.sql',
+}

compute/etc/sql_exporter/compute_pg_oldest_frozen_xid_age.sql

@@ -0,0 +1,4 @@
+SELECT datname database_name,
+  age(datfrozenxid) frozen_xid_age
+FROM pg_database
+ORDER BY frozen_xid_age DESC LIMIT 10;

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.libsonnet

@@ -0,0 +1,13 @@
+{
+  metric_name: 'compute_pg_oldest_mxid_age',
+  type: 'gauge',
+  help: 'Age of oldest MXIDs that have not been replaced by VACUUM. An indicator of how long it has been since VACUUM last ran.',
+  key_labels: [
+    'database_name',
+  ],
+  value_label: 'metric',
+  values: [
+    'min_mxid_age',
+  ],
+  query: importstr 'sql_exporter/compute_pg_oldest_mxid_age.sql',
+}

compute/etc/sql_exporter/compute_pg_oldest_mxid_age.sql

@@ -0,0 +1,4 @@
+SELECT datname database_name,
+  mxid_age(datminmxid) min_mxid_age
+FROM pg_database
+ORDER BY min_mxid_age DESC LIMIT 10;

compute/patches/anon_v2.patch

@@ -1,8 +1,8 @@
 diff --git a/sql/anon.sql b/sql/anon.sql
-index 0cdc769..f6cc950 100644
+index 0cdc769..b450327 100644
 --- a/sql/anon.sql
 +++ b/sql/anon.sql
-@@ -1141,3 +1141,8 @@ $$
+@@ -1141,3 +1141,15 @@ $$
  -- TODO : https://en.wikipedia.org/wiki/L-diversity
  
  -- TODO : https://en.wikipedia.org/wiki/T-closeness
@@ -11,6 +11,13 @@ index 0cdc769..f6cc950 100644
 +
 +GRANT ALL ON SCHEMA anon to neon_superuser;
 +GRANT ALL ON ALL TABLES IN SCHEMA anon TO neon_superuser;
++
++DO $$
++BEGIN
++    IF current_setting('server_version_num')::int >= 150000 THEN
++        GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO neon_superuser;
++    END IF;
++END $$;
 diff --git a/sql/init.sql b/sql/init.sql
 index 7da6553..9b6164b 100644
 --- a/sql/init.sql

compute/patches/pgaudit-parallel_workers-v14.patch

@@ -0,0 +1,143 @@
+commit 7220bb3a3f23fa27207d77562dcc286f9a123313
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index baa8011..a601375 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2563,6 +2563,37 @@ COMMIT;
+ NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
+ DROP TABLE part_test;
+ NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 5e6fd38..ac9ded2 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index cc1374a..1870a60 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1612,6 +1612,36 @@ COMMIT;
+ 
+ DROP TABLE part_test;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v15.patch

@@ -0,0 +1,143 @@
+commit 29dc2847f6255541992f18faf8a815dfab79631a
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index b22560b..73f0327 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2563,6 +2563,37 @@ COMMIT;
+ NOTICE:  AUDIT: SESSION,12,4,MISC,COMMIT,,,COMMIT;,<not logged>
+ DROP TABLE part_test;
+ NOTICE:  AUDIT: SESSION,13,1,DDL,DROP TABLE,,,DROP TABLE part_test;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,14,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 5e6fd38..ac9ded2 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1303,7 +1304,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1384,7 +1385,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1438,7 +1439,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1458,7 +1459,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index 8052426..7f0667b 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1612,6 +1612,36 @@ COMMIT;
+ 
+ DROP TABLE part_test;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v16.patch

@@ -0,0 +1,143 @@
+commit cc708dde7ef2af2a8120d757102d2e34c0463a0f
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index 8772054..9b66ac6 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2556,6 +2556,37 @@ DROP SERVER fdw_server;
+ NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server;,<not logged>
+ DROP EXTENSION postgres_fdw;
+ NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw;,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test;,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 004d1f9..f061164 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1339,7 +1340,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit even onto the stack */
+         stackItem = stack_push();
+@@ -1420,7 +1421,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1475,7 +1476,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1495,7 +1496,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index 6aae88b..de6d7fd 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1631,6 +1631,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
+ DROP SERVER fdw_server;
+ DROP EXTENSION postgres_fdw;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute/patches/pgaudit-parallel_workers-v17.patch

@@ -0,0 +1,143 @@
+commit 8d02e4c6c5e1e8676251b0717a46054267091cb4
+Author: Tristan Partin <tristan.partin@databricks.com>
+Date:   2025-06-23 02:09:31 +0000
+
+    Disable logging in parallel workers
+    
+    When a query uses parallel workers, pgaudit will log the same query for
+    every parallel worker. This is undesireable since it can result in log
+    amplification for queries that use parallel workers.
+    
+    Signed-off-by: Tristan Partin <tristan.partin@databricks.com>
+
+diff --git a/expected/pgaudit.out b/expected/pgaudit.out
+index d696287..4b1059a 100644
+--- a/expected/pgaudit.out
++++ b/expected/pgaudit.out
+@@ -2568,6 +2568,37 @@ DROP SERVER fdw_server;
+ NOTICE:  AUDIT: SESSION,11,1,DDL,DROP SERVER,,,DROP SERVER fdw_server,<not logged>
+ DROP EXTENSION postgres_fdw;
+ NOTICE:  AUDIT: SESSION,12,1,DDL,DROP EXTENSION,,,DROP EXTENSION postgres_fdw,<not logged>
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++SELECT count(*) FROM parallel_test;
++NOTICE:  AUDIT: SESSION,13,1,READ,SELECT,,,SELECT count(*) FROM parallel_test,<not logged>
++ count 
++-------
++  1000
++(1 row)
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';
+diff --git a/pgaudit.c b/pgaudit.c
+index 1764af1..0e48875 100644
+--- a/pgaudit.c
++++ b/pgaudit.c
+@@ -11,6 +11,7 @@
+ #include "postgres.h"
+ 
+ #include "access/htup_details.h"
++#include "access/parallel.h"
+ #include "access/sysattr.h"
+ #include "access/xact.h"
+ #include "access/relation.h"
+@@ -1406,7 +1407,7 @@ pgaudit_ExecutorStart_hook(QueryDesc *queryDesc, int eflags)
+ {
+     AuditEventStackItem *stackItem = NULL;
+ 
+-    if (!internalStatement)
++    if (!internalStatement && !IsParallelWorker())
+     {
+         /* Push the audit event onto the stack */
+         stackItem = stack_push();
+@@ -1489,7 +1490,7 @@ pgaudit_ExecutorCheckPerms_hook(List *rangeTabls, List *permInfos, bool abort)
+ 
+     /* Log DML if the audit role is valid or session logging is enabled */
+     if ((auditOid != InvalidOid || auditLogBitmap != 0) &&
+-        !IsAbortedTransactionBlockState())
++        !IsAbortedTransactionBlockState() && !IsParallelWorker())
+     {
+         /* If auditLogRows is on, wait for rows processed to be set */
+         if (auditLogRows && auditEventStack != NULL)
+@@ -1544,7 +1545,7 @@ pgaudit_ExecutorRun_hook(QueryDesc *queryDesc, ScanDirection direction, uint64 c
+     else
+         standard_ExecutorRun(queryDesc, direction, count, execute_once);
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+@@ -1564,7 +1565,7 @@ pgaudit_ExecutorEnd_hook(QueryDesc *queryDesc)
+     AuditEventStackItem *stackItem = NULL;
+     AuditEventStackItem *auditEventStackFull = NULL;
+ 
+-    if (auditLogRows && !internalStatement)
++    if (auditLogRows && !internalStatement && !IsParallelWorker())
+     {
+         /* Find an item from the stack by the query memory context */
+         stackItem = stack_find_context(queryDesc->estate->es_query_cxt);
+diff --git a/sql/pgaudit.sql b/sql/pgaudit.sql
+index e161f01..c873098 100644
+--- a/sql/pgaudit.sql
++++ b/sql/pgaudit.sql
+@@ -1637,6 +1637,36 @@ DROP USER MAPPING FOR regress_user1 SERVER fdw_server;
+ DROP SERVER fdw_server;
+ DROP EXTENSION postgres_fdw;
+ 
++--
++-- Test logging in parallel workers
++SET pgaudit.log = 'read';
++SET pgaudit.log_client = on;
++SET pgaudit.log_level = 'notice';
++
++-- Force parallel execution for testing
++SET max_parallel_workers_per_gather = 2;
++SET parallel_tuple_cost = 0;
++SET parallel_setup_cost = 0;
++SET min_parallel_table_scan_size = 0;
++SET min_parallel_index_scan_size = 0;
++
++-- Create table with enough data to trigger parallel execution
++CREATE TABLE parallel_test (id int, data text);
++INSERT INTO parallel_test SELECT generate_series(1, 1000), 'test data';
++
++SELECT count(*) FROM parallel_test;
++
++-- Cleanup parallel test
++DROP TABLE parallel_test;
++RESET max_parallel_workers_per_gather;
++RESET parallel_tuple_cost;
++RESET parallel_setup_cost;
++RESET min_parallel_table_scan_size;
++RESET min_parallel_index_scan_size;
++RESET pgaudit.log;
++RESET pgaudit.log_client;
++RESET pgaudit.log_level;
++
+ -- Cleanup
+ -- Set client_min_messages up to warning to avoid noise
+ SET client_min_messages = 'warning';

compute_tools/Cargo.toml

@@ -27,6 +27,7 @@ fail.workspace = true
 flate2.workspace = true
 futures.workspace = true
 http.workspace = true
+hostname-validator = "1.1"
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
@@ -38,6 +39,7 @@ once_cell.workspace = true
 opentelemetry.workspace = true
 opentelemetry_sdk.workspace = true
 p256 = { version = "0.13", features = ["pem"] }
+pageserver_page_api.workspace = true
 postgres.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["json"] }
@@ -53,6 +55,7 @@ tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tokio-postgres.workspace = true
 tokio-util.workspace = true
 tokio-stream.workspace = true
+tonic.workspace = true
 tower-otel.workspace = true
 tracing.workspace = true
 tracing-opentelemetry.workspace = true

compute_tools/src/bin/compute_ctl.rs

@@ -36,6 +36,8 @@
 use std::ffi::OsString;
 use std::fs::File;
 use std::process::exit;
+use std::sync::Arc;
+use std::sync::atomic::AtomicU64;
 use std::sync::mpsc;
 use std::thread;
 use std::time::Duration;
@@ -190,7 +192,9 @@ fn main() -> Result<()> {
             cgroup: cli.cgroup,
             #[cfg(target_os = "linux")]
             vm_monitor_addr: cli.vm_monitor_addr,
-            installed_extensions_collection_interval: cli.installed_extensions_collection_interval,
+            installed_extensions_collection_interval: Arc::new(AtomicU64::new(
+                cli.installed_extensions_collection_interval,
+            )),
         },
         config,
     )?;

compute_tools/src/compute.rs

@@ -6,7 +6,7 @@ use compute_api::responses::{
     LfcPrewarmState, TlsConfig,
 };
 use compute_api::spec::{
-    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PgIdent,
+    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
 };
 use futures::StreamExt;
 use futures::future::join_all;
@@ -15,27 +15,29 @@ use itertools::Itertools;
 use nix::sys::signal::{Signal, kill};
 use nix::unistd::Pid;
 use once_cell::sync::Lazy;
+use pageserver_page_api::{self as page_api, BaseBackupCompression};
 use postgres;
 use postgres::NoTls;
 use postgres::error::SqlState;
 use remote_storage::{DownloadError, RemotePath};
 use std::collections::{HashMap, HashSet};
-use std::net::SocketAddr;
 use std::os::unix::fs::{PermissionsExt, symlink};
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
-use std::sync::atomic::{AtomicU32, Ordering};
+use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
 use std::sync::{Arc, Condvar, Mutex, RwLock};
 use std::time::{Duration, Instant};
 use std::{env, fs};
-use tokio::spawn;
+use tokio::task::JoinHandle;
+use tokio::{spawn, time};
 use tracing::{Instrument, debug, error, info, instrument, warn};
 use url::Url;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 use utils::measured_stream::MeasuredReader;
 use utils::pid_file;
+use utils::shard::{ShardCount, ShardIndex, ShardNumber};
 
 use crate::configurator::launch_configurator;
 use crate::disk_quota::set_disk_quota;
@@ -69,6 +71,7 @@ pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
         .unwrap_or(BUILD_TAG_DEFAULT)
         .to_string()
 });
+const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
 
 /// Static configuration params that don't change after startup. These mostly
 /// come from the CLI args, or are derived from them.
@@ -102,9 +105,11 @@ pub struct ComputeNodeParams {
     pub remote_ext_base_url: Option<Url>,
 
     /// Interval for installed extensions collection
-    pub installed_extensions_collection_interval: u64,
+    pub installed_extensions_collection_interval: Arc<AtomicU64>,
 }
 
+type TaskHandle = Mutex<Option<JoinHandle<()>>>;
+
 /// Compute node info shared across several `compute_ctl` threads.
 pub struct ComputeNode {
     pub params: ComputeNodeParams,
@@ -125,6 +130,10 @@ pub struct ComputeNode {
     // key: ext_archive_name, value: started download time, download_completed?
     pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
     pub compute_ctl_config: ComputeCtlConfig,
+
+    /// Handle to the extension stats collection task
+    extension_stats_task: TaskHandle,
+    lfc_offload_task: TaskHandle,
 }
 
 // store some metrics about download size that might impact startup time
@@ -218,7 +227,8 @@ pub struct ParsedSpec {
     pub pageserver_connstr: String,
     pub safekeeper_connstrings: Vec<String>,
     pub storage_auth_token: Option<String>,
-    pub endpoint_storage_addr: Option<SocketAddr>,
+    /// k8s dns name and port
+    pub endpoint_storage_addr: Option<String>,
     pub endpoint_storage_token: Option<String>,
 }
 
@@ -313,13 +323,10 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
                 .or(Err("invalid timeline id"))?
         };
 
-        let endpoint_storage_addr: Option<SocketAddr> = spec
+        let endpoint_storage_addr: Option<String> = spec
             .endpoint_storage_addr
             .clone()
-            .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"))
-            .unwrap_or_default()
-            .parse()
-            .ok();
+            .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"));
         let endpoint_storage_token = spec
             .endpoint_storage_token
             .clone()
@@ -365,7 +372,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
 
 struct PostgresHandle {
     postgres: std::process::Child,
-    log_collector: tokio::task::JoinHandle<Result<()>>,
+    log_collector: JoinHandle<Result<()>>,
 }
 
 impl PostgresHandle {
@@ -379,7 +386,7 @@ struct StartVmMonitorResult {
     #[cfg(target_os = "linux")]
     token: tokio_util::sync::CancellationToken,
     #[cfg(target_os = "linux")]
-    vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
+    vm_monitor: Option<JoinHandle<Result<()>>>,
 }
 
 impl ComputeNode {
@@ -429,6 +436,8 @@ impl ComputeNode {
             state_changed: Condvar::new(),
             ext_download_progress: RwLock::new(HashMap::new()),
             compute_ctl_config: config.compute_ctl_config,
+            extension_stats_task: Mutex::new(None),
+            lfc_offload_task: Mutex::new(None),
         })
     }
 
@@ -516,6 +525,9 @@ impl ComputeNode {
             None
         };
 
+        this.terminate_extension_stats_task();
+        this.terminate_lfc_offload_task();
+
         // Terminate the vm_monitor so it releases the file watcher on
         // /sys/fs/cgroup/neon-postgres.
         // Note: the vm-monitor only runs on linux because it requires cgroups.
@@ -752,10 +764,15 @@ impl ComputeNode {
         // Configure and start rsyslog for compliance audit logging
         match pspec.spec.audit_log_level {
             ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
-                let remote_endpoint =
+                let remote_tls_endpoint =
+                    std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
+                let remote_plain_endpoint =
                     std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
-                if remote_endpoint.is_empty() {
-                    anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
+
+                if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
+                    anyhow::bail!(
+                        "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
+                    );
                 }
 
                 let log_directory_path = Path::new(&self.params.pgdata).join("log");
@@ -771,7 +788,8 @@ impl ComputeNode {
                     log_directory_path.clone(),
                     endpoint_id,
                     project_id,
-                    &remote_endpoint,
+                    &remote_plain_endpoint,
+                    &remote_tls_endpoint,
                 )?;
 
                 // Launch a background task to clean up the audit logs
@@ -838,12 +856,15 @@ impl ComputeNode {
         // Log metrics so that we can search for slow operations in logs
         info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
 
-        // Spawn the extension stats background task
         self.spawn_extension_stats_task();
 
         if pspec.spec.autoprewarm {
+            info!("autoprewarming on startup as requested");
             self.prewarm_lfc(None);
         }
+        if let Some(seconds) = pspec.spec.offload_lfc_interval_seconds {
+            self.spawn_lfc_offload_task(Duration::from_secs(seconds.into()));
+        };
         Ok(())
     }
 
@@ -998,13 +1019,80 @@ impl ComputeNode {
         Ok(())
     }
 
-    // Get basebackup from the libpq connection to pageserver using `connstr` and
-    // unarchive it to `pgdata` directory overriding all its previous content.
+    /// Fetches a basebackup from the Pageserver using the compute state's Pageserver connstring and
+    /// unarchives it to `pgdata` directory, replacing any existing contents.
     #[instrument(skip_all, fields(%lsn))]
     fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
         let spec = compute_state.pspec.as_ref().expect("spec must be set");
-        let start_time = Instant::now();
 
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let started = Instant::now();
+
+        let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
+            PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
+            PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
+        };
+
+        let mut state = self.state.lock().unwrap();
+        state.metrics.pageserver_connect_micros =
+            connected.duration_since(started).as_micros() as u64;
+        state.metrics.basebackup_bytes = size as u64;
+        state.metrics.basebackup_ms = started.elapsed().as_millis() as u64;
+
+        Ok(())
+    }
+
+    /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
+    /// the connection was established, and the (compressed) size of the basebackup.
+    fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
+        let shard0_connstr = spec
+            .pageserver_connstr
+            .split(',')
+            .next()
+            .unwrap()
+            .to_string();
+        let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
+            0 | 1 => ShardIndex::unsharded(),
+            count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
+        };
+
+        let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
+            let mut client = page_api::Client::new(
+                shard0_connstr,
+                spec.tenant_id,
+                spec.timeline_id,
+                shard_index,
+                spec.storage_auth_token.clone(),
+                None, // NB: base backups use payload compression
+            )
+            .await?;
+            let connected = Instant::now();
+            let reader = client
+                .get_base_backup(page_api::GetBaseBackupRequest {
+                    lsn: (lsn != Lsn(0)).then_some(lsn),
+                    compression: BaseBackupCompression::Gzip,
+                    replica: spec.spec.mode != ComputeMode::Primary,
+                    full: false,
+                })
+                .await?;
+            anyhow::Ok((reader, connected))
+        })?;
+
+        let mut reader = MeasuredReader::new(tokio_util::io::SyncIoBridge::new(reader));
+
+        // Set `ignore_zeros` so that unpack() reads the entire stream and doesn't just stop at the
+        // end-of-archive marker. If the server errors, the tar::Builder drop handler will write an
+        // end-of-archive marker before the error is emitted, and we would not see the error.
+        let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut reader));
+        ar.set_ignore_zeros(true);
+        ar.unpack(&self.params.pgdata)?;
+
+        Ok((connected, reader.get_byte_count()))
+    }
+
+    /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
+    /// when the connection was established, and the (compressed) size of the basebackup.
+    fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
         let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
         let mut config = postgres::Config::from_str(shard0_connstr)?;
 
@@ -1018,16 +1106,14 @@ impl ComputeNode {
         }
 
         config.application_name("compute_ctl");
-        if let Some(spec) = &compute_state.pspec {
-            config.options(&format!(
-                "-c neon.compute_mode={}",
-                spec.spec.mode.to_type_str()
-            ));
-        }
+        config.options(&format!(
+            "-c neon.compute_mode={}",
+            spec.spec.mode.to_type_str()
+        ));
 
         // Connect to pageserver
         let mut client = config.connect(NoTls)?;
-        let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
+        let connected = Instant::now();
 
         let basebackup_cmd = match lsn {
             Lsn(0) => {
@@ -1064,16 +1150,13 @@ impl ComputeNode {
         // Set `ignore_zeros` so that unpack() reads all the Copy data and
         // doesn't stop at the end-of-archive marker. Otherwise, if the server
         // sends an Error after finishing the tarball, we will not notice it.
+        // The tar::Builder drop handler will write an end-of-archive marker
+        // before emitting the error, and we would not see it otherwise.
         let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
         ar.set_ignore_zeros(true);
         ar.unpack(&self.params.pgdata)?;
 
-        // Report metrics
-        let mut state = self.state.lock().unwrap();
-        state.metrics.pageserver_connect_micros = pageserver_connect_micros;
-        state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
-        state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
-        Ok(())
+        Ok((connected, measured_reader.get_byte_count()))
     }
 
     // Gets the basebackup in a retry loop
@@ -1610,6 +1693,8 @@ impl ComputeNode {
             tls_config = self.compute_ctl_config.tls.clone();
         }
 
+        self.update_installed_extensions_collection_interval(&spec);
+
         let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
 
         // Merge-apply spec & changes to PostgreSQL state.
@@ -1674,6 +1759,8 @@ impl ComputeNode {
 
         let tls_config = self.tls_config(&spec);
 
+        self.update_installed_extensions_collection_interval(&spec);
+
         if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
             info!("tuning pgbouncer");
 
@@ -2278,24 +2365,92 @@ LIMIT 100",
     }
 
     pub fn spawn_extension_stats_task(&self) {
+        self.terminate_extension_stats_task();
+
         let conf = self.tokio_conn_conf.clone();
-        let installed_extensions_collection_interval =
-            self.params.installed_extensions_collection_interval;
-        tokio::spawn(async move {
-            // An initial sleep is added to ensure that two collections don't happen at the same time.
-            // The first collection happens during compute startup.
-            tokio::time::sleep(tokio::time::Duration::from_secs(
-                installed_extensions_collection_interval,
-            ))
-            .await;
-            let mut interval = tokio::time::interval(tokio::time::Duration::from_secs(
-                installed_extensions_collection_interval,
-            ));
+        let atomic_interval = self.params.installed_extensions_collection_interval.clone();
+        let mut installed_extensions_collection_interval =
+            2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst);
+        info!(
+            "[NEON_EXT_SPAWN] Spawning background installed extensions worker with Timeout: {}",
+            installed_extensions_collection_interval
+        );
+        let handle = tokio::spawn(async move {
             loop {
-                interval.tick().await;
+                info!(
+                    "[NEON_EXT_INT_SLEEP]: Interval: {}",
+                    installed_extensions_collection_interval
+                );
+                // Sleep at the start of the loop to ensure that two collections don't happen at the same time.
+                // The first collection happens during compute startup.
+                tokio::time::sleep(tokio::time::Duration::from_secs(
+                    installed_extensions_collection_interval,
+                ))
+                .await;
                 let _ = installed_extensions(conf.clone()).await;
+                // Acquire a read lock on the compute spec and then update the interval if necessary
+                installed_extensions_collection_interval = std::cmp::max(
+                    installed_extensions_collection_interval,
+                    2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst),
+                );
             }
         });
+
+        // Store the new task handle
+        *self.extension_stats_task.lock().unwrap() = Some(handle);
+    }
+
+    fn terminate_extension_stats_task(&self) {
+        if let Some(h) = self.extension_stats_task.lock().unwrap().take() {
+            h.abort()
+        }
+    }
+
+    pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
+        self.terminate_lfc_offload_task();
+        let secs = interval.as_secs();
+        info!("spawning lfc offload worker with {secs}s interval");
+        let this = self.clone();
+        let handle = spawn(async move {
+            let mut interval = time::interval(interval);
+            interval.tick().await; // returns immediately
+            loop {
+                interval.tick().await;
+                this.offload_lfc_async().await;
+            }
+        });
+        *self.lfc_offload_task.lock().unwrap() = Some(handle);
+    }
+
+    fn terminate_lfc_offload_task(&self) {
+        if let Some(h) = self.lfc_offload_task.lock().unwrap().take() {
+            h.abort()
+        }
+    }
+
+    fn update_installed_extensions_collection_interval(&self, spec: &ComputeSpec) {
+        // Update the interval for collecting installed extensions statistics
+        // If the value is -1, we never suspend so set the value to default collection.
+        // If the value is 0, it means default, we will just continue to use the default.
+        if spec.suspend_timeout_seconds == -1 || spec.suspend_timeout_seconds == 0 {
+            info!(
+                "[NEON_EXT_INT_UPD] Spec Timeout: {}, New Timeout: {}",
+                spec.suspend_timeout_seconds, DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL
+            );
+            self.params.installed_extensions_collection_interval.store(
+                DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL,
+                std::sync::atomic::Ordering::SeqCst,
+            );
+        } else {
+            info!(
+                "[NEON_EXT_INT_UPD] Spec Timeout: {}",
+                spec.suspend_timeout_seconds
+            );
+            self.params.installed_extensions_collection_interval.store(
+                spec.suspend_timeout_seconds as u64,
+                std::sync::atomic::Ordering::SeqCst,
+            );
+        }
     }
 }
 

compute_tools/src/compute_prewarm.rs

@@ -5,6 +5,7 @@ use compute_api::responses::LfcOffloadState;
 use compute_api::responses::LfcPrewarmState;
 use http::StatusCode;
 use reqwest::Client;
+use std::mem::replace;
 use std::sync::Arc;
 use tokio::{io::AsyncReadExt, spawn};
 use tracing::{error, info};
@@ -88,17 +89,15 @@ impl ComputeNode {
         self.state.lock().unwrap().lfc_offload_state.clone()
     }
 
-    /// Returns false if there is a prewarm request ongoing, true otherwise
+    /// If there is a prewarm request ongoing, return false, true otherwise
     pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
-        crate::metrics::LFC_PREWARM_REQUESTS.inc();
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
-            if let LfcPrewarmState::Prewarming =
-                std::mem::replace(state, LfcPrewarmState::Prewarming)
-            {
+            if let LfcPrewarmState::Prewarming = replace(state, LfcPrewarmState::Prewarming) {
                 return false;
             }
         }
+        crate::metrics::LFC_PREWARMS.inc();
 
         let cloned = self.clone();
         spawn(async move {
@@ -152,32 +151,41 @@ impl ComputeNode {
             .map(|_| ())
     }
 
-    /// Returns false if there is an offload request ongoing, true otherwise
+    /// If offload request is ongoing, return false, true otherwise
     pub fn offload_lfc(self: &Arc<Self>) -> bool {
-        crate::metrics::LFC_OFFLOAD_REQUESTS.inc();
         {
             let state = &mut self.state.lock().unwrap().lfc_offload_state;
-            if let LfcOffloadState::Offloading =
-                std::mem::replace(state, LfcOffloadState::Offloading)
-            {
+            if replace(state, LfcOffloadState::Offloading) == LfcOffloadState::Offloading {
                 return false;
             }
         }
-
         let cloned = self.clone();
-        spawn(async move {
-            let Err(err) = cloned.offload_lfc_impl().await else {
-                cloned.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
-                return;
-            };
-            error!(%err);
-            cloned.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
-                error: err.to_string(),
-            };
-        });
+        spawn(async move { cloned.offload_lfc_with_state_update().await });
         true
     }
 
+    pub async fn offload_lfc_async(self: &Arc<Self>) {
+        {
+            let state = &mut self.state.lock().unwrap().lfc_offload_state;
+            if replace(state, LfcOffloadState::Offloading) == LfcOffloadState::Offloading {
+                return;
+            }
+        }
+        self.offload_lfc_with_state_update().await
+    }
+
+    async fn offload_lfc_with_state_update(&self) {
+        crate::metrics::LFC_OFFLOADS.inc();
+        let Err(err) = self.offload_lfc_impl().await else {
+            self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
+            return;
+        };
+        error!(%err);
+        self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
+            error: err.to_string(),
+        };
+    }
+
     async fn offload_lfc_impl(&self) -> Result<()> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
         info!(%url, "requesting LFC state from postgres");

compute_tools/src/config_template/compute_audit_rsyslog_template.conf

@@ -10,7 +10,13 @@ input(type="imfile" File="{log_directory}/*.log"
   startmsg.regex="^[[:digit:]]{{4}}-[[:digit:]]{{2}}-[[:digit:]]{{2}} [[:digit:]]{{2}}:[[:digit:]]{{2}}:[[:digit:]]{{2}}.[[:digit:]]{{3}} GMT,")
 
 # the directory to store rsyslog state files
-global(workDirectory="/var/log/rsyslog")
+global(
+  workDirectory="/var/log/rsyslog"
+  DefaultNetstreamDriverCAFile="/etc/ssl/certs/ca-certificates.crt"
+)
+
+# Whether the remote syslog receiver uses tls
+set $.remote_syslog_tls = "{remote_syslog_tls}";
 
 # Construct json, endpoint_id and project_id as additional metadata
 set $.json_log!endpoint_id = "{endpoint_id}";
@@ -21,5 +27,29 @@ set $.json_log!msg = $msg;
 template(name="PgAuditLog" type="string"
     string="<%PRI%>1 %TIMESTAMP:::date-rfc3339% %HOSTNAME% - - - - %$.json_log%")
 
-# Forward to remote syslog receiver (@@<hostname>:<port>;format
-local5.info @@{remote_endpoint};PgAuditLog
+# Forward to remote syslog receiver (over TLS)
+if ( $syslogtag == 'pgaudit_log' ) then {{
+  if ( $.remote_syslog_tls == 'true' ) then {{
+    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
+      template="PgAuditLog"
+      queue.type="linkedList"
+      queue.size="1000"
+      action.ResumeRetryCount="10"
+      StreamDriver="gtls"
+      StreamDriverMode="1"
+      StreamDriverAuthMode="x509/name"
+      StreamDriverPermittedPeers="{remote_syslog_host}"
+      StreamDriver.CheckExtendedKeyPurpose="on"
+      StreamDriver.PermitExpiredCerts="off"
+    )
+    stop
+  }} else {{
+    action(type="omfwd" target="{remote_syslog_host}" port="{remote_syslog_port}" protocol="tcp"
+      template="PgAuditLog"
+      queue.type="linkedList"
+      queue.size="1000"
+      action.ResumeRetryCount="10"
+    )
+    stop
+  }}
+}}

compute_tools/src/lsn_lease.rs

@@ -4,7 +4,9 @@ use std::thread;
 use std::time::{Duration, SystemTime};
 
 use anyhow::{Result, bail};
-use compute_api::spec::ComputeMode;
+use compute_api::spec::{ComputeMode, PageserverProtocol};
+use itertools::Itertools as _;
+use pageserver_page_api as page_api;
 use postgres::{NoTls, SimpleQueryMessage};
 use tracing::{info, warn};
 use utils::id::{TenantId, TimelineId};
@@ -76,25 +78,17 @@ fn acquire_lsn_lease_with_retry(
 
     loop {
         // Note: List of pageservers is dynamic, need to re-read configs before each attempt.
-        let configs = {
+        let (connstrings, auth) = {
             let state = compute.state.lock().unwrap();
-
             let spec = state.pspec.as_ref().expect("spec must be set");
-
-            let conn_strings = spec.pageserver_connstr.split(',');
-
-            conn_strings
-                .map(|connstr| {
-                    let mut config = postgres::Config::from_str(connstr).expect("Invalid connstr");
-                    if let Some(storage_auth_token) = &spec.storage_auth_token {
-                        config.password(storage_auth_token.clone());
-                    }
-                    config
-                })
-                .collect::<Vec<_>>()
+            (
+                spec.pageserver_connstr.clone(),
+                spec.storage_auth_token.clone(),
+            )
         };
 
-        let result = try_acquire_lsn_lease(tenant_id, timeline_id, lsn, &configs);
+        let result =
+            try_acquire_lsn_lease(&connstrings, auth.as_deref(), tenant_id, timeline_id, lsn);
         match result {
             Ok(Some(res)) => {
                 return Ok(res);
@@ -116,68 +110,104 @@ fn acquire_lsn_lease_with_retry(
     }
 }
 
-/// Tries to acquire an LSN lease through PS page_service API.
+/// Tries to acquire LSN leases on all Pageserver shards.
 fn try_acquire_lsn_lease(
+    connstrings: &str,
+    auth: Option<&str>,
     tenant_id: TenantId,
     timeline_id: TimelineId,
     lsn: Lsn,
-    configs: &[postgres::Config],
 ) -> Result<Option<SystemTime>> {
-    fn get_valid_until(
-        config: &postgres::Config,
-        tenant_shard_id: TenantShardId,
-        timeline_id: TimelineId,
-        lsn: Lsn,
-    ) -> Result<Option<SystemTime>> {
-        let mut client = config.connect(NoTls)?;
-        let cmd = format!("lease lsn {tenant_shard_id} {timeline_id} {lsn} ");
-        let res = client.simple_query(&cmd)?;
-        let msg = match res.first() {
-            Some(msg) => msg,
-            None => bail!("empty response"),
-        };
-        let row = match msg {
-            SimpleQueryMessage::Row(row) => row,
-            _ => bail!("error parsing lsn lease response"),
+    let connstrings = connstrings.split(',').collect_vec();
+    let shard_count = connstrings.len();
+    let mut leases = Vec::new();
+
+    for (shard_number, &connstring) in connstrings.iter().enumerate() {
+        let tenant_shard_id = match shard_count {
+            0 | 1 => TenantShardId::unsharded(tenant_id),
+            shard_count => TenantShardId {
+                tenant_id,
+                shard_number: ShardNumber(shard_number as u8),
+                shard_count: ShardCount::new(shard_count as u8),
+            },
         };
 
-        // Note: this will be None if a lease is explicitly not granted.
-        let valid_until_str = row.get("valid_until");
-
-        let valid_until = valid_until_str.map(|s| {
-            SystemTime::UNIX_EPOCH
-                .checked_add(Duration::from_millis(u128::from_str(s).unwrap() as u64))
-                .expect("Time larger than max SystemTime could handle")
-        });
-        Ok(valid_until)
+        let lease = match PageserverProtocol::from_connstring(connstring)? {
+            PageserverProtocol::Libpq => {
+                acquire_lsn_lease_libpq(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
+            PageserverProtocol::Grpc => {
+                acquire_lsn_lease_grpc(connstring, auth, tenant_shard_id, timeline_id, lsn)?
+            }
+        };
+        leases.push(lease);
     }
 
-    let shard_count = configs.len();
+    Ok(leases.into_iter().min().flatten())
+}
 
-    let valid_until = if shard_count > 1 {
-        configs
-            .iter()
-            .enumerate()
-            .map(|(shard_number, config)| {
-                let tenant_shard_id = TenantShardId {
-                    tenant_id,
-                    shard_count: ShardCount::new(shard_count as u8),
-                    shard_number: ShardNumber(shard_number as u8),
-                };
-                get_valid_until(config, tenant_shard_id, timeline_id, lsn)
-            })
-            .collect::<Result<Vec<Option<SystemTime>>>>()?
-            .into_iter()
-            .min()
-            .unwrap()
-    } else {
-        get_valid_until(
-            &configs[0],
-            TenantShardId::unsharded(tenant_id),
-            timeline_id,
-            lsn,
-        )?
+/// Acquires an LSN lease on a single shard, using the libpq API. The connstring must use a
+/// postgresql:// scheme.
+fn acquire_lsn_lease_libpq(
+    connstring: &str,
+    auth: Option<&str>,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+    lsn: Lsn,
+) -> Result<Option<SystemTime>> {
+    let mut config = postgres::Config::from_str(connstring)?;
+    if let Some(auth) = auth {
+        config.password(auth);
+    }
+    let mut client = config.connect(NoTls)?;
+    let cmd = format!("lease lsn {tenant_shard_id} {timeline_id} {lsn} ");
+    let res = client.simple_query(&cmd)?;
+    let msg = match res.first() {
+        Some(msg) => msg,
+        None => bail!("empty response"),
+    };
+    let row = match msg {
+        SimpleQueryMessage::Row(row) => row,
+        _ => bail!("error parsing lsn lease response"),
     };
 
