Proxy release 2025-07-29 06:01 UTC

- Remove some unused code
- Use `is_multiple_of()` instead of '%'
- Collapse consecuative "if let" statements
- Elided lifetime fixes

It is enough just to review the code of your team

.github/actionlint.yml

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

.github/workflows/benchbase_tpcc.yml

@@ -0,0 +1,384 @@
+name: TPC-C like benchmark using benchbase
+
+on:
+  schedule:
+    # * is a special character in YAML so you have to quote this string
+    #          ┌───────────── minute (0 - 59)
+    #          │ ┌───────────── hour (0 - 23)
+    #          │ │ ┌───────────── day of the month (1 - 31)
+    #          │ │ │ ┌───────────── month (1 - 12 or JAN-DEC)
+    #          │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
+    - cron:   '0 6 * * *' # run once a day at 6 AM UTC
+  workflow_dispatch: # adds ability to run this manually
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+
+concurrency:
+  # Allow only one workflow globally because we do not want to be too noisy in production environment
+  group: benchbase-tpcc-workflow
+  cancel-in-progress: false
+
+permissions:
+  contents: read
+
+jobs:
+  benchbase-tpcc:
+    strategy:
+      fail-fast: false # allow other variants to continue even if one fails
+      matrix:
+        include:
+          - warehouses: 50 # defines number of warehouses and is used to compute number of terminals
+            max_rate: 800  # measured max TPS at scale factor based on experiments. Adjust if performance is better/worse
+            min_cu: 0.25   # simulate free tier plan (0.25 -2 CU)
+            max_cu: 2
+          - warehouses: 500 # serverless plan (2-8 CU)
+            max_rate: 2000
+            min_cu: 2
+            max_cu: 8
+          - warehouses: 1000 # business plan (2-16 CU)
+            max_rate: 2900
+            min_cu: 2
+            max_cu: 16
+      max-parallel: 1 # we want to run each workload size sequentially to avoid noisy neighbors
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PG_CONFIG: /tmp/neon/pg_install/v17/bin/pg_config
+      PSQL: /tmp/neon/pg_install/v17/bin/psql
+      PG_17_LIB_PATH: /tmp/neon/pg_install/v17/lib
+      POSTGRES_VERSION: 17
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    timeout-minutes: 1440
+
+    steps:
+    - name: Harden the runner (Audit all outbound calls)
+      uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
+      with:
+        egress-policy: audit
+
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+
+    - name: Configure AWS credentials # necessary to download artefacts
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours is currently max associated with IAM role
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+
+    - name: Create Neon Project
+      id: create-neon-project-tpcc
+      uses: ./.github/actions/neon-project-create
+      with:
+        region_id: aws-us-east-2
+        postgres_version: ${{ env.POSTGRES_VERSION }}
+        compute_units: '[${{ matrix.min_cu }}, ${{ matrix.max_cu }}]'
+        api_key: ${{ secrets.NEON_PRODUCTION_API_KEY_4_BENCHMARKS }}
+        api_host: console.neon.tech  # production (!)
+
+    - name: Initialize Neon project
+      env:
+          BENCHMARK_TPCC_CONNSTR: ${{ steps.create-neon-project-tpcc.outputs.dsn }}
+          PROJECT_ID: ${{ steps.create-neon-project-tpcc.outputs.project_id }}
+      run: |
+        echo "Initializing Neon project with project_id: ${PROJECT_ID}"
+        export LD_LIBRARY_PATH=${PG_17_LIB_PATH}
+        
+        # Retry logic for psql connection with 1 minute sleep between attempts
+        for attempt in {1..3}; do
+          echo "Attempt ${attempt}/3: Creating extensions in Neon project"
+          if ${PSQL} "${BENCHMARK_TPCC_CONNSTR}" -c "CREATE EXTENSION IF NOT EXISTS neon; CREATE EXTENSION IF NOT EXISTS neon_utils;"; then
+            echo "Successfully created extensions"
+            break
+          else
+            echo "Failed to create extensions on attempt ${attempt}"
+            if [ ${attempt} -lt 3 ]; then
+              echo "Waiting 60 seconds before retry..."
+              sleep 60
+            else
+              echo "All attempts failed, exiting"
+              exit 1
+            fi
+          fi
+        done
+        
+        echo "BENCHMARK_TPCC_CONNSTR=${BENCHMARK_TPCC_CONNSTR}" >> $GITHUB_ENV
+
+    - name: Generate BenchBase workload configuration
+      env:
+        WAREHOUSES: ${{ matrix.warehouses }}
+        MAX_RATE: ${{ matrix.max_rate }}
+      run: |
+        echo "Generating BenchBase configs for warehouses: ${WAREHOUSES}, max_rate: ${MAX_RATE}"
+        
+        # Extract hostname and password from connection string
+        # Format: postgresql://username:password@hostname/database?params (no port for Neon)
+        HOSTNAME=$(echo "${BENCHMARK_TPCC_CONNSTR}" | sed -n 's|.*://[^:]*:[^@]*@\([^/]*\)/.*|\1|p')
+        PASSWORD=$(echo "${BENCHMARK_TPCC_CONNSTR}" | sed -n 's|.*://[^:]*:\([^@]*\)@.*|\1|p')
+        
+        echo "Extracted hostname: ${HOSTNAME}"
+        
+        # Use runner temp (NVMe) as working directory
+        cd "${RUNNER_TEMP}"
+        
+        # Copy the generator script
+        cp "${GITHUB_WORKSPACE}/test_runner/performance/benchbase_tpc_c_helpers/generate_workload_size.py" .
+        
+        # Generate configs and scripts
+        python3 generate_workload_size.py \
+          --warehouses ${WAREHOUSES} \
+          --max-rate ${MAX_RATE} \
+          --hostname ${HOSTNAME} \
+          --password ${PASSWORD} \
+          --runner-arch ${{ runner.arch }}
+        
+        # Fix path mismatch: move generated configs and scripts to expected locations
+        mv ../configs ./configs
+        mv ../scripts ./scripts
+
+    - name: Prepare database (load data)
+      env:
+        WAREHOUSES: ${{ matrix.warehouses }}
+      run: |
+        cd "${RUNNER_TEMP}"
+        
+        echo "Loading ${WAREHOUSES} warehouses into database..."
+        
+        # Run the loader script and capture output to log file while preserving stdout/stderr
+        ./scripts/load_${WAREHOUSES}_warehouses.sh 2>&1 | tee "load_${WAREHOUSES}_warehouses.log"
+        
+        echo "Database loading completed"
+
+    - name: Run TPC-C benchmark (warmup phase, then benchmark at 70% of configuredmax TPS)
+      env:
+        WAREHOUSES: ${{ matrix.warehouses }}
+      run: |
+        cd "${RUNNER_TEMP}"
+        
+        echo "Running TPC-C benchmark with ${WAREHOUSES} warehouses..."
+        
+        # Run the optimal rate benchmark
+        ./scripts/execute_${WAREHOUSES}_warehouses_opt_rate.sh
+        
+        echo "Benchmark execution completed"
+
+    - name: Run TPC-C benchmark (warmup phase, then ramp down TPS and up again in 5 minute intervals)
+
+      env:
+          WAREHOUSES: ${{ matrix.warehouses }}
+      run: |
+        cd "${RUNNER_TEMP}"
+        
+        echo "Running TPC-C ramp-down-up with ${WAREHOUSES} warehouses..."
+        
+        # Run the optimal rate benchmark
+        ./scripts/execute_${WAREHOUSES}_warehouses_ramp_up.sh
+        
+        echo "Benchmark execution completed"
+
+    - name: Process results (upload to test results database and generate diagrams)
+      env:
+        WAREHOUSES: ${{ matrix.warehouses }}
+        MIN_CU: ${{ matrix.min_cu }}
+        MAX_CU: ${{ matrix.max_cu }}
+        PROJECT_ID: ${{ steps.create-neon-project-tpcc.outputs.project_id }}
+        REVISION: ${{ github.sha }}
+        PERF_DB_CONNSTR: ${{ secrets.PERF_TEST_RESULT_CONNSTR }}
+      run: |
+        cd "${RUNNER_TEMP}"
+        
+        echo "Creating temporary Python environment for results processing..."
+        
+        # Create temporary virtual environment
+        python3 -m venv temp_results_env
+        source temp_results_env/bin/activate
+        
+        # Install required packages in virtual environment
+        pip install matplotlib pandas psycopg2-binary
+        
+        echo "Copying results processing scripts..."
+        
+        # Copy both processing scripts
+        cp "${GITHUB_WORKSPACE}/test_runner/performance/benchbase_tpc_c_helpers/generate_diagrams.py" .
+        cp "${GITHUB_WORKSPACE}/test_runner/performance/benchbase_tpc_c_helpers/upload_results_to_perf_test_results.py" .
+        
+        echo "Processing load phase metrics..."
+        
+        # Find and process load log
+        LOAD_LOG=$(find . -name "load_${WAREHOUSES}_warehouses.log" -type f | head -1)
+        if [ -n "$LOAD_LOG" ]; then
+          echo "Processing load metrics from: $LOAD_LOG"
+          python upload_results_to_perf_test_results.py \
+            --load-log "$LOAD_LOG" \
+            --run-type "load" \
+            --warehouses "${WAREHOUSES}" \
+            --min-cu "${MIN_CU}" \
+            --max-cu "${MAX_CU}" \
+            --project-id "${PROJECT_ID}" \
+            --revision "${REVISION}" \
+            --connection-string "${PERF_DB_CONNSTR}"
+        else
+          echo "Warning: Load log file not found: load_${WAREHOUSES}_warehouses.log"
+        fi
+        
+        echo "Processing warmup results for optimal rate..."
+        
+        # Find and process warmup results
+        WARMUP_CSV=$(find results_warmup -name "*.results.csv" -type f | head -1)
+        WARMUP_JSON=$(find results_warmup -name "*.summary.json" -type f | head -1)
+        
+        if [ -n "$WARMUP_CSV" ] && [ -n "$WARMUP_JSON" ]; then
+          echo "Generating warmup diagram from: $WARMUP_CSV"
+          python generate_diagrams.py \
+            --input-csv "$WARMUP_CSV" \
+            --output-svg "warmup_${WAREHOUSES}_warehouses_performance.svg" \
+            --title-suffix "Warmup at max TPS"
+            
+          echo "Uploading warmup metrics from: $WARMUP_JSON"
+          python upload_results_to_perf_test_results.py \
+            --summary-json "$WARMUP_JSON" \
+            --results-csv "$WARMUP_CSV" \
+            --run-type "warmup" \
+            --min-cu "${MIN_CU}" \
+            --max-cu "${MAX_CU}" \
+            --project-id "${PROJECT_ID}" \
+            --revision "${REVISION}" \
+            --connection-string "${PERF_DB_CONNSTR}"
+        else
+          echo "Warning: Missing warmup results files (CSV: $WARMUP_CSV, JSON: $WARMUP_JSON)"
+        fi
+        
+        echo "Processing optimal rate results..."
+        
+        # Find and process optimal rate results  
+        OPTRATE_CSV=$(find results_opt_rate -name "*.results.csv" -type f | head -1)
+        OPTRATE_JSON=$(find results_opt_rate -name "*.summary.json" -type f | head -1)
+        
+        if [ -n "$OPTRATE_CSV" ] && [ -n "$OPTRATE_JSON" ]; then
+          echo "Generating optimal rate diagram from: $OPTRATE_CSV"
+          python generate_diagrams.py \
+            --input-csv "$OPTRATE_CSV" \
+            --output-svg "benchmark_${WAREHOUSES}_warehouses_performance.svg" \
+            --title-suffix "70% of max TPS"
+            
+          echo "Uploading optimal rate metrics from: $OPTRATE_JSON"
+          python upload_results_to_perf_test_results.py \
+            --summary-json "$OPTRATE_JSON" \
+            --results-csv "$OPTRATE_CSV" \
+            --run-type "opt-rate" \
+            --min-cu "${MIN_CU}" \
+            --max-cu "${MAX_CU}" \
+            --project-id "${PROJECT_ID}" \
+            --revision "${REVISION}" \
+            --connection-string "${PERF_DB_CONNSTR}"
+        else
+          echo "Warning: Missing optimal rate results files (CSV: $OPTRATE_CSV, JSON: $OPTRATE_JSON)"
+        fi
+
+        echo "Processing warmup 2 results for ramp down/up phase..."
+        
+        # Find and process warmup results
+        WARMUP_CSV=$(find results_warmup -name "*.results.csv" -type f | tail -1)
+        WARMUP_JSON=$(find results_warmup -name "*.summary.json" -type f | tail -1)
+        
+        if [ -n "$WARMUP_CSV" ] && [ -n "$WARMUP_JSON" ]; then
+          echo "Generating warmup diagram from: $WARMUP_CSV"
+          python generate_diagrams.py \
+            --input-csv "$WARMUP_CSV" \
+            --output-svg "warmup_2_${WAREHOUSES}_warehouses_performance.svg" \
+            --title-suffix "Warmup at max TPS"
+            
+          echo "Uploading warmup metrics from: $WARMUP_JSON"
+          python upload_results_to_perf_test_results.py \
+            --summary-json "$WARMUP_JSON" \
+            --results-csv "$WARMUP_CSV" \
+            --run-type "warmup" \
+            --min-cu "${MIN_CU}" \
+            --max-cu "${MAX_CU}" \
+            --project-id "${PROJECT_ID}" \
+            --revision "${REVISION}" \
+            --connection-string "${PERF_DB_CONNSTR}"
+        else
+          echo "Warning: Missing warmup results files (CSV: $WARMUP_CSV, JSON: $WARMUP_JSON)"
+        fi
+        
+        echo "Processing ramp results..."
+        
+        # Find and process ramp results  
+        RAMPUP_CSV=$(find results_ramp_up -name "*.results.csv" -type f | head -1)
+        RAMPUP_JSON=$(find results_ramp_up -name "*.summary.json" -type f | head -1)
+        
+        if [ -n "$RAMPUP_CSV" ] && [ -n "$RAMPUP_JSON" ]; then
+          echo "Generating ramp diagram from: $RAMPUP_CSV"
+          python generate_diagrams.py \
+            --input-csv "$RAMPUP_CSV" \
+            --output-svg "ramp_${WAREHOUSES}_warehouses_performance.svg" \
+            --title-suffix "ramp TPS down and up in 5 minute intervals"
+            
+          echo "Uploading ramp metrics from: $RAMPUP_JSON"
+          python upload_results_to_perf_test_results.py \
+            --summary-json "$RAMPUP_JSON" \
+            --results-csv "$RAMPUP_CSV" \
+            --run-type "ramp-up" \
+            --min-cu "${MIN_CU}" \
+            --max-cu "${MAX_CU}" \
+            --project-id "${PROJECT_ID}" \
+            --revision "${REVISION}" \
+            --connection-string "${PERF_DB_CONNSTR}"
+        else
+          echo "Warning: Missing ramp results files (CSV: $RAMPUP_CSV, JSON: $RAMPUP_JSON)"
+        fi
+        
+        # Deactivate and clean up virtual environment
+        deactivate
+        rm -rf temp_results_env
+        rm upload_results_to_perf_test_results.py
+        
+        echo "Results processing completed and environment cleaned up"
+
+    - name: Set date for upload
+      id: set-date
+      run: echo "date=$(date +%Y-%m-%d)" >> $GITHUB_OUTPUT
+
+    - name: Configure AWS credentials # necessary to upload results
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: us-east-2
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 900 # 900 is minimum value 
+        
+    - name: Upload benchmark results to S3
+      env:
+        S3_BUCKET: neon-public-benchmark-results
+        S3_PREFIX: benchbase-tpc-c/${{ steps.set-date.outputs.date }}/${{ github.run_id }}/${{ matrix.warehouses }}-warehouses
+      run: |
+        echo "Redacting passwords from configuration files before upload..."
+        
+        # Mask all passwords in XML config files
+        find "${RUNNER_TEMP}/configs" -name "*.xml" -type f -exec sed -i 's|<password>[^<]*</password>|<password>redacted</password>|g' {} \;
+        
+        echo "Uploading benchmark results to s3://${S3_BUCKET}/${S3_PREFIX}/"
+        
+        # Upload the entire benchmark directory recursively
+        aws s3 cp --only-show-errors --recursive "${RUNNER_TEMP}" s3://${S3_BUCKET}/${S3_PREFIX}/
+        
+        echo "Upload completed"
+        
+    - name: Delete Neon Project
+      if: ${{ always() }}
+      uses: ./.github/actions/neon-project-delete
+      with:
+        project_id: ${{ steps.create-neon-project-tpcc.outputs.project_id }}
+        api_key: ${{ secrets.NEON_PRODUCTION_API_KEY_4_BENCHMARKS }} 
+        api_host: console.neon.tech  # production (!)

