proxy: add benchmark for custom json logging vs official fmt logger

We don't want to depend on postgres_ffi in an API crate. If there is no
such dependency, we can compile stuff like `storcon_cli` without needing
a full working postgres build. Fixes regression of #12548 (before we
could compile it).

.dockerignore

@@ -27,4 +27,4 @@
 !storage_controller/
 !vendor/postgres-*/
 !workspace_hack/
-!build_tools/patches
+!build-tools/patches

.github/actionlint.yml

@@ -31,6 +31,7 @@ config-variables:
   - NEON_PROD_AWS_ACCOUNT_ID
   - PGREGRESS_PG16_PROJECT_ID
   - PGREGRESS_PG17_PROJECT_ID
+  - PREWARM_PGBENCH_SIZE
   - REMOTE_STORAGE_AZURE_CONTAINER
   - REMOTE_STORAGE_AZURE_REGION
   - SLACK_CICD_CHANNEL_ID

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

@@ -176,7 +176,11 @@ runs:
         fi
 
         if [[ $BUILD_TYPE == "debug" && $RUNNER_ARCH == 'X64' ]]; then
-          cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          # We don't use code coverage for regression tests (the step is disabled),
+          # so there's no need to collect it.
+          # Ref https://github.com/neondatabase/neon/issues/4540
+          # cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          cov_prefix=()
         else
           cov_prefix=()
         fi

.github/workflows/_build-and-test-locally.yml

@@ -150,7 +150,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
@@ -162,7 +162,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
@@ -174,7 +174,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v16
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v17 build
         id: cache_pg_17
@@ -186,7 +186,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v17
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Build all
         # Note: the Makefile picks up BUILD_TYPE and CARGO_PROFILE from the env variables

.github/workflows/benchmarking.yml

@@ -219,6 +219,7 @@ jobs:
           --ignore test_runner/performance/test_cumulative_statistics_persistence.py
           --ignore test_runner/performance/test_perf_many_relations.py
           --ignore test_runner/performance/test_perf_oltp_large_tenant.py
+          --ignore test_runner/performance/test_lfc_prewarm.py
       env:
         BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -410,6 +411,77 @@ jobs:
       env:
         SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
 
+  prewarm-test:
+    if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PGBENCH_SIZE: ${{ vars.PREWARM_PGBENCH_SIZE }}
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 17
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
+      PLATFORM: "neon-staging"
+
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    container:
+      image: ghcr.io/neondatabase/build-tools:pinned-bookworm
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+
+    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
+
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+
+    - name: Run prewarm benchmark
+      uses: ./.github/actions/run-python-test-set
+      with:
+        build_type: ${{ env.BUILD_TYPE }}
+        test_selection: performance/test_lfc_prewarm.py
+        run_in_parallel: false
+        save_perf_report: ${{ env.SAVE_PERF_REPORT }}
+        extra_params: -m remote_cluster --timeout 5400
+        pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        NEON_API_KEY: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Create Allure report
+      id: create-allure-report
+      if: ${{ !cancelled() }}
+      uses: ./.github/actions/allure-report-generate
+      with:
+        store-test-results-into-db: true
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
+
   generate-matrices:
     if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
     # Create matrices for the benchmarking jobs, so we run benchmarks on rds only once a week (on Saturday)

.github/workflows/build-build-tools-image.yml

@@ -72,7 +72,7 @@ jobs:
           ARCHS: ${{ inputs.archs || '["x64","arm64"]' }}
           DEBIANS: ${{ inputs.debians || '["bullseye","bookworm"]' }}
           IMAGE_TAG: |
-            ${{ hashFiles('build-tools.Dockerfile',
+            ${{ hashFiles('build-tools/Dockerfile',
                           '.github/workflows/build-build-tools-image.yml') }}
         run: |
           echo "archs=${ARCHS}"           | tee -a ${GITHUB_OUTPUT}
@@ -144,7 +144,7 @@ jobs:
 
       - uses: docker/build-push-action@471d1dc4e07e5cdedd4c2171150001c434f0b7a4 # v6.15.0
         with:
-          file: build-tools.Dockerfile
+          file: build-tools/Dockerfile
           context: .
           provenance: false
           push: true

.gitmodules

@@ -1,16 +1,16 @@
 [submodule "vendor/postgres-v14"]
 	path = vendor/postgres-v14
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_14_STABLE_neon
 [submodule "vendor/postgres-v15"]
 	path = vendor/postgres-v15
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_15_STABLE_neon
 [submodule "vendor/postgres-v16"]
 	path = vendor/postgres-v16
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_16_STABLE_neon
 [submodule "vendor/postgres-v17"]
 	path = vendor/postgres-v17
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_17_STABLE_neon

Cargo.lock

@@ -687,40 +687,13 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum"
-version = "0.7.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
-dependencies = [
- "async-trait",
- "axum-core 0.4.5",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "itoa",
- "matchit 0.7.3",
- "memchr",
- "mime",
- "percent-encoding",
- "pin-project-lite",
- "rustversion",
- "serde",
- "sync_wrapper 1.0.1",
- "tower 0.5.2",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum"
 version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6d6fd624c75e18b3b4c6b9caf42b1afe24437daaee904069137d8bab077be8b8"
 dependencies = [
- "axum-core 0.5.0",
+ "axum-core",
  "base64 0.22.1",
  "bytes",
  "form_urlencoded",
@@ -731,7 +704,7 @@ dependencies = [
  "hyper 1.4.1",
  "hyper-util",
  "itoa",
- "matchit 0.8.4",
+ "matchit",
  "memchr",
  "mime",
  "percent-encoding",
@@ -751,26 +724,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum-core"
-version = "0.4.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09f2bd6146b97ae3359fa0cc6d6b376d9539582c7b4220f041a33ec24c226199"
-dependencies = [
- "async-trait",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "mime",
- "pin-project-lite",
- "rustversion",
- "sync_wrapper 1.0.1",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum-core"
 version = "0.5.0"
@@ -797,8 +750,8 @@ version = "0.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fc6f625a1f7705c6cf62d0d070794e94668988b1c38111baeec177c715f7b"
 dependencies = [
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "bytes",
  "form_urlencoded",
  "futures-util",
@@ -1009,24 +962,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "bindgen"
-version = "0.70.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f49d8fed880d473ea71efb9bf597651e77201bdd4893efe54c9e5d65ae04ce6f"
-dependencies = [
- "bitflags 2.8.0",
- "cexpr",
- "clang-sys",
- "itertools 0.12.1",
- "proc-macro2",
- "quote",
- "regex",
- "rustc-hash 1.1.0",
- "shlex",
- "syn 2.0.100",
-]
-
 [[package]]
 name = "bindgen"
 version = "0.71.1"
@@ -1386,7 +1321,7 @@ dependencies = [
  "aws-sdk-kms",
  "aws-sdk-s3",
  "aws-smithy-types",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "base64 0.22.1",
  "bytes",
@@ -1401,10 +1336,8 @@ dependencies = [
  "hostname-validator",
  "http 1.1.0",
  "indexmap 2.9.0",
- "inferno 0.12.0",
  "itertools 0.10.5",
  "jsonwebtoken",
- "libproc",
  "metrics",
  "nix 0.30.1",
  "notify",
@@ -1418,8 +1351,6 @@ dependencies = [
  "postgres-types",
  "postgres_initdb",
  "postgres_versioninfo",
- "pprof 0.15.0",
- "prost 0.12.6",
  "regex",
  "remote_storage",
  "reqwest",
@@ -1431,7 +1362,6 @@ dependencies = [
  "serde_with",
  "signal-hook",
  "tar",
- "tempfile",
  "thiserror 1.0.69",
  "tokio",
  "tokio-postgres",
@@ -2152,7 +2082,7 @@ name = "endpoint_storage"
 version = "0.0.1"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "camino",
  "camino-tempfile",
@@ -2268,12 +2198,12 @@ checksum = "5443807d6dff69373d433ab9ef5378ad8df50ca6298caf15de6e52e24aaf54d5"
 
 [[package]]
 name = "errno"
-version = "0.3.13"
+version = "0.3.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "778e2ac28f6c47af28e4907f13ffd1e1ddbd400980a9abd7c8df189bf578a5ad"
+checksum = "a258e46cdc063eb8519c00b9fc845fc47bcfca4130e2f08e88665ceda8474245"
 dependencies = [
  "libc",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -2603,18 +2533,6 @@ dependencies = [
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "getrandom"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26145e563e54f2cadc477553f1ec5ee650b00862f0a58bcd12cbdc5f0ea2d2f4"
-dependencies = [
- "cfg-if",
- "libc",
- "r-efi",
- "wasi 0.14.2+wasi-0.2.4",
-]
-
 [[package]]
 name = "gettid"
 version = "0.1.3"
@@ -2872,15 +2790,6 @@ dependencies = [
  "digest",
 ]
 
-[[package]]
-name = "home"
-version = "0.5.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "589533453244b0995c858700322199b2becb13b627df2851f64a2775d024abcf"
-dependencies = [
- "windows-sys 0.59.0",
-]
-
 [[package]]
 name = "hostname"
 version = "0.4.0"
@@ -2990,7 +2899,7 @@ dependencies = [
  "jsonwebtoken",
  "metrics",
  "once_cell",
- "pprof 0.14.0",
+ "pprof",
  "regex",
  "routerify",
  "rustls 0.23.27",
@@ -3652,7 +3561,7 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 dependencies = [
- "spin 0.9.8",
+ "spin",
 ]
 
 [[package]]
@@ -3677,17 +3586,6 @@ version = "0.2.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4ec2a862134d2a7d32d7983ddcdd1c4923530833c9f2ea1a44fc5fa473989058"
 
-[[package]]
-name = "libproc"
-version = "0.14.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e78a09b56be5adbcad5aa1197371688dc6bb249a26da3bca2011ee2fb987ebfb"
-dependencies = [
- "bindgen 0.70.1",
- "errno",
- "libc",
-]
-
 [[package]]
 name = "linux-raw-sys"
 version = "0.4.14"
@@ -3700,12 +3598,6 @@ version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f0b5399f6804fbab912acbd8878ed3532d506b7c951b8f9f164ef90fef39e3f4"
 
-[[package]]
-name = "linux-raw-sys"
-version = "0.9.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd945864f07fe9f5371a27ad7b52a172b4b499999f1d97574c9fa68373937e12"
-
 [[package]]
 name = "litemap"
 version = "0.7.4"
@@ -3759,12 +3651,6 @@ dependencies = [
  "regex-automata 0.1.10",
 ]
 
-[[package]]
-name = "matchit"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"
-
 [[package]]
 name = "matchit"
 version = "0.8.4"
@@ -4410,6 +4296,7 @@ dependencies = [
  "pageserver_client",
  "pageserver_client_grpc",
  "pageserver_page_api",
+ "pprof",
  "rand 0.8.5",
  "reqwest",
  "serde",
@@ -4504,7 +4391,7 @@ dependencies = [
  "postgres_ffi_types",
  "postgres_initdb",
  "posthog_client_lite",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "procfs",
  "rand 0.8.5",
@@ -5079,7 +4966,7 @@ name = "postgres_ffi"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "bindgen 0.71.1",
+ "bindgen",
  "bytes",
  "crc32c",
  "criterion",
@@ -5090,7 +4977,7 @@ dependencies = [
  "postgres",
  "postgres_ffi_types",
  "postgres_versioninfo",
- "pprof 0.14.0",
+ "pprof",
  "regex",
  "serde",
  "thiserror 1.0.69",
@@ -5180,30 +5067,6 @@ dependencies = [
  "thiserror 1.0.69",
 ]
 
-[[package]]
-name = "pprof"
-version = "0.15.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38a01da47675efa7673b032bf8efd8214f1917d89685e07e395ab125ea42b187"
-dependencies = [
- "aligned-vec",
- "backtrace",
- "cfg-if",
- "findshlibs",
- "inferno 0.11.21",
- "libc",
- "log",
- "nix 0.26.4",
- "once_cell",
- "protobuf",
- "protobuf-codegen",
- "smallvec",
- "spin 0.10.0",
- "symbolic-demangle",
- "tempfile",
- "thiserror 2.0.11",
-]
-
 [[package]]
 name = "pprof_util"
 version = "0.7.0"
@@ -5279,7 +5142,7 @@ dependencies = [
  "hex",
  "lazy_static",
  "procfs-core",
- "rustix 0.38.41",
+ "rustix",
 ]
 
 [[package]]
@@ -5415,57 +5278,6 @@ dependencies = [
  "prost 0.13.5",
 ]
 
-[[package]]
-name = "protobuf"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d65a1d4ddae7d8b5de68153b48f6aa3bba8cb002b243dbdbc55a5afbc98f99f4"
-dependencies = [
- "once_cell",
- "protobuf-support",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "protobuf-codegen"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d3976825c0014bbd2f3b34f0001876604fe87e0c86cd8fa54251530f1544ace"
-dependencies = [
- "anyhow",
- "once_cell",
- "protobuf",
- "protobuf-parse",
- "regex",
- "tempfile",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "protobuf-parse"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b4aeaa1f2460f1d348eeaeed86aea999ce98c1bded6f089ff8514c9d9dbdc973"
-dependencies = [
- "anyhow",
- "indexmap 2.9.0",
- "log",
- "protobuf",
- "protobuf-support",
- "tempfile",
- "thiserror 1.0.69",
- "which",
-]
-
-[[package]]
-name = "protobuf-support"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e36c2f31e0a47f9280fb347ef5e461ffcd2c52dd520d8e216b52f93b0b0d7d6"
-dependencies = [
- "thiserror 1.0.69",
-]
-
 [[package]]
 name = "proxy"
 version = "0.1.0"
@@ -5478,6 +5290,7 @@ dependencies = [
  "async-trait",
  "atomic-take",
  "aws-config",
+ "aws-credential-types",
  "aws-sdk-iam",
  "aws-sigv4",
  "base64 0.22.1",
@@ -5490,6 +5303,7 @@ dependencies = [
  "clashmap",
  "compute_api",
  "consumption_metrics",
+ "criterion",
  "ecdsa 0.16.9",
  "ed25519-dalek",
  "env_logger",
@@ -5517,6 +5331,7 @@ dependencies = [
  "itoa",
  "jose-jwa",
  "jose-jwk",
+ "json",
  "lasso",
  "measured",
  "metrics",
@@ -5637,12 +5452,6 @@ dependencies = [
  "proc-macro2",
 ]
 
-[[package]]
-name = "r-efi"
-version = "5.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69cdb34c158ceb288df11e18b4bd39de994f6657d83847bdffdbd7f346754b0f"
-
 [[package]]
 name = "rand"
 version = "0.7.3"
@@ -6031,7 +5840,7 @@ dependencies = [
  "async-trait",
  "getrandom 0.2.11",
  "http 1.1.0",
- "matchit 0.8.4",
+ "matchit",
  "opentelemetry",
  "reqwest",
  "reqwest-middleware",
@@ -6222,19 +6031,6 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
-[[package]]
-name = "rustix"
-version = "1.0.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c71e83d6afe7ff64890ec6b71d6a69bb8a610ab78ce364b3352876bb4c801266"
-dependencies = [
- "bitflags 2.8.0",
- "errno",
- "libc",
- "linux-raw-sys 0.9.4",
- "windows-sys 0.59.0",
-]
-
 [[package]]
 name = "rustls"
 version = "0.21.12"
@@ -6416,8 +6212,9 @@ dependencies = [
  "postgres-protocol",
  "postgres_backend",
  "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "rand 0.8.5",
  "regex",
@@ -6460,7 +6257,7 @@ dependencies = [
  "anyhow",
  "const_format",
  "pageserver_api",
- "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
  "pq_proto",
  "serde",
@@ -7005,18 +6802,6 @@ name = "spin"
 version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6980e8d7511241f8acf4aebddbb1ff938df5eebe98691418c4468d0b72a96a67"
-dependencies = [
- "lock_api",
-]
-
-[[package]]
-name = "spin"
-version = "0.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d5fe4ccb98d9c292d56fec89a5e07da7fc4cf0dc11e156b41793132775d3e591"
-dependencies = [
- "lock_api",
-]
 
 [[package]]
 name = "spinning_top"
@@ -7211,6 +6996,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "reqwest",
+ "safekeeper_api",
  "serde_json",
  "storage_controller_client",
  "tokio",
@@ -7377,14 +7163,14 @@ dependencies = [
 
 [[package]]
 name = "tempfile"
-version = "3.20.0"
+version = "3.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8a64e3985349f2441a1a9ef0b853f869006c3855f2cda6862a94d26ebb9d6a1"
+checksum = "28cce251fcbc87fac86a866eeb0d6c2d536fc16d06f184bb61aeae11aa4cee0c"
 dependencies = [
+ "cfg-if",
  "fastrand 2.2.0",
- "getrandom 0.3.3",
  "once_cell",
- "rustix 1.0.7",
+ "rustix",
  "windows-sys 0.59.0",
 ]
 
@@ -7780,6 +7566,7 @@ dependencies = [
  "futures-core",
  "pin-project-lite",
  "tokio",
+ "tokio-util",
 ]
 
 [[package]]
@@ -7878,25 +7665,16 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
 dependencies = [
- "async-stream",
  "async-trait",
- "axum 0.7.9",
  "base64 0.22.1",
  "bytes",
- "h2 0.4.4",
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.4.1",
- "hyper-timeout",
- "hyper-util",
  "percent-encoding",
  "pin-project",
  "prost 0.13.5",
- "socket2",
- "tokio",
  "tokio-stream",
- "tower 0.4.13",
  "tower-layer",
  "tower-service",
  "tracing",
@@ -7909,7 +7687,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7e581ba15a835f4d9ea06c55ab1bd4dce26fc53752c69a04aac00703bfb49ba9"
 dependencies = [
  "async-trait",
- "axum 0.8.1",
+ "axum",
  "base64 0.22.1",
  "bytes",
  "flate2",
@@ -7970,16 +7748,11 @@ checksum = "b8fa9be0de6cf49e536ce1851f987bd21a43b771b09473c3549a6c853db37c1c"
 dependencies = [
  "futures-core",
  "futures-util",
- "indexmap 1.9.3",
  "pin-project",
  "pin-project-lite",
- "rand 0.8.5",
- "slab",
  "tokio",
- "tokio-util",
  "tower-layer",
  "tower-service",
- "tracing",
 ]
 
 [[package]]
@@ -8411,7 +8184,7 @@ dependencies = [
  "pem",
  "pin-project-lite",
  "postgres_connection",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "rand 0.8.5",
  "regex",
@@ -8463,7 +8236,7 @@ name = "vm_monitor"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "cgroups-rs",
  "clap",
  "futures",
@@ -8518,7 +8291,7 @@ dependencies = [
  "pageserver_api",
  "postgres_ffi",
  "postgres_ffi_types",
- "pprof 0.14.0",
+ "pprof",
  "prost 0.13.5",
  "remote_storage",
  "serde",
@@ -8548,7 +8321,7 @@ name = "walproposer"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "bindgen 0.71.1",
+ "bindgen",
  "postgres_ffi",
  "utils",
 ]
@@ -8575,15 +8348,6 @@ version = "0.11.0+wasi-snapshot-preview1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423"
 
-[[package]]
-name = "wasi"
-version = "0.14.2+wasi-0.2.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9683f9a5a998d873c0d21fcbe3c083009670149a8fab228644b8bd36b2c48cb3"
-dependencies = [
- "wit-bindgen-rt",
-]
-
 [[package]]
 name = "wasite"
 version = "0.1.0"
@@ -8713,18 +8477,6 @@ dependencies = [
  "rustls-pki-types",
 ]
 
-[[package]]
-name = "which"
-version = "4.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "87ba24419a2078cd2b0f2ede2691b6c66d8e47836da3b6db8265ebad47afbfc7"
-dependencies = [
- "either",
- "home",
- "once_cell",
- "rustix 0.38.41",
-]
-
 [[package]]
 name = "whoami"
 version = "1.5.1"
@@ -8953,15 +8705,6 @@ dependencies = [
  "windows-sys 0.48.0",
 ]
 
-[[package]]
-name = "wit-bindgen-rt"
-version = "0.39.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f42320e61fe2cfd34354ecb597f86f413484a798ba44a8ca1165c58d42da6c1"
-dependencies = [
- "bitflags 2.8.0",
-]
-
 [[package]]
 name = "workspace_hack"
 version = "0.1.0"
@@ -8969,19 +8712,17 @@ dependencies = [
  "ahash",
  "anstream",
  "anyhow",
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "base64 0.21.7",
  "base64ct",
  "bytes",
  "camino",
  "cc",
  "chrono",
- "clang-sys",
  "clap",
  "clap_builder",
  "const-oid",
- "criterion",
  "crypto-bigint 0.5.5",
  "der 0.7.8",
  "deranged",
@@ -9008,7 +8749,6 @@ dependencies = [
  "hyper 0.14.30",
  "hyper 1.4.1",
  "hyper-util",
- "indexmap 1.9.3",
  "indexmap 2.9.0",
  "itertools 0.12.1",
  "lazy_static",
@@ -9025,6 +8765,7 @@ dependencies = [
  "num-iter",
  "num-rational",
  "num-traits",
+ "once_cell",
  "p256 0.13.2",
  "parquet",
  "prettyplease",
@@ -9046,7 +8787,6 @@ dependencies = [
  "sha2",
  "signature 2.2.0",
  "smallvec",
- "spin 0.9.8",
  "spki 0.7.3",
  "stable_deref_trait",
  "subtle",
@@ -9061,7 +8801,6 @@ dependencies = [
  "tokio-stream",
  "tokio-util",
  "toml_edit",
- "tonic 0.12.3",
  "tower 0.5.2",
  "tracing",
  "tracing-core",

Cargo.toml

@@ -130,7 +130,6 @@ 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" }
@@ -202,7 +201,7 @@ tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.g
 tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.12.0"
 tokio-rustls = { version = "0.26.0", default-features = false, features = ["tls12", "ring"]}
-tokio-stream = "0.1"
+tokio-stream = { version = "0.1", features = ["sync"] }
 tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
@@ -279,7 +278,6 @@ 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/" }

Dockerfile

@@ -109,8 +109,6 @@ RUN set -e \
         libreadline-dev \
         libseccomp-dev \
         ca-certificates \
-        bpfcc-tools \
-        sudo \
         openssl \
         unzip \
         curl \

build-tools/Dockerfile

@@ -35,7 +35,7 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
     echo -e "retry_connrefused=on\ntimeout=15\ntries=5\nretry-on-host-error=on\n" > /root/.wgetrc && \
     echo -e "--retry-connrefused\n--connect-timeout 15\n--retry 5\n--max-time 300\n" > /root/.curlrc
 
-COPY build_tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
+COPY build-tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
 
 RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         set -e && \
@@ -61,9 +61,6 @@ RUN if [ "${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  \
@@ -108,21 +105,15 @@ 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 -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 \
+RUN set -e \
     && apt update \
     && apt install -y \
         autoconf \
         automake \
-        bc \
         bison \
         build-essential \
         ca-certificates \
         cmake \
-        cpio \
         curl \
         flex \
         gdb \
@@ -131,10 +122,8 @@ RUN set -ex \
         gzip \
         jq \
         jsonnet \
-        kmod \
         libcurl4-openssl-dev \
         libbz2-dev \
-        libelf-dev \
         libffi-dev \
         liblzma-dev \
         libncurses5-dev \
@@ -148,11 +137,6 @@ RUN set -ex \
         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 \
@@ -160,8 +144,6 @@ RUN set -ex \
         openssh-client \
         parallel \
         pkg-config \
-        rsync \
-        sudo \
         unzip \
         wget \
         xz-utils \
@@ -216,8 +198,6 @@ 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

compute/compute-node.Dockerfile

@@ -9,7 +9,7 @@
 #
 # build-tools:   This contains Rust compiler toolchain and other tools needed at compile
 #                time. This is also used for the storage builds. This image is defined in
-#                build-tools.Dockerfile.
+#                build-tools/Dockerfile.
 #
 # build-deps:    Contains C compiler, other build tools, and compile-time dependencies
 #                needed to compile PostgreSQL and most extensions. (Some extensions need
@@ -115,7 +115,7 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION
 
-# Keep in sync with build-tools.Dockerfile
+# Keep in sync with build-tools/Dockerfile
 ENV PROTOC_VERSION=25.1
 
 # Use strict mode for bash to catch errors early
@@ -149,9 +149,6 @@ 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 \
     $VERSION_INSTALLS \
@@ -1793,7 +1790,7 @@ RUN set -e \
 #########################################################################################
 FROM build-deps AS exporters
 ARG TARGETARCH
-# Keep sql_exporter version same as in build-tools.Dockerfile and
+# Keep sql_exporter version same as in build-tools/Dockerfile and
 # test_runner/regress/test_compute_metrics.py
 # See comment on the top of the file regading `echo`, `-e` and `\n`
 RUN if [ "$TARGETARCH" = "amd64" ]; then\
@@ -1991,10 +1988,6 @@ RUN apt update && \
         locales \
         lsof \
         procps \
-        bpfcc-tools \
-        libbpfcc \
-        libbpfcc-dev \
-        libclang-dev \
         rsyslog-gnutls \
         screen \
         tcpdump \

compute/vm-image-spec-bookworm.yaml

@@ -39,14 +39,6 @@ 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.
@@ -73,7 +65,7 @@ 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, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -160,8 +152,6 @@ 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

@@ -39,14 +39,6 @@ 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.
@@ -73,7 +65,7 @@ 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, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -156,8 +148,6 @@ 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

@@ -31,7 +31,6 @@ 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
@@ -50,7 +49,6 @@ 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"] }
@@ -80,10 +78,3 @@ 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/src/compute.rs

@@ -371,9 +371,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
     }
 }
 
-/// A handle to the Postgres process that is running in the compute
-/// node.
-pub struct PostgresHandle {
+struct PostgresHandle {
     postgres: std::process::Child,
     log_collector: JoinHandle<Result<()>>,
 }
@@ -1042,6 +1040,8 @@ impl ComputeNode {
             PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
         };
 
+        self.fix_zenith_signal_neon_signal()?;
+
         let mut state = self.state.lock().unwrap();
         state.metrics.pageserver_connect_micros =
             connected.duration_since(started).as_micros() as u64;
@@ -1051,6 +1051,27 @@ impl ComputeNode {
         Ok(())
     }
 
+    /// Move the Zenith signal file to Neon signal file location.
+    /// This makes Compute compatible with older PageServers that don't yet
+    /// know about the Zenith->Neon rename.
+    fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
+        let datadir = Path::new(&self.params.pgdata);
+
+        let neonsig = datadir.join("neon.signal");
+
+        if neonsig.is_file() {
+            return Ok(());
+        }
+
+        let zenithsig = datadir.join("zenith.signal");
+
+        if zenithsig.is_file() {
+            fs::copy(zenithsig, neonsig)?;
+        }
+
+        Ok(())
+    }
+
     /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
     /// the connection was established, and the (compressed) size of the basebackup.
     fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
@@ -1265,9 +1286,7 @@ impl ComputeNode {
 
         // In case of error, log and fail the check, but don't crash.
         // We're playing it safe because these errors could be transient
-        // and we don't yet retry. Also being careful here allows us to
-        // be backwards compatible with safekeepers that don't have the
-        // TIMELINE_STATUS API yet.
+        // and we don't yet retry.
         if responses.len() < quorum {
             error!(
                 "failed sync safekeepers check {:?} {:?} {:?}",
@@ -2466,7 +2485,7 @@ pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
                 serde_json::to_string(&extensions).expect("failed to serialize extensions list")
             );
         }
-        Err(err) => error!("could not get installed extensions: {err:?}"),
+        Err(err) => error!("could not get installed extensions: {err}"),
     }
     Ok(())
 }

compute_tools/src/http/routes/mod.rs

@@ -15,7 +15,6 @@ 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

@@ -1,217 +0,0 @@
-//! 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/server.rs

@@ -27,7 +27,6 @@ use super::{
     },
 };
 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
@@ -82,14 +81,8 @@ 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))
-                    .route(
-                        "/profile/cpu",
-                        get(profile::profile_status)
-                            .post(profile::profile_start)
-                            .delete(profile::profile_stop),
-                    );
+                let unauthenticated_router =
+                    Router::<Arc<ComputeNode>>::new().route("/metrics", get(metrics::get_metrics));
 
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/lfc/prewarm", get(lfc::prewarm_state).post(lfc::prewarm))

compute_tools/src/installed_extensions.rs

@@ -2,6 +2,7 @@ use std::collections::HashMap;
 
 use anyhow::Result;
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use tokio_postgres::error::Error as PostgresError;
 use tokio_postgres::{Client, Config, NoTls};
 
 use crate::metrics::INSTALLED_EXTENSIONS;
@@ -10,7 +11,7 @@ use crate::metrics::INSTALLED_EXTENSIONS;
 /// and to make database listing query here more explicit.
 ///
 /// Limit the number of databases to 500 to avoid excessive load.
-async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
+async fn list_dbs(client: &mut Client) -> Result<Vec<String>, PostgresError> {
     // `pg_database.datconnlimit = -2` means that the database is in the
     // invalid state
     let databases = client
@@ -37,7 +38,9 @@ async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
 /// Same extension can be installed in multiple databases with different versions,
 /// so we report a separate metric (number of databases where it is installed)
 /// for each extension version.
-pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExtensions> {
+pub async fn get_installed_extensions(
+    mut conf: Config,
+) -> Result<InstalledExtensions, PostgresError> {
     conf.application_name("compute_ctl:get_installed_extensions");
     let databases: Vec<String> = {
         let (mut client, connection) = conf.connect(NoTls).await?;

compute_tools/src/lib.rs

@@ -24,7 +24,6 @@ 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;

control_plane/src/broker.rs

@@ -36,7 +36,7 @@ impl StorageBroker {
     pub async fn start(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         let broker = &self.env.broker;
 
