zero-copy jwt claim validation

https://neondb.slack.com/archives/C059ZC138NR/p1727350911890989?thread_ts=1727350211.370869&cid=C059ZC138NR

.dockerignore

@@ -13,6 +13,7 @@
 # Directories
 !.cargo/
 !.config/
+!compute/
 !compute_tools/
 !control_plane/
 !libs/

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

@@ -257,7 +257,15 @@ jobs:
           ${cov_prefix} cargo nextest run $CARGO_FLAGS $CARGO_FEATURES -E 'package(remote_storage)' -E 'test(test_real_azure)'
 
       - name: Install postgres binaries
-        run: cp -a pg_install /tmp/neon/pg_install
+        run: |
+          # Use tar to copy files matching the pattern, preserving the paths in the destionation
+          tar c \
+            pg_install/v* \
+            pg_install/build/*/src/test/regress/*.so \
+            pg_install/build/*/src/test/regress/pg_regress \
+            pg_install/build/*/src/test/isolation/isolationtester \
+            pg_install/build/*/src/test/isolation/pg_isolation_regress \
+            | tar  x -C /tmp/neon
 
       - name: Upload Neon artifact
         uses: ./.github/actions/upload

.github/workflows/build_and_test.yml

@@ -120,6 +120,59 @@ jobs:
       - name: Run mypy to check types
         run: poetry run mypy .
 
+  # Check that the vendor/postgres-* submodules point to the
+  # corresponding REL_*_STABLE_neon branches.
+  check-submodules:
+    runs-on: ubuntu-22.04
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          submodules: true
+
+      - uses: dorny/paths-filter@v3
+        id: check-if-submodules-changed
+        with:
+          filters: |
+            vendor:
+              - 'vendor/**'
+
+      - name: Check vendor/postgres-v14 submodule reference
+        if: steps.check-if-submodules-changed.outputs.vendor == 'true'
+        uses: jtmullen/submodule-branch-check-action@v1
+        with:
+          path: "vendor/postgres-v14"
+          fetch_depth: "50"
+          sub_fetch_depth: "50"
+          pass_if_unchanged: true
+
+      - name: Check vendor/postgres-v15 submodule reference
+        if: steps.check-if-submodules-changed.outputs.vendor == 'true'
+        uses: jtmullen/submodule-branch-check-action@v1
+        with:
+          path: "vendor/postgres-v15"
+          fetch_depth: "50"
+          sub_fetch_depth: "50"
+          pass_if_unchanged: true
+
+      - name: Check vendor/postgres-v16 submodule reference
+        if: steps.check-if-submodules-changed.outputs.vendor == 'true'
+        uses: jtmullen/submodule-branch-check-action@v1
+        with:
+          path: "vendor/postgres-v16"
+          fetch_depth: "50"
+          sub_fetch_depth: "50"
+          pass_if_unchanged: true
+
+      - name: Check vendor/postgres-v17 submodule reference
+        if: steps.check-if-submodules-changed.outputs.vendor == 'true'
+        uses: jtmullen/submodule-branch-check-action@v1
+        with:
+          path: "vendor/postgres-v17"
+          fetch_depth: "50"
+          sub_fetch_depth: "50"
+          pass_if_unchanged: true
+
   check-codestyle-rust:
     needs: [ check-permissions, build-build-tools-image ]
     strategy:
@@ -549,7 +602,20 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version: [ v14, v15, v16, v17 ]
+        version:
+          # Much data was already generated on old PG versions with bullseye's
+          # libraries, the locales of which can cause data incompatibilities.
+          # However, new PG versions should check if they can be built on newer
+          # images, as that reduces the support burden of old and ancient
+          # distros.
+          - pg: v14
+            debian: bullseye-slim
+          - pg: v15
+            debian: bullseye-slim
+          - pg: v16
+            debian: bullseye-slim
+          - pg: v17
+            debian: bookworm-slim
         arch: [ x64, arm64 ]
 
     runs-on: ${{ fromJson(format('["self-hosted", "{0}"]', matrix.arch == 'arm64' && 'large-arm64' || 'large')) }}
@@ -592,41 +658,46 @@ jobs:
           context: .
           build-args: |
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
-            PG_VERSION=${{ matrix.version }}
+            PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
-          cache-from: type=registry,ref=cache.neon.build/compute-node-${{ matrix.version }}:cache-${{ matrix.arch }}
-          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/compute-node-{0}:cache-{1},mode=max', matrix.version, matrix.arch) || '' }}
+          file: compute/Dockerfile.compute-node
+          cache-from: type=registry,ref=cache.neon.build/compute-node-${{ matrix.version.pg }}:cache-${{ matrix.arch }}
+          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/compute-node-{0}:cache-{1},mode=max', matrix.version.pg, matrix.arch) || '' }}
           tags: |
-            neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
+            neondatabase/compute-node-${{ matrix.version.pg }}:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
       - name: Build neon extensions test image
-        if: matrix.version == 'v16'
+        if: matrix.version.pg == 'v16'
         uses: docker/build-push-action@v6
         with:
           context: .
           build-args: |
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
-            PG_VERSION=${{ matrix.version }}
+            PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
+          file: compute/Dockerfile.compute-node
           target: neon-pg-ext-test
-          cache-from: type=registry,ref=cache.neon.build/neon-test-extensions-${{ matrix.version }}:cache-${{ matrix.arch }}
-          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/neon-test-extensions-{0}:cache-{1},mode=max', matrix.version, matrix.arch) || '' }}
+          cache-from: type=registry,ref=cache.neon.build/neon-test-extensions-${{ matrix.version.pg }}:cache-${{ matrix.arch }}
+          cache-to: ${{ github.ref_name == 'main' && format('type=registry,ref=cache.neon.build/neon-test-extensions-{0}:cache-{1},mode=max', matrix.version.pg, matrix.arch) || '' }}
           tags: |
-            neondatabase/neon-test-extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}-${{ matrix.arch }}
+            neondatabase/neon-test-extensions-${{ matrix.version.pg }}:${{needs.tag.outputs.build-tag}}-${{ matrix.arch }}
 
       - name: Build compute-tools image
         # compute-tools are Postgres independent, so build it only once
-        if: matrix.version == 'v17'
+        # We pick 16, because that builds on debian 11 with older glibc (and is
+        # thus compatible with newer glibc), rather than 17 on Debian 12, as
+        # that isn't guaranteed to be compatible with Debian 11
+        if: matrix.version.pg == 'v16'
         uses: docker/build-push-action@v6
         with:
           target: compute-tools-image
@@ -635,10 +706,11 @@ jobs:
             GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
             BUILD_TAG=${{ needs.tag.outputs.build-tag }}
             TAG=${{ needs.build-build-tools-image.outputs.image-tag }}
+            DEBIAN_FLAVOR=${{ matrix.version.debian }}
           provenance: false
           push: true
           pull: true
-          file: Dockerfile.compute-node
+          file: compute/Dockerfile.compute-node
           tags: |
             neondatabase/compute-tools:${{ needs.tag.outputs.build-tag }}-${{ matrix.arch }}
 
@@ -726,7 +798,7 @@ jobs:
       - name: Build vm image
         run: |
           ./vm-builder \
-            -spec=vm-image-spec.yaml \
+            -spec=compute/vm-image-spec.yaml \
             -src=neondatabase/compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }} \
             -dst=neondatabase/vm-compute-node-${{ matrix.version }}:${{ needs.tag.outputs.build-tag }}
 
@@ -790,6 +862,9 @@ jobs:
     needs: [ check-permissions, tag, test-images, vm-compute-node-image ]
     runs-on: ubuntu-22.04
 
+    permissions:
+      id-token: write # for `aws-actions/configure-aws-credentials`
+
     env:
       VERSIONS: v14 v15 v16 v17
 
@@ -834,13 +909,19 @@ jobs:
           docker buildx imagetools create -t neondatabase/neon-test-extensions-v16:latest \
                                               neondatabase/neon-test-extensions-v16:${{ needs.tag.outputs.build-tag }}
 
+      - name: Configure AWS-prod credentials
+        if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          aws-region: eu-central-1
+          mask-aws-account-id: true
+          role-to-assume: ${{ secrets.PROD_GHA_OIDC_ROLE }}
+
       - name: Login to prod ECR
         uses: docker/login-action@v3
         if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
         with:
           registry: 093970136003.dkr.ecr.eu-central-1.amazonaws.com
-          username: ${{ secrets.PROD_GHA_RUNNER_LIMITED_AWS_ACCESS_KEY_ID }}
-          password: ${{ secrets.PROD_GHA_RUNNER_LIMITED_AWS_SECRET_ACCESS_KEY }}
 
       - name: Copy all images to prod ECR
         if: github.ref_name == 'release'|| github.ref_name == 'release-proxy'
@@ -1109,10 +1190,9 @@ jobs:
 
               files_to_promote+=("s3://${BUCKET}/${s3_key}")
 
-              # TODO Add v17
-              for pg_version in v14 v15 v16; do
+              for pg_version in v14 v15 v16 v17; do
                 # We run less tests for debug builds, so we don't need to promote them
-                if [ "${build_type}" == "debug" ] && { [ "${arch}" == "ARM64" ] || [ "${pg_version}" != "v16" ] ; }; then
+                if [ "${build_type}" == "debug" ] && { [ "${arch}" == "ARM64" ] || [ "${pg_version}" != "v17" ] ; }; then
                   continue
                 fi
 

.github/workflows/cloud-regress.yml

@@ -0,0 +1,102 @@
+name: Cloud Regression Test
+on:
+  schedule:
+    # * is a special character in YAML so you have to quote this string
+    #          ┌───────────── minute (0 - 59)
+    #          │ ┌───────────── hour (0 - 23)
+    #          │ │ ┌───────────── day of the month (1 - 31)
+    #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    - cron:  '45 1 * * *' # run once a day, timezone is utc
+  workflow_dispatch: # adds ability to run this manually
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+
+concurrency:
+  # Allow only one workflow
+  group: ${{ github.workflow }}
+  cancel-in-progress: true
+
+jobs:
+  regress:
+    env:
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 16
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+    runs-on: us-east-2
+    container:
+      image: neondatabase/build-tools:pinned
+      options: --init
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+
+      - name: Patch the test
+        run: |
+          cd "vendor/postgres-v${DEFAULT_PG_VERSION}"
+          patch -p1 < "../../compute/patches/cloud_regress_pg${DEFAULT_PG_VERSION}.patch"
+
+      - name: Generate a random password
+        id: pwgen
+        run: |
+          set +x
+          DBPASS=$(dd if=/dev/random bs=48 count=1 2>/dev/null | base64)
+          echo "::add-mask::${DBPASS//\//}"
+          echo DBPASS="${DBPASS//\//}" >> "${GITHUB_OUTPUT}"
+
+      - name: Change tests according to the generated password
+        env:
+          DBPASS: ${{ steps.pwgen.outputs.DBPASS }}
+        run: |
+          cd vendor/postgres-v"${DEFAULT_PG_VERSION}"/src/test/regress
+          for fname in sql/*.sql expected/*.out; do
+            sed -i.bak s/NEON_PASSWORD_PLACEHOLDER/"'${DBPASS}'"/ "${fname}"
+          done
+          for ph in $(grep NEON_MD5_PLACEHOLDER expected/password.out | awk '{print $3;}' | sort | uniq); do
+            USER=$(echo "${ph}" | cut -c 22-)
+            MD5=md5$(echo -n "${DBPASS}${USER}" | md5sum | awk '{print $1;}')
+            sed -i.bak "s/${ph}/${MD5}/" expected/password.out
+          done
+
+      - name: Download Neon artifact
+        uses: ./.github/actions/download
+        with:
+          name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+          path: /tmp/neon/
+          prefix: latest
+
+      - name: Run the regression tests
+        uses: ./.github/actions/run-python-test-set
+        with:
+          build_type: ${{ env.BUILD_TYPE }}
+          test_selection: cloud_regress
+          pg_version: ${{ env.DEFAULT_PG_VERSION }}
+          extra_params: -m remote_cluster
+        env:
+          BENCHMARK_CONNSTR: ${{ secrets.PG_REGRESS_CONNSTR }}
+
+      - name: Create Allure report
+        id: create-allure-report
+        if: ${{ !cancelled() }}
+        uses: ./.github/actions/allure-report-generate
+
+      - name: Post to a Slack channel
+        if: ${{ github.event.schedule && failure() }}
+        uses: slackapi/slack-github-action@v1
+        with:
+          channel-id: "C033QLM5P7D" # on-call-staging-stream
+          slack-message: |
+            Periodic pg_regress on staging: ${{ job.status }}
+            <${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}|GitHub Run>
+            <${{ steps.create-allure-report.outputs.report-url }}|Allure report>
+        env:
+          SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
+

.github/workflows/trigger-e2e-tests.yml

@@ -102,12 +102,12 @@ jobs:
           # Default set of platforms to run e2e tests on
           platforms='["docker", "k8s"]'
 
-          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
+          # If the PR changes vendor/, pgxn/ or libs/vm_monitor/ directories, or compute/Dockerfile.compute-node, add k8s-neonvm to the list of platforms.
           # If the workflow run is not a pull request, add k8s-neonvm to the list.
           if [ "$GITHUB_EVENT_NAME" == "pull_request" ]; then
             for f in $(gh api "/repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" --paginate --jq '.[].filename'); do
               case "$f" in
-                vendor/*|pgxn/*|libs/vm_monitor/*|Dockerfile.compute-node)
+                vendor/*|pgxn/*|libs/vm_monitor/*|compute/Dockerfile.compute-node)
                   platforms=$(echo "${platforms}" | jq --compact-output '. += ["k8s-neonvm"] | unique')
                   ;;
                 *)

Cargo.lock

@@ -255,12 +255,6 @@ dependencies = [
  "syn 2.0.52",
 ]
 
-[[package]]
-name = "atomic"
-version = "0.5.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c59bdb34bc650a32731b31bd8f0829cc15d24a708ee31559e0bb34f2bc320cba"
-
 [[package]]
 name = "atomic-take"
 version = "1.1.0"
@@ -295,8 +289,8 @@ dependencies = [
  "fastrand 2.0.0",
  "hex",
  "http 0.2.9",
- "hyper 0.14.26",
- "ring 0.17.6",
+ "hyper 0.14.30",
+ "ring",
  "time",
  "tokio",
  "tracing",
@@ -486,7 +480,7 @@ dependencies = [
  "once_cell",
  "p256 0.11.1",
  "percent-encoding",
- "ring 0.17.6",
+ "ring",
  "sha2",
  "subtle",
  "time",
@@ -593,7 +587,7 @@ dependencies = [
  "http 0.2.9",
  "http-body 0.4.5",
  "http-body 1.0.0",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "hyper-rustls 0.24.0",
  "once_cell",
  "pin-project-lite",
@@ -684,7 +678,7 @@ dependencies = [
  "futures-util",
  "http 0.2.9",
  "http-body 0.4.5",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "itoa",
  "matchit 0.7.0",
  "memchr",
@@ -1089,9 +1083,9 @@ dependencies = [
 
 [[package]]
 name = "ciborium"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "effd91f6c78e5a4ace8a5d3c0b6bfaec9e2baaef55f3efc00e45fb2e477ee926"
+checksum = "42e69ffd6f0917f5c029256a24d0161db17cea3997d185db0d35926308770f0e"
 dependencies = [
  "ciborium-io",
  "ciborium-ll",
@@ -1100,18 +1094,18 @@ dependencies = [
 
 [[package]]
 name = "ciborium-io"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cdf919175532b369853f5d5e20b26b43112613fd6fe7aee757e35f7a44642656"
+checksum = "05afea1e0a06c9be33d539b876f1ce3692f4afea2cb41f740e7743225ed1c757"
 
 [[package]]
 name = "ciborium-ll"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "defaa24ecc093c77630e6c15e17c51f5e187bf35ee514f4e2d67baaa96dae22b"
+checksum = "57663b653d948a338bfb3eeba9bb2fd5fcfaecb9e199e87e1eda4d9e8b240fd9"
 dependencies = [
  "ciborium-io",
- "half 1.8.2",
+ "half",
 ]
 
 [[package]]
@@ -1224,7 +1218,7 @@ dependencies = [
  "compute_api",
  "flate2",
  "futures",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "nix 0.27.1",
  "notify",
  "num_cpus",
@@ -1327,10 +1321,9 @@ dependencies = [
  "clap",
  "comfy-table",
  "compute_api",
- "git-version",
  "humantime",
  "humantime-serde",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "nix 0.27.1",
  "once_cell",
  "pageserver_api",
@@ -2304,12 +2297,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "half"
-version = "1.8.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eabb4a44450da02c90444cf74558da904edde8fb4e9035a9a6a4e15445af0bd7"
-
 [[package]]
 name = "half"
 version = "2.4.1"
@@ -2411,17 +2398,6 @@ dependencies = [
  "digest",
 ]
 
-[[package]]
-name = "hostname"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3c731c3e10504cc8ed35cfe2f1db4c9274c3d35fa486e3b31df46f068ef3e867"
-dependencies = [
- "libc",
- "match_cfg",
- "winapi",
-]
-
 [[package]]
 name = "hostname"
 version = "0.4.0"
@@ -2430,7 +2406,7 @@ checksum = "f9c7c7c8ac16c798734b8a24560c1362120597c40d5e1459f09498f8f6c8f2ba"
 dependencies = [
  "cfg-if",
  "libc",
- "windows 0.52.0",
+ "windows",
 ]
 
 [[package]]
@@ -2539,9 +2515,9 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "0.14.26"
+version = "0.14.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ab302d72a6f11a3b910431ff93aae7e773078c769f0a3ef15fb9ec692ed147d4"
+checksum = "a152ddd61dfaec7273fe8419ab357f33aee0d914c5f4efbf0d96fa749eea5ec9"
 dependencies = [
  "bytes",
  "futures-channel",
@@ -2554,7 +2530,7 @@ dependencies = [
  "httpdate",
  "itoa",
  "pin-project-lite",
- "socket2 0.4.9",
+ "socket2",
  "tokio",
  "tower-service",
  "tracing",
@@ -2589,7 +2565,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0646026eb1b3eea4cd9ba47912ea5ce9cc07713d105b1a14698f4e6433d348b7"
 dependencies = [
  "http 0.2.9",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "log",
  "rustls 0.21.11",
  "rustls-native-certs 0.6.2",
@@ -2620,7 +2596,7 @@ version = "0.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbb958482e8c7be4bc3cf272a766a2b0bf1a6755e7a6ae777f017a31d11b13b1"
 dependencies = [
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "pin-project-lite",
  "tokio",
  "tokio-io-timeout",
@@ -2639,7 +2615,7 @@ dependencies = [
  "http-body 1.0.0",
  "hyper 1.2.0",
  "pin-project-lite",
- "socket2 0.5.5",
+ "socket2",
  "tokio",
  "tower",
  "tower-service",
@@ -2648,16 +2624,16 @@ dependencies = [
 
 [[package]]
 name = "iana-time-zone"
-version = "0.1.56"
+version = "0.1.61"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0722cd7114b7de04316e7ea5456a0bbb20e4adb46fd27a3697adb812cff0f37c"
+checksum = "235e081f3925a06703c2d0117ea8b91f042756fd6e7a6e5d901e8ca1a996b220"
 dependencies = [
  "android_system_properties",
  "core-foundation-sys",
  "iana-time-zone-haiku",
  "js-sys",
  "wasm-bindgen",
- "windows 0.48.0",
+ "windows-core",
 ]
 
 [[package]]
@@ -2870,7 +2846,7 @@ dependencies = [
  "base64 0.21.1",
  "js-sys",
  "pem",
- "ring 0.17.6",
+ "ring",
  "serde",
  "serde_json",
  "simple_asn1",
@@ -2908,11 +2884,11 @@ dependencies = [
 
 [[package]]
 name = "lazy_static"
-version = "1.4.0"
+version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e2abad23fbc42b3700f2f279844dc832adb2b2eb069b2df918f455c4e18cc646"
+checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 dependencies = [
- "spin 0.5.2",
+ "spin",
 ]
 
 [[package]]
@@ -2974,12 +2950,6 @@ dependencies = [
  "hashbrown 0.14.5",
 ]
 
-[[package]]
-name = "match_cfg"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ffbee8634e0d45d258acb448e7eaab3fce7a0a467395d4d9f228e3c1f01fb2e4"
-
 [[package]]
 name = "matchers"
 version = "0.1.0"
@@ -3072,15 +3042,6 @@ dependencies = [
  "autocfg",
 ]
 
-[[package]]
-name = "memoffset"
-version = "0.8.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d61c719bcfbcf5d62b3a09efa6088de8c54bc0bfcd3ea7ae39fcc186108b8de1"
-dependencies = [
- "autocfg",
-]
-
 [[package]]
 name = "memoffset"
 version = "0.9.0"
@@ -3616,7 +3577,6 @@ dependencies = [
  "anyhow",
  "camino",
  "clap",
- "git-version",
  "humantime",
  "pageserver",
  "pageserver_api",
@@ -3655,12 +3615,11 @@ dependencies = [
  "enumset",
  "fail",
  "futures",
- "git-version",
  "hex",
  "hex-literal",
  "humantime",
  "humantime-serde",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "indoc",
  "itertools 0.10.5",
  "md5",
@@ -3775,7 +3734,6 @@ dependencies = [
  "clap",
  "criterion",
  "futures",
- "git-version",
  "hex-literal",
  "itertools 0.10.5",
  "once_cell",
@@ -3853,7 +3811,7 @@ dependencies = [
  "ahash",
  "bytes",
  "chrono",
- "half 2.4.1",
+ "half",
  "hashbrown 0.14.5",
  "num",
  "num-bigint",
@@ -4140,7 +4098,7 @@ dependencies = [
  "crc32c",
  "env_logger",
  "log",
- "memoffset 0.8.0",
+ "memoffset 0.9.0",
  "once_cell",
  "postgres",
  "regex",
@@ -4338,6 +4296,7 @@ dependencies = [
  "camino-tempfile",
  "chrono",
  "clap",
+ "compute_api",
  "consumption_metrics",
  "dashmap",
  "ecdsa 0.16.9",
@@ -4345,17 +4304,16 @@ dependencies = [
  "fallible-iterator",
  "framed-websockets",
  "futures",
- "git-version",
  "hashbrown 0.14.5",
  "hashlink",
  "hex",
  "hmac",
- "hostname 0.3.1",
+ "hostname",
  "http 1.1.0",
  "http-body-util",
  "humantime",
  "humantime-serde",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "hyper 1.2.0",
  "hyper-util",
  "indexmap 2.0.1",
@@ -4400,7 +4358,7 @@ dependencies = [
  "signature 2.2.0",
  "smallvec",
  "smol_str",
- "socket2 0.5.5",
+ "socket2",
  "subtle",
  "thiserror",
  "tikv-jemalloc-ctl",
@@ -4578,7 +4536,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "48406db8ac1f3cbc7dcdb56ec355343817958a356ff430259bb07baf7607e1e1"
 dependencies = [
  "pem",
- "ring 0.17.6",
+ "ring",
  "time",
  "yasna",
 ]
@@ -4602,7 +4560,7 @@ dependencies = [
  "rustls-pki-types",
  "ryu",
  "sha1_smol",
- "socket2 0.5.5",
+ "socket2",
  "tokio",
  "tokio-rustls 0.25.0",
  "tokio-util",
@@ -4714,7 +4672,7 @@ dependencies = [
  "futures-util",
  "http-types",
  "humantime-serde",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "itertools 0.10.5",
  "metrics",
  "once_cell",
@@ -4747,7 +4705,7 @@ dependencies = [
  "h2 0.3.26",
  "http 0.2.9",
  "http-body 0.4.5",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "hyper-rustls 0.24.0",
  "ipnet",
  "js-sys",
@@ -4905,21 +4863,6 @@ dependencies = [
  "subtle",
 ]
 
-[[package]]
-name = "ring"
-version = "0.16.20"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3053cf52e236a3ed746dfc745aa9cacf1b791d846bdaf412f60a8d7d6e17c8fc"
-dependencies = [
- "cc",
- "libc",
- "once_cell",
- "spin 0.5.2",
- "untrusted 0.7.1",
- "web-sys",
- "winapi",
-]
-
 [[package]]
 name = "ring"
 version = "0.17.6"
@@ -4929,8 +4872,8 @@ dependencies = [
  "cc",
  "getrandom 0.2.11",
  "libc",
- "spin 0.9.8",
- "untrusted 0.9.0",
+ "spin",
+ "untrusted",
  "windows-sys 0.48.0",
 ]
 
@@ -4950,7 +4893,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "496c1d3718081c45ba9c31fbfc07417900aa96f4070ff90dc29961836b7a9945"
 dependencies = [
  "http 0.2.9",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "lazy_static",
  "percent-encoding",
  "regex",
@@ -5074,7 +5017,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7fecbfb7b1444f477b345853b1fce097a2c6fb637b2bfb87e6bc5db0f043fae4"
 dependencies = [
  "log",
- "ring 0.17.6",
+ "ring",
  "rustls-webpki 0.101.7",
  "sct",
 ]
@@ -5086,7 +5029,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bf4ef73721ac7bcd79b2b315da7779d8fc09718c6b3d2d1b2d94850eb8c18432"
 dependencies = [
  "log",
- "ring 0.17.6",
+ "ring",
  "rustls-pki-types",
  "rustls-webpki 0.102.2",
  "subtle",
@@ -5143,24 +5086,14 @@ version = "1.3.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5ede67b28608b4c60685c7d54122d4400d90f62b40caee7700e700380a390fa8"
 
-[[package]]
-name = "rustls-webpki"
-version = "0.100.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e98ff011474fa39949b7e5c0428f9b4937eda7da7848bbb947786b7be0b27dab"
-dependencies = [
- "ring 0.16.20",
- "untrusted 0.7.1",
-]
-
 [[package]]
 name = "rustls-webpki"
 version = "0.101.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8b6275d1ee7a1cd780b64aca7726599a1dbc893b1e64144529e55c3c2f745765"
 dependencies = [
- "ring 0.17.6",
- "untrusted 0.9.0",
+ "ring",
+ "untrusted",
 ]
 
 [[package]]
@@ -5169,9 +5102,9 @@ version = "0.102.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "faaa0a62740bedb9b2ef5afa303da42764c012f743917351dc9a237ea1663610"
 dependencies = [
- "ring 0.17.6",
+ "ring",
  "rustls-pki-types",
- "untrusted 0.9.0",
+ "untrusted",
 ]
 
 [[package]]
@@ -5202,10 +5135,9 @@ dependencies = [
  "desim",
  "fail",
  "futures",
- "git-version",
  "hex",
  "humantime",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "metrics",
  "once_cell",
  "parking_lot 0.12.1",
@@ -5262,11 +5194,11 @@ dependencies = [
 
 [[package]]
 name = "schannel"
-version = "0.1.21"
+version = "0.1.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "713cfb06c7059f3588fb8044c0fad1d09e3c01d225e25b9220dbfdcf16dbb1b3"
+checksum = "fbc91545643bcf3a0bbb6569265615222618bdf33ce4ffbbd13c4bbd4c093534"
 dependencies = [
- "windows-sys 0.42.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -5290,8 +5222,8 @@ version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "da046153aa2352493d6cb7da4b6e5c0c057d8a1d0a9aa8560baffdd945acd414"
 dependencies = [
- "ring 0.17.6",
- "untrusted 0.9.0",
+ "ring",
+ "untrusted",
 ]
 
 [[package]]
@@ -5400,7 +5332,7 @@ version = "0.32.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eba8870c5dba2bfd9db25c75574a11429f6b95957b0a78ac02e2970dd7a5249a"
 dependencies = [
- "hostname 0.4.0",
+ "hostname",
  "libc",
  "os_info",
  "rustc_version",
@@ -5712,16 +5644,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "socket2"
-version = "0.4.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "64a4a911eed85daf18834cfaa86a79b7d266ff93ff5ba14005426219480ed662"
-dependencies = [
- "libc",
- "winapi",
-]
-
 [[package]]
 name = "socket2"
 version = "0.5.5"
@@ -5732,12 +5654,6 @@ dependencies = [
  "windows-sys 0.48.0",
 ]
 
-[[package]]
-name = "spin"
-version = "0.5.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6e63cff320ae2c57904679ba7cb63280a3dc4613885beafb148ee7bf9aa9042d"
-
 [[package]]
 name = "spin"
 version = "0.9.8"
@@ -5781,9 +5697,8 @@ dependencies = [
  "futures",
  "futures-core",
  "futures-util",
- "git-version",
  "humantime",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "metrics",
  "once_cell",
  "parking_lot 0.12.1",
@@ -5809,10 +5724,9 @@ dependencies = [
  "diesel_migrations",
  "fail",
  "futures",
- "git-version",
  "hex",
  "humantime",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "itertools 0.10.5",
  "lasso",
  "measured",
@@ -5862,7 +5776,6 @@ dependencies = [
  "either",
  "futures",
  "futures-util",
- "git-version",
  "hex",
  "humantime",
  "itertools 0.10.5",
@@ -6228,7 +6141,7 @@ dependencies = [
  "num_cpus",
  "pin-project-lite",
  "signal-hook-registry",
- "socket2 0.5.5",
+ "socket2",
  "tokio-macros",
  "windows-sys 0.48.0",
 ]
@@ -6288,7 +6201,7 @@ dependencies = [
  "pin-project-lite",
  "postgres-protocol",
  "postgres-types",
- "socket2 0.5.5",
+ "socket2",
  "tokio",
  "tokio-util",
 ]
@@ -6300,7 +6213,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0ea13f22eda7127c827983bdaf0d7fff9df21c8817bab02815ac277a21143677"
 dependencies = [
  "futures",
- "ring 0.17.6",
+ "ring",
  "rustls 0.22.4",
  "tokio",
  "tokio-postgres",
@@ -6434,7 +6347,7 @@ dependencies = [
  "h2 0.3.26",
  "http 0.2.9",
  "http-body 0.4.5",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "hyper-timeout",
  "percent-encoding",
  "pin-project",
@@ -6611,7 +6524,7 @@ dependencies = [
 name = "tracing-utils"
 version = "0.1.0"
 dependencies = [
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry-semantic-conventions",
@@ -6714,12 +6627,6 @@ version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f962df74c8c05a667b5ee8bcf162993134c104e96440b663c8daa176dc772d8c"
 
-[[package]]
-name = "untrusted"
-version = "0.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a156c684c91ea7d62626509bce3cb4e1d9ed5c4d978f7b4352658f96a4c26b4a"
-
 [[package]]
 name = "untrusted"
 version = "0.9.0"
@@ -6728,17 +6635,18 @@ checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
 
 [[package]]
 name = "ureq"
-version = "2.7.1"
+version = "2.9.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b11c96ac7ee530603dcdf68ed1557050f374ce55a5a07193ebf8cbc9f8927e9"
+checksum = "d11a831e3c0b56e438a28308e7c810799e3c118417f342d30ecec080105395cd"
 dependencies = [
- "base64 0.21.1",
+ "base64 0.22.1",
  "log",
  "once_cell",
- "rustls 0.21.11",
- "rustls-webpki 0.100.2",
+ "rustls 0.22.4",
+ "rustls-pki-types",
+ "rustls-webpki 0.102.2",
  "url",
- "webpki-roots 0.23.1",
+ "webpki-roots 0.26.1",
 ]
 
 [[package]]
@@ -6799,10 +6707,11 @@ dependencies = [
  "criterion",
  "fail",
  "futures",
+ "git-version",
  "hex",
  "hex-literal",
  "humantime",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "jsonwebtoken",
  "metrics",
  "nix 0.27.1",
@@ -6837,11 +6746,10 @@ dependencies = [
 
 [[package]]
 name = "uuid"
-version = "1.6.1"
+version = "1.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e395fcf16a7a3d8127ec99782007af141946b4795001f876d54fb0d55978560"
+checksum = "81dfa00651efa65069b0b6b651f4aaa31ba9e3c3ce0137aaad053604ee7e0314"
 dependencies = [
- "atomic",
  "getrandom 0.2.11",
  "serde",
 ]
@@ -7075,15 +6983,6 @@ dependencies = [
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "webpki-roots"
-version = "0.23.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b03058f88386e5ff5310d9111d53f48b17d732b401aeb83a8d5190f2ac459338"
-dependencies = [
- "rustls-webpki 0.100.2",
-]
-
 [[package]]
 name = "webpki-roots"
 version = "0.25.2"
@@ -7152,15 +7051,6 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
 
-[[package]]
-name = "windows"
-version = "0.48.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e686886bc078bc1b0b600cac0147aadb815089b6e4da64016cbd754b6342700f"
-dependencies = [
- "windows-targets 0.48.0",
-]
-
 [[package]]
 name = "windows"
 version = "0.52.0"
@@ -7180,21 +7070,6 @@ dependencies = [
  "windows-targets 0.52.4",
 ]
 
-[[package]]
-name = "windows-sys"
-version = "0.42.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a3e1820f08b8513f676f7ab6c1f99ff312fb97b553d30ff4dd86f9f15728aa7"
-dependencies = [
- "windows_aarch64_gnullvm 0.42.2",
- "windows_aarch64_msvc 0.42.2",
- "windows_i686_gnu 0.42.2",
- "windows_i686_msvc 0.42.2",
- "windows_x86_64_gnu 0.42.2",
- "windows_x86_64_gnullvm 0.42.2",
- "windows_x86_64_msvc 0.42.2",
-]
-
 [[package]]
 name = "windows-sys"
 version = "0.48.0"
@@ -7243,12 +7118,6 @@ dependencies = [
  "windows_x86_64_msvc 0.52.4",
 ]
 
-[[package]]
-name = "windows_aarch64_gnullvm"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "597a5118570b68bc08d8d59125332c54f1ba9d9adeedeef5b99b02ba2b0698f8"
-
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.48.0"
@@ -7261,12 +7130,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bcf46cf4c365c6f2d1cc93ce535f2c8b244591df96ceee75d8e83deb70a9cac9"
 
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e08e8864a60f06ef0d0ff4ba04124db8b0fb3be5776a5cd47641e942e58c4d43"
-
 [[package]]
 name = "windows_aarch64_msvc"
 version = "0.48.0"
@@ -7279,12 +7142,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "da9f259dd3bcf6990b55bffd094c4f7235817ba4ceebde8e6d11cd0c5633b675"
 
-[[package]]
-name = "windows_i686_gnu"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c61d927d8da41da96a81f029489353e68739737d3beca43145c8afec9a31a84f"
-
 [[package]]
 name = "windows_i686_gnu"
 version = "0.48.0"
@@ -7297,12 +7154,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b474d8268f99e0995f25b9f095bc7434632601028cf86590aea5c8a5cb7801d3"
 
-[[package]]
-name = "windows_i686_msvc"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44d840b6ec649f480a41c8d80f9c65108b92d89345dd94027bfe06ac444d1060"
-
 [[package]]
 name = "windows_i686_msvc"
 version = "0.48.0"
@@ -7315,12 +7166,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1515e9a29e5bed743cb4415a9ecf5dfca648ce85ee42e15873c3cd8610ff8e02"
 
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8de912b8b8feb55c064867cf047dda097f92d51efad5b491dfb98f6bbb70cb36"
-
 [[package]]
 name = "windows_x86_64_gnu"
 version = "0.48.0"
@@ -7333,12 +7178,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5eee091590e89cc02ad514ffe3ead9eb6b660aedca2183455434b93546371a03"
 
-[[package]]
-name = "windows_x86_64_gnullvm"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26d41b46a36d453748aedef1486d5c7a85db22e56aff34643984ea85514e94a3"
-
 [[package]]
 name = "windows_x86_64_gnullvm"
 version = "0.48.0"
@@ -7351,12 +7190,6 @@ version = "0.52.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "77ca79f2451b49fa9e2af39f0747fe999fcda4f5e241b2898624dca97a1f2177"
 
-[[package]]
-name = "windows_x86_64_msvc"
-version = "0.42.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9aec5da331524158c6d1a4ac0ab1541149c0b9505fde06423b02f5ef0106b9f0"
-
 [[package]]
 name = "windows_x86_64_msvc"
 version = "0.48.0"
@@ -7433,10 +7266,11 @@ dependencies = [
  "futures-util",
  "generic-array",
  "getrandom 0.2.11",
+ "half",
  "hashbrown 0.14.5",
  "hex",
  "hmac",
- "hyper 0.14.26",
+ "hyper 0.14.30",
  "indexmap 1.9.3",
  "itertools 0.10.5",
  "itertools 0.12.1",
@@ -7504,7 +7338,7 @@ dependencies = [
  "der 0.7.8",
  "hex",
  "pem",
- "ring 0.17.6",
+ "ring",
  "signature 2.2.0",
  "spki 0.7.3",
  "thiserror",

Cargo.toml

@@ -76,8 +76,6 @@ clap = { version = "4.0", features = ["derive"] }
 comfy-table = "7.1"
 const_format = "0.2"
 crc32c = "0.6"
-crossbeam-deque = "0.8.5"
-crossbeam-utils = "0.8.5"
 dashmap = { version = "5.5.0", features = ["raw-api"] }
 either = "1.8"
 enum-map = "2.4.2"
@@ -95,7 +93,7 @@ hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
-hostname = "0.3.1"
+hostname = "0.4"
 http = {version = "1.1.0", features = ["std"]}
 http-types = { version = "2", default-features = false }
 humantime = "2.1"
@@ -104,7 +102,6 @@ hyper = "0.14"
 tokio-tungstenite = "0.20.0"
 indexmap = "2"
 indoc = "2"
-inotify = "0.10.2"
 ipnet = "2.9.0"
 itertools = "0.10"
 jsonwebtoken = "9"
@@ -113,7 +110,7 @@ libc = "0.2"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
-memoffset = "0.8"
+memoffset = "0.9"
 nix = { version = "0.27", features = ["dir", "fs", "process", "socket", "signal", "poll"] }
 notify = "6.0.0"
 num_cpus = "1.15"
@@ -142,7 +139,6 @@ rpds = "0.13"
 rustc-hash = "1.1.0"
 rustls = "0.22"
 rustls-pemfile = "2"
-rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
 sd-notify = "0.4.1"
@@ -164,7 +160,6 @@ strum_macros = "0.26"
 svg_fmt = "0.4.3"
 sync_wrapper = "0.1.2"
 tar = "0.4"
-task-local-extensions = "0.1.4"
 test-context = "0.3"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"

compute/Dockerfile.compute-node

@@ -3,13 +3,15 @@ ARG REPOSITORY=neondatabase
 ARG IMAGE=build-tools
 ARG TAG=pinned
 ARG BUILD_TAG
+ARG DEBIAN_FLAVOR=bullseye-slim
 