.github/workflows/benchmarking.yml

@@ -418,7 +418,7 @@ jobs:
       statuses: write
       id-token: write # aws-actions/configure-aws-credentials
     env:
-      PGBENCH_SIZE: ${{ vars.PREWARM_PGBENCH_SIZE }}
+      PROJECT_ID: ${{ vars.PREWARM_PROJECT_ID }}
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
       DEFAULT_PG_VERSION: 17
       TEST_OUTPUT: /tmp/test_output

.github/workflows/pg-clients.yml

@@ -48,8 +48,20 @@ jobs:
     uses: ./.github/workflows/build-build-tools-image.yml
     secrets: inherit
 
+  generate-ch-tmppw:
+    runs-on: ubuntu-22.04
+    outputs:
+      tmp_val: ${{ steps.pwgen.outputs.tmp_val }}
+    steps:
+      - name: Generate a random password
+        id: pwgen
+        run: |
+          set +x
+          p=$(dd if=/dev/random bs=14 count=1 2>/dev/null | base64)
+          echo tmp_val="${p//\//}" >> "${GITHUB_OUTPUT}"
+
   test-logical-replication:
-    needs: [ build-build-tools-image ]
+    needs: [ build-build-tools-image, generate-ch-tmppw ]
     runs-on: ubuntu-22.04
 
     container:
@@ -60,16 +72,20 @@ jobs:
       options: --init --user root
     services:
       clickhouse:
-        image: clickhouse/clickhouse-server:24.6.3.64
+        image: clickhouse/clickhouse-server:24.8
+        env:
+          CLICKHOUSE_PASSWORD: ${{ needs.generate-ch-tmppw.outputs.tmp_val }}
         ports:
           - 9000:9000
           - 8123:8123
       zookeeper:
-        image: quay.io/debezium/zookeeper:2.7
+        image: quay.io/debezium/zookeeper:3.1.3.Final
         ports:
           - 2181:2181
+          - 2888:2888
+          - 3888:3888
       kafka:
-        image: quay.io/debezium/kafka:2.7
+        image: quay.io/debezium/kafka:3.1.3.Final
         env:
           ZOOKEEPER_CONNECT: "zookeeper:2181"
           KAFKA_ADVERTISED_LISTENERS: PLAINTEXT://kafka:9092
@@ -79,7 +95,7 @@ jobs:
         ports:
           - 9092:9092
       debezium:
-        image: quay.io/debezium/connect:2.7
+        image: quay.io/debezium/connect:3.1.3.Final
         env:
           BOOTSTRAP_SERVERS: kafka:9092
           GROUP_ID: 1
@@ -125,6 +141,7 @@ jobs:
           aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
         env:
           BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
+          CLICKHOUSE_PASSWORD: ${{ needs.generate-ch-tmppw.outputs.tmp_val }}
 
       - name: Delete Neon Project
         if: always()

.github/workflows/proxy-benchmark.yml

@@ -3,7 +3,7 @@ name: Periodic proxy performance test on unit-perf-aws-arm runners
 on:
   push: # TODO: remove after testing
     branches:
-      - test-proxy-bench # Runs on pushes to branches starting with test-proxy-bench
+      - test-proxy-bench # Runs on pushes to test-proxy-bench branch
   # schedule:
     # * is a special character in YAML so you have to quote this string
     #        ┌───────────── minute (0 - 59)
@@ -32,7 +32,7 @@ jobs:
       statuses: write
       contents: write
       pull-requests: write
-    runs-on: [self-hosted, unit-perf-aws-arm]
+    runs-on: [ self-hosted, unit-perf-aws-arm ]
     timeout-minutes: 60  # 1h timeout
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm
@@ -55,30 +55,58 @@ jobs:
         {
           echo "PROXY_BENCH_PATH=$PROXY_BENCH_PATH"
           echo "NEON_DIR=${RUNNER_TEMP}/neon"
+          echo "NEON_PROXY_PATH=${RUNNER_TEMP}/neon/bin/proxy"
           echo "TEST_OUTPUT=${PROXY_BENCH_PATH}/test_output"
           echo ""
         } >> "$GITHUB_ENV"
 
-    - name: Run proxy-bench
-      run: ${PROXY_BENCH_PATH}/run.sh
+    - name: Cache poetry deps
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/pypoetry/virtualenvs
+        key: v2-${{ runner.os }}-${{ runner.arch }}-python-deps-bookworm-${{ hashFiles('poetry.lock') }}
 
-    - name: Ingest Bench Results # neon repo script
+    - name: Install Python deps
+      shell: bash -euxo pipefail {0}
+      run: ./scripts/pysync
+
+    - name: show ulimits
+      shell: bash -euxo pipefail {0}
+      run: |
+        ulimit -a
+
+    - name: Run proxy-bench
+      working-directory: ${{ env.PROXY_BENCH_PATH }}
+      run: ./run.sh --with-grafana --bare-metal
+
+    - name: Ingest Bench Results
       if: always()
+      working-directory: ${{ env.NEON_DIR }}
       run: |
         mkdir -p $TEST_OUTPUT
         python $NEON_DIR/scripts/proxy_bench_results_ingest.py --out $TEST_OUTPUT
 
     - name: Push Metrics to Proxy perf database
+      shell: bash -euxo pipefail {0}
       if: always()
       env:
         PERF_TEST_RESULT_CONNSTR: "${{ secrets.PROXY_TEST_RESULT_CONNSTR }}"
         REPORT_FROM: $TEST_OUTPUT
+      working-directory: ${{ env.NEON_DIR }}
       run: $NEON_DIR/scripts/generate_and_push_perf_report.sh
 
-    - name: Docker cleanup
-      if: always()
-      run: docker compose down
-
     - name: Notify Failure
       if: failure()
-      run: echo "Proxy bench job failed" && exit 1
+      run: echo "Proxy bench job failed" && exit 1
+
+    - name: Cleanup Test Resources
+      if: always()
+      shell: bash -euxo pipefail {0}
+      run: |
+        # Cleanup the test resources
+        if [[ -d "${TEST_OUTPUT}" ]]; then
+          rm -rf ${TEST_OUTPUT}
+        fi
+        if [[ -d "${PROXY_BENCH_PATH}/test_output" ]]; then
+          rm -rf ${PROXY_BENCH_PATH}/test_output
+        fi

Cargo.lock

@@ -211,11 +211,11 @@ dependencies = [
 
 [[package]]
 name = "async-lock"
-version = "3.2.0"
+version = "3.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7125e42787d53db9dd54261812ef17e937c95a51e4d291373b670342fa44310c"
+checksum = "ff6e472cdea888a4bd64f342f09b3f50e1886d32afe8df3d663c01140b811b18"
 dependencies = [
- "event-listener 4.0.0",
+ "event-listener 5.4.0",
  "event-listener-strategy",
  "pin-project-lite",
 ]
@@ -1388,6 +1388,7 @@ dependencies = [
  "tower-http",
  "tower-otel",
  "tracing",
+ "tracing-appender",
  "tracing-opentelemetry",
  "tracing-subscriber",
  "tracing-utils",
@@ -1403,9 +1404,9 @@ dependencies = [
 
 [[package]]
 name = "concurrent-queue"
-version = "2.3.0"
+version = "2.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f057a694a54f12365049b0958a1685bb52d567f5593b355fbf685838e873d400"
+checksum = "4ca0197aee26d1ae37445ee532fefce43251d24cc7c166799f4d46817f1d3973"
 dependencies = [
  "crossbeam-utils",
 ]
@@ -2231,9 +2232,9 @@ checksum = "0206175f82b8d6bf6652ff7d71a1e27fd2e4efde587fd368662814d6ec1d9ce0"
 
 [[package]]
 name = "event-listener"
-version = "4.0.0"
+version = "5.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "770d968249b5d99410d61f5bf89057f3199a077a04d087092f58e7d10692baae"
+checksum = "3492acde4c3fc54c845eaab3eed8bd00c7a7d881f78bfc801e43a93dec1331ae"
 dependencies = [
  "concurrent-queue",
  "parking",
@@ -2242,11 +2243,11 @@ dependencies = [
 
 [[package]]
 name = "event-listener-strategy"
-version = "0.4.0"
+version = "0.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "958e4d70b6d5e81971bebec42271ec641e7ff4e170a6fa605f2b8a8b65cb97d3"
+checksum = "8be9f3dfaaffdae2972880079a491a1a8bb7cbed0b8dd7a347f668b4150a3b93"
 dependencies = [
- "event-listener 4.0.0",
+ "event-listener 5.4.0",
  "pin-project-lite",
 ]
 
@@ -2515,6 +2516,20 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "304de19db7028420975a296ab0fcbbc8e69438c4ed254a1e41e2a7f37d5f0e0a"
 
+[[package]]
+name = "generator"
+version = "0.8.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d18470a76cb7f8ff746cf1f7470914f900252ec36bbc40b569d74b1258446827"
+dependencies = [
+ "cc",
+ "cfg-if",
+ "libc",
+ "log",
+ "rustversion",
+ "windows 0.61.3",
+]
+
 [[package]]
 name = "generic-array"
 version = "0.14.7"
@@ -2833,7 +2848,7 @@ checksum = "f9c7c7c8ac16c798734b8a24560c1362120597c40d5e1459f09498f8f6c8f2ba"
 dependencies = [
  "cfg-if",
  "libc",
- "windows",
+ "windows 0.52.0",
 ]
 
 [[package]]
@@ -3104,7 +3119,7 @@ dependencies = [
  "iana-time-zone-haiku",
  "js-sys",
  "wasm-bindgen",
- "windows-core",
+ "windows-core 0.52.0",
 ]
 
 [[package]]
@@ -3655,6 +3670,19 @@ version = "0.4.26"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "30bde2b3dc3671ae49d8e2e9f044c7c005836e7a023ee57cffa25ab82764bb9e"
 
+[[package]]
+name = "loom"
+version = "0.7.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "419e0dc8046cb947daa77eb95ae174acfbddb7673b4151f56d1eed8e93fbfaca"
+dependencies = [
+ "cfg-if",
+ "generator",
+ "scoped-tls",
+ "tracing",
+ "tracing-subscriber",
+]
+
 [[package]]
 name = "lru"
 version = "0.12.3"
@@ -3871,6 +3899,25 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
+[[package]]
+name = "moka"
+version = "0.12.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9321642ca94a4282428e6ea4af8cc2ca4eac48ac7a6a4ea8f33f76d0ce70926"
+dependencies = [
+ "crossbeam-channel",
+ "crossbeam-epoch",
+ "crossbeam-utils",
+ "loom",
+ "parking_lot 0.12.1",
+ "portable-atomic",
+ "rustc_version",
+ "smallvec",
+ "tagptr",
+ "thiserror 1.0.69",
+ "uuid",
+]
+
 [[package]]
 name = "multimap"
 version = "0.8.3"
@@ -5384,7 +5431,6 @@ dependencies = [
  "futures",
  "gettid",
  "hashbrown 0.14.5",
- "hashlink",
  "hex",
  "hmac",
  "hostname",
@@ -5406,6 +5452,7 @@ dependencies = [
  "lasso",
  "measured",
  "metrics",
+ "moka",
  "once_cell",
  "opentelemetry",
  "ouroboros",
@@ -6419,6 +6466,12 @@ dependencies = [
  "pin-project-lite",
 ]
 
+[[package]]
+name = "scoped-tls"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1cf6437eb19a8f4a6cc0f7dca544973b0b78843adbfeb3683d1a94a0024a294"
+
 [[package]]
 name = "scopeguard"
 version = "1.1.0"
@@ -7268,6 +7321,12 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "tagptr"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7b2093cf4c8eb1e67749a6762251bc9cd836b6fc171623bd0a9d324d37af2417"
+
 [[package]]
 name = "tar"
 version = "0.4.40"
@@ -7934,11 +7993,12 @@ dependencies = [
 
 [[package]]
 name = "tracing-appender"
-version = "0.2.2"
+version = "0.2.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09d48f71a791638519505cefafe162606f706c25592e4bde4d97600c0195312e"
+checksum = "3566e8ce28cc0a3fe42519fc80e6b4c943cc4c8cef275620eb8dac2d3d4e06cf"
 dependencies = [
  "crossbeam-channel",
+ "thiserror 1.0.69",
  "time",
  "tracing-subscriber",
 ]
@@ -8636,10 +8696,32 @@ version = "0.52.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e48a53791691ab099e5e2ad123536d0fff50652600abaf43bbf952894110d0be"
 dependencies = [
- "windows-core",
+ "windows-core 0.52.0",
  "windows-targets 0.52.6",
 ]
 
+[[package]]
+name = "windows"
+version = "0.61.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9babd3a767a4c1aef6900409f85f5d53ce2544ccdfaa86dad48c91782c6d6893"
+dependencies = [
+ "windows-collections",
+ "windows-core 0.61.2",
+ "windows-future",
+ "windows-link",
+ "windows-numerics",
+]
+
+[[package]]
+name = "windows-collections"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3beeceb5e5cfd9eb1d76b381630e82c4241ccd0d27f1a39ed41b2760b255c5e8"
+dependencies = [
+ "windows-core 0.61.2",
+]
+
 [[package]]
 name = "windows-core"
 version = "0.52.0"
@@ -8649,6 +8731,86 @@ dependencies = [
  "windows-targets 0.52.6",
 ]
 
+[[package]]
+name = "windows-core"
+version = "0.61.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c0fdd3ddb90610c7638aa2b3a3ab2904fb9e5cdbecc643ddb3647212781c4ae3"
+dependencies = [
+ "windows-implement",
+ "windows-interface",
+ "windows-link",
+ "windows-result",
+ "windows-strings",
+]
+
+[[package]]
+name = "windows-future"
+version = "0.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fc6a41e98427b19fe4b73c550f060b59fa592d7d686537eebf9385621bfbad8e"
+dependencies = [
+ "windows-core 0.61.2",
+ "windows-link",
+ "windows-threading",
+]
+
+[[package]]
+name = "windows-implement"
+version = "0.60.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a47fddd13af08290e67f4acabf4b459f647552718f683a7b415d290ac744a836"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
+
+[[package]]
+name = "windows-interface"
+version = "0.59.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bd9211b69f8dcdfa817bfd14bf1c97c9188afa36f4750130fcdf3f400eca9fa8"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
+
+[[package]]
+name = "windows-link"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e6ad25900d524eaabdbbb96d20b4311e1e7ae1699af4fb28c17ae66c80d798a"
+
+[[package]]
+name = "windows-numerics"
+version = "0.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9150af68066c4c5c07ddc0ce30421554771e528bde427614c61038bc2c92c2b1"
+dependencies = [
+ "windows-core 0.61.2",
+ "windows-link",
+]
+
+[[package]]
+name = "windows-result"
+version = "0.3.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "56f42bd332cc6c8eac5af113fc0c1fd6a8fd2aa08a0119358686e5160d0586c6"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-strings"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "56e6c93f3a0c3b36176cb1327a4958a0353d5d166c2a35cb268ace15e91d3b57"
+dependencies = [
+ "windows-link",
+]
+
 [[package]]
 name = "windows-sys"
 version = "0.48.0"
@@ -8707,6 +8869,15 @@ dependencies = [
  "windows_x86_64_msvc 0.52.6",
 ]
 
+[[package]]
+name = "windows-threading"
+version = "0.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b66463ad2e0ea3bbf808b7f1d371311c80e115c0b71d60efc142cafbcfb057a6"
+dependencies = [
+ "windows-link",
+]
+
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.48.0"
@@ -8843,6 +9014,8 @@ dependencies = [
  "clap",
  "clap_builder",
  "const-oid",
+ "crossbeam-epoch",
+ "crossbeam-utils",
  "crypto-bigint 0.5.5",
  "der 0.7.8",
  "deranged",
@@ -8888,6 +9061,7 @@ dependencies = [
  "once_cell",
  "p256 0.13.2",
  "parquet",
+ "portable-atomic",
  "prettyplease",
  "proc-macro2",
  "prost 0.13.5",

Cargo.toml

@@ -46,10 +46,10 @@ members = [
     "libs/proxy/json",
     "libs/proxy/postgres-protocol2",
     "libs/proxy/postgres-types2",
+    "libs/proxy/subzero_core",
     "libs/proxy/tokio-postgres2",
     "endpoint_storage",
     "pgxn/neon/communicator",
-    "proxy/subzero_core",
 ]
 
 [workspace.package]
@@ -136,6 +136,7 @@ md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
 memoffset = "0.9"
+moka = { version = "0.12", features = ["sync"] }
 nix = { version = "0.30.1", features = ["dir", "fs", "mman", "process", "socket", "signal", "poll"] }
 # Do not update to >= 7.0.0, at least. The update will have a significant impact
 # on compute startup metrics (start_postgres_ms), >= 25% degradation.
@@ -144,8 +145,8 @@ num_cpus = "1.15"
 num-traits = "0.2.19"
 once_cell = "1.13"
 opentelemetry = "0.30"
-opentelemetry_sdk = { version = "0.30", features = ["rt-tokio", "experimental_trace_batch_span_processor_with_async_runtime"] }
-opentelemetry-otlp = { version = "0.30", default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry_sdk = "0.30"
+opentelemetry-otlp = { version = "0.30", default-features = false, features = ["http-proto", "trace", "http", "reqwest-blocking-client"] }
 opentelemetry-semantic-conventions = "0.30"
 parking_lot = "0.12"
 parquet = { version = "53", default-features = false, features = ["zstd"] }
@@ -222,6 +223,7 @@ tracing-log = "0.2"
 tracing-opentelemetry = "0.31"
 tracing-serde = "0.2.0"
 tracing-subscriber = { version = "0.3", default-features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
+tracing-appender = "0.2.3"
 try-lock = "0.2.5"
 test-log = { version = "0.2.17", default-features = false, features = ["log"] }
 twox-hash = { version = "1.6.3", default-features = false }

build-tools/Dockerfile

@@ -39,13 +39,13 @@ COPY build-tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
 
 RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         set -e && \
-        apt update && \
-        apt install -y --no-install-recommends \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
         ca-certificates wget gpg && \
         wget -qO - https://www.postgresql.org/media/keys/ACCC4CF8.asc | gpg --dearmor -o /usr/share/keyrings/postgresql-keyring.gpg && \
         echo "deb [signed-by=/usr/share/keyrings/postgresql-keyring.gpg] http://apt.postgresql.org/pub/repos/apt bookworm-pgdg main" > /etc/apt/sources.list.d/pgdg.list && \
         apt-get update && \
-        apt install -y --no-install-recommends \
+        apt-get install -y --no-install-recommends \
         build-essential \
         autotools-dev \
         libedit-dev \
@@ -89,8 +89,7 @@ RUN useradd -ms /bin/bash nonroot -b /home
 # Use strict mode for bash to catch errors early
 SHELL ["/bin/bash", "-euo", "pipefail", "-c"]
 
-RUN mkdir -p /pgcopydb/bin && \
-    mkdir -p /pgcopydb/lib && \
+RUN mkdir -p /pgcopydb/{bin,lib} && \    
     chmod -R 755 /pgcopydb && \
     chown -R nonroot:nonroot /pgcopydb
 
@@ -106,8 +105,8 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
 # 'gdb' is included so that we get backtraces of core dumps produced in
 # regression tests
 RUN set -e \
-    && apt update \
-    && apt install -y \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends \
         autoconf \
         automake \
         bison \
@@ -183,22 +182,22 @@ RUN curl -sL "https://github.com/peak/s5cmd/releases/download/v${S5CMD_VERSION}/
 ENV LLVM_VERSION=20
 RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && echo "deb http://apt.llvm.org/${DEBIAN_VERSION}/ llvm-toolchain-${DEBIAN_VERSION}-${LLVM_VERSION} main" > /etc/apt/sources.list.d/llvm.stable.list \
-    && apt update \
-    && apt install -y clang-${LLVM_VERSION} llvm-${LLVM_VERSION} \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends clang-${LLVM_VERSION} llvm-${LLVM_VERSION} \
     && bash -c 'for f in /usr/bin/clang*-${LLVM_VERSION} /usr/bin/llvm*-${LLVM_VERSION}; do ln -s "${f}" "${f%-${LLVM_VERSION}}"; done' \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
 # Install node
 ENV NODE_VERSION=24
 RUN curl -fsSL https://deb.nodesource.com/setup_${NODE_VERSION}.x | bash - \
-    && apt install -y nodejs \
+    && apt-get install -y --no-install-recommends nodejs \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
 # Install docker
 RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /usr/share/keyrings/docker-archive-keyring.gpg \
     && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/docker-archive-keyring.gpg] https://download.docker.com/linux/debian ${DEBIAN_VERSION} stable" > /etc/apt/sources.list.d/docker.list \
-    && apt update \
-    && apt install -y docker-ce docker-ce-cli \
+    && apt-get update \
+    && apt-get install -y --no-install-recommends docker-ce docker-ce-cli \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
 # Configure sudo & docker
@@ -215,12 +214,11 @@ RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-$(uname -m).zip" -o "aws
 # Mold: A Modern Linker
 ENV MOLD_VERSION=v2.37.1
 RUN set -e \
-    && git clone https://github.com/rui314/mold.git \
+    && git clone -b "${MOLD_VERSION}" --depth 1 https://github.com/rui314/mold.git \
     && mkdir mold/build \
-    && cd mold/build \
-    && git checkout ${MOLD_VERSION} \
+    && cd mold/build \    
     && cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=clang++ .. \
-    && cmake --build . -j $(nproc) \
+    && cmake --build . -j "$(nproc)" \
     && cmake --install . \
     && cd .. \
     && rm -rf mold
@@ -254,7 +252,7 @@ ENV ICU_VERSION=67.1
 ENV ICU_PREFIX=/usr/local/icu
 
