Bump postgres version

## Describe your changes

In https://github.com/neondatabase/cloud/issues/4354 we are making
scheduling of projects based on available disk space and overcommit, so
we need to know disk size and just in case instance type of the
pageserver

## Issue ticket number and link

https://github.com/neondatabase/cloud/issues/4354

## Checklist before requesting a review
- [x] I have performed a self-review of my code.
- [ ] ~If it is a core feature, I have added thorough tests.~
- [ ] ~Do we need to implement analytics? if so did you add the relevant
metrics to the dashboard?~
- [ ] ~If this PR requires public announcement, mark it with
/release-notes label and add several sentences in this section.~

.github/actions/allure-report/action.yml

@@ -15,10 +15,32 @@ outputs:
   report-url:
     description: 'Allure report URL'
     value: ${{ steps.generate-report.outputs.report-url }}
+  report-json-url:
+    description: 'Allure report JSON URL'
+    value: ${{ steps.generate-report.outputs.report-json-url }}
 
 runs:
   using: "composite"
+
   steps:
+    # We're using some of env variables quite offen, so let's set them once.
+    #
+    # It would be nice to have them set in common runs.env[0] section, but it doesn't work[1]
+    #
+    # - [0] https://docs.github.com/en/actions/creating-actions/metadata-syntax-for-github-actions#runsenv
+    # - [1] https://github.com/neondatabase/neon/pull/3907#discussion_r1154703456
+    #
+    - name: Set common environment variables
+      shell: bash -euxo pipefail {0}
+      run: |
+        echo "BUILD_TYPE=${BUILD_TYPE}"   >> $GITHUB_ENV
+        echo "BUCKET=${BUCKET}"           >> $GITHUB_ENV
+        echo "TEST_OUTPUT=${TEST_OUTPUT}" >> $GITHUB_ENV
+      env:
+        BUILD_TYPE: ${{ inputs.build_type }}
+        BUCKET: neon-github-public-dev
+        TEST_OUTPUT: /tmp/test_output
+
     - name: Validate input parameters
       shell: bash -euxo pipefail {0}
       run: |
@@ -76,16 +98,14 @@ runs:
           rm -f ${ALLURE_ZIP}
         fi
       env:
-        ALLURE_VERSION: 2.19.0
-        ALLURE_ZIP_MD5: ced21401a1a8b9dfb68cee9e4c210464
+        ALLURE_VERSION: 2.21.0
+        ALLURE_ZIP_MD5: c8db4dd8e2a7882583d569ed2c82879c
 
     - name: Upload Allure results
       if: ${{ inputs.action == 'store' }}
       env:
         REPORT_PREFIX: reports/${{ steps.calculate-vars.outputs.KEY }}/${{ inputs.build_type }}
         RAW_PREFIX: reports-raw/${{ steps.calculate-vars.outputs.KEY }}/${{ inputs.build_type }}
-        TEST_OUTPUT: /tmp/test_output
-        BUCKET: neon-github-public-dev
         TEST_SELECTION: ${{ steps.calculate-vars.outputs.TEST_SELECTION }}
       shell: bash -euxo pipefail {0}
       run: |
@@ -104,7 +124,7 @@ runs:
         EOF
         cat <<EOF > $TEST_OUTPUT/allure/results/environment.properties
           TEST_SELECTION=${{ inputs.test_selection }}
-          BUILD_TYPE=${{ inputs.build_type }}
+          BUILD_TYPE=${BUILD_TYPE}
         EOF
 
         ARCHIVE="${GITHUB_RUN_ID}-${TEST_SELECTION}-${GITHUB_RUN_ATTEMPT}-$(date +%s).tar.zst"
@@ -113,13 +133,12 @@ runs:
         tar -C ${TEST_OUTPUT}/allure/results -cf ${ARCHIVE} --zstd .
         aws s3 mv --only-show-errors ${ARCHIVE} "s3://${BUCKET}/${RAW_PREFIX}/${ARCHIVE}"
 
-    # Potentially we could have several running build for the same key (for example for the main branch),  so we use improvised lock for this
+    # Potentially we could have several running build for the same key (for example for the main branch), so we use improvised lock for this
     - name: Acquire Allure lock
       if: ${{ inputs.action == 'generate' }}
       shell: bash -euxo pipefail {0}
       env:
         LOCK_FILE: reports/${{ steps.calculate-vars.outputs.KEY }}/lock.txt
-        BUCKET: neon-github-public-dev
         TEST_SELECTION: ${{ steps.calculate-vars.outputs.TEST_SELECTION }}
       run: |
         LOCK_TIMEOUT=300 # seconds
@@ -149,8 +168,6 @@ runs:
       env:
         REPORT_PREFIX: reports/${{ steps.calculate-vars.outputs.KEY }}/${{ inputs.build_type }}
         RAW_PREFIX: reports-raw/${{ steps.calculate-vars.outputs.KEY }}/${{ inputs.build_type }}
-        TEST_OUTPUT: /tmp/test_output
-        BUCKET: neon-github-public-dev
       shell: bash -euxo pipefail {0}
       run: |
         # Get previously uploaded data for this run
@@ -186,24 +203,24 @@ runs:
         REPORT_URL=https://${BUCKET}.s3.amazonaws.com/${REPORT_PREFIX}/${GITHUB_RUN_ID}/index.html
 
         # Generate redirect
-        cat <<EOF > ./index.html
+        cat <<EOF > ${TEST_OUTPUT}/allure/index.html
           <!DOCTYPE html>
 
           <meta charset="utf-8">
           <title>Redirecting to ${REPORT_URL}</title>
           <meta http-equiv="refresh" content="0; URL=${REPORT_URL}">
         EOF
-        aws s3 cp --only-show-errors ./index.html "s3://${BUCKET}/${REPORT_PREFIX}/latest/index.html"
+        aws s3 cp --only-show-errors ${TEST_OUTPUT}/allure/index.html "s3://${BUCKET}/${REPORT_PREFIX}/latest/index.html"
 
         echo "[Allure Report](${REPORT_URL})" >> ${GITHUB_STEP_SUMMARY}
         echo "report-url=${REPORT_URL}" >> $GITHUB_OUTPUT
+        echo "report-json-url=${REPORT_URL%/index.html}/data/suites.json" >> $GITHUB_OUTPUT
 
     - name: Release Allure lock
       if: ${{ inputs.action == 'generate' && always() }}
       shell: bash -euxo pipefail {0}
       env:
         LOCK_FILE: reports/${{ steps.calculate-vars.outputs.KEY }}/lock.txt
-        BUCKET: neon-github-public-dev
         TEST_SELECTION: ${{ steps.calculate-vars.outputs.TEST_SELECTION }}
       run: |
         aws s3 cp --only-show-errors "s3://${BUCKET}/${LOCK_FILE}" ./lock.txt || exit 0
@@ -212,11 +229,16 @@ runs:
           aws s3 rm "s3://${BUCKET}/${LOCK_FILE}"
         fi
 
+    - name: Cleanup
+      if: always()
+      shell: bash -euxo pipefail {0}
+      run: |
+        rm -rf ${TEST_OUTPUT}/allure
+
     - uses: actions/github-script@v6
       if: ${{ inputs.action == 'generate' && always() }}
       env:
         REPORT_URL: ${{ steps.generate-report.outputs.report-url }}
-        BUILD_TYPE: ${{ inputs.build_type }}
         SHA: ${{ github.event.pull_request.head.sha || github.sha }}
       with:
         script: |

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

@@ -44,6 +44,10 @@ inputs:
     description: 'Secret access key'
     required: false
     default: ''
+  rerun_flaky:
+    description: 'Whether to rerun flaky tests'
+    required: false
+    default: 'false'
 
 runs:
   using: "composite"
@@ -101,6 +105,7 @@ runs:
         COMPATIBILITY_SNAPSHOT_DIR: /tmp/compatibility_snapshot_pg14
         ALLOW_BACKWARD_COMPATIBILITY_BREAKAGE: contains(github.event.pull_request.labels.*.name, 'backward compatibility breakage')
         ALLOW_FORWARD_COMPATIBILITY_BREAKAGE: contains(github.event.pull_request.labels.*.name, 'forward compatibility breakage')
+        RERUN_FLAKY: ${{ inputs.rerun_flaky }}
       shell: bash -euxo pipefail {0}
       run: |
         # PLATFORM will be embedded in the perf test report
@@ -143,6 +148,13 @@ runs:
           EXTRA_PARAMS="--out-dir $PERF_REPORT_DIR $EXTRA_PARAMS"
         fi
 
+        if [ "${RERUN_FLAKY}" == "true" ]; then
+          mkdir -p $TEST_OUTPUT
+          poetry run ./scripts/flaky_tests.py "${TEST_RESULT_CONNSTR}" --days 10 --output "$TEST_OUTPUT/flaky.json"
+
+          EXTRA_PARAMS="--flaky-tests-json $TEST_OUTPUT/flaky.json $EXTRA_PARAMS"
+        fi
+
         if [[ "${{ inputs.build_type }}" == "debug" ]]; then
           cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
         elif [[ "${{ inputs.build_type }}" == "release" ]]; then

.github/ansible/prod.ap-southeast-1.hosts.yaml

@@ -8,6 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.tech/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 85 # TODO: decrease to 80 after all pageservers are below 80
+        min_avail_bytes: 0
+        period: "10s"
+      tenant_config:
+        eviction_policy:
+          kind: "LayerAccessThreshold"
+          period: "10m"
+          threshold: &default_eviction_threshold "24h"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/ansible/prod.eu-central-1.hosts.yaml

@@ -8,6 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.tech/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 85 # TODO: decrease to 80 after all pageservers are below 80
+        min_avail_bytes: 0
+        period: "10s"
+      tenant_config:
+        eviction_policy:
+          kind: "LayerAccessThreshold"
+          period: "10m"
+          threshold: &default_eviction_threshold "24h"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/ansible/prod.us-east-2.hosts.yaml

@@ -8,6 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.tech/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 85 # TODO: decrease to 80 after all pageservers are below 80
+        min_avail_bytes: 0
+        period: "10s"
+      tenant_config:
+        eviction_policy:
+          kind: "LayerAccessThreshold"
+          period: "10m"
+          threshold: &default_eviction_threshold "24h"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/ansible/prod.us-west-2.hosts.yaml

@@ -8,6 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.tech/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 85 # TODO: decrease to 80 after all pageservers are below 80
+        min_avail_bytes: 0
+        period: "10s"
+      tenant_config:
+        eviction_policy:
+          kind: "LayerAccessThreshold"
+          period: "10m"
+          threshold: &default_eviction_threshold "24h"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/ansible/scripts/init_pageserver.sh

@@ -3,6 +3,8 @@
 # fetch params from meta-data service
 INSTANCE_ID=$(curl -s http://169.254.169.254/latest/meta-data/instance-id)
 AZ_ID=$(curl -s http://169.254.169.254/latest/meta-data/placement/availability-zone)
+INSTANCE_TYPE=$(curl -s http://169.254.169.254/latest/meta-data/instance-type)
+DISK_SIZE=$(df -B1 /storage | tail -1 | awk '{print $2}')
 
 # store fqdn hostname in var
 HOST=$(hostname -f)
@@ -18,7 +20,9 @@ cat <<EOF | tee /tmp/payload
   "http_host": "${HOST}",
   "http_port": 9898,
   "active": false,
-  "availability_zone_id": "${AZ_ID}"
+  "availability_zone_id": "${AZ_ID}",
+  "disk_size": ${DISK_SIZE},
+  "instance_type": "${INSTANCE_TYPE}"
 }
 EOF
 

.github/ansible/staging.eu-west-1.hosts.yaml

@@ -8,11 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.build/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 80
+        min_avail_bytes: 0
+        period: "10s"
       tenant_config:
         eviction_policy:
           kind: "LayerAccessThreshold"
           period: "20m"
-          threshold: "20m"
+          threshold: &default_eviction_threshold "20m"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/ansible/staging.us-east-2.hosts.yaml

@@ -8,11 +8,16 @@ storage:
       pg_distrib_dir: /usr/local
       metric_collection_endpoint: http://neon-internal-api.aws.neon.build/billing/api/v1/usage_events
       metric_collection_interval: 10min
+      disk_usage_based_eviction:
+        max_usage_pct: 80
+        min_avail_bytes: 0
+        period: "10s"
       tenant_config:
         eviction_policy:
           kind: "LayerAccessThreshold"
           period: "20m"
-          threshold: "20m"
+          threshold: &default_eviction_threshold "20m"
+      evictions_low_residence_duration_metric_threshold: *default_eviction_threshold
       remote_storage:
         bucket_name: "{{ bucket_name }}"
         bucket_region: "{{ bucket_region }}"

.github/helm-values/dev-eu-west-1-zeta.neon-proxy-scram.yaml

@@ -30,10 +30,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: dev
-  zenith_region: eu-west-1
-  zenith_region_slug: eu-west-1
+  neon_service: proxy-scram
+  neon_env: dev
+  neon_region: eu-west-1
 
 exposedService:
   annotations:

.github/helm-values/dev-us-east-2-beta.neon-proxy-link.yaml

@@ -15,10 +15,9 @@ settings:
 
 # -- Additional labels for neon-proxy-link pods
 podLabels:
-  zenith_service: proxy
-  zenith_env: dev
-  zenith_region: us-east-2
-  zenith_region_slug: us-east-2
+  neon_service: proxy
+  neon_env: dev
+  neon_region: us-east-2
 
 service:
   type: LoadBalancer

.github/helm-values/dev-us-east-2-beta.neon-proxy-scram-legacy.yaml

@@ -15,10 +15,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram-legacy
-  zenith_env: dev
-  zenith_region: us-east-2
-  zenith_region_slug: us-east-2
+  neon_service: proxy-scram-legacy
+  neon_env: dev
+  neon_region: us-east-2
 
 exposedService:
   annotations:

.github/helm-values/dev-us-east-2-beta.neon-proxy-scram.yaml

@@ -30,10 +30,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: dev
-  zenith_region: us-east-2
-  zenith_region_slug: us-east-2
+  neon_service: proxy-scram
+  neon_env: dev
+  neon_region: us-east-2
 
 exposedService:
   annotations:

.github/helm-values/prod-ap-southeast-1-epsilon.neon-proxy-scram.yaml

@@ -31,10 +31,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: prod
-  zenith_region: ap-southeast-1
-  zenith_region_slug: ap-southeast-1
+  neon_service: proxy-scram
+  neon_env: prod
+  neon_region: ap-southeast-1
 
 exposedService:
   annotations:

.github/helm-values/prod-eu-central-1-gamma.neon-proxy-scram.yaml

@@ -31,10 +31,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: prod
-  zenith_region: eu-central-1
-  zenith_region_slug: eu-central-1
+  neon_service: proxy-scram
+  neon_env: prod
+  neon_region: eu-central-1
 
 exposedService:
   annotations:

.github/helm-values/prod-us-east-2-delta.neon-proxy-link.yaml

@@ -13,10 +13,9 @@ settings:
 
 # -- Additional labels for zenith-proxy pods
 podLabels:
-  zenith_service: proxy
-  zenith_env: production
-  zenith_region: us-east-2
-  zenith_region_slug: us-east-2
+  neon_service: proxy
+  neon_env: production
+  neon_region: us-east-2
 
 service:
   type: LoadBalancer

.github/helm-values/prod-us-east-2-delta.neon-proxy-scram.yaml

@@ -31,10 +31,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: prod
-  zenith_region: us-east-2
-  zenith_region_slug: us-east-2
+  neon_service: proxy-scram
+  neon_env: prod
+  neon_region: us-east-2
 
 exposedService:
   annotations:

.github/helm-values/prod-us-west-2-eta.neon-proxy-scram-legacy.yaml

@@ -31,10 +31,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: prod
-  zenith_region: us-west-2
-  zenith_region_slug: us-west-2
+  neon_service: proxy-scram
+  neon_env: prod
+  neon_region: us-west-2
 
 exposedService:
   annotations:

.github/helm-values/prod-us-west-2-eta.neon-proxy-scram.yaml

@@ -31,10 +31,9 @@ settings:
 
 # -- Additional labels for neon-proxy pods
 podLabels:
-  zenith_service: proxy-scram
-  zenith_env: prod
-  zenith_region: us-west-2
-  zenith_region_slug: us-west-2
+  neon_service: proxy-scram
+  neon_env: prod
+  neon_region: us-west-2
 
 exposedService:
   annotations:

.github/pull_request_template.md

@@ -3,8 +3,12 @@
 ## Issue ticket number and link
 
 ## Checklist before requesting a review
+
 - [ ] I have performed a self-review of my code.
 - [ ] If it is a core feature, I have added thorough tests.
 - [ ] Do we need to implement analytics? if so did you add the relevant metrics to the dashboard?
 - [ ] If this PR requires public announcement, mark it with /release-notes label and add several sentences in this section.
 
+## Checklist before merging
+
+- [ ] Do not forget to reformat commit message to not include the above checklist

.github/workflows/build_and_test.yml

@@ -184,10 +184,10 @@ jobs:
           CARGO_FEATURES="--features testing"
           if [[ $BUILD_TYPE == "debug" ]]; then
             cov_prefix="scripts/coverage --profraw-prefix=$GITHUB_JOB --dir=/tmp/coverage run"
-            CARGO_FLAGS="--locked $CARGO_FEATURES"
+            CARGO_FLAGS="--locked"
           elif [[ $BUILD_TYPE == "release" ]]; then
             cov_prefix=""
-            CARGO_FLAGS="--locked --release $CARGO_FEATURES"
+            CARGO_FLAGS="--locked --release"
           fi
           echo "cov_prefix=${cov_prefix}" >> $GITHUB_ENV
           echo "CARGO_FEATURES=${CARGO_FEATURES}" >> $GITHUB_ENV
@@ -240,11 +240,18 @@ jobs:
 
       - name: Run cargo build
         run: |
-          ${cov_prefix} mold -run cargo build $CARGO_FLAGS --bins --tests
+          ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
 
       - name: Run cargo test
         run: |
-          ${cov_prefix} cargo test $CARGO_FLAGS
+          ${cov_prefix} cargo test $CARGO_FLAGS $CARGO_FEATURES
+
+          # Run separate tests for real S3
+          export ENABLE_REAL_S3_REMOTE_STORAGE=nonempty
+          export REMOTE_STORAGE_S3_BUCKET=neon-github-public-dev
+          export REMOTE_STORAGE_S3_REGION=eu-central-1
+          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test pagination_tests -- s3_pagination_should_work --exact
 
       - name: Install rust binaries
         run: |
@@ -268,7 +275,7 @@ jobs:
             mkdir -p /tmp/neon/test_bin/
 
             test_exe_paths=$(
-              ${cov_prefix} cargo test $CARGO_FLAGS --message-format=json --no-run |
+              ${cov_prefix} cargo test $CARGO_FLAGS $CARGO_FEATURES --message-format=json --no-run |
               jq -r '.executable | select(. != null)'
             )
             for bin in $test_exe_paths; do
@@ -328,6 +335,9 @@ jobs:
           real_s3_region: us-west-2
           real_s3_access_key_id: "${{ secrets.AWS_ACCESS_KEY_ID_CI_TESTS_S3 }}"
           real_s3_secret_access_key: "${{ secrets.AWS_SECRET_ACCESS_KEY_CI_TESTS_S3 }}"
+          rerun_flaky: true
+        env:
+          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
 
       - name: Merge and upload coverage data
         if: matrix.build_type == 'debug'
@@ -364,42 +374,90 @@ jobs:
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
-  merge-allure-report:
+  create-test-report:
     runs-on: [ self-hosted, gen3, small ]
     container:
       image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
       options: --init
     needs: [ regress-tests, benchmarks ]
     if: ${{ !cancelled() }}
-    strategy:
-      fail-fast: false
-      matrix:
-        build_type: [ debug, release ]
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v3
-        with:
-          submodules: false
 
-      - name: Create Allure report
-        id: create-allure-report
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Create Allure report (debug)
+        if: ${{ !cancelled() }}
+        id: create-allure-report-debug
         uses: ./.github/actions/allure-report
         with:
           action: generate
-          build_type: ${{ matrix.build_type }}
+          build_type: debug
+
+      - name: Create Allure report (release)
+        if: ${{ !cancelled() }}
+        id: create-allure-report-release
+        uses: ./.github/actions/allure-report
+        with:
+          action: generate
+          build_type: release
+
+      - uses: actions/github-script@v6
+        if: >
+          !cancelled() &&
+          github.event_name == 'pull_request' && (
+            steps.create-allure-report-debug.outputs.report-url ||
+            steps.create-allure-report-release.outputs.report-url
+          )
+        with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
+          script: |
+            const reports = [{
+              buildType: "debug",
+              reportUrl: "${{ steps.create-allure-report-debug.outputs.report-url }}",
+              jsonUrl:   "${{ steps.create-allure-report-debug.outputs.report-json-url }}",
+            }, {
+              buildType: "release",
+              reportUrl: "${{ steps.create-allure-report-release.outputs.report-url }}",
+              jsonUrl:   "${{ steps.create-allure-report-release.outputs.report-json-url }}",
+            }]
+
+            const script = require("./scripts/pr-comment-test-report.js")
+            await script({
+              github,
+              context,
+              fetch,
+              reports,
+            })
 
       - name: Store Allure test stat in the DB
-        if: ${{ steps.create-allure-report.outputs.report-url }}
+        if: >
+          !cancelled() && (
+            steps.create-allure-report-debug.outputs.report-url ||
+            steps.create-allure-report-release.outputs.report-url
+          )
         env:
-          BUILD_TYPE: ${{ matrix.build_type }}
           SHA: ${{ github.event.pull_request.head.sha || github.sha }}
-          REPORT_URL: ${{ steps.create-allure-report.outputs.report-url }}
+          REPORT_JSON_URL_DEBUG: ${{ steps.create-allure-report-debug.outputs.report-json-url }}
+          REPORT_JSON_URL_RELEASE: ${{ steps.create-allure-report-release.outputs.report-json-url }}
           TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
         run: |
-          curl --fail --output suites.json ${REPORT_URL%/index.html}/data/suites.json
           ./scripts/pysync
 
-          DATABASE_URL="$TEST_RESULT_CONNSTR" poetry run python3 scripts/ingest_regress_test_result.py --revision ${SHA} --reference ${GITHUB_REF} --build-type ${BUILD_TYPE} --ingest suites.json
+          for report_url in $REPORT_JSON_URL_DEBUG $REPORT_JSON_URL_RELEASE; do
+            if [ -z "$report_url" ]; then
+              continue
+            fi
+
+            if [[ "$report_url" == "$REPORT_JSON_URL_DEBUG" ]]; then
+              BUILD_TYPE=debug
+            else
+              BUILD_TYPE=release
+            fi
+
+            curl --fail --output suites.json "${report_url}"
+            DATABASE_URL="$TEST_RESULT_CONNSTR" poetry run python3 scripts/ingest_regress_test_result.py --revision ${SHA} --reference ${GITHUB_REF} --build-type ${BUILD_TYPE} --ingest suites.json
+          done
 
   coverage-report:
     runs-on: [ self-hosted, gen3, small ]
@@ -891,6 +949,16 @@ jobs:
     needs: [ push-docker-hub, tag, regress-tests ]
     if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
     steps:
+      - name: Fix git ownership
+        run: |
+          # Workaround for `fatal: detected dubious ownership in repository at ...`
+          #
+          # Use both ${{ github.workspace }} and ${GITHUB_WORKSPACE} because they're different on host and in containers
+          #   Ref https://github.com/actions/checkout/issues/785
+          #
+          git config --global --add safe.directory ${{ github.workspace }}
+          git config --global --add safe.directory ${GITHUB_WORKSPACE}
+
       - name: Checkout
         uses: actions/checkout@v3
         with:

.github/workflows/neon_extra_builds.yml

@@ -53,14 +53,14 @@ jobs:
         uses: actions/cache@v3
         with:
           path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+          key: v1-${{ runner.os }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
         uses: actions/cache@v3
         with:
           path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+          key: v1-${{ runner.os }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
 
       - name: Set extra env for macOS
         run: |

.github/workflows/release.yml

@@ -31,4 +31,3 @@ jobs:
         head: releases/${{ steps.date.outputs.date }}
         base: release
         title: Release ${{ steps.date.outputs.date }}
-        team_reviewers: release

Cargo.lock

@@ -2474,6 +2474,7 @@ dependencies = [
  "strum",
  "strum_macros",
  "svg_fmt",
+ "sync_wrapper",
  "tempfile",
  "tenant_size_model",
  "thiserror",
@@ -3085,6 +3086,7 @@ dependencies = [
  "serde",
  "serde_json",
  "tempfile",
+ "test-context",
  "tokio",
  "tokio-util",
  "toml_edit",
@@ -3888,6 +3890,27 @@ dependencies = [
  "winapi-util",
 ]
 
+[[package]]
+name = "test-context"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "055831a02a4f5aa28fede67f2902014273eb8c21b958ac5ebbd59b71ef30dbc3"
+dependencies = [
+ "async-trait",
+ "futures",
+ "test-context-macros",
+]
+
+[[package]]
+name = "test-context-macros"
+version = "0.1.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8901a55b0a7a06ebc4a674dcca925170da8e613fa3b163a1df804ed10afb154d"
+dependencies = [
+ "quote",
+ "syn",
+]
+
 [[package]]
 name = "textwrap"
 version = "0.16.0"
@@ -4534,6 +4557,7 @@ dependencies = [
  "once_cell",
  "pin-project-lite",
  "rand",
+ "regex",
  "routerify",
  "sentry",
  "serde",

Cargo.toml

@@ -97,6 +97,7 @@ strum_macros = "0.24"
 svg_fmt = "0.4.1"
 sync_wrapper = "0.1.2"
 tar = "0.4"
+test-context = "0.1"
 thiserror = "1.0"
 tls-listener = { version = "0.6", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }

Dockerfile.compute-node

@@ -38,6 +38,7 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/insert_username.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/intagg.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/moddatetime.control && \
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/pg_stat_statements.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrowlocks.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgstattuple.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/refint.control && \

README.md

@@ -40,6 +40,8 @@ pacman -S base-devel readline zlib libseccomp openssl clang \
 postgresql-libs cmake postgresql protobuf
 ```
 
+Building Neon requires 3.15+ version of `protoc` (protobuf-compiler). If your distribution provides an older version, you can install a newer version from [here](https://github.com/protocolbuffers/protobuf/releases).
+
 2. [Install Rust](https://www.rust-lang.org/tools/install)
 ```
 # recommended approach from https://www.rust-lang.org/tools/install

compute_tools/src/bin/compute_ctl.rs

@@ -203,13 +203,14 @@ fn main() -> Result<()> {
     if delay_exit {
         info!("giving control plane 30s to collect the error before shutdown");
         thread::sleep(Duration::from_secs(30));
-        info!("shutting down");
     }
 
+    info!("shutting down tracing");
     // Shutdown trace pipeline gracefully, so that it has a chance to send any
     // pending traces before we exit.
     tracing_utils::shutdown_tracing();
 
+    info!("shutting down");
     exit(exit_code.unwrap_or(1))
 }
 

compute_tools/src/pg_helpers.rs

@@ -74,18 +74,9 @@ impl GenericOption {
     /// Represent `GenericOption` as configuration option.
     pub fn to_pg_setting(&self) -> String {
         if let Some(val) = &self.value {
-            // TODO: check in the console DB that we don't have these settings
-            // set for any non-deleted project and drop this override.
-            let name = match self.name.as_str() {
-                "safekeepers" => "neon.safekeepers",
-                "wal_acceptor_reconnect" => "neon.safekeeper_reconnect_timeout",
-                "wal_acceptor_connection_timeout" => "neon.safekeeper_connection_timeout",
-                it => it,
-            };
-
             match self.vartype.as_ref() {
-                "string" => format!("{} = '{}'", name, escape_conf_value(val)),
-                _ => format!("{} = {}", name, val),
+                "string" => format!("{} = '{}'", self.name, escape_conf_value(val)),
+                _ => format!("{} = {}", self.name, val),
             }
         } else {
             self.name.to_owned()

control_plane/src/compute.rs

@@ -87,11 +87,9 @@ impl ComputeControlPlane {
             address: SocketAddr::new("127.0.0.1".parse().unwrap(), port),
             env: self.env.clone(),
             pageserver: Arc::clone(&self.pageserver),
-            is_test: false,
             timeline_id,
             lsn,
             tenant_id,
-            uses_wal_proposer: false,
             pg_version,
         });
 
@@ -113,11 +111,9 @@ pub struct PostgresNode {
     name: String,
     pub env: LocalEnv,
     pageserver: Arc<PageServerNode>,
-    is_test: bool,
     pub timeline_id: TimelineId,
     pub lsn: Option<Lsn>, // if it's a read-only node. None for primary
     pub tenant_id: TenantId,
-    uses_wal_proposer: bool,
     pg_version: u32,
 }
 
@@ -151,7 +147,6 @@ impl PostgresNode {
         let port: u16 = conf.parse_field("port", &context)?;
         let timeline_id: TimelineId = conf.parse_field("neon.timeline_id", &context)?;
         let tenant_id: TenantId = conf.parse_field("neon.tenant_id", &context)?;
-        let uses_wal_proposer = conf.get("neon.safekeepers").is_some();
 