 #########################################################################################
 #
 # Layer "build-deps"
 #
 #########################################################################################
-FROM debian:bullseye-slim AS build-deps
+FROM debian:$DEBIAN_FLAVOR AS build-deps
+ARG DEBIAN_FLAVOR
 RUN apt update &&  \
     apt install -y git autoconf automake libtool build-essential bison flex libreadline-dev \
     zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget pkg-config libssl-dev \
@@ -280,7 +282,7 @@ FROM build-deps AS vector-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-COPY patches/pgvector.patch /pgvector.patch
+COPY compute/patches/pgvector.patch /pgvector.patch
 
 # By default, pgvector Makefile uses `-march=native`. We don't want that,
 # because we build the images on different machines than where we run them.
@@ -366,7 +368,7 @@ FROM build-deps AS rum-pg-build
 ARG PG_VERSION
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-COPY patches/rum.patch /rum.patch
+COPY compute/patches/rum.patch /rum.patch
 
 RUN case "${PG_VERSION}" in "v17") \
     echo "v17 extensions are not supported yet. Quit" && exit 0;; \
@@ -1027,10 +1029,47 @@ RUN cd compute_tools && mold -run cargo build --locked --profile release-line-de
 #
 #########################################################################################
 
-FROM debian:bullseye-slim AS compute-tools-image
+FROM debian:$DEBIAN_FLAVOR AS compute-tools-image
+ARG DEBIAN_FLAVOR
 
 COPY --from=compute-tools /home/nonroot/target/release-line-debug-size-lto/compute_ctl /usr/local/bin/compute_ctl
 
+#########################################################################################
+#
+# Layer "pgbouncer"
+#
+#########################################################################################
+
+FROM debian:$DEBIAN_FLAVOR AS pgbouncer
+ARG DEBIAN_FLAVOR
+RUN set -e \
+    && apt-get update \
+    && apt-get install -y \
+        build-essential \
+        git \
+        libevent-dev \
+        libtool \
+        pkg-config
+
+# Use `dist_man_MANS=` to skip manpage generation (which requires python3/pandoc)
+ENV PGBOUNCER_TAG=pgbouncer_1_22_1
+RUN set -e \
+    && git clone --recurse-submodules --depth 1 --branch ${PGBOUNCER_TAG} https://github.com/pgbouncer/pgbouncer.git pgbouncer \
+    && cd pgbouncer \
+    && ./autogen.sh \
+    && LDFLAGS=-static ./configure --prefix=/usr/local/pgbouncer --without-openssl \
+    && make -j $(nproc) dist_man_MANS= \
+    && make install dist_man_MANS=
+
+#########################################################################################
+#
+# Layers "postgres-exporter" and "sql-exporter"
+#
+#########################################################################################
+
+FROM quay.io/prometheuscommunity/postgres-exporter:v0.12.1 AS postgres-exporter
+FROM burningalchemist/sql_exporter:0.13 AS sql-exporter
+
 #########################################################################################
 #
 # Clean up postgres folder before inclusion
@@ -1078,7 +1117,7 @@ COPY --from=pgjwt-pg-build /pgjwt.tar.gz /ext-src
 COPY --from=hypopg-pg-build /hypopg.tar.gz /ext-src
 COPY --from=pg-hashids-pg-build /pg_hashids.tar.gz /ext-src
 COPY --from=rum-pg-build /rum.tar.gz /ext-src
-COPY patches/rum.patch /ext-src
+COPY compute/patches/rum.patch /ext-src
 #COPY --from=pgtap-pg-build /pgtap.tar.gz /ext-src
 COPY --from=ip4r-pg-build /ip4r.tar.gz /ext-src
 COPY --from=prefix-pg-build /prefix.tar.gz /ext-src
@@ -1086,9 +1125,9 @@ COPY --from=hll-pg-build /hll.tar.gz /ext-src
 COPY --from=plpgsql-check-pg-build /plpgsql_check.tar.gz /ext-src
 #COPY --from=timescaledb-pg-build /timescaledb.tar.gz /ext-src
 COPY --from=pg-hint-plan-pg-build /pg_hint_plan.tar.gz /ext-src
-COPY patches/pg_hint_plan.patch /ext-src
+COPY compute/patches/pg_hint_plan.patch /ext-src
 COPY --from=pg-cron-pg-build /pg_cron.tar.gz /ext-src
-COPY patches/pg_cron.patch /ext-src
+COPY compute/patches/pg_cron.patch /ext-src
 #COPY --from=pg-pgx-ulid-build /home/nonroot/pgx_ulid.tar.gz /ext-src
 #COPY --from=rdkit-pg-build /rdkit.tar.gz /ext-src
 COPY --from=pg-uuidv7-pg-build /pg_uuidv7.tar.gz /ext-src
@@ -1097,7 +1136,7 @@ COPY --from=pg-semver-pg-build /pg_semver.tar.gz /ext-src
 #COPY --from=pg-embedding-pg-build /home/nonroot/pg_embedding-src/ /ext-src
 #COPY --from=wal2json-pg-build /wal2json_2_5.tar.gz /ext-src
 COPY --from=pg-anon-pg-build /pg_anon.tar.gz /ext-src
-COPY patches/pg_anon.patch /ext-src
+COPY compute/patches/pg_anon.patch /ext-src
 COPY --from=pg-ivm-build /pg_ivm.tar.gz /ext-src
 COPY --from=pg-partman-build /pg_partman.tar.gz /ext-src
 RUN case "${PG_VERSION}" in "v17") \
@@ -1144,7 +1183,9 @@ ENV PGDATABASE=postgres
 # Put it all together into the final image
 #
 #########################################################################################
-FROM debian:bullseye-slim
+FROM debian:$DEBIAN_FLAVOR
+ARG DEBIAN_FLAVOR
+ENV DEBIAN_FLAVOR=$DEBIAN_FLAVOR
 # Add user postgres
 RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
     echo "postgres:test_console_pass" | chpasswd && \
@@ -1160,23 +1201,50 @@ RUN mkdir /var/db && useradd -m -d /var/db/postgres postgres && \
 COPY --from=postgres-cleanup-layer --chown=postgres /usr/local/pgsql /usr/local
 COPY --from=compute-tools --chown=postgres /home/nonroot/target/release-line-debug-size-lto/compute_ctl /usr/local/bin/compute_ctl
 
+# pgbouncer and its config
+COPY --from=pgbouncer         /usr/local/pgbouncer/bin/pgbouncer /usr/local/bin/pgbouncer
+COPY --chmod=0666 --chown=postgres compute/etc/pgbouncer.ini /etc/pgbouncer.ini
+
+# Metrics exporter binaries and  configuration files
+COPY --from=postgres-exporter /bin/postgres_exporter /bin/postgres_exporter
+COPY --from=sql-exporter      /bin/sql_exporter      /bin/sql_exporter
+
+COPY --chmod=0644 compute/etc/sql_exporter.yml               /etc/sql_exporter.yml
+COPY --chmod=0644 compute/etc/neon_collector.yml             /etc/neon_collector.yml
+COPY --chmod=0644 compute/etc/sql_exporter_autoscaling.yml   /etc/sql_exporter_autoscaling.yml
+COPY --chmod=0644 compute/etc/neon_collector_autoscaling.yml /etc/neon_collector_autoscaling.yml
+
 # Create remote extension download directory
 RUN mkdir /usr/local/download_extensions && chown -R postgres:postgres /usr/local/download_extensions
 
 # Install:
 # libreadline8 for psql
-# libicu67, locales for collations (including ICU and plpgsql_check)
 # liblz4-1 for lz4
 # libossp-uuid16 for extension ossp-uuid
-# libgeos, libgdal, libsfcgal1, libproj and libprotobuf-c1 for PostGIS
+# libgeos, libsfcgal1, and libprotobuf-c1 for PostGIS
 # libxml2, libxslt1.1 for xml2
 # libzstd1 for zstd
 # libboost* for rdkit
 # ca-certificates for communicating with s3 by compute_ctl
-RUN apt update &&  \
+
+
+RUN apt update && \
+    case $DEBIAN_FLAVOR in \
+      # Version-specific installs for Bullseye (PG14-PG16):
+      # libicu67, locales for collations (including ICU and plpgsql_check)
+      # libgdal28, libproj19 for PostGIS
+      bullseye*) \
+        VERSION_INSTALLS="libicu67 libgdal28 libproj19"; \
+      ;; \
+      # Version-specific installs for Bookworm (PG17):
+      # libicu72, locales for collations (including ICU and plpgsql_check)
+      # libgdal32, libproj25 for PostGIS
+      bookworm*) \
+        VERSION_INSTALLS="libicu72 libgdal32 libproj25"; \
+      ;; \
+    esac && \
     apt install --no-install-recommends -y \
         gdb \
-        libicu67 \
         liblz4-1 \
         libreadline8 \
         libboost-iostreams1.74.0 \
@@ -1185,8 +1253,6 @@ RUN apt update &&  \
         libboost-system1.74.0 \
         libossp-uuid16 \
         libgeos-c1v5 \
-        libgdal28 \
-        libproj19 \
         libprotobuf-c1 \
         libsfcgal1 \
         libxml2 \
@@ -1195,7 +1261,8 @@ RUN apt update &&  \
         libcurl4-openssl-dev \
         locales \
         procps \
-        ca-certificates && \
+        ca-certificates \
+        $VERSION_INSTALLS && \
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* && \
     localedef -i en_US -c -f UTF-8 -A /usr/share/locale/locale.alias en_US.UTF-8
 

compute/README.md

@@ -0,0 +1,21 @@
+This directory contains files that are needed to build the compute
+images, or included in the compute images.
+
+Dockerfile.compute-node
+	To build the compute image
+
+vm-image-spec.yaml
+	Instructions for vm-builder, to turn the compute-node image into
+	corresponding vm-compute-node image.
+
+etc/
+	Configuration files included in /etc in the compute image
+
+patches/
+	Some extensions need to be patched to work with Neon. This
+	directory contains such patches. They are applied to the extension
+	sources in Dockerfile.compute-node
+
+In addition to these, postgres itself, the neon postgres extension,
+and compute_ctl are built and copied into the compute image by
+Dockerfile.compute-node.

compute/etc/neon_collector.yml

@@ -0,0 +1,246 @@
+collector_name: neon_collector
+metrics:
+- metric_name: lfc_misses
+  type: gauge
+  help: 'lfc_misses'
+  key_labels:
+  values: [lfc_misses]
+  query: |
+    select lfc_value as lfc_misses from neon.neon_lfc_stats where lfc_key='file_cache_misses';
+
+- metric_name: lfc_used
+  type: gauge
+  help: 'LFC chunks used (chunk = 1MB)'
+  key_labels:
+  values: [lfc_used]
+  query: |
+    select lfc_value as lfc_used from neon.neon_lfc_stats where lfc_key='file_cache_used';
+
+- metric_name: lfc_hits
+  type: gauge
+  help: 'lfc_hits'
+  key_labels:
+  values: [lfc_hits]
+  query: |
+    select lfc_value as lfc_hits from neon.neon_lfc_stats where lfc_key='file_cache_hits';
+
+- metric_name: lfc_writes
+  type: gauge
+  help: 'lfc_writes'
+  key_labels:
+  values: [lfc_writes]
+  query: |
+    select lfc_value as lfc_writes from neon.neon_lfc_stats where lfc_key='file_cache_writes';
+
+- metric_name: lfc_cache_size_limit
+  type: gauge
+  help: 'LFC cache size limit in bytes'
+  key_labels:
+  values: [lfc_cache_size_limit]
+  query: |
+    select pg_size_bytes(current_setting('neon.file_cache_size_limit')) as lfc_cache_size_limit;
+
+- metric_name: connection_counts
+  type: gauge
+  help: 'Connection counts'
+  key_labels:
+    - datname
+    - state
+  values: [count]
+  query: |
+    select datname, state, count(*) as count from pg_stat_activity where state <> '' group by datname, state;
+
+- metric_name: pg_stats_userdb
+  type: gauge
+  help: 'Stats for several oldest non-system dbs'
+  key_labels:
+    - datname
+  value_label: kind
+  values:
+    - db_size
+    - deadlocks
+    # Rows
+    - inserted
+    - updated
+    - deleted
+  # We export stats for 10 non-system database. Without this limit
+  # it is too easy to abuse the system by creating lots of databases.
+  query: |
+    select pg_database_size(datname) as db_size, deadlocks,
+       tup_inserted as inserted, tup_updated as updated, tup_deleted as deleted,
+       datname
+     from pg_stat_database
+     where datname IN (
+       select datname
+       from pg_database
+       where datname <> 'postgres' and not datistemplate
+       order by oid
+       limit 10
+     );
+
+- metric_name: max_cluster_size
+  type: gauge
+  help: 'neon.max_cluster_size setting'
+  key_labels:
+  values: [max_cluster_size]
+  query: |
+    select setting::int as max_cluster_size from pg_settings where name = 'neon.max_cluster_size';
+
+- metric_name: db_total_size
+  type: gauge
+  help: 'Size of all databases'
+  key_labels:
+  values: [total]
+  query: |
+    select sum(pg_database_size(datname)) as total from pg_database;
+
+# DEPRECATED
+- metric_name: lfc_approximate_working_set_size
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels:
+  values: [approximate_working_set_size]
+  query: |
+    select neon.approximate_working_set_size(false) as approximate_working_set_size;
+
+- metric_name: lfc_approximate_working_set_size_windows
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels: [duration]
+  values: [size]
+  # NOTE: This is the "public" / "human-readable" version. Here, we supply a small selection
+  # of durations in a pretty-printed form.
+  query: |
+    select
+      x as duration,
+      neon.approximate_working_set_size_seconds(extract('epoch' from x::interval)::int) as size
+    from
+      (values ('5m'),('15m'),('1h')) as t (x);
+
+- metric_name: compute_current_lsn
+  type: gauge
+  help: 'Current LSN of the database'
+  key_labels:
+  values: [lsn]
+  query: |
+    select
+      case
+        when pg_catalog.pg_is_in_recovery()
+        then (pg_last_wal_replay_lsn() - '0/0')::FLOAT8
+        else (pg_current_wal_lsn() - '0/0')::FLOAT8
+      end as lsn;
+
+- metric_name: compute_receive_lsn
+  type: gauge
+  help: 'Returns the last write-ahead log location that has been received and synced to disk by streaming replication'
+  key_labels:
+  values: [lsn]
+  query: |
+    SELECT
+      CASE
+        WHEN pg_catalog.pg_is_in_recovery()
+        THEN (pg_last_wal_receive_lsn() - '0/0')::FLOAT8
+        ELSE 0
+      END AS lsn;
+
+- metric_name: replication_delay_bytes
+  type: gauge
+  help: 'Bytes between received and replayed LSN'
+  key_labels:
+  values: [replication_delay_bytes]
+  # We use a GREATEST call here because this calculation can be negative.
+  # The calculation is not atomic, meaning after we've gotten the receive
+  # LSN, the replay LSN may have advanced past the receive LSN we
+  # are using for the calculation.
+  query: |
+    SELECT GREATEST(0, pg_wal_lsn_diff(pg_last_wal_receive_lsn(), pg_last_wal_replay_lsn())) AS replication_delay_bytes;
+
+- metric_name: replication_delay_seconds
+  type: gauge
+  help: 'Time since last LSN was replayed'
+  key_labels:
+  values: [replication_delay_seconds]
+  query: |
+    SELECT
+      CASE
+        WHEN pg_last_wal_receive_lsn() = pg_last_wal_replay_lsn() THEN 0
+        ELSE GREATEST (0, EXTRACT (EPOCH FROM now() - pg_last_xact_replay_timestamp()))
+     END AS replication_delay_seconds;
+
+- metric_name: checkpoints_req
+  type: gauge
+  help: 'Number of requested checkpoints'
+  key_labels:
+  values: [checkpoints_req]
+  query: |
+    SELECT checkpoints_req FROM pg_stat_bgwriter;
+
+- metric_name: checkpoints_timed
+  type: gauge
+  help: 'Number of scheduled checkpoints'
+  key_labels:
+  values: [checkpoints_timed]
+  query: |
+    SELECT checkpoints_timed FROM pg_stat_bgwriter;
+
+- metric_name: compute_logical_snapshot_files
+  type: gauge
+  help: 'Number of snapshot files in pg_logical/snapshot'
+  key_labels:
+    - timeline_id
+  values: [num_logical_snapshot_files]
+  query: |
+    SELECT
+      (SELECT setting FROM pg_settings WHERE name = 'neon.timeline_id') AS timeline_id,
+      -- Postgres creates temporary snapshot files of the form %X-%X.snap.%d.tmp. These
+      -- temporary snapshot files are renamed to the actual snapshot files after they are
+      -- completely built. We only WAL-log the completely built snapshot files.
+      (SELECT COUNT(*) FROM pg_ls_dir('pg_logical/snapshots') AS name WHERE name LIKE '%.snap') AS num_logical_snapshot_files;
+
+# In all the below metrics, we cast LSNs to floats because Prometheus only supports floats.
+# It's probably fine because float64 can store integers from -2^53 to +2^53 exactly.
+
+# Number of slots is limited by max_replication_slots, so collecting position for all of them shouldn't be bad.
+- metric_name: logical_slot_restart_lsn
+  type: gauge
+  help: 'restart_lsn of logical slots'
+  key_labels:
+    - slot_name
+  values: [restart_lsn]
+  query: |
+    select slot_name, (restart_lsn - '0/0')::FLOAT8 as restart_lsn
+    from pg_replication_slots
+    where slot_type = 'logical';
+
+- metric_name: compute_subscriptions_count
+  type: gauge
+  help: 'Number of logical replication subscriptions grouped by enabled/disabled'
+  key_labels:
+    - enabled
+  values: [subscriptions_count]
+  query: |
+    select subenabled::text as enabled, count(*) as subscriptions_count
+    from pg_subscription
+    group by subenabled;
+
+- metric_name: retained_wal
+  type: gauge
+  help: 'Retained WAL in inactive replication slots'
+  key_labels:
+    - slot_name
+  values: [retained_wal]
+  query: |
+    SELECT slot_name, pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn)::FLOAT8 AS retained_wal
+    FROM pg_replication_slots
+    WHERE active = false;
+
+- metric_name: wal_is_lost
+  type: gauge
+  help: 'Whether or not the replication slot wal_status is lost'
+  key_labels:
+    - slot_name
+  values: [wal_is_lost]
+  query: |
+    SELECT slot_name,
+           CASE WHEN wal_status = 'lost' THEN 1 ELSE 0 END AS wal_is_lost
+    FROM pg_replication_slots;

compute/etc/neon_collector_autoscaling.yml

@@ -0,0 +1,55 @@
+collector_name: neon_collector_autoscaling
+metrics:
+- metric_name: lfc_misses
+  type: gauge
+  help: 'lfc_misses'
+  key_labels:
+  values: [lfc_misses]
+  query: |
+    select lfc_value as lfc_misses from neon.neon_lfc_stats where lfc_key='file_cache_misses';
+
+- metric_name: lfc_used
+  type: gauge
+  help: 'LFC chunks used (chunk = 1MB)'
+  key_labels:
+  values: [lfc_used]
+  query: |
+    select lfc_value as lfc_used from neon.neon_lfc_stats where lfc_key='file_cache_used';
+
+- metric_name: lfc_hits
+  type: gauge
+  help: 'lfc_hits'
+  key_labels:
+  values: [lfc_hits]
+  query: |
+    select lfc_value as lfc_hits from neon.neon_lfc_stats where lfc_key='file_cache_hits';
+
+- metric_name: lfc_writes
+  type: gauge
+  help: 'lfc_writes'
+  key_labels:
+  values: [lfc_writes]
+  query: |
+    select lfc_value as lfc_writes from neon.neon_lfc_stats where lfc_key='file_cache_writes';
+
+- metric_name: lfc_cache_size_limit
+  type: gauge
+  help: 'LFC cache size limit in bytes'
+  key_labels:
+  values: [lfc_cache_size_limit]
+  query: |
+    select pg_size_bytes(current_setting('neon.file_cache_size_limit')) as lfc_cache_size_limit;
+
+- metric_name: lfc_approximate_working_set_size_windows
+  type: gauge
+  help: 'Approximate working set size in pages of 8192 bytes'
+  key_labels: [duration_seconds]
+  values: [size]
+  # NOTE: This is the "internal" / "machine-readable" version. This outputs the working set
+  # size looking back 1..60 minutes, labeled with the number of minutes.
+  query: |
+    select
+      x::text as duration_seconds,
+      neon.approximate_working_set_size_seconds(x) as size
+    from
+      (select generate_series * 60 as x from generate_series(1, 60)) as t (x);

compute/etc/pgbouncer.ini

@@ -0,0 +1,17 @@
+[databases]
+*=host=localhost port=5432 auth_user=cloud_admin
+[pgbouncer]
+listen_port=6432
+listen_addr=0.0.0.0
+auth_type=scram-sha-256
+auth_user=cloud_admin
+auth_dbname=postgres
+client_tls_sslmode=disable
+server_tls_sslmode=disable
+pool_mode=transaction
+max_client_conn=10000
+default_pool_size=64
+max_prepared_statements=0
+admin_users=postgres
+unix_socket_dir=/tmp/
+unix_socket_mode=0777

compute/etc/sql_exporter.yml

@@ -0,0 +1,33 @@
+# Configuration for sql_exporter
+# Global defaults.
+global:
+  # If scrape_timeout <= 0, no timeout is set unless Prometheus provides one. The default is 10s.
+  scrape_timeout: 10s
+  # Subtracted from Prometheus' scrape_timeout to give us some headroom and prevent Prometheus from timing out first.
+  scrape_timeout_offset: 500ms
+  # Minimum interval between collector runs: by default (0s) collectors are executed on every scrape.
+  min_interval: 0s
+  # Maximum number of open connections to any one target. Metric queries will run concurrently on multiple connections,
+  # as will concurrent scrapes.
+  max_connections: 1
+  # Maximum number of idle connections to any one target. Unless you use very long collection intervals, this should
+  # always be the same as max_connections.
+  max_idle_connections: 1
+  # Maximum number of maximum amount of time a connection may be reused. Expired connections may be closed lazily before reuse.
+  # If 0, connections are not closed due to a connection's age.
+  max_connection_lifetime: 5m
+
+# The target to monitor and the collectors to execute on it.
+target:
+  # Data source name always has a URI schema that matches the driver name. In some cases (e.g. MySQL)
+  # the schema gets dropped or replaced to match the driver expected DSN format.
+  data_source_name: 'postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter'
+
+  # Collectors (referenced by name) to execute on the target.
+  # Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+  collectors: [neon_collector]
+
+# Collector files specifies a list of globs. One collector definition is read from each matching file.
+# Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+collector_files:
+  - "neon_collector.yml"

compute/etc/sql_exporter_autoscaling.yml

@@ -0,0 +1,33 @@
+# Configuration for sql_exporter for autoscaling-agent
+# Global defaults.
+global:
+  # If scrape_timeout <= 0, no timeout is set unless Prometheus provides one. The default is 10s.
+  scrape_timeout: 10s
+  # Subtracted from Prometheus' scrape_timeout to give us some headroom and prevent Prometheus from timing out first.
+  scrape_timeout_offset: 500ms
+  # Minimum interval between collector runs: by default (0s) collectors are executed on every scrape.
+  min_interval: 0s
+  # Maximum number of open connections to any one target. Metric queries will run concurrently on multiple connections,
+  # as will concurrent scrapes.
+  max_connections: 1
+  # Maximum number of idle connections to any one target. Unless you use very long collection intervals, this should
+  # always be the same as max_connections.
+  max_idle_connections: 1
+  # Maximum number of maximum amount of time a connection may be reused. Expired connections may be closed lazily before reuse.
+  # If 0, connections are not closed due to a connection's age.
+  max_connection_lifetime: 5m
+
+# The target to monitor and the collectors to execute on it.
+target:
+  # Data source name always has a URI schema that matches the driver name. In some cases (e.g. MySQL)
+  # the schema gets dropped or replaced to match the driver expected DSN format.
+  data_source_name: 'postgresql://cloud_admin@127.0.0.1:5432/postgres?sslmode=disable&application_name=sql_exporter_autoscaling'
+
+  # Collectors (referenced by name) to execute on the target.
+  # Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+  collectors: [neon_collector_autoscaling]
+
+# Collector files specifies a list of globs. One collector definition is read from each matching file.
+# Glob patterns are supported (see <https://pkg.go.dev/path/filepath#Match> for syntax).
+collector_files:
+  - "neon_collector_autoscaling.yml"

compute/vm-image-spec.yaml

@@ -0,0 +1,112 @@
+# Supplemental file for neondatabase/autoscaling's vm-builder, for producing the VM compute image.
+---
+commands:
+  - name: cgconfigparser
+    user: root
+    sysvInitAction: sysinit
+    shell: 'cgconfigparser -l /etc/cgconfig.conf -s 1664'
+  # restrict permissions on /neonvm/bin/resize-swap, because we grant access to compute_ctl for
+  # running it as root.
+  - name: chmod-resize-swap
+    user: root
+    sysvInitAction: sysinit
+    shell: 'chmod 711 /neonvm/bin/resize-swap'
+  - name: pgbouncer
+    user: postgres
+    sysvInitAction: respawn
+    shell: '/usr/local/bin/pgbouncer /etc/pgbouncer.ini'
+  - name: postgres-exporter
+    user: nobody
+    sysvInitAction: respawn
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter" /bin/postgres_exporter'
+  - name: sql-exporter
+    user: nobody
+    sysvInitAction: respawn
+    shell: '/bin/sql_exporter -config.file=/etc/sql_exporter.yml -web.listen-address=:9399'
+  - name: sql-exporter-autoscaling
+    user: nobody
+    sysvInitAction: respawn
+    shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
+shutdownHook: |
+  su -p postgres --session-command '/usr/local/bin/pg_ctl stop -D /var/db/postgres/compute/pgdata -m fast --wait -t 10'
+files:
+  - filename: compute_ctl-resize-swap
+    content: |
+      # Allow postgres user (which is what compute_ctl runs as) to run /neonvm/bin/resize-swap
+      # as root without requiring entering a password (NOPASSWD), regardless of hostname (ALL)
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap
+  - filename: cgconfig.conf
+    content: |
+      # Configuration for cgroups in VM compute nodes
+      group neon-postgres {
+          perm {
+              admin {
+                  uid = postgres;
+              }
+              task {
+                  gid = users;
+              }
+          }
+          memory {}
+      }
+build: |
+  # Build cgroup-tools
+  #
+  # At time of writing (2023-03-14), debian bullseye has a version of cgroup-tools (technically
+  # libcgroup) that doesn't support cgroup v2 (version 0.41-11). Unfortunately, the vm-monitor
+  # requires cgroup v2, so we'll build cgroup-tools ourselves.
+  FROM debian:bullseye-slim as libcgroup-builder
+  ENV LIBCGROUP_VERSION=v2.0.3
+
+  RUN set -exu \
+      && apt update \
+      && apt install --no-install-recommends -y \
+          git \
+          ca-certificates \
+          automake \
+          cmake \
+          make \
+          gcc \
+          byacc \
+          flex \
+          libtool \
+          libpam0g-dev \
+      && git clone --depth 1 -b $LIBCGROUP_VERSION https://github.com/libcgroup/libcgroup \
+      && INSTALL_DIR="/libcgroup-install" \
+      && mkdir -p "$INSTALL_DIR/bin" "$INSTALL_DIR/include" \
+      && cd libcgroup \
+      # extracted from bootstrap.sh, with modified flags:
+      && (test -d m4 || mkdir m4) \
+      && autoreconf -fi \
+      && rm -rf autom4te.cache \
+      && CFLAGS="-O3" ./configure --prefix="$INSTALL_DIR" --sysconfdir=/etc --localstatedir=/var --enable-opaque-hierarchy="name=systemd" \
+      # actually build the thing...
+      && make install
+merge: |
+  # tweak nofile limits
+  RUN set -e \
+      && echo 'fs.file-max = 1048576' >>/etc/sysctl.conf \
+      && test ! -e /etc/security || ( \
+         echo '*    - nofile 1048576' >>/etc/security/limits.conf \
+      && echo 'root - nofile 1048576' >>/etc/security/limits.conf \
+         )
+
+  # Allow postgres user (compute_ctl) to run swap resizer.
+  # Need to install sudo in order to allow this.
+  #
+  # Also, remove the 'read' permission from group/other on /neonvm/bin/resize-swap, just to be safe.
+  RUN set -e \
+      && apt update \
+      && apt install --no-install-recommends -y \
+             sudo \
+      && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+  COPY compute_ctl-resize-swap /etc/sudoers.d/compute_ctl-resize-swap
+
+  COPY cgconfig.conf /etc/cgconfig.conf
+
+  RUN set -e \
+      && chmod 0644 /etc/cgconfig.conf
+
+  COPY --from=libcgroup-builder /libcgroup-install/bin/*  /usr/bin/
+  COPY --from=libcgroup-builder /libcgroup-install/lib/*  /usr/lib/
+  COPY --from=libcgroup-builder /libcgroup-install/sbin/* /usr/sbin/

compute_tools/src/configurator.rs

@@ -11,9 +11,17 @@ use crate::compute::ComputeNode;
 fn configurator_main_loop(compute: &Arc<ComputeNode>) {
     info!("waiting for reconfiguration requests");
     loop {
-        let state = compute.state.lock().unwrap();
-        let mut state = compute.state_changed.wait(state).unwrap();
+        let mut state = compute.state.lock().unwrap();
 
+        // We have to re-check the status after re-acquiring the lock because it could be that
+        // the status has changed while we were waiting for the lock, and we might not need to
+        // wait on the condition variable. Otherwise, we might end up in some soft-/deadlock, i.e.
+        // we are waiting for a condition variable that will never be signaled.
+        if state.status != ComputeStatus::ConfigurationPending {
+            state = compute.state_changed.wait(state).unwrap();
+        }
+
+        // Re-check the status after waking up
         if state.status == ComputeStatus::ConfigurationPending {
             info!("got configuration request");
             state.status = ComputeStatus::Configuration;

control_plane/Cargo.toml

@@ -9,7 +9,6 @@ anyhow.workspace = true
 camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
-git-version.workspace = true
 humantime.workspace = true
 nix.workspace = true
 once_cell.workspace = true

control_plane/src/storage_controller.rs

@@ -346,7 +346,14 @@ impl StorageController {
             let pg_log_path = pg_data_path.join("postgres.log");
 
             if !tokio::fs::try_exists(&pg_data_path).await? {
-                let initdb_args = ["-D", pg_data_path.as_ref(), "--username", &username()];
+                let initdb_args = [
+                    "-D",
+                    pg_data_path.as_ref(),
+                    "--username",
+                    &username(),
+                    "--no-sync",
+                    "--no-instructions",
+                ];
                 tracing::info!(
                     "Initializing storage controller database with args: {:?}",
                     initdb_args

control_plane/storcon_cli/src/main.rs

@@ -4,8 +4,8 @@ use std::{str::FromStr, time::Duration};
 use clap::{Parser, Subcommand};
 use pageserver_api::{
     controller_api::{
-        NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse, ShardSchedulingPolicy,
-        TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
+        AvailabilityZone, NodeAvailabilityWrapper, NodeDescribeResponse, NodeShardResponse,
+        ShardSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
     },
     models::{
         EvictionPolicy, EvictionPolicyLayerAccessThreshold, LocationConfigSecondary,
@@ -339,7 +339,7 @@ async fn main() -> anyhow::Result<()> {
                         listen_pg_port,
                         listen_http_addr,
                         listen_http_port,
-                        availability_zone_id,
+                        availability_zone_id: AvailabilityZone(availability_zone_id),
                     }),
                 )
                 .await?;

docker-compose/README.md

@@ -2,8 +2,8 @@
 # Example docker compose configuration
 
 The configuration in this directory is used for testing Neon docker images: it is
-not intended for deploying a usable system.  To run a development environment where
-you can experiment with a minature Neon system, use `cargo neon` rather than container images.
+not intended for deploying a usable system. To run a development environment where
+you can experiment with a miniature Neon system, use `cargo neon` rather than container images.
 
 This configuration does not start the storage controller, because the controller
 needs a way to reconfigure running computes, and no such thing exists in this setup.