-        print!("Starting neon broker at {}", broker.client_url());
+        println!("Starting neon broker at {}", broker.client_url());
 
         let mut args = Vec::new();
 

control_plane/src/endpoint.rs

@@ -32,7 +32,8 @@
 //!     config.json                 - passed to `compute_ctl`
 //!     pgdata/
 //!         postgresql.conf       - copy of postgresql.conf created by `compute_ctl`
-//!         zenith.signal
+//!         neon.signal
+//!         zenith.signal         - copy of neon.signal, for backward compatibility
 //!         <other PostgreSQL files>
 //! ```
 //!
@@ -463,7 +464,7 @@ impl Endpoint {
         conf.append("max_connections", "100");
         conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
-        // It also defines how often the walreciever will send a feedback message to the wal sender.
+        // It also defines how often the walreceiver will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");
         conf.append("listen_addresses", &self.pg_address.ip().to_string());
         conf.append("port", &self.pg_address.port().to_string());

control_plane/src/local_env.rs

@@ -217,6 +217,9 @@ pub struct NeonStorageControllerConf {
     pub posthog_config: Option<PostHogConfig>,
 
     pub kick_secondary_downloads: Option<bool>,
+
+    #[serde(with = "humantime_serde")]
+    pub shard_split_request_timeout: Option<Duration>,
 }
 
 impl NeonStorageControllerConf {
@@ -250,6 +253,7 @@ impl Default for NeonStorageControllerConf {
             timeline_safekeeper_count: None,
             posthog_config: None,
             kick_secondary_downloads: None,
+            shard_split_request_timeout: None,
         }
     }
 }

control_plane/src/pageserver.rs

@@ -303,7 +303,7 @@ impl PageServerNode {
     async fn start_node(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
-        print!(
+        println!(
             "Starting pageserver node {} at '{}' in {:?}, retrying for {:?}",
             self.conf.id,
             self.pg_connection_config.raw_address(),

control_plane/src/safekeeper.rs

@@ -127,7 +127,7 @@ impl SafekeeperNode {
         extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
-        print!(
+        println!(
             "Starting safekeeper at '{}' in '{}', retrying for {:?}",
             self.pg_connection_config.raw_address(),
             self.datadir_path().display(),

control_plane/src/storage_controller.rs

@@ -648,6 +648,13 @@ impl StorageController {
             args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
         }
 
+        if let Some(duration) = self.config.shard_split_request_timeout {
+            args.push(format!(
+                "--shard-split-request-timeout={}",
+                humantime::Duration::from(duration)
+            ));
+        }
+
         let mut envs = vec![
             ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
             ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
@@ -660,7 +667,7 @@ impl StorageController {
             ));
         }
 
-        println!("Starting storage controller");
+        println!("Starting storage controller at {scheme}://{host}:{listen_port}");
 
         background_process::start_process(
             COMMAND,

control_plane/storcon_cli/Cargo.toml

@@ -14,6 +14,7 @@ humantime.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 reqwest.workspace = true
+safekeeper_api.workspace=true
 serde_json = { workspace = true, features = ["raw_value"] }
 storage_controller_client.workspace = true
 tokio.workspace = true

control_plane/storcon_cli/src/main.rs

@@ -11,7 +11,7 @@ use pageserver_api::controller_api::{
     PlacementPolicy, SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest,
     ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
     SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TimelineSafekeeperMigrateRequest,
 };
 use pageserver_api::models::{
     EvictionPolicy, EvictionPolicyLayerAccessThreshold, ShardParameters, TenantConfig,
@@ -21,6 +21,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardStripeSize, TenantShardId};
 use pageserver_client::mgmt_api::{self};
 use reqwest::{Certificate, Method, StatusCode, Url};
+use safekeeper_api::models::TimelineLocateResponse;
 use storage_controller_client::control_api::Client;
 use utils::id::{NodeId, TenantId, TimelineId};
 
@@ -279,6 +280,23 @@ enum Command {
         #[arg(long)]
         concurrency: Option<usize>,
     },
+    /// Locate safekeepers for a timeline from the storcon DB.
+    TimelineLocate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+    },
+    /// Migrate a timeline to a new set of safekeepers
+    TimelineSafekeeperMigrate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+        /// Example: --new-sk-set 1,2,3
+        #[arg(long, required = true, value_delimiter = ',')]
+        new_sk_set: Vec<NodeId>,
+    },
 }
 
 #[derive(Parser)]
@@ -458,6 +476,7 @@ async fn main() -> anyhow::Result<()> {
                         listen_http_port,
                         listen_https_port,
                         availability_zone_id: AvailabilityZone(availability_zone_id),
+                        node_ip_addr: None,
                     }),
                 )
                 .await?;
@@ -1324,7 +1343,7 @@ async fn main() -> anyhow::Result<()> {
             concurrency,
         } => {
             let mut path = format!(
-                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
+                "v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
             );
 
             if let Some(c) = concurrency {
@@ -1335,6 +1354,41 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::POST, path, None)
                 .await?;
         }
+        Command::TimelineLocate {
+            tenant_id,
+            timeline_id,
+        } => {
+            let path = format!("debug/v1/tenant/{tenant_id}/timeline/{timeline_id}/locate");
+
+            let resp = storcon_client
+                .dispatch::<(), TimelineLocateResponse>(Method::GET, path, None)
+                .await?;
+
+            let sk_set = resp.sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>();
+            let new_sk_set = resp
+                .new_sk_set
+                .as_ref()
+                .map(|ids| ids.iter().map(|id| id.0 as i64).collect::<Vec<_>>());
+
+            println!("generation = {}", resp.generation);
+            println!("sk_set = {sk_set:?}");
+            println!("new_sk_set = {new_sk_set:?}");
+        }
+        Command::TimelineSafekeeperMigrate {
+            tenant_id,
+            timeline_id,
+            new_sk_set,
+        } => {
+            let path = format!("v1/tenant/{tenant_id}/timeline/{timeline_id}/safekeeper_migrate");
+
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    path,
+                    Some(TimelineSafekeeperMigrateRequest { new_sk_set }),
+                )
+                .await?;
+        }
     }
 
     Ok(())

docs/continuous-profiling.md

@@ -1,58 +0,0 @@
-# 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/core_changes.md

@@ -129,9 +129,10 @@ segment to bootstrap the WAL writing, but it doesn't contain the checkpoint reco
 changes in xlog.c, to allow starting the compute node without reading the last checkpoint record
 from WAL.
 
-This includes code to read the `zenith.signal` file, which tells the startup code the LSN to start
-at. When the `zenith.signal` file is present, the startup uses that LSN instead of the last
-checkpoint's LSN. The system is known to be consistent at that LSN, without any WAL redo.
+This includes code to read the `neon.signal` (also `zenith.signal`) file, which tells the startup 
+code the LSN to start at. When the `neon.signal` file is present, the startup uses that LSN
+instead of the last checkpoint's LSN. The system is known to be consistent at that LSN, without 
+any WAL redo.
 
 
 ### How to get rid of the patch

docs/pageserver-services.md

@@ -75,7 +75,7 @@ CLI examples:
 * AWS S3  : `env AWS_ACCESS_KEY_ID='SOMEKEYAAAAASADSAH*#' AWS_SECRET_ACCESS_KEY='SOMEsEcReTsd292v' ${PAGESERVER_BIN} -c "remote_storage={bucket_name='some-sample-bucket',bucket_region='eu-north-1', prefix_in_bucket='/test_prefix/'}"`
 
 For Amazon AWS S3, a key id and secret access key could be located in `~/.aws/credentials` if awscli was ever configured to work with the desired bucket, on the AWS Settings page for a certain user. Also note, that the bucket names does not contain any protocols when used on AWS.
-For local S3 installations, refer to the their documentation for name format and credentials.
+For local S3 installations, refer to their documentation for name format and credentials.
 
 Similar to other pageserver settings, toml config file can be used to configure either of the storages as backup targets.
 Required sections are:

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 = "3.20.0"
+tempfile = "3.14.0"

libs/pageserver_api/src/controller_api.rs

@@ -1,5 +1,6 @@
 use std::collections::{HashMap, HashSet};
 use std::fmt::Display;
+use std::net::IpAddr;
 use std::str::FromStr;
 use std::time::{Duration, Instant};
 
@@ -10,7 +11,7 @@ use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
 
-use crate::models::{PageserverUtilization, ShardParameters, TenantConfig};
+use crate::models::{PageserverUtilization, ShardParameters, TenantConfig, TimelineInfo};
 use crate::shard::{ShardStripeSize, TenantShardId};
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -60,6 +61,11 @@ pub struct NodeRegisterRequest {
     pub listen_https_port: Option<u16>,
 
     pub availability_zone_id: AvailabilityZone,
+
+    // Reachable IP address of the PS/SK registering, if known.
+    // Hadron Cluster Coordiantor will update the DNS record of the registering node
+    // with this IP address.
+    pub node_ip_addr: Option<IpAddr>,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -126,6 +132,13 @@ pub struct TenantDescribeResponse {
     pub config: TenantConfig,
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantTimelineDescribeResponse {
+    pub shards: Vec<TimelineInfo>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct NodeShardResponse {
     pub node_id: NodeId,
@@ -538,6 +551,39 @@ pub struct SafekeeperDescribeResponse {
     pub scheduling_policy: SkSchedulingPolicy,
 }
 
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineSafekeeperPeer {
+    pub node_id: NodeId,
+    pub listen_http_addr: String,
+    pub http_port: i32,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimeline {
+    // SC does not know the tenant id.
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimelinesResponse {
+    pub timelines: Vec<SCSafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimeline {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimelinesResponse {
+    pub timelines: Vec<SafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
 #[derive(Serialize, Deserialize, Clone)]
 pub struct SafekeeperSchedulingPolicyRequest {
     pub scheduling_policy: SkSchedulingPolicy,

libs/pageserver_api/src/models.rs

@@ -1622,6 +1622,9 @@ pub struct TimelineInfo {
 
     /// Whether the timeline is invisible in synthetic size calculations.
     pub is_invisible: Option<bool>,
+    // HADRON: the largest LSN below which all page updates have been included in the image layers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]

libs/postgres_ffi/build.rs

@@ -110,7 +110,6 @@ fn main() -> anyhow::Result<()> {
             .allowlist_type("XLogRecPtr")
             .allowlist_type("XLogSegNo")
             .allowlist_type("TimeLineID")
-            .allowlist_type("TimestampTz")
             .allowlist_type("MultiXactId")
             .allowlist_type("MultiXactOffset")
             .allowlist_type("MultiXactStatus")

libs/postgres_ffi/src/lib.rs

@@ -227,8 +227,7 @@ pub mod walrecord;
 // Export some widely used datatypes that are unlikely to change across Postgres versions
 pub use v14::bindings::{
     BlockNumber, CheckPoint, ControlFileData, MultiXactId, OffsetNumber, Oid, PageHeaderData,
-    RepOriginId, TimeLineID, TimestampTz, TransactionId, XLogRecPtr, XLogRecord, XLogSegNo, uint32,
-    uint64,
+    RepOriginId, TimeLineID, TransactionId, XLogRecPtr, XLogRecord, XLogSegNo, uint32, uint64,
 };
 // Likewise for these, although the assumption that these don't change is a little more iffy.
 pub use v14::bindings::{MultiXactOffset, MultiXactStatus};

libs/postgres_ffi/src/walrecord.rs

@@ -4,13 +4,14 @@
 //! TODO: Generate separate types for each supported PG version
 
 use bytes::{Buf, Bytes};
+use postgres_ffi_types::TimestampTz;
 use serde::{Deserialize, Serialize};
 use utils::bin_ser::DeserializeError;
 use utils::lsn::Lsn;
 
 use crate::{
     BLCKSZ, BlockNumber, MultiXactId, MultiXactOffset, MultiXactStatus, Oid, PgMajorVersion,
-    RepOriginId, TimestampTz, TransactionId, XLOG_SIZE_OF_XLOG_RECORD, XLogRecord, pg_constants,
+    RepOriginId, TransactionId, XLOG_SIZE_OF_XLOG_RECORD, XLogRecord, pg_constants,
 };
 
 #[repr(C)]
@@ -863,7 +864,8 @@ pub mod v17 {
         XlHeapDelete, XlHeapInsert, XlHeapLock, XlHeapMultiInsert, XlHeapUpdate, XlParameterChange,
         rm_neon,
     };
-    pub use crate::{TimeLineID, TimestampTz};
+    pub use crate::TimeLineID;
+    pub use postgres_ffi_types::TimestampTz;
 
     #[repr(C)]
     #[derive(Debug)]

libs/postgres_ffi/src/xlog_utils.rs

@@ -9,10 +9,11 @@
 
 use super::super::waldecoder::WalStreamDecoder;
 use super::bindings::{
-    CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID, TimestampTz,
+    CheckPoint, ControlFileData, DBState_DB_SHUTDOWNED, FullTransactionId, TimeLineID,
     XLogLongPageHeaderData, XLogPageHeaderData, XLogRecPtr, XLogRecord, XLogSegNo, XLOG_PAGE_MAGIC,
     MY_PGVERSION
 };
+use postgres_ffi_types::TimestampTz;
 use super::wal_generator::LogicalMessageGenerator;
 use crate::pg_constants;
 use crate::PG_TLI;

libs/postgres_ffi_types/src/lib.rs

@@ -11,3 +11,4 @@ pub mod forknum;
 
 pub type Oid = u32;
 pub type RepOriginId = u16;
+pub type TimestampTz = i64;

libs/remote_storage/src/simulate_failures.rs

@@ -31,6 +31,7 @@ pub struct UnreliableWrapper {
     /* 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.
+    // Storage will fail by probability up to attempts_to_fail times.
     attempt_failure_probability: u64,
     /* END_HADRON */
 }

libs/safekeeper_api/Cargo.toml

@@ -9,7 +9,7 @@ anyhow.workspace = true
 const_format.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-postgres_ffi.workspace = true
+postgres_ffi_types.workspace = true
 postgres_versioninfo.workspace = true
 pq_proto.workspace = true
 tokio.workspace = true

libs/safekeeper_api/src/models.rs

@@ -3,7 +3,7 @@
 use std::net::SocketAddr;
 
 use pageserver_api::shard::ShardIdentity;
-use postgres_ffi::TimestampTz;
+use postgres_ffi_types::TimestampTz;
 use postgres_versioninfo::PgVersionId;
 use serde::{Deserialize, Serialize};
 use tokio::time::Instant;
@@ -11,7 +11,7 @@ use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;
 
-use crate::membership::Configuration;
+use crate::membership::{Configuration, SafekeeperGeneration};
 use crate::{ServerInfo, Term};
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -311,3 +311,12 @@ pub struct PullTimelineResponse {
     pub safekeeper_host: Option<String>,
     // TODO: add more fields?
 }
+
+/// Response to a timeline locate request.
+/// Storcon-only API.
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineLocateResponse {
+    pub generation: SafekeeperGeneration,
+    pub sk_set: Vec<NodeId>,
+    pub new_sk_set: Option<Vec<NodeId>>,
+}

libs/utils/src/env.rs

@@ -47,6 +47,7 @@ where
 
 /* BEGIN_HADRON */
 pub enum DeploymentMode {
+    Local,
     Dev,
     Staging,
     Prod,
@@ -64,7 +65,7 @@ pub fn get_deployment_mode() -> Option<DeploymentMode> {
             }
         },
         Err(_) => {
-            tracing::error!("DEPLOYMENT_MODE not set");
+            // tracing::error!("DEPLOYMENT_MODE not set");
             None
         }
     }

libs/utils/src/ip_address.rs

@@ -0,0 +1,73 @@
+use std::env::{VarError, var};
+use std::error::Error;
+use std::net::IpAddr;
+use std::str::FromStr;
+
+/// Name of the environment variable containing the reachable IP address of the node. If set, the IP address contained in this
+/// environment variable is used as the reachable IP address of the pageserver or safekeeper node during node registration.
+/// In a Kubernetes environment, this environment variable should be set by Kubernetes to the Pod IP (specified in the Pod
+/// template).
+pub const HADRON_NODE_IP_ADDRESS: &str = "HADRON_NODE_IP_ADDRESS";
+
+/// Read the reachable IP address of this page server from env var HADRON_NODE_IP_ADDRESS.
+/// In Kubernetes this environment variable is set to the Pod IP (specified in the Pod template).
+pub fn read_node_ip_addr_from_env() -> Result<Option<IpAddr>, Box<dyn Error>> {
+    match var(HADRON_NODE_IP_ADDRESS) {
+        Ok(v) => {
+            if let Ok(addr) = IpAddr::from_str(&v) {
+                Ok(Some(addr))
+            } else {
+                Err(format!("Invalid IP address string: {v}. Cannot be parsed as either an IPv4 or an IPv6 address.").into())
+            }
+        }
+        Err(VarError::NotPresent) => Ok(None),
+        Err(e) => Err(e.into()),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::env;
+    use std::net::{Ipv4Addr, Ipv6Addr};
+
+    #[test]
+    fn test_read_node_ip_addr_from_env() {
+        // SAFETY: test code
+        unsafe {
+            // Test with a valid IPv4 address
+            env::set_var(HADRON_NODE_IP_ADDRESS, "192.168.1.1");
+            let result = read_node_ip_addr_from_env().unwrap();
+            assert_eq!(result, Some(IpAddr::V4(Ipv4Addr::new(192, 168, 1, 1))));
+
+            // Test with a valid IPv6 address
+            env::set_var(
+                HADRON_NODE_IP_ADDRESS,
+                "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
+            );
+        }
+        let result = read_node_ip_addr_from_env().unwrap();
+        assert_eq!(
+            result,
+            Some(IpAddr::V6(
+                Ipv6Addr::from_str("2001:0db8:85a3:0000:0000:8a2e:0370:7334").unwrap()
+            ))
+        );
+
+        // Test with an invalid IP address
+        // SAFETY: test code
+        unsafe {
+            env::set_var(HADRON_NODE_IP_ADDRESS, "invalid_ip");
+        }
+        let result = read_node_ip_addr_from_env();
+        assert!(result.is_err());
+
+        // Test with no environment variable set
+        // SAFETY: test code
+        unsafe {
+            env::remove_var(HADRON_NODE_IP_ADDRESS);
+        }
+        let result = read_node_ip_addr_from_env().unwrap();
+        assert_eq!(result, None);
+    }
+}

libs/utils/src/lib.rs

@@ -26,6 +26,9 @@ pub mod auth;
 // utility functions and helper traits for unified unique id generation/serialization etc.
 pub mod id;
 
+// utility functions to obtain reachable IP addresses in PS/SK nodes.
+pub mod ip_address;
+
 pub mod shard;
 
 mod hex;

libs/utils/src/logging.rs

@@ -1,4 +1,5 @@
 use std::future::Future;
+use std::pin::Pin;
 use std::str::FromStr;
 use std::time::Duration;
 
@@ -7,7 +8,7 @@ use metrics::{IntCounter, IntCounterVec};
 use once_cell::sync::Lazy;
 use strum_macros::{EnumString, VariantNames};
 use tokio::time::Instant;
-use tracing::info;
+use tracing::{info, warn};
 
 /// Logs a critical error, similarly to `tracing::error!`. This will:
 ///
@@ -377,10 +378,11 @@ impl std::fmt::Debug for SecretString {
 ///
 /// TODO: consider upgrading this to a warning, but currently it fires too often.
 #[inline]
-pub async fn log_slow<F, O>(name: &str, threshold: Duration, f: std::pin::Pin<&mut F>) -> O
-where
-    F: Future<Output = O>,
-{
+pub async fn log_slow<O>(
+    name: &str,
+    threshold: Duration,
+    f: Pin<&mut impl Future<Output = O>>,
+) -> O {
     monitor_slow_future(
         threshold,
         threshold, // period = threshold
@@ -394,16 +396,42 @@ where
             if !is_slow {
                 return;
             }
+            let elapsed = elapsed_total.as_secs_f64();
             if ready {
-                info!(
-                    "slow {name} completed after {:.3}s",
-                    elapsed_total.as_secs_f64()
-                );
+                info!("slow {name} completed after {elapsed:.3}s");
             } else {
-                info!(
-                    "slow {name} still running after {:.3}s",
-                    elapsed_total.as_secs_f64()
-                );
+                info!("slow {name} still running after {elapsed:.3}s");
+            }
+        },
+    )
+    .await
+}
+
+/// Logs a periodic warning if a future is slow to complete.
+#[inline]
+pub async fn warn_slow<O>(
+    name: &str,
+    threshold: Duration,
+    f: Pin<&mut impl Future<Output = O>>,
+) -> O {
+    monitor_slow_future(
+        threshold,
+        threshold, // period = threshold
+        f,
+        |MonitorSlowFutureCallback {
+             ready,
+             is_slow,
+             elapsed_total,
+             elapsed_since_last_callback: _,
+         }| {
+            if !is_slow {
+                return;
+            }
+            let elapsed = elapsed_total.as_secs_f64();
+            if ready {
+                warn!("slow {name} completed after {elapsed:.3}s");
+            } else {
+                warn!("slow {name} still running after {elapsed:.3}s");
             }
         },
     )
@@ -416,7 +444,7 @@ where
 pub async fn monitor_slow_future<F, O>(
     threshold: Duration,
     period: Duration,
-    mut fut: std::pin::Pin<&mut F>,
+    mut fut: Pin<&mut F>,
     mut cb: impl FnMut(MonitorSlowFutureCallback),
 ) -> O
 where

libs/wal_decoder/src/models/record.rs

@@ -2,7 +2,8 @@
 
 use bytes::Bytes;
 use postgres_ffi::walrecord::{MultiXactMember, describe_postgres_wal_record};
-use postgres_ffi::{MultiXactId, MultiXactOffset, TimestampTz, TransactionId};
+use postgres_ffi::{MultiXactId, MultiXactOffset, TransactionId};
+use postgres_ffi_types::TimestampTz;
 use serde::{Deserialize, Serialize};
 use utils::bin_ser::DeserializeError;
 

pageserver/client_grpc/src/client.rs

@@ -1,13 +1,16 @@
 use std::collections::HashMap;
 use std::num::NonZero;
+use std::pin::pin;
 use std::sync::Arc;
+use std::time::{Duration, Instant};
 
 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 tracing::{debug, instrument};
+use utils::logging::warn_slow;
 
 use crate::pool::{ChannelPool, ClientGuard, ClientPool, StreamGuard, StreamPool};
 use crate::retry::Retry;
@@ -21,28 +24,40 @@ 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.
 ///
+/// Normal requests are small, and we don't pipeline them, so we can afford a large number of
+/// streams per connection.
+///
 /// 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();
+const MAX_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(64).unwrap();
 
-/// Max number of concurrent unary request clients per shard.
-const MAX_UNARY_CLIENTS: NonZero<usize> = NonZero::new(64).unwrap();
+/// Max number of concurrent bulk GetPage streams per channel (i.e. TCP connection). These use a
+/// dedicated channel pool with a lower client limit, to avoid TCP-level head-of-line blocking and
+/// transmission delays. This also concentrates large window sizes on a smaller set of
+/// streams/connections, presumably reducing memory use.
+const MAX_BULK_CLIENTS_PER_CHANNEL: NonZero<usize> = NonZero::new(16).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();
+/// The batch size threshold at which a GetPage request will use the bulk stream pool.
+///
+/// The gRPC initial window size is 64 KB. Each page is 8 KB, so let's avoid increasing the window
+/// size for the normal stream pool, and route requests for >= 5 pages (>32 KB) to the bulk pool.
+const BULK_THRESHOLD_BATCH_SIZE: usize = 5;
 
-/// Max number of pipelined requests per stream.
-const MAX_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(2).unwrap();
+/// The overall request call timeout, including retries and pool acquisition.
+/// TODO: should we retry forever? Should the caller decide?
+const CALL_TIMEOUT: Duration = Duration::from_secs(60);
 
-/// 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();
+/// The per-request (retry attempt) timeout, including any lazy connection establishment.
+const REQUEST_TIMEOUT: Duration = Duration::from_secs(10);
 
-/// 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();
+/// The initial request retry backoff duration. The first retry does not back off.
+/// TODO: use a different backoff for ResourceExhausted (rate limiting)? Needs server support.
+const BASE_BACKOFF: Duration = Duration::from_millis(5);
+
+/// The maximum request retry backoff duration.
+const MAX_BACKOFF: Duration = Duration::from_secs(5);
+
+/// Threshold and interval for warning about slow operation.
+const SLOW_THRESHOLD: Duration = Duration::from_secs(3);
 
 /// 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:
@@ -50,10 +65,19 @@ const MAX_BULK_STREAM_QUEUE_DEPTH: NonZero<usize> = NonZero::new(4).unwrap();
 /// * 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.
+/// * Internal handling of GetPage bidirectional streams.
 /// * Automatic retries.
 /// * Observability.
 ///
+/// The client has dedicated connection/client/stream pools per shard, for resource reuse. These
+/// pools are unbounded: we allow scaling out as many concurrent streams as needed to serve all
+/// concurrent callers, which mostly eliminates head-of-line blocking. Idle streams are fairly
+/// cheap: the server task currently uses 26 KB of memory, so we can comfortably fit 100,000
+/// concurrent idle streams (2.5 GB memory). The worst case degenerates to the old libpq case with
+/// one stream per backend, but without the TCP connection overhead. In the common case we expect
+/// significantly lower stream counts due to stream sharing, driven e.g. by idle backends, LFC hits,
+/// read coalescing, sharding (backends typically only talk to one shard at a time), etc.
+///
 /// 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 {
@@ -67,8 +91,6 @@ pub struct PageserverClient {
     compression: Option<CompressionEncoding>,
     /// The shards for this tenant.
     shards: ArcSwap<Shards>,
-    /// The retry configuration.
-    retry: Retry,
 }
 
 impl PageserverClient {
@@ -94,7 +116,6 @@ impl PageserverClient {
             auth_token,
             compression,
             shards: ArcSwap::new(Arc::new(shards)),
-            retry: Retry,
         })
     }
 
@@ -142,13 +163,15 @@ impl PageserverClient {
         &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
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.check_rel_exists(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
     }
 
     /// Returns the total size of a database, as # of bytes.
@@ -157,13 +180,15 @@ impl PageserverClient {
         &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
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_db_size(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
     }
 
     /// Fetches pages. The `request_id` must be unique across all in-flight requests, and the
@@ -193,6 +218,8 @@ impl PageserverClient {
             return Err(tonic::Status::invalid_argument("request attempt must be 0"));
         }
 
+        debug!("sending request: {req:?}");
+
         // 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
@@ -201,13 +228,16 @@ impl PageserverClient {
         //
         // 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
+        let resp = Self::with_retries(CALL_TIMEOUT, async |attempt| {
+            let mut req = req.clone();
+            req.request_id.attempt = attempt as u32;
+            let shards = self.shards.load_full();
+            Self::with_timeout(REQUEST_TIMEOUT, Self::get_page_with_shards(req, &shards)).await
+        })
+        .await?;
+
+        debug!("received response: {resp:?}");
+        Ok(resp)
     }
 
     /// Fetches pages using the given shards. This uses a stable view of the shards, regardless of
@@ -246,7 +276,7 @@ impl PageserverClient {
         req: page_api::GetPageRequest,
         shard: &Shard,
     ) -> tonic::Result<page_api::GetPageResponse> {
-        let stream = shard.stream(req.request_class.is_bulk()).await;
+        let mut stream = shard.stream(Self::is_bulk(&req)).await?;
         let resp = stream.send(req.clone()).await?;
 
         // Convert per-request errors into a tonic::Status.
@@ -290,13 +320,15 @@ impl PageserverClient {
         &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
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // Relation metadata is only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_rel_size(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
     }
 
     /// Fetches an SLRU segment.
@@ -305,13 +337,50 @@ impl PageserverClient {
         &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
+        debug!("sending request: {req:?}");
+        let resp = Self::with_retries(CALL_TIMEOUT, async |_| {
+            // SLRU segments are only available on shard 0.
+            let mut client = self.shards.load_full().get_zero().client().await?;
+            Self::with_timeout(REQUEST_TIMEOUT, client.get_slru_segment(req)).await
+        })
+        .await?;
+        debug!("received response: {resp:?}");
+        Ok(resp)
+    }
+
+    /// Runs the given async closure with retries up to the given timeout. Only certain gRPC status
+    /// codes are retried, see [`Retry::should_retry`]. Returns `DeadlineExceeded` on timeout.
+    async fn with_retries<T, F, O>(timeout: Duration, f: F) -> tonic::Result<T>
+    where
+        F: FnMut(usize) -> O, // pass attempt number, starting at 0
+        O: Future<Output = tonic::Result<T>>,
+    {
+        Retry {
+            timeout: Some(timeout),
+            base_backoff: BASE_BACKOFF,
+            max_backoff: MAX_BACKOFF,
+        }
+        .with(f)
+        .await
+    }
+
+    /// Runs the given future with a timeout. Returns `DeadlineExceeded` on timeout.
+    async fn with_timeout<T>(
+        timeout: Duration,
+        f: impl Future<Output = tonic::Result<T>>,
+    ) -> tonic::Result<T> {
+        let started = Instant::now();
+        tokio::time::timeout(timeout, f).await.map_err(|_| {
+            tonic::Status::deadline_exceeded(format!(
+                "request timed out after {:.3}s",
+                started.elapsed().as_secs_f64()
+            ))
+        })?
+    }
+
+    /// Returns true if the request is considered a bulk request and should use the bulk pool.
+    fn is_bulk(req: &page_api::GetPageRequest) -> bool {
+        req.block_numbers.len() >= BULK_THRESHOLD_BATCH_SIZE
     }
 }
 