 # Download and build static ICU
-RUN wget -O /tmp/libicu-${ICU_VERSION}.tgz https://github.com/unicode-org/icu/releases/download/release-${ICU_VERSION//./-}/icu4c-${ICU_VERSION//./_}-src.tgz && \
+RUN wget -O "/tmp/libicu-${ICU_VERSION}.tgz" https://github.com/unicode-org/icu/releases/download/release-${ICU_VERSION//./-}/icu4c-${ICU_VERSION//./_}-src.tgz && \
     echo "94a80cd6f251a53bd2a997f6f1b5ac6653fe791dfab66e1eb0227740fb86d5dc /tmp/libicu-${ICU_VERSION}.tgz" | sha256sum --check && \
     mkdir /tmp/icu && \
     pushd /tmp/icu && \
@@ -265,8 +263,7 @@ RUN wget -O /tmp/libicu-${ICU_VERSION}.tgz https://github.com/unicode-org/icu/re
     make install && \
     popd && \
     rm -rf icu && \
-    rm -f /tmp/libicu-${ICU_VERSION}.tgz && \
-    popd
+    rm -f /tmp/libicu-${ICU_VERSION}.tgz
 
 # Switch to nonroot user
 USER nonroot:nonroot
@@ -279,19 +276,19 @@ ENV PYTHON_VERSION=3.11.12 \
     PYENV_ROOT=/home/nonroot/.pyenv \
     PATH=/home/nonroot/.pyenv/shims:/home/nonroot/.pyenv/bin:/home/nonroot/.poetry/bin:$PATH
 RUN set -e \
-    && cd $HOME \
+    && cd "$HOME" \
     && curl -sSO https://raw.githubusercontent.com/pyenv/pyenv-installer/master/bin/pyenv-installer \
     && chmod +x pyenv-installer \
     && ./pyenv-installer \
     && export PYENV_ROOT=/home/nonroot/.pyenv \
     && export PATH="$PYENV_ROOT/bin:$PATH" \
     && export PATH="$PYENV_ROOT/shims:$PATH" \
-    && pyenv install ${PYTHON_VERSION} \
-    && pyenv global ${PYTHON_VERSION} \
+    && pyenv install "${PYTHON_VERSION}" \
+    && pyenv global "${PYTHON_VERSION}" \
     && python --version \
-    && pip install --upgrade pip \
+    && pip install --no-cache-dir --upgrade pip \
     && pip --version \
-    && pip install pipenv wheel poetry
+    && pip install --no-cache-dir pipenv wheel poetry
 
 # Switch to nonroot user (again)
 USER nonroot:nonroot
@@ -317,13 +314,13 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     . "$HOME/.cargo/env" && \
     cargo --version && rustup --version && \
     rustup component add llvm-tools rustfmt clippy && \
-    cargo install rustfilt      --locked --version ${RUSTFILT_VERSION} && \
-    cargo install cargo-hakari  --locked --version ${CARGO_HAKARI_VERSION} && \
-    cargo install cargo-deny    --locked --version ${CARGO_DENY_VERSION} && \
-    cargo install cargo-hack    --locked --version ${CARGO_HACK_VERSION} && \
-    cargo install cargo-nextest --locked --version ${CARGO_NEXTEST_VERSION} && \
-    cargo install cargo-chef    --locked --version ${CARGO_CHEF_VERSION} && \
-    cargo install diesel_cli    --locked --version ${CARGO_DIESEL_CLI_VERSION} \
+    cargo install rustfilt      --locked --version "${RUSTFILT_VERSION}" && \
+    cargo install cargo-hakari  --locked --version "${CARGO_HAKARI_VERSION}" && \
+    cargo install cargo-deny    --locked --version "${CARGO_DENY_VERSION}" && \
+    cargo install cargo-hack    --locked --version "${CARGO_HACK_VERSION}" && \
+    cargo install cargo-nextest --locked --version "${CARGO_NEXTEST_VERSION}" && \
+    cargo install cargo-chef    --locked --version "${CARGO_CHEF_VERSION}" && \
+    cargo install diesel_cli    --locked --version "${CARGO_DIESEL_CLI_VERSION}" \
                                 --features postgres-bundled --no-default-features && \
     rm -rf /home/nonroot/.cargo/registry && \
     rm -rf /home/nonroot/.cargo/git

build-tools/package-lock.json

@@ -6,7 +6,7 @@
     "": {
       "name": "build-tools",
       "devDependencies": {
-        "@redocly/cli": "1.34.4",
+        "@redocly/cli": "1.34.5",
         "@sourcemeta/jsonschema": "10.0.0"
       }
     },
@@ -472,9 +472,9 @@
       }
     },
     "node_modules/@redocly/cli": {
-      "version": "1.34.4",
-      "resolved": "https://registry.npmjs.org/@redocly/cli/-/cli-1.34.4.tgz",
-      "integrity": "sha512-seH/GgrjSB1EeOsgJ/4Ct6Jk2N7sh12POn/7G8UQFARMyUMJpe1oHtBwT2ndfp4EFCpgBAbZ/82Iw6dwczNxEA==",
+      "version": "1.34.5",
+      "resolved": "https://registry.npmjs.org/@redocly/cli/-/cli-1.34.5.tgz",
+      "integrity": "sha512-5IEwxs7SGP5KEXjBKLU8Ffdz9by/KqNSeBk6YUVQaGxMXK//uYlTJIPntgUXbo1KAGG2d2q2XF8y4iFz6qNeiw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -484,14 +484,14 @@
         "@opentelemetry/sdk-trace-node": "1.26.0",
         "@opentelemetry/semantic-conventions": "1.27.0",
         "@redocly/config": "^0.22.0",
-        "@redocly/openapi-core": "1.34.4",
-        "@redocly/respect-core": "1.34.4",
+        "@redocly/openapi-core": "1.34.5",
+        "@redocly/respect-core": "1.34.5",
         "abort-controller": "^3.0.0",
         "chokidar": "^3.5.1",
         "colorette": "^1.2.0",
         "core-js": "^3.32.1",
         "dotenv": "16.4.7",
-        "form-data": "^4.0.0",
+        "form-data": "^4.0.4",
         "get-port-please": "^3.0.1",
         "glob": "^7.1.6",
         "handlebars": "^4.7.6",
@@ -522,9 +522,9 @@
       "license": "MIT"
     },
     "node_modules/@redocly/openapi-core": {
-      "version": "1.34.4",
-      "resolved": "https://registry.npmjs.org/@redocly/openapi-core/-/openapi-core-1.34.4.tgz",
-      "integrity": "sha512-hf53xEgpXIgWl3b275PgZU3OTpYh1RoD2LHdIfQ1JzBNTWsiNKczTEsI/4Tmh2N1oq9YcphhSMyk3lDh85oDjg==",
+      "version": "1.34.5",
+      "resolved": "https://registry.npmjs.org/@redocly/openapi-core/-/openapi-core-1.34.5.tgz",
+      "integrity": "sha512-0EbE8LRbkogtcCXU7liAyC00n9uNG9hJ+eMyHFdUsy9lB/WGqnEBgwjA9q2cyzAVcdTkQqTBBU1XePNnN3OijA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -544,21 +544,21 @@
       }
     },
     "node_modules/@redocly/respect-core": {
-      "version": "1.34.4",
-      "resolved": "https://registry.npmjs.org/@redocly/respect-core/-/respect-core-1.34.4.tgz",
-      "integrity": "sha512-MitKyKyQpsizA4qCVv+MjXL4WltfhFQAoiKiAzrVR1Kusro3VhYb6yJuzoXjiJhR0ukLP5QOP19Vcs7qmj9dZg==",
+      "version": "1.34.5",
+      "resolved": "https://registry.npmjs.org/@redocly/respect-core/-/respect-core-1.34.5.tgz",
+      "integrity": "sha512-GheC/g/QFztPe9UA9LamooSplQuy9pe0Yr8XGTqkz0ahivLDl7svoy/LSQNn1QH3XGtLKwFYMfTwFR2TAYyh5Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
         "@faker-js/faker": "^7.6.0",
         "@redocly/ajv": "8.11.2",
-        "@redocly/openapi-core": "1.34.4",
+        "@redocly/openapi-core": "1.34.5",
         "better-ajv-errors": "^1.2.0",
         "colorette": "^2.0.20",
         "concat-stream": "^2.0.0",
         "cookie": "^0.7.2",
         "dotenv": "16.4.7",
-        "form-data": "4.0.0",
+        "form-data": "^4.0.4",
         "jest-diff": "^29.3.1",
         "jest-matcher-utils": "^29.3.1",
         "js-yaml": "4.1.0",
@@ -582,21 +582,6 @@
       "dev": true,
       "license": "MIT"
     },
-    "node_modules/@redocly/respect-core/node_modules/form-data": {
-      "version": "4.0.0",
-      "resolved": "https://registry.npmjs.org/form-data/-/form-data-4.0.0.tgz",
-      "integrity": "sha512-ETEklSGi5t0QMZuiXoA/Q6vcnxcLQP5vdugSpuAyi6SVGi2clPPp+xgEhuMaHC+zGgn31Kd235W35f7Hykkaww==",
-      "dev": true,
-      "license": "MIT",
-      "dependencies": {
-        "asynckit": "^0.4.0",
-        "combined-stream": "^1.0.8",
-        "mime-types": "^2.1.12"
-      },
-      "engines": {
-        "node": ">= 6"
-      }
-    },
     "node_modules/@sinclair/typebox": {
       "version": "0.27.8",
       "resolved": "https://registry.npmjs.org/@sinclair/typebox/-/typebox-0.27.8.tgz",
@@ -1345,9 +1330,9 @@
       "license": "MIT"
     },
     "node_modules/form-data": {
-      "version": "4.0.3",
-      "resolved": "https://registry.npmjs.org/form-data/-/form-data-4.0.3.tgz",
-      "integrity": "sha512-qsITQPfmvMOSAdeyZ+12I1c+CKSstAFAwu+97zrnWAbIr5u8wfsExUzCesVLC8NgHuRUqNN4Zy6UPWUTRGslcA==",
+      "version": "4.0.4",
+      "resolved": "https://registry.npmjs.org/form-data/-/form-data-4.0.4.tgz",
+      "integrity": "sha512-KrGhL9Q4zjj0kiUt5OO4Mr/A/jlI2jDYs5eHBpYHPcBEVSiipAvn2Ko2HnPe20rmcuuvMHNdZFp+4IlGTMF0Ow==",
       "dev": true,
       "license": "MIT",
       "dependencies": {

build-tools/package.json

@@ -2,7 +2,7 @@
   "name": "build-tools",
   "private": true,
   "devDependencies": {
-    "@redocly/cli": "1.34.4",
+    "@redocly/cli": "1.34.5",
     "@sourcemeta/jsonschema": "10.0.0"
   }
 }

compute/compute-node.Dockerfile

@@ -133,7 +133,7 @@ RUN case $DEBIAN_VERSION in \
       # Install newer version (3.25) from backports.
       # libstdc++-10-dev is required for plv8
       bullseye) \
-        echo "deb http://deb.debian.org/debian bullseye-backports main" > /etc/apt/sources.list.d/bullseye-backports.list; \
+        echo "deb http://archive.debian.org/debian bullseye-backports main" > /etc/apt/sources.list.d/bullseye-backports.list; \
         VERSION_INSTALLS="cmake/bullseye-backports cmake-data/bullseye-backports libstdc++-10-dev"; \
       ;; \
       # Version-specific installs for Bookworm (PG17):

compute/vm-image-spec-bookworm.yaml

@@ -26,7 +26,13 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
+    # Turn off database collector (`--no-collector.database`), we don't use `pg_database_size_bytes` metric anyway, see
+    # https://github.com/neondatabase/flux-fleet/blob/5e19b3fd897667b70d9a7ad4aa06df0ca22b49ff/apps/base/compute-metrics/scrape-compute-pg-exporter-neon.yaml#L29
+    # but it's enabled by default and it doesn't filter out invalid databases, see
+    # https://github.com/prometheus-community/postgres_exporter/blob/06a553c8166512c9d9c5ccf257b0f9bba8751dbc/collector/pg_database.go#L67
+    # so if it hits one, it starts spamming logs
+    #   ERROR:  [NEON_SMGR] [reqid d9700000018] could not read db size of db 705302 from page server at lsn 5/A2457EB0
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --no-collector.database --config.file=/etc/postgres_exporter.yml'
   - name: pgbouncer-exporter
     user: postgres
     sysvInitAction: respawn

compute/vm-image-spec-bullseye.yaml

@@ -26,7 +26,13 @@ commands:
   - name: postgres-exporter
     user: nobody
     sysvInitAction: respawn
-    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --config.file=/etc/postgres_exporter.yml'
+    # Turn off database collector (`--no-collector.database`), we don't use `pg_database_size_bytes` metric anyway, see
+    # https://github.com/neondatabase/flux-fleet/blob/5e19b3fd897667b70d9a7ad4aa06df0ca22b49ff/apps/base/compute-metrics/scrape-compute-pg-exporter-neon.yaml#L29
+    # but it's enabled by default and it doesn't filter out invalid databases, see
+    # https://github.com/prometheus-community/postgres_exporter/blob/06a553c8166512c9d9c5ccf257b0f9bba8751dbc/collector/pg_database.go#L67
+    # so if it hits one, it starts spamming logs
+    #   ERROR:  [NEON_SMGR] [reqid d9700000018] could not read db size of db 705302 from page server at lsn 5/A2457EB0
+    shell: 'DATA_SOURCE_NAME="user=cloud_admin sslmode=disable dbname=postgres application_name=postgres-exporter pgaudit.log=none" /bin/postgres_exporter --no-collector.database --config.file=/etc/postgres_exporter.yml'
   - name: pgbouncer-exporter
     user: postgres
     sysvInitAction: respawn

compute_tools/Cargo.toml

@@ -62,6 +62,7 @@ tokio-stream.workspace = true
 tonic.workspace = true
 tower-otel.workspace = true
 tracing.workspace = true
+tracing-appender.workspace = true
 tracing-opentelemetry.workspace = true
 tracing-subscriber.workspace = true
 tracing-utils.workspace = true

