Add a test for generating pprof from collapsed

This provides the binaries and libraries required for the continuous
profiling at runtime.

.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
@@ -25,11 +26,6 @@ compaction-suite-results.*
 *.o
 *.so
 *.Po
-*.pid
 
 # pgindent typedef lists
 *.list
-
-# various files for local testing
-/proxy/.subzero
-local_proxy.json

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]
@@ -127,6 +130,7 @@ jemalloc_pprof = { version = "0.7", features = ["symbolize", "flamegraph"] }
 jsonwebtoken = "9"
 lasso = "0.7"
 libc = "0.2"
+libproc = "0.14"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
@@ -255,9 +259,11 @@ 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" }
+pageserver_client_grpc = { path = "./pageserver/client_grpc" }
 pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
 pageserver_page_api = { path = "./pageserver/page_api" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
@@ -273,6 +279,7 @@ safekeeper_api = { version = "0.1", path = "./libs/safekeeper_api" }
 safekeeper_client = { path = "./safekeeper/client" }
 storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main broker code is inside the binary crate, so linking with the library shouldn't be heavy.
 storage_controller_client = { path = "./storage_controller/client" }
+tempfile = "3"
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
@@ -284,6 +291,7 @@ 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
 
@@ -101,6 +109,8 @@ RUN set -e \
         libreadline-dev \
         libseccomp-dev \
         ca-certificates \
+        bpfcc-tools \
+        sudo \
         openssl \
         unzip \
         curl \
@@ -130,12 +140,15 @@ 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

build-tools.Dockerfile

@@ -61,6 +61,9 @@ RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         libpq5 \
         libpq-dev \
         libzstd-dev \
+        linux-perf \
+        bpfcc-tools \
+        linux-headers-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac) \
         postgresql-16 \
         postgresql-server-dev-16 \
         postgresql-common  \
@@ -105,15 +108,21 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
 #
 # 'gdb' is included so that we get backtraces of core dumps produced in
 # regression tests
-RUN set -e \
+RUN set -ex \
+    && KERNEL_VERSION="$(uname -r | cut -d'-' -f1 | sed 's/\.0$//')" \
+    && echo KERNEL_VERSION=${KERNEL_VERSION} >> /etc/environment \
+    && KERNEL_ARCH=$(uname -m | awk '{ if ($1 ~ /^(x86_64|i[3-6]86)$/) print "x86"; else if ($1 ~ /^(aarch64|arm.*)$/) print "aarch"; else print $1 }') \
+    && echo KERNEL_ARCH=${KERNEL_ARCH} >> /etc/environment \
     && apt update \
     && apt install -y \
         autoconf \
         automake \
+        bc \
         bison \
         build-essential \
         ca-certificates \
         cmake \
+        cpio \
         curl \
         flex \
         gdb \
@@ -122,8 +131,10 @@ RUN set -e \
         gzip \
         jq \
         jsonnet \
+        kmod \
         libcurl4-openssl-dev \
         libbz2-dev \
+        libelf-dev \
         libffi-dev \
         liblzma-dev \
         libncurses5-dev \
@@ -137,6 +148,11 @@ RUN set -e \
         libxml2-dev \
         libxmlsec1-dev \
         libxxhash-dev \
+        linux-perf \
+        bpfcc-tools \
+        libbpfcc \
+        libbpfcc-dev \
+        linux-headers-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac) \
         lsof \
         make \
         netcat-openbsd \
@@ -144,6 +160,8 @@ RUN set -e \
         openssh-client \
         parallel \
         pkg-config \
+        rsync \
+        sudo \
         unzip \
         wget \
         xz-utils \
@@ -198,6 +216,8 @@ RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /
 # Configure sudo & docker
 RUN usermod -aG sudo nonroot && \
     echo '%sudo ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers && \
+    mkdir -p /etc/sudoers.d && \
+    echo 'nonroot ALL=(ALL) NOPASSWD:ALL' > /etc/sudoers.d/nonroot && \
     usermod -aG docker nonroot
 
 # AWS CLI

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

@@ -149,6 +149,9 @@ RUN case $DEBIAN_VERSION in \
     ninja-build git autoconf automake libtool build-essential bison flex libreadline-dev \
     zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget ca-certificates pkg-config libssl-dev \
     libicu-dev libxslt1-dev liblz4-dev libzstd-dev zstd curl unzip g++ \
+    bpfcc-tools \
+    libbpfcc \
+    libbpfcc-dev \
     libclang-dev \
     jsonnet \
     $VERSION_INSTALLS \
@@ -1636,11 +1639,14 @@ 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 +1918,10 @@ 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
@@ -1985,6 +1991,10 @@ RUN apt update && \
         locales \
         lsof \
         procps \
+        bpfcc-tools \
+        libbpfcc \
+        libbpfcc-dev \
+        libclang-dev \
         rsyslog-gnutls \
         screen \
         tcpdump \

compute/vm-image-spec-bookworm.yaml

@@ -39,6 +39,14 @@ commands:
     user: nobody
     sysvInitAction: respawn
     shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
+  - name: enable-kernel-modules
+    user: root
+    sysvInitAction: sysinit
+    shell: mkdir -p /lib/ && ln -s /neonvm/tools/lib/modules /lib/
+  - name: enable-bpfs
+    user: root
+    sysvInitAction: sysinit
+    shell: mkdir -p /sys/kernel/debug && mount -t debugfs debugfs /sys/kernel/debug && mount -t bpf bpf /sys/fs/bpf && chmod 755 /sys/fs/bpf
   # Rsyslog by default creates a unix socket under /dev/log . That's where Postgres sends logs also.
   # We run syslog with postgres user so it can't create /dev/log. Instead we configure rsyslog to
   # use a different path for the socket. The symlink actually points to our custom path.
@@ -65,7 +73,7 @@ files:
       # regardless of hostname (ALL)
       #
       # Also allow it to shut down the VM. The fast_import job does that when it's finished.
-      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -152,6 +160,8 @@ merge: |
   RUN set -e \
       && chmod 0644 /etc/cgconfig.conf
 
+  ENV PERF_BINARY_PATH=/neonvm/tools/bin/perf
+
 
   COPY compute_rsyslog.conf /etc/compute_rsyslog.conf
   RUN chmod 0666 /etc/compute_rsyslog.conf

compute/vm-image-spec-bullseye.yaml

@@ -39,6 +39,14 @@ commands:
     user: nobody
     sysvInitAction: respawn
     shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
+  - name: enable-kernel-modules
+    user: root
+    sysvInitAction: sysinit
+    shell: mkdir -p /lib/ && ln -s /neonvm/tools/lib/modules /lib/
+  - name: enable-bpfs
+    user: root
+    sysvInitAction: sysinit
+    shell: mkdir -p /sys/kernel/debug && mount -t debugfs debugfs /sys/kernel/debug && mount -t bpf bpf /sys/fs/bpf && chmod 755 /sys/fs/bpf
   # Rsyslog by default creates a unix socket under /dev/log . That's where Postgres sends logs also.
   # We run syslog with postgres user so it can't create /dev/log. Instead we configure rsyslog to
   # use a different path for the socket. The symlink actually points to our custom path.
@@ -65,7 +73,7 @@ files:
       # regardless of hostname (ALL)
       #
       # Also allow it to shut down the VM. The fast_import job does that when it's finished.
-      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -148,6 +156,8 @@ merge: |
   RUN set -e \
       && chmod 0644 /etc/cgconfig.conf
 
+  ENV PERF_BINARY_PATH=/neonvm/tools/bin/perf
+
   COPY compute_rsyslog.conf /etc/compute_rsyslog.conf
   RUN chmod 0666 /etc/compute_rsyslog.conf
   RUN mkdir /var/log/rsyslog && chown -R postgres /var/log/rsyslog

compute_tools/Cargo.toml

@@ -31,6 +31,7 @@ hostname-validator = "1.1"
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
+libproc.workspace = true
 metrics.workspace = true
 nix.workspace = true
 notify.workspace = true
@@ -49,6 +50,7 @@ serde_with.workspace = true
 serde_json.workspace = true
 signal-hook.workspace = true
 tar.workspace = true
+tempfile.workspace = true
 tower.workspace = true
 tower-http.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
@@ -66,7 +68,7 @@ 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
@@ -78,3 +80,10 @@ zstd = "0.13"
 bytes = "1.0"
 rust-ini = "0.20.0"
 rlimit = "0.10.1"
+
+inferno = { version = "0.12", default-features = false, features = [
+    "multithreaded",
+    "nameattr",
+] }
+pprof = { version = "0.15", features = ["protobuf-codec", "flamegraph"] }
+prost = "0.12"