         // Read postgres version from PG_VERSION file to determine which postgres version binary to use.
         // If it doesn't exist, assume broken data directory and use default pg version.
@@ -171,11 +166,9 @@ impl PostgresNode {
             name,
             env: env.clone(),
             pageserver: Arc::clone(pageserver),
-            is_test: false,
             timeline_id,
             lsn: recovery_target_lsn,
             tenant_id,
-            uses_wal_proposer,
             pg_version,
         })
     }
@@ -367,7 +360,7 @@ impl PostgresNode {
     fn load_basebackup(&self, auth_token: &Option<String>) -> Result<()> {
         let backup_lsn = if let Some(lsn) = self.lsn {
             Some(lsn)
-        } else if self.uses_wal_proposer {
+        } else if !self.env.safekeepers.is_empty() {
             // LSN 0 means that it is bootstrap and we need to download just
             // latest data from the pageserver. That is a bit clumsy but whole bootstrap
             // procedure evolves quite actively right now, so let's think about it again
@@ -406,7 +399,7 @@ impl PostgresNode {
 
     fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
         let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
-        let mut cmd = Command::new(pg_ctl_path);
+        let mut cmd = Command::new(&pg_ctl_path);
         cmd.args(
             [
                 &[
@@ -435,7 +428,9 @@ impl PostgresNode {
             cmd.env("NEON_AUTH_TOKEN", token);
         }
 
-        let pg_ctl = cmd.output().context("pg_ctl failed")?;
+        let pg_ctl = cmd
+            .output()
+            .context(format!("{} failed", pg_ctl_path.display()))?;
         if !pg_ctl.status.success() {
             anyhow::bail!(
                 "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",
@@ -480,10 +475,6 @@ impl PostgresNode {
         self.pg_ctl(&["start"], auth_token)
     }
 
-    pub fn restart(&self, auth_token: &Option<String>) -> Result<()> {
-        self.pg_ctl(&["restart"], auth_token)
-    }
-
     pub fn stop(&self, destroy: bool) -> Result<()> {
         // If we are going to destroy data directory,
         // use immediate shutdown mode, otherwise,
@@ -514,26 +505,4 @@ impl PostgresNode {
             "postgres"
         )
     }
-
-    // XXX: cache that in control plane
-    pub fn whoami(&self) -> String {
-        let output = Command::new("whoami")
-            .output()
-            .expect("failed to execute whoami");
-
-        assert!(output.status.success(), "whoami failed");
-
-        String::from_utf8(output.stdout).unwrap().trim().to_string()
-    }
-}
-
-impl Drop for PostgresNode {
-    // destructor to clean up state after test is done
-    // XXX: we may detect failed test by setting some flag in catch_unwind()
-    // and checking it here. But let just clean datadirs on start.
-    fn drop(&mut self) {
-        if self.is_test {
-            let _ = self.stop(true);
-        }
-    }
 }

control_plane/src/pageserver.rs

@@ -363,6 +363,11 @@ impl PageServerNode {
                 .map(|x| serde_json::from_str(x))
                 .transpose()
                 .context("Failed to parse 'eviction_policy' json")?,
+            min_resident_size_override: settings
+                .remove("min_resident_size_override")
+                .map(|x| x.parse::<u64>())
+                .transpose()
+                .context("Failed to parse 'min_resident_size_override' as integer")?,
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -435,6 +440,11 @@ impl PageServerNode {
                     .map(|x| serde_json::from_str(x))
                     .transpose()
                     .context("Failed to parse 'eviction_policy' json")?,
+                min_resident_size_override: settings
+                    .get("min_resident_size_override")
+                    .map(|x| x.parse::<u64>())
+                    .transpose()
+                    .context("Failed to parse 'min_resident_size_override' as an integer")?,
             })
             .send()?
             .error_from_body()?;

control_plane/src/safekeeper.rs

@@ -156,7 +156,7 @@ impl SafekeeperNode {
         }
 
         background_process::start_process(
-            &format!("safekeeper {id}"),
+            &format!("safekeeper-{id}"),
             &datadir,
             &self.env.safekeeper_bin(),
             &args,

libs/pageserver_api/src/models.rs

@@ -120,6 +120,7 @@ pub struct TenantCreateRequest {
     // We might do that once the eviction feature has stabilizied.
     // For now, this field is not even documented in the openapi_spec.yml.
     pub eviction_policy: Option<serde_json::Value>,
+    pub min_resident_size_override: Option<u64>,
 }
 
 #[serde_as]
@@ -165,6 +166,7 @@ pub struct TenantConfigRequest {
     // We might do that once the eviction feature has stabilizied.
     // For now, this field is not even documented in the openapi_spec.yml.
     pub eviction_policy: Option<serde_json::Value>,
+    pub min_resident_size_override: Option<u64>,
 }
 
 impl TenantConfigRequest {
@@ -185,6 +187,7 @@ impl TenantConfigRequest {
             max_lsn_wal_lag: None,
             trace_read_requests: None,
             eviction_policy: None,
+            min_resident_size_override: None,
         }
     }
 }

libs/pq_proto/src/lib.rs

@@ -936,35 +936,40 @@ impl<'a> BeMessage<'a> {
     }
 }
 
-// Neon extension of postgres replication protocol
-// See NEON_STATUS_UPDATE_TAG_BYTE
+/// Feedback pageserver sends to safekeeper and safekeeper resends to compute.
+/// Serialized in custom flexible key/value format. In replication protocol, it
+/// is marked with NEON_STATUS_UPDATE_TAG_BYTE to differentiate from postgres
+/// Standby status update / Hot standby feedback messages.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-pub struct ReplicationFeedback {
-    // Last known size of the timeline. Used to enforce timeline size limit.
+pub struct PageserverFeedback {
+    /// Last known size of the timeline. Used to enforce timeline size limit.
     pub current_timeline_size: u64,
-    // Parts of StandbyStatusUpdate we resend to compute via safekeeper
-    pub ps_writelsn: u64,
-    pub ps_applylsn: u64,
-    pub ps_flushlsn: u64,
-    pub ps_replytime: SystemTime,
+    /// LSN last received and ingested by the pageserver.
+    pub last_received_lsn: u64,
+    /// LSN up to which data is persisted by the pageserver to its local disc.
+    pub disk_consistent_lsn: u64,
+    /// LSN up to which data is persisted by the pageserver on s3; safekeepers
+    /// consider WAL before it can be removed.
+    pub remote_consistent_lsn: u64,
+    pub replytime: SystemTime,
 }
 
-// NOTE: Do not forget to increment this number when adding new fields to ReplicationFeedback.
+// NOTE: Do not forget to increment this number when adding new fields to PageserverFeedback.
 // Do not remove previously available fields because this might be backwards incompatible.
-pub const REPLICATION_FEEDBACK_FIELDS_NUMBER: u8 = 5;
+pub const PAGESERVER_FEEDBACK_FIELDS_NUMBER: u8 = 5;
 
-impl ReplicationFeedback {
-    pub fn empty() -> ReplicationFeedback {
-        ReplicationFeedback {
+impl PageserverFeedback {
+    pub fn empty() -> PageserverFeedback {
+        PageserverFeedback {
             current_timeline_size: 0,
-            ps_writelsn: 0,
-            ps_applylsn: 0,
-            ps_flushlsn: 0,
-            ps_replytime: SystemTime::now(),
+            last_received_lsn: 0,
+            remote_consistent_lsn: 0,
+            disk_consistent_lsn: 0,
+            replytime: SystemTime::now(),
         }
     }
 
-    // Serialize ReplicationFeedback using custom format
+    // Serialize PageserverFeedback using custom format
     // to support protocol extensibility.
     //
     // Following layout is used:
@@ -974,24 +979,26 @@ impl ReplicationFeedback {
     // null-terminated string - key,
     // uint32 - value length in bytes
     // value itself
+    //
+    // TODO: change serialized fields names once all computes migrate to rename.
     pub fn serialize(&self, buf: &mut BytesMut) {
-        buf.put_u8(REPLICATION_FEEDBACK_FIELDS_NUMBER); // # of keys
+        buf.put_u8(PAGESERVER_FEEDBACK_FIELDS_NUMBER); // # of keys
         buf.put_slice(b"current_timeline_size\0");
         buf.put_i32(8);
         buf.put_u64(self.current_timeline_size);
 
         buf.put_slice(b"ps_writelsn\0");
         buf.put_i32(8);
-        buf.put_u64(self.ps_writelsn);
+        buf.put_u64(self.last_received_lsn);
         buf.put_slice(b"ps_flushlsn\0");
         buf.put_i32(8);
-        buf.put_u64(self.ps_flushlsn);
+        buf.put_u64(self.disk_consistent_lsn);
         buf.put_slice(b"ps_applylsn\0");
         buf.put_i32(8);
-        buf.put_u64(self.ps_applylsn);
+        buf.put_u64(self.remote_consistent_lsn);
 
         let timestamp = self
-            .ps_replytime
+            .replytime
             .duration_since(*PG_EPOCH)
             .expect("failed to serialize pg_replytime earlier than PG_EPOCH")
             .as_micros() as i64;
@@ -1001,9 +1008,10 @@ impl ReplicationFeedback {
         buf.put_i64(timestamp);
     }
 
-    // Deserialize ReplicationFeedback message
-    pub fn parse(mut buf: Bytes) -> ReplicationFeedback {
-        let mut rf = ReplicationFeedback::empty();
+    // Deserialize PageserverFeedback message
+    // TODO: change serialized fields names once all computes migrate to rename.
+    pub fn parse(mut buf: Bytes) -> PageserverFeedback {
+        let mut rf = PageserverFeedback::empty();
         let nfields = buf.get_u8();
         for _ in 0..nfields {
             let key = read_cstr(&mut buf).unwrap();
@@ -1016,39 +1024,39 @@ impl ReplicationFeedback {
                 b"ps_writelsn" => {
                     let len = buf.get_i32();
                     assert_eq!(len, 8);
-                    rf.ps_writelsn = buf.get_u64();
+                    rf.last_received_lsn = buf.get_u64();
                 }
                 b"ps_flushlsn" => {
                     let len = buf.get_i32();
                     assert_eq!(len, 8);
-                    rf.ps_flushlsn = buf.get_u64();
+                    rf.disk_consistent_lsn = buf.get_u64();
                 }
                 b"ps_applylsn" => {
                     let len = buf.get_i32();
                     assert_eq!(len, 8);
-                    rf.ps_applylsn = buf.get_u64();
+                    rf.remote_consistent_lsn = buf.get_u64();
                 }
                 b"ps_replytime" => {
                     let len = buf.get_i32();
                     assert_eq!(len, 8);
                     let raw_time = buf.get_i64();
                     if raw_time > 0 {
-                        rf.ps_replytime = *PG_EPOCH + Duration::from_micros(raw_time as u64);
+                        rf.replytime = *PG_EPOCH + Duration::from_micros(raw_time as u64);
                     } else {
-                        rf.ps_replytime = *PG_EPOCH - Duration::from_micros(-raw_time as u64);
+                        rf.replytime = *PG_EPOCH - Duration::from_micros(-raw_time as u64);
                     }
                 }
                 _ => {
                     let len = buf.get_i32();
                     warn!(
-                        "ReplicationFeedback parse. unknown key {} of len {len}. Skip it.",
+                        "PageserverFeedback parse. unknown key {} of len {len}. Skip it.",
                         String::from_utf8_lossy(key.as_ref())
                     );
                     buf.advance(len as usize);
                 }
             }
         }
-        trace!("ReplicationFeedback parsed is {:?}", rf);
+        trace!("PageserverFeedback parsed is {:?}", rf);
         rf
     }
 }
@@ -1059,33 +1067,33 @@ mod tests {
 
     #[test]
     fn test_replication_feedback_serialization() {
-        let mut rf = ReplicationFeedback::empty();
+        let mut rf = PageserverFeedback::empty();
         // Fill rf with some values
         rf.current_timeline_size = 12345678;
         // Set rounded time to be able to compare it with deserialized value,
         // because it is rounded up to microseconds during serialization.
-        rf.ps_replytime = *PG_EPOCH + Duration::from_secs(100_000_000);
+        rf.replytime = *PG_EPOCH + Duration::from_secs(100_000_000);
         let mut data = BytesMut::new();
         rf.serialize(&mut data);
 
-        let rf_parsed = ReplicationFeedback::parse(data.freeze());
+        let rf_parsed = PageserverFeedback::parse(data.freeze());
         assert_eq!(rf, rf_parsed);
     }
 
     #[test]
     fn test_replication_feedback_unknown_key() {
-        let mut rf = ReplicationFeedback::empty();
+        let mut rf = PageserverFeedback::empty();
         // Fill rf with some values
         rf.current_timeline_size = 12345678;
         // Set rounded time to be able to compare it with deserialized value,
         // because it is rounded up to microseconds during serialization.
-        rf.ps_replytime = *PG_EPOCH + Duration::from_secs(100_000_000);
+        rf.replytime = *PG_EPOCH + Duration::from_secs(100_000_000);
         let mut data = BytesMut::new();
         rf.serialize(&mut data);
 
         // Add an extra field to the buffer and adjust number of keys
         if let Some(first) = data.first_mut() {
-            *first = REPLICATION_FEEDBACK_FIELDS_NUMBER + 1;
+            *first = PAGESERVER_FEEDBACK_FIELDS_NUMBER + 1;
         }
 
         data.put_slice(b"new_field_one\0");
@@ -1093,7 +1101,7 @@ mod tests {
         data.put_u64(42);
 
         // Parse serialized data and check that new field is not parsed
-        let rf_parsed = ReplicationFeedback::parse(data.freeze());
+        let rf_parsed = PageserverFeedback::parse(data.freeze());
         assert_eq!(rf, rf_parsed);
     }
 

libs/remote_storage/Cargo.toml

@@ -26,3 +26,4 @@ workspace_hack.workspace = true
 
 [dev-dependencies]
 tempfile.workspace = true
+test-context.workspace = true

libs/remote_storage/src/lib.rs

@@ -39,6 +39,9 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
+/// No limits on the client side, which currenltly means 1000 for AWS S3.
+/// https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax
+pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 
@@ -64,6 +67,10 @@ impl RemotePath {
     pub fn object_name(&self) -> Option<&str> {
         self.0.file_name().and_then(|os_str| os_str.to_str())
     }
+
+    pub fn join(&self, segment: &Path) -> Self {
+        Self(self.0.join(segment))
+    }
 }
 
 /// Storage (potentially remote) API to manage its state.
@@ -71,9 +78,6 @@ impl RemotePath {
 /// providing basic CRUD operations for storage files.
 #[async_trait::async_trait]
 pub trait RemoteStorage: Send + Sync + 'static {
-    /// Lists all items the storage has right now.
-    async fn list(&self) -> anyhow::Result<Vec<RemotePath>>;
-
     /// Lists all top level subdirectories for a given prefix
     /// Note: here we assume that if the prefix is passed it was obtained via remote_object_id
     /// which already takes into account any kind of global prefix (prefix_in_bucket for S3 or storage_root for LocalFS)
@@ -266,6 +270,7 @@ pub struct S3Config {
     /// AWS S3 has various limits on its API calls, we need not to exceed those.
     /// See [`DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT`] for more details.
     pub concurrency_limit: NonZeroUsize,
+    pub max_keys_per_list_response: Option<i32>,
 }
 
 impl Debug for S3Config {
@@ -275,6 +280,10 @@ impl Debug for S3Config {
             .field("bucket_region", &self.bucket_region)
             .field("prefix_in_bucket", &self.prefix_in_bucket)
             .field("concurrency_limit", &self.concurrency_limit)
+            .field(
+                "max_keys_per_list_response",
+                &self.max_keys_per_list_response,
+            )
             .finish()
     }
 }
@@ -303,6 +312,11 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
+        let max_keys_per_list_response =
+            parse_optional_integer::<i32, _>("max_keys_per_list_response", toml)
+                .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
+                .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
+
         let storage = match (local_path, bucket_name, bucket_region) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
             (None, None, None) => return Ok(None),
@@ -324,6 +338,7 @@ impl RemoteStorageConfig {
                     .map(|endpoint| parse_toml_string("endpoint", endpoint))
                     .transpose()?,
                 concurrency_limit,
+                max_keys_per_list_response,
             }),
             (Some(local_path), None, None) => RemoteStorageKind::LocalFs(PathBuf::from(
                 parse_toml_string("local_path", local_path)?,

libs/remote_storage/src/local_fs.rs

@@ -73,10 +73,8 @@ impl LocalFs {
             Ok(None)
         }
     }
-}
 
-#[async_trait::async_trait]
-impl RemoteStorage for LocalFs {
+    #[cfg(test)]
     async fn list(&self) -> anyhow::Result<Vec<RemotePath>> {
         Ok(get_all_files(&self.storage_root, true)
             .await?
@@ -91,7 +89,10 @@ impl RemoteStorage for LocalFs {
             })
             .collect())
     }
+}
 
+#[async_trait::async_trait]
+impl RemoteStorage for LocalFs {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,

libs/remote_storage/src/s3_bucket.rs

@@ -102,6 +102,7 @@ pub struct S3Bucket {
     client: Client,
     bucket_name: String,
     prefix_in_bucket: Option<String>,
+    max_keys_per_list_response: Option<i32>,
     // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
     // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
     // The helps to ensure we don't exceed the thresholds.
@@ -164,6 +165,7 @@ impl S3Bucket {
         Ok(Self {
             client,
             bucket_name: aws_config.bucket_name.clone(),
+            max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
             concurrency_limiter: Arc::new(Semaphore::new(aws_config.concurrency_limit.get())),
         })
@@ -273,48 +275,6 @@ impl<S: AsyncRead> AsyncRead for RatelimitedAsyncRead<S> {
 
 #[async_trait::async_trait]
 impl RemoteStorage for S3Bucket {
-    async fn list(&self) -> anyhow::Result<Vec<RemotePath>> {
-        let mut document_keys = Vec::new();
-
-        let mut continuation_token = None;
-        loop {
-            let _guard = self
-                .concurrency_limiter
-                .acquire()
-                .await
-                .context("Concurrency limiter semaphore got closed during S3 list")?;
-
-            metrics::inc_list_objects();
-
-            let fetch_response = self
-                .client
-                .list_objects_v2()
-                .bucket(self.bucket_name.clone())
-                .set_prefix(self.prefix_in_bucket.clone())
-                .set_continuation_token(continuation_token)
-                .send()
-                .await
-                .map_err(|e| {
-                    metrics::inc_list_objects_fail();
-                    e
-                })?;
-            document_keys.extend(
-                fetch_response
-                    .contents
-                    .unwrap_or_default()
-                    .into_iter()
-                    .filter_map(|o| Some(self.s3_object_to_relative_path(o.key()?))),
-            );
-
-            match fetch_response.continuation_token {
-                Some(new_token) => continuation_token = Some(new_token),
-                None => break,
-            }
-        }
-
-        Ok(document_keys)
-    }
-
     /// See the doc for `RemoteStorage::list_prefixes`
     /// Note: it wont include empty "directories"
     async fn list_prefixes(
@@ -354,6 +314,7 @@ impl RemoteStorage for S3Bucket {
                 .set_prefix(list_prefix.clone())
                 .set_continuation_token(continuation_token)
                 .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string())
+                .set_max_keys(self.max_keys_per_list_response)
                 .send()
                 .await
                 .map_err(|e| {
@@ -371,7 +332,7 @@ impl RemoteStorage for S3Bucket {
                     .filter_map(|o| Some(self.s3_object_to_relative_path(o.prefix()?))),
             );
 
-            match fetch_response.continuation_token {
+            match fetch_response.next_continuation_token {
                 Some(new_token) => continuation_token = Some(new_token),
                 None => break,
             }

libs/remote_storage/src/simulate_failures.rs

@@ -20,7 +20,6 @@ pub struct UnreliableWrapper {
 /// Used to identify retries of different unique operation.
 #[derive(Debug, Hash, Eq, PartialEq)]
 enum RemoteOp {
-    List,
     ListPrefixes(Option<RemotePath>),
     Upload(RemotePath),
     Download(RemotePath),
@@ -75,12 +74,6 @@ impl UnreliableWrapper {
 
 #[async_trait::async_trait]
 impl RemoteStorage for UnreliableWrapper {
-    /// Lists all items the storage has right now.
-    async fn list(&self) -> anyhow::Result<Vec<RemotePath>> {
-        self.attempt(RemoteOp::List)?;
-        self.inner.list().await
-    }
-
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,

libs/remote_storage/tests/pagination_tests.rs

@@ -0,0 +1,275 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use remote_storage::{
+    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
+
+/// Tests that S3 client can list all prefixes, even if the response come paginated and requires multiple S3 queries.
+/// Uses real S3 and requires [`ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME`] and related S3 cred env vars specified.
+/// See the client creation in [`create_s3_client`] for details on the required env vars.
+/// If real S3 tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of S3 objects with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_s3_data`]
+/// where
+/// * `random_prefix_part` is set for the entire S3 client during the S3 client creation in [`create_s3_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real S3 enabled and `#[cfg(test)]` Rust configuration used, the S3 client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to S3,
+/// since current default AWS S3 pagination limit is 1000.
+/// (see https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax)
+///
+/// Lastly, the test attempts to clean up and remove all uploaded S3 files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledS3)]
+#[tokio::test]
+async fn s3_pagination_should_work(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledS3::Enabled(ctx) => ctx,
+        MaybeEnabledS3::Disabled => return Ok(()),
+        MaybeEnabledS3::UploadsFailed(e, _) => anyhow::bail!("S3 init failed: {e:?}"),
+    };
+
+    let test_client = Arc::clone(&ctx.client_with_excessive_pagination);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix =
+        RemotePath::new(Path::new(ctx.base_prefix_str)).context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+enum MaybeEnabledS3 {
+    Enabled(S3WithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, S3WithTestBlobs),
+}
+
+struct S3WithTestBlobs {
+    client_with_excessive_pagination: Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledS3 {
+    async fn setup() -> Self {
+        utils::logging::init(utils::logging::LogFormat::Test).expect("logging init failed");
+        if env::var(ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let client_with_excessive_pagination = create_s3_client(max_keys_in_list_response)
+            .context("S3 client creation")
+            .expect("S3 client creation failed");
+
+        let base_prefix_str = "test/";
+        match upload_s3_data(
+            &client_with_excessive_pagination,
+            base_prefix_str,
+            upload_tasks_count,
+        )
+        .await
+        {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+                Self::Enabled(S3WithTestBlobs {
+                    client_with_excessive_pagination,
+                    base_prefix_str,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to S3"),
+                S3WithTestBlobs {
+                    client_with_excessive_pagination,
+                    base_prefix_str,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.client_with_excessive_pagination, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_s3_client(max_keys_per_list_response: i32) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    let remote_storage_s3_bucket = env::var("REMOTE_STORAGE_S3_BUCKET")
+        .context("`REMOTE_STORAGE_S3_BUCKET` env var is not set, but real S3 tests are enabled")?;
+    let remote_storage_s3_region = env::var("REMOTE_STORAGE_S3_REGION")
+        .context("`REMOTE_STORAGE_S3_REGION` env var is not set, but real S3 tests are enabled")?;
+    let random_prefix_part = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random s3 test prefix part calculation")?
+        .as_millis();
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AwsS3(S3Config {
+            bucket_name: remote_storage_s3_bucket,
+            bucket_region: remote_storage_s3_region,
+            prefix_in_bucket: Some(format!("pagination_should_work_test_{random_prefix_part}/")),
+            endpoint: None,
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response: Some(max_keys_per_list_response),
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_s3_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} S3 files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = PathBuf::from(format!("{base_prefix_str}/sub_prefix_{i}/"));
+            let blob_prefix = RemotePath::new(&prefix)
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(
+                    Box::new(std::io::Cursor::new(data)),
+                    data_len,
+                    &blob_path,
+                    None,
+                )
+                .await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}

libs/utils/Cargo.toml

@@ -19,6 +19,7 @@ jsonwebtoken.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 pin-project-lite.workspace = true
+regex.workspace = true
 routerify.workspace = true
 serde.workspace = true
 serde_json.workspace = true

libs/utils/src/http/error.rs

@@ -20,6 +20,9 @@ pub enum ApiError {
     #[error("Conflict: {0}")]
     Conflict(String),
 
+    #[error("Precondition failed: {0}")]
+    PreconditionFailed(&'static str),
+
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -44,6 +47,10 @@ impl ApiError {
             ApiError::Conflict(_) => {
                 HttpErrorBody::response_from_msg_and_status(self.to_string(), StatusCode::CONFLICT)
             }
+            ApiError::PreconditionFailed(_) => HttpErrorBody::response_from_msg_and_status(
+                self.to_string(),
+                StatusCode::PRECONDITION_FAILED,
+            ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                 err.to_string(),
                 StatusCode::INTERNAL_SERVER_ERROR,

libs/utils/src/id.rs

@@ -23,7 +23,7 @@ pub enum IdError {
 struct Id([u8; 16]);
 
 impl Id {
-    pub fn get_from_buf(buf: &mut dyn bytes::Buf) -> Id {
+    pub fn get_from_buf(buf: &mut impl bytes::Buf) -> Id {
         let mut arr = [0u8; 16];
         buf.copy_to_slice(&mut arr);
         Id::from(arr)
@@ -112,7 +112,7 @@ impl fmt::Debug for Id {
 macro_rules! id_newtype {
     ($t:ident) => {
         impl $t {
-            pub fn get_from_buf(buf: &mut dyn bytes::Buf) -> $t {
+            pub fn get_from_buf(buf: &mut impl bytes::Buf) -> $t {
                 $t(Id::get_from_buf(buf))
             }
 

libs/utils/src/lib.rs

@@ -51,6 +51,9 @@ pub mod history_buffer;
 
 pub mod measured_stream;
 
+pub mod serde_percent;
+pub mod serde_regex;
+
 /// use with fail::cfg("$name", "return(2000)")
 #[macro_export]
 macro_rules! failpoint_sleep_millis_async {

libs/utils/src/serde_percent.rs

@@ -0,0 +1,91 @@
+//! A serde::Deserialize type for percentages.
+//!
+//! See [`Percent`] for details.
+
+use serde::{Deserialize, Serialize};
+
+/// If the value is not an integer between 0 and 100,
+/// deserialization fails with a descriptive error.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
+#[serde(transparent)]
+pub struct Percent(#[serde(deserialize_with = "deserialize_pct_0_to_100")] u8);
+
+impl Percent {
+    pub const fn new(pct: u8) -> Option<Self> {
+        if pct <= 100 {
+            Some(Percent(pct))
+        } else {
+            None
+        }
+    }
+
+    pub fn get(&self) -> u8 {
+        self.0
+    }
+}
+
+fn deserialize_pct_0_to_100<'de, D>(deserializer: D) -> Result<u8, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let v: u8 = serde::de::Deserialize::deserialize(deserializer)?;
+    if v > 100 {
+        return Err(serde::de::Error::custom(
+            "must be an integer between 0 and 100",
+        ));
+    }
+    Ok(v)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::Percent;
+
+    #[derive(serde::Deserialize, serde::Serialize, Debug, PartialEq, Eq)]
+    struct Foo {
+        bar: Percent,
+    }
+
+    #[test]
+    fn basics() {
+        let input = r#"{ "bar": 50 }"#;
+        let foo: Foo = serde_json::from_str(input).unwrap();
+        assert_eq!(foo.bar.get(), 50);
+    }
+    #[test]
+    fn null_handling() {
+        let input = r#"{ "bar": null }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn zero() {
+        let input = r#"{ "bar": 0 }"#;
+        let foo: Foo = serde_json::from_str(input).unwrap();
+        assert_eq!(foo.bar.get(), 0);
+    }
+    #[test]
+    fn out_of_range_above() {
+        let input = r#"{ "bar": 101 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn out_of_range_below() {
+        let input = r#"{ "bar": -1 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn float() {
+        let input = r#"{ "bar": 50.5 }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+    #[test]
+    fn string() {
+        let input = r#"{ "bar": "50 %" }"#;
+        let res: Result<Foo, _> = serde_json::from_str(input);
+        assert!(res.is_err());
+    }
+}

libs/utils/src/serde_regex.rs

@@ -0,0 +1,60 @@
+//! A `serde::{Deserialize,Serialize}` type for regexes.
+
+use std::ops::Deref;
+
+#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
+#[serde(transparent)]
+pub struct Regex(
+    #[serde(
+        deserialize_with = "deserialize_regex",
+        serialize_with = "serialize_regex"
+    )]
+    regex::Regex,
+);
+
+fn deserialize_regex<'de, D>(deserializer: D) -> Result<regex::Regex, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+{
+    let s: String = serde::de::Deserialize::deserialize(deserializer)?;
+    let re = regex::Regex::new(&s).map_err(serde::de::Error::custom)?;
+    Ok(re)
+}
+
+fn serialize_regex<S>(re: &regex::Regex, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::ser::Serializer,
+{
+    serializer.collect_str(re.as_str())
+}
+
+impl Deref for Regex {
+    type Target = regex::Regex;
+
+    fn deref(&self) -> &regex::Regex {
+        &self.0
+    }
+}
+
+impl PartialEq for Regex {
+    fn eq(&self, other: &Regex) -> bool {
+        // comparing the automatons would be quite complicated
+        self.as_str() == other.as_str()
+    }
+}
+
+impl Eq for Regex {}
+
+#[cfg(test)]
+mod tests {
+
+    #[test]
+    fn roundtrip() {
+        let input = r#""foo.*bar""#;
+        let re: super::Regex = serde_json::from_str(input).unwrap();
+        assert!(re.is_match("foo123bar"));
+        assert!(!re.is_match("foo"));
+        let output = serde_json::to_string(&re).unwrap();
+        assert_eq!(output, input);
+    }
+}

libs/utils/src/signals.rs

@@ -1,25 +1,7 @@
-use signal_hook::flag;
 use signal_hook::iterator::Signals;
-use std::sync::atomic::AtomicBool;
-use std::sync::Arc;
 
 pub use signal_hook::consts::{signal::*, TERM_SIGNALS};
 