docs/rfcs/038-independent-compute-release.md

@@ -0,0 +1,343 @@
+# Independent compute release
+
+Created at: 2024-08-30. Author: Alexey Kondratov (@ololobus)
+
+## Summary
+
+This document proposes an approach to fully independent compute release flow. It attempts to
+cover the following features:
+
+- Process is automated as much as possible to minimize human errors.
+- Compute<->storage protocol compatibility is ensured.
+- A transparent release history is available with an easy rollback strategy.
+- Although not in the scope of this document, there is a viable way to extend the proposed release
+  flow to achieve the canary and/or blue-green deployment strategies.
+
+## Motivation
+
+Previously, the compute release was tightly coupled to the storage release. This meant that once
+some storage nodes got restarted with a newer version, all new compute starts using these nodes
+automatically got a new version. Thus, two releases happen in parallel, which increases the blast
+radius and makes ownership fuzzy.
+
+Now, we practice a manual v0 independent compute release flow -- after getting a new compute release
+image and tag, we pin it region by region using Admin UI. It's better, but it still has its own flaws:
+
+1. It's a simple but fairly manual process, as you need to click through a few pages.
+2. It's prone to human errors, e.g., you could mistype or copy the wrong compute tag.
+3. We now require an additional approval in the Admin UI, which partially solves the 2.,
+   but also makes the whole process pretty annoying, as you constantly need to go back
+   and forth between two people.
+
+## Non-goals
+
+It's not the goal of this document to propose a design for some general-purpose release tool like Helm.
+The document considers how the current compute fleet is orchestrated at Neon. Even if we later
+decide to split the control plane further (e.g., introduce a separate compute controller), the proposed
+release process shouldn't change much, i.e., the releases table and API will reside in
+one of the parts.
+
+Achieving the canary and/or blue-green deploy strategies is out of the scope of this document. They
+were kept in mind, though, so it's expected that the proposed approach will lay down the foundation
+for implementing them in future iterations.
+
+## Impacted components
+
+Compute, control plane, CI, observability (some Grafana dashboards may require changes).
+
+## Prior art
+
+One of the very close examples is how Helm tracks [releases history](https://helm.sh/docs/helm/helm_history/).
+
+In the code:
+
+- [Release](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/release.go#L20-L43)
+- [Release info](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/info.go#L24-L40)
+- [Release status](https://github.com/helm/helm/blob/2b30cf4b61d587d3f7594102bb202b787b9918db/pkg/release/status.go#L18-L42)
+
+TL;DR it has several important attributes:
+
+- Revision -- unique release ID/primary key. It is not the same as the application version,
+  because the same version can be deployed several times, e.g., after a newer version rollback.
+- App version -- version of the application chart/code.
+- Config -- set of overrides to the default config of the application.
+- Status -- current status of the release in the history.
+- Timestamps -- tracks when a release was created and deployed.
+
+## Proposed implementation
+
+### Separate release branch
+
+We will use a separate release branch, `release-compute`, to have a clean history for releases and commits.
+In order to avoid confusion with storage releases, we will use a different prefix for compute [git release
+tags](https://github.com/neondatabase/neon/releases) -- `release-compute-XXXX`. We will use the same tag for
+Docker images as well. The `neondatabase/compute-node-v16:release-compute-XXXX` looks longer and a bit redundant,
+but it's better to have image and git tags in sync.
+
+Currently, control plane relies on the numeric compute and storage release versions to decide on compute->storage
+compatibility. Once we implement this proposal, we should drop this code as release numbers will be completely
+independent. The only constraint we want is that it must monotonically increase within the same release branch.
+
+### Compute config/settings manifest
+
+We will create a new sub-directory `compute` and file `compute/manifest.yaml` with a structure:
+
+```yaml
+pg_settings:
+  # Common settings for primaries and secondaries of all versions.
+  common:
+    wal_log_hints: "off"
+    max_wal_size: "1024"
+
+  per_version:
+    14:
+      # Common settings for both replica and primary of version PG 14
+      common:
+        shared_preload_libraries: "neon,pg_stat_statements,extension_x"
+    15:
+      common:
+        shared_preload_libraries: "neon,pg_stat_statements,extension_x"
+      # Settings that should be applied only to
+      replica:
+        # Available only starting Postgres 15th
+        recovery_prefetch: "off"
+    # ...
+    17:
+      common:
+        # For example, if third-party `extension_x` is not yet available for PG 17
+        shared_preload_libraries: "neon,pg_stat_statements"
+      replica:
+        recovery_prefetch: "off"
+```
+
+**N.B.** Setting value should be a string with `on|off` for booleans and a number (as a string)
+without units for all numeric settings. That's how the control plane currently operates.
+
+The priority of settings will be (a higher number is a higher priority):
+
+1. Any static and hard-coded settings in the control plane
+2. `pg_settings->common`
+3. Per-version `common`
+4. Per-version `replica`
+5. Any per-user/project/endpoint overrides in the control plane
+6. Any dynamic setting calculated based on the compute size
+
+**N.B.** For simplicity, we do not do any custom logic for `shared_preload_libraries`, so it's completely
+overridden if specified on some level. Make sure that you include all necessary extensions in it when you
+do any overrides.
+
+**N.B.** There is a tricky question about what to do with custom compute image pinning we sometimes
+do for particular projects and customers. That's usually some ad-hoc work and images are based on
+the latest compute image, so it's relatively safe to assume that we could use settings from the latest compute
+release. If for some reason that's not true, and further overrides are needed, it's also possible to do
+on the project level together with pinning the image, so it's on-call/engineer/support responsibility to
+ensure that compute starts with the specified custom image. The only real risk is that compute image will get
+stale and settings from new releases will drift away, so eventually it will get something incompatible,
+but i) this is some operational issue, as we do not want stale images anyway, and ii) base settings
+receive something really new so rarely that the chance of this happening is very low. If we want to solve it completely,
+then together with pinning the image we could also pin the matching release revision in the control plane.
+
+The compute team will own the content of `compute/manifest.yaml`.
+
+### Control plane: releases table
+
+In order to store information about releases, the control plane will use a table `compute_releases` with the following
+schema:
+
+```sql
+CREATE TABLE compute_releases (
+  -- Unique release ID
+  -- N.B. Revision won't by synchronized across all regions, because all control planes are technically independent
+  -- services. We have the same situation with Helm releases as well because they could be deployed and rolled back
+  -- independently in different clusters.
+  revision BIGSERIAL PRIMARY KEY,
+  -- Numeric version of the compute image, e.g. 9057
+  version BIGINT NOT NULL,
+  -- Compute image tag, e.g. `release-9057`
+  tag TEXT NOT NULL,
+  -- Current release status. Currently, it will be a simple enum
+  -- * `deployed` -- release is deployed and used for new compute starts.
+  --                 Exactly one release can have this status at a time.
+  -- * `superseded` -- release has been replaced by a newer one.
+  -- But we can always extend it in the future when we need more statuses
+  -- for more complex deployment strategies.
+  status TEXT NOT NULL,
+  -- Any additional metadata for compute in the corresponding release
+  manifest JSONB NOT NULL,
+  -- Timestamp when release record was created in the control plane database
+  created_at TIMESTAMP NOT NULL DEFAULT now(),
+  -- Timestamp when release deployment was finished
+  deployed_at TIMESTAMP
+);
+```
+
+We keep track of the old releases not only for the sake of audit, but also because we usually have ~30% of
+old computes started using the image from one of the previous releases. Yet, when users want to reconfigure
+them without restarting, the control plane needs to know what settings are applicable to them, so we also need
+information about the previous releases that are readily available. There could be some other auxiliary info
+needed as well: supported extensions, compute flags, etc.
+
+**N.B.** Here, we can end up in an ambiguous situation when the same compute image is deployed twice, e.g.,
+it was deployed once, then rolled back, and then deployed again, potentially with a different manifest. Yet,
+we could've started some computes with the first deployment and some with the second. Thus, when we need to
+look up the manifest for the compute by its image tag, we will see two records in the table with the same tag,
+but different revision numbers. We can assume that this could happen only in case of rollbacks, so we
+can just take the latest revision for the given tag.
+
+### Control plane: management API
+
+The control plane will implement new API methods to manage releases:
+
+1. `POST /management/api/v2/compute_releases` to create a new release. With payload
+
+   ```json
+    {
+      "version": 9057,
+      "tag": "release-9057",
+      "manifest": {}
+    }
+   ```
+
+   and response
+
+   ```json
+    {
+      "revision": 53,
+      "version": 9057,
+      "tag": "release-9057",
+      "status": "deployed",
+      "manifest": {},
+      "created_at": "2024-08-15T15:52:01.0000Z",
+      "deployed_at": "2024-08-15T15:52:01.0000Z",
+    }
+   ```
+
+   Here, we can actually mix-in custom (remote) extensions metadata into the `manifest`, so that the control plane
+   will get information about all available extensions not bundled into compute image. The corresponding
+   workflow in `neondatabase/build-custom-extensions` should produce it as an artifact and make
+   it accessible to the workflow in the `neondatabase/infra`. See the complete release flow below. Doing that,
+   we put a constraint that new custom extension requires new compute release, which is good for the safety,
+   but is not exactly what we want operational-wise (we want to be able to deploy new extensions without new
+   images). Yet, it can be solved incrementally: v0 -- do not do anything with extensions at all;
+   v1 -- put them into the same manifest; v2 -- make them separate entities with their own lifecycle.
+
+   **N.B.** This method is intended to be used in CI workflows, and CI/network can be flaky. It's reasonable
+   to assume that we could retry the request several times, even though it's already succeeded. Although it's
+   not a big deal to create several identical releases one-by-one, it's better to avoid it, so the control plane
+   should check if the latest release is identical and just return `304 Not Modified` in this case.
+
+2. `POST /management/api/v2/compute_releases/rollback` to rollback to any previously deployed release. With payload
+   including the revision of the release to rollback to:
+
+   ```json
+   {
+      "revision": 52
+   }
+   ```
+
+   Rollback marks the current release as `superseded` and creates a new release with all the same data as the
+   requested revision, but with a new revision number.
+
+   This rollback API is not strictly needed, as we can just use `infra` repo workflow to deploy any
+   available tag. It's still nice to have for on-call and any urgent matters, for example, if we need
+   to rollback and GitHub is down. It's much easier to specify only the revision number vs. crafting
+   all the necessary data for the new release payload.
+
+### Compute->storage compatibility tests
+
+In order to safely release new compute versions independently from storage, we need to ensure that the currently
+deployed storage is compatible with the new compute version. Currently, we maintain backward compatibility
+in storage, but newer computes may require a newer storage version.
+
+Remote end-to-end (e2e) tests [already accept](https://github.com/neondatabase/cloud/blob/e3468d433e0d73d02b7d7e738d027f509b522408/.github/workflows/testing.yml#L43-L48)
+`storage_image_tag` and `compute_image_tag` as separate inputs. That means that we could reuse e2e tests to ensure
+compatibility between storage and compute:
+
+1. Pick the latest storage release tag and use it as `storage_image_tag`.
+2. Pick a new compute tag built in the current compute release PR and use it as `compute_image_tag`.
+   Here, we should use a temporary ECR image tag, because the final tag will be known only after the release PR is merged.
+3. Trigger e2e tests as usual.
+
+### Release flow
+
+```mermaid
+  sequenceDiagram
+
+  actor oncall as Compute on-call person
+  participant neon as neondatabase/neon
+
+  box private
+    participant cloud as neondatabase/cloud
+    participant exts as neondatabase/build-custom-extensions
+    participant infra as neondatabase/infra
+  end
+
+  box cloud
+    participant preprod as Pre-prod control plane
+    participant prod as Production control plane
+    participant k8s as Compute k8s
+  end
+
+  oncall ->> neon: Open release PR into release-compute
+
+  activate neon
+  neon ->> cloud: CI: trigger e2e compatibility tests
+  activate cloud
+  cloud -->> neon: CI: e2e tests pass
+  deactivate cloud
+  neon ->> neon: CI: pass PR checks, get approvals
+  deactivate neon
+
+  oncall ->> neon: Merge release PR into release-compute
+
+  activate neon
+  neon ->> neon: CI: pass checks, build and push images
+  neon ->> exts: CI: trigger extensions build
+  activate exts
+  exts -->> neon: CI: extensions are ready
+  deactivate exts
+  neon ->> neon: CI: create release tag
+  neon ->> infra: Trigger release workflow using the produced tag
+  deactivate neon
+
+  activate infra
+  infra ->> infra: CI: pass checks
+  infra ->> preprod: Release new compute image to pre-prod automatically <br/> POST /management/api/v2/compute_releases
+  activate preprod
+  preprod -->> infra: 200 OK
+  deactivate preprod
+
+  infra ->> infra: CI: wait for per-region production deploy approvals
+  oncall ->> infra: CI: approve deploys region by region
+  infra ->> k8s: Prewarm new compute image
+  infra ->> prod: POST /management/api/v2/compute_releases
+  activate prod
+  prod -->> infra: 200 OK
+  deactivate prod
+  deactivate infra
+```
+
+## Further work
+
+As briefly mentioned in other sections, eventually, we would like to use more complex deployment strategies.
+For example, we can pass a fraction of the total compute starts that should use the new release. Then we can
+mark the release as `partial` or `canary` and monitor its performance. If everything is fine, we can promote it
+to `deployed` status. If not, we can roll back to the previous one.
+
+## Alternatives
+
+In theory, we can try using Helm as-is:
+
+1. Write a compute Helm chart. That will actually have only some config map, which the control plane can access and read.
+   N.B. We could reuse the control plane chart as well, but then it's not a fully independent release again and even more fuzzy.
+2. The control plane will read it and start using the new compute version for new starts.
+
+Drawbacks:
+
+1. Helm releases work best if the workload is controlled by the Helm chart itself. Then you can have different
+   deployment strategies like rolling update or canary or blue/green deployments. At Neon, the compute starts are controlled
+   by control plane, so it makes it much more tricky.
+2. Releases visibility will suffer, i.e. instead of a nice table in the control plane and Admin UI, we would need to use
+   `helm` cli and/or K8s UIs like K8sLens.
+3. We do not restart all computes shortly after the new version release. This means that for some features and compatibility
+   purpose (see above) control plane may need some auxiliary info from the previous releases.

libs/compute_api/src/spec.rs

@@ -268,6 +268,22 @@ pub struct GenericOption {
 /// declare a `trait` on it.
 pub type GenericOptions = Option<Vec<GenericOption>>;
 
+/// Configured the local-proxy application with the relevant JWKS and roles it should
+/// use for authorizing connect requests using JWT.
+#[derive(Clone, Debug, Deserialize, Serialize)]
+pub struct LocalProxySpec {
+    pub jwks: Vec<JwksSettings>,
+}
+
+#[derive(Clone, Debug, Deserialize, Serialize)]
+pub struct JwksSettings {
+    pub id: String,
+    pub role_names: Vec<String>,
+    pub jwks_url: String,
+    pub provider_name: String,
+    pub jwt_audience: Option<String>,
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/pageserver_api/src/config.rs

@@ -104,9 +104,6 @@ pub struct ConfigToml {
     pub image_compression: ImageCompressionAlgorithm,
     pub ephemeral_bytes_per_memory_kb: usize,
     pub l0_flush: Option<crate::models::L0FlushConfig>,
-    #[serde(skip_serializing)]
-    // TODO(https://github.com/neondatabase/neon/issues/8184): remove after this field is removed from all pageserver.toml's
-    pub compact_level0_phase1_value_access: serde::de::IgnoredAny,
     pub virtual_file_direct_io: crate::models::virtual_file::DirectIoMode,
     pub io_buffer_alignment: usize,
 }
@@ -384,7 +381,6 @@ impl Default for ConfigToml {
             image_compression: (DEFAULT_IMAGE_COMPRESSION),
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
             l0_flush: None,
-            compact_level0_phase1_value_access: Default::default(),
             virtual_file_direct_io: crate::models::virtual_file::DirectIoMode::default(),
 
             io_buffer_alignment: DEFAULT_IO_BUFFER_ALIGNMENT,

libs/pageserver_api/src/controller_api.rs

@@ -1,4 +1,5 @@
 use std::collections::{HashMap, HashSet};
+use std::fmt::Display;
 use std::str::FromStr;
 use std::time::{Duration, Instant};
 
@@ -57,7 +58,7 @@ pub struct NodeRegisterRequest {
     pub listen_http_addr: String,
     pub listen_http_port: u16,
 
-    pub availability_zone_id: String,
+    pub availability_zone_id: AvailabilityZone,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -74,10 +75,19 @@ pub struct TenantPolicyRequest {
     pub scheduling: Option<ShardSchedulingPolicy>,
 }
 
+#[derive(Clone, Serialize, Deserialize, PartialEq, Eq, Hash)]
+pub struct AvailabilityZone(pub String);
+
+impl Display for AvailabilityZone {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
 #[derive(Serialize, Deserialize)]
 pub struct ShardsPreferredAzsRequest {
     #[serde(flatten)]
-    pub preferred_az_ids: HashMap<TenantShardId, String>,
+    pub preferred_az_ids: HashMap<TenantShardId, AvailabilityZone>,
 }
 
 #[derive(Serialize, Deserialize)]

libs/pageserver_api/src/models.rs

@@ -37,14 +37,11 @@ use bytes::{Buf, BufMut, Bytes, BytesMut};
 /// ```mermaid
 /// stateDiagram-v2
 ///
-///     [*] --> Loading: spawn_load()
 ///     [*] --> Attaching: spawn_attach()
 ///
-///     Loading --> Activating: activate()
 ///     Attaching --> Activating: activate()
 ///     Activating --> Active: infallible
 ///
-///     Loading --> Broken: load() failure
 ///     Attaching --> Broken: attach() failure
 ///
 ///     Active --> Stopping: set_stopping(), part of shutdown & detach
@@ -68,10 +65,6 @@ use bytes::{Buf, BufMut, Bytes, BytesMut};
 )]
 #[serde(tag = "slug", content = "data")]
 pub enum TenantState {
-    /// This tenant is being loaded from local disk.
-    ///
-    /// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
-    Loading,
     /// This tenant is being attached to the pageserver.
     ///
     /// `set_stopping()` and `set_broken()` do not work in this state and wait for it to pass.
@@ -121,8 +114,6 @@ impl TenantState {
             // But, our attach task might still be fetching the remote timelines, etc.
             // So, return `Maybe` while Attaching, making Console wait for the attach task to finish.
             Self::Attaching | Self::Activating(ActivatingFrom::Attaching) => Maybe,
-            // tenant mgr startup distinguishes attaching from loading via marker file.
-            Self::Loading | Self::Activating(ActivatingFrom::Loading) => Attached,
             // We only reach Active after successful load / attach.
             // So, call atttachment status Attached.
             Self::Active => Attached,
@@ -191,10 +182,11 @@ impl LsnLease {
 }
 
 /// The only [`TenantState`] variants we could be `TenantState::Activating` from.
+///
+/// XXX: We used to have more variants here, but now it's just one, which makes this rather
+/// useless. Remove, once we've checked that there's no client code left that looks at this.
 #[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum ActivatingFrom {
-    /// Arrived to [`TenantState::Activating`] from [`TenantState::Loading`]
-    Loading,
     /// Arrived to [`TenantState::Activating`] from [`TenantState::Attaching`]
     Attaching,
 }
@@ -1562,11 +1554,8 @@ mod tests {
 
     #[test]
     fn tenantstatus_activating_serde() {
-        let states = [
-            TenantState::Activating(ActivatingFrom::Loading),
-            TenantState::Activating(ActivatingFrom::Attaching),
-        ];
-        let expected = "[{\"slug\":\"Activating\",\"data\":\"Loading\"},{\"slug\":\"Activating\",\"data\":\"Attaching\"}]";
+        let states = [TenantState::Activating(ActivatingFrom::Attaching)];
+        let expected = "[{\"slug\":\"Activating\",\"data\":\"Attaching\"}]";
 
         let actual = serde_json::to_string(&states).unwrap();
 
@@ -1581,13 +1570,7 @@ mod tests {
     fn tenantstatus_activating_strum() {
         // tests added, because we use these for metrics
         let examples = [
-            (line!(), TenantState::Loading, "Loading"),
             (line!(), TenantState::Attaching, "Attaching"),
-            (
-                line!(),
-                TenantState::Activating(ActivatingFrom::Loading),
-                "Activating",
-            ),
             (
                 line!(),
                 TenantState::Activating(ActivatingFrom::Attaching),

libs/postgres_backend/src/lib.rs

@@ -984,6 +984,7 @@ pub fn short_error(e: &QueryError) -> String {
 }
 
 fn log_query_error(query: &str, e: &QueryError) {
+    // If you want to change the log level of a specific error, also re-categorize it in `BasebackupQueryTimeOngoingRecording`.
     match e {
         QueryError::Disconnected(ConnectionError::Io(io_error)) => {
             if is_expected_io_error(io_error) {

libs/utils/Cargo.toml

@@ -19,6 +19,7 @@ bincode.workspace = true
 bytes.workspace = true
 camino.workspace = true
 chrono.workspace = true
+git-version.workspace = true
 hex = { workspace = true, features = ["serde"] }
 humantime.workspace = true
 hyper = { workspace = true, features = ["full"] }

libs/utils/src/lib.rs

@@ -92,6 +92,10 @@ pub mod toml_edit_ext;
 
 pub mod circuit_breaker;
 
+// Re-export used in macro. Avoids adding git-version as dep in target crates.
+#[doc(hidden)]
+pub use git_version;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:
@@ -131,7 +135,7 @@ macro_rules! project_git_version {
     ($const_identifier:ident) => {
         // this should try GIT_VERSION first only then git_version::git_version!
         const $const_identifier: &::core::primitive::str = {
-            const __COMMIT_FROM_GIT: &::core::primitive::str = git_version::git_version! {
+            const __COMMIT_FROM_GIT: &::core::primitive::str = $crate::git_version::git_version! {
                 prefix = "",
                 fallback = "unknown",
                 args = ["--abbrev=40", "--always", "--dirty=-modified"] // always use full sha

pageserver/Cargo.toml

@@ -27,7 +27,6 @@ crc32c.workspace = true
 either.workspace = true
 fail.workspace = true
 futures.workspace = true
-git-version.workspace = true
 hex.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true

pageserver/benches/bench_walredo.rs

@@ -1,7 +1,7 @@
 //! Quantify a single walredo manager's throughput under N concurrent callers.
 //!
 //! The benchmark implementation ([`bench_impl`]) is parametrized by
-//! - `redo_work` => [`Request::short_request`] or [`Request::medium_request`]
+//! - `redo_work` => an async closure that takes a `PostgresRedoManager` and performs one redo
 //! - `n_redos` => number of times the benchmark shell execute the `redo_work`
 //! - `nclients` => number of clients (more on this shortly).
 //!
@@ -10,7 +10,7 @@
 //! Each task executes the `redo_work` `n_redos/nclients` times.
 //!
 //! We exercise the following combinations:
-//! - `redo_work = short / medium``
+//! - `redo_work = ping / short / medium``
 //! - `nclients = [1, 2, 4, 8, 16, 32, 64, 128]`
 //!
 //! We let `criterion` determine the `n_redos` using `iter_custom`.
@@ -27,33 +27,43 @@
 //!
 //! # Reference Numbers
 //!
-//! 2024-04-15 on i3en.3xlarge
+//! 2024-09-18 on im4gn.2xlarge
 //!
 //! ```text
-//! short/1           time:   [24.584 µs 24.737 µs 24.922 µs]
-//! short/2           time:   [33.479 µs 33.660 µs 33.888 µs]
-//! short/4           time:   [42.713 µs 43.046 µs 43.440 µs]
-//! short/8           time:   [71.814 µs 72.478 µs 73.240 µs]
-//! short/16          time:   [132.73 µs 134.45 µs 136.22 µs]
-//! short/32          time:   [258.31 µs 260.73 µs 263.27 µs]
-//! short/64          time:   [511.61 µs 514.44 µs 517.51 µs]
-//! short/128         time:   [992.64 µs 998.23 µs 1.0042 ms]
-//! medium/1          time:   [110.11 µs 110.50 µs 110.96 µs]
-//! medium/2          time:   [153.06 µs 153.85 µs 154.99 µs]
-//! medium/4          time:   [317.51 µs 319.92 µs 322.85 µs]
-//! medium/8          time:   [638.30 µs 644.68 µs 652.12 µs]
-//! medium/16         time:   [1.2651 ms 1.2773 ms 1.2914 ms]
-//! medium/32         time:   [2.5117 ms 2.5410 ms 2.5720 ms]
-//! medium/64         time:   [4.8088 ms 4.8555 ms 4.9047 ms]
-//! medium/128        time:   [8.8311 ms 8.9849 ms 9.1263 ms]
+//! ping/1                  time:   [21.789 µs 21.918 µs 22.078 µs]
+//! ping/2                  time:   [27.686 µs 27.812 µs 27.970 µs]
+//! ping/4                  time:   [35.468 µs 35.671 µs 35.926 µs]
+//! ping/8                  time:   [59.682 µs 59.987 µs 60.363 µs]
+//! ping/16                 time:   [101.79 µs 102.37 µs 103.08 µs]
+//! ping/32                 time:   [184.18 µs 185.15 µs 186.36 µs]
+//! ping/64                 time:   [349.86 µs 351.45 µs 353.47 µs]
+//! ping/128                time:   [684.53 µs 687.98 µs 692.17 µs]
+//! short/1                 time:   [31.833 µs 32.126 µs 32.428 µs]
+//! short/2                 time:   [35.558 µs 35.756 µs 35.992 µs]
+//! short/4                 time:   [44.850 µs 45.138 µs 45.484 µs]
+//! short/8                 time:   [65.985 µs 66.379 µs 66.853 µs]
+//! short/16                time:   [127.06 µs 127.90 µs 128.87 µs]
+//! short/32                time:   [252.98 µs 254.70 µs 256.73 µs]
+//! short/64                time:   [497.13 µs 499.86 µs 503.26 µs]
+//! short/128               time:   [987.46 µs 993.45 µs 1.0004 ms]
+//! medium/1                time:   [137.91 µs 138.55 µs 139.35 µs]
+//! medium/2                time:   [192.00 µs 192.91 µs 194.07 µs]
+//! medium/4                time:   [389.62 µs 391.55 µs 394.01 µs]
+//! medium/8                time:   [776.80 µs 780.33 µs 784.77 µs]
+//! medium/16               time:   [1.5323 ms 1.5383 ms 1.5459 ms]
+//! medium/32               time:   [3.0120 ms 3.0226 ms 3.0350 ms]
+//! medium/64               time:   [5.7405 ms 5.7787 ms 5.8166 ms]
+//! medium/128              time:   [10.412 ms 10.574 ms 10.718 ms]
 //! ```
 
 use anyhow::Context;
 use bytes::{Buf, Bytes};
 use criterion::{BenchmarkId, Criterion};
+use once_cell::sync::Lazy;
 use pageserver::{config::PageServerConf, walrecord::NeonWalRecord, walredo::PostgresRedoManager};
 use pageserver_api::{key::Key, shard::TenantShardId};
 use std::{
+    future::Future,
     sync::Arc,
     time::{Duration, Instant},
 };
@@ -61,40 +71,59 @@ use tokio::{sync::Barrier, task::JoinSet};
 use utils::{id::TenantId, lsn::Lsn};
 
 fn bench(c: &mut Criterion) {
-    {
-        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-        for nclients in nclients {
-            let mut group = c.benchmark_group("short");
-            group.bench_with_input(
-                BenchmarkId::from_parameter(nclients),
-                &nclients,
-                |b, nclients| {
-                    let redo_work = Arc::new(Request::short_input());
-                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
-                },
-            );
-        }
-    }
-    {
-        let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
-        for nclients in nclients {
-            let mut group = c.benchmark_group("medium");
-            group.bench_with_input(
-                BenchmarkId::from_parameter(nclients),
-                &nclients,
-                |b, nclients| {
-                    let redo_work = Arc::new(Request::medium_input());
-                    b.iter_custom(|iters| bench_impl(Arc::clone(&redo_work), iters, *nclients));
-                },
-            );
-        }
+    macro_rules! bench_group {
+        ($name:expr, $redo_work:expr) => {{
+            let name: &str = $name;
+            let nclients = [1, 2, 4, 8, 16, 32, 64, 128];
+            for nclients in nclients {
+                let mut group = c.benchmark_group(name);
+                group.bench_with_input(
+                    BenchmarkId::from_parameter(nclients),
+                    &nclients,
+                    |b, nclients| {
+                        b.iter_custom(|iters| bench_impl($redo_work, iters, *nclients));
+                    },
+                );
+            }
+        }};
     }
+    //
+    // benchmark the protocol implementation
+    //
+    let pg_version = 14;
+    bench_group!(
+        "ping",
+        Arc::new(move |mgr: Arc<PostgresRedoManager>| async move {
+            let _: () = mgr.ping(pg_version).await.unwrap();
+        })
+    );
+    //
+    // benchmarks with actual record redo
+    //
+    let make_redo_work = |req: &'static Request| {
+        Arc::new(move |mgr: Arc<PostgresRedoManager>| async move {
+            let page = req.execute(&mgr).await.unwrap();
+            assert_eq!(page.remaining(), 8192);
+        })
+    };
+    bench_group!("short", {
+        static REQUEST: Lazy<Request> = Lazy::new(Request::short_input);
+        make_redo_work(&REQUEST)
+    });
+    bench_group!("medium", {
+        static REQUEST: Lazy<Request> = Lazy::new(Request::medium_input);
+        make_redo_work(&REQUEST)
+    });
 }
 criterion::criterion_group!(benches, bench);
 criterion::criterion_main!(benches);
 
 // Returns the sum of each client's wall-clock time spent executing their share of the n_redos.
-fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration {
+fn bench_impl<F, Fut>(redo_work: Arc<F>, n_redos: u64, nclients: u64) -> Duration
+where
+    F: Fn(Arc<PostgresRedoManager>) -> Fut + Send + Sync + 'static,
+    Fut: Future<Output = ()> + Send + 'static,
+{
     let repo_dir = camino_tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
 
     let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
@@ -135,17 +164,20 @@ fn bench_impl(redo_work: Arc<Request>, n_redos: u64, nclients: u64) -> Duration
     })
 }
 
-async fn client(
+async fn client<F, Fut>(
     mgr: Arc<PostgresRedoManager>,
     start: Arc<Barrier>,
-    redo_work: Arc<Request>,
+    redo_work: Arc<F>,
     n_redos: u64,
-) -> Duration {
+) -> Duration
+where
+    F: Fn(Arc<PostgresRedoManager>) -> Fut + Send + Sync + 'static,
+    Fut: Future<Output = ()> + Send + 'static,
+{
     start.wait().await;
     let start = Instant::now();
     for _ in 0..n_redos {
-        let page = redo_work.execute(&mgr).await.unwrap();
-        assert_eq!(page.remaining(), 8192);
+        redo_work(Arc::clone(&mgr)).await;
         // The real pageserver will rarely if ever do 2 walredos in a row without
         // yielding to the executor.
         tokio::task::yield_now().await;

pageserver/client/src/mgmt_api.rs

@@ -432,7 +432,7 @@ impl Client {
             self.mgmt_api_endpoint
         );
 
-        self.request(Method::POST, &uri, req)
+        self.request(Method::PUT, &uri, req)
             .await?
             .json()
             .await

pageserver/compaction/Cargo.toml

@@ -12,7 +12,6 @@ anyhow.workspace = true
 async-stream.workspace = true
 clap = { workspace = true, features = ["string"] }
 futures.workspace = true
-git-version.workspace = true
 itertools.workspace = true
 once_cell.workspace = true
 pageserver_api.workspace = true

pageserver/ctl/Cargo.toml

@@ -10,7 +10,6 @@ license.workspace = true
 anyhow.workspace = true
 camino.workspace = true
 clap = { workspace = true, features = ["string"] }
-git-version.workspace = true
 humantime.workspace = true
 pageserver = { path = ".." }
 pageserver_api.workspace = true

pageserver/src/bin/pageserver.rs

@@ -15,7 +15,7 @@ use clap::{Arg, ArgAction, Command};
 
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
 use pageserver::config::PageserverIdentity;
-use pageserver::control_plane_client::ControlPlaneClient;
+use pageserver::controller_upcall_client::ControllerUpcallClient;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::{COMPUTE_REQUEST_RUNTIME, WALRECEIVER_RUNTIME};
@@ -396,7 +396,7 @@ fn start_pageserver(
     // Set up deletion queue
     let (deletion_queue, deletion_workers) = DeletionQueue::new(
         remote_storage.clone(),
-        ControlPlaneClient::new(conf, &shutdown_pageserver),
+        ControllerUpcallClient::new(conf, &shutdown_pageserver),
         conf,
     );
     if let Some(deletion_workers) = deletion_workers {

pageserver/src/config.rs

@@ -324,7 +324,6 @@ impl PageServerConf {
             max_vectored_read_bytes,
             image_compression,
             ephemeral_bytes_per_memory_kb,
-            compact_level0_phase1_value_access: _,
             l0_flush,
             virtual_file_direct_io,
             concurrent_tenant_warmup,
@@ -535,16 +534,6 @@ mod tests {
             .expect("parse_and_validate");
     }
 
-    #[test]
-    fn test_compactl0_phase1_access_mode_is_ignored_silently() {
-        let input = indoc::indoc! {r#"
-            [compact_level0_phase1_value_access]
-            mode = "streaming-kmerge"
-            validate = "key-lsn-value"
-        "#};
-        toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(input).unwrap();
-    }
-
     /// If there's a typo in the pageserver config, we'd rather catch that typo
     /// and fail pageserver startup than silently ignoring the typo, leaving whoever
     /// made it in the believe that their config change is effective.

pageserver/src/controller_upcall_client.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 
 use futures::Future;
 use pageserver_api::{
-    controller_api::NodeRegisterRequest,
+    controller_api::{AvailabilityZone, NodeRegisterRequest},
     shard::TenantShardId,
     upcall_api::{
         ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest,
@@ -17,9 +17,12 @@ use utils::{backoff, failpoint_support, generation::Generation, id::NodeId};
 use crate::{config::PageServerConf, virtual_file::on_fatal_io_error};
 use pageserver_api::config::NodeMetadata;
 
-/// The Pageserver's client for using the control plane API: this is a small subset
-/// of the overall control plane API, for dealing with generations (see docs/rfcs/025-generation-numbers.md)
-pub struct ControlPlaneClient {
+/// The Pageserver's client for using the storage controller upcall API: this is a small API
+/// for dealing with generations (see docs/rfcs/025-generation-numbers.md).
+///
+/// The server presenting this API may either be the storage controller or some other
+/// service (such as the Neon control plane) providing a store of generation numbers.
+pub struct ControllerUpcallClient {
     http_client: reqwest::Client,
     base_url: Url,
     node_id: NodeId,
@@ -45,7 +48,7 @@ pub trait ControlPlaneGenerationsApi {
     ) -> impl Future<Output = Result<HashMap<TenantShardId, bool>, RetryForeverError>> + Send;
 }
 
-impl ControlPlaneClient {
+impl ControllerUpcallClient {
     /// A None return value indicates that the input `conf` object does not have control
     /// plane API enabled.
     pub fn new(conf: &'static PageServerConf, cancel: &CancellationToken) -> Option<Self> {
@@ -114,7 +117,7 @@ impl ControlPlaneClient {
     }
 }
 
-impl ControlPlaneGenerationsApi for ControlPlaneClient {
+impl ControlPlaneGenerationsApi for ControllerUpcallClient {
     /// Block until we get a successful response, or error out if we are shut down
     async fn re_attach(
         &self,
@@ -148,10 +151,10 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
                             .and_then(|jv| jv.as_str().map(|str| str.to_owned()));
 
                         match az_id_from_metadata {
-                            Some(az_id) => Some(az_id),
+                            Some(az_id) => Some(AvailabilityZone(az_id)),
                             None => {
                                 tracing::warn!("metadata.json does not contain an 'availability_zone_id' field");
-                                conf.availability_zone.clone()
+                                conf.availability_zone.clone().map(AvailabilityZone)
                             }
                         }
                     };
@@ -216,29 +219,38 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
             .join("validate")
             .expect("Failed to build validate path");
 
-        let request = ValidateRequest {
-            tenants: tenants
-                .into_iter()
-                .map(|(id, gen)| ValidateRequestTenant {
-                    id,
-                    gen: gen
-                        .into()
-                        .expect("Generation should always be valid for a Tenant doing deletions"),
-                })
-                .collect(),
-        };
+        // When sending validate requests, break them up into chunks so that we
+        // avoid possible edge cases of generating any HTTP requests that
+        // require database I/O across many thousands of tenants.
+        let mut result: HashMap<TenantShardId, bool> = HashMap::with_capacity(tenants.len());
+        for tenant_chunk in (tenants).chunks(128) {
+            let request = ValidateRequest {
+                tenants: tenant_chunk
+                    .iter()
+                    .map(|(id, generation)| ValidateRequestTenant {
+                        id: *id,
+                        gen: (*generation).into().expect(
+                            "Generation should always be valid for a Tenant doing deletions",
+                        ),
+                    })
+                    .collect(),
+            };
 
-        failpoint_support::sleep_millis_async!("control-plane-client-validate-sleep", &self.cancel);
-        if self.cancel.is_cancelled() {
-            return Err(RetryForeverError::ShuttingDown);
+            failpoint_support::sleep_millis_async!(
+                "control-plane-client-validate-sleep",
+                &self.cancel
+            );
+            if self.cancel.is_cancelled() {
+                return Err(RetryForeverError::ShuttingDown);
+            }
+
+            let response: ValidateResponse =
+                self.retry_http_forever(&re_attach_path, request).await?;
+            for rt in response.tenants {
+                result.insert(rt.id, rt.valid);
+            }
         }
 
-        let response: ValidateResponse = self.retry_http_forever(&re_attach_path, request).await?;
-
-        Ok(response
-            .tenants
-            .into_iter()
-            .map(|rt| (rt.id, rt.valid))
-            .collect())
+        Ok(result.into_iter().collect())
     }
 }

pageserver/src/deletion_queue.rs

@@ -6,7 +6,7 @@ use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
 
-use crate::control_plane_client::ControlPlaneGenerationsApi;
+use crate::controller_upcall_client::ControlPlaneGenerationsApi;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::remote_timeline_client::remote_timeline_path;
@@ -622,7 +622,7 @@ impl DeletionQueue {
     /// If remote_storage is None, then the returned workers will also be None.
     pub fn new<C>(
         remote_storage: GenericRemoteStorage,
-        control_plane_client: Option<C>,
+        controller_upcall_client: Option<C>,
         conf: &'static PageServerConf,
     ) -> (Self, Option<DeletionQueueWorkers<C>>)
     where
@@ -662,7 +662,7 @@ impl DeletionQueue {
                     conf,
                     backend_rx,
                     executor_tx,
-                    control_plane_client,
+                    controller_upcall_client,
                     lsn_table.clone(),
                     cancel.clone(),
                 ),
@@ -704,7 +704,7 @@ mod test {
     use tokio::task::JoinHandle;
 
     use crate::{
-        control_plane_client::RetryForeverError,
+        controller_upcall_client::RetryForeverError,
         repository::Key,
         tenant::{harness::TenantHarness, storage_layer::DeltaLayerName},
     };

pageserver/src/deletion_queue/validator.rs

@@ -25,8 +25,8 @@ use tracing::info;
 use tracing::warn;
 
 use crate::config::PageServerConf;
-use crate::control_plane_client::ControlPlaneGenerationsApi;
-use crate::control_plane_client::RetryForeverError;
+use crate::controller_upcall_client::ControlPlaneGenerationsApi;
+use crate::controller_upcall_client::RetryForeverError;
 use crate::metrics;
 use crate::virtual_file::MaybeFatalIo;
 
@@ -61,7 +61,7 @@ where
     tx: tokio::sync::mpsc::Sender<DeleterMessage>,
 
     // Client for calling into control plane API for validation of deletes
-    control_plane_client: Option<C>,
+    controller_upcall_client: Option<C>,
 
     // DeletionLists which are waiting generation validation.  Not safe to
     // execute until [`validate`] has processed them.
@@ -94,7 +94,7 @@ where
         conf: &'static PageServerConf,
         rx: tokio::sync::mpsc::Receiver<ValidatorQueueMessage>,
         tx: tokio::sync::mpsc::Sender<DeleterMessage>,
-        control_plane_client: Option<C>,
+        controller_upcall_client: Option<C>,
         lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
         cancel: CancellationToken,
     ) -> Self {
@@ -102,7 +102,7 @@ where
             conf,
             rx,
             tx,
-            control_plane_client,
+            controller_upcall_client,
             lsn_table,
             pending_lists: Vec::new(),
             validated_lists: Vec::new(),
@@ -145,8 +145,8 @@ where
             return Ok(());
         }
 
-        let tenants_valid = if let Some(control_plane_client) = &self.control_plane_client {
-            match control_plane_client
+        let tenants_valid = if let Some(controller_upcall_client) = &self.controller_upcall_client {
+            match controller_upcall_client
                 .validate(tenant_generations.iter().map(|(k, v)| (*k, *v)).collect())
                 .await
             {

pageserver/src/http/routes.rs

@@ -589,6 +589,10 @@ async fn timeline_create_handler(
                 StatusCode::SERVICE_UNAVAILABLE,
                 HttpErrorBody::from_msg(e.to_string()),
             ),
+            Err(e @ tenant::CreateTimelineError::AncestorArchived) => json_response(
+                StatusCode::NOT_ACCEPTABLE,
+                HttpErrorBody::from_msg(e.to_string()),
+            ),
             Err(tenant::CreateTimelineError::ShuttingDown) => json_response(
                 StatusCode::SERVICE_UNAVAILABLE,
                 HttpErrorBody::from_msg("tenant shutting down".to_string()),
@@ -2955,7 +2959,7 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/preserve_initdb_archive",
             |r| api_handler(r, timeline_preserve_initdb_handler),
         )
-        .post(
+        .put(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/archival_config",
             |r| api_handler(r, timeline_archival_config_handler),
         )

pageserver/src/lib.rs

@@ -6,7 +6,7 @@ pub mod basebackup;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
-pub mod control_plane_client;
+pub mod controller_upcall_client;
 pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
 pub mod http;

pageserver/src/metrics.rs

@@ -8,6 +8,8 @@ use metrics::{
 };
 use once_cell::sync::Lazy;
 use pageserver_api::shard::TenantShardId;
+use postgres_backend::{is_expected_io_error, QueryError};
+use pq_proto::framed::ConnectionError;
 use strum::{EnumCount, VariantNames};
 use strum_macros::{IntoStaticStr, VariantNames};
 use tracing::warn;
@@ -1383,7 +1385,7 @@ impl SmgrQueryTimePerTimeline {
         &'a self,
         op: SmgrQueryType,
         ctx: &'c RequestContext,
-    ) -> Option<impl Drop + '_> {
+    ) -> Option<impl Drop + 'a> {
         let start = Instant::now();
 
         self.global_started[op as usize].inc();
@@ -1508,6 +1510,7 @@ static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
 pub(crate) struct BasebackupQueryTime {
     ok: Histogram,
     error: Histogram,
+    client_error: Histogram,
 }
 
 pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
@@ -1521,6 +1524,7 @@ pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|
     BasebackupQueryTime {
         ok: vec.get_metric_with_label_values(&["ok"]).unwrap(),
         error: vec.get_metric_with_label_values(&["error"]).unwrap(),
+        client_error: vec.get_metric_with_label_values(&["client_error"]).unwrap(),
     }
 });
 