@@ -440,15 +509,23 @@ impl Shards {
     }
 }
 
-/// A single shard. Uses dedicated resource pools with the following structure:
+/// A single shard. Has 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.
+/// * Channel pool: MAX_CLIENTS_PER_CHANNEL.
+///   * Client pool: unbounded.
+///     * Stream pool: unbounded.
+/// * Bulk channel pool: MAX_BULK_CLIENTS_PER_CHANNEL.
 ///   * Bulk client pool: unbounded.
-///     * Bulk stream pool: MAX_BULK_STREAMS and MAX_BULK_STREAM_QUEUE_DEPTH.
+///     * Bulk stream pool: unbounded.
+///
+/// We use a separate bulk channel pool with a lower concurrency limit for large batch requests.
+/// This avoids TCP-level head-of-line blocking, and also concentrates large window sizes on a
+/// smaller set of streams/connections, which presumably reduces memory use. Neither of these pools
+/// are bounded, nor do they pipeline requests, so the latency characteristics should be mostly
+/// similar (except for TCP transmission time).
+///
+/// TODO: since we never use bounded pools, we could consider removing the pool limiters. However,
+/// the code is fairly trivial, so we may as well keep them around for now in case we need them.
 struct Shard {
     /// The shard ID.
     id: ShardIndex,
@@ -456,7 +533,7 @@ struct Shard {
     client_pool: Arc<ClientPool>,
     /// GetPage stream pool.
     stream_pool: Arc<StreamPool>,
-    /// GetPage stream pool for bulk requests, e.g. prefetches.
+    /// GetPage stream pool for bulk requests.
     bulk_stream_pool: Arc<StreamPool>,
 }
 
@@ -470,50 +547,30 @@ impl Shard {
         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.
+        // Shard pools for unary requests and non-bulk GetPage requests.
         let client_pool = ClientPool::new(
-            channel_pool.clone(),
+            ChannelPool::new(url.clone(), MAX_CLIENTS_PER_CHANNEL)?,
             tenant_id,
             timeline_id,
             shard_id,
             auth_token.clone(),
             compression,
-            Some(MAX_UNARY_CLIENTS),
+            None, // unbounded
         );
+        let stream_pool = StreamPool::new(client_pool.clone(), None); // unbounded
 
-        // 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.
+        // Bulk GetPage stream pool for large batches (prefetches, sequential scans, vacuum, etc.).
         let bulk_stream_pool = StreamPool::new(
             ClientPool::new(
-                ChannelPool::new(url, MAX_CLIENTS_PER_CHANNEL)?,
+                ChannelPool::new(url, MAX_BULK_CLIENTS_PER_CHANNEL)?,
                 tenant_id,
                 timeline_id,
                 shard_id,
                 auth_token,
                 compression,
-                None, // unbounded, limited by stream pool
+                None, // unbounded,
             ),
-            Some(MAX_BULK_STREAMS),
-            MAX_BULK_STREAM_QUEUE_DEPTH,
+            None, // unbounded
         );
 
         Ok(Self {
@@ -525,19 +582,23 @@ impl Shard {
     }
 
     /// Returns a pooled client for this shard.
+    #[instrument(skip_all)]
     async fn client(&self) -> tonic::Result<ClientGuard> {
-        self.client_pool
-            .get()
-            .await
-            .map_err(|err| tonic::Status::internal(format!("failed to get client: {err}")))
+        warn_slow(
+            "client pool acquisition",
+            SLOW_THRESHOLD,
+            pin!(self.client_pool.get()),
+        )
+        .await
     }
 
-    /// 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,
-        }
+    /// Returns a pooled stream for this shard. If `bulk` is `true`, uses the dedicated bulk pool.
+    #[instrument(skip_all, fields(bulk))]
+    async fn stream(&self, bulk: bool) -> tonic::Result<StreamGuard> {
+        let pool = match bulk {
+            false => &self.stream_pool,
+            true => &self.bulk_stream_pool,
+        };
+        warn_slow("stream pool acquisition", SLOW_THRESHOLD, pin!(pool.get())).await
     }
 }

pageserver/client_grpc/src/pool.rs

@@ -9,19 +9,36 @@
 //!
 //! * 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.
+//!   per-channel client limit. Channels are closed immediately when empty, and indirectly rely on
+//!   client/stream idle timeouts.
 //!
 //! * 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.
+//!   single caller at a time, and is returned to the pool when dropped. Idle clients are removed
+//!   from the pool after a while to free up resources.
 //!
 //! * 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.
+//!   ClientPool for the stream's lifetime. A stream can only be acquired by a single caller at a
+//!   time, and is returned to the pool when dropped. Idle streams are removed from the pool after
+//!   a while to free up resources.
+//!
+//!   The stream only supports sending a single, synchronous request at a time, and does not support
+//!   pipelining multiple requests from different callers onto the same stream -- instead, we scale
+//!   out concurrent streams to improve throughput. There are many reasons for this design choice:
+//!
+//!     * It (mostly) eliminates head-of-line blocking. A single stream is processed sequentially by
+//!       a single server task, which may block e.g. on layer downloads, LSN waits, etc.
+//!
+//!     * Cancellation becomes trivial, by closing the stream. Otherwise, if a caller goes away
+//!       (e.g. because of a timeout), the request would still be processed by the server and block
+//!       requests behind it in the stream. It might even block its own timeout retry.
+//!
+//!     * Stream scheduling becomes significantly simpler and cheaper.
+//!
+//!     * Individual callers can still use client-side batching for pipelining.
+//!
+//!     * Idle streams are cheap. Benchmarks show that an idle GetPage stream takes up about 26 KB
+//!       per stream (2.5 GB for 100,000 streams), so we can afford to scale out.
 //!
 //! Each channel corresponds to one TCP connection. Each client unary request and each stream
 //! corresponds to one HTTP/2 stream and server task.
@@ -29,33 +46,31 @@
 //! TODO: error handling (including custom error types).
 //! TODO: observability.
 
-use std::collections::{BTreeMap, HashMap};
+use std::collections::BTreeMap;
 use std::num::NonZero;
 use std::ops::{Deref, DerefMut};
+use std::pin::Pin;
 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 futures::{Stream, StreamExt as _};
+use tokio::sync::{OwnedSemaphorePermit, Semaphore, watch};
+use tokio_stream::wrappers::WatchStream;
 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.
+/// Reap clients/streams that have been idle for this long. Channels are reaped immediately when
+/// empty, and indirectly rely on the client/stream idle timeouts.
 ///
-/// 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.
+/// A stream's client will be reaped after 2x the idle threshold (first stream the client), but
+/// that's okay -- if the stream closes abruptly (e.g. due to timeout or cancellation), we want to
+/// keep its client around in the pool for a while.
 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
@@ -83,8 +98,6 @@ pub struct ChannelPool {
     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,
 }
@@ -96,9 +109,6 @@ struct ChannelEntry {
     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 {
@@ -108,15 +118,12 @@ impl ChannelPool {
         E: TryInto<Endpoint> + Send + Sync + 'static,
         <E as TryInto<Endpoint>>::Error: std::error::Error + Send + Sync,
     {
-        let pool = Arc::new(Self {
+        Ok(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.
@@ -137,22 +144,17 @@ impl ChannelPool {
         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
+        // order, to fill up the lower-ordered channels first. The client/stream pools also prefer
+        // 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"
-            );
+            assert_ne!(entry.clients, 0, "empty channel not reaped");
             if entry.clients < self.max_clients_per_channel.get() {
                 entry.clients += 1;
-                entry.idle_since = None;
                 return ChannelGuard {
                     pool: Arc::downgrade(self),
                     id,
@@ -169,7 +171,6 @@ impl ChannelPool {
         let entry = ChannelEntry {
             channel: channel.clone(),
             clients: 1, // account for the guard below
-            idle_since: None,
         };
         channels.insert(id, entry);
 
@@ -181,20 +182,6 @@ impl ChannelPool {
     }
 }
 
-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 {
@@ -211,7 +198,7 @@ impl ChannelGuard {
     }
 }
 
-/// Returns the channel to the pool.
+/// Returns the channel to the pool. The channel is closed when empty.
 impl Drop for ChannelGuard {
     fn drop(&mut self) {
         let Some(pool) = self.pool.upgrade() else {
@@ -220,11 +207,12 @@ impl Drop for ChannelGuard {
 
         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;
+
+        // Reap empty channels immediately.
         if entry.clients == 0 {
-            entry.idle_since = Some(Instant::now()); // mark channel as idle
+            channels.remove(&self.id);
         }
     }
 }
@@ -253,8 +241,7 @@ pub struct ClientPool {
     ///
     /// 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.
+    /// lower-ordered channels. This allows us to free up and reap higher-ordered channels.
     idle: Mutex<BTreeMap<ClientID, ClientEntry>>,
     /// Reaps idle clients.
     idle_reaper: Reaper,
@@ -310,7 +297,7 @@ impl ClientPool {
     /// 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> {
+    pub async fn get(self: &Arc<Self>) -> tonic::Result<ClientGuard> {
         // Acquire a permit if the pool is bounded.
         let mut permit = None;
         if let Some(limiter) = self.limiter.clone() {
@@ -328,7 +315,7 @@ impl ClientPool {
             });
         }
 
-        // Slow path: construct a new client.
+        // Construct a new client.
         let mut channel_guard = self.channel_pool.get();
         let client = page_api::Client::new(
             channel_guard.take(),
@@ -337,7 +324,8 @@ impl ClientPool {
             self.shard_id,
             self.auth_token.clone(),
             self.compression,
-        )?;
+        )
+        .map_err(|err| tonic::Status::internal(format!("failed to create client: {err}")))?;
 
         Ok(ClientGuard {
             pool: Arc::downgrade(self),
@@ -407,287 +395,187 @@ impl Drop for ClientGuard {
 /// 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.
+/// Individual streams only send a single request at a time, and do not pipeline multiple callers
+/// onto the same stream. Instead, we scale out the number of concurrent streams. This is primarily
+/// to eliminate head-of-line blocking. See the module documentation for more details.
 ///
 /// 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.
+    /// Idle pooled streams. Acquired streams are removed from here and returned on drop.
     ///
-    /// 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.
+    /// The first stream in the map will be acquired next. The map is sorted by stream ID, which is
+    /// equivalent to the client ID and in turn sorted by its channel ID. This way we prefer
+    /// acquiring idle streams from lower-ordered channels, which allows us to free up and reap
+    /// higher-ordered channels.
+    idle: Mutex<BTreeMap<StreamID, StreamEntry>>,
+    /// Limits the max number of concurrent streams. 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>>;
+/// The stream ID. Reuses the inner client ID.
+type StreamID = ClientID;
 
+/// A pooled stream.
 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>,
+    /// The bidirectional stream.
+    stream: BiStream,
+    /// The time when this stream was last used, i.e. when it was put back into `StreamPool::idle`.
+    idle_since: Instant,
+}
+
+/// A bidirectional GetPage stream and its client. Can send requests and receive responses.
+struct BiStream {
+    /// The owning client. Holds onto the channel slot while the stream is alive.
+    client: ClientGuard,
+    /// Stream for sending requests. Uses a watch channel, so it can only send a single request at a
+    /// time, and the caller must await the response before sending another request. This is
+    /// enforced by `StreamGuard::send`.
+    sender: watch::Sender<page_api::GetPageRequest>,
+    /// Stream for receiving responses.
+    receiver: Pin<Box<dyn Stream<Item = tonic::Result<page_api::GetPageResponse>> + Send>>,
 }
 
 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.
+    /// Creates a new stream pool, using the given client pool. It will 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> {
+    pub fn new(client_pool: Arc<ClientPool>, max_streams: Option<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: Mutex::default(),
+            limiter: max_streams.map(|max_streams| Arc::new(Semaphore::new(max_streams.get()))),
             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).
+    /// Acquires an available stream from the pool, or spins up a new stream if all streams are
+    /// full. Returns a guard that can be used to send requests and await the responses. Blocks if
+    /// the pool is full.
     ///
     /// 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 {
+    /// TODO: is a `Mutex<BTreeMap>` performant enough? Will it become too contended? We can't
+    /// trivially use e.g. DashMap or sharding, because we want to pop lower-ordered streams first
+    /// to free up higher-ordered channels.
+    pub async fn get(self: &Arc<Self>) -> tonic::Result<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,
-                };
-            }
+        // Fast path: acquire an idle stream from the pool.
+        if let Some((_, entry)) = self.idle.lock().unwrap().pop_first() {
+            return Ok(StreamGuard {
+                pool: Arc::downgrade(self),
+                stream: Some(entry.stream),
+                can_reuse: true,
+                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);
+        // Spin up a new stream. Uses a watch channel to send a single request at a time, since
+        // `StreamGuard::send` enforces this anyway and it avoids unnecessary channel overhead.
+        let mut client = self.client_pool.get().await?;
 
-        if let Some(max_streams) = self.max_streams {
-            assert!(streams.len() <= max_streams.get(), "stream overflow");
-        };
+        let (req_tx, req_rx) = watch::channel(page_api::GetPageRequest::default());
+        let req_stream = WatchStream::from_changes(req_rx);
+        let resp_stream = client.get_pages(req_stream).await?;
 
-        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 {
+        Ok(StreamGuard {
             pool: Arc::downgrade(self),
-            id,
-            sender: req_tx,
+            stream: Some(BiStream {
+                client,
+                sender: req_tx,
+                receiver: Box::pin(resp_stream),
+            }),
+            can_reuse: true,
             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
-        });
+        self.idle
+            .lock()
+            .unwrap()
+            .retain(|_, entry| entry.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.
+/// A stream acquired from the pool. Returned to the pool when dropped, unless there are still
+/// in-flight requests on the stream, or the stream failed.
 pub struct StreamGuard {
     pool: Weak<StreamPool>,
-    id: StreamID,
-    sender: RequestSender,
+    stream: Option<BiStream>,             // Some until dropped
+    can_reuse: bool,                      // returned to pool if true
     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.
+    /// Sends a request on the stream and awaits the response. If the future is dropped before it
+    /// resolves (e.g. due to a timeout or cancellation), the stream will be closed to cancel the
+    /// request and is not returned to the pool. The same is true if the stream errors, in which
+    /// case the caller can't send further requests on the stream.
     ///
-    /// The `GetPageRequest::request_id` must be unique across in-flight requests.
+    /// We only support sending a single request at a time, to eliminate head-of-line blocking. See
+    /// module documentation for details.
     ///
     /// 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,
+        &mut self,
         req: page_api::GetPageRequest,
     ) -> tonic::Result<page_api::GetPageResponse> {
-        let (resp_tx, resp_rx) = oneshot::channel();
+        let req_id = req.request_id;
+        let stream = self.stream.as_mut().expect("not dropped");
 
-        self.sender
-            .send((req, resp_tx))
-            .await
+        // Mark the stream as not reusable while the request is in flight. We can't return the
+        // stream to the pool until we receive the response, to avoid head-of-line blocking and
+        // stale responses. Failed streams can't be reused either.
+        if !self.can_reuse {
+            return Err(tonic::Status::internal("stream can't be reused"));
+        }
+        self.can_reuse = false;
+
+        // Send the request and receive the response.
+        //
+        // NB: this uses a watch channel, so it's unsafe to change this code to pipeline requests.
+        stream
+            .sender
+            .send(req)
             .map_err(|_| tonic::Status::unavailable("stream closed"))?;
 
-        resp_rx
+        let resp = stream
+            .receiver
+            .next()
             .await
-            .map_err(|_| tonic::Status::unavailable("stream closed"))?
+            .ok_or_else(|| tonic::Status::unavailable("stream closed"))??;
+
+        if resp.request_id != req_id {
+            return Err(tonic::Status::internal(format!(
+                "response ID {} does not match request ID {}",
+                resp.request_id, req_id
+            )));
+        }
+
+        // Success, mark the stream as reusable.
+        self.can_reuse = true;
+
+        Ok(resp)
     }
 }
 
@@ -697,26 +585,21 @@ impl Drop for StreamGuard {
             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
+        // If the stream isn't reusable, it can't be returned to the pool.
+        if !self.can_reuse {
+            return;
         }
 
+        // Place the idle stream back into the pool.
+        let entry = StreamEntry {
+            stream: self.stream.take().expect("dropped once"),
+            idle_since: Instant::now(),
+        };
+        pool.idle
+            .lock()
+            .unwrap()
+            .insert(entry.stream.client.id, entry);
+
         _ = self.permit; // returned on drop, referenced for visibility
     }
 }

pageserver/client_grpc/src/retry.rs

@@ -1,5 +1,6 @@
 use std::time::Duration;
 
+use futures::future::pending;
 use tokio::time::Instant;
 use tracing::{error, info, warn};
 
@@ -8,60 +9,54 @@ 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;
+pub struct Retry {
+    /// Timeout across all retry attempts. If None, retries forever.
+    pub timeout: Option<Duration>,
+    /// The initial backoff duration. The first retry does not use a backoff.
+    pub base_backoff: Duration,
+    /// The maximum backoff duration.
+    pub max_backoff: Duration,
+}
 
 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.
+    /// Runs the given async closure with timeouts and retries (exponential backoff). Logs errors,
+    /// using the current tracing span for context.
     ///
-    /// Only certain gRPC status codes are retried, see [`Self::should_retry`]. For default
-    /// timeouts, see [`Self::REQUEST_TIMEOUT`] and [`Self::TOTAL_TIMEOUT`].
+    /// Only certain gRPC status codes are retried, see [`Self::should_retry`].
     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
+        F: FnMut(usize) -> O, // pass attempt number, starting at 0
         O: Future<Output = tonic::Result<T>>,
     {
         let started = Instant::now();
-        let deadline = started + Self::TOTAL_TIMEOUT;
+        let deadline = self.timeout.map(|timeout| started + 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.
+            // Set up a future to wait for the backoff, if any, and run the closure.
             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) {
+                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()
-                        ))
-                    })?
+                f(retries).await
             };
 
-            // Wait for the backoff and request, or bail out if the total timeout is exceeded.
+            // Set up a future for the timeout, if any.
+            let timeout = async {
+                match deadline {
+                    Some(deadline) => tokio::time::sleep_until(deadline).await,
+                    None => pending().await,
+                }
+            };
+
+            // Wait for the backoff and request, or bail out if the timeout is exceeded.
             let result = tokio::select! {
                 result = backoff_and_try => result,
 
-                _ = tokio::time::sleep_until(deadline) => {
+                _ = timeout => {
                     let last_error = last_error.unwrap_or_else(|| {
                         tonic::Status::deadline_exceeded(format!(
                             "request timed out after {:.3}s",
@@ -79,7 +74,7 @@ impl Retry {
             match result {
                 // Success, return the result.
                 Ok(result) => {
-                    if retries > 0 || Self::LOG_SUCCESS {
+                    if retries > 0 {
                         info!(
                             "request succeeded after {retries} retries in {:.3}s",
                             started.elapsed().as_secs_f64(),
@@ -112,12 +107,13 @@ impl Retry {
         }
     }
 
-    /// Returns the backoff duration for the given retry attempt, or None for no backoff.
-    fn backoff_duration(retry: usize) -> Option<Duration> {
+    /// Returns the backoff duration for the given retry attempt, or None for no backoff. The first
+    /// attempt and first retry never backs off, so this returns None for 0 and 1 retries.
+    fn backoff_duration(&self, retries: usize) -> Option<Duration> {
         let backoff = exponential_backoff_duration(
-            retry as u32,
-            Self::BASE_BACKOFF.as_secs_f64(),
-            Self::MAX_BACKOFF.as_secs_f64(),
+            (retries as u32).saturating_sub(1), // first retry does not back off
+            self.base_backoff.as_secs_f64(),
+            self.max_backoff.as_secs_f64(),
         );
         (!backoff.is_zero()).then_some(backoff)
     }

pageserver/page_api/src/model.rs

@@ -49,7 +49,7 @@ impl From<ProtocolError> for tonic::Status {
 }
 
 /// The LSN a request should read at.
-#[derive(Clone, Copy, Debug)]
+#[derive(Clone, Copy, Debug, Default)]
 pub struct ReadLsn {
     /// The request's read LSN.
     pub request_lsn: Lsn,
@@ -329,7 +329,7 @@ impl From<GetDbSizeResponse> for proto::GetDbSizeResponse {
 }
 
 /// Requests one or more pages.
-#[derive(Clone, Debug)]
+#[derive(Clone, Debug, Default)]
 pub struct GetPageRequest {
     /// A request ID. Will be included in the response. Should be unique for in-flight requests on
     /// the stream.
@@ -430,12 +430,13 @@ impl From<RequestID> for proto::RequestId {
 }
 
 /// A GetPage request class.
-#[derive(Clone, Copy, Debug, strum_macros::Display)]
+#[derive(Clone, Copy, Debug, Default, 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.
     Unknown,
     /// A normal request. This is the default.
+    #[default]
     Normal,
     /// A prefetch request. NB: can only be classified on pg < 18.
     Prefetch,
@@ -443,19 +444,6 @@ 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 {

pageserver/pagebench/Cargo.toml

@@ -16,6 +16,7 @@ futures.workspace = true
 hdrhistogram.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true
+pprof.workspace = true
 rand.workspace = true
 reqwest.workspace = true
 serde.workspace = true

pageserver/pagebench/src/cmd/idle_streams.rs

@@ -0,0 +1,127 @@
+use std::sync::Arc;
+
+use anyhow::anyhow;
+use futures::StreamExt;
+use tonic::transport::Endpoint;
+use tracing::info;
+
+use pageserver_page_api::{GetPageClass, GetPageRequest, GetPageStatusCode, ReadLsn, RelTag};
+use utils::id::TenantTimelineId;
+use utils::lsn::Lsn;
+use utils::shard::ShardIndex;
+
+/// Starts a large number of idle gRPC GetPage streams.
+#[derive(clap::Parser)]
+pub(crate) struct Args {
+    /// The Pageserver to connect to. Must use grpc://.
+    #[clap(long, default_value = "grpc://localhost:51051")]
+    server: String,
+    /// The Pageserver HTTP API.
+    #[clap(long, default_value = "http://localhost:9898")]
+    http_server: String,
+    /// The number of streams to open.
+    #[clap(long, default_value = "100000")]
+    count: usize,
+    /// Number of streams per connection.
+    #[clap(long, default_value = "100")]
+    per_connection: usize,
+    /// Send a single GetPage request on each stream.
+    #[clap(long, default_value_t = false)]
+    send_request: bool,
+}
+
+pub(crate) fn main(args: Args) -> anyhow::Result<()> {
+    let rt = tokio::runtime::Builder::new_multi_thread()
+        .enable_all()
+        .build()?;
+
+    rt.block_on(main_impl(args))
+}
+
+async fn main_impl(args: Args) -> anyhow::Result<()> {
+    // Discover a tenant and timeline to use.
+    let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
+        reqwest::Client::new(),
+        args.http_server.clone(),
+        None,
+    ));
+    let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
+        &mgmt_api_client,
+        crate::util::cli::targets::Spec {
+            limit_to_first_n_targets: Some(1),
+            targets: None,
+        },
+    )
+    .await?;
+    let ttid = timelines
+        .first()
+        .ok_or_else(|| anyhow!("no timelines found"))?;
+
+    // Set up the initial client.
+    let endpoint = Endpoint::from_shared(args.server.clone())?;
+
+    let connect = async || {
+        pageserver_page_api::Client::new(
+            endpoint.connect().await?,
+            ttid.tenant_id,
+            ttid.timeline_id,
+            ShardIndex::unsharded(),
+            None,
+            None,
+        )
+    };
+
+    let mut client = connect().await?;
+    let mut streams = Vec::with_capacity(args.count);
+
+    // Create streams.
+    for i in 0..args.count {
+        if i % 100 == 0 {
+            info!("opened {}/{} streams", i, args.count);
+        }
+        if i % args.per_connection == 0 && i > 0 {
+            client = connect().await?;
+        }
+
+        let (req_tx, req_rx) = tokio::sync::mpsc::unbounded_channel();
+        let req_stream = tokio_stream::wrappers::UnboundedReceiverStream::new(req_rx);
+        let mut resp_stream = client.get_pages(req_stream).await?;
+
+        // Send request if specified.
+        if args.send_request {
+            req_tx.send(GetPageRequest {
+                request_id: 1.into(),
+                request_class: GetPageClass::Normal,
+                read_lsn: ReadLsn {
+                    request_lsn: Lsn::MAX,
+                    not_modified_since_lsn: Some(Lsn(1)),
+                },
+                rel: RelTag {
+                    spcnode: 1664, // pg_global
+                    dbnode: 0,     // shared database
+                    relnode: 1262, // pg_authid
+                    forknum: 0,    // init
+                },
+                block_numbers: vec![0],
+            })?;
+            let resp = resp_stream
+                .next()
+                .await
+                .transpose()?
+                .ok_or_else(|| anyhow!("no response"))?;
+            if resp.status_code != GetPageStatusCode::Ok {
+                return Err(anyhow!("{} response", resp.status_code));
+            }
+        }
+
+        // Hold onto streams to avoid closing them.
+        streams.push((req_tx, resp_stream));
+    }
+
+    info!("opened {} streams, sleeping", args.count);
+
+    // Block forever, to hold the idle streams open for inspection.
+    futures::future::pending::<()>().await;
+
+    Ok(())
+}

pageserver/pagebench/src/main.rs

@@ -1,4 +1,7 @@
+use std::fs::File;
+
 use clap::Parser;
+use tracing::info;
 use utils::logging;
 
 /// Re-usable pieces of code that aren't CLI-specific.
@@ -17,38 +20,73 @@ mod cmd {
     pub(super) mod aux_files;
     pub(super) mod basebackup;
     pub(super) mod getpage_latest_lsn;
+    pub(super) mod idle_streams;
     pub(super) mod ondemand_download_churn;
     pub(super) mod trigger_initial_size_calculation;
 }
 
 /// Component-level performance test for pageserver.
 #[derive(clap::Parser)]
-enum Args {
+struct Args {
+    /// Takes a client CPU profile into profile.svg. The benchmark must exit cleanly before it's
+    /// written, e.g. via --runtime.
+    #[arg(long)]
+    profile: bool,
+
+    #[command(subcommand)]
+    subcommand: Subcommand,
+}
+
+#[derive(clap::Subcommand)]
+enum Subcommand {
     Basebackup(cmd::basebackup::Args),
     GetPageLatestLsn(cmd::getpage_latest_lsn::Args),
     TriggerInitialSizeCalculation(cmd::trigger_initial_size_calculation::Args),
     OndemandDownloadChurn(cmd::ondemand_download_churn::Args),
     AuxFiles(cmd::aux_files::Args),
+    IdleStreams(cmd::idle_streams::Args),
 }
 
-fn main() {
+fn main() -> anyhow::Result<()> {
     logging::init(
         logging::LogFormat::Plain,
         logging::TracingErrorLayerEnablement::Disabled,
         logging::Output::Stderr,
-    )
-    .unwrap();
+    )?;
     logging::replace_panic_hook_with_tracing_panic_hook().forget();
 
     let args = Args::parse();
-    match args {
-        Args::Basebackup(args) => cmd::basebackup::main(args),
-        Args::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
-        Args::TriggerInitialSizeCalculation(args) => {
+
+    // Start a CPU profile if requested.
+    let mut profiler = None;
+    if args.profile {
+        profiler = Some(
+            pprof::ProfilerGuardBuilder::default()
+                .frequency(1000)
+                .blocklist(&["libc", "libgcc", "pthread", "vdso"])
+                .build()?,
+        );
+    }
+
+    match args.subcommand {
+        Subcommand::Basebackup(args) => cmd::basebackup::main(args),
+        Subcommand::GetPageLatestLsn(args) => cmd::getpage_latest_lsn::main(args),
+        Subcommand::TriggerInitialSizeCalculation(args) => {
             cmd::trigger_initial_size_calculation::main(args)
         }
-        Args::OndemandDownloadChurn(args) => cmd::ondemand_download_churn::main(args),
-        Args::AuxFiles(args) => cmd::aux_files::main(args),
+        Subcommand::OndemandDownloadChurn(args) => cmd::ondemand_download_churn::main(args),
+        Subcommand::AuxFiles(args) => cmd::aux_files::main(args),
+        Subcommand::IdleStreams(args) => cmd::idle_streams::main(args),
+    }?;
+
+    // Generate a CPU flamegraph if requested.
+    if let Some(profiler) = profiler {
+        let report = profiler.report().build()?;
+        drop(profiler); // stop profiling
+        let file = File::create("profile.svg")?;
+        report.flamegraph(file)?;
+        info!("wrote CPU profile flamegraph to profile.svg")
     }
-    .unwrap()
+
+    Ok(())
 }

pageserver/src/basebackup.rs

@@ -114,7 +114,7 @@ where
     // Compute postgres doesn't have any previous WAL files, but the first
     // record that it's going to write needs to include the LSN of the
     // previous record (xl_prev). We include prev_record_lsn in the
-    // "zenith.signal" file, so that postgres can read it during startup.
+    // "neon.signal" file, so that postgres can read it during startup.
     //
     // We don't keep full history of record boundaries in the page server,
     // however, only the predecessor of the latest record on each
@@ -751,34 +751,39 @@ where
 