compute_tools/README.md

@@ -46,11 +46,14 @@ stateDiagram-v2
   Configuration --> Failed : Failed to configure the compute
   Configuration --> Running : Compute has been configured
   Empty --> Init : Compute spec is immediately available
-  Empty --> TerminationPending : Requested termination
+  Empty --> TerminationPendingFast : Requested termination
+  Empty --> TerminationPendingImmediate : Requested termination
   Init --> Failed : Failed to start Postgres
   Init --> Running : Started Postgres
-  Running --> TerminationPending : Requested termination
-  TerminationPending --> Terminated : Terminated compute
+  Running --> TerminationPendingFast : Requested termination
+  Running --> TerminationPendingImmediate : Requested termination
+  TerminationPendingFast --> Terminated compute with 30s delay for cplane to inspect status
+  TerminationPendingImmediate --> Terminated : Terminated compute immediately
   Failed --> [*] : Compute exited
   Terminated --> [*] : Compute exited
 ```

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,
         }
     }
 
@@ -366,9 +371,11 @@ fn maybe_cgexec(cmd: &str) -> Command {
     }
 }
 
-struct PostgresHandle {
+/// A handle to the Postgres process that is running in the compute
+/// node.
+pub struct PostgresHandle {
     postgres: std::process::Child,
-    log_collector: tokio::task::JoinHandle<Result<()>>,
+    log_collector: JoinHandle<Result<()>>,
 }
 
 impl PostgresHandle {
@@ -382,7 +389,7 @@ 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 +440,7 @@ 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 +528,8 @@ 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 +859,15 @@ 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(())
     }
 
@@ -947,14 +958,20 @@ impl ComputeNode {
             None
         };
 
-        let mut delay_exit = false;
         let mut state = self.state.lock().unwrap();
         state.terminate_flush_lsn = lsn;
-        if let ComputeStatus::TerminationPending { mode } = state.status {
+
+        let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
+        if state.status == ComputeStatus::TerminationPendingFast
+            || state.status == ComputeStatus::TerminationPendingImmediate
+        {
+            info!(
+                "Changing compute status from {} to {}",
+                state.status,
+                ComputeStatus::Terminated
+            );
             state.status = ComputeStatus::Terminated;
             self.state_changed.notify_all();
-            // we were asked to terminate gracefully, don't exit to avoid restart
-            delay_exit = mode == compute_api::responses::TerminateMode::Fast
         }
         drop(state);
 
@@ -1049,7 +1066,7 @@ 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,
@@ -1796,6 +1813,8 @@ impl ComputeNode {
             tls_config,
         )?;
 
+        self.pg_reload_conf()?;
+
         if !spec.skip_pg_catalog_updates {
             let max_concurrent_connections = spec.reconfigure_concurrency;
             // Temporarily reset max_cluster_size in config
@@ -1815,10 +1834,9 @@ impl ComputeNode {
 
                 Ok(())
             })?;
+            self.pg_reload_conf()?;
         }
 
-        self.pg_reload_conf()?;
-
         let unknown_op = "unknown".to_string();
         let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
         info!(
@@ -1891,7 +1909,8 @@ impl ComputeNode {
 
                             // exit loop
                             ComputeStatus::Failed
-                            | ComputeStatus::TerminationPending { .. }
+                            | ComputeStatus::TerminationPendingFast
+                            | ComputeStatus::TerminationPendingImmediate
                             | ComputeStatus::Terminated => break 'cert_update,
 
                             // wait
@@ -2357,10 +2376,7 @@ 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 +2412,30 @@ 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 +2444,11 @@ 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};
@@ -69,7 +70,7 @@ impl ComputeNode {
             }
         };
         let row = match client
-            .query_one("select * from get_prewarm_info()", &[])
+            .query_one("select * from neon.get_prewarm_info()", &[])
             .await
         {
             Ok(row) => row,
@@ -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(),
             };
@@ -146,38 +146,48 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select prewarm_local_cache($1)", &[&uncompressed])
+            .query_one("select neon.prewarm_local_cache($1)", &[&uncompressed])
             .await
             .context("loading LFC state into postgres")
             .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");
@@ -186,7 +196,7 @@ impl ComputeNode {
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
-            .query_one("select get_local_cache_state()", &[])
+            .query_one("select neon.get_local_cache_state()", &[])
             .await
             .context("querying LFC state")?
             .try_get::<usize, &[u8]>(0)

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:
@@ -290,9 +371,28 @@ paths:
       summary: Terminate Postgres and wait for it to exit
       description: ""
       operationId: terminate
+      parameters:
+        - name: mode
+          in: query
+          description: "Terminate mode: fast (wait 30s before returning) and immediate"
+          required: false
+          schema:
+            type: string
+            enum: ["fast", "immediate"]
+            default: fast
       responses:
         200:
           description: Result
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
+        201:
+          description: Result if compute is already terminated
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/TerminateResponse"
         412:
           description: "wrong state"
           content:
@@ -335,15 +435,6 @@ 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
@@ -458,11 +549,14 @@ components:
       type: string
       enum:
         - empty
-        - init
-        - failed
-        - running
         - configuration_pending
+        - init
+        - running
         - configuration
+        - failed
+        - termination_pending_fast
+        - termination_pending_immediate
+        - terminated
       example: running
 
     ExtensionInstallRequest:
@@ -497,25 +591,69 @@ 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
@@ -544,6 +682,17 @@ components:
           description: Role name.
           example: "neon"
 
+    TerminateResponse:
+      type: object
+      required:
+        - lsn
+      properties:
+        lsn:
+          type: string
+          nullable: true
+          description: "last WAL flush LSN"
+          example: "0/028F10D8"
+
     SetRoleGrantsResponse:
       type: object
       required:

compute_tools/src/http/routes/mod.rs

@@ -14,6 +14,8 @@ 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 profile;
 pub(in crate::http) mod status;
 pub(in crate::http) mod terminate;
 

compute_tools/src/http/routes/profile.rs

@@ -0,0 +1,217 @@
+//! Contains the route for profiling the compute.
+//!
+//! Profiling the compute means generating a pprof profile of the
+//! postgres processes.
+//!
+//! The profiling is done using the `perf` tool, which is expected to be
+//! available somewhere in `$PATH`.
+use std::sync::atomic::Ordering;
+
+use axum::Json;
+use axum::response::IntoResponse;
+use http::StatusCode;
+use nix::unistd::Pid;
+use once_cell::sync::Lazy;
+use tokio::sync::Mutex;
+
+use crate::http::JsonResponse;
+
+static CANCEL_CHANNEL: Lazy<Mutex<Option<tokio::sync::broadcast::Sender<()>>>> =
+    Lazy::new(|| Mutex::new(None));
+
+fn default_sampling_frequency() -> u16 {
+    100
+}
+
+fn default_timeout_seconds() -> u8 {
+    5
+}
+
+fn deserialize_sampling_frequency<'de, D>(deserializer: D) -> Result<u16, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    use serde::Deserialize;
+
+    const MIN_SAMPLING_FREQUENCY: u16 = 1;
+    const MAX_SAMPLING_FREQUENCY: u16 = 1000;
+
+    let value = u16::deserialize(deserializer)?;
+
+    if !(MIN_SAMPLING_FREQUENCY..=MAX_SAMPLING_FREQUENCY).contains(&value) {
+        return Err(serde::de::Error::custom(format!(
+            "sampling_frequency must be between {MIN_SAMPLING_FREQUENCY} and {MAX_SAMPLING_FREQUENCY}, got {value}"
+        )));
+    }
+    Ok(value)
+}
+
+fn deserialize_profiling_timeout<'de, D>(deserializer: D) -> Result<u8, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    use serde::Deserialize;
+
+    const MIN_TIMEOUT_SECONDS: u8 = 1;
+    const MAX_TIMEOUT_SECONDS: u8 = 60;
+
+    let value = u8::deserialize(deserializer)?;
+
+    if !(MIN_TIMEOUT_SECONDS..=MAX_TIMEOUT_SECONDS).contains(&value) {
+        return Err(serde::de::Error::custom(format!(
+            "timeout_seconds must be between {MIN_TIMEOUT_SECONDS} and {MAX_TIMEOUT_SECONDS}, got {value}"
+        )));
+    }
+    Ok(value)
+}
+
+/// Request parameters for profiling the compute.
+#[derive(Debug, Clone, serde::Deserialize)]
+pub(in crate::http) struct ProfileRequest {
+    /// The profiling tool to use, currently only `perf` is supported.
+    profiler: crate::profiling::ProfileGenerator,
+    #[serde(default = "default_sampling_frequency")]
+    #[serde(deserialize_with = "deserialize_sampling_frequency")]
+    sampling_frequency: u16,
+    #[serde(default = "default_timeout_seconds")]
+    #[serde(deserialize_with = "deserialize_profiling_timeout")]
+    timeout_seconds: u8,
+    #[serde(default)]
+    archive: bool,
+}
+
+/// The HTTP request handler for reporting the profiling status of
+/// the compute.
+pub(in crate::http) async fn profile_status() -> impl IntoResponse {
+    tracing::info!("Profile status request received.");
+
+    let cancel_channel = CANCEL_CHANNEL.lock().await;
+
+    if let Some(tx) = cancel_channel.as_ref() {