@@ -1534,7 +1538,7 @@ impl BasebackupQueryTime {
     pub(crate) fn start_recording<'c: 'a, 'a>(
         &'a self,
         ctx: &'c RequestContext,
-    ) -> BasebackupQueryTimeOngoingRecording<'_, '_> {
+    ) -> BasebackupQueryTimeOngoingRecording<'a, 'a> {
         let start = Instant::now();
         match ctx.micros_spent_throttled.open() {
             Ok(()) => (),
@@ -1557,7 +1561,7 @@ impl BasebackupQueryTime {
 }
 
 impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
-    pub(crate) fn observe<T, E>(self, res: &Result<T, E>) {
+    pub(crate) fn observe<T>(self, res: &Result<T, QueryError>) {
         let elapsed = self.start.elapsed();
         let ex_throttled = self
             .ctx
@@ -1576,10 +1580,15 @@ impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
                 elapsed
             }
         };
-        let metric = if res.is_ok() {
-            &self.parent.ok
-        } else {
-            &self.parent.error
+        // If you want to change categorize of a specific error, also change it in `log_query_error`.
+        let metric = match res {
+            Ok(_) => &self.parent.ok,
+            Err(QueryError::Disconnected(ConnectionError::Io(io_error)))
+                if is_expected_io_error(io_error) =>
+            {
+                &self.parent.client_error
+            }
+            Err(_) => &self.parent.error,
         };
         metric.observe(ex_throttled.as_secs_f64());
     }
@@ -3208,45 +3217,38 @@ pub(crate) mod tenant_throttling {
 
     impl TimelineGet {
         pub(crate) fn new(tenant_shard_id: &TenantShardId) -> Self {
+            let per_tenant_label_values = &[
+                KIND,
+                &tenant_shard_id.tenant_id.to_string(),
+                &tenant_shard_id.shard_slug().to_string(),
+            ];
             TimelineGet {
                 count_accounted_start: {
                     GlobalAndPerTenantIntCounter {
                         global: COUNT_ACCOUNTED_START.with_label_values(&[KIND]),
-                        per_tenant: COUNT_ACCOUNTED_START_PER_TENANT.with_label_values(&[
-                            KIND,
-                            &tenant_shard_id.tenant_id.to_string(),
-                            &tenant_shard_id.shard_slug().to_string(),
-                        ]),
+                        per_tenant: COUNT_ACCOUNTED_START_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
                     }
                 },
                 count_accounted_finish: {
                     GlobalAndPerTenantIntCounter {
                         global: COUNT_ACCOUNTED_FINISH.with_label_values(&[KIND]),
-                        per_tenant: COUNT_ACCOUNTED_FINISH_PER_TENANT.with_label_values(&[
-                            KIND,
-                            &tenant_shard_id.tenant_id.to_string(),
-                            &tenant_shard_id.shard_slug().to_string(),
-                        ]),
+                        per_tenant: COUNT_ACCOUNTED_FINISH_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
                     }
                 },
                 wait_time: {
                     GlobalAndPerTenantIntCounter {
                         global: WAIT_USECS.with_label_values(&[KIND]),
-                        per_tenant: WAIT_USECS_PER_TENANT.with_label_values(&[
-                            KIND,
-                            &tenant_shard_id.tenant_id.to_string(),
-                            &tenant_shard_id.shard_slug().to_string(),
-                        ]),
+                        per_tenant: WAIT_USECS_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
                     }
                 },
                 count_throttled: {
                     GlobalAndPerTenantIntCounter {
                         global: WAIT_COUNT.with_label_values(&[KIND]),
-                        per_tenant: WAIT_COUNT_PER_TENANT.with_label_values(&[
-                            KIND,
-                            &tenant_shard_id.tenant_id.to_string(),
-                            &tenant_shard_id.shard_slug().to_string(),
-                        ]),
+                        per_tenant: WAIT_COUNT_PER_TENANT
+                            .with_label_values(per_tenant_label_values),
                     }
                 },
             }

pageserver/src/pgdatadir_mapping.rs

@@ -840,6 +840,36 @@ impl Timeline {
         Ok(total_size * BLCKSZ as u64)
     }
 
+    /// Get a KeySpace that covers all the Keys that are in use at AND below the given LSN. This is only used
+    /// for gc-compaction.
+    ///
+    /// gc-compaction cannot use the same `collect_keyspace` function as the legacy compaction because it
+    /// processes data at multiple LSNs and needs to be aware of the fact that some key ranges might need to
+    /// be kept only for a specific range of LSN.
+    ///
+    /// Consider the case that the user created branches at LSN 10 and 20, where the user created a table A at
+    /// LSN 10 and dropped that table at LSN 20. `collect_keyspace` at LSN 10 will return the key range
+    /// corresponding to that table, while LSN 20 won't. The keyspace info at a single LSN is not enough to
+    /// determine which keys to retain/drop for gc-compaction.
+    ///
+    /// For now, it only drops AUX-v1 keys. But in the future, the function will be extended to return the keyspace
+    /// to be retained for each of the branch LSN.
+    ///
+    /// The return value is (dense keyspace, sparse keyspace).
+    pub(crate) async fn collect_gc_compaction_keyspace(
+        &self,
+    ) -> Result<(KeySpace, SparseKeySpace), CollectKeySpaceError> {
+        let metadata_key_begin = Key::metadata_key_range().start;
+        let aux_v1_key = AUX_FILES_KEY;
+        let dense_keyspace = KeySpace {
+            ranges: vec![Key::MIN..aux_v1_key, aux_v1_key.next()..metadata_key_begin],
+        };
+        Ok((
+            dense_keyspace,
+            SparseKeySpace(KeySpace::single(Key::metadata_key_range())),
+        ))
+    }
+
     ///
     /// Get a KeySpace that covers all the Keys that are in use at the given LSN.
     /// Anything that's not listed maybe removed from the underlying storage (from

pageserver/src/tenant.rs

@@ -18,7 +18,6 @@ use camino::Utf8Path;
 use camino::Utf8PathBuf;
 use enumset::EnumSet;
 use futures::stream::FuturesUnordered;
-use futures::FutureExt;
 use futures::StreamExt;
 use pageserver_api::models;
 use pageserver_api::models::AuxFilePolicy;
@@ -34,6 +33,7 @@ use remote_storage::GenericRemoteStorage;
 use remote_storage::TimeoutOrCancel;
 use std::collections::BTreeMap;
 use std::fmt;
+use std::future::Future;
 use std::sync::Weak;
 use std::time::SystemTime;
 use storage_broker::BrokerClientChannel;
@@ -563,6 +563,8 @@ pub enum CreateTimelineError {
     AncestorLsn(anyhow::Error),
     #[error("ancestor timeline is not active")]
     AncestorNotActive,
+    #[error("ancestor timeline is archived")]
+    AncestorArchived,
     #[error("tenant shutting down")]
     ShuttingDown,
     #[error(transparent)]
@@ -1031,13 +1033,9 @@ impl Tenant {
         }
 
         Ok(TenantPreload {
-            timelines: Self::load_timeline_metadata(
-                self,
-                remote_timeline_ids,
-                remote_storage,
-                cancel,
-            )
-            .await?,
+            timelines: self
+                .load_timelines_metadata(remote_timeline_ids, remote_storage, cancel)
+                .await?,
         })
     }
 
@@ -1303,7 +1301,7 @@ impl Tenant {
         .await
     }
 
-    async fn load_timeline_metadata(
+    async fn load_timelines_metadata(
         self: &Arc<Tenant>,
         timeline_ids: HashSet<TimelineId>,
         remote_storage: &GenericRemoteStorage,
@@ -1311,33 +1309,10 @@ impl Tenant {
     ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
         let mut part_downloads = JoinSet::new();
         for timeline_id in timeline_ids {
-            let client = RemoteTimelineClient::new(
-                remote_storage.clone(),
-                self.deletion_queue_client.clone(),
-                self.conf,
-                self.tenant_shard_id,
-                timeline_id,
-                self.generation,
-            );
             let cancel_clone = cancel.clone();
             part_downloads.spawn(
-                async move {
-                    debug!("starting index part download");
-
-                    let index_part = client.download_index_file(&cancel_clone).await;
-
-                    debug!("finished index part download");
-
-                    Result::<_, anyhow::Error>::Ok(TimelinePreload {
-                        client,
-                        timeline_id,
-                        index_part,
-                    })
-                }
-                .map(move |res| {
-                    res.with_context(|| format!("download index part for timeline {timeline_id}"))
-                })
-                .instrument(info_span!("download_index_part", %timeline_id)),
+                self.load_timeline_metadata(timeline_id, remote_storage.clone(), cancel_clone)
+                    .instrument(info_span!("download_index_part", %timeline_id)),
             );
         }
 
@@ -1348,8 +1323,7 @@ impl Tenant {
                 next = part_downloads.join_next() => {
                     match next {
                         Some(result) => {
-                            let preload_result = result.context("join preload task")?;
-                            let preload = preload_result?;
+                            let preload = result.context("join preload task")?;
                             timeline_preloads.insert(preload.timeline_id, preload);
                         },
                         None => {
@@ -1366,6 +1340,36 @@ impl Tenant {
         Ok(timeline_preloads)
     }
 
+    fn load_timeline_metadata(
+        self: &Arc<Tenant>,
+        timeline_id: TimelineId,
+        remote_storage: GenericRemoteStorage,
+        cancel: CancellationToken,
+    ) -> impl Future<Output = TimelinePreload> {
+        let client = RemoteTimelineClient::new(
+            remote_storage.clone(),
+            self.deletion_queue_client.clone(),
+            self.conf,
+            self.tenant_shard_id,
+            timeline_id,
+            self.generation,
+        );
+        async move {
+            debug_assert_current_span_has_tenant_and_timeline_id();
+            debug!("starting index part download");
+
+            let index_part = client.download_index_file(&cancel).await;
+
+            debug!("finished index part download");
+
+            TimelinePreload {
+                client,
+                timeline_id,
+                index_part,
+            }
+        }
+    }
+
     pub(crate) async fn apply_timeline_archival_config(
         &self,
         timeline_id: TimelineId,
@@ -1696,6 +1700,11 @@ impl Tenant {
                     return Err(CreateTimelineError::AncestorNotActive);
                 }
 
+                if ancestor_timeline.is_archived() == Some(true) {
+                    info!("tried to branch archived timeline");
+                    return Err(CreateTimelineError::AncestorArchived);
+                }
+
                 if let Some(lsn) = ancestor_start_lsn.as_mut() {
                     *lsn = lsn.align();
 
@@ -1966,9 +1975,6 @@ impl Tenant {
                 TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
                     panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
                 }
-                TenantState::Loading => {
-                    *current_state = TenantState::Activating(ActivatingFrom::Loading);
-                }
                 TenantState::Attaching => {
                     *current_state = TenantState::Activating(ActivatingFrom::Attaching);
                 }
@@ -2149,7 +2155,7 @@ impl Tenant {
     async fn set_stopping(
         &self,
         progress: completion::Barrier,
-        allow_transition_from_loading: bool,
+        _allow_transition_from_loading: bool,
         allow_transition_from_attaching: bool,
     ) -> Result<(), SetStoppingError> {
         let mut rx = self.state.subscribe();
@@ -2164,7 +2170,6 @@ impl Tenant {
                 );
                 false
             }
-            TenantState::Loading => allow_transition_from_loading,
             TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
         })
         .await
@@ -2183,13 +2188,6 @@ impl Tenant {
                 *current_state = TenantState::Stopping { progress };
                 true
             }
-            TenantState::Loading => {
-                if !allow_transition_from_loading {
-                    unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
-                };
-                *current_state = TenantState::Stopping { progress };
-                true
-            }
             TenantState::Active => {
                 // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
                 // are created after the transition to Stopping. That's harmless, as the Timelines
@@ -2245,7 +2243,7 @@ impl Tenant {
         // The load & attach routines own the tenant state until it has reached `Active`.
         // So, wait until it's done.
         rx.wait_for(|state| match state {
-            TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
+            TenantState::Activating(_) | TenantState::Attaching => {
                 info!(
                     "waiting for {} to turn Active|Broken|Stopping",
                     <&'static str>::from(state)
@@ -2265,7 +2263,7 @@ impl Tenant {
         let reason = reason.to_string();
         self.state.send_modify(|current_state| {
             match *current_state {
-                TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
+                TenantState::Activating(_) | TenantState::Attaching => {
                     unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
                 }
                 TenantState::Active => {
@@ -2309,7 +2307,7 @@ impl Tenant {
         loop {
             let current_state = receiver.borrow_and_update().clone();
             match current_state {
-                TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
+                TenantState::Attaching | TenantState::Activating(_) => {
                     // in these states, there's a chance that we can reach ::Active
                     self.activate_now();
                     match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
@@ -3625,7 +3623,7 @@ impl Tenant {
         start_lsn: Lsn,
         ancestor: Option<Arc<Timeline>>,
         last_aux_file_policy: Option<AuxFilePolicy>,
-    ) -> anyhow::Result<UninitializedTimeline> {
+    ) -> anyhow::Result<UninitializedTimeline<'a>> {
         let tenant_shard_id = self.tenant_shard_id;
 
         let resources = self.build_timeline_resources(new_timeline_id);
@@ -4142,7 +4140,7 @@ pub(crate) mod harness {
             let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
 
             let tenant = Arc::new(Tenant::new(
-                TenantState::Loading,
+                TenantState::Attaching,
                 self.conf,
                 AttachedTenantConf::try_from(LocationConf::attached_single(
                     TenantConfOpt::from(self.tenant_conf.clone()),

pageserver/src/tenant/checks.rs

@@ -5,6 +5,7 @@ use itertools::Itertools;
 use super::storage_layer::LayerName;
 
 /// Checks whether a layer map is valid (i.e., is a valid result of the current compaction algorithm if nothing goes wrong).
+///
 /// The function checks if we can split the LSN range of a delta layer only at the LSNs of the delta layers. For example,
 ///
 /// ```plain

pageserver/src/tenant/mgr.rs

@@ -30,8 +30,8 @@ use utils::{backoff, completion, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::control_plane_client::{
-    ControlPlaneClient, ControlPlaneGenerationsApi, RetryForeverError,
+use crate::controller_upcall_client::{
+    ControlPlaneGenerationsApi, ControllerUpcallClient, RetryForeverError,
 };
 use crate::deletion_queue::DeletionQueueClient;
 use crate::http::routes::ACTIVE_TENANT_TIMEOUT;
@@ -122,7 +122,7 @@ pub(crate) enum ShardSelector {
     Known(ShardIndex),
 }
 
-/// A convenience for use with the re_attach ControlPlaneClient function: rather
+/// A convenience for use with the re_attach ControllerUpcallClient function: rather
 /// than the serializable struct, we build this enum that encapsulates
 /// the invariant that attached tenants always have generations.
 ///
@@ -341,7 +341,7 @@ async fn init_load_generations(
             "Emergency mode!  Tenants will be attached unsafely using their last known generation"
         );
         emergency_generations(tenant_confs)
-    } else if let Some(client) = ControlPlaneClient::new(conf, cancel) {
+    } else if let Some(client) = ControllerUpcallClient::new(conf, cancel) {
         info!("Calling control plane API to re-attach tenants");
         // If we are configured to use the control plane API, then it is the source of truth for what tenants to load.
         match client.re_attach(conf).await {

pageserver/src/tenant/storage_layer.rs

@@ -1,13 +1,13 @@
 //! Common traits and structs for layers
 
 pub mod delta_layer;
+pub mod filter_iterator;
 pub mod image_layer;
 pub mod inmemory_layer;
 pub(crate) mod layer;
 mod layer_desc;
 mod layer_name;
 pub mod merge_iterator;
-
 pub mod split_writer;
 
 use crate::context::{AccessStatsBehavior, RequestContext};

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -39,7 +39,7 @@ use crate::tenant::disk_btree::{
 use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
-    BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
+    BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
     VectoredReadCoalesceMode, VectoredReadPlanner,
 };
 use crate::tenant::PageReconstructError;
@@ -1021,13 +1021,30 @@ impl DeltaLayerInner {
                     continue;
                 }
             };
-
+            let view = BufView::new_slice(&blobs_buf.buf);
             for meta in blobs_buf.blobs.iter().rev() {
                 if Some(meta.meta.key) == ignore_key_with_err {
                     continue;
                 }
+                let blob_read = meta.read(&view).await;
+                let blob_read = match blob_read {
+                    Ok(buf) => buf,
+                    Err(e) => {
+                        reconstruct_state.on_key_error(
+                            meta.meta.key,
+                            PageReconstructError::Other(anyhow!(e).context(format!(
+                                "Failed to decompress blob from virtual file {}",
+                                self.file.path,
+                            ))),
+                        );
+
+                        ignore_key_with_err = Some(meta.meta.key);
+                        continue;
+                    }
+                };
+
+                let value = Value::des(&blob_read);
 
-                let value = Value::des(&blobs_buf.buf[meta.start..meta.end]);
                 let value = match value {
                     Ok(v) => v,
                     Err(e) => {
@@ -1243,21 +1260,21 @@ impl DeltaLayerInner {
                 buf.reserve(read.size());
                 let res = reader.read_blobs(&read, buf, ctx).await?;
 
+                let view = BufView::new_slice(&res.buf);
+
                 for blob in res.blobs {
                     let key = blob.meta.key;
                     let lsn = blob.meta.lsn;
-                    let data = &res.buf[blob.start..blob.end];
+
+                    let data = blob.read(&view).await?;
 
                     #[cfg(debug_assertions)]
-                    Value::des(data)
+                    Value::des(&data)
                         .with_context(|| {
                             format!(
-                                "blob failed to deserialize for {}@{}, {}..{}: {:?}",
-                                blob.meta.key,
-                                blob.meta.lsn,
-                                blob.start,
-                                blob.end,
-                                utils::Hex(data)
+                                "blob failed to deserialize for {}: {:?}",
+                                blob,
+                                utils::Hex(&data)
                             )
                         })
                         .unwrap();
@@ -1265,15 +1282,15 @@ impl DeltaLayerInner {
                     // is it an image or will_init walrecord?
                     // FIXME: this could be handled by threading the BlobRef to the
                     // VectoredReadBuilder
-                    let will_init = crate::repository::ValueBytes::will_init(data)
+                    let will_init = crate::repository::ValueBytes::will_init(&data)
                         .inspect_err(|_e| {
                             #[cfg(feature = "testing")]
-                            tracing::error!(data=?utils::Hex(data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
+                            tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
                         })
                         .unwrap_or(false);
 
                     per_blob_copy.clear();
-                    per_blob_copy.extend_from_slice(data);
+                    per_blob_copy.extend_from_slice(&data);
 
                     let (tmp, res) = writer
                         .put_value_bytes(
@@ -1538,8 +1555,11 @@ impl<'a> DeltaLayerIterator<'a> {
             .read_blobs(&plan, buf, self.ctx)
             .await?;
         let frozen_buf = blobs_buf.buf.freeze();
+        let view = BufView::new_bytes(frozen_buf);
         for meta in blobs_buf.blobs.iter() {
-            let value = Value::des(&frozen_buf[meta.start..meta.end])?;
+            let blob_read = meta.read(&view).await?;
+            let value = Value::des(&blob_read)?;
+
             next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
         }
         self.key_values_batch = next_batch;
@@ -1916,9 +1936,13 @@ pub(crate) mod test {
                 let blobs_buf = vectored_blob_reader
                     .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
                     .await?;
+                let view = BufView::new_slice(&blobs_buf.buf);
                 for meta in blobs_buf.blobs.iter() {
-                    let value = &blobs_buf.buf[meta.start..meta.end];
-                    assert_eq!(value, entries_meta.index[&(meta.meta.key, meta.meta.lsn)]);
+                    let value = meta.read(&view).await?;
+                    assert_eq!(
+                        &value[..],
+                        &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
+                    );
                 }
 
                 buf = Some(blobs_buf.buf);

pageserver/src/tenant/storage_layer/filter_iterator.rs

@@ -0,0 +1,205 @@
+use std::ops::Range;
+
+use anyhow::bail;
+use pageserver_api::{
+    key::Key,
+    keyspace::{KeySpace, SparseKeySpace},
+};
+use utils::lsn::Lsn;
+
+use crate::repository::Value;
+
+use super::merge_iterator::MergeIterator;
+
+/// A filter iterator over merge iterators (and can be easily extended to other types of iterators).
+///
+/// The iterator will skip any keys not included in the keyspace filter. In other words, the keyspace filter contains the keys
+/// to be retained.
+pub struct FilterIterator<'a> {
+    inner: MergeIterator<'a>,
+    retain_key_filters: Vec<Range<Key>>,
+    current_filter_idx: usize,
+}
+
+impl<'a> FilterIterator<'a> {
+    pub fn create(
+        inner: MergeIterator<'a>,
+        dense_keyspace: KeySpace,
+        sparse_keyspace: SparseKeySpace,
+    ) -> anyhow::Result<Self> {
+        let mut retain_key_filters = Vec::new();
+        retain_key_filters.extend(dense_keyspace.ranges);
+        retain_key_filters.extend(sparse_keyspace.0.ranges);
+        retain_key_filters.sort_by(|a, b| a.start.cmp(&b.start));
+        // Verify key filters are non-overlapping and sorted
+        for window in retain_key_filters.windows(2) {
+            if window[0].end > window[1].start {
+                bail!(
+                    "Key filters are overlapping: {:?} and {:?}",
+                    window[0],
+                    window[1]
+                );
+            }
+        }
+        Ok(Self {
+            inner,
+            retain_key_filters,
+            current_filter_idx: 0,
+        })
+    }
+
+    pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
+        while let Some(item) = self.inner.next().await? {
+            while self.current_filter_idx < self.retain_key_filters.len()
+                && item.0 >= self.retain_key_filters[self.current_filter_idx].end
+            {
+                // [filter region]    [filter region]     [filter region]
+                //                                     ^ item
+                //                    ^ current filter
+                self.current_filter_idx += 1;
+                // [filter region]    [filter region]     [filter region]
+                //                                     ^ item
+                //                                        ^ current filter
+            }
+            if self.current_filter_idx >= self.retain_key_filters.len() {
+                // We already exhausted all filters, so we should return now
+                // [filter region] [filter region] [filter region]
+                //                                                    ^ item
+                //                                                 ^ current filter (nothing)
+                return Ok(None);
+            }
+            if self.retain_key_filters[self.current_filter_idx].contains(&item.0) {
+                // [filter region]    [filter region]     [filter region]
+                //                                              ^ item
+                //                                        ^ current filter
+                return Ok(Some(item));
+            }
+            // If the key is not contained in the key retaining filters, continue to the next item.
+            // [filter region]    [filter region]     [filter region]
+            //                                     ^ item
+            //                                        ^ current filter
+        }
+        Ok(None)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use itertools::Itertools;
+    use pageserver_api::key::Key;
+    use utils::lsn::Lsn;
+
+    use crate::{
+        tenant::{
+            harness::{TenantHarness, TIMELINE_ID},
+            storage_layer::delta_layer::test::produce_delta_layer,
+        },
+        DEFAULT_PG_VERSION,
+    };
+
+    async fn assert_filter_iter_equal(
+        filter_iter: &mut FilterIterator<'_>,
+        expect: &[(Key, Lsn, Value)],
+    ) {
+        let mut expect_iter = expect.iter();
+        loop {
+            let o1 = filter_iter.next().await.unwrap();
+            let o2 = expect_iter.next();
+            assert_eq!(o1.is_some(), o2.is_some());
+            if o1.is_none() && o2.is_none() {
+                break;
+            }
+            let (k1, l1, v1) = o1.unwrap();
+            let (k2, l2, v2) = o2.unwrap();
+            assert_eq!(&k1, k2);
+            assert_eq!(l1, *l2);
+            assert_eq!(&v1, v2);
+        }
+    }
+
+    #[tokio::test]
+    async fn filter_keyspace_iterator() {
+        use crate::repository::Value;
+        use bytes::Bytes;
+
+        let harness = TenantHarness::create("filter_iterator_filter_keyspace_iterator")
+            .await
+            .unwrap();
+        let (tenant, ctx) = harness.load().await;
+
+        let tline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await
+            .unwrap();
+
+        fn get_key(id: u32) -> Key {
+            let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
+            key.field6 = id;
+            key
+        }
+        const N: usize = 100;
+        let test_deltas1 = (0..N)
+            .map(|idx| {
+                (
+                    get_key(idx as u32),
+                    Lsn(0x20 * ((idx as u64) % 10 + 1)),
+                    Value::Image(Bytes::from(format!("img{idx:05}"))),
+                )
+            })
+            .collect_vec();
+        let resident_layer_1 = produce_delta_layer(&tenant, &tline, test_deltas1.clone(), &ctx)
+            .await
+            .unwrap();
+
+        let merge_iter = MergeIterator::create(
+            &[resident_layer_1.get_as_delta(&ctx).await.unwrap()],
+            &[],
+            &ctx,
+        );
+
+        let mut filter_iter = FilterIterator::create(
+            merge_iter,
+            KeySpace {
+                ranges: vec![
+                    get_key(5)..get_key(10),
+                    get_key(20)..get_key(30),
+                    get_key(90)..get_key(110),
+                    get_key(1000)..get_key(2000),
+                ],
+            },
+            SparseKeySpace(KeySpace::default()),
+        )
+        .unwrap();
+        let mut result = Vec::new();
+        result.extend(test_deltas1[5..10].iter().cloned());
+        result.extend(test_deltas1[20..30].iter().cloned());
+        result.extend(test_deltas1[90..100].iter().cloned());
+        assert_filter_iter_equal(&mut filter_iter, &result).await;
+
+        let merge_iter = MergeIterator::create(
+            &[resident_layer_1.get_as_delta(&ctx).await.unwrap()],
+            &[],
+            &ctx,
+        );
+
+        let mut filter_iter = FilterIterator::create(
+            merge_iter,
+            KeySpace {
+                ranges: vec![
+                    get_key(0)..get_key(10),
+                    get_key(20)..get_key(30),
+                    get_key(90)..get_key(95),
+                ],
+            },
+            SparseKeySpace(KeySpace::default()),
+        )
+        .unwrap();
+        let mut result = Vec::new();
+        result.extend(test_deltas1[0..10].iter().cloned());
+        result.extend(test_deltas1[20..30].iter().cloned());
+        result.extend(test_deltas1[90..95].iter().cloned());
+        assert_filter_iter_equal(&mut filter_iter, &result).await;
+    }
+}

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -36,7 +36,8 @@ use crate::tenant::disk_btree::{
 };
 use crate::tenant::timeline::GetVectoredError;
 use crate::tenant::vectored_blob_io::{
-    BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead, VectoredReadPlanner,
+    BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
+    VectoredReadPlanner,
 };
 use crate::tenant::PageReconstructError;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
@@ -547,15 +548,15 @@ impl ImageLayerInner {
 
             let buf = BytesMut::with_capacity(buf_size);
             let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
-
             let frozen_buf = blobs_buf.buf.freeze();
+            let view = BufView::new_bytes(frozen_buf);
 
             for meta in blobs_buf.blobs.iter() {
-                let img_buf = frozen_buf.slice(meta.start..meta.end);
+                let img_buf = meta.read(&view).await?;
 
                 key_count += 1;
                 writer
-                    .put_image(meta.meta.key, img_buf, ctx)
+                    .put_image(meta.meta.key, img_buf.into_bytes(), ctx)
                     .await
                     .context(format!("Storing key {}", meta.meta.key))?;
             }
@@ -602,13 +603,28 @@ impl ImageLayerInner {
             match res {
                 Ok(blobs_buf) => {
                     let frozen_buf = blobs_buf.buf.freeze();
-
+                    let view = BufView::new_bytes(frozen_buf);
                     for meta in blobs_buf.blobs.iter() {
-                        let img_buf = frozen_buf.slice(meta.start..meta.end);
+                        let img_buf = meta.read(&view).await;
+
+                        let img_buf = match img_buf {
+                            Ok(img_buf) => img_buf,
+                            Err(e) => {
+                                reconstruct_state.on_key_error(
+                                    meta.meta.key,
+                                    PageReconstructError::Other(anyhow!(e).context(format!(
+                                        "Failed to decompress blob from virtual file {}",
+                                        self.file.path,
+                                    ))),
+                                );
+
+                                continue;
+                            }
+                        };
                         reconstruct_state.update_key(
                             &meta.meta.key,
                             self.lsn,
-                            Value::Image(img_buf),
+                            Value::Image(img_buf.into_bytes()),
                         );
                     }
                 }
@@ -1025,10 +1041,15 @@ impl<'a> ImageLayerIterator<'a> {
         let blobs_buf = vectored_blob_reader
             .read_blobs(&plan, buf, self.ctx)
             .await?;
-        let frozen_buf: Bytes = blobs_buf.buf.freeze();
+        let frozen_buf = blobs_buf.buf.freeze();
+        let view = BufView::new_bytes(frozen_buf);
         for meta in blobs_buf.blobs.iter() {
-            let img_buf = frozen_buf.slice(meta.start..meta.end);
-            next_batch.push_back((meta.meta.key, self.image_layer.lsn, Value::Image(img_buf)));
+            let img_buf = meta.read(&view).await?;
+            next_batch.push_back((
+                meta.meta.key,
+                self.image_layer.lsn,
+                Value::Image(img_buf.into_bytes()),
+            ));
         }
         self.key_values_batch = next_batch;
         Ok(())

pageserver/src/tenant/tasks.rs

@@ -481,8 +481,7 @@ async fn ingest_housekeeping_loop(tenant: Arc<Tenant>, cancel: CancellationToken
                 let allowed_rps = tenant.timeline_get_throttle.steady_rps();
                 let delta = now - prev;
                 info!(
-                    n_seconds=%format_args!("{:.3}",
-                    delta.as_secs_f64()),
+                    n_seconds=%format_args!("{:.3}", delta.as_secs_f64()),
                     count_accounted = count_accounted_finish,  // don't break existing log scraping
                     count_throttled,
                     sum_throttled_usecs,

pageserver/src/tenant/timeline.rs

@@ -112,7 +112,7 @@ use pageserver_api::reltag::RelTag;
 use pageserver_api::shard::ShardIndex;
 
 use postgres_connection::PgConnectionConfig;
-use postgres_ffi::to_pg_timestamp;
+use postgres_ffi::{to_pg_timestamp, v14::xlog_utils, WAL_SEGMENT_SIZE};
 use utils::{
     completion,
     generation::Generation,
@@ -1337,6 +1337,10 @@ impl Timeline {
         _ctx: &RequestContext,
     ) -> anyhow::Result<LsnLease> {
         let lease = {
+            // Normalize the requested LSN to be aligned, and move to the first record
+            // if it points to the beginning of the page (header).
+            let lsn = xlog_utils::normalize_lsn(lsn, WAL_SEGMENT_SIZE);
+
             let mut gc_info = self.gc_info.write().unwrap();
 
             let valid_until = SystemTime::now() + length;
@@ -3597,7 +3601,7 @@ impl Timeline {
                     ctx,
                 )
                 .await
-                .map_err(|e| FlushLayerError::from_anyhow(self, e))?;
+                .map_err(|e| FlushLayerError::from_anyhow(self, e.into()))?;
 
             if self.cancel.is_cancelled() {
                 return Err(FlushLayerError::Cancelled);
@@ -3836,16 +3840,20 @@ impl Timeline {
         partition_size: u64,
         flags: EnumSet<CompactFlags>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<((KeyPartitioning, SparseKeyPartitioning), Lsn)> {
+    ) -> Result<((KeyPartitioning, SparseKeyPartitioning), Lsn), CompactionError> {
         let Ok(mut partitioning_guard) = self.partitioning.try_lock() else {
             // NB: there are two callers, one is the compaction task, of which there is only one per struct Tenant and hence Timeline.
             // The other is the initdb optimization in flush_frozen_layer, used by `boostrap_timeline`, which runs before `.activate()`
             // and hence before the compaction task starts.
-            anyhow::bail!("repartition() called concurrently, this should not happen");
+            return Err(CompactionError::Other(anyhow!(
+                "repartition() called concurrently, this should not happen"
+            )));
         };
         let ((dense_partition, sparse_partition), partition_lsn) = &*partitioning_guard;
         if lsn < *partition_lsn {
-            anyhow::bail!("repartition() called with LSN going backwards, this should not happen");
+            return Err(CompactionError::Other(anyhow!(
+                "repartition() called with LSN going backwards, this should not happen"
+            )));
         }
 
         let distance = lsn.0 - partition_lsn.0;
@@ -4447,6 +4455,12 @@ pub(crate) enum CompactionError {
     Other(anyhow::Error),
 }
 
+impl CompactionError {
+    pub fn is_cancelled(&self) -> bool {
+        matches!(self, CompactionError::ShuttingDown)
+    }
+}
+
 impl From<CollectKeySpaceError> for CompactionError {
     fn from(err: CollectKeySpaceError) -> Self {
         match err {

pageserver/src/tenant/timeline/compaction.rs

@@ -31,6 +31,7 @@ use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder}
 use crate::page_cache;
 use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::remote_timeline_client::WaitCompletionError;
+use crate::tenant::storage_layer::filter_iterator::FilterIterator;
 use crate::tenant::storage_layer::merge_iterator::MergeIterator;
 use crate::tenant::storage_layer::split_writer::{
     SplitDeltaLayerWriter, SplitImageLayerWriter, SplitWriterResult,
@@ -389,7 +390,7 @@ impl Timeline {
                 // error but continue.
                 //
                 // Suppress error when it's due to cancellation
-                if !self.cancel.is_cancelled() {
+                if !self.cancel.is_cancelled() && !err.is_cancelled() {
                     tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
                 }
                 (1, false)
@@ -1772,6 +1773,7 @@ impl Timeline {
             gc_cutoff,
             lowest_retain_lsn
         );
+
         // Step 1: (In the future) construct a k-merge iterator over all layers. For now, simply collect all keys + LSNs.
         // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
         let mut lsn_split_point = BTreeSet::new(); // TODO: use a better data structure (range tree / range set?)
@@ -1820,7 +1822,12 @@ impl Timeline {
                 image_layers.push(layer);
             }
         }
-        let mut merge_iter = MergeIterator::create(&delta_layers, &image_layers, ctx);
+        let (dense_ks, sparse_ks) = self.collect_gc_compaction_keyspace().await?;
+        let mut merge_iter = FilterIterator::create(
+            MergeIterator::create(&delta_layers, &image_layers, ctx),
+            dense_ks,
+            sparse_ks,
+        )?;
         // Step 2: Produce images+deltas. TODO: ensure newly-produced delta does not overlap with other deltas.
         // Data of the same key.
         let mut accumulated_values = Vec::new();

pageserver/src/tenant/timeline/eviction_task.rs

@@ -30,8 +30,8 @@ use crate::{
     pgdatadir_mapping::CollectKeySpaceError,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     tenant::{
-        storage_layer::LayerVisibilityHint, tasks::BackgroundLoopKind, timeline::EvictionError,
-        LogicalSizeCalculationCause, Tenant,
+        size::CalculateSyntheticSizeError, storage_layer::LayerVisibilityHint,
+        tasks::BackgroundLoopKind, timeline::EvictionError, LogicalSizeCalculationCause, Tenant,
     },
 };
 
@@ -557,6 +557,8 @@ impl Timeline {
             gather_result = gather => {
                 match gather_result {
                     Ok(_) => {},
+                    // It can happen sometimes that we hit this instead of the cancellation token firing above
+                    Err(CalculateSyntheticSizeError::Cancelled) => {}
                     Err(e) => {
                         // We don't care about the result, but, if it failed, we should log it,
                         // since consumption metric might be hitting the cached value and

pageserver/src/tenant/vectored_blob_io.rs

@@ -16,8 +16,9 @@
 //! Note that the vectored blob api does *not* go through the page cache.
 
 use std::collections::BTreeMap;
+use std::ops::Deref;
 
-use bytes::BytesMut;
+use bytes::{Bytes, BytesMut};
 use pageserver_api::key::Key;
 use tokio::io::AsyncWriteExt;
 use tokio_epoll_uring::BoundedBuf;
@@ -35,11 +36,123 @@ pub struct BlobMeta {
     pub lsn: Lsn,
 }
 
-/// Blob offsets into [`VectoredBlobsBuf::buf`]
+/// A view into the vectored blobs read buffer.
+#[derive(Clone, Debug)]
+pub(crate) enum BufView<'a> {
+    Slice(&'a [u8]),
+    Bytes(bytes::Bytes),
+}
+
+impl<'a> BufView<'a> {
+    /// Creates a new slice-based view on the blob.
+    pub fn new_slice(slice: &'a [u8]) -> Self {
+        Self::Slice(slice)
+    }
+
+    /// Creates a new [`bytes::Bytes`]-based view on the blob.
+    pub fn new_bytes(bytes: bytes::Bytes) -> Self {
+        Self::Bytes(bytes)
+    }
+
+    /// Convert the view into `Bytes`.
+    ///
+    /// If using slice as the underlying storage, the copy will be an O(n) operation.
+    pub fn into_bytes(self) -> Bytes {
+        match self {
+            BufView::Slice(slice) => Bytes::copy_from_slice(slice),
+            BufView::Bytes(bytes) => bytes,
+        }
+    }
+
+    /// Creates a sub-view of the blob based on the range.
+    fn view(&self, range: std::ops::Range<usize>) -> Self {
+        match self {
+            BufView::Slice(slice) => BufView::Slice(&slice[range]),
+            BufView::Bytes(bytes) => BufView::Bytes(bytes.slice(range)),
+        }
+    }
+}
+
+impl<'a> Deref for BufView<'a> {
+    type Target = [u8];
+
+    fn deref(&self) -> &Self::Target {
+        match self {
+            BufView::Slice(slice) => slice,
+            BufView::Bytes(bytes) => bytes,
+        }
+    }
+}
+
+impl<'a> AsRef<[u8]> for BufView<'a> {
+    fn as_ref(&self) -> &[u8] {
+        match self {
+            BufView::Slice(slice) => slice,
+            BufView::Bytes(bytes) => bytes.as_ref(),
+        }
+    }
+}
+
+impl<'a> From<&'a [u8]> for BufView<'a> {
+    fn from(value: &'a [u8]) -> Self {
+        Self::new_slice(value)
+    }
+}
+
+impl From<Bytes> for BufView<'_> {
+    fn from(value: Bytes) -> Self {
+        Self::new_bytes(value)
+    }
+}
+
+/// Blob offsets into [`VectoredBlobsBuf::buf`]. The byte ranges is potentially compressed,
+/// subject to [`VectoredBlob::compression_bits`].
 pub struct VectoredBlob {
-    pub start: usize,
-    pub end: usize,
+    /// Blob metadata.
     pub meta: BlobMeta,
+    /// Start offset.
+    start: usize,
+    /// End offset.
+    end: usize,
+    /// Compression used on the the blob.
+    compression_bits: u8,
+}
+
+impl VectoredBlob {
+    /// Reads a decompressed view of the blob.
+    pub(crate) async fn read<'a>(&self, buf: &BufView<'a>) -> Result<BufView<'a>, std::io::Error> {
+        let view = buf.view(self.start..self.end);
+
+        match self.compression_bits {
+            BYTE_UNCOMPRESSED => Ok(view),
+            BYTE_ZSTD => {
+                let mut decompressed_vec = Vec::new();
+                let mut decoder =
+                    async_compression::tokio::write::ZstdDecoder::new(&mut decompressed_vec);
+                decoder.write_all(&view).await?;
+                decoder.flush().await?;
+                // Zero-copy conversion from `Vec` to `Bytes`
+                Ok(BufView::new_bytes(Bytes::from(decompressed_vec)))
+            }
+            bits => {
+                let error = std::io::Error::new(
+                    std::io::ErrorKind::InvalidData,
+                    format!("Failed to decompress blob for {}@{}, {}..{}: invalid compression byte {bits:x}", self.meta.key, self.meta.lsn, self.start, self.end),
+                );
+                Err(error)
+            }
+        }
+    }
+}
+
+impl std::fmt::Display for VectoredBlob {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "{}@{}, {}..{}",
+            self.meta.key, self.meta.lsn, self.start, self.end
+        )
+    }
 }
 
 /// Return type of [`VectoredBlobReader::read_blobs`]
@@ -514,7 +627,7 @@ impl<'a> VectoredBlobReader<'a> {
             );
         }
 
-        let mut buf = self
+        let buf = self
             .file
             .read_exact_at(buf.slice(0..read.size()), read.start, ctx)
             .await?
@@ -529,9 +642,6 @@ impl<'a> VectoredBlobReader<'a> {
         // of a blob is implicit: the start of the next blob if one exists
         // or the end of the read.
 