compute_tools/README.md

@@ -52,8 +52,14 @@ stateDiagram-v2
   Init --> Running : Started Postgres
   Running --> TerminationPendingFast : Requested termination
   Running --> TerminationPendingImmediate : Requested termination
+  Running --> ConfigurationPending : Received a /configure request with spec
+  Running --> RefreshConfigurationPending : Received a /refresh_configuration request, compute node will pull a new spec and reconfigure
+  RefreshConfigurationPending --> RefreshConfiguration: Received compute spec and started configuration
+  RefreshConfiguration --> Running : Compute has been re-configured
+  RefreshConfiguration --> RefreshConfigurationPending : Configuration failed and to be retried
   TerminationPendingFast --> Terminated compute with 30s delay for cplane to inspect status
   TerminationPendingImmediate --> Terminated : Terminated compute immediately
+  Failed --> RefreshConfigurationPending : Received a /refresh_configuration request
   Failed --> [*] : Compute exited
   Terminated --> [*] : Compute exited
 ```

compute_tools/src/bin/compute_ctl.rs

@@ -49,9 +49,10 @@ use compute_tools::compute::{
     BUILD_TAG, ComputeNode, ComputeNodeParams, forward_termination_signal,
 };
 use compute_tools::extension_server::get_pg_version_string;
-use compute_tools::logger::*;
 use compute_tools::params::*;
+use compute_tools::pg_isready::get_pg_isready_bin;
 use compute_tools::spec::*;
+use compute_tools::{hadron_metrics, installed_extensions, logger::*};
 use rlimit::{Resource, setrlimit};
 use signal_hook::consts::{SIGINT, SIGQUIT, SIGTERM};
 use signal_hook::iterator::Signals;
@@ -194,11 +195,19 @@ fn main() -> Result<()> {
         .build()?;
     let _rt_guard = runtime.enter();
 
-    let tracing_provider = init(cli.dev)?;
+    let mut log_dir = None;
+    if cli.lakebase_mode {
+        log_dir = std::env::var("COMPUTE_CTL_LOG_DIRECTORY").ok();
+    }
+
+    let (tracing_provider, _file_logs_guard) = init(cli.dev, log_dir)?;
 
     // enable core dumping for all child processes
     setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
 
+    installed_extensions::initialize_metrics();
+    hadron_metrics::initialize_metrics();
+
     let connstr = Url::parse(&cli.connstr).context("cannot parse connstr as a URL")?;
 
     let config = get_config(&cli)?;
@@ -226,7 +235,12 @@ fn main() -> Result<()> {
                 cli.installed_extensions_collection_interval,
             )),
             pg_init_timeout: cli.pg_init_timeout.map(Duration::from_secs),
+            pg_isready_bin: get_pg_isready_bin(&cli.pgbin),
+            instance_id: std::env::var("INSTANCE_ID").ok(),
             lakebase_mode: cli.lakebase_mode,
+            build_tag: BUILD_TAG.to_string(),
+            control_plane_uri: cli.control_plane_uri,
+            config_path_test_only: cli.config,
         },
         config,
     )?;
@@ -238,8 +252,14 @@ fn main() -> Result<()> {
     deinit_and_exit(tracing_provider, exit_code);
 }
 
-fn init(dev_mode: bool) -> Result<Option<tracing_utils::Provider>> {
-    let provider = init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
+fn init(
+    dev_mode: bool,
+    log_dir: Option<String>,
+) -> Result<(
+    Option<tracing_utils::Provider>,
+    Option<tracing_appender::non_blocking::WorkerGuard>,
+)> {
+    let (provider, file_logs_guard) = init_tracing_and_logging(DEFAULT_LOG_LEVEL, &log_dir)?;
 
     let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
     thread::spawn(move || {
@@ -250,7 +270,7 @@ fn init(dev_mode: bool) -> Result<Option<tracing_utils::Provider>> {
 
     info!("compute build_tag: {}", &BUILD_TAG.to_string());
 
-    Ok(provider)
+    Ok((provider, file_logs_guard))
 }
 
 fn get_config(cli: &Cli) -> Result<ComputeConfig> {

compute_tools/src/compute.rs

@@ -21,6 +21,7 @@ use postgres::NoTls;
 use postgres::error::SqlState;
 use remote_storage::{DownloadError, RemotePath};
 use std::collections::{HashMap, HashSet};
+use std::ffi::OsString;
 use std::os::unix::fs::{PermissionsExt, symlink};
 use std::path::Path;
 use std::process::{Command, Stdio};
@@ -40,8 +41,9 @@ use utils::shard::{ShardCount, ShardIndex, ShardNumber};
 
 use crate::configurator::launch_configurator;
 use crate::disk_quota::set_disk_quota;
+use crate::hadron_metrics::COMPUTE_ATTACHED;
 use crate::installed_extensions::get_installed_extensions;
-use crate::logger::startup_context_from_env;
+use crate::logger::{self, startup_context_from_env};
 use crate::lsn_lease::launch_lsn_lease_bg_task_for_static;
 use crate::metrics::COMPUTE_CTL_UP;
 use crate::monitor::launch_monitor;
@@ -113,11 +115,17 @@ pub struct ComputeNodeParams {
 
     /// Interval for installed extensions collection
     pub installed_extensions_collection_interval: Arc<AtomicU64>,
-
+    /// Hadron instance ID of the compute node.
+    pub instance_id: Option<String>,
     /// Timeout of PG compute startup in the Init state.
     pub pg_init_timeout: Option<Duration>,
-
+    // Path to the `pg_isready` binary.
+    pub pg_isready_bin: String,
     pub lakebase_mode: bool,
+
+    pub build_tag: String,
+    pub control_plane_uri: Option<String>,
+    pub config_path_test_only: Option<OsString>,
 }
 
 type TaskHandle = Mutex<Option<JoinHandle<()>>>;
@@ -405,6 +413,52 @@ struct StartVmMonitorResult {
     vm_monitor: Option<JoinHandle<Result<()>>>,
 }
 
+/// Databricks-specific environment variables to be passed to the `postgres` sub-process.
+pub struct DatabricksEnvVars {
+    /// The Databricks "endpoint ID" of the compute instance. Used by `postgres` to check
+    /// the token scopes of internal auth tokens.
+    pub endpoint_id: String,
+    /// Hostname of the Databricks workspace URL this compute instance belongs to.
+    /// Used by postgres to verify Databricks PAT tokens.
+    pub workspace_host: String,
+}
+
+impl DatabricksEnvVars {
+    pub fn new(compute_spec: &ComputeSpec, compute_id: Option<&String>) -> Self {
+        // compute_id is a string format of "{endpoint_id}/{compute_idx}"
+        // endpoint_id is a uuid. We only need to pass down endpoint_id to postgres.
+        // Panics if compute_id is not set or not in the expected format.
+        let endpoint_id = compute_id.unwrap().split('/').next().unwrap().to_string();
+        let workspace_host = compute_spec
+            .databricks_settings
+            .as_ref()
+            .map(|s| s.databricks_workspace_host.clone())
+            .unwrap_or("".to_string());
+        Self {
+            endpoint_id,
+            workspace_host,
+        }
+    }
+
+    /// Constants for the names of Databricks-specific postgres environment variables.
+    const DATABRICKS_ENDPOINT_ID_ENVVAR: &'static str = "DATABRICKS_ENDPOINT_ID";
+    const DATABRICKS_WORKSPACE_HOST_ENVVAR: &'static str = "DATABRICKS_WORKSPACE_HOST";
+
+    /// Convert DatabricksEnvVars to a list of string pairs that can be passed as env vars. Consumes `self`.
+    pub fn to_env_var_list(self) -> Vec<(String, String)> {
+        vec![
+            (
+                Self::DATABRICKS_ENDPOINT_ID_ENVVAR.to_string(),
+                self.endpoint_id.clone(),
+            ),
+            (
+                Self::DATABRICKS_WORKSPACE_HOST_ENVVAR.to_string(),
+                self.workspace_host.clone(),
+            ),
+        ]
+    }
+}
+
 impl ComputeNode {
     pub fn new(params: ComputeNodeParams, config: ComputeConfig) -> Result<Self> {
         let connstr = params.connstr.as_str();
@@ -486,6 +540,7 @@ impl ComputeNode {
             port: this.params.external_http_port,
             config: this.compute_ctl_config.clone(),
             compute_id: this.params.compute_id.clone(),
+            instance_id: this.params.instance_id.clone(),
         }
         .launch(&this);
 
@@ -1402,6 +1457,8 @@ impl ComputeNode {
         let pgdata_path = Path::new(&self.params.pgdata);
 
         let tls_config = self.tls_config(&pspec.spec);
+        let databricks_settings = spec.databricks_settings.as_ref();
+        let postgres_port = self.params.connstr.port();
 
         // Remove/create an empty pgdata directory and put configuration there.
         self.create_pgdata()?;
@@ -1409,8 +1466,11 @@ impl ComputeNode {
             pgdata_path,
             &self.params,
             &pspec.spec,
+            postgres_port,
             self.params.internal_http_port,
             tls_config,
+            databricks_settings,
+            self.params.lakebase_mode,
         )?;
 
         // Syncing safekeepers is only safe with primary nodes: if a primary
@@ -1450,8 +1510,28 @@ impl ComputeNode {
             )
         })?;
 
-        // Update pg_hba.conf received with basebackup.
-        update_pg_hba(pgdata_path, None)?;
+        if let Some(settings) = databricks_settings {
+            copy_tls_certificates(
+                &settings.pg_compute_tls_settings.key_file,
+                &settings.pg_compute_tls_settings.cert_file,
+                pgdata_path,
+            )?;
+
+            // Update pg_hba.conf received with basebackup including additional databricks settings.
+            update_pg_hba(pgdata_path, Some(&settings.databricks_pg_hba))?;
+            update_pg_ident(pgdata_path, Some(&settings.databricks_pg_ident))?;
+        } else {
+            // Update pg_hba.conf received with basebackup.
+            update_pg_hba(pgdata_path, None)?;
+        }
+
+        if let Some(databricks_settings) = spec.databricks_settings.as_ref() {
+            copy_tls_certificates(
+                &databricks_settings.pg_compute_tls_settings.key_file,
+                &databricks_settings.pg_compute_tls_settings.cert_file,
+                pgdata_path,
+            )?;
+        }
 
         // Place pg_dynshmem under /dev/shm. This allows us to use
         // 'dynamic_shared_memory_type = mmap' so that the files are placed in
@@ -1564,14 +1644,31 @@ impl ComputeNode {
     pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
         let pgdata_path = Path::new(&self.params.pgdata);
 
+        let env_vars: Vec<(String, String)> = if self.params.lakebase_mode {
+            let databricks_env_vars = {
+                let state = self.state.lock().unwrap();
+                let spec = &state.pspec.as_ref().unwrap().spec;
+                DatabricksEnvVars::new(spec, Some(&self.params.compute_id))
+            };
+
+            info!(
+                "Starting Postgres for databricks endpoint id: {}",
+                &databricks_env_vars.endpoint_id
+            );
+
+            let mut env_vars = databricks_env_vars.to_env_var_list();
+            env_vars.extend(storage_auth_token.map(|t| ("NEON_AUTH_TOKEN".to_string(), t)));
+            env_vars
+        } else if let Some(storage_auth_token) = &storage_auth_token {
+            vec![("NEON_AUTH_TOKEN".to_owned(), storage_auth_token.to_owned())]
+        } else {
+            vec![]
+        };
+
         // Run postgres as a child process.
         let mut pg = maybe_cgexec(&self.params.pgbin)
             .args(["-D", &self.params.pgdata])
-            .envs(if let Some(storage_auth_token) = &storage_auth_token {
-                vec![("NEON_AUTH_TOKEN", storage_auth_token)]
-            } else {
-                vec![]
-            })
+            .envs(env_vars)
             .stderr(Stdio::piped())
             .spawn()
             .expect("cannot start postgres process");
@@ -1785,6 +1882,34 @@ impl ComputeNode {
         Ok::<(), anyhow::Error>(())
     }
 
+    // Signal to the configurator to refresh the configuration by pulling a new spec from the HCC.
+    // Note that this merely triggers a notification on a condition variable the configurator thread
+    // waits on. The configurator thread (in configurator.rs) pulls the new spec from the HCC and
+    // applies it.
+    pub async fn signal_refresh_configuration(&self) -> Result<()> {
+        let states_allowing_configuration_refresh = [
+            ComputeStatus::Running,
+            ComputeStatus::Failed,
+            ComputeStatus::RefreshConfigurationPending,
+        ];
+
+        let mut state = self.state.lock().expect("state lock poisoned");
+        if states_allowing_configuration_refresh.contains(&state.status) {
+            state.status = ComputeStatus::RefreshConfigurationPending;
+            self.state_changed.notify_all();
+            Ok(())
+        } else if state.status == ComputeStatus::Init {
+            // If the compute is in Init state, we can't refresh the configuration immediately,
+            // but we should be able to do that soon.
+            Ok(())
+        } else {
+            Err(anyhow::anyhow!(
+                "Cannot refresh compute configuration in state {:?}",
+                state.status
+            ))
+        }
+    }
+
     // Wrapped this around `pg_ctl reload`, but right now we don't use
     // `pg_ctl` for start / stop.
     #[instrument(skip_all)]
@@ -1846,12 +1971,16 @@ impl ComputeNode {
 
         // Write new config
         let pgdata_path = Path::new(&self.params.pgdata);
+        let postgres_port = self.params.connstr.port();
         config::write_postgres_conf(
             pgdata_path,
             &self.params,
             &spec,
+            postgres_port,
             self.params.internal_http_port,
             tls_config,
+            spec.databricks_settings.as_ref(),
+            self.params.lakebase_mode,
         )?;
 
         self.pg_reload_conf()?;
@@ -1957,6 +2086,8 @@ impl ComputeNode {
                             // wait
                             ComputeStatus::Init
                             | ComputeStatus::Configuration
+                            | ComputeStatus::RefreshConfiguration
+                            | ComputeStatus::RefreshConfigurationPending
                             | ComputeStatus::Empty => {
                                 state = self.state_changed.wait(state).unwrap();
                             }
@@ -2513,6 +2644,34 @@ LIMIT 100",
             );
         }
     }
+
+    /// Set the compute spec and update related metrics.
+    /// This is the central place where pspec is updated.
+    pub fn set_spec(params: &ComputeNodeParams, state: &mut ComputeState, pspec: ParsedSpec) {
+        state.pspec = Some(pspec);
+        ComputeNode::update_attached_metric(params, state);
+        let _ = logger::update_ids(&params.instance_id, &Some(params.compute_id.clone()));
+    }
+
+    pub fn update_attached_metric(params: &ComputeNodeParams, state: &mut ComputeState) {
+        // Update the pg_cctl_attached gauge when all identifiers are available.
+        if let Some(instance_id) = &params.instance_id {
+            if let Some(pspec) = &state.pspec {
+                // Clear all values in the metric
+                COMPUTE_ATTACHED.reset();
+
+                // Set new metric value
+                COMPUTE_ATTACHED
+                    .with_label_values(&[
+                        &params.compute_id,
+                        instance_id,
+                        &pspec.tenant_id.to_string(),
+                        &pspec.timeline_id.to_string(),
+                    ])
+                    .set(1);
+            }
+        }
+    }
 }
 
 pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {

compute_tools/src/compute_prewarm.rs

@@ -90,6 +90,7 @@ impl ComputeNode {
     }
 
     /// If there is a prewarm request ongoing, return `false`, `true` otherwise.
+    /// Has a failpoint "compute-prewarm"
     pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
@@ -112,9 +113,8 @@ impl ComputeNode {
                 Err(err) => {
                     crate::metrics::LFC_PREWARM_ERRORS.inc();
                     error!(%err, "could not prewarm LFC");
-
                     LfcPrewarmState::Failed {
-                        error: err.to_string(),
+                        error: format!("{err:#}"),
                     }
                 }
             };
@@ -135,16 +135,20 @@ impl ComputeNode {
     async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<bool> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(from_endpoint)?;
 
+        #[cfg(feature = "testing")]
+        fail::fail_point!("compute-prewarm", |_| {
+            bail!("prewarm configured to fail because of a failpoint")
+        });
+
         info!(%url, "requesting LFC state from endpoint storage");
         let request = Client::new().get(&url).bearer_auth(token);
         let res = request.send().await.context("querying endpoint storage")?;
-        let status = res.status();
-        match status {
+        match res.status() {
             StatusCode::OK => (),
             StatusCode::NOT_FOUND => {
                 return Ok(false);
             }
-            _ => bail!("{status} querying endpoint storage"),
+            status => bail!("{status} querying endpoint storage"),
         }
 
         let mut uncompressed = Vec::new();
@@ -205,7 +209,7 @@ impl ComputeNode {
         crate::metrics::LFC_OFFLOAD_ERRORS.inc();
         error!(%err, "could not offload LFC state to endpoint storage");
         self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
-            error: err.to_string(),
+            error: format!("{err:#}"),
         };
     }
 
@@ -213,16 +217,22 @@ impl ComputeNode {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
         info!(%url, "requesting LFC state from Postgres");
 
-        let mut compressed = Vec::new();
-        ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
+        let row = ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
             .context("connecting to postgres")?
             .query_one("select neon.get_local_cache_state()", &[])
             .await
-            .context("querying LFC state")?
-            .try_get::<usize, &[u8]>(0)
-            .context("deserializing LFC state")
-            .map(ZstdEncoder::new)?
+            .context("querying LFC state")?;
+        let state = row
+            .try_get::<usize, Option<&[u8]>>(0)
+            .context("deserializing LFC state")?;
+        let Some(state) = state else {
+            info!(%url, "empty LFC state, not exporting");
+            return Ok(());
+        };
+
+        let mut compressed = Vec::new();
+        ZstdEncoder::new(state)
             .read_to_end(&mut compressed)
             .await
             .context("compressing LFC state")?;

compute_tools/src/compute_promote.rs

@@ -1,11 +1,12 @@
 use crate::compute::ComputeNode;
 use anyhow::{Context, Result, bail};
-use compute_api::{
-    responses::{LfcPrewarmState, PromoteState, SafekeepersLsn},
-    spec::ComputeMode,
-};
+use compute_api::responses::{LfcPrewarmState, PromoteConfig, PromoteState};
+use compute_api::spec::ComputeMode;
+use itertools::Itertools;
+use std::collections::HashMap;
 use std::{sync::Arc, time::Duration};
 use tokio::time::sleep;
+use tracing::info;
 use utils::lsn::Lsn;
 
 impl ComputeNode {
@@ -13,21 +14,22 @@ impl ComputeNode {
     /// and http client disconnects, this does not stop promotion, and subsequent
     /// calls block until promote finishes.
     /// Called by control plane on secondary after primary endpoint is terminated
-    pub async fn promote(self: &Arc<Self>, safekeepers_lsn: SafekeepersLsn) -> PromoteState {
+    /// Has a failpoint "compute-promotion"
+    pub async fn promote(self: &Arc<Self>, cfg: PromoteConfig) -> PromoteState {
         let cloned = self.clone();
+        let promote_fn = async move || {
+            let Err(err) = cloned.promote_impl(cfg).await else {
+                return PromoteState::Completed;
+            };
+            tracing::error!(%err, "promoting");
+            PromoteState::Failed {
+                error: format!("{err:#}"),
+            }
+        };
+
         let start_promotion = || {
             let (tx, rx) = tokio::sync::watch::channel(PromoteState::NotPromoted);
-            tokio::spawn(async move {
-                tx.send(match cloned.promote_impl(safekeepers_lsn).await {
-                    Ok(_) => PromoteState::Completed,
-                    Err(err) => {
-                        tracing::error!(%err, "promoting");
-                        PromoteState::Failed {
-                            error: err.to_string(),
-                        }
-                    }
-                })
-            });
+            tokio::spawn(async move { tx.send(promote_fn().await) });
             rx
         };
 
@@ -47,9 +49,7 @@ impl ComputeNode {
         task.borrow().clone()
     }
 
-    // Why do we have to supply safekeepers?
-    // For secondary we use primary_connection_conninfo so safekeepers field is empty