-pub fn install_shutdown_handlers() -> anyhow::Result<ShutdownSignals> {
-    let term_now = Arc::new(AtomicBool::new(false));
-    for sig in TERM_SIGNALS {
-        // When terminated by a second term signal, exit with exit code 1.
-        // This will do nothing the first time (because term_now is false).
-        flag::register_conditional_shutdown(*sig, 1, Arc::clone(&term_now))?;
-        // But this will "arm" the above for the second time, by setting it to true.
-        // The order of registering these is important, if you put this one first, it will
-        // first arm and then terminate ‒ all in the first round.
-        flag::register(*sig, Arc::clone(&term_now))?;
-    }
-
-    Ok(ShutdownSignals)
-}
-
 pub enum Signal {
     Quit,
     Interrupt,
@@ -39,10 +21,7 @@ impl Signal {
 pub struct ShutdownSignals;
 
 impl ShutdownSignals {
-    pub fn handle(
-        self,
-        mut handler: impl FnMut(Signal) -> anyhow::Result<()>,
-    ) -> anyhow::Result<()> {
+    pub fn handle(mut handler: impl FnMut(Signal) -> anyhow::Result<()>) -> anyhow::Result<()> {
         for raw_signal in Signals::new(TERM_SIGNALS)?.into_iter() {
             let signal = match raw_signal {
                 SIGINT => Signal::Interrupt,

pageserver/Cargo.toml

@@ -48,6 +48,7 @@ serde_json = { workspace = true, features = ["raw_value"] }
 serde_with.workspace = true
 signal-hook.workspace = true
 svg_fmt.workspace = true
+sync_wrapper.workspace = true
 tokio-tar.workspace = true
 thiserror.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }

pageserver/src/bin/pageserver.rs

@@ -8,6 +8,7 @@ use anyhow::{anyhow, Context};
 use clap::{Arg, ArgAction, Command};
 use fail::FailScenario;
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
+use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use remote_storage::GenericRemoteStorage;
 use tracing::*;
 
@@ -24,11 +25,9 @@ use pageserver::{
     virtual_file,
 };
 use postgres_backend::AuthType;
+use utils::signals::ShutdownSignals;
 use utils::{
-    auth::JwtAuth,
-    logging, project_git_version,
-    sentry_init::init_sentry,
-    signals::{self, Signal},
+    auth::JwtAuth, logging, project_git_version, sentry_init::init_sentry, signals::Signal,
     tcp_listener,
 };
 
@@ -263,9 +262,6 @@ fn start_pageserver(
     info!("Starting pageserver pg protocol handler on {pg_addr}");
     let pageserver_listener = tcp_listener::bind(pg_addr)?;
 
-    // Install signal handlers
-    let signals = signals::install_shutdown_handlers()?;
-
     // Launch broker client
     WALRECEIVER_RUNTIME.block_on(pageserver::broker_client::init_broker_client(conf))?;
 
@@ -319,14 +315,34 @@ fn start_pageserver(
     // Scan the local 'tenants/' directory and start loading the tenants
     BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(conf, remote_storage.clone()))?;
 
+    // shared state between the disk-usage backed eviction background task and the http endpoint
+    // that allows triggering disk-usage based eviction manually. note that the http endpoint
+    // is still accessible even if background task is not configured as long as remote storage has
+    // been configured.
+    let disk_usage_eviction_state: Arc<disk_usage_eviction_task::State> = Arc::default();
+
+    if let Some(remote_storage) = &remote_storage {
+        launch_disk_usage_global_eviction_task(
+            conf,
+            remote_storage.clone(),
+            disk_usage_eviction_state.clone(),
+        )?;
+    }
+
     // Start up the service to handle HTTP mgmt API request. We created the
     // listener earlier already.
     {
         let _rt_guard = MGMT_REQUEST_RUNTIME.enter();
 
-        let router = http::make_router(conf, launch_ts, http_auth, remote_storage)?
-            .build()
-            .map_err(|err| anyhow!(err))?;
+        let router = http::make_router(
+            conf,
+            launch_ts,
+            http_auth,
+            remote_storage,
+            disk_usage_eviction_state,
+        )?
+        .build()
+        .map_err(|err| anyhow!(err))?;
         let service = utils::http::RouterService::new(router).unwrap();
         let server = hyper::Server::from_tcp(http_listener)?
             .serve(service)
@@ -409,7 +425,7 @@ fn start_pageserver(
     }
 
     // All started up! Now just sit and wait for shutdown signal.
-    signals.handle(|signal| match signal {
+    ShutdownSignals::handle(|signal| match signal {
         Signal::Quit => {
             info!(
                 "Got {}. Terminating in immediate shutdown mode",

pageserver/src/config.rs

@@ -27,6 +27,7 @@ use utils::{
     logging::LogFormat,
 };
 
+use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{TENANT_ATTACHING_MARKER_FILENAME, TIMELINES_SEGMENT_NAME};
@@ -92,6 +93,8 @@ pub mod defaults {
 
 #evictions_low_residence_duration_metric_threshold = '{DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD}'
 
+#disk_usage_based_eviction = {{ max_usage_pct = .., min_avail_bytes = .., period = "10s"}}
+
 # [tenant_config]
 #checkpoint_distance = {DEFAULT_CHECKPOINT_DISTANCE} # in bytes
 #checkpoint_timeout = {DEFAULT_CHECKPOINT_TIMEOUT}
@@ -104,6 +107,8 @@ pub mod defaults {
 #image_creation_threshold = {DEFAULT_IMAGE_CREATION_THRESHOLD}
 #pitr_interval = '{DEFAULT_PITR_INTERVAL}'
 
+#min_resident_size_override = .. # in bytes
+
 # [remote_storage]
 
 "###
@@ -165,6 +170,10 @@ pub struct PageServerConf {
 
     /// Number of concurrent [`Tenant::gather_size_inputs`] allowed.
     pub concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore,
+    /// Limit of concurrent [`Tenant::gather_size_inputs`] issued by module `eviction_task`.
+    /// The number of permits is the same as `concurrent_tenant_size_logical_size_queries`.
+    /// See the comment in `eviction_task` for details.
+    pub eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore,
 
     // How often to collect metrics and send them to the metrics endpoint.
     pub metric_collection_interval: Duration,
@@ -176,6 +185,8 @@ pub struct PageServerConf {
     // See the corresponding metric's help string.
     pub evictions_low_residence_duration_metric_threshold: Duration,
 
+    pub disk_usage_based_eviction: Option<DiskUsageEvictionTaskConfig>,
+
     pub test_remote_failures: u64,
 
     pub ondemand_download_behavior_treat_error_as_warn: bool,
@@ -239,7 +250,7 @@ struct PageServerConfigBuilder {
 
     log_format: BuilderValue<LogFormat>,
 
-    concurrent_tenant_size_logical_size_queries: BuilderValue<ConfigurableSemaphore>,
+    concurrent_tenant_size_logical_size_queries: BuilderValue<NonZeroUsize>,
 
     metric_collection_interval: BuilderValue<Duration>,
     cached_metric_collection_interval: BuilderValue<Duration>,
@@ -248,6 +259,8 @@ struct PageServerConfigBuilder {
 
     evictions_low_residence_duration_metric_threshold: BuilderValue<Duration>,
 
+    disk_usage_based_eviction: BuilderValue<Option<DiskUsageEvictionTaskConfig>>,
+
     test_remote_failures: BuilderValue<u64>,
 
     ondemand_download_behavior_treat_error_as_warn: BuilderValue<bool>,
@@ -286,7 +299,9 @@ impl Default for PageServerConfigBuilder {
             .expect("cannot parse default keepalive interval")),
             log_format: Set(LogFormat::from_str(DEFAULT_LOG_FORMAT).unwrap()),
 
-            concurrent_tenant_size_logical_size_queries: Set(ConfigurableSemaphore::default()),
+            concurrent_tenant_size_logical_size_queries: Set(
+                ConfigurableSemaphore::DEFAULT_INITIAL,
+            ),
             metric_collection_interval: Set(humantime::parse_duration(
                 DEFAULT_METRIC_COLLECTION_INTERVAL,
             )
@@ -306,6 +321,8 @@ impl Default for PageServerConfigBuilder {
             )
             .expect("cannot parse DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD")),
 
+            disk_usage_based_eviction: Set(None),
+
             test_remote_failures: Set(0),
 
             ondemand_download_behavior_treat_error_as_warn: Set(false),
@@ -389,7 +406,7 @@ impl PageServerConfigBuilder {
         self.log_format = BuilderValue::Set(log_format)
     }
 
-    pub fn concurrent_tenant_size_logical_size_queries(&mut self, u: ConfigurableSemaphore) {
+    pub fn concurrent_tenant_size_logical_size_queries(&mut self, u: NonZeroUsize) {
         self.concurrent_tenant_size_logical_size_queries = BuilderValue::Set(u);
     }
 
@@ -425,6 +442,10 @@ impl PageServerConfigBuilder {
         self.evictions_low_residence_duration_metric_threshold = BuilderValue::Set(value);
     }
 
+    pub fn disk_usage_based_eviction(&mut self, value: Option<DiskUsageEvictionTaskConfig>) {
+        self.disk_usage_based_eviction = BuilderValue::Set(value);
+    }
+
     pub fn ondemand_download_behavior_treat_error_as_warn(
         &mut self,
         ondemand_download_behavior_treat_error_as_warn: bool,
@@ -434,6 +455,11 @@ impl PageServerConfigBuilder {
     }
 
     pub fn build(self) -> anyhow::Result<PageServerConf> {
+        let concurrent_tenant_size_logical_size_queries = self
+            .concurrent_tenant_size_logical_size_queries
+            .ok_or(anyhow!(
+                "missing concurrent_tenant_size_logical_size_queries"
+            ))?;
         Ok(PageServerConf {
             listen_pg_addr: self
                 .listen_pg_addr
@@ -481,11 +507,12 @@ impl PageServerConfigBuilder {
                 .broker_keepalive_interval
                 .ok_or(anyhow!("No broker keepalive interval provided"))?,
             log_format: self.log_format.ok_or(anyhow!("missing log_format"))?,
-            concurrent_tenant_size_logical_size_queries: self
-                .concurrent_tenant_size_logical_size_queries
-                .ok_or(anyhow!(
-                    "missing concurrent_tenant_size_logical_size_queries"
-                ))?,
+            concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::new(
+                concurrent_tenant_size_logical_size_queries,
+            ),
+            eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore::new(
+                concurrent_tenant_size_logical_size_queries,
+            ),
             metric_collection_interval: self
                 .metric_collection_interval
                 .ok_or(anyhow!("missing metric_collection_interval"))?,
@@ -503,6 +530,9 @@ impl PageServerConfigBuilder {
                 .ok_or(anyhow!(
                     "missing evictions_low_residence_duration_metric_threshold"
                 ))?,
+            disk_usage_based_eviction: self
+                .disk_usage_based_eviction
+                .ok_or(anyhow!("missing disk_usage_based_eviction"))?,
             test_remote_failures: self
                 .test_remote_failures
                 .ok_or(anyhow!("missing test_remote_failuers"))?,
@@ -680,8 +710,7 @@ impl PageServerConf {
                 "concurrent_tenant_size_logical_size_queries" => builder.concurrent_tenant_size_logical_size_queries({
                     let input = parse_toml_string(key, item)?;
                     let permits = input.parse::<usize>().context("expected a number of initial permits, not {s:?}")?;
-                    let permits = NonZeroUsize::new(permits).context("initial semaphore permits out of range: 0, use other configuration to disable a feature")?;
-                    ConfigurableSemaphore::new(permits)
+                    NonZeroUsize::new(permits).context("initial semaphore permits out of range: 0, use other configuration to disable a feature")?
                 }),
                 "metric_collection_interval" => builder.metric_collection_interval(parse_toml_duration(key, item)?),
                 "cached_metric_collection_interval" => builder.cached_metric_collection_interval(parse_toml_duration(key, item)?),
@@ -693,6 +722,12 @@ impl PageServerConf {
                     builder.synthetic_size_calculation_interval(parse_toml_duration(key, item)?),
                 "test_remote_failures" => builder.test_remote_failures(parse_toml_u64(key, item)?),
                 "evictions_low_residence_duration_metric_threshold" => builder.evictions_low_residence_duration_metric_threshold(parse_toml_duration(key, item)?),
+                "disk_usage_based_eviction" => {
+                    tracing::info!("disk_usage_based_eviction: {:#?}", &item);
+                    builder.disk_usage_based_eviction(
+                    toml_edit::de::from_item(item.clone())
+                    .context("parse disk_usage_based_eviction")?)
+                },
                 "ondemand_download_behavior_treat_error_as_warn" => builder.ondemand_download_behavior_treat_error_as_warn(parse_toml_bool(key, item)?),
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
@@ -797,6 +832,13 @@ impl PageServerConf {
             );
         }
 
+        if let Some(item) = item.get("min_resident_size_override") {
+            t_conf.min_resident_size_override = Some(
+                toml_edit::de::from_item(item.clone())
+                    .context("parse min_resident_size_override")?,
+            );
+        }
+
         Ok(t_conf)
     }
 
@@ -829,6 +871,8 @@ impl PageServerConf {
             broker_keepalive_interval: Duration::from_secs(5000),
             log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
             concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
+            eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore::default(
+            ),
             metric_collection_interval: Duration::from_secs(60),
             cached_metric_collection_interval: Duration::from_secs(60 * 60),
             metric_collection_endpoint: defaults::DEFAULT_METRIC_COLLECTION_ENDPOINT,
@@ -837,6 +881,7 @@ impl PageServerConf {
                 defaults::DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD,
             )
             .unwrap(),
+            disk_usage_based_eviction: None,
             test_remote_failures: 0,
             ondemand_download_behavior_treat_error_as_warn: false,
         }
@@ -921,6 +966,11 @@ impl ConfigurableSemaphore {
             inner: std::sync::Arc::new(tokio::sync::Semaphore::new(initial_permits.get())),
         }
     }
+
+    /// Returns the configured amount of permits.
+    pub fn initial_permits(&self) -> NonZeroUsize {
+        self.initial_permits
+    }
 }
 
 impl Default for ConfigurableSemaphore {
@@ -1025,6 +1075,8 @@ log_format = 'json'
                 )?,
                 log_format: LogFormat::from_str(defaults::DEFAULT_LOG_FORMAT).unwrap(),
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
+                eviction_task_immitated_concurrent_logical_size_queries:
+                    ConfigurableSemaphore::default(),
                 metric_collection_interval: humantime::parse_duration(
                     defaults::DEFAULT_METRIC_COLLECTION_INTERVAL
                 )?,
@@ -1038,6 +1090,7 @@ log_format = 'json'
                 evictions_low_residence_duration_metric_threshold: humantime::parse_duration(
                     defaults::DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD
                 )?,
+                disk_usage_based_eviction: None,
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
             },
@@ -1085,11 +1138,14 @@ log_format = 'json'
                 broker_keepalive_interval: Duration::from_secs(5),
                 log_format: LogFormat::Json,
                 concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore::default(),
+                eviction_task_immitated_concurrent_logical_size_queries:
+                    ConfigurableSemaphore::default(),
                 metric_collection_interval: Duration::from_secs(222),
                 cached_metric_collection_interval: Duration::from_secs(22200),
                 metric_collection_endpoint: Some(Url::parse("http://localhost:80/metrics")?),
                 synthetic_size_calculation_interval: Duration::from_secs(333),
                 evictions_low_residence_duration_metric_threshold: Duration::from_secs(444),
+                disk_usage_based_eviction: None,
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
             },
@@ -1216,6 +1272,7 @@ broker_endpoint = '{broker_endpoint}'
                         prefix_in_bucket: Some(prefix_in_bucket.clone()),
                         endpoint: Some(endpoint.clone()),
                         concurrency_limit: s3_concurrency_limit,
+                        max_keys_per_list_response: None,
                     }),
                 },
                 "Remote storage config should correctly parse the S3 config"

pageserver/src/disk_usage_eviction_task.rs

@@ -0,0 +1,728 @@
+//! This module implements the pageserver-global disk-usage-based layer eviction task.
+//!
+//! # Mechanics
+//!
+//! Function `launch_disk_usage_global_eviction_task` starts a pageserver-global background
+//! loop that evicts layers in response to a shortage of available bytes
+//! in the $repo/tenants directory's filesystem.
+//!
+//! The loop runs periodically at a configurable `period`.
+//!
+//! Each loop iteration uses `statvfs` to determine filesystem-level space usage.
+//! It compares the returned usage data against two different types of thresholds.
+//! The iteration tries to evict layers until app-internal accounting says we should be below the thresholds.
+//! We cross-check this internal accounting with the real world by making another `statvfs` at the end of the iteration.
+//! We're good if that second statvfs shows that we're _actually_ below the configured thresholds.
+//! If we're still above one or more thresholds, we emit a warning log message, leaving it to the operator to investigate further.
+//!
+//! # Eviction Policy
+//!
+//! There are two thresholds:
+//! `max_usage_pct` is the relative available space, expressed in percent of the total filesystem space.
+//! If the actual usage is higher, the threshold is exceeded.
+//! `min_avail_bytes` is the absolute available space in bytes.
+//! If the actual usage is lower, the threshold is exceeded.
+//! If either of these thresholds is exceeded, the system is considered to have "disk pressure", and eviction
+//! is performed on the next iteration, to release disk space and bring the usage below the thresholds again.
+//! The iteration evicts layers in LRU fashion, but, with a weak reservation per tenant.
+//! The reservation is to keep the most recently accessed X bytes per tenant resident.
+//! If we cannot relieve pressure by evicting layers outside of the reservation, we
+//! start evicting layers that are part of the reservation, LRU first.
+//!
+//! The value for the per-tenant reservation is referred to as `tenant_min_resident_size`
+//! throughout the code, but, no actual variable carries that name.
+//! The per-tenant default value is the `max(tenant's layer file sizes, regardless of local or remote)`.
+//! The idea is to allow at least one layer to be resident per tenant, to ensure it can make forward progress
+//! during page reconstruction.
+//! An alternative default for all tenants can be specified in the `tenant_config` section of the config.
+//! Lastly, each tenant can have an override in their respective tenant config (`min_resident_size_override`).
+
+// Implementation notes:
+// - The `#[allow(dead_code)]` above various structs are to suppress warnings about only the Debug impl
+//   reading these fields. We use the Debug impl for semi-structured logging, though.
+
+use std::{
+    collections::HashMap,
+    path::Path,
+    sync::Arc,
+    time::{Duration, SystemTime},
+};
+
+use anyhow::Context;
+use remote_storage::GenericRemoteStorage;
+use serde::{Deserialize, Serialize};
+use tokio::time::Instant;
+use tokio_util::sync::CancellationToken;
+use tracing::{debug, error, info, instrument, warn, Instrument};
+use utils::serde_percent::Percent;
+
+use crate::{
+    config::PageServerConf,
+    task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
+    tenant::{self, storage_layer::PersistentLayer, Timeline},
+};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DiskUsageEvictionTaskConfig {
+    pub max_usage_pct: Percent,
+    pub min_avail_bytes: u64,
+    #[serde(with = "humantime_serde")]
+    pub period: Duration,
+    #[cfg(feature = "testing")]
+    pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
+}
+
+#[derive(Default)]
+pub struct State {
+    /// Exclude http requests and background task from running at the same time.
+    mutex: tokio::sync::Mutex<()>,
+}
+
+pub fn launch_disk_usage_global_eviction_task(
+    conf: &'static PageServerConf,
+    storage: GenericRemoteStorage,
+    state: Arc<State>,
+) -> anyhow::Result<()> {
+    let Some(task_config) = &conf.disk_usage_based_eviction else {
+        info!("disk usage based eviction task not configured");
+        return Ok(());
+    };
+
+    info!("launching disk usage based eviction task");
+
+    task_mgr::spawn(
+        BACKGROUND_RUNTIME.handle(),
+        TaskKind::DiskUsageEviction,
+        None,
+        None,
+        "disk usage based eviction",
+        false,
+        async move {
+            disk_usage_eviction_task(
+                &state,
+                task_config,
+                storage,
+                &conf.tenants_path(),
+                task_mgr::shutdown_token(),
+            )
+            .await;
+            info!("disk usage based eviction task finishing");
+            Ok(())
+        },
+    );
+
+    Ok(())
+}
+
+#[instrument(skip_all)]
+async fn disk_usage_eviction_task(
+    state: &State,
+    task_config: &DiskUsageEvictionTaskConfig,
+    storage: GenericRemoteStorage,
+    tenants_dir: &Path,
+    cancel: CancellationToken,
+) {
+    use crate::tenant::tasks::random_init_delay;
+    {
+        if random_init_delay(task_config.period, &cancel)
+            .await
+            .is_err()
+        {
+            info!("shutting down");
+            return;
+        }
+    }
+
+    let mut iteration_no = 0;
+    loop {
+        iteration_no += 1;
+        let start = Instant::now();
+
+        async {
+            let res = disk_usage_eviction_task_iteration(
+                state,
+                task_config,
+                &storage,
+                tenants_dir,
+                &cancel,
+            )
+            .await;
+
+            match res {
+                Ok(()) => {}
+                Err(e) => {
+                    // these stat failures are expected to be very rare
+                    warn!("iteration failed, unexpected error: {e:#}");
+                }
+            }
+        }
+        .instrument(tracing::info_span!("iteration", iteration_no))
+        .await;
+
+        let sleep_until = start + task_config.period;
+        tokio::select! {
+            _ = tokio::time::sleep_until(sleep_until) => {},
+            _ = cancel.cancelled() => {
+                info!("shutting down");
+                break
+            }
+        }
+    }
+}
+
+pub trait Usage: Clone + Copy + std::fmt::Debug {
+    fn has_pressure(&self) -> bool;
+    fn add_available_bytes(&mut self, bytes: u64);
+}
+
+async fn disk_usage_eviction_task_iteration(
+    state: &State,
+    task_config: &DiskUsageEvictionTaskConfig,
+    storage: &GenericRemoteStorage,
+    tenants_dir: &Path,
+    cancel: &CancellationToken,
+) -> anyhow::Result<()> {
+    let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
+        .context("get filesystem-level disk usage before evictions")?;
+    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
+    match res {
+        Ok(outcome) => {
+            debug!(?outcome, "disk_usage_eviction_iteration finished");
+            match outcome {
+                IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
+                    // nothing to do, select statement below will handle things
+                }
+                IterationOutcome::Finished(outcome) => {
+                    // Verify with statvfs whether we made any real progress
+                    let after = filesystem_level_usage::get(tenants_dir, task_config)
+                        // It's quite unlikely to hit the error here. Keep the code simple and bail out.
+                        .context("get filesystem-level disk usage after evictions")?;
+
+                    debug!(?after, "disk usage");
+
+                    if after.has_pressure() {
+                        // Don't bother doing an out-of-order iteration here now.
+                        // In practice, the task period is set to a value in the tens-of-seconds range,
+                        // which will cause another iteration to happen soon enough.
+                        // TODO: deltas between the three different usages would be helpful,
+                        // consider MiB, GiB, TiB
+                        warn!(?outcome, ?after, "disk usage still high");
+                    } else {
+                        info!(?outcome, ?after, "disk usage pressure relieved");
+                    }
+                }
+            }
+        }
+        Err(e) => {
+            error!("disk_usage_eviction_iteration failed: {:#}", e);
+        }
+    }
+
+    Ok(())
+}
+
+#[derive(Debug, Serialize)]
+#[allow(clippy::large_enum_variant)]
+pub enum IterationOutcome<U> {
+    NoPressure,
+    Cancelled,
+    Finished(IterationOutcomeFinished<U>),
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Serialize)]
+pub struct IterationOutcomeFinished<U> {
+    /// The actual usage observed before we started the iteration.
+    before: U,
+    /// The expected value for `after`, according to internal accounting, after phase 1.
+    planned: PlannedUsage<U>,
+    /// The outcome of phase 2, where we actually do the evictions.
+    ///
+    /// If all layers that phase 1 planned to evict _can_ actually get evicted, this will
+    /// be the same as `planned`.
+    assumed: AssumedUsage<U>,
+}
+
+#[derive(Debug, Serialize)]
+#[allow(dead_code)]
+struct AssumedUsage<U> {
+    /// The expected value for `after`, after phase 2.
+    projected_after: U,
+    /// The layers we failed to evict during phase 2.
+    failed: LayerCount,
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Serialize)]
+struct PlannedUsage<U> {
+    respecting_tenant_min_resident_size: U,
+    fallback_to_global_lru: Option<U>,
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Default, Serialize)]
+struct LayerCount {
+    file_sizes: u64,
+    count: usize,
+}
+
+pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
+    state: &State,
+    storage: &GenericRemoteStorage,
+    usage_pre: U,
+    cancel: &CancellationToken,
+) -> anyhow::Result<IterationOutcome<U>> {
+    // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
+    let _g = state
+        .mutex
+        .try_lock()
+        .map_err(|_| anyhow::anyhow!("iteration is already executing"))?;
+
+    debug!(?usage_pre, "disk usage");
+
+    if !usage_pre.has_pressure() {
+        return Ok(IterationOutcome::NoPressure);
+    }
+
+    warn!(
+        ?usage_pre,
+        "running disk usage based eviction due to pressure"
+    );
+
+    let candidates = match collect_eviction_candidates(cancel).await? {
+        EvictionCandidates::Cancelled => {
+            return Ok(IterationOutcome::Cancelled);
+        }
+        EvictionCandidates::Finished(partitioned) => partitioned,
+    };
+
+    // Debug-log the list of candidates
+    let now = SystemTime::now();
+    for (i, (partition, candidate)) in candidates.iter().enumerate() {
+        debug!(
+            "cand {}/{}: size={}, no_access_for={}us, parition={:?}, tenant={} timeline={} layer={}",
+            i + 1,
+            candidates.len(),
+            candidate.layer.file_size(),
+            now.duration_since(candidate.last_activity_ts)
+                .unwrap()
+                .as_micros(),
+            partition,
+            candidate.layer.get_tenant_id(),
+            candidate.layer.get_timeline_id(),
+            candidate.layer.filename().file_name(),
+        );
+    }
+
+    // phase1: select victims to relieve pressure
+    //
+    // Walk through the list of candidates, until we have accumulated enough layers to get
+    // us back under the pressure threshold. 'usage_planned' is updated so that it tracks
+    // how much disk space would be used after evicting all the layers up to the current
+    // point in the list. The layers are collected in 'batched', grouped per timeline.
+    //
+    // If we get far enough in the list that we start to evict layers that are below
+    // the tenant's min-resident-size threshold, print a warning, and memorize the disk
+    // usage at that point, in 'usage_planned_min_resident_size_respecting'.
+    let mut batched: HashMap<_, Vec<Arc<dyn PersistentLayer>>> = HashMap::new();
+    let mut warned = None;
+    let mut usage_planned = usage_pre;
+    for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
+        if !usage_planned.has_pressure() {
+            debug!(
+                no_candidates_evicted = i,
+                "took enough candidates for pressure to be relieved"
+            );
+            break;
+        }
+
+        if partition == MinResidentSizePartition::Below && warned.is_none() {
+            warn!(?usage_pre, ?usage_planned, candidate_no=i, "tenant_min_resident_size-respecting LRU would not relieve pressure, evicting more following global LRU policy");
+            warned = Some(usage_planned);
+        }
+
+        usage_planned.add_available_bytes(candidate.layer.file_size());
+
+        batched
+            .entry(TimelineKey(candidate.timeline))
+            .or_default()
+            .push(candidate.layer);
+    }
+
+    let usage_planned = match warned {
+        Some(respecting_tenant_min_resident_size) => PlannedUsage {
+            respecting_tenant_min_resident_size,
+            fallback_to_global_lru: Some(usage_planned),
+        },
+        None => PlannedUsage {
+            respecting_tenant_min_resident_size: usage_planned,
+            fallback_to_global_lru: None,
+        },
+    };
+    debug!(?usage_planned, "usage planned");
+
+    // phase2: evict victims batched by timeline
+
+    // After the loop, `usage_assumed` is the post-eviction usage,
+    // according to internal accounting.
+    let mut usage_assumed = usage_pre;
+    let mut evictions_failed = LayerCount::default();
+    for (timeline, batch) in batched {
+        let tenant_id = timeline.tenant_id;
+        let timeline_id = timeline.timeline_id;
+        let batch_size = batch.len();
+
+        debug!(%timeline_id, "evicting batch for timeline");
+
+        async {
+            let results = timeline.evict_layers(storage, &batch, cancel.clone()).await;
+
+            match results {
+                Err(e) => {
+                    warn!("failed to evict batch: {:#}", e);
+                }
+                Ok(results) => {
+                    assert_eq!(results.len(), batch.len());
+                    for (result, layer) in results.into_iter().zip(batch.iter()) {
+                        match result {
+                            Some(Ok(true)) => {
+                                usage_assumed.add_available_bytes(layer.file_size());
+                            }
+                            Some(Ok(false)) => {
+                                // this is:
+                                // - Replacement::{NotFound, Unexpected}
+                                // - it cannot be is_remote_layer, filtered already
+                                evictions_failed.file_sizes += layer.file_size();
+                                evictions_failed.count += 1;
+                            }
+                            None => {
+                                assert!(cancel.is_cancelled());
+                                return;
+                            }
+                            Some(Err(e)) => {
+                                // we really shouldn't be getting this, precondition failure
+                                error!("failed to evict layer: {:#}", e);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size))
+        .await;
+
+        if cancel.is_cancelled() {
+            return Ok(IterationOutcome::Cancelled);
+        }
+    }
+
+    Ok(IterationOutcome::Finished(IterationOutcomeFinished {
+        before: usage_pre,
+        planned: usage_planned,
+        assumed: AssumedUsage {
+            projected_after: usage_assumed,
+            failed: evictions_failed,
+        },
+    }))
+}
+
+#[derive(Clone)]
+struct EvictionCandidate {
+    timeline: Arc<Timeline>,
+    layer: Arc<dyn PersistentLayer>,
+    last_activity_ts: SystemTime,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+enum MinResidentSizePartition {
+    Above,
+    Below,
+}
+
+enum EvictionCandidates {
+    Cancelled,
+    Finished(Vec<(MinResidentSizePartition, EvictionCandidate)>),
+}
+
+/// Gather the eviction candidates.
+///
+/// The returned `Ok(EvictionCandidates::Finished(candidates))` is sorted in eviction
+/// order. A caller that evicts in that order, until pressure is relieved, implements
+/// the eviction policy outlined in the module comment.
+///
+/// # Example
+///
+/// Imagine that there are two tenants, A and B, with five layers each, a-e.
+/// Each layer has size 100, and both tenant's min_resident_size is 150.
+/// The eviction order would be
+///
+/// ```text
+/// partition last_activity_ts    tenant/layer
+/// Above     18:30               A/c
+/// Above     19:00               A/b
+/// Above     18:29               B/c
+/// Above     19:05               B/b
+/// Above     20:00               B/a
+/// Above     20:03               A/a
+/// Below     20:30               A/d
+/// Below     20:40               B/d
+/// Below     20:45               B/e
+/// Below     20:58               A/e
+/// ```
+///
+/// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
+/// They are all in the `Above` partition, so, we respected each tenant's min_resident_size.
+///
+/// But, if we need to evict 900 bytes to relieve pressure, we'd evict
+/// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
+/// after exhauting the `Above` partition.
+/// So, we did not respect each tenant's min_resident_size.
+async fn collect_eviction_candidates(
+    cancel: &CancellationToken,
+) -> anyhow::Result<EvictionCandidates> {
+    // get a snapshot of the list of tenants
+    let tenants = tenant::mgr::list_tenants()
+        .await
+        .context("get list of tenants")?;
+
+    let mut candidates = Vec::new();
+
+    for (tenant_id, _state) in &tenants {
+        if cancel.is_cancelled() {
+            return Ok(EvictionCandidates::Cancelled);
+        }
+        let tenant = match tenant::mgr::get_tenant(*tenant_id, true).await {
+            Ok(tenant) => tenant,
+            Err(e) => {
+                // this can happen if tenant has lifecycle transition after we fetched it
+                debug!("failed to get tenant: {e:#}");
+                continue;
+            }
+        };
+
+        // collect layers from all timelines in this tenant
+        //
+        // If one of the timelines becomes `!is_active()` during the iteration,
+        // for example because we're shutting down, then `max_layer_size` can be too small.
+        // That's OK. This code only runs under a disk pressure situation, and being
+        // a little unfair to tenants during shutdown in such a situation is tolerable.
+        let mut tenant_candidates = Vec::new();
+        let mut max_layer_size = 0;
+        for tl in tenant.list_timelines() {
+            if !tl.is_active() {
+                continue;
+            }
+            let info = tl.get_local_layers_for_disk_usage_eviction();
+            debug!(tenant_id=%tl.tenant_id, timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
+            tenant_candidates.extend(
+                info.resident_layers
+                    .into_iter()
+                    .map(|layer_infos| (tl.clone(), layer_infos)),
+            );
+            max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
+
+            if cancel.is_cancelled() {
+                return Ok(EvictionCandidates::Cancelled);
+            }
+        }
+
+        // `min_resident_size` defaults to maximum layer file size of the tenant.
+        // This ensures that each tenant can have at least one layer resident at a given time,
+        // ensuring forward progress for a single Timeline::get in that tenant.
+        // It's a questionable heuristic since, usually, there are many Timeline::get
+        // requests going on for a tenant, and, at least in Neon prod, the median
+        // layer file size is much smaller than the compaction target size.
+        // We could be better here, e.g., sum of all L0 layers + most recent L1 layer.
+        // That's what's typically used by the various background loops.
+        //
+        // The default can be overriden with a fixed value in the tenant conf.
+        // A default override can be put in the default tenant conf in the pageserver.toml.
+        let min_resident_size = if let Some(s) = tenant.get_min_resident_size_override() {
+            debug!(
+                tenant_id=%tenant.tenant_id(),
+                overriden_size=s,
+                "using overridden min resident size for tenant"
+            );
+            s
+        } else {
+            debug!(
+                tenant_id=%tenant.tenant_id(),
+                max_layer_size,
+                "using max layer size as min_resident_size for tenant",
+            );
+            max_layer_size
+        };
+
+        // Sort layers most-recently-used first, then partition by
+        // cumsum above/below min_resident_size.
+        tenant_candidates
+            .sort_unstable_by_key(|(_, layer_info)| std::cmp::Reverse(layer_info.last_activity_ts));
+        let mut cumsum: i128 = 0;
+        for (timeline, layer_info) in tenant_candidates.into_iter() {
+            let file_size = layer_info.file_size();
+            let candidate = EvictionCandidate {
+                timeline,
+                last_activity_ts: layer_info.last_activity_ts,
+                layer: layer_info.layer,
+            };
+            let partition = if cumsum > min_resident_size as i128 {
+                MinResidentSizePartition::Above
+            } else {
+                MinResidentSizePartition::Below
+            };
+            candidates.push((partition, candidate));
+            cumsum += i128::from(file_size);
+        }
+    }
+
+    debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
+        "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first");
+    candidates
+        .sort_unstable_by_key(|(partition, candidate)| (*partition, candidate.last_activity_ts));
+
+    Ok(EvictionCandidates::Finished(candidates))
+}
+
+struct TimelineKey(Arc<Timeline>);
+
+impl PartialEq for TimelineKey {
+    fn eq(&self, other: &Self) -> bool {
+        Arc::ptr_eq(&self.0, &other.0)
+    }
+}
+
+impl Eq for TimelineKey {}
+
+impl std::hash::Hash for TimelineKey {
+    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
+        Arc::as_ptr(&self.0).hash(state);
+    }
+}
+
+impl std::ops::Deref for TimelineKey {
+    type Target = Timeline;
+
+    fn deref(&self) -> &Self::Target {
+        self.0.as_ref()
+    }
+}
+
+mod filesystem_level_usage {
+    use std::path::Path;
+
+    use anyhow::Context;
+
+    use crate::statvfs::Statvfs;
+
+    use super::DiskUsageEvictionTaskConfig;
+
+    #[derive(Debug, Clone, Copy)]
+    #[allow(dead_code)]
+    pub struct Usage<'a> {
+        config: &'a DiskUsageEvictionTaskConfig,
+
+        /// Filesystem capacity
+        total_bytes: u64,
+        /// Free filesystem space
+        avail_bytes: u64,
+    }
+
+    impl super::Usage for Usage<'_> {
+        fn has_pressure(&self) -> bool {
+            let usage_pct =
+                (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
+
+            let pressures = [
+                (
+                    "min_avail_bytes",
+                    self.avail_bytes < self.config.min_avail_bytes,
+                ),
+                (
+                    "max_usage_pct",
+                    usage_pct >= self.config.max_usage_pct.get() as u64,
+                ),
+            ];
+
+            pressures.into_iter().any(|(_, has_pressure)| has_pressure)
+        }
+
+        fn add_available_bytes(&mut self, bytes: u64) {
+            self.avail_bytes += bytes;
+        }
+    }
+
+    pub fn get<'a>(
+        tenants_dir: &Path,
+        config: &'a DiskUsageEvictionTaskConfig,
+    ) -> anyhow::Result<Usage<'a>> {
+        let mock_config = {
+            #[cfg(feature = "testing")]
+            {
+                config.mock_statvfs.as_ref()
+            }
+            #[cfg(not(feature = "testing"))]
+            {
+                None
+            }
+        };
+
+        let stat = Statvfs::get(tenants_dir, mock_config)
+            .context("statvfs failed, presumably directory got unlinked")?;
+
+        // https://unix.stackexchange.com/a/703650
+        let blocksize = if stat.fragment_size() > 0 {
+            stat.fragment_size()
+        } else {
+            stat.block_size()
+        };
+
+        // use blocks_available (b_avail) since, pageserver runs as unprivileged user
+        let avail_bytes = stat.blocks_available() * blocksize;
+        let total_bytes = stat.blocks() * blocksize;
+
+        Ok(Usage {
+            config,
+            total_bytes,
+            avail_bytes,
+        })
+    }
+
+    #[test]
+    fn max_usage_pct_pressure() {
+        use super::Usage as _;
+        use std::time::Duration;
+        use utils::serde_percent::Percent;
+
+        let mut usage = Usage {
+            config: &DiskUsageEvictionTaskConfig {
+                max_usage_pct: Percent::new(85).unwrap(),
+                min_avail_bytes: 0,
+                period: Duration::MAX,
+                #[cfg(feature = "testing")]
+                mock_statvfs: None,
+            },
+            total_bytes: 100_000,
+            avail_bytes: 0,
+        };
+
+        assert!(usage.has_pressure(), "expected pressure at 100%");
+
+        usage.add_available_bytes(14_000);
+        assert!(usage.has_pressure(), "expected pressure at 86%");
+
+        usage.add_available_bytes(999);
+        assert!(usage.has_pressure(), "expected pressure at 85.001%");
+
+        usage.add_available_bytes(1);
+        assert!(usage.has_pressure(), "expected pressure at precisely 85%");
+
+        usage.add_available_bytes(1);
+        assert!(!usage.has_pressure(), "no pressure at 84.999%");
+
+        usage.add_available_bytes(999);
+        assert!(!usage.has_pressure(), "no pressure at 84%");
+
+        usage.add_available_bytes(16_000);
+        assert!(!usage.has_pressure());
+    }
+}