-        // Some scratch space, put here for reusing the allocation
-        let mut decompressed_vec = Vec::new();
-
         for (blob_start, meta) in blobs_at {
             let blob_start_in_buf = blob_start - start_offset;
             let first_len_byte = buf[blob_start_in_buf as usize];
@@ -557,35 +667,14 @@ impl<'a> VectoredBlobReader<'a> {
                 )
             };
 
-            let start_raw = blob_start_in_buf + size_length;
-            let end_raw = start_raw + blob_size;
-            let (start, end);
-            if compression_bits == BYTE_UNCOMPRESSED {
-                start = start_raw as usize;
-                end = end_raw as usize;
-            } else if compression_bits == BYTE_ZSTD {
-                let mut decoder =
-                    async_compression::tokio::write::ZstdDecoder::new(&mut decompressed_vec);
-                decoder
-                    .write_all(&buf[start_raw as usize..end_raw as usize])
-                    .await?;
-                decoder.flush().await?;
-                start = buf.len();
-                buf.extend_from_slice(&decompressed_vec);
-                end = buf.len();
-                decompressed_vec.clear();
-            } else {
-                let error = std::io::Error::new(
-                    std::io::ErrorKind::InvalidData,
-                    format!("invalid compression byte {compression_bits:x}"),
-                );
-                return Err(error);
-            }
+            let start = (blob_start_in_buf + size_length) as usize;
+            let end = start + blob_size as usize;
 
             metas.push(VectoredBlob {
                 start,
                 end,
                 meta: *meta,
+                compression_bits,
             });
         }
 
@@ -1020,8 +1109,13 @@ mod tests {
             let result = vectored_blob_reader.read_blobs(&read, buf, &ctx).await?;
             assert_eq!(result.blobs.len(), 1);
             let read_blob = &result.blobs[0];
-            let read_buf = &result.buf[read_blob.start..read_blob.end];
-            assert_eq!(blob, read_buf, "mismatch for idx={idx} at offset={offset}");
+            let view = BufView::new_slice(&result.buf);
+            let read_buf = read_blob.read(&view).await?;
+            assert_eq!(
+                &blob[..],
+                &read_buf[..],
+                "mismatch for idx={idx} at offset={offset}"
+            );
             buf = result.buf;
         }
         Ok(())

pageserver/src/walredo.rs

@@ -205,6 +205,22 @@ impl PostgresRedoManager {
         }
     }
 
+    /// Do a ping request-response roundtrip.
+    ///
+    /// Not used in production, but by Rust benchmarks.
+    ///
+    /// # Cancel-Safety
+    ///
+    /// This method is cancellation-safe.
+    pub async fn ping(&self, pg_version: u32) -> Result<(), Error> {
+        self.do_with_walredo_process(pg_version, |proc| async move {
+            proc.ping(Duration::from_secs(1))
+                .await
+                .map_err(Error::Other)
+        })
+        .await
+    }
+
     pub fn status(&self) -> WalRedoManagerStatus {
         WalRedoManagerStatus {
             last_redo_at: {
@@ -297,6 +313,9 @@ impl PostgresRedoManager {
         }
     }
 
+    /// # Cancel-Safety
+    ///
+    /// This method is cancel-safe iff `closure` is cancel-safe.
     async fn do_with_walredo_process<
         F: FnOnce(Arc<Process>) -> Fut,
         Fut: Future<Output = Result<O, Error>>,
@@ -537,6 +556,17 @@ mod tests {
     use tracing::Instrument;
     use utils::{id::TenantId, lsn::Lsn};
 
+    #[tokio::test]
+    async fn test_ping() {
+        let h = RedoHarness::new().unwrap();
+
+        h.manager
+            .ping(14)
+            .instrument(h.span())
+            .await
+            .expect("ping should work");
+    }
+
     #[tokio::test]
     async fn short_v14_redo() {
         let expected = std::fs::read("test_data/short_v14_redo.page").unwrap();

pageserver/src/walredo/process.rs

@@ -6,6 +6,7 @@ use self::no_leak_child::NoLeakChild;
 use crate::{
     config::PageServerConf,
     metrics::{WalRedoKillCause, WAL_REDO_PROCESS_COUNTERS, WAL_REDO_RECORD_COUNTER},
+    page_cache::PAGE_SZ,
     span::debug_assert_current_span_has_tenant_id,
     walrecord::NeonWalRecord,
 };
@@ -237,6 +238,26 @@ impl WalRedoProcess {
         res
     }
 
+    /// Do a ping request-response roundtrip.
+    ///
+    /// Not used in production, but by Rust benchmarks.
+    pub(crate) async fn ping(&self, timeout: Duration) -> anyhow::Result<()> {
+        let mut writebuf: Vec<u8> = Vec::with_capacity(4);
+        protocol::build_ping_msg(&mut writebuf);
+        let Ok(res) = tokio::time::timeout(timeout, self.apply_wal_records0(&writebuf)).await
+        else {
+            anyhow::bail!("WAL redo ping timed out");
+        };
+        let response = res?;
+        if response.len() != PAGE_SZ {
+            anyhow::bail!(
+                "WAL redo ping response should respond with page-sized response: {}",
+                response.len()
+            );
+        }
+        Ok(())
+    }
+
     /// # Cancel-Safety
     ///
     /// When not polled to completion (e.g. because in `tokio::select!` another

pageserver/src/walredo/process/protocol.rs

@@ -55,3 +55,8 @@ pub(crate) fn build_get_page_msg(tag: BufferTag, buf: &mut Vec<u8>) {
     tag.ser_into(buf)
         .expect("serialize BufferTag should always succeed");
 }
+
+pub(crate) fn build_ping_msg(buf: &mut Vec<u8>) {
+    buf.put_u8(b'H');
+    buf.put_u32(4);
+}

pgxn/neon/Makefile

@@ -9,6 +9,8 @@ OBJS = \
 	hll.o \
 	libpagestore.o \
 	neon.o \
+	neon_pgversioncompat.o \
+	neon_perf_counters.o \
 	neon_utils.o \
 	neon_walreader.o \
 	pagestore_smgr.o \
@@ -23,7 +25,18 @@ SHLIB_LINK_INTERNAL = $(libpq)
 SHLIB_LINK = -lcurl
 
 EXTENSION = neon
-DATA = neon--1.0.sql neon--1.0--1.1.sql neon--1.1--1.2.sql neon--1.2--1.3.sql neon--1.3--1.2.sql neon--1.2--1.1.sql neon--1.1--1.0.sql  neon--1.3--1.4.sql neon--1.4--1.3.sql
+DATA = \
+	neon--1.0.sql \
+	neon--1.0--1.1.sql \
+	neon--1.1--1.2.sql \
+	neon--1.2--1.3.sql \
+	neon--1.3--1.4.sql \
+	neon--1.4--1.5.sql \
+	neon--1.5--1.4.sql \
+	neon--1.4--1.3.sql \
+	neon--1.3--1.2.sql \
+	neon--1.2--1.1.sql \
+	neon--1.1--1.0.sql
 PGFILEDESC = "neon - cloud storage for PostgreSQL"
 
 EXTRA_CLEAN = \

pgxn/neon/file_cache.c

@@ -109,6 +109,7 @@ typedef struct FileCacheControl
 								 * reenabling */
 	uint32		size;			/* size of cache file in chunks */
 	uint32		used;			/* number of used chunks */
+	uint32		used_pages;		/* number of used pages */
 	uint32		limit;			/* shared copy of lfc_size_limit */
 	uint64		hits;
 	uint64		misses;
@@ -905,6 +906,10 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 				/* Cache overflow: evict least recently used chunk */
 				FileCacheEntry *victim = dlist_container(FileCacheEntry, list_node, dlist_pop_head_node(&lfc_ctl->lru));
 	
+				for (int i = 0; i < BLOCKS_PER_CHUNK; i++)
+				{
+					lfc_ctl->used_pages -= (victim->bitmap[i >> 5] >> (i & 31)) & 1;
+				}
 				CriticalAssert(victim->access_count == 0);
 				entry->offset = victim->offset; /* grab victim's chunk */
 				hash_search_with_hash_value(lfc_hash, &victim->key, victim->hash, HASH_REMOVE, NULL);
@@ -959,6 +964,7 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 				for (int i = 0; i < blocks_in_chunk; i++)
 				{
+					lfc_ctl->used_pages += 1 - ((entry->bitmap[(chunk_offs + i) >> 5] >> ((chunk_offs + i) & 31)) & 1);
 					entry->bitmap[(chunk_offs + i) >> 5] |=
 						(1 << ((chunk_offs + i) & 31));
 				}
@@ -1051,6 +1057,11 @@ neon_get_lfc_stats(PG_FUNCTION_ARGS)
 			if (lfc_ctl)
 				value = lfc_ctl->size;
 			break;
+		case 5:
+			key = "file_cache_used_pages";
+			if (lfc_ctl)
+				value = lfc_ctl->used_pages;
+			break;
 		default:
 			SRF_RETURN_DONE(funcctx);
 	}

pgxn/neon/libpagestore.c

@@ -30,6 +30,7 @@
 #include "utils/guc.h"
 
 #include "neon.h"
+#include "neon_perf_counters.h"
 #include "neon_utils.h"
 #include "pagestore_client.h"
 #include "walproposer.h"
@@ -331,6 +332,7 @@ CLEANUP_AND_DISCONNECT(PageServer *shard)
 	}
 	if (shard->conn)
 	{
+		MyNeonCounters->pageserver_disconnects_total++;
 		PQfinish(shard->conn);
 		shard->conn = NULL;
 	}
@@ -737,6 +739,8 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 	PageServer *shard = &page_servers[shard_no];
 	PGconn	   *pageserver_conn;
 
+	MyNeonCounters->pageserver_requests_sent_total++;
+
 	/* If the connection was lost for some reason, reconnect */
 	if (shard->state == PS_Connected && PQstatus(shard->conn) == CONNECTION_BAD)
 	{
@@ -889,6 +893,7 @@ pageserver_flush(shardno_t shard_no)
 	}
 	else
 	{
+		MyNeonCounters->pageserver_send_flushes_total++;
 		if (PQflush(pageserver_conn))
 		{
 			char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
@@ -922,7 +927,7 @@ check_neon_id(char **newval, void **extra, GucSource source)
 static Size
 PagestoreShmemSize(void)
 {
-	return sizeof(PagestoreShmemState);
+	return add_size(sizeof(PagestoreShmemState), NeonPerfCountersShmemSize());
 }
 
 static bool
@@ -941,6 +946,9 @@ PagestoreShmemInit(void)
 		memset(&pagestore_shared->shard_map, 0, sizeof(ShardMap));
 		AssignPageserverConnstring(page_server_connstring, NULL);
 	}
+
+	NeonPerfCountersShmemInit();
+
 	LWLockRelease(AddinShmemInitLock);
 	return found;
 }

pgxn/neon/neon--1.4--1.5.sql

@@ -0,0 +1,39 @@
+\echo Use "ALTER EXTENSION neon UPDATE TO '1.5'" to load this file. \quit
+
+
+CREATE FUNCTION get_backend_perf_counters()
+RETURNS SETOF RECORD
+AS 'MODULE_PATHNAME', 'neon_get_backend_perf_counters'
+LANGUAGE C PARALLEL SAFE;
+
+CREATE FUNCTION get_perf_counters()
+RETURNS SETOF RECORD
+AS 'MODULE_PATHNAME', 'neon_get_perf_counters'
+LANGUAGE C PARALLEL SAFE;
+
+-- Show various metrics, for each backend. Note that the values are not reset
+-- when a backend exits. When a new backend starts with the backend ID, it will
+-- continue accumulating the values from where the old backend left. If you are
+-- only interested in the changes from your own session, store the values at the
+-- beginning of the session somewhere, and subtract them on subsequent calls.
+--
+-- For histograms, 'bucket_le' is the upper bound of the histogram bucket.
+CREATE VIEW neon_backend_perf_counters AS
+  SELECT P.procno, P.pid, P.metric, P.bucket_le, P.value
+  FROM get_backend_perf_counters() AS P (
+    procno integer,
+    pid integer,
+    metric text,
+    bucket_le float8,
+    value float8
+  );
+
+-- Summary across all backends. (This could also be implemented with
+-- an aggregate query over neon_backend_perf_counters view.)
+CREATE VIEW neon_perf_counters AS
+  SELECT P.metric, P.bucket_le, P.value
+  FROM get_perf_counters() AS P (
+    metric text,
+    bucket_le float8,
+    value float8
+  );

pgxn/neon/neon--1.5--1.4.sql

@@ -0,0 +1,4 @@
+DROP VIEW IF EXISTS neon_perf_counters;
+DROP VIEW IF EXISTS neon_backend_perf_counters;
+DROP FUNCTION IF EXISTS get_perf_counters();
+DROP FUNCTION IF EXISTS get_backend_perf_counters();

pgxn/neon/neon.control

@@ -1,5 +1,7 @@
 # neon extension
 comment = 'cloud storage for PostgreSQL'
+# TODO: bump default version to 1.5, after we are certain that we don't
+# need to rollback the compute image
 default_version = '1.4'
 module_pathname = '$libdir/neon'
 relocatable = true

pgxn/neon/neon_perf_counters.c

@@ -0,0 +1,261 @@
+/*-------------------------------------------------------------------------
+ *
+ * neon_perf_counters.c
+ *	  Collect statistics about Neon I/O
+ *
+ * Each backend has its own set of counters in shared memory.
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include <math.h>
+
+#include "funcapi.h"
+#include "miscadmin.h"
+#include "storage/proc.h"
+#include "storage/shmem.h"
+#include "utils/builtins.h"
+
+#include "neon_perf_counters.h"
+#include "neon_pgversioncompat.h"
+
+neon_per_backend_counters *neon_per_backend_counters_shared;
+
+Size
+NeonPerfCountersShmemSize(void)
+{
+	Size		size = 0;
+
+	size = add_size(size, mul_size(MaxBackends, sizeof(neon_per_backend_counters)));
+
+	return size;
+}
+
+void
+NeonPerfCountersShmemInit(void)
+{
+	bool		found;
+
+	neon_per_backend_counters_shared =
+		ShmemInitStruct("Neon perf counters",
+						mul_size(MaxBackends,
+								 sizeof(neon_per_backend_counters)),
+						&found);
+	Assert(found == IsUnderPostmaster);
+	if (!found)
+	{
+		/* shared memory is initialized to zeros, so nothing to do here */
+	}
+}
+
+/*
+ * Count a GetPage wait operation.
+ */
+void
+inc_getpage_wait(uint64 latency_us)
+{
+	int			lo = 0;
+	int			hi = NUM_GETPAGE_WAIT_BUCKETS - 1;
+
+	/* Find the right bucket with binary search */
+	while (lo < hi)
+	{
+		int			mid = (lo + hi) / 2;
+
+		if (latency_us < getpage_wait_bucket_thresholds[mid])
+			hi = mid;
+		else
+			lo = mid + 1;
+	}
+	MyNeonCounters->getpage_wait_us_bucket[lo]++;
+	MyNeonCounters->getpage_wait_us_sum += latency_us;
+	MyNeonCounters->getpage_wait_us_count++;
+}
+
+/*
+ * Support functions for the views, neon_backend_perf_counters and
+ * neon_perf_counters.
+ */
+
+typedef struct
+{
+	char	   *name;
+	bool		is_bucket;
+	double		bucket_le;
+	double		value;
+} metric_t;
+
+static metric_t *
+neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
+{
+#define NUM_METRICS (2 + NUM_GETPAGE_WAIT_BUCKETS + 8)
+	metric_t   *metrics = palloc((NUM_METRICS + 1) * sizeof(metric_t));
+	uint64		bucket_accum;
+	int			i = 0;
+	Datum		getpage_wait_str;
+
+	metrics[i].name = "getpage_wait_seconds_count";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_wait_us_count;
+	i++;
+	metrics[i].name = "getpage_wait_seconds_sum";
+	metrics[i].is_bucket = false;
+	metrics[i].value = ((double) counters->getpage_wait_us_sum) / 1000000.0;
+	i++;
+
+	bucket_accum = 0;
+	for (int bucketno = 0; bucketno < NUM_GETPAGE_WAIT_BUCKETS; bucketno++)
+	{
+		uint64		threshold = getpage_wait_bucket_thresholds[bucketno];
+
+		bucket_accum += counters->getpage_wait_us_bucket[bucketno];
+
+		metrics[i].name = "getpage_wait_seconds_bucket";
+		metrics[i].is_bucket = true;
+		metrics[i].bucket_le = (threshold == UINT64_MAX) ? INFINITY : ((double) threshold) / 1000000.0;
+		metrics[i].value = (double) bucket_accum;
+		i++;
+	}
+	metrics[i].name = "getpage_prefetch_requests_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_requests_total;
+	i++;
+	metrics[i].name = "getpage_sync_requests_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_sync_requests_total;
+	i++;
+	metrics[i].name = "getpage_prefetch_misses_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_misses_total;
+	i++;
+	metrics[i].name = "getpage_prefetch_discards_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->getpage_prefetch_discards_total;
+	i++;
+	metrics[i].name = "pageserver_requests_sent_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_requests_sent_total;
+	i++;
+	metrics[i].name = "pageserver_requests_disconnects_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_disconnects_total;
+	i++;
+	metrics[i].name = "pageserver_send_flushes_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->pageserver_send_flushes_total;
+	i++;
+	metrics[i].name = "file_cache_hits_total";
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) counters->file_cache_hits_total;
+	i++;
+
+	Assert(i == NUM_METRICS);
+
+	/* NULL entry marks end of array */
+	metrics[i].name = NULL;
+	metrics[i].value = 0;
+
+	return metrics;
+}
+
+/*
+ * Write metric to three output Datums
+ */
+static void
+metric_to_datums(metric_t *m, Datum *values, bool *nulls)
+{
+	values[0] = CStringGetTextDatum(m->name);
+	nulls[0] = false;
+	if (m->is_bucket)
+	{
+		values[1] = Float8GetDatum(m->bucket_le);
+		nulls[1] = false;
+	}
+	else
+	{
+		values[1] = (Datum) 0;
+		nulls[1] = true;
+	}
+	values[2] = Float8GetDatum(m->value);
+	nulls[2] = false;
+}
+
+PG_FUNCTION_INFO_V1(neon_get_backend_perf_counters);
+Datum
+neon_get_backend_perf_counters(PG_FUNCTION_ARGS)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Datum		values[5];
+	bool		nulls[5];
+
+	/* We put all the tuples into a tuplestore in one go. */
+	InitMaterializedSRF(fcinfo, 0);
+
+	for (int procno = 0; procno < MaxBackends; procno++)
+	{
+		PGPROC	   *proc = GetPGProcByNumber(procno);
+		int			pid = proc->pid;
+		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
+		metric_t   *metrics = neon_perf_counters_to_metrics(counters);
+
+		values[0] = Int32GetDatum(procno);
+		nulls[0] = false;
+		values[1] = Int32GetDatum(pid);
+		nulls[1] = false;
+
+		for (int i = 0; metrics[i].name != NULL; i++)
+		{
+			metric_to_datums(&metrics[i], &values[2], &nulls[2]);
+			tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+		}
+
+		pfree(metrics);
+	}
+
+	return (Datum) 0;
+}
+
+PG_FUNCTION_INFO_V1(neon_get_perf_counters);
+Datum
+neon_get_perf_counters(PG_FUNCTION_ARGS)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Datum		values[3];
+	bool		nulls[3];
+	Datum		getpage_wait_str;
+	neon_per_backend_counters totals = {0};
+	metric_t   *metrics;
+
+	/* We put all the tuples into a tuplestore in one go. */
+	InitMaterializedSRF(fcinfo, 0);
+
+	/* Aggregate the counters across all backends */
+	for (int procno = 0; procno < MaxBackends; procno++)
+	{
+		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
+
+		totals.getpage_wait_us_count += counters->getpage_wait_us_count;
+		totals.getpage_wait_us_sum += counters->getpage_wait_us_sum;
+		for (int bucketno = 0; bucketno < NUM_GETPAGE_WAIT_BUCKETS; bucketno++)
+			totals.getpage_wait_us_bucket[bucketno] += counters->getpage_wait_us_bucket[bucketno];
+		totals.getpage_prefetch_requests_total += counters->getpage_prefetch_requests_total;
+		totals.getpage_sync_requests_total += counters->getpage_sync_requests_total;
+		totals.getpage_prefetch_misses_total += counters->getpage_prefetch_misses_total;
+		totals.getpage_prefetch_discards_total += counters->getpage_prefetch_discards_total;
+		totals.pageserver_requests_sent_total += counters->pageserver_requests_sent_total;
+		totals.pageserver_disconnects_total += counters->pageserver_disconnects_total;
+		totals.pageserver_send_flushes_total += counters->pageserver_send_flushes_total;
+		totals.file_cache_hits_total += counters->file_cache_hits_total;
+	}
+
+	metrics = neon_perf_counters_to_metrics(&totals);
+	for (int i = 0; metrics[i].name != NULL; i++)
+	{
+		metric_to_datums(&metrics[i], &values[0], &nulls[0]);
+		tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+	}
+	pfree(metrics);
+
+	return (Datum) 0;
+}

pgxn/neon/neon_perf_counters.h

@@ -0,0 +1,111 @@
+/*-------------------------------------------------------------------------
+ *
+ * neon_perf_counters.h
+ *	  Performance counters for neon storage requests
+ *-------------------------------------------------------------------------
+ */
+
+#ifndef NEON_PERF_COUNTERS_H
+#define NEON_PERF_COUNTERS_H
+
+#if PG_VERSION_NUM >= 170000
+#include "storage/procnumber.h"
+#else
+#include "storage/backendid.h"
+#include "storage/proc.h"
+#endif
+
+static const uint64 getpage_wait_bucket_thresholds[] = {
+	      20,       30,       60,       100,  /* 0      -  100 us */
+	     200,      300,      600,	   1000,  /* 100 us - 1 ms */
+	    2000,     3000,     6000,     10000,  /* 1 ms   - 10 ms */
+	   20000,    30000,    60000,    100000,  /* 10 ms  - 100 ms */
+	  200000,   300000,   600000,   1000000,  /* 100 ms - 1 s */
+	 2000000,  3000000,  6000000,  10000000,  /* 1 s - 10 s */
+    20000000, 30000000, 60000000, 100000000,  /* 10 s - 100 s */
+	UINT64_MAX,
+};
+#define NUM_GETPAGE_WAIT_BUCKETS (lengthof(getpage_wait_bucket_thresholds))
+
+typedef struct
+{
+	/*
+	 * Histogram for how long an smgrread() request needs to wait for response
+	 * from pageserver. When prefetching is effective, these wait times can be
+	 * lower than the network latency to the pageserver, even zero, if the
+	 * page is already readily prefetched whenever we need to read a page.
+	 *
+	 * Note: we accumulate these in microseconds, because that's convenient in
+	 * the backend, but the 'neon_backend_perf_counters' view will convert
+	 * them to seconds, to make them more idiomatic as prometheus metrics.
+	 */
+	uint64		getpage_wait_us_count;
+	uint64		getpage_wait_us_sum;
+	uint64		getpage_wait_us_bucket[NUM_GETPAGE_WAIT_BUCKETS];
+
+	/*
+	 * Total number of speculative prefetch Getpage requests and synchronous
+	 * GetPage requests sent.
+	 */
+	uint64		getpage_prefetch_requests_total;
+	uint64		getpage_sync_requests_total;
+
+	/* XXX: It's not clear to me when these misses happen. */
+	uint64		getpage_prefetch_misses_total;
+
+	/*
+	 * Number of prefetched responses that were discarded becuase the
+	 * prefetched page was not needed or because it was concurrently fetched /
+	 * modified by another backend.
+	 */
+	uint64		getpage_prefetch_discards_total;
+
+	/*
+	 * Total number of requests send to pageserver. (prefetch_requests_total
+	 * and sync_request_total count only GetPage requests, this counts all
+	 * request types.)
+	 */
+	uint64		pageserver_requests_sent_total;
+
+	/*
+	 * Number of times the connection to the pageserver was lost and the
+	 * backend had to reconnect. Note that this doesn't count the first
+	 * connection in each backend, only reconnects.
+	 */
+	uint64		pageserver_disconnects_total;
+
+	/*
+	 * Number of network flushes to the pageserver. Synchronous requests are
+	 * flushed immediately, but when prefetching requests are sent in batches,
+	 * this can be smaller than pageserver_requests_sent_total.
+	 */
+	uint64		pageserver_send_flushes_total;
+
+	/*
+	 * Number of requests satisfied from the LFC.
+	 *
+	 * This is redundant with the server-wide file_cache_hits, but this gives
+	 * per-backend granularity, and it's handy to have this in the same place
+	 * as counters for requests that went to the pageserver. Maybe move all
+	 * the LFC stats to this struct in the future?
+	 */
+	uint64		file_cache_hits_total;
+
+} neon_per_backend_counters;
+
+/* Pointer to the shared memory array of neon_per_backend_counters structs */
+extern neon_per_backend_counters *neon_per_backend_counters_shared;
+
+#if PG_VERSION_NUM >= 170000
+#define MyNeonCounters (&neon_per_backend_counters_shared[MyProcNumber])
+#else
+#define MyNeonCounters (&neon_per_backend_counters_shared[MyProc->pgprocno])
+#endif
+
+extern void inc_getpage_wait(uint64 latency);
+
+extern Size NeonPerfCountersShmemSize(void);
+extern void NeonPerfCountersShmemInit(void);
+
+
+#endif							/* NEON_PERF_COUNTERS_H */

pgxn/neon/neon_pgversioncompat.c

@@ -0,0 +1,44 @@
+/*
+ * Support functions for the compatibility macros in neon_pgversioncompat.h
+ */
+#include "postgres.h"
+
+#include "funcapi.h"
+#include "miscadmin.h"
+#include "utils/tuplestore.h"
+
+#include "neon_pgversioncompat.h"
+
+#if PG_MAJORVERSION_NUM < 15
+void
+InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags)
+{
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	Tuplestorestate *tupstore;
+	MemoryContext old_context,
+				per_query_ctx;
+	TupleDesc	stored_tupdesc;
+
+	/* check to see if caller supports returning a tuplestore */
+	if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
+		ereport(ERROR,
+				(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+				 errmsg("set-valued function called in context that cannot accept a set")));
+
+	/*
+	 * Store the tuplestore and the tuple descriptor in ReturnSetInfo.  This
+	 * must be done in the per-query memory context.
+	 */
+	per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
+	old_context = MemoryContextSwitchTo(per_query_ctx);
+
+	if (get_call_result_type(fcinfo, NULL, &stored_tupdesc) != TYPEFUNC_COMPOSITE)
+		elog(ERROR, "return type must be a row type");
+
+	tupstore = tuplestore_begin_heap(false, false, work_mem);
+	rsinfo->returnMode = SFRM_Materialize;
+	rsinfo->setResult = tupstore;
+	rsinfo->setDesc = stored_tupdesc;
+	MemoryContextSwitchTo(old_context);
+}
+#endif

pgxn/neon/neon_pgversioncompat.h

@@ -6,6 +6,8 @@
 #ifndef NEON_PGVERSIONCOMPAT_H
 #define NEON_PGVERSIONCOMPAT_H
 
+#include "fmgr.h"
+
 #if PG_MAJORVERSION_NUM < 17
 #define NRelFileInfoBackendIsTemp(rinfo) (rinfo.backend != InvalidBackendId)
 #else
@@ -123,4 +125,8 @@
 #define AmAutoVacuumWorkerProcess() (IsAutoVacuumWorkerProcess())
 #endif
 
+#if PG_MAJORVERSION_NUM < 15
+extern void InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags);
+#endif
+
 #endif							/* NEON_PGVERSIONCOMPAT_H */

pgxn/neon/pagestore_smgr.c

@@ -66,6 +66,7 @@
 #include "storage/md.h"
 #include "storage/smgr.h"
 
+#include "neon_perf_counters.h"
 #include "pagestore_client.h"
 #include "bitmap.h"
 
@@ -289,7 +290,6 @@ static PrefetchState *MyPState;
 
 static bool compact_prefetch_buffers(void);
 static void consume_prefetch_responses(void);
-static uint64 prefetch_register_buffer(BufferTag tag, neon_request_lsns *force_request_lsns);
 static bool prefetch_read(PrefetchRequest *slot);
 static void prefetch_do_request(PrefetchRequest *slot, neon_request_lsns *force_request_lsns);
 static bool prefetch_wait_for(uint64 ring_index);
@@ -780,21 +780,27 @@ prefetch_do_request(PrefetchRequest *slot, neon_request_lsns *force_request_lsns
 }
 
 /*
- * prefetch_register_buffer() - register and prefetch buffer
+ * prefetch_register_bufferv() - register and prefetch buffers
  *
  * Register that we may want the contents of BufferTag in the near future.
+ * This is used when issuing a speculative prefetch request, but also when
+ * performing a synchronous request and need the buffer right now.
  *
  * If force_request_lsns is not NULL, those values are sent to the
  * pageserver. If NULL, we utilize the lastWrittenLsn -infrastructure
  * to calculate the LSNs to send.
  *
+ * When performing a prefetch rather than a synchronous request,
+ * is_prefetch==true. Currently, it only affects how the request is accounted
+ * in the perf counters.
+ *
  * NOTE: this function may indirectly update MyPState->pfs_hash; which
  * invalidates any active pointers into the hash table.
  */
-
 static uint64
 prefetch_register_bufferv(BufferTag tag, neon_request_lsns *frlsns,
-						  BlockNumber nblocks, const bits8 *mask)
+						  BlockNumber nblocks, const bits8 *mask,
+						  bool is_prefetch)
 {
 	uint64		min_ring_index;
 	PrefetchRequest req;
@@ -815,6 +821,7 @@ Retry:
 		PrfHashEntry *entry = NULL;
 		uint64		ring_index;
 		neon_request_lsns *lsns;
+
 		if (PointerIsValid(mask) && !BITMAP_ISSET(mask, i))
 			continue;
 
@@ -858,6 +865,7 @@ Retry:
 					prefetch_set_unused(ring_index);
 					entry = NULL;
 					slot = NULL;
+					MyNeonCounters->getpage_prefetch_discards_total++;
 				}
 			}
 
@@ -972,6 +980,11 @@ Retry:
 
 		min_ring_index = Min(min_ring_index, ring_index);
 
+		if (is_prefetch)
+			MyNeonCounters->getpage_prefetch_requests_total++;
+		else
+			MyNeonCounters->getpage_sync_requests_total++;
+
 		prefetch_do_request(slot, lsns);
 	}
 
@@ -1000,13 +1013,6 @@ Retry:
 }
 
 
-static uint64
-prefetch_register_buffer(BufferTag tag, neon_request_lsns *force_request_lsns)
-{
-	return prefetch_register_bufferv(tag, force_request_lsns, 1, NULL);
-}
-
-
 /*
  * Note: this function can get canceled and use a long jump to the next catch
  * context. Take care.
@@ -2612,7 +2618,7 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 			lfc_present[i] = ~(lfc_present[i]);
 
 		ring_index = prefetch_register_bufferv(tag, NULL, iterblocks,
-											   lfc_present);
+											   lfc_present, true);
 		nblocks -= iterblocks;
 		blocknum += iterblocks;
 
@@ -2656,7 +2662,7 @@ neon_prefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum)
 
 	CopyNRelFileInfoToBufTag(tag, InfoFromSMgrRel(reln));
 
-	ring_index = prefetch_register_buffer(tag, NULL);
+	ring_index = prefetch_register_bufferv(tag, NULL, 1, NULL, true);
 
 	Assert(ring_index < MyPState->ring_unused &&
 		   MyPState->ring_last <= ring_index);
@@ -2747,17 +2753,20 @@ neon_read_at_lsnv(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber base_block
 	 * weren't for the behaviour of the LwLsn cache that uses the highest
 	 * value of the LwLsn cache when the entry is not found.
 	 */
-	prefetch_register_bufferv(buftag, request_lsns, nblocks, mask);
+	prefetch_register_bufferv(buftag, request_lsns, nblocks, mask, false);
 
 	for (int i = 0; i < nblocks; i++)
 	{
 		void	   *buffer = buffers[i];
 		BlockNumber blockno = base_blockno + i;
 		neon_request_lsns *reqlsns = &request_lsns[i];
+		TimestampTz		start_ts, end_ts;
 
 		if (PointerIsValid(mask) && !BITMAP_ISSET(mask, i))
 			continue;
 
+		start_ts = GetCurrentTimestamp();
+
 		if (RecoveryInProgress() && MyBackendType != B_STARTUP)
 			XLogWaitForReplayOf(reqlsns[0].request_lsn);
 
@@ -2794,6 +2803,7 @@ Retry:
 				/* drop caches */
 				prefetch_set_unused(slot->my_ring_index);
 				pgBufferUsage.prefetch.expired += 1;
+				MyNeonCounters->getpage_prefetch_discards_total++;
 				/* make it look like a prefetch cache miss */
 				entry = NULL;
 			}
@@ -2804,8 +2814,9 @@ Retry:
 			if (entry == NULL)
 			{
 				pgBufferUsage.prefetch.misses += 1;
+				MyNeonCounters->getpage_prefetch_misses_total++;
 
-				ring_index = prefetch_register_bufferv(buftag, reqlsns, 1, NULL);
+				ring_index = prefetch_register_bufferv(buftag, reqlsns, 1, NULL, false);
 				Assert(ring_index != UINT64_MAX);
 				slot = GetPrfSlot(ring_index);
 			}
@@ -2860,6 +2871,9 @@ Retry:
 		/* buffer was used, clean up for later reuse */
 		prefetch_set_unused(ring_index);
 		prefetch_cleanup_trailing_unused();
+
+		end_ts = GetCurrentTimestamp();
+		inc_getpage_wait(end_ts >= start_ts ? (end_ts - start_ts) : 0);
 	}
 }
 
@@ -2913,6 +2927,7 @@ neon_read(SMgrRelation reln, ForkNumber forkNum, BlockNumber blkno, void *buffer
 	/* Try to read from local file cache */
 	if (lfc_read(InfoFromSMgrRel(reln), forkNum, blkno, buffer))
 	{
+		MyNeonCounters->file_cache_hits_total++;
 		return;
 	}
 
@@ -3097,7 +3112,7 @@ neon_readv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
 				/* assume heap */
 				RmgrTable[RM_HEAP_ID].rm_mask(mdbuf_masked, blkno);
 				RmgrTable[RM_HEAP_ID].rm_mask(pageserver_masked, blkno);
-	
+
 				if (memcmp(mdbuf_masked, pageserver_masked, BLCKSZ) != 0)
 				{
 					neon_log(PANIC, "heap buffers differ at blk %u in rel %u/%u/%u fork %u (request LSN %X/%08X):\n------ MD ------\n%s\n------ Page Server ------\n%s\n",

pgxn/neon_walredo/walredoproc.c

@@ -24,6 +24,7 @@
  * PushPage ('P'): Copy a page image (in the payload) to buffer cache
  * ApplyRecord ('A'): Apply a WAL record (in the payload)
  * GetPage ('G'): Return a page image from buffer cache.
+ * Ping ('H'): Return the input message.
  *
  * Currently, you only get a response to GetPage requests; the response is
  * simply a 8k page, without any headers. Errors are logged to stderr.
@@ -133,6 +134,7 @@ static void ApplyRecord(StringInfo input_message);
 static void apply_error_callback(void *arg);
 static bool redo_block_filter(XLogReaderState *record, uint8 block_id);
 static void GetPage(StringInfo input_message);
+static void Ping(StringInfo input_message);
 static ssize_t buffered_read(void *buf, size_t count);
 static void CreateFakeSharedMemoryAndSemaphores();
 
@@ -394,6 +396,10 @@ WalRedoMain(int argc, char *argv[])
 				GetPage(&input_message);
 				break;
 
+			case 'H': 			/* Ping */
+				Ping(&input_message);
+				break;
+
 				/*
 				 * EOF means we're done. Perform normal shutdown.
 				 */
@@ -1057,6 +1063,36 @@ GetPage(StringInfo input_message)
 }
 
 
+static void
+Ping(StringInfo input_message)
+{
+	int			tot_written;
+	/* Response: the input message */
+	tot_written = 0;
+	do {
+		ssize_t		rc;
+		/* We don't need alignment, but it's bad practice to use char[BLCKSZ] */
+#if PG_VERSION_NUM >= 160000
+		static const PGIOAlignedBlock response;
+#else
+		static const PGAlignedBlock response;
+#endif
+		rc = write(STDOUT_FILENO, &response.data[tot_written], BLCKSZ - tot_written);
+		if (rc < 0) {
+			/* If interrupted by signal, just retry */
+			if (errno == EINTR)
+				continue;
+			ereport(ERROR,
+					(errcode_for_file_access(),
+					 errmsg("could not write to stdout: %m")));
+		}
+		tot_written += rc;
+	} while (tot_written < BLCKSZ);
+
+	elog(TRACE, "Page sent back for ping");
+}
+
+
 /* Buffer used by buffered_read() */
 static char stdin_buf[16 * 1024];
 static size_t stdin_len = 0;	/* # of bytes in buffer */

proxy/Cargo.toml

@@ -24,12 +24,12 @@ bytes = { workspace = true, features = ["serde"] }
 camino.workspace = true
 chrono.workspace = true
 clap.workspace = true
+compute_api.workspace = true
 consumption_metrics.workspace = true
 dashmap.workspace = true
 env_logger.workspace = true
 framed-websockets.workspace = true
 futures.workspace = true
-git-version.workspace = true
 hashbrown.workspace = true
 hashlink.workspace = true
 hex.workspace = true

proxy/src/auth/backend.rs

@@ -80,6 +80,14 @@ pub(crate) trait TestBackend: Send + Sync + 'static {
     fn get_allowed_ips_and_secret(
         &self,
     ) -> Result<(CachedAllowedIps, Option<CachedRoleSecret>), console::errors::GetAuthInfoError>;
+    fn dyn_clone(&self) -> Box<dyn TestBackend>;
+}
+
+#[cfg(test)]
+impl Clone for Box<dyn TestBackend> {
+    fn clone(&self) -> Self {
+        TestBackend::dyn_clone(&**self)
+    }
 }
 
 impl std::fmt::Display for Backend<'_, (), ()> {
@@ -444,7 +452,7 @@ impl<'a> Backend<'a, ComputeUserInfoMaybeEndpoint, &()> {
             Self::Web(url, ()) => {
                 info!("performing web authentication");
 