+    // Note: this will be None if a lease is explicitly not granted.
+    let valid_until_str = row.get("valid_until");
+
+    let valid_until = valid_until_str.map(|s| {
+        SystemTime::UNIX_EPOCH
+            .checked_add(Duration::from_millis(u128::from_str(s).unwrap() as u64))
+            .expect("Time larger than max SystemTime could handle")
+    });
     Ok(valid_until)
 }
+
+/// Acquires an LSN lease on a single shard, using the gRPC API. The connstring must use a
+/// grpc:// scheme.
+fn acquire_lsn_lease_grpc(
+    connstring: &str,
+    auth: Option<&str>,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+    lsn: Lsn,
+) -> Result<Option<SystemTime>> {
+    tokio::runtime::Handle::current().block_on(async move {
+        let mut client = page_api::Client::new(
+            connstring.to_string(),
+            tenant_shard_id.tenant_id,
+            timeline_id,
+            tenant_shard_id.to_index(),
+            auth.map(String::from),
+            None,
+        )
+        .await?;
+
+        let req = page_api::LeaseLsnRequest { lsn };
+        match client.lease_lsn(req).await {
+            Ok(expires) => Ok(Some(expires)),
+            // Lease couldn't be acquired because the LSN has been garbage collected.
+            Err(err) if err.code() == tonic::Code::FailedPrecondition => Ok(None),
+            Err(err) => Err(err.into()),
+        }
+    })
+}

compute_tools/src/metrics.rs

@@ -97,20 +97,18 @@ pub(crate) static PG_TOTAL_DOWNTIME_MS: Lazy<GenericCounter<AtomicU64>> = Lazy::
     .expect("failed to define a metric")
 });
 
-/// Needed as neon.file_cache_prewarm_batch == 0 doesn't mean we never tried to prewarm.
-/// On the other hand, LFC_PREWARMED_PAGES is excessive as we can GET /lfc/prewarm
-pub(crate) static LFC_PREWARM_REQUESTS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static LFC_PREWARMS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
-        "compute_ctl_lfc_prewarm_requests_total",
-        "Total number of LFC prewarm requests made by compute_ctl",
+        "compute_ctl_lfc_prewarms_total",
+        "Total number of LFC prewarms requested by compute_ctl or autoprewarm option",
     )
     .expect("failed to define a metric")
 });
 
-pub(crate) static LFC_OFFLOAD_REQUESTS: Lazy<IntCounter> = Lazy::new(|| {
+pub(crate) static LFC_OFFLOADS: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
-        "compute_ctl_lfc_offload_requests_total",
-        "Total number of LFC offload requests made by compute_ctl",
+        "compute_ctl_lfc_offloads_total",
+        "Total number of LFC offloads requested by compute_ctl or lfc_offload_period_seconds option",
     )
     .expect("failed to define a metric")
 });
@@ -124,7 +122,7 @@ pub fn collect() -> Vec<MetricFamily> {
     metrics.extend(AUDIT_LOG_DIR_SIZE.collect());
     metrics.extend(PG_CURR_DOWNTIME_MS.collect());
     metrics.extend(PG_TOTAL_DOWNTIME_MS.collect());
-    metrics.extend(LFC_PREWARM_REQUESTS.collect());
-    metrics.extend(LFC_OFFLOAD_REQUESTS.collect());
+    metrics.extend(LFC_PREWARMS.collect());
+    metrics.extend(LFC_OFFLOADS.collect());
     metrics
 }

compute_tools/src/rsyslog.rs

@@ -4,8 +4,10 @@ use std::path::Path;
 use std::process::Command;
 use std::time::Duration;
 use std::{fs::OpenOptions, io::Write};
+use url::{Host, Url};
 
 use anyhow::{Context, Result, anyhow};
+use hostname_validator;
 use tracing::{error, info, instrument, warn};
 
 const POSTGRES_LOGS_CONF_PATH: &str = "/etc/rsyslog.d/postgres_logs.conf";
@@ -82,18 +84,84 @@ fn restart_rsyslog() -> Result<()> {
     Ok(())
 }
 
+fn parse_audit_syslog_address(
+    remote_plain_endpoint: &str,
+    remote_tls_endpoint: &str,
+) -> Result<(String, u16, String)> {
+    let tls;
+    let remote_endpoint = if !remote_tls_endpoint.is_empty() {
+        tls = "true".to_string();
+        remote_tls_endpoint
+    } else {
+        tls = "false".to_string();
+        remote_plain_endpoint
+    };
+    // Urlify the remote_endpoint, so parsing can be done with url::Url.
+    let url_str = format!("http://{remote_endpoint}");
+    let url = Url::parse(&url_str).map_err(|err| {
+        anyhow!("Error parsing {remote_endpoint}, expected host:port, got {err:?}")
+    })?;
+
+    let is_valid = url.scheme() == "http"
+        && url.path() == "/"
+        && url.query().is_none()
+        && url.fragment().is_none()
+        && url.username() == ""
+        && url.password().is_none();
+
+    if !is_valid {
+        return Err(anyhow!(
+            "Invalid address format {remote_endpoint}, expected host:port"
+        ));
+    }
+    let host = match url.host() {
+        Some(Host::Domain(h)) if hostname_validator::is_valid(h) => h.to_string(),
+        Some(Host::Ipv4(ip4)) => ip4.to_string(),
+        Some(Host::Ipv6(ip6)) => ip6.to_string(),
+        _ => return Err(anyhow!("Invalid host")),
+    };
+    let port = url
+        .port()
+        .ok_or_else(|| anyhow!("Invalid port in {remote_endpoint}"))?;
+
+    Ok((host, port, tls))
+}
+
+fn generate_audit_rsyslog_config(
+    log_directory: String,
+    endpoint_id: &str,
+    project_id: &str,
+    remote_syslog_host: &str,
+    remote_syslog_port: u16,
+    remote_syslog_tls: &str,
+) -> String {
+    format!(
+        include_str!("config_template/compute_audit_rsyslog_template.conf"),
+        log_directory = log_directory,
+        endpoint_id = endpoint_id,
+        project_id = project_id,
+        remote_syslog_host = remote_syslog_host,
+        remote_syslog_port = remote_syslog_port,
+        remote_syslog_tls = remote_syslog_tls
+    )
+}
+
 pub fn configure_audit_rsyslog(
     log_directory: String,
     endpoint_id: &str,
     project_id: &str,
     remote_endpoint: &str,
+    remote_tls_endpoint: &str,
 ) -> Result<()> {
-    let config_content: String = format!(
-        include_str!("config_template/compute_audit_rsyslog_template.conf"),
-        log_directory = log_directory,
-        endpoint_id = endpoint_id,
-        project_id = project_id,
-        remote_endpoint = remote_endpoint
+    let (remote_syslog_host, remote_syslog_port, remote_syslog_tls) =
+        parse_audit_syslog_address(remote_endpoint, remote_tls_endpoint).unwrap();
+    let config_content = generate_audit_rsyslog_config(
+        log_directory,
+        endpoint_id,
+        project_id,
+        &remote_syslog_host,
+        remote_syslog_port,
+        &remote_syslog_tls,
     );
 
     info!("rsyslog config_content: {}", config_content);
@@ -258,6 +326,8 @@ pub fn launch_pgaudit_gc(log_directory: String) {
 mod tests {
     use crate::rsyslog::PostgresLogsRsyslogConfig;
 
+    use super::{generate_audit_rsyslog_config, parse_audit_syslog_address};
+
     #[test]
     fn test_postgres_logs_config() {
         {
@@ -287,4 +357,146 @@ mod tests {
             assert!(res.is_err());
         }
     }
+
+    #[test]
+    fn test_parse_audit_syslog_address() {
+        {
+            // host:port format (plaintext)
+            let parsed = parse_audit_syslog_address("collector.host.tld:5555", "");
+            assert!(parsed.is_ok());
+            assert_eq!(
+                parsed.unwrap(),
+                (
+                    String::from("collector.host.tld"),
+                    5555,
+                    String::from("false")
+                )
+            );
+        }
+
+        {
+            // host:port format with ipv4 ip address (plaintext)
+            let parsed = parse_audit_syslog_address("10.0.0.1:5555", "");
+            assert!(parsed.is_ok());
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("10.0.0.1"), 5555, String::from("false"))
+            );
+        }
+
+        {
+            // host:port format with ipv6 ip address (plaintext)
+            let parsed =
+                parse_audit_syslog_address("[7e60:82ed:cb2e:d617:f904:f395:aaca:e252]:5555", "");
+            assert_eq!(
+                parsed.unwrap(),
+                (
+                    String::from("7e60:82ed:cb2e:d617:f904:f395:aaca:e252"),
+                    5555,
+                    String::from("false")
+                )
+            );
+        }
+
+        {
+            // Only TLS host:port defined
+            let parsed = parse_audit_syslog_address("", "tls.host.tld:5556");
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("tls.host.tld"), 5556, String::from("true"))
+            );
+        }
+
+        {
+            // tls host should take precedence, when both defined
+            let parsed = parse_audit_syslog_address("plaintext.host.tld:5555", "tls.host.tld:5556");
+            assert_eq!(
+                parsed.unwrap(),
+                (String::from("tls.host.tld"), 5556, String::from("true"))
+            );
+        }
+
+        {
+            // host without port (plaintext)
+            let parsed = parse_audit_syslog_address("collector.host.tld", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // port without host
+            let parsed = parse_audit_syslog_address(":5555", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // valid host with invalid port
+            let parsed = parse_audit_syslog_address("collector.host.tld:90001", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // invalid hostname with valid port
+            let parsed = parse_audit_syslog_address("-collector.host.tld:5555", "");
+            assert!(parsed.is_err());
+        }
+
+        {
+            // parse error
+            let parsed = parse_audit_syslog_address("collector.host.tld:::5555", "");
+            assert!(parsed.is_err());
+        }
+    }
+
+    #[test]
+    fn test_generate_audit_rsyslog_config() {
+        {
+            // plaintext version
+            let log_directory = "/tmp/log".to_string();
+            let endpoint_id = "ep-test-endpoint-id";
+            let project_id = "test-project-id";
+            let remote_syslog_host = "collector.host.tld";
+            let remote_syslog_port = 5555;
+            let remote_syslog_tls = "false";
+
+            let conf_str = generate_audit_rsyslog_config(
+                log_directory,
+                endpoint_id,
+                project_id,
+                remote_syslog_host,
+                remote_syslog_port,
+                remote_syslog_tls,
+            );
+
+            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "false";"#));
+            assert!(conf_str.contains(r#"type="omfwd""#));
+            assert!(conf_str.contains(r#"target="collector.host.tld""#));
+            assert!(conf_str.contains(r#"port="5555""#));
+            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
+        }
+
+        {
+            // TLS version
+            let log_directory = "/tmp/log".to_string();
+            let endpoint_id = "ep-test-endpoint-id";
+            let project_id = "test-project-id";
+            let remote_syslog_host = "collector.host.tld";
+            let remote_syslog_port = 5556;
+            let remote_syslog_tls = "true";
+
+            let conf_str = generate_audit_rsyslog_config(
+                log_directory,
+                endpoint_id,
+                project_id,
+                remote_syslog_host,
+                remote_syslog_port,
+                remote_syslog_tls,
+            );
+
+            assert!(conf_str.contains(r#"set $.remote_syslog_tls = "true";"#));
+            assert!(conf_str.contains(r#"type="omfwd""#));
+            assert!(conf_str.contains(r#"target="collector.host.tld""#));
+            assert!(conf_str.contains(r#"port="5556""#));
+            assert!(conf_str.contains(r#"StreamDriverPermittedPeers="collector.host.tld""#));
+        }
+    }
 }

compute_tools/tests/cluster_spec.json

@@ -3,7 +3,8 @@
 
   "timestamp": "2021-05-23T18:25:43.511Z",
   "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8b",
-
+  "suspend_timeout_seconds": 3600,
+  
   "cluster": {
     "cluster_id": "test-cluster-42",
     "name": "Zenith Test",

compute_tools/tests/pg_helpers_tests.rs

@@ -31,6 +31,7 @@ mod pg_helpers_tests {
 wal_level = logical
 hot_standby = on
 autoprewarm = off
+offload_lfc_interval_seconds = 20
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on
 log_connections = on

control_plane/src/bin/neon_local.rs

@@ -16,9 +16,9 @@ use std::time::Duration;
 use anyhow::{Context, Result, anyhow, bail};
 use clap::Parser;
 use compute_api::requests::ComputeClaimsScope;
-use compute_api::spec::ComputeMode;
+use compute_api::spec::{ComputeMode, PageserverProtocol};
 use control_plane::broker::StorageBroker;
-use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode, PageserverProtocol};
+use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
 use control_plane::local_env;
 use control_plane::local_env::{
@@ -64,7 +64,9 @@ const DEFAULT_PAGESERVER_ID: NodeId = NodeId(1);
 const DEFAULT_BRANCH_NAME: &str = "main";
 project_git_version!(GIT_VERSION);
 
+#[allow(dead_code)]
 const DEFAULT_PG_VERSION: PgMajorVersion = PgMajorVersion::PG17;
+const DEFAULT_PG_VERSION_NUM: &str = "17";
 
 const DEFAULT_PAGESERVER_CONTROL_PLANE_API: &str = "http://127.0.0.1:1234/upcall/v1/";
 
@@ -167,7 +169,7 @@ struct TenantCreateCmdArgs {
     #[clap(short = 'c')]
     config: Vec<String>,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version to use for the initial timeline")]
     pg_version: PgMajorVersion,
 
@@ -290,7 +292,7 @@ struct TimelineCreateCmdArgs {
     #[clap(long, help = "Human-readable alias for the new timeline")]
     branch_name: String,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version")]
     pg_version: PgMajorVersion,
 }
@@ -322,7 +324,7 @@ struct TimelineImportCmdArgs {
     #[clap(long, help = "Lsn the basebackup ends at")]
     end_lsn: Option<Lsn>,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version of the backup being imported")]
     pg_version: PgMajorVersion,
 }
@@ -601,7 +603,7 @@ struct EndpointCreateCmdArgs {
     )]
     config_only: bool,
 
-    #[arg(default_value_t = DEFAULT_PG_VERSION)]
+    #[arg(default_value = DEFAULT_PG_VERSION_NUM)]
     #[clap(long, help = "Postgres version")]
     pg_version: PgMajorVersion,
 
@@ -673,6 +675,16 @@ struct EndpointStartCmdArgs {
     #[arg(default_value = "90s")]
     start_timeout: Duration,
 
+    #[clap(
+        long,
+        help = "Download LFC cache from endpoint storage on endpoint startup",
+        default_value = "false"
+    )]
+    autoprewarm: bool,
+
+    #[clap(long, help = "Upload LFC cache to endpoint storage periodically")]
+    offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+
     #[clap(
         long,
         help = "Run in development mode, skipping VM-specific operations like process termination",
@@ -1583,22 +1595,24 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             let endpoint_storage_token = env.generate_auth_token(&claims)?;
             let endpoint_storage_addr = env.endpoint_storage.listen_addr.to_string();
 
+            let args = control_plane::endpoint::EndpointStartArgs {
+                auth_token,
+                endpoint_storage_token,
+                endpoint_storage_addr,
+                safekeepers_generation,
+                safekeepers,
+                pageservers,
+                remote_ext_base_url: remote_ext_base_url.clone(),
+                shard_stripe_size: stripe_size.0 as usize,
+                create_test_user: args.create_test_user,
+                start_timeout: args.start_timeout,
+                autoprewarm: args.autoprewarm,
+                offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
+                dev: args.dev,
+            };
+
             println!("Starting existing endpoint {endpoint_id}...");
-            endpoint
-                .start(
-                    &auth_token,
-                    endpoint_storage_token,
-                    endpoint_storage_addr,
-                    safekeepers_generation,
-                    safekeepers,
-                    pageservers,
-                    remote_ext_base_url.as_ref(),
-                    stripe_size.0 as usize,
-                    args.create_test_user,
-                    args.start_timeout,
-                    args.dev,
-                )
-                .await?;
+            endpoint.start(args).await?;
         }
         EndpointCmd::Reconfigure(args) => {
             let endpoint_id = &args.endpoint_id;
@@ -1649,7 +1663,9 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             // If --safekeepers argument is given, use only the listed
             // safekeeper nodes; otherwise all from the env.
             let safekeepers = parse_safekeepers(&args.safekeepers)?;
-            endpoint.reconfigure(pageservers, None, safekeepers).await?;
+            endpoint
+                .reconfigure(Some(pageservers), None, safekeepers, None)
+                .await?;
         }
         EndpointCmd::Stop(args) => {
             let endpoint_id = &args.endpoint_id;

control_plane/src/endpoint.rs

@@ -56,8 +56,8 @@ use compute_api::responses::{
     TlsConfig,
 };
 use compute_api::spec::{
-    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PgIdent,
-    RemoteExtSpec, Role,
+    Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PageserverProtocol,
+    PgIdent, RemoteExtSpec, Role,
 };
 use jsonwebtoken::jwk::{
     AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
@@ -373,27 +373,20 @@ impl std::fmt::Display for EndpointTerminateMode {
     }
 }
 
-/// Protocol used to connect to a Pageserver.
-#[derive(Clone, Copy, Debug)]
-pub enum PageserverProtocol {
-    Libpq,
-    Grpc,
-}
-
-impl PageserverProtocol {
-    /// Returns the URL scheme for the protocol, used in connstrings.
-    pub fn scheme(&self) -> &'static str {
-        match self {
-            Self::Libpq => "postgresql",
-            Self::Grpc => "grpc",
-        }
-    }
-}
-
-impl Display for PageserverProtocol {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.scheme())
-    }
+pub struct EndpointStartArgs {
+    pub auth_token: Option<String>,
+    pub endpoint_storage_token: String,
+    pub endpoint_storage_addr: String,
+    pub safekeepers_generation: Option<SafekeeperGeneration>,
+    pub safekeepers: Vec<NodeId>,
+    pub pageservers: Vec<(PageserverProtocol, Host, u16)>,
+    pub remote_ext_base_url: Option<String>,
+    pub shard_stripe_size: usize,
+    pub create_test_user: bool,
+    pub start_timeout: Duration,
+    pub autoprewarm: bool,
+    pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+    pub dev: bool,
 }
 
 impl Endpoint {
@@ -700,21 +693,7 @@ impl Endpoint {
         })
     }
 
-    #[allow(clippy::too_many_arguments)]
-    pub async fn start(
-        &self,
-        auth_token: &Option<String>,
-        endpoint_storage_token: String,
-        endpoint_storage_addr: String,
-        safekeepers_generation: Option<SafekeeperGeneration>,
-        safekeepers: Vec<NodeId>,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
-        remote_ext_base_url: Option<&String>,
-        shard_stripe_size: usize,
-        create_test_user: bool,
-        start_timeout: Duration,
-        dev: bool,
-    ) -> Result<()> {
+    pub async fn start(&self, args: EndpointStartArgs) -> Result<()> {
         if self.status() == EndpointStatus::Running {
             anyhow::bail!("The endpoint is already running");
         }
@@ -727,10 +706,10 @@ impl Endpoint {
             std::fs::remove_dir_all(self.pgdata())?;
         }
 
-        let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
+        let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
         assert!(!pageserver_connstring.is_empty());
 
-        let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
+        let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;
 
         // check for file remote_extensions_spec.json
         // if it is present, read it and pass to compute_ctl
@@ -758,7 +737,7 @@ impl Endpoint {
                     cluster_id: None, // project ID: not used
                     name: None,       // project name: not used
                     state: None,
-                    roles: if create_test_user {
+                    roles: if args.create_test_user {
                         vec![Role {
                             name: PgIdent::from_str("test").unwrap(),
                             encrypted_password: None,
@@ -767,7 +746,7 @@ impl Endpoint {
                     } else {
                         Vec::new()
                     },
-                    databases: if create_test_user {
+                    databases: if args.create_test_user {
                         vec![Database {
                             name: PgIdent::from_str("neondb").unwrap(),
                             owner: PgIdent::from_str("test").unwrap(),
@@ -789,20 +768,22 @@ impl Endpoint {
                 endpoint_id: Some(self.endpoint_id.clone()),
                 mode: self.mode,
                 pageserver_connstring: Some(pageserver_connstring),
-                safekeepers_generation: safekeepers_generation.map(|g| g.into_inner()),
+                safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
                 safekeeper_connstrings,
-                storage_auth_token: auth_token.clone(),
+                storage_auth_token: args.auth_token.clone(),
                 remote_extensions,
                 pgbouncer_settings: None,
-                shard_stripe_size: Some(shard_stripe_size),
+                shard_stripe_size: Some(args.shard_stripe_size),
                 local_proxy_config: None,
                 reconfigure_concurrency: self.reconfigure_concurrency,
                 drop_subscriptions_before_start: self.drop_subscriptions_before_start,
                 audit_log_level: ComputeAudit::Disabled,
                 logs_export_host: None::<String>,
-                endpoint_storage_addr: Some(endpoint_storage_addr),
-                endpoint_storage_token: Some(endpoint_storage_token),
-                autoprewarm: false,
+                endpoint_storage_addr: Some(args.endpoint_storage_addr),
+                endpoint_storage_token: Some(args.endpoint_storage_token),
+                autoprewarm: args.autoprewarm,
+                offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
+                suspend_timeout_seconds: -1, // Only used in neon_local.
             };
 
             // this strange code is needed to support respec() in tests
@@ -813,7 +794,7 @@ impl Endpoint {
                 debug!("spec.cluster {:?}", spec.cluster);
 
                 // fill missing fields again
-                if create_test_user {
+                if args.create_test_user {
                     spec.cluster.roles.push(Role {
                         name: PgIdent::from_str("test").unwrap(),
                         encrypted_password: None,
@@ -848,7 +829,7 @@ impl Endpoint {
         // Launch compute_ctl
         let conn_str = self.connstr("cloud_admin", "postgres");
         println!("Starting postgres node at '{conn_str}'");
-        if create_test_user {
+        if args.create_test_user {
             let conn_str = self.connstr("test", "neondb");
             println!("Also at '{conn_str}'");
         }
@@ -880,11 +861,11 @@ impl Endpoint {
         .stderr(logfile.try_clone()?)
         .stdout(logfile);
 
-        if let Some(remote_ext_base_url) = remote_ext_base_url {
-            cmd.args(["--remote-ext-base-url", remote_ext_base_url]);
+        if let Some(remote_ext_base_url) = args.remote_ext_base_url {
+            cmd.args(["--remote-ext-base-url", &remote_ext_base_url]);
         }
 
-        if dev {
+        if args.dev {
             cmd.arg("--dev");
         }
 
@@ -916,10 +897,11 @@ impl Endpoint {
                 Ok(state) => {
                     match state.status {
                         ComputeStatus::Init => {
-                            if Instant::now().duration_since(start_at) > start_timeout {
+                            let timeout = args.start_timeout;
+                            if Instant::now().duration_since(start_at) > timeout {
                                 bail!(
                                     "compute startup timed out {:?}; still in Init state",
-                                    start_timeout
+                                    timeout
                                 );
                             }
                             // keep retrying
@@ -947,9 +929,10 @@ impl Endpoint {
                     }
                 }
                 Err(e) => {
-                    if Instant::now().duration_since(start_at) > start_timeout {
+                    if Instant::now().duration_since(start_at) > args.start_timeout {
                         return Err(e).context(format!(
-                            "timed out {start_timeout:?} waiting to connect to compute_ctl HTTP",
+                            "timed out {:?} waiting to connect to compute_ctl HTTP",
+                            args.start_timeout
                         ));
                     }
                 }
@@ -997,12 +980,11 @@ impl Endpoint {
 
     pub async fn reconfigure(
         &self,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+        pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
         stripe_size: Option<ShardStripeSize>,
         safekeepers: Option<Vec<NodeId>>,
+        safekeeper_generation: Option<SafekeeperGeneration>,
     ) -> Result<()> {
-        anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
-
         let (mut spec, compute_ctl_config) = {
             let config_path = self.endpoint_path().join("config.json");
             let file = std::fs::File::open(config_path)?;
@@ -1014,16 +996,24 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
-        spec.pageserver_connstring = Some(pageserver_connstr);
-        if stripe_size.is_some() {
-            spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+        // If pageservers are not specified, don't change them.
+        if let Some(pageservers) = pageservers {
+            anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
+
+            let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
+            spec.pageserver_connstring = Some(pageserver_connstr);
+            if stripe_size.is_some() {
+                spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+            }
         }
 
         // If safekeepers are not specified, don't change them.
         if let Some(safekeepers) = safekeepers {
             let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
             spec.safekeeper_connstrings = safekeeper_connstrings;
+            if let Some(g) = safekeeper_generation {
+                spec.safekeepers_generation = Some(g.into_inner());
+            }
         }
 
         let client = reqwest::Client::builder()
@@ -1061,6 +1051,24 @@ impl Endpoint {
         }
     }
 
+    pub async fn reconfigure_pageservers(
+        &self,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> Result<()> {
+        self.reconfigure(Some(pageservers), stripe_size, None, None)
+            .await
+    }
+
+    pub async fn reconfigure_safekeepers(
+        &self,
+        safekeepers: Vec<NodeId>,
+        generation: SafekeeperGeneration,
+    ) -> Result<()> {
+        self.reconfigure(None, None, Some(safekeepers), Some(generation))
+            .await
+    }
+
     pub async fn stop(
         &self,
         mode: EndpointTerminateMode,

control_plane/src/local_env.rs

@@ -12,6 +12,7 @@ use std::{env, fs};
 
 use anyhow::{Context, bail};
 use clap::ValueEnum;
+use pageserver_api::config::PostHogConfig;
 use pem::Pem;
 use postgres_backend::AuthType;
 use reqwest::{Certificate, Url};
@@ -211,7 +212,9 @@ pub struct NeonStorageControllerConf {
 
     pub use_local_compute_notifications: bool,
 
-    pub timeline_safekeeper_count: Option<i64>,
+    pub timeline_safekeeper_count: Option<usize>,
+
+    pub posthog_config: Option<PostHogConfig>,
 
     pub kick_secondary_downloads: Option<bool>,
 }
@@ -245,6 +248,7 @@ impl Default for NeonStorageControllerConf {
             use_https_safekeeper_api: false,
             use_local_compute_notifications: true,
             timeline_safekeeper_count: None,
+            posthog_config: None,
             kick_secondary_downloads: None,
         }
     }

control_plane/src/storage_controller.rs

@@ -638,10 +638,28 @@ impl StorageController {
             args.push("--timelines-onto-safekeepers".to_string());
         }
 
-        if let Some(sk_cnt) = self.config.timeline_safekeeper_count {
+        // neon_local is used in test environments where we often have less than 3 safekeepers.
+        if self.config.timeline_safekeeper_count.is_some() || self.env.safekeepers.len() < 3 {
+            let sk_cnt = self
+                .config
+                .timeline_safekeeper_count
+                .unwrap_or(self.env.safekeepers.len());
+
             args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
         }
 
+        let mut envs = vec![
+            ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
+            ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
+        ];
+
+        if let Some(posthog_config) = &self.config.posthog_config {
+            envs.push((
+                "POSTHOG_CONFIG".to_string(),
+                serde_json::to_string(posthog_config)?,
+            ));
+        }
+
         println!("Starting storage controller");
 
         background_process::start_process(
@@ -649,10 +667,7 @@ impl StorageController {
             &instance_dir,
             &self.env.storage_controller_bin(),
             args,
-            vec![
-                ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
-                ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
-            ],
+            envs,
             background_process::InitialPidFile::Create(self.pid_file(start_args.instance_id)),
             &start_args.start_timeout,
             || async {

control_plane/storcon_cli/src/main.rs

@@ -65,12 +65,27 @@ enum Command {
         #[arg(long)]
         scheduling: Option<NodeSchedulingPolicy>,
     },
-    // Set a node status as deleted.
+    /// Exists for backup usage and will be removed in future.
+    /// Use [`Command::NodeStartDelete`] instead, if possible.
     NodeDelete {
         #[arg(long)]
         node_id: NodeId,
     },
+    /// Start deletion of the specified pageserver.
+    NodeStartDelete {
+        #[arg(long)]
+        node_id: NodeId,
+    },
+    /// Cancel deletion of the specified pageserver and wait for `timeout`
+    /// for the operation to be canceled. May be retried.
+    NodeCancelDelete {
+        #[arg(long)]
+        node_id: NodeId,
+        #[arg(long)]
+        timeout: humantime::Duration,
+    },
     /// Delete a tombstone of node from the storage controller.
+    /// This is used when we want to allow the node to be re-registered.
     NodeDeleteTombstone {
         #[arg(long)]
         node_id: NodeId,
@@ -912,10 +927,43 @@ async fn main() -> anyhow::Result<()> {
                 .await?;
         }
         Command::NodeDelete { node_id } => {
+            eprintln!("Warning: This command is obsolete and will be removed in a future version");
+            eprintln!("Use `NodeStartDelete` instead, if possible");
             storcon_client
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
+        Command::NodeStartDelete { node_id } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::PUT,
+                    format!("control/v1/node/{node_id}/delete"),
+                    None,
+                )
+                .await?;
+            println!("Delete started for {node_id}");
+        }
+        Command::NodeCancelDelete { node_id, timeout } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::DELETE,
+                    format!("control/v1/node/{node_id}/delete"),
+                    None,
+                )
+                .await?;
+
+            println!("Waiting for node {node_id} to quiesce on scheduling policy ...");
+
+            let final_policy =
+                wait_for_scheduling_policy(storcon_client, node_id, *timeout, |sched| {
+                    !matches!(sched, NodeSchedulingPolicy::Deleting)
+                })
+                .await?;
+
+            println!(
+                "Delete was cancelled for node {node_id}. Schedulling policy is now {final_policy:?}"
+            );
+        }
         Command::NodeDeleteTombstone { node_id } => {
             storcon_client
                 .dispatch::<(), ()>(

docker-compose/compute_wrapper/var/db/postgres/configs/config.json

@@ -4,6 +4,7 @@
 
         "timestamp": "2022-10-12T18:00:00.000Z",
         "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8c",
+        "suspend_timeout_seconds": -1,
 
         "cluster": {
             "cluster_id": "docker_compose",

docs/rfcs/035-safekeeper-dynamic-membership-change.md

@@ -20,7 +20,7 @@ In our case consensus leader is compute (walproposer), and we don't want to wake
 up all computes for the change. Neither we want to fully reimplement the leader
 logic second time outside compute. Because of that the proposed algorithm relies
 for issuing configurations on the external fault tolerant (distributed) strongly
-consisent storage with simple API: CAS (compare-and-swap) on the single key.
+consistent storage with simple API: CAS (compare-and-swap) on the single key.
 Properly configured postgres suits this.
 
 In the system consensus is implemented at the timeline level, so algorithm below
@@ -34,7 +34,7 @@ A configuration is
 
 ```
 struct Configuration {
-    generation: Generation, // a number uniquely identifying configuration
+    generation: SafekeeperGeneration, // a number uniquely identifying configuration
     sk_set: Vec<NodeId>, // current safekeeper set
     new_sk_set: Optional<Vec<NodeId>>,
 }
@@ -81,11 +81,11 @@ configuration generation in them is less than its current one. Namely, it
 refuses to vote, to truncate WAL in `handle_elected` and to accept WAL. In
 response it sends its current configuration generation to let walproposer know.
 
-Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/configuration`
-accepting `Configuration`. Safekeeper switches to the given conf it is higher than its
+Safekeeper gets `PUT /v1/tenants/{tenant_id}/timelines/{timeline_id}/membership`
+accepting `Configuration`. Safekeeper switches to the given conf if it is higher than its
 current one and ignores it otherwise. In any case it replies with
 ```
-struct ConfigurationSwitchResponse {
+struct TimelineMembershipSwitchResponse {
     conf: Configuration,
     term: Term,
     last_log_term: Term,
@@ -108,7 +108,7 @@ establishes this configuration as its own and moves to voting.
 It should stop talking to safekeepers not listed in the configuration at this
 point, though it is not unsafe to continue doing so.
 
-To be elected it must receive votes from both majorites if `new_sk_set` is present.
+To be elected it must receive votes from both majorities if `new_sk_set` is present.
 Similarly, to commit WAL it must receive flush acknowledge from both majorities.
 
 If walproposer hears from safekeeper configuration higher than his own (i.e.
@@ -130,7 +130,7 @@ storage are reachable.
 1) Fetch current timeline configuration from the configuration storage.
 2) If it is already joint one and `new_set` is different from `desired_set`
    refuse to change. However, assign join conf to (in memory) var
-   `join_conf` and proceed to step 4 to finish the ongoing change.
+   `joint_conf` and proceed to step 4 to finish the ongoing change.
 3) Else, create joint `joint_conf: Configuration`: increment current conf number
    `n` and put `desired_set` to `new_sk_set`. Persist it in the configuration
    storage by doing CAS on the current generation: change happens only if
@@ -161,11 +161,11 @@ storage are reachable.
    because `pull_timeline` already includes it and plus additionally would be
    broadcast by compute. More importantly, we may proceed to the next step
    only when `<last_log_term, flush_lsn>` on the majority of the new set reached
-   `sync_position`. Similarly, on the happy path no waiting is not needed because
+   `sync_position`. Similarly, on the happy path no waiting is needed because
    `pull_timeline` already includes it. However, we should double
     check to be safe. For example, timeline could have been created earlier e.g.
     manually or after try-to-migrate, abort, try-to-migrate-again sequence.
-7) Create `new_conf: Configuration` incrementing `join_conf` generation and having new
+7) Create `new_conf: Configuration` incrementing `joint_conf` generation and having new
    safekeeper set as `sk_set` and None `new_sk_set`. Write it to configuration
    storage under one more CAS.
 8) Call `PUT` `configuration` on safekeepers from the new set,
@@ -178,12 +178,12 @@ spec of it.
 
 Description above focuses on safety. To make the flow practical and live, here a few more
 considerations.
-1) It makes sense to ping new set to ensure it we are migrating to live node(s) before
+1) It makes sense to ping new set to ensure we are migrating to live node(s) before
   step 3.
 2) If e.g. accidentally wrong new sk set has been specified, before CAS in step `6` is completed
    it is safe to rollback to the old conf with one more CAS.
 3) On step 4 timeline might be already created on members of the new set for various reasons;
-   the simplest is the procedure restart. There are more complicated scenarious like mentioned
+   the simplest is the procedure restart. There are more complicated scenarios like mentioned
    in step 5. Deleting and re-doing `pull_timeline` is generally unsafe without involving
    generations, so seems simpler to treat existing timeline as success. However, this also
    has a disadvantage: you might imagine an surpassingly unlikely schedule where condition in
@@ -192,7 +192,7 @@ considerations.
 4) In the end timeline should be locally deleted on the safekeeper(s) which are
    in the old set but not in the new one, unless they are unreachable. To be
    safe this also should be done under generation number (deletion proceeds only if
-   current configuration is <= than one in request and safekeeper is not memeber of it).
+   current configuration is <= than one in request and safekeeper is not member of it).
 5) If current conf fetched on step 1 is already not joint and members equal to `desired_set`,
    jump to step 7, using it as `new_conf`.
 