     //
     // Add generated pg_control file and bootstrap WAL segment.
-    // Also send zenith.signal file with extra bootstrap data.
+    // Also send neon.signal and zenith.signal file with extra bootstrap data.
     //
     async fn add_pgcontrol_file(
         &mut self,
         pg_control_bytes: Bytes,
         system_identifier: u64,
     ) -> Result<(), BasebackupError> {
-        // add zenith.signal file
-        let mut zenith_signal = String::new();
+        // add neon.signal file
+        let mut neon_signal = String::new();
         if self.prev_record_lsn == Lsn(0) {
             if self.timeline.is_ancestor_lsn(self.lsn) {
-                write!(zenith_signal, "PREV LSN: none")
+                write!(neon_signal, "PREV LSN: none")
                     .map_err(|e| BasebackupError::Server(e.into()))?;
             } else {
-                write!(zenith_signal, "PREV LSN: invalid")
+                write!(neon_signal, "PREV LSN: invalid")
                     .map_err(|e| BasebackupError::Server(e.into()))?;
             }
         } else {
-            write!(zenith_signal, "PREV LSN: {}", self.prev_record_lsn)
+            write!(neon_signal, "PREV LSN: {}", self.prev_record_lsn)
                 .map_err(|e| BasebackupError::Server(e.into()))?;
         }
-        self.ar
-            .append(
-                &new_tar_header("zenith.signal", zenith_signal.len() as u64)?,
-                zenith_signal.as_bytes(),
-            )
-            .await
-            .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,zenith.signal"))?;
+
+        // TODO: Remove zenith.signal once all historical computes have been replaced
+        // ... and thus support the neon.signal file.
+        for signalfilename in ["neon.signal", "zenith.signal"] {
+            self.ar
+                .append(
+                    &new_tar_header(signalfilename, neon_signal.len() as u64)?,
+                    neon_signal.as_bytes(),
+                )
+                .await
+                .map_err(|e| BasebackupError::Client(e, "add_pgcontrol_file,neon.signal"))?;
+        }
 
         //send pg_control
         let header = new_tar_header("global/pg_control", pg_control_bytes.len() as u64)?;

pageserver/src/bin/pageserver.rs

@@ -917,11 +917,6 @@ async fn create_remote_storage_client(
     // If `test_remote_failures` is non-zero, wrap the client with a
     // wrapper that simulates failures.
     if conf.test_remote_failures > 0 {
-        if !cfg!(feature = "testing") {
-            anyhow::bail!(
-                "test_remote_failures option is not available because pageserver was compiled without the 'testing' feature"
-            );
-        }
         info!(
             "Simulating remote failures for first {} attempts of each op",
             conf.test_remote_failures

pageserver/src/controller_upcall_client.rs

@@ -194,6 +194,7 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
                         listen_http_port: m.http_port,
                         listen_https_port: m.https_port,
                         availability_zone_id: az_id.expect("Checked above"),
+                        node_ip_addr: None,
                     })
                 }
                 Err(e) => {

pageserver/src/deletion_queue/validator.rs

@@ -1,5 +1,5 @@
 //! The validator is responsible for validating DeletionLists for execution,
-//! based on whethe the generation in the DeletionList is still the latest
+//! based on whether the generation in the DeletionList is still the latest
 //! generation for a tenant.
 //!
 //! The purpose of validation is to ensure split-brain safety in the cluster

pageserver/src/http/routes.rs

@@ -10,6 +10,7 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use anyhow::{Context, Result, anyhow};
+use bytes::Bytes;
 use enumset::EnumSet;
 use futures::future::join_all;
 use futures::{StreamExt, TryFutureExt};
@@ -46,6 +47,7 @@ use pageserver_api::shard::{ShardCount, TenantShardId};
 use postgres_ffi::PgMajorVersion;
 use remote_storage::{DownloadError, GenericRemoteStorage, TimeTravelError};
 use scopeguard::defer;
+use serde::{Deserialize, Serialize};
 use serde_json::json;
 use tenant_size_model::svg::SvgBranchKind;
 use tenant_size_model::{SizeResult, StorageModel};
@@ -57,6 +59,7 @@ use utils::auth::SwappableJwtAuth;
 use utils::generation::Generation;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
+use wal_decoder::models::record::NeonWalRecord;
 
 use crate::config::PageServerConf;
 use crate::context;
@@ -77,6 +80,7 @@ use crate::tenant::remote_timeline_client::{
 };
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
+use crate::tenant::storage_layer::ValuesReconstructState;
 use crate::tenant::storage_layer::{IoConcurrency, LayerAccessStatsReset, LayerName};
 use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::timeline::offload::{OffloadError, offload_timeline};
@@ -397,6 +401,7 @@ async fn build_timeline_info(
     timeline: &Arc<Timeline>,
     include_non_incremental_logical_size: bool,
     force_await_initial_logical_size: bool,
+    include_image_consistent_lsn: bool,
     ctx: &RequestContext,
 ) -> anyhow::Result<TimelineInfo> {
     crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id();
@@ -421,6 +426,10 @@ async fn build_timeline_info(
                 .await?,
         );
     }
+    // HADRON
+    if include_image_consistent_lsn {
+        info.image_consistent_lsn = Some(timeline.compute_image_consistent_lsn().await?);
+    }
     Ok(info)
 }
 
@@ -510,6 +519,8 @@ async fn build_timeline_info_common(
         is_invisible: Some(is_invisible),
 
         walreceiver_status,
+        // HADRON
+        image_consistent_lsn: None,
     };
     Ok(info)
 }
@@ -712,6 +723,8 @@ async fn timeline_list_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let state = get_state(&request);
@@ -732,6 +745,7 @@ async fn timeline_list_handler(
                 &timeline,
                 include_non_incremental_logical_size.unwrap_or(false),
                 force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                 &ctx,
             )
             .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -760,6 +774,9 @@ async fn timeline_and_offloaded_list_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
+
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let state = get_state(&request);
@@ -780,6 +797,7 @@ async fn timeline_and_offloaded_list_handler(
                 &timeline,
                 include_non_incremental_logical_size.unwrap_or(false),
                 force_await_initial_logical_size.unwrap_or(false),
+                include_image_consistent_lsn.unwrap_or(false),
                 &ctx,
             )
             .instrument(info_span!("build_timeline_info", timeline_id = %timeline.timeline_id))
@@ -964,6 +982,9 @@ async fn timeline_detail_handler(
         parse_query_param(&request, "include-non-incremental-logical-size")?;
     let force_await_initial_logical_size: Option<bool> =
         parse_query_param(&request, "force-await-initial-logical-size")?;
+    // HADRON
+    let include_image_consistent_lsn: Option<bool> =
+        parse_query_param(&request, "include-image-consistent-lsn")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     // Logical size calculation needs downloading.
@@ -984,6 +1005,7 @@ async fn timeline_detail_handler(
             &timeline,
             include_non_incremental_logical_size.unwrap_or(false),
             force_await_initial_logical_size.unwrap_or(false),
+            include_image_consistent_lsn.unwrap_or(false),
             ctx,
         )
         .await
@@ -2690,6 +2712,16 @@ async fn deletion_queue_flush(
     }
 }
 
+/// Try if `GetPage@Lsn` is successful, useful for manual debugging.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+struct GetPageResponse {
+    pub page: Bytes,
+    pub layers_visited: u32,
+    pub delta_layers_visited: u32,
+    pub records: Vec<(Lsn, NeonWalRecord)>,
+    pub img: Option<(Lsn, Bytes)>,
+}
+
 async fn getpage_at_lsn_handler(
     request: Request<Body>,
     cancel: CancellationToken,
@@ -2740,21 +2772,24 @@ async fn getpage_at_lsn_handler_inner(
 
         // Use last_record_lsn if no lsn is provided
         let lsn = lsn.unwrap_or_else(|| timeline.get_last_record_lsn());
-        let page = timeline.get(key.0, lsn, &ctx).await?;
 
         if touch {
             json_response(StatusCode::OK, ())
         } else {
-            Result::<_, ApiError>::Ok(
-                Response::builder()
-                    .status(StatusCode::OK)
-                    .header(header::CONTENT_TYPE, "application/octet-stream")
-                    .body(hyper::Body::from(page))
-                    .unwrap(),
-            )
+            let mut reconstruct_state = ValuesReconstructState::new_with_debug(IoConcurrency::sequential());
+            let page = timeline.debug_get(key.0, lsn, &ctx, &mut reconstruct_state).await?;
+            let response = GetPageResponse {
+                page,
+                layers_visited: reconstruct_state.get_layers_visited(),
+                delta_layers_visited: reconstruct_state.get_delta_layers_visited(),
+                records: reconstruct_state.debug_state.records.clone(),
+                img: reconstruct_state.debug_state.img.clone(),
+            };
+
+            json_response(StatusCode::OK, response)
         }
     }
-    .instrument(info_span!("timeline_get", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
+    .instrument(info_span!("timeline_debug_get", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
     .await
 }
 
@@ -3643,6 +3678,7 @@ async fn activate_post_import_handler(
         let timeline_info = build_timeline_info(
             &timeline, false, // include_non_incremental_logical_size,
             false, // force_await_initial_logical_size
+            false, // include_image_consistent_lsn
             &ctx,
         )
         .await
@@ -4164,7 +4200,7 @@ pub fn make_router(
         })
         .get(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/getpage",
-            |r| testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler),
+            |r|  testing_api_handler("getpage@lsn", r, getpage_at_lsn_handler),
         )
         .get(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/touchpage",

pageserver/src/import_datadir.rs

@@ -610,13 +610,13 @@ async fn import_file(
         debug!("imported twophase file");
     } else if file_path.starts_with("pg_wal") {
         debug!("found wal file in base section. ignore it");
-    } else if file_path.starts_with("zenith.signal") {
+    } else if file_path.starts_with("zenith.signal") || file_path.starts_with("neon.signal") {
         // Parse zenith signal file to set correct previous LSN
         let bytes = read_all_bytes(reader).await?;
-        // zenith.signal format is "PREV LSN: prev_lsn"
+        // neon.signal format is "PREV LSN: prev_lsn"
         // TODO write serialization and deserialization in the same place.
-        let zenith_signal = std::str::from_utf8(&bytes)?.trim();
-        let prev_lsn = match zenith_signal {
+        let neon_signal = std::str::from_utf8(&bytes)?.trim();
+        let prev_lsn = match neon_signal {
             "PREV LSN: none" => Lsn(0),
             "PREV LSN: invalid" => Lsn(0),
             other => {
@@ -624,17 +624,17 @@ async fn import_file(
                 split[1]
                     .trim()
                     .parse::<Lsn>()
-                    .context("can't parse zenith.signal")?
+                    .context("can't parse neon.signal")?
             }
         };
 
-        // zenith.signal is not necessarily the last file, that we handle
+        // neon.signal is not necessarily the last file, that we handle
         // but it is ok to call `finish_write()`, because final `modification.commit()`
         // will update lsn once more to the final one.
         let writer = modification.tline.writer().await;
         writer.finish_write(prev_lsn);
 
-        debug!("imported zenith signal {}", prev_lsn);
+        debug!("imported neon signal {}", prev_lsn);
     } else if file_path.starts_with("pg_tblspc") {
         // TODO Backups exported from neon won't have pg_tblspc, but we will need
         // this to import arbitrary postgres databases.

pageserver/src/pgdatadir_mapping.rs

@@ -25,9 +25,9 @@ use pageserver_api::keyspace::{KeySpaceRandomAccum, SparseKeySpace};
 use pageserver_api::models::RelSizeMigration;
 use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
-use postgres_ffi::{BLCKSZ, PgMajorVersion, TimestampTz, TransactionId};
+use postgres_ffi::{BLCKSZ, PgMajorVersion, TransactionId};
 use postgres_ffi_types::forknum::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
-use postgres_ffi_types::{Oid, RepOriginId};
+use postgres_ffi_types::{Oid, RepOriginId, TimestampTz};
 use serde::{Deserialize, Serialize};
 use strum::IntoEnumIterator;
 use tokio_util::sync::CancellationToken;

pageserver/src/tenant.rs

@@ -3393,7 +3393,13 @@ impl TenantShard {
                 .collect_vec();
 
             for timeline in timelines {
-                timeline.maybe_freeze_ephemeral_layer().await;
+                // Include a span with the timeline ID. The parent span already has the tenant ID.
+                let span =
+                    info_span!("maybe_freeze_ephemeral_layer", timeline_id = %timeline.timeline_id);
+                timeline
+                    .maybe_freeze_ephemeral_layer()
+                    .instrument(span)
+                    .await;
             }
         }
 
@@ -12816,6 +12822,40 @@ mod tests {
                 },
             ]
         );
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_get_force_image_creation_lsn() -> anyhow::Result<()> {
+        let tenant_conf = pageserver_api::models::TenantConfig {
+            pitr_interval: Some(Duration::from_secs(7 * 3600)),
+            image_layer_force_creation_period: Some(Duration::from_secs(3600)),
+            ..Default::default()
+        };
+
+        let tenant_id = TenantId::generate();
+
+        let harness = TenantHarness::create_custom(
+            "test_get_force_image_creation_lsn",
+            tenant_conf,
+            tenant_id,
+            ShardIdentity::unsharded(),
+            Generation::new(1),
+        )
+        .await?;
+        let (tenant, ctx) = harness.load().await;
+        let timeline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        timeline.gc_info.write().unwrap().cutoffs.time = Some(Lsn(100));
+        {
+            let writer = timeline.writer().await;
+            writer.finish_write(Lsn(5000));
+        }
+
+        let image_creation_lsn = timeline.get_force_image_creation_lsn().unwrap();
+        assert_eq!(image_creation_lsn, Lsn(4300));
         Ok(())
     }
 }

pageserver/src/tenant/layer_map.rs

@@ -46,10 +46,11 @@
 mod historic_layer_coverage;
 mod layer_coverage;
 
-use std::collections::{HashMap, VecDeque};
+use std::collections::{BTreeMap, HashMap, VecDeque};
 use std::iter::Peekable;
 use std::ops::Range;
 use std::sync::Arc;
+use std::time::Instant;
 
 use anyhow::Result;
 use historic_layer_coverage::BufferedHistoricLayerCoverage;
@@ -904,6 +905,103 @@ impl LayerMap {
         max_stacked_deltas
     }
 
+    /* BEGIN_HADRON */
+    /**
+     * Compute the image consistent LSN, the largest LSN below which all pages have been redone successfully.
+     * It works by first finding the latest image layers and store them into a map. Then for each delta layer,
+     * find all overlapping image layers in order to potentially increase the image LSN in case there are gaps
+     * (e.g., if an image is created at LSN 100 but the delta layer spans LSN [150, 200], then we can increase
+     * image LSN to 150 because there is no WAL record in between).
+     * Finally, the image consistent LSN is computed by taking the minimum of all image layers.
+     */
+    pub fn compute_image_consistent_lsn(&self, disk_consistent_lsn: Lsn) -> Lsn {
+        struct ImageLayerInfo {
+            // creation LSN of the image layer
+            image_lsn: Lsn,
+            // the current minimum LSN of newer delta layers with overlapping key ranges
+            min_delta_lsn: Lsn,
+        }
+        let started_at = Instant::now();
+
+        let min_l0_deltas_lsn = {
+            let l0_deltas = self.level0_deltas();
+            l0_deltas
+                .iter()
+                .map(|layer| layer.get_lsn_range().start)
+                .min()
+                .unwrap_or(disk_consistent_lsn)
+        };
+        let global_key_range = Key::MIN..Key::MAX;
+
+        // step 1: collect all most recent image layers into a map
+        // map: end key to image_layer_info
+        let mut image_map: BTreeMap<Key, ImageLayerInfo> = BTreeMap::new();
+        for (img_range, img) in self.image_coverage(&global_key_range, disk_consistent_lsn) {
+            let img_lsn = img.map(|layer| layer.get_lsn_range().end).unwrap_or(Lsn(0));
+            image_map.insert(
+                img_range.end,
+                ImageLayerInfo {
+                    image_lsn: img_lsn,
+                    min_delta_lsn: min_l0_deltas_lsn,
+                },
+            );
+        }
+
+        // step 2: go through all delta layers, and update the image layer info with overlapping
+        // key ranges
+        for layer in self.historic.iter() {
+            if !layer.is_delta {
+                continue;
+            }
+            let delta_key_range = layer.get_key_range();
+            let delta_lsn_range = layer.get_lsn_range();
+            for (img_end_key, img_info) in image_map.range_mut(delta_key_range.start..Key::MAX) {
+                debug_assert!(img_end_key >= &delta_key_range.start);
+                if delta_lsn_range.end > img_info.image_lsn {
+                    // the delta layer includes WAL records after the image
+                    // it's possibel that the delta layer's start LSN < image LSN, which will be simply ignored by step 3
+                    img_info.min_delta_lsn =
+                        std::cmp::min(img_info.min_delta_lsn, delta_lsn_range.start);
+                }
+                if img_end_key >= &delta_key_range.end {
+                    // we have fully processed all overlapping image layers
+                    break;
+                }
+            }
+        }
+
+        // step 3, go through all image layers and find the image consistent LSN
+        let mut img_consistent_lsn = min_l0_deltas_lsn.checked_sub(Lsn(1)).unwrap();
+        let mut prev_key = Key::MIN;
+        for (img_key, img_info) in image_map {
+            tracing::debug!(
+                "Image layer {:?}:{} has min delta lsn {}",
+                Range {
+                    start: prev_key,
+                    end: img_key,
+                },
+                img_info.image_lsn,
+                img_info.min_delta_lsn,
+            );
+            let image_lsn = std::cmp::max(
+                img_info.image_lsn,
+                img_info.min_delta_lsn.checked_sub(Lsn(1)).unwrap_or(Lsn(0)),
+            );
+            img_consistent_lsn = std::cmp::min(img_consistent_lsn, image_lsn);
+            prev_key = img_key;
+        }
+        tracing::info!(
+            "computed image_consistent_lsn {} for disk_consistent_lsn {} in {}ms. Processed {} layrs in total.",
+            img_consistent_lsn,
+            disk_consistent_lsn,
+            started_at.elapsed().as_millis(),
+            self.historic.len()
+        );
+        img_consistent_lsn
+    }
+
+    /* END_HADRON */
+
     /// Return all L0 delta layers
     pub fn level0_deltas(&self) -> &Vec<Arc<PersistentLayerDesc>> {
         &self.l0_delta_layers
@@ -1579,6 +1677,138 @@ mod tests {
             LayerVisibilityHint::Visible
         ));
     }
+
+    /* BEGIN_HADRON */
+    #[test]
+    fn test_compute_image_consistent_lsn() {
+        let mut layer_map = LayerMap::default();
+
+        let disk_consistent_lsn = Lsn(1000);
+        // case 1: empty layer map
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(
+            disk_consistent_lsn.checked_sub(Lsn(1)).unwrap(),
+            image_consistent_lsn
+        );
+
+        // case 2: only L0 delta layer
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(900)..Lsn(990),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(100),
+                Lsn(850)..Lsn(899),
+                true,
+            ));
+        }
+
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 3: 3 images, no L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(100)..Lsn(100),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(40)..Key::from_i128(70),
+                Lsn(200)..Lsn(200),
+                false,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(70)..Key::from_i128(100),
+                Lsn(150)..Lsn(150),
+                false,
+            ));
+        }
+        // should use min L0 delta LSN - 1 as image consistent LSN
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(849), image_consistent_lsn);
+
+        // case 4: 3 images with 1 L1 delta
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(300)..Lsn(350),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(299), image_consistent_lsn);
+
+        // case 5: 3 images with 1 more L1 delta with smaller LSN
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(72),
+                Lsn(200)..Lsn(300),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 6: 3 images with more newer L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(30),
+                Lsn(400)..Lsn(500),
+                true,
+            ));
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(35)..Key::from_i128(100),
+                Lsn(450)..Lsn(600),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 7: 3 images with more older L1 deltas (no impact on final results)
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(40),
+                Lsn(0)..Lsn(50),
+                true,
+            ));
+
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(50)..Key::from_i128(100),
+                Lsn(10)..Lsn(60),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(199), image_consistent_lsn);
+
+        // case 8: 3 images with one more L1 delta with overlapping LSN range
+        {
+            let mut updates = layer_map.batch_update();
+            updates.insert_historic(PersistentLayerDesc::new_test(
+                Key::from_i128(0)..Key::from_i128(50),
+                Lsn(50)..Lsn(250),
+                true,
+            ));
+        }
+        let image_consistent_lsn = layer_map.compute_image_consistent_lsn(disk_consistent_lsn);
+        assert_eq!(Lsn(100), image_consistent_lsn);
+    }
+
+    /* END_HADRON */
 }
 
 #[cfg(test)]

pageserver/src/tenant/mgr.rs

@@ -1678,6 +1678,8 @@ impl TenantManager {
         // Phase 6: Release the InProgress on the parent shard
         drop(parent_slot_guard);
 
+        utils::pausable_failpoint!("shard-split-post-finish-pause");
+
         Ok(child_shards)
     }
 

pageserver/src/tenant/storage_layer.rs

@@ -75,7 +75,7 @@ where
 /// the same ValueReconstructState struct in the next 'get_value_reconstruct_data'
 /// call, to collect more records.
 ///
-#[derive(Debug, Default)]
+#[derive(Debug, Default, Clone)]
 pub(crate) struct ValueReconstructState {
     pub(crate) records: Vec<(Lsn, NeonWalRecord)>,
     pub(crate) img: Option<(Lsn, Bytes)>,
@@ -308,6 +308,9 @@ pub struct ValuesReconstructState {
     layers_visited: u32,
     delta_layers_visited: u32,
 
+    pub(crate) enable_debug: bool,
+    pub(crate) debug_state: ValueReconstructState,
+
     pub(crate) io_concurrency: IoConcurrency,
     num_active_ios: Arc<AtomicUsize>,
 
@@ -657,6 +660,23 @@ impl ValuesReconstructState {
             layers_visited: 0,
             delta_layers_visited: 0,
             io_concurrency,
+            enable_debug: false,
+            debug_state: ValueReconstructState::default(),
+            num_active_ios: Arc::new(AtomicUsize::new(0)),
+            read_path: None,
+        }
+    }
+
+    pub(crate) fn new_with_debug(io_concurrency: IoConcurrency) -> Self {
+        Self {
+            keys: HashMap::new(),
+            keys_done: KeySpaceRandomAccum::new(),
+            keys_with_image_coverage: None,
+            layers_visited: 0,
+            delta_layers_visited: 0,
+            io_concurrency,
+            enable_debug: true,
+            debug_state: ValueReconstructState::default(),
             num_active_ios: Arc::new(AtomicUsize::new(0)),
             read_path: None,
         }
@@ -670,6 +690,12 @@ impl ValuesReconstructState {
         self.io_concurrency.spawn_io(fut).await;
     }
 
+    pub(crate) fn set_debug_state(&mut self, debug_state: &ValueReconstructState) {
+        if self.enable_debug {
+            self.debug_state = debug_state.clone();
+        }
+    }
+
     pub(crate) fn on_layer_visited(&mut self, layer: &ReadableLayer) {
         self.layers_visited += 1;
         if let ReadableLayer::PersistentLayer(layer) = layer {

pageserver/src/tenant/timeline.rs

@@ -351,13 +351,6 @@ pub struct Timeline {
     last_image_layer_creation_check_at: AtomicLsn,
     last_image_layer_creation_check_instant: std::sync::Mutex<Option<Instant>>,
 
-    // HADRON
-    /// If a key range has writes with LSN > force_image_creation_lsn, then we should force image layer creation
-    /// on this key range.
-    force_image_creation_lsn: AtomicLsn,
-    /// The last time instant when force_image_creation_lsn is computed.
-    force_image_creation_lsn_computed_at: std::sync::Mutex<Option<Instant>>,
-
     /// Current logical size of the "datadir", at the last LSN.
     current_logical_size: LogicalSize,
 
@@ -1260,6 +1253,57 @@ impl Timeline {
         }
     }
 
+    #[inline(always)]
+    pub(crate) async fn debug_get(
+        &self,
+        key: Key,
+        lsn: Lsn,
+        ctx: &RequestContext,
+        reconstruct_state: &mut ValuesReconstructState,
+    ) -> Result<Bytes, PageReconstructError> {
+        if !lsn.is_valid() {
+            return Err(PageReconstructError::Other(anyhow::anyhow!("Invalid LSN")));
+        }
+
+        // This check is debug-only because of the cost of hashing, and because it's a double-check: we
+        // already checked the key against the shard_identity when looking up the Timeline from
+        // page_service.
+        debug_assert!(!self.shard_identity.is_key_disposable(&key));
+
+        let query = VersionedKeySpaceQuery::uniform(KeySpace::single(key..key.next()), lsn);
+        let vectored_res = self
+            .debug_get_vectored_impl(query, reconstruct_state, ctx)
+            .await;
+
+        let key_value = vectored_res?.pop_first();
+        match key_value {
+            Some((got_key, value)) => {
+                if got_key != key {
+                    error!(
+                        "Expected {}, but singular vectored get returned {}",
+                        key, got_key
+                    );
+                    Err(PageReconstructError::Other(anyhow!(
+                        "Singular vectored get returned wrong key"
+                    )))
+                } else {
+                    value
+                }
+            }
+            None => Err(PageReconstructError::MissingKey(Box::new(
+                MissingKeyError {
+                    keyspace: KeySpace::single(key..key.next()),
+                    shard: self.shard_identity.get_shard_number(&key),
+                    original_hwm_lsn: lsn,
+                    ancestor_lsn: None,
+                    backtrace: None,
+                    read_path: None,
+                    query: None,
+                },
+            ))),
+        }
+    }
+
     pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
 
     /// Look up multiple page versions at a given LSN
@@ -1554,6 +1598,98 @@ impl Timeline {
         Ok(results)
     }
 
+    // A copy of the get_vectored_impl method except that we store the image and wal records into `reconstruct_state`.
+    // This is only used in the http getpage call for debugging purpose.
+    pub(super) async fn debug_get_vectored_impl(
+        &self,
+        query: VersionedKeySpaceQuery,
+        reconstruct_state: &mut ValuesReconstructState,
+        ctx: &RequestContext,
+    ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
+        if query.is_empty() {
+            return Ok(BTreeMap::default());
+        }
+
+        let read_path = if self.conf.enable_read_path_debugging || ctx.read_path_debug() {
+            Some(ReadPath::new(
+                query.total_keyspace(),
+                query.high_watermark_lsn()?,
+            ))
+        } else {
+            None
+        };
+
+        reconstruct_state.read_path = read_path;
+
+        let traversal_res: Result<(), _> = self
+            .get_vectored_reconstruct_data(query.clone(), reconstruct_state, ctx)
+            .await;
+
+        if let Err(err) = traversal_res {
+            // Wait for all the spawned IOs to complete.
+            // See comments on `spawn_io` inside `storage_layer` for more details.
+            let mut collect_futs = std::mem::take(&mut reconstruct_state.keys)
+                .into_values()
+                .map(|state| state.collect_pending_ios())
+                .collect::<FuturesUnordered<_>>();
+            while collect_futs.next().await.is_some() {}
+            return Err(err);
+        };
+
+        let reconstruct_state = Arc::new(Mutex::new(reconstruct_state));
+        let futs = FuturesUnordered::new();
+
+        for (key, state) in std::mem::take(&mut reconstruct_state.lock().unwrap().keys) {
+            let req_lsn_for_key = query.map_key_to_lsn(&key);
+            futs.push({
+                let walredo_self = self.myself.upgrade().expect("&self method holds the arc");
+                let rc_clone = Arc::clone(&reconstruct_state);
+
+                async move {
+                    assert_eq!(state.situation, ValueReconstructSituation::Complete);
+
+                    let converted = match state.collect_pending_ios().await {
+                        Ok(ok) => ok,
+                        Err(err) => {
+                            return (key, Err(err));
+                        }
+                    };
+                    DELTAS_PER_READ_GLOBAL.observe(converted.num_deltas() as f64);
+
+                    // The walredo module expects the records to be descending in terms of Lsn.
+                    // And we submit the IOs in that order, so, there shuold be no need to sort here.
+                    debug_assert!(
+                        converted
+                            .records
+                            .is_sorted_by_key(|(lsn, _)| std::cmp::Reverse(*lsn)),
+                        "{converted:?}"
+                    );
+                    {
+                        let mut guard = rc_clone.lock().unwrap();
+                        guard.set_debug_state(&converted);
+                    }
+                    (
+                        key,
+                        walredo_self
+                            .reconstruct_value(
+                                key,
+                                req_lsn_for_key,
+                                converted,
+                                RedoAttemptType::ReadPage,
+                            )
+                            .await,
+                    )
+                }
+            });
+        }
+
+        let results = futs
+            .collect::<BTreeMap<Key, Result<Bytes, PageReconstructError>>>()
+            .await;
+
+        Ok(results)
+    }
+
     /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
     pub(crate) fn get_last_record_lsn(&self) -> Lsn {
         self.last_record_lsn.load().last
@@ -1900,6 +2036,8 @@ impl Timeline {
     // an ephemeral layer open forever when idle.  It also freezes layers if the global limit on
     // ephemeral layer bytes has been breached.
     pub(super) async fn maybe_freeze_ephemeral_layer(&self) {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
         let Ok(mut write_guard) = self.write_lock.try_lock() else {
             // If the write lock is held, there is an active wal receiver: rolling open layers
             // is their responsibility while they hold this lock.
@@ -2854,7 +2992,7 @@ impl Timeline {
     }
 
     // HADRON
-    fn get_image_creation_timeout(&self) -> Option<Duration> {
+    fn get_image_layer_force_creation_period(&self) -> Option<Duration> {
         let tenant_conf = self.tenant_conf.load();
         tenant_conf
             .tenant_conf
@@ -3134,9 +3272,6 @@ impl Timeline {
                 repartition_threshold: 0,
                 last_image_layer_creation_check_at: AtomicLsn::new(0),
                 last_image_layer_creation_check_instant: Mutex::new(None),
-                // HADRON
-                force_image_creation_lsn: AtomicLsn::new(0),
-                force_image_creation_lsn_computed_at: std::sync::Mutex::new(None),
                 last_received_wal: Mutex::new(None),
                 rel_size_latest_cache: RwLock::new(HashMap::new()),
                 rel_size_snapshot_cache: Mutex::new(LruCache::new(relsize_snapshot_cache_capacity)),
@@ -5381,13 +5516,16 @@ impl Timeline {
         }
 
