Polish API handler and refresh OpenAPI spec

With this commit one can start compute with something like
```shell
cargo run --bin compute_ctl -- -i no-compute \
 -p http://localhost:9095 \
 -D compute_pgdata \
 -C "postgresql://cloud_admin@127.0.0.1:5434/postgres" \
 -b ./pg_install/v15/bin/postgres
```
and it will hang waiting for spec.

Then send one spec
```shell
curl -d "$(cat ./compute-spec.json)" http://localhost:3080/spec
```
Postgres will be started and configured.

Then reconfigure it with
```shell
curl -d "$(cat ./compute-spec-new.json)" http://localhost:3080/spec
```

Most of safeguards and comments are added. Some polishing especially
around HTTP API is still needed.

.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/deploy.yaml

@@ -91,6 +91,15 @@
       tags:
       - pageserver
 
+    # used in `pageserver.service` template
+    - name: learn current availability_zone
+      shell:
+        cmd: "curl -s http://169.254.169.254/latest/meta-data/placement/availability-zone"
+      register: ec2_availability_zone
+
+    - set_fact: 
+        ec2_availability_zone={{ ec2_availability_zone.stdout }}
+
     - name: upload systemd service definition
       ansible.builtin.template:
         src: systemd/pageserver.service
@@ -153,6 +162,15 @@
       tags:
       - safekeeper
 
+    # used in `safekeeper.service` template
+    - name: learn current availability_zone
+      shell:
+        cmd: "curl -s http://169.254.169.254/latest/meta-data/placement/availability-zone"
+      register: ec2_availability_zone
+
+    - set_fact: 
+        ec2_availability_zone={{ ec2_availability_zone.stdout }}
+
     # in the future safekeepers should discover pageservers byself
     # but currently use first pageserver that was discovered
     - name: set first pageserver var for safekeepers

.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/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/ansible/systemd/pageserver.service

@@ -6,7 +6,7 @@ After=network.target auditd.service
 Type=simple
 User=pageserver
 Environment=RUST_BACKTRACE=1 NEON_REPO_DIR=/storage/pageserver LD_LIBRARY_PATH=/usr/local/v14/lib SENTRY_DSN={{ SENTRY_URL_PAGESERVER }} SENTRY_ENVIRONMENT={{ sentry_environment }}
-ExecStart=/usr/local/bin/pageserver -c "pg_distrib_dir='/usr/local'" -c "listen_pg_addr='0.0.0.0:6400'" -c "listen_http_addr='0.0.0.0:9898'" -c "broker_endpoint='{{ broker_endpoint }}'" -D /storage/pageserver/data
+ExecStart=/usr/local/bin/pageserver -c "pg_distrib_dir='/usr/local'" -c "listen_pg_addr='0.0.0.0:6400'" -c "listen_http_addr='0.0.0.0:9898'" -c "broker_endpoint='{{ broker_endpoint }}'" -c "availability_zone='{{ ec2_availability_zone }}'" -D /storage/pageserver/data
 ExecReload=/bin/kill -HUP $MAINPID
 KillMode=mixed
 KillSignal=SIGINT

.github/ansible/systemd/safekeeper.service

@@ -6,7 +6,7 @@ After=network.target auditd.service
 Type=simple
 User=safekeeper
 Environment=RUST_BACKTRACE=1 NEON_REPO_DIR=/storage/safekeeper/data LD_LIBRARY_PATH=/usr/local/v14/lib SENTRY_DSN={{ SENTRY_URL_SAFEKEEPER }} SENTRY_ENVIRONMENT={{ sentry_environment }}
-ExecStart=/usr/local/bin/safekeeper -l {{ inventory_hostname }}{{ hostname_suffix }}:6500 --listen-http {{ inventory_hostname }}{{ hostname_suffix }}:7676 -D /storage/safekeeper/data --broker-endpoint={{ broker_endpoint }} --remote-storage='{bucket_name="{{bucket_name}}", bucket_region="{{bucket_region}}", prefix_in_bucket="{{ safekeeper_s3_prefix }}"}'
+ExecStart=/usr/local/bin/safekeeper -l {{ inventory_hostname }}{{ hostname_suffix }}:6500 --listen-http {{ inventory_hostname }}{{ hostname_suffix }}:7676 -D /storage/safekeeper/data --broker-endpoint={{ broker_endpoint }} --remote-storage='{bucket_name="{{bucket_name}}", bucket_region="{{bucket_region}}", prefix_in_bucket="{{ safekeeper_s3_prefix }}"}' --availability-zone={{ ec2_availability_zone }}
 ExecReload=/bin/kill -HUP $MAINPID
 KillMode=mixed
 KillSignal=SIGINT

.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

@@ -74,15 +74,12 @@ jobs:
       - name: Install Python deps
         run: ./scripts/pysync
 
-      - name: Run isort to ensure code format
-        run: poetry run isort --diff --check .
+      - name: Run ruff to ensure code format
+        run: poetry run ruff .
 
       - name: Run black to ensure code format
         run: poetry run black --diff --check .
 
-      - name: Run flake8 to ensure code format
-        run: poetry run flake8 .
-
       - name: Run mypy to check types
         run: poetry run mypy .
 
@@ -187,10 +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
@@ -243,11 +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: |
@@ -271,7 +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 $test_exe_paths; do
@@ -331,6 +335,9 @@ jobs:
           real_s3_region: us-west-2
           real_s3_access_key_id: "${{ 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'
@@ -367,42 +374,88 @@ 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: ${{ !cancelled() }}
+        id: create-allure-report-release
+        uses: ./.github/actions/allure-report
+        with:
+          action: generate
+          build_type: release
+
+      - uses: actions/github-script@v6
+        if: >
+          !cancelled() &&
+          github.event_name == 'pull_request' && (
+            steps.create-allure-report-debug.outputs.report-url ||
+            steps.create-allure-report-release.outputs.report-url
+          )
+        with:
+          script: |
+            const reports = [{
+              buildType: "debug",
+              reportUrl: "${{ 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 ]
@@ -555,7 +608,7 @@ jobs:
   neon-image-depot:
     # For testing this will run side-by-side for a few merges.
     # This action is not really optimized yet, but gets the job done
-    runs-on: [ self-hosted, gen3, small ]
+    runs-on: [ self-hosted, gen3, large ]
     needs: [ tag ]
     container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
     permissions:
@@ -894,6 +947,16 @@ 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.vm-compute-node

@@ -2,26 +2,69 @@
 
 ARG SRC_IMAGE
 ARG VM_INFORMANT_VERSION=v0.1.14
+# on libcgroup update, make sure to check bootstrap.sh for changes
+ARG LIBCGROUP_VERSION=v2.0.3
 
-# Pull VM informant and set up inittab
+# Pull VM informant, to copy from later
 FROM neondatabase/vm-informant:$VM_INFORMANT_VERSION as informant
 
+# Build cgroup-tools
+#
+# At time of writing (2023-03-14), debian bullseye has a version of cgroup-tools (technically
+# libcgroup) that doesn't support cgroup v2 (version 0.41-11). Unfortunately, the vm-informant
+# requires cgroup v2, so we'll build cgroup-tools ourselves.
+FROM debian:bullseye-slim as libcgroup-builder
+ARG LIBCGROUP_VERSION
+
+RUN set -exu \
+	&& apt update \
+	&& apt install --no-install-recommends -y \
+		git \
+		ca-certificates \
+		automake \
+		cmake \
+		make \
+		gcc \
+		byacc \
+		flex \
+		libtool \
+		libpam0g-dev \
+	&& git clone --depth 1 -b $LIBCGROUP_VERSION https://github.com/libcgroup/libcgroup \
+	&& INSTALL_DIR="/libcgroup-install" \
+	&& mkdir -p "$INSTALL_DIR/bin" "$INSTALL_DIR/include" \
+	&& cd libcgroup \
+	# extracted from bootstrap.sh, with modified flags:
+	&& (test -d m4 || mkdir m4) \
+	&& autoreconf -fi \
+	&& rm -rf autom4te.cache \
+	&& CFLAGS="-O3" ./configure --prefix="$INSTALL_DIR" --sysconfdir=/etc --localstatedir=/var --enable-opaque-hierarchy="name=systemd" \
+	# actually build the thing...
+	&& make install
+
+# Combine, starting from non-VM compute node image.
+FROM $SRC_IMAGE as base
+
+# Temporarily set user back to root so we can run adduser, set inittab
+USER root
+RUN adduser vm-informant --disabled-password --no-create-home
+
 RUN set -e \
 	&& rm -f /etc/inittab \
 	&& touch /etc/inittab
 
 RUN set -e \
+	&& echo "::sysinit:cgconfigparser -l /etc/cgconfig.conf -s 1664" >> /etc/inittab \
 	&& CONNSTR="dbname=neondb user=cloud_admin sslmode=disable" \
-	&& ARGS="--auto-restart --pgconnstr=\"$CONNSTR\"" \
+	&& ARGS="--auto-restart --cgroup=neon-postgres --pgconnstr=\"$CONNSTR\"" \
 	&& echo "::respawn:su vm-informant -c '/usr/local/bin/vm-informant $ARGS'" >> /etc/inittab
 
-# Combine, starting from non-VM compute node image.
-FROM $SRC_IMAGE as base
-
-# Temporarily set user back to root so we can run adduser
-USER root
-RUN adduser vm-informant --disabled-password --no-create-home
 USER postgres
 
-COPY --from=informant /etc/inittab /etc/inittab
+ADD vm-cgconfig.conf /etc/cgconfig.conf
 COPY --from=informant /usr/bin/vm-informant /usr/local/bin/vm-informant
+
+COPY --from=libcgroup-builder /libcgroup-install/bin/* /usr/bin/
+COPY --from=libcgroup-builder /libcgroup-install/lib/* /usr/lib/
+COPY --from=libcgroup-builder /libcgroup-install/sbin/* /usr/sbin/
+
+ENTRYPOINT ["/usr/sbin/cgexec", "-g", "*:neon-postgres", "/usr/local/bin/compute_ctl"]