-    async fn promote_impl(&self, safekeepers_lsn: SafekeepersLsn) -> Result<()> {
+    async fn promote_impl(&self, mut cfg: PromoteConfig) -> Result<()> {
         {
             let state = self.state.lock().unwrap();
             let mode = &state.pspec.as_ref().unwrap().spec.mode;
@@ -73,7 +73,7 @@ impl ComputeNode {
             .await
             .context("connecting to postgres")?;
 
-        let primary_lsn = safekeepers_lsn.wal_flush_lsn;
+        let primary_lsn = cfg.wal_flush_lsn;
         let mut last_wal_replay_lsn: Lsn = Lsn::INVALID;
         const RETRIES: i32 = 20;
         for i in 0..=RETRIES {
@@ -86,7 +86,7 @@ impl ComputeNode {
             if last_wal_replay_lsn >= primary_lsn {
                 break;
             }
-            tracing::info!("Try {i}, replica lsn {last_wal_replay_lsn}, primary lsn {primary_lsn}");
+            info!("Try {i}, replica lsn {last_wal_replay_lsn}, primary lsn {primary_lsn}");
             sleep(Duration::from_secs(1)).await;
         }
         if last_wal_replay_lsn < primary_lsn {
@@ -96,7 +96,7 @@ impl ComputeNode {
         // using $1 doesn't work with ALTER SYSTEM SET
         let safekeepers_sql = format!(
             "ALTER SYSTEM SET neon.safekeepers='{}'",
-            safekeepers_lsn.safekeepers
+            cfg.spec.safekeeper_connstrings.join(",")
         );
         client
             .query(&safekeepers_sql, &[])
@@ -106,6 +106,12 @@ impl ComputeNode {
             .query("SELECT pg_reload_conf()", &[])
             .await
             .context("reloading postgres config")?;
+
+        #[cfg(feature = "testing")]
+        fail::fail_point!("compute-promotion", |_| {
+            bail!("promotion configured to fail because of a failpoint")
+        });
+
         let row = client
             .query_one("SELECT * FROM pg_promote()", &[])
             .await
@@ -125,8 +131,36 @@ impl ComputeNode {
             bail!("replica in read only mode after promotion");
         }
 
-        let mut state = self.state.lock().unwrap();
-        state.pspec.as_mut().unwrap().spec.mode = ComputeMode::Primary;
-        Ok(())
+        {
+            let mut state = self.state.lock().unwrap();
+            let spec = &mut state.pspec.as_mut().unwrap().spec;
+            spec.mode = ComputeMode::Primary;
+            let new_conf = cfg.spec.cluster.postgresql_conf.as_mut().unwrap();
+            let existing_conf = spec.cluster.postgresql_conf.as_ref().unwrap();
+            Self::merge_spec(new_conf, existing_conf);
+        }
+        info!("applied new spec, reconfiguring as primary");
+        self.reconfigure()
+    }
+
+    /// Merge old and new Postgres conf specs to apply on secondary.
+    /// Change new spec's port and safekeepers since they are supplied
+    /// differenly
+    fn merge_spec(new_conf: &mut String, existing_conf: &str) {
+        let mut new_conf_set: HashMap<&str, &str> = new_conf
+            .split_terminator('\n')
+            .map(|e| e.split_once("=").expect("invalid item"))
+            .collect();
+        new_conf_set.remove("neon.safekeepers");
+
+        let existing_conf_set: HashMap<&str, &str> = existing_conf
+            .split_terminator('\n')
+            .map(|e| e.split_once("=").expect("invalid item"))
+            .collect();
+        new_conf_set.insert("port", existing_conf_set["port"]);
+        *new_conf = new_conf_set
+            .iter()
+            .map(|(k, v)| format!("{k}={v}"))
+            .join("\n");
     }
 }

compute_tools/src/config.rs

@@ -7,11 +7,14 @@ use std::io::prelude::*;
 use std::path::Path;
 
 use compute_api::responses::TlsConfig;
-use compute_api::spec::{ComputeAudit, ComputeMode, ComputeSpec, GenericOption};
+use compute_api::spec::{
+    ComputeAudit, ComputeMode, ComputeSpec, DatabricksSettings, GenericOption,
+};
 
 use crate::compute::ComputeNodeParams;
 use crate::pg_helpers::{
-    GenericOptionExt, GenericOptionsSearch, PgOptionsSerialize, escape_conf_value,
+    DatabricksSettingsExt as _, GenericOptionExt, GenericOptionsSearch, PgOptionsSerialize,
+    escape_conf_value,
 };
 use crate::tls::{self, SERVER_CRT, SERVER_KEY};
 
@@ -40,12 +43,16 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
 }
 
 /// Create or completely rewrite configuration file specified by `path`
+#[allow(clippy::too_many_arguments)]
 pub fn write_postgres_conf(
     pgdata_path: &Path,
     params: &ComputeNodeParams,
     spec: &ComputeSpec,
+    postgres_port: Option<u16>,
     extension_server_port: u16,
     tls_config: &Option<TlsConfig>,
+    databricks_settings: Option<&DatabricksSettings>,
+    lakebase_mode: bool,
 ) -> Result<()> {
     let path = pgdata_path.join("postgresql.conf");
     // File::create() destroys the file content if it exists.
@@ -285,6 +292,24 @@ pub fn write_postgres_conf(
         writeln!(file, "log_destination='stderr,syslog'")?;
     }
 
+    if lakebase_mode {
+        // Explicitly set the port based on the connstr, overriding any previous port setting.
+        // Note: It is important that we don't specify a different port again after this.
+        let port = postgres_port.expect("port must be present in connstr");
+        writeln!(file, "port = {port}")?;
+
+        // This is databricks specific settings.
+        // This should be at the end of the file but before `compute_ctl_temp_override.conf` below
+        // so that it can override any settings above.
+        // `compute_ctl_temp_override.conf` is intended to override any settings above during specific operations.
+        // To prevent potential breakage in the future, we keep it above `compute_ctl_temp_override.conf`.
+        writeln!(file, "# Databricks settings start")?;
+        if let Some(settings) = databricks_settings {
+            writeln!(file, "{}", settings.as_pg_settings())?;
+        }
+        writeln!(file, "# Databricks settings end")?;
+    }
+
     // This is essential to keep this line at the end of the file,
     // because it is intended to override any settings above.
     writeln!(file, "include_if_exists = 'compute_ctl_temp_override.conf'")?;

compute_tools/src/configurator.rs

@@ -1,23 +1,40 @@
-use std::sync::Arc;
+use std::fs::File;
 use std::thread;
+use std::{path::Path, sync::Arc};
 
-use compute_api::responses::ComputeStatus;
+use anyhow::Result;
+use compute_api::responses::{ComputeConfig, ComputeStatus};
 use tracing::{error, info, instrument};
 
-use crate::compute::ComputeNode;
+use crate::compute::{ComputeNode, ParsedSpec};
+use crate::spec::get_config_from_control_plane;
 
 #[instrument(skip_all)]
 fn configurator_main_loop(compute: &Arc<ComputeNode>) {
     info!("waiting for reconfiguration requests");
     loop {
         let mut state = compute.state.lock().unwrap();
+        /* BEGIN_HADRON */
+        // RefreshConfiguration should only be used inside the loop
+        assert_ne!(state.status, ComputeStatus::RefreshConfiguration);
+        /* END_HADRON */
 
-        // We have to re-check the status after re-acquiring the lock because it could be that
-        // the status has changed while we were waiting for the lock, and we might not need to
-        // wait on the condition variable. Otherwise, we might end up in some soft-/deadlock, i.e.
-        // we are waiting for a condition variable that will never be signaled.
-        if state.status != ComputeStatus::ConfigurationPending {
-            state = compute.state_changed.wait(state).unwrap();
+        if compute.params.lakebase_mode {
+            while state.status != ComputeStatus::ConfigurationPending
+                && state.status != ComputeStatus::RefreshConfigurationPending
+                && state.status != ComputeStatus::Failed
+            {
+                info!("configurator: compute status: {:?}, sleeping", state.status);
+                state = compute.state_changed.wait(state).unwrap();
+            }
+        } else {
+            // We have to re-check the status after re-acquiring the lock because it could be that
+            // the status has changed while we were waiting for the lock, and we might not need to
+            // wait on the condition variable. Otherwise, we might end up in some soft-/deadlock, i.e.
+            // we are waiting for a condition variable that will never be signaled.
+            if state.status != ComputeStatus::ConfigurationPending {
+                state = compute.state_changed.wait(state).unwrap();
+            }
         }
 
         // Re-check the status after waking up
@@ -37,6 +54,133 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
             // XXX: used to test that API is blocking
             // std::thread::sleep(std::time::Duration::from_millis(10000));
 
+            compute.set_status(new_status);
+        } else if state.status == ComputeStatus::RefreshConfigurationPending {
+            info!(
+                "compute node suspects its configuration is out of date, now refreshing configuration"
+            );
+            state.set_status(ComputeStatus::RefreshConfiguration, &compute.state_changed);
+            // Drop the lock guard here to avoid holding the lock while downloading config from the control plane / HCC.
+            // This is the only thread that can move compute_ctl out of the `RefreshConfiguration` state, so it
+            // is safe to drop the lock like this.
+            drop(state);
+
+            let get_config_result: anyhow::Result<ComputeConfig> =
+                if let Some(config_path) = &compute.params.config_path_test_only {
+                    // This path is only to make testing easier. In production we always get the config from the HCC.
+                    info!(
+                        "reloading config.json from path: {}",
+                        config_path.to_string_lossy()
+                    );
+                    let path = Path::new(config_path);
+                    if let Ok(file) = File::open(path) {
+                        match serde_json::from_reader::<File, ComputeConfig>(file) {
+                            Ok(config) => Ok(config),
+                            Err(e) => {
+                                error!("could not parse config file: {}", e);
+                                Err(anyhow::anyhow!("could not parse config file: {}", e))
+                            }
+                        }
+                    } else {
+                        error!(
+                            "could not open config file at path: {:?}",
+                            config_path.to_string_lossy()
+                        );
+                        Err(anyhow::anyhow!(
+                            "could not open config file at path: {}",
+                            config_path.to_string_lossy()
+                        ))
+                    }
+                } else if let Some(control_plane_uri) = &compute.params.control_plane_uri {
+                    get_config_from_control_plane(control_plane_uri, &compute.params.compute_id)
+                } else {
+                    Err(anyhow::anyhow!("config_path_test_only is not set"))
+                };
+
+            // Parse any received ComputeSpec and transpose the result into a Result<Option<ParsedSpec>>.
+            let parsed_spec_result: Result<Option<ParsedSpec>> =
+                get_config_result.and_then(|config| {
+                    if let Some(spec) = config.spec {
+                        if let Ok(pspec) = ParsedSpec::try_from(spec) {
+                            Ok(Some(pspec))
+                        } else {
+                            Err(anyhow::anyhow!("could not parse spec"))
+                        }
+                    } else {
+                        Ok(None)
+                    }
+                });
+
+            let new_status: ComputeStatus;
+            match parsed_spec_result {
+                // Control plane (HCM) returned a spec and we were able to parse it.
+                Ok(Some(pspec)) => {
+                    {
+                        let mut state = compute.state.lock().unwrap();
+                        // Defensive programming to make sure this thread is indeed the only one that can move the compute
+                        // node out of the `RefreshConfiguration` state. Would be nice if we can encode this invariant
+                        // into the type system.
+                        assert_eq!(state.status, ComputeStatus::RefreshConfiguration);
+
+                        if state.pspec.as_ref().map(|ps| ps.pageserver_connstr.clone())
+                            == Some(pspec.pageserver_connstr.clone())
+                        {
+                            info!(
+                                "Refresh configuration: Retrieved spec is the same as the current spec. Waiting for control plane to update the spec before attempting reconfiguration."
+                            );
+                            state.status = ComputeStatus::Running;
+                            compute.state_changed.notify_all();
+                            drop(state);
+                            std::thread::sleep(std::time::Duration::from_secs(5));
+                            continue;
+                        }
+                        // state.pspec is consumed by compute.reconfigure() below. Note that compute.reconfigure() will acquire
+                        // the compute.state lock again so we need to have the lock guard go out of scope here. We could add a
+                        // "locked" variant of compute.reconfigure() that takes the lock guard as an argument to make this cleaner,
+                        // but it's not worth forking the codebase too much for this minor point alone right now.
+                        state.pspec = Some(pspec);
+                    }
+                    match compute.reconfigure() {
+                        Ok(_) => {
+                            info!("Refresh configuration: compute node configured");
+                            new_status = ComputeStatus::Running;
+                        }
+                        Err(e) => {
+                            error!(
+                                "Refresh configuration: could not configure compute node: {}",
+                                e
+                            );
+                            // Set the compute node back to the `RefreshConfigurationPending` state if the configuration
+                            // was not successful. It should be okay to treat this situation the same as if the loop
+                            // hasn't executed yet as long as the detection side keeps notifying.
+                            new_status = ComputeStatus::RefreshConfigurationPending;
+                        }
+                    }
+                }
+                // Control plane (HCM)'s response does not contain a spec. This is the "Empty" attachment case.
+                Ok(None) => {
+                    info!(
+                        "Compute Manager signaled that this compute is no longer attached to any storage. Exiting."
+                    );
+                    // We just immediately terminate the whole compute_ctl in this case. It's not necessary to attempt a
+                    // clean shutdown as Postgres is probably not responding anyway (which is why we are in this refresh
+                    // configuration state).
+                    std::process::exit(1);
+                }
+                // Various error cases:
+                // - The request to the control plane (HCM) either failed or returned a malformed spec.
+                // - compute_ctl itself is configured incorrectly (e.g., compute_id is not set).
+                Err(e) => {
+                    error!(
+                        "Refresh configuration: error getting a parsed spec: {:?}",
+                        e
+                    );
+                    new_status = ComputeStatus::RefreshConfigurationPending;
+                    // We may be dealing with an overloaded HCM if we end up in this path. Backoff 5 seconds before
+                    // retrying to avoid hammering the HCM.
+                    std::thread::sleep(std::time::Duration::from_secs(5));
+                }
+            }
             compute.set_status(new_status);
         } else if state.status == ComputeStatus::Failed {
             info!("compute node is now in Failed state, exiting");

compute_tools/src/hadron_metrics.rs

@@ -0,0 +1,60 @@
+use metrics::{
+    IntCounter, IntGaugeVec, core::Collector, proto::MetricFamily, register_int_counter,
+    register_int_gauge_vec,
+};
+use once_cell::sync::Lazy;
+
+// Counter keeping track of the number of PageStream request errors reported by Postgres.
+// An error is registered every time Postgres calls compute_ctl's /refresh_configuration API.
+// Postgres will invoke this API if it detected trouble with PageStream requests (get_page@lsn,
+// get_base_backup, etc.) it sends to any pageserver. An increase in this counter value typically
+// indicates Postgres downtime, as PageStream requests are critical for Postgres to function.
+pub static POSTGRES_PAGESTREAM_REQUEST_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pg_cctl_pagestream_request_errors_total",
+        "Number of PageStream request errors reported by the postgres process"
+    )
+    .expect("failed to define a metric")
+});
+
+// Counter keeping track of the number of compute configuration errors due to Postgres statement
+// timeouts. An error is registered every time `ComputeNode::reconfigure()` fails due to Postgres
+// error code 57014 (query cancelled). This statement timeout typically occurs when postgres is
+// stuck in a problematic retry loop when the PS is reject its connection requests (usually due
+// to PG pointing at the wrong PS). We should investigate the root cause when this counter value
+// increases by checking PG and PS logs.
+pub static COMPUTE_CONFIGURE_STATEMENT_TIMEOUT_ERRORS: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pg_cctl_configure_statement_timeout_errors_total",
+        "Number of compute configuration errors due to Postgres statement timeouts."
+    )
+    .expect("failed to define a metric")
+});
+
+pub static COMPUTE_ATTACHED: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "pg_cctl_attached",
+        "Compute node attached status (1 if attached)",
+        &[
+            "pg_compute_id",
+            "pg_instance_id",
+            "tenant_id",
+            "timeline_id"
+        ]
+    )
+    .expect("failed to define a metric")
+});
+
+pub fn collect() -> Vec<MetricFamily> {
+    let mut metrics = Vec::new();
+    metrics.extend(POSTGRES_PAGESTREAM_REQUEST_ERRORS.collect());
+    metrics.extend(COMPUTE_CONFIGURE_STATEMENT_TIMEOUT_ERRORS.collect());
+    metrics.extend(COMPUTE_ATTACHED.collect());
+    metrics
+}
+
+pub fn initialize_metrics() {
+    Lazy::force(&POSTGRES_PAGESTREAM_REQUEST_ERRORS);
+    Lazy::force(&COMPUTE_CONFIGURE_STATEMENT_TIMEOUT_ERRORS);
+    Lazy::force(&COMPUTE_ATTACHED);
+}

compute_tools/src/http/middleware/authorize.rs

@@ -16,13 +16,29 @@ use crate::http::JsonResponse;
 #[derive(Clone, Debug)]
 pub(in crate::http) struct Authorize {
     compute_id: String,
+    // BEGIN HADRON
+    // Hadron instance ID. Only set if it's a Lakebase V1 a.k.a. Hadron instance.
+    instance_id: Option<String>,
+    // END HADRON
     jwks: JwkSet,
     validation: Validation,
 }
 
 impl Authorize {
-    pub fn new(compute_id: String, jwks: JwkSet) -> Self {
+    pub fn new(compute_id: String, instance_id: Option<String>, jwks: JwkSet) -> Self {
         let mut validation = Validation::new(Algorithm::EdDSA);
+
+        // BEGIN HADRON
+        let use_rsa = jwks.keys.iter().any(|jwk| {
+            jwk.common
+                .key_algorithm
+                .is_some_and(|alg| alg == jsonwebtoken::jwk::KeyAlgorithm::RS256)
+        });
+        if use_rsa {
+            validation = Validation::new(Algorithm::RS256);
+        }
+        // END HADRON
+
         validation.validate_exp = true;
         // Unused by the control plane
         validation.validate_nbf = false;
@@ -34,6 +50,7 @@ impl Authorize {
 
         Self {
             compute_id,
+            instance_id,
             jwks,
             validation,
         }
@@ -47,10 +64,20 @@ impl AsyncAuthorizeRequest<Body> for Authorize {
 
     fn authorize(&mut self, mut request: Request<Body>) -> Self::Future {
         let compute_id = self.compute_id.clone();
+        let is_hadron_instance = self.instance_id.is_some();
