fix

## Problem

Safekeeper and pageserver metrics collection might time out. We've seen
this in both hadron and neon.

## Summary of changes

This PR moves metrics collection in PS/SK to the background so that we
will always get some metrics, despite there may be some delays. Will
leave it to the future work to reduce metrics collection time.

---------

Co-authored-by: Chen Luo <chen.luo@databricks.com>

.config/hakari.toml

@@ -33,6 +33,7 @@ workspace-members = [
     "compute_api",
     "consumption_metrics",
     "desim",
+    "json",
     "metrics",
     "pageserver_api",
     "postgres_backend",

.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

@@ -32,162 +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$(sysctl -n hw.ncpu)
-
-      - name: Upload "pg_install/${{ 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
-
-      # `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
-        run: |
-          chmod +x pg_install/v*/bin/*
-
-      - 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) PG_INSTALL_CACHED=1
-
-      - 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
@@ -203,41 +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
@@ -249,17 +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

@@ -87,6 +87,24 @@ jobs:
     uses: ./.github/workflows/build-build-tools-image.yml
     secrets: inherit
 
+  lint-openapi-spec:
+    runs-on: ubuntu-22.04
+    needs: [ meta, check-permissions ]
+    # We do need to run this in `.*-rc-pr` because of hotfixes.
+    if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
+    steps:
+      - name: Harden the runner (Audit all outbound calls)
+        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
+        with:
+          egress-policy: audit
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - run: make lint-openapi-spec
+
   check-codestyle-python:
     needs: [ meta, check-permissions, build-build-tools-image ]
     # No need to run on `main` because we this in the merge queue. We do need to run this in `.*-rc-pr` because of hotfixes.
@@ -306,14 +324,14 @@ 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:
@@ -986,6 +1004,7 @@ jobs:
       - name: Verify docker-compose example and test extensions
         timeout-minutes: 60
         env:
+          PARALLEL_COMPUTES: 3
           TAG: >-
             ${{
               needs.meta.outputs.run-kind == 'compute-rc-pr'

.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/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

.gitignore

@@ -15,6 +15,7 @@ neon.iml
 /.neon
 /integration_tests/.neon
 compaction-suite-results.*
+docker-compose/docker-compose-parallel.yml
 
 # Coverage
 *.profraw

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"
@@ -1320,6 +1348,7 @@ dependencies = [
  "p256 0.13.2",
  "pageserver_page_api",
  "postgres",
+ "postgres-types",
  "postgres_initdb",
  "postgres_versioninfo",
  "regex",
@@ -3461,6 +3490,15 @@ dependencies = [
  "wasm-bindgen",
 ]
 
+[[package]]
+name = "json"
+version = "0.1.0"
+dependencies = [
+ "futures",
+ "itoa",
+ "ryu",
+]
+
 [[package]]
 name = "json-structural-diff"
 version = "0.2.0"
@@ -4302,6 +4340,7 @@ dependencies = [
  "arc-swap",
  "async-compression",
  "async-stream",
+ "base64 0.22.1",
  "bincode",
  "bit_field",
  "byteorder",
@@ -4455,6 +4494,25 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "pageserver_client_grpc"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "bytes",
+ "compute_api",
+ "futures",
+ "pageserver_api",
+ "pageserver_page_api",
+ "tokio",
+ "tokio-stream",
+ "tokio-util",
+ "tonic 0.13.1",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "pageserver_compaction"
 version = "0.1.0"
@@ -5647,6 +5705,8 @@ dependencies = [
  "azure_identity",
  "azure_storage",
  "azure_storage_blobs",
+ "base64 0.22.1",
+ "byteorder",
  "bytes",
  "camino",
  "camino-tempfile",
@@ -8665,8 +8725,10 @@ dependencies = [
  "fail",
  "form_urlencoded",
  "futures-channel",
+ "futures-core",
  "futures-executor",
  "futures-io",
+ "futures-sink",
  "futures-util",
  "generic-array",
  "getrandom 0.2.11",
@@ -8693,6 +8755,7 @@ dependencies = [
  "num-iter",
  "num-rational",
  "num-traits",
+ "once_cell",
  "p256 0.13.2",
  "parquet",
  "prettyplease",

Cargo.toml

@@ -8,6 +8,7 @@ members = [
     "pageserver/compaction",
     "pageserver/ctl",
     "pageserver/client",
+    "pageserver/client_grpc",
     "pageserver/pagebench",
     "pageserver/page_api",
     "proxy",
@@ -42,10 +43,12 @@ members = [
     "libs/walproposer",
     "libs/wal_decoder",
     "libs/postgres_initdb",
+    "libs/proxy/json",
     "libs/proxy/postgres-protocol2",
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
     "endpoint_storage",
+    "pgxn/neon/communicator",
 ]
 
 [workspace.package]
@@ -255,6 +258,7 @@ 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" }
@@ -284,6 +288,7 @@ 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
 
@@ -38,17 +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
 
@@ -56,23 +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 \
@@ -87,10 +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
 
@@ -130,12 +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

@@ -30,11 +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
@@ -102,7 +109,7 @@ 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-install walproposer-lib cargo-target-dir
 	+@echo "Compiling Neon"
@@ -115,10 +122,13 @@ cargo-target-dir:
 	test -e target/CACHEDIR.TAG || echo "$(CACHEDIR_TAG_CONTENTS)" > target/CACHEDIR.TAG
 
 .PHONY: neon-pg-ext-%
-neon-pg-ext-%: postgres-install-%
+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
 
@@ -210,6 +220,15 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit
 
+.PHONY: lint-openapi-spec
+lint-openapi-spec:
+	# operation-2xx-response: pageserver timeline delete returns 404 on success
+	find . -iname "openapi_spec.y*ml" -exec\
+		docker run --rm -v ${PWD}:/spec ghcr.io/redocly/cli:1.34.4\
+			--skip-rule=operation-operationId --skip-rule=operation-summary --extends=minimal\
+			--skip-rule=no-server-example.com --skip-rule=operation-2xx-response\
+			lint {} \+
+
 # Targets for building PostgreSQL are defined in postgres.mk.
 #
 # But if the caller has indicated that PostgreSQL is already

clippy.toml

@@ -1,9 +1,12 @@
 disallowed-methods = [
     "tokio::task::block_in_place",
+
     # Allow this for now, to deny it later once we stop using Handle::block_on completely
     # "tokio::runtime::Handle::block_on",
-    # use tokio_epoll_uring_ext instead
-    "tokio_epoll_uring::thread_local_system",
+
+    # tokio-epoll-uring:
+    # - allow-invalid because the method doesn't exist on macOS
+    { path = "tokio_epoll_uring::thread_local_system", replacement = "tokio_epoll_uring_ext module inside pageserver crate", allow-invalid = true }
 ]
 
 disallowed-macros = [

compute/compute-node.Dockerfile

@@ -1636,11 +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"
 
 #########################################################################################
 #
@@ -1912,10 +1915,10 @@ RUN cd /ext-src/pg_repack-src && patch -p1 </ext-src/pg_repack.patch && rm -f /e
 
 COPY --chmod=755 docker-compose/run-tests.sh /run-tests.sh
 RUN echo /usr/local/pgsql/lib > /etc/ld.so.conf.d/00-neon.conf && /sbin/ldconfig
-RUN apt-get update && apt-get install -y libtap-parser-sourcehandler-pgtap-perl jq \
+RUN apt-get update && apt-get install -y libtap-parser-sourcehandler-pgtap-perl jq parallel \
    && apt clean && rm -rf /ext-src/*.tar.gz /ext-src/*.patch /var/lib/apt/lists/*
 ENV PATH=/usr/local/pgsql/bin:$PATH
-ENV PGHOST=compute
+ENV PGHOST=compute1
 ENV PGPORT=55433
 ENV PGUSER=cloud_admin
 ENV PGDATABASE=postgres

compute_tools/Cargo.toml

@@ -66,7 +66,7 @@ url.workspace = true
 uuid.workspace = true
 walkdir.workspace = true
 x509-cert.workspace = true
-
+postgres-types.workspace = true
 postgres_versioninfo.workspace = true
 postgres_initdb.workspace = true
 compute_api.workspace = true

compute_tools/src/compute.rs

@@ -3,7 +3,7 @@ use chrono::{DateTime, Utc};
 use compute_api::privilege::Privilege;
 use compute_api::responses::{
     ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus, LfcOffloadState,
-    LfcPrewarmState, TlsConfig,
+    LfcPrewarmState, PromoteState, TlsConfig,
 };
 use compute_api::spec::{
     ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
@@ -29,7 +29,7 @@ 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::{spawn, sync::watch, task::JoinHandle, time};
 use tracing::{Instrument, debug, error, info, instrument, warn};
 use url::Url;
 use utils::id::{TenantId, TimelineId};
@@ -107,6 +107,8 @@ pub struct ComputeNodeParams {
     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,
@@ -129,7 +131,8 @@ pub struct ComputeNode {
     pub compute_ctl_config: ComputeCtlConfig,
 
     /// Handle to the extension stats collection task
-    extension_stats_task: Mutex<Option<tokio::task::JoinHandle<()>>>,
+    extension_stats_task: TaskHandle,
+    lfc_offload_task: TaskHandle,
 }
 
 // store some metrics about download size that might impact startup time
@@ -171,6 +174,7 @@ pub struct ComputeState {
     /// WAL flush LSN that is set after terminating Postgres and syncing safekeepers if
     /// mode == ComputeMode::Primary. None otherwise
     pub terminate_flush_lsn: Option<Lsn>,
+    pub promote_state: Option<watch::Receiver<PromoteState>>,
 
     pub metrics: ComputeMetrics,
 }
@@ -188,6 +192,7 @@ impl ComputeState {
             lfc_prewarm_state: LfcPrewarmState::default(),
             lfc_offload_state: LfcOffloadState::default(),
             terminate_flush_lsn: None,
+            promote_state: None,
         }
     }
 
@@ -368,7 +373,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
 
 struct PostgresHandle {
     postgres: std::process::Child,
-    log_collector: tokio::task::JoinHandle<Result<()>>,
+    log_collector: JoinHandle<Result<()>>,
 }
 
 impl PostgresHandle {
@@ -382,7 +387,7 @@ 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 {
@@ -433,6 +438,7 @@ impl ComputeNode {
             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),
         })
     }
 
@@ -520,8 +526,8 @@ impl ComputeNode {
             None
         };
 
-        // Terminate the extension stats collection task
         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.
@@ -851,12 +857,15 @@ 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(())
     }
 
@@ -1049,7 +1058,7 @@ impl ComputeNode {
         };
 
         let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
-            let mut client = page_api::Client::new(
+            let mut client = page_api::Client::connect(
                 shard0_connstr,
                 spec.tenant_id,
                 spec.timeline_id,
@@ -2357,10 +2366,7 @@ LIMIT 100",
     }
 
     pub fn spawn_extension_stats_task(&self) {
-        // Cancel any existing task
-        if let Some(handle) = self.extension_stats_task.lock().unwrap().take() {
-            handle.abort();
-        }
+        self.terminate_extension_stats_task();
 
         let conf = self.tokio_conn_conf.clone();
         let atomic_interval = self.params.installed_extensions_collection_interval.clone();
@@ -2396,8 +2402,30 @@ LIMIT 100",
     }
 
     fn terminate_extension_stats_task(&self) {
-        if let Some(handle) = self.extension_stats_task.lock().unwrap().take() {
-            handle.abort();
+        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()
         }
     }
 
@@ -2406,19 +2434,11 @@ LIMIT 100",
         // 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 {
@@ -106,7 +105,8 @@ impl ComputeNode {
                 cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Completed;
                 return;
             };
-            error!(%err);
+            crate::metrics::LFC_PREWARM_ERRORS.inc();
+            error!(%err, "prewarming lfc");
             cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Failed {
                 error: err.to_string(),
             };
@@ -152,32 +152,42 @@ 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;
+        };
+        crate::metrics::LFC_OFFLOAD_ERRORS.inc();
+        error!(%err, "offloading lfc");
+        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/compute_promote.rs

@@ -0,0 +1,132 @@
+use crate::compute::ComputeNode;
+use anyhow::{Context, Result, bail};
+use compute_api::{
+    responses::{LfcPrewarmState, PromoteState, SafekeepersLsn},
+    spec::ComputeMode,
+};
+use std::{sync::Arc, time::Duration};
+use tokio::time::sleep;
+use utils::lsn::Lsn;
+
+impl ComputeNode {
+    /// Returns only when promote fails or succeeds. If a network error occurs
+    /// and http client disconnects, this does not stop promotion, and subsequent
+    /// calls block until promote finishes.
+    /// Called by control plane on secondary after primary endpoint is terminated
+    pub async fn promote(self: &Arc<Self>, safekeepers_lsn: SafekeepersLsn) -> PromoteState {
+        let cloned = self.clone();
+        let start_promotion = || {
+            let (tx, rx) = tokio::sync::watch::channel(PromoteState::NotPromoted);
+            tokio::spawn(async move {
+                tx.send(match cloned.promote_impl(safekeepers_lsn).await {
+                    Ok(_) => PromoteState::Completed,
+                    Err(err) => {
+                        tracing::error!(%err, "promoting");
+                        PromoteState::Failed {
+                            error: err.to_string(),
+                        }
+                    }
+                })
+            });
+            rx
+        };
+
+        let mut task;
+        // self.state is unlocked after block ends so we lock it in promote_impl
+        // and task.changed() is reached
+        {
+            task = self
+                .state
+                .lock()
+                .unwrap()
+                .promote_state
+                .get_or_insert_with(start_promotion)
+                .clone()
+        }
+        task.changed().await.expect("promote sender dropped");
+        task.borrow().clone()
+    }
+
+    // Why do we have to supply safekeepers?
+    // For secondary we use primary_connection_conninfo so safekeepers field is empty
+    async fn promote_impl(&self, safekeepers_lsn: SafekeepersLsn) -> Result<()> {
+        {
+            let state = self.state.lock().unwrap();
+            let mode = &state.pspec.as_ref().unwrap().spec.mode;
+            if *mode != ComputeMode::Replica {
+                bail!("{} is not replica", mode.to_type_str());
+            }
+
+            // we don't need to query Postgres so not self.lfc_prewarm_state()
+            match &state.lfc_prewarm_state {
+                LfcPrewarmState::NotPrewarmed | LfcPrewarmState::Prewarming => {
+                    bail!("prewarm not requested or pending")
+                }
+                LfcPrewarmState::Failed { error } => {
+                    tracing::warn!(%error, "replica prewarm failed")
+                }
+                _ => {}
+            }
+        }
+
+        let client = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+            .await
+            .context("connecting to postgres")?;
+
+        let primary_lsn = safekeepers_lsn.wal_flush_lsn;
+        let mut last_wal_replay_lsn: Lsn = Lsn::INVALID;
+        const RETRIES: i32 = 20;
+        for i in 0..=RETRIES {
+            let row = client
+                .query_one("SELECT pg_last_wal_replay_lsn()", &[])
+                .await
+                .context("getting last replay lsn")?;
+            let lsn: u64 = row.get::<usize, postgres_types::PgLsn>(0).into();
+            last_wal_replay_lsn = lsn.into();
+            if last_wal_replay_lsn >= primary_lsn {
+                break;
+            }
+            tracing::info!("Try {i}, replica lsn {last_wal_replay_lsn}, primary lsn {primary_lsn}");
+            sleep(Duration::from_secs(1)).await;
+        }
+        if last_wal_replay_lsn < primary_lsn {
+            bail!("didn't catch up with primary in {RETRIES} retries");
+        }
+
+        // using $1 doesn't work with ALTER SYSTEM SET
+        let safekeepers_sql = format!(
+            "ALTER SYSTEM SET neon.safekeepers='{}'",
+            safekeepers_lsn.safekeepers
+        );
+        client
+            .query(&safekeepers_sql, &[])
+            .await
+            .context("setting safekeepers")?;
+        client
+            .query("SELECT pg_reload_conf()", &[])
+            .await
+            .context("reloading postgres config")?;
+        let row = client
+            .query_one("SELECT * FROM pg_promote()", &[])
+            .await
+            .context("pg_promote")?;
+        if !row.get::<usize, bool>(0) {
+            bail!("pg_promote() returned false");
+        }
+
+        let client = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+            .await
+            .context("connecting to postgres")?;
+        let row = client
+            .query_one("SHOW transaction_read_only", &[])
+            .await
+            .context("getting transaction_read_only")?;
+        if row.get::<usize, &str>(0) == "on" {
+            bail!("replica in read only mode after promotion");
+        }
+
+        let mut state = self.state.lock().unwrap();
+        state.pspec.as_mut().unwrap().spec.mode = ComputeMode::Primary;
+        Ok(())
+    }
+}

compute_tools/src/http/openapi_spec.yaml

@@ -83,6 +83,87 @@ paths:
               schema:
                 $ref: "#/components/schemas/DbsAndRoles"
 
+  /promote:
+    post:
+      tags:
+        - Promotion
+      summary: Promote secondary replica to primary
+      description: ""
+      operationId: promoteReplica
+      requestBody:
+        description: Promote requests data
+        required: true
+        content:
+          application/json:
+            schema:
+                $ref: "#/components/schemas/SafekeepersLsn"
+      responses:
+        200:
+          description: Promote succeeded or wasn't started
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PromoteState"
+        500:
+          description: Promote failed
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PromoteState"
+
+  /lfc/prewarm:
+    post:
+      summary: Request LFC Prewarm
+      parameters:
+        - name: from_endpoint
+          in: query
+          schema:
+            type: string
+      description: ""
+      operationId: lfcPrewarm
+      responses:
+        202:
+          description: LFC prewarm started
+        429:
+          description: LFC prewarm ongoing
+    get:
+      tags:
+        - Prewarm
+      summary: Get LFC prewarm state
+      description: ""
+      operationId: getLfcPrewarmState
+      responses:
+        200:
+          description: Prewarm state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/LfcPrewarmState"
+
+  /lfc/offload:
+    post:
+      summary: Request LFC offload
+      description: ""
+      operationId: lfcOffload
+      responses:
+        202:
+          description: LFC offload started
+        429:
+          description: LFC offload ongoing
+    get:
+      tags:
+        - Prewarm
+      summary: Get LFC offloading state
+      description: ""
+      operationId: getLfcOffloadState
+      responses:
+        200:
+          description: Offload state
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/LfcOffloadState"
+
   /database_schema:
     get:
       tags:
@@ -335,15 +416,6 @@ components:
         total_startup_ms:
           type: integer
 
-    Info:
-      type: object
-      description: Information about VM/Pod.
-      required:
-        - num_cpus
-      properties:
-        num_cpus:
-          type: integer
-
     DbsAndRoles:
       type: object
       description: Databases and Roles
@@ -497,25 +569,69 @@ components:
           type: string
           example: "1.0.0"
 
-    InstalledExtensions:
+    SafekeepersLsn:
       type: object
+      required:
+        - safekeepers
+        - wal_flush_lsn
       properties:
-        extensions:
-          description: Contains list of installed extensions.
-          type: array
-          items:
-            type: object
-            properties:
-              extname:
-                type: string
-              version:
-                type: string
-                items:
-                  type: string
-              n_databases:
-                type: integer
-              owned_by_superuser:
-                type: integer
+        safekeepers:
+          description: Primary replica safekeepers
+          type: string
+        wal_flush_lsn:
+          description: Primary last WAL flush LSN
+          type: string
+
+    LfcPrewarmState:
+      type: object
+      required:
+        - status
+        - total
+        - prewarmed
+        - skipped
+      properties:
+        status:
+          description: Lfc prewarm status
+          enum: [not_prewarmed, prewarming, completed, failed]
+          type: string
+        error:
+          description: Lfc prewarm error, if any
+          type: string
+        total:
+          description: Total pages processed
+          type: integer
+        prewarmed:
+          description: Total pages prewarmed
+          type: integer
+        skipped:
+          description: Pages processed but not prewarmed
+          type: integer
+
+    LfcOffloadState:
+      type: object
+      required:
+        - status
+      properties:
+        status:
+          description: Lfc offload status
+          enum: [not_offloaded, offloading, completed, failed]
+          type: string
+        error:
+          description: Lfc offload error, if any
+          type: string
+
+    PromoteState:
+      type: object
+      required:
+        - status
+      properties:
+        status:
+          description: Promote result
+          enum: [not_promoted, completed, failed]
+          type: string
+        error:
+          description: Promote error, if any
+          type: string
 
     SetRoleGrantsRequest:
       type: object

compute_tools/src/http/routes/mod.rs

@@ -14,6 +14,7 @@ pub(in crate::http) mod insights;
 pub(in crate::http) mod lfc;
 pub(in crate::http) mod metrics;
 pub(in crate::http) mod metrics_json;
+pub(in crate::http) mod promote;
 pub(in crate::http) mod status;
 pub(in crate::http) mod terminate;
 

compute_tools/src/http/routes/promote.rs

@@ -0,0 +1,14 @@
+use crate::http::JsonResponse;
+use axum::Form;
+use http::StatusCode;
+
+pub(in crate::http) async fn promote(
+    compute: axum::extract::State<std::sync::Arc<crate::compute::ComputeNode>>,
+    Form(safekeepers_lsn): Form<compute_api::responses::SafekeepersLsn>,
+) -> axum::response::Response {
+    let state = compute.promote(safekeepers_lsn).await;
+    if let compute_api::responses::PromoteState::Failed { error } = state {
+        return JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, error);
+    }
+    JsonResponse::success(StatusCode::OK, state)
+}

compute_tools/src/http/server.rs

@@ -23,7 +23,7 @@ use super::{
     middleware::authorize::Authorize,
     routes::{
         check_writability, configure, database_schema, dbs_and_roles, extension_server, extensions,
-        grants, insights, lfc, metrics, metrics_json, status, terminate,
+        grants, insights, lfc, metrics, metrics_json, promote, status, terminate,
     },
 };
 use crate::compute::ComputeNode;
@@ -87,6 +87,7 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/lfc/prewarm", get(lfc::prewarm_state).post(lfc::prewarm))
                     .route("/lfc/offload", get(lfc::offload_state).post(lfc::offload))
+                    .route("/promote", post(promote::promote))
                     .route("/check_writability", post(check_writability::is_writable))
                     .route("/configure", post(configure::configure))
                     .route("/database_schema", get(database_schema::get_schema_dump))

compute_tools/src/lib.rs

@@ -12,6 +12,7 @@ pub mod logger;
 pub mod catalog;
 pub mod compute;
 pub mod compute_prewarm;
+pub mod compute_promote;
 pub mod disk_quota;
 pub mod extension_server;
 pub mod installed_extensions;

compute_tools/src/lsn_lease.rs

@@ -192,7 +192,7 @@ fn acquire_lsn_lease_grpc(
     lsn: Lsn,
 ) -> Result<Option<SystemTime>> {
     tokio::runtime::Handle::current().block_on(async move {
-        let mut client = page_api::Client::new(
+        let mut client = page_api::Client::connect(
             connstring.to_string(),
             tenant_shard_id.tenant_id,
             timeline_id,

compute_tools/src/metrics.rs

@@ -97,20 +97,34 @@ 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_PREWARM_ERRORS: 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_prewarm_errors_total",
+        "Total number of LFC prewarm errors",
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static LFC_OFFLOADS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "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")
+});
+
+pub(crate) static LFC_OFFLOAD_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "compute_ctl_lfc_offload_errors_total",
+        "Total number of LFC offload errors",
     )
     .expect("failed to define a metric")
 });
@@ -124,7 +138,9 @@ 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_PREWARM_ERRORS.collect());
+    metrics.extend(LFC_OFFLOADS.collect());
+    metrics.extend(LFC_OFFLOAD_ERRORS.collect());
     metrics
 }

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -0,0 +1 @@
+GRANT pg_signal_backend TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/tests/0004-grant_pg_monitor_to_neon_superuser.sql

@@ -7,13 +7,17 @@ BEGIN
         INTO monitor
         FROM pg_auth_members
         WHERE roleid = 'pg_monitor'::regrole
-            AND member = 'pg_monitor'::regrole;
+            AND member = 'neon_superuser'::regrole;
 
-    IF NOT monitor.member THEN
+    IF monitor IS NULL THEN
+        RAISE EXCEPTION 'no entry in pg_auth_members for neon_superuser and pg_monitor';
+    END IF;
+
+    IF monitor.admin IS NULL OR NOT monitor.member THEN
         RAISE EXCEPTION 'neon_superuser is not a member of pg_monitor';
     END IF;
 
-    IF NOT monitor.admin THEN
+    IF monitor.admin IS NULL OR NOT monitor.admin THEN
         RAISE EXCEPTION 'neon_superuser cannot grant pg_monitor';
     END IF;
 END $$;

compute_tools/src/migrations/tests/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -0,0 +1,23 @@
+DO $$
+DECLARE
+    signal_backend record;
+BEGIN
+    SELECT pg_has_role('neon_superuser', 'pg_signal_backend', 'member') AS member,
+            admin_option AS admin
+        INTO signal_backend
+        FROM pg_auth_members
+        WHERE roleid = 'pg_signal_backend'::regrole
+            AND member = 'neon_superuser'::regrole;
+
+    IF signal_backend IS NULL THEN
+        RAISE EXCEPTION 'no entry in pg_auth_members for neon_superuser and pg_signal_backend';
+    END IF;
+
+    IF signal_backend.member IS NULL OR NOT signal_backend.member THEN
+        RAISE EXCEPTION 'neon_superuser is not a member of pg_signal_backend';
+    END IF;
+
+    IF signal_backend.admin IS NULL OR NOT signal_backend.admin THEN
+        RAISE EXCEPTION 'neon_superuser cannot grant pg_signal_backend';
+    END IF;
+END $$;

compute_tools/src/spec.rs

@@ -197,6 +197,7 @@ pub async fn handle_migrations(client: &mut Client) -> Result<()> {
         include_str!(
             "./migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql"
         ),
+        include_str!("./migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql"),
     ];
 
     MigrationRunner::new(client, &migrations)

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

@@ -675,6 +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",
@@ -1585,22 +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;

control_plane/src/endpoint.rs

@@ -373,6 +373,22 @@ impl std::fmt::Display for EndpointTerminateMode {
     }
 }
 
+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 {
     fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
         if !entry.file_type()?.is_dir() {
@@ -677,21 +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");
         }
@@ -704,10 +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
@@ -735,7 +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,
@@ -744,7 +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(),
@@ -766,20 +768,21 @@ 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.
             };
 
@@ -791,7 +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,
@@ -826,7 +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}'");
         }
@@ -858,11 +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");
         }
 
@@ -894,10 +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
@@ -925,9 +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
                         ));
                     }
                 }

control_plane/src/pageserver.rs

@@ -452,6 +452,12 @@ impl PageServerNode {
                 .map(|x| x.parse::<usize>())
                 .transpose()
                 .context("Failed to parse 'image_creation_threshold' as non zero integer")?,
+            // HADRON
+            image_layer_force_creation_period: settings
+                .remove("image_layer_force_creation_period")
+                .map(humantime::parse_duration)
+                .transpose()
+                .context("Failed to parse 'image_layer_force_creation_period' as duration")?,
             image_layer_creation_check_threshold: settings
                 .remove("image_layer_creation_check_threshold")
                 .map(|x| x.parse::<u8>())

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/shell/compute.sh

@@ -54,14 +54,16 @@ else
     printf '%s\n' "${result}" | jq .
   fi
 
-  echo "Check if a timeline present"
-  PARAMS=(
-       -X GET
-       -H "Content-Type: application/json"
-       "http://pageserver:9898/v1/tenant/${tenant_id}/timeline"
-  )
-  timeline_id=$(curl "${PARAMS[@]}" | jq -r .[0].timeline_id)
-  if [[ -z "${timeline_id}" || "${timeline_id}" = null ]]; then
+  if [[ "${RUN_PARALLEL:-false}" != "true" ]]; then
+    echo "Check if a timeline present"
+    PARAMS=(
+         -X GET
+         -H "Content-Type: application/json"
+        "http://pageserver:9898/v1/tenant/${tenant_id}/timeline"
+    )
+    timeline_id=$(curl "${PARAMS[@]}" | jq -r .[0].timeline_id)
+  fi
+  if [[ -z "${timeline_id:-}" || "${timeline_id:-}" = null ]]; then
     generate_id timeline_id
     PARAMS=(
         -sbf

docker-compose/docker-compose.yml

@@ -142,7 +142,7 @@ services:
       - "storage_broker"
       - "--listen-addr=0.0.0.0:50051"
 
-  compute:
+  compute1:
     restart: always
     build:
       context: ./compute_wrapper/
@@ -152,6 +152,7 @@ services:
         - TAG=${COMPUTE_TAG:-${TAG:-latest}}
         - http_proxy=${http_proxy:-}
         - https_proxy=${https_proxy:-}
+    image: built-compute
     environment:
       - PG_VERSION=${PG_VERSION:-16}
       - TENANT_ID=${TENANT_ID:-}
@@ -166,6 +167,11 @@ services:
       - 3080:3080 # http endpoints
     entrypoint:
       - "/shell/compute.sh"
+    # Ad an alias for compute1 for compatibility
+    networks:
+      default:
+        aliases:
+            - compute
     depends_on:
       - safekeeper1
       - safekeeper2
@@ -174,15 +180,20 @@ services:
 
   compute_is_ready:
     image: postgres:latest
+    environment:
+      - PARALLEL_COMPUTES=1
     entrypoint:
-      - "/bin/bash"
+      - "/bin/sh"
       - "-c"
     command:
-      - "until pg_isready -h compute -p 55433 -U cloud_admin ; do
-            echo 'Waiting to start compute...' && sleep 1;
-         done"
+      - "for i in $(seq 1 $${PARALLEL_COMPUTES}); do
+           until pg_isready -h compute$$i -p 55433 -U cloud_admin ; do
+             sleep 1;
+           done;
+         done;
+         echo All computes are started"
     depends_on:
-      - compute
+      - compute1
 
   neon-test-extensions:
     profiles: ["test-extensions"]
@@ -196,4 +207,4 @@ services:
     command:
       - sleep 3600
     depends_on:
-      - compute
+      - compute1

docker-compose/docker_compose_test.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # A basic test to ensure Docker images are built correctly.
 # Build a wrapper around the compute, start all services and runs a simple SQL query.
@@ -13,9 +13,36 @@
 #
 set -eux -o pipefail
 
+cd "$(dirname "${0}")"
 export COMPOSE_FILE='docker-compose.yml'
 export COMPOSE_PROFILES=test-extensions
-cd "$(dirname "${0}")"
+export PARALLEL_COMPUTES=${PARALLEL_COMPUTES:-1}
+READY_MESSAGE="All computes are started"
+COMPUTES=()
+for i in $(seq 1 "${PARALLEL_COMPUTES}"); do
+  COMPUTES+=("compute${i}")
+done
+CURRENT_TMPDIR=$(mktemp -d)
+trap 'rm -rf ${CURRENT_TMPDIR} docker-compose-parallel.yml' EXIT
+if [[ ${PARALLEL_COMPUTES} -gt 1 ]]; then
+  export COMPOSE_FILE=docker-compose-parallel.yml
+  cp docker-compose.yml docker-compose-parallel.yml
+  # Replace the environment variable PARALLEL_COMPUTES with the actual value
+  yq eval -i ".services.compute_is_ready.environment |=  map(select(. | test(\"^PARALLEL_COMPUTES=\") | not)) + [\"PARALLEL_COMPUTES=${PARALLEL_COMPUTES}\"]" ${COMPOSE_FILE}
+  for i in $(seq 2 "${PARALLEL_COMPUTES}"); do
+    # Duplicate compute1 as compute${i} for parallel execution
+    yq eval -i ".services.compute${i} = .services.compute1" ${COMPOSE_FILE}
+    # We don't need these sections, so delete them
+    yq eval -i "(del .services.compute${i}.build) | (del .services.compute${i}.ports) | (del .services.compute${i}.networks)" ${COMPOSE_FILE}
+    # Let the compute 1 be the only dependence
+    yq eval -i ".services.compute${i}.depends_on = [\"compute1\"]" ${COMPOSE_FILE}
+    # Set RUN_PARALLEL=true for compute2. They will generate tenant_id and timeline_id to avoid using the same as other computes
+    yq eval -i ".services.compute${i}.environment += [\"RUN_PARALLEL=true\"]" ${COMPOSE_FILE}
+    # Remove TENANT_ID and TIMELINE_ID from the environment variables of the generated computes
+    # They will create new TENANT_ID and TIMELINE_ID anyway.
+    yq eval -i ".services.compute${i}.environment |= map(select(. | (test(\"^TENANT_ID=\") or test(\"^TIMELINE_ID=\")) | not))" ${COMPOSE_FILE}
+  done
+fi
 PSQL_OPTION="-h localhost -U cloud_admin -p 55433 -d postgres"
 
 function cleanup() {
@@ -27,11 +54,11 @@ function cleanup() {
 
 for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
     pg_version=${pg_version/v/}
-    echo "clean up containers if exists"
+    echo "clean up containers if exist"
     cleanup
     PG_TEST_VERSION=$((pg_version < 16 ? 16 : pg_version))
-    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose up --quiet-pull --build -d
-
+    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose build compute1
+    PG_VERSION=${pg_version} PG_TEST_VERSION=${PG_TEST_VERSION} docker compose up --quiet-pull -d
     echo "wait until the compute is ready. timeout after 60s. "
     cnt=0
     while sleep 3; do
@@ -41,45 +68,50 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
             echo "timeout before the compute is ready."
             exit 1
         fi
-        if docker compose logs "compute_is_ready" | grep -q "accepting connections"; then
+        if docker compose logs compute_is_ready | grep -q "${READY_MESSAGE}"; then
             echo "OK. The compute is ready to connect."
             echo "execute simple queries."
-            docker compose exec compute /bin/bash -c "psql ${PSQL_OPTION} -c 'SELECT 1'"
+            for compute in "${COMPUTES[@]}"; do
+              docker compose exec "${compute}" /bin/bash -c "psql ${PSQL_OPTION} -c 'SELECT 1'"
+            done
             break
         fi
     done
 
     if [[ ${pg_version} -ge 16 ]]; then
-        # This is required for the pg_hint_plan test, to prevent flaky log message causing the test to fail
-        # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
-        echo Adding dummy config
-        docker compose exec compute touch /var/db/postgres/compute/compute_ctl_temp_override.conf
-        # Prepare for the PostGIS test
-        docker compose exec compute mkdir -p /tmp/pgis_reg/pgis_reg_tmp
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/ext-src/postgis-src/raster/test "${TMPDIR}"
-        docker compose cp neon-test-extensions:/ext-src/postgis-src/regress/00-regress-install "${TMPDIR}"
-        docker compose exec compute mkdir -p /ext-src/postgis-src/raster /ext-src/postgis-src/regress /ext-src/postgis-src/regress/00-regress-install
-        docker compose cp "${TMPDIR}/test" compute:/ext-src/postgis-src/raster/test
-        docker compose cp "${TMPDIR}/00-regress-install" compute:/ext-src/postgis-src/regress
-        rm -rf "${TMPDIR}"
-        # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/ext-src/pg_hint_plan-src/data "${TMPDIR}/data"
-        docker compose cp "${TMPDIR}/data" compute:/ext-src/pg_hint_plan-src/
-        rm -rf "${TMPDIR}"
-        # The following block does the same for the contrib/file_fdw test
-        TMPDIR=$(mktemp -d)
-        docker compose cp neon-test-extensions:/postgres/contrib/file_fdw/data "${TMPDIR}/data"
-        docker compose cp "${TMPDIR}/data" compute:/postgres/contrib/file_fdw/data
-        rm -rf "${TMPDIR}"
+        mkdir "${CURRENT_TMPDIR}"/{pg_hint_plan-src,file_fdw,postgis-src}
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/raster/test "${CURRENT_TMPDIR}/postgis-src/test"
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/regress/00-regress-install "${CURRENT_TMPDIR}/postgis-src/00-regress-install"
+        docker compose cp neon-test-extensions:/ext-src/pg_hint_plan-src/data "${CURRENT_TMPDIR}/pg_hint_plan-src/data"
+        docker compose cp neon-test-extensions:/postgres/contrib/file_fdw/data "${CURRENT_TMPDIR}/file_fdw/data"
+
+        for compute in "${COMPUTES[@]}"; do
+          # This is required for the pg_hint_plan test, to prevent flaky log message causing the test to fail
+          # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
+          echo Adding dummy config on "${compute}"
+          docker compose exec "${compute}" touch /var/db/postgres/compute/compute_ctl_temp_override.conf
+          # Prepare for the PostGIS test
+          docker compose exec "${compute}" mkdir -p /tmp/pgis_reg/pgis_reg_tmp /ext-src/postgis-src/raster /ext-src/postgis-src/regress /ext-src/postgis-src/regress/00-regress-install
+          docker compose cp "${CURRENT_TMPDIR}/postgis-src/test" "${compute}":/ext-src/postgis-src/raster/test
+          docker compose cp "${CURRENT_TMPDIR}/postgis-src/00-regress-install" "${compute}":/ext-src/postgis-src/regress
+          # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
+          docker compose cp "${CURRENT_TMPDIR}/pg_hint_plan-src/data" "${compute}":/ext-src/pg_hint_plan-src/
+          # The following block does the same for the contrib/file_fdw test
+          docker compose cp "${CURRENT_TMPDIR}/file_fdw/data" "${compute}":/postgres/contrib/file_fdw/data
+        done
         # Apply patches
         docker compose exec -T neon-test-extensions bash -c "(cd /postgres && patch -p1)" <"../compute/patches/contrib_pg${pg_version}.patch"
         # We are running tests now
         rm -f testout.txt testout_contrib.txt
+        # We want to run the longest tests first to better utilize parallelization and reduce overall test time.
+        # Tests listed in the RUN_FIRST variable will be run before others.
+        # If parallelization is not used, this environment variable will be ignored.
+
         docker compose exec -e USE_PGXS=1 -e SKIP=timescaledb-src,rdkit-src,pg_jsonschema-src,kq_imcx-src,wal2json_2_5-src,rag_jina_reranker_v1_tiny_en-src,rag_bge_small_en_v15-src \
+        -e RUN_FIRST=hll-src,postgis-src,pgtap-src -e PARALLEL_COMPUTES="${PARALLEL_COMPUTES}" \
         neon-test-extensions /run-tests.sh /ext-src | tee testout.txt && EXT_SUCCESS=1 || EXT_SUCCESS=0
         docker compose exec -e SKIP=start-scripts,postgres_fdw,ltree_plpython,jsonb_plpython,jsonb_plperl,hstore_plpython,hstore_plperl,dblink,bool_plperl \
+        -e PARALLEL_COMPUTES="${PARALLEL_COMPUTES}" \
         neon-test-extensions /run-tests.sh /postgres/contrib | tee testout_contrib.txt && CONTRIB_SUCCESS=1 || CONTRIB_SUCCESS=0
         if [[ ${EXT_SUCCESS} -eq 0 || ${CONTRIB_SUCCESS} -eq 0 ]]; then
             CONTRIB_FAILED=

docker-compose/run-tests.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -x
 
 if [[ -v BENCHMARK_CONNSTR ]]; then
@@ -26,8 +26,9 @@ if [[ -v BENCHMARK_CONNSTR ]]; then
   fi
 fi
 REGULAR_USER=false
-while getopts r arg; do
-  case $arg in
+PARALLEL_COMPUTES=${PARALLEL_COMPUTES:-1}
+while getopts pr arg; do
+  case ${arg} in
   r)
     REGULAR_USER=true
     shift $((OPTIND-1))
@@ -41,26 +42,49 @@ extdir=${1}
 
 cd "${extdir}" || exit 2
 FAILED=
-LIST=$( (echo -e "${SKIP//","/"\n"}"; ls) | sort | uniq -u)
-for d in ${LIST}; do
-    [ -d "${d}" ] || continue
-    if ! psql -w -c "select 1" >/dev/null; then
-      FAILED="${d} ${FAILED}"
-      break
-    fi
-    if [[ ${REGULAR_USER} = true ]] && [ -f "${d}"/regular-test.sh ]; then
-       "${d}/regular-test.sh" || FAILED="${d} ${FAILED}"
-       continue
-    fi
+export FAILED_FILE=/tmp/failed
+rm -f ${FAILED_FILE}
+mapfile -t LIST < <( (echo -e "${SKIP//","/"\n"}"; ls) | sort | uniq -u)
+if [[ ${PARALLEL_COMPUTES} -gt 1 ]]; then
+  # Avoid errors if RUN_FIRST is not defined
+  RUN_FIRST=${RUN_FIRST:-}
+  # Move entries listed in the RUN_FIRST variable to the beginning
+  ORDERED_LIST=$(printf "%s\n" "${LIST[@]}" | grep -x -Ff <(echo -e "${RUN_FIRST//,/$'\n'}"); printf "%s\n" "${LIST[@]}" | grep -vx -Ff <(echo -e "${RUN_FIRST//,/$'\n'}"))
+  parallel -j"${PARALLEL_COMPUTES}" "[[ -d {} ]] || exit 0
+                export PGHOST=compute{%}
+                if ! psql -c 'select 1'>/dev/null; then
+                  exit 1
+                fi
+                echo Running on \${PGHOST}
+                if [[ -f ${extdir}/{}/neon-test.sh ]]; then
+                  echo Running from script
+                  ${extdir}/{}/neon-test.sh || echo {} >> ${FAILED_FILE};
+                else
+                  echo Running using make;
+                  USE_PGXS=1 make -C {} installcheck || echo {} >> ${FAILED_FILE};
+                fi" ::: ${ORDERED_LIST}
+  [[ ! -f ${FAILED_FILE} ]] && exit 0
+else
+  for d in "${LIST[@]}"; do
+      [ -d "${d}" ] || continue
+      if ! psql -w -c "select 1" >/dev/null; then
+        FAILED="${d} ${FAILED}"
+        break
+      fi
+      if [[ ${REGULAR_USER} = true ]] && [ -f "${d}"/regular-test.sh ]; then
+        "${d}/regular-test.sh" || FAILED="${d} ${FAILED}"
+        continue
+      fi
 