README.md

@@ -40,17 +40,22 @@ 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
 curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
 ```
 
-#### Installing dependencies on OSX (12.3.1)
+#### Installing dependencies on macOS (12.3.1)
 1. Install XCode and dependencies
 ```
 xcode-select --install
 brew install protobuf openssl flex bison
+
+# add openssl to PATH, required for ed25519 keys generation in neon_local
+echo 'export PATH="$(brew --prefix openssl)/bin:$PATH"' >> ~/.zshrc
 ```
 
 2. [Install Rust](https://www.rust-lang.org/tools/install)

compute_tools/src/bin/compute_ctl.rs

@@ -34,22 +34,23 @@ use std::fs::File;
 use std::panic;
 use std::path::Path;
 use std::process::exit;
-use std::sync::{Arc, RwLock};
+use std::sync::{Arc, Condvar, Mutex};
 use std::{thread, time::Duration};
 
 use anyhow::{Context, Result};
 use chrono::Utc;
 use clap::Arg;
 use tracing::{error, info};
+use url::Url;
 
 use compute_tools::compute::{ComputeMetrics, ComputeNode, ComputeState, ComputeStatus};
+use compute_tools::configurator::launch_configurator;
 use compute_tools::http::api::launch_http_server;
 use compute_tools::logger::*;
 use compute_tools::monitor::launch_monitor;
 use compute_tools::params::*;
 use compute_tools::pg_helpers::*;
 use compute_tools::spec::*;
-use url::Url;
 
 fn main() -> Result<()> {
     init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
@@ -62,7 +63,7 @@ fn main() -> Result<()> {
     let connstr = matches
         .get_one::<String>("connstr")
         .expect("Postgres connection string is required");
-    let spec = matches.get_one::<String>("spec");
+    let spec_json = matches.get_one::<String>("spec");
     let spec_path = matches.get_one::<String>("spec-path");
 
     let compute_id = matches.get_one::<String>("compute-id");
@@ -71,40 +72,110 @@ fn main() -> Result<()> {
     // Try to use just 'postgres' if no path is provided
     let pgbin = matches.get_one::<String>("pgbin").unwrap();
 
-    let spec: ComputeSpec = match spec {
+    let mut spec = Default::default();
+    let mut spec_set = false;
+    let mut live_config_allowed = false;
+    match spec_json {
         // First, try to get cluster spec from the cli argument
-        Some(json) => serde_json::from_str(json)?,
+        Some(json) => {
+            spec = serde_json::from_str(json)?;
+            spec_set = true;
+        }
         None => {
             // Second, try to read it from the file if path is provided
             if let Some(sp) = spec_path {
                 let path = Path::new(sp);
                 let file = File::open(path)?;
-                serde_json::from_reader(file)?
+                spec = serde_json::from_reader(file)?;
+                spec_set = true;
             } else if let Some(id) = compute_id {
                 if let Some(cp_base) = control_plane_uri {
-                    let cp_uri = format!("{cp_base}/management/api/v1/{id}/spec");
-                    let jwt: String = match std::env::var("NEON_CONSOLE_JWT") {
-                        Ok(v) => v,
-                        Err(_) => "".to_string(),
-                    };
-
-                    reqwest::blocking::Client::new()
-                        .get(cp_uri)
-                        .header("Authorization", jwt)
-                        .send()?
-                        .json()?
+                    live_config_allowed = true;
+                    if let Ok(s) = get_spec_from_control_plane(cp_base, id) {
+                        spec = s;
+                        spec_set = true;
+                    }
                 } else {
-                    panic!(
-                        "must specify --control-plane-uri \"{:#?}\" and --compute-id \"{:#?}\"",
-                        control_plane_uri, compute_id
-                    );
+                    panic!("must specify both --control-plane-uri and --compute-id or none");
                 }
             } else {
-                panic!("compute spec should be provided via --spec or --spec-path argument");
+                panic!(
+                    "compute spec should be provided by one of the following ways: \
+                    --spec OR --spec-path OR --control-plane-uri and --compute-id"
+                );
             }
         }
     };
 
+    let mut new_state = ComputeState::new();
+    if spec_set {
+        new_state.spec = spec;
+    }
+    // Volatile compute state under mutex and condition variable to notify everyone
+    // who is interested in the state changes.
+    let compute_state = (Mutex::new(new_state), Condvar::new());
+    let compute_node = ComputeNode {
+        start_time: Utc::now(),
+        connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
+        pgdata: pgdata.to_string(),
+        pgbin: pgbin.to_string(),
+        live_config_allowed,
+        metrics: ComputeMetrics::default(),
+        state: compute_state,
+    };
+    let compute = Arc::new(compute_node);
+
+    // Launch http service first, so we were able to serve control-plane
+    // requests, while configuration is still in progress.
+    let _http_handle = launch_http_server(&compute).expect("cannot launch http endpoint thread");
+
+    if !spec_set {
+        // No spec provided, hang waiting for it.
+        info!("no compute spec provided, waiting");
+        let (state, state_changed) = &compute.state;
+        let mut state = state.lock().unwrap();
+        while state.status != ComputeStatus::ConfigurationPending {
+            state = state_changed.wait(state).unwrap();
+
+            if state.status == ComputeStatus::ConfigurationPending {
+                info!("got spec, continue configuration");
+                // Spec is already set by the http server handler.
+                break;
+            }
+        }
+    }
+
+    // We got all we need, fill in the state.
+    let (state, _) = &compute.state;
+    let mut state = state.lock().unwrap();
+    let pageserver_connstr = state
+        .spec
+        .cluster
+        .settings
+        .find("neon.pageserver_connstring")
+        .expect("pageserver connstr should be provided");
+    let storage_auth_token = state.spec.storage_auth_token.clone();
+    let tenant = state
+        .spec
+        .cluster
+        .settings
+        .find("neon.tenant_id")
+        .expect("tenant id should be provided");
+    let timeline = state
+        .spec
+        .cluster
+        .settings
+        .find("neon.timeline_id")
+        .expect("tenant id should be provided");
+    let startup_tracing_context = state.spec.startup_tracing_context.clone();
+
+    state.pageserver_connstr = pageserver_connstr;
+    state.storage_auth_token = storage_auth_token;
+    state.tenant = tenant;
+    state.timeline = timeline;
+    state.status = ComputeStatus::Init;
+    drop(state);
+
     // Extract OpenTelemetry context for the startup actions from the spec, and
     // attach it to the current tracing context.
     //
@@ -120,7 +191,7 @@ fn main() -> Result<()> {
     // postgres is configured and up-and-running, we exit this span. Any other
     // actions that are performed on incoming HTTP requests, for example, are
     // performed in separate spans.
-    let startup_context_guard = if let Some(ref carrier) = spec.startup_tracing_context {
+    let startup_context_guard = if let Some(ref carrier) = startup_tracing_context {
         use opentelemetry::propagation::TextMapPropagator;
         use opentelemetry::sdk::propagation::TraceContextPropagator;
         Some(TraceContextPropagator::new().extract(carrier).attach())
@@ -128,40 +199,10 @@ fn main() -> Result<()> {
         None
     };
 
-    let pageserver_connstr = spec
-        .cluster
-        .settings
-        .find("neon.pageserver_connstring")
-        .expect("pageserver connstr should be provided");
-    let tenant = spec
-        .cluster
-        .settings
-        .find("neon.tenant_id")
-        .expect("tenant id should be provided");
-    let timeline = spec
-        .cluster
-        .settings
-        .find("neon.timeline_id")
-        .expect("tenant id should be provided");
-
-    let compute_state = ComputeNode {
-        start_time: Utc::now(),
-        connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
-        pgdata: pgdata.to_string(),
-        pgbin: pgbin.to_string(),
-        spec,
-        tenant,
-        timeline,
-        pageserver_connstr,
-        metrics: ComputeMetrics::default(),
-        state: RwLock::new(ComputeState::new()),
-    };
-    let compute = Arc::new(compute_state);
-
-    // Launch service threads first, so we were able to serve availability
-    // requests, while configuration is still in progress.
-    let _http_handle = launch_http_server(&compute).expect("cannot launch http endpoint thread");
+    // Launch remaining service threads
     let _monitor_handle = launch_monitor(&compute).expect("cannot launch compute monitor thread");
+    let _configurator_handle =
+        launch_configurator(&compute).expect("cannot launch configurator thread");
 
     // Start Postgres
     let mut delay_exit = false;
@@ -170,7 +211,8 @@ fn main() -> Result<()> {
         Ok(pg) => Some(pg),
         Err(err) => {
             error!("could not start the compute node: {:?}", err);
-            let mut state = compute.state.write().unwrap();
+            let (state, _) = &compute.state;
+            let mut state = state.lock().unwrap();
             state.error = Some(format!("{:?}", err));
             state.status = ComputeStatus::Failed;
             drop(state);
@@ -201,13 +243,14 @@ 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))
 }
 
@@ -259,7 +302,7 @@ fn cli() -> clap::Command {
             Arg::new("control-plane-uri")
                 .short('p')
                 .long("control-plane-uri")
-                .value_name("CONTROL_PLANE"),
+                .value_name("CONTROL_PLANE_API_BASE_URI"),
         )
 }
 

compute_tools/src/compute.rs

@@ -18,8 +18,10 @@ use std::fs;
 use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 use std::process::{Command, Stdio};
+use std::str::FromStr;
 use std::sync::atomic::{AtomicU64, Ordering};
-use std::sync::RwLock;
+// use std::sync::RwLock;
+use std::sync::{Condvar, Mutex};
 
 use anyhow::{Context, Result};
 use chrono::{DateTime, Utc};
@@ -40,15 +42,22 @@ pub struct ComputeNode {
     pub connstr: url::Url,
     pub pgdata: String,
     pub pgbin: String,
-    pub spec: ComputeSpec,
-    pub tenant: String,
-    pub timeline: String,
-    pub pageserver_connstr: String,
     pub metrics: ComputeMetrics,
-    /// Volatile part of the `ComputeNode` so should be used under `RwLock`
-    /// to allow HTTP API server to serve status requests, while configuration
-    /// is in progress.
-    pub state: RwLock<ComputeState>,
+    // We only allow live re- / configuration of the compute node if
+    // it uses 'pull model', i.e. it can go to control-plane and fetch
+    // the latest configuration. Otherwise, there could be a case:
+    // - we start compute with some spec provided as argument
+    // - we push new spec and it does reconfiguration
+    // - but then something happens and compute pod / VM is destroyed,
+    //   so k8s controller starts it again with the **old** spec
+    pub live_config_allowed: bool,
+    /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
+    /// Coupled with `Condvar` to allow notifying HTTP API and configurator
+    /// thread about state changes. To allow HTTP API server to serving status
+    /// requests, while configuration is in progress, lock should be held only
+    /// for short periods of time to do read/write, not the whole configuration
+    /// process.
+    pub state: (Mutex<ComputeState>, Condvar),
 }
 
 fn rfc3339_serialize<S>(x: &DateTime<Utc>, s: S) -> Result<S::Ok, S::Error>
@@ -58,7 +67,7 @@ where
     x.to_rfc3339().serialize(s)
 }
 
-#[derive(Serialize)]
+#[derive(Serialize, Clone)]
 #[serde(rename_all = "snake_case")]
 pub struct ComputeState {
     pub status: ComputeStatus,
@@ -66,14 +75,27 @@ pub struct ComputeState {
     #[serde(serialize_with = "rfc3339_serialize")]
     pub last_active: DateTime<Utc>,
     pub error: Option<String>,
+    #[serde(skip_serializing)]
+    pub spec: ComputeSpec,
+    pub tenant: String,
+    pub timeline: String,
+    #[serde(skip_serializing)]
+    pub pageserver_connstr: String,
+    #[serde(skip_serializing)]
+    pub storage_auth_token: Option<String>,
 }
 
 impl ComputeState {
     pub fn new() -> Self {
         Self {
-            status: ComputeStatus::Init,
+            status: ComputeStatus::Empty,
             last_active: Utc::now(),
             error: None,
+            spec: ComputeSpec::default(),
+            tenant: String::new(),
+            timeline: String::new(),
+            pageserver_connstr: String::new(),
+            storage_auth_token: None,
         }
     }
 }
@@ -84,12 +106,25 @@ impl Default for ComputeState {
     }
 }
 
-#[derive(Serialize, Clone, Copy, PartialEq, Eq)]
+#[derive(Serialize, Clone, Copy, PartialEq, Eq, Debug)]
 #[serde(rename_all = "snake_case")]
 pub enum ComputeStatus {
+    // Spec wasn't provided as start, waiting for it to be
+    // provided by control-plane.
+    Empty,
+    // Compute node has spec and initial startup and
+    // configuration is in progress.
     Init,
+    // Compute is configured and running.
     Running,
+    // Either startup or configuration failed,
+    // compute will exit soon or is waiting for
+    // control-plane to terminate it.
     Failed,
+    // Control-plane requested reconfiguration.
+    ConfigurationPending,
+    // New spec is being applied.
+    Configuration,
 }
 
 #[derive(Default, Serialize)]
@@ -102,11 +137,16 @@ pub struct ComputeMetrics {
 
 impl ComputeNode {
     pub fn set_status(&self, status: ComputeStatus) {
-        self.state.write().unwrap().status = status;
+        let (state, state_changed) = &self.state;
+        let mut state = state.lock().unwrap();
+        state.status = status;
+        state_changed.notify_all();
     }
 
     pub fn get_status(&self) -> ComputeStatus {
-        self.state.read().unwrap().status
+        // self.state.read().unwrap().status
+        let (state, _) = &self.state;
+        state.lock().unwrap().status
     }
 
     // Remove `pgdata` directory and create it again with right permissions.
@@ -122,14 +162,31 @@ impl ComputeNode {
 
     // Get basebackup from the libpq connection to pageserver using `connstr` and
     // unarchive it to `pgdata` directory overriding all its previous content.
-    #[instrument(skip(self))]
-    fn get_basebackup(&self, lsn: &str) -> Result<()> {
+    #[instrument(skip(self, compute_state))]
+    fn get_basebackup(&self, compute_state: &ComputeState, lsn: &str) -> Result<()> {
         let start_time = Utc::now();
 
-        let mut client = Client::connect(&self.pageserver_connstr, NoTls)?;
+        let mut config = postgres::Config::from_str(&compute_state.pageserver_connstr)?;
+
+        // Use the storage auth token from the config file, if given.
+        // Note: this overrides any password set in the connection string.
+        if let Some(storage_auth_token) = &compute_state.storage_auth_token {
+            info!("Got storage auth token from spec file");
+            config.password(storage_auth_token);
+        } else {
+            info!("Storage auth token not set");
+        }
+
+        let mut client = config.connect(NoTls)?;
         let basebackup_cmd = match lsn {
-            "0/0" => format!("basebackup {} {}", &self.tenant, &self.timeline), // First start of the compute
-            _ => format!("basebackup {} {} {}", &self.tenant, &self.timeline, lsn),
+            "0/0" => format!(
+                "basebackup {} {}",
+                &compute_state.tenant, &compute_state.timeline
+            ), // First start of the compute
+            _ => format!(
+                "basebackup {} {} {}",
+                &compute_state.tenant, &compute_state.timeline, lsn
+            ),
         };
         let copyreader = client.copy_out(basebackup_cmd.as_str())?;
 
@@ -156,13 +213,18 @@ impl ComputeNode {
 
     // Run `postgres` in a special mode with `--sync-safekeepers` argument
     // and return the reported LSN back to the caller.
-    #[instrument(skip(self))]
-    fn sync_safekeepers(&self) -> Result<String> {
+    #[instrument(skip(self, storage_auth_token))]
+    fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<String> {
         let start_time = Utc::now();
 
         let sync_handle = Command::new(&self.pgbin)
             .args(["--sync-safekeepers"])
             .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
+            .envs(if let Some(storage_auth_token) = &storage_auth_token {
+                vec![("NEON_AUTH_TOKEN", storage_auth_token)]
+            } else {
+                vec![]
+            })
             .stdout(Stdio::piped())
             .spawn()
             .expect("postgres --sync-safekeepers failed to start");
@@ -199,9 +261,9 @@ impl ComputeNode {
 
     /// Do all the preparations like PGDATA directory creation, configuration,
     /// safekeepers sync, basebackup, etc.
-    #[instrument(skip(self))]
-    pub fn prepare_pgdata(&self) -> Result<()> {
-        let spec = &self.spec;
+    #[instrument(skip(self, compute_state))]
+    pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
+        let spec = &compute_state.spec;
         let pgdata_path = Path::new(&self.pgdata);
 
         // Remove/create an empty pgdata directory and put configuration there.
@@ -210,18 +272,18 @@ impl ComputeNode {
 
         info!("starting safekeepers syncing");
         let lsn = self
-            .sync_safekeepers()
+            .sync_safekeepers(compute_state.storage_auth_token.clone())
             .with_context(|| "failed to sync safekeepers")?;
         info!("safekeepers synced at LSN {}", lsn);
 
         info!(
             "getting basebackup@{} from pageserver {}",
-            lsn, &self.pageserver_connstr
+            lsn, &compute_state.pageserver_connstr
         );
-        self.get_basebackup(&lsn).with_context(|| {
+        self.get_basebackup(compute_state, &lsn).with_context(|| {
             format!(
                 "failed to get basebackup@{} from pageserver {}",
-                lsn, &self.pageserver_connstr
+                lsn, &compute_state.pageserver_connstr
             )
         })?;
 
@@ -234,12 +296,20 @@ impl ComputeNode {
     /// Start Postgres as a child process and manage DBs/roles.
     /// After that this will hang waiting on the postmaster process to exit.
     #[instrument(skip(self))]
-    pub fn start_postgres(&self) -> Result<std::process::Child> {
+    pub fn start_postgres(
+        &self,
+        storage_auth_token: Option<String>,
+    ) -> Result<std::process::Child> {
         let pgdata_path = Path::new(&self.pgdata);
 
         // Run postgres as a child process.
         let mut pg = Command::new(&self.pgbin)
             .args(["-D", &self.pgdata])
+            .envs(if let Some(storage_auth_token) = &storage_auth_token {
+                vec![("NEON_AUTH_TOKEN", storage_auth_token)]
+            } else {
+                vec![]
+            })
             .spawn()
             .expect("cannot start postgres process");
 
@@ -248,8 +318,9 @@ impl ComputeNode {
         Ok(pg)
     }
 
-    #[instrument(skip(self))]
-    pub fn apply_config(&self) -> Result<()> {
+    /// Do initial configuration of the already started Postgres.
+    #[instrument(skip(self, compute_state))]
+    pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
         // If connection fails,
         // it may be the old node with `zenith_admin` superuser.
         //
@@ -280,19 +351,62 @@ impl ComputeNode {
         };
 
         // Proceed with post-startup configuration. Note, that order of operations is important.
-        handle_roles(&self.spec, &mut client)?;
-        handle_databases(&self.spec, &mut client)?;
-        handle_role_deletions(self, &mut client)?;
-        handle_grants(self, &mut client)?;
+        handle_roles(&compute_state.spec, &mut client)?;
+        handle_databases(&compute_state.spec, &mut client)?;
+        handle_role_deletions(&compute_state.spec, self.connstr.as_str(), &mut client)?;
+        handle_grants(&compute_state.spec, self.connstr.as_str(), &mut client)?;
         create_writability_check_data(&mut client)?;
-        handle_extensions(&self.spec, &mut client)?;
+        handle_extensions(&compute_state.spec, &mut client)?;
 
         // 'Close' connection
         drop(client);
 
         info!(
             "finished configuration of compute for project {}",
-            self.spec.cluster.cluster_id
+            compute_state.spec.cluster.cluster_id
+        );
+
+        Ok(())
+    }
+
+    // We could've wrapped this around `pg_ctl reload`, but right now we don't use
+    // `pg_ctl` for start / stop, so this just seems much easier to do as we already
+    // have opened connection to Postgres and superuser access.
+    #[instrument(skip(self, client))]
+    fn pg_reload_conf(&self, client: &mut Client) -> Result<()> {
+        client.simple_query("SELECT pg_reload_conf()")?;
+        Ok(())
+    }
+
+    /// Similar to `apply_config()`, but does a bit different sequence of operations,
+    /// as it's used to reconfigure a previously started and configured Postgres node.
+    #[instrument(skip(self))]
+    pub fn reconfigure(&self) -> Result<()> {
+        let (state, _) = &self.state;
+        let spec = state.lock().unwrap().spec.clone();
+
+        // Write new config
+        let pgdata_path = Path::new(&self.pgdata);
+        config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), &spec)?;
+
+        let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
+        self.pg_reload_conf(&mut client)?;
+
+        // Proceed with post-startup configuration. Note, that order of operations is important.
+        handle_roles(&spec, &mut client)?;
+        handle_databases(&spec, &mut client)?;
+        handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
+        handle_grants(&spec, self.connstr.as_str(), &mut client)?;
+        handle_extensions(&spec, &mut client)?;
+
+        // 'Close' connection
+        drop(client);
+
+        let unknown_op = "unknown".to_string();
+        let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
+        info!(
+            "finished reconfiguration of compute node for operation {}",
+            op_id
         );
 
         Ok(())
@@ -300,21 +414,24 @@ impl ComputeNode {
 
     #[instrument(skip(self))]
     pub fn start_compute(&self) -> Result<std::process::Child> {
+        // let compute_state = self.state.read().unwrap().clone();
+        let (state, _) = &self.state;
+        let compute_state = state.lock().unwrap().clone();
         info!(
             "starting compute for project {}, operation {}, tenant {}, timeline {}",
-            self.spec.cluster.cluster_id,
-            self.spec.operation_uuid.as_ref().unwrap(),
-            self.tenant,
-            self.timeline,
+            compute_state.spec.cluster.cluster_id,
+            compute_state.spec.operation_uuid.as_ref().unwrap(),
+            compute_state.tenant,
+            compute_state.timeline,
         );
 
-        self.prepare_pgdata()?;
+        self.prepare_pgdata(&compute_state)?;
 
         let start_time = Utc::now();
 
-        let pg = self.start_postgres()?;
+        let pg = self.start_postgres(compute_state.storage_auth_token.clone())?;
 
-        self.apply_config()?;
+        self.apply_config(&compute_state)?;
 
         let startup_end_time = Utc::now();
         self.metrics.config_ms.store(

compute_tools/src/configurator.rs

@@ -0,0 +1,53 @@
+use std::sync::Arc;
+use std::thread;
+
+use anyhow::Result;
+use tracing::{error, info, instrument};
+
+use crate::compute::{ComputeNode, ComputeStatus};
+
+#[instrument(skip(compute))]
+fn configurator_main_loop(compute: &Arc<ComputeNode>) {
+    info!("waiting for reconfiguration requests");
+    let (state, state_changed) = &compute.state;
+    loop {
+        let state = state.lock().unwrap();
+        let mut state = state_changed.wait(state).unwrap();
+
+        if state.status == ComputeStatus::ConfigurationPending {
+            info!("got configuration request");
+            state.status = ComputeStatus::Configuration;
+            state_changed.notify_all();
+            drop(state);
+
+            let mut new_status = ComputeStatus::Failed;
+            if let Err(e) = compute.reconfigure() {
+                error!("could not configure compute node: {}", e);
+            } else {
+                new_status = ComputeStatus::Running;
+                info!("compute node configured");
+            }
+
+            // XXX: used to test that API is blocking
+            // std::thread::sleep(std::time::Duration::from_millis(2000));
+
+            compute.set_status(new_status);
+        } else if state.status == ComputeStatus::Failed {
+            info!("compute node is now in Failed state, exiting");
+            break;
+        } else {
+            info!("woken up for compute status: {:?}, sleeping", state.status);
+        }
+    }
+}
+
+pub fn launch_configurator(compute: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>> {
+    let compute = Arc::clone(compute);
+
+    Ok(thread::Builder::new()
+        .name("compute-configurator".into())
+        .spawn(move || {
+            configurator_main_loop(&compute);
+            info!("configurator thread is exited");
+        })?)
+}

compute_tools/src/http/api.rs

@@ -3,7 +3,9 @@ use std::net::SocketAddr;
 use std::sync::Arc;
 use std::thread;
 
-use crate::compute::ComputeNode;
+use crate::compute::{ComputeNode, ComputeStatus};
+use crate::http::models::{ConfigurationRequest, GenericAPIError};
+
 use anyhow::Result;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::{Body, Method, Request, Response, Server, StatusCode};
@@ -23,7 +25,9 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
         // Serialized compute state.
         (&Method::GET, "/status") => {
             info!("serving /status GET request");
-            let state = compute.state.read().unwrap();
+            // let state = compute.state.read().unwrap();
+            let (state, _) = &compute.state;
+            let state = state.lock().unwrap();
             Response::new(Body::from(serde_json::to_string(&*state).unwrap()))
         }
 
@@ -37,12 +41,29 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
         // Collect Postgres current usage insights
         (&Method::GET, "/insights") => {
             info!("serving /insights GET request");
+            let status = compute.get_status();
+            if status != ComputeStatus::Running {
+                let msg = format!("compute is not running, current status: {:?}", status);
+                error!(msg);
+                return Response::new(Body::from(msg));
+            }
+
             let insights = compute.collect_insights().await;
             Response::new(Body::from(insights))
         }
 
         (&Method::POST, "/check_writability") => {
             info!("serving /check_writability POST request");
+            let status = compute.get_status();
+            if status != ComputeStatus::Running {
+                let msg = format!(
+                    "invalid compute status for check_writability request: {:?}",
+                    status
+                );
+                error!(msg);
+                return Response::new(Body::from(msg));
+            }
+
             let res = crate::checker::check_writability(compute).await;
             match res {
                 Ok(_) => Response::new(Body::from("true")),
@@ -61,6 +82,61 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
             ))
         }
 
+        // Accept spec in JSON format and request compute configuration from
+        // the configurator thread. If anything goes wrong after we set the
+        // compute state to `ConfigurationPending` and / or sent spec to the
+        // configurator thread, we basically leave compute in the potentially
+        // wrong state. That said, it's control-plane's responsibility to
+        // watch compute state after reconfiguration request and to clean
+        // restart in case of errors.
+        (&Method::POST, "/configure") => {
+            info!("serving /configure POST request");
+            if !compute.live_config_allowed {
+                let msg = "live reconfiguration is not allowed for this compute node";
+                error!(msg);
+                return render_json_error(msg, StatusCode::PRECONDITION_FAILED);
+            }
+
+            let body_bytes = hyper::body::to_bytes(req.into_body()).await.unwrap();
+            let spec_raw = String::from_utf8(body_bytes.to_vec()).unwrap();
+            if let Ok(request) = serde_json::from_str::<ConfigurationRequest>(&spec_raw) {
+                let spec = request.spec;
+                let (state, state_changed) = &compute.state;
+                let mut state = state.lock().unwrap();
+                if !(state.status == ComputeStatus::Empty || state.status == ComputeStatus::Running)
+                {
+                    let msg = format!(
+                        "invalid compute status for reconfiguration request: {}",
+                        serde_json::to_string(&*state).unwrap()
+                    );
+                    error!(msg);
+                    return render_json_error(&msg, StatusCode::PRECONDITION_FAILED);
+                }
+                state.spec = spec;
+                state.status = ComputeStatus::ConfigurationPending;
+                state_changed.notify_all();
+                drop(state);
+                info!("set new spec and notified configurator");
+
+                let (state, state_changed) = &compute.state;
+                let mut state = state.lock().unwrap();
+                while state.status != ComputeStatus::Running {
+                    state = state_changed.wait(state).unwrap();
+                    info!(
+                        "waiting for compute to become Running, current status: {:?}",
+                        state.status
+                    );
+                }
+
+                // Return current compute state if everything went well.
+                Response::new(Body::from(serde_json::to_string(&*state).unwrap()))
+            } else {
+                let msg = "invalid spec";
+                error!(msg);
+                render_json_error(msg, StatusCode::BAD_REQUEST)
+            }
+        }
+
         // Return the `404 Not Found` for any other routes.
         _ => {
             let mut not_found = Response::new(Body::from("404 Not Found"));
@@ -70,6 +146,16 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
     }
 }
 
+fn render_json_error(e: &str, status: StatusCode) -> Response<Body> {
+    let error = GenericAPIError {
+        error: e.to_string(),
+    };
+    Response::builder()
+        .status(status)
+        .body(Body::from(serde_json::to_string(&error).unwrap()))
+        .unwrap()
+}
+
 // Main Hyper HTTP server function that runs it and blocks waiting on it forever.
 #[tokio::main]
 async fn serve(state: Arc<ComputeNode>) {
@@ -110,7 +196,6 @@ async fn serve(state: Arc<ComputeNode>) {
 /// Launch a separate Hyper HTTP API server thread and return its `JoinHandle`.
 pub fn launch_http_server(state: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>> {
     let state = Arc::clone(state);
-
     Ok(thread::Builder::new()
         .name("http-endpoint".into())
         .spawn(move || serve(state))?)

compute_tools/src/http/mod.rs

@@ -1 +1,2 @@
 pub mod api;
+pub mod models;

compute_tools/src/http/models.rs

@@ -0,0 +1,16 @@
+use serde::{Deserialize, Serialize};
+
+use crate::spec::ComputeSpec;
+
+/// We now pass only `spec` in the configuration request, but later we can
+/// extend it and something like `restart: bool` or something else. So put
+/// `spec` into a struct initially to be more flexible in the future.
+#[derive(Deserialize, Debug)]
+pub struct ConfigurationRequest {
+    pub spec: ComputeSpec,
+}
+
+#[derive(Serialize, Debug)]
+pub struct GenericAPIError {
+    pub error: String,
+}

compute_tools/src/http/openapi_spec.yaml

@@ -11,7 +11,7 @@ paths:
     get:
       tags:
       - Info
-      summary: Get compute node internal status
+      summary: Get compute node internal status.
       description: ""
       operationId: getComputeStatus
       responses:
@@ -26,7 +26,7 @@ paths:
     get:
       tags:
       - Info
-      summary: Get compute node startup metrics in JSON format
+      summary: Get compute node startup metrics in JSON format.
       description: ""
       operationId: getComputeMetricsJSON
       responses:
@@ -41,9 +41,9 @@ paths:
     get:
       tags:
       - Info
-      summary: Get current compute insights in JSON format
+      summary: Get current compute insights in JSON format.
       description: |
-        Note, that this doesn't include any historical data
+        Note, that this doesn't include any historical data.
       operationId: getComputeInsights
       responses:
         200:
@@ -56,12 +56,12 @@ paths:
   /info:
     get:
       tags:
-      - "info"
-      summary: Get info about the compute Pod/VM
+      - Info
+      summary: Get info about the compute pod / VM.
       description: ""
       operationId: getInfo
       responses:
-        "200":
+        200:
           description: Info
           content:
             application/json:
@@ -72,7 +72,7 @@ paths:
     post:
       tags:
       - Check
-      summary: Check that we can write new data on this compute
+      summary: Check that we can write new data on this compute.
       description: ""
       operationId: checkComputeWritability
       responses:
@@ -82,9 +82,57 @@ paths:
             text/plain:
               schema:
                 type: string
-                description: Error text or 'true' if check passed
+                description: Error text or 'true' if check passed.
                 example: "true"
 
+  /configure:
+    post:
+      tags:
+      - Configure
+      summary: Request compute node configuration.
+      description: |
+        This is a blocking API endpoint, i.e. it blocks waiting until
+        compute is finished configuration and is in `Running` state.
+        Optional non-blocking mode could be added later. Currently,
+        it's also assumed that reconfiguration doesn't require restart.
+      operationId: configureCompute
+      requestBody:
+        description: Configuration request.
+        required: true
+        content:
+          application/json:
+            schema:
+              type: object
+              required:
+                - spec
+              properties:
+                spec:
+                  # XXX: I don't want to explain current spec in the OpenAPI format,
+                  # as it could be changed really soon. Consider doing it later.
+                  type: object
+      responses:
+        200:
+          description: Compute configuration finished.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ComputeState"
+        400:
+          description: Provided spec is invalid.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
+        412:
+          description: |
+            It's not possible to do live-configuration of the compute.
+            It's either in the wrong state, or compute doesn't use pull
+            mode of configuration.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
+
 components:
   securitySchemes:
     JWT:
@@ -95,7 +143,7 @@ components:
   schemas:
     ComputeMetrics:
       type: object
-      description: Compute startup metrics
+      description: Compute startup metrics.
       required:
         - sync_safekeepers_ms
         - basebackup_ms
@@ -113,7 +161,7 @@ components:
 
     Info:
       type: object
-      description: Information about VM/Pod
+      description: Information about VM/Pod.
       required:
         - num_cpus
       properties:
@@ -130,17 +178,26 @@ components:
           $ref: '#/components/schemas/ComputeStatus'
         last_active:
           type: string
-          description: The last detected compute activity timestamp in UTC and RFC3339 format
+          description: The last detected compute activity timestamp in UTC and RFC3339 format.
           example: "2022-10-12T07:20:50.52Z"
         error:
           type: string
-          description: Text of the error during compute startup, if any
+          description: Text of the error during compute startup, if any.
+          example: ""
+        tenant:
+          type: string
+          description: Identifier of the current tenant served by compute node, if any.
+          example: c9269c359e9a199fad1ea0981246a78f
+        timeline:
+          type: string
+          description: Identifier of the current timeline served by compute node, if any.
+          example: ece7de74d4b8cbe5433a68ce4d1b97b4
 
     ComputeInsights:
       type: object
       properties:
         pg_stat_statements:
-          description: Contains raw output from pg_stat_statements in JSON format
+          description: Contains raw output from pg_stat_statements in JSON format.
           type: array
           items:
             type: object
@@ -151,6 +208,19 @@ components:
         - init
         - failed
         - running
+      example: running
+
+    #
+    # Errors
+    #
+
+    GenericError:
+      type: object
+      required:
+        - error
+      properties:
+        error:
+          type: string
 
 security:
   - JWT: []

compute_tools/src/lib.rs

@@ -4,6 +4,7 @@
 //!
 pub mod checker;
 pub mod config;
+pub mod configurator;
 pub mod http;
 #[macro_use]
 pub mod logger;

compute_tools/src/monitor.rs

@@ -46,7 +46,9 @@ fn watch_compute_activity(compute: &ComputeNode) {
                             AND usename != 'cloud_admin';", // XXX: find a better way to filter other monitors?
                         &[],
                     );
-                let mut last_active = compute.state.read().unwrap().last_active;
+                // let mut last_active = compute.state.read().unwrap().last_active;
+                let (state, _) = &compute.state;
+                let mut last_active = state.lock().unwrap().last_active;
 
                 if let Ok(backs) = backends {
                     let mut idle_backs: Vec<DateTime<Utc>> = vec![];
@@ -87,7 +89,8 @@ fn watch_compute_activity(compute: &ComputeNode) {
                 }
 
                 // Update the last activity in the shared state if we got a more recent one.
-                let mut state = compute.state.write().unwrap();
+                let (state, _) = &compute.state;
+                let mut state = state.lock().unwrap();
                 if last_active > state.last_active {
                     state.last_active = last_active;
                     debug!("set the last compute activity time to: {}", last_active);

compute_tools/src/pg_helpers.rs

@@ -17,7 +17,7 @@ const POSTGRES_WAIT_TIMEOUT: Duration = Duration::from_millis(60 * 1000); // mil
 
 /// Rust representation of Postgres role info with only those fields
 /// that matter for us.
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug)]
 pub struct Role {
     pub name: PgIdent,
     pub encrypted_password: Option<String>,
@@ -26,7 +26,7 @@ pub struct Role {
 
 /// Rust representation of Postgres database info with only those fields
 /// that matter for us.
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug)]
 pub struct Database {
     pub name: PgIdent,
     pub owner: PgIdent,
@@ -36,7 +36,7 @@ pub struct Database {
 /// Common type representing both SQL statement params with or without value,
 /// like `LOGIN` or `OWNER username` in the `CREATE/ALTER ROLE`, and config
 /// options like `wal_level = logical`.
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug)]
 pub struct GenericOption {
     pub name: String,
     pub value: Option<String>,
@@ -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()

compute_tools/src/spec.rs

@@ -8,14 +8,13 @@ use postgres::{Client, NoTls};
 use serde::Deserialize;
 use tracing::{info, info_span, instrument, span_enabled, warn, Level};
 
-use crate::compute::ComputeNode;
 use crate::config;
 use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 /// Cluster spec or configuration represented as an optional number of
 /// delta operations + final cluster state description.
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug, Default)]
 pub struct ComputeSpec {
     pub format_version: f32,
     pub timestamp: String,
@@ -24,12 +23,14 @@ pub struct ComputeSpec {
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,
 
+    pub storage_auth_token: Option<String>,
+
     pub startup_tracing_context: Option<HashMap<String, String>>,
 }
 
 /// Cluster state seen from the perspective of the external tools
 /// like Rails web console.
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug, Default)]
 pub struct Cluster {
     pub cluster_id: String,
     pub name: String,
@@ -45,13 +46,36 @@ pub struct Cluster {
 /// - DROP ROLE
 /// - ALTER ROLE name RENAME TO new_name
 /// - ALTER DATABASE name RENAME TO new_name
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, Debug)]
 pub struct DeltaOp {
     pub action: String,
     pub name: PgIdent,
     pub new_name: Option<PgIdent>,
 }
 
+/// Request spec from the control-plane by compute_id. If `NEON_CONSOLE_JWT`
+/// env variable is set, it will be used for authorization.
+pub fn get_spec_from_control_plane(base_uri: &str, compute_id: &str) -> Result<ComputeSpec> {
+    let cp_uri = format!("{base_uri}/management/api/v2/computes/{compute_id}/spec");
+    let jwt: String = match std::env::var("NEON_CONSOLE_JWT") {
+        Ok(v) => v,
+        Err(_) => "".to_string(),
+    };
+    info!("getting spec from control plane: {}", cp_uri);
+
+    // TODO: check the response. We should distinguish cases when it's
+    // - network error, then retry
+    // - no spec for compute yet, then wait
+    // - compute id is unknown or any other error, then bail out
+    let spec = reqwest::blocking::Client::new()
+        .get(cp_uri)
+        .header("Authorization", jwt)
+        .send()?
+        .json()?;
+
+    Ok(spec)
+}
+
 /// It takes cluster specification and does the following:
 /// - Serialize cluster config and put it into `postgresql.conf` completely rewriting the file.
 /// - Update `pg_hba.conf` to allow external connections.
@@ -224,8 +248,8 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 
 /// Reassign all dependent objects and delete requested roles.
 #[instrument(skip_all)]
-pub fn handle_role_deletions(node: &ComputeNode, client: &mut Client) -> Result<()> {
-    if let Some(ops) = &node.spec.delta_operations {
+pub fn handle_role_deletions(spec: &ComputeSpec, connstr: &str, client: &mut Client) -> Result<()> {
+    if let Some(ops) = &spec.delta_operations {
         // First, reassign all dependent objects to db owners.
         info!("reassigning dependent objects of to-be-deleted roles");
 
@@ -242,7 +266,7 @@ pub fn handle_role_deletions(node: &ComputeNode, client: &mut Client) -> Result<
             // Check that role is still present in Postgres, as this could be a
             // restart with the same spec after role deletion.
             if op.action == "delete_role" && existing_roles.iter().any(|r| r.name == op.name) {
-                reassign_owned_objects(node, &op.name)?;
+                reassign_owned_objects(spec, connstr, &op.name)?;
             }
         }
 
@@ -266,10 +290,10 @@ pub fn handle_role_deletions(node: &ComputeNode, client: &mut Client) -> Result<
 }
 
 // Reassign all owned objects in all databases to the owner of the database.
-fn reassign_owned_objects(node: &ComputeNode, role_name: &PgIdent) -> Result<()> {
-    for db in &node.spec.cluster.databases {
+fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent) -> Result<()> {
+    for db in &spec.cluster.databases {
         if db.owner != *role_name {
-            let mut conf = Config::from_str(node.connstr.as_str())?;
+            let mut conf = Config::from_str(connstr)?;
             conf.dbname(&db.name);
 
             let mut client = conf.connect(NoTls)?;
@@ -414,9 +438,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(node: &ComputeNode, client: &mut Client) -> Result<()> {
-    let spec = &node.spec;
-
+pub fn handle_grants(spec: &ComputeSpec, connstr: &str, client: &mut Client) -> Result<()> {
     info!("cluster spec grants:");
 
     // We now have a separate `web_access` role to connect to the database
@@ -448,8 +470,8 @@ pub fn handle_grants(node: &ComputeNode, client: &mut Client) -> Result<()> {
     // Do some per-database access adjustments. We'd better do this at db creation time,
     // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
     // atomically.
-    for db in &node.spec.cluster.databases {
-        let mut conf = Config::from_str(node.connstr.as_str())?;
+    for db in &spec.cluster.databases {
+        let mut conf = Config::from_str(connstr)?;
         conf.dbname(&db.name);
 
         let mut db_client = conf.connect(NoTls)?;

control_plane/safekeepers.conf

@@ -2,7 +2,8 @@
 [pageserver]
 listen_pg_addr = '127.0.0.1:64000'
 listen_http_addr = '127.0.0.1:9898'
-auth_type = 'Trust'
+pg_auth_type = 'Trust'
+http_auth_type = 'Trust'
 
 [[safekeepers]]
 id = 1

control_plane/simple.conf

@@ -3,7 +3,8 @@
 [pageserver]
 listen_pg_addr = '127.0.0.1:64000'
 listen_http_addr = '127.0.0.1:9898'
-auth_type = 'Trust'
+pg_auth_type = 'Trust'
+http_auth_type = 'Trust'
 
 [[safekeepers]]
 id = 1

control_plane/src/bin/neon_local.rs

@@ -53,14 +53,15 @@ listen_addr = '{DEFAULT_BROKER_ADDR}'
 id = {DEFAULT_PAGESERVER_ID}
 listen_pg_addr = '{DEFAULT_PAGESERVER_PG_ADDR}'
 listen_http_addr = '{DEFAULT_PAGESERVER_HTTP_ADDR}'
-auth_type = '{pageserver_auth_type}'
+pg_auth_type = '{trust_auth}'
+http_auth_type = '{trust_auth}'
 
 [[safekeepers]]
 id = {DEFAULT_SAFEKEEPER_ID}
 pg_port = {DEFAULT_SAFEKEEPER_PG_PORT}
 http_port = {DEFAULT_SAFEKEEPER_HTTP_PORT}
 "#,
-        pageserver_auth_type = AuthType::Trust,
+        trust_auth = AuthType::Trust,
     )
 }
 
@@ -627,7 +628,7 @@ fn handle_pg(pg_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
 
             let node = cplane.nodes.get(&(tenant_id, node_name.to_string()));
 
-            let auth_token = if matches!(env.pageserver.auth_type, AuthType::NeonJWT) {
+            let auth_token = if matches!(env.pageserver.pg_auth_type, AuthType::NeonJWT) {
                 let claims = Claims::new(Some(tenant_id), Scope::Tenant);
 
                 Some(env.generate_auth_token(&claims)?)

control_plane/src/compute.rs

@@ -11,7 +11,6 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use anyhow::{Context, Result};
-use postgres_backend::AuthType;
 use utils::{
     id::{TenantId, TimelineId},
     lsn::Lsn,
@@ -88,16 +87,14 @@ 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,
         });
 
         node.create_pgdata()?;
-        node.setup_pg_conf(self.env.pageserver.auth_type)?;
+        node.setup_pg_conf()?;
 
         self.nodes
             .insert((tenant_id, node.name.clone()), Arc::clone(&node));
@@ -114,11 +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,
 }
 
@@ -152,7 +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.
@@ -172,11 +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,
         })
     }
@@ -278,7 +270,7 @@ impl PostgresNode {
 
     // Write postgresql.conf with default configuration
     // and PG_VERSION file to the data directory of a new node.
-    fn setup_pg_conf(&self, auth_type: AuthType) -> Result<()> {
+    fn setup_pg_conf(&self) -> Result<()> {
         let mut conf = PostgresConf::new();
         conf.append("max_wal_senders", "10");
         conf.append("wal_log_hints", "off");
@@ -302,29 +294,12 @@ impl PostgresNode {
             let config = &self.pageserver.pg_connection_config;
             let (host, port) = (config.host(), config.port());
 
-            // Set up authentication
-            //
-            // $NEON_AUTH_TOKEN will be replaced with value from environment
-            // variable during compute pg startup. It is done this way because
-            // otherwise user will be able to retrieve the value using SHOW
-            // command or pg_settings
-            let password = if let AuthType::NeonJWT = auth_type {
-                "$NEON_AUTH_TOKEN"
-            } else {
-                ""
-            };
-            // NOTE avoiding spaces in connection string, because it is less error prone if we forward it somewhere.
-            // Also note that not all parameters are supported here. Because in compute we substitute $NEON_AUTH_TOKEN
-            // We parse this string and build it back with token from env var, and for simplicity rebuild
-            // uses only needed variables namely host, port, user, password.
-            format!("postgresql://no_user:{password}@{host}:{port}")
+            // NOTE: avoid spaces in connection string, because it is less error prone if we forward it somewhere.
+            format!("postgresql://no_user@{host}:{port}")
         };
         conf.append("shared_preload_libraries", "neon");
         conf.append_line("");
         conf.append("neon.pageserver_connstring", &pageserver_connstr);
-        if let AuthType::NeonJWT = auth_type {
-            conf.append("neon.safekeeper_token_env", "$NEON_AUTH_TOKEN");
-        }
         conf.append("neon.tenant_id", &self.tenant_id.to_string());
         conf.append("neon.timeline_id", &self.timeline_id.to_string());
         if let Some(lsn) = self.lsn {
@@ -385,7 +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
@@ -424,7 +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(
             [
                 &[
@@ -447,11 +422,15 @@ impl PostgresNode {
             "DYLD_LIBRARY_PATH",
             self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
         );
+
+        // Pass authentication token used for the connections to pageserver and safekeepers
         if let Some(token) = auth_token {
             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: {}",
@@ -496,10 +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,
@@ -530,26 +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/local_env.rs

@@ -18,7 +18,7 @@ use std::net::SocketAddr;
 use std::path::{Path, PathBuf};
 use std::process::{Command, Stdio};
 use utils::{
-    auth::{encode_from_key_file, Claims, Scope},
+    auth::{encode_from_key_file, Claims},
     id::{NodeId, TenantId, TenantTimelineId, TimelineId},
 };
 
@@ -110,15 +110,14 @@ impl NeonBroker {
 pub struct PageServerConf {
     // node id
     pub id: NodeId,
+
     // Pageserver connection settings
     pub listen_pg_addr: String,
     pub listen_http_addr: String,
 
-    // used to determine which auth type is used
-    pub auth_type: AuthType,
-
-    // jwt auth token used for communication with pageserver
-    pub auth_token: String,
+    // auth type used for the PG and HTTP ports
+    pub pg_auth_type: AuthType,
+    pub http_auth_type: AuthType,
 }
 
 impl Default for PageServerConf {
@@ -127,8 +126,8 @@ impl Default for PageServerConf {
             id: NodeId(0),
             listen_pg_addr: String::new(),
             listen_http_addr: String::new(),
-            auth_type: AuthType::Trust,
-            auth_token: String::new(),
+            pg_auth_type: AuthType::Trust,
+            http_auth_type: AuthType::Trust,
         }
     }
 }
@@ -401,48 +400,33 @@ impl LocalEnv {
 
         fs::create_dir(base_path)?;
 
-        // generate keys for jwt
-        // openssl genrsa -out private_key.pem 2048
-        let private_key_path;
+        // Generate keypair for JWT.
+        //
+        // The keypair is only needed if authentication is enabled in any of the
+        // components. For convenience, we generate the keypair even if authentication
+        // is not enabled, so that you can easily enable it after the initialization
+        // step. However, if the key generation fails, we treat it as non-fatal if
+        // authentication was not enabled.
         if self.private_key_path == PathBuf::new() {
-            private_key_path = base_path.join("auth_private_key.pem");
-            let keygen_output = Command::new("openssl")
-                .arg("genrsa")
-                .args(["-out", private_key_path.to_str().unwrap()])
-                .arg("2048")
-                .stdout(Stdio::null())
-                .output()
-                .context("failed to generate auth private key")?;
-            if !keygen_output.status.success() {
-                bail!(
-                    "openssl failed: '{}'",
-                    String::from_utf8_lossy(&keygen_output.stderr)
-                );
-            }
-            self.private_key_path = PathBuf::from("auth_private_key.pem");
-
-            let public_key_path = base_path.join("auth_public_key.pem");
-            // openssl rsa -in private_key.pem -pubout -outform PEM -out public_key.pem
-            let keygen_output = Command::new("openssl")
-                .arg("rsa")
-                .args(["-in", private_key_path.to_str().unwrap()])
-                .arg("-pubout")
-                .args(["-outform", "PEM"])
-                .args(["-out", public_key_path.to_str().unwrap()])
-                .stdout(Stdio::null())
-                .output()
-                .context("failed to generate auth private key")?;
-            if !keygen_output.status.success() {
-                bail!(
-                    "openssl failed: '{}'",
-                    String::from_utf8_lossy(&keygen_output.stderr)
-                );
+            match generate_auth_keys(
+                base_path.join("auth_private_key.pem").as_path(),
+                base_path.join("auth_public_key.pem").as_path(),
+            ) {
+                Ok(()) => {
+                    self.private_key_path = PathBuf::from("auth_private_key.pem");
+                }
+                Err(e) => {
+                    if !self.auth_keys_needed() {
+                        eprintln!("Could not generate keypair for JWT authentication: {e}");
+                        eprintln!("Continuing anyway because authentication was not enabled");
+                        self.private_key_path = PathBuf::from("auth_private_key.pem");
+                    } else {
+                        return Err(e);
+                    }
+                }
             }
         }
 
-        self.pageserver.auth_token =
-            self.generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
-
         fs::create_dir_all(self.pg_data_dirs_path())?;
 
         for safekeeper in &self.safekeepers {
@@ -451,6 +435,12 @@ impl LocalEnv {
 
         self.persist_config(base_path)
     }
+
+    fn auth_keys_needed(&self) -> bool {
+        self.pageserver.pg_auth_type == AuthType::NeonJWT
+            || self.pageserver.http_auth_type == AuthType::NeonJWT
+            || self.safekeepers.iter().any(|sk| sk.auth_enabled)
+    }
 }
 
 fn base_path() -> PathBuf {
@@ -460,6 +450,43 @@ fn base_path() -> PathBuf {
     }
 }
 
+/// Generate a public/private key pair for JWT authentication
+fn generate_auth_keys(private_key_path: &Path, public_key_path: &Path) -> anyhow::Result<()> {
+    // Generate the key pair
+    //
+    // openssl genpkey -algorithm ed25519 -out auth_private_key.pem
+    let keygen_output = Command::new("openssl")
+        .arg("genpkey")
+        .args(["-algorithm", "ed25519"])
+        .args(["-out", private_key_path.to_str().unwrap()])
+        .stdout(Stdio::null())
+        .output()
+        .context("failed to generate auth private key")?;
+    if !keygen_output.status.success() {
+        bail!(
+            "openssl failed: '{}'",
+            String::from_utf8_lossy(&keygen_output.stderr)
+        );
+    }
+    // Extract the public key from the private key file
+    //
+    // openssl pkey -in auth_private_key.pem -pubout -out auth_public_key.pem
+    let keygen_output = Command::new("openssl")
+        .arg("pkey")
+        .args(["-in", private_key_path.to_str().unwrap()])
+        .arg("-pubout")
+        .args(["-out", public_key_path.to_str().unwrap()])
+        .output()
+        .context("failed to extract public key from private key")?;
+    if !keygen_output.status.success() {
+        bail!(
+            "openssl failed: '{}'",
+            String::from_utf8_lossy(&keygen_output.stderr)
+        );
+    }
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

control_plane/src/pageserver.rs

@@ -82,15 +82,8 @@ impl PageServerNode {
         let (host, port) = parse_host_port(&env.pageserver.listen_pg_addr)
             .expect("Unable to parse listen_pg_addr");
         let port = port.unwrap_or(5432);
-        let password = if env.pageserver.auth_type == AuthType::NeonJWT {
-            Some(env.pageserver.auth_token.clone())
-        } else {
-            None
-        };
-
         Self {
-            pg_connection_config: PgConnectionConfig::new_host_port(host, port)
-                .set_password(password),
+            pg_connection_config: PgConnectionConfig::new_host_port(host, port),
             env: env.clone(),
             http_client: Client::new(),
             http_base_url: format!("http://{}/v1", env.pageserver.listen_http_addr),
@@ -106,25 +99,32 @@ impl PageServerNode {
             self.env.pg_distrib_dir_raw().display()
         );
 
-        let authg_type_param = format!("auth_type='{}'", self.env.pageserver.auth_type);
+        let http_auth_type_param =
+            format!("http_auth_type='{}'", self.env.pageserver.http_auth_type);
         let listen_http_addr_param = format!(
             "listen_http_addr='{}'",
             self.env.pageserver.listen_http_addr
         );
+
+        let pg_auth_type_param = format!("pg_auth_type='{}'", self.env.pageserver.pg_auth_type);
         let listen_pg_addr_param =
             format!("listen_pg_addr='{}'", self.env.pageserver.listen_pg_addr);
+
         let broker_endpoint_param = format!("broker_endpoint='{}'", self.env.broker.client_url());
 
         let mut overrides = vec![
             id,
             pg_distrib_dir_param,
-            authg_type_param,
+            http_auth_type_param,
+            pg_auth_type_param,
             listen_http_addr_param,
             listen_pg_addr_param,
             broker_endpoint_param,
         ];
 
-        if self.env.pageserver.auth_type != AuthType::Trust {
+        if self.env.pageserver.http_auth_type != AuthType::Trust
+            || self.env.pageserver.pg_auth_type != AuthType::Trust
+        {
             overrides.push("auth_validation_public_key_path='auth_public_key.pem'".to_owned());
         }
         overrides
@@ -247,7 +247,10 @@ impl PageServerNode {
     }
 
     fn pageserver_env_variables(&self) -> anyhow::Result<Vec<(String, String)>> {
-        Ok(if self.env.pageserver.auth_type != AuthType::Trust {
+        // FIXME: why is this tied to pageserver's auth type? Whether or not the safekeeper
+        // needs a token, and how to generate that token, seems independent to whether
+        // the pageserver requires a token in incoming requests.
+        Ok(if self.env.pageserver.http_auth_type != AuthType::Trust {
             // Generate a token to connect from the pageserver to a safekeeper
             let token = self
                 .env
@@ -270,27 +273,30 @@ impl PageServerNode {
         background_process::stop_process(immediate, "pageserver", &self.pid_file())
     }
 
-    pub fn page_server_psql(&self, sql: &str) -> Vec<postgres::SimpleQueryMessage> {
-        let mut client = self.pg_connection_config.connect_no_tls().unwrap();
-
-        println!("Pageserver query: '{sql}'");
-        client.simple_query(sql).unwrap()
-    }
-
-    pub fn page_server_psql_client(&self) -> result::Result<postgres::Client, postgres::Error> {
-        self.pg_connection_config.connect_no_tls()
-    }
-
-    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> RequestBuilder {
-        let mut builder = self.http_client.request(method, url);
-        if self.env.pageserver.auth_type == AuthType::NeonJWT {
-            builder = builder.bearer_auth(&self.env.pageserver.auth_token)
+    pub fn page_server_psql_client(&self) -> anyhow::Result<postgres::Client> {
+        let mut config = self.pg_connection_config.clone();
+        if self.env.pageserver.pg_auth_type == AuthType::NeonJWT {
+            let token = self
+                .env
+                .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
+            config = config.set_password(Some(token));
         }
-        builder
+        Ok(config.connect_no_tls()?)
+    }
+
+    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> anyhow::Result<RequestBuilder> {
+        let mut builder = self.http_client.request(method, url);
+        if self.env.pageserver.http_auth_type == AuthType::NeonJWT {
+            let token = self
+                .env
+                .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
+            builder = builder.bearer_auth(token)
+        }
+        Ok(builder)
     }
 
     pub fn check_status(&self) -> Result<()> {
-        self.http_request(Method::GET, format!("{}/status", self.http_base_url))
+        self.http_request(Method::GET, format!("{}/status", self.http_base_url))?
             .send()?
             .error_from_body()?;
         Ok(())
@@ -298,7 +304,7 @@ impl PageServerNode {
 
     pub fn tenant_list(&self) -> Result<Vec<TenantInfo>> {
         Ok(self
-            .http_request(Method::GET, format!("{}/tenant", self.http_base_url))
+            .http_request(Method::GET, format!("{}/tenant", self.http_base_url))?
             .send()?
             .error_from_body()?
             .json()?)
@@ -352,11 +358,21 @@ impl PageServerNode {
                 .map(|x| x.parse::<bool>())
                 .transpose()
                 .context("Failed to parse 'trace_read_requests' as bool")?,
+            eviction_policy: settings
+                .get("eviction_policy")
+                .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:?}")
         }
-        self.http_request(Method::POST, format!("{}/tenant", self.http_base_url))
+        self.http_request(Method::POST, format!("{}/tenant", self.http_base_url))?
             .json(&request)
             .send()?
             .error_from_body()?
@@ -373,7 +389,7 @@ impl PageServerNode {
     }
 
     pub fn tenant_config(&self, tenant_id: TenantId, settings: HashMap<&str, &str>) -> Result<()> {
-        self.http_request(Method::PUT, format!("{}/tenant/config", self.http_base_url))
+        self.http_request(Method::PUT, format!("{}/tenant/config", self.http_base_url))?
             .json(&TenantConfigRequest {
                 tenant_id,
                 checkpoint_distance: settings
@@ -424,6 +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()?;
@@ -436,7 +457,7 @@ impl PageServerNode {
             .http_request(
                 Method::GET,
                 format!("{}/tenant/{}/timeline", self.http_base_url, tenant_id),
-            )
+            )?
             .send()?
             .error_from_body()?
             .json()?;
@@ -455,7 +476,7 @@ impl PageServerNode {
         self.http_request(
             Method::POST,
             format!("{}/tenant/{}/timeline", self.http_base_url, tenant_id),
-        )
+        )?
         .json(&TimelineCreateRequest {
             new_timeline_id,
             ancestor_start_lsn,
@@ -492,7 +513,7 @@ impl PageServerNode {
         pg_wal: Option<(Lsn, PathBuf)>,
         pg_version: u32,
     ) -> anyhow::Result<()> {
-        let mut client = self.pg_connection_config.connect_no_tls().unwrap();
+        let mut client = self.page_server_psql_client()?;
 
         // Init base reader
         let (start_lsn, base_tarfile_path) = base;

control_plane/src/safekeeper.rs

@@ -1,7 +1,6 @@
 use std::io::Write;
 use std::path::PathBuf;
 use std::process::Child;
-use std::sync::Arc;
 use std::{io, result};
 
 use anyhow::Context;
@@ -11,7 +10,6 @@ use reqwest::{IntoUrl, Method};
 use thiserror::Error;
 use utils::{http::error::HttpErrorBody, id::NodeId};
 
-use crate::pageserver::PageServerNode;
 use crate::{
     background_process,
     local_env::{LocalEnv, SafekeeperConf},
@@ -65,14 +63,10 @@ pub struct SafekeeperNode {
     pub env: LocalEnv,
     pub http_client: Client,
     pub http_base_url: String,
-
-    pub pageserver: Arc<PageServerNode>,
 }
 
 impl SafekeeperNode {
     pub fn from_env(env: &LocalEnv, conf: &SafekeeperConf) -> SafekeeperNode {
-        let pageserver = Arc::new(PageServerNode::from_env(env));
-
         SafekeeperNode {
             id: conf.id,
             conf: conf.clone(),
@@ -80,7 +74,6 @@ impl SafekeeperNode {
             env: env.clone(),
             http_client: Client::new(),
             http_base_url: format!("http://127.0.0.1:{}/v1", conf.http_port),
-            pageserver,
         }
     }
 
@@ -115,6 +108,10 @@ impl SafekeeperNode {
         let datadir = self.datadir_path();
 
         let id_string = id.to_string();
+        // TODO: add availability_zone to the config.
+        // Right now we just specify any value here and use it to check metrics in tests.
+        let availability_zone = format!("sk-{}", id_string);
+
         let mut args = vec![
             "-D",
             datadir.to_str().with_context(|| {
@@ -126,6 +123,8 @@ impl SafekeeperNode {
             &listen_pg,
             "--listen-http",
             &listen_http,
+            "--availability-zone",
+            &availability_zone,
         ];
         if !self.conf.sync {
             args.push("--no-sync");
@@ -157,7 +156,7 @@ impl SafekeeperNode {
         }
 
         background_process::start_process(
-            &format!("safekeeper {id}"),
+            &format!("safekeeper-{id}"),
             &datadir,
             &self.env.safekeeper_bin(),
             &args,

docker-compose/docker-compose.yml

@@ -160,6 +160,7 @@ services:
     build:
       context: ./compute_wrapper/
       args:
+        - REPOSITORY=${REPOSITORY:-neondatabase}
         - COMPUTE_IMAGE=compute-node-v${PG_VERSION:-14}
         - TAG=${TAG:-latest}
         - http_proxy=$http_proxy

docs/authentication.md

@@ -29,15 +29,22 @@ These components should not have access to the private key and may only get toke
 The key pair is generated once for an installation of compute/pageserver/safekeeper, e.g. by `neon_local init`.
 There is currently no way to rotate the key without bringing down all components.
 
+### Best practices
+
+See [RFC 8725: JSON Web Token Best Current Practices](https://www.rfc-editor.org/rfc/rfc8725)
+
+
 ### Token format
 
-The JWT tokens in Neon use RSA as the algorithm. Example:
+The JWT tokens in Neon use "EdDSA" as the algorithm (defined in [RFC8037](https://www.rfc-editor.org/rfc/rfc8037)).
+
+Example:
 
 Header:
 