         let jwks = self.jwks.clone();
         let validation = self.validation.clone();
 
         Box::pin(async move {
+            // BEGIN HADRON
+            // In Hadron deployments the "external" HTTP endpoint on compute_ctl can only be
+            // accessed by trusted components (enforced by dblet network policy), so we can bypass
+            // all auth here.
+            if is_hadron_instance {
+                return Ok(request);
+            }
+            // END HADRON
+
             let TypedHeader(Authorization(bearer)) = request
                 .extract_parts::<TypedHeader<Authorization<Bearer>>>()
                 .await

compute_tools/src/http/openapi_spec.yaml

@@ -96,7 +96,7 @@ paths:
         content:
           application/json:
             schema:
-                $ref: "#/components/schemas/SafekeepersLsn"
+              $ref: "#/components/schemas/ComputeSchemaWithLsn"
       responses:
         200:
           description: Promote succeeded or wasn't started
@@ -297,14 +297,7 @@ paths:
         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
+              $ref: "#/components/schemas/ComputeSchema"
       responses:
         200:
           description: Compute configuration finished.
@@ -591,18 +584,25 @@ components:
           type: string
           example: "1.0.0"
 
-    SafekeepersLsn:
+    ComputeSchema:
       type: object
       required:
-        - safekeepers
+        - spec
+      properties:
+        spec:
+          type: object
+    ComputeSchemaWithLsn:
+      type: object
+      required:
+        - spec
         - wal_flush_lsn
       properties:
-        safekeepers:
-          description: Primary replica safekeepers
-          type: string
+        spec:
+          $ref: "#/components/schemas/ComputeState"
         wal_flush_lsn:
-          description: Primary last WAL flush LSN
           type: string
+          description: "last WAL flush LSN"
+          example: "0/028F10D8"
 
     LfcPrewarmState:
       type: object

compute_tools/src/http/routes/configure.rs

@@ -43,7 +43,12 @@ pub(in crate::http) async fn configure(
         // configure request for tracing purposes.
         state.startup_span = Some(tracing::Span::current());
 
-        state.pspec = Some(pspec);
+        if compute.params.lakebase_mode {
+            ComputeNode::set_spec(&compute.params, &mut state, pspec);
+        } else {
+            state.pspec = Some(pspec);
+        }
+
         state.set_status(ComputeStatus::ConfigurationPending, &compute.state_changed);
         drop(state);
     }

compute_tools/src/http/routes/hadron_liveness_probe.rs

@@ -0,0 +1,34 @@
+use crate::pg_isready::pg_isready;
+use crate::{compute::ComputeNode, http::JsonResponse};
+use axum::{extract::State, http::StatusCode, response::Response};
+use std::sync::Arc;
+
+/// NOTE: NOT ENABLED YET
+/// Detect if the compute is alive.
+/// Called by the liveness probe of the compute container.
+pub(in crate::http) async fn hadron_liveness_probe(
+    State(compute): State<Arc<ComputeNode>>,
+) -> Response {
+    let port = match compute.params.connstr.port() {
+        Some(port) => port,
+        None => {
+            return JsonResponse::error(
+                StatusCode::INTERNAL_SERVER_ERROR,
+                "Failed to get the port from the connection string",
+            );
+        }
+    };
+    match pg_isready(&compute.params.pg_isready_bin, port) {
+        Ok(_) => {
+            // The connection is successful, so the compute is alive.
+            // Return a 200 OK response.
+            JsonResponse::success(StatusCode::OK, "ok")
+        }
+        Err(e) => {
+            tracing::error!("Hadron liveness probe failed: {}", e);
+            // The connection failed, so the compute is not alive.
+            // Return a 500 Internal Server Error response.
+            JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, e)
+        }
+    }
+}

compute_tools/src/http/routes/metrics.rs

@@ -13,6 +13,7 @@ use metrics::{Encoder, TextEncoder};
 
 use crate::communicator_socket_client::connect_communicator_socket;
 use crate::compute::ComputeNode;
+use crate::hadron_metrics;
 use crate::http::JsonResponse;
 use crate::metrics::collect;
 
@@ -21,11 +22,18 @@ pub(in crate::http) async fn get_metrics() -> Response {
     // When we call TextEncoder::encode() below, it will immediately return an
     // error if a metric family has no metrics, so we need to preemptively
     // filter out metric families with no metrics.
-    let metrics = collect()
+    let mut metrics = collect()
         .into_iter()
         .filter(|m| !m.get_metric().is_empty())
         .collect::<Vec<MetricFamily>>();
 
+    // Add Hadron metrics.
+    let hadron_metrics: Vec<MetricFamily> = hadron_metrics::collect()
+        .into_iter()
+        .filter(|m| !m.get_metric().is_empty())
+        .collect();
+    metrics.extend(hadron_metrics);
+
     let encoder = TextEncoder::new();
     let mut buffer = vec![];
 

compute_tools/src/http/routes/mod.rs

@@ -10,11 +10,13 @@ pub(in crate::http) mod extension_server;
 pub(in crate::http) mod extensions;
 pub(in crate::http) mod failpoints;
 pub(in crate::http) mod grants;
+pub(in crate::http) mod hadron_liveness_probe;
 pub(in crate::http) mod insights;
 pub(in crate::http) mod lfc;
 pub(in crate::http) mod metrics;
 pub(in crate::http) mod metrics_json;
 pub(in crate::http) mod promote;
+pub(in crate::http) mod refresh_configuration;
 pub(in crate::http) mod status;
 pub(in crate::http) mod terminate;
 

compute_tools/src/http/routes/promote.rs

@@ -1,14 +1,14 @@
 use crate::http::JsonResponse;
-use axum::Form;
+use axum::extract::Json;
 use http::StatusCode;
 
 pub(in crate::http) async fn promote(
     compute: axum::extract::State<std::sync::Arc<crate::compute::ComputeNode>>,
-    Form(safekeepers_lsn): Form<compute_api::responses::SafekeepersLsn>,
+    Json(cfg): Json<compute_api::responses::PromoteConfig>,
 ) -> axum::response::Response {
-    let state = compute.promote(safekeepers_lsn).await;
-    if let compute_api::responses::PromoteState::Failed { error } = state {
-        return JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, error);
+    let state = compute.promote(cfg).await;
+    if let compute_api::responses::PromoteState::Failed { error: _ } = state {
+        return JsonResponse::create_response(StatusCode::INTERNAL_SERVER_ERROR, state);
     }
     JsonResponse::success(StatusCode::OK, state)
 }

compute_tools/src/http/routes/refresh_configuration.rs

@@ -0,0 +1,29 @@
+// This file is added by Hadron
+
+use std::sync::Arc;
+
+use axum::{
+    extract::State,
+    response::{IntoResponse, Response},
+};
+use http::StatusCode;
+
+use crate::compute::ComputeNode;
+use crate::hadron_metrics::POSTGRES_PAGESTREAM_REQUEST_ERRORS;
+use crate::http::JsonResponse;
+
+/// The /refresh_configuration POST method is used to nudge compute_ctl to pull a new spec
+/// from the HCC and attempt to reconfigure Postgres with the new spec. The method does not wait
+/// for the reconfiguration to complete. Rather, it simply delivers a signal that will cause
+/// configuration to be reloaded in a best effort manner. Invocation of this method does not
+/// guarantee that a reconfiguration will occur. The caller should consider keep sending this
+/// request while it believes that the compute configuration is out of date.
+pub(in crate::http) async fn refresh_configuration(
+    State(compute): State<Arc<ComputeNode>>,
+) -> Response {
+    POSTGRES_PAGESTREAM_REQUEST_ERRORS.inc();
+    match compute.signal_refresh_configuration().await {
+        Ok(_) => StatusCode::OK.into_response(),
+        Err(e) => JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, e),
+    }
+}

compute_tools/src/http/routes/terminate.rs

@@ -1,7 +1,7 @@
 use crate::compute::{ComputeNode, forward_termination_signal};
 use crate::http::JsonResponse;
 use axum::extract::State;
-use axum::response::Response;
+use axum::response::{IntoResponse, Response};
 use axum_extra::extract::OptionalQuery;
 use compute_api::responses::{ComputeStatus, TerminateMode, TerminateResponse};
 use http::StatusCode;
@@ -33,7 +33,29 @@ pub(in crate::http) async fn terminate(
         if !matches!(state.status, ComputeStatus::Empty | ComputeStatus::Running) {
             return JsonResponse::invalid_status(state.status);
         }
+
+        // If compute is Empty, there's no Postgres to terminate. The regular compute_ctl termination path
+        // assumes Postgres to be configured and running, so we just special-handle this case by exiting
+        // the process directly.
+        if compute.params.lakebase_mode && state.status == ComputeStatus::Empty {
+            drop(state);
+            info!("terminating empty compute - will exit process");
+
+            // Queue a task to exit the process after 5 seconds. The 5-second delay aims to
+            // give enough time for the HTTP response to be sent so that HCM doesn't get an abrupt
+            // connection termination.
+            tokio::spawn(async {
+                tokio::time::sleep(tokio::time::Duration::from_secs(5)).await;
+                info!("exiting process after terminating empty compute");
+                std::process::exit(0);
+            });
+
+            return StatusCode::OK.into_response();
+        }
+
+        // For Running status, proceed with normal termination
         state.set_status(mode.into(), &compute.state_changed);
+        drop(state);
     }
 
     forward_termination_signal(false);

compute_tools/src/http/server.rs

@@ -23,7 +23,8 @@ use super::{
     middleware::authorize::Authorize,
     routes::{
         check_writability, configure, database_schema, dbs_and_roles, extension_server, extensions,
-        grants, insights, lfc, metrics, metrics_json, promote, status, terminate,
+        grants, hadron_liveness_probe, insights, lfc, metrics, metrics_json, promote,
+        refresh_configuration, status, terminate,
     },
 };
 use crate::compute::ComputeNode;
@@ -43,6 +44,7 @@ pub enum Server {
         port: u16,
         config: ComputeCtlConfig,
         compute_id: String,
+        instance_id: Option<String>,
     },
 }
 
@@ -67,7 +69,12 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                         post(extension_server::download_extension),
                     )
                     .route("/extensions", post(extensions::install_extension))
-                    .route("/grants", post(grants::add_grant));
+                    .route("/grants", post(grants::add_grant))
+                    // Hadron: Compute-initiated configuration refresh
+                    .route(
+                        "/refresh_configuration",
+                        post(refresh_configuration::refresh_configuration),
+                    );
 
                 // Add in any testing support
                 if cfg!(feature = "testing") {
@@ -79,7 +86,10 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                 router
             }
             Server::External {
-                config, compute_id, ..
+                config,
+                compute_id,
+                instance_id,
+                ..
             } => {
                 let unauthenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/metrics", get(metrics::get_metrics))
@@ -100,8 +110,13 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                     .route("/metrics.json", get(metrics_json::get_metrics))
                     .route("/status", get(status::get_status))
                     .route("/terminate", post(terminate::terminate))
+                    .route(
+                        "/hadron_liveness_probe",
+                        get(hadron_liveness_probe::hadron_liveness_probe),
+                    )
                     .layer(AsyncRequireAuthorizationLayer::new(Authorize::new(
                         compute_id.clone(),
+                        instance_id.clone(),
                         config.jwks.clone(),
                     )));
 

compute_tools/src/installed_extensions.rs

@@ -2,6 +2,7 @@ use std::collections::HashMap;
 
 use anyhow::Result;
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use once_cell::sync::Lazy;
 use tokio_postgres::error::Error as PostgresError;
 use tokio_postgres::{Client, Config, NoTls};
 
@@ -119,3 +120,7 @@ pub async fn get_installed_extensions(
         extensions: extensions_map.into_values().collect(),
     })
 }
+
+pub fn initialize_metrics() {
+    Lazy::force(&INSTALLED_EXTENSIONS);
+}

compute_tools/src/lib.rs

@@ -16,6 +16,7 @@ pub mod compute_prewarm;
 pub mod compute_promote;
 pub mod disk_quota;
 pub mod extension_server;
+pub mod hadron_metrics;
 pub mod installed_extensions;
 pub mod local_proxy;
 pub mod lsn_lease;
@@ -24,6 +25,7 @@ mod migration;
 pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
+pub mod pg_isready;
 pub mod pgbouncer;
 pub mod rsyslog;
 pub mod spec;

compute_tools/src/logger.rs

@@ -1,7 +1,10 @@
 use std::collections::HashMap;
+use std::sync::{LazyLock, RwLock};
+use tracing::Subscriber;
 use tracing::info;
-use tracing_subscriber::layer::SubscriberExt;
+use tracing_appender;
 use tracing_subscriber::prelude::*;
+use tracing_subscriber::{fmt, layer::SubscriberExt, registry::LookupSpan};
 
 /// Initialize logging to stderr, and OpenTelemetry tracing and exporter.
 ///
@@ -15,16 +18,44 @@ use tracing_subscriber::prelude::*;
 ///
 pub fn init_tracing_and_logging(
     default_log_level: &str,
-) -> anyhow::Result<Option<tracing_utils::Provider>> {
+    log_dir_opt: &Option<String>,
+) -> anyhow::Result<(
+    Option<tracing_utils::Provider>,
+    Option<tracing_appender::non_blocking::WorkerGuard>,
+)> {
     // Initialize Logging
     let env_filter = tracing_subscriber::EnvFilter::try_from_default_env()
         .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new(default_log_level));
 
+    // Standard output streams
     let fmt_layer = tracing_subscriber::fmt::layer()
         .with_ansi(false)
         .with_target(false)
         .with_writer(std::io::stderr);
 
+    // Logs with file rotation. Files in `$log_dir/pgcctl.yyyy-MM-dd`
+    let (json_to_file_layer, _file_logs_guard) = if let Some(log_dir) = log_dir_opt {
+        std::fs::create_dir_all(log_dir)?;
+        let file_logs_appender = tracing_appender::rolling::RollingFileAppender::builder()
+            .rotation(tracing_appender::rolling::Rotation::DAILY)
+            .filename_prefix("pgcctl")
+            // Lib appends to existing files, so we will keep files for up to 2 days even on restart loops.
+            // At minimum, log-daemon will have 1 day to detect and upload a file (if created right before midnight).
+            .max_log_files(2)
+            .build(log_dir)
+            .expect("Initializing rolling file appender should succeed");
+        let (file_logs_writer, _file_logs_guard) =
+            tracing_appender::non_blocking(file_logs_appender);
+        let json_to_file_layer = tracing_subscriber::fmt::layer()
+            .with_ansi(false)
+            .with_target(false)
+            .event_format(PgJsonLogShapeFormatter)
+            .with_writer(file_logs_writer);
+        (Some(json_to_file_layer), Some(_file_logs_guard))
+    } else {
+        (None, None)
+    };
+
     // Initialize OpenTelemetry
     let provider =
         tracing_utils::init_tracing("compute_ctl", tracing_utils::ExportConfig::default());
@@ -35,12 +66,13 @@ pub fn init_tracing_and_logging(
         .with(env_filter)
         .with(otlp_layer)
         .with(fmt_layer)
+        .with(json_to_file_layer)
         .init();
     tracing::info!("logging and tracing started");
 
     utils::logging::replace_panic_hook_with_tracing_panic_hook().forget();
 
-    Ok(provider)
+    Ok((provider, _file_logs_guard))
 }
 
 /// Replace all newline characters with a special character to make it
@@ -95,3 +127,157 @@ pub fn startup_context_from_env() -> Option<opentelemetry::Context> {
         None
     }
 }
+
+/// Track relevant id's
+const UNKNOWN_IDS: &str = r#""pg_instance_id": "", "pg_compute_id": """#;
+static IDS: LazyLock<RwLock<String>> = LazyLock::new(|| RwLock::new(UNKNOWN_IDS.to_string()));
+
+pub fn update_ids(instance_id: &Option<String>, compute_id: &Option<String>) -> anyhow::Result<()> {
+    let ids = format!(
+        r#""pg_instance_id": "{}", "pg_compute_id": "{}""#,
+        instance_id.as_ref().map(|s| s.as_str()).unwrap_or_default(),
+        compute_id.as_ref().map(|s| s.as_str()).unwrap_or_default()
+    );
+    let mut guard = IDS
+        .write()
+        .map_err(|e| anyhow::anyhow!("Log set id's rwlock poisoned: {}", e))?;
+    *guard = ids;
+    Ok(())
+}
+
+/// Massage compute_ctl logs into PG json log shape so we can use the same Lumberjack setup.
+struct PgJsonLogShapeFormatter;
+impl<S, N> fmt::format::FormatEvent<S, N> for PgJsonLogShapeFormatter
+where
+    S: Subscriber + for<'a> LookupSpan<'a>,
+    N: for<'a> fmt::format::FormatFields<'a> + 'static,
+{
+    fn format_event(
+        &self,
+        ctx: &fmt::FmtContext<'_, S, N>,
+        mut writer: fmt::format::Writer<'_>,
+        event: &tracing::Event<'_>,
+    ) -> std::fmt::Result {
+        // Format values from the event's metadata, and open message string
+        let metadata = event.metadata();
+        {
+            let ids_guard = IDS.read();
+            let ids = ids_guard
+                .as_ref()
+                .map(|guard| guard.as_str())
+                // Surpress so that we don't lose all uploaded/ file logs if something goes super wrong. We would notice the missing id's.
+                .unwrap_or(UNKNOWN_IDS);
+            write!(
+                &mut writer,
+                r#"{{"timestamp": "{}", "error_severity": "{}", "file_name": "{}", "backend_type": "compute_ctl_self", {}, "message": "#,
+                chrono::Utc::now().format("%Y-%m-%d %H:%M:%S%.3f GMT"),
+                metadata.level(),
+                metadata.target(),
+                ids
+            )?;
+        }
+
+        let mut message = String::new();
+        let message_writer = fmt::format::Writer::new(&mut message);
+
+        // Gather the message
+        ctx.field_format().format_fields(message_writer, event)?;
+
+        // TODO: any better options than to copy-paste this OSS span formatter?
+        // impl<S, N, T> FormatEvent<S, N> for Format<Full, T>
+        // https://docs.rs/tracing-subscriber/latest/tracing_subscriber/fmt/trait.FormatEvent.html#impl-FormatEvent%3CS,+N%3E-for-Format%3CFull,+T%3E
+
+        // write message, close bracket, and new line
+        writeln!(writer, "{}}}", serde_json::to_string(&message).unwrap())
+    }
+}
+
+#[cfg(feature = "testing")]
+#[cfg(test)]
+mod test {
+    use super::*;
+    use std::{cell::RefCell, io};
+
+    // Use thread_local! instead of Mutex for test isolation
+    thread_local! {
+        static WRITER_OUTPUT: RefCell<String> = const { RefCell::new(String::new()) };
+    }
+
+    #[derive(Clone, Default)]
+    struct StaticStringWriter;
+
+    impl io::Write for StaticStringWriter {
+        fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+            let output = String::from_utf8(buf.to_vec()).expect("Invalid UTF-8 in test output");
+            WRITER_OUTPUT.with(|s| s.borrow_mut().push_str(&output));
+            Ok(buf.len())
+        }
+
+        fn flush(&mut self) -> io::Result<()> {
+            Ok(())
+        }
+    }
+
+    impl fmt::MakeWriter<'_> for StaticStringWriter {
+        type Writer = Self;
+
+        fn make_writer(&self) -> Self::Writer {
+            Self
+        }
+    }
+
+    #[test]
+    fn test_log_pg_json_shape_formatter() {
+        // Use a scoped subscriber to prevent global state pollution
+        let subscriber = tracing_subscriber::registry().with(
+            tracing_subscriber::fmt::layer()
+                .with_ansi(false)
+                .with_target(false)
+                .event_format(PgJsonLogShapeFormatter)
+                .with_writer(StaticStringWriter),
+        );
+
+        let _ = update_ids(&Some("000".to_string()), &Some("111".to_string()));
+
+        // Clear any previous test state
+        WRITER_OUTPUT.with(|s| s.borrow_mut().clear());
+
+        let messages = [
+            "test message",
+            r#"json escape check:  name="BatchSpanProcessor.Flush.ExportError" reason="Other(reqwest::Error { kind: Request, url: \"http://localhost:4318/v1/traces\", source: hyper_
+            util::client::legacy::Error(Connect, ConnectError(\"tcp connect error\", Os { code: 111, kind: ConnectionRefused, message: \"Connection refused\" })) })" Failed during the export process"#,