-                let info = web::authenticate(ctx, &url, client).await?;
+                let info = web::authenticate(ctx, config, &url, client).await?;
 
                 Backend::Web(url, info)
             }
@@ -557,7 +565,7 @@ mod tests {
         stream::{PqStream, Stream},
     };
 
-    use super::{auth_quirks, AuthRateLimiter};
+    use super::{auth_quirks, jwt::JwkCache, AuthRateLimiter};
 
     struct Auth {
         ips: Vec<IpPattern>,
@@ -585,6 +593,14 @@ mod tests {
             ))
         }
 
+        async fn get_endpoint_jwks(
+            &self,
+            _ctx: &RequestMonitoring,
+            _endpoint: crate::EndpointId,
+        ) -> anyhow::Result<Vec<super::jwt::AuthRule>> {
+            unimplemented!()
+        }
+
         async fn wake_compute(
             &self,
             _ctx: &RequestMonitoring,
@@ -595,12 +611,15 @@ mod tests {
     }
 
     static CONFIG: Lazy<AuthenticationConfig> = Lazy::new(|| AuthenticationConfig {
+        jwks_cache: JwkCache::default(),
         thread_pool: ThreadPool::new(1),
         scram_protocol_timeout: std::time::Duration::from_secs(5),
         rate_limiter_enabled: true,
         rate_limiter: AuthRateLimiter::new(&RateBucketInfo::DEFAULT_AUTH_SET),
         rate_limit_ip_subnet: 64,
         ip_allowlist_check_enabled: true,
+        is_auth_broker: false,
+        accept_jwts: false,
     });
 
     async fn read_message(r: &mut (impl AsyncRead + Unpin), b: &mut BytesMut) -> PgMessage {

proxy/src/auth/backend/jwt.rs

@@ -1,4 +1,5 @@
 use std::{
+    borrow::Cow,
     future::Future,
     sync::Arc,
     time::{Duration, SystemTime},
@@ -8,11 +9,17 @@ use anyhow::{bail, ensure, Context};
 use arc_swap::ArcSwapOption;
 use dashmap::DashMap;
 use jose_jwk::crypto::KeyInfo;
-use serde::{Deserialize, Deserializer};
+use serde::{
+    de::{DeserializeSeed, IgnoredAny, Visitor},
+    Deserializer,
+};
 use signature::Verifier;
 use tokio::time::Instant;
 
-use crate::{context::RequestMonitoring, http::parse_json_body_with_limit, EndpointId, RoleName};
+use crate::{
+    context::RequestMonitoring, http::parse_json_body_with_limit, intern::RoleNameInt, EndpointId,
+    RoleName,
+};
 
 // TODO(conrad): make these configurable.
 const CLOCK_SKEW_LEEWAY: Duration = Duration::from_secs(30);
@@ -27,18 +34,19 @@ pub(crate) trait FetchAuthRules: Clone + Send + Sync + 'static {
         &self,
         ctx: &RequestMonitoring,
         endpoint: EndpointId,
-        role_name: RoleName,
     ) -> impl Future<Output = anyhow::Result<Vec<AuthRule>>> + Send;
 }
 
+#[derive(Debug, Clone)]
 pub(crate) struct AuthRule {
     pub(crate) id: String,
     pub(crate) jwks_url: url::Url,
     pub(crate) audience: Option<String>,
+    pub(crate) role_names: Vec<RoleNameInt>,
 }
 
 #[derive(Default)]
-pub(crate) struct JwkCache {
+pub struct JwkCache {
     client: reqwest::Client,
 
     map: DashMap<(EndpointId, RoleName), Arc<JwkCacheEntryLock>>,
@@ -54,18 +62,28 @@ pub(crate) struct JwkCacheEntry {
 }
 
 impl JwkCacheEntry {
-    fn find_jwk_and_audience(&self, key_id: &str) -> Option<(&jose_jwk::Jwk, Option<&str>)> {
-        self.key_sets.values().find_map(|key_set| {
-            key_set
-                .find_key(key_id)
-                .map(|jwk| (jwk, key_set.audience.as_deref()))
-        })
+    fn find_jwk_and_audience(
+        &self,
+        key_id: &str,
+        role_name: &RoleName,
+    ) -> Option<(&jose_jwk::Jwk, Option<&str>)> {
+        self.key_sets
+            .values()
+            // make sure our requested role has access to the key set
+            .filter(|key_set| key_set.role_names.iter().any(|role| **role == **role_name))
+            // try and find the requested key-id in the key set
+            .find_map(|key_set| {
+                key_set
+                    .find_key(key_id)
+                    .map(|jwk| (jwk, key_set.audience.as_deref()))
+            })
     }
 }
 
 struct KeySet {
     jwks: jose_jwk::JwkSet,
     audience: Option<String>,
+    role_names: Vec<RoleNameInt>,
 }
 
 impl KeySet {
@@ -106,7 +124,6 @@ impl JwkCacheEntryLock {
         ctx: &RequestMonitoring,
         client: &reqwest::Client,
         endpoint: EndpointId,
-        role_name: RoleName,
         auth_rules: &F,
     ) -> anyhow::Result<Arc<JwkCacheEntry>> {
         // double check that no one beat us to updating the cache.
@@ -119,11 +136,10 @@ impl JwkCacheEntryLock {
             }
         }
 
-        let rules = auth_rules
-            .fetch_auth_rules(ctx, endpoint, role_name)
-            .await?;
+        let rules = auth_rules.fetch_auth_rules(ctx, endpoint).await?;
         let mut key_sets =
             ahash::HashMap::with_capacity_and_hasher(rules.len(), ahash::RandomState::new());
+
         // TODO(conrad): run concurrently
         // TODO(conrad): strip the JWKs urls (should be checked by cplane as well - cloud#16284)
         for rule in rules {
@@ -151,6 +167,7 @@ impl JwkCacheEntryLock {
                                 KeySet {
                                     jwks,
                                     audience: rule.audience,
+                                    role_names: rule.role_names,
                                 },
                             );
                         }
@@ -173,7 +190,6 @@ impl JwkCacheEntryLock {
         ctx: &RequestMonitoring,
         client: &reqwest::Client,
         endpoint: EndpointId,
-        role_name: RoleName,
         fetch: &F,
     ) -> Result<Arc<JwkCacheEntry>, anyhow::Error> {
         let now = Instant::now();
@@ -183,9 +199,7 @@ impl JwkCacheEntryLock {
         let Some(cached) = guard else {
             let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
             let permit = self.acquire_permit().await;
-            return self
-                .renew_jwks(permit, ctx, client, endpoint, role_name, fetch)
-                .await;
+            return self.renew_jwks(permit, ctx, client, endpoint, fetch).await;
         };
 
         let last_update = now.duration_since(cached.last_retrieved);
@@ -196,9 +210,7 @@ impl JwkCacheEntryLock {
             let permit = self.acquire_permit().await;
 
             // it's been too long since we checked the keys. wait for them to update.
-            return self
-                .renew_jwks(permit, ctx, client, endpoint, role_name, fetch)
-                .await;
+            return self.renew_jwks(permit, ctx, client, endpoint, fetch).await;
         }
 
         // every 5 minutes we should spawn a job to eagerly update the token.
@@ -212,7 +224,7 @@ impl JwkCacheEntryLock {
                 let ctx = ctx.clone();
                 tokio::spawn(async move {
                     if let Err(e) = entry
-                        .renew_jwks(permit, &ctx, &client, endpoint, role_name, &fetch)
+                        .renew_jwks(permit, &ctx, &client, endpoint, &fetch)
                         .await
                     {
                         tracing::warn!(error=?e, "could not fetch JWKs in background job");
@@ -232,7 +244,7 @@ impl JwkCacheEntryLock {
         jwt: &str,
         client: &reqwest::Client,
         endpoint: EndpointId,
-        role_name: RoleName,
+        role_name: &RoleName,
         fetch: &F,
     ) -> Result<(), anyhow::Error> {
         // JWT compact form is defined to be
@@ -254,30 +266,26 @@ impl JwkCacheEntryLock {
         let sig = base64::decode_config(signature, base64::URL_SAFE_NO_PAD)
             .context("Provided authentication token is not a valid JWT encoding")?;
 
-        ensure!(header.typ == "JWT");
+        ensure!(
+            header.typ == "JWT",
+            "Provided authentication token is not a valid JWT encoding"
+        );
         let kid = header.key_id.context("missing key id")?;
 
         let mut guard = self
-            .get_or_update_jwk_cache(ctx, client, endpoint.clone(), role_name.clone(), fetch)
+            .get_or_update_jwk_cache(ctx, client, endpoint.clone(), fetch)
             .await?;
 
         // get the key from the JWKs if possible. If not, wait for the keys to update.
         let (jwk, expected_audience) = loop {
-            match guard.find_jwk_and_audience(kid) {
+            match guard.find_jwk_and_audience(kid, role_name) {
                 Some(jwk) => break jwk,
                 None if guard.last_retrieved.elapsed() > MIN_RENEW => {
                     let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
 
                     let permit = self.acquire_permit().await;
                     guard = self
-                        .renew_jwks(
-                            permit,
-                            ctx,
-                            client,
-                            endpoint.clone(),
-                            role_name.clone(),
-                            fetch,
-                        )
+                        .renew_jwks(permit, ctx, client, endpoint.clone(), fetch)
                         .await?;
                 }
                 _ => {
@@ -300,32 +308,21 @@ impl JwkCacheEntryLock {
             }
             key => bail!("unsupported key type {key:?}"),
         };
+        tracing::debug!("JWT signature valid");
 
         let payload = base64::decode_config(payload, base64::URL_SAFE_NO_PAD)
             .context("Provided authentication token is not a valid JWT encoding")?;
-        let payload = serde_json::from_slice::<JwtPayload<'_>>(&payload)
-            .context("Provided authentication token is not a valid JWT encoding")?;
 
-        tracing::debug!(?payload, "JWT signature valid with claims");
+        let validator = JwtValidator {
+            expected_audience,
+            current_time: SystemTime::now(),
+            clock_skew_leeway: CLOCK_SKEW_LEEWAY,
+        };
 
-        match (expected_audience, payload.audience) {
-            // check the audience matches
-            (Some(aud1), Some(aud2)) => ensure!(aud1 == aud2, "invalid JWT token audience"),
-            // the audience is expected but is missing
-            (Some(_), None) => bail!("invalid JWT token audience"),
-            // we don't care for the audience field
-            (None, _) => {}
-        }
+        let payload = validator
+            .deserialize(&mut serde_json::Deserializer::from_slice(&payload))?;
 
-        let now = SystemTime::now();
-
-        if let Some(exp) = payload.expiration {
-            ensure!(now < exp + CLOCK_SKEW_LEEWAY);
-        }
-
-        if let Some(nbf) = payload.not_before {
-            ensure!(nbf < now + CLOCK_SKEW_LEEWAY);
-        }
+        tracing::debug!(?payload, "JWT claims valid");
 
         Ok(())
     }
@@ -336,7 +333,7 @@ impl JwkCache {
         &self,
         ctx: &RequestMonitoring,
         endpoint: EndpointId,
-        role_name: RoleName,
+        role_name: &RoleName,
         fetch: &F,
         jwt: &str,
     ) -> Result<(), anyhow::Error> {
@@ -413,37 +410,184 @@ struct JwtHeader<'a> {
     key_id: Option<&'a str>,
 }
 
-/// <https://datatracker.ietf.org/doc/html/rfc7519#section-4.1>
-#[derive(serde::Deserialize, serde::Serialize, Debug)]
-struct JwtPayload<'a> {
-    /// Audience - Recipient for which the JWT is intended
-    #[serde(rename = "aud")]
-    audience: Option<&'a str>,
-    /// Expiration - Time after which the JWT expires
-    #[serde(deserialize_with = "numeric_date_opt", rename = "exp", default)]
-    expiration: Option<SystemTime>,
-    /// Not before - Time after which the JWT expires
-    #[serde(deserialize_with = "numeric_date_opt", rename = "nbf", default)]
-    not_before: Option<SystemTime>,
-
-    // the following entries are only extracted for the sake of debug logging.
-    /// Issuer of the JWT
-    #[serde(rename = "iss")]
-    issuer: Option<&'a str>,
-    /// Subject of the JWT (the user)
-    #[serde(rename = "sub")]
-    subject: Option<&'a str>,
-    /// Unique token identifier
-    #[serde(rename = "jti")]
-    jwt_id: Option<&'a str>,
-    /// Unique session identifier
-    #[serde(rename = "sid")]
-    session_id: Option<&'a str>,
+struct JwtValidator<'a> {
+    expected_audience: Option<&'a str>,
+    current_time: SystemTime,
+    clock_skew_leeway: Duration,
 }
 
-fn numeric_date_opt<'de, D: Deserializer<'de>>(d: D) -> Result<Option<SystemTime>, D::Error> {
-    let d = <Option<u64>>::deserialize(d)?;
-    Ok(d.map(|n| SystemTime::UNIX_EPOCH + Duration::from_secs(n)))
+impl<'de> DeserializeSeed<'de> for JwtValidator<'_> {
+    type Value = JwtPayload<'de>;
+
+    fn deserialize<D>(self, deserializer: D) -> Result<Self::Value, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        impl<'de> Visitor<'de> for JwtValidator<'_> {
+            type Value = JwtPayload<'de>;
+
+            fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+                formatter.write_str("a JWT payload")
+            }
+
+            fn visit_map<A>(self, mut map: A) -> Result<Self::Value, A::Error>
+            where
+                A: serde::de::MapAccess<'de>,
+            {
+                let mut payload = JwtPayload {
+                    issuer: None,
+                    subject: None,
+                    jwt_id: None,
+                    session_id: None,
+                };
+
+                let mut aud = false;
+
+                while let Some(key) = map.next_key()? {
+                    match key {
+                        "iss" if payload.issuer.is_none() => {
+                            payload.issuer = Some(map.next_value()?);
+                        }
+                        "sub" if payload.subject.is_none() => {
+                            payload.subject = Some(map.next_value()?);
+                        }
+                        "jit" if payload.jwt_id.is_none() => {
+                            payload.jwt_id = Some(map.next_value()?);
+                        }
+                        "sid" if payload.session_id.is_none() => {
+                            payload.session_id = Some(map.next_value()?);
+                        }
+                        "exp" => {
+                            let exp = map.next_value::<u64>()?;
+                            let exp = SystemTime::UNIX_EPOCH + Duration::from_secs(exp);
+
+                            if self.current_time > exp + self.clock_skew_leeway {
+                                return Err(serde::de::Error::custom("JWT token has expired"));
+                            }
+                        }
+                        "nbf" => {
+                            let nbf = map.next_value::<u64>()?;
+                            let nbf = SystemTime::UNIX_EPOCH + Duration::from_secs(nbf);
+
+                            if self.current_time + self.clock_skew_leeway < nbf {
+                                return Err(serde::de::Error::custom(
+                                    "JWT token is not yet ready to use",
+                                ));
+                            }
+                        }
+                        "aud" => {
+                            if let Some(expected_audience) = self.expected_audience {
+                                map.next_value_seed(AudienceValidator { expected_audience })?;
+                                aud = true;
+                            } else {
+                                map.next_value::<IgnoredAny>()?;
+                            }
+                        }
+                        _ => map.next_value::<IgnoredAny>().map(|IgnoredAny| ())?,
+                    }
+                }
+
+                if self.expected_audience.is_some() && !aud {
+                    return Err(serde::de::Error::custom("invalid JWT token audience"));
+                }
+
+                Ok(payload)
+            }
+        }
+
+        deserializer.deserialize_map(self)
+    }
+}
+
+struct AudienceValidator<'a> {
+    expected_audience: &'a str,
+}
+
+impl<'de> DeserializeSeed<'de> for AudienceValidator<'_> {
+    type Value = ();
+
+    fn deserialize<D>(self, deserializer: D) -> Result<Self::Value, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        impl<'de> Visitor<'de> for AudienceValidator<'_> {
+            type Value = ();
+
+            fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+                formatter.write_str("a single string or an array of strings")
+            }
+
+            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
+            where
+                E: serde::de::Error,
+            {
+                if self.expected_audience == v {
+                    Ok(())
+                } else {
+                    Err(E::custom("invalid JWT token audience"))
+                }
+            }
+
+            fn visit_seq<A>(self, mut seq: A) -> Result<Self::Value, A::Error>
+            where
+                A: serde::de::SeqAccess<'de>,
+            {
+                while let Some(v) = seq.next_element_seed(SingleAudienceValidator {
+                    expected_audience: self.expected_audience,
+                })? {
+                    if v {
+                        return Ok(());
+                    }
+                }
+                Err(serde::de::Error::custom("invalid JWT token audience"))
+            }
+        }
+        deserializer.deserialize_any(self)
+    }
+}
+
+struct SingleAudienceValidator<'a> {
+    expected_audience: &'a str,
+}
+
+impl<'de> DeserializeSeed<'de> for SingleAudienceValidator<'_> {
+    type Value = bool;
+
+    fn deserialize<D>(self, deserializer: D) -> Result<Self::Value, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        impl<'de> Visitor<'de> for SingleAudienceValidator<'_> {
+            type Value = bool;
+
+            fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+                formatter.write_str("a single audience string")
+            }
+
+            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
+            where
+                E: serde::de::Error,
+            {
+                Ok(self.expected_audience == v)
+            }
+        }
+        deserializer.deserialize_any(self)
+    }
+}
+
+/// <https://datatracker.ietf.org/doc/html/rfc7519#section-4.1>
+// the following entries are only extracted for the sake of debug logging.
+#[derive(Debug)]
+#[allow(dead_code)]
+struct JwtPayload<'a> {
+    /// Issuer of the JWT
+    issuer: Option<Cow<'a, str>>,
+    /// Subject of the JWT (the user)
+    subject: Option<Cow<'a, str>>,
+    /// Unique token identifier
+    jwt_id: Option<Cow<'a, str>>,
+    /// Unique session identifier
+    session_id: Option<Cow<'a, str>>,
 }
 
 struct JwkRenewalPermit<'a> {
@@ -524,6 +668,8 @@ mod tests {
     use hyper_util::rt::TokioIo;
     use rand::rngs::OsRng;
     use rsa::pkcs8::DecodePrivateKey;
+    use serde::Serialize;
+    use serde_json::json;
     use signature::Signer;
     use tokio::net::TcpListener;
 
@@ -556,23 +702,41 @@ mod tests {
     }
 
     fn build_jwt_payload(kid: String, sig: jose_jwa::Signing) -> String {
+        let now = SystemTime::now()
+            .duration_since(SystemTime::UNIX_EPOCH)
+            .unwrap()
+            .as_secs();
+        let body = typed_json::json! {{
+            "exp": now + 3600,
+            "nbf": now,
+            "aud": ["audience1", "neon", "audience2"],
+            "sub": "user1",
+            "sid": "session1",
+            "jti": "token1",
+            "iss": "neon-testing",
+        }};
+        build_custom_jwt_payload(kid, body, sig)
+    }
+
+    fn build_custom_jwt_payload(
+        kid: String,
+        body: impl Serialize,
+        sig: jose_jwa::Signing,
+    ) -> String {
         let header = JwtHeader {
             typ: "JWT",
             algorithm: jose_jwa::Algorithm::Signing(sig),
             key_id: Some(&kid),
         };
-        let body = typed_json::json! {{
-            "exp": SystemTime::now().duration_since(SystemTime::UNIX_EPOCH).unwrap().as_secs() + 3600,
-        }};
 
         let header =
             base64::encode_config(serde_json::to_string(&header).unwrap(), URL_SAFE_NO_PAD);
-        let body = base64::encode_config(body.to_string(), URL_SAFE_NO_PAD);
+        let body = base64::encode_config(serde_json::to_string(&body).unwrap(), URL_SAFE_NO_PAD);
 
         format!("{header}.{body}")
     }
 
-    fn new_ec_jwt(kid: String, key: p256::SecretKey) -> String {
+    fn new_ec_jwt(kid: String, key: &p256::SecretKey) -> String {
         use p256::ecdsa::{Signature, SigningKey};
 
         let payload = build_jwt_payload(kid, jose_jwa::Signing::Es256);
@@ -582,6 +746,16 @@ mod tests {
         format!("{payload}.{sig}")
     }
 
+    fn new_custom_ec_jwt(kid: String, key: &p256::SecretKey, body: impl Serialize) -> String {
+        use p256::ecdsa::{Signature, SigningKey};
+
+        let payload = build_custom_jwt_payload(kid, body, jose_jwa::Signing::Es256);
+        let sig: Signature = SigningKey::from(key).sign(payload.as_bytes());
+        let sig = base64::encode_config(sig.to_bytes(), URL_SAFE_NO_PAD);
+
+        format!("{payload}.{sig}")
+    }
+
     fn new_rsa_jwt(kid: String, key: rsa::RsaPrivateKey) -> String {
         use rsa::pkcs1v15::SigningKey;
         use rsa::signature::SignatureEncoding;
@@ -653,42 +827,34 @@ X0n5X2/pBLJzxZc62ccvZYVnctBiFs6HbSnxpuMQCfkt/BcR/ttIepBQQIW86wHL
 -----END PRIVATE KEY-----
 ";
 
-    #[tokio::test]
-    async fn renew() {
-        let (rs1, jwk1) = new_rsa_jwk(RS1, "1".into());
-        let (rs2, jwk2) = new_rsa_jwk(RS2, "2".into());
-        let (ec1, jwk3) = new_ec_jwk("3".into());
-        let (ec2, jwk4) = new_ec_jwk("4".into());
+    #[derive(Clone)]
+    struct Fetch(Vec<AuthRule>);
 
-        let jwt1 = new_rsa_jwt("1".into(), rs1);
-        let jwt2 = new_rsa_jwt("2".into(), rs2);
-        let jwt3 = new_ec_jwt("3".into(), ec1);
-        let jwt4 = new_ec_jwt("4".into(), ec2);
-
-        let foo_jwks = jose_jwk::JwkSet {
-            keys: vec![jwk1, jwk3],
-        };
-        let bar_jwks = jose_jwk::JwkSet {
-            keys: vec![jwk2, jwk4],
-        };
+    impl FetchAuthRules for Fetch {
+        async fn fetch_auth_rules(
+            &self,
+            _ctx: &RequestMonitoring,
+            _endpoint: EndpointId,
+        ) -> anyhow::Result<Vec<AuthRule>> {
+            Ok(self.0.clone())
+        }
+    }
 
+    async fn jwks_server(
+        router: impl for<'a> Fn(&'a str) -> Option<Vec<u8>> + Send + Sync + 'static,
+    ) -> SocketAddr {
+        let router = Arc::new(router);
         let service = service_fn(move |req| {
-            let foo_jwks = foo_jwks.clone();
-            let bar_jwks = bar_jwks.clone();
+            let router = Arc::clone(&router);
             async move {
-                let jwks = match req.uri().path() {
-                    "/foo" => &foo_jwks,
-                    "/bar" => &bar_jwks,
-                    _ => {
-                        return Response::builder()
-                            .status(404)
-                            .body(Full::new(Bytes::new()));
-                    }
-                };
-                let body = serde_json::to_vec(jwks).unwrap();
-                Response::builder()
-                    .status(200)
-                    .body(Full::new(Bytes::from(body)))
+                match router(req.uri().path()) {
+                    Some(body) => Response::builder()
+                        .status(200)
+                        .body(Full::new(Bytes::from(body))),
+                    None => Response::builder()
+                        .status(404)
+                        .body(Full::new(Bytes::new())),
+                }
             }
         });
 
@@ -703,50 +869,257 @@ X0n5X2/pBLJzxZc62ccvZYVnctBiFs6HbSnxpuMQCfkt/BcR/ttIepBQQIW86wHL
             }
         });
 
-        let client = reqwest::Client::new();
+        addr
+    }
 
-        #[derive(Clone)]
-        struct Fetch(SocketAddr);
+    #[tokio::test]
+    async fn check_jwt_happy_path() {
+        let (rs1, jwk1) = new_rsa_jwk(RS1, "rs1".into());
+        let (rs2, jwk2) = new_rsa_jwk(RS2, "rs2".into());
+        let (ec1, jwk3) = new_ec_jwk("ec1".into());
+        let (ec2, jwk4) = new_ec_jwk("ec2".into());
 
-        impl FetchAuthRules for Fetch {
-            async fn fetch_auth_rules(
-                &self,
-                _ctx: &RequestMonitoring,
-                _endpoint: EndpointId,
-                _role_name: RoleName,
-            ) -> anyhow::Result<Vec<AuthRule>> {
-                Ok(vec![
-                    AuthRule {
-                        id: "foo".to_owned(),
-                        jwks_url: format!("http://{}/foo", self.0).parse().unwrap(),
-                        audience: None,
-                    },
-                    AuthRule {
-                        id: "bar".to_owned(),
-                        jwks_url: format!("http://{}/bar", self.0).parse().unwrap(),
-                        audience: None,
-                    },
-                ])
-            }
-        }
+        let foo_jwks = jose_jwk::JwkSet {
+            keys: vec![jwk1, jwk3],
+        };
+        let bar_jwks = jose_jwk::JwkSet {
+            keys: vec![jwk2, jwk4],
+        };
+
+        let jwks_addr = jwks_server(move |path| match path {
+            "/foo" => Some(serde_json::to_vec(&foo_jwks).unwrap()),
+            "/bar" => Some(serde_json::to_vec(&bar_jwks).unwrap()),
+            _ => None,
+        })
+        .await;
+
+        let role_name1 = RoleName::from("anonymous");
+        let role_name2 = RoleName::from("authenticated");
+
+        let roles = vec![
+            RoleNameInt::from(&role_name1),
+            RoleNameInt::from(&role_name2),
+        ];
+        let rules = vec![
+            AuthRule {
+                id: "foo".to_owned(),
+                jwks_url: format!("http://{jwks_addr}/foo").parse().unwrap(),
+                audience: None,
+                role_names: roles.clone(),
+            },
+            AuthRule {
+                id: "bar".to_owned(),
+                jwks_url: format!("http://{jwks_addr}/bar").parse().unwrap(),
+                audience: None,
+                role_names: roles.clone(),
+            },
+        ];
+
+        let fetch = Fetch(rules);
+        let jwk_cache = JwkCache::default();
 
-        let role_name = RoleName::from("user");
         let endpoint = EndpointId::from("ep");
 
-        let jwk_cache = Arc::new(JwkCacheEntryLock::default());
+        let jwt1 = new_rsa_jwt("rs1".into(), rs1);
+        let jwt2 = new_rsa_jwt("rs2".into(), rs2);
+        let jwt3 = new_ec_jwt("ec1".into(), &ec1);
+        let jwt4 = new_ec_jwt("ec2".into(), &ec2);
 
-        for token in [jwt1, jwt2, jwt3, jwt4] {
-            jwk_cache
-                .check_jwt(
-                    &RequestMonitoring::test(),
-                    &token,
-                    &client,
-                    endpoint.clone(),
-                    role_name.clone(),
-                    &Fetch(addr),
-                )
+        let tokens = [jwt1, jwt2, jwt3, jwt4];
+        let role_names = [role_name1, role_name2];
+        for role in &role_names {
+            for token in &tokens {
+                jwk_cache
+                    .check_jwt(
+                        &RequestMonitoring::test(),
+                        endpoint.clone(),
+                        role,
+                        &fetch,
+                        token,
+                    )
+                    .await
+                    .unwrap();
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn check_jwt_invalid_signature() {
+        let (_, jwk) = new_ec_jwk("1".into());
+        let (key, _) = new_ec_jwk("1".into());
+
+        // has a matching kid, but signed by the wrong key
+        let bad_jwt = new_ec_jwt("1".into(), &key);
+
+        let jwks = jose_jwk::JwkSet { keys: vec![jwk] };
+        let jwks_addr = jwks_server(move |path| match path {
+            "/" => Some(serde_json::to_vec(&jwks).unwrap()),
+            _ => None,
+        })
+        .await;
+
+        let role = RoleName::from("authenticated");
+
+        let rules = vec![AuthRule {
+            id: String::new(),
+            jwks_url: format!("http://{jwks_addr}/").parse().unwrap(),
+            audience: None,
+            role_names: vec![RoleNameInt::from(&role)],
+        }];
+
+        let fetch = Fetch(rules);
+        let jwk_cache = JwkCache::default();
+
+        let ep = EndpointId::from("ep");
+
+        let ctx = RequestMonitoring::test();
+        let err = jwk_cache
+            .check_jwt(&ctx, ep, &role, &fetch, &bad_jwt)
+            .await
+            .unwrap_err();
+        assert!(
+            err.to_string().contains("signature error"),
+            "expected \"signature error\", got {err:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn check_jwt_unknown_role() {
+        let (key, jwk) = new_rsa_jwk(RS1, "1".into());
+        let jwt = new_rsa_jwt("1".into(), key);
+
+        let jwks = jose_jwk::JwkSet { keys: vec![jwk] };
+        let jwks_addr = jwks_server(move |path| match path {
+            "/" => Some(serde_json::to_vec(&jwks).unwrap()),
+            _ => None,
+        })
+        .await;
+
+        let role = RoleName::from("authenticated");
+        let rules = vec![AuthRule {
+            id: String::new(),
+            jwks_url: format!("http://{jwks_addr}/").parse().unwrap(),
+            audience: None,
+            role_names: vec![RoleNameInt::from(&role)],
+        }];
+
+        let fetch = Fetch(rules);
+        let jwk_cache = JwkCache::default();
+
+        let ep = EndpointId::from("ep");
+
+        // this role_name is not accepted
+        let bad_role_name = RoleName::from("cloud_admin");
+
+        let ctx = RequestMonitoring::test();
+        let err = jwk_cache
+            .check_jwt(&ctx, ep, &bad_role_name, &fetch, &jwt)
+            .await
+            .unwrap_err();
+
+        assert!(
+            err.to_string().contains("jwk not found"),
+            "expected \"jwk not found\", got {err:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn check_jwt_invalid_claims() {
+        let (key, jwk) = new_ec_jwk("1".into());
+
+        let jwks = jose_jwk::JwkSet { keys: vec![jwk] };
+        let jwks_addr = jwks_server(move |path| match path {
+            "/" => Some(serde_json::to_vec(&jwks).unwrap()),
+            _ => None,
+        })
+        .await;
+
+        let now = SystemTime::now()
+            .duration_since(SystemTime::UNIX_EPOCH)
+            .unwrap()
+            .as_secs();
+
+        struct Test {
+            body: serde_json::Value,
+            error: &'static str,
+        }
+
+        let table = vec![
+            Test {
+                body: json! {{
+                    "nbf": now + 60,
+                    "aud": "neon",
+                }},
+                error: "JWT token is not yet ready to use",
+            },
+            Test {
+                body: json! {{
+                    "exp": now - 60,
+                    "aud": ["neon"],
+                }},
+                error: "JWT token has expired",
+            },
+            Test {
+                body: json! {{
+                }},
+                error: "invalid JWT token audience",
+            },
+            Test {
+                body: json! {{
+                    "aud": [],
+                }},
+                error: "invalid JWT token audience",
+            },
+            Test {
+                body: json! {{
+                    "aud": "foo",
+                }},
+                error: "invalid JWT token audience",
+            },
+            Test {
+                body: json! {{
+                    "aud": ["foo"],
+                }},
+                error: "invalid JWT token audience",
+            },
+            Test {
+                body: json! {{
+                    "aud": ["foo", "bar"],
+                }},
+                error: "invalid JWT token audience",
+            },
+        ];
+
+        let role = RoleName::from("authenticated");
+
+        let rules = vec![AuthRule {
+            id: String::new(),
+            jwks_url: format!("http://{jwks_addr}/").parse().unwrap(),
+            audience: Some("neon".to_string()),
+            role_names: vec![RoleNameInt::from(&role)],
+        }];
+
+        let fetch = Fetch(rules);
+        let jwk_cache = JwkCache::default();
+
+        let ep = EndpointId::from("ep");
+
+        let ctx = RequestMonitoring::test();
+        for test in table {
+            let jwt = new_custom_ec_jwt("1".into(), &key, test.body);
+
+            match jwk_cache
+                .check_jwt(&ctx, ep.clone(), &role, &fetch, &jwt)
                 .await
-                .unwrap();
+            {
+                Err(err) if err.to_string().contains(test.error) => {}
+                Err(err) => {
+                    panic!("expected {:?}, got {err:?}", test.error)
+                }
+                Ok(()) => {
+                    panic!("expected {:?}, got ok", test.error)
+                }
+            }
         }
     }
 }

proxy/src/auth/backend/local.rs

@@ -1,4 +1,4 @@
-use std::{collections::HashMap, net::SocketAddr};
+use std::net::SocketAddr;
 
 use anyhow::Context;
 use arc_swap::ArcSwapOption;
@@ -10,21 +10,19 @@ use crate::{
         NodeInfo,
     },
     context::RequestMonitoring,
-    intern::{BranchIdInt, BranchIdTag, EndpointIdTag, InternId, ProjectIdInt, ProjectIdTag},
-    EndpointId, RoleName,
+    intern::{BranchIdTag, EndpointIdTag, InternId, ProjectIdTag},
+    EndpointId,
 };
 
-use super::jwt::{AuthRule, FetchAuthRules, JwkCache};
+use super::jwt::{AuthRule, FetchAuthRules};
 
 pub struct LocalBackend {
-    pub(crate) jwks_cache: JwkCache,
     pub(crate) node_info: NodeInfo,
 }
 
 impl LocalBackend {
     pub fn new(postgres_addr: SocketAddr) -> Self {
         LocalBackend {
-            jwks_cache: JwkCache::default(),
             node_info: NodeInfo {
                 config: {
                     let mut cfg = ConnCfg::new();
@@ -48,26 +46,17 @@ impl LocalBackend {
 #[derive(Clone, Copy)]
 pub(crate) struct StaticAuthRules;
 
-pub static JWKS_ROLE_MAP: ArcSwapOption<JwksRoleSettings> = ArcSwapOption::const_empty();
-
-#[derive(Debug, Clone)]
-pub struct JwksRoleSettings {
-    pub roles: HashMap<RoleName, EndpointJwksResponse>,
-    pub project_id: ProjectIdInt,
-    pub branch_id: BranchIdInt,
-}
+pub static JWKS_ROLE_MAP: ArcSwapOption<EndpointJwksResponse> = ArcSwapOption::const_empty();
 
 impl FetchAuthRules for StaticAuthRules {
     async fn fetch_auth_rules(
         &self,
         _ctx: &RequestMonitoring,
         _endpoint: EndpointId,
-        role_name: RoleName,
     ) -> anyhow::Result<Vec<AuthRule>> {
         let mappings = JWKS_ROLE_MAP.load();
         let role_mappings = mappings
             .as_deref()
-            .and_then(|m| m.roles.get(&role_name))
             .context("JWKs settings for this role were not configured")?;
         let mut rules = vec![];
         for setting in &role_mappings.jwks {
@@ -75,6 +64,7 @@ impl FetchAuthRules for StaticAuthRules {
                 id: setting.id.clone(),
                 jwks_url: setting.jwks_url.clone(),
                 audience: setting.jwt_audience.clone(),
+                role_names: setting.role_names.clone(),
             });
         }
 

proxy/src/auth/backend/web.rs

@@ -1,5 +1,6 @@
 use crate::{
     auth, compute,
+    config::AuthenticationConfig,
     console::{self, provider::NodeInfo},
     context::RequestMonitoring,
     error::{ReportableError, UserFacingError},
@@ -58,6 +59,7 @@ pub(crate) fn new_psql_session_id() -> String {
 
 pub(super) async fn authenticate(
     ctx: &RequestMonitoring,
+    auth_config: &'static AuthenticationConfig,
     link_uri: &reqwest::Url,
     client: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
 ) -> auth::Result<NodeInfo> {
@@ -89,6 +91,14 @@ pub(super) async fn authenticate(
     info!(parent: &span, "waiting for console's reply...");
     let db_info = waiter.await.map_err(WebAuthError::from)?;
 
+    if auth_config.ip_allowlist_check_enabled {
+        if let Some(allowed_ips) = &db_info.allowed_ips {
+            if !auth::check_peer_addr_is_in_list(&ctx.peer_addr(), allowed_ips) {
+                return Err(auth::AuthError::ip_address_not_allowed(ctx.peer_addr()));
+            }
+        }
+    }
+
     client.write_message_noflush(&Be::NoticeResponse("Connecting to database."))?;
 
     // This config should be self-contained, because we won't

proxy/src/bin/local_proxy.rs

@@ -1,34 +1,38 @@
-use std::{
-    net::SocketAddr,
-    path::{Path, PathBuf},
-    pin::pin,
-    sync::Arc,
-    time::Duration,
-};
+use std::{net::SocketAddr, pin::pin, str::FromStr, sync::Arc, time::Duration};
 
-use anyhow::{bail, ensure};
+use anyhow::{bail, ensure, Context};
+use camino::{Utf8Path, Utf8PathBuf};
+use compute_api::spec::LocalProxySpec;
 use dashmap::DashMap;
-use futures::{future::Either, FutureExt};
+use futures::future::Either;
 use proxy::{
-    auth::backend::local::{JwksRoleSettings, LocalBackend, JWKS_ROLE_MAP},
+    auth::backend::{
+        jwt::JwkCache,
+        local::{LocalBackend, JWKS_ROLE_MAP},
+    },
     cancellation::CancellationHandlerMain,
     config::{self, AuthenticationConfig, HttpConfig, ProxyConfig, RetryConfig},
-    console::{locks::ApiLocks, messages::JwksRoleMapping},
+    console::{
+        locks::ApiLocks,
+        messages::{EndpointJwksResponse, JwksSettings},
+    },
     http::health_server::AppMetrics,
+    intern::RoleNameInt,
     metrics::{Metrics, ThreadPoolMetrics},
     rate_limiter::{BucketRateLimiter, EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo},
     scram::threadpool::ThreadPool,
     serverless::{self, cancel_set::CancelSet, GlobalConnPoolOptions},
+    RoleName,
 };
 
 project_git_version!(GIT_VERSION);
 project_build_tag!(BUILD_TAG);
 
 use clap::Parser;
-use tokio::{net::TcpListener, task::JoinSet};
+use tokio::{net::TcpListener, sync::Notify, task::JoinSet};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};
-use utils::{project_build_tag, project_git_version, sentry_init::init_sentry};
+use utils::{pid_file, project_build_tag, project_git_version, sentry_init::init_sentry};
 