 ```
 {
-  "alg": "RS512",     # RS256, RS384, or RS512
+  "alg": "EdDSA",
   "typ": "JWT"
 }
 ```
@@ -68,8 +75,8 @@ Currently also used for connection from any pageserver to any safekeeper.
 CLI generates a key pair during call to `neon_local init` with the following commands:
 
 ```bash
-openssl genrsa -out auth_private_key.pem 2048
-openssl rsa -in auth_private_key.pem -pubout -outform PEM -out auth_public_key.pem
+openssl genpkey -algorithm ed25519 -out auth_private_key.pem
+openssl pkey -in auth_private_key.pem -pubout -out auth_public_key.pem
 ```
 
 Configuration files for all components point to `public_key.pem` for JWT validation.
@@ -99,20 +106,22 @@ Their authentication is just plain PostgreSQL authentication and out of scope fo
 There is no administrative API except those provided by PostgreSQL.
 
 #### Outgoing connections
-Compute connects to Pageserver for getting pages.
-The connection string is configured by the `neon.pageserver_connstring` PostgreSQL GUC, e.g. `postgresql://no_user:$NEON_AUTH_TOKEN@localhost:15028`.
-The environment variable inside the connection string is substituted with
-the JWT token.
+Compute connects to Pageserver for getting pages. The connection string is
+configured by the `neon.pageserver_connstring` PostgreSQL GUC,
+e.g. `postgresql://no_user@localhost:15028`. If the `$NEON_AUTH_TOKEN`
+environment variable is set, it is used as the password for the connection. (The
+pageserver uses JWT tokens for authentication, so the password is really a
+token.)
 
-Compute connects to Safekeepers to write and commit data.
-The token is the same for all safekeepers.
-It's stored in an environment variable, whose name is configured
-by the `neon.safekeeper_token_env` PostgreSQL GUC.
-If the GUC is unset, no token is passed.
+Compute connects to Safekeepers to write and commit data. The list of safekeeper
+addresses is given in the `neon.safekeepers` GUC. The connections to the
+safekeepers take the password from the `$NEON_AUTH_TOKEN` environment
+variable, if set.
 
-Note that both tokens can be (and typically are) the same;
-the scope is the tenant and the token is usually passed through the
-`$NEON_AUTH_TOKEN` environment variable.
+The `compute_ctl` binary that runs before the PostgreSQL server, and launches
+PostgreSQL, also makes a connection to the pageserver. It uses it to fetch the
+initial "base backup" dump, to initialize the PostgreSQL data directory. It also
+uses `$NEON_AUTH_TOKEN` as the password for the connection.
 
 ### Pageserver
 #### Overview
@@ -137,10 +146,12 @@ Each compute should present a token valid for the timeline's tenant.
 Pageserver also has HTTP API: some parts are per-tenant,
 some parts are server-wide, these are different scopes.
 
-The `auth_type` configuration variable in Pageserver's config may have
-either of three values:
+Authentication can be enabled separately for the HTTP mgmt API, and
+for the libpq connections from compute. The `http_auth_type` and
+`pg_auth_type` configuration variables in Pageserver's config may
+have one of these values:
 
-* `Trust` removes all authentication. The outdated `MD5` value does likewise
+* `Trust` removes all authentication.
 * `NeonJWT` enables JWT validation.
    Tokens are validated using the public key which lies in a PEM file
    specified in the `auth_validation_public_key_path` config.

docs/docker.md

@@ -37,9 +37,9 @@ You can specify version of neon cluster using following environment values.
 - PG_VERSION: postgres version for compute (default is 14)
 - TAG: the tag version of [docker image](https://registry.hub.docker.com/r/neondatabase/neon/tags) (default is latest), which is tagged in [CI test](/.github/workflows/build_and_test.yml)
 ```
-$ cd docker-compose/docker-compose.yml
+$ cd docker-compose/
 $ docker-compose down   # remove the conainers if exists
-$ PG_VERSION=15 TAG=2221 docker-compose up --build -d  # You can specify the postgres and image version
+$ PG_VERSION=15 TAG=2937 docker-compose up --build -d  # You can specify the postgres and image version
 Creating network "dockercompose_default" with the default driver
 Creating docker-compose_storage_broker_1       ... done
 (...omit...)