+        ];
+
+        tracing::subscriber::with_default(subscriber, || {
+            for message in messages {
+                tracing::info!(message);
+            }
+        });
+        tracing::info!("not test message");
+
+        // Get captured output
+        let output = WRITER_OUTPUT.with(|s| s.borrow().clone());
+
+        let json_strings: Vec<&str> = output.lines().collect();
+        assert_eq!(
+            json_strings.len(),
+            messages.len(),
+            "Log didn't have the expected number of json strings."
+        );
+
+        let json_string_shape_regex = regex::Regex::new(
+            r#"\{"timestamp": "\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\.\d{3} GMT", "error_severity": "INFO", "file_name": ".+", "backend_type": "compute_ctl_self", "pg_instance_id": "000", "pg_compute_id": "111", "message": ".+"\}"#
+        ).unwrap();
+
+        for (i, expected_message) in messages.iter().enumerate() {
+            let json_string = json_strings[i];
+            assert!(
+                json_string_shape_regex.is_match(json_string),
+                "Json log didn't match expected pattern:\n{json_string}",
+            );
+            let parsed_json: serde_json::Value = serde_json::from_str(json_string).unwrap();
+            let actual_message = parsed_json["message"].as_str().unwrap();
+            assert_eq!(*expected_message, actual_message);
+        }
+    }
+}

compute_tools/src/pg_isready.rs

@@ -0,0 +1,30 @@
+use anyhow::{Context, anyhow};
+
+// Run `/usr/local/bin/pg_isready -p {port}`
+// Check the connectivity of PG
+// Success means PG is listening on the port and accepting connections
+// Note that PG does not need to authenticate the connection, nor reserve a connection quota for it.
+// See https://www.postgresql.org/docs/current/app-pg-isready.html
+pub fn pg_isready(bin: &str, port: u16) -> anyhow::Result<()> {
+    let child_result = std::process::Command::new(bin)
+        .arg("-p")
+        .arg(port.to_string())
+        .spawn();
+
+    child_result
+        .context("spawn() failed")
+        .and_then(|mut child| child.wait().context("wait() failed"))
+        .and_then(|status| match status.success() {
+            true => Ok(()),
+            false => Err(anyhow!("process exited with {status}")),
+        })
+        // wrap any prior error with the overall context that we couldn't run the command
+        .with_context(|| format!("could not run `{bin} --port {port}`"))
+}
+
+// It's safe to assume pg_isready is under the same directory with postgres,
+// because it is a PG util bin installed along with postgres
+pub fn get_pg_isready_bin(pgbin: &str) -> String {
+    let split = pgbin.split("/").collect::<Vec<&str>>();
+    split[0..split.len() - 1].join("/") + "/pg_isready"
+}

compute_tools/src/spec.rs

@@ -142,7 +142,7 @@ pub fn update_pg_hba(pgdata_path: &Path, databricks_pg_hba: Option<&String>) ->
     // Update pg_hba to contains databricks specfic settings before adding neon settings
     // PG uses the first record that matches to perform authentication, so we need to have
     // our rules before the default ones from neon.
-    // See https://www.postgresql.org/docs/16/auth-pg-hba-conf.html
+    // See https://www.postgresql.org/docs/current/auth-pg-hba-conf.html
     if let Some(databricks_pg_hba) = databricks_pg_hba {
         if config::line_in_file(
             &pghba_path,

control_plane/src/bin/neon_local.rs

@@ -560,7 +560,9 @@ enum EndpointCmd {
     Create(EndpointCreateCmdArgs),
     Start(EndpointStartCmdArgs),
     Reconfigure(EndpointReconfigureCmdArgs),
+    RefreshConfiguration(EndpointRefreshConfigurationArgs),
     Stop(EndpointStopCmdArgs),
+    UpdatePageservers(EndpointUpdatePageserversCmdArgs),
     GenerateJwt(EndpointGenerateJwtCmdArgs),
 }
 
@@ -721,6 +723,13 @@ struct EndpointReconfigureCmdArgs {
     safekeepers: Option<String>,
 }
 
+#[derive(clap::Args)]
+#[clap(about = "Refresh the endpoint's configuration by forcing it reload it's spec")]
+struct EndpointRefreshConfigurationArgs {
+    #[clap(help = "Postgres endpoint id")]
+    endpoint_id: String,
+}
+
 #[derive(clap::Args)]
 #[clap(about = "Stop an endpoint")]
 struct EndpointStopCmdArgs {
@@ -738,6 +747,16 @@ struct EndpointStopCmdArgs {
     mode: EndpointTerminateMode,
 }
 
+#[derive(clap::Args)]
+#[clap(about = "Update the pageservers in the spec file of the compute endpoint")]
+struct EndpointUpdatePageserversCmdArgs {
+    #[clap(help = "Postgres endpoint id")]
+    endpoint_id: String,
+
+    #[clap(short = 'p', long, help = "Specified pageserver id")]
+    pageserver_id: Option<NodeId>,
+}
+
 #[derive(clap::Args)]
 #[clap(about = "Generate a JWT for an endpoint")]
 struct EndpointGenerateJwtCmdArgs {
@@ -1517,7 +1536,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             let endpoint = cplane
                 .endpoints
                 .get(endpoint_id.as_str())
-                .ok_or_else(|| anyhow::anyhow!("endpoint {endpoint_id} not found"))?;
+                .ok_or_else(|| anyhow!("endpoint {endpoint_id} not found"))?;
 
             if !args.allow_multiple {
                 cplane.check_conflicting_endpoints(
@@ -1625,6 +1644,44 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
             println!("Starting existing endpoint {endpoint_id}...");
             endpoint.start(args).await?;
         }
+        EndpointCmd::UpdatePageservers(args) => {
+            let endpoint_id = &args.endpoint_id;
+            let endpoint = cplane
+                .endpoints
+                .get(endpoint_id.as_str())
+                .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
+            let pageservers = match args.pageserver_id {
+                Some(pageserver_id) => {
+                    let pageserver =
+                        PageServerNode::from_env(env, env.get_pageserver_conf(pageserver_id)?);
+
+                    vec![(
+                        PageserverProtocol::Libpq,
+                        pageserver.pg_connection_config.host().clone(),
+                        pageserver.pg_connection_config.port(),
+                    )]
+                }
+                None => {
+                    let storage_controller = StorageController::from_env(env);
+                    storage_controller
+                        .tenant_locate(endpoint.tenant_id)
+                        .await?
+                        .shards
+                        .into_iter()
+                        .map(|shard| {
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr)
+                                    .expect("Storage controller reported malformed host"),
+                                shard.listen_pg_port,
+                            )
+                        })
+                        .collect::<Vec<_>>()
+                }
+            };
+
+            endpoint.update_pageservers_in_config(pageservers).await?;
+        }
         EndpointCmd::Reconfigure(args) => {
             let endpoint_id = &args.endpoint_id;
             let endpoint = cplane
@@ -1678,6 +1735,14 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 .reconfigure(Some(pageservers), None, safekeepers, None)
                 .await?;
         }
+        EndpointCmd::RefreshConfiguration(args) => {
+            let endpoint_id = &args.endpoint_id;
+            let endpoint = cplane
+                .endpoints
+                .get(endpoint_id.as_str())
+                .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
+            endpoint.refresh_configuration().await?;
+        }
         EndpointCmd::Stop(args) => {
             let endpoint_id = &args.endpoint_id;
             let endpoint = cplane

control_plane/src/endpoint.rs

@@ -793,6 +793,7 @@ impl Endpoint {
                 autoprewarm: args.autoprewarm,
                 offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
                 suspend_timeout_seconds: -1, // Only used in neon_local.
+                databricks_settings: None,
             };
 
             // this strange code is needed to support respec() in tests
@@ -937,7 +938,9 @@ impl Endpoint {
                         | ComputeStatus::Configuration
                         | ComputeStatus::TerminationPendingFast
                         | ComputeStatus::TerminationPendingImmediate
-                        | ComputeStatus::Terminated => {
+                        | ComputeStatus::Terminated
+                        | ComputeStatus::RefreshConfigurationPending
+                        | ComputeStatus::RefreshConfiguration => {
                             bail!("unexpected compute status: {:?}", state.status)
                         }
                     }
@@ -960,6 +963,29 @@ impl Endpoint {
         Ok(())
     }
 
+    // Update the pageservers in the spec file of the endpoint. This is useful to test the spec refresh scenario.
+    pub async fn update_pageservers_in_config(
+        &self,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+    ) -> Result<()> {
+        let config_path = self.endpoint_path().join("config.json");
+        let mut config: ComputeConfig = {
+            let file = std::fs::File::open(&config_path)?;
+            serde_json::from_reader(file)?
+        };
+
+        let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
+        assert!(!pageserver_connstring.is_empty());
+        let mut spec = config.spec.unwrap();
+        spec.pageserver_connstring = Some(pageserver_connstring);
+        config.spec = Some(spec);
+
+        let file = std::fs::File::create(&config_path)?;
+        serde_json::to_writer_pretty(file, &config)?;
+
+        Ok(())
+    }
+
     // Call the /status HTTP API
     pub async fn get_status(&self) -> Result<ComputeStatusResponse> {
         let client = reqwest::Client::new();
@@ -1125,6 +1151,33 @@ impl Endpoint {
         Ok(response)
     }
 
+    pub async fn refresh_configuration(&self) -> Result<()> {
+        let client = reqwest::Client::builder()
+            .timeout(Duration::from_secs(30))
+            .build()
+            .unwrap();
+        let response = client
+            .post(format!(
+                "http://{}:{}/refresh_configuration",
+                self.internal_http_address.ip(),
+                self.internal_http_address.port()
+            ))
+            .send()
+            .await?;
+
+        let status = response.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            Ok(())
+        } else {
+            let url = response.url().to_owned();
+            let msg = match response.text().await {
+                Ok(err_body) => format!("Error: {err_body}"),
+                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
+            };
+            Err(anyhow::anyhow!(msg))
+        }
+    }
+
     pub fn connstr(&self, user: &str, db_name: &str) -> String {
         format!(
             "postgresql://{}@{}:{}/{}",

libs/compute_api/src/responses.rs

@@ -108,11 +108,10 @@ pub enum PromoteState {
     Failed { error: String },
 }
 
-#[derive(Deserialize, Serialize, Default, Debug, Clone)]
+#[derive(Deserialize, Default, Debug)]
 #[serde(rename_all = "snake_case")]
-/// Result of /safekeepers_lsn
-pub struct SafekeepersLsn {
-    pub safekeepers: String,
+pub struct PromoteConfig {
+    pub spec: ComputeSpec,
     pub wal_flush_lsn: utils::lsn::Lsn,
 }
 
@@ -173,6 +172,11 @@ pub enum ComputeStatus {
     TerminationPendingImmediate,
     // Terminated Postgres
     Terminated,
+    // A spec refresh is being requested
+    RefreshConfigurationPending,
+    // A spec refresh is being applied. We cannot refresh configuration again until the current
+    // refresh is done, i.e., signal_refresh_configuration() will return 500 error.
+    RefreshConfiguration,
 }
 
 #[derive(Deserialize, Serialize)]
@@ -185,6 +189,10 @@ impl Display for ComputeStatus {
         match self {
             ComputeStatus::Empty => f.write_str("empty"),
             ComputeStatus::ConfigurationPending => f.write_str("configuration-pending"),
+            ComputeStatus::RefreshConfiguration => f.write_str("refresh-configuration"),
+            ComputeStatus::RefreshConfigurationPending => {
+                f.write_str("refresh-configuration-pending")
+            }
             ComputeStatus::Init => f.write_str("init"),
             ComputeStatus::Running => f.write_str("running"),
             ComputeStatus::Configuration => f.write_str("configuration"),

libs/compute_api/src/spec.rs

@@ -193,6 +193,9 @@ pub struct ComputeSpec {
     ///
     /// We use this value to derive other values, such as the installed extensions metric.
     pub suspend_timeout_seconds: i64,
+
+    // Databricks specific options for compute instance.
+    pub databricks_settings: Option<DatabricksSettings>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.

libs/http-utils/src/endpoint.rs

@@ -558,11 +558,11 @@ async fn add_request_id_header_to_response(
     mut res: Response<Body>,
     req_info: RequestInfo,
 ) -> Result<Response<Body>, ApiError> {
-    if let Some(request_id) = req_info.context::<RequestId>() {
-        if let Ok(request_header_value) = HeaderValue::from_str(&request_id.0) {
-            res.headers_mut()
-                .insert(&X_REQUEST_ID_HEADER, request_header_value);
-        };
+    if let Some(request_id) = req_info.context::<RequestId>()
+        && let Ok(request_header_value) = HeaderValue::from_str(&request_id.0)
+    {
+        res.headers_mut()
+            .insert(&X_REQUEST_ID_HEADER, request_header_value);
     };
 
     Ok(res)

libs/http-utils/src/server.rs

@@ -72,10 +72,10 @@ impl Server {
             if err.is_incomplete_message() || err.is_closed() || err.is_timeout() {
                 return true;
             }
-            if let Some(inner) = err.source() {
-                if let Some(io) = inner.downcast_ref::<std::io::Error>() {
-                    return suppress_io_error(io);
-                }
+            if let Some(inner) = err.source()
+                && let Some(io) = inner.downcast_ref::<std::io::Error>()
+            {
+                return suppress_io_error(io);
             }
             false
         }

libs/metrics/src/lib.rs

@@ -129,6 +129,12 @@ impl<L: LabelGroup> InfoMetric<L> {
     }
 }
 
+impl<L: LabelGroup + Default> Default for InfoMetric<L, GaugeState> {
+    fn default() -> Self {
+        InfoMetric::new(L::default())
+    }
+}
+
 impl<L: LabelGroup, M: MetricType<Metadata = ()>> InfoMetric<L, M> {
     pub fn with_metric(label: L, metric: M) -> Self {
         Self {

libs/neon-shmem/src/hash.rs

@@ -363,7 +363,7 @@ where
     // TODO: An Iterator might be nicer. The communicator's clock algorithm needs to
     // _slowly_ iterate through all buckets with its clock hand,  without holding a lock.
     // If we switch to an Iterator, it must not hold the lock.
-    pub fn get_at_bucket(&self, pos: usize) -> Option<ValueReadGuard<(K, V)>> {
+    pub fn get_at_bucket(&self, pos: usize) -> Option<ValueReadGuard<'_, (K, V)>> {
         let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
         if pos >= map.buckets.len() {
             return None;

libs/pageserver_api/src/models.rs

@@ -1500,6 +1500,7 @@ pub struct TimelineArchivalConfigRequest {
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone)]
 pub struct TimelinePatchIndexPartRequest {
     pub rel_size_migration: Option<RelSizeMigration>,
+    pub rel_size_migrated_at: Option<Lsn>,
     pub gc_compaction_last_completed_lsn: Option<Lsn>,
     pub applied_gc_cutoff_lsn: Option<Lsn>,
     #[serde(default)]
@@ -1533,10 +1534,10 @@ pub enum RelSizeMigration {
     /// `None` is the same as `Some(RelSizeMigration::Legacy)`.
     Legacy,
     /// The tenant is migrating to the new rel_size format. Both old and new rel_size format are
-    /// persisted in the index part. The read path will read both formats and merge them.
+    /// persisted in the storage. The read path will read both formats and validate them.
     Migrating,
     /// The tenant has migrated to the new rel_size format. Only the new rel_size format is persisted
-    /// in the index part, and the read path will not read the old format.
+    /// in the storage, and the read path will not read the old format.
     Migrated,
 }
 
@@ -1619,6 +1620,7 @@ pub struct TimelineInfo {
 
     /// The status of the rel_size migration.
     pub rel_size_migration: Option<RelSizeMigration>,
+    pub rel_size_migrated_at: Option<Lsn>,
 
     /// Whether the timeline is invisible in synthetic size calculations.
     pub is_invisible: Option<bool>,

libs/proxy/tokio-postgres2/src/client.rs

@@ -15,6 +15,7 @@ use tokio::sync::mpsc;
 use crate::cancel_token::RawCancelToken;
 use crate::codec::{BackendMessages, FrontendMessage, RecordNotices};
 use crate::config::{Host, SslMode};
+use crate::connection::gc_bytesmut;
 use crate::query::RowStream;
 use crate::simple_query::SimpleQueryStream;
 use crate::types::{Oid, Type};
@@ -95,20 +96,13 @@ impl InnerClient {
         Ok(PartialQuery(Some(self)))
     }
 
-    // pub fn send_with_sync<F>(&mut self, f: F) -> Result<&mut Responses, Error>
-    // where
-    //     F: FnOnce(&mut BytesMut) -> Result<(), Error>,
-    // {
-    //     self.start()?.send_with_sync(f)
-    // }
-
     pub fn send_simple_query(&mut self, query: &str) -> Result<&mut Responses, Error> {
         self.responses.waiting += 1;
 
         self.buffer.clear();
         // simple queries do not need sync.
         frontend::query(query, &mut self.buffer).map_err(Error::encode)?;
-        let buf = self.buffer.split().freeze();
+        let buf = self.buffer.split();
         self.send_message(FrontendMessage::Raw(buf))
     }
 
@@ -125,7 +119,7 @@ impl Drop for PartialQuery<'_> {
         if let Some(client) = self.0.take() {
             client.buffer.clear();
             frontend::sync(&mut client.buffer);
-            let buf = client.buffer.split().freeze();
+            let buf = client.buffer.split();
             let _ = client.send_message(FrontendMessage::Raw(buf));
         }
     }
@@ -141,7 +135,7 @@ impl<'a> PartialQuery<'a> {
         client.buffer.clear();
         f(&mut client.buffer)?;
         frontend::flush(&mut client.buffer);
-        let buf = client.buffer.split().freeze();
+        let buf = client.buffer.split();
         client.send_message(FrontendMessage::Raw(buf))
     }
 
@@ -154,7 +148,7 @@ impl<'a> PartialQuery<'a> {
         client.buffer.clear();
         f(&mut client.buffer)?;
         frontend::sync(&mut client.buffer);
-        let buf = client.buffer.split().freeze();
+        let buf = client.buffer.split();
         let _ = client.send_message(FrontendMessage::Raw(buf));
 
         Ok(&mut self.0.take().unwrap().responses)
@@ -191,6 +185,7 @@ impl Client {
         ssl_mode: SslMode,
         process_id: i32,
         secret_key: i32,
+        write_buf: BytesMut,
     ) -> Client {
         Client {
             inner: InnerClient {
@@ -201,7 +196,7 @@ impl Client {
                     waiting: 0,
                     received: 0,
                 },
-                buffer: Default::default(),
+                buffer: write_buf,
             },
             cached_typeinfo: Default::default(),
 
@@ -292,8 +287,35 @@ impl Client {
         simple_query::batch_execute(self.inner_mut(), query).await
     }
 
-    pub async fn discard_all(&mut self) -> Result<ReadyForQueryStatus, Error> {
-        self.batch_execute("discard all").await
+    /// Similar to `discard_all`, but it does not clear any query plans
+    ///
+    /// This runs in the background, so it can be executed without `await`ing.
+    pub fn reset_session_background(&mut self) -> Result<(), Error> {
+        // "CLOSE ALL": closes any cursors
+        // "SET SESSION AUTHORIZATION DEFAULT": resets the current_user back to the session_user
+        // "RESET ALL": resets any GUCs back to their session defaults.
+        // "DEALLOCATE ALL": deallocates any prepared statements
+        // "UNLISTEN *": stops listening on all channels
+        // "SELECT pg_advisory_unlock_all();": unlocks all advisory locks
+        // "DISCARD TEMP;": drops all temporary tables
+        // "DISCARD SEQUENCES;": deallocates all cached sequence state
+
+        let _responses = self.inner_mut().send_simple_query(
+            "ROLLBACK;
+            CLOSE ALL;
+            SET SESSION AUTHORIZATION DEFAULT;
+            RESET ALL;
+            DEALLOCATE ALL;
+            UNLISTEN *;