-    if [ -f "${d}/neon-test.sh" ]; then
-       "${d}/neon-test.sh" || FAILED="${d} ${FAILED}"
-    else
-       USE_PGXS=1 make -C "${d}" installcheck || FAILED="${d} ${FAILED}"
-    fi
-done
-[ -z "${FAILED}" ] && exit 0
-for d in ${FAILED}; do
+      if [ -f "${d}/neon-test.sh" ]; then
+        "${d}/neon-test.sh" || FAILED="${d} ${FAILED}"
+      else
+        USE_PGXS=1 make -C "${d}" installcheck || FAILED="${d} ${FAILED}"
+      fi
+  done
+  [[ -z ${FAILED} ]]  && exit 0
+fi
+for d in ${FAILED} $([[ ! -f ${FAILED_FILE} ]] || cat ${FAILED_FILE}); do
   cat "$(find $d -name regression.diffs)"
 done
 for postgis_diff in /tmp/pgis_reg/*_diff; do
@@ -68,4 +92,5 @@ for postgis_diff in /tmp/pgis_reg/*_diff; do
   cat "${postgis_diff}"
 done
 echo "${FAILED}"
+cat ${FAILED_FILE}
 exit 1

docker-compose/test_extensions_upgrade.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -eux -o pipefail
 cd "$(dirname "${0}")"
 # Takes a variable name as argument. The result is stored in that variable.
@@ -60,8 +60,8 @@ function check_timeline() {
 # Restarts the compute node with the required compute tag and timeline.
 # Accepts the tag for the compute node and the timeline as parameters.
 function restart_compute() {
-  docker compose down compute compute_is_ready
-  COMPUTE_TAG=${1} TENANT_ID=${tenant_id} TIMELINE_ID=${2} docker compose up --quiet-pull -d --build compute compute_is_ready
+  docker compose down compute1 compute_is_ready
+  COMPUTE_TAG=${1} TENANT_ID=${tenant_id} TIMELINE_ID=${2} docker compose up --quiet-pull -d --build compute1 compute_is_ready
   wait_for_ready
   check_timeline ${2}
 }

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.

endpoint_storage/src/app.rs

@@ -13,6 +13,8 @@ use utils::backoff::retry;
 pub fn app(state: Arc<Storage>) -> Router<()> {
     use axum::routing::{delete as _delete, get as _get};
     let delete_prefix = _delete(delete_prefix);
+    // NB: On any changes do not forget to update the OpenAPI spec
+    // in /endpoint_storage/src/openapi_spec.yml.
     Router::new()
         .route(
             "/{tenant_id}/{timeline_id}/{endpoint_id}/{*path}",

endpoint_storage/src/openapi_spec.yml

@@ -0,0 +1,146 @@
+openapi: "3.0.2"
+info:
+  title: Endpoint Storage API
+  description: Endpoint Storage API
+  version: "1.0"
+  license:
+    name: "Apache"
+    url: https://github.com/neondatabase/neon/blob/main/LICENSE
+servers:
+  - url: ""
+paths:
+  /status:
+    description: Healthcheck endpoint
+    get:
+      description: Healthcheck
+      security: []
+      responses:
+        "200":
+          description: OK
+
+  /{tenant_id}/{timeline_id}/{endpoint_id}/{key}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: endpoint_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: key
+        in: path
+        required: true
+        schema:
+          type: string
+    get:
+      description: Get file from blob storage
+      responses:
+        "200":
+          description: "File stream from blob storage"
+          content:
+            application/octet-stream:
+              schema:
+                type: string
+                format: binary
+        "400":
+          description: File was not found
+        "403":
+          description: JWT does not authorize request to this route
+    put:
+      description: Insert file into blob storage. If file exists, override it
+      requestBody:
+        content:
+          application/octet-stream:
+            schema:
+              type: string
+              format: binary
+      responses:
+        "200":
+          description: File was inserted successfully
+        "403":
+          description: JWT does not authorize request to this route
+    delete:
+      description: Delete file from blob storage
+      responses:
+        "200":
+          description: File was successfully deleted or not found
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}/{timeline_id}/{endpoint_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: endpoint_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete endpoint data from blob storage
+      responses:
+        "200":
+          description: Endpoint data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}/{timeline_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete timeline data from blob storage
+      responses:
+        "200":
+          description: Timeline data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+  /{tenant_id}:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+    delete:
+      description: Delete tenant data from blob storage
+      responses:
+        "200":
+          description: Tenant data was deleted
+        "403":
+          description: JWT does not authorize request to this route
+
+components:
+  securitySchemes:
+    JWT:
+      type: http
+      scheme: bearer
+      bearerFormat: JWT
+
+security:
+  - JWT: []

libs/compute_api/src/responses.rs

@@ -46,7 +46,7 @@ pub struct ExtensionInstallResponse {
     pub version: ExtVersion,
 }
 
-#[derive(Serialize, Default, Debug, Clone)]
+#[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcPrewarmState {
     #[default]
@@ -58,7 +58,18 @@ pub enum LfcPrewarmState {
     },
 }
 
-#[derive(Serialize, Default, Debug, Clone)]
+impl Display for LfcPrewarmState {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            LfcPrewarmState::NotPrewarmed => f.write_str("NotPrewarmed"),
+            LfcPrewarmState::Prewarming => f.write_str("Prewarming"),
+            LfcPrewarmState::Completed => f.write_str("Completed"),
+            LfcPrewarmState::Failed { error } => write!(f, "Error({error})"),
+        }
+    }
+}
+
+#[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcOffloadState {
     #[default]
@@ -70,6 +81,23 @@ pub enum LfcOffloadState {
     },
 }
 
+#[derive(Serialize, Debug, Clone, PartialEq)]
+#[serde(tag = "status", rename_all = "snake_case")]
+/// Response of /promote
+pub enum PromoteState {
+    NotPromoted,
+    Completed,
+    Failed { error: String },
+}
+
+#[derive(Deserialize, Serialize, Default, Debug, Clone)]
+#[serde(rename_all = "snake_case")]
+/// Result of /safekeepers_lsn
+pub struct SafekeepersLsn {
+    pub safekeepers: String,
+    pub wal_flush_lsn: utils::lsn::Lsn,
+}
+
 /// Response of the /status API
 #[derive(Serialize, Debug, Deserialize)]
 #[serde(rename_all = "snake_case")]

libs/compute_api/src/spec.rs

@@ -181,10 +181,14 @@ 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.
@@ -438,7 +442,7 @@ pub struct JwksSettings {
 }
 
 /// Protocol used to connect to a Pageserver. Parsed from the connstring scheme.
-#[derive(Clone, Copy, Debug, Default)]
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub enum PageserverProtocol {
     /// The original protocol based on libpq and COPY. Uses postgresql:// or postgres:// scheme.
     #[default]

libs/compute_api/tests/cluster_spec.json

@@ -90,6 +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/http-utils/src/endpoint.rs

@@ -7,7 +7,7 @@ use anyhow::{Context, anyhow};
 use bytes::{Bytes, BytesMut};
 use hyper::header::{AUTHORIZATION, CONTENT_DISPOSITION, CONTENT_TYPE, HeaderName};
 use hyper::http::HeaderValue;
-use hyper::{Body, Method, Request, Response};
+use hyper::{Body, Request, Response};
 use jsonwebtoken::TokenData;
 use metrics::{Encoder, IntCounter, TextEncoder, register_int_counter};
 use once_cell::sync::Lazy;
@@ -18,8 +18,9 @@ use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
 use tokio::sync::{Mutex, Notify, mpsc};
 use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::io::ReaderStream;
-use tracing::{Instrument, debug, info, info_span, warn};
+use tracing::{Instrument, info, info_span, warn};
 use utils::auth::{AuthError, Claims, SwappableJwtAuth};
+use utils::metrics_collector::{METRICS_COLLECTOR, METRICS_STALE_MILLIS};
 
 use crate::error::{ApiError, api_error_handler, route_error_handler};
 use crate::request::{get_query_param, parse_query_param};
@@ -80,14 +81,10 @@ where
     let path = request.uri().path();
     let request_span = info_span!("request", %method, %path, %request_id);
 
-    let log_quietly = method == Method::GET;
     async move {
         let cancellation_guard = RequestCancelled::warn_when_dropped_without_responding();
-        if log_quietly {
-            debug!("Handling request");
-        } else {
-            info!("Handling request");
-        }
+
+        info!("Handling request");
 
         // No special handling for panics here. There's a `tracing_panic_hook` from another
         // module to do that globally.
@@ -108,11 +105,7 @@ where
         match res {
             Ok(response) => {
                 let response_status = response.status();
-                if log_quietly && response_status.is_success() {
-                    debug!("Request handled, status: {response_status}");
-                } else {
-                    info!("Request handled, status: {response_status}");
-                }
+                info!("Request handled, status: {response_status}");
                 Ok(response)
             }
             Err(err) => Ok(api_error_handler(err)),
@@ -250,9 +243,28 @@ impl std::io::Write for ChannelWriter {
     }
 }
 
-pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+pub async fn prometheus_metrics_handler(
+    req: Request<Body>,
+    force_metric_collection_on_scrape: bool,
+) -> Result<Response<Body>, ApiError> {
     SERVE_METRICS_COUNT.inc();
 
+    // HADRON
+    let requested_use_latest = parse_query_param(&req, "use_latest")?;
+
+    let use_latest = match requested_use_latest {
+        None => force_metric_collection_on_scrape,
+        Some(true) => true,
+        Some(false) => {
+            if force_metric_collection_on_scrape {
+                // We don't cache in this case
+                true
+            } else {
+                false
+            }
+        }
+    };
+
     let started_at = std::time::Instant::now();
 
     let (tx, rx) = mpsc::channel(1);
@@ -277,12 +289,18 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
 
         let _span = span.entered();
 
-        let metrics = metrics::gather();
+        // HADRON
+        let collected = if use_latest {
+            // Skip caching the results if we always force metric collection on scrape.
+            METRICS_COLLECTOR.run_once(!force_metric_collection_on_scrape)
+        } else {
+            METRICS_COLLECTOR.last_collected()
+        };
 
         let gathered_at = std::time::Instant::now();
 
         let res = encoder
-            .encode(&metrics, &mut writer)
+            .encode(&collected.metrics, &mut writer)
             .and_then(|_| writer.flush().map_err(|e| e.into()));
 
         // this instant is not when we finally got the full response sent, sending is done by hyper
@@ -295,6 +313,10 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
         let encoded_in = encoded_at - gathered_at - writer.wait_time();
         let total = encoded_at - started_at;
 
+        // HADRON
+        let staleness_ms = (encoded_at - collected.collected_at).as_millis();
+        METRICS_STALE_MILLIS.set(staleness_ms as i64);
+
         match res {
             Ok(()) => {
                 tracing::info!(
@@ -303,6 +325,7 @@ pub async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<
                     spawning_ms = spawned_in.as_millis(),
                     collection_ms = collected_in.as_millis(),
                     encoding_ms = encoded_in.as_millis(),
+                    stalenss_ms = staleness_ms,
                     "responded /metrics"
                 );
             }

libs/http-utils/src/request.rs

@@ -41,17 +41,35 @@ pub fn get_query_param<'a>(
         Some(q) => q,
         None => return Ok(None),
     };
-    let mut values = url::form_urlencoded::parse(query.as_bytes())
+    let values = url::form_urlencoded::parse(query.as_bytes())
         .filter_map(|(k, v)| if k == param_name { Some(v) } else { None })
         // we call .next() twice below. If it's None the first time, .fuse() ensures it's None afterwards
         .fuse();
 
-    let value1 = values.next();
-    if values.next().is_some() {
-        return Err(ApiError::BadRequest(anyhow!(
-            "param {param_name} specified more than once"
-        )));
-    }
+    // Work around an issue with Alloy's pyroscope scrape where the "seconds"
+    // parameter is added several times. https://github.com/grafana/alloy/issues/3026
+    // TODO: revert after Alloy is fixed.
+    let value1 = values
+        .map(Ok)
+        .reduce(|acc, i| {
+            match acc {
+                Err(_) => acc,
+
+                // It's okay to have duplicates as along as they have the same value.
+                Ok(ref a) if a == &i.unwrap() => acc,
+
+                _ => Err(ApiError::BadRequest(anyhow!(
+                    "param {param_name} specified more than once"
+                ))),
+            }
+        })
+        .transpose()?;
+    // if values.next().is_some() {
+    //     return Err(ApiError::BadRequest(anyhow!(
+    //         "param {param_name} specified more than once"
+    //     )));
+    // }
+
     Ok(value1)
 }
 
@@ -92,3 +110,39 @@ pub async fn ensure_no_body(request: &mut Request<Body>) -> Result<(), ApiError>
         None => Ok(()),
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_get_query_param_duplicate() {
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert_eq!(value.unwrap(), "1");
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1&testparam=1")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert_eq!(value.unwrap(), "1");
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam").unwrap();
+        assert!(value.is_none());
+
+        let req = Request::builder()
+            .uri("http://localhost:12345/testuri?testparam=1&testparam=2&testparam=3")
+            .body(hyper::Body::empty())
+            .unwrap();
+        let value = get_query_param(&req, "testparam");
+        assert!(value.is_err());
+    }
+}

libs/pageserver_api/src/config.rs

@@ -5,6 +5,7 @@ mod tests;
 
 use const_format::formatcp;
 use posthog_client_lite::PostHogClientConfig;
+use utils::serde_percent::Percent;
 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;
@@ -223,8 +224,9 @@ pub struct ConfigToml {
     pub metric_collection_bucket: Option<RemoteStorageConfig>,
     #[serde(with = "humantime_serde")]
     pub synthetic_size_calculation_interval: Duration,
-    pub disk_usage_based_eviction: Option<DiskUsageEvictionTaskConfig>,
+    pub disk_usage_based_eviction: DiskUsageEvictionTaskConfig,
     pub test_remote_failures: u64,
+    pub test_remote_failures_probability: u64,
     pub ondemand_download_behavior_treat_error_as_warn: bool,
     #[serde(with = "humantime_serde")]
     pub background_task_maximum_delay: Duration,
@@ -270,9 +272,13 @@ pub struct ConfigToml {
     pub timeline_import_config: TimelineImportConfig,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub basebackup_cache_config: Option<BasebackupCacheConfig>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_layer_generation_large_timeline_threshold: Option<u64>,
+    pub force_metric_collection_on_scrape: bool,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(default)]
 pub struct DiskUsageEvictionTaskConfig {
     pub max_usage_pct: utils::serde_percent::Percent,
     pub min_avail_bytes: u64,
@@ -283,6 +289,21 @@ pub struct DiskUsageEvictionTaskConfig {
     /// Select sorting for evicted layers
     #[serde(default)]
     pub eviction_order: EvictionOrder,
+    pub enabled: bool,
+}
+
+impl Default for DiskUsageEvictionTaskConfig {
+    fn default() -> Self {
+        Self {
+            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: EvictionOrder::default(),
+            enabled: true,
+        }
+    }
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -543,6 +564,11 @@ pub struct TenantConfigToml {
     pub gc_period: Duration,
     // Delta layer churn threshold to create L1 image layers.
     pub image_creation_threshold: usize,
+    // HADRON
+    // When the timeout is reached, PageServer will (1) force compact any remaining L0 deltas and
+    // (2) create image layers if there are any L1 deltas.
+    #[serde(with = "humantime_serde")]
+    pub image_layer_force_creation_period: Option<Duration>,
     // Determines how much history is retained, to allow
     // branching and read replicas at an older point in time.
     // The unit is time.
@@ -738,9 +764,10 @@ impl Default for ConfigToml {
 
             metric_collection_bucket: (None),
 
-            disk_usage_based_eviction: (None),
+            disk_usage_based_eviction: DiskUsageEvictionTaskConfig::default(),
 
             test_remote_failures: (0),
+            test_remote_failures_probability: (100),
 
             ondemand_download_behavior_treat_error_as_warn: (false),
 
@@ -804,6 +831,8 @@ impl Default for ConfigToml {
             },
             basebackup_cache_config: None,
             posthog_config: None,
+            image_layer_generation_large_timeline_threshold: Some(2 * 1024 * 1024 * 1024),
+            force_metric_collection_on_scrape: true,
         }
     }
 }
@@ -897,6 +926,7 @@ impl Default for TenantConfigToml {
             gc_period: humantime::parse_duration(DEFAULT_GC_PERIOD)
                 .expect("cannot parse default gc period"),
             image_creation_threshold: DEFAULT_IMAGE_CREATION_THRESHOLD,
+            image_layer_force_creation_period: None,
             pitr_interval: humantime::parse_duration(DEFAULT_PITR_INTERVAL)
                 .expect("cannot parse default PITR interval"),
             walreceiver_connect_timeout: humantime::parse_duration(

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()
     }

libs/pageserver_api/src/models.rs

@@ -384,7 +384,7 @@ pub struct SafekeepersInfo {
     pub safekeepers: Vec<SafekeeperInfo>,
 }
 
-#[derive(Serialize, Deserialize, Clone)]
+#[derive(Serialize, Deserialize, Clone, Debug)]
 pub struct SafekeeperInfo {
     pub id: NodeId,
     pub hostname: String,
@@ -597,6 +597,9 @@ pub struct TenantConfigPatch {
     pub gc_period: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub image_creation_threshold: FieldPatch<usize>,
+    // HADRON
+    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
+    pub image_layer_force_creation_period: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub pitr_interval: FieldPatch<String>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
@@ -700,6 +703,11 @@ pub struct TenantConfig {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub image_creation_threshold: Option<usize>,
 
+    // HADRON
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(with = "humantime_serde")]
+    pub image_layer_force_creation_period: Option<Duration>,
+
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(with = "humantime_serde")]
     pub pitr_interval: Option<Duration>,
@@ -798,6 +806,7 @@ impl TenantConfig {
             mut gc_horizon,
             mut gc_period,
             mut image_creation_threshold,
+            mut image_layer_force_creation_period,
             mut pitr_interval,
             mut walreceiver_connect_timeout,
             mut lagging_wal_timeout,
@@ -861,6 +870,11 @@ impl TenantConfig {
         patch
             .image_creation_threshold
             .apply(&mut image_creation_threshold);
+        // HADRON
+        patch
+            .image_layer_force_creation_period
+            .map(|v| humantime::parse_duration(&v))?
+            .apply(&mut image_layer_force_creation_period);
         patch
             .pitr_interval
             .map(|v| humantime::parse_duration(&v))?
@@ -942,6 +956,7 @@ impl TenantConfig {
             gc_horizon,
             gc_period,
             image_creation_threshold,
+            image_layer_force_creation_period,
             pitr_interval,
             walreceiver_connect_timeout,
             lagging_wal_timeout,
@@ -1016,6 +1031,9 @@ impl TenantConfig {
             image_creation_threshold: self
                 .image_creation_threshold
                 .unwrap_or(global_conf.image_creation_threshold),
+            image_layer_force_creation_period: self
+                .image_layer_force_creation_period
+                .or(global_conf.image_layer_force_creation_period),
             pitr_interval: self.pitr_interval.unwrap_or(global_conf.pitr_interval),
             walreceiver_connect_timeout: self
                 .walreceiver_connect_timeout

libs/pageserver_api/src/shard.rs

@@ -332,7 +332,11 @@ fn hash_combine(mut a: u32, mut b: u32) -> u32 {
 ///
 /// The mapping of key to shard is not stable across changes to ShardCount: this is intentional
 /// and will be handled at higher levels when shards are split.
-fn key_to_shard_number(count: ShardCount, stripe_size: ShardStripeSize, key: &Key) -> ShardNumber {
+pub fn key_to_shard_number(
+    count: ShardCount,
+    stripe_size: ShardStripeSize,
+    key: &Key,
+) -> ShardNumber {
     // Fast path for un-sharded tenants or broadcast keys
     if count < ShardCount(2) || key_is_shard0(key) {
         return ShardNumber(0);

libs/proxy/json/Cargo.toml

@@ -0,0 +1,12 @@
+[package]
+name = "json"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+ryu = "1"
+itoa = "1"
+
+[dev-dependencies]
+futures = "0.3"

libs/proxy/json/src/lib.rs

@@ -0,0 +1,412 @@
+//! A JSON serialization lib, designed for more flexibility than `serde_json` offers.
+//!
+//! Features:
+//!
+//! ## Dynamic construction
+//!
+//! Sometimes you have dynamic values you want to serialize, that are not already in a serde-aware model like a struct or a Vec etc.
+//! To achieve this with serde, you need to implement a lot of different traits on a lot of different new-types.
+//! Because of this, it's often easier to give-in and pull all the data into a serde-aware model (`serde_json::Value` or some intermediate struct),
+//! but that is often not very efficient.
+//!
+//! This crate allows full control over the JSON encoding without needing to implement any extra traits. Just call the
+//! relevant functions, and it will guarantee a correctly encoded JSON value.
+//!
+//! ## Async construction
+//!
+//! Similar to the above, sometimes the values arrive asynchronously. Often collecting those values in memory
+//! is more expensive than writing them as JSON, since the overheads of `Vec` and `String` is much higher, however
+//! there are exceptions.
+//!
+//! Serializing to JSON all in one go is also more CPU intensive and can cause lag spikes,
+//! whereas serializing values incrementally spreads out the CPU load and reduces lag.
+//!
+//! ## Examples
+//!
+//! To represent the following JSON as a compact string
+//!
+//! ```json
+//! {
+//!   "results": {
+//!     "rows": [
+//!       {
+//!         "id": 1,
+//!         "value": null
+//!       },
+//!       {
+//!         "id": 2,
+//!         "value": "hello"
+//!       }
+//!     ]
+//!   }
+//! }
+//! ```
+//!
+//! We can use the following code:
+//!
+//! ```
+//! // create the outer object
+//! let s = json::value_to_string!(|v| json::value_as_object!(|v| {
+//!     // create an entry with key "results" and start an object value associated with it.
+//!     let results = v.key("results");
+//!     json::value_as_object!(|results| {
+//!         // create an entry with key "rows" and start an list value associated with it.
+//!         let rows = results.key("rows");
+//!         json::value_as_list!(|rows| {
+//!             // create a list entry and start an object value associated with it.
+//!             let row = rows.entry();
+//!             json::value_as_object!(|row| {
+//!                 // add entry "id": 1
+//!                 row.entry("id", 1);
+//!                 // add entry "value": null
+//!                 row.entry("value", json::Null);
+//!             });
+//!
+//!             // create a list entry and start an object value associated with it.
+//!             let row = rows.entry();
+//!             json::value_as_object!(|row| {
+//!                 // add entry "id": 2
+//!                 row.entry("id", 2);
+//!                 // add entry "value": "hello"
+//!                 row.entry("value", "hello");
+//!             });
+//!         });
+//!     });
+//! }));
+//!
+//! assert_eq!(s, r#"{"results":{"rows":[{"id":1,"value":null},{"id":2,"value":"hello"}]}}"#);
+//! ```
+
+mod macros;
+mod str;
+mod value;
+
+pub use value::{Null, ValueEncoder};
+
+#[must_use]
+/// Serialize a single json value.
+pub struct ValueSer<'buf> {
+    buf: &'buf mut Vec<u8>,
+    start: usize,
+}
+
+impl<'buf> ValueSer<'buf> {
+    /// Create a new json value serializer.
+    pub fn new(buf: &'buf mut Vec<u8>) -> Self {
+        Self { buf, start: 0 }
+    }
+
+    /// Borrow the underlying buffer
+    pub fn as_buffer(&self) -> &[u8] {
+        self.buf
+    }
+
+    #[inline]
+    pub fn value(self, e: impl ValueEncoder) {
+        e.encode(self);
+    }
+
+    /// Write raw bytes to the buf. This must be already JSON encoded.
+    #[inline]
+    pub fn write_raw_json(self, data: &[u8]) {
+        self.buf.extend_from_slice(data);
+        self.finish();
+    }
+
+    /// Start a new object serializer.
+    #[inline]
+    pub fn object(self) -> ObjectSer<'buf> {
+        ObjectSer::new(self)
+    }
+
+    /// Start a new list serializer.
+    #[inline]
+    pub fn list(self) -> ListSer<'buf> {
+        ListSer::new(self)
+    }
+
+    /// Finish the value ser.
+    #[inline]
+    fn finish(self) {
+        // don't trigger the drop handler which triggers a rollback.
+        // this won't cause memory leaks because `ValueSet` owns no allocations.
+        std::mem::forget(self);
+    }
+}
+
+impl Drop for ValueSer<'_> {
+    fn drop(&mut self) {
+        self.buf.truncate(self.start);
+    }
+}
+
+#[must_use]
+/// Serialize a json object.
+pub struct ObjectSer<'buf> {
+    value: ValueSer<'buf>,
+    start: usize,
+}
+
+impl<'buf> ObjectSer<'buf> {
+    /// Start a new object serializer.
+    #[inline]
+    pub fn new(value: ValueSer<'buf>) -> Self {
+        value.buf.push(b'{');
+        let start = value.buf.len();
+        Self { value, start }
+    }
+
+    /// Borrow the underlying buffer
+    pub fn as_buffer(&self) -> &[u8] {
+        self.value.as_buffer()
+    }
+
+    /// Start a new object entry with the given string key, returning a [`ValueSer`] for the associated value.
+    #[inline]
+    pub fn key(&mut self, key: impl KeyEncoder) -> ValueSer<'_> {
+        key.write_key(self)
+    }
+
+    /// Write an entry (key-value pair) to the object.
+    #[inline]
+    pub fn entry(&mut self, key: impl KeyEncoder, val: impl ValueEncoder) {
+        self.key(key).value(val);
+    }
+
+    #[inline]
+    fn entry_inner(&mut self, f: impl FnOnce(&mut Vec<u8>)) -> ValueSer<'_> {
+        // track before the separator so we the value is rolled back it also removes the separator.
+        let start = self.value.buf.len();
+
+        // push separator if necessary
+        if self.value.buf.len() > self.start {
+            self.value.buf.push(b',');
+        }
+        // push key
+        f(self.value.buf);
+        // push value separator
+        self.value.buf.push(b':');
+
+        // return value writer.
+        ValueSer {
+            buf: self.value.buf,
+            start,
+        }
+    }
+
+    /// Reset the buffer back to before this object was started.
+    #[inline]
+    pub fn rollback(self) -> ValueSer<'buf> {
+        // Do not fully reset the value, only reset it to before the `{`.
+        // This ensures any `,` before this value are not clobbered.
+        self.value.buf.truncate(self.start - 1);
+        self.value
+    }
+
+    /// Finish the object ser.
+    #[inline]
+    pub fn finish(self) {
+        self.value.buf.push(b'}');
+        self.value.finish();
+    }
+}
+
+pub trait KeyEncoder {
+    fn write_key<'a>(self, obj: &'a mut ObjectSer) -> ValueSer<'a>;
+}
+
+#[must_use]
+/// Serialize a json object.
+pub struct ListSer<'buf> {
+    value: ValueSer<'buf>,
+    start: usize,
+}
+
+impl<'buf> ListSer<'buf> {
+    /// Start a new list serializer.
+    #[inline]
+    pub fn new(value: ValueSer<'buf>) -> Self {
+        value.buf.push(b'[');
+        let start = value.buf.len();
+        Self { value, start }
+    }
+
+    /// Borrow the underlying buffer
+    pub fn as_buffer(&self) -> &[u8] {
+        self.value.as_buffer()
+    }
+
+    /// Write an value to the list.
+    #[inline]
+    pub fn push(&mut self, val: impl ValueEncoder) {
+        self.entry().value(val);
+    }
+
+    /// Start a new value entry in this list.
+    #[inline]
+    pub fn entry(&mut self) -> ValueSer<'_> {
+        // track before the separator so we the value is rolled back it also removes the separator.
+        let start = self.value.buf.len();
+
+        // push separator if necessary
+        if self.value.buf.len() > self.start {
+            self.value.buf.push(b',');
+        }
+
+        // return value writer.
+        ValueSer {
+            buf: self.value.buf,
+            start,
+        }
+    }
+
+    /// Reset the buffer back to before this object was started.
+    #[inline]
+    pub fn rollback(self) -> ValueSer<'buf> {
+        // Do not fully reset the value, only reset it to before the `[`.
+        // This ensures any `,` before this value are not clobbered.
+        self.value.buf.truncate(self.start - 1);
+        self.value
+    }
+
+    /// Finish the object ser.
+    #[inline]
+    pub fn finish(self) {
+        self.value.buf.push(b']');
+        self.value.finish();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::{Null, ValueSer};
+
+    #[test]
+    fn object() {
+        let mut buf = vec![];
+        let mut object = ValueSer::new(&mut buf).object();
+        object.entry("foo", "bar");
+        object.entry("baz", Null);
+        object.finish();
+
+        assert_eq!(buf, br#"{"foo":"bar","baz":null}"#);
+    }
+
+    #[test]
+    fn list() {
+        let mut buf = vec![];
+        let mut list = ValueSer::new(&mut buf).list();
+        list.entry().value("bar");
+        list.entry().value(Null);
+        list.finish();
+
+        assert_eq!(buf, br#"["bar",null]"#);
+    }
+
+    #[test]
+    fn object_macro() {
+        let res = crate::value_to_string!(|obj| {
+            crate::value_as_object!(|obj| {
+                obj.entry("foo", "bar");
+                obj.entry("baz", Null);
+            })
+        });
+
+        assert_eq!(res, r#"{"foo":"bar","baz":null}"#);
+    }
+
+    #[test]
+    fn list_macro() {
+        let res = crate::value_to_string!(|list| {
+            crate::value_as_list!(|list| {
+                list.entry().value("bar");
+                list.entry().value(Null);
+            })
+        });
+
+        assert_eq!(res, r#"["bar",null]"#);
+    }
+
+    #[test]
+    fn rollback_on_drop() {
+        let res = crate::value_to_string!(|list| {
+            crate::value_as_list!(|list| {
+                list.entry().value("bar");
+
+                'cancel: {
+                    let nested_list = list.entry();
+                    crate::value_as_list!(|nested_list| {
+                        nested_list.entry().value(1);
+
+                        assert_eq!(nested_list.as_buffer(), br#"["bar",[1"#);
+                        if true {
+                            break 'cancel;
+                        }
+                    })
+                }
+
+                assert_eq!(list.as_buffer(), br#"["bar""#);
+
+                list.entry().value(Null);
+            })
+        });
+
+        assert_eq!(res, r#"["bar",null]"#);
+    }
+
+    #[test]
+    fn rollback_object() {
+        let res = crate::value_to_string!(|obj| {
+            crate::value_as_object!(|obj| {
+                let entry = obj.key("1");
+                entry.value(1_i32);
+
+                let entry = obj.key("2");
+                let entry = {
+                    let mut nested_obj = entry.object();
+                    nested_obj.entry("foo", "bar");
+                    nested_obj.rollback()
+                };
+
+                entry.value(2_i32);
+            })
+        });
+
+        assert_eq!(res, r#"{"1":1,"2":2}"#);
+    }
+
+    #[test]
+    fn rollback_list() {
+        let res = crate::value_to_string!(|list| {
+            crate::value_as_list!(|list| {
+                let entry = list.entry();
+                entry.value(1_i32);
+
+                let entry = list.entry();
+                let entry = {
+                    let mut nested_list = entry.list();
+                    nested_list.push("foo");
+                    nested_list.rollback()
+                };
+
+                entry.value(2_i32);
+            })
+        });
+
+        assert_eq!(res, r#"[1,2]"#);
+    }
+
+    #[test]
+    fn string_escaping() {
+        let mut buf = vec![];
+        let mut object = ValueSer::new(&mut buf).object();
+
+        let key = "hello";
+        let value = "\n world";
+
+        object.entry(format_args!("{key:?}"), value);
+        object.finish();
+
+        assert_eq!(buf, br#"{"\"hello\"":"\n world"}"#);
+    }
+}