+        if tx.receiver_count() > 0 {
+            return JsonResponse::create_response(
+                StatusCode::OK,
+                "Profiling is currently in progress.",
+            );
+        }
+    }
+
+    JsonResponse::create_response(StatusCode::NO_CONTENT, "Profiling is not in progress.")
+}
+
+/// The HTTP request handler for stopping profiling the compute.
+pub(in crate::http) async fn profile_stop() -> impl IntoResponse {
+    tracing::info!("Profile stop request received.");
+
+    match CANCEL_CHANNEL.lock().await.take() {
+        Some(tx) => {
+            if tx.send(()).is_err() {
+                tracing::error!("Failed to send cancellation signal.");
+                return JsonResponse::create_response(
+                    StatusCode::INTERNAL_SERVER_ERROR,
+                    "Failed to send cancellation signal",
+                );
+            }
+            JsonResponse::create_response(StatusCode::OK, "Profiling stopped successfully.")
+        }
+        None => JsonResponse::create_response(
+            StatusCode::PRECONDITION_FAILED,
+            "Profiling is not in progress, there is nothing to stop.",
+        ),
+    }
+}
+
+/// The HTTP request handler for starting profiling the compute.
+pub(in crate::http) async fn profile_start(
+    Json(request): Json<ProfileRequest>,
+) -> impl IntoResponse {
+    tracing::info!("Profile start request received: {request:?}");
+
+    let tx = tokio::sync::broadcast::Sender::<()>::new(1);
+
+    {
+        let mut cancel_channel = CANCEL_CHANNEL.lock().await;
+
+        if cancel_channel.is_some() {
+            return JsonResponse::create_response(
+                StatusCode::CONFLICT,
+                "Profiling is already in progress.",
+            );
+        }
+        *cancel_channel = Some(tx.clone());
+    }
+
+    tracing::info!("Profiling will start with parameters: {request:?}");
+    let pg_pid = Pid::from_raw(crate::compute::PG_PID.load(Ordering::SeqCst) as _);
+
+    let run_with_sudo = !cfg!(feature = "testing");
+
+    let options = crate::profiling::ProfileGenerationOptions {
+        profiler: request.profiler,
+        run_with_sudo,
+        pids: [pg_pid].into_iter().collect(),
+        follow_forks: true,
+        sampling_frequency: request.sampling_frequency as u32,
+        blocklist_symbols: vec![
+            "libc".to_owned(),
+            "libgcc".to_owned(),
+            "pthread".to_owned(),
+            "vdso".to_owned(),
+        ],
+        archive: request.archive,
+    };
+
+    let options = crate::profiling::ProfileGenerationTaskOptions {
+        options,
+        timeout: std::time::Duration::from_secs(request.timeout_seconds as u64),
+        should_stop: Some(tx),
+    };
+
+    let pprof_data = crate::profiling::generate_pprof_profile(options).await;
+
+    if CANCEL_CHANNEL.lock().await.take().is_none() {
+        tracing::error!("Profiling was cancelled from another request.");
+
+        return JsonResponse::create_response(
+            StatusCode::NO_CONTENT,
+            "Profiling was cancelled from another request.",
+        );
+    }
+
+    let pprof_data = match pprof_data {
+        Ok(data) => data,
+        Err(e) => {
+            tracing::error!(error = ?e, "failed to generate pprof data");
+            return JsonResponse::create_response(
+                StatusCode::INTERNAL_SERVER_ERROR,
+                format!("Failed to generate pprof data: {e:?}"),
+            );
+        }
+    };
+
+    tracing::info!("Profiling has completed successfully.");
+
+    let mut headers = http::HeaderMap::new();
+
+    if request.archive {
+        headers.insert(
+            http::header::CONTENT_TYPE,
+            http::HeaderValue::from_static("application/gzip"),
+        );
+        headers.insert(
+            http::header::CONTENT_DISPOSITION,
+            http::HeaderValue::from_static("attachment; filename=\"profile.pb.gz\""),
+        );
+    } else {
+        headers.insert(
+            http::header::CONTENT_TYPE,
+            http::HeaderValue::from_static("application/octet-stream"),
+        );
+        headers.insert(
+            http::header::CONTENT_DISPOSITION,
+            http::HeaderValue::from_static("attachment; filename=\"profile.pb\""),
+        );
+    }
+
+    (headers, pprof_data.0).into_response()
+}

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/routes/terminate.rs

@@ -3,7 +3,7 @@ use crate::http::JsonResponse;
 use axum::extract::State;
 use axum::response::Response;
 use axum_extra::extract::OptionalQuery;
-use compute_api::responses::{ComputeStatus, TerminateResponse};
+use compute_api::responses::{ComputeStatus, TerminateMode, TerminateResponse};
 use http::StatusCode;
 use serde::Deserialize;
 use std::sync::Arc;
@@ -12,7 +12,7 @@ use tracing::info;
 
 #[derive(Deserialize, Default)]
 pub struct TerminateQuery {
-    mode: compute_api::responses::TerminateMode,
+    mode: TerminateMode,
 }
 
 /// Terminate the compute.
@@ -24,16 +24,16 @@ pub(in crate::http) async fn terminate(
     {
         let mut state = compute.state.lock().unwrap();
         if state.status == ComputeStatus::Terminated {
-            return JsonResponse::success(StatusCode::CREATED, state.terminate_flush_lsn);
+            let response = TerminateResponse {
+                lsn: state.terminate_flush_lsn,
+            };
+            return JsonResponse::success(StatusCode::CREATED, response);
         }
 
         if !matches!(state.status, ComputeStatus::Empty | ComputeStatus::Running) {
             return JsonResponse::invalid_status(state.status);
         }
-        state.set_status(
-            ComputeStatus::TerminationPending { mode },
-            &compute.state_changed,
-        );
+        state.set_status(mode.into(), &compute.state_changed);
     }
 
     forward_termination_signal(false);

compute_tools/src/http/server.rs

@@ -23,10 +23,11 @@ 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;
+use crate::http::routes::profile;
 
 /// `compute_ctl` has two servers: internal and external. The internal server
 /// binds to the loopback interface and handles communication from clients on
@@ -81,12 +82,19 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
             Server::External {
                 config, compute_id, ..
             } => {
-                let unauthenticated_router =
-                    Router::<Arc<ComputeNode>>::new().route("/metrics", get(metrics::get_metrics));
+                let unauthenticated_router = Router::<Arc<ComputeNode>>::new()
+                    .route("/metrics", get(metrics::get_metrics))
+                    .route(
+                        "/profile/cpu",
+                        get(profile::profile_status)
+                            .post(profile::profile_start)
+                            .delete(profile::profile_stop),
+                    );
 
                 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;
@@ -23,6 +24,7 @@ pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
 pub mod pgbouncer;
+pub mod profiling;
 pub mod rsyslog;
 pub mod spec;
 mod spec_apply;

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/0002-alter_roles.sql

@@ -1,3 +1,16 @@
+-- On December 8th, 2023, an engineering escalation (INC-110) was opened after
+-- it was found that BYPASSRLS was being applied to all roles.
+--
+-- PR that introduced the issue: https://github.com/neondatabase/neon/pull/5657
+-- Subsequent commit on main: https://github.com/neondatabase/neon/commit/ad99fa5f0393e2679e5323df653c508ffa0ac072
+--
+-- NOBYPASSRLS and INHERIT are the defaults for a Postgres role, but because it
+-- isn't easy to know if a Postgres cluster is affected by the issue, we need to
+-- keep the migration around for a long time, if not indefinitely, so any
+-- cluster can be fixed.
+--
+-- Branching is the gift that keeps on giving...
+
 DO $$
 DECLARE
     role_name text;

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/monitor.rs

@@ -84,7 +84,8 @@ impl ComputeMonitor {
         if matches!(
             compute_status,
             ComputeStatus::Terminated
-                | ComputeStatus::TerminationPending { .. }
+                | ComputeStatus::TerminationPendingFast
+                | ComputeStatus::TerminationPendingImmediate
                 | ComputeStatus::Failed
         ) {
             info!(

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
@@ -918,16 +922,18 @@ impl Endpoint {
                         ComputeStatus::Empty
                         | ComputeStatus::ConfigurationPending
                         | ComputeStatus::Configuration
-                        | ComputeStatus::TerminationPending { .. }
+                        | ComputeStatus::TerminationPendingFast
+                        | ComputeStatus::TerminationPendingImmediate
                         | ComputeStatus::Terminated => {
                             bail!("unexpected compute status: {:?}", state.status)
                         }
                     }
                 }
                 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/continuous-profiling.md

@@ -0,0 +1,58 @@
+# Continuous Crofiling (Compute)
+
+The continuous profiling of the compute node is performed by `perf` or `bcc-tools`, the latter is preferred.
+
+The executables profiled are all the postgres-related ones only, excluding the actual compute code (Rust). This can be done as well but
+was not the main goal.
+
+## Tools
+
+The aforementioned tools are available within the same Docker image as
+the compute node itself, but the corresponding dependencies linux the
+linux kernel headers and the linux kernel itself are not and can't be
+for obvious reasons. To solve that, as we run the compute nodes as a
+virtual machine (qemu), we need to deliver these dependencies to it.
+This is done by the `autoscaling` part, which builds and deploys the