pageserver/src/http/openapi_spec.yml

@@ -27,6 +27,31 @@ paths:
                   id:
                     type: integer
 
+  /v1/disk_usage_eviction/run:
+    put:
+      description: Do an iteration of disk-usage-based eviction to evict a given amount of disk space.
+      security: []
+      requestBody:
+        content:
+          application/json:
+            schema:
+              type: object
+              required:
+                - evict_bytes
+              properties:
+                evict_bytes:
+                  type: integer
+      responses:
+        "200":
+          description: |
+            The run completed.
+            This does not necessarily mean that we actually evicted `evict_bytes`.
+            Examine the returned object for detail, or, just watch the actual effect of the call using `du` or `df`.
+          content:
+            application/json:
+              schema:
+                type: object
+
   /v1/tenant/{tenant_id}:
     parameters:
       - name: tenant_id
@@ -183,6 +208,19 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/ForbiddenError"
+        "404":
+          description: Timeline not found
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
+        "412":
+          description: Tenant is missing
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PreconditionFailedError"
+
         "500":
           description: Generic operation error
           content:
@@ -383,6 +421,12 @@ paths:
             application/json:
               schema:
                 $ref: "#/components/schemas/ForbiddenError"
+        "404":
+          description: Tenant not found
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
         "500":
           description: Generic operation error
           content:
@@ -854,13 +898,9 @@ components:
       type: object
       properties:
         tenant_specific_overrides:
-          type: object
-          schema:
-            $ref: "#/components/schemas/TenantConfigInfo"
+          $ref: "#/components/schemas/TenantConfigInfo"
         effective_config:
-          type: object
-          schema:
-            $ref: "#/components/schemas/TenantConfigInfo"
+          $ref: "#/components/schemas/TenantConfigInfo"
     TimelineInfo:
       type: object
       required:
@@ -946,6 +986,13 @@ components:
       properties:
         msg:
           type: string
+    PreconditionFailedError:
+      type: object
+      required:
+        - msg
+      properties:
+        msg:
+          type: string
 
 security:
   - JWT: []

pageserver/src/http/routes.rs

@@ -18,6 +18,7 @@ use super::models::{
     TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
 };
 use crate::context::{DownloadBehavior, RequestContext};
+use crate::disk_usage_eviction_task;
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::TenantConfOpt;
@@ -48,6 +49,7 @@ struct State {
     auth: Option<Arc<JwtAuth>>,
     allowlist_routes: Vec<Uri>,
     remote_storage: Option<GenericRemoteStorage>,
+    disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 }
 
 impl State {
@@ -55,6 +57,7 @@ impl State {
         conf: &'static PageServerConf,
         auth: Option<Arc<JwtAuth>>,
         remote_storage: Option<GenericRemoteStorage>,
+        disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     ) -> anyhow::Result<Self> {
         let allowlist_routes = ["/v1/status", "/v1/doc", "/swagger.yml"]
             .iter()
@@ -65,6 +68,7 @@ impl State {
             auth,
             allowlist_routes,
             remote_storage,
+            disk_usage_eviction_state,
         })
     }
 }
@@ -131,6 +135,34 @@ impl From<TenantStateError> for ApiError {
     }
 }
 
+impl From<crate::tenant::DeleteTimelineError> for ApiError {
+    fn from(value: crate::tenant::DeleteTimelineError) -> Self {
+        use crate::tenant::DeleteTimelineError::*;
+        match value {
+            NotFound => ApiError::NotFound(anyhow::anyhow!("timeline not found")),
+            HasChildren => ApiError::BadRequest(anyhow::anyhow!(
+                "Cannot delete timeline which has child timelines"
+            )),
+            Other(e) => ApiError::InternalServerError(e),
+        }
+    }
+}
+
+impl From<crate::tenant::mgr::DeleteTimelineError> for ApiError {
+    fn from(value: crate::tenant::mgr::DeleteTimelineError) -> Self {
+        use crate::tenant::mgr::DeleteTimelineError::*;
+        match value {
+            // Report Precondition failed so client can distinguish between
+            // "tenant is missing" case from "timeline is missing"
+            Tenant(TenantStateError::NotFound(..)) => {
+                ApiError::PreconditionFailed("Requested tenant is missing")
+            }
+            Tenant(t) => ApiError::from(t),
+            Timeline(t) => ApiError::from(t),
+        }
+    }
+}
+
 // Helper function to construct a TimelineInfo struct for a timeline
 async fn build_timeline_info(
     timeline: &Arc<Timeline>,
@@ -747,6 +779,8 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
         );
     }
 
+    tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
+
     let target_tenant_id = request_data
         .new_tenant_id
         .map(TenantId::from)
@@ -878,6 +912,8 @@ async fn update_tenant_config_handler(
         );
     }
 
+    tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
+
     let state = get_state(&request);
     mgr::set_new_tenant_config(state.conf, tenant_conf, tenant_id)
         .instrument(info_span!("tenant_config", tenant = ?tenant_id))
@@ -886,6 +922,20 @@ async fn update_tenant_config_handler(
     json_response(StatusCode::OK, ())
 }
 
+/// Testing helper to transition a tenant to [`crate::tenant::TenantState::Broken`].
+#[cfg(feature = "testing")]
+async fn handle_tenant_break(r: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&r, "tenant_id")?;
+
+    let tenant = crate::tenant::mgr::get_tenant(tenant_id, true)
+        .await
+        .map_err(|_| ApiError::Conflict(String::from("no active tenant found")))?;
+
+    tenant.set_broken("broken from test");
+
+    json_response(StatusCode::OK, ())
+}
+
 #[cfg(feature = "testing")]
 async fn failpoints_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
     if !fail::has_failpoints() {
@@ -1035,6 +1085,89 @@ async fn always_panic_handler(req: Request<Body>) -> Result<Response<Body>, ApiE
     json_response(StatusCode::NO_CONTENT, ())
 }
 
+async fn disk_usage_eviction_run(mut r: Request<Body>) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
+
+    #[derive(Debug, Clone, Copy, serde::Serialize, serde::Deserialize)]
+    struct Config {
+        /// How many bytes to evict before reporting that pressure is relieved.
+        evict_bytes: u64,
+    }
+
+    #[derive(Debug, Clone, Copy, serde::Serialize)]
+    struct Usage {
+        // remains unchanged after instantiation of the struct
+        config: Config,
+        // updated by `add_available_bytes`
+        freed_bytes: u64,
+    }
+
+    impl crate::disk_usage_eviction_task::Usage for Usage {
+        fn has_pressure(&self) -> bool {
+            self.config.evict_bytes > self.freed_bytes
+        }
+
+        fn add_available_bytes(&mut self, bytes: u64) {
+            self.freed_bytes += bytes;
+        }
+    }
+
+    let config = json_request::<Config>(&mut r)
+        .await
+        .map_err(|_| ApiError::BadRequest(anyhow::anyhow!("invalid JSON body")))?;
+
+    let usage = Usage {
+        config,
+        freed_bytes: 0,
+    };
+
+    use crate::task_mgr::MGMT_REQUEST_RUNTIME;
+
+    let (tx, rx) = tokio::sync::oneshot::channel();
+
+    let state = get_state(&r);
+
+    let Some(storage) = state.remote_storage.clone() else {
+        return Err(ApiError::InternalServerError(anyhow::anyhow!(
+            "remote storage not configured, cannot run eviction iteration"
+        )))
+    };
+
+    let state = state.disk_usage_eviction_state.clone();
+
+    let cancel = CancellationToken::new();
+    let child_cancel = cancel.clone();
+    let _g = cancel.drop_guard();
+
+    crate::task_mgr::spawn(
+        MGMT_REQUEST_RUNTIME.handle(),
+        TaskKind::DiskUsageEviction,
+        None,
+        None,
+        "ondemand disk usage eviction",
+        false,
+        async move {
+            let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
+                &state,
+                &storage,
+                usage,
+                &child_cancel,
+            )
+            .await;
+
+            info!(?res, "disk_usage_eviction_task_iteration_impl finished");
+
+            let _ = tx.send(res);
+            Ok(())
+        }
+        .in_current_span(),
+    );
+
+    let response = rx.await.unwrap().map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, response)
+}
+
 async fn handler_404(_: Request<Body>) -> Result<Response<Body>, ApiError> {
     json_response(
         StatusCode::NOT_FOUND,
@@ -1047,6 +1180,7 @@ pub fn make_router(
     launch_ts: &'static LaunchTimestamp,
     auth: Option<Arc<JwtAuth>>,
     remote_storage: Option<GenericRemoteStorage>,
+    disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 ) -> anyhow::Result<RouterBuilder<hyper::Body, ApiError>> {
     let spec = include_bytes!("openapi_spec.yml");
     let mut router = attach_openapi_ui(endpoint::make_router(), spec, "/swagger.yml", "/v1/doc");
@@ -1091,7 +1225,8 @@ pub fn make_router(
 
     Ok(router
         .data(Arc::new(
-            State::new(conf, auth, remote_storage).context("Failed to initialize router state")?,
+            State::new(conf, auth, remote_storage, disk_usage_eviction_state)
+                .context("Failed to initialize router state")?,
         ))
         .get("/v1/status", |r| RequestSpan(status_handler).handle(r))
         .put(
@@ -1172,6 +1307,13 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/layer/:layer_file_name",
             |r| RequestSpan(evict_timeline_layer_handler).handle(r),
         )
+        .put("/v1/disk_usage_eviction/run", |r| {
+            RequestSpan(disk_usage_eviction_run).handle(r)
+        })
+        .put(
+            "/v1/tenant/:tenant_id/break",
+            testing_api!("set tenant state to broken", handle_tenant_break),
+        )
         .get("/v1/panic", |r| RequestSpan(always_panic_handler).handle(r))
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -4,6 +4,7 @@ pub mod broker_client;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
+pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;
 pub mod keyspace;
@@ -12,6 +13,7 @@ pub mod page_cache;
 pub mod page_service;
 pub mod pgdatadir_mapping;
 pub mod repository;
+pub(crate) mod statvfs;
 pub mod task_mgr;
 pub mod tenant;
 pub mod trace;

pageserver/src/metrics.rs

@@ -257,7 +257,7 @@ impl EvictionsWithLowResidenceDuration {
     }
 
     pub fn observe(&self, observed_value: Duration) {
-        if self.threshold < observed_value {
+        if observed_value < self.threshold {
             self.counter
                 .as_ref()
                 .expect("nobody calls this function after `remove_from_vec`")
@@ -586,7 +586,6 @@ pub struct TimelineMetrics {
     pub flush_time_histo: StorageTimeMetrics,
     pub compact_time_histo: StorageTimeMetrics,
     pub create_images_time_histo: StorageTimeMetrics,
-    pub init_logical_size_histo: StorageTimeMetrics,
     pub logical_size_histo: StorageTimeMetrics,
     pub load_layer_map_histo: StorageTimeMetrics,
     pub garbage_collect_histo: StorageTimeMetrics,
@@ -619,8 +618,6 @@ impl TimelineMetrics {
         let compact_time_histo = StorageTimeMetrics::new("compact", &tenant_id, &timeline_id);
         let create_images_time_histo =
             StorageTimeMetrics::new("create images", &tenant_id, &timeline_id);
-        let init_logical_size_histo =
-            StorageTimeMetrics::new("init logical size", &tenant_id, &timeline_id);
         let logical_size_histo = StorageTimeMetrics::new("logical size", &tenant_id, &timeline_id);
         let load_layer_map_histo =
             StorageTimeMetrics::new("load layer map", &tenant_id, &timeline_id);
@@ -657,7 +654,6 @@ impl TimelineMetrics {
             flush_time_histo,
             compact_time_histo,
             create_images_time_histo,
-            init_logical_size_histo,
             logical_size_histo,
             garbage_collect_histo,
             load_layer_map_histo,

pageserver/src/page_service.rs

@@ -27,6 +27,7 @@ use pq_proto::FeStartupPacket;
 use pq_proto::{BeMessage, FeMessage, RowDescriptor};
 use std::io;
 use std::net::TcpListener;
+use std::pin::pin;
 use std::str;
 use std::str::FromStr;
 use std::sync::Arc;
@@ -466,8 +467,7 @@ impl PageServerHandler {
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
 
-        let copyin_reader = StreamReader::new(copyin_stream(pgb));
-        tokio::pin!(copyin_reader);
+        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
         timeline
             .import_basebackup_from_tar(&mut copyin_reader, base_lsn, &ctx)
             .await?;
@@ -512,8 +512,7 @@ impl PageServerHandler {
         info!("importing wal");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
-        let copyin_reader = StreamReader::new(copyin_stream(pgb));
-        tokio::pin!(copyin_reader);
+        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
         import_wal_from_tar(&timeline, &mut copyin_reader, start_lsn, end_lsn, &ctx).await?;
         info!("wal import complete");
 

pageserver/src/statvfs.rs

@@ -0,0 +1,150 @@
+//! Wrapper around nix::sys::statvfs::Statvfs that allows for mocking.
+
+use std::path::Path;
+
+pub enum Statvfs {
+    Real(nix::sys::statvfs::Statvfs),
+    Mock(mock::Statvfs),
+}
+
+// NB: on macOS, the block count type of struct statvfs is u32.
+// The workaround seems to be to use the non-standard statfs64 call.
+// Sincce it should only be a problem on > 2TiB disks, let's ignore
+// the problem for now and upcast to u64.
+impl Statvfs {
+    pub fn get(tenants_dir: &Path, mocked: Option<&mock::Behavior>) -> nix::Result<Self> {
+        if let Some(mocked) = mocked {
+            Ok(Statvfs::Mock(mock::get(tenants_dir, mocked)?))
+        } else {
+            Ok(Statvfs::Real(nix::sys::statvfs::statvfs(tenants_dir)?))
+        }
+    }
+
+    // NB: allow() because the block count type is u32 on macOS.
+    #[allow(clippy::useless_conversion)]
+    pub fn blocks(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => u64::try_from(stat.blocks()).unwrap(),
+            Statvfs::Mock(stat) => stat.blocks,
+        }
+    }
+
+    // NB: allow() because the block count type is u32 on macOS.
+    #[allow(clippy::useless_conversion)]
+    pub fn blocks_available(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => u64::try_from(stat.blocks_available()).unwrap(),
+            Statvfs::Mock(stat) => stat.blocks_available,
+        }
+    }
+
+    pub fn fragment_size(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => stat.fragment_size(),
+            Statvfs::Mock(stat) => stat.fragment_size,
+        }
+    }
+
+    pub fn block_size(&self) -> u64 {
+        match self {
+            Statvfs::Real(stat) => stat.block_size(),
+            Statvfs::Mock(stat) => stat.block_size,
+        }
+    }
+}
+
+pub mod mock {
+    use anyhow::Context;
+    use regex::Regex;
+    use std::path::Path;
+    use tracing::log::info;
+
+    #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+    #[serde(tag = "type")]
+    pub enum Behavior {
+        Success {
+            blocksize: u64,
+            total_blocks: u64,
+            name_filter: Option<utils::serde_regex::Regex>,
+        },
+        Failure {
+            mocked_error: MockedError,
+        },
+    }
+
+    #[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+    #[allow(clippy::upper_case_acronyms)]
+    pub enum MockedError {
+        EIO,
+    }
+
+    impl From<MockedError> for nix::Error {
+        fn from(e: MockedError) -> Self {
+            match e {
+                MockedError::EIO => nix::Error::EIO,
+            }
+        }
+    }
+
+    pub fn get(tenants_dir: &Path, behavior: &Behavior) -> nix::Result<Statvfs> {
+        info!("running mocked statvfs");
+
+        match behavior {
+            Behavior::Success {
+                blocksize,
+                total_blocks,
+                ref name_filter,
+            } => {
+                let used_bytes = walk_dir_disk_usage(tenants_dir, name_filter.as_deref()).unwrap();
+
+                // round it up to the nearest block multiple
+                let used_blocks = (used_bytes + (blocksize - 1)) / blocksize;
+
+                if used_blocks > *total_blocks {
+                    panic!(
+                        "mocking error: used_blocks > total_blocks: {used_blocks} > {total_blocks}"
+                    );
+                }
+
+                let avail_blocks = total_blocks - used_blocks;
+
+                Ok(Statvfs {
+                    blocks: *total_blocks,
+                    blocks_available: avail_blocks,
+                    fragment_size: *blocksize,
+                    block_size: *blocksize,
+                })
+            }
+            Behavior::Failure { mocked_error } => Err((*mocked_error).into()),
+        }
+    }
+
+    fn walk_dir_disk_usage(path: &Path, name_filter: Option<&Regex>) -> anyhow::Result<u64> {
+        let mut total = 0;
+        for entry in walkdir::WalkDir::new(path) {
+            let entry = entry?;
+            if !entry.file_type().is_file() {
+                continue;
+            }
+            if !name_filter
+                .as_ref()
+                .map(|filter| filter.is_match(entry.file_name().to_str().unwrap()))
+                .unwrap_or(true)
+            {
+                continue;
+            }
+            total += entry
+                .metadata()
+                .with_context(|| format!("get metadata of {:?}", entry.path()))?
+                .len();
+        }
+        Ok(total)
+    }
+
+    pub struct Statvfs {
+        pub blocks: u64,
+        pub blocks_available: u64,
+        pub fragment_size: u64,
+        pub block_size: u64,
+    }
+}

pageserver/src/task_mgr.rs

@@ -234,6 +234,9 @@ pub enum TaskKind {
     // Eviction. One per timeline.
     Eviction,
 
+    /// See [`crate::disk_usage_eviction_task`].
+    DiskUsageEviction,
+
     // Initial logical size calculation
     InitialLogicalSizeCalculation,
 
@@ -481,13 +484,25 @@ pub async fn shutdown_tasks(
     for task in victim_tasks {
         let join_handle = {
             let mut task_mut = task.mutable.lock().unwrap();
-            info!("waiting for {} to shut down", task.name);
-            let join_handle = task_mut.join_handle.take();
-            drop(task_mut);
-            join_handle
+            task_mut.join_handle.take()
         };
-        if let Some(join_handle) = join_handle {
-            let _ = join_handle.await;
+        if let Some(mut join_handle) = join_handle {
+            let completed = tokio::select! {
+                _ = &mut join_handle => { true },
+                _ = tokio::time::sleep(std::time::Duration::from_secs(1)) => {
+                    // allow some time to elapse before logging to cut down the number of log
+                    // lines.
+                    info!("waiting for {} to shut down", task.name);
+                    false
+                }
+            };
+            if !completed {
+                // we never handled this return value, but:
+                // - we don't deschedule which would lead to is_cancelled
+                // - panics are already logged (is_panicked)
+                // - task errors are already logged in the wrapper
+                let _ = join_handle.await;
+            }
         } else {
             // Possibly one of:
             //  * The task had not even fully started yet.

pageserver/src/tenant.rs

@@ -46,6 +46,7 @@ use std::time::{Duration, Instant};
 use self::config::TenantConf;
 use self::metadata::TimelineMetadata;
 use self::remote_timeline_client::RemoteTimelineClient;
+use self::timeline::EvictionTaskTenantState;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir;
@@ -94,7 +95,7 @@ mod timeline;
 
 pub mod size;
 
-pub use timeline::{PageReconstructError, Timeline};
+pub use timeline::{LocalLayerInfoForDiskUsageEviction, PageReconstructError, Timeline};
 
 // re-export this function so that page_cache.rs can use it.
 pub use crate::tenant::ephemeral_file::writeback as writeback_ephemeral_file;
@@ -142,6 +143,8 @@ pub struct Tenant {
     /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
     cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
     cached_synthetic_tenant_size: Arc<AtomicU64>,
+
+    eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
 }
 
 /// A timeline with some of its files on disk, being initialized.
@@ -431,6 +434,16 @@ remote:
     }
 }
 
+#[derive(Debug, thiserror::Error)]
+pub enum DeleteTimelineError {
+    #[error("NotFound")]
+    NotFound,
+    #[error("HasChildren")]
+    HasChildren,
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
 struct RemoteStartupData {
     index_part: IndexPart,
     remote_metadata: TimelineMetadata,
@@ -1307,7 +1320,7 @@ impl Tenant {
         &self,
         timeline_id: TimelineId,
         _ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), DeleteTimelineError> {
         // Transition the timeline into TimelineState::Stopping.
         // This should prevent new operations from starting.
         let timeline = {
@@ -1319,13 +1332,13 @@ impl Tenant {
                 .iter()
                 .any(|(_, entry)| entry.get_ancestor_timeline_id() == Some(timeline_id));
 