libs/proxy/json/src/macros.rs

@@ -0,0 +1,86 @@
+//! # Examples
+//!
+//! ```
+//! use futures::{StreamExt, TryStream, TryStreamExt};
+//!
+//! async fn stream_to_json_list<S, T, E>(mut s: S) -> Result<String, E>
+//! where
+//!     S: TryStream<Ok = T, Error = E> + Unpin,
+//!     T: json::ValueEncoder
+//! {
+//!     Ok(json::value_to_string!(|val| json::value_as_list!(|val| {
+//!         // note how we can use `.await` and `?` in here.
+//!         while let Some(value) = s.try_next().await? {
+//!             val.push(value);
+//!         }
+//!     })))
+//! }
+//!
+//! let stream = futures::stream::iter([1, 2, 3]).map(Ok::<i32, ()>);
+//! let json_string = futures::executor::block_on(stream_to_json_list(stream)).unwrap();
+//! assert_eq!(json_string, "[1,2,3]");
+//! ```
+
+/// A helper to create a new JSON vec.
+///
+/// Implemented as a macro to preserve all control flow.
+#[macro_export]
+macro_rules! value_to_vec {
+    (|$val:ident| $body:expr) => {{
+        let mut buf = vec![];
+        let $val = $crate::ValueSer::new(&mut buf);
+        let _: () = $body;
+        buf
+    }};
+}
+
+/// A helper to create a new JSON string.
+///
+/// Implemented as a macro to preserve all control flow.
+#[macro_export]
+macro_rules! value_to_string {
+    (|$val:ident| $body:expr) => {{
+        ::std::string::String::from_utf8($crate::value_to_vec!(|$val| $body))
+            .expect("json should be valid utf8")
+    }};
+}
+
+/// A helper that ensures the [`ObjectSer::finish`](crate::ObjectSer::finish) method is called on completion.
+///
+/// Consumes `$val` and assigns it as an [`ObjectSer`](crate::ObjectSer) serializer.
+/// The serializer is only 'finished' if the body completes.
+/// The serializer is rolled back if `break`/`return` escapes the body.
+///
+/// Implemented as a macro to preserve all control flow.
+#[macro_export]
+macro_rules! value_as_object {
+    (|$val:ident| $body:expr) => {{
+        let mut obj = $crate::ObjectSer::new($val);
+
+        let $val = &mut obj;
+        let res = $body;
+
+        obj.finish();
+        res
+    }};
+}
+
+/// A helper that ensures the [`ListSer::finish`](crate::ListSer::finish) method is called on completion.
+///
+/// Consumes `$val` and assigns it as an [`ListSer`](crate::ListSer) serializer.
+/// The serializer is only 'finished' if the body completes.
+/// The serializer is rolled back if `break`/`return` escapes the body.
+///
+/// Implemented as a macro to preserve all control flow.
+#[macro_export]
+macro_rules! value_as_list {
+    (|$val:ident| $body:expr) => {{
+        let mut list = $crate::ListSer::new($val);
+
+        let $val = &mut list;
+        let res = $body;
+
+        list.finish();
+        res
+    }};
+}

libs/proxy/json/src/str.rs

@@ -0,0 +1,166 @@
+//! Helpers for serializing escaped strings.
+//!
+//! ## License
+//!
+//! <https://github.com/serde-rs/json/blob/c1826ebcccb1a520389c6b78ad3da15db279220d/src/ser.rs#L1514-L1552>
+//! <https://github.com/serde-rs/json/blob/c1826ebcccb1a520389c6b78ad3da15db279220d/src/ser.rs#L2081-L2157>
+//! Licensed by David Tolnay under MIT or Apache-2.0.
+//!
+//! With modifications by Conrad Ludgate on behalf of Databricks.
+
+use std::fmt::{self, Write};
+
+/// Represents a character escape code in a type-safe manner.
+pub enum CharEscape {
+    /// An escaped quote `"`
+    Quote,
+    /// An escaped reverse solidus `\`
+    ReverseSolidus,
+    // /// An escaped solidus `/`
+    // Solidus,
+    /// An escaped backspace character (usually escaped as `\b`)
+    Backspace,
+    /// An escaped form feed character (usually escaped as `\f`)
+    FormFeed,
+    /// An escaped line feed character (usually escaped as `\n`)
+    LineFeed,
+    /// An escaped carriage return character (usually escaped as `\r`)
+    CarriageReturn,
+    /// An escaped tab character (usually escaped as `\t`)
+    Tab,
+    /// An escaped ASCII plane control character (usually escaped as
+    /// `\u00XX` where `XX` are two hex characters)
+    AsciiControl(u8),
+}
+
+impl CharEscape {
+    #[inline]
+    fn from_escape_table(escape: u8, byte: u8) -> CharEscape {
+        match escape {
+            self::BB => CharEscape::Backspace,
+            self::TT => CharEscape::Tab,
+            self::NN => CharEscape::LineFeed,
+            self::FF => CharEscape::FormFeed,
+            self::RR => CharEscape::CarriageReturn,
+            self::QU => CharEscape::Quote,
+            self::BS => CharEscape::ReverseSolidus,
+            self::UU => CharEscape::AsciiControl(byte),
+            _ => unreachable!(),
+        }
+    }
+}
+
+pub(crate) fn format_escaped_str(writer: &mut Vec<u8>, value: &str) {
+    writer.reserve(2 + value.len());
+
+    writer.push(b'"');
+
+    let rest = format_escaped_str_contents(writer, value);
+    writer.extend_from_slice(rest);
+
+    writer.push(b'"');
+}
+
+pub(crate) fn format_escaped_fmt(writer: &mut Vec<u8>, args: fmt::Arguments) {
+    writer.push(b'"');
+
+    Collect { buf: writer }
+        .write_fmt(args)
+        .expect("formatting should not error");
+
+    writer.push(b'"');
+}
+
+struct Collect<'buf> {
+    buf: &'buf mut Vec<u8>,
+}
+
+impl fmt::Write for Collect<'_> {
+    fn write_str(&mut self, s: &str) -> fmt::Result {
+        let last = format_escaped_str_contents(self.buf, s);
+        self.buf.extend(last);
+        Ok(())
+    }
+}
+
+// writes any escape sequences, and returns the suffix still needed to be written.
+fn format_escaped_str_contents<'a>(writer: &mut Vec<u8>, value: &'a str) -> &'a [u8] {
+    let bytes = value.as_bytes();
+
+    let mut start = 0;
+
+    for (i, &byte) in bytes.iter().enumerate() {
+        let escape = ESCAPE[byte as usize];
+        if escape == 0 {
+            continue;
+        }
+
+        writer.extend_from_slice(&bytes[start..i]);
+
+        let char_escape = CharEscape::from_escape_table(escape, byte);
+        write_char_escape(writer, char_escape);
+
+        start = i + 1;
+    }
+
+    &bytes[start..]
+}
+
+const BB: u8 = b'b'; // \x08
+const TT: u8 = b't'; // \x09
+const NN: u8 = b'n'; // \x0A
+const FF: u8 = b'f'; // \x0C
+const RR: u8 = b'r'; // \x0D
+const QU: u8 = b'"'; // \x22
+const BS: u8 = b'\\'; // \x5C
+const UU: u8 = b'u'; // \x00...\x1F except the ones above
+const __: u8 = 0;
+
+// Lookup table of escape sequences. A value of b'x' at index i means that byte
+// i is escaped as "\x" in JSON. A value of 0 means that byte i is not escaped.
+static ESCAPE: [u8; 256] = [
+    //   1   2   3   4   5   6   7   8   9   A   B   C   D   E   F
+    UU, UU, UU, UU, UU, UU, UU, UU, BB, TT, NN, UU, FF, RR, UU, UU, // 0
+    UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, UU, // 1
+    __, __, QU, __, __, __, __, __, __, __, __, __, __, __, __, __, // 2
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 3
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 4
+    __, __, __, __, __, __, __, __, __, __, __, __, BS, __, __, __, // 5
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 6
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 7
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 8
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // 9
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // A
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // B
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // C
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // D
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // E
+    __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, __, // F
+];
+
+fn write_char_escape(writer: &mut Vec<u8>, char_escape: CharEscape) {
+    let s = match char_escape {
+        CharEscape::Quote => b"\\\"",
+        CharEscape::ReverseSolidus => b"\\\\",
+        // CharEscape::Solidus => b"\\/",
+        CharEscape::Backspace => b"\\b",
+        CharEscape::FormFeed => b"\\f",
+        CharEscape::LineFeed => b"\\n",
+        CharEscape::CarriageReturn => b"\\r",
+        CharEscape::Tab => b"\\t",
+        CharEscape::AsciiControl(byte) => {
+            static HEX_DIGITS: [u8; 16] = *b"0123456789abcdef";
+            let bytes = &[
+                b'\\',
+                b'u',
+                b'0',
+                b'0',
+                HEX_DIGITS[(byte >> 4) as usize],
+                HEX_DIGITS[(byte & 0xF) as usize],
+            ];
+            return writer.extend_from_slice(bytes);
+        }
+    };
+
+    writer.extend_from_slice(s);
+}

libs/proxy/json/src/value.rs

@@ -0,0 +1,168 @@
+use core::fmt;
+use std::collections::{BTreeMap, HashMap};
+
+use crate::str::{format_escaped_fmt, format_escaped_str};
+use crate::{KeyEncoder, ObjectSer, ValueSer, value_as_list, value_as_object};
+
+/// Write a value to the underlying json representation.
+pub trait ValueEncoder {
+    fn encode(self, v: ValueSer<'_>);
+}
+
+pub(crate) fn write_int(x: impl itoa::Integer, b: &mut Vec<u8>) {
+    b.extend_from_slice(itoa::Buffer::new().format(x).as_bytes());
+}
+
+pub(crate) fn write_float(x: impl ryu::Float, b: &mut Vec<u8>) {
+    b.extend_from_slice(ryu::Buffer::new().format(x).as_bytes());
+}
+
+impl<T: Copy + ValueEncoder> ValueEncoder for &T {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        T::encode(*self, v);
+    }
+}
+
+impl ValueEncoder for &str {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        format_escaped_str(v.buf, self);
+        v.finish();
+    }
+}
+
+impl ValueEncoder for fmt::Arguments<'_> {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        if let Some(s) = self.as_str() {
+            format_escaped_str(v.buf, s);
+        } else {
+            format_escaped_fmt(v.buf, self);
+        }
+        v.finish();
+    }
+}
+
+macro_rules! int {
+    [$($t:ty),*] => {
+        $(
+            impl ValueEncoder for $t {
+                #[inline]
+                fn encode(self, v: ValueSer<'_>) {
+                    write_int(self, v.buf);
+                    v.finish();
+                }
+            }
+        )*
+    };
+}
+
+int![u8, u16, u32, u64, usize, u128];
+int![i8, i16, i32, i64, isize, i128];
+
+macro_rules! float {
+    [$($t:ty),*] => {
+        $(
+            impl ValueEncoder for $t {
+                #[inline]
+                fn encode(self, v: ValueSer<'_>) {
+                    write_float(self, v.buf);
+                    v.finish();
+                }
+            }
+        )*
+    };
+}
+
+float![f32, f64];
+
+impl ValueEncoder for bool {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        v.write_raw_json(if self { b"true" } else { b"false" });
+    }
+}
+
+impl<T: ValueEncoder> ValueEncoder for Option<T> {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        match self {
+            Some(value) => value.encode(v),
+            None => Null.encode(v),
+        }
+    }
+}
+
+impl KeyEncoder for &str {
+    #[inline]
+    fn write_key<'a>(self, obj: &'a mut ObjectSer) -> ValueSer<'a> {
+        let obj = &mut *obj;
+        obj.entry_inner(|b| format_escaped_str(b, self))
+    }
+}
+
+impl KeyEncoder for fmt::Arguments<'_> {
+    #[inline]
+    fn write_key<'a>(self, obj: &'a mut ObjectSer) -> ValueSer<'a> {
+        if let Some(key) = self.as_str() {
+            obj.entry_inner(|b| format_escaped_str(b, key))
+        } else {
+            obj.entry_inner(|b| format_escaped_fmt(b, self))
+        }
+    }
+}
+
+/// Represents the JSON null value.
+pub struct Null;
+
+impl ValueEncoder for Null {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        v.write_raw_json(b"null");
+    }
+}
+
+impl<T: ValueEncoder> ValueEncoder for Vec<T> {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        value_as_list!(|v| {
+            for t in self {
+                v.entry().value(t);
+            }
+        });
+    }
+}
+
+impl<T: Copy + ValueEncoder> ValueEncoder for &[T] {
+    #[inline]
+    fn encode(self, v: ValueSer<'_>) {
+        value_as_list!(|v| {
+            for t in self {
+                v.entry().value(t);
+            }
+        });
+    }
+}
+
+impl<K: KeyEncoder, V: ValueEncoder, S> ValueEncoder for HashMap<K, V, S> {
+    #[inline]
+    fn encode(self, o: ValueSer<'_>) {
+        value_as_object!(|o| {
+            for (k, v) in self {
+                o.entry(k, v);
+            }
+        });
+    }
+}
+
+impl<K: KeyEncoder, V: ValueEncoder> ValueEncoder for BTreeMap<K, V> {
+    #[inline]
+    fn encode(self, o: ValueSer<'_>) {
+        value_as_object!(|o| {
+            for (k, v) in self {
+                o.entry(k, v);
+            }
+        });
+    }
+}

libs/remote_storage/Cargo.toml

@@ -13,6 +13,7 @@ aws-smithy-async.workspace = true
 aws-smithy-types.workspace = true
 aws-config.workspace = true
 aws-sdk-s3.workspace = true
+base64.workspace = true
 bytes.workspace = true
 camino = { workspace = true, features = ["serde1"] }
 humantime-serde.workspace = true
@@ -41,6 +42,9 @@ http-body-util.workspace = true
 itertools.workspace = true
 sync_wrapper = { workspace = true, features = ["futures"] }
 
+byteorder = "1.4"
+rand = "0.8.5"
+
 [dev-dependencies]
 camino-tempfile.workspace = true
 test-context.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -14,17 +14,25 @@ use anyhow::{Context, Result, anyhow};
 use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
 use azure_core::{Continuable, HttpClient, RetryOptions, TransportOptions};
 use azure_storage::StorageCredentials;
-use azure_storage_blobs::blob::operations::GetBlobBuilder;
+use azure_storage_blobs::blob::BlobBlockType;
+use azure_storage_blobs::blob::BlockList;
 use azure_storage_blobs::blob::{Blob, CopyStatus};
 use azure_storage_blobs::container::operations::ListBlobsBuilder;
-use azure_storage_blobs::prelude::{ClientBuilder, ContainerClient};
+use azure_storage_blobs::prelude::ClientBuilder;
+use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerClient};
+use base64::{Engine as _, engine::general_purpose::URL_SAFE};
+use byteorder::{BigEndian, ByteOrder};
 use bytes::Bytes;
+use camino::Utf8Path;
 use futures::FutureExt;
 use futures::future::Either;
 use futures::stream::Stream;
 use futures_util::{StreamExt, TryStreamExt};
 use http_types::{StatusCode, Url};
 use scopeguard::ScopeGuard;
+use tokio::fs::File;
+use tokio::io::AsyncReadExt;
+use tokio::io::AsyncSeekExt;
 use tokio_util::sync::CancellationToken;
 use tracing::debug;
 use utils::backoff;
@@ -51,6 +59,9 @@ pub struct AzureBlobStorage {
 
     // Alternative timeout used for metadata objects which are expected to be small
     pub small_timeout: Duration,
+    /* BEGIN_HADRON */
+    pub put_block_size_mb: Option<usize>,
+    /* END_HADRON */
 }
 
 impl AzureBlobStorage {
@@ -107,6 +118,9 @@ impl AzureBlobStorage {
             concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
             timeout,
             small_timeout,
+            /* BEGIN_HADRON */
+            put_block_size_mb: azure_config.put_block_size_mb,
+            /* END_HADRON */
         })
     }
 
@@ -583,31 +597,137 @@ impl RemoteStorage for AzureBlobStorage {
 
         let started_at = start_measuring_requests(kind);
 
-        let op = async {
+        let mut metadata_map = metadata.unwrap_or([].into());
+        let timeline_file_path = metadata_map.0.remove("databricks_azure_put_block");
+
+        /* BEGIN_HADRON */
+        let op = async move {
             let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+            let put_block_size = self.put_block_size_mb.unwrap_or(0) * 1024 * 1024;
+            if timeline_file_path.is_none() || put_block_size == 0 {
+                // Use put_block_blob directly.
+                let from: Pin<
+                    Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>,
+                > = Box::pin(from);
+                let from = NonSeekableStream::new(from, data_size_bytes);
+                let body = azure_core::Body::SeekableStream(Box::new(from));
 
-            let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
-                Box::pin(from);
+                let mut builder = blob_client.put_block_blob(body);
+                if !metadata_map.0.is_empty() {
+                    builder = builder.metadata(to_azure_metadata(metadata_map));
+                }
+                let fut = builder.into_future();
+                let fut = tokio::time::timeout(self.timeout, fut);
+                let result = fut.await;
+                match result {
+                    Ok(Ok(_response)) => return Ok(()),
+                    Ok(Err(azure)) => return Err(azure.into()),
+                    Err(_timeout) => return Err(TimeoutOrCancel::Timeout.into()),
+                };
+            }
+            // Upload chunks concurrently using Put Block.
+            // Each PutBlock uploads put_block_size bytes of the file.
+            let mut upload_futures: Vec<tokio::task::JoinHandle<Result<(), azure_core::Error>>> =
+                vec![];
+            let mut block_list = BlockList::default();
+            let mut start_bytes = 0u64;
+            let mut remaining_bytes = data_size_bytes;
+            let mut block_list_count = 0;
 
-            let from = NonSeekableStream::new(from, data_size_bytes);
+            while remaining_bytes > 0 {
+                let block_size = std::cmp::min(remaining_bytes, put_block_size);
+                let end_bytes = start_bytes + block_size as u64;
+                let block_id = block_list_count;
+                let timeout = self.timeout;
+                let blob_client = blob_client.clone();
+                let timeline_file = timeline_file_path.clone().unwrap().clone();
 
-            let body = azure_core::Body::SeekableStream(Box::new(from));
+                let mut encoded_block_id = [0u8; 8];
+                BigEndian::write_u64(&mut encoded_block_id, block_id);
+                URL_SAFE.encode(encoded_block_id);
 
-            let mut builder = blob_client.put_block_blob(body);
+                // Put one block.
+                let part_fut = async move {
+                    let mut file = File::open(Utf8Path::new(&timeline_file.clone())).await?;
+                    file.seek(io::SeekFrom::Start(start_bytes)).await?;
+                    let limited_reader = file.take(block_size as u64);
+                    let file_chunk_stream =
+                        tokio_util::io::ReaderStream::with_capacity(limited_reader, 1024 * 1024);
+                    let file_chunk_stream_pin: Pin<
+                        Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>,
+                    > = Box::pin(file_chunk_stream);
+                    let stream_wrapper = NonSeekableStream::new(file_chunk_stream_pin, block_size);
+                    let body = azure_core::Body::SeekableStream(Box::new(stream_wrapper));
+                    // Azure put block takes URL-encoded block ids and all blocks must have the same byte length.
+                    // https://learn.microsoft.com/en-us/rest/api/storageservices/put-block?tabs=microsoft-entra-id#uri-parameters
+                    let builder = blob_client.put_block(encoded_block_id.to_vec(), body);
+                    let fut = builder.into_future();
+                    let fut = tokio::time::timeout(timeout, fut);
+                    let result = fut.await;
+                    tracing::debug!(
+                        "azure put block id-{} size {} start {} end {} file {} response {:#?}",
+                        block_id,
+                        block_size,
+                        start_bytes,
+                        end_bytes,
+                        timeline_file,
+                        result
+                    );
+                    match result {
+                        Ok(Ok(_response)) => Ok(()),
+                        Ok(Err(azure)) => Err(azure),
+                        Err(_timeout) => Err(azure_core::Error::new(
+                            azure_core::error::ErrorKind::Io,
+                            std::io::Error::new(
+                                std::io::ErrorKind::TimedOut,
+                                "Operation timed out",
+                            ),
+                        )),
+                    }
+                };
+                upload_futures.push(tokio::spawn(part_fut));
 
-            if let Some(metadata) = metadata {
-                builder = builder.metadata(to_azure_metadata(metadata));
+                block_list_count += 1;
+                remaining_bytes -= block_size;
+                start_bytes += block_size as u64;
+
+                block_list
+                    .blocks
+                    .push(BlobBlockType::Uncommitted(encoded_block_id.to_vec().into()));
             }
 
+            tracing::debug!(
+                "azure put blocks {} total MB: {} chunk size MB: {}",
+                block_list_count,
+                data_size_bytes / 1024 / 1024,
+                put_block_size / 1024 / 1024
+            );
+            // Wait for all blocks to be uploaded.
+            let upload_results = futures::future::try_join_all(upload_futures).await;
+            if upload_results.is_err() {
+                return Err(anyhow::anyhow!(format!(
+                    "Failed to upload all blocks {:#?}",
+                    upload_results.unwrap_err()
+                )));
+            }
+
+            // Commit the blocks.
+            let mut builder = blob_client.put_block_list(block_list);
+            if !metadata_map.0.is_empty() {
+                builder = builder.metadata(to_azure_metadata(metadata_map));
+            }
             let fut = builder.into_future();
             let fut = tokio::time::timeout(self.timeout, fut);
+            let result = fut.await;
+            tracing::debug!("azure put block list response {:#?}", result);
 
-            match fut.await {
+            match result {
                 Ok(Ok(_response)) => Ok(()),
                 Ok(Err(azure)) => Err(azure.into()),
                 Err(_timeout) => Err(TimeoutOrCancel::Timeout.into()),
             }
         };
+        /* END_HADRON */
 
         let res = tokio::select! {
             res = op => res,
@@ -622,7 +742,6 @@ impl RemoteStorage for AzureBlobStorage {
         crate::metrics::BUCKET_METRICS
             .req_seconds
             .observe_elapsed(kind, outcome, started_at);
-
         res
     }
 

libs/remote_storage/src/config.rs

@@ -195,8 +195,19 @@ pub struct AzureConfig {
     pub max_keys_per_list_response: Option<i32>,
     #[serde(default = "default_azure_conn_pool_size")]
     pub conn_pool_size: usize,
+    /* BEGIN_HADRON */
+    #[serde(default = "default_azure_put_block_size_mb")]
+    pub put_block_size_mb: Option<usize>,
+    /* END_HADRON */
 }
 
+/* BEGIN_HADRON */
+fn default_azure_put_block_size_mb() -> Option<usize> {
+    // Disable parallel upload by default.
+    Some(0)
+}
+/* END_HADRON */
+
 fn default_remote_storage_azure_concurrency_limit() -> NonZeroUsize {
     NonZeroUsize::new(DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT).unwrap()
 }
@@ -213,6 +224,9 @@ impl Debug for AzureConfig {
                 "max_keys_per_list_response",
                 &self.max_keys_per_list_response,
             )
+            /* BEGIN_HADRON */
+            .field("put_block_size_mb", &self.put_block_size_mb)
+            /* END_HADRON */
             .finish()
     }
 }
@@ -352,6 +366,7 @@ timeout = '5s'";
     upload_storage_class = 'INTELLIGENT_TIERING'
     timeout = '7s'
     conn_pool_size = 8
+    put_block_size_mb = 1024
     ";
 
         let config = parse(toml).unwrap();
@@ -367,6 +382,9 @@ timeout = '5s'";
                     concurrency_limit: default_remote_storage_azure_concurrency_limit(),
                     max_keys_per_list_response: DEFAULT_MAX_KEYS_PER_LIST_RESPONSE,
                     conn_pool_size: 8,
+                    /* BEGIN_HADRON */
+                    put_block_size_mb: Some(1024),
+                    /* END_HADRON */
                 }),
                 timeout: Duration::from_secs(7),
                 small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT

libs/remote_storage/src/lib.rs

@@ -732,9 +732,15 @@ impl GenericRemoteStorage {
         })
     }
 
-    pub fn unreliable_wrapper(s: Self, fail_first: u64) -> Self {
-        Self::Unreliable(Arc::new(UnreliableWrapper::new(s, fail_first)))
+    /* BEGIN_HADRON */
+    pub fn unreliable_wrapper(s: Self, fail_first: u64, fail_probability: u64) -> Self {
+        Self::Unreliable(Arc::new(UnreliableWrapper::new(
+            s,
+            fail_first,
+            fail_probability,
+        )))
     }
+    /* END_HADRON */
 
     /// See [`RemoteStorage::upload`], which this method calls with `None` as metadata.
     pub async fn upload_storage_object(

libs/remote_storage/src/simulate_failures.rs

@@ -1,6 +1,8 @@
 //! This module provides a wrapper around a real RemoteStorage implementation that
 //! causes the first N attempts at each upload or download operatio to fail. For
 //! testing purposes.
+use rand::Rng;
+use std::cmp;
 use std::collections::HashMap;
 use std::collections::hash_map::Entry;
 use std::num::NonZeroU32;
@@ -25,6 +27,12 @@ pub struct UnreliableWrapper {
 
     // Tracks how many failed attempts of each operation has been made.
     attempts: Mutex<HashMap<RemoteOp, u64>>,
+
+    /* BEGIN_HADRON */
+    // This the probability of failure for each operation, ranged from [0, 100].
+    // The probability is default to 100, which means that all operations will fail.
+    attempt_failure_probability: u64,
+    /* END_HADRON */
 }
 
 /// Used to identify retries of different unique operation.
@@ -40,7 +48,11 @@ enum RemoteOp {
 }
 
 impl UnreliableWrapper {
-    pub fn new(inner: crate::GenericRemoteStorage, attempts_to_fail: u64) -> Self {
+    pub fn new(
+        inner: crate::GenericRemoteStorage,
+        attempts_to_fail: u64,
+        attempt_failure_probability: u64,
+    ) -> Self {
         assert!(attempts_to_fail > 0);
         let inner = match inner {
             GenericRemoteStorage::AwsS3(s) => GenericRemoteStorage::AwsS3(s),
@@ -51,9 +63,11 @@ impl UnreliableWrapper {
                 panic!("Can't wrap unreliable wrapper unreliably")
             }
         };
+        let actual_attempt_failure_probability = cmp::min(attempt_failure_probability, 100);
         UnreliableWrapper {
             inner,
             attempts_to_fail,
+            attempt_failure_probability: actual_attempt_failure_probability,
             attempts: Mutex::new(HashMap::new()),
         }
     }
@@ -66,6 +80,7 @@ impl UnreliableWrapper {
     ///
     fn attempt(&self, op: RemoteOp) -> anyhow::Result<u64> {
         let mut attempts = self.attempts.lock().unwrap();
+        let mut rng = rand::thread_rng();
 
         match attempts.entry(op) {
             Entry::Occupied(mut e) => {
@@ -75,15 +90,19 @@ impl UnreliableWrapper {
                     *p
                 };
 
-                if attempts_before_this >= self.attempts_to_fail {
-                    // let it succeed
-                    e.remove();
-                    Ok(attempts_before_this)
-                } else {
+                /* BEGIN_HADRON */
+                // If there are more attempts to fail, fail the request by probability.
+                if (attempts_before_this < self.attempts_to_fail)
+                    && (rng.gen_range(0..=100) < self.attempt_failure_probability)
+                {
                     let error =
                         anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());
                     Err(error)
+                } else {
+                    e.remove();
+                    Ok(attempts_before_this)
                 }
+                /* END_HADRON */
             }
             Entry::Vacant(e) => {
                 let error = anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());

libs/remote_storage/tests/common/mod.rs

@@ -165,10 +165,42 @@ pub(crate) async fn upload_remote_data(
 
             let (data, data_len) =
                 upload_stream(format!("remote blob data {i}").into_bytes().into());
+
+            /* BEGIN_HADRON */
+            let mut metadata = None;
+            if matches!(&*task_client, GenericRemoteStorage::AzureBlob(_)) {
+                let file_path = "/tmp/dbx_upload_tmp_file.txt";
+                {
+                    // Open the file in append mode
+                    let mut file = std::fs::OpenOptions::new()
+                        .append(true)
+                        .create(true) // Create the file if it doesn't exist
+                        .open(file_path)?;
+                    // Append some bytes to the file
+                    std::io::Write::write_all(
+                        &mut file,
+                        &format!("remote blob data {i}").into_bytes(),
+                    )?;
+                    file.sync_all()?;
+                }
+                metadata = Some(remote_storage::StorageMetadata::from([(
+                    "databricks_azure_put_block",
+                    file_path,
+                )]));
+            }
+            /* END_HADRON */
+
             task_client
-                .upload(data, data_len, &blob_path, None, &cancel)
+                .upload(data, data_len, &blob_path, metadata, &cancel)
                 .await?;
 
+            // TODO: Check upload is using the put_block upload.
+            // We cannot consume data here since data is moved inside the upload.
+            // let total_bytes = data.fold(0, |acc, chunk| async move {
+            //     acc + chunk.map(|bytes| bytes.len()).unwrap_or(0)
+            // }).await;
+            // assert_eq!(total_bytes, data_len);
+
             Ok::<_, anyhow::Error>((blob_prefix, blob_path))
         });
     }

libs/remote_storage/tests/test_real_azure.rs

@@ -219,6 +219,9 @@ async fn create_azure_client(
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
             conn_pool_size: 8,
+            /* BEGIN_HADRON */
+            put_block_size_mb: Some(1),
+            /* END_HADRON */
         }),
         timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
         small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,

libs/safekeeper_api/src/models.rs

@@ -221,7 +221,7 @@ pub struct TimelineMembershipSwitchRequest {
 pub struct TimelineMembershipSwitchResponse {
     pub previous_conf: Configuration,
     pub current_conf: Configuration,
-    pub term: Term,
+    pub last_log_term: Term,
     pub flush_lsn: Lsn,
 }
 

libs/utils/src/env.rs

@@ -44,3 +44,62 @@ where
         }
     }
 }
+
+/* BEGIN_HADRON */
+pub enum DeploymentMode {
+    Dev,
+    Staging,
+    Prod,
+}
+
+pub fn get_deployment_mode() -> Option<DeploymentMode> {
+    match std::env::var("DEPLOYMENT_MODE") {
+        Ok(env) => match env.as_str() {
+            "development" => Some(DeploymentMode::Dev),
+            "staging" => Some(DeploymentMode::Staging),
+            "production" => Some(DeploymentMode::Prod),
+            _ => {
+                tracing::error!("Unexpected DEPLOYMENT_MODE: {}", env);
+                None
+            }
+        },
+        Err(_) => {
+            tracing::error!("DEPLOYMENT_MODE not set");
+            None
+        }
+    }
+}
+
+pub fn is_dev_or_staging() -> bool {
+    matches!(
+        get_deployment_mode(),
+        Some(DeploymentMode::Dev) | Some(DeploymentMode::Staging)
+    )
+}
+
+pub enum TestingMode {
+    Chaos,
+    Stress,
+}
+
+pub fn get_test_mode() -> Option<TestingMode> {
+    match std::env::var("HADRON_TEST_MODE") {
+        Ok(env) => match env.as_str() {
+            "chaos" => Some(TestingMode::Chaos),
+            "stress" => Some(TestingMode::Stress),
+            _ => {
+                tracing::error!("Unexpected HADRON_TEST_MODE: {}", env);
+                None
+            }
+        },
+        Err(_) => {
+            tracing::error!("HADRON_TEST_MODE not set");
+            None
+        }
+    }
+}
+
+pub fn is_chaos_testing() -> bool {
+    matches!(get_test_mode(), Some(TestingMode::Chaos))
+}
+/* END_HADRON */

libs/utils/src/lib.rs

@@ -99,6 +99,8 @@ pub mod elapsed_accum;
 #[cfg(target_os = "linux")]
 pub mod linux_socket_ioctl;
 