docs/rfcs/022-pageserver-delete-from-s3.md

@@ -0,0 +1,269 @@
+# Deleting pageserver part of tenants data from s3
+
+Created on 08.03.23
+
+## Motivation
+
+Currently we dont delete pageserver part of the data from s3 when project is deleted. (The same is true for safekeepers, but this outside of the scope of this RFC).
+
+This RFC aims to spin a discussion to come to a robust deletion solution that wont put us in into a corner for features like postponed deletion (when we keep data for user to be able to restore a project if it was deleted by accident)
+
+## Summary
+
+TLDR; There are two options, one based on control plane issuing actual delete requests to s3 and the other one that keeps s3 stuff bound to pageserver. Each one has its pros and cons.
+
+The decision is to stick with pageserver centric approach. For motivation see [Decision](#decision).
+
+## Components
+
+pageserver, control-plane
+
+## Requirements
+
+Deletion should successfully finish (eventually) without leaving dangling files in presense of:
+
+- component restarts
+- component outage
+- pageserver loss
+
+## Proposed implementation
+
+Before the options are discussed, note that deletion can be quite long process. For deletion from s3 the obvious choice is [DeleteObjects](https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObjects.html) API call. It allows to batch deletion of up to 1k objects in one API call. So deletion operation linearly depends on number of layer files.
+
+Another design limitation is that there is no cheap `mv` operation available for s3. `mv` from `aws s3 mv` uses `copy(src, dst) + delete(src)`. So `mv`-like operation is not feasible as a building block because it actually amplifies the problem with both duration and resulting cost of the operation.
+
+The case when there are multiple pageservers handling the same tenants is largely out of scope of the RFC. We still consider case with migration from one PS to another, but do not consider case when tenant exists on multiple pageservers for extended period of time. The case with multiple pageservers can be reduced to case with one pageservers by calling detach on all pageservers except the last one, for it actual delete needs to be called.
+
+For simplicity lets look into deleting tenants. Differences in deletion process between tenants and timelines are mentioned in paragraph ["Differences between tenants and timelines"](#differences-between-tenants-and-timelines)
+
+### 1. Pageserver owns deletion machinery
+
+#### The sequence
+
+TLDR; With this approach control plane needs to call delete on a tenant and poll for progress. As much as possible is handled on pageserver. Lets see the sequence.
+
+Happy path:
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant CP as Control Plane
+    participant PS as Pageserver
+    participant S3
+
+    CP->>PS: Delete tenant
+    PS->>S3: Create deleted mark file at <br> /tenant/meta/deleted
+    PS->>PS: Create deleted mark file locally
+    PS->>CP: Accepted
+    PS->>PS: delete local files other than deleted mark
+    loop Delete layers for each timeline
+        PS->>S3: delete(..)
+        CP->>PS: Finished?
+        PS->>CP: False
+    end
+    PS->>S3: Delete mark file
+    PS->>PS: Delete local mark file
+
+    loop Poll for status
+        CP->>PS: Finished?
+        PS->>CP: True or False
+    end
+```
+
+Why two mark files?
+Remote one is needed for cases when pageserver is lost during deletion so other pageserver can learn the deletion from s3 during attach.
+
+Why local mark file is needed?
+
+If we dont have one, we have two choices, delete local data before deleting the remote part or do that after.
+
+If we delete local data before remote then during restart pageserver wont pick up remote tenant at all because nothing is available locally (pageserver looks for remote conuterparts of locally available tenants).
+
+If we delete local data after remote then at the end of the sequence when remote mark file is deleted if pageserver restart happens then the state is the same to situation when pageserver just missing data on remote without knowing the fact that this data is intended to be deleted. In this case the current behavior is upload everything local-only to remote.
+
+Thus we need local record of tenant being deleted as well.
+
+##### Handle pageserver crashes
+
+Lets explore sequences with various crash points.
+
+Pageserver crashes before `deleted` mark file is persisted in s3:
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant CP as Control Plane
+    participant PS as Pageserver
+    participant S3
+
+    CP->>PS: Delete tenant
+    note over PS: Crash point 1.
+    CP->>PS: Retry delete request
+
+    PS->>S3: Create deleted mark file at <br> /tenant/meta/deleted
+    PS->>PS: Create deleted mark file locally
+
+    PS->>CP: Accepted
+
+    PS->>PS: delete local files other than deleted mark
+
+    loop Delete layers for each timeline
+        PS->>S3: delete(..)
+        CP->>PS: Finished?
+        PS->>CP: False
+    end
+    PS->>S3: Delete mark file
+    PS->>PS: Delete local mark file
+
+    CP->>PS: Finished?
+    PS->>CP: True
+```
+
+Pageserver crashed when deleted mark was about to be persisted in s3, before Control Plane gets a response:
+
+```mermaid
+sequenceDiagram
+    autonumber
+    participant CP as Control Plane
+    participant PS as Pageserver
+    participant S3
+
+    CP->>PS: Delete tenant
+    PS->>S3: Create deleted mark file at <br> /tenant/meta/deleted
+
+    note over PS: Crash point 2.
+    note over PS: During startup we reconcile <br> with remote and see <br> whether the remote mark exists
+    alt Remote mark exists
+        PS->>PS: create local mark if its missing
+        PS->>PS: delete local files other than deleted mark
+        loop Delete layers for each timeline
+            PS->>S3: delete(..)
+        end
+
+        note over CP: Eventually console should <br> retry delete request
+
+        CP->>PS: Retry delete tenant
+        PS->>CP: Not modified
+    else Mark is missing
+        note over PS: Continue to operate the tenant as if deletion didnt happen
+
+        note over CP: Eventually console should <br> retry delete request
+
+        CP->>PS: Retry delete tenant
+        PS->>S3: Create deleted mark file at <br> /tenant/meta/deleted
+        PS->>CP: Delete tenant
+    end
+
+    PS->>PS: Continue with layer file deletions
+    loop Delete layers for each timeline
+        PS->>S3: delete(..)
+        CP->>PS: Finished?
+        PS->>CP: False
+    end
+
+    PS->>S3: Delete mark file
+    PS->>PS: Delete local mark file
+
+    CP->>PS: Finished?
+    PS->>CP: True
+```
+
+Similar sequence applies when both local and remote marks were persisted but Control Plane still didnt receive a response.
+
+If pageserver crashes after both mark files were deleted then it will reply to control plane status poll request with 404 which should be treated by control plane as success.
+
+The same applies if pageserver crashes in the end, when remote mark is deleted but before local one gets deleted. In this case on restart pageserver moves forward with deletion of local mark and Control Plane will receive 404.
+
+##### Differences between tenants and timelines
+
+For timeline the sequence is the same with the following differences:
+
+- remote delete mark file can be replaced with a boolean "deleted" flag in index_part.json
+- local deletion mark is not needed, because whole tenant is kept locally so situation described in motivation for local mark is impossible
+
+##### Handle pageserver loss
+
+If pageseserver is lost then the deleted tenant should be attached to different pageserver and delete request needs to be retried against new pageserver. Then attach logic is shared with one described for pageserver restarts (local deletion mark wont be available so needs to be created).
+
+##### Restrictions for tenant that is in progress of being deleted
+
+I propose to add another state to tenant/timeline - PendingDelete. This state shouldnt allow executing any operations aside from polling the deletion status.
+
+#### Summary
+
+Pros:
+
+- Storage is not dependent on control plane. Storage can be restarted even if control plane is not working.
+- Allows for easier dogfooding, console can use Neon backed database as primary operational data store. If storage depends on control plane and control plane depends on storage we're stuck.
+- No need to share inner s3 workings with control plane. Pageserver presents api contract and S3 paths are not part of this contract.
+- No need to pass list of alive timelines to attach call. This will be solved by pageserver observing deleted flag. See
+
+Cons:
+
+- Logic is a tricky, needs good testing
+- Anything else?
+
+### 2. Control plane owns deletion machinery
+
+In this case the only action performed on pageserver is removal of local files.
+
+Everything else is done by control plane. The steps are as follows:
+
+1. Control plane marks tenant as "delete pending" in its database
+2. It lists the s3 for all the files and repeatedly calls delete until nothing is left behind
+3. When no files are left marks deletion as completed
+
+In case of restart it selects all tenants marked as "delete pending" and continues the deletion.
+
+For tenants it is simple. For timelines there are caveats.
+
+Assume that the same workflow is used for timelines.
+
+If a tenant gets relocated during timeline deletion the attach call with its current logic will pick up deleted timeline in its half deleted state.
+
+Available options:
+
+- require list of alive timelines to be passed to attach call
+- use the same schema with flag in index_part.json (again part of the caveats around pageserver restart applies). In this case nothing stops pageserver from implementing deletion inside if we already have these deletion marks.
+
+With first option the following problem becomes apparent:
+
+Who is the source of truth regarding timeline liveness?
+
+Imagine:
+PS1 fails.
+PS2 gets assigned the tenant.
+New branch gets created
+PS1 starts up (is it possible or we just recycle it?)
+PS1 is unaware of the new branch. It can either fall back to s3 ls, or ask control plane.
+
+So here comes the dependency of storage on control plane. During restart storage needs to know which timelines are valid for operation. If there is nothing on s3 that can answer that question storage neeeds to ask control plane.
+
+### Summary
+
+Cons:
+
+- Potential thundering herd-like problem during storage restart (requests to control plane)
+- Potential increase in storage startup time (additional request to control plane)
+- Storage startup starts to depend on console
+- Erroneous attach call can attach tenant in half deleted state
+
+Pros:
+
+- Easier to reason about if you dont have to account for pageserver restarts
+
+### Extra notes
+
+There was a concern that having deletion code in pageserver is a littlebit scary, but we need to have this code somewhere. So to me it is equally scary to have that in whatever place it ends up at.
+
+Delayed deletion can be done with both approaches. As discussed with Anna (@stepashka) this is only relevant for tenants (projects) not for timelines. For first approach detach can be called immediately and deletion can be done later with attach + delete. With second approach control plane needs to start the deletion whenever necessary.
+
+## Decision
+
+After discussion in comments I see that we settled on two options (though a bit different from ones described in rfc). First one is the same - pageserver owns as much as possible. The second option is that pageserver owns markers thing, but actual deletion happens in control plane by repeatedly calling ls + delete.
+
+To my mind the only benefit of the latter approach is possible code reuse between safekeepers and pageservers. Otherwise poking around integrating s3 library into control plane, configuring shared knowledge abouth paths in s3 - are the downsides. Another downside of relying on control plane is the testing process. Control plane resides in different repository so it is quite hard to test pageserver related changes there. e2e test suite there doesnt support shutting down pageservers, which are separate docker containers there instead of just processes.
+
+With pageserver owning everything we still give the retry logic to control plane but its easier to duplicate if needed compared to sharing inner s3 workings. We will have needed tests for retry logic in neon repo.
+
+So the decision is to proceed with pageserver centric approach.

docs/sourcetree.md

@@ -129,13 +129,12 @@ Run `poetry shell` to activate the virtual environment.
 Alternatively, use `poetry run` to run a single command in the venv, e.g. `poetry run pytest`.
 
 ### Obligatory checks
-We force code formatting via `black`, `isort` and type hints via `mypy`.
+We force code formatting via `black`, `ruff`, and type hints via `mypy`.
 Run the following commands in the repository's root (next to `pyproject.toml`):
 
 ```bash
-poetry run isort .  # Imports are reformatted
 poetry run black .  # All code is reformatted
-poetry run flake8 .  # Python linter
+poetry run ruff .  # Python linter
 poetry run mypy .  # Ensure there are no typing errors
 ```
 

libs/pageserver_api/src/models.rs

@@ -115,6 +115,12 @@ pub struct TenantCreateRequest {
     pub lagging_wal_timeout: Option<String>,
     pub max_lsn_wal_lag: Option<NonZeroU64>,
     pub trace_read_requests: Option<bool>,
+    // We defer the parsing of the eviction_policy field to the request handler.
+    // Otherwise we'd have to move the types for eviction policy into this package.
+    // 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]
@@ -160,6 +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 {
@@ -180,6 +187,7 @@ impl TenantConfigRequest {
             max_lsn_wal_lag: None,
             trace_read_requests: None,
             eviction_policy: None,
+            min_resident_size_override: None,
         }
     }
 }
@@ -341,7 +349,7 @@ pub enum InMemoryLayerInfo {
 pub enum HistoricLayerInfo {
     Delta {
         layer_file_name: String,
-        layer_file_size: Option<u64>,
+        layer_file_size: u64,
 
         #[serde_as(as = "DisplayFromStr")]
         lsn_start: Lsn,
@@ -352,7 +360,7 @@ pub enum HistoricLayerInfo {
     },
     Image {
         layer_file_name: String,
-        layer_file_size: Option<u64>,
+        layer_file_size: u64,
 
         #[serde_as(as = "DisplayFromStr")]
         lsn_start: Lsn,

libs/postgres_backend/src/lib.rs

@@ -767,7 +767,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
 
         let err_to_send_and_errcode = match &end {
             ServerInitiated(_) => Some((end.to_string(), SQLSTATE_SUCCESSFUL_COMPLETION)),
-            Other(_) => Some((end.to_string(), SQLSTATE_INTERNAL_ERROR)),
+            Other(_) => Some((format!("{end:#}"), SQLSTATE_INTERNAL_ERROR)),
             // Note: CopyFail in duplex copy is somewhat unexpected (at least to
             // PG walsender; evidently and per my docs reading client should
             // finish it with CopyDone). It is not a problem to recover from it

libs/postgres_ffi/build.rs

@@ -63,10 +63,7 @@ fn main() -> anyhow::Result<()> {
             pg_install_dir_versioned = cwd.join("..").join("..").join(pg_install_dir_versioned);
         }
 
-        let pg_config_bin = pg_install_dir_versioned
-            .join(pg_version)
-            .join("bin")
-            .join("pg_config");
+        let pg_config_bin = pg_install_dir_versioned.join("bin").join("pg_config");
         let inc_server_path: String = if pg_config_bin.exists() {
             let output = Command::new(pg_config_bin)
                 .arg("--includedir-server")

libs/pq_proto/src/lib.rs

@@ -265,11 +265,9 @@ impl FeMessage {
             b'c' => Ok(Some(FeMessage::CopyDone)),
             b'f' => Ok(Some(FeMessage::CopyFail)),
             b'p' => Ok(Some(FeMessage::PasswordMessage(msg))),
-            tag => {
-                return Err(ProtocolError::Protocol(format!(
-                    "unknown message tag: {tag},'{msg:?}'"
-                )))
-            }
+            tag => Err(ProtocolError::Protocol(format!(
+                "unknown message tag: {tag},'{msg:?}'"
+            ))),
         }
     }
 }
@@ -938,35 +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:
@@ -976,24 +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;
@@ -1003,9 +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();
@@ -1018,39 +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
     }
 }
@@ -1061,33 +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");
@@ -1095,7 +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/auth.rs

@@ -1,7 +1,4 @@
 // For details about authentication see docs/authentication.md
-//
-// TODO: use ed25519 keys
-// Relevant issue: https://github.com/Keats/jsonwebtoken/issues/162
 
 use serde;
 use std::fs;
@@ -9,26 +6,15 @@ use std::path::Path;
 
 use anyhow::Result;
 use jsonwebtoken::{
-    decode, encode, Algorithm, Algorithm::*, DecodingKey, EncodingKey, Header, TokenData,
-    Validation,
+    decode, encode, Algorithm, DecodingKey, EncodingKey, Header, TokenData, Validation,
 };
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 
 use crate::id::TenantId;
 
-/// Algorithms accepted during validation.
-///
-/// Accept all RSA-based algorithms. We pass this list to jsonwebtoken::decode,
-/// which checks that the algorithm in the token is one of these.
-///
-/// XXX: It also fails the validation if there are any algorithms in this list that belong
-/// to different family than the token's algorithm. In other words, we can *not* list any
-/// non-RSA algorithms here, or the validation always fails with InvalidAlgorithm error.
-const ACCEPTED_ALGORITHMS: &[Algorithm] = &[RS256, RS384, RS512];
-
-/// Algorithm to use when generating a new token in [`encode_from_key_file`]
-const ENCODE_ALGORITHM: Algorithm = Algorithm::RS256;
+/// Algorithm to use. We require EdDSA.
+const STORAGE_TOKEN_ALGORITHM: Algorithm = Algorithm::EdDSA;
 
 #[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
 #[serde(rename_all = "lowercase")]
@@ -69,7 +55,7 @@ pub struct JwtAuth {
 impl JwtAuth {
     pub fn new(decoding_key: DecodingKey) -> Self {
         let mut validation = Validation::default();
-        validation.algorithms = ACCEPTED_ALGORITHMS.into();
+        validation.algorithms = vec![STORAGE_TOKEN_ALGORITHM];
         // The default 'required_spec_claims' is 'exp'. But we don't want to require
         // expiration.
         validation.required_spec_claims = [].into();
@@ -81,7 +67,7 @@ impl JwtAuth {
 
     pub fn from_key_path(key_path: &Path) -> Result<Self> {
         let public_key = fs::read(key_path)?;
-        Ok(Self::new(DecodingKey::from_rsa_pem(&public_key)?))
+        Ok(Self::new(DecodingKey::from_ed_pem(&public_key)?))
     }
 
     pub fn decode(&self, token: &str) -> Result<TokenData<Claims>> {
@@ -99,8 +85,8 @@ impl std::fmt::Debug for JwtAuth {
 
 // this function is used only for testing purposes in CLI e g generate tokens during init
 pub fn encode_from_key_file(claims: &Claims, key_data: &[u8]) -> Result<String> {
-    let key = EncodingKey::from_rsa_pem(key_data)?;
-    Ok(encode(&Header::new(ENCODE_ALGORITHM), claims, &key)?)
+    let key = EncodingKey::from_ed_pem(key_data)?;
+    Ok(encode(&Header::new(STORAGE_TOKEN_ALGORITHM), claims, &key)?)
 }
 
 #[cfg(test)]
@@ -108,49 +94,19 @@ mod tests {
     use super::*;
     use std::str::FromStr;
 
-    // generated with:
+    // Generated with:
     //
-    // openssl genpkey -algorithm rsa -out storage-auth-priv.pem
-    // openssl pkey -in storage-auth-priv.pem -pubout -out storage-auth-pub.pem
-    const TEST_PUB_KEY_RSA: &[u8] = br#"
+    // openssl genpkey -algorithm ed25519 -out ed25519-priv.pem
+    // openssl pkey -in ed25519-priv.pem -pubout -out ed25519-pub.pem
+    const TEST_PUB_KEY_ED25519: &[u8] = br#"
 -----BEGIN PUBLIC KEY-----
-MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAy6OZ+/kQXcueVJA/KTzO
-v4ljxylc/Kcb0sXWuXg1GB8k3nDA1gK66LFYToH0aTnqrnqG32Vu6wrhwuvqsZA7
-jQvP0ZePAbWhpEqho7EpNunDPcxZ/XDy5TQlB1P58F9I3lkJXDC+DsHYLuuzwhAv
-vo2MtWRdYlVHblCVLyZtANHhUMp2HUhgjHnJh5UrLIKOl4doCBxkM3rK0wjKsNCt
-M92PCR6S9rvYzldfeAYFNppBkEQrXt2CgUqZ4KaS4LXtjTRUJxljijA4HWffhxsr
-euRu3ufq8kVqie7fum0rdZZSkONmce0V0LesQ4aE2jB+2Sn48h6jb4dLXGWdq8TV
-wQIDAQAB
+MCowBQYDK2VwAyEARYwaNBayR+eGI0iXB4s3QxE3Nl2g1iWbr6KtLWeVD/w=
 -----END PUBLIC KEY-----
 "#;
-    const TEST_PRIV_KEY_RSA: &[u8] = br#"
+
+    const TEST_PRIV_KEY_ED25519: &[u8] = br#"
 -----BEGIN PRIVATE KEY-----
-MIIEvQIBADANBgkqhkiG9w0BAQEFAASCBKcwggSjAgEAAoIBAQDLo5n7+RBdy55U
-kD8pPM6/iWPHKVz8pxvSxda5eDUYHyTecMDWArrosVhOgfRpOequeobfZW7rCuHC
-6+qxkDuNC8/Rl48BtaGkSqGjsSk26cM9zFn9cPLlNCUHU/nwX0jeWQlcML4Owdgu
-67PCEC++jYy1ZF1iVUduUJUvJm0A0eFQynYdSGCMecmHlSssgo6Xh2gIHGQzesrT
-CMqw0K0z3Y8JHpL2u9jOV194BgU2mkGQRCte3YKBSpngppLgte2NNFQnGWOKMDgd
-Z9+HGyt65G7e5+ryRWqJ7t+6bSt1llKQ42Zx7RXQt6xDhoTaMH7ZKfjyHqNvh0tc
-ZZ2rxNXBAgMBAAECggEAVz3u4Wlx3o02dsoZlSQs+xf0PEX3RXKeU+1YMbtTG9Nz
-6yxpIQaoZrpbt76rJE2gwkFR+PEu1NmjoOuLb6j4KlQuI4AHz1auOoGSwFtM6e66
-K4aZ4x95oEJ3vqz2fkmEIWYJwYpMUmwvnuJx76kZm0xvROMLsu4QHS2+zCVtO5Tr
-hvS05IMVuZ2TdQBZw0+JaFdwXbgDjQnQGY5n9MoTWSx1a4s/FF4Eby65BbDutcpn
-Vt3jQAOmO1X2kbPeWSGuPJRzyUs7Kg8qfeglBIR3ppGP3vPYAdWX+ho00bmsVkSp
-Q8vjul6C3WiM+kjwDxotHSDgbl/xldAl7OqPh0bfAQKBgQDnycXuq14Vg8nZvyn9
-rTnvucO8RBz5P6G+FZ+44cAS2x79+85onARmMnm+9MKYLSMo8fOvsK034NDI68XM
-04QQ/vlfouvFklMTGJIurgEImTZbGCmlMYCvFyIxaEWixon8OpeI4rFe4Hmbiijh
-PxhxWg221AwvBS2sco8J/ylEkQKBgQDg6Rh2QYb/j0Wou1rJPbuy3NhHofd5Rq35
-4YV3f2lfVYcPrgRhwe3T9SVII7Dx8LfwzsX5TAlf48ESlI3Dzv40uOCDM+xdtBRI
-r96SfSm+jup6gsXU3AsdNkrRK3HoOG9Z/TkrUp213QAIlVnvIx65l4ckFMlpnPJ0
-lo1LDXZWMQKBgFArzjZ7N5OhfdO+9zszC3MLgdRAivT7OWqR+CjujIz5FYMr8Xzl
-WfAvTUTrS9Nu6VZkObFvHrrRG+YjBsuN7YQjbQXTSFGSBwH34bgbn2fl9pMTjHQC
-50uoaL9GHa/rlBaV/YvvPQJgCi/uXa1rMX0jdNLkDULGO8IF7cu7Yf7BAoGBAIUU
-J29BkpmAst0GDs/ogTlyR18LTR0rXyHt+UUd1MGeH859TwZw80JpWWf4BmkB4DTS
-hH3gKePdJY7S65ci0XNsuRupC4DeXuorde0DtkGU2tUmr9wlX0Ynq9lcdYfMbMa4
-eK1TsxG69JwfkxlWlIWITWRiEFM3lJa7xlrUWmLhAoGAFpKWF/hn4zYg3seU9gai
-EYHKSbhxA4mRb+F0/9IlCBPMCqFrL5yftUsYIh2XFKn8+QhO97Nmk8wJSK6TzQ5t
-ZaSRmgySrUUhx4nZ/MgqWCFv8VUbLM5MBzwxPKhXkSTfR4z2vLYLJwVY7Tb4kZtp
-8ismApXVGHpOCstzikV9W7k=
+MC4CAQAwBQYDK2VwBCIEID/Drmc1AA6U/znNRWpF3zEGegOATQxfkdWxitcOMsIH
 -----END PRIVATE KEY-----
 "#;
 
@@ -161,8 +117,7 @@ ZaSRmgySrUUhx4nZ/MgqWCFv8VUbLM5MBzwxPKhXkSTfR4z2vLYLJwVY7Tb4kZtp
             scope: Scope::Tenant,
         };
 
-        // Here are tokens containing the following payload, signed using TEST_PRIV_KEY_RSA
-        // using RS512, RS384 and RS256 algorithms:
+        // A test token containing the following payload, signed using TEST_PRIV_KEY_ED25519:
         //
         // ```
         // {
@@ -174,21 +129,13 @@ ZaSRmgySrUUhx4nZ/MgqWCFv8VUbLM5MBzwxPKhXkSTfR4z2vLYLJwVY7Tb4kZtp
         // }
         // ```
         //
-        // These were encoded with the online debugger at https://jwt.io
-        //
-        let encoded_rs512 = "eyJhbGciOiJSUzUxMiIsInR5cCI6IkpXVCJ9.eyJzY29wZSI6InRlbmFudCIsInRlbmFudF9pZCI6IjNkMWY3NTk1YjQ2ODIzMDMwNGUwYjczY2VjYmNiMDgxIiwiaXNzIjoibmVvbi5jb250cm9scGxhbmUiLCJleHAiOjE3MDkyMDA4NzksImlhdCI6MTY3ODQ0MjQ3OX0.QmqfteDQmDGoxQ5EFkasbt35Lx0W0Nh63muQnYZvFq93DSh4ZbOG9Mc4yaiXZoiS5HgeKtFKv3mbWkDqjz3En06aY17hWwguBtAsGASX48lYeCPADYGlGAuaWnOnVRwe3iiOC7tvPFvwX_45S84X73sNUXyUiXv6nLdcDqVXudtNrGST_DnZDnjuUJX11w7sebtKqQQ8l9-iGHiXOl5yevpMCoB1OcTWcT6DfDtffoNuMHDC3fyhmEGG5oKAt1qBybqAIiyC9-UBAowRZXhdfxrzUl-I9jzKWvk85c5ulhVRwbPeP6TTTlPKwFzBNHg1i2U-1GONew5osQ3aoptwsA";
+        let encoded_eddsa = "eyJhbGciOiJFZERTQSIsInR5cCI6IkpXVCJ9.eyJzY29wZSI6InRlbmFudCIsInRlbmFudF9pZCI6IjNkMWY3NTk1YjQ2ODIzMDMwNGUwYjczY2VjYmNiMDgxIiwiaXNzIjoibmVvbi5jb250cm9scGxhbmUiLCJleHAiOjE3MDkyMDA4NzksImlhdCI6MTY3ODQ0MjQ3OX0.U3eA8j-uU-JnhzeO3EDHRuXLwkAUFCPxtGHEgw6p7Ccc3YRbFs2tmCdbD9PZEXP-XsxSeBQi1FY0YPcT3NXADw";
 