@@ -261,14 +261,14 @@ Timeline (branch) creation in cplane should call storage_controller POST
 Response should be augmented with `safekeepers_generation` and `safekeepers`
 fields like described in `/notify-safekeepers` above. Initially (currently)
 these fields may be absent; in this case cplane chooses safekeepers on its own
-like it currently does. The call should be retried until succeeds.
+like it currently does. The call should be retried until it succeeds.
 
 Timeline deletion and tenant deletion in cplane should call appropriate
 storage_controller endpoints like it currently does for sharded tenants. The
 calls should be retried until they succeed.
 
-When compute receives safekeepers list from control plane it needs to know the
-generation to checked whether it should be updated (note that compute may get
+When compute receives safekeeper list from control plane it needs to know the
+generation to check whether it should be updated (note that compute may get
 safekeeper list from either cplane or safekeepers). Currently `neon.safekeepers`
 GUC is just a comma separates list of `host:port`. Let's prefix it with
 `g#<generation>:` to this end, so it will look like
@@ -305,8 +305,8 @@ enum MigrationRequest {
 ```
 
 `FinishPending` requests to run the procedure to ensure state is clean: current
-configuration is not joint and majority of safekeepers are aware of it, but do
-not attempt to migrate anywhere. If current configuration fetched on step 1 is
+configuration is not joint and the majority of safekeepers are aware of it, but do
+not attempt to migrate anywhere. If the current configuration fetched on step 1 is
 not joint it jumps to step 7. It should be run at startup for all timelines (but
 similarly, in the first version it is ok to trigger it manually).
 
@@ -315,7 +315,7 @@ similarly, in the first version it is ok to trigger it manually).
 `safekeepers` table mirroring current `nodes` should be added, except that for
 `scheduling_policy`: it is enough to have at least in the beginning only 3
 fields: 1) `active` 2) `paused` (initially means only not assign new tlis there
-3) `decomissioned` (node is removed).
+3) `decommissioned` (node is removed).
 
 `timelines` table:
 ```
@@ -326,9 +326,10 @@ table! {
         tenant_id -> Varchar,
         start_lsn -> pg_lsn,
         generation -> Int4,
-        sk_set -> Array<Int4>, // list of safekeeper ids
+        sk_set -> Array<Int8>, // list of safekeeper ids
         new_sk_set -> Nullable<Array<Int8>>, // list of safekeeper ids, null if not joint conf
         cplane_notified_generation -> Int4,
+        sk_set_notified_generation -> Int4, // the generation a quorum of sk_set knows about
         deleted_at -> Nullable<Timestamptz>,
     }
 }
@@ -338,13 +339,23 @@ table! {
 might also want to add ancestor_timeline_id to preserve the hierarchy, but for
 this RFC it is not needed.
 
+`cplane_notified_generation` and `sk_set_notified_generation` fields are used to
+track the last stage of the algorithm, when we need to notify safekeeper set and cplane
+with the final configuration after it's already committed to DB.
+
+The timeline is up-to-date (no migration in progress) if `new_sk_set` is null and
+`*_notified_generation` fields are up to date with `generation`. 
+
+It's possible to replace `*_notified_generation` with one boolean field `migration_completed`,
+but for better observability it's nice to have them separately.
+
 #### API
 
 Node management is similar to pageserver:
-1) POST `/control/v1/safekeepers` inserts safekeeper.
-2) GET `/control/v1/safekeepers` lists safekeepers.
-3) GET `/control/v1/safekeepers/:node_id` gets safekeeper.
-4) PUT `/control/v1/safekepers/:node_id/status` changes status to e.g.
+1) POST `/control/v1/safekeeper` inserts safekeeper.
+2) GET `/control/v1/safekeeper` lists safekeepers.
+3) GET `/control/v1/safekeeper/:node_id` gets safekeeper.
+4) PUT `/control/v1/safekeper/:node_id/scheduling_policy` changes status to e.g.
    `offline` or `decomissioned`. Initially it is simpler not to schedule any
     migrations here.
 
@@ -368,8 +379,8 @@ Migration API: the first version is the simplest and the most imperative:
 all timelines from one safekeeper to another. It accepts json
 ```
 {
-    "src_sk": u32,
-    "dst_sk": u32,
+    "src_sk": NodeId,
+    "dst_sk": NodeId,
     "limit": Optional<u32>,
 }
 ```
@@ -379,12 +390,15 @@ Returns list of scheduled requests.
 2) PUT `/control/v1/tenant/:tenant_id/timeline/:timeline_id/safekeeper_migrate` schedules `MigrationRequest`
    to move single timeline to given set of safekeepers:
 ```
-{
-    "desired_set": Vec<u32>,
+struct TimelineSafekeeperMigrateRequest {
+    "new_sk_set": Vec<NodeId>,
 }
 ```
 
-Returns scheduled request.
+In the first version the handler migrates the timeline to `new_sk_set` synchronously.
+Should be retried until success.
+
+In the future we might change it to asynchronous API and return scheduled request.
 
 Similar call should be added for the tenant.
 