-            anyhow::ensure!(
-                !children_exist,
-                "Cannot delete timeline which has child timelines"
-            );
+            if children_exist {
+                return Err(DeleteTimelineError::HasChildren);
+            }
+
             let timeline_entry = match timelines.entry(timeline_id) {
                 Entry::Occupied(e) => e,
-                Entry::Vacant(_) => bail!("timeline not found"),
+                Entry::Vacant(_) => return Err(DeleteTimelineError::NotFound),
             };
 
             let timeline = Arc::clone(timeline_entry.get());
@@ -1693,6 +1706,13 @@ impl Tenant {
             .unwrap_or(self.conf.default_tenant_conf.trace_read_requests)
     }
 
+    pub fn get_min_resident_size_override(&self) -> Option<u64> {
+        let tenant_conf = self.tenant_conf.read().unwrap();
+        tenant_conf
+            .min_resident_size_override
+            .or(self.conf.default_tenant_conf.min_resident_size_override)
+    }
+
     pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
         *self.tenant_conf.write().unwrap() = new_tenant_conf;
     }
@@ -1771,6 +1791,7 @@ impl Tenant {
             state,
             cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
             cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
+            eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
         }
     }
 
@@ -2769,6 +2790,7 @@ pub mod harness {
                 max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
                 trace_read_requests: Some(tenant_conf.trace_read_requests),
                 eviction_policy: Some(tenant_conf.eviction_policy),
+                min_resident_size_override: tenant_conf.min_resident_size_override,
             }
         }
     }

pageserver/src/tenant/config.rs

@@ -92,6 +92,7 @@ pub struct TenantConf {
     pub max_lsn_wal_lag: NonZeroU64,
     pub trace_read_requests: bool,
     pub eviction_policy: EvictionPolicy,
+    pub min_resident_size_override: Option<u64>,
 }
 
 /// Same as TenantConf, but this struct preserves the information about
@@ -159,6 +160,10 @@ pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub eviction_policy: Option<EvictionPolicy>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(default)]
+    pub min_resident_size_override: Option<u64>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
@@ -220,6 +225,9 @@ impl TenantConfOpt {
                 .trace_read_requests
                 .unwrap_or(global_conf.trace_read_requests),
             eviction_policy: self.eviction_policy.unwrap_or(global_conf.eviction_policy),
+            min_resident_size_override: self
+                .min_resident_size_override
+                .or(global_conf.min_resident_size_override),
         }
     }
 }
@@ -251,6 +259,7 @@ impl Default for TenantConf {
                 .expect("cannot parse default max walreceiver Lsn wal lag"),
             trace_read_requests: false,
             eviction_policy: EvictionPolicy::NoEviction,
+            min_resident_size_override: None,
         }
     }
 }

pageserver/src/tenant/mgr.rs

@@ -321,11 +321,20 @@ pub async fn get_tenant(
     }
 }
 
+#[derive(Debug, thiserror::Error)]
+pub enum DeleteTimelineError {
+    #[error("Tenant {0}")]
+    Tenant(#[from] TenantStateError),
+
+    #[error("Timeline {0}")]
+    Timeline(#[from] crate::tenant::DeleteTimelineError),
+}
+
 pub async fn delete_timeline(
     tenant_id: TenantId,
     timeline_id: TimelineId,
     ctx: &RequestContext,
-) -> Result<(), TenantStateError> {
+) -> Result<(), DeleteTimelineError> {
     let tenant = get_tenant(tenant_id, true).await?;
     tenant.delete_timeline(timeline_id, ctx).await?;
     Ok(())

pageserver/src/tenant/size.rs

@@ -6,6 +6,7 @@ use std::sync::Arc;
 use anyhow::{bail, Context};
 use tokio::sync::oneshot::error::RecvError;
 use tokio::sync::Semaphore;
+use tokio_util::sync::CancellationToken;
 
 use crate::context::RequestContext;
 use crate::pgdatadir_mapping::CalculateLogicalSizeError;
@@ -352,6 +353,10 @@ async fn fill_logical_sizes(
     // our advantage with `?` error handling.
     let mut joinset = tokio::task::JoinSet::new();
 
+    let cancel = tokio_util::sync::CancellationToken::new();
+    // be sure to cancel all spawned tasks if we are dropped
+    let _dg = cancel.clone().drop_guard();
+
     // For each point that would benefit from having a logical size available,
     // spawn a Task to fetch it, unless we have it cached already.
     for seg in segments.iter() {
@@ -373,6 +378,7 @@ async fn fill_logical_sizes(
                     timeline,
                     lsn,
                     ctx,
+                    cancel.child_token(),
                 ));
             }
             e.insert(cached_size);
@@ -477,13 +483,14 @@ async fn calculate_logical_size(
     timeline: Arc<crate::tenant::Timeline>,
     lsn: utils::lsn::Lsn,
     ctx: RequestContext,
+    cancel: CancellationToken,
 ) -> Result<TimelineAtLsnSizeResult, RecvError> {
     let _permit = tokio::sync::Semaphore::acquire_owned(limit)
         .await
         .expect("global semaphore should not had been closed");
 
     let size_res = timeline
-        .spawn_ondemand_logical_size_calculation(lsn, ctx)
+        .spawn_ondemand_logical_size_calculation(lsn, ctx, cancel)
         .instrument(info_span!("spawn_ondemand_logical_size_calculation"))
         .await?;
     Ok(TimelineAtLsnSizeResult(timeline, lsn, size_res))

pageserver/src/tenant/storage_layer.rs

@@ -121,10 +121,10 @@ struct LayerAccessStatsInner {
 }
 
 #[derive(Debug, Clone, Copy)]
-pub(super) struct LayerAccessStatFullDetails {
-    pub(super) when: SystemTime,
-    pub(super) task_kind: TaskKind,
-    pub(super) access_kind: LayerAccessKind,
+pub(crate) struct LayerAccessStatFullDetails {
+    pub(crate) when: SystemTime,
+    pub(crate) task_kind: TaskKind,
+    pub(crate) access_kind: LayerAccessKind,
 }
 
 #[derive(Clone, Copy, strum_macros::EnumString)]
@@ -255,7 +255,7 @@ impl LayerAccessStats {
         ret
     }
 
-    pub(super) fn most_recent_access_or_residence_event(
+    fn most_recent_access_or_residence_event(
         &self,
     ) -> Either<LayerAccessStatFullDetails, LayerResidenceEvent> {
         let locked = self.0.lock().unwrap();
@@ -268,6 +268,13 @@ impl LayerAccessStats {
             }
         }
     }
+
+    pub(crate) fn latest_activity(&self) -> SystemTime {
+        match self.most_recent_access_or_residence_event() {
+            Either::Left(mra) => mra.when,
+            Either::Right(re) => re.timestamp,
+        }
+    }
 }
 
 /// Supertrait of the [`Layer`] trait that captures the bare minimum interface

pageserver/src/tenant/tasks.rs

@@ -244,14 +244,12 @@ pub(crate) async fn random_init_delay(
 ) -> Result<(), Cancelled> {
     use rand::Rng;
 
+    if period == Duration::ZERO {
+        return Ok(());
+    }
+
     let d = {
         let mut rng = rand::thread_rng();
-
-        // gen_range asserts that the range cannot be empty, which it could be because period can
-        // be set to zero to disable gc or compaction, so lets set it to be at least 10s.
-        let period = std::cmp::max(period, Duration::from_secs(10));
-
-        // semi-ok default as the source of jitter
         rng.gen_range(Duration::ZERO..=period)
     };
 

pageserver/src/tenant/timeline.rs

@@ -13,6 +13,7 @@ use pageserver_api::models::{
     DownloadRemoteLayersTaskInfo, DownloadRemoteLayersTaskSpawnRequest,
     DownloadRemoteLayersTaskState, LayerMapInfo, LayerResidenceStatus, TimelineState,
 };
+use remote_storage::GenericRemoteStorage;
 use tokio::sync::{oneshot, watch, Semaphore, TryAcquireError};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -24,6 +25,7 @@ use std::collections::HashMap;
 use std::fs;
 use std::ops::{Deref, Range};
 use std::path::{Path, PathBuf};
+use std::pin::pin;
 use std::sync::atomic::{AtomicI64, Ordering as AtomicOrdering};
 use std::sync::{Arc, Mutex, MutexGuard, RwLock, Weak};
 use std::time::{Duration, Instant, SystemTime};
@@ -71,6 +73,9 @@ use crate::ZERO_PAGE;
 use crate::{is_temporary, task_mgr};
 use walreceiver::spawn_connection_manager_task;
 
+pub(super) use self::eviction_task::EvictionTaskTenantState;
+use self::eviction_task::EvictionTaskTimelineState;
+
 use super::layer_map::BatchedUpdates;
 use super::remote_timeline_client::index::IndexPart;
 use super::remote_timeline_client::RemoteTimelineClient;
@@ -216,6 +221,8 @@ pub struct Timeline {
     download_all_remote_layers_task_info: RwLock<Option<DownloadRemoteLayersTaskInfo>>,
 
     state: watch::Sender<TimelineState>,
+
+    eviction_task_timeline_state: tokio::sync::Mutex<EvictionTaskTimelineState>,
 }
 
 /// Internal structure to hold all data needed for logical size calculation.
@@ -328,9 +335,13 @@ impl LogicalSize {
             .fetch_add(delta, AtomicOrdering::SeqCst);
     }
 
-    /// Returns the initialized (already calculated) value, if any.
-    fn initialized_size(&self) -> Option<u64> {
-        self.initial_logical_size.get().copied()
+    /// Make the value computed by initial logical size computation
+    /// available for re-use. This doesn't contain the incremental part.
+    fn initialized_size(&self, lsn: Lsn) -> Option<u64> {
+        match self.initial_part_end {
+            Some(v) if v == lsn => self.initial_logical_size.get().copied(),
+            _ => None,
+        }
     }
 }
 
@@ -668,8 +679,7 @@ impl Timeline {
 
             let mut failed = 0;
 
-            let cancelled = task_mgr::shutdown_watcher();
-            tokio::pin!(cancelled);
+            let mut cancelled = pin!(task_mgr::shutdown_watcher());
 
             loop {
                 tokio::select! {
@@ -806,11 +816,11 @@ impl Timeline {
 
         let mut is_exact = true;
         let size = current_size.size();
-        if let (CurrentLogicalSize::Approximate(_), Some(init_lsn)) =
+        if let (CurrentLogicalSize::Approximate(_), Some(initial_part_end)) =
             (current_size, self.current_logical_size.initial_part_end)
         {
             is_exact = false;
-            self.try_spawn_size_init_task(init_lsn, ctx);
+            self.try_spawn_size_init_task(initial_part_end, ctx);
         }
 
         Ok((size, is_exact))
@@ -948,6 +958,25 @@ impl Timeline {
         }
     }
 
+    /// Evict a batch of layers.
+    ///
+    /// GenericRemoteStorage reference is required as a witness[^witness_article] for "remote storage is configured."
+    ///
+    /// [^witness_article]: https://willcrichton.net/rust-api-type-patterns/witnesses.html
+    pub async fn evict_layers(
+        &self,
+        _: &GenericRemoteStorage,
+        layers_to_evict: &[Arc<dyn PersistentLayer>],
+        cancel: CancellationToken,
+    ) -> anyhow::Result<Vec<Option<anyhow::Result<bool>>>> {
+        let remote_client = self.remote_client.clone().expect(
+            "GenericRemoteStorage is configured, so timeline must have RemoteTimelineClient",
+        );
+
+        self.evict_layer_batch(&remote_client, layers_to_evict, cancel)
+            .await
+    }
+
     /// Evict multiple layers at once, continuing through errors.
     ///
     /// Try to evict the given `layers_to_evict` by
@@ -985,6 +1014,15 @@ impl Timeline {
         // now lock out layer removal (compaction, gc, timeline deletion)
         let layer_removal_guard = self.layer_removal_cs.lock().await;
 
+        {
+            // to avoid racing with detach and delete_timeline
+            let state = self.current_state();
+            anyhow::ensure!(
+                state == TimelineState::Active,
+                "timeline is not active but {state:?}"
+            );
+        }
+
         // start the batch update
         let mut layer_map = self.layers.write().unwrap();
         let mut batch_updates = layer_map.batch_update();
@@ -1018,6 +1056,8 @@ impl Timeline {
         use super::layer_map::Replacement;
 
         if local_layer.is_remote_layer() {
+            // TODO(issue #3851): consider returning an err here instead of false,
+            // which is the same out the match later
             return Ok(false);
         }
 
@@ -1087,6 +1127,9 @@ impl Timeline {
                     self.metrics
                         .evictions_with_low_residence_duration
                         .observe(delta);
+                    info!(layer=%local_layer.short_id(), residence_millis=delta.as_millis(), "evicted layer after known residence period");
+                } else {
+                    info!(layer=%local_layer.short_id(), "evicted layer after unknown residence period");
                 }
 
                 true
@@ -1248,6 +1291,10 @@ impl Timeline {
                 download_all_remote_layers_task_info: RwLock::new(None),
 
                 state,
+
+                eviction_task_timeline_state: tokio::sync::Mutex::new(
+                    EvictionTaskTimelineState::default(),
+                ),
             };
             result.repartition_threshold = result.get_checkpoint_distance() / 10;
             result
@@ -1688,7 +1735,7 @@ impl Timeline {
         Ok(())
     }
 
-    fn try_spawn_size_init_task(self: &Arc<Self>, init_lsn: Lsn, ctx: &RequestContext) {
+    fn try_spawn_size_init_task(self: &Arc<Self>, lsn: Lsn, ctx: &RequestContext) {
         let permit = match Arc::clone(&self.current_logical_size.initial_size_computation)
             .try_acquire_owned()
         {
@@ -1725,8 +1772,11 @@ impl Timeline {
             false,
             // NB: don't log errors here, task_mgr will do that.
             async move {
+                // no cancellation here, because nothing really waits for this to complete compared
+                // to spawn_ondemand_logical_size_calculation.
+                let cancel = CancellationToken::new();
                 let calculated_size = match self_clone
-                    .logical_size_calculation_task(init_lsn, &background_ctx)
+                    .logical_size_calculation_task(lsn, &background_ctx, cancel)
                     .await
                 {
                     Ok(s) => s,
@@ -1781,6 +1831,7 @@ impl Timeline {
         self: &Arc<Self>,
         lsn: Lsn,
         ctx: RequestContext,
+        cancel: CancellationToken,
     ) -> oneshot::Receiver<Result<u64, CalculateLogicalSizeError>> {
         let (sender, receiver) = oneshot::channel();
         let self_clone = Arc::clone(self);
@@ -1800,7 +1851,9 @@ impl Timeline {
             "ondemand logical size calculation",
             false,
             async move {
-                let res = self_clone.logical_size_calculation_task(lsn, &ctx).await;
+                let res = self_clone
+                    .logical_size_calculation_task(lsn, &ctx, cancel)
+                    .await;
                 let _ = sender.send(res).ok();
                 Ok(()) // Receiver is responsible for handling errors
             },
@@ -1811,20 +1864,20 @@ impl Timeline {
     #[instrument(skip_all, fields(tenant = %self.tenant_id, timeline = %self.timeline_id))]
     async fn logical_size_calculation_task(
         self: &Arc<Self>,
-        init_lsn: Lsn,
+        lsn: Lsn,
         ctx: &RequestContext,
+        cancel: CancellationToken,
     ) -> Result<u64, CalculateLogicalSizeError> {
         let mut timeline_state_updates = self.subscribe_for_state_updates();
         let self_calculation = Arc::clone(self);
-        let cancel = CancellationToken::new();
 
-        let calculation = async {
+        let mut calculation = pin!(async {
             let cancel = cancel.child_token();
             let ctx = ctx.attached_child();
             self_calculation
-                .calculate_logical_size(init_lsn, cancel, &ctx)
+                .calculate_logical_size(lsn, cancel, &ctx)
                 .await
-        };
+        });
         let timeline_state_cancellation = async {
             loop {
                 match timeline_state_updates.changed().await {
@@ -1853,7 +1906,6 @@ impl Timeline {
             "aborted because task_mgr shutdown requested".to_string()
         };
 
-        tokio::pin!(calculation);
         loop {
             tokio::select! {
                 res = &mut calculation => { return res }
@@ -1906,21 +1958,12 @@ impl Timeline {
             // need to return something
             Ok(0)
         });
-        let timer = if up_to_lsn == self.initdb_lsn {
-            if let Some(size) = self.current_logical_size.initialized_size() {
-                if size != 0 {
-                    // non-zero size means that the size has already been calculated by this method
-                    // after startup. if the logical size is for a new timeline without layers the
-                    // size will be zero, and we cannot use that, or this caching strategy until
-                    // pageserver restart.
-                    return Ok(size);
-                }
-            }
-
-            self.metrics.init_logical_size_histo.start_timer()
-        } else {
-            self.metrics.logical_size_histo.start_timer()
-        };
+        // See if we've already done the work for initial size calculation.
+        // This is a short-cut for timelines that are mostly unused.
+        if let Some(size) = self.current_logical_size.initialized_size(up_to_lsn) {
+            return Ok(size);
+        }
+        let timer = self.metrics.logical_size_histo.start_timer();
         let logical_size = self
             .get_current_logical_size_non_incremental(up_to_lsn, cancel, ctx)
             .await?;
@@ -4003,6 +4046,67 @@ impl Timeline {
     }
 }
 
+pub struct DiskUsageEvictionInfo {
+    /// Timeline's largest layer (remote or resident)
+    pub max_layer_size: Option<u64>,
+    /// Timeline's resident layers
+    pub resident_layers: Vec<LocalLayerInfoForDiskUsageEviction>,
+}
+
+pub struct LocalLayerInfoForDiskUsageEviction {
+    pub layer: Arc<dyn PersistentLayer>,
+    pub last_activity_ts: SystemTime,
+}
+
+impl std::fmt::Debug for LocalLayerInfoForDiskUsageEviction {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        // format the tv_sec, tv_nsec into rfc3339 in case someone is looking at it
+        // having to allocate a string to this is bad, but it will rarely be formatted
+        let ts = chrono::DateTime::<chrono::Utc>::from(self.last_activity_ts);
+        let ts = ts.to_rfc3339_opts(chrono::SecondsFormat::Nanos, true);
+        f.debug_struct("LocalLayerInfoForDiskUsageEviction")
+            .field("layer", &self.layer)
+            .field("last_activity", &ts)
+            .finish()
+    }
+}
+
+impl LocalLayerInfoForDiskUsageEviction {
+    pub fn file_size(&self) -> u64 {
+        self.layer.file_size()
+    }
+}
+
+impl Timeline {
+    pub(crate) fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
+        let layers = self.layers.read().unwrap();
+
+        let mut max_layer_size: Option<u64> = None;
+        let mut resident_layers = Vec::new();
+
+        for l in layers.iter_historic_layers() {
+            let file_size = l.file_size();
+            max_layer_size = max_layer_size.map_or(Some(file_size), |m| Some(m.max(file_size)));
+
+            if l.is_remote_layer() {
+                continue;
+            }
+
+            let last_activity_ts = l.access_stats().latest_activity();
+
+            resident_layers.push(LocalLayerInfoForDiskUsageEviction {
+                layer: l,
+                last_activity_ts,
+            });
+        }
+
+        DiskUsageEvictionInfo {
+            max_layer_size,
+            resident_layers,
+        }
+    }
+}
+
 type TraversalPathItem = (
     ValueReconstructResult,
     Lsn,

pageserver/src/tenant/timeline/eviction_task.rs

@@ -14,12 +14,12 @@
 //!
 //! See write-up on restart on-demand download spike: <https://gist.github.com/problame/2265bf7b8dc398be834abfead36c76b5>
 use std::{
+    collections::HashMap,
     ops::ControlFlow,
     sync::Arc,
     time::{Duration, SystemTime},
 };
 
-use either::Either;
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, instrument, warn};
@@ -30,11 +30,22 @@ use crate::{
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
         storage_layer::PersistentLayer,
+        Tenant,
     },
 };
 
 use super::Timeline;
 
+#[derive(Default)]
+pub struct EvictionTaskTimelineState {
+    last_layer_access_imitation: Option<tokio::time::Instant>,
+}
+
+#[derive(Default)]
+pub struct EvictionTaskTenantState {
+    last_layer_access_imitation: Option<Instant>,
+}
+
 impl Timeline {
     pub(super) fn launch_eviction_task(self: &Arc<Self>) {
         let self_clone = Arc::clone(self);
@@ -122,6 +133,35 @@ impl Timeline {
     ) -> ControlFlow<()> {
         let now = SystemTime::now();
 
+        // If we evict layers but keep cached values derived from those layers, then
+        // we face a storm of on-demand downloads after pageserver restart.
+        // The reason is that the restart empties the caches, and so, the values
+        // need to be re-computed by accessing layers, which we evicted while the
+        // caches were filled.
+        //
+        // Solutions here would be one of the following:
+        // 1. Have a persistent cache.
+        // 2. Count every access to a cached value to the access stats of all layers
+        //    that were accessed to compute the value in the first place.
+        // 3. Invalidate the caches at a period of < p.threshold/2, so that the values
+        //    get re-computed from layers, thereby counting towards layer access stats.
+        // 4. Make the eviction task imitate the layer accesses that typically hit caches.
+        //
+        // We follow approach (4) here because in Neon prod deployment:
+        // - page cache is quite small => high churn => low hit rate
+        //   => eviction gets correct access stats
+        // - value-level caches such as logical size & repatition have a high hit rate,
+        //   especially for inactive tenants
+        //   => eviction sees zero accesses for these
+        //   => they cause the on-demand download storm on pageserver restart
+        //
+        // We should probably move to persistent caches in the future, or avoid
+        // having inactive tenants attached to pageserver in the first place.
+        match self.imitate_layer_accesses(p, cancel, ctx).await {
+            ControlFlow::Break(()) => return ControlFlow::Break(()),
+            ControlFlow::Continue(()) => (),
+        }
+
         #[allow(dead_code)]
         #[derive(Debug, Default)]
         struct EvictionStats {
@@ -132,19 +172,6 @@ impl Timeline {
             skipped_for_shutdown: usize,
         }
 
-        // what we want is to invalidate any caches which haven't been accessed for `p.threshold`,
-        // but we cannot actually do it for current limitations except by restarting pageserver. we
-        // just recompute the values which would be recomputed on startup.
-        //
-        // for active tenants this will likely materialized page cache or in-memory layers. for
-        // inactive tenants it will refresh the last_access timestamps so that we will not evict
-        // and re-download on restart these layers.
-        self.refresh_layers_required_in_restart(cancel, ctx).await;
-
-        if cancel.is_cancelled() {
-            return ControlFlow::Break(());
-        }
-
         let mut stats = EvictionStats::default();
         // Gather layers for eviction.
         // NB: all the checks can be invalidated as soon as we release the layer map lock.
@@ -157,13 +184,7 @@ impl Timeline {
                 if hist_layer.is_remote_layer() {
                     continue;
                 }
-                let last_activity_ts = match hist_layer
-                    .access_stats()
-                    .most_recent_access_or_residence_event()
-                {
-                    Either::Left(mra) => mra.when,
-                    Either::Right(re) => re.timestamp,
-                };
+                let last_activity_ts = hist_layer.access_stats().latest_activity();
                 let no_activity_for = match now.duration_since(last_activity_ts) {
                     Ok(d) => d,
                     Err(_e) => {
@@ -248,8 +269,55 @@ impl Timeline {
         ControlFlow::Continue(())
     }
 
+    async fn imitate_layer_accesses(
+        &self,
+        p: &EvictionPolicyLayerAccessThreshold,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
+    ) -> ControlFlow<()> {
+        let mut state = self.eviction_task_timeline_state.lock().await;
+        match state.last_layer_access_imitation {
+            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
+            _ => {
+                self.imitate_timeline_cached_layer_accesses(cancel, ctx)
+                    .await;
+                state.last_layer_access_imitation = Some(tokio::time::Instant::now())
+            }
+        }
+        drop(state);
+
+        if cancel.is_cancelled() {
+            return ControlFlow::Break(());
+        }
+
+        // This task is timeline-scoped, but the synthetic size calculation is tenant-scoped.
+        // Make one of the tenant's timelines draw the short straw and run the calculation.
+        // The others wait until the calculation is done so that they take into account the
+        // imitated accesses that the winner made.
+        let Ok(tenant) = crate::tenant::mgr::get_tenant(self.tenant_id, true).await else {
+            // likely, we're shutting down
+            return ControlFlow::Break(());
+        };
+        let mut state = tenant.eviction_task_tenant_state.lock().await;
+        match state.last_layer_access_imitation {
+            Some(ts) if ts.elapsed() < p.threshold => { /* no need to run */ }
+            _ => {
+                self.imitate_synthetic_size_calculation_worker(&tenant, ctx, cancel)
+                    .await;
+                state.last_layer_access_imitation = Some(tokio::time::Instant::now());
+            }
+        }
+        drop(state);
+
+        if cancel.is_cancelled() {
+            return ControlFlow::Break(());
+        }
+
+        ControlFlow::Continue(())
+    }
+
     /// Recompute the values which would cause on-demand downloads during restart.
-    async fn refresh_layers_required_in_restart(
+    async fn imitate_timeline_cached_layer_accesses(
         &self,
         cancel: &CancellationToken,
         ctx: &RequestContext,
@@ -283,4 +351,61 @@ impl Timeline {
             }
         }
     }
+
+    // Imitate the synthetic size calculation done by the consumption_metrics module.
+    async fn imitate_synthetic_size_calculation_worker(
+        &self,
+        tenant: &Arc<Tenant>,
+        ctx: &RequestContext,
+        cancel: &CancellationToken,
+    ) {
+        if self.conf.metric_collection_endpoint.is_none() {
+            // We don't start the consumption metrics task if this is not set in the config.
+            // So, no need to imitate the accesses in that case.
+            return;
+        }
+
+        // The consumption metrics are collected on a per-tenant basis, by a single
+        // global background loop.
+        // It limits the number of synthetic size calculations using the global
+        // `concurrent_tenant_size_logical_size_queries` semaphore to not overload
+        // the pageserver. (size calculation is somewhat expensive in terms of CPU and IOs).
+        //
+        // If we used that same semaphore here, then we'd compete for the
+        // same permits, which may impact timeliness of consumption metrics.
+        // That is a no-go, as consumption metrics are much more important
+        // than what we do here.
+        //
+        // So, we have a separate semaphore, initialized to the same
+        // number of permits as the `concurrent_tenant_size_logical_size_queries`.
+        // In the worst, we would have twice the amount of concurrenct size calculations.
+        // But in practice, the `p.threshold` >> `consumption metric interval`, and
+        // we spread out the eviction task using `random_init_delay`.
+        // So, the chance of the worst case is quite low in practice.
+        // It runs as a per-tenant task, but the eviction_task.rs is per-timeline.
+        // So, we must coordinate with other with other eviction tasks of this tenant.
+        let limit = self
+            .conf
+            .eviction_task_immitated_concurrent_logical_size_queries
+            .inner();
+
+        let mut throwaway_cache = HashMap::new();
+        let gather =
+            crate::tenant::size::gather_inputs(tenant, limit, None, &mut throwaway_cache, ctx);
+
+        tokio::select! {
+            _ = cancel.cancelled() => {}
+            gather_result = gather => {
+                match gather_result {
+                    Ok(_) => {},
+                    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
+                        // thus not encountering this error.
+                        warn!("failed to imitate synthetic size calculation accesses: {e:#}")
+                    }
+                }
+           }
+        }
+    }
 }

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

@@ -237,11 +237,7 @@ async fn connection_manager_loop_step(
         if let Some(new_candidate) = walreceiver_state.next_connection_candidate() {
             info!("Switching to new connection candidate: {new_candidate:?}");
             walreceiver_state
-                .change_connection(
-                    new_candidate.safekeeper_id,
-                    new_candidate.wal_source_connconf,
-                    ctx,
-                )
+                .change_connection(new_candidate, ctx)
                 .await
         }
     }
@@ -346,6 +342,8 @@ struct WalConnection {
     started_at: NaiveDateTime,
     /// Current safekeeper pageserver is connected to for WAL streaming.
     sk_id: NodeId,
+    /// Availability zone of the safekeeper.
+    availability_zone: Option<String>,
     /// Status of the connection.
     status: WalConnectionStatus,
     /// WAL streaming task handle.
@@ -405,12 +403,7 @@ impl WalreceiverState {
     }
 