+kernel headers, needed modules, and the `perf` binary into an ext4-fs
+disk image, which is later attached to the VM and is symlinked to be
+made available for the compute node.
+
+## Output
+
+The output of the profiling is always a binary file in the same format
+of `pprof`. It can, however, be archived by `gzip` additionally, if the
+corresponding argument is provided in the JSON request.
+
+## REST API
+
+### Test profiling
+
+One can test the profiling after connecting to the VM and running:
+
+```sh
+curl -X POST -H "Content-Type: application/json" http://localhost:3080/profile/cpu -d '{"profiler": {"BccProfile": null}, "sampling_frequency": 99, "timeout_seconds": 5, "archive": false}' -v --output profile.pb
+```
+
+This uses the `Bcc` profiler and does not archive the output. The
+profiling data will be saved into the `profile.pb` file locally.
+
+**Only one profiling session can be run at a time.**
+
+To check the profiling status (to see whether it is already running or
+not), one can perform the `GET` request:
+
+```sh
+curl http://localhost:3080/profile/cpu -v
+```
+
+The profiling can be stopped by performing the `DELETE` request:
+
+```sh
+curl -X DELETE http://localhost:3080/profile/cpu -v
+```
+
+## Supported profiling data
+
+For now, only the CPU profiling is done and ther is no heap profiling.
+Also, only the postgres-related executables are tracked, the compute
+(Rust) part itself **is not tracked**.

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")]
@@ -93,6 +121,15 @@ pub enum TerminateMode {
     Immediate,
 }
 
+impl From<TerminateMode> for ComputeStatus {
+    fn from(mode: TerminateMode) -> Self {
+        match mode {
+            TerminateMode::Fast => ComputeStatus::TerminationPendingFast,
+            TerminateMode::Immediate => ComputeStatus::TerminationPendingImmediate,
+        }
+    }
+}
+
 #[derive(Serialize, Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "snake_case")]
 pub enum ComputeStatus {
@@ -113,7 +150,9 @@ pub enum ComputeStatus {
     // control-plane to terminate it.
     Failed,
     // Termination requested
-    TerminationPending { mode: TerminateMode },
+    TerminationPendingFast,
+    // Termination requested, without waiting 30s before returning from /terminate
+    TerminationPendingImmediate,
     // Terminated Postgres
     Terminated,
 }
@@ -132,7 +171,10 @@ impl Display for ComputeStatus {
             ComputeStatus::Running => f.write_str("running"),
             ComputeStatus::Configuration => f.write_str("configuration"),
             ComputeStatus::Failed => f.write_str("failed"),
-            ComputeStatus::TerminationPending { .. } => f.write_str("termination-pending"),
+            ComputeStatus::TerminationPendingFast => f.write_str("termination-pending-fast"),
+            ComputeStatus::TerminationPendingImmediate => {
+                f.write_str("termination-pending-immediate")
+            }
             ComputeStatus::Terminated => f.write_str("terminated"),
         }
     }

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

@@ -20,6 +20,7 @@ use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::io::ReaderStream;
 use tracing::{Instrument, debug, 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};
@@ -250,9 +251,28 @@ 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 +297,18 @@ 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 +321,10 @@ 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 +333,7 @@ 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/neon-shmem/Cargo.toml

@@ -6,8 +6,8 @@ license.workspace = true
 
 [dependencies]
 thiserror.workspace = true
-nix.workspace=true
+nix.workspace = true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
 
 [target.'cfg(target_os = "macos")'.dependencies]
-tempfile = "3.14.0"
+tempfile = "3.20.0"

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!(

libs/walproposer/src/api_bindings.rs

@@ -428,6 +428,12 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         shard_number: 0,
     };
 
+    let empty_wal_rate_limiter = crate::bindings::WalRateLimiter {
+        should_limit: crate::bindings::pg_atomic_uint32 { value: 0 },
+        sent_bytes: 0,
+        last_recorded_time_us: 0,
+    };
+
     crate::bindings::WalproposerShmemState {
         propEpochStartLsn: crate::bindings::pg_atomic_uint64 { value: 0 },
         donor_name: [0; 64],
@@ -441,6 +447,7 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         num_shards: 0,
         replica_promote: false,
         min_ps_feedback: empty_feedback,
+        wal_rate_limiter: empty_wal_rate_limiter,
     }
 }
 

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/src/mgmt_api.rs

@@ -1,4 +1,4 @@
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap};
 use std::error::Error as _;
 use std::time::Duration;
 
@@ -251,6 +251,70 @@ impl Client {
         Ok(())
     }
 
+    pub async fn tenant_timeline_compact(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        force_image_layer_creation: bool,
+        must_force_image_layer_creation: bool,
+        scheduled: bool,
+        wait_until_done: bool,
+    ) -> Result<()> {
+        let mut path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/compact",
+            self.mgmt_api_endpoint
+        ))
+        .expect("Cannot build URL");
+
+        if force_image_layer_creation {
+            path.query_pairs_mut()
+                .append_pair("force_image_layer_creation", "true");
+        }
+
+        if must_force_image_layer_creation {
+            path.query_pairs_mut()
+                .append_pair("must_force_image_layer_creation", "true");
+        }
+
+        if scheduled {
+            path.query_pairs_mut().append_pair("scheduled", "true");
+        }
+        if wait_until_done {
+            path.query_pairs_mut()
+                .append_pair("wait_until_scheduled_compaction_done", "true");
+            path.query_pairs_mut()
+                .append_pair("wait_until_uploaded", "true");
+        }
+        self.request(Method::PUT, path, ()).await?;
+        Ok(())
+    }
+
+    /* BEGIN_HADRON */
+    pub async fn tenant_timeline_describe(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+    ) -> Result<TimelineInfo> {
+        let mut path = reqwest::Url::parse(&format!(
+            "{}/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}",
+            self.mgmt_api_endpoint
+        ))
+        .expect("Cannot build URL");
+        path.query_pairs_mut()
+            .append_pair("include-image-consistent-lsn", "true");
+
+        let response: reqwest::Response = self.request(Method::GET, path, ()).await?;
+        let body = response.json().await.map_err(Error::ReceiveBody)?;
+        Ok(body)
+    }
+
+    pub async fn list_tenant_visible_size(&self) -> Result<BTreeMap<TenantShardId, u64>> {
+        let uri = format!("{}/v1/list_tenant_visible_size", self.mgmt_api_endpoint);
+        let resp = self.get(&uri).await?;
+        resp.json().await.map_err(Error::ReceiveBody)
+    }
+    /* END_HADRON */
+
     pub async fn tenant_scan_remote_storage(
         &self,
         tenant_id: TenantId,

pageserver/client_grpc/Cargo.toml

@@ -0,0 +1,24 @@
+[package]
+name = "pageserver_client_grpc"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[features]
+testing = ["pageserver_api/testing"]
+
+[dependencies]
+anyhow.workspace = true
+arc-swap.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 +1,543 @@
+use std::collections::HashMap;
+use std::num::NonZero;
+use std::sync::Arc;
+
+use anyhow::anyhow;
+use arc_swap::ArcSwap;
+use futures::stream::FuturesUnordered;
+use futures::{FutureExt as _, StreamExt as _};
+use tonic::codec::CompressionEncoding;
+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 {
+    /// The tenant ID.
+    tenant_id: TenantId,
+    /// The timeline ID.
+    timeline_id: TimelineId,
+    /// The JWT auth token for this tenant, if any.
+    auth_token: Option<String>,
+    /// The compression to use, if any.
+    compression: Option<CompressionEncoding>,
+    /// The shards for this tenant.
+    shards: ArcSwap<Shards>,
+    /// The retry configuration.
+    retry: Retry,
+}
+
+impl PageserverClient {
+    /// Creates a new Pageserver client for a given tenant and timeline. Uses the Pageservers given
+    /// in the shard spec, which must be complete and must use gRPC URLs.
+    pub fn new(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_spec: ShardSpec,
+        auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self> {
+        let shards = Shards::new(
+            tenant_id,
+            timeline_id,
+            shard_spec,
+            auth_token.clone(),
+            compression,
+        )?;
+        Ok(Self {
+            tenant_id,
+            timeline_id,
+            auth_token,
+            compression,
+            shards: ArcSwap::new(Arc::new(shards)),
+            retry: Retry,
+        })
+    }
+
+    /// Updates the shards from the given shard spec. In-flight requests will complete using the
+    /// existing shards, but may retry with the new shards if they fail.
+    ///
+    /// TODO: verify that in-flight requests are allowed to complete, and that the old pools are
+    /// properly spun down and dropped afterwards.
+    pub fn update_shards(&self, shard_spec: ShardSpec) -> anyhow::Result<()> {
+        // Validate the shard spec. We should really use `ArcSwap::rcu` for this, to avoid races
+        // with concurrent updates, but that involves creating a new `Shards` on every attempt,
+        // which spins up a bunch of Tokio tasks and such. These should already be checked elsewhere
+        // in the stack, and if they're violated then we already have problems elsewhere, so a
+        // best-effort but possibly-racy check is okay here.
+        let old = self.shards.load_full();
+        if shard_spec.count < old.count {
+            return Err(anyhow!(