         // HADRON
+        // for child timelines, we consider all pages up to ancestor_LSN are redone successfully by the parent timeline
+        min_image_lsn = min_image_lsn.max(self.get_ancestor_lsn());
         if min_image_lsn < force_image_creation_lsn.unwrap_or(Lsn(0)) && max_deltas > 0 {
             info!(
-                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}",
+                "forcing image creation for partitioned range {}-{}. Min image LSN: {}, force image creation LSN: {}, num deltas: {}",
                 partition.ranges[0].start,
                 partition.ranges[0].end,
                 min_image_lsn,
-                force_image_creation_lsn.unwrap()
+                force_image_creation_lsn.unwrap(),
+                max_deltas
             );
             return true;
         }
@@ -5611,10 +5749,11 @@ impl Timeline {
     /// Predicate function which indicates whether we should check if new image layers
     /// are required. Since checking if new image layers are required is expensive in
     /// terms of CPU, we only do it in the following cases:
-    /// 1. If the timeline has ingested sufficient WAL to justify the cost
+    /// 1. If the timeline has ingested sufficient WAL to justify the cost or ...
     /// 2. If enough time has passed since the last check:
     ///     1. For large tenants, we wish to perform the check more often since they
-    ///        suffer from the lack of image layers
+    ///        suffer from the lack of image layers. Note that we assume sharded tenants
+    ///        to be large since non-zero shards do not track the logical size.
     ///     2. For small tenants (that can mostly fit in RAM), we use a much longer interval
     fn should_check_if_image_layers_required(self: &Arc<Timeline>, lsn: Lsn) -> bool {
         let large_timeline_threshold = self.conf.image_layer_generation_large_timeline_threshold;
@@ -5628,30 +5767,39 @@ impl Timeline {
 
         let distance_based_decision = distance.0 >= min_distance;
 
-        let mut time_based_decision = false;
         let mut last_check_instant = self.last_image_layer_creation_check_instant.lock().unwrap();
-        if let CurrentLogicalSize::Exact(logical_size) = self.current_logical_size.current_size() {
-            let check_required_after =
-                if Some(Into::<u64>::into(&logical_size)) >= large_timeline_threshold {
-                    self.get_checkpoint_timeout()
-                } else {
-                    Duration::from_secs(3600 * 48)
-                };
-
-            time_based_decision = match *last_check_instant {
-                Some(last_check) => {
-                    let elapsed = last_check.elapsed();
-                    elapsed >= check_required_after
+        let check_required_after = (|| {
+            if self.shard_identity.is_unsharded() {
+                if let CurrentLogicalSize::Exact(logical_size) =
+                    self.current_logical_size.current_size()
+                {
+                    if Some(Into::<u64>::into(&logical_size)) < large_timeline_threshold {
+                        return Duration::from_secs(3600 * 48);
+                    }
                 }
-                None => true,
-            };
-        }
+            }
+
+            self.get_checkpoint_timeout()
+        })();
+
+        let time_based_decision = match *last_check_instant {
+            Some(last_check) => {
+                let elapsed = last_check.elapsed();
+                elapsed >= check_required_after
+            }
+            None => true,
+        };
 
         // Do the expensive delta layer counting only if this timeline has ingested sufficient
         // WAL since the last check or a checkpoint timeout interval has elapsed since the last
         // check.
         let decision = distance_based_decision || time_based_decision;
-
+        tracing::info!(
+            "Decided to check image layers: {}. Distance-based decision: {}, time-based decision: {}",
+            decision,
+            distance_based_decision,
+            time_based_decision
+        );
         if decision {
             self.last_image_layer_creation_check_at.store(lsn);
             *last_check_instant = Some(Instant::now());
@@ -7153,6 +7301,19 @@ impl Timeline {
             .unwrap()
             .clone()
     }
+
+    /* BEGIN_HADRON */
+    pub(crate) async fn compute_image_consistent_lsn(&self) -> anyhow::Result<Lsn> {
+        let guard = self
+            .layers
+            .read(LayerManagerLockHolder::ComputeImageConsistentLsn)
+            .await;
+        let layer_map = guard.layer_map()?;
+        let disk_consistent_lsn = self.get_disk_consistent_lsn();
+
+        Ok(layer_map.compute_image_consistent_lsn(disk_consistent_lsn))
+    }
+    /* END_HADRON */
 }
 
 impl Timeline {

pageserver/src/tenant/timeline/compaction.rs

@@ -8,7 +8,7 @@ use std::cmp::min;
 use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
-use std::time::{Duration, Instant, SystemTime};
+use std::time::{Duration, Instant};
 
 use super::layer_manager::LayerManagerLockHolder;
 use super::{
@@ -34,7 +34,6 @@ use pageserver_api::models::{CompactInfoResponse, CompactKeyRange};
 use pageserver_api::shard::{ShardCount, ShardIdentity, TenantShardId};
 use pageserver_compaction::helpers::{fully_contains, overlaps_with};
 use pageserver_compaction::interface::*;
-use postgres_ffi::to_pg_timestamp;
 use serde::Serialize;
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
 use tokio_util::sync::CancellationToken;
@@ -47,7 +46,6 @@ use wal_decoder::models::value::Value;
 
 use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::page_cache;
-use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::statvfs::Statvfs;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::gc_block::GcBlock;
@@ -1271,10 +1269,7 @@ impl Timeline {
         // Define partitioning schema if needed
 
         // HADRON
-        let force_image_creation_lsn = self
-            .get_or_compute_force_image_creation_lsn(cancel, ctx)
-            .await
-            .map_err(CompactionError::Other)?;
+        let force_image_creation_lsn = self.get_force_image_creation_lsn();
 
         // 1. L0 Compact
         let l0_outcome = {
@@ -1484,59 +1479,37 @@ impl Timeline {
     }
 
     /* BEGIN_HADRON */
-    // Get the force image creation LSN. Compute it if the last computed LSN is too old.
-    async fn get_or_compute_force_image_creation_lsn(
-        self: &Arc<Self>,
-        cancel: &CancellationToken,
-        ctx: &RequestContext,
-    ) -> anyhow::Result<Option<Lsn>> {
-        const FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL: Duration = Duration::from_secs(10 * 60); // 10 minutes
-        let image_layer_force_creation_period = self.get_image_creation_timeout();
-        if image_layer_force_creation_period.is_none() {
-            return Ok(None);
+    // Get the force image creation LSN based on gc_cutoff_lsn.
+    // Note that this is an estimation and the workload rate may suddenly change. When that happens,
+    // the force image creation may be too early or too late, but eventually it should be able to catch up.
+    pub(crate) fn get_force_image_creation_lsn(self: &Arc<Self>) -> Option<Lsn> {
+        let image_creation_period = self.get_image_layer_force_creation_period()?;
+        let current_lsn = self.get_last_record_lsn();
+        let pitr_lsn = self.gc_info.read().unwrap().cutoffs.time?;
+        let pitr_interval = self.get_pitr_interval();
+        if pitr_lsn == Lsn::INVALID || pitr_interval.is_zero() {
+            tracing::warn!(
+                "pitr LSN/interval not found, skipping force image creation LSN calculation"
+            );
+            return None;
         }
 
-        let image_layer_force_creation_period = image_layer_force_creation_period.unwrap();
-        let force_image_creation_lsn_computed_at =
-            *self.force_image_creation_lsn_computed_at.lock().unwrap();
-        if force_image_creation_lsn_computed_at.is_none()
-            || force_image_creation_lsn_computed_at.unwrap().elapsed()
-                > FORCE_IMAGE_CREATION_LSN_COMPUTE_INTERVAL
-        {
-            let now: SystemTime = SystemTime::now();
-            let timestamp = now
-                .checked_sub(image_layer_force_creation_period)
-                .ok_or_else(|| {
-                    anyhow::anyhow!(
-                        "image creation timeout is too large: {image_layer_force_creation_period:?}"
-                    )
-                })?;
-            let timestamp = to_pg_timestamp(timestamp);
-            let force_image_creation_lsn = match self
-                .find_lsn_for_timestamp(timestamp, cancel, ctx)
-                .await?
-            {
-                LsnForTimestamp::Present(lsn) | LsnForTimestamp::Future(lsn) => lsn,
-                _ => {
-                    let gc_lsn = *self.get_applied_gc_cutoff_lsn();
-                    tracing::info!(
-                        "no LSN found for timestamp {timestamp:?}, using latest GC cutoff LSN {}",
-                        gc_lsn
-                    );
-                    gc_lsn
-                }
-            };
-            self.force_image_creation_lsn
-                .store(force_image_creation_lsn);
-            *self.force_image_creation_lsn_computed_at.lock().unwrap() = Some(Instant::now());
-            tracing::info!(
-                "computed force image creation LSN: {}",
-                force_image_creation_lsn
-            );
-            Ok(Some(force_image_creation_lsn))
-        } else {
-            Ok(Some(self.force_image_creation_lsn.load()))
-        }
+        let delta_lsn = current_lsn.checked_sub(pitr_lsn).unwrap().0
+            * image_creation_period.as_secs()
+            / pitr_interval.as_secs();
+        let force_image_creation_lsn = current_lsn.checked_sub(delta_lsn).unwrap_or(Lsn(0));
+
+        tracing::info!(
+            "Tenant shard {} computed force_image_creation_lsn: {}. Current lsn: {}, image_layer_force_creation_period: {:?}, GC cutoff: {}, PITR interval: {:?}",
+            self.tenant_shard_id,
+            force_image_creation_lsn,
+            current_lsn,
+            image_creation_period,
+            pitr_lsn,
+            pitr_interval
+        );
+
+        Some(force_image_creation_lsn)
     }
     /* END_HADRON */
 

pageserver/src/tenant/timeline/handle.rs

@@ -359,14 +359,14 @@ impl<T: Types> Cache<T> {
                 Err(e) => {
                     // Retry on tenant manager error to handle tenant split more gracefully
                     if attempt < GET_MAX_RETRIES {
-                        tracing::warn!(
-                            "Fail to resolve tenant shard in attempt {}: {:?}. Retrying...",
-                            attempt,
-                            e
-                        );
                         tokio::time::sleep(RETRY_BACKOFF).await;
                         continue;
                     } else {
+                        tracing::warn!(
+                            "Failed to resolve tenant shard after {} attempts: {:?}",
+                            GET_MAX_RETRIES,
+                            e
+                        );
                         return Err(e);
                     }
                 }

pageserver/src/tenant/timeline/layer_manager.rs

@@ -47,6 +47,7 @@ pub(crate) enum LayerManagerLockHolder {
     ImportPgData,
     DetachAncestor,
     Eviction,
+    ComputeImageConsistentLsn,
     #[cfg(test)]
     Testing,
 }

pageserver/src/utilization.rs

@@ -1,6 +1,6 @@
 //! An utilization metric which is used to decide on which pageserver to put next tenant.
 //!
-//! The metric is exposed via `GET /v1/utilization`. Refer and maintain it's openapi spec as the
+//! The metric is exposed via `GET /v1/utilization`. Refer and maintain its openapi spec as the
 //! truth.
 
 use std::path::Path;

pageserver/src/walingest.rs

@@ -32,9 +32,10 @@ use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::walrecord::*;
 use postgres_ffi::{
-    PgMajorVersion, TimestampTz, TransactionId, dispatch_pgversion, enum_pgversion,
-    enum_pgversion_dispatch, fsm_logical_to_physical, pg_constants,
+    PgMajorVersion, TransactionId, dispatch_pgversion, enum_pgversion, enum_pgversion_dispatch,
+    fsm_logical_to_physical, pg_constants,
 };
+use postgres_ffi_types::TimestampTz;
 use postgres_ffi_types::forknum::{FSM_FORKNUM, INIT_FORKNUM, MAIN_FORKNUM, VISIBILITYMAP_FORKNUM};
 use tracing::*;
 use utils::bin_ser::{DeserializeError, SerializeError};
@@ -1069,7 +1070,7 @@ impl WalIngest {
         // NB: In PostgreSQL, the next-multi-xid stored in the control file is allowed to
         // go to 0, and it's fixed up by skipping to FirstMultiXactId in functions that
         // read it, like GetNewMultiXactId(). This is different from how nextXid is
-        // incremented! nextXid skips over < FirstNormalTransactionId when the the value
+        // incremented! nextXid skips over < FirstNormalTransactionId when the value
         // is stored, so it's never 0 in a checkpoint.
         //
         // I don't know why it's done that way, it seems less error-prone to skip over 0

pageserver/src/walredo.rs

@@ -147,6 +147,16 @@ pub enum RedoAttemptType {
     GcCompaction,
 }
 
+impl std::fmt::Display for RedoAttemptType {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            RedoAttemptType::ReadPage => write!(f, "read page"),
+            RedoAttemptType::LegacyCompaction => write!(f, "legacy compaction"),
+            RedoAttemptType::GcCompaction => write!(f, "gc compaction"),
+        }
+    }
+}
+
 ///
 /// Public interface of WAL redo manager
 ///
@@ -199,6 +209,7 @@ impl PostgresRedoManager {
                         self.conf.wal_redo_timeout,
                         pg_version,
                         max_retry_attempts,
+                        redo_attempt_type,
                     )
                     .await
                 };
@@ -221,6 +232,7 @@ impl PostgresRedoManager {
                 self.conf.wal_redo_timeout,
                 pg_version,
                 max_retry_attempts,
+                redo_attempt_type,
             )
             .await
         }
@@ -445,6 +457,7 @@ impl PostgresRedoManager {
         wal_redo_timeout: Duration,
         pg_version: PgMajorVersion,
         max_retry_attempts: u32,
+        redo_attempt_type: RedoAttemptType,
     ) -> Result<Bytes, Error> {
         *(self.last_redo_at.lock().unwrap()) = Some(Instant::now());
 
@@ -485,17 +498,28 @@ impl PostgresRedoManager {
                 );
 
                 if let Err(e) = result.as_ref() {
-                    error!(
-                        "error applying {} WAL records {}..{} ({} bytes) to key {key}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
-                        records.len(),
-                        records.first().map(|p| p.0).unwrap_or(Lsn(0)),
-                        records.last().map(|p| p.0).unwrap_or(Lsn(0)),
-                        nbytes,
-                        base_img_lsn,
-                        lsn,
-                        n_attempts,
-                        e,
-                    );
+                    macro_rules! message {
+                        ($level:tt) => {
+                            $level!(
+                                "error applying {} WAL records {}..{} ({} bytes) to key {} during {}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
+                                records.len(),
+                                records.first().map(|p| p.0).unwrap_or(Lsn(0)),
+                                records.last().map(|p| p.0).unwrap_or(Lsn(0)),
+                                nbytes,
+                                key,
+                                redo_attempt_type,
+                                base_img_lsn,
+                                lsn,
+                                n_attempts,
+                                e,
+                            )
+                        }
+                    }
+                    match redo_attempt_type {
+                        RedoAttemptType::ReadPage => message!(error),
+                        RedoAttemptType::LegacyCompaction => message!(error),
+                        RedoAttemptType::GcCompaction => message!(warn),
+                    }
                 }
 
                 result.map_err(Error::Other)

pgxn/neon/communicator.c

@@ -421,7 +421,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 {
 	if (resp->tag != T_NeonGetPageResponse && resp->tag != T_NeonErrorResponse)
 	{
-		neon_shard_log(slot->shard_no, PANIC, "Unexpected prefetch response %d, ring_receive=%ld, ring_flush=%ld, ring_unused=%ld",
+		neon_shard_log(slot->shard_no, PANIC, "Unexpected prefetch response %d, ring_receive=" UINT64_FORMAT ", ring_flush=" UINT64_FORMAT ", ring_unused=" UINT64_FORMAT "",
 					   resp->tag, MyPState->ring_receive, MyPState->ring_flush, MyPState->ring_unused);
 	}
 	if (neon_protocol_version >= 3)
@@ -438,7 +438,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 				getpage_resp->req.blkno != slot->buftag.blockNum)
 			{
 				NEON_PANIC_CONNECTION_STATE(slot->shard_no, PANIC,
-											"Receive unexpected getpage response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u} to get page request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u}",
+											"Receive unexpected getpage response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u} to get page request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u, block=%u}",
 											resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(getpage_resp->req.rinfo), getpage_resp->req.forknum, getpage_resp->req.blkno,
 											slot->reqid, LSN_FORMAT_ARGS(slot->request_lsns.request_lsn), LSN_FORMAT_ARGS(slot->request_lsns.not_modified_since), RelFileInfoFmt(rinfo), slot->buftag.forkNum, slot->buftag.blockNum);
 			}
@@ -447,7 +447,7 @@ check_getpage_response(PrefetchRequest* slot, NeonResponse* resp)
 				 resp->lsn != slot->request_lsns.request_lsn ||
 				 resp->not_modified_since != slot->request_lsns.not_modified_since)
 		{
-			elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+			elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 				 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 				 slot->reqid, LSN_FORMAT_ARGS(slot->request_lsns.request_lsn), LSN_FORMAT_ARGS(slot->request_lsns.not_modified_since));
 		}
@@ -496,9 +496,9 @@ communicator_prefetch_pump_state(void)
 			slot->my_ring_index != MyPState->ring_receive)
 		{
 			neon_shard_log(slot->shard_no, PANIC,
-						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 						   slot->status, slot->response,
-						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+						   slot->my_ring_index, MyPState->ring_receive);
 		}
 		/* update prefetch state */
 		MyPState->n_responses_buffered += 1;
@@ -789,9 +789,9 @@ prefetch_read(PrefetchRequest *slot)
 		slot->my_ring_index != MyPState->ring_receive)
 	{
 		neon_shard_log(slot->shard_no, PANIC,
-					   "Incorrect prefetch read: status=%d response=%p my=%lu receive=%lu",
+					   "Incorrect prefetch read: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 					   slot->status, slot->response,
-					   (long)slot->my_ring_index, (long)MyPState->ring_receive);
+					   slot->my_ring_index, MyPState->ring_receive);
 	}
 
 	/*
@@ -816,9 +816,9 @@ prefetch_read(PrefetchRequest *slot)
 			slot->my_ring_index != MyPState->ring_receive)
 		{
 			neon_shard_log(shard_no, PANIC,
-						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=" UINT64_FORMAT " receive=" UINT64_FORMAT "",
 						   slot->status, slot->response,
-						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+						   slot->my_ring_index, MyPState->ring_receive);
 		}
 
 		/* update prefetch state */
@@ -852,8 +852,8 @@ prefetch_read(PrefetchRequest *slot)
 		 * and the prefetch queue was flushed during the receive call
 		 */
 		neon_shard_log(shard_no, LOG,
-					   "No response from reading prefetch entry %lu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
-					   (long) my_ring_index,
+					   "No response from reading prefetch entry " UINT64_FORMAT ": %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
+					   my_ring_index,
 					   RelFileInfoFmt(BufTagGetNRelFileInfo(buftag)),
 					   buftag.forkNum, buftag.blockNum);
 		return false;
@@ -1844,7 +1844,7 @@ nm_to_string(NeonMessage *msg)
 				NeonDbSizeResponse *msg_resp = (NeonDbSizeResponse *) msg;
 
 				appendStringInfoString(&s, "{\"type\": \"NeonDbSizeResponse\"");
-				appendStringInfo(&s, ", \"db_size\": %ld}",
+				appendStringInfo(&s, ", \"db_size\": " INT64_FORMAT "}",
 								 msg_resp->db_size);
 				appendStringInfoChar(&s, '}');
 
@@ -2045,7 +2045,7 @@ communicator_exists(NRelFileInfo rinfo, ForkNumber forkNum, neon_request_lsns *r
 						exists_resp->req.forknum != request.forknum)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to exits request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to exits request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(exists_resp->req.rinfo), exists_resp->req.forknum,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), RelFileInfoFmt(request.rinfo), request.forknum);
 					}
@@ -2058,14 +2058,14 @@ communicator_exists(NRelFileInfo rinfo, ForkNumber forkNum, neon_request_lsns *r
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match exists request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read relation existence of rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read relation existence of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								resp->reqid,
 								RelFileInfoFmt(rinfo),
 								forkNum,
@@ -2241,7 +2241,7 @@ Retry:
 			case T_NeonErrorResponse:
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[shard %d, reqid %lx] could not read block %u in rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[shard %d, reqid " UINT64_HEX_FORMAT "] could not read block %u in rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								slot->shard_no, resp->reqid, blockno, RelFileInfoFmt(rinfo),
 								forkNum, LSN_FORMAT_ARGS(reqlsns->effective_request_lsn)),
 						 errdetail("page server returned error: %s",
@@ -2294,7 +2294,7 @@ communicator_nblocks(NRelFileInfo rinfo, ForkNumber forknum, neon_request_lsns *
 						relsize_resp->req.forknum != forknum)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to get relsize request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u} to get relsize request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, rel=%u/%u/%u.%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), RelFileInfoFmt(relsize_resp->req.rinfo), relsize_resp->req.forknum,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), RelFileInfoFmt(request.rinfo), forknum);
 					}
@@ -2307,14 +2307,14 @@ communicator_nblocks(NRelFileInfo rinfo, ForkNumber forknum, neon_request_lsns *
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get relsize request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get relsize request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read relation size of rel %u/%u/%u.%u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read relation size of rel %u/%u/%u.%u from page server at lsn %X/%08X",
 								resp->reqid,
 								RelFileInfoFmt(rinfo),
 								forknum,
@@ -2364,7 +2364,7 @@ communicator_dbsize(Oid dbNode, neon_request_lsns *request_lsns)
 						dbsize_resp->req.dbNode != dbNode)
 					{
 						NEON_PANIC_CONNECTION_STATE(0, PANIC,
-													"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, dbNode=%u} to get DB size request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, dbNode=%u}",
+													"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, dbNode=%u} to get DB size request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, dbNode=%u}",
 													resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), dbsize_resp->req.dbNode,
 													request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), dbNode);
 					}
@@ -2377,14 +2377,14 @@ communicator_dbsize(Oid dbNode, neon_request_lsns *request_lsns)
 				{
 					if (!equal_requests(resp, &request.hdr))
 					{
-						elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get DB size request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+						elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get DB size request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 							 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 							 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 					}
 				}
 				ereport(ERROR,
 						(errcode(ERRCODE_IO_ERROR),
-						 errmsg(NEON_TAG "[reqid %lx] could not read db size of db %u from page server at lsn %X/%08X",
+						 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read db size of db %u from page server at lsn %X/%08X",
 								resp->reqid,
 								dbNode, LSN_FORMAT_ARGS(request_lsns->effective_request_lsn)),
 						 errdetail("page server returned error: %s",
@@ -2455,7 +2455,7 @@ communicator_read_slru_segment(SlruKind kind, int64 segno, neon_request_lsns *re
 					slru_resp->req.segno != segno)
 				{
 					NEON_PANIC_CONNECTION_STATE(0, PANIC,
-												"Unexpect response {reqid=%lx,lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%u} to get SLRU segment request {reqid=%lx,lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%lluu}",
+												"Unexpect response {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%u} to get SLRU segment request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X, kind=%u, segno=%lluu}",
 												resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since), slru_resp->req.kind, slru_resp->req.segno,
 												request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since), kind, (unsigned long long) segno);
 				}
@@ -2469,14 +2469,14 @@ communicator_read_slru_segment(SlruKind kind, int64 segno, neon_request_lsns *re
 			{
 				if (!equal_requests(resp, &request.hdr))
 				{
-					elog(WARNING, NEON_TAG "Error message {reqid=%lx,lsn=%X/%08X, since=%X/%08X} doesn't match get SLRU segment request {reqid=%lx,lsn=%X/%08X, since=%X/%08X}",
+					elog(WARNING, NEON_TAG "Error message {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X} doesn't match get SLRU segment request {reqid=" UINT64_HEX_FORMAT ",lsn=%X/%08X, since=%X/%08X}",
 						 resp->reqid, LSN_FORMAT_ARGS(resp->lsn), LSN_FORMAT_ARGS(resp->not_modified_since),
 						 request.hdr.reqid, LSN_FORMAT_ARGS(request.hdr.lsn), LSN_FORMAT_ARGS(request.hdr.not_modified_since));
 				}
 			}
 			ereport(ERROR,
 					(errcode(ERRCODE_IO_ERROR),
-					 errmsg(NEON_TAG "[reqid %lx] could not read SLRU %d segment %llu at lsn %X/%08X",
+					 errmsg(NEON_TAG "[reqid " UINT64_HEX_FORMAT "] could not read SLRU %d segment %llu at lsn %X/%08X",
 							resp->reqid,
 							kind,
 							(unsigned long long) segno,

pgxn/neon/file_cache.c

@@ -162,8 +162,34 @@ typedef struct FileCacheControl
 	dlist_head	lru;			/* double linked list for LRU replacement
 								 * algorithm */
 	dlist_head  holes;          /* double linked list of punched holes */
-	HyperLogLogState wss_estimation; /* estimation of working set size */
+
 	ConditionVariable cv[N_COND_VARS]; /* turnstile of condition variables */
+
+	/*
+	 * Estimation of working set size.
+	 *
+	 * This is not guarded by the lock. No locking is needed because all the
+	 * writes to the "registers" are simple 64-bit stores, to update a
+	 * timestamp. We assume that:
+	 *
+	 * - 64-bit stores are atomic. We could enforce that by using
+	 *   pg_atomic_uint64 instead of TimestampTz as the datatype in hll.h, but
+	 *   for now we just rely on it implicitly.
+	 *
+	 * - Even if they're not, and there is a race between two stores, it
+	 *   doesn't matter much which one wins because they're both updating the
+	 *   register with the current timestamp. Or you have a race between
+	 *   resetting the register and updating it, in which case it also doesn't
+	 *   matter much which one wins.
+	 *
+	 * - If they're not atomic, you might get an occasional "torn write" if
+	 *   you're really unlucky, but we tolerate that too. It just means that
+	 *   the estimate will be a little off, until the register is updated
+	 *   again.
+	 */
+	HyperLogLogState wss_estimation;
+
+	/* Prewarmer state */
 	PrewarmWorkerState prewarm_workers[MAX_PREWARM_WORKERS];
 	size_t n_prewarm_workers;
 	size_t n_prewarm_entries;
@@ -205,6 +231,8 @@ bool AmPrewarmWorker;
 
 #define LFC_ENABLED() (lfc_ctl->limit != 0)
 
+PGDLLEXPORT void lfc_prewarm_main(Datum main_arg);
+
 /*
  * Close LFC file if opened.
  * All backends should close their LFC files once LFC is disabled.
@@ -1142,6 +1170,13 @@ lfc_readv_select(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
 
+	/* Update working set size estimate for the blocks */
+	for (int i = 0; i < nblocks; i++)
+	{
+		tag.blockNum = blkno + i;
+		addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
+	}
+
 	/*
 	 * For every chunk that has blocks we're interested in, we
 	 * 1. get the chunk header
@@ -1220,14 +1255,6 @@ lfc_readv_select(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		}
 
 		entry = hash_search_with_hash_value(lfc_hash, &tag, hash, HASH_FIND, NULL);
-
-		/* Approximate working set for the blocks assumed in this entry */
-		for (int i = 0; i < blocks_in_chunk; i++)
-		{
-			tag.blockNum = blkno + i;
-			addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
-		}
-
 		if (entry == NULL)
 		{
 			/* Pages are not cached */
@@ -1504,9 +1531,15 @@ lfc_prefetch(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
 		return false;
 
 	CopyNRelFileInfoToBufTag(tag, rinfo);
+	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
 	tag.forkNum = forknum;
 
-	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
+	/* Update working set size estimate for the blocks */
+	if (lfc_prewarm_update_ws_estimation)
+	{
+		tag.blockNum = blkno;
+		addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
+	}
 
 	tag.blockNum = blkno - chunk_offs;
 	hash = get_hash_value(lfc_hash, &tag);
@@ -1524,19 +1557,13 @@ lfc_prefetch(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blkno,
 
 	if (lwlsn > lsn)
 	{
-		elog(DEBUG1, "Skip LFC write for %d because LwLSN=%X/%X is greater than not_nodified_since LSN %X/%X",
+		elog(DEBUG1, "Skip LFC write for %u because LwLSN=%X/%X is greater than not_nodified_since LSN %X/%X",
 			 blkno, LSN_FORMAT_ARGS(lwlsn), LSN_FORMAT_ARGS(lsn));
 		LWLockRelease(lfc_lock);
 		return false;
 	}
 
 	entry = hash_search_with_hash_value(lfc_hash, &tag, hash, HASH_ENTER, &found);
-
-	if (lfc_prewarm_update_ws_estimation)
-	{
-		tag.blockNum = blkno;
-		addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
-	}
 	if (found)
 	{
 		state = GET_STATE(entry, chunk_offs);
@@ -1649,9 +1676,15 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		return;
 
 	CopyNRelFileInfoToBufTag(tag, rinfo);
+	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
 	tag.forkNum = forkNum;
 
-	CriticalAssert(BufTagGetRelNumber(&tag) != InvalidRelFileNumber);
+	/* Update working set size estimate for the blocks */
+	for (int i = 0; i < nblocks; i++)
+	{
+		tag.blockNum = blkno + i;
+		addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
+	}
 
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
 
@@ -1692,14 +1725,6 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		cv = &lfc_ctl->cv[hash % N_COND_VARS];
 
 		entry = hash_search_with_hash_value(lfc_hash, &tag, hash, HASH_ENTER, &found);
-
-		/* Approximate working set for the blocks assumed in this entry */
-		for (int i = 0; i < blocks_in_chunk; i++)
-		{
-			tag.blockNum = blkno + i;
-			addSHLL(&lfc_ctl->wss_estimation, hash_bytes((uint8_t const*)&tag, sizeof(tag)));
-		}
-
 		if (found)
 		{
 			/*
@@ -2135,40 +2160,23 @@ local_cache_pages(PG_FUNCTION_ARGS)
 		SRF_RETURN_DONE(funcctx);
 }
 