     /// Shuts down the current connection (if any) and immediately starts another one with the given connection string.
-    async fn change_connection(
-        &mut self,
-        new_sk_id: NodeId,
-        new_wal_source_connconf: PgConnectionConfig,
-        ctx: &RequestContext,
-    ) {
+    async fn change_connection(&mut self, new_sk: NewWalConnectionCandidate, ctx: &RequestContext) {
         self.drop_old_connection(true).await;
 
         let id = self.id;
@@ -424,7 +417,7 @@ impl WalreceiverState {
             async move {
                 super::walreceiver_connection::handle_walreceiver_connection(
                     timeline,
-                    new_wal_source_connconf,
+                    new_sk.wal_source_connconf,
                     events_sender,
                     cancellation,
                     connect_timeout,
@@ -433,13 +426,16 @@ impl WalreceiverState {
                 .await
                 .context("walreceiver connection handling failure")
             }
-            .instrument(info_span!("walreceiver_connection", id = %id, node_id = %new_sk_id))
+            .instrument(
+                info_span!("walreceiver_connection", id = %id, node_id = %new_sk.safekeeper_id),
+            )
         });
 
         let now = Utc::now().naive_utc();
         self.wal_connection = Some(WalConnection {
             started_at: now,
-            sk_id: new_sk_id,
+            sk_id: new_sk.safekeeper_id,
+            availability_zone: new_sk.availability_zone,
             status: WalConnectionStatus {
                 is_connected: false,
                 has_processed_wal: false,
@@ -546,6 +542,7 @@ impl WalreceiverState {
     /// * if connected safekeeper is not present, pick the candidate
     /// * if we haven't received any updates for some time, pick the candidate
     /// * if the candidate commit_lsn is much higher than the current one, pick the candidate
+    /// * if the candidate commit_lsn is same, but candidate is located in the same AZ as the pageserver, pick the candidate
     /// * if connected safekeeper stopped sending us new WAL which is available on other safekeeper, pick the candidate
     ///
     /// This way we ensure to keep up with the most up-to-date safekeeper and don't try to jump from one safekeeper to another too frequently.
@@ -559,6 +556,7 @@ impl WalreceiverState {
 
                 let (new_sk_id, new_safekeeper_broker_data, new_wal_source_connconf) =
                     self.select_connection_candidate(Some(connected_sk_node))?;
+                let new_availability_zone = new_safekeeper_broker_data.availability_zone.clone();
 
                 let now = Utc::now().naive_utc();
                 if let Ok(latest_interaciton) =
@@ -569,6 +567,7 @@ impl WalreceiverState {
                         return Some(NewWalConnectionCandidate {
                             safekeeper_id: new_sk_id,
                             wal_source_connconf: new_wal_source_connconf,
+                            availability_zone: new_availability_zone,
                             reason: ReconnectReason::NoKeepAlives {
                                 last_keep_alive: Some(
                                     existing_wal_connection.status.latest_connection_update,
@@ -594,6 +593,7 @@ impl WalreceiverState {
                                 return Some(NewWalConnectionCandidate {
                                     safekeeper_id: new_sk_id,
                                     wal_source_connconf: new_wal_source_connconf,
+                                    availability_zone: new_availability_zone,
                                     reason: ReconnectReason::LaggingWal {
                                         current_commit_lsn,
                                         new_commit_lsn,
@@ -601,6 +601,20 @@ impl WalreceiverState {
                                     },
                                 });
                             }
+                            // If we have a candidate with the same commit_lsn as the current one, which is in the same AZ as pageserver,
+                            // and the current one is not, switch to the new one.
+                            if self.availability_zone.is_some()
+                                && existing_wal_connection.availability_zone
+                                    != self.availability_zone
+                                && self.availability_zone == new_availability_zone
+                            {
+                                return Some(NewWalConnectionCandidate {
+                                    safekeeper_id: new_sk_id,
+                                    availability_zone: new_availability_zone,
+                                    wal_source_connconf: new_wal_source_connconf,
+                                    reason: ReconnectReason::SwitchAvailabilityZone,
+                                });
+                            }
                         }
                         None => debug!(
                             "Best SK candidate has its commit_lsn behind connected SK's commit_lsn"
@@ -668,6 +682,7 @@ impl WalreceiverState {
                             return Some(NewWalConnectionCandidate {
                                 safekeeper_id: new_sk_id,
                                 wal_source_connconf: new_wal_source_connconf,
+                                availability_zone: new_availability_zone,
                                 reason: ReconnectReason::NoWalTimeout {
                                     current_lsn,
                                     current_commit_lsn,
@@ -686,10 +701,11 @@ impl WalreceiverState {
                 self.wal_connection.as_mut().unwrap().discovered_new_wal = discovered_new_wal;
             }
             None => {
-                let (new_sk_id, _, new_wal_source_connconf) =
+                let (new_sk_id, new_safekeeper_broker_data, new_wal_source_connconf) =
                     self.select_connection_candidate(None)?;
                 return Some(NewWalConnectionCandidate {
                     safekeeper_id: new_sk_id,
+                    availability_zone: new_safekeeper_broker_data.availability_zone.clone(),
                     wal_source_connconf: new_wal_source_connconf,
                     reason: ReconnectReason::NoExistingConnection,
                 });
@@ -794,6 +810,7 @@ impl WalreceiverState {
 struct NewWalConnectionCandidate {
     safekeeper_id: NodeId,
     wal_source_connconf: PgConnectionConfig,
+    availability_zone: Option<String>,
     // This field is used in `derive(Debug)` only.
     #[allow(dead_code)]
     reason: ReconnectReason,
@@ -808,6 +825,7 @@ enum ReconnectReason {
         new_commit_lsn: Lsn,
         threshold: NonZeroU64,
     },
+    SwitchAvailabilityZone,
     NoWalTimeout {
         current_lsn: Lsn,
         current_commit_lsn: Lsn,
@@ -873,6 +891,7 @@ mod tests {
                 peer_horizon_lsn: 0,
                 local_start_lsn: 0,
                 safekeeper_connstr: safekeeper_connstr.to_owned(),
+                availability_zone: None,
             },
             latest_update,
         }
@@ -933,6 +952,7 @@ mod tests {
         state.wal_connection = Some(WalConnection {
             started_at: now,
             sk_id: connected_sk_id,
+            availability_zone: None,
             status: connection_status,
             connection_task: TaskHandle::spawn(move |sender, _| async move {
                 sender
@@ -1095,6 +1115,7 @@ mod tests {
         state.wal_connection = Some(WalConnection {
             started_at: now,
             sk_id: connected_sk_id,
+            availability_zone: None,
             status: connection_status,
             connection_task: TaskHandle::spawn(move |sender, _| async move {
                 sender
@@ -1160,6 +1181,7 @@ mod tests {
         state.wal_connection = Some(WalConnection {
             started_at: now,
             sk_id: NodeId(1),
+            availability_zone: None,
             status: connection_status,
             connection_task: TaskHandle::spawn(move |sender, _| async move {
                 sender
@@ -1222,6 +1244,7 @@ mod tests {
         state.wal_connection = Some(WalConnection {
             started_at: now,
             sk_id: NodeId(1),
+            availability_zone: None,
             status: connection_status,
             connection_task: TaskHandle::spawn(move |_, _| async move { Ok(()) }),
             discovered_new_wal: Some(NewCommittedWAL {
@@ -1289,4 +1312,74 @@ mod tests {
             availability_zone: None,
         }
     }
+
+    #[tokio::test]
+    async fn switch_to_same_availability_zone() -> anyhow::Result<()> {
+        // Pageserver and one of safekeepers will be in the same availability zone
+        // and pageserver should prefer to connect to it.
+        let test_az = Some("test_az".to_owned());
+
+        let harness = TenantHarness::create("switch_to_same_availability_zone")?;
+        let mut state = dummy_state(&harness).await;
+        state.availability_zone = test_az.clone();
+        let current_lsn = Lsn(100_000).align();
+        let now = Utc::now().naive_utc();
+
+        let connected_sk_id = NodeId(0);
+
+        let connection_status = WalConnectionStatus {
+            is_connected: true,
+            has_processed_wal: true,
+            latest_connection_update: now,
+            latest_wal_update: now,
+            commit_lsn: Some(current_lsn),
+            streaming_lsn: Some(current_lsn),
+        };
+
+        state.wal_connection = Some(WalConnection {
+            started_at: now,
+            sk_id: connected_sk_id,
+            availability_zone: None,
+            status: connection_status,
+            connection_task: TaskHandle::spawn(move |sender, _| async move {
+                sender
+                    .send(TaskStateUpdate::Progress(connection_status))
+                    .ok();
+                Ok(())
+            }),
+            discovered_new_wal: None,
+        });
+
+        // We have another safekeeper with the same commit_lsn, and it have the same availability zone as
+        // the current pageserver.
+        let mut same_az_sk = dummy_broker_sk_timeline(current_lsn.0, "same_az", now);
+        same_az_sk.timeline.availability_zone = test_az.clone();
+
+        state.wal_stream_candidates = HashMap::from([
+            (
+                connected_sk_id,
+                dummy_broker_sk_timeline(current_lsn.0, DUMMY_SAFEKEEPER_HOST, now),
+            ),
+            (NodeId(1), same_az_sk),
+        ]);
+
+        // We expect that pageserver will switch to the safekeeper in the same availability zone,
+        // even if it has the same commit_lsn.
+        let next_candidate = state.next_connection_candidate().expect(
+            "Expected one candidate selected out of multiple valid data options, but got none",
+        );
+
+        assert_eq!(next_candidate.safekeeper_id, NodeId(1));
+        assert_eq!(
+            next_candidate.reason,
+            ReconnectReason::SwitchAvailabilityZone,
+            "Should switch to the safekeeper in the same availability zone, if it has the same commit_lsn"
+        );
+        assert_eq!(
+            next_candidate.wal_source_connconf.host(),
+            &Host::Domain("same_az".to_owned())
+        );
+
+        Ok(())
+    }
 }

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

@@ -2,6 +2,7 @@
 
 use std::{
     error::Error,
+    pin::pin,
     str::FromStr,
     sync::Arc,
     time::{Duration, SystemTime},
@@ -17,7 +18,7 @@ use postgres_ffi::v14::xlog_utils::normalize_lsn;
 use postgres_ffi::WAL_SEGMENT_SIZE;
 use postgres_protocol::message::backend::ReplicationMessage;
 use postgres_types::PgLsn;
-use tokio::{pin, select, sync::watch, time};
+use tokio::{select, sync::watch, time};
 use tokio_postgres::{replication::ReplicationStream, Client};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, trace, warn};
@@ -36,7 +37,7 @@ use crate::{
 use postgres_backend::is_expected_io_error;
 use postgres_connection::PgConnectionConfig;
 use postgres_ffi::waldecoder::WalStreamDecoder;
-use pq_proto::ReplicationFeedback;
+use pq_proto::PageserverFeedback;
 use utils::lsn::Lsn;
 
 /// Status of the connection.
@@ -187,8 +188,7 @@ pub async fn handle_walreceiver_connection(
     let query = format!("START_REPLICATION PHYSICAL {startpoint}");
 
     let copy_stream = replication_client.copy_both_simple(&query).await?;
-    let physical_stream = ReplicationStream::new(copy_stream);
-    pin!(physical_stream);
+    let mut physical_stream = pin!(ReplicationStream::new(copy_stream));
 
     let mut waldecoder = WalStreamDecoder::new(startpoint, timeline.pg_version);
 
@@ -319,12 +319,12 @@ pub async fn handle_walreceiver_connection(
                 timeline.get_remote_consistent_lsn().unwrap_or(Lsn(0));
 
             // The last LSN we processed. It is not guaranteed to survive pageserver crash.
-            let write_lsn = u64::from(last_lsn);
+            let last_received_lsn = u64::from(last_lsn);
             // `disk_consistent_lsn` is the LSN at which page server guarantees local persistence of all received data
-            let flush_lsn = u64::from(timeline.get_disk_consistent_lsn());
+            let disk_consistent_lsn = u64::from(timeline.get_disk_consistent_lsn());
             // The last LSN that is synced to remote storage and is guaranteed to survive pageserver crash
             // Used by safekeepers to remove WAL preceding `remote_consistent_lsn`.
-            let apply_lsn = u64::from(timeline_remote_consistent_lsn);
+            let remote_consistent_lsn = u64::from(timeline_remote_consistent_lsn);
             let ts = SystemTime::now();
 
             // Update the status about what we just received. This is shown in the mgmt API.
@@ -343,12 +343,12 @@ pub async fn handle_walreceiver_connection(
             let (timeline_logical_size, _) = timeline
                 .get_current_logical_size(&ctx)
                 .context("Status update creation failed to get current logical size")?;
-            let status_update = ReplicationFeedback {
+            let status_update = PageserverFeedback {
                 current_timeline_size: timeline_logical_size,
-                ps_writelsn: write_lsn,
-                ps_flushlsn: flush_lsn,
-                ps_applylsn: apply_lsn,
-                ps_replytime: ts,
+                last_received_lsn,
+                disk_consistent_lsn,
+                remote_consistent_lsn,
+                replytime: ts,
             };
 
             debug!("neon_status_update {status_update:?}");

pgxn/neon/file_cache.c

@@ -14,6 +14,7 @@
  */
 
 #include <sys/file.h>
+#include <sys/statvfs.h>
 #include <unistd.h>
 #include <fcntl.h>
 
@@ -34,6 +35,9 @@
 #include "storage/fd.h"
 #include "storage/pg_shmem.h"
 #include "storage/buf_internals.h"
+#include "storage/procsignal.h"
+#include "postmaster/bgworker.h"
+#include "postmaster/interrupt.h"
 
 /*
  * Local file cache is used to temporary store relations pages in local file system.
@@ -59,6 +63,9 @@
 
 #define SIZE_MB_TO_CHUNKS(size) ((uint32)((size) * MB / BLCKSZ / BLOCKS_PER_CHUNK))
 
+#define MAX_MONITOR_INTERVAL_USEC 1000000 /* 1 second */
+#define MAX_DISK_WRITE_RATE       1000 /* MB/sec */
+
 typedef struct FileCacheEntry
 {
 	BufferTag	key;
@@ -71,6 +78,7 @@ typedef struct FileCacheEntry
 typedef struct FileCacheControl
 {
 	uint32 size; /* size of cache file in chunks */
+	uint32 used; /* number of used chunks */
 	dlist_head lru; /* double linked list for LRU replacement algorithm */
 } FileCacheControl;
 
@@ -79,12 +87,14 @@ static int   lfc_desc;
 static LWLockId lfc_lock;
 static int   lfc_max_size;
 static int   lfc_size_limit;
+static int   lfc_free_space_watermark;
 static char* lfc_path;
 static  FileCacheControl* lfc_ctl;
 static shmem_startup_hook_type prev_shmem_startup_hook;
 #if PG_VERSION_NUM>=150000
 static shmem_request_hook_type prev_shmem_request_hook;
 #endif
+static int   lfc_shrinking_factor; /* power of two by which local cache size will be shrinked when lfc_free_space_watermark is reached */
 
 static void
 lfc_shmem_startup(void)
@@ -112,6 +122,7 @@ lfc_shmem_startup(void)
 								 &info,
 								 HASH_ELEM | HASH_BLOBS);
 		lfc_ctl->size = 0;
+		lfc_ctl->used = 0;
 		dlist_init(&lfc_ctl->lru);
 
 		/* Remove file cache on restart */
@@ -165,7 +176,7 @@ lfc_change_limit_hook(int newval, void *extra)
 		}
 	}
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
-	while (new_size < lfc_ctl->size && !dlist_is_empty(&lfc_ctl->lru))
+	while (new_size < lfc_ctl->used && !dlist_is_empty(&lfc_ctl->lru))
 	{
 		/* Shrink cache by throwing away least recently accessed chunks and returning their space to file system */
 		FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
@@ -175,12 +186,86 @@ lfc_change_limit_hook(int newval, void *extra)
 			elog(LOG, "Failed to punch hole in file: %m");
 #endif
 		hash_search(lfc_hash, &victim->key, HASH_REMOVE, NULL);
-		lfc_ctl->size -= 1;
+		lfc_ctl->used -= 1;
 	}
 	elog(LOG, "set local file cache limit to %d", new_size);
 	LWLockRelease(lfc_lock);
 }
 
+/*
+ * Local file system state monitor check available free space.
+ * If it is lower than lfc_free_space_watermark then we shrink size of local cache
+ * but throwing away least recently accessed chunks.
+ * First time low space watermark is reached cache size is divided by two,
+ * second time by four,... Finally we remove all chunks from local cache.
+ *
+ * Please notice that we are not changing lfc_cache_size: it is used to be adjusted by autoscaler.
+ * We only throw away cached chunks but do not prevent from filling cache by new chunks.
+ *
+ * Interval of poooling cache state is calculated as minimal time needed to consume lfc_free_space_watermark
+ * disk space with maximal possible disk write speed (1Gb/sec). But not larger than 1 second.
+ * Calling statvfs each second should not add any noticeable overhead.
+ */
+void
+FileCacheMonitorMain(Datum main_arg)
+{
+	/*
+	 * Choose file system state monitor interval so that space can not be exosted
+	 * during this period but not longer than  MAX_MONITOR_INTERVAL (10 sec)
+	 */
+	uint64 monitor_interval = Min(MAX_MONITOR_INTERVAL_USEC, lfc_free_space_watermark*MB/MAX_DISK_WRITE_RATE);
+
+	/* Establish signal handlers. */
+	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
+	pqsignal(SIGHUP, SignalHandlerForConfigReload);
+	pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
+	BackgroundWorkerUnblockSignals();
+
+	/* Periodically dump buffers until terminated. */
+	while (!ShutdownRequestPending)
+	{
+		if (lfc_size_limit != 0)
+		{
+			struct statvfs sfs;
+			if (statvfs(lfc_path, &sfs) < 0)
+			{
+				elog(WARNING, "Failed to obtain status of %s: %m", lfc_path);
+			}
+			else
+			{
+				if (sfs.f_bavail*sfs.f_bsize < lfc_free_space_watermark*MB)
+				{
+					if (lfc_shrinking_factor < 31) {
+						lfc_shrinking_factor += 1;
+					}
+					lfc_change_limit_hook(lfc_size_limit >> lfc_shrinking_factor, NULL);
+				}
+				else
+					lfc_shrinking_factor = 0; /* reset to initial value */
+			}
+		}
+		pg_usleep(monitor_interval);
+	}
+}
+
+static void
+lfc_register_free_space_monitor(void)
+{
+	BackgroundWorker bgw;
+	memset(&bgw, 0, sizeof(bgw));
+	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
+	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
+	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "neon");
+	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "FileCacheMonitorMain");
+	snprintf(bgw.bgw_name, BGW_MAXLEN, "Local free space monitor");
+	snprintf(bgw.bgw_type, BGW_MAXLEN, "Local free space monitor");
+	bgw.bgw_restart_time = 5;
+	bgw.bgw_notify_pid = 0;
+	bgw.bgw_main_arg = (Datum) 0;
+
+	RegisterBackgroundWorker(&bgw);
+}
+
 void
 lfc_init(void)
 {
@@ -217,6 +302,19 @@ lfc_init(void)
 							lfc_change_limit_hook,
 							NULL);
 
+	DefineCustomIntVariable("neon.free_space_watermark",
+							"Minimal free space in local file system after reaching which local file cache will be truncated",
+							NULL,
+							&lfc_free_space_watermark,
+							1024, /* 1GB */
+							0,
+							INT_MAX,
+							PGC_SIGHUP,
+							GUC_UNIT_MB,
+							NULL,
+							NULL,
+							NULL);
+
 	DefineCustomStringVariable("neon.file_cache_path",
 							   "Path to local file cache (can be raw device)",
 							   NULL,
@@ -231,6 +329,9 @@ lfc_init(void)
 	if (lfc_max_size == 0)
 		return;
 
+	if (lfc_free_space_watermark != 0)
+		lfc_register_free_space_monitor();
+
 	prev_shmem_startup_hook = shmem_startup_hook;
 	shmem_startup_hook = lfc_shmem_startup;
 #if PG_VERSION_NUM>=150000
@@ -380,7 +481,7 @@ lfc_write(RelFileNode rnode, ForkNumber forkNum, BlockNumber blkno,
 		 * there are should be very large number of concurrent IO operations and them are limited by max_connections,
 		 * we prefer not to complicate code and use second approach.
 		 */
-		if (lfc_ctl->size >= SIZE_MB_TO_CHUNKS(lfc_size_limit) && !dlist_is_empty(&lfc_ctl->lru))
+		if (lfc_ctl->used >= SIZE_MB_TO_CHUNKS(lfc_size_limit) && !dlist_is_empty(&lfc_ctl->lru))
 		{
 			/* Cache overflow: evict least recently used chunk */
 			FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
@@ -390,7 +491,10 @@ lfc_write(RelFileNode rnode, ForkNumber forkNum, BlockNumber blkno,
 			elog(LOG, "Swap file cache page");
 		}
 		else
+		{
+			lfc_ctl->used += 1;
 			entry->offset = lfc_ctl->size++; /* allocate new chunk at end of file */
+		}
 		entry->access_count = 1;
 		memset(entry->bitmap, 0, sizeof entry->bitmap);
 	}

pgxn/neon/walproposer.c

@@ -1872,9 +1872,9 @@ RecvAppendResponses(Safekeeper *sk)
 	return sk->state == SS_ACTIVE;
 }
 
-/* Parse a ReplicationFeedback message, or the ReplicationFeedback part of an AppendResponse */
+/* Parse a PageserverFeedback message, or the PageserverFeedback part of an AppendResponse */
 void
-ParseReplicationFeedbackMessage(StringInfo reply_message, ReplicationFeedback * rf)
+ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback * rf)
 {
 	uint8		nkeys;
 	int			i;
@@ -1892,45 +1892,45 @@ ParseReplicationFeedbackMessage(StringInfo reply_message, ReplicationFeedback *
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
 			rf->currentClusterSize = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParseReplicationFeedbackMessage: current_timeline_size %lu",
+			elog(DEBUG2, "ParsePageserverFeedbackMessage: current_timeline_size %lu",
 				 rf->currentClusterSize);
 		}
-		else if (strcmp(key, "ps_writelsn") == 0)
+		else if ((strcmp(key, "ps_writelsn") == 0) || (strcmp(key, "last_received_lsn") == 0))
 		{
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
-			rf->ps_writelsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParseReplicationFeedbackMessage: ps_writelsn %X/%X",
-				 LSN_FORMAT_ARGS(rf->ps_writelsn));
+			rf->last_received_lsn = pq_getmsgint64(reply_message);
+			elog(DEBUG2, "ParsePageserverFeedbackMessage: last_received_lsn %X/%X",
+				 LSN_FORMAT_ARGS(rf->last_received_lsn));
 		}
-		else if (strcmp(key, "ps_flushlsn") == 0)
+		else if ((strcmp(key, "ps_flushlsn") == 0) || (strcmp(key, "disk_consistent_lsn") == 0))
 		{
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
-			rf->ps_flushlsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParseReplicationFeedbackMessage: ps_flushlsn %X/%X",
-				 LSN_FORMAT_ARGS(rf->ps_flushlsn));
+			rf->disk_consistent_lsn = pq_getmsgint64(reply_message);
+			elog(DEBUG2, "ParsePageserverFeedbackMessage: disk_consistent_lsn %X/%X",
+				 LSN_FORMAT_ARGS(rf->disk_consistent_lsn));
 		}
-		else if (strcmp(key, "ps_applylsn") == 0)
+		else if ((strcmp(key, "ps_applylsn") == 0) || (strcmp(key, "remote_consistent_lsn") == 0))
 		{
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
-			rf->ps_applylsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParseReplicationFeedbackMessage: ps_applylsn %X/%X",
-				 LSN_FORMAT_ARGS(rf->ps_applylsn));
+			rf->remote_consistent_lsn = pq_getmsgint64(reply_message);
+			elog(DEBUG2, "ParsePageserverFeedbackMessage: remote_consistent_lsn %X/%X",
+				 LSN_FORMAT_ARGS(rf->remote_consistent_lsn));
 		}
-		else if (strcmp(key, "ps_replytime") == 0)
+		else if ((strcmp(key, "ps_replytime") == 0) || (strcmp(key, "replytime") == 0))
 		{
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
-			rf->ps_replytime = pq_getmsgint64(reply_message);
+			rf->replytime = pq_getmsgint64(reply_message);
 			{
 				char	   *replyTimeStr;
 
 				/* Copy because timestamptz_to_str returns a static buffer */
-				replyTimeStr = pstrdup(timestamptz_to_str(rf->ps_replytime));
-				elog(DEBUG2, "ParseReplicationFeedbackMessage: ps_replytime %lu reply_time: %s",
-					 rf->ps_replytime, replyTimeStr);
+				replyTimeStr = pstrdup(timestamptz_to_str(rf->replytime));
+				elog(DEBUG2, "ParsePageserverFeedbackMessage: replytime %lu reply_time: %s",
+					 rf->replytime, replyTimeStr);
 
 				pfree(replyTimeStr);
 			}
@@ -1944,7 +1944,7 @@ ParseReplicationFeedbackMessage(StringInfo reply_message, ReplicationFeedback *
 			 * Skip unknown keys to support backward compatibile protocol
 			 * changes
 			 */
-			elog(LOG, "ParseReplicationFeedbackMessage: unknown key: %s len %d", key, len);
+			elog(LOG, "ParsePageserverFeedbackMessage: unknown key: %s len %d", key, len);
 			pq_getmsgbytes(reply_message, len);
 		};
 	}
@@ -2024,7 +2024,7 @@ GetAcknowledgedByQuorumWALPosition(void)
 }
 
 /*
- * ReplicationFeedbackShmemSize --- report amount of shared memory space needed
+ * WalproposerShmemSize --- report amount of shared memory space needed
  */
 Size
 WalproposerShmemSize(void)
@@ -2054,10 +2054,10 @@ WalproposerShmemInit(void)
 }
 
 void
-replication_feedback_set(ReplicationFeedback * rf)
+replication_feedback_set(PageserverFeedback * rf)
 {
 	SpinLockAcquire(&walprop_shared->mutex);
-	memcpy(&walprop_shared->feedback, rf, sizeof(ReplicationFeedback));
+	memcpy(&walprop_shared->feedback, rf, sizeof(PageserverFeedback));
 	SpinLockRelease(&walprop_shared->mutex);
 }
 
@@ -2065,43 +2065,43 @@ void
 replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn)
 {
 	SpinLockAcquire(&walprop_shared->mutex);
-	*writeLsn = walprop_shared->feedback.ps_writelsn;
-	*flushLsn = walprop_shared->feedback.ps_flushlsn;
-	*applyLsn = walprop_shared->feedback.ps_applylsn;
+	*writeLsn = walprop_shared->feedback.last_received_lsn;
+	*flushLsn = walprop_shared->feedback.disk_consistent_lsn;
+	*applyLsn = walprop_shared->feedback.remote_consistent_lsn;
 	SpinLockRelease(&walprop_shared->mutex);
 }
 
 /*
- * Get ReplicationFeedback fields from the most advanced safekeeper
+ * Get PageserverFeedback fields from the most advanced safekeeper
  */
 static void
-GetLatestNeonFeedback(ReplicationFeedback * rf)
+GetLatestNeonFeedback(PageserverFeedback * rf)
 {
 	int			latest_safekeeper = 0;
-	XLogRecPtr	ps_writelsn = InvalidXLogRecPtr;
+	XLogRecPtr	last_received_lsn = InvalidXLogRecPtr;
 
 	for (int i = 0; i < n_safekeepers; i++)
 	{
-		if (safekeeper[i].appendResponse.rf.ps_writelsn > ps_writelsn)
+		if (safekeeper[i].appendResponse.rf.last_received_lsn > last_received_lsn)
 		{
 			latest_safekeeper = i;
-			ps_writelsn = safekeeper[i].appendResponse.rf.ps_writelsn;
+			last_received_lsn = safekeeper[i].appendResponse.rf.last_received_lsn;
 		}
 	}
 
 	rf->currentClusterSize = safekeeper[latest_safekeeper].appendResponse.rf.currentClusterSize;
-	rf->ps_writelsn = safekeeper[latest_safekeeper].appendResponse.rf.ps_writelsn;
-	rf->ps_flushlsn = safekeeper[latest_safekeeper].appendResponse.rf.ps_flushlsn;
-	rf->ps_applylsn = safekeeper[latest_safekeeper].appendResponse.rf.ps_applylsn;
-	rf->ps_replytime = safekeeper[latest_safekeeper].appendResponse.rf.ps_replytime;
+	rf->last_received_lsn = safekeeper[latest_safekeeper].appendResponse.rf.last_received_lsn;
+	rf->disk_consistent_lsn = safekeeper[latest_safekeeper].appendResponse.rf.disk_consistent_lsn;
+	rf->remote_consistent_lsn = safekeeper[latest_safekeeper].appendResponse.rf.remote_consistent_lsn;
+	rf->replytime = safekeeper[latest_safekeeper].appendResponse.rf.replytime;
 
 	elog(DEBUG2, "GetLatestNeonFeedback: currentClusterSize %lu,"
-		 " ps_writelsn %X/%X, ps_flushlsn %X/%X, ps_applylsn %X/%X, ps_replytime %lu",
+		 " last_received_lsn %X/%X, disk_consistent_lsn %X/%X, remote_consistent_lsn %X/%X, replytime %lu",
 		 rf->currentClusterSize,
-		 LSN_FORMAT_ARGS(rf->ps_writelsn),
-		 LSN_FORMAT_ARGS(rf->ps_flushlsn),
-		 LSN_FORMAT_ARGS(rf->ps_applylsn),
-		 rf->ps_replytime);
+		 LSN_FORMAT_ARGS(rf->last_received_lsn),
+		 LSN_FORMAT_ARGS(rf->disk_consistent_lsn),
+		 LSN_FORMAT_ARGS(rf->remote_consistent_lsn),
+		 rf->replytime);
 
 	replication_feedback_set(rf);
 }
@@ -2115,16 +2115,16 @@ HandleSafekeeperResponse(void)
 	XLogRecPtr	minFlushLsn;
 
 	minQuorumLsn = GetAcknowledgedByQuorumWALPosition();
-	diskConsistentLsn = quorumFeedback.rf.ps_flushlsn;
+	diskConsistentLsn = quorumFeedback.rf.disk_consistent_lsn;
 
 	if (!syncSafekeepers)
 	{
-		/* Get ReplicationFeedback fields from the most advanced safekeeper */
+		/* Get PageserverFeedback fields from the most advanced safekeeper */
 		GetLatestNeonFeedback(&quorumFeedback.rf);
 		SetZenithCurrentClusterSize(quorumFeedback.rf.currentClusterSize);
 	}
 