-        let encoded_rs384 = "eyJhbGciOiJSUzM4NCIsInR5cCI6IkpXVCJ9.eyJzY29wZSI6InRlbmFudCIsInRlbmFudF9pZCI6IjNkMWY3NTk1YjQ2ODIzMDMwNGUwYjczY2VjYmNiMDgxIiwiaXNzIjoibmVvbi5jb250cm9scGxhbmUiLCJleHAiOjE3MDkyMDA4NzksImlhdCI6MTY3ODQ0MjQ3OX0.qqk4nkxKzOJP38c_g57_w_SfdQVmCsDT_bsLmdFj_N6LIB22gr6U6_P_5mvk3pIAsp0VCTDwPrCU908TxqjibEkwvQoJwbogHamSGHpD7eJBxGblSnA-Nr3MlEMxpFtec8QokSm6C5mH7DoBYjB2xzeOlxAmpR2GAzInKiMkU4kZ_OcqqrmVcMXY_6VnbxZWMekuw56zE1-PP_qNF1HvYOH-P08ONP8qdo5UPtBG7QBEFlCqZXJZCFihQaI4Vzil9rDuZGCm3I7xQJ8-yh1PX3BTbGo8EzqLdRyBeTpr08UTuRbp_MJDWevHpP3afvJetAItqZXIoZQrbJjcByHqKw";
+        // Check it can be validated with the public key
+        let auth = JwtAuth::new(DecodingKey::from_ed_pem(TEST_PUB_KEY_ED25519)?);
+        let claims_from_token = auth.decode(encoded_eddsa)?.claims;
+        assert_eq!(claims_from_token, expected_claims);
 
-        let encoded_rs256 = "eyJhbGciOiJSUzI1NiIsInR5cCI6IkpXVCJ9.eyJzY29wZSI6InRlbmFudCIsInRlbmFudF9pZCI6IjNkMWY3NTk1YjQ2ODIzMDMwNGUwYjczY2VjYmNiMDgxIiwiaXNzIjoibmVvbi5jb250cm9scGxhbmUiLCJleHAiOjE3MDkyMDA4NzksImlhdCI6MTY3ODQ0MjQ3OX0.dF2N9KXG8ftFKHYbd5jQtXMQqv0Ej8FISGp1b_dmqOCotXj5S1y2AWjwyB_EXHM77JXfbEoJPAPrFFBNfd8cWtkCSTvpxWoHaecGzegDFGv5ZSc5AECFV1Daahc3PI3jii9wEiGkFOiwiBNfZ5INomOAsV--XXxlqIwKbTcgSYI7lrOTfecXAbAHiMKQlQYiIBSGnytRCgafhRkyGzPAL8ismthFJ9RHfeejyskht-9GbVHURw02bUyijuHEulpf9eEY3ZiB28de6jnCdU7ftIYaUMaYWt0nZQGkzxKPSfSLZNy14DTOYLDS04DVstWQPqnCUW_ojg0wJETOOfo9Zw";
-
-        // Check that RS512, RS384 and RS256 tokens can all be validated
-        let auth = JwtAuth::new(DecodingKey::from_rsa_pem(TEST_PUB_KEY_RSA)?);
-
-        for encoded in [encoded_rs512, encoded_rs384, encoded_rs256] {
-            let claims_from_token = auth.decode(encoded)?.claims;
-            assert_eq!(claims_from_token, expected_claims);
-        }
         Ok(())
     }
 
@@ -199,10 +146,10 @@ ZaSRmgySrUUhx4nZ/MgqWCFv8VUbLM5MBzwxPKhXkSTfR4z2vLYLJwVY7Tb4kZtp
             scope: Scope::Tenant,
         };
 
-        let encoded = encode_from_key_file(&claims, TEST_PRIV_KEY_RSA)?;
+        let encoded = encode_from_key_file(&claims, TEST_PRIV_KEY_ED25519)?;
 
         // decode it back
-        let auth = JwtAuth::new(DecodingKey::from_rsa_pem(TEST_PUB_KEY_RSA)?);
+        let auth = JwtAuth::new(DecodingKey::from_ed_pem(TEST_PUB_KEY_ED25519)?);
         let decoded = auth.decode(&encoded)?;
 
         assert_eq!(decoded.claims, claims);

libs/utils/src/fs_ext.rs

@@ -11,7 +11,7 @@ where
     P: AsRef<Path>,
 {
     fn is_empty_dir(&self) -> io::Result<bool> {
-        Ok(fs::read_dir(self)?.into_iter().next().is_none())
+        Ok(fs::read_dir(self)?.next().is_none())
     }
 }
 

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 +1,83 @@
+//! 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 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,
 };
 
@@ -263,22 +262,35 @@ 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))?;
 
     // Initialize authentication for incoming connections
-    let auth = match &conf.auth_type {
-        AuthType::Trust => None,
-        AuthType::NeonJWT => {
-            // unwrap is ok because check is performed when creating config, so path is set and file exists
-            let key_path = conf.auth_validation_public_key_path.as_ref().unwrap();
-            Some(JwtAuth::from_key_path(key_path)?.into())
-        }
-    };
-    info!("Using auth: {:#?}", conf.auth_type);
+    let http_auth;
+    let pg_auth;
+    if conf.http_auth_type == AuthType::NeonJWT || conf.pg_auth_type == AuthType::NeonJWT {
+        // unwrap is ok because check is performed when creating config, so path is set and file exists
+        let key_path = conf.auth_validation_public_key_path.as_ref().unwrap();
+        info!(
+            "Loading public key for verifying JWT tokens from {:#?}",
+            key_path
+        );
+        let auth: Arc<JwtAuth> = Arc::new(JwtAuth::from_key_path(key_path)?);
+
+        http_auth = match &conf.http_auth_type {
+            AuthType::Trust => None,
+            AuthType::NeonJWT => Some(auth.clone()),
+        };
+        pg_auth = match &conf.pg_auth_type {
+            AuthType::Trust => None,
+            AuthType::NeonJWT => Some(auth),
+        };
+    } else {
+        http_auth = None;
+        pg_auth = None;
+    }
+    info!("Using auth for http API: {:#?}", conf.http_auth_type);
+    info!("Using auth for pg connections: {:#?}", conf.pg_auth_type);
 
     match var("NEON_AUTH_TOKEN") {
         Ok(v) => {
@@ -303,14 +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, auth.clone(), 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)
@@ -382,9 +414,9 @@ fn start_pageserver(
             async move {
                 page_service::libpq_listener_main(
                     conf,
-                    auth,
+                    pg_auth,
                     pageserver_listener,
-                    conf.auth_type,
+                    conf.pg_auth_type,
                     libpq_ctx,
                 )
                 .await
@@ -393,7 +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};
@@ -61,6 +62,7 @@ pub mod defaults {
     pub const DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL: &str = "1 hour";
     pub const DEFAULT_METRIC_COLLECTION_ENDPOINT: Option<reqwest::Url> = None;
     pub const DEFAULT_SYNTHETIC_SIZE_CALCULATION_INTERVAL: &str = "10 min";
+    pub const DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD: &str = "24 hour";
 
     ///
     /// Default built-in configuration file.
@@ -89,6 +91,10 @@ pub mod defaults {
 #cached_metric_collection_interval = '{DEFAULT_CACHED_METRIC_COLLECTION_INTERVAL}'
 #synthetic_size_calculation_interval = '{DEFAULT_SYNTHETIC_SIZE_CALCULATION_INTERVAL}'
 
+#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}
@@ -101,6 +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]
 
 "###
@@ -118,6 +126,9 @@ pub struct PageServerConf {
     /// Example (default): 127.0.0.1:9898
     pub listen_http_addr: String,
 
+    /// Current availability zone. Used for traffic metrics.
+    pub availability_zone: Option<String>,
+
     // Timeout when waiting for WAL receiver to catch up to an LSN given in a GetPage@LSN call.
     pub wait_lsn_timeout: Duration,
     // How long to wait for WAL redo to complete.
@@ -138,9 +149,15 @@ pub struct PageServerConf {
 
     pub pg_distrib_dir: PathBuf,
 
-    pub auth_type: AuthType,
-
+    // Authentication
+    /// authentication method for the HTTP mgmt API
+    pub http_auth_type: AuthType,
+    /// authentication method for libpq connections from compute
+    pub pg_auth_type: AuthType,
+    /// Path to a file containing public key for verifying JWT tokens.
+    /// Used for both mgmt and compute auth, if enabled.
     pub auth_validation_public_key_path: Option<PathBuf>,
+
     pub remote_storage_config: Option<RemoteStorageConfig>,
 
     pub default_tenant_conf: TenantConf,
@@ -153,6 +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,
@@ -161,6 +182,11 @@ pub struct PageServerConf {
     pub metric_collection_endpoint: Option<Url>,
     pub synthetic_size_calculation_interval: Duration,
 
+    // 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,
@@ -196,6 +222,8 @@ struct PageServerConfigBuilder {
 
     listen_http_addr: BuilderValue<String>,
 
+    availability_zone: BuilderValue<Option<String>>,
+
     wait_lsn_timeout: BuilderValue<Duration>,
     wal_redo_timeout: BuilderValue<Duration>,
 
@@ -208,7 +236,8 @@ struct PageServerConfigBuilder {
 
     pg_distrib_dir: BuilderValue<PathBuf>,
 
-    auth_type: BuilderValue<AuthType>,
+    http_auth_type: BuilderValue<AuthType>,
+    pg_auth_type: BuilderValue<AuthType>,
 
     //
     auth_validation_public_key_path: BuilderValue<Option<PathBuf>>,
@@ -221,13 +250,17 @@ 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>,
     metric_collection_endpoint: BuilderValue<Option<Url>>,
     synthetic_size_calculation_interval: BuilderValue<Duration>,
 
+    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>,
@@ -240,6 +273,7 @@ impl Default for PageServerConfigBuilder {
         Self {
             listen_pg_addr: Set(DEFAULT_PG_LISTEN_ADDR.to_string()),
             listen_http_addr: Set(DEFAULT_HTTP_LISTEN_ADDR.to_string()),
+            availability_zone: Set(None),
             wait_lsn_timeout: Set(humantime::parse_duration(DEFAULT_WAIT_LSN_TIMEOUT)
                 .expect("cannot parse default wait lsn timeout")),
             wal_redo_timeout: Set(humantime::parse_duration(DEFAULT_WAL_REDO_TIMEOUT)
@@ -251,7 +285,8 @@ impl Default for PageServerConfigBuilder {
             pg_distrib_dir: Set(env::current_dir()
                 .expect("cannot access current directory")
                 .join("pg_install")),
-            auth_type: Set(AuthType::Trust),
+            http_auth_type: Set(AuthType::Trust),
+            pg_auth_type: Set(AuthType::Trust),
             auth_validation_public_key_path: Set(None),
             remote_storage_config: Set(None),
             id: NotSet,
@@ -264,7 +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,
             )
@@ -279,6 +316,13 @@ impl Default for PageServerConfigBuilder {
             .expect("cannot parse default synthetic size calculation interval")),
             metric_collection_endpoint: Set(DEFAULT_METRIC_COLLECTION_ENDPOINT),
 
+            evictions_low_residence_duration_metric_threshold: Set(humantime::parse_duration(
+                DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD,
+            )
+            .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),
@@ -295,6 +339,10 @@ impl PageServerConfigBuilder {
         self.listen_http_addr = BuilderValue::Set(listen_http_addr)
     }
 
+    pub fn availability_zone(&mut self, availability_zone: Option<String>) {
+        self.availability_zone = BuilderValue::Set(availability_zone)
+    }
+
     pub fn wait_lsn_timeout(&mut self, wait_lsn_timeout: Duration) {
         self.wait_lsn_timeout = BuilderValue::Set(wait_lsn_timeout)
     }
@@ -323,8 +371,12 @@ impl PageServerConfigBuilder {
         self.pg_distrib_dir = BuilderValue::Set(pg_distrib_dir)
     }
 
-    pub fn auth_type(&mut self, auth_type: AuthType) {
-        self.auth_type = BuilderValue::Set(auth_type)
+    pub fn http_auth_type(&mut self, auth_type: AuthType) {
+        self.http_auth_type = BuilderValue::Set(auth_type)
+    }
+
+    pub fn pg_auth_type(&mut self, auth_type: AuthType) {
+        self.pg_auth_type = BuilderValue::Set(auth_type)
     }
 
     pub fn auth_validation_public_key_path(
@@ -354,7 +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);
     }
 
@@ -386,6 +438,14 @@ impl PageServerConfigBuilder {
         self.test_remote_failures = BuilderValue::Set(fail_first);
     }
 
+    pub fn evictions_low_residence_duration_metric_threshold(&mut self, value: Duration) {
+        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,
@@ -395,6 +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
@@ -402,6 +467,9 @@ impl PageServerConfigBuilder {
             listen_http_addr: self
                 .listen_http_addr
                 .ok_or(anyhow!("missing listen_http_addr"))?,
+            availability_zone: self
+                .availability_zone
+                .ok_or(anyhow!("missing availability_zone"))?,
             wait_lsn_timeout: self
                 .wait_lsn_timeout
                 .ok_or(anyhow!("missing wait_lsn_timeout"))?,
@@ -419,7 +487,10 @@ impl PageServerConfigBuilder {
             pg_distrib_dir: self
                 .pg_distrib_dir
                 .ok_or(anyhow!("missing pg_distrib_dir"))?,
-            auth_type: self.auth_type.ok_or(anyhow!("missing auth_type"))?,
+            http_auth_type: self
+                .http_auth_type
+                .ok_or(anyhow!("missing http_auth_type"))?,
+            pg_auth_type: self.pg_auth_type.ok_or(anyhow!("missing pg_auth_type"))?,
             auth_validation_public_key_path: self
                 .auth_validation_public_key_path
                 .ok_or(anyhow!("missing auth_validation_public_key_path"))?,
@@ -436,11 +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"))?,
@@ -453,6 +525,14 @@ impl PageServerConfigBuilder {
             synthetic_size_calculation_interval: self
                 .synthetic_size_calculation_interval
                 .ok_or(anyhow!("missing synthetic_size_calculation_interval"))?,
+            evictions_low_residence_duration_metric_threshold: self
+                .evictions_low_residence_duration_metric_threshold
+                .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"))?,
@@ -599,6 +679,7 @@ impl PageServerConf {
             match key {
                 "listen_pg_addr" => builder.listen_pg_addr(parse_toml_string(key, item)?),
                 "listen_http_addr" => builder.listen_http_addr(parse_toml_string(key, item)?),
+                "availability_zone" => builder.availability_zone(Some(parse_toml_string(key, item)?)),
                 "wait_lsn_timeout" => builder.wait_lsn_timeout(parse_toml_duration(key, item)?),
                 "wal_redo_timeout" => builder.wal_redo_timeout(parse_toml_duration(key, item)?),
                 "initial_superuser_name" => builder.superuser(parse_toml_string(key, item)?),
@@ -612,7 +693,8 @@ impl PageServerConf {
                 "auth_validation_public_key_path" => builder.auth_validation_public_key_path(Some(
                     PathBuf::from(parse_toml_string(key, item)?),
                 )),
-                "auth_type" => builder.auth_type(parse_toml_from_str(key, item)?),
+                "http_auth_type" => builder.http_auth_type(parse_toml_from_str(key, item)?),
+                "pg_auth_type" => builder.pg_auth_type(parse_toml_from_str(key, item)?),
                 "remote_storage" => {
                     builder.remote_storage_config(RemoteStorageConfig::from_toml(item)?)
                 }
@@ -628,8 +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)?),
@@ -640,6 +721,13 @@ impl PageServerConf {
                 "synthetic_size_calculation_interval" =>
                     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}'"),
             }
@@ -647,7 +735,7 @@ impl PageServerConf {
 
         let mut conf = builder.build().context("invalid config")?;
 
-        if conf.auth_type == AuthType::NeonJWT {
+        if conf.http_auth_type == AuthType::NeonJWT || conf.pg_auth_type == AuthType::NeonJWT {
             let auth_validation_public_key_path = conf
                 .auth_validation_public_key_path
                 .get_or_insert_with(|| workdir.join("auth_public_key.pem"));
@@ -744,6 +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)
     }
 
@@ -763,10 +858,12 @@ impl PageServerConf {
             max_file_descriptors: defaults::DEFAULT_MAX_FILE_DESCRIPTORS,
             listen_pg_addr: defaults::DEFAULT_PG_LISTEN_ADDR.to_string(),
             listen_http_addr: defaults::DEFAULT_HTTP_LISTEN_ADDR.to_string(),
+            availability_zone: None,
             superuser: "cloud_admin".to_string(),
             workdir: repo_dir,
             pg_distrib_dir,
-            auth_type: AuthType::Trust,
+            http_auth_type: AuthType::Trust,
+            pg_auth_type: AuthType::Trust,
             auth_validation_public_key_path: None,
             remote_storage_config: None,
             default_tenant_conf: TenantConf::default(),
@@ -774,10 +871,17 @@ 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,
             synthetic_size_calculation_interval: Duration::from_secs(60),
+            evictions_low_residence_duration_metric_threshold: humantime::parse_duration(
+                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,
         }
@@ -862,6 +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 {
@@ -919,6 +1028,9 @@ metric_collection_interval = '222 s'
 cached_metric_collection_interval = '22200 s'
 metric_collection_endpoint = 'http://localhost:80/metrics'
 synthetic_size_calculation_interval = '333 s'
+
+evictions_low_residence_duration_metric_threshold = '444 s'
+
 log_format = 'json'
 
 "#;
@@ -944,6 +1056,7 @@ log_format = 'json'
                 id: NodeId(10),
                 listen_pg_addr: defaults::DEFAULT_PG_LISTEN_ADDR.to_string(),
                 listen_http_addr: defaults::DEFAULT_HTTP_LISTEN_ADDR.to_string(),
+                availability_zone: None,
                 wait_lsn_timeout: humantime::parse_duration(defaults::DEFAULT_WAIT_LSN_TIMEOUT)?,
                 wal_redo_timeout: humantime::parse_duration(defaults::DEFAULT_WAL_REDO_TIMEOUT)?,
                 superuser: defaults::DEFAULT_SUPERUSER.to_string(),
@@ -951,7 +1064,8 @@ log_format = 'json'
                 max_file_descriptors: defaults::DEFAULT_MAX_FILE_DESCRIPTORS,
                 workdir,
                 pg_distrib_dir,
-                auth_type: AuthType::Trust,
+                http_auth_type: AuthType::Trust,
+                pg_auth_type: AuthType::Trust,
                 auth_validation_public_key_path: None,
                 remote_storage_config: None,
                 default_tenant_conf: TenantConf::default(),
@@ -961,6 +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
                 )?,
@@ -971,6 +1087,10 @@ log_format = 'json'
                 synthetic_size_calculation_interval: humantime::parse_duration(
                     defaults::DEFAULT_SYNTHETIC_SIZE_CALCULATION_INTERVAL
                 )?,
+                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,
             },
@@ -1001,6 +1121,7 @@ log_format = 'json'
                 id: NodeId(10),
                 listen_pg_addr: "127.0.0.1:64000".to_string(),
                 listen_http_addr: "127.0.0.1:9898".to_string(),
+                availability_zone: None,
                 wait_lsn_timeout: Duration::from_secs(111),
                 wal_redo_timeout: Duration::from_secs(111),
                 superuser: "zzzz".to_string(),
@@ -1008,7 +1129,8 @@ log_format = 'json'
                 max_file_descriptors: 333,
                 workdir,
                 pg_distrib_dir,
-                auth_type: AuthType::Trust,
+                http_auth_type: AuthType::Trust,
+                pg_auth_type: AuthType::Trust,
                 auth_validation_public_key_path: None,
                 remote_storage_config: None,
                 default_tenant_conf: TenantConf::default(),
@@ -1016,10 +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,
             },
@@ -1146,6 +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 +1,689 @@
+//! 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,
+        })
+    }
+}

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:
@@ -351,6 +389,13 @@ paths:
         schema:
           type: string
           format: hex
+      - name: detach_ignored
+        in: query
+        required: false
+        schema:
+          type: boolean
+        description: |
+          When true, allow to detach a tenant which state is ignored.
     post:
       description: |
         Remove tenant data (including all corresponding timelines) from pageserver's memory and file system.
@@ -376,6 +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:
@@ -847,13 +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:
@@ -939,6 +986,13 @@ components:
       properties:
         msg:
           type: string
+    PreconditionFailedError:
+      type: object
+      required:
+        - msg
+      properties:
+        msg:
+          type: string
 
 security:
   - JWT: []

pageserver/src/http/routes.rs

@@ -18,10 +18,11 @@ 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;
-use crate::tenant::mgr::TenantMapInsertError;
+use crate::tenant::mgr::{TenantMapInsertError, TenantStateError};
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
 use crate::tenant::{PageReconstructError, Timeline};
@@ -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,
         })
     }
 }
@@ -82,38 +86,80 @@ fn get_config(request: &Request<Body>) -> &'static PageServerConf {
     get_state(request).conf
 }
 
+/// Check that the requester is authorized to operate on given tenant
 fn check_permission(request: &Request<Body>, tenant_id: Option<TenantId>) -> Result<(), ApiError> {
     check_permission_with(request, |claims| {
         crate::auth::check_permission(claims, tenant_id)
     })
 }
 
-fn apierror_from_prerror(err: PageReconstructError) -> ApiError {
-    match err {
-        PageReconstructError::Other(err) => ApiError::InternalServerError(err),
-        PageReconstructError::NeedsDownload(_, _) => {
-            // This shouldn't happen, because we use a RequestContext that requests to
-            // download any missing layer files on-demand.
-            ApiError::InternalServerError(anyhow::anyhow!("need to download remote layer file"))
-        }
-        PageReconstructError::Cancelled => {
-            ApiError::InternalServerError(anyhow::anyhow!("request was cancelled"))
-        }
-        PageReconstructError::WalRedo(err) => {
-            ApiError::InternalServerError(anyhow::Error::new(err))
+impl From<PageReconstructError> for ApiError {
+    fn from(pre: PageReconstructError) -> ApiError {
+        match pre {
+            PageReconstructError::Other(pre) => ApiError::InternalServerError(pre),
+            PageReconstructError::NeedsDownload(_, _) => {
+                // This shouldn't happen, because we use a RequestContext that requests to
+                // download any missing layer files on-demand.
+                ApiError::InternalServerError(anyhow::anyhow!("need to download remote layer file"))
+            }
+            PageReconstructError::Cancelled => {
+                ApiError::InternalServerError(anyhow::anyhow!("request was cancelled"))
+            }
+            PageReconstructError::WalRedo(pre) => {
+                ApiError::InternalServerError(anyhow::Error::new(pre))
+            }
         }
     }
 }
 
-fn apierror_from_tenant_map_insert_error(e: TenantMapInsertError) -> ApiError {
-    match e {
-        TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
-            ApiError::InternalServerError(anyhow::Error::new(e))
+impl From<TenantMapInsertError> for ApiError {
+    fn from(tmie: TenantMapInsertError) -> ApiError {
+        match tmie {
+            TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
+                ApiError::InternalServerError(anyhow::Error::new(tmie))
+            }
+            TenantMapInsertError::TenantAlreadyExists(id, state) => {
+                ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
+            }
+            TenantMapInsertError::Closure(e) => ApiError::InternalServerError(e),
         }
-        TenantMapInsertError::TenantAlreadyExists(id, state) => {
-            ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
+    }
+}
+
+impl From<TenantStateError> for ApiError {
+    fn from(tse: TenantStateError) -> ApiError {
+        match tse {
+            TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid)),
+            _ => ApiError::InternalServerError(anyhow::Error::new(tse)),
+        }
+    }
+}
+
+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),
         }
-        TenantMapInsertError::Closure(e) => ApiError::InternalServerError(e),
     }
 }
 
@@ -171,7 +217,7 @@ fn build_timeline_info_common(
             None
         }
     };
-    let current_physical_size = Some(timeline.layer_size_sum().approximate_is_ok());
+    let current_physical_size = Some(timeline.layer_size_sum());
     let state = timeline.current_state();
     let remote_consistent_lsn = timeline.get_remote_consistent_lsn().unwrap_or(Lsn(0));
 
@@ -217,9 +263,7 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Error);
 