@@ -434,6 +448,9 @@ table! {
 }
 ```
 
+We load all pending ops from the table on startup into the memory.
+The table is needed only to preserve the state between restarts.
+
 `op_type` can be `include` (seed from peers and ensure generation is up to
 date), `exclude` (remove locally) and `delete`. Field is actually not strictly
 needed as it can be computed from current configuration, but gives more explicit
@@ -474,7 +491,7 @@ actions must be idempotent. Now, a tricky point here is timeline start LSN. For
 the initial (tenant creation) call cplane doesn't know it. However, setting
 start_lsn on safekeepers during creation is a good thing -- it provides a
 guarantee that walproposer can always find a common point in WAL histories of
-safekeeper and its own, and so absense of it would be a clear sign of
+safekeeper and its own, and so absence of it would be a clear sign of
 corruption. The following sequence works:
 1) Create timeline (or observe that it exists) on pageserver,
    figuring out last_record_lsn in response.
@@ -497,11 +514,9 @@ corruption. The following sequence works:
    retries the call until 200 response.
 
    There is a small question how request handler (timeline creation in this
-   case) would interact with per sk reconciler. As always I prefer to do the
-   simplest possible thing and here it seems to be just waking it up so it
-   re-reads the db for work to do. Passing work in memory is faster, but
-   that shouldn't matter, and path to scan db for work will exist anyway, 
-   simpler to reuse it.
+   case) would interact with per sk reconciler. In the current implementation
+   we first persist the request in the DB, and then send an in-memory request
+   to each safekeeper reconciler to process it.
 
 For pg version / wal segment size: while we may persist them in `timelines`
 table, it is not necessary as initial creation at step 3 can take them from
@@ -509,30 +524,40 @@ pageserver or cplane creation call and later pull_timeline will carry them
 around.
 
 Timeline migration.
-1) CAS to the db to create joint conf, and in the same transaction create
-   `safekeeper_timeline_pending_ops` `include` entries to initialize new members
-   as well as deliver this conf to current ones; poke per sk reconcilers to work
-   on it. Also any conf change should also poke cplane notifier task(s).
-2) Once it becomes possible per alg description above, get out of joint conf
-   with another CAS. Task should get wakeups from per sk reconcilers because 
-   conf switch is required for advancement; however retries should be sleep
-   based as well as LSN advancement might be needed, though in happy path 
-   it isn't. To see whether further transition is possible on wakup migration
-   executor polls safekeepers per the algorithm. CAS creating new conf with only
-   new members should again insert entries to `safekeeper_timeline_pending_ops`
-   to switch them there, as well as `exclude` rows to remove timeline from 
-   old members.
+1) CAS to the db to create joint conf. Since this moment the migration is considered to be 
+   "in progress". We can detect all "in-progress" migrations looking into the database.
+2) Do steps 4-6 from the algorithm, including `pull_timeline` onto `new_sk_set`, update membership
+   configuration on all safekeepers, notify cplane, etc. All operations are idempotent,
+   so we don't need to persist anything in the database at this stage. If any errors occur,
+   it's safe to retry or abort the migration.
+3) Once it becomes possible per alg description above, get out of joint conf
+   with another CAS. Also should insert `exclude` entries into `safekeeper_timeline_pending_ops`
+   in the same DB transaction. Adding `exclude` entries atomically is nesessary because after
+   CAS we don't have the list of excluded safekeepers in the `timelines` table anymore, but we
+   need to have them persisted somewhere in case the migration is interrupted right after the CAS.
+4) Finish the migration. The final membership configuration is committed to the DB at this stage.
+   So, the migration can not be aborted anymore. But it can still be retried if the migration fails
+   past stage 3. To finish the migration we need to send the new membership configuration to
+   a new quorum of safekeepers, notify cplane with the new safekeeper list and schedule the `exclude`
+   requests to in-memory queue for safekeeper reconciler. If the algrorithm is retried, it's
+   possible that we have already committed `exclude` requests to DB, but didn't send them to
+   the in-memory queue. In this case we need to read them from `safekeeper_timeline_pending_ops`
+   because it's the only place where they are persistent. The fields `sk_set_notified_generation`
+   and `cplane_notified_generation` are updated after each step. The migration is considered
+   fully completed when they match the `generation` field.
+
+In practice, we can report "success" after stage 3 and do the "finish" step in per-timeline
+reconciler (if we implement it). But it's wise to at least try to finish them synchronously,
+so the timeline is always in a "good state" and doesn't require an old quorum to commit
+WAL after the migration reported "success".
 
 Timeline deletion: just set `deleted_at` on the timeline row and insert
 `safekeeper_timeline_pending_ops` entries in the same xact, the rest is done by
 per sk reconcilers.
 
-When node is removed (set to `decomissioned`), `safekeeper_timeline_pending_ops`
+When node is removed (set to `decommissioned`), `safekeeper_timeline_pending_ops`
 for it must be cleared in the same transaction.
 
-One more task pool should infinitely retry notifying control plane about changed
-safekeeper sets (trying making `cplane_notified_generation` equal `generation`).
-
 #### Dealing with multiple instances of storage_controller
 
 Operations described above executed concurrently might create some errors but do
@@ -541,7 +566,7 @@ of storage_controller it is fine to have it temporarily, e.g. during redeploy.
 
 To harden against some controller instance creating some work in
 `safekeeper_timeline_pending_ops` and then disappearing without anyone pickup up
-the job per sk reconcilers apart from explicit wakups should scan for work
+the job per sk reconcilers apart from explicit wakeups should scan for work
 periodically. It is possible to remove that though if all db updates are
 protected with leadership token/term -- then such scans are needed only after
 leadership is acquired.
@@ -563,7 +588,7 @@ There should be following layers of tests:
    safekeeper communication and pull_timeline need to be mocked and main switch
    procedure wrapped to as a node (thread) in simulation tests, using these
    mocks. Test would inject migrations like it currently injects
-   safekeeper/walproposer restars. Main assert is the same -- committed WAL must
+   safekeeper/walproposer restarts. Main assert is the same -- committed WAL must
    not be lost.
 
 3) Since simulation testing injects at relatively high level points (not
@@ -613,7 +638,7 @@ Let's have the following implementation bits for gradual rollout:
   `notify-safekeepers`.
 
 Then the rollout for a region would be:
-- Current situation: safekeepers are choosen by control_plane.
+- Current situation: safekeepers are chosen by control_plane.
 - We manually migrate some timelines, test moving them around.
 - Then we enable `--set-safekeepers` so that all new timelines
   are on storage controller.

docs/rfcs/040-Endpoint-Persistent-Unlogged-Files-Storage.md

@@ -0,0 +1,396 @@
+# Memo: Endpoint Persistent Unlogged Files Storage
+Created on 2024-11-05
+Implemented on N/A
+
+## Summary
+A design for a storage system that allows storage of files required to make
+Neon's Endpoints have a better experience at or after a reboot.
+
+## Motivation
+Several systems inside PostgreSQL (and Neon) need some persistent storage for
+optimal workings across reboots and restarts, but still work without.
+Examples are the query-level statistics files of `pg_stat_statements` in
+`pg_stat/pg_stat_statements.stat`, and `pg_prewarm`'s `autoprewarm.blocks`.
+We need a storage system that can store and manage these files for each
+Endpoint, without necessarily granting users access to an unlimited storage
+device.
+
+## Goals
+- Store known files for Endpoints with reasonable persistence.  
+  _Data loss in this service, while annoying and bad for UX, won't lose any
+  customer's data._
+
+## Non Goals (if relevant)
+- This storage system does not need branching, file versioning, or other such
+  features. The files are as ephemeral to the timeline of the data as the
+  Endpoints that host the data.
+- This storage system does not need to store _all_ user files, only 'known'
+  user files.
+- This storage system does not need to be hosted fully inside Computes.  
+  _Instead, this will be a separate component similar to Pageserver,
+  SafeKeeper, the S3 proxy used for dynamically loaded extensions, etc._
+
+## Impacted components
+- Compute needs new code to load and store these files in its lifetime.
+- Control Plane needs to consider this new storage system when signalling
+  the deletion of an Endpoint, Timeline, or Tenant.
+- Control Plane needs to consider this new storage system when it resets
+  or re-assigns an endpoint's timeline/branch state.
+
+A new service is created: the Endpoint Persistent Unlogged Files Storage
+service.  This could be integrated in e.g. Pageserver or Control Plane, or a
+separately hosted service.
+
+## Proposed implementation
+Endpoint-related data files are managed by a newly designed service (which
+optionally is integrated in an existing service like Pageserver or Control
+Plane), which stores data directly into S3 or any blob storage of choice.
+
+Upon deletion of the Endpoint, or reassignment of the endpoint to a different
+branch, this ephemeral data is dropped: the data stored may not match the
+state of the branch's data after reassignment, and on endpoint deletion the
+data won't have any use to the user.
+
+Compute gets credentials (JWT token with Tenant, Timeline & Endpoint claims)
+which it can use to authenticate to this new service and retrieve and store
+data associated with this endpoint.  This limited scope reduces leaks of data
+across endpoints and timeline resets, and limits the ability of endpoints to
+mess with other endpoints' data.
+
+The path of this endpoint data in S3 is initially as follows:
+
+    s3://<regional-epufs-bucket>/
+      tenants/
+        <hex-tenant-id>/
+          tenants/
+            <hex-timeline-id>/
+              endpoints/
+                <endpoint-id>/
+                  pgdata/
+                    <file_path_in_pgdatadir>
+
+For other blob storages an equivalent or similar path can be constructed.
+
+### Reliability, failure modes and corner cases (if relevant)
+Reliability is important, but not critical to the workings of Neon.  The data
+stored in this service will, when lost, reduce performance, but won't be a
+cause of permanent data loss - only operational metadata is stored.
+
+Most, if not all, blob storage services have sufficiently high persistence
+guarantees to cater our need for persistence and uptime. The only concern with
+blob storages is that the access latency is generally higher than local disk,
+but for the object types stored (cache state, ...) I don't think this will be
+much of an issue.
+
+### Interaction/Sequence diagram (if relevant)
+
+In these diagrams you can replace S3 with any persistent storage device of
+choice, but S3 is chosen as representative name: The well-known and short name
+of AWS' blob storage. Azure Blob Storage should work too, but it has a much
+longer name making it less practical for the diagrams.
+
+Write data:
+
+```http
+POST /tenants/<tenant-id>/timelines/<tl-id>/endpoints/<endpoint-id>/pgdata/<the-pgdata-path>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "version": "<opaque>", # opaque file version token, changes when the file contents change
+  "size": <bytes>,
+}
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    co-->ep: Connect with credentials
+    co->>+ep: Store Unlogged Persistent File
+    opt is authenticated
+        ep->>s3: Write UPF to S3
+    end
+    ep->>-co: OK / Failure / Auth Failure
+    co-->ep: Cancel connection
+```
+
+Read data: (optional with cache-relevant request parameters, e.g. If-Modified-Since)
+```http
+GET /tenants/<tenant-id>/timelines/<tl-id>/endpoints/<endpoint-id>/pgdata/<the-pgdata-path>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+
+<file data>
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    co->>+ep: Read Unlogged Persistent File
+    opt is authenticated
+        ep->>+s3: Request UPF from storage
+        s3->>-ep: Receive UPF from storage
+    end
+    ep->>-co: OK(response) / Failure(storage, auth, ...)
+```
+
+Compute Startup:
+```mermaid
+sequenceDiagram
+    autonumber
+    participant co as Compute
+    participant ps as Pageserver
+    participant ep as EPUFS
+    participant es as Extension server
+
+    note over co: Bind endpoint ep-xxx
+    par Get basebackup
+        co->>+ps: Request basebackup @ LSN
+        ps-)ps: Construct basebackup
+        ps->>-co: Receive basebackup TAR @ LSN
+    and Get startup-critical Unlogged Persistent Files
+        co->>+ep: Get all UPFs of endpoint ep-xxx
+        ep-)ep: Retrieve and gather all UPFs
+        ep->>-co: TAR of UPFs
+    and Get startup-critical extensions
+        loop For every startup-critical extension
+            co->>es: Get critical extension
+            es->>co: Receive critical extension
+        end
+    end
+    note over co: Start compute
+```
+
+CPlane ops:
+```http
+DELETE /tenants/<tenant-id>/timelines/<timeline-id>/endpoints/<endpoint-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "timeline": "<timeline-id>",
+  "endpoint": "<endpoint-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```http
+DELETE /tenants/<tenant-id>/timelines/<timeline-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "timeline": "<timeline-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```http
+DELETE /tenants/<tenant-id>
+Host: epufs.svc.neon.local
+
+<<<
+
+200 OK
+{
+  "tenant": "<tenant-id>",
+  "deleted": {
+    "files": <count>,
+    "bytes": <count>,
+  },
+}
+```
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant cp as Control Plane
+    participant ep as EPUFS
+    participant s3 as Blob Storage
+
+    alt Tenant deleted
+        cp-)ep: Tenant deleted
+        loop For every object associated with removed tenant
+            ep->>s3: Remove data of deleted tenant from Storage
+        end
+        opt
+            ep-)cp: Tenant cleanup complete
+        end
+    alt Timeline deleted
+        cp-)ep: Timeline deleted
+        loop For every object associated with removed timeline
+            ep->>s3: Remove data of deleted timeline from Storage
+        end
+        opt
+            ep-)cp: Timeline cleanup complete
+        end
+    else Endpoint reassigned or removed
+        cp->>+ep: Endpoint reassigned
+        loop For every object associated with reassigned/removed endpoint
+            ep->>s3: Remove data from Storage
+        end
+        ep->>-cp: Cleanup complete
+    end
+```
+
+### Scalability (if relevant)
+
+Provisionally:  As this service is going to be part of compute startup, this
+service should be able to quickly respond to all requests.  Therefore this
+service is deployed to every AZ we host Computes in, and Computes communicate
+(generally) only to the EPUFS endpoint of the AZ they're hosted in.
+
+Local caching of frequently restarted endpoints' data or metadata may be
+needed for best performance.  However, due to the regional nature of stored
+data but zonal nature of the service deployment, we should be careful when we
+implement any local caching, as it is possible that computes in AZ 1 will
+update data originally written and thus cached by AZ 2.  Cache version tests
+and invalidation is therefore required if we want to roll out caching to this
+service, which is too broad a scope for an MVC.  This is why caching is left
+out of scope for this RFC, and should be considered separately after this RFC
+is implemented.
+
+### Security implications (if relevant)
+This service must be able to authenticate users at least by Tenant ID,
+Timeline ID and Endpoint ID. This will use the existing JWT infrastructure of
+Compute, which will be upgraded to the extent needed to support Timeline- and
+Endpoint-based claims.
+
+The service requires unlimited access to (a prefix of) a blob storage bucket,
+and thus must be hosted outside the Compute VM sandbox.
+
+A service that generates pre-signed request URLs for Compute to download the
+data from that URL is likely problematic, too:  Compute would be able to write
+unlimited data to the bucket, or exfiltrate this signed URL to get read/write
+access to specific objects in this bucket, which would still effectively give
+users access to the S3 bucket (but with improved access logging).
+
+There may be a use case for transferring data associated with one endpoint to
+another endpoint (e.g. to make one endpoint warm its caches with the state of
+another endpoint), but that's not currently in scope, and specific needs may
+be solved through out-of-line communication of data or pre-signed URLs.
+
+### Unresolved questions (if relevant)
+Caching of files is not in the implementation scope of the document, but
+should at some future point be considered to maximize performance.
+
+## Alternative implementation (if relevant)
+Several ideas have come up to solve this issue:
+
+### Use AUXfile
+One prevalent idea was to WAL-log the files using our AUXfile mechanism.
+
+Benefits:
+
++ We already have this storage mechanism
+
+Demerits:
+
+- It isn't available on read replicas
+- Additional WAL will be consumed during shutdown and after the shutdown
+  checkpoint, which needs PG modifications to work without panics.
+- It increases the data we need to manage in our versioned storage, thus
+  causing higher storage costs with higher retention due to duplication at
+  the storage layer.
+
+### Sign URLs for read/write operations, instead of proxying them
+
+Benefits:
+
++ The service can be implemented with a much reduced IO budget
+
+Demerits:
+
+- Users could get access to these signed credentials
+- Not all blob storage services may implement URL signing
+
+### Give endpoints each their own directly accessed block volume
+
+Benefits:
+
++ Easier to integrate for PostgreSQL
+
+Demerits:
+
+- Little control on data size and contents
+- Potentially problematic as we'd need to store data all across the pgdata
+  directory.
+- EBS is not a good candidate
+   - Attaches in 10s of seconds, if not more; i.e. too cold to start
+   - Shared EBS volumes are a no-go, as you'd have to schedule the endpoint
+     with users of the same EBS volumes, which can't work with VM migration
+   - EBS storage costs are very high (>80$/kilotenant when using a
+     volume/tenant)
+   - EBS volumes can't be mounted across AZ boundaries
+- Bucket per endpoint is unfeasible
+   - S3 buckets are priced at $20/month per 1k, which we could better spend
+     on developers.
+   - Allocating service accounts takes time (100s of ms), and service accounts
+     are a limited resource, too; so they're not a good candidate to allocate
+     on a per-endpoint basis.
+   - Giving credentials limited to prefix has similar issues as the pre-signed
+     URL approach.
+   - Bucket DNS lookup will fill DNS caches and put pressure on DNS lookup
+     much more than our current systems would.
+- Volumes bound by hypervisor are unlikely
+   - This requires significant investment and increased software on the
+     hypervisor.
+   - It is unclear if we can attach volumes after boot, i.e. for pooled
+     instances.
+
+### Put the files into a table
+
+Benefits:
+
+ + Mostly already available in PostgreSQL
+
+Demerits:
+
+ - Uses WAL
+   - Can't be used after shutdown checkpoint
+   - Needs a RW endpoint, and table & catalog access to write to this data
+ - Gets hit with DB size limitations
+ - Depending on user acces:
+   - Inaccessible:  
+     The user doesn't have control over database size caused by
+     these systems.
+   - Accessible:  
+     The user can corrupt these files and cause the system to crash while
+     user-corrupted files are present, thus increasing on-call overhead.
+
+## Definition of Done (if relevant)
+
+This project is done if we have:
+
+- One S3 bucket equivalent per region, which stores this per-endpoint data.
+- A new service endpoint in at least every AZ, which indirectly grants
+  endpoints access to the data stored for these endpoints in these buckets.
+- Compute writes & reads temp-data at shutdown and startup, respectively, for
+  at least the pg_prewarm or lfc_prewarm state files.
+- Cleanup of endpoint data is triggered when the endpoint is deleted or is
+  detached from its current timeline.

docs/rfcs/044-feature-flag.md

@@ -0,0 +1,179 @@
+# Storage Feature Flags
+
+In this RFC, we will describe how we will implement per-tenant feature flags.
+
+## PostHog as Feature Flag Service
+
+Before we start, let's talk about how current feature flag services work. PostHog is the feature flag service we are currently using across multiple user-facing components in the company. PostHog has two modes of operation: HTTP evaluation and server-side local evaluation.
+
+Let's assume we have a storage feature flag called gc-compaction and we want to roll it out to scale-tier users with resident size >= 10GB and <= 100GB.
+
+### Define User Profiles
+
+The first step is to synchronize our user profiles to the PostHog service. We can simply assume that each tenant is a user in PostHog. Each user profile has some properties associated with it. In our case, it will be: plan type (free, scale, enterprise, etc); resident size (in bytes); primary pageserver (string); region (string).
+
+### Define Feature Flags
+
+We would create a feature flag called gc-compaction in PostHog with 4 variants: disabled, stage-1, stage-2, fully-enabled. We will flip the feature flags from disabled to fully-enabled stage by stage for some percentage of our users.
+
+### Option 1: HTTP Evaluation Mode
+
+When using PostHog's HTTP evaluation mode, the client will make request to the PostHog service, asking for the value of a feature flag for a specific user.
+
+* Control plane will report the plan type to PostHog each time it attaches a tenant to the storcon or when the user upgrades/downgrades. It calls the PostHog profile API to associate tenant ID with the plan type. Assume we have X active tenants and such attach or plan change event happens each week, that would be 4X profile update requests per month.
+* Pageservers will report the resident size and the primary pageserver to the PostHog service. Assume we report resident size every 24 hours, that would be 30X requests per month.
+* Each tenant will request the state of the feature flag every 1 hour, that's 720X requests per month.
+* The Rust client would be easy to implement as we only need to call the `/decide` API on PostHog.
+
+Using the HTTP evaluation mode we will issue 754X requests a month.
+
+### Option 2: Local Evaluation Mode
+
+When using PostHog's HTTP evaluation mode, the client (usually the server in a browser/server architecture) will poll the feature flag configuration every 30s (default in the Python client) from PostHog. Such configuration contains data like:
+
+<details>
+
+<summary>Example JSON response from the PostHog local evaluation API</summary>
+
+```
+[
+    {
+        "id": 1,
+        "name": "Beta Feature",
+        "key": "person-flag",
+        "is_simple_flag": True,
+        "active": True,
+        "filters": {
+            "groups": [
+                {
+                    "properties": [
+                        {
+                            "key": "location",
+                            "operator": "exact",
+                            "value": ["Straße"],
+                            "type": "person",
+                        }
+                    ],
+                    "rollout_percentage": 100,
+                },
+                {
+                    "properties": [
+                        {
+                            "key": "star",
+                            "operator": "exact",
+                            "value": ["ſun"],
+                            "type": "person",
+                        }
+                    ],
+                    "rollout_percentage": 100,
+                },
+            ],
+        },
+    }
+]
+```
+
+</details>
+
+Note that the API only contains information like "under what condition => rollout percentage". The user is responsible to provide the properties required to the client for local evaluation, and the PostHog service (web UI) cannot know if a feature is enabled for the tenant or not until the client uses the `capture` API to report the result back. To control the rollout percentage, the user ID gets mapped to a float number in `[0, 1)` on a consistent hash ring. All values <= the percentage will get the feature enabled or set to the desired value.
+
+To use the local evaluation mode, the system needs:
+
+* Assume each pageserver will poll PostHog for the local evaluation JSON every 5 minutes (instead of the 30s default as it's too frequent). That's 8640Y per month, Y is the number of pageservers. Local evaluation requests cost 10x more than the normal decide request, so that's 86400Y request units to bill.
+* Storcon needs to store the plan type in the database and pass that information to the pageserver when attaching the tenant.
+* Storcon also needs to update PostHog with the active tenants, for example, when the tenant gets detached/attached. Assume each active tenant gets detached/attached every week, that would be 4X requests per month.
+* We do not need to update bill type or resident size to PostHog as all these are evaluated locally.
+* After each local evaluation of the feature flag, we need to call PostHog's capture event API to update the result of the evaluation that the feature is enabled. We can do this when the flag gets changed compared with the last cached state in memory. That would be at least 4X (assume we do deployment every week so the cache gets cleared) and maybe an additional multiplifier of 10 assume we have 10 active features.
+
+In this case, we will issue 86400Y + 40X requests per month.
+
+Assume X = 1,000,000 and Y = 100,
+
+|   | HTTP Evaluation  | Local Evaluation  |
+|---|---|---|
+| Latency of propagating the conditions/properties for feature flag  | 24 hours  | available locally  |
+| Latency of applying the feature flag  | 1 hour  | 5 minutes  |
+| Can properties be reported from different services |  Yes |  No  |
+| Do we need to sync billing info etc to pageserver |  No |  Yes  |
+| Cost | 75400$ / month | 4864$ / month |
+
+# Our Solution
+
+We will use PostHog _only_ as an UI to configure the feature flags. Whether a feature is enabled or not can only be queried through storcon/pageserver instead of using the PostHog UI. (We could report it back to PostHog via `capture_event` but it costs $$$.) This allows us to ramp up the feature flag functionality fast at first. At the same time, it would also give us the option to migrate to our own solution once we want to have more properties and more complex evaluation rules in our system.
+
+* We will create several fake users (tenants) in PostHog that contains all the properties we will use for evaluating a feature flag (i.e., resident size, billing type, pageserver id, etc.)
+* We will use PostHog's local evaluation API to poll the configuration of the feature flags and evaluate them locally on each of the pageserver.
+* The evaluation result will not be reported back to PostHog.
+* Storcon needs to pull some information from cplane database.
+* To know if a feature is currently enabled or not, we need to call the storcon/pageserver API; and we won't be able to know if a feature has been enabled on a tenant before easily: we need to look at the Grafana logs.
+
+We only need to pay for the 86400Y local evaluation requests (that would be setting Y=0 in solution 2 => $864/month, and even less if we proxy it through storcon).
+
+## Implementation
+
+* Pageserver: implement a PostHog local evaluation client. The client will be shared across all tenants on the pageserver with a single API: `evaluate(tenant_id, feature_flag, properties) -> json`.
+* Storcon: if we need plan type as the evaluation condition, pull it from cplane database.
+* Storcon/Pageserver: implement an HTTP API `:tenant_id/feature/:feature` to retrieve the current feature flag status.
+* Storcon/Pageserver: a loop to update the feature flag spec on both storcon and pageserver. Pageserver loop will only be activated if storcon does not push the specs to the pageserver.
+
+## Difference from Tenant Config
+
+* Feature flags can be modified by percentage, and the default config for each feature flag can be modified in UI without going through the release process.
+* Feature flags are more flexible and won't be persisted anywhere and will be passed as plain JSON over the wire so that do not need to handle backward/forward compatibility as in tenant config.
+* The expectation of tenant config is that once we add a flag we cannot remove it (or it will be hard to remove), but feature flags are more flexible.
+
+# Final Implementation
+
+* We added a new crate `posthog_lite_client` that supports local feature evaluations.
+* We set up two projects "Storage (staging)" and "Storage (production)" in the PostHog console.
+* Each pageserver reports 10 fake tenants to PostHog so that we can get all combinations of regions (and other properties) in the PostHog UI.
+* Supported properties: AZ, neon_region, pageserver, tenant_id.
+* You may use "Pageserver Feature Flags" dashboard to see the evaluation status.
+* The feature flag spec is polled on storcon every 30s (in each of the region) and storcon will propagate the spec to the pageservers.
+* The pageserver housekeeping loop updates the tenant-specific properties (e.g., remote size) for evaluation.
+
+Each tenant has a `feature_resolver` object. After you add a feature flag in the PostHog console, you can retrieve it with:
+
+```rust
+// Boolean flag
+self
+    .feature_resolver
+    .evaluate_boolean("flag")
+    .is_ok()
+// Multivariate flag
+self
+    .feature_resolver
+    .evaluate_multivariate("gc-comapction-strategy")
+    .ok();
+```
+
+The user needs to handle the case where the evaluation result is an error. This can occur in a variety of cases:
+
+* During the pageserver start, the feature flag spec has not been retrieved.
+* No condition group is matched.
+* The feature flag spec contains an operand/operation not supported by the lite PostHog library.
+
+For boolean flags, the return value is `Result<(), Error>`. `Ok(())` means the flag is evaluated to true. Otherwise,
+there is either an error in evaluation or it does not match any groups.
+
+For multivariate flags, the return value is `Result<String, Error>`. `Ok(variant)` indicates the flag is evaluated
+to a variant. Otherwise, there is either an error in evaluation or it does not match any groups.
+
+The evaluation logic is documented in the PostHog lite library. It compares the consistent hash of a flag key + tenant_id
+with the rollout percentage and determines which tenant to roll out a specific feature.
+
+Users can use the feature flag evaluation API to get the flag evaluation result of a specific tenant for debugging purposes.
+
+```
+curl http://localhost:9898/v1/tenant/:tenant_id/feature_flag?flag=:key&as=multivariate/boolean"
+```
+
+By default, the storcon pushes the feature flag specs to the pageservers every 30 seconds, which means that a change in feature flag in the
+PostHog UI will propagate to the pageservers within 30 seconds.
+
+# Future Works
+
+* Support dynamic tenant properties like logical size as the evaluation condition.
+* Support properties like `plan_type` (needs cplane to pass it down).
+* Report feature flag evaluation result back to PostHog (if the cost is okay).
+* Fast feature flag evaluation cache on critical paths (e.g., cache a feature flag result in `AtomicBool` and use it on the read path).

docs/rfcs/2025-03-17-compute-prewarm.md

@@ -0,0 +1,399 @@
+# Compute rolling restart with prewarm
+
+Created on 2025-03-17
+Implemented on _TBD_
+Author: Alexey Kondratov (@ololobus)
+
+## Summary
+
+This RFC describes an approach to reduce performance degradation due to missing caches after compute node restart, i.e.:
+
+1. Rolling restart of the running instance via 'warm' replica.
+2. Auto-prewarm compute caches after unplanned restart or scale-to-zero.
+
+## Motivation
+
+Neon currently implements several features that guarantee high uptime of compute nodes:
+
+1. Storage high-availability (HA), i.e. each tenant shard has a secondary pageserver location, so we can quickly switch over compute to it in case of primary pageserver failure.
+2. Fast compute provisioning, i.e. we have a fleet of pre-created empty computes, that are ready to serve workload, so restarting unresponsive compute is very fast.
+3. Preemptive NeonVM compute provisioning in case of k8s node unavailability.
+
+This helps us to be well-within the uptime SLO of 99.95% most of the time. Problems begin when we go up to multi-TB workloads and 32-64 CU computes.
+During restart, compute loses all caches: LFC, shared buffers, file system cache. Depending on the workload, it can take a lot of time to warm up the caches,
+so that performance could be degraded and might be even unacceptable for certain workloads. The latter means that although current approach works well for small to
+medium workloads, we still have to do some additional work to avoid performance degradation after restart of large instances.
+
+## Non Goals
+
+- Details of the persistence storage for prewarm data are out of scope, there is a separate RFC for that: <https://github.com/neondatabase/neon/pull/9661>.
+- Complete compute/Postgres HA setup and flow. Although it was originally in scope of this RFC, during preliminary research it appeared to be a rabbit hole, so it's worth of a separate RFC.
+- Low-level implementation details for Postgres replica-to-primary promotion. There are a lot of things to think and care about: how to start walproposer, [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html), and so on, but it's worth of at least a separate one-pager design document if not RFC.
+
+## Impacted components
+
+Postgres, compute_ctl, Control plane, Endpoint storage for unlogged storage of compute files.
+For the latter, we will need to implement a uniform abstraction layer on top of S3, ABS, etc., but
+S3 is used in text interchangeably with 'endpoint storage' for simplicity.
+
+## Proposed implementation
+
+### compute_ctl spec changes and auto-prewarm
+
+We are going to extend the current compute spec with the following attributes
+
+```rust
+struct ComputeSpec {
+    /// [All existing attributes]
+    ...
+    /// Whether to do auto-prewarm at start or not.
+    /// Default to `false`.
+    pub lfc_auto_prewarm: bool
+    /// Interval in seconds between automatic dumps of
+    /// LFC state into S3. Default `None`, which means 'off'.
+    pub lfc_dump_interval_sec: Option<i32>
+}
+```
+
+When `lfc_dump_interval_sec` is set to `N`, `compute_ctl` will periodically dump the LFC state
+and store it in S3, so that it could be used either for auto-prewarm after restart or by replica
+during the rolling restart. For enabling periodic dumping, we should consider the following value
+`lfc_dump_interval_sec=300` (5 minutes), same as in the upstream's `pg_prewarm.autoprewarm_interval`.
+
+When `lfc_auto_prewarm` is set to `true`, `compute_ctl` will start prewarming the LFC upon restart
+iif some of the previous states is present in S3.
+
+### compute_ctl API
+
+1. `POST /store_lfc_state` -- dump LFC state using Postgres SQL interface and store result in S3.
+    This has to be a blocking call, i.e. it will return only after the state is stored in S3.
+    If there is any concurrent request in progress, we should return `429 Too Many Requests`,
+    and let the caller to retry.
+
+2. `GET /dump_lfc_state` -- dump LFC state using Postgres SQL interface and return it as is
+    in text format suitable for the future restore/prewarm. This API is not strictly needed at
+    the end state, but could be useful for a faster prototyping of a complete rolling restart flow
+    with prewarm, as it doesn't require persistent for LFC state storage.
+
+3. `POST /restore_lfc_state` -- restore/prewarm LFC state with request
+
+    ```yaml
+    RestoreLFCStateRequest:
+      oneOf:
+        - type: object
+          required:
+            - lfc_state
+          properties:
+            lfc_state:
+              type: string
+              description: Raw LFC content dumped with GET `/dump_lfc_state`
+        - type: object
+          required:
+            - lfc_cache_key
+          properties:
+            lfc_cache_key:
+              type: string
+              description: |
+                endpoint_id of the source endpoint on the same branch
+                to use as a 'donor' for LFC content. Compute will look up
+                LFC content dump in S3 using this key and do prewarm.
+    ```
+
+    where `lfc_state` and `lfc_cache_key` are mutually exclusive.
+
+    The actual prewarming will happen asynchronously, so the caller need to check the
+    prewarm status using the compute's standard `GET /status` API.
+
+4. `GET /status` -- extend existing API with following attributes
+
+    ```rust
+    struct ComputeStatusResponse {
+        // [All existing attributes]
+        ...
+        pub prewarm_state: PrewarmState
+    }
+
+    /// Compute prewarm state. Will be stored in the shared Compute state
+    /// in compute_ctl
+    struct PrewarmState {
+        pub status: PrewarmStatus
+        /// Total number of pages to prewarm
+        pub pages_total: i64
+        /// Number of pages prewarmed so far
+        pub pages_processed: i64
+        /// Optional prewarm error
+        pub error: Option<String>
+    }
+
+    pub enum PrewarmStatus {
+        /// Prewarming was never requested on this compute
+        Off,
+        /// Prewarming was requested, but not started yet
+        Pending,
+        /// Prewarming is in progress. The caller should follow
+        /// `PrewarmState::progress`.
+        InProgress,
+        /// Prewarming has been successfully completed
+        Completed,
+        /// Prewarming failed. The caller should look at
+        /// `PrewarmState::error` for the reason.
+        Failed,
+        /// It is intended to be used by auto-prewarm if none of
+        /// the previous LFC states is available in S3.
+        /// This is a distinct state from the `Failed` because
+        /// technically it's not a failure and could happen if
+        /// compute was restart before it dumped anything into S3,
+        /// or just after the initial rollout of the feature.
+        Skipped,
+    }
+    ```
+
+5. `POST /promote` -- this is a **blocking** API call to promote compute replica into primary.
+    This API should be very similar to the existing `POST /configure` API, i.e. accept the
+    spec (primary spec, because originally compute was started as replica). It's a distinct
+    API method because semantics and response codes are different:
+
+    - If promotion is done successfully, it will return `200 OK`.
+    - If compute is already primary, the call will be no-op and `compute_ctl`
+      will return `412 Precondition Failed`.
+    - If, for some reason, second request reaches compute that is in progress of promotion,
+      it will respond with `429 Too Many Requests`.
+    - If compute hit any permanent failure during promotion `500 Internal Server Error`
+      will be returned.
+
+### Control plane operations
+
+The complete flow will be present as a sequence diagram in the next section, but here
+we just want to list some important steps that have to be done by control plane during
+the rolling restart via warm replica, but without much of low-level implementation details.
+
+1. Register the 'intent' of the instance restart, but not yet interrupt any workload at
+    primary and also accept new connections. This may require some endpoint state machine
+    changes, e.g. introduction of the `pending_restart` state. Being in this state also
+    **mustn't prevent any other operations except restart**: suspend, live-reconfiguration
+    (e.g. due to notify-attach call from the storage controller), deletion.
+
+2. Start new replica compute on the same timeline and start prewarming it. This process
+    may take quite a while, so the same concurrency considerations as in 1. should be applied
+    here as well.
+
+3. When warm replica is ready, control plane should:
+
+    3.1. Terminate the primary compute. Starting from here, **this is a critical section**,
+        if anything goes off, the only option is to start the primary normally and proceed
+        with auto-prewarm.
+
+    3.2. Send cache invalidation message to all proxies, notifying them that all new connections
+        should request and wait for the new connection details. At this stage, proxy has to also
+        drop any existing connections to the old primary, so they didn't do stale reads.
+
+    3.3. Attach warm replica compute to the primary endpoint inside control plane metadata
+        database.
+
+    3.4. Promote replica to primary.
+
+    3.5. When everything is done, finalize the endpoint state to be just `active`.
+
+### Complete rolling restart flow
+
+```mermaid
+  sequenceDiagram
+
+  autonumber
+
+  participant proxy as Neon proxy
+
+  participant cplane as Control plane
+
+  participant primary as Compute (primary)
+  box Compute (replica)
+    participant ctl as compute_ctl
+    participant pg as Postgres
+  end
+
+  box Endpoint unlogged storage
+    participant s3proxy as Endpoint storage service
+    participant s3 as S3/ABS/etc.
+  end
+
+
+  cplane ->> primary: POST /store_lfc_state
+  primary -->> cplane: 200 OK
+
+  cplane ->> ctl: POST /restore_lfc_state
+  activate ctl
+  ctl -->> cplane: 202 Accepted
+
+  activate cplane
+  cplane ->> ctl: GET /status: poll prewarm status
+  ctl ->> s3proxy: GET /read_file
+  s3proxy ->> s3: read file
+  s3 -->> s3proxy: file content
+  s3proxy -->> ctl: 200 OK: file content
+
+  proxy ->> cplane: GET /proxy_wake_compute
+  cplane -->> proxy: 200 OK: old primary conninfo
+
+  ctl ->> pg: prewarm LFC
+  activate pg
+  pg -->> ctl: prewarm is completed
+  deactivate pg
+
+  ctl -->> cplane: 200 OK: prewarm is completed
+  deactivate ctl
+  deactivate cplane
+
+  cplane -->> cplane: reassign replica compute to endpoint,<br>start terminating the old primary compute
+  activate cplane
+  cplane ->> proxy: invalidate caches
+
+  proxy ->> cplane: GET /proxy_wake_compute
+
+  cplane -x primary: POST /terminate
+  primary -->> cplane: 200 OK
+  note over primary: old primary<br>compute terminated
+
+  cplane ->> ctl: POST /promote
+  activate ctl
+  ctl ->> pg: pg_ctl promote
+  activate pg
+  pg -->> ctl: done
+  deactivate pg
+  ctl -->> cplane: 200 OK
+  deactivate ctl
+
+  cplane -->> cplane: finalize operation
+  cplane -->> proxy: 200 OK: new primary conninfo
+  deactivate cplane
+```
+
+### Network bandwidth and prewarm speed
+
+It's currently known that pageserver can sustain about 3000 RPS per shard for a few running computes.
+Large tenants are usually split into 8 shards, so the final formula may look like this:
+
+```text
+8 shards * 3000 RPS * 8 KB =~ 190 MB/s
+```
+
+so depending on the LFC size, prewarming will take at least:
+
+- ~5s for 1 GB
+- ~50s for 10 GB
+- ~5m for 100 GB
+- \>1h for 1 TB
+
+In total, one pageserver is normally capped by 30k RPS, so it obviously can't sustain many computes
+doing prewarm at the same time. Later, we may need an additional mechanism for computes to throttle
+the prewarming requests gracefully.
+
+### Reliability, failure modes and corner cases
+
+We consider following failures while implementing this RFC:
+
+1. Compute got interrupted/crashed/restarted during prewarm. The caller -- control plane -- should
+    detect that and start prewarm from the beginning.
+
+2. Control plane promotion request timed out or hit network issues. If it never reached the
+    compute, control plane should just repeat it. If it did reach the compute, then during
+    retry control plane can hit `409` as previous request triggered the promotion already.
+    In this case, control plane need to retry until either `200` or
+    permanent error `500` is returned.
+
+3. Compute got interrupted/crashed/restarted during promotion. At restart it will ask for
+    a spec from control plane, and its content should signal compute to start as **primary**,
+    so it's expected that control plane will continue polling for certain period of time and
+    will discover that compute is ready to accept connections if restart is fast enough.
+
+4. Any other unexpected failure or timeout during prewarming. This **failure mustn't be fatal**,
+    control plane has to report failure, terminate replica and keep primary running.
+
+5. Any other unexpected failure or timeout during promotion. Unfortunately, at this moment
+    we already have the primary node stopped, so the only option is to start primary again
+    and proceed with auto-prewarm.
+
+6. Any unexpected failure during auto-prewarm. This **failure mustn't be fatal**,
+    `compute_ctl` has to report the failure, but do not crash the compute.
+
+7. Control plane failed to confirm that old primary has terminated. This can happen, especially
+    in the future HA setup. In this case, control plane has to ensure that it sent VM deletion
+    and pod termination requests to k8s, so long-term we do not have two running primaries
+    on the same timeline.
+
+### Security implications
+
+There are two security implications to consider:
+
+1. Access to `compute_ctl` API. It has to be accessible from the outside of compute, so all
+    new API methods have to be exposed on the **external** HTTP port and **must** be authenticated
+    with JWT.
+
+2. Read/write only your own LFC state data in S3. Although it's not really a security concern,
+    since LFC state is just a mapping of blocks present in LFC at certain moment in time;
+    it still has to be highly restricted, so that i) only computes on the same timeline can
+    read S3 state; ii) each compute can only write to the path that contains it's `endpoint_id`.
+    Both of this must be validated by Endpoint storage service using the JWT token provided by `compute_ctl`.
+
+### Unresolved questions
+
+#### Billing, metrics and monitoring
+
+Currently, we only label computes with `endpoint_id` after attaching them to the endpoint.
+In this proposal, this means that temporary replica will remain unlabelled until it's promoted
+to primary. We can also hide it from users in the control plane API, but what to do with
+billing and monitoring is still unclear.
+
+We can probably mark it as 'billable' and tag with `project_id`, so it will be billed, but
+not interfere in any way with the current primary monitoring.
+
+Another thing to consider is how logs and metrics export will switch to the new compute.
+It's expected that OpenTelemetry collector will auto-discover the new compute and start
+scraping metrics from it.
+
+#### Auto-prewarm
+
+It's still an open question whether we need auto-prewarm at all. The author's gut-feeling is
+that yes, we need it, but might be not for all workloads, so it could end up exposed as a
+user-controllable knob on the endpoint. There are two arguments for that:
+
+1. Auto-prewarm existing in upstream's `pg_prewarm`, _probably for a reason_.
+
+2. There are still could be 2 flows when we cannot perform the rolling restart via the warm
+    replica: i) any failure or interruption during promotion; ii) wake up after scale-to-zero.
+    The latter might be challenged as well, i.e. one can argue that auto-prewarm may and will
+    compete with user-workload for storage resources. This is correct, but it might as well
+    reduce the time to get warm LFC and good performance.
+
+#### Low-level details of the replica promotion
+
+There are many things to consider here, but three items just off the top of my head:
+
+1. How to properly start the `walproposer` inside Postgres.
+
+2. What to do with logical replication. Currently, we do not include logical replication slots
+    inside basebackup, because nobody advances them at replica, so they just prevent the WAL
+    deletion. Yet, we do need to have them at primary after promotion. Starting with Postgres 17,
+    there is a new feature called
+    [logical replication failover](https://www.postgresql.org/docs/current/logical-replication-failover.html)
+    and `synchronized_standby_slots` setting, but we need a plan for the older versions. Should we
+    request a new basebackup during promotion?
+
+3. How do we guarantee that replica will receive all the latest WAL from safekeepers? Do some
+    'shallow' version of sync safekeepers without data copying? Or just a standard version of
+    sync safekeepers?
+
+## Alternative implementation
+
+The proposal already assumes one of the alternatives -- do not have any persistent storage for
+LFC state. This is possible to implement faster with the proposed API, but it means that
+we do not implement auto-prewarm yet.
+
+## Definition of Done
+
+At the end of implementing this RFC we should have two high-level settings that enable:
+
+1. Auto-prewarm of user computes upon restart.
+2. Perform primary compute restart via the warm replica promotion.
+
+It also has to be decided what's the criteria for enabling one or both of these flows for
+certain clients.

libs/compute_api/Cargo.toml

@@ -12,6 +12,7 @@ jsonwebtoken.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 regex.workspace = true
+url.workspace = true
 
 utils = { path = "../utils" }
 remote_storage = { version = "0.1", path = "../remote_storage/" }

libs/compute_api/src/responses.rs

@@ -58,7 +58,7 @@ pub enum LfcPrewarmState {
     },
 }
 
-#[derive(Serialize, Default, Debug, Clone)]
+#[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcOffloadState {
     #[default]

libs/compute_api/src/spec.rs

@@ -4,11 +4,14 @@
 //! provide it by calling the compute_ctl's `/compute_ctl` endpoint, or
 //! compute_ctl can fetch it by calling the control plane's API.
 use std::collections::HashMap;
+use std::fmt::Display;
 
+use anyhow::anyhow;
 use indexmap::IndexMap;
 use regex::Regex;
 use remote_storage::RemotePath;
 use serde::{Deserialize, Serialize};
+use url::Url;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -178,9 +181,18 @@ pub struct ComputeSpec {
     /// JWT for authorizing requests to endpoint storage service
     pub endpoint_storage_token: Option<String>,
 
-    /// Download LFC state from endpoint_storage and pass it to Postgres on startup
     #[serde(default)]
+    /// Download LFC state from endpoint storage and pass it to Postgres on compute startup
     pub autoprewarm: bool,
+
+    #[serde(default)]
+    /// Upload LFC state to endpoint storage periodically. Default value (None) means "don't upload"
+    pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
+
+    /// Suspend timeout in seconds.
+    ///
+    /// We use this value to derive other values, such as the installed extensions metric.
+    pub suspend_timeout_seconds: i64,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -429,6 +441,47 @@ pub struct JwksSettings {
     pub jwt_audience: Option<String>,
 }
 
+/// Protocol used to connect to a Pageserver. Parsed from the connstring scheme.
+#[derive(Clone, Copy, Debug, Default)]
+pub enum PageserverProtocol {
+    /// The original protocol based on libpq and COPY. Uses postgresql:// or postgres:// scheme.
+    #[default]
+    Libpq,
+    /// A newer, gRPC-based protocol. Uses grpc:// scheme.
+    Grpc,
+}
+
+impl PageserverProtocol {
+    /// Parses the protocol from a connstring scheme. Defaults to Libpq if no scheme is given.
+    /// Errors if the connstring is an invalid URL.
+    pub fn from_connstring(connstring: &str) -> anyhow::Result<Self> {
+        let scheme = match Url::parse(connstring) {
+            Ok(url) => url.scheme().to_lowercase(),
+            Err(url::ParseError::RelativeUrlWithoutBase) => return Ok(Self::default()),
+            Err(err) => return Err(anyhow!("invalid connstring URL: {err}")),
+        };
+        match scheme.as_str() {
+            "postgresql" | "postgres" => Ok(Self::Libpq),
+            "grpc" => Ok(Self::Grpc),
+            scheme => Err(anyhow!("invalid protocol scheme: {scheme}")),
+        }
+    }
+
+    /// Returns the URL scheme for the protocol, for use in connstrings.
+    pub fn scheme(&self) -> &'static str {
+        match self {
+            Self::Libpq => "postgresql",
+            Self::Grpc => "grpc",
+        }
+    }
+}
+
+impl Display for PageserverProtocol {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.scheme())
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use std::fs::File;

libs/compute_api/tests/cluster_spec.json

@@ -3,6 +3,7 @@
 
     "timestamp": "2021-05-23T18:25:43.511Z",
     "operation_uuid": "0f657b36-4b0f-4a2d-9c2e-1dcd615e7d8b",
+    "suspend_timeout_seconds": 3600,
 
     "cluster": {
         "cluster_id": "test-cluster-42",
@@ -89,6 +90,11 @@
                 "value": "off",
                 "vartype": "bool"
             },
+            {
+                "name": "offload_lfc_interval_seconds",
+                "value": "20",
+                "vartype": "integer"
+            },
             {
                 "name": "neon.safekeepers",
                 "value": "127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501",

libs/pageserver_api/Cargo.toml

@@ -19,6 +19,7 @@ byteorder.workspace = true
 utils.workspace = true
 postgres_ffi_types.workspace = true
 postgres_versioninfo.workspace = true
+posthog_client_lite.workspace = true
 enum-map.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
@@ -29,12 +30,13 @@ humantime-serde.workspace = true
 chrono = { workspace = true, features = ["serde"] }
 itertools.workspace = true
 storage_broker.workspace = true
-camino = {workspace = true, features = ["serde1"]}
+camino = { workspace = true, features = ["serde1"] }
 remote_storage.workspace = true
 postgres_backend.workspace = true
-nix = {workspace = true, optional = true}
+nix = { workspace = true, optional = true }
 reqwest.workspace = true
 rand.workspace = true
+tracing.workspace = true
 tracing-utils.workspace = true
 once_cell.workspace = true
 

libs/pageserver_api/src/config.rs

@@ -4,6 +4,7 @@ use camino::Utf8PathBuf;
 mod tests;
 
 use const_format::formatcp;
+use posthog_client_lite::PostHogClientConfig;
 pub const DEFAULT_PG_LISTEN_PORT: u16 = 64000;
 pub const DEFAULT_PG_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_PG_LISTEN_PORT}");
 pub const DEFAULT_HTTP_LISTEN_PORT: u16 = 9898;
@@ -63,25 +64,66 @@ impl Display for NodeMetadata {
     }
 }
 
-/// PostHog integration config.
+/// PostHog integration config. This is used in pageserver, storcon, and neon_local.
+/// Ensure backward compatibility when adding new fields.
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub struct PostHogConfig {
     /// PostHog project ID
-    pub project_id: String,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub project_id: Option<String>,
     /// Server-side (private) API key
-    pub server_api_key: String,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub server_api_key: Option<String>,
     /// Client-side (public) API key
-    pub client_api_key: String,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub client_api_key: Option<String>,
     /// Private API URL
-    pub private_api_url: String,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub private_api_url: Option<String>,
     /// Public API URL
-    pub public_api_url: String,
-    /// Refresh interval for the feature flag spec
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub public_api_url: Option<String>,
+    /// Refresh interval for the feature flag spec.
+    /// The storcon will push the feature flag spec to the pageserver. If the pageserver does not receive
+    /// the spec for `refresh_interval`, it will fetch the spec from the PostHog API.
+    #[serde(default)]
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(with = "humantime_serde")]
     pub refresh_interval: Option<Duration>,
 }
 
+impl PostHogConfig {
+    pub fn try_into_posthog_config(self) -> Result<PostHogClientConfig, &'static str> {
+        let Some(project_id) = self.project_id else {
+            return Err("project_id is required");
+        };
+        let Some(server_api_key) = self.server_api_key else {
+            return Err("server_api_key is required");
+        };
+        let Some(client_api_key) = self.client_api_key else {
+            return Err("client_api_key is required");
+        };
+        let Some(private_api_url) = self.private_api_url else {
+            return Err("private_api_url is required");
+        };
+        let Some(public_api_url) = self.public_api_url else {
+            return Err("public_api_url is required");
+        };
+        Ok(PostHogClientConfig {
+            project_id,
+            server_api_key,
+            client_api_key,
+            private_api_url,
+            public_api_url,
+        })
+    }
+}
+
 /// `pageserver.toml`
 ///
 /// We use serde derive with `#[serde(default)]` to generate a deserializer
@@ -367,6 +409,9 @@ pub struct BasebackupCacheConfig {
     // TODO(diko): support max_entry_size_bytes.
     // pub max_entry_size_bytes: u64,
     pub max_size_entries: usize,
+    /// Size of the channel used to send prepare requests to the basebackup cache worker.
+    /// If exceeded, new prepare requests will be dropped.
+    pub prepare_channel_size: usize,
 }
 
 impl Default for BasebackupCacheConfig {
@@ -375,7 +420,8 @@ impl Default for BasebackupCacheConfig {
             cleanup_period: Duration::from_secs(60),
             max_total_size_bytes: 1024 * 1024 * 1024, // 1 GiB
             // max_entry_size_bytes: 16 * 1024 * 1024,   // 16 MiB
-            max_size_entries: 1000,
+            max_size_entries: 10000,
+            prepare_channel_size: 100,
         }
     }
 }

libs/pageserver_api/src/controller_api.rs

@@ -386,6 +386,7 @@ pub enum NodeSchedulingPolicy {
     Pause,
     PauseForRestart,
     Draining,
+    Deleting,
 }
 
 impl FromStr for NodeSchedulingPolicy {
@@ -398,6 +399,7 @@ impl FromStr for NodeSchedulingPolicy {
             "pause" => Ok(Self::Pause),
             "pause_for_restart" => Ok(Self::PauseForRestart),
             "draining" => Ok(Self::Draining),
+            "deleting" => Ok(Self::Deleting),
             _ => Err(anyhow::anyhow!("Unknown scheduling state '{s}'")),
         }
     }
@@ -412,6 +414,7 @@ impl From<NodeSchedulingPolicy> for String {
             Pause => "pause",
             PauseForRestart => "pause_for_restart",
             Draining => "draining",
+            Deleting => "deleting",
         }
         .to_string()
     }
@@ -420,6 +423,7 @@ impl From<NodeSchedulingPolicy> for String {
 #[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
 pub enum SkSchedulingPolicy {
     Active,
+    Activating,
     Pause,
     Decomissioned,
 }
@@ -430,6 +434,7 @@ impl FromStr for SkSchedulingPolicy {
     fn from_str(s: &str) -> Result<Self, Self::Err> {
         Ok(match s {
             "active" => Self::Active,
+            "activating" => Self::Activating,
             "pause" => Self::Pause,
             "decomissioned" => Self::Decomissioned,
             _ => {
@@ -446,6 +451,7 @@ impl From<SkSchedulingPolicy> for String {
         use SkSchedulingPolicy::*;
         match value {
             Active => "active",
+            Activating => "activating",
             Pause => "pause",
             Decomissioned => "decomissioned",
         }
@@ -546,6 +552,11 @@ pub struct TimelineImportRequest {
     pub sk_set: Vec<NodeId>,
 }
 
+#[derive(serde::Serialize, serde::Deserialize, Clone)]
+pub struct TimelineSafekeeperMigrateRequest {
+    pub new_sk_set: Vec<NodeId>,
+}
+
 #[cfg(test)]
 mod test {
     use serde_json;

libs/pageserver_api/src/models.rs

@@ -21,7 +21,9 @@ use utils::{completion, serde_system_time};
 
 use crate::config::Ratio;
 use crate::key::{CompactKey, Key};
-use crate::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardStripeSize, TenantShardId};
+use crate::shard::{
+    DEFAULT_STRIPE_SIZE, ShardCount, ShardIdentity, ShardStripeSize, TenantShardId,
+};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -475,7 +477,7 @@ pub struct TenantShardSplitResponse {
 }
 
 /// Parameters that apply to all shards in a tenant.  Used during tenant creation.
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Clone, Copy, Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct ShardParameters {
     pub count: ShardCount,
@@ -497,6 +499,15 @@ impl Default for ShardParameters {
     }
 }
 
+impl From<ShardIdentity> for ShardParameters {
+    fn from(identity: ShardIdentity) -> Self {
+        Self {
+            count: identity.count,
+            stripe_size: identity.stripe_size,
+        }
+    }
+}
+
 #[derive(Debug, Default, Clone, Eq, PartialEq)]
 pub enum FieldPatch<T> {
     Upsert(T),
@@ -1591,6 +1602,10 @@ pub struct TimelineInfo {
     /// The status of the rel_size migration.
     pub rel_size_migration: Option<RelSizeMigration>,
 
+    /// Whether the timeline is read-only (doesn't support WAL advancing).
+    /// The property is not recursive; child timelines can be non-read-only.
+    pub read_only: bool,
+
     /// Whether the timeline is invisible in synthetic size calculations.
     pub is_invisible: Option<bool>,
 }

libs/pageserver_api/src/shard.rs

@@ -37,6 +37,7 @@ use std::hash::{Hash, Hasher};
 pub use ::utils::shard::*;
 use postgres_ffi_types::forknum::INIT_FORKNUM;
 use serde::{Deserialize, Serialize};
+use utils::critical;
 
 use crate::key::Key;
 use crate::models::ShardParameters;
@@ -179,7 +180,7 @@ impl ShardIdentity {
 
     /// For use when creating ShardIdentity instances for new shards, where a creation request
     /// specifies the ShardParameters that apply to all shards.
-    pub fn from_params(number: ShardNumber, params: &ShardParameters) -> Self {
+    pub fn from_params(number: ShardNumber, params: ShardParameters) -> Self {
         Self {
             number,
             count: params.count,
@@ -188,6 +189,17 @@ impl ShardIdentity {
         }
     }
 
+    /// Asserts that the given shard identities are equal. Changes to shard parameters will likely
+    /// result in data corruption.
+    pub fn assert_equal(&self, other: ShardIdentity) {
+        if self != &other {
+            // TODO: for now, we're conservative and just log errors in production. Turn this into a
+            // real assertion when we're confident it doesn't misfire, and also reject requests that
+            // attempt to change it with an error response.
+            critical!("shard identity mismatch: {self:?} != {other:?}");
+        }
+    }
+
     fn is_broken(&self) -> bool {
         self.layout == LAYOUT_BROKEN
     }

libs/postgres_backend/src/lib.rs

@@ -78,7 +78,13 @@ pub fn is_expected_io_error(e: &io::Error) -> bool {
     use io::ErrorKind::*;
     matches!(
         e.kind(),
-        BrokenPipe | ConnectionRefused | ConnectionAborted | ConnectionReset | TimedOut
+        HostUnreachable
+            | NetworkUnreachable
+            | BrokenPipe
+            | ConnectionRefused
+            | ConnectionAborted
+            | ConnectionReset
+            | TimedOut,
     )
 }
 

libs/posthog_client_lite/src/background_loop.rs

@@ -1,17 +1,22 @@
 //! A background loop that fetches feature flags from PostHog and updates the feature store.
 
-use std::{sync::Arc, time::Duration};
+use std::{
+    sync::Arc,
+    time::{Duration, SystemTime},
+};
 
 use arc_swap::ArcSwap;
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, info_span};
 
-use crate::{CaptureEvent, FeatureStore, PostHogClient, PostHogClientConfig};
+use crate::{
+    CaptureEvent, FeatureStore, LocalEvaluationResponse, PostHogClient, PostHogClientConfig,
+};
 
 /// A background loop that fetches feature flags from PostHog and updates the feature store.
 pub struct FeatureResolverBackgroundLoop {
     posthog_client: PostHogClient,
-    feature_store: ArcSwap<FeatureStore>,
+    feature_store: ArcSwap<(SystemTime, Arc<FeatureStore>)>,
     cancel: CancellationToken,
 }
 
@@ -19,11 +24,35 @@ impl FeatureResolverBackgroundLoop {
     pub fn new(config: PostHogClientConfig, shutdown_pageserver: CancellationToken) -> Self {
         Self {
             posthog_client: PostHogClient::new(config),
-            feature_store: ArcSwap::new(Arc::new(FeatureStore::new())),
+            feature_store: ArcSwap::new(Arc::new((
+                SystemTime::UNIX_EPOCH,
+                Arc::new(FeatureStore::new()),
+            ))),
             cancel: shutdown_pageserver,
         }
     }
 
+    /// Update the feature store with a new feature flag spec bypassing the normal refresh loop.
+    pub fn update(&self, spec: String) -> anyhow::Result<()> {
+        let resp: LocalEvaluationResponse = serde_json::from_str(&spec)?;
+        self.update_feature_store_nofail(resp, "http_propagate");
+        Ok(())
+    }
+
+    fn update_feature_store_nofail(&self, resp: LocalEvaluationResponse, source: &'static str) {
+        let project_id = self.posthog_client.config.project_id.parse::<u64>().ok();
+        match FeatureStore::new_with_flags(resp.flags, project_id) {
+            Ok(feature_store) => {
+                self.feature_store
+                    .store(Arc::new((SystemTime::now(), Arc::new(feature_store))));
+                tracing::info!("Feature flag updated from {}", source);
+            }
+            Err(e) => {
+                tracing::warn!("Cannot process feature flag spec from {}: {}", source, e);
+            }
+        }
+    }
+
     pub fn spawn(
         self: Arc<Self>,
         handle: &tokio::runtime::Handle,
@@ -47,6 +76,17 @@ impl FeatureResolverBackgroundLoop {
                         _ = ticker.tick() => {}
                         _ = cancel.cancelled() => break
                     }
+                    {
+                        let last_update = this.feature_store.load().0;
+                        if let Ok(elapsed) = last_update.elapsed() {
+                            if elapsed < refresh_period {
+                                tracing::debug!(
+                                    "Skipping feature flag refresh because it's too soon"
+                                );
+                                continue;
+                            }
+                        }
+                    }
                     let resp = match this
                         .posthog_client
                         .get_feature_flags_local_evaluation()
@@ -58,16 +98,7 @@ impl FeatureResolverBackgroundLoop {
                             continue;
                         }
                     };
-                    let project_id = this.posthog_client.config.project_id.parse::<u64>().ok();
-                    match FeatureStore::new_with_flags(resp.flags, project_id) {
-                        Ok(feature_store) => {
-                            this.feature_store.store(Arc::new(feature_store));
-                            tracing::info!("Feature flag updated");
-                        }
-                        Err(e) => {
-                            tracing::warn!("Cannot process feature flag spec: {}", e);
-                        }
-                    }
+                    this.update_feature_store_nofail(resp, "refresh_loop");
                 }
                 tracing::info!("PostHog feature resolver stopped");
             }
@@ -92,6 +123,6 @@ impl FeatureResolverBackgroundLoop {
     }
 
     pub fn feature_store(&self) -> Arc<FeatureStore> {
-        self.feature_store.load_full()
+        self.feature_store.load().1.clone()
     }
 }

libs/posthog_client_lite/src/lib.rs

@@ -544,17 +544,8 @@ impl PostHogClient {
         self.config.server_api_key.starts_with("phs_")
     }
 
-    /// Fetch the feature flag specs from the server.
-    ///
-    /// This is unfortunately an undocumented API at:
-    /// - <https://posthog.com/docs/api/feature-flags#get-api-projects-project_id-feature_flags-local_evaluation>
-    /// - <https://posthog.com/docs/feature-flags/local-evaluation>
-    ///
-    /// The handling logic in [`FeatureStore`] mostly follows the Python API implementation.
-    /// See `_compute_flag_locally` in <https://github.com/PostHog/posthog-python/blob/master/posthog/client.py>
-    pub async fn get_feature_flags_local_evaluation(
-        &self,
-    ) -> anyhow::Result<LocalEvaluationResponse> {
+    /// Get the raw JSON spec, same as `get_feature_flags_local_evaluation` but without parsing.
+    pub async fn get_feature_flags_local_evaluation_raw(&self) -> anyhow::Result<String> {
         // BASE_URL/api/projects/:project_id/feature_flags/local_evaluation
         // with bearer token of self.server_api_key
         // OR
@@ -588,7 +579,22 @@ impl PostHogClient {
                 body
             ));
         }
-        Ok(serde_json::from_str(&body)?)
+        Ok(body)
+    }
+
+    /// Fetch the feature flag specs from the server.
+    ///
+    /// This is unfortunately an undocumented API at:
+    /// - <https://posthog.com/docs/api/feature-flags#get-api-projects-project_id-feature_flags-local_evaluation>
+    /// - <https://posthog.com/docs/feature-flags/local-evaluation>
+    ///
+    /// The handling logic in [`FeatureStore`] mostly follows the Python API implementation.
+    /// See `_compute_flag_locally` in <https://github.com/PostHog/posthog-python/blob/master/posthog/client.py>
+    pub async fn get_feature_flags_local_evaluation(
+        &self,
+    ) -> Result<LocalEvaluationResponse, anyhow::Error> {
+        let raw = self.get_feature_flags_local_evaluation_raw().await?;
+        Ok(serde_json::from_str(&raw)?)
     }
 
     /// Capture an event. This will only be used to report the feature flag usage back to PostHog, though

libs/proxy/postgres-protocol2/src/authentication/sasl.rs

@@ -52,7 +52,7 @@ pub(crate) async fn hi(str: &[u8], salt: &[u8], iterations: u32) -> [u8; 32] {
         }
         // yield every ~250us
         // hopefully reduces tail latencies
-        if i % 1024 == 0 {
+        if i.is_multiple_of(1024) {
             yield_now().await
         }
     }

libs/proxy/tokio-postgres2/src/client.rs

@@ -90,7 +90,7 @@ pub struct InnerClient {
 }
 
 impl InnerClient {
-    pub fn start(&mut self) -> Result<PartialQuery, Error> {
+    pub fn start(&mut self) -> Result<PartialQuery<'_>, Error> {
         self.responses.waiting += 1;
         Ok(PartialQuery(Some(self)))
     }
@@ -227,7 +227,7 @@ impl Client {
         &mut self,
         statement: &str,
         params: I,
-    ) -> Result<RowStream, Error>
+    ) -> Result<RowStream<'_>, Error>
     where
         S: AsRef<str>,
         I: IntoIterator<Item = Option<S>>,
@@ -262,7 +262,7 @@ impl Client {
     pub(crate) async fn simple_query_raw(
         &mut self,
         query: &str,
-    ) -> Result<SimpleQueryStream, Error> {
+    ) -> Result<SimpleQueryStream<'_>, Error> {
         simple_query::simple_query(self.inner_mut(), query).await
     }
 

libs/proxy/tokio-postgres2/src/config.rs

@@ -12,7 +12,9 @@ use tokio::net::TcpStream;
 
 use crate::connect::connect;
 use crate::connect_raw::{RawConnection, connect_raw};
-use crate::tls::{MakeTlsConnect, TlsConnect};
+use crate::connect_tls::connect_tls;
+use crate::maybe_tls_stream::MaybeTlsStream;
+use crate::tls::{MakeTlsConnect, TlsConnect, TlsStream};
 use crate::{Client, Connection, Error};
 
 /// TLS configuration.
@@ -238,7 +240,7 @@ impl Config {
         connect(tls, self).await
     }
 
-    pub async fn connect_raw<S, T>(
+    pub async fn tls_and_authenticate<S, T>(
         &self,
         stream: S,
         tls: T,
@@ -247,7 +249,19 @@ impl Config {
         S: AsyncRead + AsyncWrite + Unpin,
         T: TlsConnect<S>,
     {
-        connect_raw(stream, tls, self).await
+        let stream = connect_tls(stream, self.ssl_mode, tls).await?;
+        connect_raw(stream, self).await
+    }
+
+    pub async fn authenticate<S, T>(
+        &self,
+        stream: MaybeTlsStream<S, T>,
+    ) -> Result<RawConnection<S, T>, Error>
+    where
+        S: AsyncRead + AsyncWrite + Unpin,
+        T: TlsStream + Unpin,
+    {
+        connect_raw(stream, self).await
     }
 }
 

libs/proxy/tokio-postgres2/src/connect.rs

@@ -9,6 +9,7 @@ use crate::codec::BackendMessage;
 use crate::config::Host;
 use crate::connect_raw::connect_raw;
 use crate::connect_socket::connect_socket;
+use crate::connect_tls::connect_tls;
 use crate::tls::{MakeTlsConnect, TlsConnect};
 use crate::{Client, Config, Connection, Error, RawConnection};
 
@@ -44,13 +45,14 @@ where
     T: TlsConnect<TcpStream>,
 {
     let socket = connect_socket(host_addr, host, port, config.connect_timeout).await?;
+    let stream = connect_tls(socket, config.ssl_mode, tls).await?;
     let RawConnection {
         stream,
         parameters,
         delayed_notice,
         process_id,
         secret_key,
-    } = connect_raw(socket, tls, config).await?;
+    } = connect_raw(stream, config).await?;
 
     let socket_config = SocketConfig {
         host_addr,

libs/proxy/tokio-postgres2/src/connect_raw.rs

@@ -16,9 +16,8 @@ use tokio_util::codec::Framed;
 use crate::Error;
 use crate::codec::{BackendMessage, BackendMessages, FrontendMessage, PostgresCodec};
 use crate::config::{self, AuthKeys, Config};
-use crate::connect_tls::connect_tls;
 use crate::maybe_tls_stream::MaybeTlsStream;
-use crate::tls::{TlsConnect, TlsStream};
+use crate::tls::TlsStream;
 
 pub struct StartupStream<S, T> {
     inner: Framed<MaybeTlsStream<S, T>, PostgresCodec>,
@@ -87,16 +86,13 @@ pub struct RawConnection<S, T> {
 }
 
 pub async fn connect_raw<S, T>(
-    stream: S,
-    tls: T,
+    stream: MaybeTlsStream<S, T>,
     config: &Config,
-) -> Result<RawConnection<S, T::Stream>, Error>
+) -> Result<RawConnection<S, T>, Error>
 where
     S: AsyncRead + AsyncWrite + Unpin,
-    T: TlsConnect<S>,
+    T: TlsStream + Unpin,
 {
-    let stream = connect_tls(stream, config.ssl_mode, tls).await?;
-
     let mut stream = StartupStream {
         inner: Framed::new(stream, PostgresCodec),
         buf: BackendMessages::empty(),

libs/proxy/tokio-postgres2/src/generic_client.rs

@@ -12,7 +12,11 @@ mod private {
 /// This trait is "sealed", and cannot be implemented outside of this crate.
 pub trait GenericClient: private::Sealed {
     /// Like `Client::query_raw_txt`.
-    async fn query_raw_txt<S, I>(&mut self, statement: &str, params: I) -> Result<RowStream, Error>
+    async fn query_raw_txt<S, I>(
+        &mut self,
+        statement: &str,
+        params: I,
+    ) -> Result<RowStream<'_>, Error>
     where
         S: AsRef<str> + Sync + Send,
         I: IntoIterator<Item = Option<S>> + Sync + Send,
@@ -22,7 +26,11 @@ pub trait GenericClient: private::Sealed {
 impl private::Sealed for Client {}
 
 impl GenericClient for Client {
-    async fn query_raw_txt<S, I>(&mut self, statement: &str, params: I) -> Result<RowStream, Error>
+    async fn query_raw_txt<S, I>(
+        &mut self,
+        statement: &str,
+        params: I,
+    ) -> Result<RowStream<'_>, Error>
     where
         S: AsRef<str> + Sync + Send,
         I: IntoIterator<Item = Option<S>> + Sync + Send,
@@ -35,7 +43,11 @@ impl GenericClient for Client {
 impl private::Sealed for Transaction<'_> {}
 
 impl GenericClient for Transaction<'_> {
-    async fn query_raw_txt<S, I>(&mut self, statement: &str, params: I) -> Result<RowStream, Error>
+    async fn query_raw_txt<S, I>(
+        &mut self,
+        statement: &str,
+        params: I,
+    ) -> Result<RowStream<'_>, Error>
     where
         S: AsRef<str> + Sync + Send,
         I: IntoIterator<Item = Option<S>> + Sync + Send,

libs/proxy/tokio-postgres2/src/transaction.rs

@@ -47,7 +47,7 @@ impl<'a> Transaction<'a> {
         &mut self,
         statement: &str,
         params: I,
-    ) -> Result<RowStream, Error>
+    ) -> Result<RowStream<'_>, Error>
     where
         S: AsRef<str>,
         I: IntoIterator<Item = Option<S>>,

libs/safekeeper_api/src/models.rs

@@ -210,7 +210,7 @@ pub struct TimelineStatus {
 }
 
 /// Request to switch membership configuration.
-#[derive(Serialize, Deserialize)]
+#[derive(Clone, Serialize, Deserialize)]
 #[serde(transparent)]
 pub struct TimelineMembershipSwitchRequest {
     pub mconf: Configuration,
@@ -221,6 +221,8 @@ pub struct TimelineMembershipSwitchRequest {
 pub struct TimelineMembershipSwitchResponse {
     pub previous_conf: Configuration,
     pub current_conf: Configuration,
+    pub last_log_term: Term,
+    pub flush_lsn: Lsn,
 }
 
 #[derive(Clone, Copy, Serialize, Deserialize)]

libs/utils/src/logging.rs

@@ -24,12 +24,28 @@ macro_rules! critical {
         if cfg!(debug_assertions) {
             panic!($($arg)*);
         }
+        // Increment both metrics
         $crate::logging::TRACING_EVENT_COUNT_METRIC.inc_critical();
         let backtrace = std::backtrace::Backtrace::capture();
         tracing::error!("CRITICAL: {}\n{backtrace}", format!($($arg)*));
     }};
 }
 
+#[macro_export]
+macro_rules! critical_timeline {
+    ($tenant_shard_id:expr, $timeline_id:expr, $($arg:tt)*) => {{
+        if cfg!(debug_assertions) {
+            panic!($($arg)*);
+        }
+        // Increment both metrics
+        $crate::logging::TRACING_EVENT_COUNT_METRIC.inc_critical();
+        $crate::logging::HADRON_CRITICAL_STORAGE_EVENT_COUNT_METRIC.inc(&$tenant_shard_id.to_string(), &$timeline_id.to_string());
+        let backtrace = std::backtrace::Backtrace::capture();
+        tracing::error!("CRITICAL: [tenant_shard_id: {}, timeline_id: {}] {}\n{backtrace}",
+                       $tenant_shard_id, $timeline_id, format!($($arg)*));
+    }};
+}
+
 #[derive(EnumString, strum_macros::Display, VariantNames, Eq, PartialEq, Debug, Clone, Copy)]
 #[strum(serialize_all = "snake_case")]
 pub enum LogFormat {
@@ -61,6 +77,36 @@ pub struct TracingEventCountMetric {
     trace: IntCounter,
 }
 
+// Begin Hadron: Add a HadronCriticalStorageEventCountMetric metric that is sliced by tenant_id and timeline_id
+pub struct HadronCriticalStorageEventCountMetric {
+    critical: IntCounterVec,
+}
+
+pub static HADRON_CRITICAL_STORAGE_EVENT_COUNT_METRIC: Lazy<HadronCriticalStorageEventCountMetric> =
+    Lazy::new(|| {
+        let vec = metrics::register_int_counter_vec!(
+            "hadron_critical_storage_event_count",
+            "Number of critical storage events, by tenant_id and timeline_id",
+            &["tenant_shard_id", "timeline_id"]
+        )
+        .expect("failed to define metric");
+        HadronCriticalStorageEventCountMetric::new(vec)
+    });
+
+impl HadronCriticalStorageEventCountMetric {
+    fn new(vec: IntCounterVec) -> Self {
+        Self { critical: vec }
+    }
+
+    // Allow public access from `critical!` macro.
+    pub fn inc(&self, tenant_shard_id: &str, timeline_id: &str) {
+        self.critical
+            .with_label_values(&[tenant_shard_id, timeline_id])
+            .inc();
+    }
+}
+// End Hadron
+
 pub static TRACING_EVENT_COUNT_METRIC: Lazy<TracingEventCountMetric> = Lazy::new(|| {
     let vec = metrics::register_int_counter_vec!(
         "libmetrics_tracing_event_count",

libs/utils/src/sync/gate.rs

@@ -86,6 +86,14 @@ pub enum GateError {
     GateClosed,
 }
 
+impl GateError {
+    pub fn is_cancel(&self) -> bool {
+        match self {
+            GateError::GateClosed => true,
+        }
+    }
+}
+
 impl Default for Gate {
     fn default() -> Self {
         Self {

pageserver/client/src/mgmt_api.rs

@@ -844,4 +844,13 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn update_feature_flag_spec(&self, spec: String) -> Result<()> {
+        let uri = format!("{}/v1/feature_flag_spec", self.mgmt_api_endpoint);
+        self.request(Method::POST, uri, spec)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/page_api/Cargo.toml

@@ -9,12 +9,14 @@ anyhow.workspace = true
 bytes.workspace = true
 futures.workspace = true
 pageserver_api.workspace = true
-postgres_ffi.workspace = true
+postgres_ffi_types.workspace = true
 prost.workspace = true
+prost-types.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
+tokio-util.workspace = true
 tonic.workspace = true
 utils.workspace = true
 workspace_hack.workspace = true

pageserver/page_api/proto/page_service.proto

@@ -35,6 +35,8 @@
 syntax = "proto3";
 package page_api;
 
+import "google/protobuf/timestamp.proto";
+
 service PageService {
   // Returns whether a relation exists.
   rpc CheckRelExists(CheckRelExistsRequest) returns (CheckRelExistsResponse);
@@ -64,6 +66,10 @@ service PageService {
 
   // Fetches an SLRU segment.
   rpc GetSlruSegment (GetSlruSegmentRequest) returns (GetSlruSegmentResponse);
+
+  // Acquires or extends a lease on the given LSN. This guarantees that the Pageserver won't garbage
+  // collect the LSN until the lease expires. Must be acquired on all relevant shards.
+  rpc LeaseLsn (LeaseLsnRequest) returns (LeaseLsnResponse);
 }
 
 // The LSN a request should read at.
@@ -110,6 +116,19 @@ message GetBaseBackupRequest {
   bool replica = 2;
   // If true, include relation files in the base backup. Mainly for debugging and tests.
   bool full = 3;
+  // Compression algorithm to use. Base backups send a compressed payload instead of using gRPC
+  // compression, so that we can cache compressed backups on the server.
+  BaseBackupCompression compression = 4;
+}
+
+// Base backup compression algorithms.
+enum BaseBackupCompression {
+  // Unknown algorithm. Used when clients send an unsupported algorithm.
+  BASE_BACKUP_COMPRESSION_UNKNOWN = 0;
+  // No compression.
+  BASE_BACKUP_COMPRESSION_NONE = 1;
+  // GZIP compression.
+  BASE_BACKUP_COMPRESSION_GZIP = 2;
 }
 
 // Base backup response chunk, returned as an ordered stream.
@@ -239,3 +258,17 @@ message GetSlruSegmentRequest {
 message GetSlruSegmentResponse {
   bytes segment = 1;
 }
+
+// Acquires or extends a lease on the given LSN. This guarantees that the Pageserver won't garbage
+// collect the LSN until the lease expires. Must be acquired on all relevant shards.
+message LeaseLsnRequest {
+  // The LSN to lease. Can't be 0 or below the current GC cutoff.
+  uint64 lsn = 1;
+}
+
+// Lease acquisition response. If the lease could not be granted because the LSN has already been
+// garbage collected, a FailedPrecondition status will be returned instead.
+message LeaseLsnResponse {
+  // The lease expiration time.
+  google.protobuf.Timestamp expires = 1;
+}

pageserver/page_api/src/client.rs

@@ -1,8 +1,7 @@
-use std::convert::TryInto;
-
-use bytes::Bytes;
-use futures::TryStreamExt;
-use futures::{Stream, StreamExt};
+use anyhow::Result;
+use futures::{Stream, StreamExt as _, TryStreamExt as _};
+use tokio::io::AsyncRead;
+use tokio_util::io::StreamReader;
 use tonic::metadata::AsciiMetadataValue;
 use tonic::metadata::errors::InvalidMetadataValue;
 use tonic::transport::Channel;
@@ -12,8 +11,6 @@ use utils::id::TenantId;
 use utils::id::TimelineId;
 use utils::shard::ShardIndex;
 
-use anyhow::Result;
-
 use crate::model;
 use crate::proto;
 
@@ -69,6 +66,7 @@ impl tonic::service::Interceptor for AuthInterceptor {
         Ok(req)
     }
 }
+
 #[derive(Clone)]
 pub struct Client {
     client: proto::PageServiceClient<
@@ -95,7 +93,6 @@ impl Client {
 
         if let Some(compression) = compression {
             // TODO: benchmark this (including network latency).
-            // TODO: consider enabling compression by default.
             client = client
                 .accept_compressed(compression)
                 .send_compressed(compression);
@@ -121,22 +118,15 @@ impl Client {
     pub async fn get_base_backup(
         &mut self,
         req: model::GetBaseBackupRequest,
-    ) -> Result<impl Stream<Item = Result<Bytes, tonic::Status>> + 'static, tonic::Status> {
-        let proto_req = proto::GetBaseBackupRequest::from(req);
-
-        let response_stream: Streaming<proto::GetBaseBackupResponseChunk> =
-            self.client.get_base_backup(proto_req).await?.into_inner();
-
-        // TODO: Consider dechunking internally
-        let domain_stream = response_stream.map(|chunk_res| {
-            chunk_res.and_then(|proto_chunk| {
-                proto_chunk.try_into().map_err(|e| {
-                    tonic::Status::internal(format!("Failed to convert response chunk: {e}"))
-                })
-            })
-        });
-
-        Ok(domain_stream)
+    ) -> Result<impl AsyncRead + use<>, tonic::Status> {
+        let req = proto::GetBaseBackupRequest::from(req);
+        let chunks = self.client.get_base_backup(req).await?.into_inner();
+        let reader = StreamReader::new(
+            chunks
+                .map_ok(|resp| resp.chunk)
+                .map_err(std::io::Error::other),
+        );
+        Ok(reader)
     }
 
     /// Returns the total size of a database, as # of bytes.
@@ -197,4 +187,17 @@ impl Client {
         let response = self.client.get_slru_segment(proto_req).await?;
         Ok(response.into_inner().try_into()?)
     }
+
+    /// Acquires or extends a lease on the given LSN. This guarantees that the Pageserver won't
+    /// garbage collect the LSN until the lease expires. Must be acquired on all relevant shards.
+    ///
+    /// Returns the lease expiration time, or a FailedPrecondition status if the lease could not be
+    /// acquired because the LSN has already been garbage collected.
+    pub async fn lease_lsn(
+        &mut self,
+        req: model::LeaseLsnRequest,
+    ) -> Result<model::LeaseLsnResponse, tonic::Status> {
+        let req = proto::LeaseLsnRequest::from(req);
+        Ok(self.client.lease_lsn(req).await?.into_inner().try_into()?)
+    }
 }

pageserver/page_api/src/model.rs

@@ -16,10 +16,11 @@
 //! stream combinators without dealing with errors, and avoids validating the same message twice.
 
 use std::fmt::Display;
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
 use bytes::Bytes;
-use postgres_ffi::Oid;
-// TODO: split out Lsn, RelTag, SlruKind, Oid and other basic types to a separate crate, to avoid
+use postgres_ffi_types::Oid;
+// TODO: split out Lsn, RelTag, SlruKind and other basic types to a separate crate, to avoid
 // pulling in all of their other crate dependencies when building the client.
 use utils::lsn::Lsn;
 
@@ -191,15 +192,21 @@ pub struct GetBaseBackupRequest {
     pub replica: bool,
     /// If true, include relation files in the base backup. Mainly for debugging and tests.
     pub full: bool,
+    /// Compression algorithm to use. Base backups send a compressed payload instead of using gRPC
+    /// compression, so that we can cache compressed backups on the server.
+    pub compression: BaseBackupCompression,
 }
 
-impl From<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
-    fn from(pb: proto::GetBaseBackupRequest) -> Self {
-        Self {
+impl TryFrom<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetBaseBackupRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
             lsn: (pb.lsn != 0).then_some(Lsn(pb.lsn)),
             replica: pb.replica,
             full: pb.full,
-        }
+            compression: pb.compression.try_into()?,
+        })
     }
 }
 
@@ -209,10 +216,55 @@ impl From<GetBaseBackupRequest> for proto::GetBaseBackupRequest {
             lsn: request.lsn.unwrap_or_default().0,
             replica: request.replica,
             full: request.full,
+            compression: request.compression.into(),
         }
     }
 }
 
+/// Base backup compression algorithm.
+#[derive(Clone, Copy, Debug)]
+pub enum BaseBackupCompression {
+    None,
+    Gzip,
+}
+
+impl TryFrom<proto::BaseBackupCompression> for BaseBackupCompression {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::BaseBackupCompression) -> Result<Self, Self::Error> {
+        match pb {