+pub mod metrics_collector;
+
 // Re-export used in macro. Avoids adding git-version as dep in target crates.
 #[doc(hidden)]
 pub use git_version;

libs/utils/src/metrics_collector.rs

@@ -0,0 +1,75 @@
+use std::{
+    sync::{Arc, RwLock},
+    time::{Duration, Instant},
+};
+
+use metrics::{IntGauge, proto::MetricFamily, register_int_gauge};
+use once_cell::sync::Lazy;
+
+pub static METRICS_STALE_MILLIS: Lazy<IntGauge> = Lazy::new(|| {
+    register_int_gauge!(
+        "metrics_metrics_stale_milliseconds",
+        "The current metrics stale time in milliseconds"
+    )
+    .expect("failed to define a metric")
+});
+
+#[derive(Debug)]
+pub struct CollectedMetrics {
+    pub metrics: Vec<MetricFamily>,
+    pub collected_at: Instant,
+}
+
+impl CollectedMetrics {
+    fn new(metrics: Vec<MetricFamily>) -> Self {
+        Self {
+            metrics,
+            collected_at: Instant::now(),
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct MetricsCollector {
+    last_collected: RwLock<Arc<CollectedMetrics>>,
+}
+
+impl MetricsCollector {
+    pub fn new() -> Self {
+        Self {
+            last_collected: RwLock::new(Arc::new(CollectedMetrics::new(vec![]))),
+        }
+    }
+
+    #[tracing::instrument(name = "metrics_collector", skip_all)]
+    pub fn run_once(&self, cache_metrics: bool) -> Arc<CollectedMetrics> {
+        let started = Instant::now();
+        let metrics = metrics::gather();
+        let collected = Arc::new(CollectedMetrics::new(metrics));
+        if cache_metrics {
+            let mut guard = self.last_collected.write().unwrap();
+            *guard = collected.clone();
+        }
+        tracing::info!(
+            "Collected {} metric families in {} ms",
+            collected.metrics.len(),
+            started.elapsed().as_millis()
+        );
+        collected
+    }
+
+    pub fn last_collected(&self) -> Arc<CollectedMetrics> {
+        self.last_collected.read().unwrap().clone()
+    }
+}
+
+impl Default for MetricsCollector {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+// Interval for metrics collection. Currently hard-coded to be the same as the metrics scape interval from the obs agent
+pub static METRICS_COLLECTION_INTERVAL: Duration = Duration::from_secs(30);
+
+pub static METRICS_COLLECTOR: Lazy<MetricsCollector> = Lazy::new(MetricsCollector::default);

libs/utils/src/shard.rs

@@ -171,6 +171,12 @@ impl std::fmt::Display for ShardNumber {
     }
 }
 
+impl std::fmt::Display for ShardCount {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        self.0.fmt(f)
+    }
+}
+
 impl std::fmt::Display for ShardSlug<'_> {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(

pageserver/Cargo.toml

@@ -112,6 +112,7 @@ twox-hash.workspace = true
 procfs.workspace = true
 
 [dev-dependencies]
+base64.workspace = true
 criterion.workspace = true
 hex-literal.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time", "test-util"] }

pageserver/client_grpc/Cargo.toml

@@ -0,0 +1,23 @@
+[package]
+name = "pageserver_client_grpc"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[features]
+testing = ["pageserver_api/testing"]
+
+[dependencies]
+anyhow.workspace = true
+bytes.workspace = true
+compute_api.workspace = true
+futures.workspace = true
+pageserver_api.workspace = true
+pageserver_page_api.workspace = true
+tokio.workspace = true
+tokio-stream.workspace = true
+tokio-util.workspace = true
+tonic.workspace = true
+tracing.workspace = true
+utils.workspace = true
+workspace_hack.workspace = true

pageserver/client_grpc/src/client.rs

@@ -0,0 +1,403 @@
+use std::collections::HashMap;
+use std::num::NonZero;
+use std::sync::Arc;
+
+use anyhow::anyhow;
+use futures::stream::FuturesUnordered;
+use futures::{FutureExt as _, StreamExt as _};
+use tracing::instrument;
+
+use crate::pool::{ChannelPool, ClientGuard, ClientPool, StreamGuard, StreamPool};
+use crate::retry::Retry;
+use crate::split::GetPageSplitter;
+use compute_api::spec::PageserverProtocol;
+use pageserver_api::shard::ShardStripeSize;
+use pageserver_page_api as page_api;
+use utils::id::{TenantId, TimelineId};
+use utils::shard::{ShardCount, ShardIndex, ShardNumber};
+
+/// Max number of concurrent clients per channel (i.e. TCP connection). New channels will be spun up
+/// when full.
+///
+/// TODO: tune all of these constants, and consider making them configurable.
+/// TODO: consider separate limits for unary and streaming clients, so we don't fill up channels
+/// with only streams.
+const MAX_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(16).unwrap();
+
+/// Max number of concurrent unary request clients per shard.
+const MAX_UNARY_CLIENTS: NonZero<usize> = NonZero::new(64).unwrap();
+
+/// Max number of concurrent GetPage streams per shard. The max number of concurrent GetPage
+/// requests is given by `MAX_STREAMS * MAX_STREAM_QUEUE_DEPTH`.
+const MAX_STREAMS: NonZero<usize> = NonZero::new(64).unwrap();
+
+/// Max number of pipelined requests per stream.
+const MAX_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(2).unwrap();
+
+/// Max number of concurrent bulk GetPage streams per shard, used e.g. for prefetches. Because these
+/// are more throughput-oriented, we have a smaller limit but higher queue depth.
+const MAX_BULK_STREAMS: NonZero<usize> = NonZero::new(16).unwrap();
+
+/// Max number of pipelined requests per bulk stream. These are more throughput-oriented and thus
+/// get a larger queue depth.
+const MAX_BULK_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(4).unwrap();
+
+/// A rich Pageserver gRPC client for a single tenant timeline. This client is more capable than the
+/// basic `page_api::Client` gRPC client, and supports:
+///
+/// * Sharded tenants across multiple Pageservers.
+/// * Pooling of connections, clients, and streams for efficient resource use.
+/// * Concurrent use by many callers.
+/// * Internal handling of GetPage bidirectional streams, with pipelining and error handling.
+/// * Automatic retries.
+/// * Observability.
+///
+/// TODO: this client does not support base backups or LSN leases, as these are only used by
+/// compute_ctl. Consider adding this, but LSN leases need concurrent requests on all shards.
+pub struct PageserverClient {
+    // TODO: support swapping out the shard map, e.g. via an ArcSwap.
+    shards: Shards,
+    retry: Retry,
+}
+
+impl PageserverClient {
+    /// Creates a new Pageserver client for a given tenant and timeline. Uses the Pageservers given
+    /// in the shard map, which must be complete and must use gRPC URLs.
+    pub fn new(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_map: HashMap<ShardIndex, String>,
+        stripe_size: ShardStripeSize,
+        auth_token: Option<String>,
+    ) -> anyhow::Result<Self> {
+        let shards = Shards::new(tenant_id, timeline_id, shard_map, stripe_size, auth_token)?;
+        Ok(Self {
+            shards,
+            retry: Retry,
+        })
+    }
+
+    /// Returns whether a relation exists.
+    #[instrument(skip_all, fields(rel=%req.rel, lsn=%req.read_lsn))]
+    pub async fn check_rel_exists(
+        &self,
+        req: page_api::CheckRelExistsRequest,
+    ) -> tonic::Result<page_api::CheckRelExistsResponse> {
+        self.retry
+            .with(async || {
+                // Relation metadata is only available on shard 0.
+                let mut client = self.shards.get_zero().client().await?;
+                client.check_rel_exists(req).await
+            })
+            .await
+    }
+
+    /// Returns the total size of a database, as # of bytes.
+    #[instrument(skip_all, fields(db_oid=%req.db_oid, lsn=%req.read_lsn))]
+    pub async fn get_db_size(
+        &self,
+        req: page_api::GetDbSizeRequest,
+    ) -> tonic::Result<page_api::GetDbSizeResponse> {
+        self.retry
+            .with(async || {
+                // Relation metadata is only available on shard 0.
+                let mut client = self.shards.get_zero().client().await?;
+                client.get_db_size(req).await
+            })
+            .await
+    }
+
+    /// Fetches pages. The `request_id` must be unique across all in-flight requests. Automatically
+    /// splits requests that straddle shard boundaries, and assembles the responses.
+    ///
+    /// Unlike `page_api::Client`, this automatically converts `status_code` into `tonic::Status`
+    /// errors. All responses will have `GetPageStatusCode::Ok`.
+    #[instrument(skip_all, fields(
+        req_id = %req.request_id,
+        class = %req.request_class,
+        rel = %req.rel,
+        blkno = %req.block_numbers[0],
+        blks = %req.block_numbers.len(),
+        lsn = %req.read_lsn,
+    ))]
+    pub async fn get_page(
+        &self,
+        req: page_api::GetPageRequest,
+    ) -> tonic::Result<page_api::GetPageResponse> {
+        // Make sure we have at least one page.
+        if req.block_numbers.is_empty() {
+            return Err(tonic::Status::invalid_argument("no block number"));
+        }
+
+        // Fast path: request is for a single shard.
+        if let Some(shard_id) =
+            GetPageSplitter::is_single_shard(&req, self.shards.count, self.shards.stripe_size)
+        {
+            return self.get_page_for_shard(shard_id, req).await;
+        }
+
+        // Request spans multiple shards. Split it, dispatch concurrent per-shard requests, and
+        // reassemble the responses.
+        //
+        // TODO: when we support shard map updates, we need to detect when it changes and re-split
+        // the request on errors.
+        let mut splitter = GetPageSplitter::split(req, self.shards.count, self.shards.stripe_size);
+
+        let mut shard_requests: FuturesUnordered<_> = splitter
+            .drain_requests()
+            .map(|(shard_id, shard_req)| {
+                // NB: each request will retry internally.
+                self.get_page_for_shard(shard_id, shard_req)
+                    .map(move |result| result.map(|resp| (shard_id, resp)))
+            })
+            .collect();
+
+        while let Some((shard_id, shard_response)) = shard_requests.next().await.transpose()? {
+            splitter.add_response(shard_id, shard_response)?;
+        }
+
+        splitter.assemble_response()
+    }
+
+    /// Fetches pages that belong to the given shard.
+    #[instrument(skip_all, fields(shard = %shard_id))]
+    async fn get_page_for_shard(
+        &self,
+        shard_id: ShardIndex,
+        req: page_api::GetPageRequest,
+    ) -> tonic::Result<page_api::GetPageResponse> {
+        let resp = self
+            .retry
+            .with(async || {
+                let stream = self
+                    .shards
+                    .get(shard_id)?
+                    .stream(req.request_class.is_bulk())
+                    .await;
+                let resp = stream.send(req.clone()).await?;
+
+                // Convert per-request errors into a tonic::Status.
+                if resp.status_code != page_api::GetPageStatusCode::Ok {
+                    return Err(tonic::Status::new(
+                        resp.status_code.into(),
+                        resp.reason.unwrap_or_else(|| String::from("unknown error")),
+                    ));
+                }
+
+                Ok(resp)
+            })
+            .await?;
+
+        // Make sure we got the right number of pages.
+        // NB: check outside of the retry loop, since we don't want to retry this.
+        let (expected, actual) = (req.block_numbers.len(), resp.page_images.len());
+        if expected != actual {
+            return Err(tonic::Status::internal(format!(
+                "expected {expected} pages for shard {shard_id}, got {actual}",
+            )));
+        }
+
+        Ok(resp)
+    }
+
+    /// Returns the size of a relation, as # of blocks.
+    #[instrument(skip_all, fields(rel=%req.rel, lsn=%req.read_lsn))]
+    pub async fn get_rel_size(
+        &self,
+        req: page_api::GetRelSizeRequest,
+    ) -> tonic::Result<page_api::GetRelSizeResponse> {
+        self.retry
+            .with(async || {
+                // Relation metadata is only available on shard 0.
+                let mut client = self.shards.get_zero().client().await?;
+                client.get_rel_size(req).await
+            })
+            .await
+    }
+
+    /// Fetches an SLRU segment.
+    #[instrument(skip_all, fields(kind=%req.kind, segno=%req.segno, lsn=%req.read_lsn))]
+    pub async fn get_slru_segment(
+        &self,
+        req: page_api::GetSlruSegmentRequest,
+    ) -> tonic::Result<page_api::GetSlruSegmentResponse> {
+        self.retry
+            .with(async || {
+                // SLRU segments are only available on shard 0.
+                let mut client = self.shards.get_zero().client().await?;
+                client.get_slru_segment(req).await
+            })
+            .await
+    }
+}
+
+/// Tracks the tenant's shards.
+struct Shards {
+    /// The shard count.
+    ///
+    /// NB: this is 0 for unsharded tenants, following `ShardIndex::unsharded()` convention.
+    count: ShardCount,
+    /// The stripe size. Only used for sharded tenants.
+    stripe_size: ShardStripeSize,
+    /// Shards by shard index.
+    ///
+    /// NB: unsharded tenants use count 0, like `ShardIndex::unsharded()`.
+    ///
+    /// INVARIANT: every shard 0..count is present.
+    /// INVARIANT: shard 0 is always present.
+    map: HashMap<ShardIndex, Shard>,
+}
+
+impl Shards {
+    /// Creates a new set of shards based on a shard map.
+    fn new(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_map: HashMap<ShardIndex, String>,
+        stripe_size: ShardStripeSize,
+        auth_token: Option<String>,
+    ) -> anyhow::Result<Self> {
+        let count = match shard_map.len() {
+            0 => return Err(anyhow!("no shards provided")),
+            1 => ShardCount::new(0), // NB: unsharded tenants use 0, like `ShardIndex::unsharded()`
+            n if n > u8::MAX as usize => return Err(anyhow!("too many shards: {n}")),
+            n => ShardCount::new(n as u8),
+        };
+
+        let mut map = HashMap::new();
+        for (shard_id, url) in shard_map {
+            // The shard index must match the computed shard count, even for unsharded tenants.
+            if shard_id.shard_count != count {
+                return Err(anyhow!("invalid shard index {shard_id}, expected {count}"));
+            }
+            // The shard index' number and count must be consistent.
+            if !shard_id.is_unsharded() && shard_id.shard_number.0 >= shard_id.shard_count.0 {
+                return Err(anyhow!("invalid shard index {shard_id}"));
+            }
+            // The above conditions guarantee that we have all shards 0..count: len() matches count,
+            // shard number < count, and numbers are unique (via hashmap).
+            let shard = Shard::new(url, tenant_id, timeline_id, shard_id, auth_token.clone())?;
+            map.insert(shard_id, shard);
+        }
+
+        Ok(Self {
+            count,
+            stripe_size,
+            map,
+        })
+    }
+
+    /// Looks up the given shard.
+    #[allow(clippy::result_large_err)] // TODO: check perf impact
+    fn get(&self, shard_id: ShardIndex) -> tonic::Result<&Shard> {
+        self.map
+            .get(&shard_id)
+            .ok_or_else(|| tonic::Status::not_found(format!("unknown shard {shard_id}")))
+    }
+
+    /// Returns shard 0.
+    fn get_zero(&self) -> &Shard {
+        self.get(ShardIndex::new(ShardNumber(0), self.count))
+            .expect("always present")
+    }
+}
+
+/// A single shard. Uses dedicated resource pools with the following structure:
+///
+/// * Channel pool: unbounded.
+///   * Unary client pool: MAX_UNARY_CLIENTS.
+///   * Stream client pool: unbounded.
+///     * Stream pool: MAX_STREAMS and MAX_STREAM_QUEUE_DEPTH.
+/// * Bulk channel pool: unbounded.
+///   * Bulk client pool: unbounded.
+///     * Bulk stream pool: MAX_BULK_STREAMS and MAX_BULK_STREAM_QUEUE_DEPTH.
+struct Shard {
+    /// Unary gRPC client pool.
+    client_pool: Arc<ClientPool>,
+    /// GetPage stream pool.
+    stream_pool: Arc<StreamPool>,
+    /// GetPage stream pool for bulk requests, e.g. prefetches.
+    bulk_stream_pool: Arc<StreamPool>,
+}
+
+impl Shard {
+    /// Creates a new shard. It has its own dedicated resource pools.
+    fn new(
+        url: String,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_id: ShardIndex,
+        auth_token: Option<String>,
+    ) -> anyhow::Result<Self> {
+        // Sanity-check that the URL uses gRPC.
+        if PageserverProtocol::from_connstring(&url)? != PageserverProtocol::Grpc {
+            return Err(anyhow!("invalid shard URL {url}: must use gRPC"));
+        }
+
+        // Common channel pool for unary and stream requests. Bounded by client/stream pools.
+        let channel_pool = ChannelPool::new(url.clone(), MAX_CLIENTS_PER_CHANNEL)?;
+
+        // Client pool for unary requests.
+        let client_pool = ClientPool::new(
+            channel_pool.clone(),
+            tenant_id,
+            timeline_id,
+            shard_id,
+            auth_token.clone(),
+            Some(MAX_UNARY_CLIENTS),
+        );
+
+        // GetPage stream pool. Uses a dedicated client pool to avoid starving out unary clients,
+        // but shares a channel pool with it (as it's unbounded).
+        let stream_pool = StreamPool::new(
+            ClientPool::new(
+                channel_pool.clone(),
+                tenant_id,
+                timeline_id,
+                shard_id,
+                auth_token.clone(),
+                None, // unbounded, limited by stream pool
+            ),
+            Some(MAX_STREAMS),
+            MAX_STREAM_QUEUE_DEPTH,
+        );
+
+        // Bulk GetPage stream pool, e.g. for prefetches. Uses dedicated channel/client/stream pools
+        // to avoid head-of-line blocking of latency-sensitive requests.
+        let bulk_stream_pool = StreamPool::new(
+            ClientPool::new(
+                ChannelPool::new(url, MAX_CLIENTS_PER_CHANNEL)?,
+                tenant_id,
+                timeline_id,
+                shard_id,
+                auth_token,
+                None, // unbounded, limited by stream pool
+            ),
+            Some(MAX_BULK_STREAMS),
+            MAX_BULK_STREAM_QUEUE_DEPTH,
+        );
+
+        Ok(Self {
+            client_pool,
+            stream_pool,
+            bulk_stream_pool,
+        })
+    }
+
+    /// Returns a pooled client for this shard.
+    async fn client(&self) -> tonic::Result<ClientGuard> {
+        self.client_pool
+            .get()
+            .await
+            .map_err(|err| tonic::Status::internal(format!("failed to get client: {err}")))
+    }
+
+    /// Returns a pooled stream for this shard. If `bulk` is `true`, uses the dedicated bulk stream
+    /// pool (e.g. for prefetches).
+    async fn stream(&self, bulk: bool) -> StreamGuard {
+        match bulk {
+            false => self.stream_pool.get().await,
+            true => self.bulk_stream_pool.get().await,
+        }
+    }
+}

pageserver/client_grpc/src/lib.rs

@@ -0,0 +1,6 @@
+mod client;
+mod pool;
+mod retry;
+mod split;
+
+pub use client::PageserverClient;

pageserver/client_grpc/src/pool.rs

@@ -0,0 +1,761 @@
+//! This module provides various Pageserver gRPC client resource pools.
+//!
+//! These pools are designed to reuse gRPC resources (connections, clients, and streams) across
+//! multiple concurrent callers (i.e. Postgres backends). This avoids the resource cost and latency
+//! of creating dedicated TCP connections and server tasks for every Postgres backend.
+//!
+//! Each resource has its own, nested pool. The pools are custom-built for the properties of each
+//! resource -- they are different enough that a generic pool isn't suitable.
+//!
+//! * ChannelPool: manages gRPC channels (TCP connections) to a single Pageserver. Multiple clients
+//!   can acquire and use the same channel concurrently (via HTTP/2 stream multiplexing), up to a
+//!   per-channel client limit. Channels may be closed when they are no longer used by any clients.
+//!
+//! * ClientPool: manages gRPC clients for a single tenant shard. Each client acquires a (shared)
+//!   channel from the ChannelPool for the client's lifetime. A client can only be acquired by a
+//!   single caller at a time, and is returned to the pool when dropped. Idle clients may be removed
+//!   from the pool after some time, to free up the channel.
+//!
+//! * StreamPool: manages bidirectional gRPC GetPage streams. Each stream acquires a client from the
+//!   ClientPool for the stream's lifetime. Internal streams are not exposed to callers; instead, it
+//!   returns a guard that can be used to send a single request, to properly enforce queue depth and
+//!   route responses. Internally, the pool will reuse or spin up a suitable stream for the request,
+//!   possibly pipelining multiple requests from multiple callers on the same stream (up to some
+//!   queue depth). Idle streams may be removed from the pool after a while to free up the client.
+//!
+//! Each channel corresponds to one TCP connection. Each client unary request and each stream
+//! corresponds to one HTTP/2 stream and server task.
+//!
+//! TODO: error handling (including custom error types).
+//! TODO: observability.
+
+use std::collections::{BTreeMap, HashMap};
+use std::num::NonZero;
+use std::ops::{Deref, DerefMut};
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex, Weak};
+use std::time::{Duration, Instant};
+
+use futures::StreamExt as _;
+use tokio::sync::mpsc::{Receiver, Sender};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore, mpsc, oneshot};
+use tokio_util::sync::CancellationToken;
+use tonic::transport::{Channel, Endpoint};
+use tracing::{error, warn};
+
+use pageserver_page_api as page_api;
+use utils::id::{TenantId, TimelineId};
+use utils::shard::ShardIndex;
+
+/// Reap channels/clients/streams that have been idle for this long.
+///
+/// TODO: this is per-pool. For nested pools, it can take up to 3x as long for a TCP connection to
+/// be reaped. First, we must wait for an idle stream to be reaped, which marks its client as idle.
+/// Then, we must wait for the idle client to be reaped, which marks its channel as idle. Then, we
+/// must wait for the idle channel to be reaped. Is that a problem? Maybe not, we just have to
+/// account for it when setting the reap threshold. Alternatively, we can immediately reap empty
+/// channels, and/or stream pool clients.
+const REAP_IDLE_THRESHOLD: Duration = match cfg!(any(test, feature = "testing")) {
+    false => Duration::from_secs(180),
+    true => Duration::from_secs(1), // exercise reaping in tests
+};
+
+/// Reap idle resources with this interval.
+const REAP_IDLE_INTERVAL: Duration = match cfg!(any(test, feature = "testing")) {
+    false => Duration::from_secs(10),
+    true => Duration::from_secs(1), // exercise reaping in tests
+};
+
+/// A gRPC channel pool, for a single Pageserver. A channel is shared by many clients (via HTTP/2
+/// stream multiplexing), up to `clients_per_channel` -- a new channel will be spun up beyond this.
+/// The pool does not limit the number of channels, and instead relies on `ClientPool` or
+/// `StreamPool` to limit the number of concurrent clients.
+///
+/// The pool is always wrapped in an outer `Arc`, to allow long-lived guards across tasks/threads.
+///
+/// TODO: consider prewarming a set of channels, to avoid initial connection latency.
+/// TODO: consider adding a circuit breaker for errors and fail fast.
+pub struct ChannelPool {
+    /// Pageserver endpoint to connect to.
+    endpoint: Endpoint,
+    /// Max number of clients per channel. Beyond this, a new channel will be created.
+    max_clients_per_channel: NonZero<usize>,
+    /// Open channels.
+    channels: Mutex<BTreeMap<ChannelID, ChannelEntry>>,
+    /// Reaps idle channels.
+    idle_reaper: Reaper,
+    /// Channel ID generator.
+    next_channel_id: AtomicUsize,
+}
+
+type ChannelID = usize;
+
+struct ChannelEntry {
+    /// The gRPC channel (i.e. TCP connection). Shared by multiple clients.
+    channel: Channel,
+    /// Number of clients using this channel.
+    clients: usize,
+    /// The channel has been idle (no clients) since this time. None if channel is in use.
+    /// INVARIANT: Some if clients == 0, otherwise None.
+    idle_since: Option<Instant>,
+}
+
+impl ChannelPool {
+    /// Creates a new channel pool for the given Pageserver endpoint.
+    pub fn new<E>(endpoint: E, max_clients_per_channel: NonZero<usize>) -> anyhow::Result<Arc<Self>>
+    where
+        E: TryInto<Endpoint> + Send + Sync + 'static,
+        <E as TryInto<Endpoint>>::Error: std::error::Error + Send + Sync,
+    {
+        let pool = Arc::new(Self {
+            endpoint: endpoint.try_into()?,
+            max_clients_per_channel,
+            channels: Mutex::default(),
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
+            next_channel_id: AtomicUsize::default(),
+        });
+        pool.idle_reaper.spawn(&pool);
+        Ok(pool)
+    }
+
+    /// Acquires a gRPC channel for a client. Multiple clients may acquire the same channel.
+    ///
+    /// This never blocks (except for mutex acquisition). The channel is connected lazily on first
+    /// use, and the `ChannelPool` does not have a channel limit. Channels will be re-established
+    /// automatically on failure (TODO: verify).
+    ///
+    /// Callers should not clone the returned channel, and must hold onto the returned guard as long
+    /// as the channel is in use. It is unfortunately not possible to enforce this: the Protobuf
+    /// client requires an owned `Channel` and we don't have access to the channel's internal
+    /// refcount.
+    ///
+    /// This is not performance-sensitive. It is only called when creating a new client, and clients
+    /// are pooled and reused by `ClientPool`. The total number of channels will also be small. O(n)
+    /// performance is therefore okay.
+    pub fn get(self: &Arc<Self>) -> ChannelGuard {
+        let mut channels = self.channels.lock().unwrap();
+
+        // Try to find an existing channel with available capacity. We check entries in BTreeMap
+        // order, to fill up the lower-ordered channels first. The ClientPool also prefers clients
+        // with lower-ordered channel IDs first. This will cluster clients in lower-ordered
+        // channels, and free up higher-ordered channels such that they can be reaped.
+        for (&id, entry) in channels.iter_mut() {
+            assert!(
+                entry.clients <= self.max_clients_per_channel.get(),
+                "channel overflow"
+            );
+            assert_eq!(
+                entry.idle_since.is_some(),
+                entry.clients == 0,
+                "incorrect channel idle state"
+            );
+            if entry.clients < self.max_clients_per_channel.get() {
+                entry.clients += 1;
+                entry.idle_since = None;
+                return ChannelGuard {
+                    pool: Arc::downgrade(self),
+                    id,
+                    channel: Some(entry.channel.clone()),
+                };
+            }
+        }
+
+        // Create a new channel. We connect lazily on first use, such that we don't block here and
+        // other clients can join onto the same channel while it's connecting.
+        let channel = self.endpoint.connect_lazy();
+
+        let id = self.next_channel_id.fetch_add(1, Ordering::Relaxed);
+        let entry = ChannelEntry {
+            channel: channel.clone(),
+            clients: 1, // account for the guard below
+            idle_since: None,
+        };
+        channels.insert(id, entry);
+
+        ChannelGuard {
+            pool: Arc::downgrade(self),
+            id,
+            channel: Some(channel),
+        }
+    }
+}
+
+impl Reapable for ChannelPool {
+    /// Reaps channels that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.channels.lock().unwrap().retain(|_, entry| {
+            let Some(idle_since) = entry.idle_since else {
+                assert_ne!(entry.clients, 0, "empty channel not marked idle");
+                return true;
+            };
+            assert_eq!(entry.clients, 0, "idle channel has clients");
+            idle_since >= cutoff
+        })
+    }
+}
+
+/// Tracks a channel acquired from the pool. The owned inner channel can be obtained with `take()`,
+/// since the gRPC client requires an owned `Channel`.
+pub struct ChannelGuard {
+    pool: Weak<ChannelPool>,
+    id: ChannelID,
+    channel: Option<Channel>,
+}
+
+impl ChannelGuard {
+    /// Returns the inner owned channel. Panics if called more than once. The caller must hold onto
+    /// the guard as long as the channel is in use, and should not clone it.
+    pub fn take(&mut self) -> Channel {
+        self.channel.take().expect("channel already taken")
+    }
+}
+
+/// Returns the channel to the pool.
+impl Drop for ChannelGuard {
+    fn drop(&mut self) {
+        let Some(pool) = self.pool.upgrade() else {
+            return; // pool was dropped
+        };
+
+        let mut channels = pool.channels.lock().unwrap();
+        let entry = channels.get_mut(&self.id).expect("unknown channel");
+        assert!(entry.idle_since.is_none(), "active channel marked idle");
+        assert!(entry.clients > 0, "channel underflow");
+        entry.clients -= 1;
+        if entry.clients == 0 {
+            entry.idle_since = Some(Instant::now()); // mark channel as idle
+        }
+    }
+}
+
+/// A pool of gRPC clients for a single tenant shard. Each client acquires a channel from the inner
+/// `ChannelPool`. A client is only given out to single caller at a time. The pool limits the total
+/// number of concurrent clients to `max_clients` via semaphore.
+///
+/// The pool is always wrapped in an outer `Arc`, to allow long-lived guards across tasks/threads.
+pub struct ClientPool {
+    /// Tenant ID.
+    tenant_id: TenantId,
+    /// Timeline ID.
+    timeline_id: TimelineId,
+    /// Shard ID.
+    shard_id: ShardIndex,
+    /// Authentication token, if any.
+    auth_token: Option<String>,
+    /// Channel pool to acquire channels from.
+    channel_pool: Arc<ChannelPool>,
+    /// Limits the max number of concurrent clients for this pool. None if the pool is unbounded.
+    limiter: Option<Arc<Semaphore>>,
+    /// Idle pooled clients. Acquired clients are removed from here and returned on drop.
+    ///
+    /// The first client in the map will be acquired next. The map is sorted by client ID, which in
+    /// turn is sorted by its channel ID, such that we prefer acquiring idle clients from
+    /// lower-ordered channels. This allows us to free up and reap higher-numbered channels as idle
+    /// clients are reaped.
+    idle: Mutex<BTreeMap<ClientID, ClientEntry>>,
+    /// Reaps idle clients.
+    idle_reaper: Reaper,
+    /// Unique client ID generator.
+    next_client_id: AtomicUsize,
+}
+
+type ClientID = (ChannelID, usize);
+
+struct ClientEntry {
+    /// The pooled gRPC client.
+    client: page_api::Client,
+    /// The channel guard for the channel used by the client.
+    channel_guard: ChannelGuard,
+    /// The client has been idle since this time. All clients in `ClientPool::idle` are idle by
+    /// definition, so this is the time when it was added back to the pool.
+    idle_since: Instant,
+}
+
+impl ClientPool {
+    /// Creates a new client pool for the given tenant shard. Channels are acquired from the given
+    /// `ChannelPool`, which must point to a Pageserver that hosts the tenant shard. Allows up to
+    /// `max_clients` concurrent clients, or unbounded if None.
+    pub fn new(
+        channel_pool: Arc<ChannelPool>,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_id: ShardIndex,
+        auth_token: Option<String>,
+        max_clients: Option<NonZero<usize>>,
+    ) -> Arc<Self> {
+        let pool = Arc::new(Self {
+            tenant_id,
+            timeline_id,
+            shard_id,
+            auth_token,
+            channel_pool,
+            idle: Mutex::default(),
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
+            limiter: max_clients.map(|max| Arc::new(Semaphore::new(max.get()))),
+            next_client_id: AtomicUsize::default(),
+        });
+        pool.idle_reaper.spawn(&pool);
+        pool
+    }
+
+    /// Gets a client from the pool, or creates a new one if necessary. Connections are established
+    /// lazily and do not block, but this call can block if the pool is at `max_clients`. The client
+    /// is returned to the pool when the guard is dropped.
+    ///
+    /// This is moderately performance-sensitive. It is called for every unary request, but these
+    /// establish a new gRPC stream per request so they're already expensive. GetPage requests use
+    /// the `StreamPool` instead.
+    pub async fn get(self: &Arc<Self>) -> anyhow::Result<ClientGuard> {
+        // Acquire a permit if the pool is bounded.
+        let mut permit = None;
+        if let Some(limiter) = self.limiter.clone() {
+            permit = Some(limiter.acquire_owned().await.expect("never closed"));
+        }
+
+        // Fast path: acquire an idle client from the pool.
+        if let Some((id, entry)) = self.idle.lock().unwrap().pop_first() {
+            return Ok(ClientGuard {
+                pool: Arc::downgrade(self),
+                id,
+                client: Some(entry.client),
+                channel_guard: Some(entry.channel_guard),
+                permit,
+            });
+        }
+
+        // Slow path: construct a new client.
+        let mut channel_guard = self.channel_pool.get();
+        let client = page_api::Client::new(
+            channel_guard.take(),
+            self.tenant_id,
+            self.timeline_id,
+            self.shard_id,
+            self.auth_token.clone(),
+            None,
+        )?;
+
+        Ok(ClientGuard {
+            pool: Arc::downgrade(self),
+            id: (
+                channel_guard.id,
+                self.next_client_id.fetch_add(1, Ordering::Relaxed),
+            ),
+            client: Some(client),
+            channel_guard: Some(channel_guard),
+            permit,
+        })
+    }
+}
+
+impl Reapable for ClientPool {
+    /// Reaps clients that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.idle
+            .lock()
+            .unwrap()
+            .retain(|_, entry| entry.idle_since >= cutoff)
+    }
+}
+
+/// A client acquired from the pool. The inner client can be accessed via Deref. The client is
+/// returned to the pool when dropped.
+pub struct ClientGuard {
+    pool: Weak<ClientPool>,
+    id: ClientID,
+    client: Option<page_api::Client>,     // Some until dropped
+    channel_guard: Option<ChannelGuard>,  // Some until dropped
+    permit: Option<OwnedSemaphorePermit>, // None if pool is unbounded
+}
+
+impl Deref for ClientGuard {
+    type Target = page_api::Client;
+
+    fn deref(&self) -> &Self::Target {
+        self.client.as_ref().expect("not dropped")
+    }
+}
+
+impl DerefMut for ClientGuard {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.client.as_mut().expect("not dropped")
+    }
+}
+
+/// Returns the client to the pool.
+impl Drop for ClientGuard {
+    fn drop(&mut self) {
+        let Some(pool) = self.pool.upgrade() else {
+            return; // pool was dropped
+        };
+
+        let entry = ClientEntry {
+            client: self.client.take().expect("dropped once"),
+            channel_guard: self.channel_guard.take().expect("dropped once"),
+            idle_since: Instant::now(),
+        };
+        pool.idle.lock().unwrap().insert(self.id, entry);
+
+        _ = self.permit; // returned on drop, referenced for visibility
+    }
+}
+
+/// A pool of bidirectional gRPC streams. Currently only used for GetPage streams. Each stream
+/// acquires a client from the inner `ClientPool` for the stream's lifetime.
+///
+/// Individual streams are not exposed to callers -- instead, the returned guard can be used to send
+/// a single request and await the response. Internally, requests are multiplexed across streams and
+/// channels. This allows proper queue depth enforcement and response routing.
+///
+/// TODO: consider making this generic over request and response types; not currently needed.
+pub struct StreamPool {
+    /// The client pool to acquire clients from. Must be unbounded.
+    client_pool: Arc<ClientPool>,
+    /// All pooled streams.
+    ///
+    /// Incoming requests will be sent over an existing stream with available capacity. If all
+    /// streams are full, a new one is spun up and added to the pool (up to `max_streams`). Each
+    /// stream has an associated Tokio task that processes requests and responses.
+    streams: Mutex<HashMap<StreamID, StreamEntry>>,
+    /// The max number of concurrent streams, or None if unbounded.
+    max_streams: Option<NonZero<usize>>,
+    /// The max number of concurrent requests per stream.
+    max_queue_depth: NonZero<usize>,
+    /// Limits the max number of concurrent requests, given by `max_streams * max_queue_depth`.
+    /// None if the pool is unbounded.
+    limiter: Option<Arc<Semaphore>>,
+    /// Reaps idle streams.
+    idle_reaper: Reaper,
+    /// Stream ID generator.
+    next_stream_id: AtomicUsize,
+}
+
+type StreamID = usize;
+type RequestSender = Sender<(page_api::GetPageRequest, ResponseSender)>;
+type RequestReceiver = Receiver<(page_api::GetPageRequest, ResponseSender)>;
+type ResponseSender = oneshot::Sender<tonic::Result<page_api::GetPageResponse>>;
+
+struct StreamEntry {
+    /// Sends caller requests to the stream task. The stream task exits when this is dropped.
+    sender: RequestSender,
+    /// Number of in-flight requests on this stream.
+    queue_depth: usize,
+    /// The time when this stream went idle (queue_depth == 0).
+    /// INVARIANT: Some if queue_depth == 0, otherwise None.
+    idle_since: Option<Instant>,
+}
+
+impl StreamPool {
+    /// Creates a new stream pool, using the given client pool. It will send up to `max_queue_depth`
+    /// concurrent requests on each stream, and use up to `max_streams` concurrent streams.
+    ///
+    /// The client pool must be unbounded. The stream pool will enforce its own limits, and because
+    /// streams are long-lived they can cause persistent starvation if they exhaust the client pool.
+    /// The stream pool should generally have its own dedicated client pool (but it can share a
+    /// channel pool with others since these are always unbounded).
+    pub fn new(
+        client_pool: Arc<ClientPool>,
+        max_streams: Option<NonZero<usize>>,
+        max_queue_depth: NonZero<usize>,
+    ) -> Arc<Self> {
+        assert!(client_pool.limiter.is_none(), "bounded client pool");
+        let pool = Arc::new(Self {
+            client_pool,
+            streams: Mutex::default(),
+            limiter: max_streams.map(|max_streams| {
+                Arc::new(Semaphore::new(max_streams.get() * max_queue_depth.get()))
+            }),
+            max_streams,
+            max_queue_depth,
+            idle_reaper: Reaper::new(REAP_IDLE_THRESHOLD, REAP_IDLE_INTERVAL),
+            next_stream_id: AtomicUsize::default(),
+        });
+        pool.idle_reaper.spawn(&pool);
+        pool
+    }
+
+    /// Acquires an available stream from the pool, or spins up a new stream async if all streams
+    /// are full. Returns a guard that can be used to send a single request on the stream and await
+    /// the response, with queue depth quota already acquired. Blocks if the pool is at capacity
+    /// (i.e. `CLIENT_LIMIT * STREAM_QUEUE_DEPTH` requests in flight).
+    ///
+    /// This is very performance-sensitive, as it is on the GetPage hot path.
+    ///
+    /// TODO: this must do something more sophisticated for performance. We want:
+    ///
+    /// * Cheap, concurrent access in the common case where we can use a pooled stream.
+    /// * Quick acquisition of pooled streams with available capacity.
+    /// * Prefer streams that belong to lower-numbered channels, to reap idle channels.
+    /// * Prefer filling up existing streams' queue depth before spinning up new streams.
+    /// * Don't hold a lock while spinning up new streams.
+    /// * Allow concurrent clients to join onto streams while they're spun up.
+    /// * Allow spinning up multiple streams concurrently, but don't overshoot limits.
+    ///
+    /// For now, we just do something simple but inefficient (linear scan under mutex).
+    pub async fn get(self: &Arc<Self>) -> StreamGuard {
+        // Acquire a permit if the pool is bounded.
+        let mut permit = None;
+        if let Some(limiter) = self.limiter.clone() {
+            permit = Some(limiter.acquire_owned().await.expect("never closed"));
+        }
+        let mut streams = self.streams.lock().unwrap();
+
+        // Look for a pooled stream with available capacity.
+        for (&id, entry) in streams.iter_mut() {
+            assert!(
+                entry.queue_depth <= self.max_queue_depth.get(),
+                "stream queue overflow"
+            );
+            assert_eq!(
+                entry.idle_since.is_some(),
+                entry.queue_depth == 0,
+                "incorrect stream idle state"
+            );
+            if entry.queue_depth < self.max_queue_depth.get() {
+                entry.queue_depth += 1;
+                entry.idle_since = None;
+                return StreamGuard {
+                    pool: Arc::downgrade(self),
+                    id,
+                    sender: entry.sender.clone(),
+                    permit,
+                };
+            }
+        }
+
+        // No available stream, spin up a new one. We install the stream entry in the pool first and
+        // return the guard, while spinning up the stream task async. This allows other callers to
+        // join onto this stream and also create additional streams concurrently if this fills up.
+        let id = self.next_stream_id.fetch_add(1, Ordering::Relaxed);
+        let (req_tx, req_rx) = mpsc::channel(self.max_queue_depth.get());
+        let entry = StreamEntry {
+            sender: req_tx.clone(),
+            queue_depth: 1, // reserve quota for this caller
+            idle_since: None,
+        };
+        streams.insert(id, entry);
+
+        if let Some(max_streams) = self.max_streams {
+            assert!(streams.len() <= max_streams.get(), "stream overflow");
+        };
+
+        let client_pool = self.client_pool.clone();
+        let pool = Arc::downgrade(self);
+
+        tokio::spawn(async move {
+            if let Err(err) = Self::run_stream(client_pool, req_rx).await {
+                error!("stream failed: {err}");
+            }
+            // Remove stream from pool on exit. Weak reference to avoid holding the pool alive.
+            if let Some(pool) = pool.upgrade() {
+                let entry = pool.streams.lock().unwrap().remove(&id);
+                assert!(entry.is_some(), "unknown stream ID: {id}");
+            }
+        });
+
+        StreamGuard {
+            pool: Arc::downgrade(self),
+            id,
+            sender: req_tx,
+            permit,
+        }
+    }
+
+    /// Runs a stream task. This acquires a client from the `ClientPool` and establishes a
+    /// bidirectional GetPage stream, then forwards requests and responses between callers and the
+    /// stream. It does not track or enforce queue depths -- that's done by `get()` since it must be
+    /// atomic with pool stream acquisition.
+    ///
+    /// The task exits when the request channel is closed, or on a stream error. The caller is
+    /// responsible for removing the stream from the pool on exit.
+    async fn run_stream(
+        client_pool: Arc<ClientPool>,
+        mut caller_rx: RequestReceiver,
+    ) -> anyhow::Result<()> {
+        // Acquire a client from the pool and create a stream.
+        let mut client = client_pool.get().await?;
+
+        // NB: use an unbounded channel such that the stream send never blocks. Otherwise, we could
+        // theoretically deadlock if both the client and server block on sends (since we're not
+        // reading responses while sending). This is unlikely to happen due to gRPC/TCP buffers and
+        // low queue depths, but it was seen to happen with the libpq protocol so better safe than
+        // sorry. It should never buffer more than the queue depth anyway, but using an unbounded
+        // channel guarantees that it will never block.
+        let (req_tx, req_rx) = mpsc::unbounded_channel();
+        let req_stream = tokio_stream::wrappers::UnboundedReceiverStream::new(req_rx);
+        let mut resp_stream = client.get_pages(req_stream).await?;
+
+        // Track caller response channels by request ID. If the task returns early, these response
+        // channels will be dropped and the waiting callers will receive an error.
+        let mut callers = HashMap::new();
+
+        // Process requests and responses.
+        loop {
+            tokio::select! {
+                // Receive requests from callers and send them to the stream.
+                req = caller_rx.recv() => {
+                    // Shut down if request channel is closed.
+                    let Some((req, resp_tx)) = req else {
+                        return Ok(());
+                    };
+
+                    // Store the response channel by request ID.
+                    if callers.contains_key(&req.request_id) {
+                        // Error on request ID duplicates. Ignore callers that went away.
+                        _ = resp_tx.send(Err(tonic::Status::invalid_argument(
+                            format!("duplicate request ID: {}", req.request_id),
+                        )));
+                        continue;
+                    }
+                    callers.insert(req.request_id, resp_tx);
+
+                    // Send the request on the stream. Bail out if the stream is closed.
+                    req_tx.send(req).map_err(|_| {
+                        tonic::Status::unavailable("stream closed")
+                    })?;
+                }
+
+                // Receive responses from the stream and send them to callers.
+                resp = resp_stream.next() => {
+                    // Shut down if the stream is closed, and bail out on stream errors.
+                    let Some(resp) = resp.transpose()? else {
+                        return Ok(())
+                    };
+
+                    // Send the response to the caller. Ignore errors if the caller went away.
+                    let Some(resp_tx) = callers.remove(&resp.request_id) else {
+                        warn!("received response for unknown request ID: {}", resp.request_id);
+                        continue;
+                    };
+                    _ = resp_tx.send(Ok(resp));
+                }
+            }
+        }
+    }
+}
+
+impl Reapable for StreamPool {
+    /// Reaps streams that have been idle since before the cutoff.
+    fn reap_idle(&self, cutoff: Instant) {
+        self.streams.lock().unwrap().retain(|_, entry| {
+            let Some(idle_since) = entry.idle_since else {
+                assert_ne!(entry.queue_depth, 0, "empty stream not marked idle");
+                return true;
+            };
+            assert_eq!(entry.queue_depth, 0, "idle stream has requests");
+            idle_since >= cutoff
+        });
+    }
+}
+
+/// A pooled stream reference. Can be used to send a single request, to properly enforce queue
+/// depth. Queue depth is already reserved and will be returned on drop.
+pub struct StreamGuard {
+    pool: Weak<StreamPool>,
+    id: StreamID,
+    sender: RequestSender,
+    permit: Option<OwnedSemaphorePermit>, // None if pool is unbounded
+}
+
+impl StreamGuard {
+    /// Sends a request on the stream and awaits the response. Consumes the guard, since it's only
+    /// valid for a single request (to enforce queue depth). This also drops the guard on return and
+    /// returns the queue depth quota to the pool.
+    ///
+    /// The `GetPageRequest::request_id` must be unique across in-flight requests.
+    ///
+    /// NB: errors are often returned as `GetPageResponse::status_code` instead of `tonic::Status`
+    /// to avoid tearing down the stream for per-request errors. Callers must check this.
+    pub async fn send(
+        self,
+        req: page_api::GetPageRequest,
+    ) -> tonic::Result<page_api::GetPageResponse> {
+        let (resp_tx, resp_rx) = oneshot::channel();
+
+        self.sender
+            .send((req, resp_tx))
+            .await
+            .map_err(|_| tonic::Status::unavailable("stream closed"))?;
+
+        resp_rx
+            .await
+            .map_err(|_| tonic::Status::unavailable("stream closed"))?
+    }
+}
+
+impl Drop for StreamGuard {
+    fn drop(&mut self) {
+        let Some(pool) = self.pool.upgrade() else {
+            return; // pool was dropped
+        };
+
+        // Release the queue depth reservation on drop. This can prematurely decrement it if dropped
+        // before the response is received, but that's okay.
+        let mut streams = pool.streams.lock().unwrap();
+        let entry = streams.get_mut(&self.id).expect("unknown stream");
+        assert!(entry.idle_since.is_none(), "active stream marked idle");
+        assert!(entry.queue_depth > 0, "stream queue underflow");
+        entry.queue_depth -= 1;
+        if entry.queue_depth == 0 {
+            entry.idle_since = Some(Instant::now()); // mark stream as idle
+        }
+
+        _ = self.permit; // returned on drop, referenced for visibility
+    }
+}
+
+/// Periodically reaps idle resources from a pool.
+struct Reaper {
+    /// The task check interval.
+    interval: Duration,
+    /// The threshold for reaping idle resources.
+    threshold: Duration,
+    /// Cancels the reaper task. Cancelled when the reaper is dropped.
+    cancel: CancellationToken,
+}
+
+impl Reaper {
+    /// Creates a new reaper.
+    pub fn new(threshold: Duration, interval: Duration) -> Self {
+        Self {
+            cancel: CancellationToken::new(),
+            threshold,
+            interval,
+        }
+    }
+
+    /// Spawns a task to periodically reap idle resources from the given task pool. The task is
+    /// cancelled when the reaper is dropped.
+    pub fn spawn(&self, pool: &Arc<impl Reapable>) {
+        // NB: hold a weak pool reference, otherwise the task will prevent dropping the pool.
+        let pool = Arc::downgrade(pool);
+        let cancel = self.cancel.clone();
+        let (interval, threshold) = (self.interval, self.threshold);
+
+        tokio::spawn(async move {
+            loop {
+                tokio::select! {
+                    _ = tokio::time::sleep(interval) => {
+                        let Some(pool) = pool.upgrade() else {
+                            return; // pool was dropped
+                        };
+                        pool.reap_idle(Instant::now() - threshold);
+                    }
+
+                    _ = cancel.cancelled() => return,
+                }
+            }
+        });
+    }
+}
+
+impl Drop for Reaper {
+    fn drop(&mut self) {
+        self.cancel.cancel(); // cancel reaper task
+    }
+}
+
+/// A reapable resource pool.
+trait Reapable: Send + Sync + 'static {
+    /// Reaps resources that have been idle since before the given cutoff.
+    fn reap_idle(&self, cutoff: Instant);
+}