-    let tenant = mgr::get_tenant(tenant_id, true)
-        .await
-        .map_err(ApiError::NotFound)?;
+    let tenant = mgr::get_tenant(tenant_id, true).await?;
     match tenant.create_timeline(
         new_timeline_id,
         request_data.ancestor_timeline_id.map(TimelineId::from),
@@ -249,9 +293,7 @@ async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>,
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
 
     let response_data = async {
-        let tenant = mgr::get_tenant(tenant_id, true)
-            .await
-            .map_err(ApiError::NotFound)?;
+        let tenant = mgr::get_tenant(tenant_id, true).await?;
         let timelines = tenant.list_timelines();
 
         let mut response_data = Vec::with_capacity(timelines.len());
@@ -267,7 +309,7 @@ async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>,
 
             response_data.push(timeline_info);
         }
-        Ok(response_data)
+        Ok::<Vec<TimelineInfo>, ApiError>(response_data)
     }
     .instrument(info_span!("timeline_list", tenant = %tenant_id))
     .await?;
@@ -286,9 +328,7 @@ async fn timeline_detail_handler(request: Request<Body>) -> Result<Response<Body
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
 
     let timeline_info = async {
-        let tenant = mgr::get_tenant(tenant_id, true)
-            .await
-            .map_err(ApiError::NotFound)?;
+        let tenant = mgr::get_tenant(tenant_id, true).await?;
 
         let timeline = tenant
             .get_timeline(timeline_id, false)
@@ -324,10 +364,7 @@ async fn get_lsn_by_timestamp_handler(request: Request<Body>) -> Result<Response
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
     let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
-    let result = timeline
-        .find_lsn_for_timestamp(timestamp_pg, &ctx)
-        .await
-        .map_err(apierror_from_prerror)?;
+    let result = timeline.find_lsn_for_timestamp(timestamp_pg, &ctx).await?;
 
     let result = match result {
         LsnForTimestamp::Present(lsn) => format!("{lsn}"),
@@ -352,8 +389,7 @@ async fn tenant_attach_handler(request: Request<Body>) -> Result<Response<Body>,
     if let Some(remote_storage) = &state.remote_storage {
         mgr::attach_tenant(state.conf, tenant_id, remote_storage.clone(), &ctx)
             .instrument(info_span!("tenant_attach", tenant = %tenant_id))
-            .await
-            .map_err(apierror_from_tenant_map_insert_error)?;
+            .await?;
     } else {
         return Err(ApiError::BadRequest(anyhow!(
             "attach_tenant is not possible because pageserver was configured without remote storage"
@@ -372,11 +408,7 @@ async fn timeline_delete_handler(request: Request<Body>) -> Result<Response<Body
 
     mgr::delete_timeline(tenant_id, timeline_id, &ctx)
         .instrument(info_span!("timeline_delete", tenant = %tenant_id, timeline = %timeline_id))
-        .await
-        // FIXME: Errors from `delete_timeline` can occur for a number of reasons, incuding both
-        // user and internal errors. Replace this with better handling once the error type permits
-        // it.
-        .map_err(ApiError::InternalServerError)?;
+        .await?;
 
     json_response(StatusCode::OK, ())
 }
@@ -384,15 +416,13 @@ async fn timeline_delete_handler(request: Request<Body>) -> Result<Response<Body
 async fn tenant_detach_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     check_permission(&request, Some(tenant_id))?;
+    let detach_ignored: Option<bool> = parse_query_param(&request, "detach_ignored")?;
 
     let state = get_state(&request);
     let conf = state.conf;
-    mgr::detach_tenant(conf, tenant_id)
+    mgr::detach_tenant(conf, tenant_id, detach_ignored.unwrap_or(false))
         .instrument(info_span!("tenant_detach", tenant = %tenant_id))
-        .await
-        // FIXME: Errors from `detach_tenant` can be caused by both both user and internal errors.
-        // Replace this with better handling once the error type permits it.
-        .map_err(ApiError::InternalServerError)?;
+        .await?;
 
     json_response(StatusCode::OK, ())
 }
@@ -406,8 +436,7 @@ async fn tenant_load_handler(request: Request<Body>) -> Result<Response<Body>, A
     let state = get_state(&request);
     mgr::load_tenant(state.conf, tenant_id, state.remote_storage.clone(), &ctx)
         .instrument(info_span!("load", tenant = %tenant_id))
-        .await
-        .map_err(apierror_from_tenant_map_insert_error)?;
+        .await?;
 
     json_response(StatusCode::ACCEPTED, ())
 }
@@ -420,10 +449,7 @@ async fn tenant_ignore_handler(request: Request<Body>) -> Result<Response<Body>,
     let conf = state.conf;
     mgr::ignore_tenant(conf, tenant_id)
         .instrument(info_span!("ignore_tenant", tenant = %tenant_id))
-        .await
-        // FIXME: Errors from `ignore_tenant` can be caused by both both user and internal errors.
-        // Replace this with better handling once the error type permits it.
-        .map_err(ApiError::InternalServerError)?;
+        .await?;
 
     json_response(StatusCode::OK, ())
 }
@@ -458,7 +484,7 @@ async fn tenant_status(request: Request<Body>) -> Result<Response<Body>, ApiErro
         // Calculate total physical size of all timelines
         let mut current_physical_size = 0;
         for timeline in tenant.list_timelines().iter() {
-            current_physical_size += timeline.layer_size_sum().approximate_is_ok();
+            current_physical_size += timeline.layer_size_sum();
         }
 
         let state = tenant.current_state();
@@ -497,9 +523,7 @@ async fn tenant_size_handler(request: Request<Body>) -> Result<Response<Body>, A
     let headers = request.headers();
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let tenant = mgr::get_tenant(tenant_id, true)
-        .await
-        .map_err(ApiError::InternalServerError)?;
+    let tenant = mgr::get_tenant(tenant_id, true).await?;
 
     // this can be long operation
     let inputs = tenant
@@ -747,6 +771,16 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
         );
     }
 
+    if let Some(eviction_policy) = request_data.eviction_policy {
+        tenant_conf.eviction_policy = Some(
+            serde_json::from_value(eviction_policy)
+                .context("parse field `eviction_policy`")
+                .map_err(ApiError::BadRequest)?,
+        );
+    }
+
+    tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
+
     let target_tenant_id = request_data
         .new_tenant_id
         .map(TenantId::from)
@@ -762,8 +796,7 @@ async fn tenant_create_handler(mut request: Request<Body>) -> Result<Response<Bo
         &ctx,
     )
     .instrument(info_span!("tenant_create", tenant = ?target_tenant_id))
-    .await
-    .map_err(apierror_from_tenant_map_insert_error)?;
+    .await?;
 
     // We created the tenant. Existing API semantics are that the tenant
     // is Active when this function returns.
@@ -787,9 +820,7 @@ async fn get_tenant_config_handler(request: Request<Body>) -> Result<Response<Bo
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     check_permission(&request, Some(tenant_id))?;
 
-    let tenant = mgr::get_tenant(tenant_id, false)
-        .await
-        .map_err(ApiError::NotFound)?;
+    let tenant = mgr::get_tenant(tenant_id, false).await?;
 
     let response = HashMap::from([
         (
@@ -881,13 +912,26 @@ 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))
+        .await?;
+
+    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
-        // FIXME: `update_tenant_config` can fail because of both user and internal errors.
-        // Replace this `map_err` with better error handling once the type permits it
-        .map_err(ApiError::InternalServerError)?;
+        .map_err(|_| ApiError::Conflict(String::from("no active tenant found")))?;
+
+    tenant.set_broken("broken from test");
 
     json_response(StatusCode::OK, ())
 }
@@ -1024,9 +1068,7 @@ async fn active_timeline_of_active_tenant(
     tenant_id: TenantId,
     timeline_id: TimelineId,
 ) -> Result<Arc<Timeline>, ApiError> {
-    let tenant = mgr::get_tenant(tenant_id, true)
-        .await
-        .map_err(ApiError::NotFound)?;
+    let tenant = mgr::get_tenant(tenant_id, true).await?;
     tenant
         .get_timeline(timeline_id, true)
         .map_err(ApiError::NotFound)
@@ -1043,6 +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,
@@ -1055,6 +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");
@@ -1099,7 +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(
@@ -1180,6 +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

@@ -9,22 +9,18 @@ use once_cell::sync::Lazy;
 use pageserver_api::models::state;
 use utils::id::{TenantId, TimelineId};
 
-/// Prometheus histogram buckets (in seconds) that capture the majority of
-/// latencies in the microsecond range but also extend far enough up to distinguish
-/// "bad" from "really bad".
-fn get_buckets_for_critical_operations() -> Vec<f64> {
-    let buckets_per_digit = 5;
-    let min_exponent = -6;
-    let max_exponent = 2;
-
-    let mut buckets = vec![];
-    // Compute 10^(exp / buckets_per_digit) instead of 10^(1/buckets_per_digit)^exp
-    // because it's more numerically stable and doesn't result in numbers like 9.999999
-    for exp in (min_exponent * buckets_per_digit)..=(max_exponent * buckets_per_digit) {
-        buckets.push(10_f64.powf(exp as f64 / buckets_per_digit as f64))
-    }
-    buckets
-}
+/// Prometheus histogram buckets (in seconds) for operations in the critical
+/// path. In other words, operations that directly affect that latency of user
+/// queries.
+///
+/// The buckets capture the majority of latencies in the microsecond and
+/// millisecond range but also extend far enough up to distinguish "bad" from
+/// "really bad".
+const CRITICAL_OP_BUCKETS: &[f64] = &[
+    0.000_001, 0.000_010, 0.000_100, // 1 us, 10 us, 100 us
+    0.001_000, 0.010_000, 0.100_000, // 1 ms, 10 ms, 100 ms
+    1.0, 10.0, 100.0, // 1 s, 10 s, 100 s
+];
 
 // Metrics collected on operations on the storage repository.
 const STORAGE_TIME_OPERATIONS: &[&str] = &[
@@ -55,12 +51,15 @@ pub static STORAGE_TIME_COUNT_PER_TIMELINE: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+// Buckets for background operations like compaction, GC, size calculation
+const STORAGE_OP_BUCKETS: &[f64] = &[0.010, 0.100, 1.0, 10.0, 100.0, 1000.0];
+
 pub static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_storage_operations_seconds_global",
         "Time spent on storage operations",
         &["operation"],
-        get_buckets_for_critical_operations(),
+        STORAGE_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
@@ -71,7 +70,7 @@ static RECONSTRUCT_TIME: Lazy<HistogramVec> = Lazy::new(|| {
         "pageserver_getpage_reconstruct_seconds",
         "Time spent in reconstruct_value",
         &["tenant_id", "timeline_id"],
-        get_buckets_for_critical_operations(),
+        CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
@@ -90,7 +89,7 @@ static WAIT_LSN_TIME: Lazy<HistogramVec> = Lazy::new(|| {
         "pageserver_wait_lsn_seconds",
         "Time spent waiting for WAL to arrive",
         &["tenant_id", "timeline_id"],
-        get_buckets_for_critical_operations(),
+        CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
@@ -195,15 +194,101 @@ static PERSISTENT_BYTES_WRITTEN: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static EVICTIONS: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_evictions",
+        "Number of layers evicted from the pageserver",
+        &["tenant_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
+static EVICTIONS_WITH_LOW_RESIDENCE_DURATION: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_evictions_with_low_residence_duration",
+        "If a layer is evicted that was resident for less than `low_threshold`, it is counted to this counter. \
+         Residence duration is determined using the `residence_duration_data_source`.",
+        &["tenant_id", "timeline_id", "residence_duration_data_source", "low_threshold_secs"]
+    )
+    .expect("failed to define a metric")
+});
+
+/// Each [`Timeline`]'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
+#[derive(Debug)]
+pub struct EvictionsWithLowResidenceDuration {
+    data_source: &'static str,
+    threshold: Duration,
+    counter: Option<IntCounter>,
+}
+
+pub struct EvictionsWithLowResidenceDurationBuilder {
+    data_source: &'static str,
+    threshold: Duration,
+}
+
+impl EvictionsWithLowResidenceDurationBuilder {
+    pub fn new(data_source: &'static str, threshold: Duration) -> Self {
+        Self {
+            data_source,
+            threshold,
+        }
+    }
+
+    fn build(&self, tenant_id: &str, timeline_id: &str) -> EvictionsWithLowResidenceDuration {
+        let counter = EVICTIONS_WITH_LOW_RESIDENCE_DURATION
+            .get_metric_with_label_values(&[
+                tenant_id,
+                timeline_id,
+                self.data_source,
+                &EvictionsWithLowResidenceDuration::threshold_label_value(self.threshold),
+            ])
+            .unwrap();
+        EvictionsWithLowResidenceDuration {
+            data_source: self.data_source,
+            threshold: self.threshold,
+            counter: Some(counter),
+        }
+    }
+}
+
+impl EvictionsWithLowResidenceDuration {
+    fn threshold_label_value(threshold: Duration) -> String {
+        format!("{}", threshold.as_secs())
+    }
+
+    pub fn observe(&self, observed_value: Duration) {
+        if observed_value < self.threshold {
+            self.counter
+                .as_ref()
+                .expect("nobody calls this function after `remove_from_vec`")
+                .inc();
+        }
+    }
+
+    // This could be a `Drop` impl, but, we need the `tenant_id` and `timeline_id`.
+    fn remove(&mut self, tenant_id: &str, timeline_id: &str) {
+        let Some(_counter) = self.counter.take() else {
+            return;
+        };
+        EVICTIONS_WITH_LOW_RESIDENCE_DURATION
+            .remove_label_values(&[
+                tenant_id,
+                timeline_id,
+                self.data_source,
+                &Self::threshold_label_value(self.threshold),
+            ])
+            .expect("we own the metric, no-one else should remove it");
+    }
+}
+
 // Metrics collected on disk IO operations
+//
+// Roughly logarithmic scale.
 const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
-    0.000001, // 1 usec
-    0.00001,  // 10 usec
-    0.0001,   // 100 usec
-    0.001,    // 1 msec
-    0.01,     // 10 msec
-    0.1,      // 100 msec
-    1.0,      // 1 sec
+    0.000030, // 30 usec
+    0.001000, // 1000 usec
+    0.030,    // 30 ms
+    1.000,    // 1000 ms
 ];
 
 const STORAGE_IO_TIME_OPERATIONS: &[&str] = &[
@@ -238,20 +323,12 @@ const SMGR_QUERY_TIME_OPERATIONS: &[&str] = &[
     "get_db_size",
 ];
 
-const SMGR_QUERY_TIME_BUCKETS: &[f64] = &[
-    0.00001, // 1/100000 s
-    0.0001, 0.00015, 0.0002, 0.00025, 0.0003, 0.00035, 0.0005, 0.00075, // 1/10000 s
-    0.001, 0.0025, 0.005, 0.0075, // 1/1000 s
-    0.01, 0.0125, 0.015, 0.025, 0.05, // 1/100 s
-    0.1,  // 1/10 s
-];
-
 pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
     register_histogram_vec!(
         "pageserver_smgr_query_seconds",
         "Time spent on smgr query handling",
         &["smgr_query_type", "tenant_id", "timeline_id"],
-        SMGR_QUERY_TIME_BUCKETS.into()
+        CRITICAL_OP_BUCKETS.into(),
     )
     .expect("failed to define a metric")
 });
@@ -509,7 +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,
@@ -520,10 +596,16 @@ pub struct TimelineMetrics {
     pub current_logical_size_gauge: UIntGauge,
     pub num_persistent_files_created: IntCounter,
     pub persistent_bytes_written: IntCounter,
+    pub evictions: IntCounter,
+    pub evictions_with_low_residence_duration: EvictionsWithLowResidenceDuration,
 }
 
 impl TimelineMetrics {
-    pub fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
+    pub fn new(
+        tenant_id: &TenantId,
+        timeline_id: &TimelineId,
+        evictions_with_low_residence_duration_builder: EvictionsWithLowResidenceDurationBuilder,
+    ) -> Self {
         let tenant_id = tenant_id.to_string();
         let timeline_id = timeline_id.to_string();
         let reconstruct_time_histo = RECONSTRUCT_TIME
@@ -536,8 +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);
@@ -560,6 +640,11 @@ impl TimelineMetrics {
         let persistent_bytes_written = PERSISTENT_BYTES_WRITTEN
             .get_metric_with_label_values(&[&tenant_id, &timeline_id])
             .unwrap();
+        let evictions = EVICTIONS
+            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
+            .unwrap();
+        let evictions_with_low_residence_duration =
+            evictions_with_low_residence_duration_builder.build(&tenant_id, &timeline_id);
 
         TimelineMetrics {
             tenant_id,
@@ -569,7 +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,
@@ -579,6 +663,8 @@ impl TimelineMetrics {
             current_logical_size_gauge,
             num_persistent_files_created,
             persistent_bytes_written,
+            evictions,
+            evictions_with_low_residence_duration,
         }
     }
 }
@@ -595,7 +681,9 @@ impl Drop for TimelineMetrics {
         let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
         let _ = NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, timeline_id]);
         let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, timeline_id]);
-
+        let _ = EVICTIONS.remove_label_values(&[tenant_id, timeline_id]);
+        self.evictions_with_low_residence_duration
+            .remove(tenant_id, timeline_id);
         for op in STORAGE_TIME_OPERATIONS {
             let _ =
                 STORAGE_TIME_SUM_PER_TIMELINE.remove_label_values(&[op, tenant_id, timeline_id]);
@@ -630,7 +718,7 @@ use std::collections::HashMap;
 use std::pin::Pin;
 use std::sync::{Arc, Mutex};
 use std::task::{Context, Poll};
-use std::time::Instant;
+use std::time::{Duration, Instant};
 
 pub struct RemoteTimelineClientMetrics {
     tenant_id: String,

pageserver/src/page_service.rs

@@ -12,7 +12,7 @@
 use anyhow::Context;
 use bytes::Buf;
 use bytes::Bytes;
-use futures::{Stream, StreamExt};
+use futures::Stream;
 use pageserver_api::models::TenantState;
 use pageserver_api::models::{
     PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
@@ -27,10 +27,12 @@ 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;
 use std::time::Duration;
+use tokio_util::io::StreamReader;
 use tracing::*;
 use utils::id::ConnectionId;
 use utils::{
@@ -115,6 +117,49 @@ fn copyin_stream(pgb: &mut PostgresBackendTCP) -> impl Stream<Item = io::Result<
     }
 }
 
+/// Read the end of a tar archive.
+///
+/// A tar archive normally ends with two consecutive blocks of zeros, 512 bytes each.
+/// `tokio_tar` already read the first such block. Read the second all-zeros block,
+/// and check that there is no more data after the EOF marker.
+///
+/// XXX: Currently, any trailing data after the EOF marker prints a warning.
+/// Perhaps it should be a hard error?
+async fn read_tar_eof(mut reader: (impl tokio::io::AsyncRead + Unpin)) -> anyhow::Result<()> {
+    use tokio::io::AsyncReadExt;
+    let mut buf = [0u8; 512];
+
+    // Read the all-zeros block, and verify it
+    let mut total_bytes = 0;
+    while total_bytes < 512 {
+        let nbytes = reader.read(&mut buf[total_bytes..]).await?;
+        total_bytes += nbytes;
+        if nbytes == 0 {
+            break;
+        }
+    }
+    if total_bytes < 512 {
+        anyhow::bail!("incomplete or invalid tar EOF marker");
+    }
+    if !buf.iter().all(|&x| x == 0) {
+        anyhow::bail!("invalid tar EOF marker");
+    }
+
+    // Drain any data after the EOF marker
+    let mut trailing_bytes = 0;
+    loop {
+        let nbytes = reader.read(&mut buf).await?;
+        trailing_bytes += nbytes;
+        if nbytes == 0 {
+            break;
+        }
+    }
+    if trailing_bytes > 0 {
+        warn!("ignored {trailing_bytes} unexpected bytes after the tar archive");
+    }
+    Ok(())
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 
 ///
@@ -422,19 +467,13 @@ impl PageServerHandler {
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
 
-        let mut copyin_stream = Box::pin(copyin_stream(pgb));
+        let mut copyin_reader = pin!(StreamReader::new(copyin_stream(pgb)));
         timeline
-            .import_basebackup_from_tar(&mut copyin_stream, base_lsn, &ctx)
+            .import_basebackup_from_tar(&mut copyin_reader, base_lsn, &ctx)
             .await?;
 
-        // Drain the rest of the Copy data
-        let mut bytes_after_tar = 0;
-        while let Some(bytes) = copyin_stream.next().await {
-            bytes_after_tar += bytes?.len();
-        }
-        if bytes_after_tar > 0 {
-            warn!("ignored {bytes_after_tar} unexpected bytes after the tar archive");
-        }
+        // Read the end of the tar archive.
+        read_tar_eof(copyin_reader).await?;
 
         // TODO check checksum
         // Meanwhile you can verify client-side by taking fullbackup
@@ -473,19 +512,12 @@ impl PageServerHandler {
         info!("importing wal");
         pgb.write_message_noflush(&BeMessage::CopyInResponse)?;
         pgb.flush().await?;
-        let mut copyin_stream = Box::pin(copyin_stream(pgb));
-        let mut reader = tokio_util::io::StreamReader::new(&mut copyin_stream);
-        import_wal_from_tar(&timeline, &mut reader, start_lsn, end_lsn, &ctx).await?;
+        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");
 
-        // Drain the rest of the Copy data
-        let mut bytes_after_tar = 0;
-        while let Some(bytes) = copyin_stream.next().await {
-            bytes_after_tar += bytes?.len();
-        }
-        if bytes_after_tar > 0 {
-            warn!("ignored {bytes_after_tar} unexpected bytes after the tar archive");
-        }
+        // Read the end of the tar archive.
+        read_tar_eof(copyin_reader).await?;
 
         // TODO Does it make sense to overshoot?
         if timeline.get_last_record_lsn() < end_lsn {
@@ -1074,7 +1106,10 @@ async fn get_active_tenant_with_timeout(
     tenant_id: TenantId,
     _ctx: &RequestContext, /* require get a context to support cancellation in the future */
 ) -> Result<Arc<Tenant>, GetActiveTenantError> {
-    let tenant = mgr::get_tenant(tenant_id, false).await?;
+    let tenant = match mgr::get_tenant(tenant_id, false).await {
+        Ok(tenant) => tenant,
+        Err(e) => return Err(GetActiveTenantError::Other(e.into())),
+    };
     let wait_time = Duration::from_secs(30);
     match tokio::time::timeout(wait_time, tenant.wait_to_become_active()).await {
         Ok(Ok(())) => Ok(tenant),

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

@@ -12,9 +12,7 @@
 //!
 
 use anyhow::{bail, Context};
-use bytes::Bytes;
 use futures::FutureExt;
-use futures::Stream;
 use pageserver_api::models::TimelineState;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
@@ -48,6 +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;
@@ -96,7 +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;
@@ -144,6 +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.
@@ -239,14 +240,13 @@ impl UninitializedTimeline<'_> {
     /// Prepares timeline data by loading it from the basebackup archive.
     pub async fn import_basebackup_from_tar(
         self,
-        copyin_stream: &mut (impl Stream<Item = io::Result<Bytes>> + Sync + Send + Unpin),
+        copyin_read: &mut (impl tokio::io::AsyncRead + Send + Sync + Unpin),
         base_lsn: Lsn,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Timeline>> {
         let raw_timeline = self.raw_timeline()?;
 
-        let mut reader = tokio_util::io::StreamReader::new(copyin_stream);
-        import_datadir::import_basebackup_from_tar(raw_timeline, &mut reader, base_lsn, ctx)
+        import_datadir::import_basebackup_from_tar(raw_timeline, copyin_read, base_lsn, ctx)
             .await
             .context("Failed to import basebackup")?;
 
@@ -434,6 +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,
@@ -481,7 +491,7 @@ impl Tenant {
 
             let dummy_timeline = self.create_timeline_data(
                 timeline_id,
-                up_to_date_metadata.clone(),
+                up_to_date_metadata,
                 ancestor.clone(),
                 remote_client,
             )?;
@@ -506,7 +516,7 @@ impl Tenant {
                     let broken_timeline = self
                         .create_timeline_data(
                             timeline_id,
-                            up_to_date_metadata.clone(),
+                            up_to_date_metadata,
                             ancestor.clone(),
                             None,
                         )
@@ -1145,7 +1155,7 @@ impl Tenant {
         );
         self.prepare_timeline(
             new_timeline_id,
-            new_metadata,
+            &new_metadata,
             timeline_uninit_mark,
             true,
             None,
@@ -1243,11 +1253,8 @@ impl Tenant {
             "Cannot run GC iteration on inactive tenant"
         );
 
-        let gc_result = self
-            .gc_iteration_internal(target_timeline_id, horizon, pitr, ctx)
-            .await;
-
-        gc_result
+        self.gc_iteration_internal(target_timeline_id, horizon, pitr, ctx)
+            .await
     }
 
     /// Perform one compaction iteration.
@@ -1313,7 +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 = {
@@ -1325,13 +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());
@@ -1699,6 +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;
     }
@@ -1706,7 +1720,7 @@ impl Tenant {
     fn create_timeline_data(
         &self,
         new_timeline_id: TimelineId,
-        new_metadata: TimelineMetadata,
+        new_metadata: &TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
         remote_client: Option<RemoteTimelineClient>,
     ) -> anyhow::Result<Arc<Timeline>> {
@@ -1777,6 +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()),
         }
     }
 
@@ -2166,13 +2181,25 @@ impl Tenant {
         let new_timeline = self
             .prepare_timeline(
                 dst_id,
-                metadata,
+                &metadata,
                 timeline_uninit_mark,
                 false,
                 Some(Arc::clone(src_timeline)),
             )?
             .initialize_with_lock(&mut timelines, true, true)?;
         drop(timelines);
+
+        // Root timeline gets its layers during creation and uploads them along with the metadata.
+        // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
+        // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
+        // could get incorrect information and remove more layers, than needed.
+        // See also https://github.com/neondatabase/neon/issues/3865
+        if let Some(remote_client) = new_timeline.remote_client.as_ref() {
+            remote_client
+                .schedule_index_upload_for_metadata_update(&metadata)
+                .context("branch initial metadata upload")?;
+        }
+
         info!("branched timeline {dst_id} from {src_id} at {start_lsn}");
 