+                "can't reduce shard count from {} to {}",
+                old.count,
+                shard_spec.count
+            ));
+        }
+        if !old.count.is_unsharded() && shard_spec.stripe_size != old.stripe_size {
+            return Err(anyhow!(
+                "can't change stripe size from {} to {}",
+                old.stripe_size,
+                shard_spec.stripe_size
+            ));
+        }
+
+        let shards = Shards::new(
+            self.tenant_id,
+            self.timeline_id,
+            shard_spec,
+            self.auth_token.clone(),
+            self.compression,
+        )?;
+        self.shards.store(Arc::new(shards));
+        Ok(())
+    }
+
+    /// 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.load_full().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.load_full().get_zero().client().await?;
+                client.get_db_size(req).await
+            })
+            .await
+    }
+
+    /// Fetches pages. The `request_id` must be unique across all in-flight requests, and the
+    /// `attempt` must be 0 (incremented on retry). 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"));
+        }
+        // The request attempt must be 0. The client will increment it internally.
+        if req.request_id.attempt != 0 {
+            return Err(tonic::Status::invalid_argument("request attempt must be 0"));
+        }
+
+        // The shards may change while we're fetching pages. We execute the request using a stable
+        // view of the shards (especially important for requests that span shards), but retry the
+        // top-level (pre-split) request to pick up shard changes. This can lead to unnecessary
+        // retries and re-splits in some cases where requests span shards, but these are expected to
+        // be rare.
+        //
+        // TODO: the gRPC server and client doesn't yet properly support shard splits. Revisit this
+        // once we figure out how to handle these.
+        self.retry
+            .with(async |attempt| {
+                let mut req = req.clone();
+                req.request_id.attempt = attempt as u32;
+                Self::get_page_with_shards(req, &self.shards.load_full()).await
+            })
+            .await
+    }
+
+    /// Fetches pages using the given shards. This uses a stable view of the shards, regardless of
+    /// concurrent shard updates. Does not retry internally, but is retried by `get_page()`.
+    async fn get_page_with_shards(
+        req: page_api::GetPageRequest,
+        shards: &Shards,
+    ) -> tonic::Result<page_api::GetPageResponse> {
+        // Fast path: request is for a single shard.
+        if let Some(shard_id) =
+            GetPageSplitter::for_single_shard(&req, shards.count, shards.stripe_size)
+        {
+            return Self::get_page_with_shard(req, shards.get(shard_id)?).await;
+        }
+
+        // Request spans multiple shards. Split it, dispatch concurrent per-shard requests, and
+        // reassemble the responses.
+        let mut splitter = GetPageSplitter::split(req, shards.count, shards.stripe_size);
+
+        let mut shard_requests = FuturesUnordered::new();
+        for (shard_id, shard_req) in splitter.drain_requests() {
+            let future = Self::get_page_with_shard(shard_req, shards.get(shard_id)?)
+                .map(move |result| result.map(|resp| (shard_id, resp)));
+            shard_requests.push(future);
+        }
+
+        while let Some((shard_id, shard_response)) = shard_requests.next().await.transpose()? {
+            splitter.add_response(shard_id, shard_response)?;
+        }
+
+        splitter.get_response()
+    }
+
+    /// Fetches pages on the given shard. Does not retry internally.
+    async fn get_page_with_shard(
+        req: page_api::GetPageRequest,
+        shard: &Shard,
+    ) -> tonic::Result<page_api::GetPageResponse> {
+        let stream = shard.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")),
+            ));
+        }
+
+        // Check that we received the expected pages.
+        if req.rel != resp.rel {
+            return Err(tonic::Status::internal(format!(
+                "shard {} returned wrong relation, expected {} got {}",
+                shard.id, req.rel, resp.rel
+            )));
+        }
+        if !req
+            .block_numbers
+            .iter()
+            .copied()
+            .eq(resp.pages.iter().map(|p| p.block_number))
+        {
+            return Err(tonic::Status::internal(format!(
+                "shard {} returned wrong pages, expected {:?} got {:?}",
+                shard.id,
+                req.block_numbers,
+                resp.pages
+                    .iter()
+                    .map(|page| page.block_number)
+                    .collect::<Vec<_>>()
+            )));
+        }
+
+        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.load_full().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.load_full().get_zero().client().await?;
+                client.get_slru_segment(req).await
+            })
+            .await
+    }
+}
+
+/// Shard specification for a PageserverClient.
+pub struct ShardSpec {
+    /// Maps shard indices to gRPC URLs.
+    ///
+    /// INVARIANT: every shard 0..count is present, and shard 0 is always present.
+    /// INVARIANT: every URL is valid and uses grpc:// scheme.
+    urls: HashMap<ShardIndex, String>,
+    /// The shard count.
+    ///
+    /// NB: this is 0 for unsharded tenants, following `ShardIndex::unsharded()` convention.
+    count: ShardCount,
+    /// The stripe size for these shards.
+    stripe_size: ShardStripeSize,
+}
+
+impl ShardSpec {
+    /// Creates a new shard spec with the given URLs and stripe size. All shards must be given.
+    /// The stripe size may be omitted for unsharded tenants.
+    pub fn new(
+        urls: HashMap<ShardIndex, String>,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> anyhow::Result<Self> {
+        // Compute the shard count.
+        let count = match urls.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),
+        };
+
+        // Determine the stripe size. It doesn't matter for unsharded tenants.
+        if stripe_size.is_none() && !count.is_unsharded() {
+            return Err(anyhow!("stripe size must be given for sharded tenants"));
+        }
+        let stripe_size = stripe_size.unwrap_or_default();
+
+        // Validate the shard spec.
+        for (shard_id, url) in &urls {
+            // 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).
+
+            // Validate the URL.
+            if PageserverProtocol::from_connstring(url)? != PageserverProtocol::Grpc {
+                return Err(anyhow!("invalid shard URL {url}: must use gRPC"));
+            }
+        }
+
+        Ok(Self {
+            urls,
+            count,
+            stripe_size,
+        })
+    }
+}
+
+/// Tracks the tenant's shards.
+struct Shards {
+    /// Shards by shard index.
+    ///
+    /// INVARIANT: every shard 0..count is present.
+    /// INVARIANT: shard 0 is always present.
+    by_index: HashMap<ShardIndex, Shard>,
+    /// 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,
+}
+
+impl Shards {
+    /// Creates a new set of shards based on a shard spec.
+    fn new(
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        shard_spec: ShardSpec,
+        auth_token: Option<String>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self> {
+        // NB: the shard spec has already been validated when constructed.
+        let mut shards = HashMap::with_capacity(shard_spec.urls.len());
+        for (shard_id, url) in shard_spec.urls {
+            shards.insert(
+                shard_id,
+                Shard::new(
+                    url,
+                    tenant_id,
+                    timeline_id,
+                    shard_id,
+                    auth_token.clone(),
+                    compression,
+                )?,
+            );
+        }
+
+        Ok(Self {
+            by_index: shards,
+            count: shard_spec.count,
+            stripe_size: shard_spec.stripe_size,
+        })
+    }
+
+    /// Looks up the given shard.
+    #[allow(clippy::result_large_err)] // TODO: check perf impact
+    fn get(&self, shard_id: ShardIndex) -> tonic::Result<&Shard> {
+        self.by_index
+            .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 {
+    /// The shard ID.
+    id: ShardIndex,
+    /// 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>,
+        compression: Option<CompressionEncoding>,
+    ) -> anyhow::Result<Self> {
+        // 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(),
+            compression,
+            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(),
+                compression,
+                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,
+                compression,
+                None, // unbounded, limited by stream pool
+            ),
+            Some(MAX_BULK_STREAMS),
+            MAX_BULK_STREAM_QUEUE_DEPTH,
+        );
+
+        Ok(Self {
+            id: shard_id,
+            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, ShardSpec};

pageserver/client_grpc/src/pool.rs

@@ -0,0 +1,779 @@
+//! 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::codec::CompressionEncoding;
+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>,
+    /// Compression to use.
+    compression: Option<CompressionEncoding>,
+    /// 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>,
+        compression: Option<CompressionEncoding>,
+        max_clients: Option<NonZero<usize>>,
+    ) -> Arc<Self> {
+        let pool = Arc::new(Self {
+            tenant_id,
+            timeline_id,
+            shard_id,
+            auth_token,
+            compression,
+            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(),