pageserver/client_grpc/src/retry.rs

@@ -0,0 +1,151 @@
+use std::time::Duration;
+
+use tokio::time::Instant;
+use tracing::{error, info, warn};
+
+use utils::backoff::exponential_backoff_duration;
+
+/// A retry handler for Pageserver gRPC requests.
+///
+/// This is used instead of backoff::retry for better control and observability.
+pub struct Retry;
+
+impl Retry {
+    /// The per-request timeout.
+    // TODO: tune these, and/or make them configurable. Should we retry forever?
+    const REQUEST_TIMEOUT: Duration = Duration::from_secs(10);
+    /// The total timeout across all attempts
+    const TOTAL_TIMEOUT: Duration = Duration::from_secs(60);
+    /// The initial backoff duration.
+    const BASE_BACKOFF: Duration = Duration::from_millis(10);
+    /// The maximum backoff duration.
+    const MAX_BACKOFF: Duration = Duration::from_secs(10);
+    /// If true, log successful requests. For debugging.
+    const LOG_SUCCESS: bool = false;
+
+    /// Runs the given async closure with timeouts and retries (exponential backoff). Logs errors,
+    /// using the current tracing span for context.
+    ///
+    /// Only certain gRPC status codes are retried, see [`Self::should_retry`]. For default
+    /// timeouts, see [`Self::REQUEST_TIMEOUT`] and [`Self::TOTAL_TIMEOUT`].
+    pub async fn with<T, F, O>(&self, mut f: F) -> tonic::Result<T>
+    where
+        F: FnMut() -> O,
+        O: Future<Output = tonic::Result<T>>,
+    {
+        let started = Instant::now();
+        let deadline = started + Self::TOTAL_TIMEOUT;
+        let mut last_error = None;
+        let mut retries = 0;
+        loop {
+            // Set up a future to wait for the backoff (if any) and run the request with a timeout.
+            let backoff_and_try = async {
+                // NB: sleep() always sleeps 1ms, even when given a 0 argument. See:
+                // https://github.com/tokio-rs/tokio/issues/6866
+                if let Some(backoff) = Self::backoff_duration(retries) {
+                    tokio::time::sleep(backoff).await;
+                }
+
+                let request_started = Instant::now();
+                tokio::time::timeout(Self::REQUEST_TIMEOUT, f())
+                    .await
+                    .map_err(|_| {
+                        tonic::Status::deadline_exceeded(format!(
+                            "request timed out after {:.3}s",
+                            request_started.elapsed().as_secs_f64()
+                        ))
+                    })?
+            };
+
+            // Wait for the backoff and request, or bail out if the total timeout is exceeded.
+            let result = tokio::select! {
+                result = backoff_and_try => result,
+
+                _ = tokio::time::sleep_until(deadline) => {
+                    let last_error = last_error.unwrap_or_else(|| {
+                        tonic::Status::deadline_exceeded(format!(
+                            "request timed out after {:.3}s",
+                            started.elapsed().as_secs_f64()
+                        ))
+                    });
+                    error!(
+                        "giving up after {:.3}s and {retries} retries, last error {:?}: {}",
+                        started.elapsed().as_secs_f64(), last_error.code(), last_error.message(),
+                    );
+                    return Err(last_error);
+                }
+            };
+
+            match result {
+                // Success, return the result.
+                Ok(result) => {
+                    if retries > 0 || Self::LOG_SUCCESS {
+                        info!(
+                            "request succeeded after {retries} retries in {:.3}s",
+                            started.elapsed().as_secs_f64(),
+                        );
+                    }
+
+                    return Ok(result);
+                }
+
+                // Error, retry or bail out.
+                Err(status) => {
+                    let (code, message) = (status.code(), status.message());
+                    let attempt = retries + 1;
+
+                    if !Self::should_retry(code) {
+                        // NB: include the attempt here too. This isn't necessarily the first
+                        // attempt, because the error may change between attempts.
+                        error!(
+                            "request failed with {code:?}: {message}, not retrying (attempt {attempt})"
+                        );
+                        return Err(status);
+                    }
+
+                    warn!("request failed with {code:?}: {message}, retrying (attempt {attempt})");
+
+                    retries += 1;
+                    last_error = Some(status);
+                }
+            }
+        }
+    }
+
+    /// Returns the backoff duration for the given retry attempt, or None for no backoff.
+    fn backoff_duration(retry: usize) -> Option<Duration> {
+        let backoff = exponential_backoff_duration(
+            retry as u32,
+            Self::BASE_BACKOFF.as_secs_f64(),
+            Self::MAX_BACKOFF.as_secs_f64(),
+        );
+        (!backoff.is_zero()).then_some(backoff)
+    }
+
+    /// Returns true if the given status code should be retries.
+    fn should_retry(code: tonic::Code) -> bool {
+        match code {
+            tonic::Code::Ok => panic!("unexpected Ok status code"),
+
+            // These codes are transient, so retry them.
+            tonic::Code::Aborted => true,
+            tonic::Code::Cancelled => true,
+            tonic::Code::DeadlineExceeded => true, // maybe transient slowness
+            tonic::Code::Internal => true,         // maybe transient failure?
+            tonic::Code::ResourceExhausted => true,
+            tonic::Code::Unavailable => true,
+
+            // The following codes will like continue to fail, so don't retry.
+            tonic::Code::AlreadyExists => false,
+            tonic::Code::DataLoss => false,
+            tonic::Code::FailedPrecondition => false,
+            tonic::Code::InvalidArgument => false,
+            tonic::Code::NotFound => false,
+            tonic::Code::OutOfRange => false,
+            tonic::Code::PermissionDenied => false,
+            tonic::Code::Unauthenticated => false,
+            tonic::Code::Unimplemented => false,
+            tonic::Code::Unknown => false,
+        }
+    }
+}

pageserver/client_grpc/src/split.rs

@@ -0,0 +1,172 @@
+use std::collections::HashMap;
+
+use bytes::Bytes;
+
+use pageserver_api::key::rel_block_to_key;
+use pageserver_api::shard::{ShardStripeSize, key_to_shard_number};
+use pageserver_page_api as page_api;
+use utils::shard::{ShardCount, ShardIndex};
+
+/// Splits GetPageRequests that straddle shard boundaries and assembles the responses.
+/// TODO: add tests for this.
+pub struct GetPageSplitter {
+    /// The original request ID. Used for all shard requests.
+    request_id: page_api::RequestID,
+    /// Split requests by shard index.
+    requests: HashMap<ShardIndex, page_api::GetPageRequest>,
+    /// Maps the offset in `GetPageRequest::block_numbers` to the owning shard. Used to assemble
+    /// the response pages in the same order as the original request.
+    block_shards: Vec<ShardIndex>,
+    /// Page responses by shard index. Will be assembled into a single response.
+    responses: HashMap<ShardIndex, Vec<Bytes>>,
+}
+
+impl GetPageSplitter {
+    /// Checks if the given request only touches a single shard, and returns the shard ID. This is
+    /// the common case, so we check first in order to avoid unnecessary allocations and overhead.
+    /// The caller must ensure that the request has at least one block number, or this will panic.
+    pub fn is_single_shard(
+        req: &page_api::GetPageRequest,
+        count: ShardCount,
+        stripe_size: ShardStripeSize,
+    ) -> Option<ShardIndex> {
+        // Fast path: unsharded tenant.
+        if count.is_unsharded() {
+            return Some(ShardIndex::unsharded());
+        }
+
+        // Find the base shard index for the first page, and compare with the rest.
+        let key = rel_block_to_key(req.rel, *req.block_numbers.first().expect("no pages"));
+        let shard_number = key_to_shard_number(count, stripe_size, &key);
+
+        req.block_numbers
+            .iter()
+            .skip(1) // computed above
+            .all(|&blkno| {
+                let key = rel_block_to_key(req.rel, blkno);
+                key_to_shard_number(count, stripe_size, &key) == shard_number
+            })
+            .then_some(ShardIndex::new(shard_number, count))
+    }
+
+    /// Splits the given request.
+    pub fn split(
+        req: page_api::GetPageRequest,
+        count: ShardCount,
+        stripe_size: ShardStripeSize,
+    ) -> Self {
+        // The caller should make sure we don't split requests unnecessarily.
+        debug_assert!(
+            Self::is_single_shard(&req, count, stripe_size).is_none(),
+            "unnecessary request split"
+        );
+
+        // Split the requests by shard index.
+        let mut requests = HashMap::with_capacity(2); // common case
+        let mut block_shards = Vec::with_capacity(req.block_numbers.len());
+        for blkno in req.block_numbers {
+            let key = rel_block_to_key(req.rel, blkno);
+            let shard_number = key_to_shard_number(count, stripe_size, &key);
+            let shard_id = ShardIndex::new(shard_number, count);
+
+            let shard_req = requests
+                .entry(shard_id)
+                .or_insert_with(|| page_api::GetPageRequest {
+                    request_id: req.request_id,
+                    request_class: req.request_class,
+                    rel: req.rel,
+                    read_lsn: req.read_lsn,
+                    block_numbers: Vec::new(),
+                });
+            shard_req.block_numbers.push(blkno);
+            block_shards.push(shard_id);
+        }
+
+        Self {
+            request_id: req.request_id,
+            responses: HashMap::with_capacity(requests.len()),
+            requests,
+            block_shards,
+        }
+    }
+
+    /// Drains the per-shard requests, moving them out of the hashmap to avoid extra allocations.
+    pub fn drain_requests(
+        &mut self,
+    ) -> impl Iterator<Item = (ShardIndex, page_api::GetPageRequest)> {
+        self.requests.drain()
+    }
+
+    /// Adds a response from the given shard.
+    #[allow(clippy::result_large_err)]
+    pub fn add_response(
+        &mut self,
+        shard_id: ShardIndex,
+        response: page_api::GetPageResponse,
+    ) -> tonic::Result<()> {
+        // The caller should already have converted status codes into tonic::Status.
+        assert_eq!(response.status_code, page_api::GetPageStatusCode::Ok);
+
+        // Make sure the response matches the request ID.
+        if response.request_id != self.request_id {
+            return Err(tonic::Status::internal(format!(
+                "response ID {} does not match request ID {}",
+                response.request_id, self.request_id
+            )));
+        }
+
+        // Add the response data to the map.
+        let old = self.responses.insert(shard_id, response.page_images);
+
+        if old.is_some() {
+            return Err(tonic::Status::internal(format!(
+                "duplicate response for shard {shard_id}",
+            )));
+        }
+
+        Ok(())
+    }
+
+    /// Assembles the shard responses into a single response. Responses must be present for all
+    /// relevant shards, and the total number of pages must match the original request.
+    #[allow(clippy::result_large_err)]
+    pub fn assemble_response(self) -> tonic::Result<page_api::GetPageResponse> {
+        let mut response = page_api::GetPageResponse {
+            request_id: self.request_id,
+            status_code: page_api::GetPageStatusCode::Ok,
+            reason: None,
+            page_images: Vec::with_capacity(self.block_shards.len()),
+        };
+
+        // Set up per-shard page iterators we can pull from.
+        let mut shard_responses = HashMap::with_capacity(self.responses.len());
+        for (shard_id, responses) in self.responses {
+            shard_responses.insert(shard_id, responses.into_iter());
+        }
+
+        // Reassemble the responses in the same order as the original request.
+        for shard_id in &self.block_shards {
+            let page = shard_responses
+                .get_mut(shard_id)
+                .ok_or_else(|| {
+                    tonic::Status::internal(format!("missing response for shard {shard_id}"))
+                })?
+                .next()
+                .ok_or_else(|| {
+                    tonic::Status::internal(format!("missing page from shard {shard_id}"))
+                })?;
+            response.page_images.push(page);
+        }
+
+        // Make sure there are no additional pages.
+        for (shard_id, mut pages) in shard_responses {
+            if pages.next().is_some() {
+                return Err(tonic::Status::internal(format!(
+                    "extra pages returned from shard {shard_id}"
+                )));
+            }
+        }
+
+        Ok(response)
+    }
+}

pageserver/ctl/src/index_part.rs

@@ -1,10 +1,101 @@
+use std::str::FromStr;
+
 use anyhow::Context;
 use camino::Utf8PathBuf;
-use pageserver::tenant::IndexPart;
+use pageserver::tenant::{
+    IndexPart,
+    layer_map::{LayerMap, SearchResult},
+    remote_timeline_client::remote_layer_path,
+    storage_layer::{PersistentLayerDesc, ReadableLayerWeak},
+};
+use pageserver_api::key::Key;
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+    shard::TenantShardId,
+};
 
 #[derive(clap::Subcommand)]
 pub(crate) enum IndexPartCmd {
-    Dump { path: Utf8PathBuf },
+    Dump {
+        path: Utf8PathBuf,
+    },
+    /// Find all layers that need to be searched to construct the given page at the given LSN.
+    Search {
+        #[arg(long)]
+        tenant_id: String,
+        #[arg(long)]
+        timeline_id: String,
+        #[arg(long)]
+        path: Utf8PathBuf,
+        #[arg(long)]
+        key: String,
+        #[arg(long)]
+        lsn: String,
+    },
+}
+
+async fn search_layers(
+    tenant_id: &str,
+    timeline_id: &str,
+    path: &Utf8PathBuf,
+    key: &str,
+    lsn: &str,
+) -> anyhow::Result<()> {
+    let tenant_id = TenantId::from_str(tenant_id).unwrap();
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+    let timeline_id = TimelineId::from_str(timeline_id).unwrap();
+    let index_json = {
+        let bytes = tokio::fs::read(path).await?;
+        IndexPart::from_json_bytes(&bytes).unwrap()
+    };
+    let mut layer_map = LayerMap::default();
+    {
+        let mut updates = layer_map.batch_update();
+        for (key, value) in index_json.layer_metadata.iter() {
+            updates.insert_historic(PersistentLayerDesc::from_filename(
+                tenant_shard_id,
+                timeline_id,
+                key.clone(),
+                value.file_size,
+            ));
+        }
+    }
+    let key = Key::from_hex(key)?;
+
+    let lsn = Lsn::from_str(lsn).unwrap();
+    let mut end_lsn = lsn;
+    loop {
+        let result = layer_map.search(key, end_lsn);
+        match result {
+            Some(SearchResult { layer, lsn_floor }) => {
+                let disk_layer = match layer {
+                    ReadableLayerWeak::PersistentLayer(layer) => layer,
+                    ReadableLayerWeak::InMemoryLayer(_) => {
+                        anyhow::bail!("unexpected in-memory layer")
+                    }
+                };
+
+                let metadata = index_json
+                    .layer_metadata
+                    .get(&disk_layer.layer_name())
+                    .unwrap();
+                println!(
+                    "{}",
+                    remote_layer_path(
+                        &tenant_id,
+                        &timeline_id,
+                        metadata.shard,
+                        &disk_layer.layer_name(),
+                        metadata.generation
+                    )
+                );
+                end_lsn = lsn_floor;
+            }
+            None => break,
+        }
+    }
+    Ok(())
 }
 
 pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
@@ -16,5 +107,12 @@ pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
             println!("{output}");
             Ok(())
         }
+        IndexPartCmd::Search {
+            tenant_id,
+            timeline_id,
+            path,
+            key,
+            lsn,
+        } => search_layers(tenant_id, timeline_id, path, key, lsn).await,
     }
 }

pageserver/page_api/src/client.rs

@@ -1,23 +1,151 @@
-use anyhow::Result;
+use anyhow::Context as _;
 use futures::{Stream, StreamExt as _, TryStreamExt as _};
 use tokio::io::AsyncRead;
 use tokio_util::io::StreamReader;
+use tonic::codec::CompressionEncoding;
 use tonic::metadata::AsciiMetadataValue;
-use tonic::metadata::errors::InvalidMetadataValue;
-use tonic::transport::Channel;
-use tonic::{Request, Streaming};
+use tonic::service::Interceptor;
+use tonic::service::interceptor::InterceptedService;
+use tonic::transport::{Channel, Endpoint};
 
-use utils::id::TenantId;
-use utils::id::TimelineId;
+use utils::id::{TenantId, TimelineId};
 use utils::shard::ShardIndex;
 
-use crate::model;
+use crate::model::*;
 use crate::proto;
 
-///
-/// AuthInterceptor adds tenant, timeline, and auth header to the channel. These
-/// headers are required at the pageserver.
-///
+/// A basic Pageserver gRPC client, for a single tenant shard. This API uses native Rust domain
+/// types from `model` rather than generated Protobuf types.
+pub struct Client {
+    inner: proto::PageServiceClient<InterceptedService<Channel, AuthInterceptor>>,
+}
+
+impl Client {
+    /// Connects to the given gRPC endpoint.
+    pub async fn connect<E>(
+        endpoint: E,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_id: ShardIndex,
+        auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self>
+    where
+        E: TryInto<Endpoint> + Send + Sync + 'static,
+        <E as TryInto<Endpoint>>::Error: std::error::Error + Send + Sync,
+    {
+        let endpoint: Endpoint = endpoint.try_into().context("invalid endpoint")?;
+        let channel = endpoint.connect().await?;
+        Self::new(
+            channel,
+            tenant_id,
+            timeline_id,
+            shard_id,
+            auth_token,
+            compression,
+        )
+    }
+
+    /// Creates a new client using the given gRPC channel.
+    pub fn new(
+        channel: Channel,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_id: ShardIndex,
+        auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self> {
+        let auth = AuthInterceptor::new(tenant_id, timeline_id, shard_id, auth_token)?;
+        let mut inner = proto::PageServiceClient::with_interceptor(channel, auth);
+
+        if let Some(compression) = compression {
+            // TODO: benchmark this (including network latency).
+            inner = inner
+                .accept_compressed(compression)
+                .send_compressed(compression);
+        }
+
+        Ok(Self { inner })
+    }
+
+    /// Returns whether a relation exists.
+    pub async fn check_rel_exists(
+        &mut self,
+        req: CheckRelExistsRequest,
+    ) -> tonic::Result<CheckRelExistsResponse> {
+        let req = proto::CheckRelExistsRequest::from(req);
+        let resp = self.inner.check_rel_exists(req).await?.into_inner();
+        Ok(resp.into())
+    }
+
+    /// Fetches a base backup.
+    pub async fn get_base_backup(
+        &mut self,
+        req: GetBaseBackupRequest,
+    ) -> tonic::Result<impl AsyncRead + use<>> {
+        let req = proto::GetBaseBackupRequest::from(req);
+        let chunks = self.inner.get_base_backup(req).await?.into_inner();
+        Ok(StreamReader::new(
+            chunks
+                .map_ok(|resp| resp.chunk)
+                .map_err(std::io::Error::other),
+        ))
+    }
+
+    /// Returns the total size of a database, as # of bytes.
+    pub async fn get_db_size(&mut self, req: GetDbSizeRequest) -> tonic::Result<GetDbSizeResponse> {
+        let req = proto::GetDbSizeRequest::from(req);
+        let resp = self.inner.get_db_size(req).await?.into_inner();
+        Ok(resp.into())
+    }
+
+    /// Fetches pages.
+    ///
+    /// This is implemented as a bidirectional streaming RPC for performance. Per-request errors are
+    /// typically returned as status_code instead of errors, to avoid tearing down the entire stream
+    /// via a tonic::Status error.
+    pub async fn get_pages(
+        &mut self,
+        reqs: impl Stream<Item = GetPageRequest> + Send + 'static,
+    ) -> tonic::Result<impl Stream<Item = tonic::Result<GetPageResponse>> + Send + 'static> {
+        let reqs = reqs.map(proto::GetPageRequest::from);
+        let resps = self.inner.get_pages(reqs).await?.into_inner();
+        Ok(resps.map_ok(GetPageResponse::from))
+    }
+
+    /// Returns the size of a relation, as # of blocks.
+    pub async fn get_rel_size(
+        &mut self,
+        req: GetRelSizeRequest,
+    ) -> tonic::Result<GetRelSizeResponse> {
+        let req = proto::GetRelSizeRequest::from(req);
+        let resp = self.inner.get_rel_size(req).await?.into_inner();
+        Ok(resp.into())
+    }
+
+    /// Fetches an SLRU segment.
+    pub async fn get_slru_segment(
+        &mut self,
+        req: GetSlruSegmentRequest,
+    ) -> tonic::Result<GetSlruSegmentResponse> {
+        let req = proto::GetSlruSegmentRequest::from(req);
+        let resp = self.inner.get_slru_segment(req).await?.into_inner();
+        Ok(resp.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: LeaseLsnRequest) -> tonic::Result<LeaseLsnResponse> {
+        let req = proto::LeaseLsnRequest::from(req);
+        let resp = self.inner.lease_lsn(req).await?.into_inner();
+        Ok(resp.try_into()?)
+    }
+}
+
+/// Adds authentication metadata to gRPC requests.
 #[derive(Clone)]
 struct AuthInterceptor {
     tenant_id: AsciiMetadataValue,
@@ -30,174 +158,29 @@ impl AuthInterceptor {
     fn new(
         tenant_id: TenantId,
         timeline_id: TimelineId,
-        auth_token: Option<String>,
         shard_id: ShardIndex,
-    ) -> Result<Self, InvalidMetadataValue> {
-        let tenant_ascii: AsciiMetadataValue = tenant_id.to_string().try_into()?;
-        let timeline_ascii: AsciiMetadataValue = timeline_id.to_string().try_into()?;
-        let shard_ascii: AsciiMetadataValue = shard_id.to_string().try_into()?;
-
-        let auth_header: Option<AsciiMetadataValue> = match auth_token {
-            Some(token) => Some(format!("Bearer {token}").try_into()?),
-            None => None,
-        };
-
+        auth_token: Option<String>,
+    ) -> anyhow::Result<Self> {
         Ok(Self {
-            tenant_id: tenant_ascii,
-            shard_id: shard_ascii,
-            timeline_id: timeline_ascii,
-            auth_header,
+            tenant_id: tenant_id.to_string().try_into()?,
+            timeline_id: timeline_id.to_string().try_into()?,
+            shard_id: shard_id.to_string().try_into()?,
+            auth_header: auth_token
+                .map(|token| format!("Bearer {token}").try_into())
+                .transpose()?,
         })
     }
 }
 
-impl tonic::service::Interceptor for AuthInterceptor {
-    fn call(&mut self, mut req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
-        req.metadata_mut()
-            .insert("neon-tenant-id", self.tenant_id.clone());
-        req.metadata_mut()
-            .insert("neon-shard-id", self.shard_id.clone());
-        req.metadata_mut()
-            .insert("neon-timeline-id", self.timeline_id.clone());
-        if let Some(auth_header) = &self.auth_header {
-            req.metadata_mut()
-                .insert("authorization", auth_header.clone());
+impl Interceptor for AuthInterceptor {
+    fn call(&mut self, mut req: tonic::Request<()>) -> tonic::Result<tonic::Request<()>> {
+        let metadata = req.metadata_mut();
+        metadata.insert("neon-tenant-id", self.tenant_id.clone());
+        metadata.insert("neon-timeline-id", self.timeline_id.clone());
+        metadata.insert("neon-shard-id", self.shard_id.clone());
+        if let Some(ref auth_header) = self.auth_header {
+            metadata.insert("authorization", auth_header.clone());
         }
         Ok(req)
     }
 }
-
-#[derive(Clone)]
-pub struct Client {
-    client: proto::PageServiceClient<
-        tonic::service::interceptor::InterceptedService<Channel, AuthInterceptor>,
-    >,
-}
-
-impl Client {
-    pub async fn new<T: TryInto<tonic::transport::Endpoint> + Send + Sync + 'static>(
-        into_endpoint: T,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        shard_id: ShardIndex,
-        auth_header: Option<String>,
-        compression: Option<tonic::codec::CompressionEncoding>,
-    ) -> anyhow::Result<Self> {
-        let endpoint: tonic::transport::Endpoint = into_endpoint
-            .try_into()
-            .map_err(|_e| anyhow::anyhow!("failed to convert endpoint"))?;
-        let channel = endpoint.connect().await?;
-        let auth = AuthInterceptor::new(tenant_id, timeline_id, auth_header, shard_id)
-            .map_err(|e| anyhow::anyhow!(e.to_string()))?;
-        let mut client = proto::PageServiceClient::with_interceptor(channel, auth);
-
-        if let Some(compression) = compression {
-            // TODO: benchmark this (including network latency).
-            client = client
-                .accept_compressed(compression)
-                .send_compressed(compression);
-        }
-
-        Ok(Self { client })
-    }
-
-    /// Returns whether a relation exists.
-    pub async fn check_rel_exists(
-        &mut self,
-        req: model::CheckRelExistsRequest,
-    ) -> Result<model::CheckRelExistsResponse, tonic::Status> {
-        let proto_req = proto::CheckRelExistsRequest::from(req);
-
-        let response = self.client.check_rel_exists(proto_req).await?;
-
-        let proto_resp = response.into_inner();
-        Ok(proto_resp.into())
-    }
-
-    /// Fetches a base backup.
-    pub async fn get_base_backup(
-        &mut self,
-        req: model::GetBaseBackupRequest,
-    ) -> 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.
-    pub async fn get_db_size(
-        &mut self,
-        req: model::GetDbSizeRequest,
-    ) -> Result<u64, tonic::Status> {
-        let proto_req = proto::GetDbSizeRequest::from(req);
-
-        let response = self.client.get_db_size(proto_req).await?;
-        Ok(response.into_inner().into())
-    }
-
-    /// Fetches pages.
-    ///
-    /// This is implemented as a bidirectional streaming RPC for performance.
-    /// Per-request errors are often returned as status_code instead of errors,
-    /// to avoid tearing down the entire stream via tonic::Status.
-    pub async fn get_pages<ReqSt>(
-        &mut self,
-        inbound: ReqSt,
-    ) -> Result<
-        impl Stream<Item = Result<model::GetPageResponse, tonic::Status>> + Send + 'static,
-        tonic::Status,
-    >
-    where
-        ReqSt: Stream<Item = model::GetPageRequest> + Send + 'static,
-    {
-        let outbound_proto = inbound.map(|domain_req| domain_req.into());
-
-        let req_new = Request::new(outbound_proto);
-
-        let response_stream: Streaming<proto::GetPageResponse> =
-            self.client.get_pages(req_new).await?.into_inner();
-
-        let domain_stream = response_stream.map_ok(model::GetPageResponse::from);
-
-        Ok(domain_stream)
-    }
-
-    /// Returns the size of a relation, as # of blocks.
-    pub async fn get_rel_size(
-        &mut self,
-        req: model::GetRelSizeRequest,
-    ) -> Result<model::GetRelSizeResponse, tonic::Status> {
-        let proto_req = proto::GetRelSizeRequest::from(req);
-        let response = self.client.get_rel_size(proto_req).await?;
-        let proto_resp = response.into_inner();
-        Ok(proto_resp.into())
-    }
-
-    /// Fetches an SLRU segment.
-    pub async fn get_slru_segment(
-        &mut self,
-        req: model::GetSlruSegmentRequest,
-    ) -> Result<model::GetSlruSegmentResponse, tonic::Status> {
-        let proto_req = proto::GetSlruSegmentRequest::from(req);
-        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

@@ -384,7 +384,7 @@ impl From<GetPageRequest> for proto::GetPageRequest {
 pub type RequestID = u64;
 
 /// A GetPage request class.
-#[derive(Clone, Copy, Debug)]
+#[derive(Clone, Copy, Debug, strum_macros::Display)]
 pub enum GetPageClass {
     /// Unknown class. For backwards compatibility: used when an older client version sends a class
     /// that a newer server version has removed.
@@ -397,6 +397,19 @@ pub enum GetPageClass {
     Background,
 }
 
+impl GetPageClass {
+    /// Returns true if this is considered a bulk request (i.e. more throughput-oriented rather than
+    /// latency-sensitive).
+    pub fn is_bulk(&self) -> bool {
+        match self {
+            Self::Unknown => false,
+            Self::Normal => false,
+            Self::Prefetch => true,
+            Self::Background => true,
+        }
+    }
+}
+
 impl From<proto::GetPageClass> for GetPageClass {
     fn from(pb: proto::GetPageClass) -> Self {
         match pb {
@@ -602,6 +615,21 @@ impl TryFrom<tonic::Code> for GetPageStatusCode {
     }
 }
 
+impl From<GetPageStatusCode> for tonic::Code {
+    fn from(status_code: GetPageStatusCode) -> Self {
+        use tonic::Code;
+
+        match status_code {
+            GetPageStatusCode::Unknown => Code::Unknown,
+            GetPageStatusCode::Ok => Code::Ok,
+            GetPageStatusCode::NotFound => Code::NotFound,
+            GetPageStatusCode::InvalidRequest => Code::InvalidArgument,
+            GetPageStatusCode::InternalError => Code::Internal,
+            GetPageStatusCode::SlowDown => Code::ResourceExhausted,
+        }
+    }
+}
+
 // Fetches the size of a relation at a given LSN, as # of blocks. Only valid on shard 0, other
 // shards will error.
 #[derive(Clone, Copy, Debug)]

pageserver/pagebench/src/cmd/basebackup.rs

@@ -326,7 +326,7 @@ impl GrpcClient {
         ttid: TenantTimelineId,
         compression: bool,
     ) -> anyhow::Result<Self> {
-        let inner = page_api::Client::new(
+        let inner = page_api::Client::connect(
             connstring.to_string(),
             ttid.tenant_id,
             ttid.timeline_id,