         Ok(new_timeline)
@@ -2235,7 +2262,7 @@ impl Tenant {
             pg_version,
         );
         let raw_timeline =
-            self.prepare_timeline(timeline_id, new_metadata, timeline_uninit_mark, true, None)?;
+            self.prepare_timeline(timeline_id, &new_metadata, timeline_uninit_mark, true, None)?;
 
         let tenant_id = raw_timeline.owning_tenant.tenant_id;
         let unfinished_timeline = raw_timeline.raw_timeline()?;
@@ -2289,7 +2316,7 @@ impl Tenant {
     fn prepare_timeline(
         &self,
         new_timeline_id: TimelineId,
-        new_metadata: TimelineMetadata,
+        new_metadata: &TimelineMetadata,
         uninit_mark: TimelineUninitMark,
         init_layers: bool,
         ancestor: Option<Arc<Timeline>>,
@@ -2303,7 +2330,7 @@ impl Tenant {
                 tenant_id,
                 new_timeline_id,
             );
-            remote_client.init_upload_queue_for_empty_remote(&new_metadata)?;
+            remote_client.init_upload_queue_for_empty_remote(new_metadata)?;
             Some(remote_client)
         } else {
             None
@@ -2342,17 +2369,12 @@ impl Tenant {
         &self,
         timeline_path: &Path,
         new_timeline_id: TimelineId,
-        new_metadata: TimelineMetadata,
+        new_metadata: &TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
         remote_client: Option<RemoteTimelineClient>,
     ) -> anyhow::Result<Arc<Timeline>> {
         let timeline_data = self
-            .create_timeline_data(
-                new_timeline_id,
-                new_metadata.clone(),
-                ancestor,
-                remote_client,
-            )
+            .create_timeline_data(new_timeline_id, new_metadata, ancestor, remote_client)
             .context("Failed to create timeline data structure")?;
         crashsafe::create_dir_all(timeline_path).context("Failed to create timeline directory")?;
 
@@ -2364,7 +2386,7 @@ impl Tenant {
             self.conf,
             new_timeline_id,
             self.tenant_id,
-            &new_metadata,
+            new_metadata,
             true,
         )
         .context("Failed to create timeline metadata")?;
@@ -2768,6 +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,
             }
         }
     }
@@ -3176,6 +3199,44 @@ mod tests {
     }
      */
 
+    #[tokio::test]
+    async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
+        let (tenant, ctx) =
+            TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")?
+                .load()
+                .await;
+        let tline = tenant
+            .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)?
+            .initialize(&ctx)?;
+        make_some_layers(tline.as_ref(), Lsn(0x20)).await?;
+
+        tenant
+            .branch_timeline(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
+            .await?;
+        let newtline = tenant
+            .get_timeline(NEW_TIMELINE_ID, true)
+            .expect("Should have a local timeline");
+
+        make_some_layers(newtline.as_ref(), Lsn(0x60)).await?;
+
+        tline.set_state(TimelineState::Broken);
+
+        tenant
+            .gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO, &ctx)
+            .await?;
+
+        assert_eq!(
+            newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
+            TEST_IMG(&format!("foo at {}", Lsn(0x40)))
+        );
+
+        let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
+        assert_eq!(branchpoints.len(), 1);
+        assert_eq!(branchpoints[0], Lsn(0x40));
+
+        Ok(())
+    }
+
     #[tokio::test]
     async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
         let (tenant, ctx) =

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)]
@@ -220,48 +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),
-        }
-    }
-
-    pub fn update(&mut self, other: &TenantConfOpt) {
-        if let Some(checkpoint_distance) = other.checkpoint_distance {
-            self.checkpoint_distance = Some(checkpoint_distance);
-        }
-        if let Some(checkpoint_timeout) = other.checkpoint_timeout {
-            self.checkpoint_timeout = Some(checkpoint_timeout);
-        }
-        if let Some(compaction_target_size) = other.compaction_target_size {
-            self.compaction_target_size = Some(compaction_target_size);
-        }
-        if let Some(compaction_period) = other.compaction_period {
-            self.compaction_period = Some(compaction_period);
-        }
-        if let Some(compaction_threshold) = other.compaction_threshold {
-            self.compaction_threshold = Some(compaction_threshold);
-        }
-        if let Some(gc_horizon) = other.gc_horizon {
-            self.gc_horizon = Some(gc_horizon);
-        }
-        if let Some(gc_period) = other.gc_period {
-            self.gc_period = Some(gc_period);
-        }
-        if let Some(image_creation_threshold) = other.image_creation_threshold {
-            self.image_creation_threshold = Some(image_creation_threshold);
-        }
-        if let Some(pitr_interval) = other.pitr_interval {
-            self.pitr_interval = Some(pitr_interval);
-        }
-        if let Some(walreceiver_connect_timeout) = other.walreceiver_connect_timeout {
-            self.walreceiver_connect_timeout = Some(walreceiver_connect_timeout);
-        }
-        if let Some(lagging_wal_timeout) = other.lagging_wal_timeout {
-            self.lagging_wal_timeout = Some(lagging_wal_timeout);
-        }
-        if let Some(max_lsn_wal_lag) = other.max_lsn_wal_lag {
-            self.max_lsn_wal_lag = Some(max_lsn_wal_lag);
-        }
-        if let Some(trace_read_requests) = other.trace_read_requests {
-            self.trace_read_requests = Some(trace_read_requests);
+            min_resident_size_override: self
+                .min_resident_size_override
+                .or(global_conf.min_resident_size_override),
         }
     }
 }
@@ -293,6 +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/layer_map.rs

@@ -154,11 +154,7 @@ where
         expected: &Arc<L>,
         new: Arc<L>,
     ) -> anyhow::Result<Replacement<Arc<L>>> {
-        fail::fail_point!("layermap-replace-notfound", |_| Ok(
-            // this is not what happens if an L0 layer was not found a anyhow error but perhaps
-            // that should be changed. this is good enough to show a replacement failure.
-            Replacement::NotFound
-        ));
+        fail::fail_point!("layermap-replace-notfound", |_| Ok(Replacement::NotFound));
 
         self.layer_map.replace_historic_noflush(expected, new)
     }
@@ -340,12 +336,15 @@ where
 
         let l0_index = if expected_l0 {
             // find the index in case replace worked, we need to replace that as well
-            Some(
-                self.l0_delta_layers
-                    .iter()
-                    .position(|slot| Self::compare_arced_layers(slot, expected))
-                    .ok_or_else(|| anyhow::anyhow!("existing l0 delta layer was not found"))?,
-            )
+            let pos = self
+                .l0_delta_layers
+                .iter()
+                .position(|slot| Self::compare_arced_layers(slot, expected));
+
+            if pos.is_none() {
+                return Ok(Replacement::NotFound);
+            }
+            pos
         } else {
             None
         };
@@ -804,6 +803,26 @@ mod tests {
             )
         }
 
+        #[test]
+        fn replacing_missing_l0_is_notfound() {
+            // original impl had an oversight, and L0 was an anyhow::Error. anyhow::Error should
+            // however only happen for precondition failures.
+
+            let layer = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69";
+            let layer = LayerFileName::from_str(layer).unwrap();
+            let layer = LayerDescriptor::from(layer);
+
+            // same skeletan construction; see scenario below
+            let not_found: Arc<dyn Layer> = Arc::new(layer.clone());
+            let new_version: Arc<dyn Layer> = Arc::new(layer);
+
+            let mut map = LayerMap::default();
+
+            let res = map.batch_update().replace_historic(&not_found, new_version);
+
+            assert!(matches!(res, Ok(Replacement::NotFound)), "{res:?}");
+        }
+
         fn l0_delta_layers_updated_scenario(layer_name: &str, expected_l0: bool) {
             let name = LayerFileName::from_str(layer_name).unwrap();
             let skeleton = LayerDescriptor::from(name);
@@ -813,7 +832,8 @@ mod tests {
 
             let mut map = LayerMap::default();
 
-            // two disjoint Arcs in different lifecycle phases.
+            // two disjoint Arcs in different lifecycle phases. even if it seems they must be the
+            // same layer, we use LayerMap::compare_arced_layers as the identity of layers.
             assert!(!LayerMap::compare_arced_layers(&remote, &downloaded));
 
             let expected_in_counts = (1, usize::from(expected_l0));

pageserver/src/tenant/mgr.rs

@@ -289,7 +289,7 @@ pub async fn set_new_tenant_config(
     conf: &'static PageServerConf,
     new_tenant_conf: TenantConfOpt,
     tenant_id: TenantId,
-) -> anyhow::Result<()> {
+) -> Result<(), TenantStateError> {
     info!("configuring tenant {tenant_id}");
     let tenant = get_tenant(tenant_id, true).await?;
 
@@ -306,50 +306,84 @@ pub async fn set_new_tenant_config(
 
 /// Gets the tenant from the in-memory data, erroring if it's absent or is not fitting to the query.
 /// `active_only = true` allows to query only tenants that are ready for operations, erroring on other kinds of tenants.
-pub async fn get_tenant(tenant_id: TenantId, active_only: bool) -> anyhow::Result<Arc<Tenant>> {
+pub async fn get_tenant(
+    tenant_id: TenantId,
+    active_only: bool,
+) -> Result<Arc<Tenant>, TenantStateError> {
     let m = TENANTS.read().await;
     let tenant = m
         .get(&tenant_id)
-        .with_context(|| format!("Tenant {tenant_id} not found in the local state"))?;
+        .ok_or(TenantStateError::NotFound(tenant_id))?;
     if active_only && !tenant.is_active() {
-        anyhow::bail!(
-            "Tenant {tenant_id} is not active. Current state: {:?}",
-            tenant.current_state()
-        )
+        Err(TenantStateError::NotActive(tenant_id))
     } else {
         Ok(Arc::clone(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,
-) -> anyhow::Result<()> {
-    match get_tenant(tenant_id, true).await {
-        Ok(tenant) => {
-            tenant.delete_timeline(timeline_id, ctx).await?;
-        }
-        Err(e) => anyhow::bail!("Cannot access tenant {tenant_id} in local tenant state: {e:?}"),
-    }
-
+) -> Result<(), DeleteTimelineError> {
+    let tenant = get_tenant(tenant_id, true).await?;
+    tenant.delete_timeline(timeline_id, ctx).await?;
     Ok(())
 }
 
+#[derive(Debug, thiserror::Error)]
+pub enum TenantStateError {
+    #[error("Tenant {0} not found")]
+    NotFound(TenantId),
+    #[error("Tenant {0} is stopping")]
+    IsStopping(TenantId),
+    #[error("Tenant {0} is not active")]
+    NotActive(TenantId),
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
 pub async fn detach_tenant(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
-) -> anyhow::Result<()> {
-    remove_tenant_from_memory(tenant_id, async {
-        let local_tenant_directory = conf.tenant_path(&tenant_id);
+    detach_ignored: bool,
+) -> Result<(), TenantStateError> {
+    let local_files_cleanup_operation = |tenant_id_to_clean| async move {
+        let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
         fs::remove_dir_all(&local_tenant_directory)
             .await
             .with_context(|| {
-                format!("Failed to remove local tenant directory {local_tenant_directory:?}")
+                format!("local tenant directory {local_tenant_directory:?} removal")
             })?;
         Ok(())
-    })
-    .await
+    };
+
+    let removal_result =
+        remove_tenant_from_memory(tenant_id, local_files_cleanup_operation(tenant_id)).await;
+
+    // Ignored tenants are not present in memory and will bail the removal from memory operation.
+    // Before returning the error, check for ignored tenant removal case — we only need to clean its local files then.
+    if detach_ignored && matches!(removal_result, Err(TenantStateError::NotFound(_))) {
+        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(tenant_id);
+        if tenant_ignore_mark.exists() {
+            info!("Detaching an ignored tenant");
+            local_files_cleanup_operation(tenant_id)
+                .await
+                .with_context(|| format!("Ignored tenant {tenant_id} local files cleanup"))?;
+            return Ok(());
+        }
+    }
+
+    removal_result
 }
 
 pub async fn load_tenant(
@@ -379,7 +413,7 @@ pub async fn load_tenant(
 pub async fn ignore_tenant(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
-) -> anyhow::Result<()> {
+) -> Result<(), TenantStateError> {
     remove_tenant_from_memory(tenant_id, async {
         let ignore_mark_file = conf.tenant_ignore_mark_file_path(tenant_id);
         fs::File::create(&ignore_mark_file)
@@ -489,7 +523,7 @@ where
 async fn remove_tenant_from_memory<V, F>(
     tenant_id: TenantId,
     tenant_cleanup: F,
-) -> anyhow::Result<V>
+) -> Result<V, TenantStateError>
 where
     F: std::future::Future<Output = anyhow::Result<V>>,
 {
@@ -505,11 +539,9 @@ where
                 | TenantState::Loading
                 | TenantState::Broken
                 | TenantState::Active => tenant.set_stopping(),
-                TenantState::Stopping => {
-                    anyhow::bail!("Tenant {tenant_id} is stopping already")
-                }
+                TenantState::Stopping => return Err(TenantStateError::IsStopping(tenant_id)),
             },
-            None => anyhow::bail!("Tenant not found for id {tenant_id}"),
+            None => return Err(TenantStateError::NotFound(tenant_id)),
         }
     }
 
@@ -532,10 +564,15 @@ where
         Err(e) => {
             let tenants_accessor = TENANTS.read().await;
             match tenants_accessor.get(&tenant_id) {
-                Some(tenant) => tenant.set_broken(&e.to_string()),
-                None => warn!("Tenant {tenant_id} got removed from memory"),
+                Some(tenant) => {
+                    tenant.set_broken(&e.to_string());
+                }
+                None => {
+                    warn!("Tenant {tenant_id} got removed from memory");
+                    return Err(TenantStateError::NotFound(tenant_id));
+                }
             }
-            Err(e)
+            Err(TenantStateError::Other(e))
         }
     }
 }
@@ -555,7 +592,7 @@ pub async fn immediate_gc(
     let tenant = guard
         .get(&tenant_id)
         .map(Arc::clone)
-        .with_context(|| format!("Tenant {tenant_id} not found"))
+        .with_context(|| format!("tenant {tenant_id}"))
         .map_err(ApiError::NotFound)?;
 
     let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
@@ -605,7 +642,7 @@ pub async fn immediate_compact(
     let tenant = guard
         .get(&tenant_id)
         .map(Arc::clone)
-        .with_context(|| format!("Tenant {tenant_id} not found"))
+        .with_context(|| format!("tenant {tenant_id}"))
         .map_err(ApiError::NotFound)?;
 
     let timeline = tenant

pageserver/src/tenant/remote_timeline_client.rs

@@ -210,7 +210,6 @@ pub use download::{is_temp_download_file, list_remote_timelines};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
 
-use anyhow::ensure;
 use remote_storage::{DownloadError, GenericRemoteStorage};
 use std::ops::DerefMut;
 use tokio::runtime::Runtime;
@@ -347,7 +346,7 @@ impl RemoteTimelineClient {
                 .layer_metadata
                 .values()
                 // If we don't have the file size for the layer, don't account for it in the metric.
-                .map(|ilmd| ilmd.file_size.unwrap_or(0))
+                .map(|ilmd| ilmd.file_size)
                 .sum()
         } else {
             0
@@ -420,34 +419,6 @@ impl RemoteTimelineClient {
             .await?
         };
 
-        // Update the metadata for given layer file. The remote index file
-        // might be missing some information for the file; this allows us
-        // to fill in the missing details.
-        if layer_metadata.file_size().is_none() {
-            let new_metadata = LayerFileMetadata::new(downloaded_size);
-            let mut guard = self.upload_queue.lock().unwrap();
-            let upload_queue = guard.initialized_mut()?;
-            if let Some(upgraded) = upload_queue.latest_files.get_mut(layer_file_name) {
-                if upgraded.merge(&new_metadata) {
-                    upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-                }
-                // If we don't do an index file upload inbetween here and restart,
-                // the value will go back down after pageserver restart, since we will
-                // have lost this data point.
-                // But, we upload index part fairly frequently, and restart pageserver rarely.
-                // So, by accounting eagerly, we present a most-of-the-time-more-accurate value sooner.
-                self.metrics
-                    .remote_physical_size_gauge()
-                    .add(downloaded_size);
-            } else {
-                // The file should exist, since we just downloaded it.
-                warn!(
-                    "downloaded file {:?} not found in local copy of the index file",
-                    layer_file_name
-                );
-            }
-        }
-
         REMOTE_ONDEMAND_DOWNLOADED_LAYERS.inc();
         REMOTE_ONDEMAND_DOWNLOADED_BYTES.inc_by(downloaded_size);
 
@@ -550,13 +521,6 @@ impl RemoteTimelineClient {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
-        // The file size can be missing for files that were created before we tracked that
-        // in the metadata, but it should be present for any new files we create.
-        ensure!(
-            layer_metadata.file_size().is_some(),
-            "file size not initialized in metadata"
-        );
-
         upload_queue
             .latest_files
             .insert(layer_file_name.clone(), layer_metadata.clone());

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -21,7 +21,7 @@ use remote_storage::{DownloadError, GenericRemoteStorage};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
 
-use super::index::{IndexPart, IndexPartUnclean, LayerFileMetadata};
+use super::index::{IndexPart, LayerFileMetadata};
 use super::{FAILED_DOWNLOAD_RETRIES, FAILED_DOWNLOAD_WARN_THRESHOLD};
 
 async fn fsync_path(path: impl AsRef<std::path::Path>) -> Result<(), std::io::Error> {
@@ -113,16 +113,11 @@ pub async fn download_layer_file<'a>(
         })
         .map_err(DownloadError::Other)?;
 
-    match layer_metadata.file_size() {
-        Some(expected) if expected != bytes_amount => {
-            return Err(DownloadError::Other(anyhow!(
-                "According to layer file metadata should have downloaded {expected} bytes but downloaded {bytes_amount} bytes into file '{}'",
-                temp_file_path.display()
-            )));
-        }
-        Some(_) | None => {
-            // matches, or upgrading from an earlier IndexPart version
-        }
+    let expected = layer_metadata.file_size();
+    if expected != bytes_amount {
+        return Err(DownloadError::Other(anyhow!(
+            "According to layer file metadata should have downloaded {expected} bytes but downloaded {bytes_amount} bytes into file {temp_file_path:?}",
+        )));
     }
 
     // not using sync_data because it can lose file size update
@@ -261,14 +256,12 @@ pub(super) async fn download_index_part(
     )
     .await?;
 
-    let index_part: IndexPartUnclean = serde_json::from_slice(&index_part_bytes)
+    let index_part: IndexPart = serde_json::from_slice(&index_part_bytes)
         .with_context(|| {
             format!("Failed to deserialize index part file into file {index_part_path:?}")
         })
         .map_err(DownloadError::Other)?;
 
-    let index_part = index_part.remove_unclean_layer_file_names();
-
     Ok(index_part)
 }
 

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -6,7 +6,6 @@ use std::collections::{HashMap, HashSet};
 
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
-use tracing::warn;
 
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::storage_layer::LayerFileName;
@@ -20,7 +19,7 @@ use utils::lsn::Lsn;
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
 #[cfg_attr(test, derive(Default))]
 pub struct LayerFileMetadata {
-    file_size: Option<u64>,
+    file_size: u64,
 }
 
 impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
@@ -33,36 +32,16 @@ impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
 
 impl LayerFileMetadata {
     pub fn new(file_size: u64) -> Self {
-        LayerFileMetadata {
-            file_size: Some(file_size),
-        }
+        LayerFileMetadata { file_size }
     }
 
-    /// This is used to initialize the metadata for remote layers, for which
-    /// the metadata was missing from the index part file.
-    pub const MISSING: Self = LayerFileMetadata { file_size: None };
-
-    pub fn file_size(&self) -> Option<u64> {
+    pub fn file_size(&self) -> u64 {
         self.file_size
     }
-
-    /// Metadata has holes due to version upgrades. This method is called to upgrade self with the
-    /// other value.
-    ///
-    /// This is called on the possibly outdated version. Returns true if any changes
-    /// were made.
-    pub fn merge(&mut self, other: &Self) -> bool {
-        let mut changed = false;
-
-        if self.file_size != other.file_size {
-            self.file_size = other.file_size.or(self.file_size);
-            changed = true;
-        }
-
-        changed
-    }
 }
 
+// TODO seems like another part of the remote storage file format
+// compatibility issue, see https://github.com/neondatabase/neon/issues/3072
 /// In-memory representation of an `index_part.json` file
 ///
 /// Contains the data about all files in the timeline, present remotely and its metadata.
@@ -71,10 +50,7 @@ impl LayerFileMetadata {
 /// remember to add a test case for the changed version.
 #[serde_as]
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
-pub struct IndexPartImpl<L>
-where
-    L: std::hash::Hash + PartialEq + Eq,
-{
+pub struct IndexPart {
     /// Debugging aid describing the version of this type.
     #[serde(default)]
     version: usize,
@@ -82,14 +58,13 @@ where
     /// Layer names, which are stored on the remote storage.
     ///
     /// Additional metadata can might exist in `layer_metadata`.
-    pub timeline_layers: HashSet<L>,
+    pub timeline_layers: HashSet<LayerFileName>,
 
     /// Per layer file name metadata, which can be present for a present or missing layer file.
     ///
     /// Older versions of `IndexPart` will not have this property or have only a part of metadata
     /// that latest version stores.
-    #[serde(default = "HashMap::default")]
-    pub layer_metadata: HashMap<L, IndexLayerMetadata>,
+    pub layer_metadata: HashMap<LayerFileName, IndexLayerMetadata>,
 
     // 'disk_consistent_lsn' is a copy of the 'disk_consistent_lsn' in the metadata.
     // It's duplicated here for convenience.
@@ -98,101 +73,6 @@ where
     metadata_bytes: Vec<u8>,
 }
 
-// TODO seems like another part of the remote storage file format
-// compatibility issue, see https://github.com/neondatabase/neon/issues/3072
-pub type IndexPart = IndexPartImpl<LayerFileName>;
-
-pub type IndexPartUnclean = IndexPartImpl<UncleanLayerFileName>;
-
-#[derive(Debug, PartialEq, Eq, Hash, Clone)]
-pub enum UncleanLayerFileName {
-    Clean(LayerFileName),
-    BackupFile(String),
-}
-
-impl<'de> serde::Deserialize<'de> for UncleanLayerFileName {
-    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
-    where
-        D: serde::Deserializer<'de>,
-    {
-        deserializer.deserialize_string(UncleanLayerFileNameVisitor)
-    }
-}
-
-struct UncleanLayerFileNameVisitor;
-
-impl<'de> serde::de::Visitor<'de> for UncleanLayerFileNameVisitor {
-    type Value = UncleanLayerFileName;
-
-    fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
-        write!(
-            formatter,
-            "a string that is a valid LayerFileName or '.old' backup file name"
-        )
-    }
-
-    fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
-    where
-        E: serde::de::Error,
-    {
-        let maybe_clean: Result<LayerFileName, _> = v.parse();
-        match maybe_clean {
-            Ok(clean) => Ok(UncleanLayerFileName::Clean(clean)),
-            Err(e) => {
-                if v.ends_with(".old") || v == "metadata_backup" {
-                    Ok(UncleanLayerFileName::BackupFile(v.to_owned()))
-                } else {
-                    Err(E::custom(e))
-                }
-            }
-        }
-    }
-}
-
-impl UncleanLayerFileName {
-    fn into_clean(self) -> Option<LayerFileName> {
-        match self {
-            UncleanLayerFileName::Clean(clean) => Some(clean),
-            UncleanLayerFileName::BackupFile(_) => None,
-        }
-    }
-}
-
-impl IndexPartUnclean {
-    pub fn remove_unclean_layer_file_names(self) -> IndexPart {
-        let IndexPartUnclean {
-            version,
-            timeline_layers,
-            layer_metadata,
-            disk_consistent_lsn,
-            metadata_bytes,
-        } = self;
-
-        IndexPart {
-            version,
-            timeline_layers: timeline_layers
-                .into_iter()
-                .filter_map(|unclean_file_name| match unclean_file_name {
-                    UncleanLayerFileName::Clean(clean_name) => Some(clean_name),
-                    UncleanLayerFileName::BackupFile(backup_file_name) => {
-                        // For details see https://github.com/neondatabase/neon/issues/3024
-                        warn!(
-                            "got backup file on the remote storage, ignoring it {backup_file_name}"
-                        );
-                        None
-                    }
-                })
-                .collect(),
-            layer_metadata: layer_metadata
-                .into_iter()
-                .filter_map(|(l, m)| l.into_clean().map(|l| (l, m)))
-                .collect(),
-            disk_consistent_lsn,
-            metadata_bytes,
-        }
-    }
-}
-
 impl IndexPart {
     /// When adding or modifying any parts of `IndexPart`, increment the version so that it can be
     /// used to understand later versions.
@@ -232,7 +112,7 @@ impl IndexPart {
 /// Serialized form of [`LayerFileMetadata`].
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Default)]
 pub struct IndexLayerMetadata {
-    pub(super) file_size: Option<u64>,
+    pub(super) file_size: u64,
 }
 
 impl From<&'_ LayerFileMetadata> for IndexLayerMetadata {
@@ -247,27 +127,6 @@ impl From<&'_ LayerFileMetadata> for IndexLayerMetadata {
 mod tests {
     use super::*;
 
-    #[test]
-    fn v0_indexpart_is_parsed() {
-        let example = r#"{
-            "timeline_layers":["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9"],
-            "disk_consistent_lsn":"0/16960E8",
-            "metadata_bytes":[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
-        }"#;
-
-        let expected = IndexPart {
-            version: 0,
-            timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
-            layer_metadata: HashMap::default(),
-            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
-            metadata_bytes: [113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
-        };
-
-        let part: IndexPartUnclean = serde_json::from_str(example).unwrap();
-        let part = part.remove_unclean_layer_file_names();
-        assert_eq!(part, expected);
-    }
-
     #[test]
     fn v1_indexpart_is_parsed() {
         let example = r#"{
@@ -287,21 +146,19 @@ mod tests {
             timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
-                    file_size: Some(25600000),
+                    file_size: 25600000,
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
-                    file_size: Some(9007199254741001),
+                    file_size: 9007199254741001,
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             metadata_bytes: [113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
         };
 