-PG_FUNCTION_INFO_V1(approximate_working_set_size_seconds);
 
-Datum
-approximate_working_set_size_seconds(PG_FUNCTION_ARGS)
+/*
+ * Internal implementation of the approximate_working_set_size_seconds()
+ * function.
+ */
+int32
+lfc_approximate_working_set_size_seconds(time_t duration, bool reset)
 {
-	if (lfc_size_limit != 0)
-	{
-		int32 dc;
-		time_t duration = PG_ARGISNULL(0) ? (time_t)-1 : PG_GETARG_INT32(0);
-		LWLockAcquire(lfc_lock, LW_SHARED);
-		dc = (int32) estimateSHLL(&lfc_ctl->wss_estimation, duration);
-		LWLockRelease(lfc_lock);
-		PG_RETURN_INT32(dc);
-	}
-	PG_RETURN_NULL();
-}
+	int32		dc;
 
-PG_FUNCTION_INFO_V1(approximate_working_set_size);
+	if (lfc_size_limit == 0)
+		return -1;
 
-Datum
-approximate_working_set_size(PG_FUNCTION_ARGS)
-{
-	if (lfc_size_limit != 0)
-	{
-		int32 dc;
-		bool reset = PG_GETARG_BOOL(0);
-		LWLockAcquire(lfc_lock, reset ? LW_EXCLUSIVE : LW_SHARED);
-		dc = (int32) estimateSHLL(&lfc_ctl->wss_estimation, (time_t)-1);
-		if (reset)
-			memset(lfc_ctl->wss_estimation.regs, 0, sizeof lfc_ctl->wss_estimation.regs);
-		LWLockRelease(lfc_lock);
-		PG_RETURN_INT32(dc);
-	}
-	PG_RETURN_NULL();
+	dc = (int32) estimateSHLL(&lfc_ctl->wss_estimation, duration);
+	if (reset)
+		memset(lfc_ctl->wss_estimation.regs, 0, sizeof lfc_ctl->wss_estimation.regs);
+	return dc;
 }
 
 PG_FUNCTION_INFO_V1(get_local_cache_state);

pgxn/neon/file_cache.h

@@ -47,7 +47,8 @@ extern bool lfc_prefetch(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blk
 extern FileCacheState* lfc_get_state(size_t max_entries);
 extern void lfc_prewarm(FileCacheState* fcs, uint32 n_workers);
 
-PGDLLEXPORT void lfc_prewarm_main(Datum main_arg);
+extern int32 lfc_approximate_working_set_size_seconds(time_t duration, bool reset);
+
 
 static inline bool
 lfc_read(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,

pgxn/neon/neon.c

@@ -561,6 +561,8 @@ _PG_init(void)
 PG_FUNCTION_INFO_V1(pg_cluster_size);
 PG_FUNCTION_INFO_V1(backpressure_lsns);
 PG_FUNCTION_INFO_V1(backpressure_throttling_time);
+PG_FUNCTION_INFO_V1(approximate_working_set_size_seconds);
+PG_FUNCTION_INFO_V1(approximate_working_set_size);
 
 Datum
 pg_cluster_size(PG_FUNCTION_ARGS)
@@ -607,6 +609,34 @@ backpressure_throttling_time(PG_FUNCTION_ARGS)
 	PG_RETURN_UINT64(BackpressureThrottlingTime());
 }
 
+Datum
+approximate_working_set_size_seconds(PG_FUNCTION_ARGS)
+{
+	time_t		duration;
+	int32		dc;
+
+	duration = PG_ARGISNULL(0) ? (time_t) -1 : PG_GETARG_INT32(0);
+
+	dc = lfc_approximate_working_set_size_seconds(duration, false);
+	if (dc < 0)
+		PG_RETURN_NULL();
+	else
+		PG_RETURN_INT32(dc);
+}
+
+Datum
+approximate_working_set_size(PG_FUNCTION_ARGS)
+{
+	bool		reset = PG_GETARG_BOOL(0);
+	int32		dc;
+
+	dc = lfc_approximate_working_set_size_seconds(-1, reset);
+	if (dc < 0)
+		PG_RETURN_NULL();
+	else
+		PG_RETURN_INT32(dc);
+}
+
 #if PG_MAJORVERSION_NUM >= 16
 static void
 neon_shmem_startup_hook(void)

pgxn/neon/neon_pgversioncompat.h

@@ -165,4 +165,8 @@ extern void InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags);
 extern TimeLineID GetWALInsertionTimeLine(void);
 #endif
 
+/* format codes not present in PG17-; but available in PG18+ */
+#define INT64_HEX_FORMAT "%" INT64_MODIFIER "x"
+#define UINT64_HEX_FORMAT "%" INT64_MODIFIER "x"
+
 #endif							/* NEON_PGVERSIONCOMPAT_H */

pgxn/neon_test_utils/neontest.c

@@ -236,13 +236,13 @@ clear_buffer_cache(PG_FUNCTION_ARGS)
 	bool		save_neon_test_evict;
 
 	/*
-	 * Temporarily set the zenith_test_evict GUC, so that when we pin and
+	 * Temporarily set the neon_test_evict GUC, so that when we pin and
 	 * unpin a buffer, the buffer is evicted. We use that hack to evict all
 	 * buffers, as there is no explicit "evict this buffer" function in the
 	 * buffer manager.
 	 */
-	save_neon_test_evict = zenith_test_evict;
-	zenith_test_evict = true;
+	save_neon_test_evict = neon_test_evict;
+	neon_test_evict = true;
 	PG_TRY();
 	{
 		/* Scan through all the buffers */
@@ -273,7 +273,7 @@ clear_buffer_cache(PG_FUNCTION_ARGS)
 
 			/*
 			 * Pin the buffer, and release it again. Because we have
-			 * zenith_test_evict==true, this will evict the page from the
+			 * neon_test_evict==true, this will evict the page from the
 			 * buffer cache if no one else is holding a pin on it.
 			 */
 			if (isvalid)
@@ -286,7 +286,7 @@ clear_buffer_cache(PG_FUNCTION_ARGS)
 	PG_FINALLY();
 	{
 		/* restore the GUC */
-		zenith_test_evict = save_neon_test_evict;
+		neon_test_evict = save_neon_test_evict;
 	}
 	PG_END_TRY();
 

pgxn/typedefs.list

@@ -2953,17 +2953,17 @@ XmlTableBuilderData
 YYLTYPE
 YYSTYPE
 YY_BUFFER_STATE
-ZenithErrorResponse
-ZenithExistsRequest
-ZenithExistsResponse
-ZenithGetPageRequest
-ZenithGetPageResponse
-ZenithMessage
-ZenithMessageTag
-ZenithNblocksRequest
-ZenithNblocksResponse
-ZenithRequest
-ZenithResponse
+NeonErrorResponse
+NeonExistsRequest
+NeonExistsResponse
+NeonGetPageRequest
+NeonGetPageResponse
+NeonMessage
+NeonMessageTag
+NeonNblocksRequest
+NeonNblocksResponse
+NeonRequest
+NeonResponse
 _SPI_connection
 _SPI_plan
 __AssignProcessToJobObject

poetry.lock

@@ -1,128 +1,124 @@
-# This file is automatically @generated by Poetry 2.1.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry 2.1.1 and should not be changed by hand.
 
 [[package]]
 name = "aiohappyeyeballs"
-version = "2.3.5"
+version = "2.6.1"
 description = "Happy Eyeballs for asyncio"
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "aiohappyeyeballs-2.3.5-py3-none-any.whl", hash = "sha256:4d6dea59215537dbc746e93e779caea8178c866856a721c9c660d7a5a7b8be03"},
-    {file = "aiohappyeyeballs-2.3.5.tar.gz", hash = "sha256:6fa48b9f1317254f122a07a131a86b71ca6946ca989ce6326fff54a99a920105"},
+    {file = "aiohappyeyeballs-2.6.1-py3-none-any.whl", hash = "sha256:f349ba8f4b75cb25c99c5c2d84e997e485204d2902a9597802b0371f09331fb8"},
+    {file = "aiohappyeyeballs-2.6.1.tar.gz", hash = "sha256:c3f9d0113123803ccadfdf3f0faa505bc78e6a72d1cc4806cbd719826e943558"},
 ]
 
 [[package]]
 name = "aiohttp"
-version = "3.10.11"
+version = "3.12.14"
 description = "Async http client/server framework (asyncio)"
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "aiohttp-3.10.11-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:5077b1a5f40ffa3ba1f40d537d3bec4383988ee51fbba6b74aa8fb1bc466599e"},
-    {file = "aiohttp-3.10.11-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8d6a14a4d93b5b3c2891fca94fa9d41b2322a68194422bef0dd5ec1e57d7d298"},
-    {file = "aiohttp-3.10.11-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ffbfde2443696345e23a3c597049b1dd43049bb65337837574205e7368472177"},
-    {file = "aiohttp-3.10.11-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:20b3d9e416774d41813bc02fdc0663379c01817b0874b932b81c7f777f67b217"},
-    {file = "aiohttp-3.10.11-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2b943011b45ee6bf74b22245c6faab736363678e910504dd7531a58c76c9015a"},
-    {file = "aiohttp-3.10.11-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:48bc1d924490f0d0b3658fe5c4b081a4d56ebb58af80a6729d4bd13ea569797a"},
-    {file = "aiohttp-3.10.11-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e12eb3f4b1f72aaaf6acd27d045753b18101524f72ae071ae1c91c1cd44ef115"},
-    {file = "aiohttp-3.10.11-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f14ebc419a568c2eff3c1ed35f634435c24ead2fe19c07426af41e7adb68713a"},
-    {file = "aiohttp-3.10.11-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:72b191cdf35a518bfc7ca87d770d30941decc5aaf897ec8b484eb5cc8c7706f3"},
-    {file = "aiohttp-3.10.11-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5ab2328a61fdc86424ee540d0aeb8b73bbcad7351fb7cf7a6546fc0bcffa0038"},
-    {file = "aiohttp-3.10.11-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:aa93063d4af05c49276cf14e419550a3f45258b6b9d1f16403e777f1addf4519"},
-    {file = "aiohttp-3.10.11-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:30283f9d0ce420363c24c5c2421e71a738a2155f10adbb1a11a4d4d6d2715cfc"},
-    {file = "aiohttp-3.10.11-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:e5358addc8044ee49143c546d2182c15b4ac3a60be01c3209374ace05af5733d"},
-    {file = "aiohttp-3.10.11-cp310-cp310-win32.whl", hash = "sha256:e1ffa713d3ea7cdcd4aea9cddccab41edf6882fa9552940344c44e59652e1120"},
-    {file = "aiohttp-3.10.11-cp310-cp310-win_amd64.whl", hash = "sha256:778cbd01f18ff78b5dd23c77eb82987ee4ba23408cbed233009fd570dda7e674"},
-    {file = "aiohttp-3.10.11-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:80ff08556c7f59a7972b1e8919f62e9c069c33566a6d28586771711e0eea4f07"},
-    {file = "aiohttp-3.10.11-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:2c8f96e9ee19f04c4914e4e7a42a60861066d3e1abf05c726f38d9d0a466e695"},
-    {file = "aiohttp-3.10.11-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:fb8601394d537da9221947b5d6e62b064c9a43e88a1ecd7414d21a1a6fba9c24"},
-    {file = "aiohttp-3.10.11-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2ea224cf7bc2d8856d6971cea73b1d50c9c51d36971faf1abc169a0d5f85a382"},
-    {file = "aiohttp-3.10.11-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:db9503f79e12d5d80b3efd4d01312853565c05367493379df76d2674af881caa"},
-    {file = "aiohttp-3.10.11-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0f449a50cc33f0384f633894d8d3cd020e3ccef81879c6e6245c3c375c448625"},
-    {file = "aiohttp-3.10.11-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:82052be3e6d9e0c123499127782a01a2b224b8af8c62ab46b3f6197035ad94e9"},
-    {file = "aiohttp-3.10.11-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:20063c7acf1eec550c8eb098deb5ed9e1bb0521613b03bb93644b810986027ac"},
-    {file = "aiohttp-3.10.11-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:489cced07a4c11488f47aab1f00d0c572506883f877af100a38f1fedaa884c3a"},
-    {file = "aiohttp-3.10.11-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:ea9b3bab329aeaa603ed3bf605f1e2a6f36496ad7e0e1aa42025f368ee2dc07b"},
-    {file = "aiohttp-3.10.11-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:ca117819d8ad113413016cb29774b3f6d99ad23c220069789fc050267b786c16"},
-    {file = "aiohttp-3.10.11-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:2dfb612dcbe70fb7cdcf3499e8d483079b89749c857a8f6e80263b021745c730"},
-    {file = "aiohttp-3.10.11-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f9b615d3da0d60e7d53c62e22b4fd1c70f4ae5993a44687b011ea3a2e49051b8"},
-    {file = "aiohttp-3.10.11-cp311-cp311-win32.whl", hash = "sha256:29103f9099b6068bbdf44d6a3d090e0a0b2be6d3c9f16a070dd9d0d910ec08f9"},
-    {file = "aiohttp-3.10.11-cp311-cp311-win_amd64.whl", hash = "sha256:236b28ceb79532da85d59aa9b9bf873b364e27a0acb2ceaba475dc61cffb6f3f"},
-    {file = "aiohttp-3.10.11-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:7480519f70e32bfb101d71fb9a1f330fbd291655a4c1c922232a48c458c52710"},
-    {file = "aiohttp-3.10.11-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:f65267266c9aeb2287a6622ee2bb39490292552f9fbf851baabc04c9f84e048d"},
-    {file = "aiohttp-3.10.11-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:7400a93d629a0608dc1d6c55f1e3d6e07f7375745aaa8bd7f085571e4d1cee97"},
-    {file = "aiohttp-3.10.11-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f34b97e4b11b8d4eb2c3a4f975be626cc8af99ff479da7de49ac2c6d02d35725"},
-    {file = "aiohttp-3.10.11-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1e7b825da878464a252ccff2958838f9caa82f32a8dbc334eb9b34a026e2c636"},
-    {file = "aiohttp-3.10.11-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f9f92a344c50b9667827da308473005f34767b6a2a60d9acff56ae94f895f385"},
-    {file = "aiohttp-3.10.11-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc6f1ab987a27b83c5268a17218463c2ec08dbb754195113867a27b166cd6087"},
-    {file = "aiohttp-3.10.11-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:1dc0f4ca54842173d03322793ebcf2c8cc2d34ae91cc762478e295d8e361e03f"},
-    {file = "aiohttp-3.10.11-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:7ce6a51469bfaacff146e59e7fb61c9c23006495d11cc24c514a455032bcfa03"},
-    {file = "aiohttp-3.10.11-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:aad3cd91d484d065ede16f3cf15408254e2469e3f613b241a1db552c5eb7ab7d"},
-    {file = "aiohttp-3.10.11-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:f4df4b8ca97f658c880fb4b90b1d1ec528315d4030af1ec763247ebfd33d8b9a"},
-    {file = "aiohttp-3.10.11-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:2e4e18a0a2d03531edbc06c366954e40a3f8d2a88d2b936bbe78a0c75a3aab3e"},
-    {file = "aiohttp-3.10.11-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:6ce66780fa1a20e45bc753cda2a149daa6dbf1561fc1289fa0c308391c7bc0a4"},
-    {file = "aiohttp-3.10.11-cp312-cp312-win32.whl", hash = "sha256:a919c8957695ea4c0e7a3e8d16494e3477b86f33067478f43106921c2fef15bb"},
-    {file = "aiohttp-3.10.11-cp312-cp312-win_amd64.whl", hash = "sha256:b5e29706e6389a2283a91611c91bf24f218962717c8f3b4e528ef529d112ee27"},
-    {file = "aiohttp-3.10.11-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:703938e22434d7d14ec22f9f310559331f455018389222eed132808cd8f44127"},
-    {file = "aiohttp-3.10.11-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:9bc50b63648840854e00084c2b43035a62e033cb9b06d8c22b409d56eb098413"},
-    {file = "aiohttp-3.10.11-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:5f0463bf8b0754bc744e1feb61590706823795041e63edf30118a6f0bf577461"},
-    {file = "aiohttp-3.10.11-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f6c6dec398ac5a87cb3a407b068e1106b20ef001c344e34154616183fe684288"},
-    {file = "aiohttp-3.10.11-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:bcaf2d79104d53d4dcf934f7ce76d3d155302d07dae24dff6c9fffd217568067"},
-    {file = "aiohttp-3.10.11-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:25fd5470922091b5a9aeeb7e75be609e16b4fba81cdeaf12981393fb240dd10e"},
-    {file = "aiohttp-3.10.11-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bbde2ca67230923a42161b1f408c3992ae6e0be782dca0c44cb3206bf330dee1"},
-    {file = "aiohttp-3.10.11-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:249c8ff8d26a8b41a0f12f9df804e7c685ca35a207e2410adbd3e924217b9006"},
-    {file = "aiohttp-3.10.11-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:878ca6a931ee8c486a8f7b432b65431d095c522cbeb34892bee5be97b3481d0f"},
-    {file = "aiohttp-3.10.11-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:8663f7777ce775f0413324be0d96d9730959b2ca73d9b7e2c2c90539139cbdd6"},
-    {file = "aiohttp-3.10.11-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:6cd3f10b01f0c31481fba8d302b61603a2acb37b9d30e1d14e0f5a58b7b18a31"},
-    {file = "aiohttp-3.10.11-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:4e8d8aad9402d3aa02fdc5ca2fe68bcb9fdfe1f77b40b10410a94c7f408b664d"},
-    {file = "aiohttp-3.10.11-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:38e3c4f80196b4f6c3a85d134a534a56f52da9cb8d8e7af1b79a32eefee73a00"},
-    {file = "aiohttp-3.10.11-cp313-cp313-win32.whl", hash = "sha256:fc31820cfc3b2863c6e95e14fcf815dc7afe52480b4dc03393c4873bb5599f71"},
-    {file = "aiohttp-3.10.11-cp313-cp313-win_amd64.whl", hash = "sha256:4996ff1345704ffdd6d75fb06ed175938c133425af616142e7187f28dc75f14e"},
-    {file = "aiohttp-3.10.11-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:74baf1a7d948b3d640badeac333af581a367ab916b37e44cf90a0334157cdfd2"},
-    {file = "aiohttp-3.10.11-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:473aebc3b871646e1940c05268d451f2543a1d209f47035b594b9d4e91ce8339"},
-    {file = "aiohttp-3.10.11-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:c2f746a6968c54ab2186574e15c3f14f3e7f67aef12b761e043b33b89c5b5f95"},
-    {file = "aiohttp-3.10.11-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d110cabad8360ffa0dec8f6ec60e43286e9d251e77db4763a87dcfe55b4adb92"},
-    {file = "aiohttp-3.10.11-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e0099c7d5d7afff4202a0c670e5b723f7718810000b4abcbc96b064129e64bc7"},
-    {file = "aiohttp-3.10.11-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:0316e624b754dbbf8c872b62fe6dcb395ef20c70e59890dfa0de9eafccd2849d"},
-    {file = "aiohttp-3.10.11-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5a5f7ab8baf13314e6b2485965cbacb94afff1e93466ac4d06a47a81c50f9cca"},
-    {file = "aiohttp-3.10.11-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c891011e76041e6508cbfc469dd1a8ea09bc24e87e4c204e05f150c4c455a5fa"},
-    {file = "aiohttp-3.10.11-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:9208299251370ee815473270c52cd3f7069ee9ed348d941d574d1457d2c73e8b"},
-    {file = "aiohttp-3.10.11-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:459f0f32c8356e8125f45eeff0ecf2b1cb6db1551304972702f34cd9e6c44658"},
-    {file = "aiohttp-3.10.11-cp38-cp38-musllinux_1_2_ppc64le.whl", hash = "sha256:14cdc8c1810bbd4b4b9f142eeee23cda528ae4e57ea0923551a9af4820980e39"},
-    {file = "aiohttp-3.10.11-cp38-cp38-musllinux_1_2_s390x.whl", hash = "sha256:971aa438a29701d4b34e4943e91b5e984c3ae6ccbf80dd9efaffb01bd0b243a9"},
-    {file = "aiohttp-3.10.11-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:9a309c5de392dfe0f32ee57fa43ed8fc6ddf9985425e84bd51ed66bb16bce3a7"},
-    {file = "aiohttp-3.10.11-cp38-cp38-win32.whl", hash = "sha256:9ec1628180241d906a0840b38f162a3215114b14541f1a8711c368a8739a9be4"},
-    {file = "aiohttp-3.10.11-cp38-cp38-win_amd64.whl", hash = "sha256:9c6e0ffd52c929f985c7258f83185d17c76d4275ad22e90aa29f38e211aacbec"},
-    {file = "aiohttp-3.10.11-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:cdc493a2e5d8dc79b2df5bec9558425bcd39aff59fc949810cbd0832e294b106"},
-    {file = "aiohttp-3.10.11-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:b3e70f24e7d0405be2348da9d5a7836936bf3a9b4fd210f8c37e8d48bc32eca6"},
-    {file = "aiohttp-3.10.11-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:968b8fb2a5eee2770eda9c7b5581587ef9b96fbdf8dcabc6b446d35ccc69df01"},
-    {file = "aiohttp-3.10.11-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:deef4362af9493d1382ef86732ee2e4cbc0d7c005947bd54ad1a9a16dd59298e"},
-    {file = "aiohttp-3.10.11-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:686b03196976e327412a1b094f4120778c7c4b9cff9bce8d2fdfeca386b89829"},
-    {file = "aiohttp-3.10.11-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3bf6d027d9d1d34e1c2e1645f18a6498c98d634f8e373395221121f1c258ace8"},
-    {file = "aiohttp-3.10.11-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:099fd126bf960f96d34a760e747a629c27fb3634da5d05c7ef4d35ef4ea519fc"},
-    {file = "aiohttp-3.10.11-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c73c4d3dae0b4644bc21e3de546530531d6cdc88659cdeb6579cd627d3c206aa"},
-    {file = "aiohttp-3.10.11-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:0c5580f3c51eea91559db3facd45d72e7ec970b04528b4709b1f9c2555bd6d0b"},
-    {file = "aiohttp-3.10.11-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:fdf6429f0caabfd8a30c4e2eaecb547b3c340e4730ebfe25139779b9815ba138"},
-    {file = "aiohttp-3.10.11-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:d97187de3c276263db3564bb9d9fad9e15b51ea10a371ffa5947a5ba93ad6777"},
-    {file = "aiohttp-3.10.11-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:0acafb350cfb2eba70eb5d271f55e08bd4502ec35e964e18ad3e7d34d71f7261"},
-    {file = "aiohttp-3.10.11-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:c13ed0c779911c7998a58e7848954bd4d63df3e3575f591e321b19a2aec8df9f"},
-    {file = "aiohttp-3.10.11-cp39-cp39-win32.whl", hash = "sha256:22b7c540c55909140f63ab4f54ec2c20d2635c0289cdd8006da46f3327f971b9"},
-    {file = "aiohttp-3.10.11-cp39-cp39-win_amd64.whl", hash = "sha256:7b26b1551e481012575dab8e3727b16fe7dd27eb2711d2e63ced7368756268fb"},
-    {file = "aiohttp-3.10.11.tar.gz", hash = "sha256:9dc2b8f3dcab2e39e0fa309c8da50c3b55e6f34ab25f1a71d3288f24924d33a7"},
+    {file = "aiohttp-3.12.14-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:906d5075b5ba0dd1c66fcaaf60eb09926a9fef3ca92d912d2a0bbdbecf8b1248"},
+    {file = "aiohttp-3.12.14-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c875bf6fc2fd1a572aba0e02ef4e7a63694778c5646cdbda346ee24e630d30fb"},
+    {file = "aiohttp-3.12.14-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:fbb284d15c6a45fab030740049d03c0ecd60edad9cd23b211d7e11d3be8d56fd"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:38e360381e02e1a05d36b223ecab7bc4a6e7b5ab15760022dc92589ee1d4238c"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:aaf90137b5e5d84a53632ad95ebee5c9e3e7468f0aab92ba3f608adcb914fa95"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:e532a25e4a0a2685fa295a31acf65e027fbe2bea7a4b02cdfbbba8a064577663"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:eab9762c4d1b08ae04a6c77474e6136da722e34fdc0e6d6eab5ee93ac29f35d1"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:abe53c3812b2899889a7fca763cdfaeee725f5be68ea89905e4275476ffd7e61"},
+    {file = "aiohttp-3.12.14-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5760909b7080aa2ec1d320baee90d03b21745573780a072b66ce633eb77a8656"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:02fcd3f69051467bbaa7f84d7ec3267478c7df18d68b2e28279116e29d18d4f3"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:4dcd1172cd6794884c33e504d3da3c35648b8be9bfa946942d353b939d5f1288"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:224d0da41355b942b43ad08101b1b41ce633a654128ee07e36d75133443adcda"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:e387668724f4d734e865c1776d841ed75b300ee61059aca0b05bce67061dcacc"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:dec9cde5b5a24171e0b0a4ca064b1414950904053fb77c707efd876a2da525d8"},
+    {file = "aiohttp-3.12.14-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:bbad68a2af4877cc103cd94af9160e45676fc6f0c14abb88e6e092b945c2c8e3"},
+    {file = "aiohttp-3.12.14-cp310-cp310-win32.whl", hash = "sha256:ee580cb7c00bd857b3039ebca03c4448e84700dc1322f860cf7a500a6f62630c"},
+    {file = "aiohttp-3.12.14-cp310-cp310-win_amd64.whl", hash = "sha256:cf4f05b8cea571e2ccc3ca744e35ead24992d90a72ca2cf7ab7a2efbac6716db"},
+    {file = "aiohttp-3.12.14-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:f4552ff7b18bcec18b60a90c6982049cdb9dac1dba48cf00b97934a06ce2e597"},
+    {file = "aiohttp-3.12.14-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:8283f42181ff6ccbcf25acaae4e8ab2ff7e92b3ca4a4ced73b2c12d8cd971393"},
+    {file = "aiohttp-3.12.14-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:040afa180ea514495aaff7ad34ec3d27826eaa5d19812730fe9e529b04bb2179"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b413c12f14c1149f0ffd890f4141a7471ba4b41234fe4fd4a0ff82b1dc299dbb"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:1d6f607ce2e1a93315414e3d448b831238f1874b9968e1195b06efaa5c87e245"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:565e70d03e924333004ed101599902bba09ebb14843c8ea39d657f037115201b"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4699979560728b168d5ab63c668a093c9570af2c7a78ea24ca5212c6cdc2b641"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ad5fdf6af93ec6c99bf800eba3af9a43d8bfd66dce920ac905c817ef4a712afe"},
+    {file = "aiohttp-3.12.14-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4ac76627c0b7ee0e80e871bde0d376a057916cb008a8f3ffc889570a838f5cc7"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:798204af1180885651b77bf03adc903743a86a39c7392c472891649610844635"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:4f1205f97de92c37dd71cf2d5bcfb65fdaed3c255d246172cce729a8d849b4da"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:76ae6f1dd041f85065d9df77c6bc9c9703da9b5c018479d20262acc3df97d419"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:a194ace7bc43ce765338ca2dfb5661489317db216ea7ea700b0332878b392cab"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:16260e8e03744a6fe3fcb05259eeab8e08342c4c33decf96a9dad9f1187275d0"},
+    {file = "aiohttp-3.12.14-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:8c779e5ebbf0e2e15334ea404fcce54009dc069210164a244d2eac8352a44b28"},
+    {file = "aiohttp-3.12.14-cp311-cp311-win32.whl", hash = "sha256:a289f50bf1bd5be227376c067927f78079a7bdeccf8daa6a9e65c38bae14324b"},
+    {file = "aiohttp-3.12.14-cp311-cp311-win_amd64.whl", hash = "sha256:0b8a69acaf06b17e9c54151a6c956339cf46db4ff72b3ac28516d0f7068f4ced"},
+    {file = "aiohttp-3.12.14-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:a0ecbb32fc3e69bc25efcda7d28d38e987d007096cbbeed04f14a6662d0eee22"},
+    {file = "aiohttp-3.12.14-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:0400f0ca9bb3e0b02f6466421f253797f6384e9845820c8b05e976398ac1d81a"},
+    {file = "aiohttp-3.12.14-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:a56809fed4c8a830b5cae18454b7464e1529dbf66f71c4772e3cfa9cbec0a1ff"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:27f2e373276e4755691a963e5d11756d093e346119f0627c2d6518208483fb6d"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:ca39e433630e9a16281125ef57ece6817afd1d54c9f1bf32e901f38f16035869"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9c748b3f8b14c77720132b2510a7d9907a03c20ba80f469e58d5dfd90c079a1c"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f0a568abe1b15ce69d4cc37e23020720423f0728e3cb1f9bcd3f53420ec3bfe7"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9888e60c2c54eaf56704b17feb558c7ed6b7439bca1e07d4818ab878f2083660"},
+    {file = "aiohttp-3.12.14-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3006a1dc579b9156de01e7916d38c63dc1ea0679b14627a37edf6151bc530088"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:aa8ec5c15ab80e5501a26719eb48a55f3c567da45c6ea5bb78c52c036b2655c7"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:39b94e50959aa07844c7fe2206b9f75d63cc3ad1c648aaa755aa257f6f2498a9"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:04c11907492f416dad9885d503fbfc5dcb6768d90cad8639a771922d584609d3"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:88167bd9ab69bb46cee91bd9761db6dfd45b6e76a0438c7e884c3f8160ff21eb"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:791504763f25e8f9f251e4688195e8b455f8820274320204f7eafc467e609425"},
+    {file = "aiohttp-3.12.14-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:2785b112346e435dd3a1a67f67713a3fe692d288542f1347ad255683f066d8e0"},
+    {file = "aiohttp-3.12.14-cp312-cp312-win32.whl", hash = "sha256:15f5f4792c9c999a31d8decf444e79fcfd98497bf98e94284bf390a7bb8c1729"},
+    {file = "aiohttp-3.12.14-cp312-cp312-win_amd64.whl", hash = "sha256:3b66e1a182879f579b105a80d5c4bd448b91a57e8933564bf41665064796a338"},
+    {file = "aiohttp-3.12.14-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:3143a7893d94dc82bc409f7308bc10d60285a3cd831a68faf1aa0836c5c3c767"},
+    {file = "aiohttp-3.12.14-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:3d62ac3d506cef54b355bd34c2a7c230eb693880001dfcda0bf88b38f5d7af7e"},
+    {file = "aiohttp-3.12.14-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:48e43e075c6a438937c4de48ec30fa8ad8e6dfef122a038847456bfe7b947b63"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:077b4488411a9724cecc436cbc8c133e0d61e694995b8de51aaf351c7578949d"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:d8c35632575653f297dcbc9546305b2c1133391089ab925a6a3706dfa775ccab"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6b8ce87963f0035c6834b28f061df90cf525ff7c9b6283a8ac23acee6502afd4"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f0a2cf66e32a2563bb0766eb24eae7e9a269ac0dc48db0aae90b575dc9583026"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cdea089caf6d5cde975084a884c72d901e36ef9c2fd972c9f51efbbc64e96fbd"},
+    {file = "aiohttp-3.12.14-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8a7865f27db67d49e81d463da64a59365ebd6b826e0e4847aa111056dcb9dc88"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0ab5b38a6a39781d77713ad930cb5e7feea6f253de656a5f9f281a8f5931b086"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:9b3b15acee5c17e8848d90a4ebc27853f37077ba6aec4d8cb4dbbea56d156933"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:e4c972b0bdaac167c1e53e16a16101b17c6d0ed7eac178e653a07b9f7fad7151"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:7442488b0039257a3bdbc55f7209587911f143fca11df9869578db6c26feeeb8"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:f68d3067eecb64c5e9bab4a26aa11bd676f4c70eea9ef6536b0a4e490639add3"},
+    {file = "aiohttp-3.12.14-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f88d3704c8b3d598a08ad17d06006cb1ca52a1182291f04979e305c8be6c9758"},
+    {file = "aiohttp-3.12.14-cp313-cp313-win32.whl", hash = "sha256:a3c99ab19c7bf375c4ae3debd91ca5d394b98b6089a03231d4c580ef3c2ae4c5"},
+    {file = "aiohttp-3.12.14-cp313-cp313-win_amd64.whl", hash = "sha256:3f8aad695e12edc9d571f878c62bedc91adf30c760c8632f09663e5f564f4baa"},
+    {file = "aiohttp-3.12.14-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:b8cc6b05e94d837bcd71c6531e2344e1ff0fb87abe4ad78a9261d67ef5d83eae"},
+    {file = "aiohttp-3.12.14-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:d1dcb015ac6a3b8facd3677597edd5ff39d11d937456702f0bb2b762e390a21b"},
+    {file = "aiohttp-3.12.14-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:3779ed96105cd70ee5e85ca4f457adbce3d9ff33ec3d0ebcdf6c5727f26b21b3"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:717a0680729b4ebd7569c1dcd718c46b09b360745fd8eb12317abc74b14d14d0"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:b5dd3a2ef7c7e968dbbac8f5574ebeac4d2b813b247e8cec28174a2ba3627170"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4710f77598c0092239bc12c1fcc278a444e16c7032d91babf5abbf7166463f7b"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f3e9f75ae842a6c22a195d4a127263dbf87cbab729829e0bd7857fb1672400b2"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5f9c8d55d6802086edd188e3a7d85a77787e50d56ce3eb4757a3205fa4657922"},
+    {file = "aiohttp-3.12.14-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:79b29053ff3ad307880d94562cca80693c62062a098a5776ea8ef5ef4b28d140"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:23e1332fff36bebd3183db0c7a547a1da9d3b4091509f6d818e098855f2f27d3"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:a564188ce831fd110ea76bcc97085dd6c625b427db3f1dbb14ca4baa1447dcbc"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:a7a1b4302f70bb3ec40ca86de82def532c97a80db49cac6a6700af0de41af5ee"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:1b07ccef62950a2519f9bfc1e5b294de5dd84329f444ca0b329605ea787a3de5"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:938bd3ca6259e7e48b38d84f753d548bd863e0c222ed6ee6ace3fd6752768a84"},
+    {file = "aiohttp-3.12.14-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:8bc784302b6b9f163b54c4e93d7a6f09563bd01ff2b841b29ed3ac126e5040bf"},
+    {file = "aiohttp-3.12.14-cp39-cp39-win32.whl", hash = "sha256:a3416f95961dd7d5393ecff99e3f41dc990fb72eda86c11f2a60308ac6dcd7a0"},
+    {file = "aiohttp-3.12.14-cp39-cp39-win_amd64.whl", hash = "sha256:196858b8820d7f60578f8b47e5669b3195c21d8ab261e39b1d705346458f445f"},
+    {file = "aiohttp-3.12.14.tar.gz", hash = "sha256:6e06e120e34d93100de448fd941522e11dafa78ef1a893c179901b7d66aa29f2"},
 ]
 