-	if (minQuorumLsn > quorumFeedback.flushLsn || diskConsistentLsn != quorumFeedback.rf.ps_flushlsn)
+	if (minQuorumLsn > quorumFeedback.flushLsn || diskConsistentLsn != quorumFeedback.rf.disk_consistent_lsn)
 	{
 
 		if (minQuorumLsn > quorumFeedback.flushLsn)
@@ -2142,7 +2142,7 @@ HandleSafekeeperResponse(void)
 			 * apply_lsn - This is what processed and durably saved at*
 			 * pageserver.
 			 */
-								quorumFeedback.rf.ps_flushlsn,
+								quorumFeedback.rf.disk_consistent_lsn,
 								GetCurrentTimestamp(), false);
 	}
 
@@ -2326,7 +2326,7 @@ AsyncReadMessage(Safekeeper *sk, AcceptorProposerMessage * anymsg)
 				msg->hs.xmin.value = pq_getmsgint64_le(&s);
 				msg->hs.catalog_xmin.value = pq_getmsgint64_le(&s);
 				if (buf_size > APPENDRESPONSE_FIXEDPART_SIZE)
-					ParseReplicationFeedbackMessage(&s, &msg->rf);
+					ParsePageserverFeedbackMessage(&s, &msg->rf);
 				pq_getmsgend(&s);
 				return true;
 			}
@@ -2462,7 +2462,7 @@ backpressure_lag_impl(void)
 		replication_feedback_get_lsns(&writePtr, &flushPtr, &applyPtr);
 #define MB ((XLogRecPtr)1024 * 1024)
 
-		elog(DEBUG2, "current flushLsn %X/%X ReplicationFeedback: write %X/%X flush %X/%X apply %X/%X",
+		elog(DEBUG2, "current flushLsn %X/%X PageserverFeedback: write %X/%X flush %X/%X apply %X/%X",
 			 LSN_FORMAT_ARGS(myFlushLsn),
 			 LSN_FORMAT_ARGS(writePtr),
 			 LSN_FORMAT_ARGS(flushPtr),

pgxn/neon/walproposer.h

@@ -280,21 +280,21 @@ typedef struct HotStandbyFeedback
 	FullTransactionId catalog_xmin;
 }			HotStandbyFeedback;
 
-typedef struct ReplicationFeedback
+typedef struct PageserverFeedback
 {
 	/* current size of the timeline on pageserver */
 	uint64		currentClusterSize;
 	/* standby_status_update fields that safekeeper received from pageserver */
-	XLogRecPtr	ps_writelsn;
-	XLogRecPtr	ps_flushlsn;
-	XLogRecPtr	ps_applylsn;
-	TimestampTz ps_replytime;
-}			ReplicationFeedback;
+	XLogRecPtr	last_received_lsn;
+	XLogRecPtr	disk_consistent_lsn;
+	XLogRecPtr	remote_consistent_lsn;
+	TimestampTz replytime;
+}			PageserverFeedback;
 
 typedef struct WalproposerShmemState
 {
 	slock_t		mutex;
-	ReplicationFeedback feedback;
+	PageserverFeedback feedback;
 	term_t		mineLastElectedTerm;
 	pg_atomic_uint64 backpressureThrottlingTime;
 }			WalproposerShmemState;
@@ -320,10 +320,10 @@ typedef struct AppendResponse
 	/* Feedback recieved from pageserver includes standby_status_update fields */
 	/* and custom neon feedback. */
 	/* This part of the message is extensible. */
-	ReplicationFeedback rf;
+	PageserverFeedback rf;
 }			AppendResponse;
 
-/*  ReplicationFeedback is extensible part of the message that is parsed separately */
+/*  PageserverFeedback is extensible part of the message that is parsed separately */
 /*  Other fields are fixed part */
 #define APPENDRESPONSE_FIXEDPART_SIZE offsetof(AppendResponse, rf)
 
@@ -383,13 +383,13 @@ extern void WalProposerSync(int argc, char *argv[]);
 extern void WalProposerMain(Datum main_arg);
 extern void WalProposerBroadcast(XLogRecPtr startpos, XLogRecPtr endpos);
 extern void WalProposerPoll(void);
-extern void ParseReplicationFeedbackMessage(StringInfo reply_message,
-											ReplicationFeedback *rf);
+extern void ParsePageserverFeedbackMessage(StringInfo reply_message,
+											PageserverFeedback *rf);
 extern void StartProposerReplication(StartReplicationCmd *cmd);
 
 extern Size WalproposerShmemSize(void);
 extern bool WalproposerShmemInit(void);
-extern void replication_feedback_set(ReplicationFeedback *rf);
+extern void replication_feedback_set(PageserverFeedback *rf);
 extern void replication_feedback_get_lsns(XLogRecPtr *writeLsn, XLogRecPtr *flushLsn, XLogRecPtr *applyLsn);
 
 /* libpqwalproposer hooks & helper type */

poetry.lock

@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.4.0 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.4.1 and should not be changed by hand.
 
 [[package]]
 name = "aiohttp"
@@ -79,37 +79,35 @@ sa = ["sqlalchemy[postgresql-psycopg2binary] (>=1.3,<1.5)"]
 
 [[package]]
 name = "allure-pytest"
-version = "2.10.0"
+version = "2.13.1"
 description = "Allure pytest integration"
 category = "main"
 optional = false
 python-versions = "*"
 files = [
-    {file = "allure-pytest-2.10.0.tar.gz", hash = "sha256:3b2ab67629f4cbd8617abd817d2b22292c6eb7efd5584f992d1af8143aea6ee7"},
-    {file = "allure_pytest-2.10.0-py3-none-any.whl", hash = "sha256:08274096594758447db54c3b2c382526ee04f1fe12119cdaee92d2d93c84b530"},
+    {file = "allure-pytest-2.13.1.tar.gz", hash = "sha256:68d69456eeb65af4061ec06a80bc941163b0616e8216554d36b070a6bf070e08"},
+    {file = "allure_pytest-2.13.1-py3-none-any.whl", hash = "sha256:a8de2fc3b3effe2d8f98801646920de3f055b779710f4c806dbee7c613c24633"},
 ]
 
 [package.dependencies]
-allure-python-commons = "2.10.0"
+allure-python-commons = "2.13.1"
 pytest = ">=4.5.0"
-six = ">=1.9.0"
 
 [[package]]
 name = "allure-python-commons"
-version = "2.10.0"
+version = "2.13.1"
 description = "Common module for integrate allure with python-based frameworks"
 category = "main"
 optional = false
-python-versions = ">=3.5"
+python-versions = ">=3.6"
 files = [
-    {file = "allure-python-commons-2.10.0.tar.gz", hash = "sha256:d4d31344b0f0037a4a11e16b91b28cf0eeb23ffa0e50c27fcfc6aabe72212d3c"},
-    {file = "allure_python_commons-2.10.0-py3-none-any.whl", hash = "sha256:2a717e8ca8d296bf89cd57f38fc3c21893bd7ea8cd02a6ae5420e6d1a6eda5d0"},
+    {file = "allure-python-commons-2.13.1.tar.gz", hash = "sha256:3fc13e1da8ebb23f9ab5c9c72ad04595023cdd5078dbb8604939997faebed5cb"},
+    {file = "allure_python_commons-2.13.1-py3-none-any.whl", hash = "sha256:d08e04867bddf44fef55def3d67f4bc25af58a1bf9fcffcf4ec3331f7f2ef0d0"},
 ]
 
 [package.dependencies]
 attrs = ">=16.0.0"
 pluggy = ">=0.4.0"
-six = ">=1.9.0"
 
 [[package]]
 name = "async-timeout"
@@ -1932,6 +1930,22 @@ pytest = [
     {version = ">=6.2.4", markers = "python_version >= \"3.10\""},
 ]
 
+[[package]]
+name = "pytest-rerunfailures"
+version = "11.1.2"
+description = "pytest plugin to re-run tests to eliminate flaky failures"
+category = "main"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "pytest-rerunfailures-11.1.2.tar.gz", hash = "sha256:55611661e873f1cafa384c82f08d07883954f4b76435f4b8a5b470c1954573de"},
+    {file = "pytest_rerunfailures-11.1.2-py3-none-any.whl", hash = "sha256:d21fe2e46d9774f8ad95f1aa799544ae95cac3a223477af94aa985adfae92b7e"},
+]
+
+[package.dependencies]
+packaging = ">=17.1"
+pytest = ">=5.3"
+
 [[package]]
 name = "pytest-timeout"
 version = "2.1.0"
@@ -2597,4 +2611,4 @@ testing = ["func-timeout", "jaraco.itertools", "pytest (>=6)", "pytest-black (>=
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.9"
-content-hash = "2515a9320c2960076012fbc036fb33c4f6a23515c8d143785931dc18c6722d91"
+content-hash = "b689ffd6eae32b966f1744b5ac3343fe0dd26b31ee1f50e13daf5045ee0623e1"

proxy/src/auth/backend.rs

@@ -140,7 +140,7 @@ async fn auth_quirks(
 
 impl BackendType<'_, ClientCredentials<'_>> {
     /// Authenticate the client via the requested backend, possibly using credentials.
-    #[tracing::instrument(fields(allow_cleartext), skip_all)]
+    #[tracing::instrument(fields(allow_cleartext = allow_cleartext), skip_all)]
     pub async fn authenticate(
         &mut self,
         extra: &ConsoleReqExtra<'_>,

proxy/src/auth/backend/hacks.rs

@@ -53,7 +53,7 @@ pub async fn password_hack(
         .await?;
 
     info!(project = &payload.project, "received missing parameter");
-    creds.project = Some(payload.project.into());
+    creds.project = Some(payload.project);
 
     let mut node = api.wake_compute(extra, creds).await?;
     node.config.password(payload.password);

proxy/src/auth/credentials.rs

@@ -2,7 +2,7 @@
 
 use crate::error::UserFacingError;
 use pq_proto::StartupMessageParams;
-use std::borrow::Cow;
+use std::collections::HashSet;
 use thiserror::Error;
 use tracing::info;
 
@@ -19,11 +19,10 @@ pub enum ClientCredsParseError {
     InconsistentProjectNames { domain: String, option: String },
 
     #[error(
-        "SNI ('{}') inconsistently formatted with respect to common name ('{}'). \
-         SNI should be formatted as '<project-name>.{}'.",
-        .sni, .cn, .cn,
+        "Common name inferred from SNI ('{}') is not known",
+        .cn,
     )]
-    InconsistentSni { sni: String, cn: String },
+    UnknownCommonName { cn: String },
 
     #[error("Project name ('{0}') must contain only alphanumeric characters and hyphen.")]
     MalformedProjectName(String),
@@ -37,7 +36,7 @@ impl UserFacingError for ClientCredsParseError {}
 pub struct ClientCredentials<'a> {
     pub user: &'a str,
     // TODO: this is a severe misnomer! We should think of a new name ASAP.
-    pub project: Option<Cow<'a, str>>,
+    pub project: Option<String>,
 }
 
 impl ClientCredentials<'_> {
@@ -51,7 +50,7 @@ impl<'a> ClientCredentials<'a> {
     pub fn parse(
         params: &'a StartupMessageParams,
         sni: Option<&str>,
-        common_name: Option<&str>,
+        common_names: Option<HashSet<String>>,
     ) -> Result<Self, ClientCredsParseError> {
         use ClientCredsParseError::*;
 
@@ -60,37 +59,43 @@ impl<'a> ClientCredentials<'a> {
         let user = get_param("user")?;
 
         // Project name might be passed via PG's command-line options.
-        let project_option = params.options_raw().and_then(|mut options| {
-            options
-                .find_map(|opt| opt.strip_prefix("project="))
-                .map(Cow::Borrowed)
-        });
+        let project_option = params
+            .options_raw()
+            .and_then(|mut options| options.find_map(|opt| opt.strip_prefix("project=")))
+            .map(|name| name.to_string());
 
-        // Alternative project name is in fact a subdomain from SNI.
-        // NOTE: we do not consider SNI if `common_name` is missing.
-        let project_domain = sni
-            .zip(common_name)
-            .map(|(sni, cn)| {
-                subdomain_from_sni(sni, cn)
-                    .ok_or_else(|| InconsistentSni {
-                        sni: sni.into(),
-                        cn: cn.into(),
+        let project_from_domain = if let Some(sni_str) = sni {
+            if let Some(cn) = common_names {
+                let common_name_from_sni = sni_str.split_once('.').map(|(_, domain)| domain);
+
+                let project = common_name_from_sni
+                    .and_then(|domain| {
+                        if cn.contains(domain) {
+                            subdomain_from_sni(sni_str, domain)
+                        } else {
+                            None
+                        }
                     })
-                    .map(Cow::<'static, str>::Owned)
-            })
-            .transpose()?;
+                    .ok_or_else(|| UnknownCommonName {
+                        cn: common_name_from_sni.unwrap_or("").into(),
+                    })?;
 
-        let project = match (project_option, project_domain) {
+                Some(project)
+            } else {
+                None
+            }
+        } else {
+            None
+        };
+
+        let project = match (project_option, project_from_domain) {
             // Invariant: if we have both project name variants, they should match.
             (Some(option), Some(domain)) if option != domain => {
-                Some(Err(InconsistentProjectNames {
-                    domain: domain.into(),
-                    option: option.into(),
-                }))
+                Some(Err(InconsistentProjectNames { domain, option }))
             }
             // Invariant: project name may not contain certain characters.
             (a, b) => a.or(b).map(|name| match project_name_valid(&name) {
-                false => Err(MalformedProjectName(name.into())),
+                false => Err(MalformedProjectName(name)),
                 true => Ok(name),
             }),
         }
@@ -149,9 +154,9 @@ mod tests {
         let options = StartupMessageParams::new([("user", "john_doe")]);
 
         let sni = Some("foo.localhost");
-        let common_name = Some("localhost");
+        let common_names = Some(["localhost".into()].into());
 
-        let creds = ClientCredentials::parse(&options, sni, common_name)?;
+        let creds = ClientCredentials::parse(&options, sni, common_names)?;
         assert_eq!(creds.user, "john_doe");
         assert_eq!(creds.project.as_deref(), Some("foo"));
 
@@ -177,24 +182,41 @@ mod tests {
         let options = StartupMessageParams::new([("user", "john_doe"), ("options", "project=baz")]);
 
         let sni = Some("baz.localhost");
-        let common_name = Some("localhost");
+        let common_names = Some(["localhost".into()].into());
 
-        let creds = ClientCredentials::parse(&options, sni, common_name)?;
+        let creds = ClientCredentials::parse(&options, sni, common_names)?;
         assert_eq!(creds.user, "john_doe");
         assert_eq!(creds.project.as_deref(), Some("baz"));
 
         Ok(())
     }
 
+    #[test]
+    fn parse_multi_common_names() -> anyhow::Result<()> {
+        let options = StartupMessageParams::new([("user", "john_doe")]);
+
+        let common_names = Some(["a.com".into(), "b.com".into()].into());
+        let sni = Some("p1.a.com");
+        let creds = ClientCredentials::parse(&options, sni, common_names)?;
+        assert_eq!(creds.project.as_deref(), Some("p1"));
+
+        let common_names = Some(["a.com".into(), "b.com".into()].into());
+        let sni = Some("p1.b.com");
+        let creds = ClientCredentials::parse(&options, sni, common_names)?;
+        assert_eq!(creds.project.as_deref(), Some("p1"));
+
+        Ok(())
+    }
+
     #[test]
     fn parse_projects_different() {
         let options =
             StartupMessageParams::new([("user", "john_doe"), ("options", "project=first")]);
 
         let sni = Some("second.localhost");
-        let common_name = Some("localhost");
+        let common_names = Some(["localhost".into()].into());
 
-        let err = ClientCredentials::parse(&options, sni, common_name).expect_err("should fail");
+        let err = ClientCredentials::parse(&options, sni, common_names).expect_err("should fail");
         match err {
             InconsistentProjectNames { domain, option } => {
                 assert_eq!(option, "first");
@@ -209,13 +231,12 @@ mod tests {
         let options = StartupMessageParams::new([("user", "john_doe")]);
 
         let sni = Some("project.localhost");
-        let common_name = Some("example.com");
+        let common_names = Some(["example.com".into()].into());
 
-        let err = ClientCredentials::parse(&options, sni, common_name).expect_err("should fail");
+        let err = ClientCredentials::parse(&options, sni, common_names).expect_err("should fail");
         match err {
-            InconsistentSni { sni, cn } => {
-                assert_eq!(sni, "project.localhost");
-                assert_eq!(cn, "example.com");
+            UnknownCommonName { cn } => {
+                assert_eq!(cn, "localhost");
             }
             _ => panic!("bad error: {err:?}"),
         }

proxy/src/config.rs

@@ -1,6 +1,12 @@
 use crate::auth;
-use anyhow::{bail, ensure, Context};
-use std::{str::FromStr, sync::Arc, time::Duration};
+use anyhow::{bail, ensure, Context, Ok};
+use rustls::sign;
+use std::{
+    collections::{HashMap, HashSet},
+    str::FromStr,
+    sync::Arc,
+    time::Duration,
+};
 
 pub struct ProxyConfig {
     pub tls_config: Option<TlsConfig>,
@@ -16,7 +22,7 @@ pub struct MetricCollectionConfig {
 
 pub struct TlsConfig {
     pub config: Arc<rustls::ServerConfig>,
-    pub common_name: Option<String>,
+    pub common_names: Option<HashSet<String>>,
 }
 
 impl TlsConfig {
@@ -26,28 +32,34 @@ impl TlsConfig {
 }
 
 /// Configure TLS for the main endpoint.
-pub fn configure_tls(key_path: &str, cert_path: &str) -> anyhow::Result<TlsConfig> {
-    let key = {
-        let key_bytes = std::fs::read(key_path).context("TLS key file")?;
-        let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..])
-            .context(format!("Failed to read TLS keys at '{key_path}'"))?;
+pub fn configure_tls(
+    key_path: &str,
+    cert_path: &str,
+    certs_dir: Option<&String>,
+) -> anyhow::Result<TlsConfig> {
+    let mut cert_resolver = CertResolver::new();
 
-        ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
-        keys.pop().map(rustls::PrivateKey).unwrap()
-    };
+    // add default certificate
+    cert_resolver.add_cert(key_path, cert_path)?;
 
-    let cert_chain_bytes = std::fs::read(cert_path)
-        .context(format!("Failed to read TLS cert file at '{cert_path}.'"))?;
+    // add extra certificates
+    if let Some(certs_dir) = certs_dir {
+        for entry in std::fs::read_dir(certs_dir)? {
+            let entry = entry?;
+            let path = entry.path();
+            if path.is_dir() {
+                // file names aligned with default cert-manager names
+                let key_path = path.join("tls.key");
+                let cert_path = path.join("tls.crt");
+                if key_path.exists() && cert_path.exists() {
+                    cert_resolver
+                        .add_cert(&key_path.to_string_lossy(), &cert_path.to_string_lossy())?;
+                }
+            }
+        }
+    }
 
-    let cert_chain = {
-        rustls_pemfile::certs(&mut &cert_chain_bytes[..])
-            .context(format!(
-                "Failed to read TLS certificate chain from bytes from file at '{cert_path}'."
-            ))?
-            .into_iter()
-            .map(rustls::Certificate)
-            .collect()
-    };
+    let common_names = cert_resolver.get_common_names();
 
     let config = rustls::ServerConfig::builder()
         .with_safe_default_cipher_suites()
@@ -55,27 +67,105 @@ pub fn configure_tls(key_path: &str, cert_path: &str) -> anyhow::Result<TlsConfi
         // allow TLS 1.2 to be compatible with older client libraries
         .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])?
         .with_no_client_auth()
-        .with_single_cert(cert_chain, key)?
+        .with_cert_resolver(Arc::new(cert_resolver))
         .into();
 
-    // determine common name from tls-cert (-c server.crt param).
-    // used in asserting project name formatting invariant.
-    let common_name = {
-        let pem = x509_parser::pem::parse_x509_pem(&cert_chain_bytes)
-            .context(format!(
-                "Failed to parse PEM object from bytes from file at '{cert_path}'."
-            ))?
-            .1;
-        let common_name = pem.parse_x509()?.subject().to_string();
-        common_name.strip_prefix("CN=*.").map(|s| s.to_string())
-    };
-
     Ok(TlsConfig {
         config,
-        common_name,
+        common_names: Some(common_names),
     })
 }
 
+struct CertResolver {
+    certs: HashMap<String, Arc<rustls::sign::CertifiedKey>>,
+}
+
+impl CertResolver {
+    fn new() -> Self {
+        Self {
+            certs: HashMap::new(),
+        }
+    }
+
+    fn add_cert(&mut self, key_path: &str, cert_path: &str) -> anyhow::Result<()> {
+        let priv_key = {
+            let key_bytes = std::fs::read(key_path).context("TLS key file")?;
+            let mut keys = rustls_pemfile::pkcs8_private_keys(&mut &key_bytes[..])
+                .context(format!("Failed to read TLS keys at '{key_path}'"))?;
+
+            ensure!(keys.len() == 1, "keys.len() = {} (should be 1)", keys.len());
+            keys.pop().map(rustls::PrivateKey).unwrap()
+        };
+
+        let key = sign::any_supported_type(&priv_key).context("invalid private key")?;
+
+        let cert_chain_bytes = std::fs::read(cert_path)
+            .context(format!("Failed to read TLS cert file at '{cert_path}.'"))?;
+
+        let cert_chain = {
+            rustls_pemfile::certs(&mut &cert_chain_bytes[..])
+                .context(format!(
+                    "Failed to read TLS certificate chain from bytes from file at '{cert_path}'."
+                ))?
+                .into_iter()
+                .map(rustls::Certificate)
+                .collect()
+        };
+
+        let common_name = {
+            let pem = x509_parser::pem::parse_x509_pem(&cert_chain_bytes)
+                .context(format!(
+                    "Failed to parse PEM object from bytes from file at '{cert_path}'."
+                ))?
+                .1;
+            let common_name = pem.parse_x509()?.subject().to_string();
+            common_name.strip_prefix("CN=*.").map(|s| s.to_string())
+        }
+        .context(format!(
+            "Failed to parse common name from certificate at '{cert_path}'."
+        ))?;
+
+        self.certs.insert(
+            common_name,
+            Arc::new(rustls::sign::CertifiedKey::new(cert_chain, key)),
+        );
+
+        Ok(())
+    }
+
+    fn get_common_names(&self) -> HashSet<String> {
+        self.certs.keys().map(|s| s.to_string()).collect()
+    }
+}
+
+impl rustls::server::ResolvesServerCert for CertResolver {
+    fn resolve(
+        &self,
+        _client_hello: rustls::server::ClientHello,
+    ) -> Option<Arc<rustls::sign::CertifiedKey>> {
+        // loop here and cut off more and more subdomains until we find
+        // a match to get a proper wildcard support. OTOH, we now do not
+        // use nested domains, so keep this simple for now.
+        //
+        // With the current coding foo.com will match *.foo.com and that
+        // repeats behavior of the old code.
+        if let Some(mut sni_name) = _client_hello.server_name() {
+            loop {
+                if let Some(cert) = self.certs.get(sni_name) {
+                    return Some(cert.clone());
+                }
+                if let Some((_, rest)) = sni_name.split_once('.') {
+                    sni_name = rest;
+                } else {
+                    return None;
+                }
+            }
+        } else {
+            None
+        }
+    }
+}
+
 /// Helper for cmdline cache options parsing.
 pub struct CacheOptions {
     /// Max number of entries.

proxy/src/main.rs

@@ -132,7 +132,11 @@ fn build_config(args: &clap::ArgMatches) -> anyhow::Result<&'static ProxyConfig>
         args.get_one::<String>("tls-key"),
         args.get_one::<String>("tls-cert"),
     ) {
-        (Some(key_path), Some(cert_path)) => Some(config::configure_tls(key_path, cert_path)?),
+        (Some(key_path), Some(cert_path)) => Some(config::configure_tls(
+            key_path,
+            cert_path,
+            args.get_one::<String>("certs-dir"),
+        )?),
         (None, None) => None,
         _ => bail!("either both or neither tls-key and tls-cert must be specified"),
     };
@@ -254,6 +258,12 @@ fn cli() -> clap::Command {
                 .alias("ssl-cert") // backwards compatibility
                 .help("path to TLS cert for client postgres connections"),
         )
+        // tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
+        .arg(
+            Arg::new("certs-dir")
+                .long("certs-dir")
+                .help("path to directory with TLS certificates for client postgres connections"),
+        )
         .arg(
             Arg::new("metric-collection-endpoint")
                 .long("metric-collection-endpoint")

proxy/src/proxy.rs

@@ -98,7 +98,7 @@ pub async fn task_main(
 }
 
 // TODO(tech debt): unite this with its twin below.
-#[tracing::instrument(fields(session_id), skip_all)]
+#[tracing::instrument(fields(session_id = ?session_id), skip_all)]
 pub async fn handle_ws_client(
     config: &'static ProxyConfig,
     cancel_map: &CancelMap,
@@ -124,11 +124,11 @@ pub async fn handle_ws_client(
 
     // Extract credentials which we're going to use for auth.
     let creds = {
-        let common_name = tls.and_then(|tls| tls.common_name.as_deref());
+        let common_names = tls.and_then(|tls| tls.common_names.clone());
         let result = config
             .auth_backend
             .as_ref()
-            .map(|_| auth::ClientCredentials::parse(&params, hostname, common_name))
+            .map(|_| auth::ClientCredentials::parse(&params, hostname, common_names))
             .transpose();
 
         async { result }.or_else(|e| stream.throw_error(e)).await?
@@ -140,7 +140,7 @@ pub async fn handle_ws_client(
         .await
 }
 
-#[tracing::instrument(fields(session_id), skip_all)]
+#[tracing::instrument(fields(session_id = ?session_id), skip_all)]
 async fn handle_client(
     config: &'static ProxyConfig,
     cancel_map: &CancelMap,
@@ -163,11 +163,11 @@ async fn handle_client(
     // Extract credentials which we're going to use for auth.
     let creds = {
         let sni = stream.get_ref().sni_hostname();
-        let common_name = tls.and_then(|tls| tls.common_name.as_deref());
+        let common_names = tls.and_then(|tls| tls.common_names.clone());
         let result = config
             .auth_backend
             .as_ref()
-            .map(|_| auth::ClientCredentials::parse(&params, sni, common_name))
+            .map(|_| auth::ClientCredentials::parse(&params, sni, common_names))
             .transpose();
 
         async { result }.or_else(|e| stream.throw_error(e)).await?

proxy/src/proxy/tests.rs

@@ -54,9 +54,11 @@ fn generate_tls_config<'a>(
             .with_single_cert(vec![cert], key)?
             .into();
 
+        let common_names = Some([common_name.to_owned()].iter().cloned().collect());
+
         TlsConfig {
             config,
-            common_name: Some(common_name.to_string()),
+            common_names,
         }
     };
 

pyproject.toml

@@ -26,7 +26,7 @@ prometheus-client = "^0.14.1"
 pytest-timeout = "^2.1.0"
 Werkzeug = "^2.2.3"
 pytest-order = "^1.0.1"
-allure-pytest = "^2.10.0"
+allure-pytest = "^2.13.1"
 pytest-asyncio = "^0.19.0"
 toml = "^0.10.2"
 psutil = "^5.9.4"
@@ -34,6 +34,7 @@ types-psutil = "^5.9.5.4"
 types-toml = "^0.10.8"
 pytest-httpserver = "^1.0.6"
 aiohttp = "3.7.4"
+pytest-rerunfailures = "^11.1.2"
 
 [tool.poetry.group.dev.dependencies]
 black = "^23.1.0"
@@ -69,6 +70,9 @@ strict = true
 module = [
     "asyncpg.*",
     "pg8000.*",
+    "allure.*",
+    "allure_commons.*",
+    "allure_pytest.*",
 ]
 ignore_missing_imports = true
 

rust-toolchain.toml

@@ -1,5 +1,5 @@
 [toolchain]
-channel = "1.66.1"
+channel = "1.68.2"
 profile = "default"
 # The default profile includes rustc, rust-std, cargo, rust-docs, rustfmt and clippy.
 # https://rust-lang.github.io/rustup/concepts/profiles.html

safekeeper/src/bin/safekeeper.rs

@@ -5,6 +5,7 @@ use anyhow::{bail, Context, Result};
 use clap::Parser;
 use remote_storage::RemoteStorageConfig;
 use toml_edit::Document;
+use utils::signals::ShutdownSignals;
 
 use std::fs::{self, File};
 use std::io::{ErrorKind, Write};
@@ -39,7 +40,7 @@ use utils::{
     logging::{self, LogFormat},
     project_git_version,
     sentry_init::init_sentry,
-    signals, tcp_listener,
+    tcp_listener,
 };
 
 const PID_FILE_NAME: &str = "safekeeper.pid";
@@ -216,7 +217,6 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
     let timeline_collector = safekeeper::metrics::TimelineCollector::new();
     metrics::register_internal(Box::new(timeline_collector))?;
 
-    let signals = signals::install_shutdown_handlers()?;
     let mut threads = vec![];
     let (wal_backup_launcher_tx, wal_backup_launcher_rx) = mpsc::channel(100);
 
@@ -274,15 +274,12 @@ fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
 
     set_build_info_metric(GIT_VERSION);
     // TODO: put more thoughts into handling of failed threads
-    // We probably should restart them.
+    // We should catch & die if they are in trouble.
 