 #[global_allocator]
 static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
@@ -72,9 +76,12 @@ struct LocalProxyCliArgs {
     /// Address of the postgres server
     #[clap(long, default_value = "127.0.0.1:5432")]
     compute: SocketAddr,
-    /// File address of the local proxy config file
+    /// Path of the local proxy config file
     #[clap(long, default_value = "./localproxy.json")]
-    config_path: PathBuf,
+    config_path: Utf8PathBuf,
+    /// Path of the local proxy PID file
+    #[clap(long, default_value = "./localproxy.pid")]
+    pid_path: Utf8PathBuf,
 }
 
 #[derive(clap::Args, Clone, Copy, Debug)]
@@ -126,6 +133,24 @@ async fn main() -> anyhow::Result<()> {
     let args = LocalProxyCliArgs::parse();
     let config = build_config(&args)?;
 
+    // before we bind to any ports, write the process ID to a file
+    // so that compute-ctl can find our process later
+    // in order to trigger the appropriate SIGHUP on config change.
+    //
+    // This also claims a "lock" that makes sure only one instance
+    // of local-proxy runs at a time.
+    let _process_guard = loop {
+        match pid_file::claim_for_current_process(&args.pid_path) {
+            Ok(guard) => break guard,
+            Err(e) => {
+                // compute-ctl might have tried to read the pid-file to let us
+                // know about some config change. We should try again.
+                error!(path=?args.pid_path, "could not claim PID file guard: {e:?}");
+                tokio::time::sleep(Duration::from_secs(1)).await;
+            }
+        }
+    };
+
     let metrics_listener = TcpListener::bind(args.metrics).await?.into_std()?;
     let http_listener = TcpListener::bind(args.http).await?;
     let shutdown = CancellationToken::new();
@@ -139,12 +164,30 @@ async fn main() -> anyhow::Result<()> {
         16,
     ));
 
-    refresh_config(args.config_path.clone()).await;
+    // write the process ID to a file so that compute-ctl can find our process later
+    // in order to trigger the appropriate SIGHUP on config change.
+    let pid = std::process::id();
+    info!("process running in PID {pid}");
+    std::fs::write(args.pid_path, format!("{pid}\n")).context("writing PID to file")?;
 
     let mut maintenance_tasks = JoinSet::new();
-    maintenance_tasks.spawn(proxy::handle_signals(shutdown.clone(), move || {
-        refresh_config(args.config_path.clone()).map(Ok)
+
+    let refresh_config_notify = Arc::new(Notify::new());
+    maintenance_tasks.spawn(proxy::handle_signals(shutdown.clone(), {
+        let refresh_config_notify = Arc::clone(&refresh_config_notify);
+        move || {
+            refresh_config_notify.notify_one();
+        }
     }));
+
+    // trigger the first config load **after** setting up the signal hook
+    // to avoid the race condition where:
+    // 1. No config file registered when local-proxy starts up
+    // 2. The config file is written but the signal hook is not yet received
+    // 3. local-proxy completes startup but has no config loaded, despite there being a registerd config.
+    refresh_config_notify.notify_one();
+    tokio::spawn(refresh_config_loop(args.config_path, refresh_config_notify));
+
     maintenance_tasks.spawn(proxy::http::health_server::task_main(
         metrics_listener,
         AppMetrics {
@@ -227,12 +270,15 @@ fn build_config(args: &LocalProxyCliArgs) -> anyhow::Result<&'static ProxyConfig
         allow_self_signed_compute: false,
         http_config,
         authentication_config: AuthenticationConfig {
+            jwks_cache: JwkCache::default(),
             thread_pool: ThreadPool::new(0),
             scram_protocol_timeout: Duration::from_secs(10),
             rate_limiter_enabled: false,
             rate_limiter: BucketRateLimiter::new(vec![]),
             rate_limit_ip_subnet: 64,
             ip_allowlist_check_enabled: true,
+            is_auth_broker: false,
+            accept_jwts: true,
         },
         require_client_ip: false,
         handshake_timeout: Duration::from_secs(10),
@@ -245,81 +291,84 @@ fn build_config(args: &LocalProxyCliArgs) -> anyhow::Result<&'static ProxyConfig
     })))
 }
 
-async fn refresh_config(path: PathBuf) {
-    match refresh_config_inner(&path).await {
-        Ok(()) => {}
-        Err(e) => {
-            error!(error=?e, ?path, "could not read config file");
+async fn refresh_config_loop(path: Utf8PathBuf, rx: Arc<Notify>) {
+    loop {
+        rx.notified().await;
+
+        match refresh_config_inner(&path).await {
+            Ok(()) => {}
+            Err(e) => {
+                error!(error=?e, ?path, "could not read config file");
+            }
         }
     }
 }
 
-async fn refresh_config_inner(path: &Path) -> anyhow::Result<()> {
+async fn refresh_config_inner(path: &Utf8Path) -> anyhow::Result<()> {
     let bytes = tokio::fs::read(&path).await?;
-    let mut data: JwksRoleMapping = serde_json::from_slice(&bytes)?;
+    let data: LocalProxySpec = serde_json::from_slice(&bytes)?;
 
-    let mut settings = None;
+    let mut jwks_set = vec![];
 
-    for mapping in data.roles.values_mut() {
-        for jwks in &mut mapping.jwks {
-            ensure!(
-                jwks.jwks_url.has_authority()
-                    && (jwks.jwks_url.scheme() == "http" || jwks.jwks_url.scheme() == "https"),
-                "Invalid JWKS url. Must be HTTP",
-            );
+    for jwks in data.jwks {
+        let mut jwks_url = url::Url::from_str(&jwks.jwks_url).context("parsing JWKS url")?;
 
-            ensure!(
-                jwks.jwks_url
-                    .host()
-                    .is_some_and(|h| h != url::Host::Domain("")),
-                "Invalid JWKS url. No domain listed",
-            );
+        ensure!(
+            jwks_url.has_authority()
+                && (jwks_url.scheme() == "http" || jwks_url.scheme() == "https"),
+            "Invalid JWKS url. Must be HTTP",
+        );
 
-            // clear username, password and ports
-            jwks.jwks_url.set_username("").expect(
+        ensure!(
+            jwks_url.host().is_some_and(|h| h != url::Host::Domain("")),
+            "Invalid JWKS url. No domain listed",
+        );
+
+        // clear username, password and ports
+        jwks_url
+            .set_username("")
+            .expect("url can be a base and has a valid host and is not a file. should not error");
+        jwks_url
+            .set_password(None)
+            .expect("url can be a base and has a valid host and is not a file. should not error");
+        // local testing is hard if we need to have a specific restricted port
+        if cfg!(not(feature = "testing")) {
+            jwks_url.set_port(None).expect(
                 "url can be a base and has a valid host and is not a file. should not error",
             );
-            jwks.jwks_url.set_password(None).expect(
-                "url can be a base and has a valid host and is not a file. should not error",
-            );
-            // local testing is hard if we need to have a specific restricted port
-            if cfg!(not(feature = "testing")) {
-                jwks.jwks_url.set_port(None).expect(
-                    "url can be a base and has a valid host and is not a file. should not error",
-                );
-            }
-
-            // clear query params
-            jwks.jwks_url.set_fragment(None);
-            jwks.jwks_url.query_pairs_mut().clear().finish();
-
-            if jwks.jwks_url.scheme() != "https" {
-                // local testing is hard if we need to set up https support.
-                if cfg!(not(feature = "testing")) {
-                    jwks.jwks_url
-                        .set_scheme("https")
-                        .expect("should not error to set the scheme to https if it was http");
-                } else {
-                    warn!(scheme = jwks.jwks_url.scheme(), "JWKS url is not HTTPS");
-                }
-            }
-
-            let (pr, br) = settings.get_or_insert((jwks.project_id, jwks.branch_id));
-            ensure!(
-                *pr == jwks.project_id,
-                "inconsistent project IDs configured"
-            );
-            ensure!(*br == jwks.branch_id, "inconsistent branch IDs configured");
         }
+
+        // clear query params
+        jwks_url.set_fragment(None);
+        jwks_url.query_pairs_mut().clear().finish();
+
+        if jwks_url.scheme() != "https" {
+            // local testing is hard if we need to set up https support.
+            if cfg!(not(feature = "testing")) {
+                jwks_url
+                    .set_scheme("https")
+                    .expect("should not error to set the scheme to https if it was http");
+            } else {
+                warn!(scheme = jwks_url.scheme(), "JWKS url is not HTTPS");
+            }
+        }
+
+        jwks_set.push(JwksSettings {
+            id: jwks.id,
+            jwks_url,
+            provider_name: jwks.provider_name,
+            jwt_audience: jwks.jwt_audience,
+            role_names: jwks
+                .role_names
+                .into_iter()
+                .map(RoleName::from)
+                .map(|s| RoleNameInt::from(&s))
+                .collect(),
+        })
     }
 
-    if let Some((project_id, branch_id)) = settings {
-        JWKS_ROLE_MAP.store(Some(Arc::new(JwksRoleSettings {
-            roles: data.roles,
-            project_id,
-            branch_id,
-        })));
-    }
+    info!("successfully loaded new config");
+    JWKS_ROLE_MAP.store(Some(Arc::new(EndpointJwksResponse { jwks: jwks_set })));
 
     Ok(())
 }

proxy/src/bin/pg_sni_router.rs

@@ -133,9 +133,7 @@ async fn main() -> anyhow::Result<()> {
         proxy_listener,
         cancellation_token.clone(),
     ));
-    let signals_task = tokio::spawn(proxy::handle_signals(cancellation_token, || async {
-        Ok(())
-    }));
+    let signals_task = tokio::spawn(proxy::handle_signals(cancellation_token, || {}));
 
     // the signal task cant ever succeed.
     // the main task can error, or can succeed on cancellation.

proxy/src/bin/proxy.rs

@@ -8,6 +8,7 @@ use aws_config::web_identity_token::WebIdentityTokenCredentialsProvider;
 use aws_config::Region;
 use futures::future::Either;
 use proxy::auth;
+use proxy::auth::backend::jwt::JwkCache;
 use proxy::auth::backend::AuthRateLimiter;
 use proxy::auth::backend::MaybeOwned;
 use proxy::cancellation::CancelMap;
@@ -102,6 +103,9 @@ struct ProxyCliArgs {
         default_value = "http://localhost:3000/authenticate_proxy_request/"
     )]
     auth_endpoint: String,
+    /// if this is not local proxy, this toggles whether we accept jwt or passwords for http
+    #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
+    is_auth_broker: bool,
     /// path to TLS key for client postgres connections
     ///
     /// tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
@@ -382,9 +386,27 @@ async fn main() -> anyhow::Result<()> {
     info!("Starting mgmt on {mgmt_address}");
     let mgmt_listener = TcpListener::bind(mgmt_address).await?;
 
-    let proxy_address: SocketAddr = args.proxy.parse()?;
-    info!("Starting proxy on {proxy_address}");
-    let proxy_listener = TcpListener::bind(proxy_address).await?;
+    let proxy_listener = if !args.is_auth_broker {
+        let proxy_address: SocketAddr = args.proxy.parse()?;
+        info!("Starting proxy on {proxy_address}");
+
+        Some(TcpListener::bind(proxy_address).await?)
+    } else {
+        None
+    };
+
+    // TODO: rename the argument to something like serverless.
+    // It now covers more than just websockets, it also covers SQL over HTTP.
+    let serverless_listener = if let Some(serverless_address) = args.wss {
+        let serverless_address: SocketAddr = serverless_address.parse()?;
+        info!("Starting wss on {serverless_address}");
+        Some(TcpListener::bind(serverless_address).await?)
+    } else if args.is_auth_broker {
+        bail!("wss arg must be present for auth-broker")
+    } else {
+        None
+    };
+
     let cancellation_token = CancellationToken::new();
 
     let cancel_map = CancelMap::default();
@@ -430,21 +452,17 @@ async fn main() -> anyhow::Result<()> {
     // client facing tasks. these will exit on error or on cancellation
     // cancellation returns Ok(())
     let mut client_tasks = JoinSet::new();
-    client_tasks.spawn(proxy::proxy::task_main(
-        config,
-        proxy_listener,
-        cancellation_token.clone(),
-        cancellation_handler.clone(),
-        endpoint_rate_limiter.clone(),
-    ));
-
-    // TODO: rename the argument to something like serverless.
-    // It now covers more than just websockets, it also covers SQL over HTTP.
-    if let Some(serverless_address) = args.wss {
-        let serverless_address: SocketAddr = serverless_address.parse()?;
-        info!("Starting wss on {serverless_address}");
-        let serverless_listener = TcpListener::bind(serverless_address).await?;
+    if let Some(proxy_listener) = proxy_listener {
+        client_tasks.spawn(proxy::proxy::task_main(
+            config,
+            proxy_listener,
+            cancellation_token.clone(),
+            cancellation_handler.clone(),
+            endpoint_rate_limiter.clone(),
+        ));
+    }
 
+    if let Some(serverless_listener) = serverless_listener {
         client_tasks.spawn(serverless::task_main(
             config,
             serverless_listener,
@@ -461,10 +479,7 @@ async fn main() -> anyhow::Result<()> {
 
     // maintenance tasks. these never return unless there's an error
     let mut maintenance_tasks = JoinSet::new();
-    maintenance_tasks.spawn(proxy::handle_signals(
-        cancellation_token.clone(),
-        || async { Ok(()) },
-    ));
+    maintenance_tasks.spawn(proxy::handle_signals(cancellation_token.clone(), || {}));
     maintenance_tasks.spawn(http::health_server::task_main(
         http_listener,
         AppMetrics {
@@ -677,7 +692,7 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
     )?;
 
     let http_config = HttpConfig {
-        accept_websockets: true,
+        accept_websockets: !args.is_auth_broker,
         pool_options: GlobalConnPoolOptions {
             max_conns_per_endpoint: args.sql_over_http.sql_over_http_pool_max_conns_per_endpoint,
             gc_epoch: args.sql_over_http.sql_over_http_pool_gc_epoch,
@@ -692,12 +707,15 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
         max_response_size_bytes: args.sql_over_http.sql_over_http_max_response_size_bytes,
     };
     let authentication_config = AuthenticationConfig {
+        jwks_cache: JwkCache::default(),
         thread_pool,
         scram_protocol_timeout: args.scram_protocol_timeout,
         rate_limiter_enabled: args.auth_rate_limit_enabled,
         rate_limiter: AuthRateLimiter::new(args.auth_rate_limit.clone()),
         rate_limit_ip_subnet: args.auth_rate_limit_ip_subnet,
         ip_allowlist_check_enabled: !args.is_private_access_proxy,
+        is_auth_broker: args.is_auth_broker,
+        accept_jwts: args.is_auth_broker,
     };
 
     let config = Box::leak(Box::new(ProxyConfig {

proxy/src/config.rs

@@ -1,5 +1,8 @@
 use crate::{
-    auth::{self, backend::AuthRateLimiter},
+    auth::{
+        self,
+        backend::{jwt::JwkCache, AuthRateLimiter},
+    },
     console::locks::ApiLocks,
     rate_limiter::{RateBucketInfo, RateLimitAlgorithm, RateLimiterConfig},
     scram::threadpool::ThreadPool,
@@ -67,6 +70,9 @@ pub struct AuthenticationConfig {
     pub rate_limiter: AuthRateLimiter,
     pub rate_limit_ip_subnet: u8,
     pub ip_allowlist_check_enabled: bool,
+    pub jwks_cache: JwkCache,
+    pub is_auth_broker: bool,
+    pub accept_jwts: bool,
 }
 
 impl TlsConfig {
@@ -250,18 +256,26 @@ impl CertResolver {
 
         let common_name = pem.subject().to_string();
 
-        // We only use non-wildcard certificates in web auth proxy so it seems okay to treat them the same as
-        // wildcard ones as we don't use SNI there. That treatment only affects certificate selection, so
-        // verify-full will still check wildcard match. Old coding here just ignored non-wildcard common names
-        // and passed None instead, which blows up number of cases downstream code should handle. Proper coding
-        // here should better avoid Option for common_names, and do wildcard-based certificate selection instead
-        // of cutting off '*.' parts.
-        let common_name = if common_name.starts_with("CN=*.") {
-            common_name.strip_prefix("CN=*.").map(|s| s.to_string())
+        // We need to get the canonical name for this certificate so we can match them against any domain names
+        // seen within the proxy codebase.
+        //
+        // In scram-proxy we use wildcard certificates only, with the database endpoint as the wildcard subdomain, taken from SNI.
+        // We need to remove the wildcard prefix for the purposes of certificate selection.
+        //
+        // auth-broker does not use SNI and instead uses the Neon-Connection-String header.
+        // Auth broker has the subdomain `apiauth` we need to remove for the purposes of validating the Neon-Connection-String.
+        //
+        // Console Web proxy does not use any wildcard domains and does not need any certificate selection or conn string
+        // validation, so let's we can continue with any common-name
+        let common_name = if let Some(s) = common_name.strip_prefix("CN=*.") {
+            s.to_string()
+        } else if let Some(s) = common_name.strip_prefix("CN=apiauth.") {
+            s.to_string()
+        } else if let Some(s) = common_name.strip_prefix("CN=") {
+            s.to_string()
         } else {
-            common_name.strip_prefix("CN=").map(|s| s.to_string())
-        }
-        .context("Failed to parse common name from certificate")?;
+            bail!("Failed to parse common name from certificate")
+        };
 
         let cert = Arc::new(rustls::sign::CertifiedKey::new(cert_chain, key));
 

proxy/src/console/messages.rs

@@ -1,13 +1,11 @@
 use measured::FixedCardinalityLabel;
 use serde::{Deserialize, Serialize};
-use std::collections::HashMap;
 use std::fmt::{self, Display};
 
 use crate::auth::IpPattern;
 
-use crate::intern::{BranchIdInt, EndpointIdInt, ProjectIdInt};
+use crate::intern::{BranchIdInt, EndpointIdInt, ProjectIdInt, RoleNameInt};
 use crate::proxy::retry::CouldRetry;
-use crate::RoleName;
 
 /// Generic error response with human-readable description.
 /// Note that we can't always present it to user as is.
@@ -284,6 +282,8 @@ pub(crate) struct DatabaseInfo {
     /// be inconvenient for debug with local PG instance.
     pub(crate) password: Option<Box<str>>,
     pub(crate) aux: MetricsAuxInfo,
+    #[serde(default)]
+    pub(crate) allowed_ips: Option<Vec<IpPattern>>,
 }
 
 // Manually implement debug to omit sensitive info.
@@ -294,6 +294,7 @@ impl fmt::Debug for DatabaseInfo {
             .field("port", &self.port)
             .field("dbname", &self.dbname)
             .field("user", &self.user)
+            .field("allowed_ips", &self.allowed_ips)
             .finish_non_exhaustive()
     }
 }
@@ -345,11 +346,6 @@ impl ColdStartInfo {
     }
 }
 
-#[derive(Debug, Deserialize, Clone)]
-pub struct JwksRoleMapping {
-    pub roles: HashMap<RoleName, EndpointJwksResponse>,
-}
-
 #[derive(Debug, Deserialize, Clone)]
 pub struct EndpointJwksResponse {
     pub jwks: Vec<JwksSettings>,
@@ -358,11 +354,10 @@ pub struct EndpointJwksResponse {
 #[derive(Debug, Deserialize, Clone)]
 pub struct JwksSettings {
     pub id: String,
-    pub project_id: ProjectIdInt,
-    pub branch_id: BranchIdInt,
     pub jwks_url: url::Url,
     pub provider_name: String,
     pub jwt_audience: Option<String>,
+    pub role_names: Vec<RoleNameInt>,
 }
 
 #[cfg(test)]
@@ -432,6 +427,22 @@ mod tests {
             "aux": dummy_aux(),
         }))?;
 
+        // with allowed_ips
+        let dbinfo = serde_json::from_value::<DatabaseInfo>(json!({
+            "host": "localhost",
+            "port": 5432,
+            "dbname": "postgres",
+            "user": "john_doe",
+            "password": "password",
+            "aux": dummy_aux(),
+            "allowed_ips": ["127.0.0.1"],
+        }))?;
+
+        assert_eq!(
+            dbinfo.allowed_ips,
+            Some(vec![IpPattern::Single("127.0.0.1".parse()?)])
+        );
+
         Ok(())
     }
 

proxy/src/console/provider.rs

@@ -5,7 +5,10 @@ pub mod neon;
 use super::messages::{ConsoleError, MetricsAuxInfo};
 use crate::{
     auth::{
-        backend::{ComputeCredentialKeys, ComputeUserInfo},
+        backend::{
+            jwt::{AuthRule, FetchAuthRules},
+            ComputeCredentialKeys, ComputeUserInfo,
+        },
         IpPattern,
     },
     cache::{endpoints::EndpointsCache, project_info::ProjectInfoCacheImpl, Cached, TimedLru},
@@ -16,7 +19,7 @@ use crate::{
     intern::ProjectIdInt,
     metrics::ApiLockMetrics,
     rate_limiter::{DynamicLimiter, Outcome, RateLimiterConfig, Token},
-    scram, EndpointCacheKey,
+    scram, EndpointCacheKey, EndpointId,
 };
 use dashmap::DashMap;
 use std::{hash::Hash, sync::Arc, time::Duration};
@@ -334,6 +337,12 @@ pub(crate) trait Api {
         user_info: &ComputeUserInfo,
     ) -> Result<(CachedAllowedIps, Option<CachedRoleSecret>), errors::GetAuthInfoError>;
 
+    async fn get_endpoint_jwks(
+        &self,
+        ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>>;
+
     /// Wake up the compute node and return the corresponding connection info.
     async fn wake_compute(
         &self,
@@ -343,6 +352,7 @@ pub(crate) trait Api {
 }
 
 #[non_exhaustive]
+#[derive(Clone)]
 pub enum ConsoleBackend {
     /// Current Cloud API (V2).
     Console(neon::Api),
@@ -386,6 +396,20 @@ impl Api for ConsoleBackend {
         }
     }
 
+    async fn get_endpoint_jwks(
+        &self,
+        ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>> {
+        match self {
+            Self::Console(api) => api.get_endpoint_jwks(ctx, endpoint).await,
+            #[cfg(any(test, feature = "testing"))]
+            Self::Postgres(api) => api.get_endpoint_jwks(ctx, endpoint).await,
+            #[cfg(test)]
+            Self::Test(_api) => Ok(vec![]),
+        }
+    }
+
     async fn wake_compute(
         &self,
         ctx: &RequestMonitoring,
@@ -552,3 +576,13 @@ impl WakeComputePermit {
         res
     }
 }
+
+impl FetchAuthRules for ConsoleBackend {
+    async fn fetch_auth_rules(
+        &self,
+        ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>> {
+        self.get_endpoint_jwks(ctx, endpoint).await
+    }
+}

proxy/src/console/provider/mock.rs

@@ -4,7 +4,9 @@ use super::{
     errors::{ApiError, GetAuthInfoError, WakeComputeError},
     AuthInfo, AuthSecret, CachedNodeInfo, NodeInfo,
 };
-use crate::context::RequestMonitoring;
+use crate::{
+    auth::backend::jwt::AuthRule, context::RequestMonitoring, intern::RoleNameInt, RoleName,
+};
 use crate::{auth::backend::ComputeUserInfo, compute, error::io_error, scram, url::ApiUrl};
 use crate::{auth::IpPattern, cache::Cached};
 use crate::{
@@ -118,6 +120,39 @@ impl Api {
         })
     }
 
+    async fn do_get_endpoint_jwks(&self, endpoint: EndpointId) -> anyhow::Result<Vec<AuthRule>> {
+        let (client, connection) =
+            tokio_postgres::connect(self.endpoint.as_str(), tokio_postgres::NoTls).await?;
+
+        let connection = tokio::spawn(connection);
+
+        let res = client.query(
+                "select id, jwks_url, audience, role_names from neon_control_plane.endpoint_jwks where endpoint_id = $1",
+                &[&endpoint.as_str()],
+            )
+            .await?;
+
+        let mut rows = vec![];
+        for row in res {
+            rows.push(AuthRule {
+                id: row.get("id"),
+                jwks_url: url::Url::parse(row.get("jwks_url"))?,
+                audience: row.get("audience"),
+                role_names: row
+                    .get::<_, Vec<String>>("role_names")
+                    .into_iter()
+                    .map(RoleName::from)
+                    .map(|s| RoleNameInt::from(&s))
+                    .collect(),
+            });
+        }
+
+        drop(client);
+        connection.await??;
+
+        Ok(rows)
+    }
+
     async fn do_wake_compute(&self) -> Result<NodeInfo, WakeComputeError> {
         let mut config = compute::ConnCfg::new();
         config
@@ -185,6 +220,14 @@ impl super::Api for Api {
         ))
     }
 
+    async fn get_endpoint_jwks(
+        &self,
+        _ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>> {
+        self.do_get_endpoint_jwks(endpoint).await
+    }
+
     #[tracing::instrument(skip_all)]
     async fn wake_compute(
         &self,

proxy/src/console/provider/neon.rs

@@ -7,27 +7,33 @@ use super::{
     NodeInfo,
 };
 use crate::{
-    auth::backend::ComputeUserInfo,
+    auth::backend::{jwt::AuthRule, ComputeUserInfo},
     compute,
-    console::messages::{ColdStartInfo, Reason},
+    console::messages::{ColdStartInfo, EndpointJwksResponse, Reason},
     http,
     metrics::{CacheOutcome, Metrics},
     rate_limiter::WakeComputeRateLimiter,
-    scram, EndpointCacheKey,
+    scram, EndpointCacheKey, EndpointId,
 };
 use crate::{cache::Cached, context::RequestMonitoring};
+use ::http::{header::AUTHORIZATION, HeaderName};
+use anyhow::bail;
 use futures::TryFutureExt;
 use std::{sync::Arc, time::Duration};
 use tokio::time::Instant;
 use tokio_postgres::config::SslMode;
 use tracing::{debug, error, info, info_span, warn, Instrument};
 
+const X_REQUEST_ID: HeaderName = HeaderName::from_static("x-request-id");
+
+#[derive(Clone)]
 pub struct Api {
     endpoint: http::Endpoint,
     pub caches: &'static ApiCaches,
     pub(crate) locks: &'static ApiLocks<EndpointCacheKey>,
     pub(crate) wake_compute_endpoint_rate_limiter: Arc<WakeComputeRateLimiter>,
-    jwt: String,
+    // put in a shared ref so we don't copy secrets all over in memory
+    jwt: Arc<str>,
 }
 
 impl Api {
@@ -38,7 +44,9 @@ impl Api {
         locks: &'static ApiLocks<EndpointCacheKey>,
         wake_compute_endpoint_rate_limiter: Arc<WakeComputeRateLimiter>,
     ) -> Self {
-        let jwt = std::env::var("NEON_PROXY_TO_CONTROLPLANE_TOKEN").unwrap_or_default();
+        let jwt = std::env::var("NEON_PROXY_TO_CONTROLPLANE_TOKEN")
+            .unwrap_or_default()
+            .into();
         Self {
             endpoint,
             caches,
@@ -71,9 +79,9 @@ impl Api {
         async {
             let request = self
                 .endpoint
-                .get("proxy_get_role_secret")
-                .header("X-Request-ID", &request_id)
-                .header("Authorization", format!("Bearer {}", &self.jwt))
+                .get_path("proxy_get_role_secret")
+                .header(X_REQUEST_ID, &request_id)
+                .header(AUTHORIZATION, format!("Bearer {}", &self.jwt))
                 .query(&[("session_id", ctx.session_id())])
                 .query(&[
                     ("application_name", application_name.as_str()),
@@ -125,6 +133,61 @@ impl Api {
         .await
     }
 
+    async fn do_get_endpoint_jwks(
+        &self,
+        ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>> {
+        if !self
+            .caches
+            .endpoints_cache
+            .is_valid(ctx, &endpoint.normalize())
+            .await
+        {
+            bail!("endpoint not found");
+        }
+        let request_id = ctx.session_id().to_string();
+        async {
+            let request = self
+                .endpoint
+                .get_with_url(|url| {
+                    url.path_segments_mut()
+                        .push("endpoints")
+                        .push(endpoint.as_str())
+                        .push("jwks");
+                })
+                .header(X_REQUEST_ID, &request_id)
+                .header(AUTHORIZATION, format!("Bearer {}", &self.jwt))
+                .query(&[("session_id", ctx.session_id())])
+                .build()?;
+
+            info!(url = request.url().as_str(), "sending http request");
+            let start = Instant::now();
+            let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Cplane);
+            let response = self.endpoint.execute(request).await?;
+            drop(pause);
+            info!(duration = ?start.elapsed(), "received http response");
+
+            let body = parse_body::<EndpointJwksResponse>(response).await?;
+
+            let rules = body
+                .jwks
+                .into_iter()
+                .map(|jwks| AuthRule {
+                    id: jwks.id,
+                    jwks_url: jwks.jwks_url,
+                    audience: jwks.jwt_audience,
+                    role_names: jwks.role_names,
+                })
+                .collect();
+
+            Ok(rules)
+        }
+        .map_err(crate::error::log_error)
+        .instrument(info_span!("http", id = request_id))
+        .await
+    }
+
     async fn do_wake_compute(
         &self,
         ctx: &RequestMonitoring,
@@ -135,7 +198,7 @@ impl Api {
         async {
             let mut request_builder = self
                 .endpoint
-                .get("proxy_wake_compute")
+                .get_path("proxy_wake_compute")
                 .header("X-Request-ID", &request_id)
                 .header("Authorization", format!("Bearer {}", &self.jwt))
                 .query(&[("session_id", ctx.session_id())])
@@ -262,6 +325,15 @@ impl super::Api for Api {
         ))
     }
 
+    #[tracing::instrument(skip_all)]
+    async fn get_endpoint_jwks(
+        &self,
+        ctx: &RequestMonitoring,
+        endpoint: EndpointId,
+    ) -> anyhow::Result<Vec<AuthRule>> {
+        self.do_get_endpoint_jwks(ctx, endpoint).await
+    }
+
     #[tracing::instrument(skip_all)]
     async fn wake_compute(
         &self,

proxy/src/http.rs

@@ -86,9 +86,17 @@ impl Endpoint {
 
     /// Return a [builder](RequestBuilder) for a `GET` request,
     /// appending a single `path` segment to the base endpoint URL.
-    pub(crate) fn get(&self, path: &str) -> RequestBuilder {
+    pub(crate) fn get_path(&self, path: &str) -> RequestBuilder {
+        self.get_with_url(|u| {
+            u.path_segments_mut().push(path);
+        })
+    }
+
+    /// Return a [builder](RequestBuilder) for a `GET` request,
+    /// accepting a closure to modify the url path segments for more complex paths queries.
+    pub(crate) fn get_with_url(&self, f: impl for<'a> FnOnce(&'a mut ApiUrl)) -> RequestBuilder {
         let mut url = self.endpoint.clone();
-        url.path_segments_mut().push(path);
+        f(&mut url);
         self.client.get(url.into_inner())
     }
 
@@ -144,7 +152,7 @@ mod tests {
 
         // Validate that this pattern makes sense.
         let req = endpoint
-            .get("frobnicate")
+            .get_path("frobnicate")
             .query(&[
                 ("foo", Some("10")), // should be just `foo=10`
                 ("bar", None),       // shouldn't be passed at all
@@ -162,7 +170,7 @@ mod tests {
         let endpoint = Endpoint::new(url, Client::new());
 
         let req = endpoint
-            .get("frobnicate")
+            .get_path("frobnicate")
             .query(&[("session_id", uuid::Uuid::nil())])
             .build()?;
 

proxy/src/intern.rs

@@ -1,5 +1,6 @@
 use std::{
-    hash::BuildHasherDefault, marker::PhantomData, num::NonZeroUsize, ops::Index, sync::OnceLock,
+    any::type_name, hash::BuildHasherDefault, marker::PhantomData, num::NonZeroUsize, ops::Index,
+    sync::OnceLock,
 };
 
 use lasso::{Capacity, MemoryLimits, Spur, ThreadedRodeo};
@@ -16,12 +17,21 @@ pub struct StringInterner<Id> {
     _id: PhantomData<Id>,
 }
 
-#[derive(PartialEq, Debug, Clone, Copy, Eq, Hash)]
+#[derive(PartialEq, Clone, Copy, Eq, Hash)]
 pub struct InternedString<Id> {
     inner: Spur,
     _id: PhantomData<Id>,
 }
 
+impl<Id: InternId> std::fmt::Debug for InternedString<Id> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_tuple("InternedString")
+            .field(&type_name::<Id>())
+            .field(&self.as_str())
+            .finish()
+    }
+}
+
 impl<Id: InternId> std::fmt::Display for InternedString<Id> {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         self.as_str().fmt(f)
@@ -130,14 +140,14 @@ impl<Id: InternId> Default for StringInterner<Id> {
 }
 
 #[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)]
-pub(crate) struct RoleNameTag;
+pub struct RoleNameTag;
 impl InternId for RoleNameTag {
     fn get_interner() -> &'static StringInterner<Self> {
         static ROLE_NAMES: OnceLock<StringInterner<RoleNameTag>> = OnceLock::new();
         ROLE_NAMES.get_or_init(Default::default)
     }
 }
-pub(crate) type RoleNameInt = InternedString<RoleNameTag>;
+pub type RoleNameInt = InternedString<RoleNameTag>;
 impl From<&RoleName> for RoleNameInt {
     fn from(value: &RoleName) -> Self {
         RoleNameTag::get_interner().get_or_intern(value)

proxy/src/lib.rs

@@ -82,7 +82,7 @@
     impl_trait_overcaptures,
 )]
 
-use std::{convert::Infallible, future::Future};
+use std::convert::Infallible;
 
 use anyhow::{bail, Context};
 use intern::{EndpointIdInt, EndpointIdTag, InternId};
@@ -117,13 +117,12 @@ pub mod usage_metrics;
 pub mod waiters;
 
 /// Handle unix signals appropriately.
-pub async fn handle_signals<F, Fut>(
+pub async fn handle_signals<F>(
     token: CancellationToken,
     mut refresh_config: F,
 ) -> anyhow::Result<Infallible>
 where
-    F: FnMut() -> Fut,
-    Fut: Future<Output = anyhow::Result<()>>,
+    F: FnMut(),
 {
     use tokio::signal::unix::{signal, SignalKind};
 
@@ -136,7 +135,7 @@ where
             // Hangup is commonly used for config reload.
             _ = hangup.recv() => {
                 warn!("received SIGHUP");
-                refresh_config().await?;
+                refresh_config();
             }
             // Shut down the whole application.
             _ = interrupt.recv() => {

proxy/src/proxy/tests.rs

@@ -525,6 +525,10 @@ impl TestBackend for TestConnectMechanism {
     {
         unimplemented!("not used in tests")
     }
+
+    fn dyn_clone(&self) -> Box<dyn TestBackend> {
+        Box::new(self.clone())
+    }
 }
 
 fn helper_create_cached_node_info(cache: &'static NodeInfoCache) -> CachedNodeInfo {

proxy/src/scram/threadpool.rs

@@ -43,6 +43,13 @@ impl ThreadPool {
     pub fn new(n_workers: u8) -> Arc<Self> {
         // rayon would be nice here, but yielding in rayon does not work well afaict.
 