+            self.compression,
+        )?;
+
+        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.
+        //
+        // NB: this will leak entries if the server doesn't respond to a request (by request ID).
+        // It shouldn't happen, and if it does it will often hold onto queue depth quota anyway and
+        // block further use. But we could consider reaping closed channels after some time.
+        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.
+        //
+        // TODO: actually, it's probably not okay. Queue depth release should be moved into the
+        // stream task, such that it continues to account for the queue depth slot until the server
+        // responds. Otherwise, if a slow request times out and keeps blocking the stream, the
+        // server will keep waiting on it and we can pile on subsequent requests (including the
+        // timeout retry) in the same stream and get blocked. But we may also want to avoid blocking
+        // requests on e.g. LSN waits and layer downloads, instead returning early to free up the
+        // stream. Or just scale out streams with a queue depth of 1 to sidestep all head-of-line
+        // blocking. TBD.
+        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 +1,154 @@
+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), passing the
+    /// attempt number starting at 0. 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(usize) -> O, // takes attempt number, starting at 0
+        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(retries))
+                    .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::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,
+            // NB: don't retry Internal. It is intended for serious errors such as invariant
+            // violations, and is also used for client-side invariant checks that would otherwise
+            // result in retry loops.
+            tonic::Code::Internal => 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 +1,209 @@
+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, ShardNumber};
+
+/// Splits GetPageRequests that straddle shard boundaries and assembles the responses.
+/// TODO: add tests for this.
+pub struct GetPageSplitter {
+    /// Split requests by shard index.
+    requests: HashMap<ShardIndex, page_api::GetPageRequest>,
+    /// The response being assembled. Preallocated with empty pages, to be filled in.
+    response: page_api::GetPageResponse,
+    /// Maps the offset in `request.block_numbers` and `response.pages` to the owning shard. Used
+    /// to assemble the response pages in the same order as the original request.
+    block_shards: Vec<ShardIndex>,
+}
+
+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.
+    pub fn for_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 first page's shard, for comparison. If there are no pages, just return the first
+        // shard (caller likely checked already, otherwise the server will reject it).
+        let Some(&first_page) = req.block_numbers.first() else {
+            return Some(ShardIndex::new(ShardNumber(0), count));
+        };
+        let key = rel_block_to_key(req.rel, first_page);
+        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::for_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);
+
+            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(),
+                })
+                .block_numbers
+                .push(blkno);
+            block_shards.push(shard_id);
+        }
+
+        // Construct a response to be populated by shard responses. Preallocate empty page slots
+        // with the expected block numbers.
+        let response = page_api::GetPageResponse {
+            request_id: req.request_id,
+            status_code: page_api::GetPageStatusCode::Ok,
+            reason: None,
+            rel: req.rel,
+            pages: req
+                .block_numbers
+                .into_iter()
+                .map(|block_number| {
+                    page_api::Page {
+                        block_number,
+                        image: Bytes::new(), // empty page slot to be filled in
+                    }
+                })
+                .collect(),
+        };
+
+        Self {
+            requests,
+            response,
+            block_shards,
+        }
+    }
+
+    /// Drains the per-shard requests, moving them out of the splitter 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. The response must match the request ID and have an OK
+    /// status code. A response must not already exist for the given shard ID.
+    #[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.
+        if response.status_code != page_api::GetPageStatusCode::Ok {
+            return Err(tonic::Status::internal(format!(
+                "unexpected non-OK response for shard {shard_id}: {} {}",
+                response.status_code,
+                response.reason.unwrap_or_default()
+            )));
+        }
+
+        if response.request_id != self.response.request_id {
+            return Err(tonic::Status::internal(format!(
+                "response ID mismatch for shard {shard_id}: expected {}, got {}",
+                self.response.request_id, response.request_id
+            )));
+        }
+
+        // Place the shard response pages into the assembled response, in request order.
+        let mut pages = response.pages.into_iter();
+
+        for (i, &s) in self.block_shards.iter().enumerate() {
+            if shard_id != s {
+                continue;
+            }
+
+            let Some(slot) = self.response.pages.get_mut(i) else {
+                return Err(tonic::Status::internal(format!(
+                    "no block_shards slot {i} for shard {shard_id}"
+                )));
+            };
+            let Some(page) = pages.next() else {
+                return Err(tonic::Status::internal(format!(
+                    "missing page {} in shard {shard_id} response",
+                    slot.block_number
+                )));
+            };
+            if page.block_number != slot.block_number {
+                return Err(tonic::Status::internal(format!(
+                    "shard {shard_id} returned wrong page at index {i}, expected {} got {}",
+                    slot.block_number, page.block_number
+                )));
+            }
+            if !slot.image.is_empty() {
+                return Err(tonic::Status::internal(format!(
+                    "shard {shard_id} returned duplicate page {} at index {i}",
+                    slot.block_number
+                )));
+            }
+
+            *slot = page;
+        }
+
+        // Make sure we've consumed all pages from the shard response.
+        if let Some(extra_page) = pages.next() {
+            return Err(tonic::Status::internal(format!(
+                "shard {shard_id} returned extra page: {}",
+                extra_page.block_number
+            )));
+        }
+
+        Ok(())
+    }
+
+    /// Fetches the final, assembled response.
+    #[allow(clippy::result_large_err)]
+    pub fn get_response(self) -> tonic::Result<page_api::GetPageResponse> {
+        // Check that the response is complete.
+        for (i, page) in self.response.pages.iter().enumerate() {
+            if page.image.is_empty() {
+                return Err(tonic::Status::internal(format!(
+                    "missing page {} for shard {}",
+                    page.block_number,
+                    self.block_shards
+                        .get(i)
+                        .map(|s| s.to_string())
+                        .unwrap_or_else(|| "?".to_string())
+                )));
+            }
+        }
+
+        Ok(self.response)
+    }
+}

pageserver/ctl/Cargo.toml

@@ -17,6 +17,7 @@ pageserver = { path = ".." }
 pageserver_api.workspace = true
 remote_storage = { path = "../../libs/remote_storage" }
 postgres_ffi.workspace = true
+serde.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
 tokio-util.workspace = true

pageserver/ctl/src/download_remote_object.rs

@@ -0,0 +1,85 @@
+use camino::Utf8PathBuf;
+use clap::Parser;
+use tokio_util::sync::CancellationToken;
+
+/// Download a specific object from remote storage to a local file.
+///
+/// The remote storage configuration is supplied via the `REMOTE_STORAGE_CONFIG` environment
+/// variable, in the same TOML format that the pageserver itself understands. This allows the
+/// command to work with any cloud supported by the `remote_storage` crate (currently AWS S3,
+/// Azure Blob Storage and local files), as long as the credentials are available via the
+/// standard environment variables expected by the underlying SDKs.
+///
+/// Examples for setting the environment variable:
+///
+/// ```bash
+/// # AWS S3 (region can also be provided via AWS_REGION)
+/// export REMOTE_STORAGE_CONFIG='remote_storage = { bucket_name = "my-bucket", bucket_region = "us-east-2" }'
+///
+/// # Azure Blob Storage (account key picked up from AZURE_STORAGE_ACCOUNT_KEY)
+/// export REMOTE_STORAGE_CONFIG='remote_storage = { container = "my-container", account = "my-account" }'
+/// ```
+#[derive(Parser)]
+pub(crate) struct DownloadRemoteObjectCmd {
+    /// Key / path of the object to download (relative to the remote storage prefix).
+    ///
+    /// Examples:
+    ///   "wal/3aa8f.../00000001000000000000000A"
+    ///   "pageserver/v1/tenants/<tenant_id>/timelines/<timeline_id>/layer_12345"
+    pub remote_path: String,
+
+    /// Path of the local file to create. Existing file will be overwritten.
+    ///
+    /// Examples:
+    ///   "./segment"
+    ///   "/tmp/layer_12345.parquet"
+    pub output_file: Utf8PathBuf,
+}
+
+pub(crate) async fn main(cmd: &DownloadRemoteObjectCmd) -> anyhow::Result<()> {
+    use remote_storage::{DownloadOpts, GenericRemoteStorage, RemotePath, RemoteStorageConfig};