+            proto::BaseBackupCompression::Unknown => Err(ProtocolError::invalid("compression", pb)),
+            proto::BaseBackupCompression::None => Ok(Self::None),
+            proto::BaseBackupCompression::Gzip => Ok(Self::Gzip),
+        }
+    }
+}
+
+impl TryFrom<i32> for BaseBackupCompression {
+    type Error = ProtocolError;
+
+    fn try_from(compression: i32) -> Result<Self, Self::Error> {
+        proto::BaseBackupCompression::try_from(compression)
+            .map_err(|_| ProtocolError::invalid("compression", compression))
+            .and_then(Self::try_from)
+    }
+}
+
+impl From<BaseBackupCompression> for proto::BaseBackupCompression {
+    fn from(compression: BaseBackupCompression) -> Self {
+        match compression {
+            BaseBackupCompression::None => Self::None,
+            BaseBackupCompression::Gzip => Self::Gzip,
+        }
+    }
+}
+
+impl From<BaseBackupCompression> for i32 {
+    fn from(compression: BaseBackupCompression) -> Self {
+        proto::BaseBackupCompression::from(compression).into()
+    }
+}
+
 pub type GetBaseBackupResponseChunk = Bytes;
 
 impl TryFrom<proto::GetBaseBackupResponseChunk> for GetBaseBackupResponseChunk {
@@ -652,3 +704,54 @@ impl From<GetSlruSegmentResponse> for proto::GetSlruSegmentResponse {
 
 // SlruKind is defined in pageserver_api::reltag.
 pub type SlruKind = pageserver_api::reltag::SlruKind;
+
+/// Acquires or extends a lease on the given LSN. This guarantees that the Pageserver won't garbage
+/// collect the LSN until the lease expires.
+pub struct LeaseLsnRequest {
+    /// The LSN to lease.
+    pub lsn: Lsn,
+}
+
+impl TryFrom<proto::LeaseLsnRequest> for LeaseLsnRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::LeaseLsnRequest) -> Result<Self, Self::Error> {
+        if pb.lsn == 0 {
+            return Err(ProtocolError::Missing("lsn"));
+        }
+        Ok(Self { lsn: Lsn(pb.lsn) })
+    }
+}
+
+impl From<LeaseLsnRequest> for proto::LeaseLsnRequest {
+    fn from(request: LeaseLsnRequest) -> Self {
+        Self { lsn: request.lsn.0 }
+    }
+}
+
+/// Lease expiration time. If the lease could not be granted because the LSN has already been
+/// garbage collected, a FailedPrecondition status will be returned instead.
+pub type LeaseLsnResponse = SystemTime;
+
+impl TryFrom<proto::LeaseLsnResponse> for LeaseLsnResponse {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::LeaseLsnResponse) -> Result<Self, Self::Error> {
+        let expires = pb.expires.ok_or(ProtocolError::Missing("expires"))?;
+        UNIX_EPOCH
+            .checked_add(Duration::new(expires.seconds as u64, expires.nanos as u32))
+            .ok_or_else(|| ProtocolError::invalid("expires", expires))
+    }
+}
+
+impl From<LeaseLsnResponse> for proto::LeaseLsnResponse {
+    fn from(response: LeaseLsnResponse) -> Self {
+        let expires = response.duration_since(UNIX_EPOCH).unwrap_or_default();
+        Self {
+            expires: Some(prost_types::Timestamp {
+                seconds: expires.as_secs() as i64,
+                nanos: expires.subsec_nanos() as i32,
+            }),
+        }
+    }
+}

pageserver/pagebench/src/cmd/basebackup.rs

@@ -317,6 +317,7 @@ impl Client for LibpqClient {
 /// A gRPC Pageserver client.
 struct GrpcClient {
     inner: page_api::Client,
+    compression: page_api::BaseBackupCompression,
 }
 
 impl GrpcClient {
@@ -331,10 +332,14 @@ impl GrpcClient {
             ttid.timeline_id,
             ShardIndex::unsharded(),
             None,
-            compression.then_some(tonic::codec::CompressionEncoding::Zstd),
+            None, // NB: uses payload compression
         )
         .await?;
-        Ok(Self { inner })
+        let compression = match compression {
+            true => page_api::BaseBackupCompression::Gzip,
+            false => page_api::BaseBackupCompression::None,
+        };
+        Ok(Self { inner, compression })
     }
 }
 
@@ -348,10 +353,8 @@ impl Client for GrpcClient {
             lsn,
             replica: false,
             full: false,
+            compression: self.compression,
         };
-        let stream = self.inner.get_base_backup(req).await?;
-        Ok(Box::pin(StreamReader::new(
-            stream.map_err(std::io::Error::other),
-        )))
+        Ok(Box::pin(self.inner.get_base_backup(req).await?))
     }
 }

pageserver/src/basebackup.rs

@@ -14,6 +14,7 @@ use std::fmt::Write as FmtWrite;
 use std::time::{Instant, SystemTime};
 
 use anyhow::{Context, anyhow};
+use async_compression::tokio::write::GzipEncoder;
 use bytes::{BufMut, Bytes, BytesMut};
 use fail::fail_point;
 use pageserver_api::key::{Key, rel_block_to_key};
@@ -25,8 +26,7 @@ use postgres_ffi::{
 };
 use postgres_ffi_types::constants::{DEFAULTTABLESPACE_OID, GLOBALTABLESPACE_OID};
 use postgres_ffi_types::forknum::{INIT_FORKNUM, MAIN_FORKNUM};
-use tokio::io;
-use tokio::io::AsyncWrite;
+use tokio::io::{self, AsyncWrite, AsyncWriteExt as _};
 use tokio_tar::{Builder, EntryType, Header};
 use tracing::*;
 use utils::lsn::Lsn;
@@ -97,6 +97,7 @@ impl From<BasebackupError> for tonic::Status {
 ///  * When working without safekeepers. In this situation it is important to match the lsn
 ///    we are taking basebackup on with the lsn that is used in pageserver's walreceiver
 ///    to start the replication.
+#[allow(clippy::too_many_arguments)]
 pub async fn send_basebackup_tarball<'a, W>(
     write: &'a mut W,
     timeline: &'a Timeline,
@@ -104,6 +105,7 @@ pub async fn send_basebackup_tarball<'a, W>(
     prev_lsn: Option<Lsn>,
     full_backup: bool,
     replica: bool,
+    gzip_level: Option<async_compression::Level>,
     ctx: &'a RequestContext,
 ) -> Result<(), BasebackupError>
 where
@@ -122,7 +124,7 @@ where
     // prev_lsn value; that happens if the timeline was just branched from
     // an old LSN and it doesn't have any WAL of its own yet. We will set
     // prev_lsn to Lsn(0) if we cannot provide the correct value.
-    let (backup_prev, backup_lsn) = if let Some(req_lsn) = req_lsn {
+    let (backup_prev, lsn) = if let Some(req_lsn) = req_lsn {
         // Backup was requested at a particular LSN. The caller should've
         // already checked that it's a valid LSN.
 
@@ -143,7 +145,7 @@ where
     };
 
     // Consolidate the derived and the provided prev_lsn values
-    let prev_lsn = if let Some(provided_prev_lsn) = prev_lsn {
+    let prev_record_lsn = if let Some(provided_prev_lsn) = prev_lsn {
         if backup_prev != Lsn(0) && backup_prev != provided_prev_lsn {
             return Err(BasebackupError::Server(anyhow!(
                 "backup_prev {backup_prev} != provided_prev_lsn {provided_prev_lsn}"
@@ -155,30 +157,55 @@ where
     };
 
     info!(
-        "taking basebackup lsn={}, prev_lsn={} (full_backup={}, replica={})",
-        backup_lsn, prev_lsn, full_backup, replica
+        "taking basebackup lsn={lsn}, prev_lsn={prev_record_lsn} \
+        (full_backup={full_backup}, replica={replica}, gzip={gzip_level:?})",
+    );
+    let span = info_span!("send_tarball", backup_lsn=%lsn);
+
+    let io_concurrency = IoConcurrency::spawn_from_conf(
+        timeline.conf.get_vectored_concurrent_io,
+        timeline
+            .gate
+            .enter()
+            .map_err(|_| BasebackupError::Shutdown)?,
     );
 
-    let basebackup = Basebackup {
-        ar: Builder::new_non_terminated(write),
-        timeline,
-        lsn: backup_lsn,
-        prev_record_lsn: prev_lsn,
-        full_backup,
-        replica,
-        ctx,
-        io_concurrency: IoConcurrency::spawn_from_conf(
-            timeline.conf.get_vectored_concurrent_io,
-            timeline
-                .gate
-                .enter()
-                .map_err(|_| BasebackupError::Shutdown)?,
-        ),
-    };
-    basebackup
+    if let Some(gzip_level) = gzip_level {
+        let mut encoder = GzipEncoder::with_quality(write, gzip_level);
+        Basebackup {
+            ar: Builder::new_non_terminated(&mut encoder),
+            timeline,
+            lsn,
+            prev_record_lsn,
+            full_backup,
+            replica,
+            ctx,
+            io_concurrency,
+        }
         .send_tarball()
-        .instrument(info_span!("send_tarball", backup_lsn=%backup_lsn))
-        .await
+        .instrument(span)
+        .await?;
+        encoder
+            .shutdown()
+            .await
+            .map_err(|err| BasebackupError::Client(err, "gzip"))?;
+    } else {
+        Basebackup {
+            ar: Builder::new_non_terminated(write),
+            timeline,
+            lsn,
+            prev_record_lsn,
+            full_backup,
+            replica,
+            ctx,
+            io_concurrency,
+        }
+        .send_tarball()
+        .instrument(span)
+        .await?;
+    }
+
+    Ok(())
 }
 
 /// This is short-living object only for the time of tarball creation,

pageserver/src/basebackup_cache.rs

@@ -1,13 +1,12 @@
 use std::{collections::HashMap, sync::Arc};
 
 use anyhow::Context;
-use async_compression::tokio::write::GzipEncoder;
 use camino::{Utf8Path, Utf8PathBuf};
 use metrics::core::{AtomicU64, GenericCounter};
 use pageserver_api::{config::BasebackupCacheConfig, models::TenantState};
 use tokio::{
     io::{AsyncWriteExt, BufWriter},
-    sync::mpsc::{UnboundedReceiver, UnboundedSender},
+    sync::mpsc::{Receiver, Sender, error::TrySendError},
 };
 use tokio_util::sync::CancellationToken;
 use utils::{
@@ -20,8 +19,8 @@ use crate::{
     basebackup::send_basebackup_tarball,
     context::{DownloadBehavior, RequestContext},
     metrics::{
-        BASEBACKUP_CACHE_ENTRIES, BASEBACKUP_CACHE_PREPARE, BASEBACKUP_CACHE_READ,
-        BASEBACKUP_CACHE_SIZE,
+        BASEBACKUP_CACHE_ENTRIES, BASEBACKUP_CACHE_PREPARE, BASEBACKUP_CACHE_PREPARE_QUEUE_SIZE,
+        BASEBACKUP_CACHE_READ, BASEBACKUP_CACHE_SIZE,
     },
     task_mgr::TaskKind,
     tenant::{
@@ -36,8 +35,8 @@ pub struct BasebackupPrepareRequest {
     pub lsn: Lsn,
 }
 
-pub type BasebackupPrepareSender = UnboundedSender<BasebackupPrepareRequest>;
-pub type BasebackupPrepareReceiver = UnboundedReceiver<BasebackupPrepareRequest>;
+pub type BasebackupPrepareSender = Sender<BasebackupPrepareRequest>;
+pub type BasebackupPrepareReceiver = Receiver<BasebackupPrepareRequest>;
 
 #[derive(Clone)]
 struct CacheEntry {
@@ -61,40 +60,65 @@ struct CacheEntry {
 /// and ~1 RPS for get requests.
 pub struct BasebackupCache {
     data_dir: Utf8PathBuf,
+    config: Option<BasebackupCacheConfig>,
 
     entries: std::sync::Mutex<HashMap<TenantTimelineId, CacheEntry>>,
 
+    prepare_sender: BasebackupPrepareSender,
+
     read_hit_count: GenericCounter<AtomicU64>,
     read_miss_count: GenericCounter<AtomicU64>,
     read_err_count: GenericCounter<AtomicU64>,
+
+    prepare_skip_count: GenericCounter<AtomicU64>,
 }
 
 impl BasebackupCache {
-    /// Creates a BasebackupCache and spawns the background task.
-    /// The initialization of the cache is performed in the background and does not
-    /// block the caller. The cache will return `None` for any get requests until
-    /// initialization is complete.
-    pub fn spawn(
-        runtime_handle: &tokio::runtime::Handle,
+    /// Create a new BasebackupCache instance.
+    /// Also returns a BasebackupPrepareReceiver which is needed to start
+    /// the background task.
+    /// The cache is initialized from the data_dir in the background task.
+    /// The cache will return `None` for any get requests until the initialization is complete.
+    /// The background task is spawned separately using [`Self::spawn_background_task`]
+    /// to avoid a circular dependency between the cache and the tenant manager.
+    pub fn new(
         data_dir: Utf8PathBuf,
         config: Option<BasebackupCacheConfig>,
-        prepare_receiver: BasebackupPrepareReceiver,
-        tenant_manager: Arc<TenantManager>,
-        cancel: CancellationToken,
-    ) -> Arc<Self> {
+    ) -> (Arc<Self>, BasebackupPrepareReceiver) {
+        let chan_size = config.as_ref().map(|c| c.max_size_entries).unwrap_or(1);
+
+        let (prepare_sender, prepare_receiver) = tokio::sync::mpsc::channel(chan_size);
+
         let cache = Arc::new(BasebackupCache {
             data_dir,
-
+            config,
             entries: std::sync::Mutex::new(HashMap::new()),
+            prepare_sender,
 
             read_hit_count: BASEBACKUP_CACHE_READ.with_label_values(&["hit"]),
             read_miss_count: BASEBACKUP_CACHE_READ.with_label_values(&["miss"]),
             read_err_count: BASEBACKUP_CACHE_READ.with_label_values(&["error"]),
+
+            prepare_skip_count: BASEBACKUP_CACHE_PREPARE.with_label_values(&["skip"]),
         });
 
-        if let Some(config) = config {
+        (cache, prepare_receiver)
+    }
+
+    /// Spawns the background task.
+    /// The background task initializes the cache from the disk,
+    /// processes prepare requests, and cleans up outdated cache entries.
+    /// Noop if the cache is disabled (config is None).
+    pub fn spawn_background_task(
+        self: Arc<Self>,
+        runtime_handle: &tokio::runtime::Handle,
+        prepare_receiver: BasebackupPrepareReceiver,
+        tenant_manager: Arc<TenantManager>,
+        cancel: CancellationToken,
+    ) {
+        if let Some(config) = self.config.clone() {
             let background = BackgroundTask {
-                c: cache.clone(),
+                c: self,
 
                 config,
                 tenant_manager,
@@ -109,8 +133,45 @@ impl BasebackupCache {
             };
             runtime_handle.spawn(background.run(prepare_receiver));
         }
+    }
 
-        cache
+    /// Send a basebackup prepare request to the background task.
+    /// The basebackup will be prepared asynchronously, it does not block the caller.
+    /// The request will be skipped if any cache limits are exceeded.
+    pub fn send_prepare(&self, tenant_shard_id: TenantShardId, timeline_id: TimelineId, lsn: Lsn) {
+        let req = BasebackupPrepareRequest {
+            tenant_shard_id,
+            timeline_id,
+            lsn,
+        };
+
+        BASEBACKUP_CACHE_PREPARE_QUEUE_SIZE.inc();
+        let res = self.prepare_sender.try_send(req);
+
+        if let Err(e) = res {
+            BASEBACKUP_CACHE_PREPARE_QUEUE_SIZE.dec();
+            self.prepare_skip_count.inc();
+            match e {
+                TrySendError::Full(_) => {
+                    // Basebackup prepares are pretty rare, normally we should not hit this.
+                    tracing::info!(
+                        tenant_id = %tenant_shard_id.tenant_id,
+                        %timeline_id,
+                        %lsn,
+                        "Basebackup prepare channel is full, skipping the request"
+                    );
+                }
+                TrySendError::Closed(_) => {
+                    // Normal during shutdown, not critical.
+                    tracing::info!(
+                        tenant_id = %tenant_shard_id.tenant_id,
+                        %timeline_id,
+                        %lsn,
+                        "Basebackup prepare channel is closed, skipping the request"
+                    );
+                }
+            }
+        }
     }
 
     /// Gets a basebackup entry from the cache.
@@ -123,6 +184,10 @@ impl BasebackupCache {
         timeline_id: TimelineId,
         lsn: Lsn,
     ) -> Option<tokio::fs::File> {
+        if !self.is_enabled() {
+            return None;
+        }
+
         // Fast path. Check if the entry exists using the in-memory state.
         let tti = TenantTimelineId::new(tenant_id, timeline_id);
         if self.entries.lock().unwrap().get(&tti).map(|e| e.lsn) != Some(lsn) {
@@ -150,6 +215,10 @@ impl BasebackupCache {
         }
     }
 
+    pub fn is_enabled(&self) -> bool {
+        self.config.is_some()
+    }
+
     // Private methods.
 
     fn entry_filename(tenant_id: TenantId, timeline_id: TimelineId, lsn: Lsn) -> String {
@@ -367,6 +436,7 @@ impl BackgroundTask {
         loop {
             tokio::select! {
                 Some(req) = prepare_receiver.recv() => {
+                    BASEBACKUP_CACHE_PREPARE_QUEUE_SIZE.dec();
                     if let Err(err) = self.prepare_basebackup(
                         req.tenant_shard_id,
                         req.timeline_id,
@@ -594,13 +664,6 @@ impl BackgroundTask {
         let file = tokio::fs::File::create(entry_tmp_path).await?;
         let mut writer = BufWriter::new(file);
 
-        let mut encoder = GzipEncoder::with_quality(
-            &mut writer,
-            // Level::Best because compression is not on the hot path of basebackup requests.
-            // The decompression is almost not affected by the compression level.
-            async_compression::Level::Best,
-        );
-
         // We may receive a request before the WAL record is applied to the timeline.
         // Wait for the requested LSN to be applied.
         timeline
@@ -613,17 +676,19 @@ impl BackgroundTask {
             .await?;
 
         send_basebackup_tarball(
-            &mut encoder,
+            &mut writer,
             timeline,
             Some(req_lsn),
             None,
             false,
             false,
+            // Level::Best because compression is not on the hot path of basebackup requests.
+            // The decompression is almost not affected by the compression level.
+            Some(async_compression::Level::Best),
             &ctx,
         )
         .await?;
 
-        encoder.shutdown().await?;
         writer.flush().await?;
         writer.into_inner().sync_all().await?;
 

pageserver/src/bin/pageserver.rs

@@ -569,8 +569,10 @@ fn start_pageserver(
         pageserver::l0_flush::L0FlushGlobalState::new(conf.l0_flush.clone());
 
     // Scan the local 'tenants/' directory and start loading the tenants
-    let (basebackup_prepare_sender, basebackup_prepare_receiver) =
-        tokio::sync::mpsc::unbounded_channel();
+    let (basebackup_cache, basebackup_prepare_receiver) = BasebackupCache::new(
+        conf.basebackup_cache_dir(),
+        conf.basebackup_cache_config.clone(),
+    );
     let deletion_queue_client = deletion_queue.new_client();
     let background_purges = mgr::BackgroundPurges::default();
 
@@ -582,7 +584,7 @@ fn start_pageserver(
             remote_storage: remote_storage.clone(),
             deletion_queue_client,
             l0_flush_global_state,
-            basebackup_prepare_sender,
+            basebackup_cache: Arc::clone(&basebackup_cache),
             feature_resolver: feature_resolver.clone(),
         },
         shutdown_pageserver.clone(),
@@ -590,10 +592,8 @@ fn start_pageserver(
     let tenant_manager = Arc::new(tenant_manager);
     BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(tenant_manager.clone(), order))?;
 
-    let basebackup_cache = BasebackupCache::spawn(
+    basebackup_cache.spawn_background_task(
         BACKGROUND_RUNTIME.handle(),
-        conf.basebackup_cache_dir(),
-        conf.basebackup_cache_config.clone(),
         basebackup_prepare_receiver,
         Arc::clone(&tenant_manager),
         shutdown_pageserver.child_token(),
@@ -806,7 +806,6 @@ fn start_pageserver(
         } else {
             None
         },
-        basebackup_cache,
     );
 
     // Spawn a Pageserver gRPC server task. It will spawn separate tasks for

pageserver/src/bin/test_helper_slow_client_reads.rs

@@ -37,7 +37,7 @@ async fn main() -> anyhow::Result<()> {
                 not_modified_since: Lsn(23),
             },
             batch_key: 42,
-            message: format!("message {}", msg),
+            message: format!("message {msg}"),
         }));
         let Ok(res) = tokio::time::timeout(Duration::from_secs(10), fut).await else {
             eprintln!("pipe seems full");

pageserver/src/config.rs

@@ -28,6 +28,7 @@ use reqwest::Url;
 use storage_broker::Uri;
 use utils::id::{NodeId, TimelineId};
 use utils::logging::{LogFormat, SecretString};
+use utils::serde_percent::Percent;
 
 use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
 use crate::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
@@ -459,7 +460,16 @@ impl PageServerConf {
             metric_collection_endpoint,
             metric_collection_bucket,
             synthetic_size_calculation_interval,
-            disk_usage_based_eviction,
+            disk_usage_based_eviction: Some(disk_usage_based_eviction.unwrap_or(
+                DiskUsageEvictionTaskConfig {
+                    max_usage_pct: Percent::new(80).unwrap(),
+                    min_avail_bytes: 2_000_000_000,
+                    period: Duration::from_secs(60),
+                    #[cfg(feature = "testing")]
+                    mock_statvfs: None,
+                    eviction_order: Default::default(),
+                },
+            )),
             test_remote_failures,
             ondemand_download_behavior_treat_error_as_warn,
             background_task_maximum_delay,
@@ -697,6 +707,8 @@ impl ConfigurableSemaphore {
 #[cfg(test)]
 mod tests {
 
+    use std::time::Duration;
+
     use camino::Utf8PathBuf;
     use rstest::rstest;
     use utils::id::NodeId;
@@ -762,4 +774,56 @@ mod tests {
         let result = PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir);
         assert_eq!(result.is_ok(), is_valid);
     }
+
+    #[test]
+    fn test_config_posthog_config_is_valid() {
+        let input = r#"
+            control_plane_api = "http://localhost:6666"
+
+            [posthog_config]
+            server_api_key = "phs_AAA"
+            client_api_key = "phc_BBB"
+            project_id = "000"
+            private_api_url = "https://us.posthog.com"
+            public_api_url = "https://us.i.posthog.com"
+        "#;
+        let config_toml = toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(input)
+            .expect("posthogconfig is valid");
+        let workdir = Utf8PathBuf::from("/nonexistent");
+        PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir)
+            .expect("parse_and_validate");
+    }
+
+    #[test]
+    fn test_config_posthog_incomplete_config_is_valid() {
+        let input = r#"
+            control_plane_api = "http://localhost:6666"
+
+            [posthog_config]
+            server_api_key = "phs_AAA"
+            private_api_url = "https://us.posthog.com"
+            public_api_url = "https://us.i.posthog.com"
+        "#;
+        let config_toml = toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(input)
+            .expect("posthogconfig is valid");
+        let workdir = Utf8PathBuf::from("/nonexistent");
+        PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir)
+            .expect("parse_and_validate");
+    }
+
+    #[test]
+    fn test_config_disk_usage_based_eviction_is_valid() {
+        let input = r#"
+            control_plane_api = "http://localhost:6666"
+        "#;
+        let config_toml = toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(input)
+            .expect("disk_usage_based_eviction is valid");
+        let workdir = Utf8PathBuf::from("/nonexistent");
+        let config = PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir).unwrap();
+        let disk_usage_based_eviction = config.disk_usage_based_eviction.unwrap();
+        assert_eq!(disk_usage_based_eviction.max_usage_pct.get(), 80);
+        assert_eq!(disk_usage_based_eviction.min_avail_bytes, 2_000_000_000);
+        assert_eq!(disk_usage_based_eviction.period, Duration::from_secs(60));
+        assert_eq!(disk_usage_based_eviction.eviction_order, Default::default());
+    }
 }

pageserver/src/consumption_metrics/upload.rs

@@ -99,7 +99,7 @@ pub(super) async fn upload_metrics_bucket(
 
     // Compose object path
     let datetime: DateTime<Utc> = SystemTime::now().into();
-    let ts_prefix = datetime.format("year=%Y/month=%m/day=%d/%H:%M:%SZ");
+    let ts_prefix = datetime.format("year=%Y/month=%m/day=%d/hour=%H/%H:%M:%SZ");
     let path = RemotePath::from_string(&format!("{ts_prefix}_{node_id}.ndjson.gz"))?;
 
     // Set up a gzip writer into a buffer
@@ -109,7 +109,7 @@ pub(super) async fn upload_metrics_bucket(
 
     // Serialize and write into compressed buffer
     let started_at = std::time::Instant::now();
-    for res in serialize_in_chunks(CHUNK_SIZE, metrics, idempotency_keys) {
+    for res in serialize_in_chunks_ndjson(CHUNK_SIZE, metrics, idempotency_keys) {
         let (_chunk, body) = res?;
         gzip_writer.write_all(&body).await?;
     }
@@ -216,6 +216,86 @@ fn serialize_in_chunks<'a>(
     }
 }
 
+/// Serializes the input metrics as NDJSON in chunks of chunk_size. Each event
+/// is serialized as a separate JSON object on its own line. The provided
+/// idempotency keys are injected into the corresponding metric events (reused
+/// across different metrics sinks), and must have the same length as input.
+fn serialize_in_chunks_ndjson<'a>(
+    chunk_size: usize,
+    input: &'a [NewRawMetric],
+    idempotency_keys: &'a [IdempotencyKey<'a>],
+) -> impl ExactSizeIterator<Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>> + 'a
+{
+    use bytes::BufMut;
+
+    assert_eq!(input.len(), idempotency_keys.len());
+
+    struct Iter<'a> {
+        inner: std::slice::Chunks<'a, NewRawMetric>,
+        idempotency_keys: std::slice::Iter<'a, IdempotencyKey<'a>>,
+        chunk_size: usize,
+
+        // write to a BytesMut so that we can cheaply clone the frozen Bytes for retries
+        buffer: bytes::BytesMut,
+        // chunk amount of events are reused to produce the serialized document
+        scratch: Vec<Event<Ids, Name>>,
+    }
+
+    impl<'a> Iterator for Iter<'a> {
+        type Item = Result<(&'a [NewRawMetric], bytes::Bytes), serde_json::Error>;
+
+        fn next(&mut self) -> Option<Self::Item> {
+            let chunk = self.inner.next()?;
+
+            if self.scratch.is_empty() {
+                // first round: create events with N strings
+                self.scratch.extend(
+                    chunk
+                        .iter()
+                        .zip(&mut self.idempotency_keys)
+                        .map(|(raw_metric, key)| raw_metric.as_event(key)),
+                );
+            } else {
+                // next rounds: update_in_place to reuse allocations
+                assert_eq!(self.scratch.len(), self.chunk_size);
+                itertools::izip!(self.scratch.iter_mut(), chunk, &mut self.idempotency_keys)
+                    .for_each(|(slot, raw_metric, key)| raw_metric.update_in_place(slot, key));
+            }
+
+            // Serialize each event as NDJSON (one JSON object per line)
+            for event in self.scratch[..chunk.len()].iter() {
+                let res = serde_json::to_writer((&mut self.buffer).writer(), event);
+                if let Err(e) = res {
+                    return Some(Err(e));
+                }
+                // Add newline after each event to follow NDJSON format
+                self.buffer.put_u8(b'\n');
+            }
+
+            Some(Ok((chunk, self.buffer.split().freeze())))
+        }
+
+        fn size_hint(&self) -> (usize, Option<usize>) {
+            self.inner.size_hint()
+        }
+    }
+
+    impl ExactSizeIterator for Iter<'_> {}
+
+    let buffer = bytes::BytesMut::new();
+    let inner = input.chunks(chunk_size);
+    let idempotency_keys = idempotency_keys.iter();
+    let scratch = Vec::new();
+
+    Iter {
+        inner,
+        idempotency_keys,
+        chunk_size,
+        buffer,
+        scratch,
+    }
+}
+
 trait RawMetricExt {
     fn as_event(&self, key: &IdempotencyKey<'_>) -> Event<Ids, Name>;
     fn update_in_place(&self, event: &mut Event<Ids, Name>, key: &IdempotencyKey<'_>);
@@ -479,6 +559,43 @@ mod tests {
         }
     }
 
+    #[test]
+    fn chunked_serialization_ndjson() {
+        let examples = metric_samples();
+        assert!(examples.len() > 1);
+
+        let now = Utc::now();
+        let idempotency_keys = (0..examples.len())
+            .map(|i| FixedGen::new(now, "1", i as u16).generate())
+            .collect::<Vec<_>>();
+
+        // Parse NDJSON format - each line is a separate JSON object
+        let parse_ndjson = |body: &[u8]| -> Vec<Event<Ids, Name>> {
+            let body_str = std::str::from_utf8(body).unwrap();
+            body_str
+                .trim_end_matches('\n')
+                .lines()
+                .filter(|line| !line.is_empty())
+                .map(|line| serde_json::from_str::<Event<Ids, Name>>(line).unwrap())
+                .collect()
+        };
+
+        let correct = serialize_in_chunks_ndjson(examples.len(), &examples, &idempotency_keys)
+            .map(|res| res.unwrap().1)
+            .flat_map(|body| parse_ndjson(&body))
+            .collect::<Vec<_>>();
+
+        for chunk_size in 1..examples.len() {
+            let actual = serialize_in_chunks_ndjson(chunk_size, &examples, &idempotency_keys)
+                .map(|res| res.unwrap().1)
+                .flat_map(|body| parse_ndjson(&body))
+                .collect::<Vec<_>>();
+
+            // if these are equal, it means that multi-chunking version works as well
+            assert_eq!(correct, actual);
+        }
+    }
+
     #[derive(Clone, Copy)]
     struct FixedGen<'a>(chrono::DateTime<chrono::Utc>, &'a str, u16);
 

pageserver/src/feature_resolver.rs

@@ -3,15 +3,16 @@ use std::{collections::HashMap, sync::Arc, time::Duration};
 use arc_swap::ArcSwap;
 use pageserver_api::config::NodeMetadata;
 use posthog_client_lite::{
-    CaptureEvent, FeatureResolverBackgroundLoop, PostHogClientConfig, PostHogEvaluationError,
+    CaptureEvent, FeatureResolverBackgroundLoop, PostHogEvaluationError,
     PostHogFlagFilterPropertyValue,
 };
+use rand::Rng;
 use remote_storage::RemoteStorageKind;
 use serde_json::json;
 use tokio_util::sync::CancellationToken;
 use utils::id::TenantId;
 
-use crate::{config::PageServerConf, metrics::FEATURE_FLAG_EVALUATION};
+use crate::{config::PageServerConf, metrics::FEATURE_FLAG_EVALUATION, tenant::TenantShard};
 
 const DEFAULT_POSTHOG_REFRESH_INTERVAL: Duration = Duration::from_secs(600);
 
@@ -31,6 +32,13 @@ impl FeatureResolver {
         }
     }
 
+    pub fn update(&self, spec: String) -> anyhow::Result<()> {
+        if let Some(inner) = &self.inner {
+            inner.update(spec)?;
+        }
+        Ok(())
+    }
+
     pub fn spawn(
         conf: &PageServerConf,
         shutdown_pageserver: CancellationToken,
@@ -38,16 +46,24 @@ impl FeatureResolver {
     ) -> anyhow::Result<Self> {
         // DO NOT block in this function: make it return as fast as possible to avoid startup delays.
         if let Some(posthog_config) = &conf.posthog_config {
-            let inner = FeatureResolverBackgroundLoop::new(
-                PostHogClientConfig {
-                    server_api_key: posthog_config.server_api_key.clone(),
-                    client_api_key: posthog_config.client_api_key.clone(),
-                    project_id: posthog_config.project_id.clone(),
-                    private_api_url: posthog_config.private_api_url.clone(),
-                    public_api_url: posthog_config.public_api_url.clone(),
-                },
-                shutdown_pageserver,
-            );
+            let posthog_client_config = match posthog_config.clone().try_into_posthog_config() {
+                Ok(config) => config,
+                Err(e) => {
+                    tracing::warn!(
+                        "invalid posthog config, skipping posthog integration: {}",
+                        e
+                    );
+                    return Ok(FeatureResolver {
+                        inner: None,
+                        internal_properties: None,
+                        force_overrides_for_testing: Arc::new(ArcSwap::new(Arc::new(
+                            HashMap::new(),
+                        ))),
+                    });
+                }
+            };
+            let inner =
+                FeatureResolverBackgroundLoop::new(posthog_client_config, shutdown_pageserver);
             let inner = Arc::new(inner);
 
             // The properties shared by all tenants on this pageserver.
@@ -123,6 +139,7 @@ impl FeatureResolver {
                 }
                 Arc::new(properties)
             };
+
             let fake_tenants = {
                 let mut tenants = Vec::new();
                 for i in 0..10 {
@@ -132,9 +149,16 @@ impl FeatureResolver {
                         conf.id,
                         i
                     );
+
+                    let tenant_properties = PerTenantProperties {
+                        remote_size_mb: Some(rand::thread_rng().gen_range(100.0..1000000.00)),
+                    }
+                    .into_posthog_properties();
+
                     let properties = Self::collect_properties_inner(
                         distinct_id.clone(),
                         Some(&internal_properties),
+                        &tenant_properties,
                     );
                     tenants.push(CaptureEvent {
                         event: "initial_tenant_report".to_string(),
@@ -168,6 +192,7 @@ impl FeatureResolver {
     fn collect_properties_inner(
         tenant_id: String,
         internal_properties: Option<&HashMap<String, PostHogFlagFilterPropertyValue>>,
+        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
         let mut properties = HashMap::new();
         if let Some(internal_properties) = internal_properties {
@@ -179,6 +204,9 @@ impl FeatureResolver {
             "tenant_id".to_string(),
             PostHogFlagFilterPropertyValue::String(tenant_id),
         );
+        for (key, value) in tenant_properties.iter() {
+            properties.insert(key.clone(), value.clone());
+        }
         properties
     }
 
@@ -186,8 +214,13 @@ impl FeatureResolver {
     pub(crate) fn collect_properties(
         &self,
         tenant_id: TenantId,
+        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
-        Self::collect_properties_inner(tenant_id.to_string(), self.internal_properties.as_deref())
+        Self::collect_properties_inner(
+            tenant_id.to_string(),
+            self.internal_properties.as_deref(),
+            tenant_properties,
+        )
     }
 
     /// Evaluate a multivariate feature flag. Currently, we do not support any properties.
@@ -199,6 +232,7 @@ impl FeatureResolver {
         &self,
         flag_key: &str,
         tenant_id: TenantId,
+        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> Result<String, PostHogEvaluationError> {
         let force_overrides = self.force_overrides_for_testing.load();
         if let Some(value) = force_overrides.get(flag_key) {
@@ -209,7 +243,7 @@ impl FeatureResolver {
             let res = inner.feature_store().evaluate_multivariate(
                 flag_key,
                 &tenant_id.to_string(),
-                &self.collect_properties(tenant_id),
+                &self.collect_properties(tenant_id, tenant_properties),
             );
             match &res {
                 Ok(value) => {
@@ -242,6 +276,7 @@ impl FeatureResolver {
         &self,
         flag_key: &str,
         tenant_id: TenantId,
+        tenant_properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> Result<(), PostHogEvaluationError> {
         let force_overrides = self.force_overrides_for_testing.load();
         if let Some(value) = force_overrides.get(flag_key) {
@@ -256,7 +291,7 @@ impl FeatureResolver {
             let res = inner.feature_store().evaluate_boolean(
                 flag_key,
                 &tenant_id.to_string(),
-                &self.collect_properties(tenant_id),
+                &self.collect_properties(tenant_id, tenant_properties),
             );
             match &res {
                 Ok(()) => {
@@ -302,3 +337,78 @@ impl FeatureResolver {
             .store(Arc::new(force_overrides));
     }
 }
+
+struct PerTenantProperties {
+    pub remote_size_mb: Option<f64>,
+}
+
+impl PerTenantProperties {
+    pub fn into_posthog_properties(self) -> HashMap<String, PostHogFlagFilterPropertyValue> {
+        let mut properties = HashMap::new();
+        if let Some(remote_size_mb) = self.remote_size_mb {
+            properties.insert(
+                "tenant_remote_size_mb".to_string(),
+                PostHogFlagFilterPropertyValue::Number(remote_size_mb),
+            );
+        }
+        properties
+    }
+}
+
+#[derive(Clone)]
+pub struct TenantFeatureResolver {
+    inner: FeatureResolver,
+    tenant_id: TenantId,
+    cached_tenant_properties: Arc<ArcSwap<HashMap<String, PostHogFlagFilterPropertyValue>>>,
+}
+
+impl TenantFeatureResolver {
+    pub fn new(inner: FeatureResolver, tenant_id: TenantId) -> Self {
+        Self {
+            inner,
+            tenant_id,
+            cached_tenant_properties: Arc::new(ArcSwap::new(Arc::new(HashMap::new()))),
+        }
+    }
+
+    pub fn evaluate_multivariate(&self, flag_key: &str) -> Result<String, PostHogEvaluationError> {
+        self.inner.evaluate_multivariate(
+            flag_key,
+            self.tenant_id,
+            &self.cached_tenant_properties.load(),
+        )
+    }
+
+    pub fn evaluate_boolean(&self, flag_key: &str) -> Result<(), PostHogEvaluationError> {
+        self.inner.evaluate_boolean(
+            flag_key,
+            self.tenant_id,
+            &self.cached_tenant_properties.load(),
+        )
+    }
+
+    pub fn collect_properties(&self) -> HashMap<String, PostHogFlagFilterPropertyValue> {
+        self.inner
+            .collect_properties(self.tenant_id, &self.cached_tenant_properties.load())
+    }
+
+    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
+        self.inner.is_feature_flag_boolean(flag_key)
+    }
+
+    pub fn update_cached_tenant_properties(&self, tenant_shard: &TenantShard) {
+        let mut remote_size_mb = None;
+        for timeline in tenant_shard.list_timelines() {
+            let size = timeline.metrics.resident_physical_size_get();
+            if size == 0 {
+                remote_size_mb = None;
+            }
+            if let Some(ref mut remote_size_mb) = remote_size_mb {
+                *remote_size_mb += size as f64 / 1024.0 / 1024.0;
+            }
+        }
+        self.cached_tenant_properties.store(Arc::new(
+            PerTenantProperties { remote_size_mb }.into_posthog_properties(),
+        ));
+    }
+}

pageserver/src/http/routes.rs

@@ -452,6 +452,7 @@ async fn build_timeline_info_common(
     let state = timeline.current_state();
     // Report is_archived = false if the timeline is still loading
     let is_archived = timeline.is_archived().unwrap_or(false);
+    let read_only = timeline.is_read_only().unwrap_or(false);
     let remote_consistent_lsn_projected = timeline
         .get_remote_consistent_lsn_projected()
         .unwrap_or(Lsn(0));
@@ -503,6 +504,7 @@ async fn build_timeline_info_common(
 
         state,
         is_archived: Some(is_archived),
+        read_only,
         rel_size_migration: Some(timeline.get_rel_size_v2_status()),
         is_invisible: Some(is_invisible),
 
@@ -591,12 +593,13 @@ async fn timeline_create_handler(
         TimelineCreateRequestMode::Branch {
             ancestor_timeline_id,
             ancestor_start_lsn,
-            read_only: _,
+            read_only,
             pg_version: _,
         } => tenant::CreateTimelineParams::Branch(tenant::CreateTimelineParamsBranch {
             new_timeline_id,
             ancestor_timeline_id,
             ancestor_start_lsn,
+            read_only,
         }),
         TimelineCreateRequestMode::ImportPgdata {
             import_pgdata:
@@ -1893,9 +1896,13 @@ async fn update_tenant_config_handler(
     let location_conf = LocationConf::attached_single(
         new_tenant_conf.clone(),
         tenant.get_generation(),
-        &ShardParameters::default(),
+        ShardParameters::from(tenant.get_shard_identity()),
     );
 
+    tenant
+        .get_shard_identity()
+        .assert_equal(location_conf.shard); // not strictly necessary since we construct it above
+
     crate::tenant::TenantShard::persist_tenant_config(state.conf, &tenant_shard_id, &location_conf)
         .await
         .map_err(|e| ApiError::InternalServerError(anyhow::anyhow!(e)))?;
@@ -1937,9 +1944,13 @@ async fn patch_tenant_config_handler(
     let location_conf = LocationConf::attached_single(
         updated,
         tenant.get_generation(),
-        &ShardParameters::default(),
+        ShardParameters::from(tenant.get_shard_identity()),
     );
 
+    tenant
+        .get_shard_identity()
+        .assert_equal(location_conf.shard); // not strictly necessary since we construct it above
+
     crate::tenant::TenantShard::persist_tenant_config(state.conf, &tenant_shard_id, &location_conf)
         .await
         .map_err(|e| ApiError::InternalServerError(anyhow::anyhow!(e)))?;
@@ -2430,6 +2441,7 @@ async fn timeline_offload_handler(
             .map_err(|e| {
                 match e {
                     OffloadError::Cancelled => ApiError::ResourceUnavailable("Timeline shutting down".into()),
+                    OffloadError::AlreadyInProgress => ApiError::Conflict("Timeline already being offloaded or deleted".into()),
                     _ => ApiError::InternalServerError(anyhow!(e))
                 }
             })?;
@@ -3689,23 +3701,25 @@ async fn tenant_evaluate_feature_flag(
         let tenant = state
             .tenant_manager
             .get_attached_tenant_shard(tenant_shard_id)?;
-        let properties = tenant.feature_resolver.collect_properties(tenant_shard_id.tenant_id);
+        // TODO: the properties we get here might be stale right after it is collected. But such races are rare (updated every 10s)
+        // and we don't need to worry about it for now.
+        let properties = tenant.feature_resolver.collect_properties();
         if as_type.as_deref() == Some("boolean") {
-            let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+            let result = tenant.feature_resolver.evaluate_boolean(&flag);
             let result = result.map(|_| true).map_err(|e| e.to_string());
             json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
         } else if as_type.as_deref() == Some("multivariate") {
-            let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+            let result = tenant.feature_resolver.evaluate_multivariate(&flag).map_err(|e| e.to_string());
             json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
         } else {
             // Auto infer the type of the feature flag.
             let is_boolean = tenant.feature_resolver.is_feature_flag_boolean(&flag).map_err(|e| ApiError::InternalServerError(anyhow::anyhow!("{e}")))?;
             if is_boolean {
-                let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+                let result = tenant.feature_resolver.evaluate_boolean(&flag);
                 let result = result.map(|_| true).map_err(|e| e.to_string());
                 json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
             } else {
-                let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+                let result = tenant.feature_resolver.evaluate_multivariate(&flag).map_err(|e| e.to_string());
                 json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
             }
         }
@@ -3743,6 +3757,20 @@ async fn force_override_feature_flag_for_testing_delete(
     json_response(StatusCode::OK, ())
 }
 
+async fn update_feature_flag_spec(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&request, None)?;
+    let body = json_request(&mut request).await?;
+    let state = get_state(&request);
+    state
+        .feature_resolver
+        .update(body)
+        .map_err(ApiError::InternalServerError)?;
+    json_response(StatusCode::OK, ())
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -4128,5 +4156,8 @@ pub fn make_router(
         .delete("/v1/feature_flag/:flag_key", |r| {
             testing_api_handler("force override feature flag - delete", r, force_override_feature_flag_for_testing_delete)
         })
+        .post("/v1/feature_flag_spec", |r| {
+            api_handler(r, update_feature_flag_spec)
+        })