 [package.dependencies]
-aiohappyeyeballs = ">=2.3.0"
-aiosignal = ">=1.1.2"
+aiohappyeyeballs = ">=2.5.0"
+aiosignal = ">=1.4.0"
 attrs = ">=17.3.0"
 frozenlist = ">=1.1.1"
 multidict = ">=4.5,<7.0"
-yarl = ">=1.12.0,<2.0"
+propcache = ">=0.2.0"
+yarl = ">=1.17.0,<2.0"
 
 [package.extras]
-speedups = ["Brotli ; platform_python_implementation == \"CPython\"", "aiodns (>=3.2.0) ; sys_platform == \"linux\" or sys_platform == \"darwin\"", "brotlicffi ; platform_python_implementation != \"CPython\""]
+speedups = ["Brotli ; platform_python_implementation == \"CPython\"", "aiodns (>=3.3.0)", "brotlicffi ; platform_python_implementation != \"CPython\""]
 
 [[package]]
 name = "aiopg"
@@ -145,18 +141,19 @@ sa = ["sqlalchemy[postgresql-psycopg2binary] (>=1.3,<1.5)"]
 
 [[package]]
 name = "aiosignal"
-version = "1.3.1"
+version = "1.4.0"
 description = "aiosignal: a list of registered asynchronous callbacks"
 optional = false
-python-versions = ">=3.7"
+python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "aiosignal-1.3.1-py3-none-any.whl", hash = "sha256:f8376fb07dd1e86a584e4fcdec80b36b7f81aac666ebc724e2c090300dd83b17"},
-    {file = "aiosignal-1.3.1.tar.gz", hash = "sha256:54cd96e15e1649b75d6c87526a6ff0b6c1b0dd3459f43d9ca11d48c339b68cfc"},
+    {file = "aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e"},
+    {file = "aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7"},
 ]
 
 [package.dependencies]
 frozenlist = ">=1.1.0"
+typing-extensions = {version = ">=4.2", markers = "python_version < \"3.13\""}
 
 [[package]]
 name = "allure-pytest"
@@ -2326,25 +2323,6 @@ files = [
     {file = "propcache-0.2.0.tar.gz", hash = "sha256:df81779732feb9d01e5d513fad0122efb3d53bbc75f61b2a4f29a020bc985e70"},
 ]
 
-[[package]]
-name = "protobuf"
-version = "6.31.1"
-description = ""
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "protobuf-6.31.1-cp310-abi3-win32.whl", hash = "sha256:7fa17d5a29c2e04b7d90e5e32388b8bfd0e7107cd8e616feef7ed3fa6bdab5c9"},
-    {file = "protobuf-6.31.1-cp310-abi3-win_amd64.whl", hash = "sha256:426f59d2964864a1a366254fa703b8632dcec0790d8862d30034d8245e1cd447"},
-    {file = "protobuf-6.31.1-cp39-abi3-macosx_10_9_universal2.whl", hash = "sha256:6f1227473dc43d44ed644425268eb7c2e488ae245d51c6866d19fe158e207402"},
-    {file = "protobuf-6.31.1-cp39-abi3-manylinux2014_aarch64.whl", hash = "sha256:a40fc12b84c154884d7d4c4ebd675d5b3b5283e155f324049ae396b95ddebc39"},
-    {file = "protobuf-6.31.1-cp39-abi3-manylinux2014_x86_64.whl", hash = "sha256:4ee898bf66f7a8b0bd21bce523814e6fbd8c6add948045ce958b73af7e8878c6"},
-    {file = "protobuf-6.31.1-cp39-cp39-win32.whl", hash = "sha256:0414e3aa5a5f3ff423828e1e6a6e907d6c65c1d5b7e6e975793d5590bdeecc16"},
-    {file = "protobuf-6.31.1-cp39-cp39-win_amd64.whl", hash = "sha256:8764cf4587791e7564051b35524b72844f845ad0bb011704c3736cce762d8fe9"},
-    {file = "protobuf-6.31.1-py3-none-any.whl", hash = "sha256:720a6c7e6b77288b85063569baae8536671b39f15cc22037ec7045658d80489e"},
-    {file = "protobuf-6.31.1.tar.gz", hash = "sha256:d8cac4c982f0b957a4dc73a80e2ea24fab08e679c0de9deb835f4a12d69aca9a"},
-]
-
 [[package]]
 name = "psutil"
 version = "5.9.4"
@@ -3328,18 +3306,6 @@ files = [
 [package.dependencies]
 cryptography = "*"
 
-[[package]]
-name = "types-protobuf"
-version = "6.30.2.20250516"
-description = "Typing stubs for protobuf"
-optional = false
-python-versions = ">=3.9"
-groups = ["main"]
-files = [
-    {file = "types_protobuf-6.30.2.20250516-py3-none-any.whl", hash = "sha256:8c226d05b5e8b2623111765fa32d6e648bbc24832b4c2fddf0fa340ba5d5b722"},
-    {file = "types_protobuf-6.30.2.20250516.tar.gz", hash = "sha256:aecd1881770a9bb225ede66872ef7f0da4505edd0b193108edd9892e48d49a41"},
-]
-
 [[package]]
 name = "types-psutil"
 version = "5.9.5.12"
@@ -3878,4 +3844,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.1"
 python-versions = "^3.11"
-content-hash = "7cc735f57c2760db6c994575a98d4f0e2670497ad9e909135a3bc67d479f5edf"
+content-hash = "6a1e8ba06b8194bf28d87fd5e184e2ddc2b4a19dffcbe3953b26da3d55c9212f"

proxy/Cargo.toml

@@ -16,6 +16,7 @@ async-compression.workspace = true
 async-trait.workspace = true
 atomic-take.workspace = true
 aws-config.workspace = true
+aws-credential-types.workspace = true
 aws-sdk-iam.workspace = true
 aws-sigv4.workspace = true
 base64.workspace = true
@@ -48,6 +49,7 @@ indexmap = { workspace = true, features = ["serde"] }
 ipnet.workspace = true
 itertools.workspace = true
 itoa.workspace = true
+json = { path = "../libs/proxy/json" }
 lasso = { workspace = true, features = ["multi-threaded"] }
 measured = { workspace = true, features = ["lasso"] }
 metrics.workspace = true
@@ -118,6 +120,7 @@ workspace_hack.workspace = true
 [dev-dependencies]
 assert-json-diff.workspace = true
 camino-tempfile.workspace = true
+criterion.workspace = true
 fallible-iterator.workspace = true
 flate2.workspace = true
 tokio-tungstenite.workspace = true
@@ -127,4 +130,9 @@ rstest.workspace = true
 walkdir.workspace = true
 rand_distr = "0.4"
 tokio-postgres.workspace = true
-tracing-test = "0.2"
+tracing-test = "0.2"
+
+[[bench]]
+name = "logging"
+harness = false
+

proxy/README.md

@@ -123,6 +123,11 @@ docker exec -it proxy-postgres psql -U postgres -c "CREATE TABLE neon_control_pl
 docker exec -it proxy-postgres psql -U postgres -c "CREATE ROLE proxy WITH SUPERUSER LOGIN PASSWORD 'password';"
 ```
 
+If you want to test query cancellation, redis is also required:
+```sh
+docker run --detach --name proxy-redis --publish 6379:6379 redis:7.0
+```
+
 Let's create self-signed certificate by running:
 ```sh
 openssl req -new -x509 -days 365 -nodes -text -out server.crt -keyout server.key -subj "/CN=*.local.neon.build"
@@ -130,7 +135,10 @@ openssl req -new -x509 -days 365 -nodes -text -out server.crt -keyout server.key
 
 Then we need to build proxy with 'testing' feature and run, e.g.:
 ```sh
-RUST_LOG=proxy LOGFMT=text cargo run -p proxy --bin proxy --features testing -- --auth-backend postgres --auth-endpoint 'postgresql://postgres:proxy-postgres@127.0.0.1:5432/postgres' -c server.crt -k server.key
+RUST_LOG=proxy LOGFMT=text cargo run -p proxy --bin proxy --features testing -- \
+  --auth-backend postgres --auth-endpoint 'postgresql://postgres:proxy-postgres@127.0.0.1:5432/postgres' \
+  --redis-auth-type="plain" --redis-plain="redis://127.0.0.1:6379" \
+  -c server.crt -k server.key
 ```
 
 Now from client you can start a new session:

proxy/benches/logging.rs

@@ -0,0 +1,127 @@
+use std::io;
+
+use criterion::{Criterion, criterion_group, criterion_main};
+use proxy::logging::{Clock, JsonLoggingLayer};
+use tracing_subscriber::prelude::*;
+
+struct DevNullWriter;
+
+impl proxy::logging::MakeWriter for DevNullWriter {
+    fn make_writer(&self) -> impl io::Write {
+        DevNullWriter
+    }
+}
+
+impl<'a> tracing_subscriber::fmt::MakeWriter<'a> for DevNullWriter {
+    type Writer = DevNullWriter;
+    fn make_writer(&'a self) -> Self::Writer {
+        DevNullWriter
+    }
+}
+
+impl io::Write for DevNullWriter {
+    #[inline]
+    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+        Ok(criterion::black_box(buf).len())
+    }
+
+    #[inline(always)]
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
+
+struct FixedClock;
+
+impl Clock for FixedClock {
+    fn now(&self) -> chrono::DateTime<chrono::Utc> {
+        const { chrono::DateTime::from_timestamp_nanos(1747859990_000_000_000).to_utc() }
+    }
+}
+
+pub fn bench_logging(c: &mut Criterion) {
+    c.bench_function("text fmt current", |b| {
+        let registry = tracing_subscriber::Registry::default().with(
+            tracing_subscriber::fmt::layer()
+                .with_ansi(false)
+                .with_target(false)
+                .with_writer(DevNullWriter),
+        );
+
+        tracing::subscriber::with_default(registry, || {
+            tracing::info_span!("span1", a = 42, b = true, c = "string").in_scope(|| {
+                tracing::info_span!("span2", a = 42, b = true, c = "string").in_scope(|| {
+                    b.iter(|| {
+                        tracing::error!(a = 42, b = true, c = "string", "message field");
+                    })
+                });
+            });
+        });
+    });
+
+    c.bench_function("text fmt full", |b| {
+        let registry = tracing_subscriber::Registry::default().with(
+            tracing_subscriber::fmt::layer()
+                .with_level(true)
+                .with_file(true)
+                .with_line_number(true)
+                .with_target(true)
+                .with_thread_ids(true)
+                .with_writer(DevNullWriter),
+        );
+
+        tracing::subscriber::with_default(registry, || {
+            tracing::info_span!("span1", a = 42, b = true, c = "string").in_scope(|| {
+                tracing::info_span!("span2", a = 42, b = true, c = "string").in_scope(|| {
+                    b.iter(|| {
+                        tracing::error!(a = 42, b = true, c = "string", "message field");
+                    })
+                });
+            });
+        });
+    });
+
+    c.bench_function("json fmt", |b| {
+        let registry = tracing_subscriber::Registry::default().with(
+            tracing_subscriber::fmt::layer()
+                .with_level(true)
+                .with_file(true)
+                .with_line_number(true)
+                .with_target(true)
+                .with_thread_ids(true)
+                .with_writer(DevNullWriter)
+                .json(),
+        );
+
+        tracing::subscriber::with_default(registry, || {
+            tracing::info_span!("span1", a = 42, b = true, c = "string").in_scope(|| {
+                tracing::info_span!("span2", a = 42, b = true, c = "string").in_scope(|| {
+                    b.iter(|| {
+                        tracing::error!(a = 42, b = true, c = "string", "message field");
+                    })
+                });
+            });
+        });
+    });
+
+    c.bench_function("json custom", |b| {
+        let registry = tracing_subscriber::Registry::default().with(JsonLoggingLayer::new(
+            FixedClock,
+            DevNullWriter,
+            &["a"],
+        ));
+
+        tracing::subscriber::with_default(registry, || {
+            tracing::info_span!("span1", a = 42, b = true, c = "string").in_scope(|| {
+                tracing::info_span!("span2", a = 42, b = true, c = "string").in_scope(|| {
+                    b.iter(|| {
+                        tracing::error!(a = 42, b = true, c = "string", "message field");
+                    })
+                });
+            });
+        });
+    });
+}
+
+criterion_group!(benches, bench_logging);
+criterion_main!(benches);

proxy/src/batch.rs

@@ -7,13 +7,17 @@ use std::pin::pin;
 use std::sync::Mutex;
 
 use scopeguard::ScopeGuard;
+use tokio::sync::oneshot;
 use tokio::sync::oneshot::error::TryRecvError;
 
 use crate::ext::LockExt;
 
+type ProcResult<P> = Result<<P as QueueProcessing>::Res, <P as QueueProcessing>::Err>;
+
 pub trait QueueProcessing: Send + 'static {
     type Req: Send + 'static;
     type Res: Send;
+    type Err: Send + Clone;
 
     /// Get the desired batch size.
     fn batch_size(&self, queue_size: usize) -> usize;
@@ -24,7 +28,18 @@ pub trait QueueProcessing: Send + 'static {
     /// If this apply can error, it's expected that errors be forwarded to each Self::Res.
     ///
     /// Batching does not need to happen atomically.
-    fn apply(&mut self, req: Vec<Self::Req>) -> impl Future<Output = Vec<Self::Res>> + Send;
+    fn apply(
+        &mut self,
+        req: Vec<Self::Req>,
+    ) -> impl Future<Output = Result<Vec<Self::Res>, Self::Err>> + Send;
+}
+
+#[derive(thiserror::Error)]
+pub enum BatchQueueError<E: Clone, C> {
+    #[error(transparent)]
+    Result(E),
+    #[error(transparent)]
+    Cancelled(C),
 }
 
 pub struct BatchQueue<P: QueueProcessing> {
@@ -34,7 +49,7 @@ pub struct BatchQueue<P: QueueProcessing> {
 
 struct BatchJob<P: QueueProcessing> {
     req: P::Req,
-    res: tokio::sync::oneshot::Sender<P::Res>,
+    res: tokio::sync::oneshot::Sender<Result<P::Res, P::Err>>,
 }
 
 impl<P: QueueProcessing> BatchQueue<P> {
@@ -55,11 +70,11 @@ impl<P: QueueProcessing> BatchQueue<P> {
         &self,
         req: P::Req,
         cancelled: impl Future<Output = R>,
-    ) -> Result<P::Res, R> {
+    ) -> Result<P::Res, BatchQueueError<P::Err, R>> {
         let (id, mut rx) = self.inner.lock_propagate_poison().register_job(req);
 
         let mut cancelled = pin!(cancelled);
-        let resp = loop {
+        let resp: Option<Result<P::Res, P::Err>> = loop {
             // try become the leader, or try wait for success.
             let mut processor = tokio::select! {
                 // try become leader.
@@ -72,7 +87,7 @@ impl<P: QueueProcessing> BatchQueue<P> {
                     if inner.queue.remove(&id).is_some() {
                         tracing::warn!("batched task cancelled before completion");
                     }
-                    return Err(cancel);
+                    return Err(BatchQueueError::Cancelled(cancel));
                 },
             };
 
@@ -96,18 +111,30 @@ impl<P: QueueProcessing> BatchQueue<P> {
             // good: we didn't get cancelled.
             ScopeGuard::into_inner(cancel_safety);
 
-            if values.len() != resps.len() {
-                tracing::error!(
-                    "batch: invalid response size, expected={}, got={}",
-                    resps.len(),
-                    values.len()
-                );
-            }
+            match values {
+                Ok(values) => {
+                    if values.len() != resps.len() {
+                        tracing::error!(
+                            "batch: invalid response size, expected={}, got={}",
+                            resps.len(),
+                            values.len()
+                        );
+                    }
 
-            // send response values.
-            for (tx, value) in std::iter::zip(resps, values) {
-                if tx.send(value).is_err() {
-                    // receiver hung up but that's fine.
+                    // send response values.
+                    for (tx, value) in std::iter::zip(resps, values) {
+                        if tx.send(Ok(value)).is_err() {
+                            // receiver hung up but that's fine.
+                        }
+                    }
+                }
+
+                Err(err) => {
+                    for tx in resps {
+                        if tx.send(Err(err.clone())).is_err() {
+                            // receiver hung up but that's fine.
+                        }
+                    }
                 }
             }
 
@@ -129,7 +156,8 @@ impl<P: QueueProcessing> BatchQueue<P> {
 
         tracing::debug!(id, "batch: job completed");
 
-        Ok(resp.expect("no response found. batch processer should not panic"))
+        resp.expect("no response found. batch processer should not panic")
+            .map_err(BatchQueueError::Result)
     }
 }
 
@@ -139,8 +167,8 @@ struct BatchQueueInner<P: QueueProcessing> {
 }
 
 impl<P: QueueProcessing> BatchQueueInner<P> {
-    fn register_job(&mut self, req: P::Req) -> (u64, tokio::sync::oneshot::Receiver<P::Res>) {
-        let (tx, rx) = tokio::sync::oneshot::channel();
+    fn register_job(&mut self, req: P::Req) -> (u64, oneshot::Receiver<ProcResult<P>>) {
+        let (tx, rx) = oneshot::channel();
 
         let id = self.version;
 
@@ -158,7 +186,7 @@ impl<P: QueueProcessing> BatchQueueInner<P> {
         (id, rx)
     }
 
-    fn get_batch(&mut self, p: &P) -> (Vec<P::Req>, Vec<tokio::sync::oneshot::Sender<P::Res>>) {
+    fn get_batch(&mut self, p: &P) -> (Vec<P::Req>, Vec<oneshot::Sender<ProcResult<P>>>) {
         let batch_size = p.batch_size(self.queue.len());
         let mut reqs = Vec::with_capacity(batch_size);
         let mut resps = Vec::with_capacity(batch_size);

proxy/src/binary/proxy.rs

@@ -522,15 +522,7 @@ pub async fn run() -> anyhow::Result<()> {
         maintenance_tasks.spawn(usage_metrics::task_main(metrics_config));
     }
 
-    if let Either::Left(auth::Backend::ControlPlane(api, ())) = &auth_backend
-        && let crate::control_plane::client::ControlPlaneClient::ProxyV1(api) = &**api
-        && let Some(client) = redis_client
-    {
-        // project info cache and invalidation of that cache.
-        let cache = api.caches.project_info.clone();
-        maintenance_tasks.spawn(notifications::task_main(client.clone(), cache.clone()));
-        maintenance_tasks.spawn(async move { cache.clone().gc_worker().await });
-
+    if let Some(client) = redis_client {
         // Try to connect to Redis 3 times with 1 + (0..0.1) second interval.
         // This prevents immediate exit and pod restart,
         // which can cause hammering of the redis in case of connection issues.
@@ -560,6 +552,16 @@ pub async fn run() -> anyhow::Result<()> {
                 }
             }
         }
+
+        #[allow(irrefutable_let_patterns)]
+        if let Either::Left(auth::Backend::ControlPlane(api, ())) = &auth_backend
+            && let crate::control_plane::client::ControlPlaneClient::ProxyV1(api) = &**api
+        {
+            // project info cache and invalidation of that cache.
+            let cache = api.caches.project_info.clone();
+            maintenance_tasks.spawn(notifications::task_main(client, cache.clone()));
+            maintenance_tasks.spawn(async move { cache.gc_worker().await });
+        }
     }
 
     let maintenance = loop {

proxy/src/cache/timed_lru.rs

@@ -14,8 +14,8 @@ use std::time::{Duration, Instant};
 use hashlink::{LruCache, linked_hash_map::RawEntryMut};
 use tracing::debug;
 
+use super::Cache;
 use super::common::Cached;
-use super::{Cache, timed_lru};
 
 /// An implementation of timed LRU cache with fixed capacity.
 /// Key properties:
@@ -30,7 +30,7 @@ use super::{Cache, timed_lru};
 ///
 /// * There's an API for immediate invalidation (removal) of a cache entry;
 ///   It's useful in case we know for sure that the entry is no longer correct.
-///   See [`timed_lru::Cached`] for more information.
+///   See [`Cached`] for more information.
 ///
 /// * Expired entries are kept in the cache, until they are evicted by the LRU policy,
 ///   or by a successful lookup (i.e. the entry hasn't expired yet).
@@ -217,15 +217,18 @@ impl<K: Hash + Eq + Clone, V: Clone> TimedLru<K, V> {
 }
 
 impl<K: Hash + Eq, V: Clone> TimedLru<K, V> {
-    /// Retrieve a cached entry in convenient wrapper.
-    pub(crate) fn get<Q>(&self, key: &Q) -> Option<timed_lru::Cached<&Self>>
+    /// Retrieve a cached entry in convenient wrapper, alongside timing information.
+    pub(crate) fn get_with_created_at<Q>(
+        &self,
+        key: &Q,
+    ) -> Option<Cached<&Self, (<Self as Cache>::Value, Instant)>>
     where
         K: Borrow<Q> + Clone,
         Q: Hash + Eq + ?Sized,
     {
         self.get_raw(key, |key, entry| Cached {
             token: Some((self, key.clone())),
-            value: entry.value.clone(),
+            value: (entry.value.clone(), entry.created_at),
         })
     }
 }

proxy/src/cancellation.rs

@@ -4,12 +4,11 @@ use std::pin::pin;
 use std::sync::{Arc, OnceLock};
 use std::time::Duration;
 
-use anyhow::anyhow;
 use futures::FutureExt;
 use ipnet::{IpNet, Ipv4Net, Ipv6Net};
 use postgres_client::RawCancelToken;
 use postgres_client::tls::MakeTlsConnect;
-use redis::{Cmd, FromRedisValue, Value};
+use redis::{Cmd, FromRedisValue, SetExpiry, SetOptions, Value};
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
 use tokio::net::TcpStream;
@@ -18,7 +17,7 @@ use tracing::{debug, error, info};
 
 use crate::auth::AuthError;
 use crate::auth::backend::ComputeUserInfo;
-use crate::batch::{BatchQueue, QueueProcessing};
+use crate::batch::{BatchQueue, BatchQueueError, QueueProcessing};
 use crate::config::ComputeConfig;
 use crate::context::RequestContext;
 use crate::control_plane::ControlPlaneApi;
@@ -28,23 +27,39 @@ use crate::metrics::{CancelChannelSizeGuard, CancellationRequest, Metrics, Redis
 use crate::pqproto::CancelKeyData;
 use crate::rate_limiter::LeakyBucketRateLimiter;
 use crate::redis::keys::KeyPrefix;
-use crate::redis::kv_ops::RedisKVClient;
+use crate::redis::kv_ops::{RedisKVClient, RedisKVClientError};
+use crate::util::run_until;
 
 type IpSubnetKey = IpNet;
 
-const CANCEL_KEY_TTL: std::time::Duration = std::time::Duration::from_secs(600);
-const CANCEL_KEY_REFRESH: std::time::Duration = std::time::Duration::from_secs(570);
+const CANCEL_KEY_TTL: Duration = Duration::from_secs(600);
+const CANCEL_KEY_REFRESH: Duration = Duration::from_secs(570);
 
 // Message types for sending through mpsc channel
 pub enum CancelKeyOp {
-    StoreCancelKey {
+    Store {
         key: CancelKeyData,
         value: Box<str>,
-        expire: std::time::Duration,
+        expire: Duration,
     },
-    GetCancelData {
+    Refresh {
+        key: CancelKeyData,
+        expire: Duration,
+    },
+    Get {
         key: CancelKeyData,
     },
+    GetOld {
+        key: CancelKeyData,
+    },
+}
+
+#[derive(thiserror::Error, Debug, Clone)]
+pub enum PipelineError {
+    #[error("could not send cmd to redis: {0}")]
+    RedisKVClient(Arc<RedisKVClientError>),
+    #[error("incorrect number of responses from redis")]
+    IncorrectNumberOfResponses,
 }
 
 pub struct Pipeline {
@@ -60,7 +75,7 @@ impl Pipeline {
         }
     }
 
-    async fn execute(self, client: &mut RedisKVClient) -> Vec<anyhow::Result<Value>> {
+    async fn execute(self, client: &mut RedisKVClient) -> Result<Vec<Value>, PipelineError> {
         let responses = self.replies;
         let batch_size = self.inner.len();
 