+        if n_workers == 0 {
+            return Arc::new(Self {
+                runtime: None,
+                metrics: Arc::new(ThreadPoolMetrics::new(n_workers as usize)),
+            });
+        }
+
         Arc::new_cyclic(|pool| {
             let pool = pool.clone();
             let worker_id = AtomicUsize::new(0);

proxy/src/serverless.rs

@@ -5,6 +5,7 @@
 mod backend;
 pub mod cancel_set;
 mod conn_pool;
+mod http_conn_pool;
 mod http_util;
 mod json;
 mod sql_over_http;
@@ -19,7 +20,8 @@ use anyhow::Context;
 use futures::future::{select, Either};
 use futures::TryFutureExt;
 use http::{Method, Response, StatusCode};
-use http_body_util::Full;
+use http_body_util::combinators::BoxBody;
+use http_body_util::{BodyExt, Empty};
 use hyper1::body::Incoming;
 use hyper_util::rt::TokioExecutor;
 use hyper_util::server::conn::auto::Builder;
@@ -81,7 +83,28 @@ pub async fn task_main(
         }
     });
 
+    let http_conn_pool = http_conn_pool::GlobalConnPool::new(&config.http_config);
+    {
+        let http_conn_pool = Arc::clone(&http_conn_pool);
+        tokio::spawn(async move {
+            http_conn_pool.gc_worker(StdRng::from_entropy()).await;
+        });
+    }
+
+    // shutdown the connection pool
+    tokio::spawn({
+        let cancellation_token = cancellation_token.clone();
+        let http_conn_pool = http_conn_pool.clone();
+        async move {
+            cancellation_token.cancelled().await;
+            tokio::task::spawn_blocking(move || http_conn_pool.shutdown())
+                .await
+                .unwrap();
+        }
+    });
+
     let backend = Arc::new(PoolingBackend {
+        http_conn_pool: Arc::clone(&http_conn_pool),
         pool: Arc::clone(&conn_pool),
         config,
         endpoint_rate_limiter: Arc::clone(&endpoint_rate_limiter),
@@ -342,7 +365,7 @@ async fn request_handler(
     // used to cancel in-flight HTTP requests. not used to cancel websockets
     http_cancellation_token: CancellationToken,
     endpoint_rate_limiter: Arc<EndpointRateLimiter>,
-) -> Result<Response<Full<Bytes>>, ApiError> {
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, ApiError> {
     let host = request
         .headers()
         .get("host")
@@ -386,7 +409,7 @@ async fn request_handler(
         );
 
         // Return the response so the spawned future can continue.
-        Ok(response.map(|_: http_body_util::Empty<Bytes>| Full::new(Bytes::new())))
+        Ok(response.map(|b| b.map_err(|x| match x {}).boxed()))
     } else if request.uri().path() == "/sql" && *request.method() == Method::POST {
         let ctx = RequestMonitoring::new(
             session_id,
@@ -409,7 +432,7 @@ async fn request_handler(
             )
             .header("Access-Control-Max-Age", "86400" /* 24 hours */)
             .status(StatusCode::OK) // 204 is also valid, but see: https://developer.mozilla.org/en-US/docs/Web/HTTP/Methods/OPTIONS#status_code
-            .body(Full::new(Bytes::new()))
+            .body(Empty::new().map_err(|x| match x {}).boxed())
             .map_err(|e| ApiError::InternalServerError(e.into()))
     } else {
         json_response(StatusCode::BAD_REQUEST, "query is not supported")

proxy/src/serverless/backend.rs

@@ -1,6 +1,8 @@
-use std::{sync::Arc, time::Duration};
+use std::{io, sync::Arc, time::Duration};
 
 use async_trait::async_trait;
+use hyper_util::rt::{TokioExecutor, TokioIo, TokioTimer};
+use tokio::net::{lookup_host, TcpStream};
 use tracing::{field::display, info};
 
 use crate::{
@@ -27,9 +29,13 @@ use crate::{
     Host,
 };
 
-use super::conn_pool::{poll_client, Client, ConnInfo, GlobalConnPool};
+use super::{
+    conn_pool::{poll_client, Client, ConnInfo, GlobalConnPool},
+    http_conn_pool::{self, poll_http2_client},
+};
 
 pub(crate) struct PoolingBackend {
+    pub(crate) http_conn_pool: Arc<super::http_conn_pool::GlobalConnPool>,
     pub(crate) pool: Arc<GlobalConnPool<tokio_postgres::Client>>,
     pub(crate) config: &'static ProxyConfig,
     pub(crate) endpoint_rate_limiter: Arc<EndpointRateLimiter>,
@@ -103,32 +109,44 @@ impl PoolingBackend {
     pub(crate) async fn authenticate_with_jwt(
         &self,
         ctx: &RequestMonitoring,
+        config: &AuthenticationConfig,
         user_info: &ComputeUserInfo,
-        jwt: &str,
-    ) -> Result<ComputeCredentials, AuthError> {
+        jwt: String,
+    ) -> Result<(), AuthError> {
         match &self.config.auth_backend {
-            crate::auth::Backend::Console(_, ()) => {
-                Err(AuthError::auth_failed("JWT login is not yet supported"))
-            }
-            crate::auth::Backend::Web(_, ()) => Err(AuthError::auth_failed(
-                "JWT login over web auth proxy is not supported",
-            )),
-            crate::auth::Backend::Local(cache) => {
-                cache
+            crate::auth::Backend::Console(console, ()) => {
+                config
                     .jwks_cache
                     .check_jwt(
                         ctx,
                         user_info.endpoint.clone(),
-                        user_info.user.clone(),
-                        &StaticAuthRules,
-                        jwt,
+                        &user_info.user,
+                        &**console,
+                        &jwt,
                     )
                     .await
                     .map_err(|e| AuthError::auth_failed(e.to_string()))?;
-                Ok(ComputeCredentials {
-                    info: user_info.clone(),
-                    keys: crate::auth::backend::ComputeCredentialKeys::None,
-                })
+
+                Ok(())
+            }
+            crate::auth::Backend::Web(_, ()) => Err(AuthError::auth_failed(
+                "JWT login over web auth proxy is not supported",
+            )),
+            crate::auth::Backend::Local(_) => {
+                config
+                    .jwks_cache
+                    .check_jwt(
+                        ctx,
+                        user_info.endpoint.clone(),
+                        &user_info.user,
+                        &StaticAuthRules,
+                        &jwt,
+                    )
+                    .await
+                    .map_err(|e| AuthError::auth_failed(e.to_string()))?;
+
+                // todo: rewrite JWT signature with key shared somehow between local proxy and postgres
+                Ok(())
             }
         }
     }
@@ -174,14 +192,55 @@ impl PoolingBackend {
         )
         .await
     }
+
+    // Wake up the destination if needed
+    #[tracing::instrument(fields(pid = tracing::field::Empty), skip_all)]
+    pub(crate) async fn connect_to_local_proxy(
+        &self,
+        ctx: &RequestMonitoring,
+        conn_info: ConnInfo,
+    ) -> Result<http_conn_pool::Client, HttpConnError> {
+        info!("pool: looking for an existing connection");
+        if let Some(client) = self.http_conn_pool.get(ctx, &conn_info) {
+            return Ok(client);
+        }
+
+        let conn_id = uuid::Uuid::new_v4();
+        tracing::Span::current().record("conn_id", display(conn_id));
+        info!(%conn_id, "pool: opening a new connection '{conn_info}'");
+        let backend = self
+            .config
+            .auth_backend
+            .as_ref()
+            .map(|()| ComputeCredentials {
+                info: conn_info.user_info.clone(),
+                keys: crate::auth::backend::ComputeCredentialKeys::None,
+            });
+        crate::proxy::connect_compute::connect_to_compute(
+            ctx,
+            &HyperMechanism {
+                conn_id,
+                conn_info,
+                pool: self.http_conn_pool.clone(),
+                locks: &self.config.connect_compute_locks,
+            },
+            &backend,
+            false, // do not allow self signed compute for http flow
+            self.config.wake_compute_retry_config,
+            self.config.connect_to_compute_retry_config,
+        )
+        .await
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum HttpConnError {
     #[error("pooled connection closed at inconsistent state")]
     ConnectionClosedAbruptly(#[from] tokio::sync::watch::error::SendError<uuid::Uuid>),
-    #[error("could not connection to compute")]
-    ConnectionError(#[from] tokio_postgres::Error),
+    #[error("could not connection to postgres in compute")]
+    PostgresConnectionError(#[from] tokio_postgres::Error),
+    #[error("could not connection to local-proxy in compute")]
+    LocalProxyConnectionError(#[from] LocalProxyConnError),
 
     #[error("could not get auth info")]
     GetAuthInfo(#[from] GetAuthInfoError),
@@ -193,11 +252,20 @@ pub(crate) enum HttpConnError {
     TooManyConnectionAttempts(#[from] ApiLockError),
 }
 
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum LocalProxyConnError {
+    #[error("error with connection to local-proxy")]
+    Io(#[source] std::io::Error),
+    #[error("could not establish h2 connection")]
+    H2(#[from] hyper1::Error),
+}
+
 impl ReportableError for HttpConnError {
     fn get_error_kind(&self) -> ErrorKind {
         match self {
             HttpConnError::ConnectionClosedAbruptly(_) => ErrorKind::Compute,
-            HttpConnError::ConnectionError(p) => p.get_error_kind(),
+            HttpConnError::PostgresConnectionError(p) => p.get_error_kind(),
+            HttpConnError::LocalProxyConnectionError(_) => ErrorKind::Compute,
             HttpConnError::GetAuthInfo(a) => a.get_error_kind(),
             HttpConnError::AuthError(a) => a.get_error_kind(),
             HttpConnError::WakeCompute(w) => w.get_error_kind(),
@@ -210,7 +278,8 @@ impl UserFacingError for HttpConnError {
     fn to_string_client(&self) -> String {
         match self {
             HttpConnError::ConnectionClosedAbruptly(_) => self.to_string(),
-            HttpConnError::ConnectionError(p) => p.to_string(),
+            HttpConnError::PostgresConnectionError(p) => p.to_string(),
+            HttpConnError::LocalProxyConnectionError(p) => p.to_string(),
             HttpConnError::GetAuthInfo(c) => c.to_string_client(),
             HttpConnError::AuthError(c) => c.to_string_client(),
             HttpConnError::WakeCompute(c) => c.to_string_client(),
@@ -224,7 +293,8 @@ impl UserFacingError for HttpConnError {
 impl CouldRetry for HttpConnError {
     fn could_retry(&self) -> bool {
         match self {
-            HttpConnError::ConnectionError(e) => e.could_retry(),
+            HttpConnError::PostgresConnectionError(e) => e.could_retry(),
+            HttpConnError::LocalProxyConnectionError(e) => e.could_retry(),
             HttpConnError::ConnectionClosedAbruptly(_) => false,
             HttpConnError::GetAuthInfo(_) => false,
             HttpConnError::AuthError(_) => false,
@@ -236,7 +306,7 @@ impl CouldRetry for HttpConnError {
 impl ShouldRetryWakeCompute for HttpConnError {
     fn should_retry_wake_compute(&self) -> bool {
         match self {
-            HttpConnError::ConnectionError(e) => e.should_retry_wake_compute(),
+            HttpConnError::PostgresConnectionError(e) => e.should_retry_wake_compute(),
             // we never checked cache validity
             HttpConnError::TooManyConnectionAttempts(_) => false,
             _ => true,
@@ -244,6 +314,38 @@ impl ShouldRetryWakeCompute for HttpConnError {
     }
 }
 
+impl ReportableError for LocalProxyConnError {
+    fn get_error_kind(&self) -> ErrorKind {
+        match self {
+            LocalProxyConnError::Io(_) => ErrorKind::Compute,
+            LocalProxyConnError::H2(_) => ErrorKind::Compute,
+        }
+    }
+}
+
+impl UserFacingError for LocalProxyConnError {
+    fn to_string_client(&self) -> String {
+        "Could not establish HTTP connection to the database".to_string()
+    }
+}
+
+impl CouldRetry for LocalProxyConnError {
+    fn could_retry(&self) -> bool {
+        match self {
+            LocalProxyConnError::Io(_) => false,
+            LocalProxyConnError::H2(_) => false,
+        }
+    }
+}
+impl ShouldRetryWakeCompute for LocalProxyConnError {
+    fn should_retry_wake_compute(&self) -> bool {
+        match self {
+            LocalProxyConnError::Io(_) => false,
+            LocalProxyConnError::H2(_) => false,
+        }
+    }
+}
+
 struct TokioMechanism {
     pool: Arc<GlobalConnPool<tokio_postgres::Client>>,
     conn_info: ConnInfo,
@@ -293,3 +395,99 @@ impl ConnectMechanism for TokioMechanism {
 
     fn update_connect_config(&self, _config: &mut compute::ConnCfg) {}
 }
+
+struct HyperMechanism {
+    pool: Arc<http_conn_pool::GlobalConnPool>,
+    conn_info: ConnInfo,
+    conn_id: uuid::Uuid,
+
+    /// connect_to_compute concurrency lock
+    locks: &'static ApiLocks<Host>,
+}
+
+#[async_trait]
+impl ConnectMechanism for HyperMechanism {
+    type Connection = http_conn_pool::Client;
+    type ConnectError = HttpConnError;
+    type Error = HttpConnError;
+
+    async fn connect_once(
+        &self,
+        ctx: &RequestMonitoring,
+        node_info: &CachedNodeInfo,
+        timeout: Duration,
+    ) -> Result<Self::Connection, Self::ConnectError> {
+        let host = node_info.config.get_host()?;
+        let permit = self.locks.get_permit(&host).await?;
+
+        let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
+
+        // let port = node_info.config.get_ports().first().unwrap_or_else(10432);
+        let res = connect_http2(&host, 10432, timeout).await;
+        drop(pause);
+        let (client, connection) = permit.release_result(res)?;
+
+        Ok(poll_http2_client(
+            self.pool.clone(),
+            ctx,
+            &self.conn_info,
+            client,
+            connection,
+            self.conn_id,
+            node_info.aux.clone(),
+        ))
+    }
+
+    fn update_connect_config(&self, _config: &mut compute::ConnCfg) {}
+}
+
+async fn connect_http2(
+    host: &str,
+    port: u16,
+    timeout: Duration,
+) -> Result<(http_conn_pool::Send, http_conn_pool::Connect), LocalProxyConnError> {
+    // assumption: host is an ip address so this should not actually perform any requests.
+    // todo: add that assumption as a guarantee in the control-plane API.
+    let mut addrs = lookup_host((host, port))
+        .await
+        .map_err(LocalProxyConnError::Io)?;
+
+    let mut last_err = None;
+
+    let stream = loop {
+        let Some(addr) = addrs.next() else {
+            return Err(last_err.unwrap_or_else(|| {
+                LocalProxyConnError::Io(io::Error::new(
+                    io::ErrorKind::InvalidInput,
+                    "could not resolve any addresses",
+                ))
+            }));
+        };
+
+        match tokio::time::timeout(timeout, TcpStream::connect(addr)).await {
+            Ok(Ok(stream)) => {
+                stream.set_nodelay(true).map_err(LocalProxyConnError::Io)?;
+                break stream;
+            }
+            Ok(Err(e)) => {
+                last_err = Some(LocalProxyConnError::Io(e));
+            }
+            Err(e) => {
+                last_err = Some(LocalProxyConnError::Io(io::Error::new(
+                    io::ErrorKind::TimedOut,
+                    e,
+                )));
+            }
+        };
+    };
+
+    let (client, connection) = hyper1::client::conn::http2::Builder::new(TokioExecutor::new())
+        .timer(TokioTimer::new())
+        .keep_alive_interval(Duration::from_secs(20))
+        .keep_alive_while_idle(true)
+        .keep_alive_timeout(Duration::from_secs(5))
+        .handshake(TokioIo::new(stream))
+        .await?;
+
+    Ok((client, connection))
+}

proxy/src/serverless/http_conn_pool.rs

@@ -0,0 +1,335 @@
+use dashmap::DashMap;
+use hyper1::client::conn::http2;
+use hyper_util::rt::{TokioExecutor, TokioIo};
+use parking_lot::RwLock;
+use rand::Rng;
+use std::collections::VecDeque;
+use std::sync::atomic::{self, AtomicUsize};
+use std::{sync::Arc, sync::Weak};
+use tokio::net::TcpStream;
+
+use crate::console::messages::{ColdStartInfo, MetricsAuxInfo};
+use crate::metrics::{HttpEndpointPoolsGuard, Metrics};
+use crate::usage_metrics::{Ids, MetricCounter, USAGE_METRICS};
+use crate::{context::RequestMonitoring, EndpointCacheKey};
+
+use tracing::{debug, error};
+use tracing::{info, info_span, Instrument};
+
+use super::conn_pool::ConnInfo;
+
+pub(crate) type Send = http2::SendRequest<hyper1::body::Incoming>;
+pub(crate) type Connect =
+    http2::Connection<TokioIo<TcpStream>, hyper1::body::Incoming, TokioExecutor>;
+
+#[derive(Clone)]
+struct ConnPoolEntry {
+    conn: Send,
+    conn_id: uuid::Uuid,
+    aux: MetricsAuxInfo,
+}
+
+// Per-endpoint connection pool
+// Number of open connections is limited by the `max_conns_per_endpoint`.
+pub(crate) struct EndpointConnPool {
+    conns: VecDeque<ConnPoolEntry>,
+    _guard: HttpEndpointPoolsGuard<'static>,
+    global_connections_count: Arc<AtomicUsize>,
+}
+
+impl EndpointConnPool {
+    fn get_conn_entry(&mut self) -> Option<ConnPoolEntry> {
+        let Self { conns, .. } = self;
+
+        let conn = conns.pop_front()?;
+        conns.push_back(conn.clone());
+        Some(conn)
+    }
+
+    fn remove_conn(&mut self, conn_id: uuid::Uuid) -> bool {
+        let Self {
+            conns,
+            global_connections_count,
+            ..
+        } = self;
+
+        let old_len = conns.len();
+        conns.retain(|conn| conn.conn_id != conn_id);
+        let new_len = conns.len();
+        let removed = old_len - new_len;
+        if removed > 0 {
+            global_connections_count.fetch_sub(removed, atomic::Ordering::Relaxed);
+            Metrics::get()
+                .proxy
+                .http_pool_opened_connections
+                .get_metric()
+                .dec_by(removed as i64);
+        }
+        removed > 0
+    }
+}
+
+impl Drop for EndpointConnPool {
+    fn drop(&mut self) {
+        if !self.conns.is_empty() {
+            self.global_connections_count
+                .fetch_sub(self.conns.len(), atomic::Ordering::Relaxed);
+            Metrics::get()
+                .proxy
+                .http_pool_opened_connections
+                .get_metric()
+                .dec_by(self.conns.len() as i64);
+        }
+    }
+}
+
+pub(crate) struct GlobalConnPool {
+    // endpoint -> per-endpoint connection pool
+    //
+    // That should be a fairly conteded map, so return reference to the per-endpoint
+    // pool as early as possible and release the lock.
+    global_pool: DashMap<EndpointCacheKey, Arc<RwLock<EndpointConnPool>>>,
+
+    /// Number of endpoint-connection pools
+    ///
+    /// [`DashMap::len`] iterates over all inner pools and acquires a read lock on each.
+    /// That seems like far too much effort, so we're using a relaxed increment counter instead.
+    /// It's only used for diagnostics.
+    global_pool_size: AtomicUsize,
+
+    /// Total number of connections in the pool
+    global_connections_count: Arc<AtomicUsize>,
+
+    config: &'static crate::config::HttpConfig,
+}
+
+impl GlobalConnPool {
+    pub(crate) fn new(config: &'static crate::config::HttpConfig) -> Arc<Self> {
+        let shards = config.pool_options.pool_shards;
+        Arc::new(Self {
+            global_pool: DashMap::with_shard_amount(shards),
+            global_pool_size: AtomicUsize::new(0),
+            config,
+            global_connections_count: Arc::new(AtomicUsize::new(0)),
+        })
+    }
+
+    pub(crate) fn shutdown(&self) {
+        // drops all strong references to endpoint-pools
+        self.global_pool.clear();
+    }
+
+    pub(crate) async fn gc_worker(&self, mut rng: impl Rng) {
+        let epoch = self.config.pool_options.gc_epoch;
+        let mut interval = tokio::time::interval(epoch / (self.global_pool.shards().len()) as u32);
+        loop {
+            interval.tick().await;
+
+            let shard = rng.gen_range(0..self.global_pool.shards().len());
+            self.gc(shard);
+        }
+    }
+
+    fn gc(&self, shard: usize) {
+        debug!(shard, "pool: performing epoch reclamation");
+
+        // acquire a random shard lock
+        let mut shard = self.global_pool.shards()[shard].write();
+
+        let timer = Metrics::get()
+            .proxy
+            .http_pool_reclaimation_lag_seconds
+            .start_timer();
+        let current_len = shard.len();
+        let mut clients_removed = 0;
+        shard.retain(|endpoint, x| {
+            // if the current endpoint pool is unique (no other strong or weak references)
+            // then it is currently not in use by any connections.
+            if let Some(pool) = Arc::get_mut(x.get_mut()) {
+                let EndpointConnPool { conns, .. } = pool.get_mut();
+
+                let old_len = conns.len();
+
+                conns.retain(|conn| !conn.conn.is_closed());
+
+                let new_len = conns.len();
+                let removed = old_len - new_len;
+                clients_removed += removed;
+
+                // we only remove this pool if it has no active connections
+                if conns.is_empty() {
+                    info!("pool: discarding pool for endpoint {endpoint}");
+                    return false;
+                }
+            }
+
+            true
+        });
+
+        let new_len = shard.len();
+        drop(shard);
+        timer.observe();
+
+        // Do logging outside of the lock.
+        if clients_removed > 0 {
+            let size = self
+                .global_connections_count
+                .fetch_sub(clients_removed, atomic::Ordering::Relaxed)
+                - clients_removed;
+            Metrics::get()
+                .proxy
+                .http_pool_opened_connections
+                .get_metric()
+                .dec_by(clients_removed as i64);
+            info!("pool: performed global pool gc. removed {clients_removed} clients, total number of clients in pool is {size}");
+        }
+        let removed = current_len - new_len;
+
+        if removed > 0 {
+            let global_pool_size = self
+                .global_pool_size
+                .fetch_sub(removed, atomic::Ordering::Relaxed)
+                - removed;
+            info!("pool: performed global pool gc. size now {global_pool_size}");
+        }
+    }
+
+    pub(crate) fn get(
+        self: &Arc<Self>,
+        ctx: &RequestMonitoring,
+        conn_info: &ConnInfo,
+    ) -> Option<Client> {
+        let endpoint = conn_info.endpoint_cache_key()?;
+        let endpoint_pool = self.get_or_create_endpoint_pool(&endpoint);
+        let client = endpoint_pool.write().get_conn_entry()?;
+
+        if client.conn.is_closed() {
+            info!("pool: cached connection '{conn_info}' is closed, opening a new one");
+            return None;
+        }
+        tracing::Span::current().record("conn_id", tracing::field::display(client.conn_id));
+        info!(
+            cold_start_info = ColdStartInfo::HttpPoolHit.as_str(),
+            "pool: reusing connection '{conn_info}'"
+        );
+        ctx.set_cold_start_info(ColdStartInfo::HttpPoolHit);
+        ctx.success();
+        Some(Client::new(client.conn, client.aux))
+    }
+
+    fn get_or_create_endpoint_pool(
+        self: &Arc<Self>,
+        endpoint: &EndpointCacheKey,
+    ) -> Arc<RwLock<EndpointConnPool>> {
+        // fast path
+        if let Some(pool) = self.global_pool.get(endpoint) {
+            return pool.clone();
+        }
+
+        // slow path
+        let new_pool = Arc::new(RwLock::new(EndpointConnPool {
+            conns: VecDeque::new(),
+            _guard: Metrics::get().proxy.http_endpoint_pools.guard(),
+            global_connections_count: self.global_connections_count.clone(),
+        }));
+
+        // find or create a pool for this endpoint
+        let mut created = false;
+        let pool = self
+            .global_pool
+            .entry(endpoint.clone())
+            .or_insert_with(|| {
+                created = true;
+                new_pool
+            })
+            .clone();
+
+        // log new global pool size
+        if created {
+            let global_pool_size = self
+                .global_pool_size
+                .fetch_add(1, atomic::Ordering::Relaxed)
+                + 1;
+            info!(
+                "pool: created new pool for '{endpoint}', global pool size now {global_pool_size}"
+            );
+        }
+
+        pool
+    }
+}
+
+pub(crate) fn poll_http2_client(
+    global_pool: Arc<GlobalConnPool>,
+    ctx: &RequestMonitoring,
+    conn_info: &ConnInfo,
+    client: Send,
+    connection: Connect,
+    conn_id: uuid::Uuid,
+    aux: MetricsAuxInfo,
+) -> Client {
+    let conn_gauge = Metrics::get().proxy.db_connections.guard(ctx.protocol());
+    let session_id = ctx.session_id();
+
+    let span = info_span!(parent: None, "connection", %conn_id);
+    let cold_start_info = ctx.cold_start_info();
+    span.in_scope(|| {
+        info!(cold_start_info = cold_start_info.as_str(), %conn_info, %session_id, "new connection");
+    });
+
+    let pool = match conn_info.endpoint_cache_key() {
+        Some(endpoint) => {
+            let pool = global_pool.get_or_create_endpoint_pool(&endpoint);
+
+            pool.write().conns.push_back(ConnPoolEntry {
+                conn: client.clone(),
+                conn_id,
+                aux: aux.clone(),
+            });
+
+            Arc::downgrade(&pool)
+        }
+        None => Weak::new(),
+    };
+
+    // let idle = global_pool.get_idle_timeout();
+
+    tokio::spawn(
+        async move {
+            let _conn_gauge = conn_gauge;
+            let res = connection.await;
+            match res {
+                Ok(()) => info!("connection closed"),
+                Err(e) => error!(%session_id, "connection error: {}", e),
+            }
+
+            // remove from connection pool
+            if let Some(pool) = pool.clone().upgrade() {
+                if pool.write().remove_conn(conn_id) {
+                    info!("closed connection removed");
+                }
+            }
+        }
+        .instrument(span),
+    );
+
+    Client::new(client, aux)
+}
+
+pub(crate) struct Client {
+    pub(crate) inner: Send,
+    aux: MetricsAuxInfo,
+}
+
+impl Client {
+    pub(self) fn new(inner: Send, aux: MetricsAuxInfo) -> Self {
+        Self { inner, aux }
+    }
+
+    pub(crate) fn metrics(&self) -> Arc<MetricCounter> {
+        USAGE_METRICS.register(Ids {
+            endpoint_id: self.aux.endpoint_id,
+            branch_id: self.aux.branch_id,
+        })
+    }
+}

proxy/src/serverless/http_util.rs

@@ -5,13 +5,13 @@ use bytes::Bytes;
 
 use anyhow::Context;
 use http::{Response, StatusCode};
-use http_body_util::Full;
+use http_body_util::{combinators::BoxBody, BodyExt, Full};
 
 use serde::Serialize;
 use utils::http::error::ApiError;
 
 /// Like [`ApiError::into_response`]
-pub(crate) fn api_error_into_response(this: ApiError) -> Response<Full<Bytes>> {
+pub(crate) fn api_error_into_response(this: ApiError) -> Response<BoxBody<Bytes, hyper1::Error>> {
     match this {
         ApiError::BadRequest(err) => HttpErrorBody::response_from_msg_and_status(
             format!("{err:#?}"), // use debug printing so that we give the cause
@@ -64,17 +64,24 @@ struct HttpErrorBody {
 
 impl HttpErrorBody {
     /// Same as [`utils::http::error::HttpErrorBody::response_from_msg_and_status`]
-    fn response_from_msg_and_status(msg: String, status: StatusCode) -> Response<Full<Bytes>> {
+    fn response_from_msg_and_status(
+        msg: String,
+        status: StatusCode,
+    ) -> Response<BoxBody<Bytes, hyper1::Error>> {
         HttpErrorBody { msg }.to_response(status)
     }
 
     /// Same as [`utils::http::error::HttpErrorBody::to_response`]
-    fn to_response(&self, status: StatusCode) -> Response<Full<Bytes>> {
+    fn to_response(&self, status: StatusCode) -> Response<BoxBody<Bytes, hyper1::Error>> {
         Response::builder()
             .status(status)
             .header(http::header::CONTENT_TYPE, "application/json")
             // we do not have nested maps with non string keys so serialization shouldn't fail
-            .body(Full::new(Bytes::from(serde_json::to_string(self).unwrap())))
+            .body(
+                Full::new(Bytes::from(serde_json::to_string(self).unwrap()))
+                    .map_err(|x| match x {})
+                    .boxed(),
+            )
             .unwrap()
     }
 }
@@ -83,14 +90,14 @@ impl HttpErrorBody {
 pub(crate) fn json_response<T: Serialize>(
     status: StatusCode,
     data: T,
-) -> Result<Response<Full<Bytes>>, ApiError> {
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, ApiError> {
     let json = serde_json::to_string(&data)
         .context("Failed to serialize JSON response")
         .map_err(ApiError::InternalServerError)?;
     let response = Response::builder()
         .status(status)
         .header(http::header::CONTENT_TYPE, "application/json")
-        .body(Full::new(Bytes::from(json)))
+        .body(Full::new(Bytes::from(json)).map_err(|x| match x {}).boxed())
         .map_err(|e| ApiError::InternalServerError(e.into()))?;
     Ok(response)
 }

proxy/src/serverless/sql_over_http.rs

@@ -8,6 +8,8 @@ use futures::future::Either;
 use futures::StreamExt;
 use futures::TryFutureExt;
 use http::header::AUTHORIZATION;
+use http::Method;
+use http_body_util::combinators::BoxBody;
 use http_body_util::BodyExt;
 use http_body_util::Full;
 use hyper1::body::Body;
@@ -38,9 +40,11 @@ use url::Url;
 use urlencoding;
 use utils::http::error::ApiError;
 
+use crate::auth::backend::ComputeCredentials;
 use crate::auth::backend::ComputeUserInfo;
 use crate::auth::endpoint_sni;
 use crate::auth::ComputeUserInfoParseError;
+use crate::config::AuthenticationConfig;
 use crate::config::ProxyConfig;
 use crate::config::TlsConfig;
 use crate::context::RequestMonitoring;
@@ -56,6 +60,7 @@ use crate::usage_metrics::MetricCounterRecorder;
 use crate::DbName;
 use crate::RoleName;
 
+use super::backend::LocalProxyConnError;
 use super::backend::PoolingBackend;
 use super::conn_pool::AuthData;
 use super::conn_pool::Client;
@@ -123,8 +128,8 @@ pub(crate) enum ConnInfoError {
     MissingUsername,
     #[error("invalid username: {0}")]
     InvalidUsername(#[from] std::string::FromUtf8Error),
-    #[error("missing password")]
-    MissingPassword,
+    #[error("missing authentication credentials: {0}")]
+    MissingCredentials(Credentials),
     #[error("missing hostname")]
     MissingHostname,
     #[error("invalid hostname: {0}")]
@@ -133,6 +138,14 @@ pub(crate) enum ConnInfoError {
     MalformedEndpoint,
 }
 
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum Credentials {
+    #[error("required password")]
+    Password,
+    #[error("required authorization bearer token in JWT format")]
+    BearerJwt,
+}
+
 impl ReportableError for ConnInfoError {
     fn get_error_kind(&self) -> ErrorKind {
         ErrorKind::User
@@ -146,6 +159,7 @@ impl UserFacingError for ConnInfoError {
 }
 
 fn get_conn_info(
+    config: &'static AuthenticationConfig,
     ctx: &RequestMonitoring,
     headers: &HeaderMap,
     tls: Option<&TlsConfig>,
@@ -181,21 +195,32 @@ fn get_conn_info(
     ctx.set_user(username.clone());
 
     let auth = if let Some(auth) = headers.get(&AUTHORIZATION) {
+        if !config.accept_jwts {
+            return Err(ConnInfoError::MissingCredentials(Credentials::Password));
+        }
+
         let auth = auth
             .to_str()
             .map_err(|_| ConnInfoError::InvalidHeader(&AUTHORIZATION))?;
         AuthData::Jwt(
             auth.strip_prefix("Bearer ")
-                .ok_or(ConnInfoError::MissingPassword)?
+                .ok_or(ConnInfoError::MissingCredentials(Credentials::BearerJwt))?
                 .into(),
         )
     } else if let Some(pass) = connection_url.password() {
+        // wrong credentials provided
+        if config.accept_jwts {
+            return Err(ConnInfoError::MissingCredentials(Credentials::BearerJwt));
+        }
+
         AuthData::Password(match urlencoding::decode_binary(pass.as_bytes()) {
             std::borrow::Cow::Borrowed(b) => b.into(),
             std::borrow::Cow::Owned(b) => b.into(),
         })
+    } else if config.accept_jwts {
+        return Err(ConnInfoError::MissingCredentials(Credentials::BearerJwt));
     } else {
-        return Err(ConnInfoError::MissingPassword);
+        return Err(ConnInfoError::MissingCredentials(Credentials::Password));
     };
 
     let endpoint = match connection_url.host() {
@@ -247,7 +272,7 @@ pub(crate) async fn handle(
     request: Request<Incoming>,
     backend: Arc<PoolingBackend>,
     cancel: CancellationToken,
-) -> Result<Response<Full<Bytes>>, ApiError> {
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, ApiError> {
     let result = handle_inner(cancel, config, &ctx, request, backend).await;
 
     let mut response = match result {
@@ -279,7 +304,7 @@ pub(crate) async fn handle(
 
             let mut message = e.to_string_client();
             let db_error = match &e {
-                SqlOverHttpError::ConnectCompute(HttpConnError::ConnectionError(e))
+                SqlOverHttpError::ConnectCompute(HttpConnError::PostgresConnectionError(e))
                 | SqlOverHttpError::Postgres(e) => e.as_db_error(),
                 _ => None,
             };
@@ -504,7 +529,7 @@ async fn handle_inner(
     ctx: &RequestMonitoring,
     request: Request<Incoming>,
     backend: Arc<PoolingBackend>,
-) -> Result<Response<Full<Bytes>>, SqlOverHttpError> {
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, SqlOverHttpError> {
     let _requeset_gauge = Metrics::get()
         .proxy
         .connection_requests
@@ -514,18 +539,50 @@ async fn handle_inner(
         "handling interactive connection from client"
     );
 
-    //
-    // Determine the destination and connection params
-    //
-    let headers = request.headers();
-
-    // TLS config should be there.
-    let conn_info = get_conn_info(ctx, headers, config.tls_config.as_ref())?;
+    let conn_info = get_conn_info(
+        &config.authentication_config,
+        ctx,
+        request.headers(),
+        config.tls_config.as_ref(),
+    )?;
     info!(
         user = conn_info.conn_info.user_info.user.as_str(),
         "credentials"
     );
 
+    match conn_info.auth {
+        AuthData::Jwt(jwt) if config.authentication_config.is_auth_broker => {
+            handle_auth_broker_inner(config, ctx, request, conn_info.conn_info, jwt, backend).await
+        }
+        auth => {
+            handle_db_inner(
+                cancel,
+                config,
+                ctx,
+                request,
+                conn_info.conn_info,
+                auth,
+                backend,
+            )
+            .await
+        }
+    }
+}
+
+async fn handle_db_inner(
+    cancel: CancellationToken,
+    config: &'static ProxyConfig,
+    ctx: &RequestMonitoring,
+    request: Request<Incoming>,
+    conn_info: ConnInfo,
+    auth: AuthData,
+    backend: Arc<PoolingBackend>,
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, SqlOverHttpError> {
+    //
+    // Determine the destination and connection params
+    //
+    let headers = request.headers();
+
     // Allow connection pooling only if explicitly requested
     // or if we have decided that http pool is no longer opt-in
     let allow_pool = !config.http_config.pool_options.opt_in
@@ -563,26 +620,36 @@ async fn handle_inner(
 
     let authenticate_and_connect = Box::pin(
         async {
-            let keys = match &conn_info.auth {
+            let keys = match auth {
                 AuthData::Password(pw) => {
                     backend
                         .authenticate_with_password(
                             ctx,
                             &config.authentication_config,
-                            &conn_info.conn_info.user_info,
-                            pw,
+                            &conn_info.user_info,
+                            &pw,
                         )
                         .await?
                 }
                 AuthData::Jwt(jwt) => {
                     backend
-                        .authenticate_with_jwt(ctx, &conn_info.conn_info.user_info, jwt)
-                        .await?
+                        .authenticate_with_jwt(
+                            ctx,
+                            &config.authentication_config,
+                            &conn_info.user_info,
+                            jwt,
+                        )
+                        .await?;
+
+                    ComputeCredentials {
+                        info: conn_info.user_info.clone(),
+                        keys: crate::auth::backend::ComputeCredentialKeys::None,
+                    }
                 }
             };
 
             let client = backend
-                .connect_to_compute(ctx, conn_info.conn_info, keys, !allow_pool)
+                .connect_to_compute(ctx, conn_info, keys, !allow_pool)
                 .await?;
             // not strictly necessary to mark success here,
             // but it's just insurance for if we forget it somewhere else
@@ -640,7 +707,11 @@ async fn handle_inner(
 
     let len = json_output.len();
     let response = response
-        .body(Full::new(Bytes::from(json_output)))
+        .body(
+            Full::new(Bytes::from(json_output))
+                .map_err(|x| match x {})
+                .boxed(),
+        )
         // only fails if invalid status code or invalid header/values are given.
         // these are not user configurable so it cannot fail dynamically
         .expect("building response payload should not fail");
@@ -656,6 +727,65 @@ async fn handle_inner(
     Ok(response)
 }
 
+static HEADERS_TO_FORWARD: &[&HeaderName] = &[
+    &AUTHORIZATION,
+    &CONN_STRING,
+    &RAW_TEXT_OUTPUT,
+    &ARRAY_MODE,
+    &TXN_ISOLATION_LEVEL,
+    &TXN_READ_ONLY,
+    &TXN_DEFERRABLE,
+];
+
+async fn handle_auth_broker_inner(
+    config: &'static ProxyConfig,
+    ctx: &RequestMonitoring,
+    request: Request<Incoming>,
+    conn_info: ConnInfo,
+    jwt: String,
+    backend: Arc<PoolingBackend>,
+) -> Result<Response<BoxBody<Bytes, hyper1::Error>>, SqlOverHttpError> {
+    backend
+        .authenticate_with_jwt(
+            ctx,
+            &config.authentication_config,
+            &conn_info.user_info,
+            jwt,
+        )
+        .await
+        .map_err(HttpConnError::from)?;
+
+    let mut client = backend.connect_to_local_proxy(ctx, conn_info).await?;
+
+    let local_proxy_uri = ::http::Uri::from_static("http://proxy.local/sql");
+
+    let (mut parts, body) = request.into_parts();
+    let mut req = Request::builder().method(Method::POST).uri(local_proxy_uri);
+
+    // todo(conradludgate): maybe auth-broker should parse these and re-serialize
+    // these instead just to ensure they remain normalised.
+    for &h in HEADERS_TO_FORWARD {
+        if let Some(hv) = parts.headers.remove(h) {
+            req = req.header(h, hv);
+        }
+    }
+
+    let req = req
+        .body(body)
+        .expect("all headers and params received via hyper should be valid for request");
+
+    // todo: map body to count egress
+    let _metrics = client.metrics();
+
+    Ok(client
+        .inner
+        .send_request(req)
+        .await
+        .map_err(LocalProxyConnError::from)
+        .map_err(HttpConnError::from)?
+        .map(|b| b.boxed()))
+}
+
 impl QueryData {
     async fn process(
         self,
@@ -705,7 +835,9 @@ impl QueryData {
                     // query failed or was cancelled.
                     Ok(Err(error)) => {
                         let db_error = match &error {
-                            SqlOverHttpError::ConnectCompute(HttpConnError::ConnectionError(e))
+                            SqlOverHttpError::ConnectCompute(
+                                HttpConnError::PostgresConnectionError(e),
+                            )
                             | SqlOverHttpError::Postgres(e) => e.as_db_error(),
                             _ => None,
                         };

safekeeper/Cargo.toml

@@ -21,7 +21,6 @@ chrono.workspace = true
 clap = { workspace = true, features = ["derive"] }
 crc32c.workspace = true
 fail.workspace = true
-git-version.workspace = true
 hex.workspace = true
 humantime.workspace = true
 hyper.workspace = true

safekeeper/src/bin/safekeeper.rs

@@ -374,14 +374,16 @@ type JoinTaskRes = Result<anyhow::Result<()>, JoinError>;
 
 async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
     // fsync the datadir to make sure we have a consistent state on disk.
-    let dfd = File::open(&conf.workdir).context("open datadir for syncfs")?;
-    let started = Instant::now();
-    utils::crashsafe::syncfs(dfd)?;
-    let elapsed = started.elapsed();
-    info!(
-        elapsed_ms = elapsed.as_millis(),
-        "syncfs data directory done"
-    );
+    if !conf.no_sync {
+        let dfd = File::open(&conf.workdir).context("open datadir for syncfs")?;
+        let started = Instant::now();
+        utils::crashsafe::syncfs(dfd)?;
+        let elapsed = started.elapsed();
+        info!(
+            elapsed_ms = elapsed.as_millis(),
+            "syncfs data directory done"
+        );
+    }
 
     info!("starting safekeeper WAL service on {}", conf.listen_pg_addr);
     let pg_listener = tcp_listener::bind(conf.listen_pg_addr.clone()).map_err(|e| {