         .any(handler_404))
 }

pageserver/src/metrics.rs

@@ -4439,6 +4439,14 @@ pub(crate) static BASEBACKUP_CACHE_SIZE: Lazy<UIntGauge> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+pub(crate) static BASEBACKUP_CACHE_PREPARE_QUEUE_SIZE: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_basebackup_cache_prepare_queue_size",
+        "Number of requests in the basebackup prepare channel"
+    )
+    .expect("failed to define a metric")
+});
+
 static PAGESERVER_CONFIG_IGNORED_ITEMS: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_config_ignored_items",

pageserver/src/page_service.rs

@@ -12,9 +12,9 @@ use std::task::{Context, Poll};
 use std::time::{Duration, Instant, SystemTime};
 use std::{io, str};
 
-use anyhow::{Context as _, anyhow, bail};
-use async_compression::tokio::write::GzipEncoder;
+use anyhow::{Context as _, bail};
 use bytes::{Buf as _, BufMut as _, BytesMut};
+use chrono::Utc;
 use futures::future::BoxFuture;
 use futures::{FutureExt, Stream};
 use itertools::Itertools;
@@ -63,7 +63,6 @@ use utils::{failpoint_support, span_record};
 
 use crate::auth::check_permission;
 use crate::basebackup::{self, BasebackupError};
-use crate::basebackup_cache::BasebackupCache;
 use crate::config::PageServerConf;
 use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
@@ -138,7 +137,6 @@ pub fn spawn(
     perf_trace_dispatch: Option<Dispatch>,
     tcp_listener: tokio::net::TcpListener,
     tls_config: Option<Arc<rustls::ServerConfig>>,
-    basebackup_cache: Arc<BasebackupCache>,
 ) -> Listener {
     let cancel = CancellationToken::new();
     let libpq_ctx = RequestContext::todo_child(
@@ -160,7 +158,6 @@ pub fn spawn(
             conf.pg_auth_type,
             tls_config,
             conf.page_service_pipelining.clone(),
-            basebackup_cache,
             libpq_ctx,
             cancel.clone(),
         )
@@ -219,7 +216,6 @@ pub async fn libpq_listener_main(
     auth_type: AuthType,
     tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
-    basebackup_cache: Arc<BasebackupCache>,
     listener_ctx: RequestContext,
     listener_cancel: CancellationToken,
 ) -> Connections {
@@ -263,7 +259,6 @@ pub async fn libpq_listener_main(
                     auth_type,
                     tls_config.clone(),
                     pipelining_config.clone(),
-                    Arc::clone(&basebackup_cache),
                     connection_ctx,
                     connections_cancel.child_token(),
                     gate_guard,
@@ -306,7 +301,6 @@ async fn page_service_conn_main(
     auth_type: AuthType,
     tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
-    basebackup_cache: Arc<BasebackupCache>,
     connection_ctx: RequestContext,
     cancel: CancellationToken,
     gate_guard: GateGuard,
@@ -372,7 +366,6 @@ async fn page_service_conn_main(
         pipelining_config,
         conf.get_vectored_concurrent_io,
         perf_span_fields,
-        basebackup_cache,
         connection_ctx,
         cancel.clone(),
         gate_guard,
@@ -426,8 +419,6 @@ struct PageServerHandler {
     pipelining_config: PageServicePipeliningConfig,
     get_vectored_concurrent_io: GetVectoredConcurrentIo,
 
-    basebackup_cache: Arc<BasebackupCache>,
-
     gate_guard: GateGuard,
 }
 
@@ -913,7 +904,6 @@ impl PageServerHandler {
         pipelining_config: PageServicePipeliningConfig,
         get_vectored_concurrent_io: GetVectoredConcurrentIo,
         perf_span_fields: ConnectionPerfSpanFields,
-        basebackup_cache: Arc<BasebackupCache>,
         connection_ctx: RequestContext,
         cancel: CancellationToken,
         gate_guard: GateGuard,
@@ -927,7 +917,6 @@ impl PageServerHandler {
             cancel,
             pipelining_config,
             get_vectored_concurrent_io,
-            basebackup_cache,
             gate_guard,
         }
     }
@@ -2613,26 +2602,16 @@ impl PageServerHandler {
                 prev_lsn,
                 full_backup,
                 replica,
+                None,
                 &ctx,
             )
             .await?;
         } else {
             let mut writer = BufWriter::new(pgb.copyout_writer());
 
-            let cached = {
-                // Basebackup is cached only for this combination of parameters.
-                if timeline.is_basebackup_cache_enabled()
-                    && gzip
-                    && lsn.is_some()
-                    && prev_lsn.is_none()
-                {
-                    self.basebackup_cache
-                        .get(tenant_id, timeline_id, lsn.unwrap())
-                        .await
-                } else {
-                    None
-                }
-            };
+            let cached = timeline
+                .get_cached_basebackup_if_enabled(lsn, prev_lsn, full_backup, replica, gzip)
+                .await;
 
             if let Some(mut cached) = cached {
                 from_cache = true;
@@ -2641,31 +2620,6 @@ impl PageServerHandler {
                     .map_err(|err| {
                         BasebackupError::Client(err, "handle_basebackup_request,cached,copy")
                     })?;
-            } else if gzip {
-                let mut encoder = GzipEncoder::with_quality(
-                    &mut writer,
-                    // NOTE using fast compression because it's on the critical path
-                    //      for compute startup. For an empty database, we get
-                    //      <100KB with this method. The Level::Best compression method
-                    //      gives us <20KB, but maybe we should add basebackup caching
-                    //      on compute shutdown first.
-                    async_compression::Level::Fastest,
-                );
-                basebackup::send_basebackup_tarball(
-                    &mut encoder,
-                    &timeline,
-                    lsn,
-                    prev_lsn,
-                    full_backup,
-                    replica,
-                    &ctx,
-                )
-                .await?;
-                // shutdown the encoder to ensure the gzip footer is written
-                encoder
-                    .shutdown()
-                    .await
-                    .map_err(|e| QueryError::Disconnected(ConnectionError::Io(e)))?;
             } else {
                 basebackup::send_basebackup_tarball(
                     &mut writer,
@@ -2674,6 +2628,11 @@ impl PageServerHandler {
                     prev_lsn,
                     full_backup,
                     replica,
+                    // NB: using fast compression because it's on the critical path for compute
+                    // startup. For an empty database, we get <100KB with this method. The
+                    // Level::Best compression method gives us <20KB, but maybe we should add
+                    // basebackup caching on compute shutdown first.
+                    gzip.then_some(async_compression::Level::Fastest),
                     &ctx,
                 )
                 .await?;
@@ -3553,7 +3512,7 @@ impl proto::PageService for GrpcPageServiceHandler {
         if timeline.is_archived() == Some(true) {
             return Err(tonic::Status::failed_precondition("timeline is archived"));
         }
-        let req: page_api::GetBaseBackupRequest = req.into_inner().into();
+        let req: page_api::GetBaseBackupRequest = req.into_inner().try_into()?;
 
         span_record!(lsn=?req.lsn);
 
@@ -3579,20 +3538,50 @@ impl proto::PageService for GrpcPageServiceHandler {
         let span = Span::current();
         let (mut simplex_read, mut simplex_write) = tokio::io::simplex(CHUNK_SIZE);
         let jh = tokio::spawn(async move {
-            let result = basebackup::send_basebackup_tarball(
-                &mut simplex_write,
-                &timeline,
-                req.lsn,
-                None,
-                req.full,
-                req.replica,
-                &ctx,
-            )
-            .instrument(span) // propagate request span
-            .await;
-            simplex_write.shutdown().await.map_err(|err| {
-                BasebackupError::Server(anyhow!("simplex shutdown failed: {err}"))
-            })?;
+            let gzip_level = match req.compression {
+                page_api::BaseBackupCompression::None => None,
+                // NB: using fast compression because it's on the critical path for compute
+                // startup. For an empty database, we get <100KB with this method. The
+                // Level::Best compression method gives us <20KB, but maybe we should add
+                // basebackup caching on compute shutdown first.
+                page_api::BaseBackupCompression::Gzip => Some(async_compression::Level::Fastest),
+            };
+
+            // Check for a cached basebackup.
+            let cached = timeline
+                .get_cached_basebackup_if_enabled(
+                    req.lsn,
+                    None,
+                    req.full,
+                    req.replica,
+                    gzip_level.is_some(),
+                )
+                .await;
+
+            let result = if let Some(mut cached) = cached {
+                // If we have a cached basebackup, send it.
+                tokio::io::copy(&mut cached, &mut simplex_write)
+                    .await
+                    .map(|_| ())
+                    .map_err(|err| BasebackupError::Client(err, "cached,copy"))
+            } else {
+                basebackup::send_basebackup_tarball(
+                    &mut simplex_write,
+                    &timeline,
+                    req.lsn,
+                    None,
+                    req.full,
+                    req.replica,
+                    gzip_level,
+                    &ctx,
+                )
+                .instrument(span) // propagate request span
+                .await
+            };
+            simplex_write
+                .shutdown()
+                .await
+                .map_err(|err| BasebackupError::Client(err, "simplex_write"))?;
             result
         });
 
@@ -3772,6 +3761,36 @@ impl proto::PageService for GrpcPageServiceHandler {
         let resp: page_api::GetSlruSegmentResponse = resp.segment;
         Ok(tonic::Response::new(resp.into()))
     }
+
+    #[instrument(skip_all, fields(lsn))]
+    async fn lease_lsn(
+        &self,
+        req: tonic::Request<proto::LeaseLsnRequest>,
+    ) -> Result<tonic::Response<proto::LeaseLsnResponse>, tonic::Status> {
+        let timeline = self.get_request_timeline(&req).await?;
+        let ctx = self.ctx.with_scope_timeline(&timeline);
+
+        // Validate and convert the request, and decorate the span.
+        let req: page_api::LeaseLsnRequest = req.into_inner().try_into()?;
+
+        span_record!(lsn=%req.lsn);
+
+        // Attempt to acquire a lease. Return FailedPrecondition if the lease could not be granted.
+        let lease_length = timeline.get_lsn_lease_length();
+        let expires = match timeline.renew_lsn_lease(req.lsn, lease_length, &ctx) {
+            Ok(lease) => lease.valid_until,
+            Err(err) => return Err(tonic::Status::failed_precondition(format!("{err}"))),
+        };
+
+        // TODO: is this spammy? Move it compute-side?
+        info!(
+            "acquired lease for {} until {}",
+            req.lsn,
+            chrono::DateTime::<Utc>::from(expires).to_rfc3339()
+        );
+
+        Ok(tonic::Response::new(expires.into()))
+    }
 }
 
 /// gRPC middleware layer that handles observability concerns:

pageserver/src/pgdatadir_mapping.rs

@@ -3015,7 +3015,7 @@ mod tests {
         // This shard will get the even blocks
         let shard = ShardIdentity::from_params(
             ShardNumber(0),
-            &ShardParameters {
+            ShardParameters {
                 count: ShardCount(2),
                 stripe_size: ShardStripeSize(1),
             },

pageserver/src/tenant.rs

@@ -80,13 +80,13 @@ use self::timeline::uninit::{TimelineCreateGuard, TimelineExclusionError, Uninit
 use self::timeline::{
     EvictionTaskTenantState, GcCutoffs, TimelineDeleteProgress, TimelineResources, WaitLsnError,
 };
-use crate::basebackup_cache::BasebackupPrepareSender;
+use crate::basebackup_cache::BasebackupCache;
 use crate::config::PageServerConf;
 use crate::context;
 use crate::context::RequestContextBuilder;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
-use crate::feature_resolver::FeatureResolver;
+use crate::feature_resolver::{FeatureResolver, TenantFeatureResolver};
 use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::{
     BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
@@ -162,7 +162,7 @@ pub struct TenantSharedResources {
     pub remote_storage: GenericRemoteStorage,
     pub deletion_queue_client: DeletionQueueClient,
     pub l0_flush_global_state: L0FlushGlobalState,
-    pub basebackup_prepare_sender: BasebackupPrepareSender,
+    pub basebackup_cache: Arc<BasebackupCache>,
     pub feature_resolver: FeatureResolver,
 }
 
@@ -331,7 +331,7 @@ pub struct TenantShard {
     deletion_queue_client: DeletionQueueClient,
 
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
-    basebackup_prepare_sender: BasebackupPrepareSender,
+    basebackup_cache: Arc<BasebackupCache>,
 
     /// Cached logical sizes updated updated on each [`TenantShard::gather_size_inputs`].
     cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
@@ -386,7 +386,7 @@ pub struct TenantShard {
 
     l0_flush_global_state: L0FlushGlobalState,
 
-    pub(crate) feature_resolver: FeatureResolver,
+    pub(crate) feature_resolver: TenantFeatureResolver,
 }
 impl std::fmt::Debug for TenantShard {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -943,6 +943,7 @@ pub(crate) struct CreateTimelineParamsBranch {
     pub(crate) new_timeline_id: TimelineId,
     pub(crate) ancestor_timeline_id: TimelineId,
     pub(crate) ancestor_start_lsn: Option<Lsn>,
+    pub(crate) read_only: bool,
 }
 
 #[derive(Debug)]
@@ -1363,7 +1364,7 @@ impl TenantShard {
             remote_storage,
             deletion_queue_client,
             l0_flush_global_state,
-            basebackup_prepare_sender,
+            basebackup_cache,
             feature_resolver,
         } = resources;
 
@@ -1380,7 +1381,7 @@ impl TenantShard {
             remote_storage.clone(),
             deletion_queue_client,
             l0_flush_global_state,
-            basebackup_prepare_sender,
+            basebackup_cache,
             feature_resolver,
         ));
 
@@ -2572,6 +2573,7 @@ impl TenantShard {
             initdb_lsn,
             pg_version,
         );
+        let read_only = false;
         self.prepare_new_timeline(
             new_timeline_id,
             &new_metadata,
@@ -2579,6 +2581,7 @@ impl TenantShard {
             initdb_lsn,
             None,
             None,
+            read_only,
             ctx,
         )
         .await
@@ -2723,6 +2726,7 @@ impl TenantShard {
                 new_timeline_id,
                 ancestor_timeline_id,
                 mut ancestor_start_lsn,
+                read_only,
             }) => {
                 let ancestor_timeline = self
                     .get_timeline(ancestor_timeline_id, false)
@@ -2775,8 +2779,14 @@ impl TenantShard {
                         })?;
                 }
 
-                self.branch_timeline(&ancestor_timeline, new_timeline_id, ancestor_start_lsn, ctx)
-                    .await?
+                self.branch_timeline(
+                    &ancestor_timeline,
+                    new_timeline_id,
+                    ancestor_start_lsn,
+                    read_only,
+                    ctx,
+                )
+                .await?
             }
             CreateTimelineParams::ImportPgdata(params) => {
                 self.create_timeline_import_pgdata(params, ctx).await?
@@ -2901,6 +2911,7 @@ impl TenantShard {
                     initdb_lsn,
                     PgMajorVersion::PG15,
                 );
+                let read_only = false;
                 this.prepare_new_timeline(
                     new_timeline_id,
                     &new_metadata,
@@ -2908,6 +2919,7 @@ impl TenantShard {
                     initdb_lsn,
                     None,
                     None,
+                    read_only,
                     ctx,
                 )
                 .await
@@ -3263,7 +3275,7 @@ impl TenantShard {
                 };
                 let gc_compaction_strategy = self
                     .feature_resolver
-                    .evaluate_multivariate("gc-comapction-strategy", self.tenant_shard_id.tenant_id)
+                    .evaluate_multivariate("gc-comapction-strategy")
                     .ok();
                 let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
                     info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
@@ -3285,6 +3297,7 @@ impl TenantShard {
                     .or_else(|err| match err {
                         // Ignore this, we likely raced with unarchival.
                         OffloadError::NotArchived => Ok(()),
+                        OffloadError::AlreadyInProgress => Ok(()),
                         err => Err(err),
                     })?;
             }
@@ -3408,6 +3421,9 @@ impl TenantShard {
         if let Some(ref walredo_mgr) = self.walredo_mgr {
             walredo_mgr.maybe_quiesce(WALREDO_IDLE_TIMEOUT);
         }
+
+        // Update the feature resolver with the latest tenant-spcific data.
+        self.feature_resolver.update_cached_tenant_properties(self);
     }
 
     pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
@@ -3872,6 +3888,10 @@ impl TenantShard {
         &self.tenant_shard_id
     }
 
+    pub(crate) fn get_shard_identity(&self) -> ShardIdentity {
+        self.shard_identity
+    }
+
     pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
         self.shard_identity.stripe_size
     }
@@ -4380,7 +4400,7 @@ impl TenantShard {
         remote_storage: GenericRemoteStorage,
         deletion_queue_client: DeletionQueueClient,
         l0_flush_global_state: L0FlushGlobalState,
-        basebackup_prepare_sender: BasebackupPrepareSender,
+        basebackup_cache: Arc<BasebackupCache>,
         feature_resolver: FeatureResolver,
     ) -> TenantShard {
         assert!(!attached_conf.location.generation.is_none());
@@ -4485,8 +4505,11 @@ impl TenantShard {
             ongoing_timeline_detach: std::sync::Mutex::default(),
             gc_block: Default::default(),
             l0_flush_global_state,
-            basebackup_prepare_sender,
-            feature_resolver,
+            basebackup_cache,
+            feature_resolver: TenantFeatureResolver::new(
+                feature_resolver,
+                tenant_shard_id.tenant_id,
+            ),
         }
     }
 
@@ -4525,6 +4548,10 @@ impl TenantShard {
         Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
     }
 
+    /// Stores a tenant location config to disk.
+    ///
+    /// NB: make sure to call `ShardIdentity::assert_equal` before persisting a new config, to avoid
+    /// changes to shard parameters that may result in data corruption.
     #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
     pub(super) async fn persist_tenant_config(
         conf: &'static PageServerConf,
@@ -4942,9 +4969,10 @@ impl TenantShard {
         src_timeline: &Arc<Timeline>,
         dst_id: TimelineId,
         start_lsn: Option<Lsn>,
+        read_only: bool,
         ctx: &RequestContext,
     ) -> Result<CreateTimelineResult, CreateTimelineError> {
-        self.branch_timeline_impl(src_timeline, dst_id, start_lsn, ctx)
+        self.branch_timeline_impl(src_timeline, dst_id, start_lsn, read_only, ctx)
             .await
     }
 
@@ -4953,6 +4981,7 @@ impl TenantShard {
         src_timeline: &Arc<Timeline>,
         dst_id: TimelineId,
         start_lsn: Option<Lsn>,
+        read_only: bool,
         ctx: &RequestContext,
     ) -> Result<CreateTimelineResult, CreateTimelineError> {
         let src_id = src_timeline.timeline_id;
@@ -5064,6 +5093,7 @@ impl TenantShard {
                 start_lsn + 1,
                 Some(Arc::clone(src_timeline)),
                 Some(src_timeline.get_rel_size_v2_status()),
+                read_only,
                 ctx,
             )
             .await?;
@@ -5328,6 +5358,7 @@ impl TenantShard {
             }
         }
         let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
+        let read_only = false;
 
         // Import the contents of the data directory at the initial checkpoint
         // LSN, and any WAL after that.
@@ -5350,6 +5381,7 @@ impl TenantShard {
                 pgdata_lsn,
                 None,
                 None,
+                read_only,
                 ctx,
             )
             .await?;
@@ -5414,7 +5446,7 @@ impl TenantShard {
             pagestream_throttle_metrics: self.pagestream_throttle_metrics.clone(),
             l0_compaction_trigger: self.l0_compaction_trigger.clone(),
             l0_flush_global_state: self.l0_flush_global_state.clone(),
-            basebackup_prepare_sender: self.basebackup_prepare_sender.clone(),
+            basebackup_cache: self.basebackup_cache.clone(),
             feature_resolver: self.feature_resolver.clone(),
         }
     }
@@ -5433,14 +5465,17 @@ impl TenantShard {
         start_lsn: Lsn,
         ancestor: Option<Arc<Timeline>>,
         rel_size_v2_status: Option<RelSizeMigration>,
+        read_only: bool,
         ctx: &RequestContext,
     ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
         let tenant_shard_id = self.tenant_shard_id;
 
         let resources = self.build_timeline_resources(new_timeline_id);
-        resources
-            .remote_client
-            .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
+        resources.remote_client.init_upload_queue_for_empty_remote(
+            new_metadata,
+            rel_size_v2_status.clone(),
+            read_only,
+        )?;
 
         let (timeline_struct, timeline_ctx) = self
             .create_timeline_struct(
@@ -6000,7 +6035,7 @@ pub(crate) mod harness {
         ) -> anyhow::Result<Arc<TenantShard>> {
             let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
 
-            let (basebackup_requst_sender, _) = tokio::sync::mpsc::unbounded_channel();
+            let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
 
             let tenant = Arc::new(TenantShard::new(
                 TenantState::Attaching,
@@ -6008,7 +6043,7 @@ pub(crate) mod harness {
                 AttachedTenantConf::try_from(LocationConf::attached_single(
                     self.tenant_conf.clone(),
                     self.generation,
-                    &ShardParameters::default(),
+                    ShardParameters::default(),
                 ))
                 .unwrap(),
                 self.shard_identity,
@@ -6018,7 +6053,7 @@ pub(crate) mod harness {
                 self.deletion_queue.new_client(),
                 // TODO: ideally we should run all unit tests with both configs
                 L0FlushGlobalState::new(L0FlushConfig::default()),
-                basebackup_requst_sender,
+                basebackup_cache,
                 FeatureResolver::new_disabled(),
             ));
 
@@ -11429,11 +11464,11 @@ mod tests {
         if left != right {
             eprintln!("---LEFT---");
             for left in left.iter() {
-                eprintln!("{}", left);
+                eprintln!("{left}");
             }
             eprintln!("---RIGHT---");
             for right in right.iter() {
-                eprintln!("{}", right);
+                eprintln!("{right}");
             }
             assert_eq!(left, right);
         }

pageserver/src/tenant/config.rs

@@ -12,6 +12,7 @@
 use pageserver_api::models;
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
 use serde::{Deserialize, Serialize};
+use utils::critical;
 use utils::generation::Generation;
 
 #[derive(Debug, Copy, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -136,7 +137,7 @@ impl LocationConf {
     pub(crate) fn attached_single(
         tenant_conf: pageserver_api::models::TenantConfig,
         generation: Generation,
-        shard_params: &models::ShardParameters,
+        shard_params: models::ShardParameters,
     ) -> Self {
         Self {
             mode: LocationMode::Attached(AttachedLocationConfig {
@@ -171,6 +172,16 @@ impl LocationConf {
             }
         }
 
+        // This should never happen.
+        // TODO: turn this into a proper assertion.
+        if stripe_size != self.shard.stripe_size {
+            critical!(
+                "stripe size mismatch: {} != {}",
+                self.shard.stripe_size,
+                stripe_size,
+            );
+        }
+
         self.shard.stripe_size = stripe_size;
     }
 

pageserver/src/tenant/mgr.rs

@@ -880,6 +880,9 @@ impl TenantManager {
         // phase of writing config and/or waiting for flush, before returning.
         match fast_path_taken {
             Some(FastPathModified::Attached(tenant)) => {
+                tenant
+                    .shard_identity
+                    .assert_equal(new_location_config.shard);
                 TenantShard::persist_tenant_config(
                     self.conf,
                     &tenant_shard_id,
@@ -914,7 +917,10 @@ impl TenantManager {
 
                 return Ok(Some(tenant));
             }
-            Some(FastPathModified::Secondary(_secondary_tenant)) => {
+            Some(FastPathModified::Secondary(secondary_tenant)) => {
+                secondary_tenant
+                    .shard_identity
+                    .assert_equal(new_location_config.shard);
                 TenantShard::persist_tenant_config(
                     self.conf,
                     &tenant_shard_id,
@@ -948,6 +954,10 @@ impl TenantManager {
 
         match slot_guard.get_old_value() {
             Some(TenantSlot::Attached(tenant)) => {
+                tenant
+                    .shard_identity
+                    .assert_equal(new_location_config.shard);
+
                 // The case where we keep a Tenant alive was covered above in the special case
                 // for Attached->Attached transitions in the same generation.  By this point,
                 // if we see an attached tenant we know it will be discarded and should be
@@ -981,9 +991,13 @@ impl TenantManager {
                 // rather than assuming it to be empty.
                 spawn_mode = SpawnMode::Eager;
             }
-            Some(TenantSlot::Secondary(state)) => {
+            Some(TenantSlot::Secondary(secondary_tenant)) => {
+                secondary_tenant
+                    .shard_identity
+                    .assert_equal(new_location_config.shard);
+
                 info!("Shutting down secondary tenant");
-                state.shutdown().await;
+                secondary_tenant.shutdown().await;
             }
             Some(TenantSlot::InProgress(_)) => {
                 // This should never happen: acquire_slot should error out
@@ -2200,7 +2214,7 @@ impl TenantManager {
         selector: ShardSelector,
     ) -> ShardResolveResult {
         let tenants = self.tenants.read().unwrap();
-        let mut want_shard = None;
+        let mut want_shard: Option<ShardIndex> = None;
         let mut any_in_progress = None;
 
         match &*tenants {
@@ -2225,14 +2239,23 @@ impl TenantManager {
                             return ShardResolveResult::Found(tenant.clone());
                         }
                         ShardSelector::Page(key) => {
-                            // First slot we see for this tenant, calculate the expected shard number
-                            // for the key: we will use this for checking if this and subsequent
-                            // slots contain the key, rather than recalculating the hash each time.
-                            if want_shard.is_none() {
-                                want_shard = Some(tenant.shard_identity.get_shard_number(&key));
+                            // Each time we find an attached slot with a different shard count,
+                            // recompute the expected shard number: during shard splits we might
+                            // have multiple shards with the old shard count.
+                            if want_shard.is_none()
+                                || want_shard.unwrap().shard_count != tenant.shard_identity.count
+                            {
+                                want_shard = Some(ShardIndex {
+                                    shard_number: tenant.shard_identity.get_shard_number(&key),
+                                    shard_count: tenant.shard_identity.count,
+                                });
                             }
 
-                            if Some(tenant.shard_identity.number) == want_shard {
+                            if Some(ShardIndex {
+                                shard_number: tenant.shard_identity.number,
+                                shard_count: tenant.shard_identity.count,
+                            }) == want_shard
+                            {
                                 return ShardResolveResult::Found(tenant.clone());
                             }
                         }
@@ -2891,14 +2914,18 @@ mod tests {
     use std::collections::BTreeMap;
     use std::sync::Arc;
 
+    use camino::Utf8PathBuf;
     use storage_broker::BrokerClientChannel;
     use tracing::Instrument;
 
     use super::super::harness::TenantHarness;
     use super::TenantsMap;
-    use crate::tenant::{
-        TenantSharedResources,
-        mgr::{BackgroundPurges, TenantManager, TenantSlot},
+    use crate::{
+        basebackup_cache::BasebackupCache,
+        tenant::{
+            TenantSharedResources,
+            mgr::{BackgroundPurges, TenantManager, TenantSlot},
+        },
     };
 
     #[tokio::test(start_paused = true)]
@@ -2924,9 +2951,7 @@ mod tests {
         // Invoke remove_tenant_from_memory with a cleanup hook that blocks until we manually
         // permit it to proceed: that will stick the tenant in InProgress
 
-        let (basebackup_prepare_sender, _) = tokio::sync::mpsc::unbounded_channel::<
-            crate::basebackup_cache::BasebackupPrepareRequest,
-        >();
+        let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
 
         let tenant_manager = TenantManager {
             tenants: std::sync::RwLock::new(TenantsMap::Open(tenants)),
@@ -2940,7 +2965,7 @@ mod tests {
                 l0_flush_global_state: crate::l0_flush::L0FlushGlobalState::new(
                     h.conf.l0_flush.clone(),
                 ),
-                basebackup_prepare_sender,
+                basebackup_cache,
                 feature_resolver: crate::feature_resolver::FeatureResolver::new_disabled(),
             },
             cancel: tokio_util::sync::CancellationToken::new(),

pageserver/src/tenant/remote_timeline_client.rs

@@ -444,6 +444,7 @@ impl RemoteTimelineClient {
         &self,
         local_metadata: &TimelineMetadata,
         rel_size_v2_status: Option<RelSizeMigration>,
+        read_only: bool,
     ) -> anyhow::Result<()> {
         // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
         // certainly no point in starting more upload tasks than this.
@@ -456,6 +457,9 @@ impl RemoteTimelineClient {
         let initialized_queue =
             upload_queue.initialize_empty_remote(local_metadata, inprogress_limit)?;
         initialized_queue.dirty.rel_size_migration = rel_size_v2_status;
+        if read_only {
+            initialized_queue.dirty.read_only = Some(read_only);
+        }
         self.update_remote_physical_size_gauge(None);
         info!("initialized upload queue as empty");
         Ok(())
@@ -583,6 +587,17 @@ impl RemoteTimelineClient {
             .ok()
     }
 
+    /// Returns whether the timeline is archived.
+    /// Return None if the remote index_part hasn't been downloaded yet.
+    pub(crate) fn is_read_only(&self) -> Option<bool> {
+        self.upload_queue
+            .lock()
+            .unwrap()
+            .initialized_mut()
+            .map(|q| q.clean.0.read_only.unwrap_or_default())
+            .ok()
+    }
+
     /// Returns `Ok(Some(timestamp))` if the timeline has been archived, `Ok(None)` if the timeline hasn't been archived.
     ///
     /// Return Err(_) if the remote index_part hasn't been downloaded yet, or the timeline hasn't been stopped yet.

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -114,6 +114,12 @@ pub struct IndexPart {
     /// The timestamp when the timeline was marked invisible in synthetic size calculations.
     #[serde(skip_serializing_if = "Option::is_none", default)]
     pub(crate) marked_invisible_at: Option<NaiveDateTime>,
+
+    /// Whether the timeline is read only or not.
+    ///
+    /// The property is non-recursive, so child timelines can be non-read-only,
+    /// but it can't be changed for an individual timeline once the timeline has been created.
+    pub(crate) read_only: Option<bool>,
 }
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
@@ -142,6 +148,7 @@ impl IndexPart {
     /// - 12: +l2_lsn
     /// - 13: +gc_compaction
     /// - 14: +marked_invisible_at
+    /// - 15: +read_only
     const LATEST_VERSION: usize = 14;
 
     // Versions we may see when reading from a bucket.
@@ -165,6 +172,7 @@ impl IndexPart {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         }
     }
 
@@ -475,6 +483,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -524,6 +533,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -574,6 +584,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -675,6 +686,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -726,6 +738,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -782,6 +795,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -843,6 +857,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -905,6 +920,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -972,6 +988,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1052,6 +1069,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1133,6 +1151,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1220,6 +1239,7 @@ mod tests {
                 last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             }),
             marked_invisible_at: None,
+            read_only: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1308,6 +1328,97 @@ mod tests {
                 last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             }),
             marked_invisible_at: Some(parse_naive_datetime("2023-07-31T09:00:00.123000000")),
+            read_only: None,
+        };
+
+        let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v15_read_only_is_parsed() {
+        let example = r#"{
+            "version": 14,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata": {
+                "disk_consistent_lsn": "0/16960E8",
+                "prev_record_lsn": "0/1696070",
+                "ancestor_timeline": "e45a7f37d3ee2ff17dc14bf4f4e3f52e",
+                "ancestor_lsn": "0/0",
+                "latest_gc_cutoff_lsn": "0/1696070",
+                "initdb_lsn": "0/1696070",
+                "pg_version": 14
+            },
+            "gc_blocking": {
+                "started_at": "2024-07-19T09:00:00.123",
+                "reasons": ["DetachAncestor"]
+            },
+            "import_pgdata": {
+                "V1": {
+                    "Done": {
+                        "idempotency_key": "specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5",
+                        "started_at": "2024-11-13T09:23:42.123",
+                        "finished_at": "2024-11-13T09:42:23.123"
+                    }
+                }
+            },
+            "rel_size_migration": "legacy",
+            "l2_lsn": "0/16960E8",
+            "gc_compaction": {
+                "last_completed_lsn": "0/16960E8"
+            },
+            "marked_invisible_at": "2023-07-31T09:00:00.123",
+            "read_only": true,
+        }"#;
+
+        let expected = IndexPart {
+            version: 14,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
+                    file_size: 9007199254741001,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::new(
+                Lsn::from_str("0/16960E8").unwrap(),
+                Some(Lsn::from_str("0/1696070").unwrap()),
+                Some(TimelineId::from_str("e45a7f37d3ee2ff17dc14bf4f4e3f52e").unwrap()),
+                Lsn::INVALID,
+                Lsn::from_str("0/1696070").unwrap(),
+                Lsn::from_str("0/1696070").unwrap(),
+                PgMajorVersion::PG14,
+            ).with_recalculated_checksum().unwrap(),
+            deleted_at: None,
+            lineage: Default::default(),
+            gc_blocking: Some(GcBlocking {
+                started_at: parse_naive_datetime("2024-07-19T09:00:00.123000000"),
+                reasons: enumset::EnumSet::from_iter([GcBlockingReason::DetachAncestor]),
+            }),
+            last_aux_file_policy: Default::default(),
+            archived_at: None,
+            import_pgdata: Some(import_pgdata::index_part_format::Root::V1(import_pgdata::index_part_format::V1::Done(import_pgdata::index_part_format::Done{
+                started_at: parse_naive_datetime("2024-11-13T09:23:42.123000000"),
+                finished_at: parse_naive_datetime("2024-11-13T09:42:23.123000000"),
+                idempotency_key: import_pgdata::index_part_format::IdempotencyKey::new("specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5".to_string()),
+            }))),
+            rel_size_migration: Some(RelSizeMigration::Legacy),
+            l2_lsn: Some("0/16960E8".parse::<Lsn>().unwrap()),
+            gc_compaction: Some(GcCompactionState {
+                last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            }),
+            marked_invisible_at: Some(parse_naive_datetime("2023-07-31T09:00:00.123000000")),
+            read_only: Some(true),
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();

pageserver/src/tenant/secondary.rs

@@ -101,7 +101,7 @@ pub(crate) struct SecondaryTenant {
     // Secondary mode does not need the full shard identity or the pageserver_api::models::TenantConfig.  However,
     // storing these enables us to report our full LocationConf, enabling convenient reconciliation
     // by the control plane (see [`Self::get_location_conf`])
-    shard_identity: ShardIdentity,
+    pub(crate) shard_identity: ShardIdentity,
     tenant_conf: std::sync::Mutex<pageserver_api::models::TenantConfig>,
 
     // Internal state used by the Downloader.

pageserver/src/tenant/storage_layer/batch_split_writer.rs

@@ -55,11 +55,11 @@ pub struct BatchLayerWriter {
 }
 
 impl BatchLayerWriter {
-    pub async fn new(conf: &'static PageServerConf) -> anyhow::Result<Self> {
-        Ok(Self {
+    pub fn new(conf: &'static PageServerConf) -> Self {
+        Self {
             generated_layer_writers: Vec::new(),
             conf,
-        })
+        }
     }
 
     pub fn add_unfinished_image_writer(
@@ -182,7 +182,7 @@ impl BatchLayerWriter {
 /// An image writer that takes images and produces multiple image layers.
 #[must_use]
 pub struct SplitImageLayerWriter<'a> {
-    inner: ImageLayerWriter,
+    inner: Option<ImageLayerWriter>,
     target_layer_size: u64,
     lsn: Lsn,
     conf: &'static PageServerConf,
@@ -196,7 +196,7 @@ pub struct SplitImageLayerWriter<'a> {
 
 impl<'a> SplitImageLayerWriter<'a> {
     #[allow(clippy::too_many_arguments)]
-    pub async fn new(
+    pub fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
@@ -205,30 +205,19 @@ impl<'a> SplitImageLayerWriter<'a> {
         target_layer_size: u64,
         gate: &'a utils::sync::gate::Gate,
         cancel: CancellationToken,
-        ctx: &RequestContext,
-    ) -> anyhow::Result<Self> {
-        Ok(Self {
+    ) -> Self {
+        Self {
             target_layer_size,
-            inner: ImageLayerWriter::new(
-                conf,
-                timeline_id,
-                tenant_shard_id,
-                &(start_key..Key::MAX),
-                lsn,
-                gate,
-                cancel.clone(),
-                ctx,
-            )
-            .await?,
+            inner: None,
             conf,
             timeline_id,
             tenant_shard_id,
-            batches: BatchLayerWriter::new(conf).await?,
+            batches: BatchLayerWriter::new(conf),
             lsn,
             start_key,
             gate,
             cancel,
-        })
+        }
     }
 
     pub async fn put_image(
@@ -237,12 +226,31 @@ impl<'a> SplitImageLayerWriter<'a> {
         img: Bytes,
         ctx: &RequestContext,
     ) -> Result<(), PutError> {
+        if self.inner.is_none() {
+            self.inner = Some(
+                ImageLayerWriter::new(
+                    self.conf,
+                    self.timeline_id,
+                    self.tenant_shard_id,
+                    &(self.start_key..Key::MAX),
+                    self.lsn,
+                    self.gate,
+                    self.cancel.clone(),
+                    ctx,
+                )
+                .await
+                .map_err(PutError::Other)?,
+            );
+        }
+
+        let inner = self.inner.as_mut().unwrap();
+
         // The current estimation is an upper bound of the space that the key/image could take
         // because we did not consider compression in this estimation. The resulting image layer
         // could be smaller than the target size.
         let addition_size_estimation = KEY_SIZE as u64 + img.len() as u64;
-        if self.inner.num_keys() >= 1
-            && self.inner.estimated_size() + addition_size_estimation >= self.target_layer_size
+        if inner.num_keys() >= 1
+            && inner.estimated_size() + addition_size_estimation >= self.target_layer_size
         {
             let next_image_writer = ImageLayerWriter::new(
                 self.conf,
@@ -256,7 +264,7 @@ impl<'a> SplitImageLayerWriter<'a> {
             )
             .await
             .map_err(PutError::Other)?;
-            let prev_image_writer = std::mem::replace(&mut self.inner, next_image_writer);
+            let prev_image_writer = std::mem::replace(inner, next_image_writer);
             self.batches.add_unfinished_image_writer(
                 prev_image_writer,
                 self.start_key..key,
@@ -264,7 +272,7 @@ impl<'a> SplitImageLayerWriter<'a> {
             );
             self.start_key = key;
         }
-        self.inner.put_image(key, img, ctx).await
+        inner.put_image(key, img, ctx).await
     }
 
     pub(crate) async fn finish_with_discard_fn<D, F>(
@@ -281,8 +289,10 @@ impl<'a> SplitImageLayerWriter<'a> {
         let Self {
             mut batches, inner, ..
         } = self;
-        if inner.num_keys() != 0 {
-            batches.add_unfinished_image_writer(inner, self.start_key..end_key, self.lsn);
+        if let Some(inner) = inner {
+            if inner.num_keys() != 0 {
+                batches.add_unfinished_image_writer(inner, self.start_key..end_key, self.lsn);
+            }
         }
         batches.finish_with_discard_fn(tline, ctx, discard_fn).await
     }
@@ -319,7 +329,7 @@ pub struct SplitDeltaLayerWriter<'a> {
 }
 
 impl<'a> SplitDeltaLayerWriter<'a> {
-    pub async fn new(
+    pub fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
@@ -327,8 +337,8 @@ impl<'a> SplitDeltaLayerWriter<'a> {
         target_layer_size: u64,
         gate: &'a utils::sync::gate::Gate,
         cancel: CancellationToken,
-    ) -> anyhow::Result<Self> {
-        Ok(Self {
+    ) -> Self {
+        Self {
             target_layer_size,
             inner: None,
             conf,
@@ -336,10 +346,10 @@ impl<'a> SplitDeltaLayerWriter<'a> {
             tenant_shard_id,
             lsn_range,
             last_key_written: Key::MIN,
-            batches: BatchLayerWriter::new(conf).await?,
+            batches: BatchLayerWriter::new(conf),
             gate,
             cancel,
-        })
+        }
     }
 
     pub async fn put_value(