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -625,7 +625,7 @@ impl GrpcClient {
         ttid: TenantTimelineId,
         compression: bool,
     ) -> anyhow::Result<Self> {
-        let mut client = page_api::Client::new(
+        let mut client = page_api::Client::connect(
             connstring.to_string(),
             ttid.tenant_id,
             ttid.timeline_id,

pageserver/src/bin/pageserver.rs

@@ -29,8 +29,8 @@ use pageserver::task_mgr::{
 };
 use pageserver::tenant::{TenantSharedResources, mgr, secondary};
 use pageserver::{
-    CancellableTask, ConsumptionMetricsTasks, HttpEndpointListener, HttpsEndpointListener, http,
-    page_cache, page_service, task_mgr, virtual_file,
+    CancellableTask, ConsumptionMetricsTasks, HttpEndpointListener, HttpsEndpointListener,
+    MetricsCollectionTask, http, page_cache, page_service, task_mgr, virtual_file,
 };
 use postgres_backend::AuthType;
 use remote_storage::GenericRemoteStorage;
@@ -41,6 +41,7 @@ use tracing_utils::OtelGuard;
 use utils::auth::{JwtAuth, SwappableJwtAuth};
 use utils::crashsafe::syncfs;
 use utils::logging::TracingErrorLayerEnablement;
+use utils::metrics_collector::{METRICS_COLLECTION_INTERVAL, METRICS_COLLECTOR};
 use utils::sentry_init::init_sentry;
 use utils::{failpoint_support, logging, project_build_tag, project_git_version, tcp_listener};
 
@@ -763,6 +764,41 @@ fn start_pageserver(
         (http_task, https_task)
     };
 
+    /* BEGIN_HADRON */
+    let metrics_collection_task = {
+        let cancel = shutdown_pageserver.child_token();
+        let task = crate::BACKGROUND_RUNTIME.spawn({
+            let cancel = cancel.clone();
+            let background_jobs_barrier = background_jobs_barrier.clone();
+            async move {
+                if conf.force_metric_collection_on_scrape {
+                    return;
+                }
+
+                // first wait until background jobs are cleared to launch.
+                tokio::select! {
+                    _ = cancel.cancelled() => { return; },
+                    _ = background_jobs_barrier.wait() => {}
+                };
+                let mut interval = tokio::time::interval(METRICS_COLLECTION_INTERVAL);
+                loop {
+                    tokio::select! {
+                        _ = cancel.cancelled() => {
+                            tracing::info!("cancelled metrics collection task, exiting...");
+                             break;
+                        },
+                        _ = interval.tick() => {}
+                    }
+                    tokio::task::spawn_blocking(|| {
+                        METRICS_COLLECTOR.run_once(true);
+                    });
+                }
+            }
+        });
+        MetricsCollectionTask(CancellableTask { task, cancel })
+    };
+    /* END_HADRON */
+
     let consumption_metrics_tasks = {
         let cancel = shutdown_pageserver.child_token();
         let task = crate::BACKGROUND_RUNTIME.spawn({
@@ -844,6 +880,7 @@ fn start_pageserver(
             https_endpoint_listener,
             page_service,
             page_service_grpc,
+            metrics_collection_task,
             consumption_metrics_tasks,
             disk_usage_eviction_task,
             &tenant_manager,
@@ -889,8 +926,11 @@ async fn create_remote_storage_client(
             "Simulating remote failures for first {} attempts of each op",
             conf.test_remote_failures
         );
-        remote_storage =
-            GenericRemoteStorage::unreliable_wrapper(remote_storage, conf.test_remote_failures);
+        remote_storage = GenericRemoteStorage::unreliable_wrapper(
+            remote_storage,
+            conf.test_remote_failures,
+            conf.test_remote_failures_probability,
+        );
     }
 
     Ok(remote_storage)

pageserver/src/config.rs

@@ -145,9 +145,13 @@ pub struct PageServerConf {
     pub metric_collection_bucket: Option<RemoteStorageConfig>,
     pub synthetic_size_calculation_interval: Duration,
 
-    pub disk_usage_based_eviction: Option<DiskUsageEvictionTaskConfig>,
+    pub disk_usage_based_eviction: DiskUsageEvictionTaskConfig,
 
+    // The number of allowed failures in remote storage operations.
     pub test_remote_failures: u64,
+    // The probability of failure in remote storage operations. Only works when test_remote_failures > 1.
+    // Use 100 for 100% failure, 0 for no failure.
+    pub test_remote_failures_probability: u64,
 
     pub ondemand_download_behavior_treat_error_as_warn: bool,
 
@@ -248,6 +252,14 @@ pub struct PageServerConf {
     pub timeline_import_config: pageserver_api::config::TimelineImportConfig,
 
     pub basebackup_cache_config: Option<pageserver_api::config::BasebackupCacheConfig>,
+
+    /// Defines what is a big tenant for the purpose of image layer generation.
+    /// See Timeline::should_check_if_image_layers_required
+    pub image_layer_generation_large_timeline_threshold: Option<u64>,
+
+    /// Controls whether to collect all metrics on each scrape or to return potentially stale
+    /// results.
+    pub force_metric_collection_on_scrape: bool,
 }
 
 /// Token for authentication to safekeepers
@@ -392,6 +404,7 @@ impl PageServerConf {
             synthetic_size_calculation_interval,
             disk_usage_based_eviction,
             test_remote_failures,
+            test_remote_failures_probability,
             ondemand_download_behavior_treat_error_as_warn,
             background_task_maximum_delay,
             control_plane_api,
@@ -427,6 +440,8 @@ impl PageServerConf {
             posthog_config,
             timeline_import_config,
             basebackup_cache_config,
+            image_layer_generation_large_timeline_threshold,
+            force_metric_collection_on_scrape,
         } = config_toml;
 
         let mut conf = PageServerConf {
@@ -461,6 +476,7 @@ impl PageServerConf {
             synthetic_size_calculation_interval,
             disk_usage_based_eviction,
             test_remote_failures,
+            test_remote_failures_probability,
             ondemand_download_behavior_treat_error_as_warn,
             background_task_maximum_delay,
             control_plane_api: control_plane_api
@@ -484,6 +500,8 @@ impl PageServerConf {
             dev_mode,
             timeline_import_config,
             basebackup_cache_config,
+            image_layer_generation_large_timeline_threshold,
+            force_metric_collection_on_scrape,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling
@@ -625,7 +643,7 @@ impl PageServerConf {
     pub fn dummy_conf(repo_dir: Utf8PathBuf) -> Self {
         let pg_distrib_dir = Utf8PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../pg_install");
 
-        let config_toml = pageserver_api::config::ConfigToml {
+        let mut config_toml = pageserver_api::config::ConfigToml {
             wait_lsn_timeout: Duration::from_secs(60),
             wal_redo_timeout: Duration::from_secs(60),
             pg_distrib_dir: Some(pg_distrib_dir),
@@ -637,6 +655,15 @@ impl PageServerConf {
             control_plane_api: Some(Url::parse("http://localhost:6666").unwrap()),
             ..Default::default()
         };
+
+        // Test authors tend to forget about the default 10min initial lease deadline
+        // when writing tests, which turns their immediate gc requests via mgmt API
+        // into no-ops. Override the binary default here, such that there is no initial
+        // lease deadline by default in tests. Tests that care can always override it
+        // themselves.
+        // Cf https://databricks.atlassian.net/browse/LKB-92?focusedCommentId=6722329
+        config_toml.tenant_config.lsn_lease_length = Duration::from_secs(0);
+
         PageServerConf::parse_and_validate(NodeId(0), config_toml, &repo_dir).unwrap()
     }
 }
@@ -697,9 +724,12 @@ impl ConfigurableSemaphore {
 #[cfg(test)]
 mod tests {
 
+    use std::time::Duration;
+
     use camino::Utf8PathBuf;
+    use pageserver_api::config::{DiskUsageEvictionTaskConfig, EvictionOrder};
     use rstest::rstest;
-    use utils::id::NodeId;
+    use utils::{id::NodeId, serde_percent::Percent};
 
     use super::PageServerConf;
 
@@ -798,4 +828,70 @@ mod tests {
         PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir)
             .expect("parse_and_validate");
     }
+
+    #[rstest]
+    #[
+        case::omit_the_whole_config(
+            DiskUsageEvictionTaskConfig {
+                max_usage_pct: Percent::new(80).unwrap(),
+                min_avail_bytes: 2_000_000_000,
+                period: Duration::from_secs(60),
+                eviction_order: Default::default(),
+                #[cfg(feature = "testing")]
+                mock_statvfs: None,
+                enabled: true,
+            },
+        r#"
+            control_plane_api = "http://localhost:6666"
+        "#,
+    )]
+    #[
+        case::omit_enabled_field(
+            DiskUsageEvictionTaskConfig {
+                max_usage_pct: Percent::new(80).unwrap(),
+                min_avail_bytes: 1_000_000_000,
+                period: Duration::from_secs(60),
+                eviction_order: EvictionOrder::RelativeAccessed {
+                    highest_layer_count_loses_first: true,
+                },
+                #[cfg(feature = "testing")]
+                mock_statvfs: None,
+                enabled: true,
+            },
+        r#"
+            control_plane_api = "http://localhost:6666"
+            disk_usage_based_eviction = { max_usage_pct = 80, min_avail_bytes = 1000000000, period = "60s" }
+        "#,
+    )]
+    #[case::disabled(
+        DiskUsageEvictionTaskConfig {
+            max_usage_pct: Percent::new(80).unwrap(),
+            min_avail_bytes: 2_000_000_000,
+            period: Duration::from_secs(60),
+            eviction_order: EvictionOrder::RelativeAccessed {
+                highest_layer_count_loses_first: true,
+            },
+            #[cfg(feature = "testing")]
+            mock_statvfs: None,
+            enabled: false,
+        },
+        r#"
+            control_plane_api = "http://localhost:6666"
+            disk_usage_based_eviction = { enabled = false }
+        "#
+    )]
+    fn test_config_disk_usage_based_eviction_is_valid(
+        #[case] expected_disk_usage_based_eviction: DiskUsageEvictionTaskConfig,
+        #[case] input: &str,
+    ) {
+        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;
+        assert_eq!(
+            expected_disk_usage_based_eviction,
+            disk_usage_based_eviction
+        );
+    }
 }

pageserver/src/disk_usage_eviction_task.rs

@@ -171,7 +171,8 @@ pub fn launch_disk_usage_global_eviction_task(
     tenant_manager: Arc<TenantManager>,
     background_jobs_barrier: completion::Barrier,
 ) -> Option<DiskUsageEvictionTask> {
-    let Some(task_config) = &conf.disk_usage_based_eviction else {
+    let task_config = &conf.disk_usage_based_eviction;
+    if !task_config.enabled {
         info!("disk usage based eviction task not configured");
         return None;
     };
@@ -458,6 +459,9 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                 match next {
                     Ok(Ok(file_size)) => {
                         METRICS.layers_evicted.inc();
+                        /*BEGIN_HADRON */
+                        METRICS.bytes_evicted.inc_by(file_size);
+                        /*END_HADRON */
                         usage_assumed.add_available_bytes(file_size);
                     }
                     Ok(Err((
@@ -1265,6 +1269,7 @@ mod filesystem_level_usage {
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
                 eviction_order: pageserver_api::config::EvictionOrder::default(),
+                enabled: true,
             },
             total_bytes: 100_000,
             avail_bytes: 0,

pageserver/src/feature_resolver.rs

@@ -1,4 +1,8 @@
-use std::{collections::HashMap, sync::Arc, time::Duration};
+use std::{
+    collections::HashMap,
+    sync::{Arc, atomic::AtomicBool},
+    time::Duration,
+};
 
 use arc_swap::ArcSwap;
 use pageserver_api::config::NodeMetadata;
@@ -355,11 +359,17 @@ impl PerTenantProperties {
     }
 }
 
-#[derive(Clone)]
 pub struct TenantFeatureResolver {
     inner: FeatureResolver,
     tenant_id: TenantId,
-    cached_tenant_properties: Arc<ArcSwap<HashMap<String, PostHogFlagFilterPropertyValue>>>,
+    cached_tenant_properties: ArcSwap<HashMap<String, PostHogFlagFilterPropertyValue>>,
+
+    // Add feature flag on the critical path below.
+    //
+    // If a feature flag will be used on the critical path, we will update it in the tenant housekeeping loop insetad of
+    // resolving directly by calling `evaluate_multivariate` or `evaluate_boolean`. Remember to update the flag in the
+    // housekeeping loop. The user should directly read this atomic flag instead of using the set of evaluate functions.
+    pub feature_test_remote_size_flag: AtomicBool,
 }
 
 impl TenantFeatureResolver {
@@ -367,7 +377,8 @@ impl TenantFeatureResolver {
         Self {
             inner,
             tenant_id,
-            cached_tenant_properties: Arc::new(ArcSwap::new(Arc::new(HashMap::new()))),
+            cached_tenant_properties: ArcSwap::new(Arc::new(HashMap::new())),
+            feature_test_remote_size_flag: AtomicBool::new(false),
         }
     }
 
@@ -396,12 +407,14 @@ impl TenantFeatureResolver {
         self.inner.is_feature_flag_boolean(flag_key)
     }
 
-    pub fn update_cached_tenant_properties(&self, tenant_shard: &TenantShard) {
-        let mut remote_size_mb = None;
+    /// Refresh the cached properties and flags on the critical path.
+    pub fn refresh_properties_and_flags(&self, tenant_shard: &TenantShard) {
+        let mut remote_size_mb = Some(0.0);
         for timeline in tenant_shard.list_timelines() {
             let size = timeline.metrics.resident_physical_size_get();
             if size == 0 {
                 remote_size_mb = None;
+                break;
             }
             if let Some(ref mut remote_size_mb) = remote_size_mb {
                 *remote_size_mb += size as f64 / 1024.0 / 1024.0;
@@ -410,5 +423,12 @@ impl TenantFeatureResolver {
         self.cached_tenant_properties.store(Arc::new(
             PerTenantProperties { remote_size_mb }.into_posthog_properties(),
         ));
+
+        // BEGIN: Update the feature flag on the critical path.
+        self.feature_test_remote_size_flag.store(
+            self.evaluate_boolean("test-remote-size-flag").is_ok(),
+            std::sync::atomic::Ordering::Relaxed,
+        );
+        // END: Update the feature flag on the critical path.
     }
 }

pageserver/src/http/openapi_spec.yml

@@ -116,26 +116,6 @@ paths:
               schema:
                 type: string
 
-  /v1/tenant/{tenant_id}/timeline:
-    parameters:
-      - name: tenant_id
-        in: path
-        required: true
-        schema:
-          type: string
-    get:
-      description: Get timelines for tenant
-      responses:
-        "200":
-          description: TimelineInfo
-          content:
-            application/json:
-              schema:
-                type: array
-                items:
-                  $ref: "#/components/schemas/TimelineInfo"
-
-
   /v1/tenant/{tenant_id}/timeline/{timeline_id}:
     parameters:
       - name: tenant_id
@@ -618,7 +598,7 @@ paths:
               schema:
                 $ref: "#/components/schemas/SecondaryProgress"
 
-  /v1/tenant/{tenant_id}/timeline/:
+  /v1/tenant/{tenant_id}/timeline:
     parameters:
       - name: tenant_id
         in: path
@@ -685,6 +665,17 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/Error"
+    get:
+      description: Get timelines for tenant
+      responses:
+        "200":
+          description: TimelineInfo
+          content:
+            application/json:
+              schema:
+                type: array
+                items:
+                  $ref: "#/components/schemas/TimelineInfo"
 
   /v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/detach_ancestor:
     parameters:
@@ -767,7 +758,7 @@ paths:
                 $ref: "#/components/schemas/ServiceUnavailableError"
 
 
-  /v1/tenant/:
+  /v1/tenant:
     get:
       description: Get tenants list
       responses:
@@ -847,7 +838,7 @@ paths:
                 items:
                   $ref: "#/components/schemas/TenantInfo"
 
-  /v1/tenant/{tenant_id}/config/:
+  /v1/tenant/{tenant_id}/config:
     parameters:
       - name: tenant_id
         in: path

pageserver/src/http/routes.rs

@@ -61,6 +61,7 @@ use crate::context;
 use crate::context::{DownloadBehavior, RequestContext, RequestContextBuilder};
 use crate::deletion_queue::DeletionQueueClient;
 use crate::feature_resolver::FeatureResolver;
+use crate::metrics::LOCAL_DATA_LOSS_SUSPECTED;
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::LocationConf;
@@ -78,8 +79,8 @@ use crate::tenant::storage_layer::{IoConcurrency, LayerAccessStatsReset, LayerNa
 use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::timeline::offload::{OffloadError, offload_timeline};
 use crate::tenant::timeline::{
-    CompactFlags, CompactOptions, CompactRequest, CompactionError, MarkInvisibleRequest, Timeline,
-    WaitLsnTimeout, WaitLsnWaiter, import_pgdata,
+    CompactFlags, CompactOptions, CompactRequest, MarkInvisibleRequest, Timeline, WaitLsnTimeout,
+    WaitLsnWaiter, import_pgdata,
 };
 use crate::tenant::{
     GetTimelineError, LogicalSizeCalculationCause, OffloadedTimeline, PageReconstructError,
@@ -2499,12 +2500,10 @@ async fn timeline_checkpoint_handler(
                 .compact(&cancel, flags, &ctx)
                 .await
                 .map_err(|e|
-                    match e {
-                        CompactionError::ShuttingDown => ApiError::ShuttingDown,
-                        CompactionError::Offload(e) => ApiError::InternalServerError(anyhow::anyhow!(e)),
-                        CompactionError::CollectKeySpaceError(e) => ApiError::InternalServerError(anyhow::anyhow!(e)),
-                        CompactionError::Other(e) => ApiError::InternalServerError(e),
-                        CompactionError::AlreadyRunning(_) => ApiError::InternalServerError(anyhow::anyhow!(e)),
+                    if e.is_cancel() {
+                        ApiError::ShuttingDown
+                    } else {
+                        ApiError::InternalServerError(e.into_anyhow())
                     }
                 )?;
         }
@@ -3630,6 +3629,17 @@ async fn activate_post_import_handler(
     .await
 }
 
+// [Hadron] Reset gauge metrics that are used to raised alerts. We need this API as a stop-gap measure to reset alerts
+// after we manually rectify situations such as local SSD data loss. We will eventually automate this.
+async fn hadron_reset_alert_gauges(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&request, None)?;
+    LOCAL_DATA_LOSS_SUSPECTED.set(0);
+    json_response(StatusCode::OK, ())
+}
+
 /// Read the end of a tar archive.
 ///
 /// A tar archive normally ends with two consecutive blocks of zeros, 512 bytes each.
@@ -3682,6 +3692,23 @@ async fn read_tar_eof(mut reader: (impl tokio::io::AsyncRead + Unpin)) -> anyhow
     Ok(())
 }
 
+async fn force_refresh_feature_flag(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let state = get_state(&request);
+    let tenant = state
+        .tenant_manager
+        .get_attached_tenant_shard(tenant_shard_id)?;
+    tenant
+        .feature_resolver
+        .refresh_properties_and_flags(&tenant);
+    json_response(StatusCode::OK, ())
+}
+
 async fn tenant_evaluate_feature_flag(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -3698,7 +3725,7 @@ async fn tenant_evaluate_feature_flag(
         let tenant = state
             .tenant_manager
             .get_attached_tenant_shard(tenant_shard_id)?;
-        // TODO: the properties we get here might be stale right after it is collected. But such races are rare (updated every 10s) 
+        // 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") {
@@ -3911,9 +3938,14 @@ pub fn make_router(
         .expect("construct launch timestamp header middleware"),
     );
 
+    let force_metric_collection_on_scrape = state.conf.force_metric_collection_on_scrape;
+
+    let prometheus_metrics_handler_wrapper =
+        move |req| prometheus_metrics_handler(req, force_metric_collection_on_scrape);
+
     Ok(router
         .data(state)
-        .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
+        .get("/metrics", move |r| request_span(r, prometheus_metrics_handler_wrapper))
         .get("/profile/cpu", |r| request_span(r, profile_cpu_handler))
         .get("/profile/heap", |r| request_span(r, profile_heap_handler))
         .get("/v1/status", |r| api_handler(r, status_handler))
@@ -4147,6 +4179,9 @@ pub fn make_router(
         .get("/v1/tenant/:tenant_shard_id/feature_flag/:flag_key", |r| {
             api_handler(r, tenant_evaluate_feature_flag)
         })
+        .post("/v1/tenant/:tenant_shard_id/force_refresh_feature_flag", |r| {
+            api_handler(r, force_refresh_feature_flag)
+        })
         .put("/v1/feature_flag/:flag_key", |r| {
             testing_api_handler("force override feature flag - put", r, force_override_feature_flag_for_testing_put)
         })
@@ -4156,5 +4191,8 @@ pub fn make_router(
         .post("/v1/feature_flag_spec", |r| {
             api_handler(r, update_feature_flag_spec)
         })
+        .post("/hadron-internal/reset_alert_gauges", |r| {
+            api_handler(r, hadron_reset_alert_gauges)
+        })
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -73,6 +73,9 @@ pub struct HttpEndpointListener(pub CancellableTask);
 pub struct HttpsEndpointListener(pub CancellableTask);
 pub struct ConsumptionMetricsTasks(pub CancellableTask);
 pub struct DiskUsageEvictionTask(pub CancellableTask);
+// HADRON
+pub struct MetricsCollectionTask(pub CancellableTask);
+
 impl CancellableTask {
     pub async fn shutdown(self) {
         self.cancel.cancel();
@@ -87,6 +90,7 @@ pub async fn shutdown_pageserver(
     https_listener: Option<HttpsEndpointListener>,
     page_service: page_service::Listener,
     grpc_task: Option<CancellableTask>,
+    metrics_collection_task: MetricsCollectionTask,
     consumption_metrics_worker: ConsumptionMetricsTasks,
     disk_usage_eviction_task: Option<DiskUsageEvictionTask>,
     tenant_manager: &TenantManager,
@@ -211,6 +215,14 @@ pub async fn shutdown_pageserver(
     // Best effort to persist any outstanding deletions, to avoid leaking objects
     deletion_queue.shutdown(Duration::from_secs(5)).await;
 
+    // HADRON
+    timed(
+        metrics_collection_task.0.shutdown(),
+        "shutdown metrics collections metrics",
+        Duration::from_secs(1),
+    )
+    .await;
+
     timed(
         consumption_metrics_worker.0.shutdown(),
         "shutdown consumption metrics",

pageserver/src/metrics.rs

@@ -1,3 +1,4 @@
+use std::cell::Cell;
 use std::collections::HashMap;
 use std::num::NonZeroUsize;
 use std::os::fd::RawFd;
@@ -102,7 +103,18 @@ pub(crate) static STORAGE_TIME_COUNT_PER_TIMELINE: Lazy<IntCounterVec> = Lazy::n
     .expect("failed to define a metric")
 });
 
-// Buckets for background operation duration in seconds, like compaction, GC, size calculation.
+/* BEGIN_HADRON */
+pub(crate) static STORAGE_ACTIVE_COUNT_PER_TIMELINE: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "pageserver_active_storage_operations_count",
+        "Count of active storage operations with operation, tenant and timeline dimensions",
+        &["operation", "tenant_id", "shard_id", "timeline_id"],
+    )
+    .expect("failed to define a metric")
+});
+/*END_HADRON */
+
+// Buckets for background operations like compaction, GC, size calculation
 const STORAGE_OP_BUCKETS: &[f64] = &[0.010, 0.100, 1.0, 10.0, 100.0, 1000.0];
 
 pub(crate) static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
@@ -2810,6 +2822,31 @@ pub(crate) static WALRECEIVER_CANDIDATES_ADDED: Lazy<IntCounter> =
 pub(crate) static WALRECEIVER_CANDIDATES_REMOVED: Lazy<IntCounter> =
     Lazy::new(|| WALRECEIVER_CANDIDATES_EVENTS.with_label_values(&["remove"]));
 
+pub(crate) static LOCAL_DATA_LOSS_SUSPECTED: Lazy<IntGauge> = Lazy::new(|| {
+    register_int_gauge!(
+        "pageserver_local_data_loss_suspected",
+        "Non-zero value indicates that pageserver local data loss is suspected (and highly likely)."
+    )
+    .expect("failed to define a metric")
+});
+
+// Counter keeping track of misrouted PageStream requests. Spelling out PageStream requests here to distinguish
+// it from other types of reqeusts (SK wal replication, http requests, etc.). PageStream requests are used by
+// Postgres compute to fetch data from pageservers.
+// A misrouted PageStream request is registered if the pageserver cannot find the tenant identified in the
+// request, or if the pageserver is not the "primary" serving the tenant shard. These error almost always identify
+// issues with compute configuration, caused by either the compute node itself being stuck in the wrong
+// configuration or Storage Controller reconciliation bugs. Misrouted requests are expected during tenant migration
+// and/or during recovery following a pageserver failure, but persistently high rates of misrouted requests
+// are indicative of bugs (and unavailability).
+pub(crate) static MISROUTED_PAGESTREAM_REQUESTS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_misrouted_pagestream_requests_total",
+        "Number of pageserver pagestream requests that were routed to the wrong pageserver"
+    )
+    .expect("failed to define a metric")
+});
+
 // Metrics collected on WAL redo operations
 //
 // We collect the time spent in actual WAL redo ('redo'), and time waiting
@@ -3048,13 +3085,19 @@ pub(crate) static WAL_REDO_PROCESS_COUNTERS: Lazy<WalRedoProcessCounters> =
 pub(crate) struct StorageTimeMetricsTimer {
     metrics: StorageTimeMetrics,
     start: Instant,
+    stopped: Cell<bool>,
 }
 
 impl StorageTimeMetricsTimer {
     fn new(metrics: StorageTimeMetrics) -> Self {
+        /*BEGIN_HADRON */
+        // record the active operation as the timer starts
+        metrics.timeline_active_count.inc();
+        /*END_HADRON */
         Self {
             metrics,
             start: Instant::now(),
+            stopped: Cell::new(false),
         }
     }
 
@@ -3070,6 +3113,10 @@ impl StorageTimeMetricsTimer {
         self.metrics.timeline_sum.inc_by(seconds);
         self.metrics.timeline_count.inc();
         self.metrics.global_histogram.observe(seconds);
+        /* BEGIN_HADRON*/
+        self.stopped.set(true);
+        self.metrics.timeline_active_count.dec();
+        /*END_HADRON */
         duration
     }
 
@@ -3080,6 +3127,16 @@ impl StorageTimeMetricsTimer {
     }
 }
 
+/*BEGIN_HADRON */
+impl Drop for StorageTimeMetricsTimer {
+    fn drop(&mut self) {
+        if !self.stopped.get() {
+            self.metrics.timeline_active_count.dec();
+        }
+    }
+}
+/*END_HADRON */
+
 pub(crate) struct AlwaysRecordingStorageTimeMetricsTimer(Option<StorageTimeMetricsTimer>);
 
 impl Drop for AlwaysRecordingStorageTimeMetricsTimer {
@@ -3105,6 +3162,10 @@ pub(crate) struct StorageTimeMetrics {
     timeline_sum: Counter,
     /// Number of oeprations, per operation, tenant_id and timeline_id
     timeline_count: IntCounter,
+    /*BEGIN_HADRON */
+    /// Number of active operations per operation, tenant_id, and timeline_id
+    timeline_active_count: IntGauge,
+    /*END_HADRON */
     /// Global histogram having only the "operation" label.
     global_histogram: Histogram,
 }
@@ -3124,6 +3185,11 @@ impl StorageTimeMetrics {
         let timeline_count = STORAGE_TIME_COUNT_PER_TIMELINE
             .get_metric_with_label_values(&[operation, tenant_id, shard_id, timeline_id])
             .unwrap();
+        /*BEGIN_HADRON */
+        let timeline_active_count = STORAGE_ACTIVE_COUNT_PER_TIMELINE
+            .get_metric_with_label_values(&[operation, tenant_id, shard_id, timeline_id])
+            .unwrap();
+        /*END_HADRON */
         let global_histogram = STORAGE_TIME_GLOBAL
             .get_metric_with_label_values(&[operation])
             .unwrap();
@@ -3131,6 +3197,7 @@ impl StorageTimeMetrics {
         StorageTimeMetrics {
             timeline_sum,
             timeline_count,
+            timeline_active_count,
             global_histogram,
         }
     }
@@ -3544,6 +3611,14 @@ impl TimelineMetrics {
                 shard_id,
                 timeline_id,
             ]);
+            /* BEGIN_HADRON */
+            let _ = STORAGE_ACTIVE_COUNT_PER_TIMELINE.remove_label_values(&[
+                op,
+                tenant_id,
+                shard_id,
+                timeline_id,
+            ]);
+            /*END_HADRON */
         }
 
         for op in StorageIoSizeOperation::VARIANTS {
@@ -4336,6 +4411,9 @@ pub(crate) mod disk_usage_based_eviction {
         pub(crate) layers_collected: IntCounter,
         pub(crate) layers_selected: IntCounter,
         pub(crate) layers_evicted: IntCounter,
+        /*BEGIN_HADRON */
+        pub(crate) bytes_evicted: IntCounter,
+        /*END_HADRON */
     }
 
     impl Default for Metrics {
@@ -4372,12 +4450,21 @@ pub(crate) mod disk_usage_based_eviction {
             )
             .unwrap();
 
+            /*BEGIN_HADRON */
+            let bytes_evicted = register_int_counter!(
+                "pageserver_disk_usage_based_eviction_evicted_bytes_total",
+                "Amount of bytes successfully evicted"
+            )
+            .unwrap();
+            /*END_HADRON */
+
             Self {
                 tenant_collection_time,
                 tenant_layer_count,
                 layers_collected,
                 layers_selected,
                 layers_evicted,
+                bytes_evicted,
             }
         }
     }
@@ -4497,6 +4584,7 @@ pub fn preinitialize_metrics(
         &CIRCUIT_BREAKERS_UNBROKEN,
         &PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS_GLOBAL,
         &WAIT_LSN_IN_PROGRESS_GLOBAL_MICROS,
+        &MISROUTED_PAGESTREAM_REQUESTS,
     ]
     .into_iter()
     .for_each(|c| {
@@ -4534,6 +4622,7 @@ pub fn preinitialize_metrics(
 
     // gauges
     WALRECEIVER_ACTIVE_MANAGERS.get();
+    LOCAL_DATA_LOSS_SUSPECTED.get();
 
     // histograms
     [

pageserver/src/page_service.rs

@@ -50,6 +50,7 @@ use tokio::io::{AsyncRead, AsyncReadExt as _, AsyncWrite, AsyncWriteExt as _, Bu
 use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
 use tonic::service::Interceptor as _;
+use tonic::transport::server::TcpConnectInfo;
 use tracing::*;
 use utils::auth::{Claims, Scope, SwappableJwtAuth};
 use utils::id::{TenantId, TenantTimelineId, TimelineId};
@@ -69,7 +70,7 @@ use crate::context::{
 };
 use crate::metrics::{
     self, COMPUTE_COMMANDS_COUNTERS, ComputeCommandKind, GetPageBatchBreakReason, LIVE_CONNECTIONS,
-    SmgrOpTimer, TimelineMetrics,
+    MISROUTED_PAGESTREAM_REQUESTS, SmgrOpTimer, TimelineMetrics,
 };
 use crate::pgdatadir_mapping::{LsnRange, Version};
 use crate::span::{
@@ -90,7 +91,8 @@ use crate::{CancellableTask, PERF_TRACE_TARGET, timed_after_cancellation};
 /// is not yet in state [`TenantState::Active`].
 ///
 /// NB: this is a different value than [`crate::http::routes::ACTIVE_TENANT_TIMEOUT`].
-const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
+/// HADRON: reduced timeout and we will retry in Cache::get().
+const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
 
 /// Threshold at which to log slow GetPage requests.
 const LOG_SLOW_GETPAGE_THRESHOLD: Duration = Duration::from_secs(30);
@@ -1127,6 +1129,7 @@ impl PageServerHandler {
                                 // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
                                 // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
                                 // and talk to a different pageserver.
+                                MISROUTED_PAGESTREAM_REQUESTS.inc();
                                 return respond_error!(
                                     span,
                                     PageStreamError::Reconnect(
@@ -3350,6 +3353,8 @@ impl GrpcPageServiceHandler {
     /// NB: errors returned from here are intercepted in get_pages(), and may be converted to a
     /// GetPageResponse with an appropriate status code to avoid terminating the stream.
     ///
+    /// TODO: verify that the requested pages belong to this shard.
+    ///
     /// TODO: get_vectored() currently enforces a batch limit of 32. Postgres will typically send
     /// batches up to effective_io_concurrency = 100. Either we have to accept large batches, or
     /// split them up in the client or server.
@@ -3685,8 +3690,15 @@ impl proto::PageService for GrpcPageServiceHandler {
                 yield match result {
                     Ok(resp) => resp,
                     // Convert per-request errors to GetPageResponses as appropriate, or terminate
-                    // the stream with a tonic::Status.
-                    Err(err) => page_api::GetPageResponse::try_from_status(err, req_id)?.into(),
+                    // the stream with a tonic::Status. Log the error regardless, since
+                    // ObservabilityLayer can't automatically log stream errors.
+                    Err(status) => {
+                        // TODO: it would be nice if we could propagate the get_page() fields here.
+                        span.in_scope(|| {
+                            warn!("request failed with {:?}: {}", status.code(), status.message());
+                        });
+                        page_api::GetPageResponse::try_from_status(status, req_id)?.into()
+                    }
                 }
             }
         };
@@ -3824,40 +3836,85 @@ impl<S: tonic::server::NamedService> tonic::server::NamedService for Observabili
     const NAME: &'static str = S::NAME; // propagate inner service name
 }
 
-impl<S, B> tower::Service<http::Request<B>> for ObservabilityLayerService<S>
+impl<S, Req, Resp> tower::Service<http::Request<Req>> for ObservabilityLayerService<S>
 where
-    S: tower::Service<http::Request<B>>,
+    S: tower::Service<http::Request<Req>, Response = http::Response<Resp>> + Send,
     S::Future: Send + 'static,
 {
     type Response = S::Response;
     type Error = S::Error;
     type Future = BoxFuture<'static, Result<Self::Response, Self::Error>>;
 
-    fn call(&mut self, mut req: http::Request<B>) -> Self::Future {
+    fn call(&mut self, mut req: http::Request<Req>) -> Self::Future {
         // Record the request start time as a request extension.
         //
         // TODO: we should start a timer here instead, but it currently requires a timeline handle
         // and SmgrQueryType, which we don't have yet. Refactor it to provide it later.
         req.extensions_mut().insert(ReceivedAt(Instant::now()));
 
-        // Create a basic tracing span. Enter the span for the current thread (to use it for inner
-        // sync code like interceptors), and instrument the future (to use it for inner async code
-        // like the page service itself).
+        // Extract the peer address and gRPC method.
+        let peer = req
+            .extensions()
+            .get::<TcpConnectInfo>()
+            .and_then(|info| info.remote_addr())
+            .map(|addr| addr.to_string())
+            .unwrap_or_default();
+
+        let method = req
+            .uri()
+            .path()
+            .split('/')
+            .nth(2)
+            .unwrap_or(req.uri().path())
+            .to_string();
+
+        // Create a basic tracing span.
         //
-        // The instrument() call below is not sufficient. It only affects the returned future, and
-        // only takes effect when the caller polls it. Any sync code executed when we call
-        // self.inner.call() below (such as interceptors) runs outside of the returned future, and
-        // is not affected by it. We therefore have to enter the span on the current thread too.
+        // Enter the span for the current thread and instrument the future. It is not sufficient to
+        // only instrument the future, since it only takes effect after the future is returned and
+        // polled, not when the inner service is called below (e.g. during interceptor execution).
         let span = info_span!(
             "grpc:pageservice",
-            // Set by TenantMetadataInterceptor.
+            // These will be populated by TenantMetadataInterceptor.
             tenant_id = field::Empty,
             timeline_id = field::Empty,
             shard_id = field::Empty,
+            // NB: empty fields must be listed first above. Otherwise, the field names will be
+            // clobbered when the empty fields are populated. They will be output last regardless.
+            %peer,
+            %method,
         );
         let _guard = span.enter();
 