+
+    // Fetch remote storage configuration from the environment
+    let config_str = std::env::var("REMOTE_STORAGE_CONFIG").map_err(|_| {
+        anyhow::anyhow!(
+            "'REMOTE_STORAGE_CONFIG' environment variable must be set to a valid remote storage TOML config"
+        )
+    })?;
+
+    let config = RemoteStorageConfig::from_toml_str(&config_str)?;
+
+    // Initialise remote storage client
+    let storage = GenericRemoteStorage::from_config(&config).await?;
+
+    // RemotePath must be relative – leading slashes confuse the parser.
+    let remote_path_str = cmd.remote_path.trim_start_matches('/');
+    let remote_path = RemotePath::from_string(remote_path_str)?;
+
+    let cancel = CancellationToken::new();
+
+    println!(
+        "Downloading '{remote_path}' from remote storage bucket {:?} ...",
+        config.storage.bucket_name()
+    );
+
+    // Start the actual download
+    let download = storage
+        .download(&remote_path, &DownloadOpts::default(), &cancel)
+        .await?;
+
+    // Stream to file
+    let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
+    let tmp_path = cmd.output_file.with_extension("tmp");
+    let mut file = tokio::fs::File::create(&tmp_path).await?;
+    tokio::io::copy(&mut reader, &mut file).await?;
+    file.sync_all().await?;
+    // Atomically move into place
+    tokio::fs::rename(&tmp_path, &cmd.output_file).await?;
+
+    println!(
+        "Downloaded to '{}'. Last modified: {:?}, etag: {}",
+        cmd.output_file, download.last_modified, download.etag
+    );
+
+    Ok(())
+}

pageserver/ctl/src/index_part.rs

@@ -1,10 +1,180 @@
-use anyhow::Context;
+use std::str::FromStr;
+
+use anyhow::{Context, Ok};
 use camino::Utf8PathBuf;
-use pageserver::tenant::IndexPart;
+use pageserver::tenant::{
+    IndexPart,
+    layer_map::{LayerMap, SearchResult},
+    remote_timeline_client::{index::LayerFileMetadata, remote_layer_path},
+    storage_layer::{LayerName, LayerVisibilityHint, PersistentLayerDesc, ReadableLayerWeak},
+};
+use pageserver_api::key::Key;
+use serde::Serialize;
+use std::collections::BTreeMap;
+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,
+    },
+    /// List all visible delta and image layers at the latest LSN.
+    ListVisibleLayers {
+        #[arg(long)]
+        path: Utf8PathBuf,
+    },
+}
+
+fn create_layer_map_from_index_part(
+    index_part: &IndexPart,
+    tenant_shard_id: TenantShardId,
+    timeline_id: TimelineId,
+) -> LayerMap {
+    let mut layer_map = LayerMap::default();
+    {
+        let mut updates = layer_map.batch_update();
+        for (key, value) in index_part.layer_metadata.iter() {
+            updates.insert_historic(PersistentLayerDesc::from_filename(
+                tenant_shard_id,
+                timeline_id,
+                key.clone(),
+                value.file_size,
+            ));
+        }
+    }
+    layer_map
+}
+
+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 layer_map = create_layer_map_from_index_part(&index_json, tenant_shard_id, timeline_id);
+    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(())
+}
+
+#[derive(Debug, Clone, Serialize)]
+struct VisibleLayers {
+    pub total_images: u64,
+    pub total_image_bytes: u64,
+    pub total_deltas: u64,
+    pub total_delta_bytes: u64,
+    pub layer_metadata: BTreeMap<LayerName, LayerFileMetadata>,
+}
+
+impl VisibleLayers {
+    pub fn new() -> Self {
+        Self {
+            layer_metadata: BTreeMap::new(),
+            total_images: 0,
+            total_image_bytes: 0,
+            total_deltas: 0,
+            total_delta_bytes: 0,
+        }
+    }
+
+    pub fn add_layer(&mut self, name: LayerName, layer: LayerFileMetadata) {
+        match name {
+            LayerName::Image(_) => {
+                self.total_images += 1;
+                self.total_image_bytes += layer.file_size;
+            }
+            LayerName::Delta(_) => {
+                self.total_deltas += 1;
+                self.total_delta_bytes += layer.file_size;
+            }
+        }
+        self.layer_metadata.insert(name, layer);
+    }
+}
+
+async fn list_visible_layers(path: &Utf8PathBuf) -> anyhow::Result<()> {
+    let tenant_id = TenantId::generate();
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+    let timeline_id = TimelineId::generate();
+
+    let bytes = tokio::fs::read(path).await.context("read file")?;
+    let index_part = IndexPart::from_json_bytes(&bytes).context("deserialize")?;
+    let layer_map = create_layer_map_from_index_part(&index_part, tenant_shard_id, timeline_id);
+    let mut visible_layers = VisibleLayers::new();
+    let (layers, _key_space) = layer_map.get_visibility(Vec::new());
+    for (layer, visibility) in layers {
+        if visibility == LayerVisibilityHint::Visible {
+            visible_layers.add_layer(
+                layer.layer_name(),
+                index_part
+                    .layer_metadata
+                    .get(&layer.layer_name())
+                    .unwrap()
+                    .clone(),
+            );
+        }
+    }
+    let output = serde_json::to_string_pretty(&visible_layers).context("serialize output")?;
+    println!("{output}");
+
+    Ok(())
 }
 
 pub(crate) async fn main(cmd: &IndexPartCmd) -> anyhow::Result<()> {
@@ -16,5 +186,13 @@ 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,
+        IndexPartCmd::ListVisibleLayers { path } => list_visible_layers(path).await,
     }
 }

pageserver/ctl/src/main.rs

@@ -4,6 +4,7 @@
 //!
 //! Separate, `metadata` subcommand allows to print and update pageserver's metadata file.
 
+mod download_remote_object;
 mod draw_timeline_dir;
 mod index_part;
 mod key;
@@ -16,6 +17,7 @@ use std::time::{Duration, SystemTime};
 
 use camino::{Utf8Path, Utf8PathBuf};
 use clap::{Parser, Subcommand};
+use download_remote_object::DownloadRemoteObjectCmd;
 use index_part::IndexPartCmd;
 use layers::LayerCmd;
 use page_trace::PageTraceCmd;
@@ -63,6 +65,7 @@ enum Commands {
     /// Debug print a hex key found from logs
     Key(key::DescribeKeyCommand),
     PageTrace(PageTraceCmd),
+    DownloadRemoteObject(DownloadRemoteObjectCmd),
 }
 
 /// Read and update pageserver metadata file
@@ -185,6 +188,9 @@ async fn main() -> anyhow::Result<()> {
         }
         Commands::Key(dkc) => dkc.execute(),
         Commands::PageTrace(cmd) => page_trace::main(&cmd)?,
+        Commands::DownloadRemoteObject(cmd) => {
+            download_remote_object::main(&cmd).await?;
+        }
     };
     Ok(())
 }

pageserver/page_api/proto/page_service.proto

@@ -153,7 +153,7 @@ message GetDbSizeResponse {
 message GetPageRequest {
   // A request ID. Will be included in the response. Should be unique for
   // in-flight requests on the stream.
-  uint64 request_id = 1;
+  RequestID request_id = 1;
   // The request class.
   GetPageClass request_class = 2;
   // The LSN to read at.
@@ -177,6 +177,14 @@ message GetPageRequest {
   repeated uint32 block_number = 5;
 }
 
+// A Request ID. Should be unique for in-flight requests on a stream. Included in the response.
+message RequestID {
+  // The base request ID.
+  uint64 id = 1;
+  // The request attempt. Starts at 0, incremented on each retry.
+  uint32 attempt = 2;
+}
+
 // A GetPageRequest class. Primarily intended for observability, but may also be
 // used for prioritization in the future.
 enum GetPageClass {
@@ -199,13 +207,26 @@ enum GetPageClass {
 // the entire batch is ready, so no one can make use of the individual pages.
 message GetPageResponse {
   // The original request's ID.
-  uint64 request_id = 1;
-  // The response status code.
+  RequestID request_id = 1;
+  // The response status code. If not OK, the rel and page fields will be empty.
   GetPageStatusCode status_code = 2;
   // A string describing the status, if any.
   string reason = 3;
-  // The 8KB page images, in the same order as the request. Empty if status_code != OK.
-  repeated bytes page_image = 4;
+  // The relation that the pages belong to.
+  RelTag rel = 4;
+  // The page(s), in the same order as the request.
+  repeated Page page = 5;
+}
+
+// A page.
+//
+// TODO: it would be slightly more efficient (but less convenient) to have separate arrays of block
+// numbers and images, but given the 8KB page size it's probably negligible. Benchmark it anyway.
+message Page {
+  // The page number.
+  uint32 block_number = 1;
+  // The materialized page image, as an 8KB byte vector.
+  bytes image = 2;
 }
 
 // A GetPageResponse status code.

pageserver/page_api/src/client.rs

@@ -1,23 +1,152 @@
-use anyhow::Result;
+use anyhow::Context as _;
+use futures::future::ready;
 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.and_then(|resp| ready(GetPageResponse::try_from(resp).map_err(|err| err.into()))))
+    }
+
+    /// 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 +159,29 @@ 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

@@ -356,7 +356,10 @@ impl TryFrom<proto::GetPageRequest> for GetPageRequest {
             return Err(ProtocolError::Missing("block_number"));
         }
         Ok(Self {
-            request_id: pb.request_id,
+            request_id: pb
+                .request_id
+                .ok_or(ProtocolError::Missing("request_id"))?
+                .into(),
             request_class: pb.request_class.into(),
             read_lsn: pb
                 .read_lsn
@@ -371,7 +374,7 @@ impl TryFrom<proto::GetPageRequest> for GetPageRequest {
 impl From<GetPageRequest> for proto::GetPageRequest {
     fn from(request: GetPageRequest) -> Self {
         Self {
-            request_id: request.request_id,
+            request_id: Some(request.request_id.into()),
             request_class: request.request_class.into(),
             read_lsn: Some(request.read_lsn.into()),
             rel: Some(request.rel.into()),
@@ -380,11 +383,54 @@ impl From<GetPageRequest> for proto::GetPageRequest {