-        let part = serde_json::from_str::<IndexPartUnclean>(example)
-            .unwrap()
-            .remove_unclean_layer_file_names();
+        let part = serde_json::from_str::<IndexPart>(example).unwrap();
         assert_eq!(part, expected);
     }
 
@@ -325,20 +182,64 @@ mod tests {
             timeline_layers: HashSet::from(["000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap()]),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
-                    file_size: Some(25600000),
+                    file_size: 25600000,
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
-                    file_size: Some(9007199254741001),
+                    file_size: 9007199254741001,
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             metadata_bytes: [112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
         };
 
-        let part = serde_json::from_str::<IndexPartUnclean>(example).unwrap();
-        let part = part.remove_unclean_layer_file_names();
+        let part = serde_json::from_str::<IndexPart>(example).unwrap();
         assert_eq!(part, expected);
     }
+
+    #[test]
+    fn empty_layers_are_parsed() {
+        let empty_layers_json = r#"{
+            "version":1,
+            "timeline_layers":[],
+            "layer_metadata":{},
+            "disk_consistent_lsn":"0/2532648",
+            "metadata_bytes":[136,151,49,208,0,70,0,4,0,0,0,0,2,83,38,72,1,0,0,0,0,2,83,38,32,1,87,198,240,135,97,119,45,125,38,29,155,161,140,141,255,210,0,0,0,0,2,83,38,72,0,0,0,0,1,73,240,192,0,0,0,0,1,73,240,192,0,0,0,15,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
+        }"#;
+
+        let expected = IndexPart {
+            version: 1,
+            timeline_layers: HashSet::new(),
+            layer_metadata: HashMap::new(),
+            disk_consistent_lsn: "0/2532648".parse::<Lsn>().unwrap(),
+            metadata_bytes: [
+                136, 151, 49, 208, 0, 70, 0, 4, 0, 0, 0, 0, 2, 83, 38, 72, 1, 0, 0, 0, 0, 2, 83,
+                38, 32, 1, 87, 198, 240, 135, 97, 119, 45, 125, 38, 29, 155, 161, 140, 141, 255,
+                210, 0, 0, 0, 0, 2, 83, 38, 72, 0, 0, 0, 0, 1, 73, 240, 192, 0, 0, 0, 0, 1, 73,
+                240, 192, 0, 0, 0, 15, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                0, 0,
+            ]
+            .to_vec(),
+        };
+
+        let empty_layers_parsed = serde_json::from_str::<IndexPart>(empty_layers_json).unwrap();
+
+        assert_eq!(empty_layers_parsed, expected);
+    }
 }

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -64,13 +64,9 @@ pub(super) async fn upload_timeline_layer<'a>(
         })?
         .len();
 
-    // FIXME: this looks bad
-    if let Some(metadata_size) = known_metadata.file_size() {
-        if metadata_size != fs_size {
-            bail!("File {source_path:?} has its current FS size {fs_size} diferent from initially determined {metadata_size}");
-        }
-    } else {
-        // this is a silly state we would like to avoid
+    let metadata_size = known_metadata.file_size();
+    if metadata_size != fs_size {
+        bail!("File {source_path:?} has its current FS size {fs_size} diferent from initially determined {metadata_size}");
     }
 
     let fs_size = usize::try_from(fs_size).with_context(|| {

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
@@ -378,7 +385,7 @@ pub trait PersistentLayer: Layer {
     ///
     /// Should not change over the lifetime of the layer object because
     /// current_physical_size is computed as the som of this value.
-    fn file_size(&self) -> Option<u64>;
+    fn file_size(&self) -> u64;
 
     fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo;
 

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -444,8 +444,8 @@ impl PersistentLayer for DeltaLayer {
         Ok(())
     }
 
-    fn file_size(&self) -> Option<u64> {
-        Some(self.file_size)
+    fn file_size(&self) -> u64 {
+        self.file_size
     }
 
     fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
@@ -456,7 +456,7 @@ impl PersistentLayer for DeltaLayer {
 
         HistoricLayerInfo::Delta {
             layer_file_name,
-            layer_file_size: Some(self.file_size),
+            layer_file_size: self.file_size,
             lsn_start: lsn_range.start,
             lsn_end: lsn_range.end,
             remote: false,

pageserver/src/tenant/storage_layer/filename.rs

@@ -258,6 +258,15 @@ impl serde::Serialize for LayerFileName {
     }
 }
 
+impl<'de> serde::Deserialize<'de> for LayerFileName {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        deserializer.deserialize_string(LayerFileNameVisitor)
+    }
+}
+
 struct LayerFileNameVisitor;
 
 impl<'de> serde::de::Visitor<'de> for LayerFileNameVisitor {

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -258,8 +258,8 @@ impl PersistentLayer for ImageLayer {
         Ok(())
     }
 
-    fn file_size(&self) -> Option<u64> {
-        Some(self.file_size)
+    fn file_size(&self) -> u64 {
+        self.file_size
     }
 
     fn info(&self, reset: LayerAccessStatsReset) -> HistoricLayerInfo {
@@ -268,7 +268,7 @@ impl PersistentLayer for ImageLayer {
 
         HistoricLayerInfo::Image {
             layer_file_name,
-            layer_file_size: Some(self.file_size),
+            layer_file_size: self.file_size,
             lsn_start: lsn_range.start,
             remote: false,
             access_stats: self.access_stats.as_api_model(reset),

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -167,7 +167,7 @@ impl PersistentLayer for RemoteLayer {
         true
     }
 
-    fn file_size(&self) -> Option<u64> {
+    fn file_size(&self) -> u64 {
         self.layer_metadata.file_size()
     }
 

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.
@@ -328,27 +335,12 @@ 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()
-    }
-}
-
-/// Returned by [`Timeline::layer_size_sum`]
-pub enum LayerSizeSum {
-    /// The result is accurate.
-    Accurate(u64),
-    // We don't know the layer file size of one or more layers.
-    // They contribute to the sum with a value of 0.
-    // Hence, the sum is a lower bound for the actualy layer file size sum.
-    ApproximateLowerBound(u64),
-}
-
-impl LayerSizeSum {
-    pub fn approximate_is_ok(self) -> u64 {
-        match self {
-            LayerSizeSum::Accurate(v) => v,
-            LayerSizeSum::ApproximateLowerBound(v) => v,
+    /// 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,
         }
     }
 }
@@ -550,20 +542,13 @@ impl Timeline {
     /// The sum of the file size of all historic layers in the layer map.
     /// This method makes no distinction between local and remote layers.
     /// Hence, the result **does not represent local filesystem usage**.
-    pub fn layer_size_sum(&self) -> LayerSizeSum {
+    pub fn layer_size_sum(&self) -> u64 {
         let layer_map = self.layers.read().unwrap();
         let mut size = 0;
-        let mut no_size_cnt = 0;
         for l in layer_map.iter_historic_layers() {
-            let (l_size, l_no_size) = l.file_size().map(|s| (s, 0)).unwrap_or((0, 1));
-            size += l_size;
-            no_size_cnt += l_no_size;
-        }
-        if no_size_cnt == 0 {
-            LayerSizeSum::Accurate(size)
-        } else {
-            LayerSizeSum::ApproximateLowerBound(size)
+            size += l.file_size();
         }
+        size
     }
 
     pub fn get_resident_physical_size(&self) -> u64 {
@@ -694,8 +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! {
@@ -832,11 +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))
@@ -974,6 +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
@@ -1011,6 +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();
@@ -1044,12 +1056,28 @@ 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);
         }
 
-        let layer_file_size = local_layer
-            .file_size()
-            .expect("Local layer should have a file size");
+        let layer_file_size = local_layer.file_size();
+
+        let local_layer_mtime = local_layer
+            .local_path()
+            .expect("local layer should have a local path")
+            .metadata()
+            .context("get local layer file stat")?
+            .modified()
+            .context("get mtime of layer file")?;
+        let local_layer_residence_duration =
+            match SystemTime::now().duration_since(local_layer_mtime) {
+                Err(e) => {
+                    warn!("layer mtime is in the future: {}", e);
+                    None
+                }
+                Ok(delta) => Some(delta),
+            };
 
         let layer_metadata = LayerFileMetadata::new(layer_file_size);
 
@@ -1093,6 +1121,17 @@ impl Timeline {
                     .resident_physical_size_gauge
                     .sub(layer_file_size);
 
+                self.metrics.evictions.inc();
+
+                if let Some(delta) = local_layer_residence_duration {
+                    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
             }
             Replacement::NotFound => {
@@ -1167,7 +1206,7 @@ impl Timeline {
     pub(super) fn new(
         conf: &'static PageServerConf,
         tenant_conf: Arc<RwLock<TenantConfOpt>>,
-        metadata: TimelineMetadata,
+        metadata: &TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
         timeline_id: TimelineId,
         tenant_id: TenantId,
@@ -1208,7 +1247,14 @@ impl Timeline {
                 ancestor_timeline: ancestor,
                 ancestor_lsn: metadata.ancestor_lsn(),
 
-                metrics: TimelineMetrics::new(&tenant_id, &timeline_id),
+                metrics: TimelineMetrics::new(
+                    &tenant_id,
+                    &timeline_id,
+                    crate::metrics::EvictionsWithLowResidenceDurationBuilder::new(
+                        "mtime",
+                        conf.evictions_low_residence_duration_metric_threshold,
+                    ),
+                ),
 
                 flush_loop_state: Mutex::new(FlushLoopState::NotStarted),
 
@@ -1245,6 +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
@@ -1335,6 +1385,7 @@ impl Timeline {
             lagging_wal_timeout,
             max_lsn_wal_lag,
             crate::config::SAFEKEEPER_AUTH_TOKEN.get().cloned(),
+            self.conf.availability_zone.clone(),
             background_ctx,
         );
     }
@@ -1482,7 +1533,12 @@ impl Timeline {
                 .layer_metadata
                 .get(remote_layer_name)
                 .map(LayerFileMetadata::from)
-                .unwrap_or(LayerFileMetadata::MISSING);
+                .with_context(|| {
+                    format!(
+                        "No remote layer metadata found for layer {}",
+                        remote_layer_name.file_name()
+                    )
+                })?;
 
             // Is the local layer's size different from the size stored in the
             // remote index file?
@@ -1498,34 +1554,27 @@ impl Timeline {
                     local_layer_path.display()
                 );
 
-                if let Some(remote_size) = remote_layer_metadata.file_size() {
-                    let metadata = local_layer_path.metadata().with_context(|| {
-                        format!(
-                            "get file size of local layer {}",
-                            local_layer_path.display()
-                        )
-                    })?;
-                    let local_size = metadata.len();
-                    if local_size != remote_size {
-                        warn!("removing local file {local_layer_path:?} because it has unexpected length {local_size}; length in remote index is {remote_size}");
-                        if let Err(err) = rename_to_backup(&local_layer_path) {
-                            assert!(local_layer_path.exists(), "we would leave the local_layer without a file if this does not hold: {}", local_layer_path.display());
-                            anyhow::bail!("could not rename file {local_layer_path:?}: {err:?}");
-                        } else {
-                            self.metrics.resident_physical_size_gauge.sub(local_size);
-                            updates.remove_historic(local_layer);
-                            // fall-through to adding the remote layer
-                        }
+                let remote_size = remote_layer_metadata.file_size();
+                let metadata = local_layer_path.metadata().with_context(|| {
+                    format!(
+                        "get file size of local layer {}",
+                        local_layer_path.display()
+                    )
+                })?;
+                let local_size = metadata.len();
+                if local_size != remote_size {
+                    warn!("removing local file {local_layer_path:?} because it has unexpected length {local_size}; length in remote index is {remote_size}");
+                    if let Err(err) = rename_to_backup(&local_layer_path) {
+                        assert!(local_layer_path.exists(), "we would leave the local_layer without a file if this does not hold: {}", local_layer_path.display());
+                        anyhow::bail!("could not rename file {local_layer_path:?}: {err:?}");
                     } else {
-                        debug!(
-                            "layer is present locally and file size matches remote, using it: {}",
-                            local_layer_path.display()
-                        );
-                        continue;
+                        self.metrics.resident_physical_size_gauge.sub(local_size);
+                        updates.remove_historic(local_layer);
+                        // fall-through to adding the remote layer
                     }
                 } else {
                     debug!(
-                        "layer is present locally and remote does not have file size, using it: {}",
+                        "layer is present locally and file size matches remote, using it: {}",
                         local_layer_path.display()
                     );
                     continue;
@@ -1627,6 +1676,8 @@ impl Timeline {
             .map(|l| (l.filename(), l))
             .collect::<HashMap<_, _>>();
 
+        // If no writes happen, new branches do not have any layers, only the metadata file.
+        let has_local_layers = !local_layers.is_empty();
         let local_only_layers = match index_part {
             Some(index_part) => {
                 info!(
@@ -1644,28 +1695,47 @@ impl Timeline {
             }
         };
 
-        // Are there local files that don't exist remotely? Schedule uploads for them
-        for (layer_name, layer) in &local_only_layers {
-            // XXX solve this in the type system
-            let layer_path = layer
-                .local_path()
-                .expect("local_only_layers only contains local layers");
-            let layer_size = layer_path
-                .metadata()
-                .with_context(|| format!("failed to get file {layer_path:?} metadata"))?
-                .len();
-            info!("scheduling {layer_path:?} for upload");
-            remote_client
-                .schedule_layer_file_upload(layer_name, &LayerFileMetadata::new(layer_size))?;
+        if has_local_layers {
+            // Are there local files that don't exist remotely? Schedule uploads for them.
+            // Local timeline metadata will get uploaded to remove along witht he layers.
+            for (layer_name, layer) in &local_only_layers {
+                // XXX solve this in the type system
+                let layer_path = layer
+                    .local_path()
+                    .expect("local_only_layers only contains local layers");
+                let layer_size = layer_path
+                    .metadata()
+                    .with_context(|| format!("failed to get file {layer_path:?} metadata"))?
+                    .len();
+                info!("scheduling {layer_path:?} for upload");
+                remote_client
+                    .schedule_layer_file_upload(layer_name, &LayerFileMetadata::new(layer_size))?;
+            }
+            remote_client.schedule_index_upload_for_file_changes()?;
+        } else if index_part.is_none() {
+            // No data on the remote storage, no local layers, local metadata file.
+            //
+            // TODO https://github.com/neondatabase/neon/issues/3865
+            // Currently, console does not wait for the timeline data upload to the remote storage
+            // and considers the timeline created, expecting other pageserver nodes to work with it.
+            // Branch metadata upload could get interrupted (e.g pageserver got killed),
+            // hence any locally existing branch metadata with no remote counterpart should be uploaded,
+            // otherwise any other pageserver won't see the branch on `attach`.
+            //
+            // After the issue gets implemented, pageserver should rather remove the branch,
+            // since absence on S3 means we did not acknowledge the branch creation and console will have to retry,
+            // no need to keep the old files.
+            remote_client.schedule_index_upload_for_metadata_update(up_to_date_metadata)?;
+        } else {
+            // Local timeline has a metadata file, remote one too, both have no layers to sync.
         }
-        remote_client.schedule_index_upload_for_file_changes()?;
 
         info!("Done");
 
         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()
         {
@@ -1702,8 +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,
@@ -1758,6 +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);
@@ -1777,7 +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
             },
@@ -1788,20 +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 {
@@ -1830,7 +1906,6 @@ impl Timeline {
             "aborted because task_mgr shutdown requested".to_string()
         };
 
-        tokio::pin!(calculation);
         loop {
             tokio::select! {
                 res = &mut calculation => { return res }
@@ -1883,21 +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?;
@@ -1952,9 +2018,7 @@ impl Timeline {
     ) -> anyhow::Result<()> {
         if !layer.is_remote_layer() {
             layer.delete_resident_layer_file()?;
-            let layer_file_size = layer
-                .file_size()
-                .expect("Local layer should have a file size");
+            let layer_file_size = layer.file_size();
             self.metrics
                 .resident_physical_size_gauge
                 .sub(layer_file_size);
@@ -2715,10 +2779,22 @@ impl Timeline {
     ) -> Result<HashMap<LayerFileName, LayerFileMetadata>, PageReconstructError> {
         let timer = self.metrics.create_images_time_histo.start_timer();
         let mut image_layers: Vec<ImageLayer> = Vec::new();
+
+        // We need to avoid holes between generated image layers.
+        // Otherwise LayerMap::image_layer_exists will return false if key range of some layer is covered by more than one
+        // image layer with hole between them. In this case such layer can not be utilized by GC.
+        //
+        // How such hole between partitions can appear?
+        // if we have relation with relid=1 and size 100 and relation with relid=2 with size 200 then result of
+        // KeySpace::partition may contain partitions <100000000..100000099> and <200000000..200000199>.
+        // If there is delta layer <100000000..300000000> then it never be garbage collected because
+        // image layers  <100000000..100000099> and <200000000..200000199> are not completely covering it.
+        let mut start = Key::MIN;
+
         for partition in partitioning.parts.iter() {
+            let img_range = start..partition.ranges.last().unwrap().end;
+            start = img_range.end;
             if force || self.time_for_new_image_layer(partition, lsn)? {
-                let img_range =
-                    partition.ranges.first().unwrap().start..partition.ranges.last().unwrap().end;
                 let mut image_layer_writer = ImageLayerWriter::new(
                     self.conf,
                     self.timeline_id,
@@ -2732,7 +2808,6 @@ impl Timeline {
                         "failpoint image-layer-writer-fail-before-finish"
                     )))
                 });
-
                 for range in &partition.ranges {
                     let mut key = range.start;
                     while key < range.end {
@@ -3147,9 +3222,7 @@ impl Timeline {
             }
 
             fail_point!("delta-layer-writer-fail-before-finish", |_| {
-                return Err(
-                    anyhow::anyhow!("failpoint delta-layer-writer-fail-before-finish").into(),
-                );
+                Err(anyhow::anyhow!("failpoint delta-layer-writer-fail-before-finish").into())
             });
 
             writer.as_mut().unwrap().put_value(key, lsn, value)?;
@@ -3973,6 +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

@@ -1,26 +1,51 @@
-//! The per-timeline layer eviction task.
-
+//! The per-timeline layer eviction task, which evicts data which has not been accessed for more
+//! than a given threshold.
+//!
+//! Data includes all kinds of caches, namely:
+//! - (in-memory layers)
+//! - on-demand downloaded layer files on disk
+//! - (cached layer file pages)
+//! - derived data from layer file contents, namely:
+//!     - initial logical size
+//!     - partitioning
+//!     - (other currently missing unknowns)
+//!
+//! Items with parentheses are not (yet) touched by this task.
+//!
+//! 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};
 
 use crate::{
+    context::{DownloadBehavior, RequestContext},
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     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);
@@ -54,9 +79,10 @@ impl Timeline {
             }
         }
 
+        let ctx = RequestContext::new(TaskKind::Eviction, DownloadBehavior::Warn);
         loop {
             let policy = self.get_eviction_policy();
-            let cf = self.eviction_iteration(&policy, cancel.clone()).await;
+            let cf = self.eviction_iteration(&policy, &cancel, &ctx).await;
 
             match cf {
                 ControlFlow::Break(()) => break,
@@ -77,7 +103,8 @@ impl Timeline {
     async fn eviction_iteration(
         self: &Arc<Self>,
         policy: &EvictionPolicy,
-        cancel: CancellationToken,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
     ) -> ControlFlow<(), Instant> {
         debug!("eviction iteration: {policy:?}");
         match policy {
@@ -87,7 +114,7 @@ impl Timeline {
             }
             EvictionPolicy::LayerAccessThreshold(p) => {
                 let start = Instant::now();
-                match self.eviction_iteration_threshold(p, cancel).await {
+                match self.eviction_iteration_threshold(p, cancel, ctx).await {
                     ControlFlow::Break(()) => return ControlFlow::Break(()),
                     ControlFlow::Continue(()) => (),
                 }
@@ -101,10 +128,40 @@ impl Timeline {
     async fn eviction_iteration_threshold(
         self: &Arc<Self>,
         p: &EvictionPolicyLayerAccessThreshold,
-        cancel: CancellationToken,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
     ) -> 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 {
@@ -114,6 +171,7 @@ impl Timeline {
             not_evictable: usize,
             skipped_for_shutdown: usize,
         }
+
         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.
@@ -126,13 +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) => {
@@ -174,7 +226,7 @@ impl Timeline {
         };
 
         let results = match self
-            .evict_layer_batch(remote_client, &candidates[..], cancel)
+            .evict_layer_batch(remote_client, &candidates[..], cancel.clone())
             .await
         {
             Err(pre_err) => {
@@ -216,4 +268,144 @@ 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 imitate_timeline_cached_layer_accesses(
+        &self,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
+    ) {
+        let lsn = self.get_last_record_lsn();
+
+        // imitiate on-restart initial logical size
+        let size = self.calculate_logical_size(lsn, cancel.clone(), ctx).await;
+
+        match &size {
+            Ok(_size) => {
+                // good, don't log it to avoid confusion
+            }
+            Err(_) => {
+                // we have known issues for which we already log this on consumption metrics,
+                // gc, and compaction. leave logging out for now.
+                //
+                // https://github.com/neondatabase/neon/issues/2539
+            }
+        }
+
+        // imitiate repartiting on first compactation
+        if let Err(e) = self.collect_keyspace(lsn, ctx).await {
+            // if this failed, we probably failed logical size because these use the same keys
+            if size.is_err() {
+                // ignore, see above comment
+            } else {
+                warn!(
+                    "failed to collect keyspace but succeeded in calculating logical size: {e:#}"
+                );
+            }
+        }
+    }
+
+    // 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

@@ -45,6 +45,7 @@ pub fn spawn_connection_manager_task(
     lagging_wal_timeout: Duration,
     max_lsn_wal_lag: NonZeroU64,
     auth_token: Option<Arc<String>>,
+    availability_zone: Option<String>,
     ctx: RequestContext,
 ) {
     let mut broker_client = get_broker_client().clone();
@@ -67,6 +68,7 @@ pub fn spawn_connection_manager_task(
                 lagging_wal_timeout,
                 max_lsn_wal_lag,
                 auth_token,
+                availability_zone,
             );
             loop {
                 select! {
@@ -235,11 +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
         }
     }
@@ -334,6 +332,7 @@ struct WalreceiverState {
     /// Data about all timelines, available for connection, fetched from storage broker, grouped by their corresponding safekeeper node id.
     wal_stream_candidates: HashMap<NodeId, BrokerSkTimeline>,
     auth_token: Option<Arc<String>>,
+    availability_zone: Option<String>,
 }
 
 /// Current connection data.
@@ -343,6 +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.
@@ -381,6 +382,7 @@ impl WalreceiverState {
         lagging_wal_timeout: Duration,
         max_lsn_wal_lag: NonZeroU64,
         auth_token: Option<Arc<String>>,
+        availability_zone: Option<String>,
     ) -> Self {
         let id = TenantTimelineId {
             tenant_id: timeline.tenant_id,
@@ -396,16 +398,12 @@ impl WalreceiverState {
             wal_stream_candidates: HashMap::new(),
             wal_connection_retries: HashMap::new(),
             auth_token,
+            availability_zone,
         }
     }
 
     /// 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;
@@ -419,7 +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,
@@ -428,13 +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,
@@ -541,6 +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.
@@ -554,6 +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) =
@@ -564,6 +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,
@@ -589,6 +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,
@@ -596,6 +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"
@@ -663,6 +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,
@@ -681,10 +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,
                 });
@@ -740,6 +761,7 @@ impl WalreceiverState {
                         None => None,
                         Some(x) => Some(x),
                     },
+                    self.availability_zone.as_deref(),
                 ) {
                     Ok(connstr) => Some((*sk_id, info, connstr)),
                     Err(e) => {
@@ -788,6 +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,
@@ -802,6 +825,7 @@ enum ReconnectReason {
         new_commit_lsn: Lsn,
         threshold: NonZeroU64,
     },
+    SwitchAvailabilityZone,
     NoWalTimeout {
         current_lsn: Lsn,
         current_commit_lsn: Lsn,
@@ -824,17 +848,24 @@ fn wal_stream_connection_config(
     }: TenantTimelineId,
     listen_pg_addr_str: &str,
     auth_token: Option<&str>,
+    availability_zone: Option<&str>,
 ) -> anyhow::Result<PgConnectionConfig> {
     let (host, port) =
         parse_host_port(listen_pg_addr_str).context("Unable to parse listen_pg_addr_str")?;
     let port = port.unwrap_or(5432);
-    Ok(PgConnectionConfig::new_host_port(host, port)
+    let mut connstr = PgConnectionConfig::new_host_port(host, port)
         .extend_options([
             "-c".to_owned(),
             format!("timeline_id={}", timeline_id),
             format!("tenant_id={}", tenant_id),
         ])
-        .set_password(auth_token.map(|s| s.to_owned())))
+        .set_password(auth_token.map(|s| s.to_owned()));
+
+    if let Some(availability_zone) = availability_zone {
+        connstr = connstr.extend_options([format!("availability_zone={}", availability_zone)]);
+    }
+
+    Ok(connstr)
 }
 
 #[cfg(test)]
@@ -860,6 +891,7 @@ mod tests {
                 peer_horizon_lsn: 0,
                 local_start_lsn: 0,
                 safekeeper_connstr: safekeeper_connstr.to_owned(),
+                availability_zone: None,
             },
             latest_update,
         }
@@ -920,6 +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
@@ -1082,6 +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
@@ -1147,6 +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
@@ -1209,6 +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 {
@@ -1273,6 +1309,77 @@ mod tests {
             wal_stream_candidates: HashMap::new(),
             wal_connection_retries: HashMap::new(),
             auth_token: None,
+            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(())
+    }
 }