@@ -78,43 +93,44 @@ impl Pipeline {
                     batch_size,
                     responses, "successfully completed cancellation jobs",
                 );
-                values.into_iter().map(Ok).collect()
+                Ok(values.into_iter().collect())
             }
             Ok(value) => {
                 error!(batch_size, ?value, "unexpected redis return value");
-                std::iter::repeat_with(|| Err(anyhow!("incorrect response type from redis")))
-                    .take(responses)
-                    .collect()
-            }
-            Err(err) => {
-                std::iter::repeat_with(|| Err(anyhow!("could not send cmd to redis: {err}")))
-                    .take(responses)
-                    .collect()
+                Err(PipelineError::IncorrectNumberOfResponses)
             }
+            Err(err) => Err(PipelineError::RedisKVClient(Arc::new(err))),
         }
     }
 
-    fn add_command_with_reply(&mut self, cmd: Cmd) {
+    fn add_command(&mut self, cmd: Cmd) {
         self.inner.add_command(cmd);
         self.replies += 1;
     }
-
-    fn add_command_no_reply(&mut self, cmd: Cmd) {
-        self.inner.add_command(cmd).ignore();
-    }
 }
 
 impl CancelKeyOp {
     fn register(&self, pipe: &mut Pipeline) {
         match self {
-            CancelKeyOp::StoreCancelKey { key, value, expire } => {
+            CancelKeyOp::Store { key, value, expire } => {
                 let key = KeyPrefix::Cancel(*key).build_redis_key();
-                pipe.add_command_with_reply(Cmd::hset(&key, "data", &**value));
-                pipe.add_command_no_reply(Cmd::expire(&key, expire.as_secs() as i64));
+                pipe.add_command(Cmd::set_options(
+                    &key,
+                    &**value,
+                    SetOptions::default().with_expiration(SetExpiry::EX(expire.as_secs())),
+                ));
             }
-            CancelKeyOp::GetCancelData { key } => {
+            CancelKeyOp::Refresh { key, expire } => {
                 let key = KeyPrefix::Cancel(*key).build_redis_key();
-                pipe.add_command_with_reply(Cmd::hget(key, "data"));
+                pipe.add_command(Cmd::expire(&key, expire.as_secs() as i64));
+            }
+            CancelKeyOp::GetOld { key } => {
+                let key = KeyPrefix::Cancel(*key).build_redis_key();
+                pipe.add_command(Cmd::hget(key, "data"));
+            }
+            CancelKeyOp::Get { key } => {
+                let key = KeyPrefix::Cancel(*key).build_redis_key();
+                pipe.add_command(Cmd::get(key));
             }
         }
     }
@@ -127,13 +143,14 @@ pub struct CancellationProcessor {
 
 impl QueueProcessing for CancellationProcessor {
     type Req = (CancelChannelSizeGuard<'static>, CancelKeyOp);
-    type Res = anyhow::Result<redis::Value>;
+    type Res = redis::Value;
+    type Err = PipelineError;
 
     fn batch_size(&self, _queue_size: usize) -> usize {
         self.batch_size
     }
 
-    async fn apply(&mut self, batch: Vec<Self::Req>) -> Vec<Self::Res> {
+    async fn apply(&mut self, batch: Vec<Self::Req>) -> Result<Vec<Self::Res>, Self::Err> {
         if !self.client.credentials_refreshed() {
             // this will cause a timeout for cancellation operations
             tracing::debug!(
@@ -244,18 +261,18 @@ impl CancellationHandler {
         &self,
         key: CancelKeyData,
     ) -> Result<Option<CancelClosure>, CancelError> {
-        let guard = Metrics::get()
-            .proxy
-            .cancel_channel_size
-            .guard(RedisMsgKind::HGet);
-        let op = CancelKeyOp::GetCancelData { key };
+        const TIMEOUT: Duration = Duration::from_secs(5);
 
         let Some(tx) = self.tx.get() else {
             tracing::warn!("cancellation handler is not available");
             return Err(CancelError::InternalError);
         };
 
-        const TIMEOUT: Duration = Duration::from_secs(5);
+        let guard = Metrics::get()
+            .proxy
+            .cancel_channel_size
+            .guard(RedisMsgKind::Get);
+        let op = CancelKeyOp::Get { key };
         let result = timeout(
             TIMEOUT,
             tx.call((guard, op), std::future::pending::<Infallible>()),
@@ -264,10 +281,37 @@ impl CancellationHandler {
         .map_err(|_| {
             tracing::warn!("timed out waiting to receive GetCancelData response");
             CancelError::RateLimit
-        })?
-        // cannot be cancelled
-        .unwrap_or_else(|x| match x {})
-        .map_err(|e| {
+        })?;
+
+        // We may still have cancel keys set with HSET <key> "data".
+        // Check error type and retry with HGET.
+        // TODO: remove code after HSET is not used anymore.
+        let result = if let Err(err) = result.as_ref()
+            && let BatchQueueError::Result(err) = err
+            && let PipelineError::RedisKVClient(err) = err
+            && let RedisKVClientError::Redis(err) = &**err
+            && let Some(errcode) = err.code()
+            && errcode == "WRONGTYPE"
+        {
+            let guard = Metrics::get()
+                .proxy
+                .cancel_channel_size
+                .guard(RedisMsgKind::HGet);
+            let op = CancelKeyOp::GetOld { key };
+            timeout(
+                TIMEOUT,
+                tx.call((guard, op), std::future::pending::<Infallible>()),
+            )
+            .await
+            .map_err(|_| {
+                tracing::warn!("timed out waiting to receive GetCancelData response");
+                CancelError::RateLimit
+            })?
+        } else {
+            result
+        };
+
+        let result = result.map_err(|e| {
             tracing::warn!("failed to receive GetCancelData response: {e}");
             CancelError::InternalError
         })?;
@@ -438,39 +482,94 @@ impl Session {
 
         let mut cancel = pin!(cancel);
 
+        enum State {
+            Set,
+            Refresh,
+        }
+        let mut state = State::Set;
+
         loop {
-            let guard = Metrics::get()
-                .proxy
-                .cancel_channel_size
-                .guard(RedisMsgKind::HSet);
-            let op = CancelKeyOp::StoreCancelKey {
-                key: self.key,
-                value: closure_json.clone(),
-                expire: CANCEL_KEY_TTL,
+            let guard_op = match state {
+                State::Set => {
+                    let guard = Metrics::get()
+                        .proxy
+                        .cancel_channel_size
+                        .guard(RedisMsgKind::Set);
+                    let op = CancelKeyOp::Store {
+                        key: self.key,
+                        value: closure_json.clone(),
+                        expire: CANCEL_KEY_TTL,
+                    };
+                    tracing::debug!(
+                        src=%self.key,
+                        dest=?cancel_closure.cancel_token,
+                        "registering cancellation key"
+                    );
+                    (guard, op)
+                }
+
+                State::Refresh => {
+                    let guard = Metrics::get()
+                        .proxy
+                        .cancel_channel_size
+                        .guard(RedisMsgKind::Expire);
+                    let op = CancelKeyOp::Refresh {
+                        key: self.key,
+                        expire: CANCEL_KEY_TTL,
+                    };
+                    tracing::debug!(
+                        src=%self.key,
+                        dest=?cancel_closure.cancel_token,
+                        "refreshing cancellation key"
+                    );
+                    (guard, op)
+                }
             };
 
-            tracing::debug!(
-                src=%self.key,
-                dest=?cancel_closure.cancel_token,
-                "registering cancellation key"
-            );
-
-            match tx.call((guard, op), cancel.as_mut()).await {
-                Ok(Ok(_)) => {
+            match tx.call(guard_op, cancel.as_mut()).await {
+                // SET returns OK
+                Ok(Value::Okay) => {
                     tracing::debug!(
                         src=%self.key,
                         dest=?cancel_closure.cancel_token,
                         "registered cancellation key"
                     );
+                    state = State::Refresh;
+                }
 
-                    // wait before continuing.
-                    tokio::time::sleep(CANCEL_KEY_REFRESH).await;
+                // EXPIRE returns 1
+                Ok(Value::Int(1)) => {
+                    tracing::debug!(
+                        src=%self.key,
+                        dest=?cancel_closure.cancel_token,
+                        "refreshed cancellation key"
+                    );
                 }
+
+                Ok(_) => {
+                    // Any other response likely means the key expired.
+                    tracing::warn!(src=%self.key, "refreshing cancellation key failed");
+                    // Re-enter the SET loop to repush full data.
+                    state = State::Set;
+                }
+
                 // retry immediately.
-                Ok(Err(error)) => {
-                    tracing::warn!(?error, "error registering cancellation key");
+                Err(BatchQueueError::Result(error)) => {
+                    tracing::warn!(?error, "error refreshing cancellation key");
+                    // Small delay to prevent busy loop with high cpu and logging.
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                    continue;
                 }
-                Err(Err(_cancelled)) => break,
+
+                Err(BatchQueueError::Cancelled(Err(_cancelled))) => break,
+            }
+
+            // wait before continuing. break immediately if cancelled.
+            if run_until(tokio::time::sleep(CANCEL_KEY_REFRESH), cancel.as_mut())
+                .await
+                .is_err()
+            {
+                break;
             }
         }
 

proxy/src/control_plane/client/cplane_proxy_v1.rs

@@ -23,12 +23,13 @@ use crate::control_plane::errors::{
     ControlPlaneError, GetAuthInfoError, GetEndpointJwksError, WakeComputeError,
 };
 use crate::control_plane::locks::ApiLocks;
-use crate::control_plane::messages::{ColdStartInfo, EndpointJwksResponse, Reason};
+use crate::control_plane::messages::{ColdStartInfo, EndpointJwksResponse};
 use crate::control_plane::{
     AccessBlockerFlags, AuthInfo, AuthSecret, CachedNodeInfo, EndpointAccessControl, NodeInfo,
     RoleAccessControl,
 };
 use crate::metrics::Metrics;
+use crate::proxy::retry::CouldRetry;
 use crate::rate_limiter::WakeComputeRateLimiter;
 use crate::types::{EndpointCacheKey, EndpointId, RoleName};
 use crate::{compute, http, scram};
@@ -382,16 +383,31 @@ impl super::ControlPlaneApi for NeonControlPlaneClient {
 
         macro_rules! check_cache {
             () => {
-                if let Some(cached) = self.caches.node_info.get(&key) {
-                    let (cached, info) = cached.take_value();
-                    let info = info.map_err(|c| {
-                        info!(key = &*key, "found cached wake_compute error");
-                        WakeComputeError::ControlPlane(ControlPlaneError::Message(Box::new(*c)))
-                    })?;
+                if let Some(cached) = self.caches.node_info.get_with_created_at(&key) {
+                    let (cached, (info, created_at)) = cached.take_value();
+                    return match info {
+                        Err(mut msg) => {
+                            info!(key = &*key, "found cached wake_compute error");
 
-                    debug!(key = &*key, "found cached compute node info");
-                    ctx.set_project(info.aux.clone());
-                    return Ok(cached.map(|()| info));
+                            // if retry_delay_ms is set, reduce it by the amount of time it spent in cache
+                            if let Some(status) = &mut msg.status {
+                                if let Some(retry_info) = &mut status.details.retry_info {
+                                    retry_info.retry_delay_ms = retry_info
+                                        .retry_delay_ms
+                                        .saturating_sub(created_at.elapsed().as_millis() as u64)
+                                }
+                            }
+
+                            Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
+                                msg,
+                            )))
+                        }
+                        Ok(info) => {
+                            debug!(key = &*key, "found cached compute node info");
+                            ctx.set_project(info.aux.clone());
+                            Ok(cached.map(|()| info))
+                        }
+                    };
                 }
             };
         }
@@ -434,42 +450,29 @@ impl super::ControlPlaneApi for NeonControlPlaneClient {
                 Ok(cached.map(|()| node))
             }
             Err(err) => match err {
-                WakeComputeError::ControlPlane(ControlPlaneError::Message(err)) => {
-                    let Some(status) = &err.status else {
-                        return Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
-                            err,
-                        )));
-                    };
+                WakeComputeError::ControlPlane(ControlPlaneError::Message(ref msg)) => {
+                    let retry_info = msg.status.as_ref().and_then(|s| s.details.retry_info);
 
-                    let reason = status
-                        .details
-                        .error_info
-                        .map_or(Reason::Unknown, |x| x.reason);
-
-                    // if we can retry this error, do not cache it.
-                    if reason.can_retry() {
-                        return Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
-                            err,
-                        )));
+                    // If we can retry this error, do not cache it,
+                    // unless we were given a retry delay.
+                    if msg.could_retry() && retry_info.is_none() {
+                        return Err(err);
                     }
 
-                    // at this point, we should only have quota errors.
                     debug!(
                         key = &*key,
                         "created a cache entry for the wake compute error"
                     );
 
-                    self.caches.node_info.insert_ttl(
-                        key,
-                        Err(err.clone()),
-                        Duration::from_secs(30),
-                    );
+                    let ttl = retry_info.map_or(Duration::from_secs(30), |r| {
+                        Duration::from_millis(r.retry_delay_ms)
+                    });
 
-                    Err(WakeComputeError::ControlPlane(ControlPlaneError::Message(
-                        err,
-                    )))
+                    self.caches.node_info.insert_ttl(key, Err(msg.clone()), ttl);
+
+                    Err(err)
                 }
-                err => return Err(err),
+                err => Err(err),
             },
         }
     }

proxy/src/control_plane/errors.rs

@@ -43,28 +43,35 @@ impl UserFacingError for ControlPlaneError {
 }
 
 impl ReportableError for ControlPlaneError {
-    fn get_error_kind(&self) -> crate::error::ErrorKind {
+    fn get_error_kind(&self) -> ErrorKind {
         match self {
             ControlPlaneError::Message(e) => match e.get_reason() {
-                Reason::RoleProtected => ErrorKind::User,
-                Reason::ResourceNotFound => ErrorKind::User,
-                Reason::ProjectNotFound => ErrorKind::User,
-                Reason::EndpointNotFound => ErrorKind::User,
-                Reason::BranchNotFound => ErrorKind::User,
+                Reason::RoleProtected
+                | Reason::ResourceNotFound
+                | Reason::ProjectNotFound
+                | Reason::EndpointNotFound
+                | Reason::EndpointDisabled
+                | Reason::BranchNotFound
+                | Reason::InvalidEphemeralEndpointOptions => ErrorKind::User,
+
                 Reason::RateLimitExceeded => ErrorKind::ServiceRateLimit,
-                Reason::NonDefaultBranchComputeTimeExceeded => ErrorKind::Quota,
-                Reason::ActiveTimeQuotaExceeded => ErrorKind::Quota,
-                Reason::ComputeTimeQuotaExceeded => ErrorKind::Quota,
-                Reason::WrittenDataQuotaExceeded => ErrorKind::Quota,
-                Reason::DataTransferQuotaExceeded => ErrorKind::Quota,
-                Reason::LogicalSizeQuotaExceeded => ErrorKind::Quota,
-                Reason::ConcurrencyLimitReached => ErrorKind::ControlPlane,
-                Reason::LockAlreadyTaken => ErrorKind::ControlPlane,
-                Reason::RunningOperations => ErrorKind::ControlPlane,
-                Reason::ActiveEndpointsLimitExceeded => ErrorKind::ControlPlane,
-                Reason::Unknown => ErrorKind::ControlPlane,
+
+                Reason::NonDefaultBranchComputeTimeExceeded
+                | Reason::ActiveTimeQuotaExceeded
+                | Reason::ComputeTimeQuotaExceeded
+                | Reason::WrittenDataQuotaExceeded
+                | Reason::DataTransferQuotaExceeded
+                | Reason::LogicalSizeQuotaExceeded
+                | Reason::ActiveEndpointsLimitExceeded => ErrorKind::Quota,
+
+                Reason::ConcurrencyLimitReached
+                | Reason::LockAlreadyTaken
+                | Reason::RunningOperations
+                | Reason::EndpointIdle
+                | Reason::ProjectUnderMaintenance
+                | Reason::Unknown => ErrorKind::ControlPlane,
             },
-            ControlPlaneError::Transport(_) => crate::error::ErrorKind::ControlPlane,
+            ControlPlaneError::Transport(_) => ErrorKind::ControlPlane,
         }
     }
 }
@@ -120,10 +127,10 @@ impl UserFacingError for GetAuthInfoError {
 }
 
 impl ReportableError for GetAuthInfoError {
-    fn get_error_kind(&self) -> crate::error::ErrorKind {
+    fn get_error_kind(&self) -> ErrorKind {
         match self {
-            Self::BadSecret => crate::error::ErrorKind::ControlPlane,
-            Self::ApiError(_) => crate::error::ErrorKind::ControlPlane,
+            Self::BadSecret => ErrorKind::ControlPlane,
+            Self::ApiError(_) => ErrorKind::ControlPlane,
         }
     }
 }

proxy/src/control_plane/messages.rs

@@ -126,10 +126,16 @@ pub(crate) enum Reason {
     /// or that the subject doesn't have enough permissions to access the requested endpoint.
     #[serde(rename = "ENDPOINT_NOT_FOUND")]
     EndpointNotFound,
+    /// EndpointDisabled indicates that the endpoint has been disabled and does not accept connections.
+    #[serde(rename = "ENDPOINT_DISABLED")]
+    EndpointDisabled,
     /// BranchNotFound indicates that the branch wasn't found, usually due to the provided ID not being correct,
     /// or that the subject doesn't have enough permissions to access the requested branch.
     #[serde(rename = "BRANCH_NOT_FOUND")]
     BranchNotFound,
+    /// InvalidEphemeralEndpointOptions indicates that the specified LSN or timestamp are wrong.
+    #[serde(rename = "INVALID_EPHEMERAL_OPTIONS")]
+    InvalidEphemeralEndpointOptions,
     /// RateLimitExceeded indicates that the rate limit for the operation has been exceeded.
     #[serde(rename = "RATE_LIMIT_EXCEEDED")]
     RateLimitExceeded,
@@ -152,6 +158,9 @@ pub(crate) enum Reason {
     /// LogicalSizeQuotaExceeded indicates that the logical size quota was exceeded.
     #[serde(rename = "LOGICAL_SIZE_QUOTA_EXCEEDED")]
     LogicalSizeQuotaExceeded,
+    /// ActiveEndpointsLimitExceeded indicates that the limit of concurrently active endpoints was exceeded.
+    #[serde(rename = "ACTIVE_ENDPOINTS_LIMIT_EXCEEDED")]
+    ActiveEndpointsLimitExceeded,
     /// RunningOperations indicates that the project already has some running operations
     /// and scheduling of new ones is prohibited.
     #[serde(rename = "RUNNING_OPERATIONS")]
@@ -162,9 +171,13 @@ pub(crate) enum Reason {
     /// LockAlreadyTaken indicates that the we attempted to take a lock that was already taken.
     #[serde(rename = "LOCK_ALREADY_TAKEN")]
     LockAlreadyTaken,
-    /// ActiveEndpointsLimitExceeded indicates that the limit of concurrently active endpoints was exceeded.
-    #[serde(rename = "ACTIVE_ENDPOINTS_LIMIT_EXCEEDED")]
-    ActiveEndpointsLimitExceeded,
+    /// EndpointIdle indicates that the endpoint cannot become active, because it's idle.
+    #[serde(rename = "ENDPOINT_IDLE")]
+    EndpointIdle,
+    /// ProjectUnderMaintenance indicates that the project is currently ongoing maintenance,
+    /// and thus cannot accept connections.
+    #[serde(rename = "PROJECT_UNDER_MAINTENANCE")]
+    ProjectUnderMaintenance,
     #[default]
     #[serde(other)]
     Unknown,
@@ -184,13 +197,15 @@ impl Reason {
     pub(crate) fn can_retry(self) -> bool {
         match self {
             // do not retry role protected errors
-            // not a transitive error
+            // not a transient error
             Reason::RoleProtected => false,
-            // on retry, it will still not be found
+            // on retry, it will still not be found or valid
             Reason::ResourceNotFound
             | Reason::ProjectNotFound
             | Reason::EndpointNotFound
-            | Reason::BranchNotFound => false,
+            | Reason::EndpointDisabled
+            | Reason::BranchNotFound
+            | Reason::InvalidEphemeralEndpointOptions => false,
             // we were asked to go away
             Reason::RateLimitExceeded
             | Reason::NonDefaultBranchComputeTimeExceeded
@@ -200,11 +215,13 @@ impl Reason {
             | Reason::DataTransferQuotaExceeded
             | Reason::LogicalSizeQuotaExceeded
             | Reason::ActiveEndpointsLimitExceeded => false,
-            // transitive error. control plane is currently busy
+            // transient error. control plane is currently busy
             // but might be ready soon
             Reason::RunningOperations
             | Reason::ConcurrencyLimitReached
-            | Reason::LockAlreadyTaken => true,
+            | Reason::LockAlreadyTaken
+            | Reason::EndpointIdle
+            | Reason::ProjectUnderMaintenance => true,
             // unknown error. better not retry it.
             Reason::Unknown => false,
         }

proxy/src/lib.rs

@@ -93,7 +93,7 @@ mod ext;
 mod http;
 mod intern;
 mod jemalloc;
-mod logging;
+pub mod logging;
 mod metrics;
 mod parse;
 mod pglb;

proxy/src/logging.rs

@@ -148,11 +148,11 @@ impl LogFormat {
     }
 }
 
-trait MakeWriter {
+pub trait MakeWriter {
     fn make_writer(&self) -> impl io::Write;
 }
 
-struct StderrWriter {
+pub struct StderrWriter {
     stderr: io::Stderr,
 }
 
@@ -164,11 +164,11 @@ impl MakeWriter for StderrWriter {
 }
 
 // TODO: move into separate module or even separate crate.
-trait Clock {
+pub trait Clock {
     fn now(&self) -> DateTime<Utc>;
 }
 
-struct RealClock;
+pub struct RealClock;
 
 impl Clock for RealClock {
     #[inline]
@@ -203,7 +203,7 @@ type CallsiteMap<T> =
     papaya::HashMap<callsite::Identifier, T, std::hash::BuildHasherDefault<rustc_hash::FxHasher>>;
 
 /// Implements tracing layer to handle events specific to logging.
-struct JsonLoggingLayer<C: Clock, W: MakeWriter> {
+pub struct JsonLoggingLayer<C: Clock, W: MakeWriter> {
     clock: C,
     writer: W,
 
@@ -217,7 +217,7 @@ struct JsonLoggingLayer<C: Clock, W: MakeWriter> {
 }
 
 impl<C: Clock, W: MakeWriter> JsonLoggingLayer<C, W> {
-    fn new(clock: C, writer: W, extract_fields: &'static [&'static str]) -> Self {
+    pub fn new(clock: C, writer: W, extract_fields: &'static [&'static str]) -> Self {
         JsonLoggingLayer {
             clock,
             skipped_field_indices: CallsiteMap::default(),

proxy/src/metrics.rs

@@ -374,11 +374,10 @@ pub enum Waiting {
 #[label(singleton = "kind")]
 #[allow(clippy::enum_variant_names)]
 pub enum RedisMsgKind {
-    HSet,
-    HSetMultiple,
+    Set,
+    Get,
+    Expire,
     HGet,
-    HGetAll,
-    HDel,
 }
 
 #[derive(Default, Clone)]

proxy/src/proxy/connect_compute.rs

@@ -110,7 +110,7 @@ where
     debug!(error = ?err, COULD_NOT_CONNECT);
 
     let node_info = if !node_info.cached() || !err.should_retry_wake_compute() {
-        // If we just recieved this from cplane and didn't get it from cache, we shouldn't retry.
+        // If we just received this from cplane and not from the cache, we shouldn't retry.
         // Do not need to retrieve a new node_info, just return the old one.
         if !should_retry(&err, num_retries, compute.retry) {
             Metrics::get().proxy.retries_metric.observe(

proxy/src/proxy/mod.rs

@@ -195,15 +195,18 @@ impl NeonOptions {
     // proxy options:
 
     /// `PARAMS_COMPAT` allows opting in to forwarding all startup parameters from client to compute.
-    pub const PARAMS_COMPAT: &str = "proxy_params_compat";
+    pub const PARAMS_COMPAT: &'static str = "proxy_params_compat";
 
     // cplane options:
 
     /// `LSN` allows provisioning an ephemeral compute with time-travel to the provided LSN.
-    const LSN: &str = "lsn";
+    const LSN: &'static str = "lsn";
+
+    /// `TIMESTAMP` allows provisioning an ephemeral compute with time-travel to the provided timestamp.
+    const TIMESTAMP: &'static str = "timestamp";
 
     /// `ENDPOINT_TYPE` allows configuring an ephemeral compute to be read_only or read_write.
-    const ENDPOINT_TYPE: &str = "endpoint_type";
+    const ENDPOINT_TYPE: &'static str = "endpoint_type";
 
     pub(crate) fn parse_params(params: &StartupMessageParams) -> Self {
         params
@@ -228,6 +231,7 @@ impl NeonOptions {
             // This is not a cplane option, we know it does not create ephemeral computes.
             Self::PARAMS_COMPAT => false,
             Self::LSN => true,
+            Self::TIMESTAMP => true,
             Self::ENDPOINT_TYPE => true,
             // err on the side of caution. any cplane options we don't know about
             // might lead to ephemeral computes.

proxy/src/redis/connection_with_credentials_provider.rs

@@ -4,11 +4,12 @@ use std::time::Duration;
 
 use futures::FutureExt;
 use redis::aio::{ConnectionLike, MultiplexedConnection};
-use redis::{ConnectionInfo, IntoConnectionInfo, RedisConnectionInfo, RedisResult};
+use redis::{ConnectionInfo, IntoConnectionInfo, RedisConnectionInfo, RedisError, RedisResult};
 use tokio::task::AbortHandle;
 use tracing::{error, info, warn};
 
 use super::elasticache::CredentialsProvider;
+use crate::redis::elasticache::CredentialsProviderError;
 
 enum Credentials {
     Static(ConnectionInfo),
@@ -26,6 +27,14 @@ impl Clone for Credentials {
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+pub enum ConnectionProviderError {
+    #[error(transparent)]
+    Redis(#[from] RedisError),
+    #[error(transparent)]
+    CredentialsProvider(#[from] CredentialsProviderError),
+}
+
 /// A wrapper around `redis::MultiplexedConnection` that automatically refreshes the token.
 /// Provides PubSub connection without credentials refresh.
 pub struct ConnectionWithCredentialsProvider {
@@ -86,15 +95,18 @@ impl ConnectionWithCredentialsProvider {
         }
     }
 
-    async fn ping(con: &mut MultiplexedConnection) -> RedisResult<()> {
-        redis::cmd("PING").query_async(con).await
+    async fn ping(con: &mut MultiplexedConnection) -> Result<(), ConnectionProviderError> {
+        redis::cmd("PING")
+            .query_async(con)
+            .await
+            .map_err(Into::into)
     }
 
     pub(crate) fn credentials_refreshed(&self) -> bool {
         self.credentials_refreshed.load(Ordering::Relaxed)
     }
 
-    pub(crate) async fn connect(&mut self) -> anyhow::Result<()> {
+    pub(crate) async fn connect(&mut self) -> Result<(), ConnectionProviderError> {
         let _guard = self.mutex.lock().await;
         if let Some(con) = self.con.as_mut() {
             match Self::ping(con).await {
@@ -141,7 +153,7 @@ impl ConnectionWithCredentialsProvider {
         Ok(())
     }
 
-    async fn get_connection_info(&self) -> anyhow::Result<ConnectionInfo> {
+    async fn get_connection_info(&self) -> Result<ConnectionInfo, ConnectionProviderError> {
         match &self.credentials {
             Credentials::Static(info) => Ok(info.clone()),
             Credentials::Dynamic(provider, addr) => {
@@ -160,7 +172,7 @@ impl ConnectionWithCredentialsProvider {
         }
     }
 
-    async fn get_client(&self) -> anyhow::Result<redis::Client> {
+    async fn get_client(&self) -> Result<redis::Client, ConnectionProviderError> {
         let client = redis::Client::open(self.get_connection_info().await?)?;
         self.credentials_refreshed.store(true, Ordering::Relaxed);
         Ok(client)

proxy/src/redis/elasticache.rs

@@ -9,10 +9,12 @@ use aws_config::meta::region::RegionProviderChain;
 use aws_config::profile::ProfileFileCredentialsProvider;
 use aws_config::provider_config::ProviderConfig;
 use aws_config::web_identity_token::WebIdentityTokenCredentialsProvider;
+use aws_credential_types::provider::error::CredentialsError;
 use aws_sdk_iam::config::ProvideCredentials;
 use aws_sigv4::http_request::{
-    self, SignableBody, SignableRequest, SignatureLocation, SigningSettings,
+    self, SignableBody, SignableRequest, SignatureLocation, SigningError, SigningSettings,
 };
+use aws_sigv4::sign::v4::signing_params::BuildError;
 use tracing::info;
 
 #[derive(Debug)]
@@ -40,6 +42,18 @@ impl AWSIRSAConfig {
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+pub enum CredentialsProviderError {
+    #[error(transparent)]
+    AwsCredentials(#[from] CredentialsError),
+    #[error(transparent)]
+    AwsSigv4Build(#[from] BuildError),
+    #[error(transparent)]
+    AwsSigv4Singing(#[from] SigningError),
+    #[error(transparent)]
+    Http(#[from] http::Error),
+}
+
 /// Credentials provider for AWS elasticache authentication.
 ///
 /// Official documentation:
@@ -92,7 +106,9 @@ impl CredentialsProvider {
         })
     }
 
-    pub(crate) async fn provide_credentials(&self) -> anyhow::Result<(String, String)> {
+    pub(crate) async fn provide_credentials(
+        &self,
+    ) -> Result<(String, String), CredentialsProviderError> {
         let aws_credentials = self
             .credentials_provider
             .provide_credentials()