     }
 }
 
-/// A GetPage request ID.
-pub type RequestID = u64;
+/// A GetPage request ID and retry attempt. Should be unique for in-flight requests on a stream.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Hash, PartialOrd, Ord)]
+pub struct RequestID {
+    /// The base request ID.
+    pub id: u64,
+    // The request attempt. Starts at 0, incremented on each retry.
+    pub attempt: u32,
+}
+
+impl RequestID {
+    /// Creates a new RequestID with the given ID and an initial attempt of 0.
+    pub fn new(id: u64) -> Self {
+        Self { id, attempt: 0 }
+    }
+}
+
+impl Display for RequestID {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}.{}", self.id, self.attempt)
+    }
+}
+
+impl From<proto::RequestId> for RequestID {
+    fn from(pb: proto::RequestId) -> Self {
+        Self {
+            id: pb.id,
+            attempt: pb.attempt,
+        }
+    }
+}
+
+impl From<u64> for RequestID {
+    fn from(id: u64) -> Self {
+        Self::new(id)
+    }
+}
+
+impl From<RequestID> for proto::RequestId {
+    fn from(request_id: RequestID) -> Self {
+        Self {
+            id: request_id.id,
+            attempt: request_id.attempt,
+        }
+    }
+}
 
 /// 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 +443,19 @@ 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 {
@@ -443,32 +502,41 @@ impl From<GetPageClass> for i32 {
 pub struct GetPageResponse {
     /// The original request's ID.
     pub request_id: RequestID,
-    /// The response status code.
+    /// The response status code. If not OK, the `rel` and `pages` fields will be empty.
     pub status_code: GetPageStatusCode,
     /// A string describing the status, if any.
     pub reason: Option<String>,
-    /// The 8KB page images, in the same order as the request. Empty if status != OK.
-    pub page_images: Vec<Bytes>,
+    /// The relation that the pages belong to.
+    pub rel: RelTag,
+    // The page(s), in the same order as the request.
+    pub pages: Vec<Page>,
 }
 
-impl From<proto::GetPageResponse> for GetPageResponse {
-    fn from(pb: proto::GetPageResponse) -> Self {
-        Self {
-            request_id: pb.request_id,
+impl TryFrom<proto::GetPageResponse> for GetPageResponse {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetPageResponse) -> Result<Self, ProtocolError> {
+        Ok(Self {
+            request_id: pb
+                .request_id
+                .ok_or(ProtocolError::Missing("request_id"))?
+                .into(),
             status_code: pb.status_code.into(),
             reason: Some(pb.reason).filter(|r| !r.is_empty()),
-            page_images: pb.page_image,
-        }
+            rel: pb.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
+            pages: pb.page.into_iter().map(Page::from).collect(),
+        })
     }
 }
 
 impl From<GetPageResponse> for proto::GetPageResponse {
     fn from(response: GetPageResponse) -> Self {
         Self {
-            request_id: response.request_id,
+            request_id: Some(response.request_id.into()),
             status_code: response.status_code.into(),
             reason: response.reason.unwrap_or_default(),
-            page_image: response.page_images,
+            rel: Some(response.rel.into()),
+            page: response.pages.into_iter().map(proto::Page::from).collect(),
         }
     }
 }
@@ -501,11 +569,39 @@ impl GetPageResponse {
             request_id,
             status_code,
             reason: Some(status.message().to_string()),
-            page_images: Vec::new(),
+            rel: RelTag::default(),
+            pages: Vec::new(),
         })
     }
 }
 
+// A page.
+#[derive(Clone, Debug)]
+pub struct Page {
+    /// The page number.
+    pub block_number: u32,
+    /// The materialized page image, as an 8KB byte vector.
+    pub image: Bytes,
+}
+
+impl From<proto::Page> for Page {
+    fn from(pb: proto::Page) -> Self {
+        Self {
+            block_number: pb.block_number,
+            image: pb.image,
+        }
+    }
+}
+
+impl From<Page> for proto::Page {
+    fn from(page: Page) -> Self {
+        Self {
+            block_number: page.block_number,
+            image: page.image,
+        }
+    }
+}
+
 /// A GetPage response status code.
 ///
 /// These are effectively equivalent to gRPC statuses. However, we use a bidirectional stream
@@ -602,6 +698,21 @@ 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/Cargo.toml

@@ -27,8 +27,9 @@ tokio-util.workspace = true
 tonic.workspace = true
 url.workspace = true
 
-pageserver_client.workspace = true
 pageserver_api.workspace = true
+pageserver_client.workspace = true
+pageserver_client_grpc.workspace = true
 pageserver_page_api.workspace = true
 utils = { path = "../../libs/utils/" }
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }

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

@@ -10,12 +10,14 @@ use anyhow::Context;
 use async_trait::async_trait;
 use bytes::Bytes;
 use camino::Utf8PathBuf;
+use futures::stream::FuturesUnordered;
 use futures::{Stream, StreamExt as _};
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpaceAccum;
 use pageserver_api::pagestream_api::{PagestreamGetPageRequest, PagestreamRequest};
 use pageserver_api::reltag::RelTag;
 use pageserver_api::shard::TenantShardId;
+use pageserver_client_grpc::{self as client_grpc, ShardSpec};
 use pageserver_page_api as page_api;
 use rand::prelude::*;
 use tokio::task::JoinSet;
@@ -37,6 +39,10 @@ pub(crate) struct Args {
     /// Pageserver connection string. Supports postgresql:// and grpc:// protocols.
     #[clap(long, default_value = "postgres://postgres@localhost:64000")]
     page_service_connstring: String,
+    /// Use the rich gRPC Pageserver client `client_grpc::PageserverClient`, rather than the basic
+    /// no-frills `page_api::Client`. Only valid with grpc:// connstrings.
+    #[clap(long)]
+    rich_client: bool,
     #[clap(long)]
     pageserver_jwt: Option<String>,
     #[clap(long, default_value = "1")]
@@ -332,6 +338,7 @@ async fn main_impl(
             let client: Box<dyn Client> = match scheme.as_str() {
                 "postgresql" | "postgres" => {
                     assert!(!args.compression, "libpq does not support compression");
+                    assert!(!args.rich_client, "rich client requires grpc://");
                     Box::new(
                         LibpqClient::new(&args.page_service_connstring, worker_id.timeline)
                             .await
@@ -339,6 +346,16 @@ async fn main_impl(
                     )
                 }
 
+                "grpc" if args.rich_client => Box::new(
+                    RichGrpcClient::new(
+                        &args.page_service_connstring,
+                        worker_id.timeline,
+                        args.compression,
+                    )
+                    .await
+                    .unwrap(),
+                ),
+
                 "grpc" => Box::new(
                     GrpcClient::new(
                         &args.page_service_connstring,
@@ -625,7 +642,7 @@ 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,
@@ -657,7 +674,7 @@ impl Client for GrpcClient {
         blks: Vec<u32>,
     ) -> anyhow::Result<()> {
         let req = page_api::GetPageRequest {
-            request_id: req_id,
+            request_id: req_id.into(),
             request_class: page_api::GetPageClass::Normal,
             read_lsn: page_api::ReadLsn {
                 request_lsn: req_lsn,
@@ -677,6 +694,79 @@ impl Client for GrpcClient {
             "unexpected status code: {}",
             resp.status_code,
         );
-        Ok((resp.request_id, resp.page_images))
+        Ok((
+            resp.request_id.id,
+            resp.pages.into_iter().map(|p| p.image).collect(),
+        ))
+    }
+}
+
+/// A rich gRPC Pageserver client.
+struct RichGrpcClient {
+    inner: Arc<client_grpc::PageserverClient>,
+    requests: FuturesUnordered<
+        Pin<Box<dyn Future<Output = anyhow::Result<page_api::GetPageResponse>> + Send>>,
+    >,
+}
+
+impl RichGrpcClient {
+    async fn new(
+        connstring: &str,
+        ttid: TenantTimelineId,
+        compression: bool,
+    ) -> anyhow::Result<Self> {
+        let inner = Arc::new(client_grpc::PageserverClient::new(
+            ttid.tenant_id,
+            ttid.timeline_id,
+            ShardSpec::new(
+                [(ShardIndex::unsharded(), connstring.to_string())].into(),
+                None,
+            )?,
+            None,
+            compression.then_some(tonic::codec::CompressionEncoding::Zstd),
+        )?);
+        Ok(Self {
+            inner,
+            requests: FuturesUnordered::new(),
+        })
+    }
+}
+
+#[async_trait]
+impl Client for RichGrpcClient {
+    async fn send_get_page(
+        &mut self,
+        req_id: u64,
+        req_lsn: Lsn,
+        mod_lsn: Lsn,
+        rel: RelTag,
+        blks: Vec<u32>,
+    ) -> anyhow::Result<()> {
+        let req = page_api::GetPageRequest {
+            request_id: req_id.into(),
+            request_class: page_api::GetPageClass::Normal,
+            read_lsn: page_api::ReadLsn {
+                request_lsn: req_lsn,
+                not_modified_since_lsn: Some(mod_lsn),
+            },
+            rel,
+            block_numbers: blks,
+        };
+        let inner = self.inner.clone();
+        self.requests.push(Box::pin(async move {
+            inner
+                .get_page(req)
+                .await
+                .map_err(|err| anyhow::anyhow!("{err}"))
+        }));
+        Ok(())
+    }
+
+    async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
+        let resp = self.requests.next().await.unwrap()?;
+        Ok((
+            resp.request_id.id,
+            resp.pages.into_iter().map(|p| p.image).collect(),
+        ))
     }
 }

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,