-        Box::pin(self.inner.call(req).instrument(span.clone()))
+        // Construct a future for calling the inner service, but don't await it. This avoids having
+        // to clone the inner service into the future below.
+        let call = self.inner.call(req);
+
+        async move {
+            // Await the inner service call.
+            let result = call.await;
+
+            // Log gRPC error statuses. This won't include request info from handler spans, but it
+            // will catch all errors (even those emitted before handler spans are constructed). Only
+            // unary request errors are logged here, not streaming response errors.
+            if let Ok(ref resp) = result
+                && let Some(status) = tonic::Status::from_header_map(resp.headers())
+                && status.code() != tonic::Code::Ok
+            {
+                // TODO: it would be nice if we could propagate the handler span's request fields
+                // here. This could e.g. be done by attaching the request fields to
+                // tonic::Status::metadata via a proc macro.
+                warn!(
+                    "request failed with {:?}: {}",
+                    status.code(),
+                    status.message()
+                );
+            }
+
+            result
+        }
+        .instrument(span.clone())
+        .boxed()
     }
 
     fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {

pageserver/src/pgdatadir_mapping.rs

@@ -141,6 +141,23 @@ pub(crate) enum CollectKeySpaceError {
     Cancelled,
 }
 
+impl CollectKeySpaceError {
+    pub(crate) fn is_cancel(&self) -> bool {
+        match self {
+            CollectKeySpaceError::Decode(_) => false,
+            CollectKeySpaceError::PageRead(e) => e.is_cancel(),
+            CollectKeySpaceError::Cancelled => true,
+        }
+    }
+    pub(crate) fn into_anyhow(self) -> anyhow::Error {
+        match self {
+            CollectKeySpaceError::Decode(e) => anyhow::Error::new(e),
+            CollectKeySpaceError::PageRead(e) => anyhow::Error::new(e),
+            CollectKeySpaceError::Cancelled => anyhow::Error::new(self),
+        }
+    }
+}
+
 impl From<PageReconstructError> for CollectKeySpaceError {
     fn from(err: PageReconstructError) -> Self {
         match err {

pageserver/src/tenant.rs

@@ -34,7 +34,7 @@ use once_cell::sync::Lazy;
 pub use pageserver_api::models::TenantState;
 use pageserver_api::models::{self, RelSizeMigration};
 use pageserver_api::models::{
-    CompactInfoResponse, LsnLease, TimelineArchivalState, TimelineState, TopTenantShardItem,
+    CompactInfoResponse, TimelineArchivalState, TimelineState, TopTenantShardItem,
     WalRedoManagerStatus,
 };
 use pageserver_api::shard::{ShardIdentity, ShardStripeSize, TenantShardId};
@@ -142,6 +142,9 @@ mod gc_block;
 mod gc_result;
 pub(crate) mod throttle;
 
+#[cfg(test)]
+pub mod debug;
+
 pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
 
 pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
@@ -180,6 +183,7 @@ pub(super) struct AttachedTenantConf {
 
 impl AttachedTenantConf {
     fn new(
+        conf: &'static PageServerConf,
         tenant_conf: pageserver_api::models::TenantConfig,
         location: AttachedLocationConfig,
     ) -> Self {
@@ -191,9 +195,7 @@ impl AttachedTenantConf {
         let lsn_lease_deadline = if location.attach_mode == AttachmentMode::Single {
             Some(
                 tokio::time::Instant::now()
-                    + tenant_conf
-                        .lsn_lease_length
-                        .unwrap_or(LsnLease::DEFAULT_LENGTH),
+                    + TenantShard::get_lsn_lease_length_impl(conf, &tenant_conf),
             )
         } else {
             // We don't use `lsn_lease_deadline` to delay GC in AttachedMulti and AttachedStale
@@ -208,10 +210,13 @@ impl AttachedTenantConf {
         }
     }
 
-    fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
+    fn try_from(
+        conf: &'static PageServerConf,
+        location_conf: LocationConf,
+    ) -> anyhow::Result<Self> {
         match &location_conf.mode {
             LocationMode::Attached(attach_conf) => {
-                Ok(Self::new(location_conf.tenant_conf, *attach_conf))
+                Ok(Self::new(conf, location_conf.tenant_conf, *attach_conf))
             }
             LocationMode::Secondary(_) => {
                 anyhow::bail!(
@@ -386,7 +391,7 @@ pub struct TenantShard {
 
     l0_flush_global_state: L0FlushGlobalState,
 
-    pub(crate) feature_resolver: TenantFeatureResolver,
+    pub(crate) feature_resolver: Arc<TenantFeatureResolver>,
 }
 impl std::fmt::Debug for TenantShard {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -3286,7 +3291,9 @@ impl TenantShard {
                         // Ignore this, we likely raced with unarchival.
                         OffloadError::NotArchived => Ok(()),
                         OffloadError::AlreadyInProgress => Ok(()),
-                        err => Err(err),
+                        OffloadError::Cancelled => Err(CompactionError::new_cancelled()),
+                        // don't break the anyhow chain
+                        OffloadError::Other(err) => Err(CompactionError::Other(err)),
                     })?;
             }
 
@@ -3314,27 +3321,13 @@ impl TenantShard {
 
     /// Trips the compaction circuit breaker if appropriate.
     pub(crate) fn maybe_trip_compaction_breaker(&self, err: &CompactionError) {
-        match err {
-            err if err.is_cancel() => {}
-            CompactionError::ShuttingDown => (),
-            // Offload failures don't trip the circuit breaker, since they're cheap to retry and
-            // shouldn't block compaction.
-            CompactionError::Offload(_) => {}
-            CompactionError::CollectKeySpaceError(err) => {
-                // CollectKeySpaceError::Cancelled and PageRead::Cancelled are handled in `err.is_cancel` branch.
-                self.compaction_circuit_breaker
-                    .lock()
-                    .unwrap()
-                    .fail(&CIRCUIT_BREAKERS_BROKEN, err);
-            }
-            CompactionError::Other(err) => {
-                self.compaction_circuit_breaker
-                    .lock()
-                    .unwrap()
-                    .fail(&CIRCUIT_BREAKERS_BROKEN, err);
-            }
-            CompactionError::AlreadyRunning(_) => {}
+        if err.is_cancel() {
+            return;
         }
+        self.compaction_circuit_breaker
+            .lock()
+            .unwrap()
+            .fail(&CIRCUIT_BREAKERS_BROKEN, err);
     }
 
     /// Cancel scheduled compaction tasks
@@ -3411,7 +3404,7 @@ impl TenantShard {
         }
 
         // Update the feature resolver with the latest tenant-spcific data.
-        self.feature_resolver.update_cached_tenant_properties(self);
+        self.feature_resolver.refresh_properties_and_flags(self);
     }
 
     pub fn timeline_has_no_attached_children(&self, timeline_id: TimelineId) -> bool {
@@ -4178,6 +4171,15 @@ impl TenantShard {
             .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
     }
 
+    // HADRON
+    pub fn get_image_creation_timeout(&self) -> Option<Duration> {
+        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
+        tenant_conf.image_layer_force_creation_period.or(self
+            .conf
+            .default_tenant_conf
+            .image_layer_force_creation_period)
+    }
+
     pub fn get_pitr_interval(&self) -> Duration {
         let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
         tenant_conf
@@ -4205,10 +4207,16 @@ impl TenantShard {
     }
 
     pub fn get_lsn_lease_length(&self) -> Duration {
-        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
+        Self::get_lsn_lease_length_impl(self.conf, &self.tenant_conf.load().tenant_conf)
+    }
+
+    pub fn get_lsn_lease_length_impl(
+        conf: &'static PageServerConf,
+        tenant_conf: &pageserver_api::models::TenantConfig,
+    ) -> Duration {
         tenant_conf
             .lsn_lease_length
-            .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
+            .unwrap_or(conf.default_tenant_conf.lsn_lease_length)
     }
 
     pub fn get_timeline_offloading_enabled(&self) -> bool {
@@ -4494,10 +4502,10 @@ impl TenantShard {
             gc_block: Default::default(),
             l0_flush_global_state,
             basebackup_cache,
-            feature_resolver: TenantFeatureResolver::new(
+            feature_resolver: Arc::new(TenantFeatureResolver::new(
                 feature_resolver,
                 tenant_shard_id.tenant_id,
-            ),
+            )),
         }
     }
 
@@ -6009,22 +6017,24 @@ pub(crate) mod harness {
         }
 
         #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
-        pub(crate) async fn do_try_load(
+        pub(crate) async fn do_try_load_with_redo(
             &self,
+            walredo_mgr: Arc<WalRedoManager>,
             ctx: &RequestContext,
         ) -> anyhow::Result<Arc<TenantShard>> {
-            let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
-
             let (basebackup_cache, _) = BasebackupCache::new(Utf8PathBuf::new(), None);
 
             let tenant = Arc::new(TenantShard::new(
                 TenantState::Attaching,
                 self.conf,
-                AttachedTenantConf::try_from(LocationConf::attached_single(
-                    self.tenant_conf.clone(),
-                    self.generation,
-                    ShardParameters::default(),
-                ))
+                AttachedTenantConf::try_from(
+                    self.conf,
+                    LocationConf::attached_single(
+                        self.tenant_conf.clone(),
+                        self.generation,
+                        ShardParameters::default(),
+                    ),
+                )
                 .unwrap(),
                 self.shard_identity,
                 Some(walredo_mgr),
@@ -6049,6 +6059,14 @@ pub(crate) mod harness {
             Ok(tenant)
         }
 
+        pub(crate) async fn do_try_load(
+            &self,
+            ctx: &RequestContext,
+        ) -> anyhow::Result<Arc<TenantShard>> {
+            let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
+            self.do_try_load_with_redo(walredo_mgr, ctx).await
+        }
+
         pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
             self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
         }
@@ -6125,7 +6143,7 @@ mod tests {
     use pageserver_api::keyspace::KeySpace;
     #[cfg(feature = "testing")]
     use pageserver_api::keyspace::KeySpaceRandomAccum;
-    use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
+    use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings, LsnLease};
     use pageserver_compaction::helpers::overlaps_with;
     #[cfg(feature = "testing")]
     use rand::SeedableRng;
@@ -6675,17 +6693,13 @@ mod tests {
         tline.freeze_and_flush().await.map_err(|e| e.into())
     }
 
-    #[tokio::test(start_paused = true)]
+    #[tokio::test]
     async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
         let (tenant, ctx) =
             TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
                 .await?
                 .load()
                 .await;
-        // Advance to the lsn lease deadline so that GC is not blocked by
-        // initial transition into AttachedSingle.
-        tokio::time::advance(tenant.get_lsn_lease_length()).await;
-        tokio::time::resume();
         let tline = tenant
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
@@ -9384,17 +9398,21 @@ mod tests {
         Ok(())
     }
 
-    #[tokio::test(start_paused = true)]
+    #[tokio::test]
     async fn test_lsn_lease() -> anyhow::Result<()> {
         let (tenant, ctx) = TenantHarness::create("test_lsn_lease")
             .await
             .unwrap()
             .load()
             .await;
-        // Advance to the lsn lease deadline so that GC is not blocked by
-        // initial transition into AttachedSingle.
-        tokio::time::advance(tenant.get_lsn_lease_length()).await;
-        tokio::time::resume();
+        // set a non-zero lease length to test the feature
+        tenant
+            .update_tenant_config(|mut conf| {
+                conf.lsn_lease_length = Some(LsnLease::DEFAULT_LENGTH);
+                Ok(conf)
+            })
+            .unwrap();
+
         let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
 
         let end_lsn = Lsn(0x100);

pageserver/src/tenant/debug.rs

@@ -0,0 +1,366 @@
+use std::{ops::Range, str::FromStr, sync::Arc};
+
+use crate::walredo::RedoAttemptType;
+use base64::{Engine as _, engine::general_purpose::STANDARD};
+use bytes::{Bytes, BytesMut};
+use camino::Utf8PathBuf;
+use clap::Parser;
+use itertools::Itertools;
+use pageserver_api::{
+    key::Key,
+    keyspace::KeySpace,
+    shard::{ShardIdentity, ShardStripeSize},
+};
+use postgres_ffi::PgMajorVersion;
+use postgres_ffi::{BLCKSZ, page_is_new, page_set_lsn};
+use tracing::Instrument;
+use utils::{
+    generation::Generation,
+    id::{TenantId, TimelineId},
+    lsn::Lsn,
+    shard::{ShardCount, ShardIndex, ShardNumber},
+};
+use wal_decoder::models::record::NeonWalRecord;
+
+use crate::{
+    context::{DownloadBehavior, RequestContext},
+    task_mgr::TaskKind,
+    tenant::storage_layer::ValueReconstructState,
+    walredo::harness::RedoHarness,
+};
+
+use super::{
+    WalRedoManager, WalredoManagerId,
+    harness::TenantHarness,
+    remote_timeline_client::LayerFileMetadata,
+    storage_layer::{AsLayerDesc, IoConcurrency, Layer, LayerName, ValuesReconstructState},
+};
+
+fn process_page_image(next_record_lsn: Lsn, is_fpw: bool, img_bytes: Bytes) -> Bytes {
+    // To match the logic in libs/wal_decoder/src/serialized_batch.rs
+    let mut new_image: BytesMut = img_bytes.into();
+    if is_fpw && !page_is_new(&new_image) {
+        page_set_lsn(&mut new_image, next_record_lsn);
+    }
+    assert_eq!(new_image.len(), BLCKSZ as usize);
+    new_image.freeze()
+}
+
+async fn redo_wals(input: &str, key: Key) -> anyhow::Result<()> {
+    let tenant_id = TenantId::generate();
+    let timeline_id = TimelineId::generate();
+    let redo_harness = RedoHarness::new()?;
+    let span = redo_harness.span();
+    let tenant_conf = pageserver_api::models::TenantConfig {
+        ..Default::default()
+    };
+
+    let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
+    let tenant = TenantHarness::create_custom(
+        "search_key",
+        tenant_conf,
+        tenant_id,
+        ShardIdentity::unsharded(),
+        Generation::new(1),
+    )
+    .await?
+    .do_try_load_with_redo(
+        Arc::new(WalRedoManager::Prod(
+            WalredoManagerId::next(),
+            redo_harness.manager,
+        )),
+        &ctx,
+    )
+    .await
+    .unwrap();
+    let timeline = tenant
+        .create_test_timeline(timeline_id, Lsn(0x10), PgMajorVersion::PG16, &ctx)
+        .await?;
+    let contents = tokio::fs::read_to_string(input)
+        .await
+        .map_err(|e| anyhow::Error::msg(format!("Failed to read input file {input}: {e}")))
+        .unwrap();
+    let lines = contents.lines();
+    let mut last_wal_lsn: Option<Lsn> = None;
+    let state = {
+        let mut state = ValueReconstructState::default();
+        let mut is_fpw = false;
+        let mut is_first_line = true;
+        for line in lines {
+            if is_first_line {
+                is_first_line = false;
+                if line.trim() == "FPW" {
+                    is_fpw = true;
+                }
+                continue; // Skip the first line.
+            }
+            // Each input line is in the "<next_record_lsn>,<base64>" format.
+            let (lsn_str, payload_b64) = line
+                .split_once(',')
+                .expect("Invalid input format: expected '<lsn>,<base64>'");
+
+            // Parse the LSN and decode the payload.
+            let lsn = Lsn::from_str(lsn_str.trim()).expect("Invalid LSN format");
+            let bytes = Bytes::from(
+                STANDARD
+                    .decode(payload_b64.trim())
+                    .expect("Invalid base64 payload"),
+            );
+
+            // The first line is considered the base image, the rest are WAL records.
+            if state.img.is_none() {
+                state.img = Some((lsn, process_page_image(lsn, is_fpw, bytes)));
+            } else {
+                let wal_record = NeonWalRecord::Postgres {
+                    will_init: false,
+                    rec: bytes,
+                };
+                state.records.push((lsn, wal_record));
+                last_wal_lsn.replace(lsn);
+            }
+        }
+        state
+    };
+
+    assert!(state.img.is_some(), "No base image found");
+    assert!(!state.records.is_empty(), "No WAL records found");
+    let result = timeline
+        .reconstruct_value(key, last_wal_lsn.unwrap(), state, RedoAttemptType::ReadPage)
+        .instrument(span.clone())
+        .await?;
+
+    eprintln!("final image: {:?}", STANDARD.encode(result));
+
+    Ok(())
+}
+
+async fn search_key(
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+    dir: String,
+    key: Key,
+    lsn: Lsn,
+) -> anyhow::Result<()> {
+    let shard_index = ShardIndex {
+        shard_number: ShardNumber(0),
+        shard_count: ShardCount(4),
+    };
+
+    let redo_harness = RedoHarness::new()?;
+    let span = redo_harness.span();
+    let tenant_conf = pageserver_api::models::TenantConfig {
+        ..Default::default()
+    };
+    let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
+    let tenant = TenantHarness::create_custom(
+        "search_key",
+        tenant_conf,
+        tenant_id,
+        ShardIdentity::new(
+            shard_index.shard_number,
+            shard_index.shard_count,
+            ShardStripeSize(32768),
+        )
+        .unwrap(),
+        Generation::new(1),
+    )
+    .await?
+    .do_try_load_with_redo(
+        Arc::new(WalRedoManager::Prod(
+            WalredoManagerId::next(),
+            redo_harness.manager,
+        )),
+        &ctx,
+    )
+    .await
+    .unwrap();
+
+    let timeline = tenant
+        .create_test_timeline(timeline_id, Lsn(0x10), PgMajorVersion::PG16, &ctx)
+        .await?;
+
+    let mut delta_layers: Vec<Layer> = Vec::new();
+    let mut img_layer: Option<Layer> = Option::None;
+    let mut dir = tokio::fs::read_dir(dir).await?;
+    loop {
+        let entry = dir.next_entry().await?;
+        if entry.is_none() || !entry.as_ref().unwrap().file_type().await?.is_file() {
+            break;
+        }
+        let path = Utf8PathBuf::from_path_buf(entry.unwrap().path()).unwrap();
+        let layer_name = match LayerName::from_str(path.file_name().unwrap()) {
+            Ok(name) => name,
+            Err(_) => {
+                eprintln!("Skipped invalid layer: {path}");
+                continue;
+            }
+        };
+        let layer = Layer::for_resident(
+            tenant.conf,
+            &timeline,
+            path.clone(),
+            layer_name,
+            LayerFileMetadata::new(
+                tokio::fs::metadata(path.clone()).await?.len(),
+                Generation::new(1),
+                shard_index,
+            ),
+        );
+        if layer.layer_desc().is_delta() {
+            delta_layers.push(layer.into());
+        } else if img_layer.is_none() {
+            img_layer = Some(layer.into());
+        } else {
+            anyhow::bail!("Found multiple image layers");
+        }
+    }
+    // sort delta layers based on the descending order of LSN
+    delta_layers.sort_by(|a, b| {
+        b.layer_desc()
+            .get_lsn_range()
+            .start
+            .cmp(&a.layer_desc().get_lsn_range().start)
+    });
+
+    let mut state = ValuesReconstructState::new(IoConcurrency::Sequential);
+
+    let key_space = KeySpace::single(Range {
+        start: key,
+        end: key.next(),
+    });
+    let lsn_range = Range {
+        start: img_layer
+            .as_ref()
+            .map_or(Lsn(0x00), |img| img.layer_desc().image_layer_lsn()),
+        end: lsn,
+    };
+    for delta_layer in delta_layers.iter() {
+        delta_layer
+            .get_values_reconstruct_data(key_space.clone(), lsn_range.clone(), &mut state, &ctx)
+            .await?;
+    }
+
+    img_layer
+        .as_ref()
+        .unwrap()
+        .get_values_reconstruct_data(key_space.clone(), lsn_range.clone(), &mut state, &ctx)
+        .await?;
+
+    for (_key, result) in std::mem::take(&mut state.keys) {
+        let state = result.collect_pending_ios().await?;
+        if state.img.is_some() {
+            eprintln!(
+                "image: {}: {:x?}",
+                state.img.as_ref().unwrap().0,
+                STANDARD.encode(state.img.as_ref().unwrap().1.clone())
+            );
+        }
+        for delta in state.records.iter() {
+            match &delta.1 {
+                NeonWalRecord::Postgres { will_init, rec } => {
+                    eprintln!(
+                        "delta: {}: will_init: {}, {:x?}",
+                        delta.0,
+                        will_init,
+                        STANDARD.encode(rec)
+                    );
+                }
+                _ => {
+                    eprintln!("delta: {}: {:x?}", delta.0, delta.1);
+                }
+            }
+        }
+
+        let result = timeline
+            .reconstruct_value(key, lsn_range.end, state, RedoAttemptType::ReadPage)
+            .instrument(span.clone())
+            .await?;
+        eprintln!("final image: {lsn} : {result:?}");
+    }
+
+    Ok(())
+}
+
+/// Redo all WALs against the base image in the input file. Return the base64 encoded final image.
+/// Each line in the input file must be in the form "<lsn>,<base64>" where:
+///   * `<lsn>` is a PostgreSQL LSN in hexadecimal notation, e.g. `0/16ABCDE`.
+///   * `<base64>` is the base64‐encoded page image (first line) or WAL record (subsequent lines).
+///
+/// The first line provides the base image of a page. The LSN is the LSN of "next record" following
+/// the record containing the FPI. For example, if the FPI was extracted from a WAL record occuping
+/// [0/1, 0/200) in the WAL stream, the LSN appearing along side the page image here should be 0/200.
+///
+/// The subsequent lines are WAL records, ordered from the oldest to the newest. The LSN is the
+/// record LSN of the WAL record, not the "next record" LSN. For example, if the WAL record here
+/// occupies [0/1, 0/200) in the WAL stream, the LSN appearing along side the WAL record here should
+/// be 0/1.
+#[derive(Parser)]
+struct RedoWalsCmd {
+    #[clap(long)]
+    input: String,
+    #[clap(long)]
+    key: String,
+}
+
+#[tokio::test]
+async fn test_redo_wals() -> anyhow::Result<()> {
+    let args = std::env::args().collect_vec();
+    let pos = args
+        .iter()
+        .position(|arg| arg == "--")
+        .unwrap_or(args.len());
+    let slice = &args[pos..args.len()];
+    let cmd = match RedoWalsCmd::try_parse_from(slice) {
+        Ok(cmd) => cmd,
+        Err(err) => {
+            eprintln!("{err}");
+            return Ok(());
+        }
+    };
+
+    let key = Key::from_hex(&cmd.key).unwrap();
+    redo_wals(&cmd.input, key).await?;
+
+    Ok(())
+}
+
+/// Search for a page at the given LSN in all layers of the data_dir.
+/// Return the base64-encoded image and all WAL records, as well as the final reconstructed image.
+#[derive(Parser)]
+struct SearchKeyCmd {
+    #[clap(long)]
+    tenant_id: String,
+    #[clap(long)]
+    timeline_id: String,
+    #[clap(long)]
+    data_dir: String,
+    #[clap(long)]
+    key: String,
+    #[clap(long)]
+    lsn: String,
+}
+
+#[tokio::test]
+async fn test_search_key() -> anyhow::Result<()> {
+    let args = std::env::args().collect_vec();
+    let pos = args
+        .iter()
+        .position(|arg| arg == "--")
+        .unwrap_or(args.len());
+    let slice = &args[pos..args.len()];
+    let cmd = match SearchKeyCmd::try_parse_from(slice) {
+        Ok(cmd) => cmd,
+        Err(err) => {
+            eprintln!("{err}");
+            return Ok(());
+        }
+    };
+
+    let tenant_id = TenantId::from_str(&cmd.tenant_id).unwrap();
+    let timeline_id = TimelineId::from_str(&cmd.timeline_id).unwrap();
+    let key = Key::from_hex(&cmd.key).unwrap();
+    let lsn = Lsn::from_str(&cmd.lsn).unwrap();
+    search_key(tenant_id, timeline_id, cmd.data_dir, key, lsn).await?;
+
+    Ok(())
+}

pageserver/src/tenant/mgr.rs

@@ -43,7 +43,7 @@ use crate::controller_upcall_client::{
 };
 use crate::deletion_queue::DeletionQueueClient;
 use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
-use crate::metrics::{TENANT, TENANT_MANAGER as METRICS};
+use crate::metrics::{LOCAL_DATA_LOSS_SUSPECTED, TENANT, TENANT_MANAGER as METRICS};
 use crate::task_mgr::{BACKGROUND_RUNTIME, TaskKind};
 use crate::tenant::config::{
     AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, SecondaryLocationConfig,
@@ -538,6 +538,21 @@ pub async fn init_tenant_mgr(
     // Determine which tenants are to be secondary or attached, and in which generation
     let tenant_modes = init_load_generations(conf, &tenant_configs, resources, cancel).await?;
 
+    // Hadron local SSD check: Raise an alert if our local filesystem does not contain any tenants but the re-attach request returned tenants.
+    // This can happen if the PS suffered a Kubernetes node failure resulting in loss of all local data, but recovered quickly on another node
+    // so the Storage Controller has not had the time to move tenants out.
+    let data_loss_suspected = if let Some(tenant_modes) = &tenant_modes {
+        tenant_configs.is_empty() && !tenant_modes.is_empty()
+    } else {
+        false
+    };
+    if data_loss_suspected {
+        tracing::error!(
+            "Local data loss suspected: no tenants found on local filesystem, but re-attach request returned tenants"
+        );
+    }
+    LOCAL_DATA_LOSS_SUSPECTED.set(if data_loss_suspected { 1 } else { 0 });
+
     tracing::info!(
         "Attaching {} tenants at startup, warming up {} at a time",
         tenant_configs.len(),
@@ -664,7 +679,7 @@ pub async fn init_tenant_mgr(
                     tenant_shard_id,
                     &tenant_dir_path,
                     resources.clone(),
-                    AttachedTenantConf::new(location_conf.tenant_conf, attached_conf),
+                    AttachedTenantConf::new(conf, location_conf.tenant_conf, attached_conf),
                     shard_identity,
                     Some(init_order.clone()),
                     SpawnMode::Lazy,
@@ -842,8 +857,11 @@ impl TenantManager {
                             // take our fast path and just provide the updated configuration
                             // to the tenant.
                             tenant.set_new_location_config(
-                                AttachedTenantConf::try_from(new_location_config.clone())
-                                    .map_err(UpsertLocationError::BadRequest)?,
+                                AttachedTenantConf::try_from(
+                                    self.conf,
+                                    new_location_config.clone(),
+                                )
+                                .map_err(UpsertLocationError::BadRequest)?,
                             );
 
                             Some(FastPathModified::Attached(tenant.clone()))
@@ -1046,7 +1064,7 @@ impl TenantManager {
                 // Testing hack: if we are configured with no control plane, then drop the generation
                 // from upserts.  This enables creating generation-less tenants even though neon_local
                 // always uses generations when calling the location conf API.
-                let attached_conf = AttachedTenantConf::try_from(new_location_config)
+                let attached_conf = AttachedTenantConf::try_from(self.conf, new_location_config)
                     .map_err(UpsertLocationError::BadRequest)?;
 
                 let tenant = tenant_spawn(
@@ -1250,7 +1268,7 @@ impl TenantManager {
             tenant_shard_id,
             &tenant_path,
             self.resources.clone(),
-            AttachedTenantConf::try_from(config)?,
+            AttachedTenantConf::try_from(self.conf, config)?,
             shard_identity,
             None,
             SpawnMode::Eager,
@@ -2131,7 +2149,7 @@ impl TenantManager {
                 tenant_shard_id,
                 &tenant_path,
                 self.resources.clone(),
-                AttachedTenantConf::try_from(config).map_err(Error::DetachReparent)?,
+                AttachedTenantConf::try_from(self.conf, config).map_err(Error::DetachReparent)?,
                 shard_identity,
                 None,
                 SpawnMode::Eager,

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -141,11 +141,29 @@ pub(super) async fn upload_timeline_layer<'a>(
 
     let fs_size = usize::try_from(fs_size)
         .with_context(|| format!("convert {local_path:?} size {fs_size} usize"))?;
-
+    /* BEGIN_HADRON */
+    let mut metadata = None;
+    match storage {
+        // Pass the file path as a storage metadata to minimize changes to neon.
+        // Otherwise, we need to change the upload interface.
+        GenericRemoteStorage::AzureBlob(s) => {
+            let block_size_mb = s.put_block_size_mb.unwrap_or(0);
+            if block_size_mb > 0 && fs_size > block_size_mb * 1024 * 1024 {
+                metadata = Some(remote_storage::StorageMetadata::from([(
+                    "databricks_azure_put_block",
+                    local_path.as_str(),
+                )]));
+            }
+        }
+        GenericRemoteStorage::LocalFs(_) => {}
+        GenericRemoteStorage::AwsS3(_) => {}
+        GenericRemoteStorage::Unreliable(_) => {}
+    };
+    /* END_HADRON */
     let reader = tokio_util::io::ReaderStream::with_capacity(source_file, super::BUFFER_SIZE);
 
     storage
-        .upload(reader, fs_size, remote_path, None, cancel)
+        .upload(reader, fs_size, remote_path, metadata, cancel)
         .await
         .with_context(|| format!("upload layer from local path '{local_path}'"))
 }

pageserver/src/tenant/tasks.rs

@@ -17,23 +17,35 @@ 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).
 ///
 /// We use 3/4 Tokio threads, to avoid blocking all threads in case we do any CPU-heavy work.
 static CONCURRENT_BACKGROUND_TASKS: Lazy<Semaphore> = Lazy::new(|| {
     let total_threads = TOKIO_WORKER_THREADS.get();
+
+    /*BEGIN_HADRON*/
+    // ideally we should run at least one compaction task per tenant in order to (1) maximize
+    // compaction throughput (2) avoid head-of-line blocking of large compactions. However doing
+    // that may create too many compaction tasks with lots of memory overheads. So we limit the
+    // number of compaction tasks based on the available CPU core count.
+    // Need to revisit.
+    // let tasks_per_thread = std::env::var("BG_TASKS_PER_THREAD")
+    //     .ok()
+    //     .and_then(|s| s.parse().ok())
+    //     .unwrap_or(4);
+    // let permits = usize::max(1, total_threads * tasks_per_thread);
+    // // assert!(permits < total_threads, "need threads for other work");
+    /*END_HADRON*/
+
     let permits = max(1, (total_threads * 3).checked_div(4).unwrap_or(0));
     assert_ne!(permits, 0, "we will not be adding in permits later");
     assert!(permits < total_threads, "need threads for other work");
@@ -295,48 +307,12 @@ pub(crate) fn log_compaction_error(
     task_cancelled: bool,
     degrade_to_warning: bool,
 ) {
-    use CompactionError::*;
+    let is_cancel = err.is_cancel();
 
-    use crate::tenant::PageReconstructError;
-    use crate::tenant::upload_queue::NotInitialized;
-
-    let level = match err {
-        e if e.is_cancel() => return,
-        ShuttingDown => return,
-        Offload(_) => Level::ERROR,
-        AlreadyRunning(_) => Level::ERROR,
-        CollectKeySpaceError(_) => Level::ERROR,
-        _ if task_cancelled => Level::INFO,
-        Other(err) => {
-            let root_cause = err.root_cause();
-
-            let upload_queue = root_cause
-                .downcast_ref::<NotInitialized>()
-                .is_some_and(|e| e.is_stopping());
-            let timeline = root_cause
-                .downcast_ref::<PageReconstructError>()
-                .is_some_and(|e| e.is_stopping());
-            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
-            } else {
-                Level::ERROR
-            }
-        }
+    let level = if is_cancel || task_cancelled {
+        Level::INFO
+    } else {
+        Level::ERROR
     };
 
     if let Some((error_count, sleep_duration)) = retry_info {

pageserver/src/tenant/timeline.rs

@@ -40,7 +40,6 @@ use layer_manager::{
     Shutdown,
 };
 
-use offload::OffloadError;
 use once_cell::sync::Lazy;
 use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
 use pageserver_api::key::{
@@ -119,7 +118,6 @@ use crate::pgdatadir_mapping::{
     MAX_AUX_FILE_V2_DELTAS, MetricsUpdate,
 };
 use crate::task_mgr::TaskKind;
-use crate::tenant::config::AttachmentMode;
 use crate::tenant::gc_result::GcResult;
 use crate::tenant::layer_map::LayerMap;
 use crate::tenant::metadata::TimelineMetadata;
@@ -202,7 +200,7 @@ pub struct TimelineResources {
     pub l0_compaction_trigger: Arc<Notify>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
     pub basebackup_cache: Arc<BasebackupCache>,
-    pub feature_resolver: TenantFeatureResolver,
+    pub feature_resolver: Arc<TenantFeatureResolver>,
 }
 
 pub struct Timeline {
@@ -353,6 +351,13 @@ pub struct Timeline {
     last_image_layer_creation_check_at: AtomicLsn,
     last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
 
+    // HADRON
+    /// If a key range has writes with LSN > force_image_creation_lsn, then we should force image layer creation
+    /// on this key range.
+    force_image_creation_lsn: AtomicLsn,
+    /// The last time instant when force_image_creation_lsn is computed.
+    force_image_creation_lsn_computed_at: std::sync::Mutex<Option<Instant>>,
+
     /// Current logical size of the "datadir", at the last LSN.
     current_logical_size: LogicalSize,
 
@@ -450,7 +455,7 @@ pub struct Timeline {
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
     basebackup_cache: Arc<BasebackupCache>,
 
-    feature_resolver: TenantFeatureResolver,
+    feature_resolver: Arc<TenantFeatureResolver>,
 }
 
 pub(crate) enum PreviousHeatmap {
@@ -587,6 +592,28 @@ pub(crate) enum PageReconstructError {
     MissingKey(Box<MissingKeyError>),
 }
 
+impl PageReconstructError {
+    pub(crate) fn is_cancel(&self) -> bool {
+        match self {
+            PageReconstructError::Other(_) => false,
+            PageReconstructError::AncestorLsnTimeout(e) => e.is_cancel(),
+            PageReconstructError::Cancelled => true,
+            PageReconstructError::WalRedo(_) => false,
+            PageReconstructError::MissingKey(_) => false,
+        }
+    }
+    #[allow(dead_code)] // we use the is_cancel + into_anyhow pattern in quite a few places, this one will follow soon enough
+    pub(crate) fn into_anyhow(self) -> anyhow::Error {
+        match self {
+            PageReconstructError::Other(e) => e,
+            PageReconstructError::AncestorLsnTimeout(e) => e.into_anyhow(),
+            PageReconstructError::Cancelled => anyhow::Error::new(self),
+            PageReconstructError::WalRedo(e) => e,
+            PageReconstructError::MissingKey(_) => anyhow::Error::new(self),
+        }
+    }
+}
+
 impl From<anyhow::Error> for PageReconstructError {
     fn from(value: anyhow::Error) -> Self {
         // with walingest.rs many PageReconstructError are wrapped in as anyhow::Error
@@ -740,17 +767,6 @@ impl std::fmt::Display for MissingKeyError {
     }
 }
 
-impl PageReconstructError {
-    /// Returns true if this error indicates a tenant/timeline shutdown alike situation
-    pub(crate) fn is_stopping(&self) -> bool {
-        use PageReconstructError::*;
-        match self {
-            Cancelled => true,
-            Other(_) | AncestorLsnTimeout(_) | WalRedo(_) | MissingKey(_) => false,
-        }
-    }
-}
-
 #[derive(thiserror::Error, Debug)]
 pub(crate) enum CreateImageLayersError {
     #[error("timeline shutting down")]
@@ -953,13 +969,35 @@ pub enum WaitLsnError {
     Timeout(String),
 }
 
+impl WaitLsnError {
+    pub(crate) fn is_cancel(&self) -> bool {
+        match self {
+            WaitLsnError::Shutdown => true,
+            WaitLsnError::BadState(timeline_state) => match timeline_state {
+                TimelineState::Loading => false,
+                TimelineState::Active => false,
+                TimelineState::Stopping => true,
+                TimelineState::Broken { .. } => false,
+            },
+            WaitLsnError::Timeout(_) => false,
+        }
+    }
+    pub(crate) fn into_anyhow(self) -> anyhow::Error {
+        match self {
+            WaitLsnError::Shutdown => anyhow::Error::new(self),
+            WaitLsnError::BadState(_) => anyhow::Error::new(self),
+            WaitLsnError::Timeout(_) => anyhow::Error::new(self),
+        }
+    }
+}
+
 impl From<WaitLsnError> for tonic::Status {
     fn from(err: WaitLsnError) -> Self {
         use tonic::Code;
-        let code = match &err {
-            WaitLsnError::Timeout(_) => Code::Internal,
-            WaitLsnError::BadState(_) => Code::Internal,
-            WaitLsnError::Shutdown => Code::Unavailable,
+        let code = if err.is_cancel() {
+            Code::Unavailable
+        } else {
+            Code::Internal
         };
         tonic::Status::new(code, err.to_string())
     }
@@ -971,7 +1009,7 @@ impl From<WaitLsnError> for tonic::Status {
 impl From<CreateImageLayersError> for CompactionError {
     fn from(e: CreateImageLayersError) -> Self {
         match e {
-            CreateImageLayersError::Cancelled => CompactionError::ShuttingDown,
+            CreateImageLayersError::Cancelled => CompactionError::new_cancelled(),
             CreateImageLayersError::Other(e) => {
                 CompactionError::Other(e.context("create image layers"))
             }
@@ -1086,6 +1124,26 @@ enum ImageLayerCreationOutcome {
     Skip,
 }
 