-    // NOTE: we still have to handle signals like SIGQUIT to prevent coredumps
-    signals.handle(|signal| {
-        // TODO: implement graceful shutdown with joining threads etc
-        info!(
-            "received {}, terminating in immediate shutdown mode",
-            signal.name()
-        );
+    // On any shutdown signal, log receival and exit. Additionally, handling
+    // SIGQUIT prevents coredump.
+    ShutdownSignals::handle(|signal| {
+        info!("received {}, terminating", signal.name());
         std::process::exit(0);
     })
 }

safekeeper/src/http/routes.rs

@@ -242,6 +242,7 @@ async fn record_safekeeper_info(mut request: Request<Body>) -> Result<Response<B
         safekeeper_connstr: sk_info.safekeeper_connstr.unwrap_or_else(|| "".to_owned()),
         backup_lsn: sk_info.backup_lsn.0,
         local_start_lsn: sk_info.local_start_lsn.0,
+        availability_zone: None,
     };
 
     let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;

safekeeper/src/metrics.rs

@@ -255,7 +255,7 @@ pub struct TimelineCollector {
     epoch_start_lsn: GenericGaugeVec<AtomicU64>,
     peer_horizon_lsn: GenericGaugeVec<AtomicU64>,
     remote_consistent_lsn: GenericGaugeVec<AtomicU64>,
-    feedback_ps_write_lsn: GenericGaugeVec<AtomicU64>,
+    ps_last_received_lsn: GenericGaugeVec<AtomicU64>,
     feedback_last_time_seconds: GenericGaugeVec<AtomicU64>,
     timeline_active: GenericGaugeVec<AtomicU64>,
     wal_backup_active: GenericGaugeVec<AtomicU64>,
@@ -339,15 +339,15 @@ impl TimelineCollector {
         .unwrap();
         descs.extend(remote_consistent_lsn.desc().into_iter().cloned());
 
-        let feedback_ps_write_lsn = GenericGaugeVec::new(
+        let ps_last_received_lsn = GenericGaugeVec::new(
             Opts::new(
-                "safekeeper_feedback_ps_write_lsn",
+                "safekeeper_ps_last_received_lsn",
                 "Last LSN received by the pageserver, acknowledged in the feedback",
             ),
             &["tenant_id", "timeline_id"],
         )
         .unwrap();
-        descs.extend(feedback_ps_write_lsn.desc().into_iter().cloned());
+        descs.extend(ps_last_received_lsn.desc().into_iter().cloned());
 
         let feedback_last_time_seconds = GenericGaugeVec::new(
             Opts::new(
@@ -458,7 +458,7 @@ impl TimelineCollector {
             epoch_start_lsn,
             peer_horizon_lsn,
             remote_consistent_lsn,
-            feedback_ps_write_lsn,
+            ps_last_received_lsn,
             feedback_last_time_seconds,
             timeline_active,
             wal_backup_active,
@@ -489,7 +489,7 @@ impl Collector for TimelineCollector {
         self.epoch_start_lsn.reset();
         self.peer_horizon_lsn.reset();
         self.remote_consistent_lsn.reset();
-        self.feedback_ps_write_lsn.reset();
+        self.ps_last_received_lsn.reset();
         self.feedback_last_time_seconds.reset();
         self.timeline_active.reset();
         self.wal_backup_active.reset();
@@ -514,11 +514,11 @@ impl Collector for TimelineCollector {
             let timeline_id = tli.ttid.timeline_id.to_string();
             let labels = &[tenant_id.as_str(), timeline_id.as_str()];
 
-            let mut most_advanced: Option<pq_proto::ReplicationFeedback> = None;
+            let mut most_advanced: Option<pq_proto::PageserverFeedback> = None;
             for replica in tli.replicas.iter() {
                 if let Some(replica_feedback) = replica.pageserver_feedback {
                     if let Some(current) = most_advanced {
-                        if current.ps_writelsn < replica_feedback.ps_writelsn {
+                        if current.last_received_lsn < replica_feedback.last_received_lsn {
                             most_advanced = Some(replica_feedback);
                         }
                     } else {
@@ -568,11 +568,10 @@ impl Collector for TimelineCollector {
                 .set(tli.wal_storage.flush_wal_seconds);
 
             if let Some(feedback) = most_advanced {
-                self.feedback_ps_write_lsn
+                self.ps_last_received_lsn
                     .with_label_values(labels)
-                    .set(feedback.ps_writelsn);
-                if let Ok(unix_time) = feedback.ps_replytime.duration_since(SystemTime::UNIX_EPOCH)
-                {
+                    .set(feedback.last_received_lsn);
+                if let Ok(unix_time) = feedback.replytime.duration_since(SystemTime::UNIX_EPOCH) {
                     self.feedback_last_time_seconds
                         .with_label_values(labels)
                         .set(unix_time.as_secs());
@@ -599,7 +598,7 @@ impl Collector for TimelineCollector {
         mfs.extend(self.epoch_start_lsn.collect());
         mfs.extend(self.peer_horizon_lsn.collect());
         mfs.extend(self.remote_consistent_lsn.collect());
-        mfs.extend(self.feedback_ps_write_lsn.collect());
+        mfs.extend(self.ps_last_received_lsn.collect());
         mfs.extend(self.feedback_last_time_seconds.collect());
         mfs.extend(self.timeline_active.collect());
         mfs.extend(self.wal_backup_active.collect());

safekeeper/src/safekeeper.rs

@@ -18,7 +18,7 @@ use crate::control_file;
 use crate::send_wal::HotStandbyFeedback;
 
 use crate::wal_storage;
-use pq_proto::{ReplicationFeedback, SystemId};
+use pq_proto::{PageserverFeedback, SystemId};
 use utils::{
     bin_ser::LeSer,
     id::{NodeId, TenantId, TenantTimelineId, TimelineId},
@@ -360,7 +360,7 @@ pub struct AppendResponse {
     // a criterion for walproposer --sync mode exit
     pub commit_lsn: Lsn,
     pub hs_feedback: HotStandbyFeedback,
-    pub pageserver_feedback: ReplicationFeedback,
+    pub pageserver_feedback: PageserverFeedback,
 }
 
 impl AppendResponse {
@@ -370,7 +370,7 @@ impl AppendResponse {
             flush_lsn: Lsn(0),
             commit_lsn: Lsn(0),
             hs_feedback: HotStandbyFeedback::empty(),
-            pageserver_feedback: ReplicationFeedback::empty(),
+            pageserver_feedback: PageserverFeedback::empty(),
         }
     }
 }
@@ -708,7 +708,7 @@ where
             commit_lsn: self.state.commit_lsn,
             // will be filled by the upper code to avoid bothering safekeeper
             hs_feedback: HotStandbyFeedback::empty(),
-            pageserver_feedback: ReplicationFeedback::empty(),
+            pageserver_feedback: PageserverFeedback::empty(),
         };
         trace!("formed AppendResponse {:?}", ar);
         ar

safekeeper/src/send_wal.rs

@@ -11,7 +11,7 @@ use postgres_backend::PostgresBackend;
 use postgres_backend::{CopyStreamHandlerEnd, PostgresBackendReader, QueryError};
 use postgres_ffi::get_current_timestamp;
 use postgres_ffi::{TimestampTz, MAX_SEND_SIZE};
-use pq_proto::{BeMessage, ReplicationFeedback, WalSndKeepAlive, XLogDataBody};
+use pq_proto::{BeMessage, PageserverFeedback, WalSndKeepAlive, XLogDataBody};
 use serde::{Deserialize, Serialize};
 use tokio::io::{AsyncRead, AsyncWrite};
 
@@ -319,11 +319,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
                 // pageserver sends this.
                 // Note: deserializing is on m[9..] because we skip the tag byte and len bytes.
                 let buf = Bytes::copy_from_slice(&msg[9..]);
-                let reply = ReplicationFeedback::parse(buf);
+                let reply = PageserverFeedback::parse(buf);
 
-                trace!("ReplicationFeedback is {:?}", reply);
-                // Only pageserver sends ReplicationFeedback, so set the flag.
-                // This replica is the source of information to resend to compute.
+                trace!("PageserverFeedback is {:?}", reply);
                 self.feedback.pageserver_feedback = Some(reply);
 
                 self.tli

safekeeper/src/timeline.rs

@@ -4,7 +4,7 @@
 use anyhow::{anyhow, bail, Result};
 use parking_lot::{Mutex, MutexGuard};
 use postgres_ffi::XLogSegNo;
-use pq_proto::ReplicationFeedback;
+use pq_proto::PageserverFeedback;
 use serde::Serialize;
 use std::cmp::{max, min};
 use std::path::PathBuf;
@@ -91,7 +91,7 @@ pub struct ReplicaState {
     /// combined hot standby feedback from all replicas
     pub hs_feedback: HotStandbyFeedback,
     /// Replication specific feedback received from pageserver, if any
-    pub pageserver_feedback: Option<ReplicationFeedback>,
+    pub pageserver_feedback: Option<PageserverFeedback>,
 }
 
 impl Default for ReplicaState {
@@ -276,7 +276,7 @@ impl SharedState {
             //
             if let Some(pageserver_feedback) = state.pageserver_feedback {
                 if let Some(acc_feedback) = acc.pageserver_feedback {
-                    if acc_feedback.ps_writelsn < pageserver_feedback.ps_writelsn {
+                    if acc_feedback.last_received_lsn < pageserver_feedback.last_received_lsn {
                         warn!("More than one pageserver is streaming WAL for the timeline. Feedback resolving is not fully supported yet.");
                         acc.pageserver_feedback = Some(pageserver_feedback);
                     }
@@ -287,12 +287,12 @@ impl SharedState {
                 // last lsn received by pageserver
                 // FIXME if multiple pageservers are streaming WAL, last_received_lsn must be tracked per pageserver.
                 // See https://github.com/neondatabase/neon/issues/1171
-                acc.last_received_lsn = Lsn::from(pageserver_feedback.ps_writelsn);
+                acc.last_received_lsn = Lsn::from(pageserver_feedback.last_received_lsn);
 
                 // When at least one pageserver has preserved data up to remote_consistent_lsn,
                 // safekeeper is free to delete it, so choose max of all pageservers.
                 acc.remote_consistent_lsn = max(
-                    Lsn::from(pageserver_feedback.ps_applylsn),
+                    Lsn::from(pageserver_feedback.remote_consistent_lsn),
                     acc.remote_consistent_lsn,
                 );
             }
@@ -337,6 +337,7 @@ impl SharedState {
             safekeeper_connstr: conf.listen_pg_addr.clone(),
             backup_lsn: self.sk.inmem.backup_lsn.0,
             local_start_lsn: self.sk.state.local_start_lsn.0,
+            availability_zone: conf.availability_zone.clone(),
         }
     }
 }
@@ -584,7 +585,7 @@ impl Timeline {
             let replica_state = shared_state.replicas[replica_id].unwrap();
             let reported_remote_consistent_lsn = replica_state
                 .pageserver_feedback
-                .map(|f| Lsn(f.ps_applylsn))
+                .map(|f| Lsn(f.remote_consistent_lsn))
                 .unwrap_or(Lsn::INVALID);
             let stop = shared_state.sk.inmem.commit_lsn == Lsn(0) || // no data at all yet
             (reported_remote_consistent_lsn!= Lsn::MAX && // Lsn::MAX means that we don't know the latest LSN yet.

scripts/flaky_tests.py

@@ -0,0 +1,87 @@
+#! /usr/bin/env python3
+
+import argparse
+import json
+import logging
+from collections import defaultdict
+from typing import DefaultDict, Dict
+
+import psycopg2
+import psycopg2.extras
+
+# We call the test "flaky" if it failed at least once on the main branch in the last N=10 days.
+FLAKY_TESTS_QUERY = """
+    SELECT
+        DISTINCT parent_suite, suite, test
+    FROM
+        (
+            SELECT
+                revision,
+                jsonb_array_elements(data -> 'children') -> 'name' as parent_suite,
+                jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'name' as suite,
+                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'name' as test,
+                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'status' as status,
+                to_timestamp((jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'time' -> 'start')::bigint / 1000)::date as timestamp
+            FROM
+                regress_test_results
+            WHERE
+                reference = 'refs/heads/main'
+        ) data
+    WHERE
+        timestamp > CURRENT_DATE - INTERVAL '%s' day
+        AND status::text IN ('"failed"', '"broken"')
+    ;
+"""
+
+
+def main(args: argparse.Namespace):
+    connstr = args.connstr
+    interval_days = args.days
+    output = args.output
+
+    res: DefaultDict[str, DefaultDict[str, Dict[str, bool]]]
+    res = defaultdict(lambda: defaultdict(dict))
+
+    logging.info("connecting to the database...")
+    with psycopg2.connect(connstr, connect_timeout=10) as conn:
+        with conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
+            logging.info("fetching flaky tests...")
+            cur.execute(FLAKY_TESTS_QUERY, (interval_days,))
+            rows = cur.fetchall()
+
+    for row in rows:
+        logging.info(f"\t{row['parent_suite'].replace('.', '/')}/{row['suite']}.py::{row['test']}")
+        res[row["parent_suite"]][row["suite"]][row["test"]] = True
+
+    logging.info(f"saving results to {output.name}")
+    json.dump(res, output, indent=2)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Detect flaky tests in the last N days")
+    parser.add_argument(
+        "--output",
+        type=argparse.FileType("w"),
+        default="flaky.json",
+        help="path to output json file (default: flaky.json)",
+    )
+    parser.add_argument(
+        "--days",
+        required=False,
+        default=10,
+        type=int,
+        help="how many days to look back for flaky tests (default: 10)",
+    )
+    parser.add_argument(
+        "connstr",
+        help="connection string to the test results database",
+    )
+    args = parser.parse_args()
+
+    level = logging.INFO
+    logging.basicConfig(
+        format="%(message)s",
+        level=level,
+    )
+
+    main(args)

scripts/pr-comment-test-report.js

@@ -0,0 +1,125 @@
+//
+// The script parses Allure reports and posts a comment with a summary of the test results to the PR.
+// It accepts an array of items and creates a comment with a summary for each one (for "release" and "debug", together or separately if any of them failed to be generated).
+//
+// The comment is updated on each run with the latest results.
+//
+// It is designed to be used with actions/github-script from GitHub Workflows:
+// - uses: actions/github-script@v6
+//   with:
+//     script: |
+//       const script = require("./scripts/pr-comment-test-report.js")
+//       await script({
+//         github,
+//         context,
+//         fetch,
+//         reports: [{...}, ...], // each report is expected to have "buildType", "reportUrl", and "jsonUrl" properties
+//       })
+//
+
+module.exports = async ({ github, context, fetch, reports }) => {
+    // Marker to find the comment in the subsequent runs
+    const startMarker = `<!--AUTOMATIC COMMENT START #${context.payload.number}-->`
+    // GitHub bot id taken from (https://api.github.com/users/github-actions[bot])
+    const githubActionsBotId = 41898282
+    // The latest commit in the PR URL
+    const commitUrl = `${context.serverUrl}/${context.repo.owner}/${context.repo.repo}/pull/${context.payload.number}/commits/${context.payload.pull_request.head.sha}`
+    // Commend body itself
+    let commentBody = `${startMarker}\n### Test results for ${commitUrl}:\n___\n`
+
+    // Common parameters for GitHub API requests
+    const ownerRepoParams = {
+        owner: context.repo.owner,
+        repo: context.repo.repo,
+    }
+
+    for (const report of reports) {
+        const {buildType, reportUrl, jsonUrl} = report
+
+        if (!reportUrl || !jsonUrl) {
+            console.warn(`"reportUrl" or "jsonUrl" aren't set for ${buildType} build`)
+            continue
+        }
+
+        const suites = await (await fetch(jsonUrl)).json()
+
+        // Allure distinguishes "failed" (with an assertion error) and "broken" (with any other error) tests.
+        // For this report it's ok to treat them in the same way (as failed).
+        failedTests = []
+        passedTests = []
+        skippedTests = []
+
+        retriedTests = []
+        retriedStatusChangedTests = []
+
+        for (const parentSuite of suites.children) {
+            for (const suite of parentSuite.children) {
+                for (const test of suite.children) {
+                    pytestName = `${parentSuite.name.replace(".", "/")}/${suite.name}.py::${test.name}`
+                    test.pytestName = pytestName
+
+                    if (test.status === "passed") {
+                        passedTests.push(test);
+                    } else if (test.status === "failed" || test.status === "broken") {
+                        failedTests.push(test);
+                    } else if (test.status === "skipped") {
+                        skippedTests.push(test);
+                    }
+
+                    if (test.retriesCount > 0) {
+                        retriedTests.push(test);
+
+                        if (test.retriedStatusChangedTests) {
+                            retriedStatusChangedTests.push(test);
+                        }
+                    }
+                }
+            }
+        }
+
+        const totalTestsCount = failedTests.length + passedTests.length + skippedTests.length
+        commentBody += `#### ${buildType} build: ${totalTestsCount} tests run: ${passedTests.length} passed, ${failedTests.length} failed, ${skippedTests.length} ([full report](${reportUrl}))\n`
+        if (failedTests.length > 0) {
+            commentBody += `Failed tests:\n`
+            for (const test of failedTests) {
+                const allureLink = `${reportUrl}#suites/${test.parentUid}/${test.uid}`
+
+                commentBody += `- [\`${test.pytestName}\`](${allureLink})`
+                if (test.retriesCount > 0) {
+                    commentBody += ` (ran [${test.retriesCount + 1} times](${allureLink}/retries))`
+                }
+                commentBody += "\n"
+            }
+            commentBody += "\n"
+        }
+        if (retriedStatusChangedTests > 0) {
+            commentBody += `Flaky tests:\n`
+            for (const test of retriedStatusChangedTests) {
+                const status = test.status === "passed" ? ":white_check_mark:" : ":x:"
+                commentBody += `- ${status} [\`${test.pytestName}\`](${reportUrl}#suites/${test.parentUid}/${test.uid}/retries)\n`
+            }
+            commentBody += "\n"
+        }
+        commentBody += "___\n"
+    }
+
+    const { data: comments } = await github.rest.issues.listComments({
+        issue_number: context.payload.number,
+        ...ownerRepoParams,
+    })
+
+    const comment = comments.find(comment => comment.user.id === githubActionsBotId && comment.body.startsWith(startMarker))
+    if (comment) {
+        await github.rest.issues.updateComment({
+            comment_id: comment.id,
+            body: commentBody,
+            ...ownerRepoParams,
+        })
+    } else {
+        await github.rest.issues.createComment({
+            issue_number: context.payload.number,
+            body: commentBody,
+            ...ownerRepoParams,
+        })
+    }
+}

scripts/sk_collect_dumps/.gitignore

@@ -0,0 +1,2 @@
+result
+*.json

scripts/sk_collect_dumps/readme.md

@@ -0,0 +1,25 @@
+# Collect /v1/debug_dump from all safekeeper nodes
+
+1. Run ansible playbooks to collect .json dumps from all safekeepers and store them in `./result` directory.
+2. Run `DB_CONNSTR=... ./upload.sh prod_feb30` to upload dumps to `prod_feb30` table in specified postgres database.
+
+## How to use ansible (staging)
+
+```
+AWS_DEFAULT_PROFILE=dev ansible-playbook -i ../../.github/ansible/staging.us-east-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+
+AWS_DEFAULT_PROFILE=dev ansible-playbook -i ../../.github/ansible/staging.eu-west-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+```
+
+## How to use ansible (prod)
+
+```
+AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.us-west-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+
+AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.us-east-2.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+
+AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.eu-central-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+
+AWS_DEFAULT_PROFILE=prod ansible-playbook -i ../../.github/ansible/prod.ap-southeast-1.hosts.yaml -e @../../.github/ansible/ssm_config remote.yaml
+```
+

scripts/sk_collect_dumps/remote.yaml

@@ -0,0 +1,18 @@
+- name: Fetch state dumps from safekeepers
+  hosts: safekeepers
+  gather_facts: False
+  remote_user: "{{ remote_user }}"
+    
+  tasks:
+    - name: Download file
+      get_url:
+        url: "http://{{ inventory_hostname }}:7676/v1/debug_dump?dump_all=true&dump_disk_content=false"
+        dest: "/tmp/{{ inventory_hostname }}.json"
+
+    - name: Fetch file from remote hosts
+      fetch:
+        src: "/tmp/{{ inventory_hostname }}.json"
+        dest: "./result/{{ inventory_hostname }}.json"
+        flat: yes
+        fail_on_missing: no
+

scripts/sk_collect_dumps/upload.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+
+if [ -z "$DB_CONNSTR" ]; then
+    echo "DB_CONNSTR is not set"
+    exit 1
+fi
+
+# Create a temporary table for JSON data
+psql $DB_CONNSTR -c 'DROP TABLE IF EXISTS tmp_json'
+psql $DB_CONNSTR -c 'CREATE TABLE tmp_json (data jsonb)'
+
+for file in ./result/*.json; do
+    echo "$file"
+    SK_ID=$(jq '.config.id' $file)
+    echo "SK_ID: $SK_ID"
+    jq -c ".timelines[] |  . + {\"sk_id\": $SK_ID}" $file | psql $DB_CONNSTR -c "\\COPY tmp_json (data) FROM STDIN"
+done
+
+TABLE_NAME=$1
+
+if [ -z "$TABLE_NAME" ]; then
+    echo "TABLE_NAME is not set, skipping conversion to table with typed columns"
+    echo "Usage: ./upload.sh TABLE_NAME"
+    exit 0
+fi
+
+psql $DB_CONNSTR <<EOF
+CREATE TABLE $TABLE_NAME AS
+SELECT
+  (data->>'sk_id')::bigint AS sk_id,
+  (data->>'tenant_id') AS tenant_id,
+  (data->>'timeline_id') AS timeline_id,
+  (data->'memory'->>'active')::bool AS active,
+  (data->'memory'->>'flush_lsn')::bigint AS flush_lsn,
+  (data->'memory'->'mem_state'->>'backup_lsn')::bigint AS backup_lsn,
+  (data->'memory'->'mem_state'->>'commit_lsn')::bigint AS commit_lsn,
+  (data->'memory'->'mem_state'->>'peer_horizon_lsn')::bigint AS peer_horizon_lsn,
+  (data->'memory'->'mem_state'->>'remote_consistent_lsn')::bigint AS remote_consistent_lsn,
+  (data->'memory'->>'write_lsn')::bigint AS write_lsn,
+  (data->'memory'->>'num_computes')::bigint AS num_computes,
+  (data->'memory'->>'epoch_start_lsn')::bigint AS epoch_start_lsn,
+  (data->'memory'->>'last_removed_segno')::bigint AS last_removed_segno,
+  (data->'memory'->>'is_cancelled')::bool AS is_cancelled,
+  (data->'control_file'->>'backup_lsn')::bigint AS disk_backup_lsn,
+  (data->'control_file'->>'commit_lsn')::bigint AS disk_commit_lsn,
+  (data->'control_file'->'acceptor_state'->>'term')::bigint AS disk_term,
+  (data->'control_file'->>'local_start_lsn')::bigint AS local_start_lsn,
+  (data->'control_file'->>'peer_horizon_lsn')::bigint AS disk_peer_horizon_lsn,
+  (data->'control_file'->>'timeline_start_lsn')::bigint AS timeline_start_lsn,
+  (data->'control_file'->>'remote_consistent_lsn')::bigint AS disk_remote_consistent_lsn
+FROM tmp_json
+EOF

storage_broker/benches/rps.rs

@@ -133,6 +133,7 @@ async fn publish(client: Option<BrokerClientChannel>, n_keys: u64) {
                 peer_horizon_lsn: 5,
                 safekeeper_connstr: "zenith-1-sk-1.local:7676".to_owned(),
                 local_start_lsn: 0,
+                availability_zone: None,
             };
             counter += 1;
             yield info;

storage_broker/proto/broker.proto

@@ -36,9 +36,11 @@ message SafekeeperTimelineInfo {
     uint64 local_start_lsn = 9;
     // A connection string to use for WAL receiving.
     string safekeeper_connstr = 10;
+    // Availability zone of a safekeeper.
+    optional string availability_zone = 11;
 }
 
 message TenantTimelineId {
     bytes tenant_id = 1;
     bytes timeline_id = 2;
-}
+}

storage_broker/src/bin/storage_broker.rs

@@ -33,6 +33,7 @@ use tonic::transport::server::Connected;
 use tonic::Code;
 use tonic::{Request, Response, Status};
 use tracing::*;
+use utils::signals::ShutdownSignals;
 
 use metrics::{Encoder, TextEncoder};
 use storage_broker::metrics::{NUM_PUBS, NUM_SUBS_ALL, NUM_SUBS_TIMELINE};
@@ -437,6 +438,14 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
     info!("version: {GIT_VERSION}");
     ::metrics::set_build_info_metric(GIT_VERSION);
 
+    // On any shutdown signal, log receival and exit.
+    std::thread::spawn(move || {
+        ShutdownSignals::handle(|signal| {
+            info!("received {}, terminating", signal.name());
+            std::process::exit(0);
+        })
+    });
+
     let registry = Registry {
         shared_state: Arc::new(RwLock::new(SharedState::new(args.all_keys_chan_size))),
         timeline_chan_size: args.timeline_chan_size,
@@ -516,6 +525,7 @@ mod tests {
             peer_horizon_lsn: 5,
             safekeeper_connstr: "neon-1-sk-1.local:7676".to_owned(),
             local_start_lsn: 0,
+            availability_zone: None,
         }
     }
 

test_runner/conftest.py

@@ -4,4 +4,5 @@ pytest_plugins = (
     "fixtures.pg_stats",
     "fixtures.compare_fixtures",
     "fixtures.slow",
+    "fixtures.flaky",
 )

test_runner/fixtures/flaky.py

@@ -0,0 +1,58 @@
+import json
+from pathlib import Path
+from typing import List
+
+import pytest
+from _pytest.config import Config
+from _pytest.config.argparsing import Parser
+from allure_commons.types import LabelType
+from allure_pytest.utils import allure_name, allure_suite_labels
+
+from fixtures.log_helper import log
+
+"""
+The plugin reruns flaky tests.
+It uses `pytest.mark.flaky` provided by `pytest-rerunfailures` plugin and flaky tests detected by `scripts/flaky_tests.py`
+
+Note: the logic of getting flaky tests is extracted to a separate script to avoid running it for each of N xdist workers
+"""
+
+
+def pytest_addoption(parser: Parser):
+    parser.addoption(
+        "--flaky-tests-json",
+        action="store",
+        type=Path,
+        help="Path to json file with flaky tests generated by scripts/flaky_tests.py",
+    )
+
+
+def pytest_collection_modifyitems(config: Config, items: List[pytest.Item]):
+    if not config.getoption("--flaky-tests-json"):
+        return
+
+    # Any error with getting flaky tests aren't critical, so just do not rerun any tests
+    flaky_json = config.getoption("--flaky-tests-json")
+    if not flaky_json.exists():
+        return
+
+    content = flaky_json.read_text()
+    try:
+        flaky_tests = json.loads(content)
+    except ValueError:
+        log.error(f"Can't parse {content} as json")
+        return
+
+    for item in items:
+        # Use the same logic for constructing test name as Allure does (we store allure-provided data in DB)
+        # Ref https://github.com/allure-framework/allure-python/blob/2.13.1/allure-pytest/src/listener.py#L98-L100
+        allure_labels = dict(allure_suite_labels(item))
+        parent_suite = str(allure_labels.get(LabelType.PARENT_SUITE))
+        suite = str(allure_labels.get(LabelType.SUITE))
+        params = item.callspec.params if hasattr(item, "callspec") else {}
+        name = allure_name(item, params)
+
+        if flaky_tests.get(parent_suite, {}).get(suite, {}).get(name, False):
+            # Rerun 3 times = 1 original run + 2 reruns
+            log.info(f"Marking {item.nodeid} as flaky. It will be rerun up to 3 times")
+            item.add_marker(pytest.mark.flaky(reruns=2))

test_runner/fixtures/neon_fixtures.py

@@ -1220,6 +1220,28 @@ class PageserverHttpClient(requests.Session):
         self.verbose_error(res)
         return TenantConfig.from_json(res.json())
 
+    def set_tenant_config(self, tenant_id: TenantId, config: dict[str, Any]):
+        assert "tenant_id" not in config.keys()
+        res = self.put(
+            f"http://localhost:{self.port}/v1/tenant/config",
+            json={**config, "tenant_id": str(tenant_id)},
+        )
+        self.verbose_error(res)
+
+    def patch_tenant_config_client_side(
+        self,
+        tenant_id: TenantId,
+        inserts: Optional[Dict[str, Any]] = None,
+        removes: Optional[List[str]] = None,
+    ):
+        current = self.tenant_config(tenant_id).tenant_specific_overrides
+        if inserts is not None:
+            current.update(inserts)
+        if removes is not None:
+            for key in removes:
+                del current[key]
+        self.set_tenant_config(tenant_id, current)
+
     def tenant_size(self, tenant_id: TenantId) -> int:
         return self.tenant_size_and_modelinputs(tenant_id)[0]
 
@@ -1536,6 +1558,18 @@ class PageserverHttpClient(requests.Session):
         for layer in info.historic_layers:
             self.evict_layer(tenant_id, timeline_id, layer.layer_file_name)
 
+    def disk_usage_eviction_run(self, request: dict[str, Any]):
+        res = self.put(
+            f"http://localhost:{self.port}/v1/disk_usage_eviction/run",
+            json=request,
+        )
+        self.verbose_error(res)
+        return res.json()
+
+    def tenant_break(self, tenant_id: TenantId):
+        res = self.put(f"http://localhost:{self.port}/v1/tenant/{tenant_id}/break")
+        self.verbose_error(res)
+
 
 @dataclass
 class TenantConfig:

test_runner/fixtures/utils.py

@@ -7,7 +7,7 @@ import time
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Tuple, TypeVar
 
-import allure  # type: ignore
+import allure
 from psycopg2.extensions import cursor
 
 from fixtures.log_helper import log