@@ -497,10 +507,7 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-            &ctx,
-        )
-        .await
-        .unwrap();
+        );
 
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,
@@ -510,9 +517,7 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        )
-        .await
-        .unwrap();
+        );
 
         image_writer
             .put_image(get_key(0), get_img(0), &ctx)
@@ -578,10 +583,7 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-            &ctx,
-        )
-        .await
-        .unwrap();
+        );
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,
             tline.timeline_id,
@@ -590,9 +592,7 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        )
-        .await
-        .unwrap();
+        );
         const N: usize = 2000;
         for i in 0..N {
             let i = i as u32;
@@ -679,10 +679,7 @@ mod tests {
             4 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-            &ctx,
-        )
-        .await
-        .unwrap();
+        );
 
         let mut delta_writer = SplitDeltaLayerWriter::new(
             tenant.conf,
@@ -692,9 +689,7 @@ mod tests {
             4 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        )
-        .await
-        .unwrap();
+        );
 
         image_writer
             .put_image(get_key(0), get_img(0), &ctx)
@@ -770,9 +765,7 @@ mod tests {
             4 * 1024 * 1024,
             &tline.gate,
             tline.cancel.clone(),
-        )
-        .await
-        .unwrap();
+        );
 
         for i in 0..N {
             let i = i as u32;

pageserver/src/tenant/tasks.rs

@@ -17,14 +17,17 @@ use tracing::*;
 use utils::backoff::exponential_backoff_duration;
 use utils::completion::Barrier;
 use utils::pausable_failpoint;
+use utils::sync::gate::GateError;
 
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::metrics::{self, BackgroundLoopSemaphoreMetricsRecorder, TENANT_TASK_EVENTS};
 use crate::task_mgr::{self, BACKGROUND_RUNTIME, TOKIO_WORKER_THREADS, TaskKind};
+use crate::tenant::blob_io::WriteBlobError;
 use crate::tenant::throttle::Stats;
 use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::compaction::CompactionOutcome;
 use crate::tenant::{TenantShard, TenantState};
+use crate::virtual_file::owned_buffers_io::write::FlushTaskError;
 
 /// Semaphore limiting concurrent background tasks (across all tenants).
 ///
@@ -313,7 +316,20 @@ pub(crate) fn log_compaction_error(
             let timeline = root_cause
                 .downcast_ref::<PageReconstructError>()
                 .is_some_and(|e| e.is_stopping());
-            let is_stopping = upload_queue || timeline;
+            let buffered_writer_flush_task_canelled = root_cause
+                .downcast_ref::<FlushTaskError>()
+                .is_some_and(|e| e.is_cancel());
+            let write_blob_cancelled = root_cause
+                .downcast_ref::<WriteBlobError>()
+                .is_some_and(|e| e.is_cancel());
+            let gate_closed = root_cause
+                .downcast_ref::<GateError>()
+                .is_some_and(|e| e.is_cancel());
+            let is_stopping = upload_queue
+                || timeline
+                || buffered_writer_flush_task_canelled
+                || write_blob_cancelled
+                || gate_closed;
 
             if is_stopping {
                 Level::INFO

pageserver/src/tenant/timeline.rs

@@ -78,7 +78,7 @@ use utils::rate_limit::RateLimit;
 use utils::seqwait::SeqWait;
 use utils::simple_rcu::{Rcu, RcuReadGuard};
 use utils::sync::gate::{Gate, GateGuard};
-use utils::{completion, critical, fs_ext, pausable_failpoint};
+use utils::{completion, critical_timeline, fs_ext, pausable_failpoint};
 #[cfg(test)]
 use wal_decoder::models::value::Value;
 use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
@@ -95,18 +95,18 @@ use super::storage_layer::{LayerFringe, LayerVisibilityHint, ReadableLayer};
 use super::tasks::log_compaction_error;
 use super::upload_queue::NotInitialized;
 use super::{
-    AttachedTenantConf, BasebackupPrepareSender, GcError, HeatMapTimeline, MaybeOffloaded,
+    AttachedTenantConf, GcError, HeatMapTimeline, MaybeOffloaded,
     debug_assert_current_span_has_tenant_and_timeline_id,
 };
 use crate::PERF_TRACE_TARGET;
 use crate::aux_file::AuxFileSizeEstimator;
-use crate::basebackup_cache::BasebackupPrepareRequest;
+use crate::basebackup_cache::BasebackupCache;
 use crate::config::PageServerConf;
 use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
 };
 use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
-use crate::feature_resolver::FeatureResolver;
+use crate::feature_resolver::TenantFeatureResolver;
 use crate::keyspace::{KeyPartitioning, KeySpace};
 use crate::l0_flush::{self, L0FlushGlobalState};
 use crate::metrics::{
@@ -201,8 +201,8 @@ pub struct TimelineResources {
     pub pagestream_throttle_metrics: Arc<crate::metrics::tenant_throttling::Pagestream>,
     pub l0_compaction_trigger: Arc<Notify>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
-    pub basebackup_prepare_sender: BasebackupPrepareSender,
-    pub feature_resolver: FeatureResolver,
+    pub basebackup_cache: Arc<BasebackupCache>,
+    pub feature_resolver: TenantFeatureResolver,
 }
 
 pub struct Timeline {
@@ -448,9 +448,9 @@ pub struct Timeline {
     wait_lsn_log_slow: tokio::sync::Semaphore,
 
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
-    basebackup_prepare_sender: BasebackupPrepareSender,
+    basebackup_cache: Arc<BasebackupCache>,
 
-    feature_resolver: FeatureResolver,
+    feature_resolver: TenantFeatureResolver,
 }
 
 pub(crate) enum PreviousHeatmap {
@@ -763,7 +763,7 @@ pub(crate) enum CreateImageLayersError {
     PageReconstructError(#[source] PageReconstructError),
 
     #[error(transparent)]
-    Other(#[from] anyhow::Error),
+    Other(anyhow::Error),
 }
 
 impl From<layer_manager::Shutdown> for CreateImageLayersError {
@@ -2144,14 +2144,31 @@ impl Timeline {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
         // Regardless of whether we're going to try_freeze_and_flush
-        // or not, stop ingesting any more data.
+        // cancel walreceiver to stop ingesting more data asap.
+        //
+        // Note that we're accepting a race condition here where we may
+        // do the final flush below, before walreceiver observes the
+        // cancellation and exits.
+        // This means we may open a new InMemoryLayer after the final flush below.
+        // Flush loop is also still running for a short while, so, in theory, it
+        // could also make its way into the upload queue.
+        //
+        // If we wait for the shutdown of the walreceiver before moving on to the
+        // flush, then that would be avoided. But we don't do it because the
+        // walreceiver entertains reads internally, which means that it possibly
+        // depends on the download of layers. Layer download is only sensitive to
+        // the cancellation of the entire timeline, so cancelling the walreceiver
+        // will have no effect on the individual get requests.
+        // This would cause problems when there is a lot of ongoing downloads or
+        // there is S3 unavailabilities, i.e. detach, deletion, etc would hang,
+        // and we can't deallocate resources of the timeline, etc.
         let walreceiver = self.walreceiver.lock().unwrap().take();
         tracing::debug!(
             is_some = walreceiver.is_some(),
             "Waiting for WalReceiverManager..."
         );
         if let Some(walreceiver) = walreceiver {
-            walreceiver.shutdown().await;
+            walreceiver.cancel().await;
         }
         // ... and inform any waiters for newer LSNs that there won't be any.
         self.last_record_lsn.shutdown();
@@ -2332,6 +2349,10 @@ impl Timeline {
         self.remote_client.is_invisible()
     }
 
+    pub(crate) fn is_read_only(&self) -> Option<bool> {
+        self.remote_client.is_read_only()
+    }
+
     pub(crate) fn is_stopping(&self) -> bool {
         self.current_state() == TimelineState::Stopping
     }
@@ -2500,6 +2521,37 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.basebackup_cache_enabled)
     }
 
+    /// Try to get a basebackup from the on-disk cache.
+    pub(crate) async fn get_cached_basebackup(&self, lsn: Lsn) -> Option<tokio::fs::File> {
+        self.basebackup_cache
+            .get(self.tenant_shard_id.tenant_id, self.timeline_id, lsn)
+            .await
+    }
+
+    /// Convenience method to attempt fetching a basebackup for the timeline if enabled and safe for
+    /// the given request parameters.
+    ///
+    /// TODO: consider moving this onto GrpcPageServiceHandler once the libpq handler is gone.
+    pub async fn get_cached_basebackup_if_enabled(
+        &self,
+        lsn: Option<Lsn>,
+        prev_lsn: Option<Lsn>,
+        full: bool,
+        replica: bool,
+        gzip: bool,
+    ) -> Option<tokio::fs::File> {
+        if !self.is_basebackup_cache_enabled() || !self.basebackup_cache.is_enabled() {
+            return None;
+        }
+        // We have to know which LSN to fetch the basebackup for.
+        let lsn = lsn?;
+        // We only cache gzipped, non-full basebackups for primary computes with automatic prev_lsn.
+        if prev_lsn.is_some() || full || replica || !gzip {
+            return None;
+        }
+        self.get_cached_basebackup(lsn).await
+    }
+
     /// Prepare basebackup for the given LSN and store it in the basebackup cache.
     /// The method is asynchronous and returns immediately.
     /// The actual basebackup preparation is performed in the background
@@ -2521,17 +2573,8 @@ impl Timeline {
             return;
         }
 
-        let res = self
-            .basebackup_prepare_sender
-            .send(BasebackupPrepareRequest {
-                tenant_shard_id: self.tenant_shard_id,
-                timeline_id: self.timeline_id,
-                lsn,
-            });
-        if let Err(e) = res {
-            // May happen during shutdown, it's not critical.
-            info!("Failed to send shutdown checkpoint: {e:#}");
-        }
+        self.basebackup_cache
+            .send_prepare(self.tenant_shard_id, self.timeline_id, lsn);
     }
 }
 
@@ -3088,7 +3131,7 @@ impl Timeline {
 
                 wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
 
-                basebackup_prepare_sender: resources.basebackup_prepare_sender,
+                basebackup_cache: resources.basebackup_cache,
 
                 feature_resolver: resources.feature_resolver,
             };
@@ -4658,6 +4701,16 @@ impl Timeline {
         mut layer_flush_start_rx: tokio::sync::watch::Receiver<(u64, Lsn)>,
         ctx: &RequestContext,
     ) {
+        // Always notify waiters about the flush loop exiting since the loop might stop
+        // when the timeline hasn't been cancelled.
+        let scopeguard_rx = layer_flush_start_rx.clone();
+        scopeguard::defer! {
+            let (flush_counter, _) = *scopeguard_rx.borrow();
+            let _ = self
+                .layer_flush_done_tx
+                .send_replace((flush_counter, Err(FlushLayerError::Cancelled)));
+        }
+
         // Subscribe to L0 delta layer updates, for compaction backpressure.
         let mut watch_l0 = match self
             .layers
@@ -4687,9 +4740,6 @@ impl Timeline {
             let result = loop {
                 if self.cancel.is_cancelled() {
                     info!("dropping out of flush loop for timeline shutdown");
-                    // Note: we do not bother transmitting into [`layer_flush_done_tx`], because
-                    // anyone waiting on that will respect self.cancel as well: they will stop
-                    // waiting at the same time we as drop out of this loop.
                     return;
                 }
 
@@ -4700,7 +4750,7 @@ impl Timeline {
                 }
 
                 // Fetch the next layer to flush, if any.
-                let (layer, l0_count, frozen_count, frozen_size) = {
+                let (layer, l0_count, frozen_count, frozen_size, open_layer_size) = {
                     let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
                     let Ok(lm) = layers.layer_map() else {
                         info!("dropping out of flush loop for timeline shutdown");
@@ -4713,8 +4763,13 @@ impl Timeline {
                         .iter()
                         .map(|l| l.estimated_in_mem_size())
                         .sum();
+                    let open_layer_size: u64 = lm
+                        .open_layer
+                        .as_ref()
+                        .map(|l| l.estimated_in_mem_size())
+                        .unwrap_or(0);
                     let layer = lm.frozen_layers.front().cloned();
-                    (layer, l0_count, frozen_count, frozen_size)
+                    (layer, l0_count, frozen_count, frozen_size, open_layer_size)
                     // drop 'layers' lock
                 };
                 let Some(layer) = layer else {
@@ -4727,7 +4782,7 @@ impl Timeline {
                     if l0_count >= stall_threshold {
                         warn!(
                             "stalling layer flushes for compaction backpressure at {l0_count} \
-                            L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
+                            L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
                         );
                         let stall_timer = self
                             .metrics
@@ -4780,7 +4835,7 @@ impl Timeline {
                         let delay = flush_duration.as_secs_f64();
                         info!(
                             "delaying layer flush by {delay:.3}s for compaction backpressure at \
-                            {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes)"
+                            {l0_count} L0 layers ({frozen_count} frozen layers with {frozen_size} bytes, {open_layer_size} bytes in open layer)"
                         );
                         let _delay_timer = self
                             .metrics
@@ -5279,6 +5334,7 @@ impl Timeline {
         ctx: &RequestContext,
         img_range: Range<Key>,
         io_concurrency: IoConcurrency,
+        progress: Option<(usize, usize)>,
     ) -> Result<ImageLayerCreationOutcome, CreateImageLayersError> {
         let mut wrote_keys = false;
 
@@ -5355,11 +5411,15 @@ impl Timeline {
             }
         }
 
+        let progress_report = progress
+            .map(|(idx, total)| format!("({idx}/{total}) "))
+            .unwrap_or_default();
         if wrote_keys {
             // Normal path: we have written some data into the new image layer for this
             // partition, so flush it to disk.
             info!(
-                "produced image layer for rel {}",
+                "{} produced image layer for rel {}",
+                progress_report,
                 ImageLayerName {
                     key_range: img_range.clone(),
                     lsn
@@ -5369,7 +5429,12 @@ impl Timeline {
                 unfinished_image_layer: image_layer_writer,
             })
         } else {
-            tracing::debug!("no data in range {}-{}", img_range.start, img_range.end);
+            tracing::debug!(
+                "{} no data in range {}-{}",
+                progress_report,
+                img_range.start,
+                img_range.end
+            );
             Ok(ImageLayerCreationOutcome::Empty)
         }
     }
@@ -5561,7 +5626,7 @@ impl Timeline {
                 self.should_check_if_image_layers_required(lsn)
             };
 
-        let mut batch_image_writer = BatchLayerWriter::new(self.conf).await?;
+        let mut batch_image_writer = BatchLayerWriter::new(self.conf);
 
         let mut all_generated = true;
 
@@ -5604,7 +5669,8 @@ impl Timeline {
             }
         }
 
-        for partition in partition_parts.iter() {
+        let total = partition_parts.len();
+        for (idx, partition) in partition_parts.iter().enumerate() {
             if self.cancel.is_cancelled() {
                 return Err(CreateImageLayersError::Cancelled);
             }
@@ -5665,7 +5731,8 @@ impl Timeline {
                 self.cancel.clone(),
                 ctx,
             )
-            .await?;
+            .await
+            .map_err(CreateImageLayersError::Other)?;
 
             fail_point!("image-layer-writer-fail-before-finish", |_| {
                 Err(CreateImageLayersError::Other(anyhow::anyhow!(
@@ -5688,6 +5755,7 @@ impl Timeline {
                     ctx,
                     img_range.clone(),
                     io_concurrency,
+                    Some((idx, total)),
                 )
                 .await?
             } else {
@@ -5760,7 +5828,10 @@ impl Timeline {
             }
         }
 
-        let image_layers = batch_image_writer.finish(self, ctx).await?;
+        let image_layers = batch_image_writer
+            .finish(self, ctx)
+            .await
+            .map_err(CreateImageLayersError::Other)?;
 
         let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
 
@@ -6774,7 +6845,11 @@ impl Timeline {
                     Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
                     Err(walredo::Error::Other(err)) => {
                         if fire_critical_error {
-                            critical!("walredo failure during page reconstruction: {err:?}");
+                            critical_timeline!(
+                                self.tenant_shard_id,
+                                self.timeline_id,
+                                "walredo failure during page reconstruction: {err:?}"
+                            );
                         }
                         return Err(PageReconstructError::WalRedo(
                             err.context("reconstruct a page image"),

pageserver/src/tenant/timeline/compaction.rs

@@ -9,7 +9,7 @@ use std::ops::{Deref, Range};
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
-use super::layer_manager::{LayerManagerLockHolder, LayerManagerReadGuard};
+use super::layer_manager::LayerManagerLockHolder;
 use super::{
     CompactFlags, CompactOptions, CompactionError, CreateImageLayersError, DurationRecorder,
     GetVectoredError, ImageLayerCreationMode, LastImageLayerCreationStatus, RecordedDuration,
@@ -36,7 +36,7 @@ use serde::Serialize;
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, debug, error, info, info_span, trace, warn};
-use utils::critical;
+use utils::critical_timeline;
 use utils::id::TimelineId;
 use utils::lsn::Lsn;
 use wal_decoder::models::record::NeonWalRecord;
@@ -101,7 +101,11 @@ pub enum GcCompactionQueueItem {
         /// Whether the compaction is triggered automatically (determines whether we need to update L2 LSN)
         auto: bool,
     },
-    SubCompactionJob(CompactOptions),
+    SubCompactionJob {
+        i: usize,
+        total: usize,
+        options: CompactOptions,
+    },
     Notify(GcCompactionJobId, Option<Lsn>),
 }
 
@@ -163,7 +167,7 @@ impl GcCompactionQueueItem {
                 running,
                 job_id: id.0,
             }),
-            GcCompactionQueueItem::SubCompactionJob(options) => Some(CompactInfoResponse {
+            GcCompactionQueueItem::SubCompactionJob { options, .. } => Some(CompactInfoResponse {
                 compact_key_range: options.compact_key_range,
                 compact_lsn_range: options.compact_lsn_range,
                 sub_compaction: options.sub_compaction,
@@ -489,7 +493,7 @@ impl GcCompactionQueue {
                 .map(|job| job.compact_lsn_range.end)
                 .max()
                 .unwrap();
-            for job in jobs {
+            for (i, job) in jobs.into_iter().enumerate() {
                 // Unfortunately we need to convert the `GcCompactJob` back to `CompactionOptions`
                 // until we do further refactors to allow directly call `compact_with_gc`.
                 let mut flags: EnumSet<CompactFlags> = EnumSet::default();
@@ -507,7 +511,11 @@ impl GcCompactionQueue {
                     compact_lsn_range: Some(job.compact_lsn_range.into()),
                     sub_compaction_max_job_size_mb: None,
                 };
-                pending_tasks.push(GcCompactionQueueItem::SubCompactionJob(options));
+                pending_tasks.push(GcCompactionQueueItem::SubCompactionJob {
+                    options,
+                    i,
+                    total: jobs_len,
+                });
             }
 
             if !auto {
@@ -651,7 +659,7 @@ impl GcCompactionQueue {
                     }
                 }
             }
-            GcCompactionQueueItem::SubCompactionJob(options) => {
+            GcCompactionQueueItem::SubCompactionJob { options, i, total } => {
                 // TODO: error handling, clear the queue if any task fails?
                 let _gc_guard = match gc_block.start().await {
                     Ok(guard) => guard,
@@ -663,6 +671,7 @@ impl GcCompactionQueue {
                         )));
                     }
                 };
+                info!("running gc-compaction subcompaction job {}/{}", i, total);
                 let res = timeline.compact_with_options(cancel, options, ctx).await;
                 let compaction_result = match res {
                     Ok(res) => res,
@@ -1310,7 +1319,7 @@ impl Timeline {
             || cfg!(feature = "testing")
             || self
                 .feature_resolver
-                .evaluate_boolean("image-compaction-boundary", self.tenant_shard_id.tenant_id)
+                .evaluate_boolean("image-compaction-boundary")
                 .is_ok()
         {
             let last_repartition_lsn = self.partitioning.read().1;
@@ -1381,7 +1390,11 @@ impl Timeline {
                             GetVectoredError::MissingKey(_),
                         ) = err
                         {
-                            critical!("missing key during compaction: {err:?}");
+                            critical_timeline!(
+                                self.tenant_shard_id,
+                                self.timeline_id,
+                                "missing key during compaction: {err:?}"
+                            );
                         }
                     })?;
 
@@ -1409,7 +1422,11 @@ impl Timeline {
 
             // Alert on critical errors that indicate data corruption.
             Err(err) if err.is_critical() => {
-                critical!("could not compact, repartitioning keyspace failed: {err:?}");
+                critical_timeline!(
+                    self.tenant_shard_id,
+                    self.timeline_id,
+                    "could not compact, repartitioning keyspace failed: {err:?}"
+                );
             }
 
             // Log other errors. No partitioning? This is normal, if the timeline was just created
@@ -1591,13 +1608,15 @@ impl Timeline {
         let started = Instant::now();
 
         let mut replace_image_layers = Vec::new();
+        let total = layers_to_rewrite.len();
 
-        for layer in layers_to_rewrite {
+        for (i, layer) in layers_to_rewrite.into_iter().enumerate() {
             if self.cancel.is_cancelled() {
                 return Err(CompactionError::ShuttingDown);
             }
 
-            info!(layer=%layer, "rewriting layer after shard split");
+            info!(layer=%layer, "rewriting layer after shard split: {}/{}", i, total);
+
             let mut image_layer_writer = ImageLayerWriter::new(
                 self.conf,
                 self.timeline_id,
@@ -1779,20 +1798,14 @@ impl Timeline {
         } = {
             let phase1_span = info_span!("compact_level0_phase1");
             let ctx = ctx.attached_child();
-            let mut stats = CompactLevel0Phase1StatsBuilder {
+            let stats = CompactLevel0Phase1StatsBuilder {
                 version: Some(2),
                 tenant_id: Some(self.tenant_shard_id),
                 timeline_id: Some(self.timeline_id),
                 ..Default::default()
             };
 
-            let begin = tokio::time::Instant::now();
-            let phase1_layers_locked = self.layers.read(LayerManagerLockHolder::Compaction).await;
-            let now = tokio::time::Instant::now();
-            stats.read_lock_acquisition_micros =
-                DurationRecorder::Recorded(RecordedDuration(now - begin), now);
             self.compact_level0_phase1(
-                phase1_layers_locked,
                 stats,
                 target_file_size,
                 force_compaction_ignore_threshold,
@@ -1813,16 +1826,19 @@ impl Timeline {
     }
 
     /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
-    async fn compact_level0_phase1<'a>(
-        self: &'a Arc<Self>,
-        guard: LayerManagerReadGuard<'a>,
+    async fn compact_level0_phase1(
+        self: &Arc<Self>,
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
         ctx: &RequestContext,
     ) -> Result<CompactLevel0Phase1Result, CompactionError> {
-        stats.read_lock_held_spawn_blocking_startup_micros =
-            stats.read_lock_acquisition_micros.till_now(); // set by caller
+        let begin = tokio::time::Instant::now();
+        let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
+        let now = tokio::time::Instant::now();
+        stats.read_lock_acquisition_micros =
+            DurationRecorder::Recorded(RecordedDuration(now - begin), now);
+
         let layers = guard.layer_map()?;
         let level0_deltas = layers.level0_deltas();
         stats.level0_deltas_count = Some(level0_deltas.len());
@@ -1857,6 +1873,12 @@ impl Timeline {
             .map(|x| guard.get_from_desc(x))
             .collect::<Vec<_>>();
 
+        drop_layer_manager_rlock(guard);
+
+        // The is the last LSN that we have seen for L0 compaction in the timeline. This LSN might be updated
+        // by the time we finish the compaction. So we need to get it here.
+        let l0_last_record_lsn = self.get_last_record_lsn();
+
         // Gather the files to compact in this iteration.
         //
         // Start with the oldest Level 0 delta file, and collect any other
@@ -1944,9 +1966,7 @@ impl Timeline {
         // we don't accidentally use it later in the function.
         drop(level0_deltas);
 
-        stats.read_lock_held_prerequisites_micros = stats
-            .read_lock_held_spawn_blocking_startup_micros
-            .till_now();
+        stats.compaction_prerequisites_micros = stats.read_lock_acquisition_micros.till_now();
 
         // TODO: replace with streaming k-merge
         let all_keys = {
@@ -1968,7 +1988,7 @@ impl Timeline {
             all_keys
         };
 
-        stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
+        stats.read_lock_held_key_sort_micros = stats.compaction_prerequisites_micros.till_now();
 
         // Determine N largest holes where N is number of compacted layers. The vec is sorted by key range start.
         //
@@ -2002,7 +2022,6 @@ impl Timeline {
                 }
             }
             let max_holes = deltas_to_compact.len();
-            let last_record_lsn = self.get_last_record_lsn();
             let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
             let min_hole_coverage_size = 3; // TODO: something more flexible?
             // min-heap (reserve space for one more element added before eviction)
@@ -2021,8 +2040,12 @@ impl Timeline {
                         // has not so much sense, because largest holes will corresponds field1/field2 changes.
                         // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
                         // That is why it is better to measure size of hole as number of covering image layers.
-                        let coverage_size =
-                            layers.image_coverage(&key_range, last_record_lsn).len();
+                        let coverage_size = {
+                            // TODO: optimize this with copy-on-write layer map.
+                            let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
+                            let layers = guard.layer_map()?;
+                            layers.image_coverage(&key_range, l0_last_record_lsn).len()
+                        };
                         if coverage_size >= min_hole_coverage_size {
                             heap.push(Hole {
                                 key_range,
@@ -2041,7 +2064,6 @@ impl Timeline {
             holes
         };
         stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
-        drop_layer_manager_rlock(guard);
 
         if self.cancel.is_cancelled() {
             return Err(CompactionError::ShuttingDown);
@@ -2382,9 +2404,8 @@ struct CompactLevel0Phase1StatsBuilder {
     tenant_id: Option<TenantShardId>,
     timeline_id: Option<TimelineId>,
     read_lock_acquisition_micros: DurationRecorder,
-    read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
     read_lock_held_key_sort_micros: DurationRecorder,
-    read_lock_held_prerequisites_micros: DurationRecorder,
+    compaction_prerequisites_micros: DurationRecorder,
     read_lock_held_compute_holes_micros: DurationRecorder,
     read_lock_drop_micros: DurationRecorder,
     write_layer_files_micros: DurationRecorder,
@@ -2399,9 +2420,8 @@ struct CompactLevel0Phase1Stats {
     tenant_id: TenantShardId,
     timeline_id: TimelineId,
     read_lock_acquisition_micros: RecordedDuration,
-    read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
     read_lock_held_key_sort_micros: RecordedDuration,
-    read_lock_held_prerequisites_micros: RecordedDuration,
+    compaction_prerequisites_micros: RecordedDuration,
     read_lock_held_compute_holes_micros: RecordedDuration,
     read_lock_drop_micros: RecordedDuration,
     write_layer_files_micros: RecordedDuration,
@@ -2426,16 +2446,12 @@ impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
                 .read_lock_acquisition_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
-            read_lock_held_spawn_blocking_startup_micros: value
-                .read_lock_held_spawn_blocking_startup_micros
-                .into_recorded()
-                .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
             read_lock_held_key_sort_micros: value
                 .read_lock_held_key_sort_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
-            read_lock_held_prerequisites_micros: value
-                .read_lock_held_prerequisites_micros
+            compaction_prerequisites_micros: value
+                .compaction_prerequisites_micros
                 .into_recorded()
                 .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
             read_lock_held_compute_holes_micros: value
@@ -3503,22 +3519,16 @@ impl Timeline {
         // Only create image layers when there is no ancestor branches. TODO: create covering image layer
         // when some condition meet.
         let mut image_layer_writer = if !has_data_below {
-            Some(
-                SplitImageLayerWriter::new(
-                    self.conf,
-                    self.timeline_id,
-                    self.tenant_shard_id,
-                    job_desc.compaction_key_range.start,
-                    lowest_retain_lsn,
-                    self.get_compaction_target_size(),
-                    &self.gate,
-                    self.cancel.clone(),
-                    ctx,
-                )
-                .await
-                .context("failed to create image layer writer")
-                .map_err(CompactionError::Other)?,
-            )
+            Some(SplitImageLayerWriter::new(
+                self.conf,
+                self.timeline_id,
+                self.tenant_shard_id,
+                job_desc.compaction_key_range.start,
+                lowest_retain_lsn,
+                self.get_compaction_target_size(),
+                &self.gate,
+                self.cancel.clone(),
+            ))
         } else {
             None
         };
@@ -3531,10 +3541,7 @@ impl Timeline {
             self.get_compaction_target_size(),
             &self.gate,
             self.cancel.clone(),
-        )
-        .await
-        .context("failed to create delta layer writer")
-        .map_err(CompactionError::Other)?;
+        );
 
         #[derive(Default)]
         struct RewritingLayers {
@@ -4330,7 +4337,8 @@ impl TimelineAdaptor {
             self.timeline.cancel.clone(),
             ctx,
         )
-        .await?;
+        .await
+        .map_err(CreateImageLayersError::Other)?;
 
         fail_point!("image-layer-writer-fail-before-finish", |_| {
             Err(CreateImageLayersError::Other(anyhow::anyhow!(
@@ -4339,7 +4347,10 @@ impl TimelineAdaptor {
         });
 
         let keyspace = KeySpace {
-            ranges: self.get_keyspace(key_range, lsn, ctx).await?,
+            ranges: self
+                .get_keyspace(key_range, lsn, ctx)
+                .await
+                .map_err(CreateImageLayersError::Other)?,
         };
         // TODO set proper (stateful) start. The create_image_layer_for_rel_blocks function mostly
         let outcome = self
@@ -4351,6 +4362,7 @@ impl TimelineAdaptor {
                 ctx,
                 key_range.clone(),
                 IoConcurrency::sequential(),
+                None,
             )
             .await?;
 
@@ -4358,9 +4370,13 @@ impl TimelineAdaptor {
             unfinished_image_layer,
         } = outcome
         {
-            let (desc, path) = unfinished_image_layer.finish(ctx).await?;
+            let (desc, path) = unfinished_image_layer
+                .finish(ctx)
+                .await
+                .map_err(CreateImageLayersError::Other)?;
             let image_layer =
-                Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)?;
+                Layer::finish_creating(self.timeline.conf, &self.timeline, desc, &path)
+                    .map_err(CreateImageLayersError::Other)?;
             self.new_images.push(image_layer);
         }
 

pageserver/src/tenant/timeline/delete.rs

@@ -241,8 +241,17 @@ impl DeleteTimelineFlow {
                 {
                     Ok(r) => r,
                     Err(DownloadError::NotFound) => {
-                        // Deletion is already complete
+                        // Deletion is already complete.
+                        // As we came here, we will need to remove the timeline from the tenant though.
                         tracing::info!("Timeline already deleted in remote storage");
+                        if let TimelineOrOffloaded::Offloaded(_) = &timeline {
+                            // We only supoprt this for offloaded timelines, as we don't know which state non-offloaded timelines are in.
+                            tracing::info!(
+                                "Timeline with gone index part is offloaded timeline. Removing from tenant."
+                            );
+                            remove_maybe_offloaded_timeline_from_tenant(tenant, &timeline, &guard)
+                                .await?;
+                        }
                         return Ok(());
                     }
                     Err(e) => {

pageserver/src/tenant/timeline/detach_ancestor.rs

@@ -182,6 +182,7 @@ pub(crate) async fn generate_tombstone_image_layer(
     detached: &Arc<Timeline>,
     ancestor: &Arc<Timeline>,
     ancestor_lsn: Lsn,
+    historic_layers_to_copy: &Vec<Layer>,
     ctx: &RequestContext,
 ) -> Result<Option<ResidentLayer>, Error> {
     tracing::info!(
@@ -199,6 +200,20 @@ pub(crate) async fn generate_tombstone_image_layer(
     let image_lsn = ancestor_lsn;
 
     {
+        for layer in historic_layers_to_copy {
+            let desc = layer.layer_desc();
+            if !desc.is_delta
+                && desc.lsn_range.start == image_lsn
+                && overlaps_with(&key_range, &desc.key_range)
+            {
+                tracing::info!(
+                    layer=%layer, "will copy tombstone from ancestor instead of creating a new one"
+                );
+
+                return Ok(None);
+            }
+        }
+
         let layers = detached
             .layers
             .read(LayerManagerLockHolder::DetachAncestor)
@@ -450,7 +465,8 @@ pub(super) async fn prepare(
         Vec::with_capacity(straddling_branchpoint.len() + rest_of_historic.len() + 1);
 
     if let Some(tombstone_layer) =
-        generate_tombstone_image_layer(detached, &ancestor, ancestor_lsn, ctx).await?
+        generate_tombstone_image_layer(detached, &ancestor, ancestor_lsn, &rest_of_historic, ctx)
+            .await?
     {
         new_layers.push(tombstone_layer.into());
     }
@@ -885,7 +901,7 @@ async fn remote_copy(
                 }
                 tracing::info!("Deleting orphan layer file to make way for hard linking");
                 // Delete orphan layer file and try again, to ensure this layer has a well understood source
-                std::fs::remove_file(adopted_path)
+                std::fs::remove_file(&adoptee_path)
                     .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
                 std::fs::hard_link(adopted_path, &adoptee_path)
                     .map_err(|e| Error::launder(e.into(), Error::Prepare))?;

pageserver/src/tenant/timeline/handle.rs

@@ -887,7 +887,7 @@ mod tests {
             .expect("we still have it");
     }
 
-    fn make_relation_key_for_shard(shard: ShardNumber, params: &ShardParameters) -> Key {
+    fn make_relation_key_for_shard(shard: ShardNumber, params: ShardParameters) -> Key {
         rel_block_to_key(
             RelTag {
                 spcnode: 1663,
@@ -917,14 +917,14 @@ mod tests {
         let child0 = Arc::new_cyclic(|myself| StubTimeline {
             gate: Default::default(),
             id: timeline_id,
-            shard: ShardIdentity::from_params(ShardNumber(0), &child_params),
+            shard: ShardIdentity::from_params(ShardNumber(0), child_params),
             per_timeline_state: PerTimelineState::default(),
             myself: myself.clone(),
         });
         let child1 = Arc::new_cyclic(|myself| StubTimeline {
             gate: Default::default(),
             id: timeline_id,
-            shard: ShardIdentity::from_params(ShardNumber(1), &child_params),
+            shard: ShardIdentity::from_params(ShardNumber(1), child_params),
             per_timeline_state: PerTimelineState::default(),
             myself: myself.clone(),
         });
@@ -937,7 +937,7 @@ mod tests {
             let handle = cache
                 .get(
                     timeline_id,
-                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), &child_params)),
+                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), child_params)),
                     &StubManager {
                         shards: vec![parent.clone()],
                     },
@@ -961,7 +961,7 @@ mod tests {
             let handle = cache
                 .get(
                     timeline_id,
-                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), &child_params)),
+                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), child_params)),
                     &StubManager {
                         shards: vec![], // doesn't matter what's in here, the cache is fully loaded
                     },
@@ -978,7 +978,7 @@ mod tests {
         let parent_handle = cache
             .get(
                 timeline_id,
-                ShardSelector::Page(make_relation_key_for_shard(ShardNumber(0), &child_params)),
+                ShardSelector::Page(make_relation_key_for_shard(ShardNumber(0), child_params)),
                 &StubManager {
                     shards: vec![parent.clone()],
                 },
@@ -995,7 +995,7 @@ mod tests {
             let handle = cache
                 .get(
                     timeline_id,
-                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), &child_params)),
+                    ShardSelector::Page(make_relation_key_for_shard(ShardNumber(i), child_params)),
                     &StubManager {
                         shards: vec![child0.clone(), child1.clone()], // <====== this changed compared to previous loop
                     },

pageserver/src/tenant/timeline/offload.rs

@@ -19,6 +19,8 @@ pub(crate) enum OffloadError {
     NotArchived,
     #[error(transparent)]
     RemoteStorage(anyhow::Error),
+    #[error("Offload or deletion already in progress")]
+    AlreadyInProgress,
     #[error("Unexpected offload error: {0}")]
     Other(anyhow::Error),
 }
@@ -44,20 +46,26 @@ pub(crate) async fn offload_timeline(
         timeline.timeline_id,
         TimelineDeleteGuardKind::Offload,
     );
-    if let Err(DeleteTimelineError::HasChildren(children)) = delete_guard_res {
-        let is_archived = timeline.is_archived();
-        if is_archived == Some(true) {
-            tracing::error!("timeline is archived but has non-archived children: {children:?}");
+    let (timeline, guard) = match delete_guard_res {
+        Ok(timeline_and_guard) => timeline_and_guard,
+        Err(DeleteTimelineError::HasChildren(children)) => {
+            let is_archived = timeline.is_archived();
+            if is_archived == Some(true) {
+                tracing::error!("timeline is archived but has non-archived children: {children:?}");
+                return Err(OffloadError::NotArchived);
+            }
+            tracing::info!(
+                ?is_archived,
+                "timeline is not archived and has unarchived children"
+            );
             return Err(OffloadError::NotArchived);
         }
-        tracing::info!(
-            ?is_archived,
-            "timeline is not archived and has unarchived children"
-        );
-        return Err(OffloadError::NotArchived);
+        Err(DeleteTimelineError::AlreadyInProgress(_)) => {
+            tracing::info!("timeline offload or deletion already in progress");
+            return Err(OffloadError::AlreadyInProgress);
+        }
+        Err(e) => return Err(OffloadError::Other(anyhow::anyhow!(e))),
     };
-    let (timeline, guard) =
-        delete_guard_res.map_err(|e| OffloadError::Other(anyhow::anyhow!(e)))?;
 
     let TimelineOrOffloaded::Timeline(timeline) = timeline else {
         tracing::error!("timeline already offloaded, but given timeline object");

pageserver/src/tenant/timeline/walreceiver.rs

@@ -63,7 +63,6 @@ pub struct WalReceiver {
     /// All task spawned by [`WalReceiver::start`] and its children are sensitive to this token.
     /// It's a child token of [`Timeline`] so that timeline shutdown can cancel WalReceiver tasks early for `freeze_and_flush=true`.
     cancel: CancellationToken,
-    task: tokio::task::JoinHandle<()>,
 }
 
 impl WalReceiver {
@@ -80,7 +79,7 @@ impl WalReceiver {
         let loop_status = Arc::new(std::sync::RwLock::new(None));
         let manager_status = Arc::clone(&loop_status);
         let cancel = timeline.cancel.child_token();
-        let task = WALRECEIVER_RUNTIME.spawn({
+        let _task = WALRECEIVER_RUNTIME.spawn({
             let cancel = cancel.clone();
             async move {
                 debug_assert_current_span_has_tenant_and_timeline_id();
@@ -121,25 +120,14 @@ impl WalReceiver {
         Self {
             manager_status,
             cancel,
-            task,
         }
     }
 
     #[instrument(skip_all, level = tracing::Level::DEBUG)]
-    pub async fn shutdown(self) {
+    pub async fn cancel(self) {
         debug_assert_current_span_has_tenant_and_timeline_id();
         debug!("cancelling walreceiver tasks");
         self.cancel.cancel();
-        match self.task.await {
-            Ok(()) => debug!("Shutdown success"),
-            Err(je) if je.is_cancelled() => unreachable!("not used"),
-            Err(je) if je.is_panic() => {
-                // already logged by panic hook
-            }
-            Err(je) => {
-                error!("shutdown walreceiver task join error: {je}")
-            }
-        }
     }
 
     pub(crate) fn status(&self) -> Option<ConnectionManagerStatus> {