+enum RepartitionError {
+    Other(anyhow::Error),
+    CollectKeyspace(CollectKeySpaceError),
+}
+
+impl RepartitionError {
+    fn is_cancel(&self) -> bool {
+        match self {
+            RepartitionError::Other(_) => false,
+            RepartitionError::CollectKeyspace(e) => e.is_cancel(),
+        }
+    }
+    fn into_anyhow(self) -> anyhow::Error {
+        match self {
+            RepartitionError::Other(e) => e,
+            RepartitionError::CollectKeyspace(e) => e.into_anyhow(),
+        }
+    }
+}
+
 /// Public interface functions
 impl Timeline {
     /// Get the LSN where this branch was created
@@ -1772,30 +1830,31 @@ impl Timeline {
                     existing_lease.clone()
                 }
                 Entry::Vacant(vacant) => {
-                    // Reject already GC-ed LSN if we are in AttachedSingle and
-                    // not blocked by the lsn lease deadline.
+                    // Never allow a lease to be requested for an LSN below the applied GC cutoff. The data could have been deleted.
+                    let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
+                    if lsn < *latest_gc_cutoff_lsn {
+                        bail!(
+                            "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
+                            lsn,
+                            *latest_gc_cutoff_lsn
+                        );
+                    }
+
+                    // We allow create lease for those below the planned gc cutoff if we are still within the grace period
+                    // of GC blocking.
                     let validate = {
                         let conf = self.tenant_conf.load();
-                        conf.location.attach_mode == AttachmentMode::Single
-                            && !conf.is_gc_blocked_by_lsn_lease_deadline()
+                        !conf.is_gc_blocked_by_lsn_lease_deadline()
                     };
 
-                    if init || validate {
-                        let latest_gc_cutoff_lsn = self.get_applied_gc_cutoff_lsn();
-                        if lsn < *latest_gc_cutoff_lsn {
-                            bail!(
-                                "tried to request an lsn lease for an lsn below the latest gc cutoff. requested at {} gc cutoff {}",
-                                lsn,
-                                *latest_gc_cutoff_lsn
-                            );
-                        }
-                        if lsn < planned_cutoff {
-                            bail!(
-                                "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
-                                lsn,
-                                planned_cutoff
-                            );
-                        }
+                    // Do not allow initial lease creation to be below the planned gc cutoff. The client (compute_ctl) determines
+                    // whether it is a initial lease creation or a renewal.
+                    if (init || validate) && lsn < planned_cutoff {
+                        bail!(
+                            "tried to request an lsn lease for an lsn below the planned gc cutoff. requested at {} planned gc cutoff {}",
+                            lsn,
+                            planned_cutoff
+                        );
                     }
 
                     let dt: DateTime<Utc> = valid_until.into();
@@ -2065,22 +2124,7 @@ impl Timeline {
         match &result {
             Ok(_) => self.compaction_failed.store(false, AtomicOrdering::Relaxed),
             Err(e) if e.is_cancel() => {}
-            Err(CompactionError::ShuttingDown) => {
-                // Covered by the `Err(e) if e.is_cancel()` branch.
-            }
-            Err(CompactionError::AlreadyRunning(_)) => {
-                // Covered by the `Err(e) if e.is_cancel()` branch.
-            }
-            Err(CompactionError::Other(_)) => {
-                self.compaction_failed.store(true, AtomicOrdering::Relaxed)
-            }
-            Err(CompactionError::CollectKeySpaceError(_)) => {
-                // Cancelled errors are covered by the `Err(e) if e.is_cancel()` branch.
-                self.compaction_failed.store(true, AtomicOrdering::Relaxed)
-            }
-            // Don't change the current value on offload failure or shutdown. We don't want to
-            // abruptly stall nor resume L0 flushes in these cases.
-            Err(CompactionError::Offload(_)) => {}
+            Err(_) => self.compaction_failed.store(true, AtomicOrdering::Relaxed),
         };
 
         result
@@ -2144,14 +2188,31 @@ 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();
@@ -2792,6 +2853,18 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
     }
 
+    // HADRON
+    fn get_image_creation_timeout(&self) -> Option<Duration> {
+        let tenant_conf = self.tenant_conf.load();
+        tenant_conf
+            .tenant_conf
+            .image_layer_force_creation_period
+            .or(self
+                .conf
+                .default_tenant_conf
+                .image_layer_force_creation_period)
+    }
+
     fn get_compaction_algorithm_settings(&self) -> CompactionAlgorithmSettings {
         let tenant_conf = &self.tenant_conf.load();
         tenant_conf
@@ -3061,7 +3134,9 @@ impl Timeline {
                 repartition_threshold: 0,
                 last_image_layer_creation_check_at: AtomicLsn::new(0),
                 last_image_layer_creation_check_instant: Mutex::new(None),
-
+                // HADRON
+                force_image_creation_lsn: AtomicLsn::new(0),
+                force_image_creation_lsn_computed_at: std::sync::Mutex::new(None),
                 last_received_wal: Mutex::new(None),
                 rel_size_latest_cache: RwLock::new(HashMap::new()),
                 rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
@@ -3112,7 +3187,7 @@ impl Timeline {
 
                 basebackup_cache: resources.basebackup_cache,
 
-                feature_resolver: resources.feature_resolver,
+                feature_resolver: resources.feature_resolver.clone(),
             };
 
             result.repartition_threshold =
@@ -4953,7 +5028,7 @@ impl Timeline {
                     ctx,
                 )
                 .await
-                .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
+                .map_err(|e| FlushLayerError::from_anyhow(self, e.into_anyhow()))?;
 
             if self.cancel.is_cancelled() {
                 return Err(FlushLayerError::Cancelled);
@@ -4982,6 +5057,7 @@ impl Timeline {
                 .create_image_layers(
                     &partitions,
                     self.initdb_lsn,
+                    None,
                     ImageLayerCreationMode::Initial,
                     ctx,
                     LastImageLayerCreationStatus::Initial,
@@ -5203,18 +5279,18 @@ impl Timeline {
         partition_size: u64,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
-    ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
+    ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), RepartitionError> {
         let Ok(mut guard) = self.partitioning.try_write_guard() else {
             // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
             // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
             // and hence before the compaction task starts.
-            return Err(CompactionError::Other(anyhow!(
+            return Err(RepartitionError::Other(anyhow!(
                 "repartition() called concurrently"
             )));
         };
         let ((dense_partition, sparse_partition), partition_lsn) = &*guard.read();
         if lsn < *partition_lsn {
-            return Err(CompactionError::Other(anyhow!(
+            return Err(RepartitionError::Other(anyhow!(
                 "repartition() called with LSN going backwards, this should not happen"
             )));
         }
@@ -5235,7 +5311,10 @@ impl Timeline {
             ));
         }
 
-        let (dense_ks, sparse_ks) = self.collect_keyspace(lsn, ctx).await?;
+        let (dense_ks, sparse_ks) = self
+            .collect_keyspace(lsn, ctx)
+            .await
+            .map_err(RepartitionError::CollectKeyspace)?;
         let dense_partitioning = dense_ks.partition(
             &self.shard_identity,
             partition_size,
@@ -5250,14 +5329,19 @@ impl Timeline {
     }
 
     // Is it time to create a new image layer for the given partition? True if we want to generate.
-    async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
+    async fn time_for_new_image_layer(
+        &self,
+        partition: &KeySpace,
+        lsn: Lsn,
+        force_image_creation_lsn: Option<Lsn>,
+    ) -> bool {
         let threshold = self.get_image_creation_threshold();
 
         let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
         let Ok(layers) = guard.layer_map() else {
             return false;
         };
-
+        let mut min_image_lsn: Lsn = Lsn::MAX;
         let mut max_deltas = 0;
         for part_range in &partition.ranges {
             let image_coverage = layers.image_coverage(part_range, lsn);
@@ -5292,9 +5376,22 @@ impl Timeline {
                         return true;
                     }
                 }
+                min_image_lsn = min(min_image_lsn, img_lsn);
             }
         }
 
+        // HADRON
+        if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
+            info!(
+                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}",
+                partition.ranges[0].start,
+                partition.ranges[0].end,
+                min_image_lsn,
+                force_image_creation_lsn.unwrap()
+            );
+            return true;
+        }
+
         debug!(
             max_deltas,
             "none of the partitioned ranges had >= {threshold} deltas"
@@ -5520,7 +5617,7 @@ impl Timeline {
     ///        suffer from the lack of image layers
     ///     2. For small tenants (that can mostly fit in RAM), we use a much longer interval
     fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
-        const LARGE_TENANT_THRESHOLD: u64 = 2 * 1024 * 1024 * 1024;
+        let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
 
         let last_checks_at = self.last_image_layer_creation_check_at.load();
         let distance = lsn
@@ -5534,12 +5631,12 @@ impl Timeline {
         let mut time_based_decision = false;
         let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
         if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
-            let check_required_after = if Into::<u64>::into(&logical_size) >= LARGE_TENANT_THRESHOLD
-            {
-                self.get_checkpoint_timeout()
-            } else {
-                Duration::from_secs(3600 * 48)
-            };
+            let check_required_after =
+                if Some(Into::<u64>::into(&logical_size)) >= large_timeline_threshold {
+                    self.get_checkpoint_timeout()
+                } else {
+                    Duration::from_secs(3600 * 48)
+                };
 
             time_based_decision = match *last_check_instant {
                 Some(last_check) => {
@@ -5567,10 +5664,12 @@ impl Timeline {
     /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
     ///
     /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
+    #[allow(clippy::too_many_arguments)]
     async fn create_image_layers(
         self: &Arc<Timeline>,
         partitioning: &KeyPartitioning,
         lsn: Lsn,
+        force_image_creation_lsn: Option<Lsn>,
         mode: ImageLayerCreationMode,
         ctx: &RequestContext,
         last_status: LastImageLayerCreationStatus,
@@ -5674,7 +5773,11 @@ impl Timeline {
             } else if let ImageLayerCreationMode::Try = mode {
                 // check_for_image_layers = false -> skip
                 // check_for_image_layers = true -> check time_for_new_image_layer -> skip/generate
-                if !check_for_image_layers || !self.time_for_new_image_layer(partition, lsn).await {
+                if !check_for_image_layers
+                    || !self
+                        .time_for_new_image_layer(partition, lsn, force_image_creation_lsn)
+                        .await
+                {
                     start = img_range.end;
                     continue;
                 }
@@ -5995,57 +6098,88 @@ impl Drop for Timeline {
     }
 }
 
-/// Top-level failure to compact.
-#[derive(Debug, thiserror::Error)]
-pub(crate) enum CompactionError {
-    #[error("The timeline or pageserver is shutting down")]
-    ShuttingDown,
-    /// Compaction tried to offload a timeline and failed
-    #[error("Failed to offload timeline: {0}")]
-    Offload(OffloadError),
-    /// Compaction cannot be done right now; page reconstruction and so on.
-    #[error("Failed to collect keyspace: {0}")]
-    CollectKeySpaceError(#[from] CollectKeySpaceError),
-    #[error(transparent)]
-    Other(anyhow::Error),
-    #[error("Compaction already running: {0}")]
-    AlreadyRunning(&'static str),
-}
+pub(crate) use compaction_error::CompactionError;
+/// In a private mod to enforce that [`CompactionError::is_cancel`] is used
+/// instead of `match`ing on [`CompactionError::ShuttingDown`].
+mod compaction_error {
+    use utils::sync::gate::GateError;
 
-impl CompactionError {
-    /// Errors that can be ignored, i.e., cancel and shutdown.
-    pub fn is_cancel(&self) -> bool {
-        matches!(
-            self,
-            Self::ShuttingDown
-                | Self::AlreadyRunning(_)
-                | Self::CollectKeySpaceError(CollectKeySpaceError::Cancelled)
-                | Self::CollectKeySpaceError(CollectKeySpaceError::PageRead(
-                    PageReconstructError::Cancelled
-                ))
-                | Self::Offload(OffloadError::Cancelled)
-        )
+    use crate::{
+        pgdatadir_mapping::CollectKeySpaceError,
+        tenant::{PageReconstructError, blob_io::WriteBlobError, upload_queue::NotInitialized},
+        virtual_file::owned_buffers_io::write::FlushTaskError,
+    };
+
+    /// Top-level failure to compact. Use [`Self::is_cancel`].
+    #[derive(Debug, thiserror::Error)]
+    pub(crate) enum CompactionError {
+        /// Use [`Self::is_cancel`] instead of checking for this variant.
+        #[error("The timeline or pageserver is shutting down")]
+        #[allow(private_interfaces)]
+        ShuttingDown(ForbidMatching), // private ForbidMatching enforces use of [`Self::is_cancel`].
+        #[error(transparent)]
+        Other(anyhow::Error),
     }
 
-    /// Critical errors that indicate data corruption.
-    pub fn is_critical(&self) -> bool {
-        matches!(
-            self,
-            Self::CollectKeySpaceError(
-                CollectKeySpaceError::Decode(_)
-                    | CollectKeySpaceError::PageRead(
-                        PageReconstructError::MissingKey(_) | PageReconstructError::WalRedo(_),
-                    )
-            )
-        )
-    }
-}
+    #[derive(Debug)]
+    struct ForbidMatching;
 
-impl From<OffloadError> for CompactionError {
-    fn from(e: OffloadError) -> Self {
-        match e {
-            OffloadError::Cancelled => Self::ShuttingDown,
-            _ => Self::Offload(e),
+    impl CompactionError {
+        pub fn new_cancelled() -> Self {
+            Self::ShuttingDown(ForbidMatching)
+        }
+        /// Errors that can be ignored, i.e., cancel and shutdown.
+        pub fn is_cancel(&self) -> bool {
+            let other = match self {
+                CompactionError::ShuttingDown(_) => return true,
+                CompactionError::Other(other) => other,
+            };
+
+            // The write path of compaction in particular often lacks differentiated
+            // handling errors stemming from cancellation from other errors.
+            // So, if requested, we also check the ::Other variant by downcasting.
+            // The list below has been found empirically from flaky tests and production logs.
+            // The process is simple: on ::Other(), compaction will print the enclosed
+            // anyhow::Error in debug mode, i.e., with backtrace. That backtrace contains the
+            // line where the write path / compaction code does undifferentiated error handling
+            // from a non-anyhow type to an anyhow type. Add the type to the list of downcasts
+            // below, following the same is_cancel() pattern.
+
+            let root_cause = other.root_cause();
+
+            let upload_queue = root_cause
+                .downcast_ref::<NotInitialized>()
+                .is_some_and(|e| e.is_stopping());
+            let timeline = root_cause
+                .downcast_ref::<PageReconstructError>()
+                .is_some_and(|e| e.is_cancel());
+            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());
+            upload_queue
+                || timeline
+                || buffered_writer_flush_task_canelled
+                || write_blob_cancelled
+                || gate_closed
+        }
+        pub fn into_anyhow(self) -> anyhow::Error {
+            match self {
+                CompactionError::ShuttingDown(ForbidMatching) => anyhow::Error::new(self),
+                CompactionError::Other(e) => e,
+            }
+        }
+        pub fn from_collect_keyspace(err: CollectKeySpaceError) -> Self {
+            if err.is_cancel() {
+                Self::new_cancelled()
+            } else {
+                Self::Other(err.into_anyhow())
+            }
         }
     }
 }
@@ -6057,7 +6191,7 @@ impl From<super::upload_queue::NotInitialized> for CompactionError {
                 CompactionError::Other(anyhow::anyhow!(value))
             }
             super::upload_queue::NotInitialized::ShuttingDown
-            | super::upload_queue::NotInitialized::Stopped => CompactionError::ShuttingDown,
+            | super::upload_queue::NotInitialized::Stopped => CompactionError::new_cancelled(),
         }
     }
 }
@@ -6067,7 +6201,7 @@ impl From<super::storage_layer::layer::DownloadError> for CompactionError {
         match e {
             super::storage_layer::layer::DownloadError::TimelineShutdown
             | super::storage_layer::layer::DownloadError::DownloadCancelled => {
-                CompactionError::ShuttingDown
+                CompactionError::new_cancelled()
             }
             super::storage_layer::layer::DownloadError::ContextAndConfigReallyDeniesDownloads
             | super::storage_layer::layer::DownloadError::DownloadRequired
@@ -6086,14 +6220,14 @@ impl From<super::storage_layer::layer::DownloadError> for CompactionError {
 
 impl From<layer_manager::Shutdown> for CompactionError {
     fn from(_: layer_manager::Shutdown) -> Self {
-        CompactionError::ShuttingDown
+        CompactionError::new_cancelled()
     }
 }
 
 impl From<super::storage_layer::errors::PutError> for CompactionError {
     fn from(e: super::storage_layer::errors::PutError) -> Self {
         if e.is_cancel() {
-            CompactionError::ShuttingDown
+            CompactionError::new_cancelled()
         } else {
             CompactionError::Other(e.into_anyhow())
         }
@@ -6192,7 +6326,7 @@ impl Timeline {
         let mut guard = tokio::select! {
             guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
             _ = self.cancel.cancelled() => {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
         };
 
@@ -6748,7 +6882,7 @@ impl Timeline {
     }
 
     /// Reconstruct a value, using the given base image and WAL records in 'data'.
-    async fn reconstruct_value(
+    pub(crate) async fn reconstruct_value(
         &self,
         key: Key,
         request_lsn: Lsn,

pageserver/src/tenant/timeline/compaction.rs

@@ -4,10 +4,11 @@
 //!
 //! The old legacy algorithm is implemented directly in `timeline.rs`.
 
+use std::cmp::min;
 use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
-use std::time::{Duration, Instant};
+use std::time::{Duration, Instant, SystemTime};
 
 use super::layer_manager::LayerManagerLockHolder;
 use super::{
@@ -16,7 +17,8 @@ use super::{
     Timeline,
 };
 
-use crate::tenant::timeline::DeltaEntry;
+use crate::pgdatadir_mapping::CollectKeySpaceError;
+use crate::tenant::timeline::{DeltaEntry, RepartitionError};
 use crate::walredo::RedoAttemptType;
 use anyhow::{Context, anyhow};
 use bytes::Bytes;
@@ -32,6 +34,7 @@ use pageserver_api::models::{CompactInfoResponse, CompactKeyRange};
 use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
 use pageserver_compaction::helpers::{fully_contains, overlaps_with};
 use pageserver_compaction::interface::*;
+use postgres_ffi::to_pg_timestamp;
 use serde::Serialize;
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
 use tokio_util::sync::CancellationToken;
@@ -44,6 +47,7 @@ use wal_decoder::models::value::Value;
 
 use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::page_cache;
+use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::statvfs::Statvfs;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::gc_block::GcBlock;
@@ -64,7 +68,7 @@ use crate::tenant::timeline::{
     DeltaLayerWriter, ImageLayerCreationOutcome, ImageLayerWriter, IoConcurrency, Layer,
     ResidentLayer, drop_layer_manager_rlock,
 };
-use crate::tenant::{DeltaLayer, MaybeOffloaded};
+use crate::tenant::{DeltaLayer, MaybeOffloaded, PageReconstructError};
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
 
 /// Maximum number of deltas before generating an image layer in bottom-most compaction.
@@ -571,7 +575,7 @@ impl GcCompactionQueue {
         }
         match res {
             Ok(res) => Ok(res),
-            Err(CompactionError::ShuttingDown) => Err(CompactionError::ShuttingDown),
+            Err(e) if e.is_cancel() => Err(e),
             Err(_) => {
                 // There are some cases where traditional gc might collect some layer
                 // files causing gc-compaction cannot read the full history of the key.
@@ -591,9 +595,9 @@ impl GcCompactionQueue {
         timeline: &Arc<Timeline>,
     ) -> Result<CompactionOutcome, CompactionError> {
         let Ok(_one_op_at_a_time_guard) = self.consumer_lock.try_lock() else {
-            return Err(CompactionError::AlreadyRunning(
-                "cannot run gc-compaction because another gc-compaction is running. This should not happen because we only call this function from the gc-compaction queue.",
-            ));
+            return Err(CompactionError::Other(anyhow::anyhow!(
+                "cannot run gc-compaction because another gc-compaction is running. This should not happen because we only call this function from the gc-compaction queue."
+            )));
         };
         let has_pending_tasks;
         let mut yield_for_l0 = false;
@@ -1259,13 +1263,19 @@ impl Timeline {
         // Is the timeline being deleted?
         if self.is_stopping() {
             trace!("Dropping out of compaction on timeline shutdown");
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
         let target_file_size = self.get_checkpoint_distance();
 
         // Define partitioning schema if needed
 
+        // HADRON
+        let force_image_creation_lsn = self
+            .get_or_compute_force_image_creation_lsn(cancel, ctx)
+            .await
+            .map_err(CompactionError::Other)?;
+
         // 1. L0 Compact
         let l0_outcome = {
             let timer = self.metrics.compact_time_histo.start_timer();
@@ -1273,6 +1283,7 @@ impl Timeline {
                 .compact_level0(
                     target_file_size,
                     options.flags.contains(CompactFlags::ForceL0Compaction),
+                    force_image_creation_lsn,
                     ctx,
                 )
                 .await?;
@@ -1375,6 +1386,7 @@ impl Timeline {
                     .create_image_layers(
                         &partitioning,
                         lsn,
+                        force_image_creation_lsn,
                         mode,
                         &image_ctx,
                         self.last_image_layer_creation_status
@@ -1417,22 +1429,33 @@ impl Timeline {
             }
 
             // Suppress errors when cancelled.
-            Err(_) if self.cancel.is_cancelled() => {}
+            //
+            // Log other errors but continue. Failure to repartition is normal, if the timeline was just created
+            // as an empty timeline. Also in unit tests, when we use the timeline as a simple
+            // key-value store, ignoring the datadir layout. Log the error but continue.
+            //
+            // TODO:
+            // 1. shouldn't we return early here if we observe cancellation
+            // 2. Experiment: can we stop checking self.cancel here?
+            Err(_) if self.cancel.is_cancelled() => {} // TODO: try how we fare removing this branch
             Err(err) if err.is_cancel() => {}
-
-            // Alert on critical errors that indicate data corruption.
-            Err(err) if err.is_critical() => {
+            Err(RepartitionError::CollectKeyspace(
+                e @ CollectKeySpaceError::Decode(_)
+                | e @ CollectKeySpaceError::PageRead(
+                    PageReconstructError::MissingKey(_) | PageReconstructError::WalRedo(_),
+                ),
+            )) => {
+                // Alert on critical errors that indicate data corruption.
                 critical_timeline!(
                     self.tenant_shard_id,
                     self.timeline_id,
-                    "could not compact, repartitioning keyspace failed: {err:?}"
+                    "could not compact, repartitioning keyspace failed: {e:?}"
                 );
             }
-
-            // Log other errors. No partitioning? This is normal, if the timeline was just created
-            // as an empty timeline. Also in unit tests, when we use the timeline as a simple
-            // key-value store, ignoring the datadir layout. Log the error but continue.
-            Err(err) => error!("could not compact, repartitioning keyspace failed: {err:?}"),
+            Err(e) => error!(
+                "could not compact, repartitioning keyspace failed: {:?}",
+                e.into_anyhow()
+            ),
         };
 
         let partition_count = self.partitioning.read().0.0.parts.len();
@@ -1460,6 +1483,63 @@ impl Timeline {
         Ok(CompactionOutcome::Done)
     }
 
+    /* BEGIN_HADRON */
+    // Get the force image creation LSN. Compute it if the last computed LSN is too old.
+    async fn get_or_compute_force_image_creation_lsn(
+        self: &Arc<Self>,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Option<Lsn>> {
+        const FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL: Duration = Duration::from_secs(10 * 60); // 10 minutes
+        let image_layer_force_creation_period = self.get_image_creation_timeout();
+        if image_layer_force_creation_period.is_none() {
+            return Ok(None);
+        }
+
+        let image_layer_force_creation_period = image_layer_force_creation_period.unwrap();
+        let force_image_creation_lsn_computed_at =
+            *self.force_image_creation_lsn_computed_at.lock().unwrap();
+        if force_image_creation_lsn_computed_at.is_none()
+            || force_image_creation_lsn_computed_at.unwrap().elapsed()
+                > FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL
+        {
+            let now: SystemTime = SystemTime::now();
+            let timestamp = now
+                .checked_sub(image_layer_force_creation_period)
+                .ok_or_else(|| {
+                    anyhow::anyhow!(
+                        "image creation timeout is too large: {image_layer_force_creation_period:?}"
+                    )
+                })?;
+            let timestamp = to_pg_timestamp(timestamp);
+            let force_image_creation_lsn = match self
+                .find_lsn_for_timestamp(timestamp, cancel, ctx)
+                .await?
+            {
+                LsnForTimestamp::Present(lsn) | LsnForTimestamp::Future(lsn) => lsn,
+                _ => {
+                    let gc_lsn = *self.get_applied_gc_cutoff_lsn();
+                    tracing::info!(
+                        "no LSN found for timestamp {timestamp:?}, using latest GC cutoff LSN {}",
+                        gc_lsn
+                    );
+                    gc_lsn
+                }
+            };
+            self.force_image_creation_lsn
+                .store(force_image_creation_lsn);
+            *self.force_image_creation_lsn_computed_at.lock().unwrap() = Some(Instant::now());
+            tracing::info!(
+                "computed force image creation LSN: {}",
+                force_image_creation_lsn
+            );
+            Ok(Some(force_image_creation_lsn))
+        } else {
+            Ok(Some(self.force_image_creation_lsn.load()))
+        }
+    }
+    /* END_HADRON */
+
     /// Check for layers that are elegible to be rewritten:
     /// - Shard splitting: After a shard split, ancestor layers beyond pitr_interval, so that
     ///   we don't indefinitely retain keys in this shard that aren't needed.
@@ -1612,7 +1692,7 @@ impl Timeline {
 
         for (i, layer) in layers_to_rewrite.into_iter().enumerate() {
             if self.cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             info!(layer=%layer, "rewriting layer after shard split: {}/{}", i, total);
@@ -1710,7 +1790,7 @@ impl Timeline {
                     Ok(()) => {},
                     Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
                     Err(WaitCompletionError::UploadQueueShutDownOrStopped) => {
-                        return Err(CompactionError::ShuttingDown);
+                        return Err(CompactionError::new_cancelled());
                     }
                 },
                 // Don't wait if there's L0 compaction to do. We don't need to update the outcome
@@ -1789,6 +1869,7 @@ impl Timeline {
         self: &Arc<Self>,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
+        force_compaction_lsn: Option<Lsn>,
         ctx: &RequestContext,
     ) -> Result<CompactionOutcome, CompactionError> {
         let CompactLevel0Phase1Result {
@@ -1809,6 +1890,7 @@ impl Timeline {
                 stats,
                 target_file_size,
                 force_compaction_ignore_threshold,
+                force_compaction_lsn,
                 &ctx,
             )
             .instrument(phase1_span)
@@ -1831,6 +1913,7 @@ impl Timeline {
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
+        force_compaction_lsn: Option<Lsn>,
         ctx: &RequestContext,
     ) -> Result<CompactLevel0Phase1Result, CompactionError> {
         let begin = tokio::time::Instant::now();
@@ -1860,11 +1943,28 @@ impl Timeline {
                     return Ok(CompactLevel0Phase1Result::default());
                 }
             } else {
-                debug!(
-                    level0_deltas = level0_deltas.len(),
-                    threshold, "too few deltas to compact"
-                );
-                return Ok(CompactLevel0Phase1Result::default());
+                // HADRON
+                let min_lsn = level0_deltas
+                    .iter()
+                    .map(|a| a.get_lsn_range().start)
+                    .reduce(min);
+                if force_compaction_lsn.is_some()
+                    && min_lsn.is_some()
+                    && min_lsn.unwrap() < force_compaction_lsn.unwrap()
+                {
+                    info!(
+                        "forcing L0 compaction of {} L0 deltas. Min lsn: {}, force compaction lsn: {}",
+                        level0_deltas.len(),
+                        min_lsn.unwrap(),
+                        force_compaction_lsn.unwrap()
+                    );
+                } else {
+                    debug!(
+                        level0_deltas = level0_deltas.len(),
+                        threshold, "too few deltas to compact"
+                    );
+                    return Ok(CompactLevel0Phase1Result::default());
+                }
             }
         }
 
@@ -1973,7 +2073,7 @@ impl Timeline {
             let mut all_keys = Vec::new();
             for l in deltas_to_compact.iter() {
                 if self.cancel.is_cancelled() {
-                    return Err(CompactionError::ShuttingDown);
+                    return Err(CompactionError::new_cancelled());
                 }
                 let delta = l.get_as_delta(ctx).await.map_err(CompactionError::Other)?;
                 let keys = delta
@@ -2066,7 +2166,7 @@ impl Timeline {
         stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
 
         if self.cancel.is_cancelled() {
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
         stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
@@ -2174,7 +2274,7 @@ impl Timeline {
                 // avoid hitting the cancellation token on every key. in benches, we end up
                 // shuffling an order of million keys per layer, this means we'll check it
                 // around tens of times per layer.
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             let same_key = prev_key == Some(key);
@@ -2259,7 +2359,7 @@ impl Timeline {
                 if writer.is_none() {
                     if self.cancel.is_cancelled() {
                         // to be somewhat responsive to cancellation, check for each new layer
-                        return Err(CompactionError::ShuttingDown);
+                        return Err(CompactionError::new_cancelled());
                     }
                     // Create writer if not initiaized yet
                     writer = Some(
@@ -2515,10 +2615,13 @@ impl Timeline {
         // Is the timeline being deleted?
         if self.is_stopping() {
             trace!("Dropping out of compaction on timeline shutdown");
-            return Err(CompactionError::ShuttingDown);
+            return Err(CompactionError::new_cancelled());
         }
 
-        let (dense_ks, _sparse_ks) = self.collect_keyspace(end_lsn, ctx).await?;
+        let (dense_ks, _sparse_ks) = self
+            .collect_keyspace(end_lsn, ctx)
+            .await
+            .map_err(CompactionError::from_collect_keyspace)?;
         // TODO(chi): ignore sparse_keyspace for now, compact it in the future.
         let mut adaptor = TimelineAdaptor::new(self, (end_lsn, dense_ks));
 
@@ -3174,7 +3277,7 @@ impl Timeline {
         let gc_lock = async {
             tokio::select! {
                 guard = self.gc_lock.lock() => Ok(guard),
-                _ = cancel.cancelled() => Err(CompactionError::ShuttingDown),
+                _ = cancel.cancelled() => Err(CompactionError::new_cancelled()),
             }
         };
 
@@ -3447,7 +3550,7 @@ impl Timeline {
             }
             total_layer_size += layer.layer_desc().file_size;
             if cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
             let should_yield = yield_for_l0
                 && self
@@ -3594,7 +3697,7 @@ impl Timeline {
             }
 
             if cancel.is_cancelled() {
-                return Err(CompactionError::ShuttingDown);
+                return Err(CompactionError::new_cancelled());
             }
 
             let should_yield = yield_for_l0

pageserver/src/tenant/timeline/handle.rs

@@ -212,8 +212,12 @@
 //! to the parent shard during a shard split. Eventually, the shard split task will
 //! shut down the parent => case (1).
 
-use std::collections::{HashMap, hash_map};
-use std::sync::{Arc, Mutex, Weak};
+use std::collections::HashMap;
+use std::collections::hash_map;
+use std::sync::Arc;
+use std::sync::Mutex;
+use std::sync::Weak;
+use std::time::Duration;
 
 use pageserver_api::shard::ShardIdentity;
 use tracing::{instrument, trace};
@@ -333,6 +337,44 @@ enum RoutingResult<T: Types> {
 }
 
 impl<T: Types> Cache<T> {
+    /* BEGIN_HADRON */
+    /// A wrapper of do_get to resolve the tenant shard for a get page request.
+    #[instrument(level = "trace", skip_all)]
+    pub(crate) async fn get(
+        &mut self,
+        timeline_id: TimelineId,
+        shard_selector: ShardSelector,
+        tenant_manager: &T::TenantManager,
+    ) -> Result<Handle<T>, GetError<T>> {
+        const GET_MAX_RETRIES: usize = 10;
+        const RETRY_BACKOFF: Duration = Duration::from_millis(100);
+        let mut attempt = 0;
+        loop {
+            attempt += 1;
+            match self
+                .do_get(timeline_id, shard_selector, tenant_manager)
+                .await
+            {
+                Ok(handle) => return Ok(handle),
+                Err(e) => {
+                    // Retry on tenant manager error to handle tenant split more gracefully
+                    if attempt < GET_MAX_RETRIES {
+                        tracing::warn!(
+                            "Fail to resolve tenant shard in attempt {}: {:?}. Retrying...",
+                            attempt,
+                            e
+                        );
+                        tokio::time::sleep(RETRY_BACKOFF).await;
+                        continue;
+                    } else {
+                        return Err(e);
+                    }
+                }
+            }
+        }
+    }
+    /* END_HADRON */
+
     /// See module-level comment for details.
     ///
     /// Does NOT check for the shutdown state of [`Types::Timeline`].
@@ -341,7 +383,7 @@ impl<T: Types> Cache<T> {
     /// and if so, return an error that causes the page service to
     /// close the connection.
     #[instrument(level = "trace", skip_all)]
-    pub(crate) async fn get(
+    async fn do_get(
         &mut self,
         timeline_id: TimelineId,
         shard_selector: ShardSelector,
@@ -879,6 +921,7 @@ mod tests {
             .await
             .err()
             .expect("documented behavior: can't get new handle after shutdown");
+
         assert_eq!(cache.map.len(), 1, "next access cleans up the cache");
 
         cache

pageserver/src/tenant/timeline/offload.rs

@@ -17,8 +17,6 @@ pub(crate) enum OffloadError {
     Cancelled,
     #[error("Timeline is not archived")]
     NotArchived,
-    #[error(transparent)]
-    RemoteStorage(anyhow::Error),
     #[error("Offload or deletion already in progress")]
     AlreadyInProgress,
     #[error("Unexpected offload error: {0}")]
@@ -29,7 +27,7 @@ impl From<TenantManifestError> for OffloadError {
     fn from(e: TenantManifestError) -> Self {
         match e {
             TenantManifestError::Cancelled => Self::Cancelled,
-            TenantManifestError::RemoteStorage(e) => Self::RemoteStorage(e),
+            TenantManifestError::RemoteStorage(e) => Self::Other(e),
         }
     }
 }

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> {

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

@@ -100,6 +100,7 @@ pub(super) async fn connection_manager_loop_step(
     // with other streams on this client (other connection managers). When
     // object goes out of scope, stream finishes in drop() automatically.
     let mut broker_subscription = subscribe_for_timeline_updates(broker_client, id, cancel).await?;
+    let mut broker_reset_interval = tokio::time::interval(tokio::time::Duration::from_secs(30));
     debug!("Subscribed for broker timeline updates");
 
     loop {
@@ -156,7 +157,10 @@ pub(super) async fn connection_manager_loop_step(
             // Got a new update from the broker
             broker_update = broker_subscription.message() /* TODO: review cancellation-safety */ => {
                 match broker_update {
-                    Ok(Some(broker_update)) => connection_manager_state.register_timeline_update(broker_update),
+                    Ok(Some(broker_update)) => {
+                        broker_reset_interval.reset();
+                        connection_manager_state.register_timeline_update(broker_update);
+                    },
                     Err(status) => {
                         match status.code() {
                             Code::Unknown if status.message().contains("stream closed because of a broken pipe") || status.message().contains("connection reset") || status.message().contains("error reading a body from connection") => {
@@ -178,6 +182,21 @@ pub(super) async fn connection_manager_loop_step(
                 }
             },
 
+            // If we've not received any updates from the broker from a while, are waiting for WAL
+            // and have no safekeeper connection or connection candidates, then it might be that
+            // the broker subscription is wedged. Drop the currrent subscription and re-subscribe
+            // with the goal of unblocking it.
+            _ = broker_reset_interval.tick() => {
+                let awaiting_lsn = wait_lsn_status.borrow().is_some();
+                let no_candidates = connection_manager_state.wal_stream_candidates.is_empty();
+                let no_connection = connection_manager_state.wal_connection.is_none();
+
+                if awaiting_lsn && no_candidates && no_connection {
+                    tracing::warn!("No broker updates received for a while, but waiting for WAL. Re-setting stream ...");
+                    broker_subscription = subscribe_for_timeline_updates(broker_client, id, cancel).await?;
+                }
+            },
+
             new_event = async {
                 // Reminder: this match arm needs to be cancellation-safe.
                 loop {