+            SELECT pg_advisory_unlock_all();
+            DISCARD TEMP;
+            DISCARD SEQUENCES;",
+        )?;
+
+        // Clean up memory usage.
+        gc_bytesmut(&mut self.inner_mut().buffer);
+
+        Ok(())
     }
 
     /// Begins a new database transaction.

libs/proxy/tokio-postgres2/src/codec.rs

@@ -1,13 +1,13 @@
 use std::io;
 
-use bytes::{Bytes, BytesMut};
+use bytes::BytesMut;
 use fallible_iterator::FallibleIterator;
 use postgres_protocol2::message::backend;
 use tokio::sync::mpsc::UnboundedSender;
 use tokio_util::codec::{Decoder, Encoder};
 
 pub enum FrontendMessage {
-    Raw(Bytes),
+    Raw(BytesMut),
     RecordNotices(RecordNotices),
 }
 
@@ -17,7 +17,10 @@ pub struct RecordNotices {
 }
 
 pub enum BackendMessage {
-    Normal { messages: BackendMessages },
+    Normal {
+        messages: BackendMessages,
+        ready: bool,
+    },
     Async(backend::Message),
 }
 
@@ -40,11 +43,11 @@ impl FallibleIterator for BackendMessages {
 
 pub struct PostgresCodec;
 
-impl Encoder<Bytes> for PostgresCodec {
+impl Encoder<BytesMut> for PostgresCodec {
     type Error = io::Error;
 
-    fn encode(&mut self, item: Bytes, dst: &mut BytesMut) -> io::Result<()> {
-        dst.extend_from_slice(&item);
+    fn encode(&mut self, item: BytesMut, dst: &mut BytesMut) -> io::Result<()> {
+        dst.unsplit(item);
         Ok(())
     }
 }
@@ -56,6 +59,7 @@ impl Decoder for PostgresCodec {
     fn decode(&mut self, src: &mut BytesMut) -> Result<Option<BackendMessage>, io::Error> {
         let mut idx = 0;
 
+        let mut ready = false;
         while let Some(header) = backend::Header::parse(&src[idx..])? {
             let len = header.len() as usize + 1;
             if src[idx..].len() < len {
@@ -79,6 +83,7 @@ impl Decoder for PostgresCodec {
             idx += len;
 
             if header.tag() == backend::READY_FOR_QUERY_TAG {
+                ready = true;
                 break;
             }
         }
@@ -88,6 +93,7 @@ impl Decoder for PostgresCodec {
         } else {
             Ok(Some(BackendMessage::Normal {
                 messages: BackendMessages(src.split_to(idx)),
+                ready,
             }))
         }
     }

libs/proxy/tokio-postgres2/src/config.rs

@@ -11,9 +11,8 @@ use tokio::io::{AsyncRead, AsyncWrite};
 use tokio::net::TcpStream;
 
 use crate::connect::connect;
-use crate::connect_raw::{RawConnection, connect_raw};
+use crate::connect_raw::{self, StartupStream};
 use crate::connect_tls::connect_tls;
-use crate::maybe_tls_stream::MaybeTlsStream;
 use crate::tls::{MakeTlsConnect, TlsConnect, TlsStream};
 use crate::{Client, Connection, Error};
 
@@ -244,24 +243,27 @@ impl Config {
         &self,
         stream: S,
         tls: T,
-    ) -> Result<RawConnection<S, T::Stream>, Error>
+    ) -> Result<StartupStream<S, T::Stream>, Error>
     where
         S: AsyncRead + AsyncWrite + Unpin,
         T: TlsConnect<S>,
     {
         let stream = connect_tls(stream, self.ssl_mode, tls).await?;
-        connect_raw(stream, self).await
+        let mut stream = StartupStream::new(stream);
+        connect_raw::authenticate(&mut stream, self).await?;
+
+        Ok(stream)
     }
 
-    pub async fn authenticate<S, T>(
+    pub fn authenticate<S, T>(
         &self,
-        stream: MaybeTlsStream<S, T>,
-    ) -> Result<RawConnection<S, T>, Error>
+        stream: &mut StartupStream<S, T>,
+    ) -> impl Future<Output = Result<(), Error>>
     where
         S: AsyncRead + AsyncWrite + Unpin,
         T: TlsStream + Unpin,
     {
-        connect_raw(stream, self).await
+        connect_raw::authenticate(stream, self)
     }
 }
 

libs/proxy/tokio-postgres2/src/connect.rs

@@ -1,15 +1,17 @@
 use std::net::IpAddr;
 
+use futures_util::TryStreamExt;
+use postgres_protocol2::message::backend::Message;
+use tokio::io::{AsyncRead, AsyncWrite};
 use tokio::net::TcpStream;
 use tokio::sync::mpsc;
 
 use crate::client::SocketConfig;
-use crate::config::Host;
-use crate::connect_raw::connect_raw;
+use crate::config::{Host, SslMode};
+use crate::connect_raw::StartupStream;
 use crate::connect_socket::connect_socket;
-use crate::connect_tls::connect_tls;
 use crate::tls::{MakeTlsConnect, TlsConnect};
-use crate::{Client, Config, Connection, Error, RawConnection};
+use crate::{Client, Config, Connection, Error};
 
 pub async fn connect<T>(
     tls: &T,
@@ -43,34 +45,78 @@ where
     T: TlsConnect<TcpStream>,
 {
     let socket = connect_socket(host_addr, host, port, config.connect_timeout).await?;
-    let stream = connect_tls(socket, config.ssl_mode, tls).await?;
-    let RawConnection {
+    let stream = config.tls_and_authenticate(socket, tls).await?;
+    managed(
         stream,
-        parameters: _,
-        delayed_notice: _,
-        process_id,
-        secret_key,
-    } = connect_raw(stream, config).await?;
+        host_addr,
+        host.clone(),
+        port,
+        config.ssl_mode,
+        config.connect_timeout,
+    )
+    .await
+}
+
+pub async fn managed<TlsStream>(
+    mut stream: StartupStream<TcpStream, TlsStream>,
+    host_addr: Option<IpAddr>,
+    host: Host,
+    port: u16,
+    ssl_mode: SslMode,
+    connect_timeout: Option<std::time::Duration>,
+) -> Result<(Client, Connection<TcpStream, TlsStream>), Error>
+where
+    TlsStream: AsyncRead + AsyncWrite + Unpin,
+{
+    let (process_id, secret_key) = wait_until_ready(&mut stream).await?;
 
     let socket_config = SocketConfig {
         host_addr,
-        host: host.clone(),
+        host,
         port,
-        connect_timeout: config.connect_timeout,
+        connect_timeout,
     };
 
+    let mut stream = stream.into_framed();
+    let write_buf = std::mem::take(stream.write_buffer_mut());
+
     let (client_tx, conn_rx) = mpsc::unbounded_channel();
     let (conn_tx, client_rx) = mpsc::channel(4);
     let client = Client::new(
         client_tx,
         client_rx,
         socket_config,
-        config.ssl_mode,
+        ssl_mode,
         process_id,
         secret_key,
+        write_buf,
     );
 
     let connection = Connection::new(stream, conn_tx, conn_rx);
 
     Ok((client, connection))
 }
+
+async fn wait_until_ready<S, T>(stream: &mut StartupStream<S, T>) -> Result<(i32, i32), Error>
+where
+    S: AsyncRead + AsyncWrite + Unpin,
+    T: AsyncRead + AsyncWrite + Unpin,
+{
+    let mut process_id = 0;
+    let mut secret_key = 0;
+
+    loop {
+        match stream.try_next().await.map_err(Error::io)? {
+            Some(Message::BackendKeyData(body)) => {
+                process_id = body.process_id();
+                secret_key = body.secret_key();
+            }
+            // These values are currently not used by `Client`/`Connection`. Ignore them.
+            Some(Message::ParameterStatus(_)) | Some(Message::NoticeResponse(_)) => {}
+            Some(Message::ReadyForQuery(_)) => return Ok((process_id, secret_key)),
+            Some(Message::ErrorResponse(body)) => return Err(Error::db(body)),
+            Some(_) => return Err(Error::unexpected_message()),
+            None => return Err(Error::closed()),
+        }
+    }
+}

libs/proxy/tokio-postgres2/src/connect_raw.rs

@@ -1,52 +1,27 @@
-use std::collections::HashMap;
 use std::io;
 use std::pin::Pin;
-use std::task::{Context, Poll};
+use std::task::{Context, Poll, ready};
 
-use bytes::{Bytes, BytesMut};
+use bytes::BytesMut;
 use fallible_iterator::FallibleIterator;
-use futures_util::{Sink, SinkExt, Stream, TryStreamExt, ready};
+use futures_util::{SinkExt, Stream, TryStreamExt};
 use postgres_protocol2::authentication::sasl;
 use postgres_protocol2::authentication::sasl::ScramSha256;
-use postgres_protocol2::message::backend::{AuthenticationSaslBody, Message, NoticeResponseBody};
+use postgres_protocol2::message::backend::{AuthenticationSaslBody, Message};
 use postgres_protocol2::message::frontend;
-use tokio::io::{AsyncRead, AsyncWrite};
-use tokio_util::codec::Framed;
+use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
+use tokio_util::codec::{Framed, FramedParts};
 
 use crate::Error;
-use crate::codec::{BackendMessage, BackendMessages, PostgresCodec};
+use crate::codec::PostgresCodec;
 use crate::config::{self, AuthKeys, Config};
+use crate::connection::{GC_THRESHOLD, INITIAL_CAPACITY};
 use crate::maybe_tls_stream::MaybeTlsStream;
 use crate::tls::TlsStream;
 
 pub struct StartupStream<S, T> {
     inner: Framed<MaybeTlsStream<S, T>, PostgresCodec>,
-    buf: BackendMessages,
-    delayed_notice: Vec<NoticeResponseBody>,
-}
-
-impl<S, T> Sink<Bytes> for StartupStream<S, T>
-where
-    S: AsyncRead + AsyncWrite + Unpin,
-    T: AsyncRead + AsyncWrite + Unpin,
-{
-    type Error = io::Error;
-
-    fn poll_ready(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.inner).poll_ready(cx)
-    }
-
-    fn start_send(mut self: Pin<&mut Self>, item: Bytes) -> io::Result<()> {
-        Pin::new(&mut self.inner).start_send(item)
-    }
-
-    fn poll_flush(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.inner).poll_flush(cx)
-    }
-
-    fn poll_close(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.inner).poll_close(cx)
-    }
+    read_buf: BytesMut,
 }
 
 impl<S, T> Stream for StartupStream<S, T>
@@ -56,78 +31,109 @@ where
 {
     type Item = io::Result<Message>;
 
-    fn poll_next(
-        mut self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-    ) -> Poll<Option<io::Result<Message>>> {
-        loop {
-            match self.buf.next() {
-                Ok(Some(message)) => return Poll::Ready(Some(Ok(message))),
-                Ok(None) => {}
-                Err(e) => return Poll::Ready(Some(Err(e))),
-            }
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        // We don't use `self.inner.poll_next()` as that might over-read into the read buffer.
 
-            match ready!(Pin::new(&mut self.inner).poll_next(cx)) {
-                Some(Ok(BackendMessage::Normal { messages, .. })) => self.buf = messages,
-                Some(Ok(BackendMessage::Async(message))) => return Poll::Ready(Some(Ok(message))),
-                Some(Err(e)) => return Poll::Ready(Some(Err(e))),
-                None => return Poll::Ready(None),
-            }
+        // read 1 byte tag, 4 bytes length.
+        let header = ready!(self.as_mut().poll_fill_buf_exact(cx, 5)?);
+
+        let len = u32::from_be_bytes(header[1..5].try_into().unwrap());
+        if len < 4 {
+            return Poll::Ready(Some(Err(std::io::Error::other(
+                "postgres message too small",
+            ))));
         }
+        if len >= 65536 {
+            return Poll::Ready(Some(Err(std::io::Error::other(
+                "postgres message too large",
+            ))));
+        }
+
+        // the tag is an additional byte.
+        let _message = ready!(self.as_mut().poll_fill_buf_exact(cx, len as usize + 1)?);
+
+        // Message::parse will remove the all the bytes from the buffer.
+        Poll::Ready(Message::parse(&mut self.read_buf).transpose())
     }
 }
 
-pub struct RawConnection<S, T> {
-    pub stream: Framed<MaybeTlsStream<S, T>, PostgresCodec>,
-    pub parameters: HashMap<String, String>,
-    pub delayed_notice: Vec<NoticeResponseBody>,
-    pub process_id: i32,
-    pub secret_key: i32,
-}
-
-pub async fn connect_raw<S, T>(
-    stream: MaybeTlsStream<S, T>,
-    config: &Config,
-) -> Result<RawConnection<S, T>, Error>
-where
-    S: AsyncRead + AsyncWrite + Unpin,
-    T: TlsStream + Unpin,
-{
-    let mut stream = StartupStream {
-        inner: Framed::new(stream, PostgresCodec),
-        buf: BackendMessages::empty(),
-        delayed_notice: Vec::new(),
-    };
-
-    startup(&mut stream, config).await?;
-    authenticate(&mut stream, config).await?;
-    let (process_id, secret_key, parameters) = read_info(&mut stream).await?;
-
-    Ok(RawConnection {
-        stream: stream.inner,
-        parameters,
-        delayed_notice: stream.delayed_notice,
-        process_id,
-        secret_key,
-    })
-}
-
-async fn startup<S, T>(stream: &mut StartupStream<S, T>, config: &Config) -> Result<(), Error>
+impl<S, T> StartupStream<S, T>
 where
     S: AsyncRead + AsyncWrite + Unpin,
     T: AsyncRead + AsyncWrite + Unpin,
 {
-    let mut buf = BytesMut::new();
-    frontend::startup_message(&config.server_params, &mut buf).map_err(Error::encode)?;
+    /// Fill the buffer until it's the exact length provided. No additional data will be read from the socket.
+    ///
+    /// If the current buffer length is greater, nothing happens.
+    fn poll_fill_buf_exact(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        len: usize,
+    ) -> Poll<Result<&[u8], std::io::Error>> {
+        let this = self.get_mut();
+        let mut stream = Pin::new(this.inner.get_mut());
 
-    stream.send(buf.freeze()).await.map_err(Error::io)
+        let mut n = this.read_buf.len();
+        while n < len {
+            this.read_buf.resize(len, 0);
+
+            let mut buf = ReadBuf::new(&mut this.read_buf[..]);
+            buf.set_filled(n);
+
+            if stream.as_mut().poll_read(cx, &mut buf)?.is_pending() {
+                this.read_buf.truncate(n);
+                return Poll::Pending;
+            }
+
+            if buf.filled().len() == n {
+                return Poll::Ready(Err(std::io::Error::new(
+                    std::io::ErrorKind::UnexpectedEof,
+                    "early eof",
+                )));
+            }
+            n = buf.filled().len();
+
+            this.read_buf.truncate(n);
+        }
+
+        Poll::Ready(Ok(&this.read_buf[..len]))
+    }
+
+    pub fn into_framed(mut self) -> Framed<MaybeTlsStream<S, T>, PostgresCodec> {
+        *self.inner.read_buffer_mut() = self.read_buf;
+        self.inner
+    }
+
+    pub fn new(io: MaybeTlsStream<S, T>) -> Self {
+        let mut parts = FramedParts::new(io, PostgresCodec);
+        parts.write_buf = BytesMut::with_capacity(INITIAL_CAPACITY);
+
+        let mut inner = Framed::from_parts(parts);
+
+        // This is the default already, but nice to be explicit.
+        // We divide by two because writes will overshoot the boundary.
+        // We don't want constant overshoots to cause us to constantly re-shrink the buffer.
+        inner.set_backpressure_boundary(GC_THRESHOLD / 2);
+
+        Self {
+            inner,
+            read_buf: BytesMut::with_capacity(INITIAL_CAPACITY),
+        }
+    }
 }
 
-async fn authenticate<S, T>(stream: &mut StartupStream<S, T>, config: &Config) -> Result<(), Error>
+pub(crate) async fn authenticate<S, T>(
+    stream: &mut StartupStream<S, T>,
+    config: &Config,
+) -> Result<(), Error>
 where
     S: AsyncRead + AsyncWrite + Unpin,
     T: TlsStream + Unpin,
 {
+    frontend::startup_message(&config.server_params, stream.inner.write_buffer_mut())
+        .map_err(Error::encode)?;
+
+    stream.inner.flush().await.map_err(Error::io)?;
     match stream.try_next().await.map_err(Error::io)? {
         Some(Message::AuthenticationOk) => {
             can_skip_channel_binding(config)?;
@@ -141,7 +147,8 @@ where
                 .as_ref()
                 .ok_or_else(|| Error::config("password missing".into()))?;
 
-            authenticate_password(stream, pass).await?;
+            frontend::password_message(pass, stream.inner.write_buffer_mut())
+                .map_err(Error::encode)?;
         }
         Some(Message::AuthenticationSasl(body)) => {
             authenticate_sasl(stream, body, config).await?;
@@ -160,6 +167,7 @@ where
         None => return Err(Error::closed()),
     }
 
+    stream.inner.flush().await.map_err(Error::io)?;
     match stream.try_next().await.map_err(Error::io)? {
         Some(Message::AuthenticationOk) => Ok(()),
         Some(Message::ErrorResponse(body)) => Err(Error::db(body)),
@@ -177,20 +185,6 @@ fn can_skip_channel_binding(config: &Config) -> Result<(), Error> {
     }
 }
 
-async fn authenticate_password<S, T>(
-    stream: &mut StartupStream<S, T>,
-    password: &[u8],
-) -> Result<(), Error>
-where
-    S: AsyncRead + AsyncWrite + Unpin,
-    T: AsyncRead + AsyncWrite + Unpin,
-{
-    let mut buf = BytesMut::new();
-    frontend::password_message(password, &mut buf).map_err(Error::encode)?;
-
-    stream.send(buf.freeze()).await.map_err(Error::io)
-}
-
 async fn authenticate_sasl<S, T>(
     stream: &mut StartupStream<S, T>,
     body: AuthenticationSaslBody,
@@ -245,10 +239,10 @@ where
         return Err(Error::config("password or auth keys missing".into()));
     };
 
-    let mut buf = BytesMut::new();
-    frontend::sasl_initial_response(mechanism, scram.message(), &mut buf).map_err(Error::encode)?;
-    stream.send(buf.freeze()).await.map_err(Error::io)?;
+    frontend::sasl_initial_response(mechanism, scram.message(), stream.inner.write_buffer_mut())
+        .map_err(Error::encode)?;
 
+    stream.inner.flush().await.map_err(Error::io)?;
     let body = match stream.try_next().await.map_err(Error::io)? {
         Some(Message::AuthenticationSaslContinue(body)) => body,
         Some(Message::ErrorResponse(body)) => return Err(Error::db(body)),
@@ -261,10 +255,10 @@ where
         .await
         .map_err(|e| Error::authentication(e.into()))?;
 
-    let mut buf = BytesMut::new();
-    frontend::sasl_response(scram.message(), &mut buf).map_err(Error::encode)?;
-    stream.send(buf.freeze()).await.map_err(Error::io)?;
+    frontend::sasl_response(scram.message(), stream.inner.write_buffer_mut())
+        .map_err(Error::encode)?;
 
+    stream.inner.flush().await.map_err(Error::io)?;
     let body = match stream.try_next().await.map_err(Error::io)? {
         Some(Message::AuthenticationSaslFinal(body)) => body,
         Some(Message::ErrorResponse(body)) => return Err(Error::db(body)),
@@ -278,35 +272,3 @@ where
 
     Ok(())
 }
-
-async fn read_info<S, T>(
-    stream: &mut StartupStream<S, T>,
-) -> Result<(i32, i32, HashMap<String, String>), Error>
-where
-    S: AsyncRead + AsyncWrite + Unpin,
-    T: AsyncRead + AsyncWrite + Unpin,
-{
-    let mut process_id = 0;
-    let mut secret_key = 0;
-    let mut parameters = HashMap::new();
-
-    loop {
-        match stream.try_next().await.map_err(Error::io)? {
-            Some(Message::BackendKeyData(body)) => {
-                process_id = body.process_id();
-                secret_key = body.secret_key();
-            }
-            Some(Message::ParameterStatus(body)) => {
-                parameters.insert(
-                    body.name().map_err(Error::parse)?.to_string(),
-                    body.value().map_err(Error::parse)?.to_string(),
-                );
-            }
-            Some(Message::NoticeResponse(body)) => stream.delayed_notice.push(body),
-            Some(Message::ReadyForQuery(_)) => return Ok((process_id, secret_key, parameters)),
-            Some(Message::ErrorResponse(body)) => return Err(Error::db(body)),
-            Some(_) => return Err(Error::unexpected_message()),
-            None => return Err(Error::closed()),
-        }
-    }
-}

libs/proxy/tokio-postgres2/src/connection.rs

@@ -44,6 +44,27 @@ pub struct Connection<S, T> {
     state: State,
 }
 
+pub const INITIAL_CAPACITY: usize = 2 * 1024;
+pub const GC_THRESHOLD: usize = 16 * 1024;
+
+/// Gargabe collect the [`BytesMut`] if it has too much spare capacity.
+pub fn gc_bytesmut(buf: &mut BytesMut) {
+    // We use a different mode to shrink the buf when above the threshold.
+    // When above the threshold, we only re-allocate when the buf has 2x spare capacity.
+    let reclaim = GC_THRESHOLD.checked_sub(buf.len()).unwrap_or(buf.len());
+
+    // `try_reclaim` tries to get the capacity from any shared `BytesMut`s,
+    // before then comparing the length against the capacity.
+    if buf.try_reclaim(reclaim) {
+        let capacity = usize::max(buf.len(), INITIAL_CAPACITY);
+
+        // Allocate a new `BytesMut` so that we deallocate the old version.
+        let mut new = BytesMut::with_capacity(capacity);
+        new.extend_from_slice(buf);
+        *buf = new;
+    }
+}
+
 pub enum Never {}
 
 impl<S, T> Connection<S, T>
@@ -86,7 +107,14 @@ where
                             continue;
                         }
                         BackendMessage::Async(_) => continue,
-                        BackendMessage::Normal { messages } => messages,
+                        BackendMessage::Normal { messages, ready } => {
+                            // if we read a ReadyForQuery from postgres, let's try GC the read buffer.
+                            if ready {
+                                gc_bytesmut(self.stream.read_buffer_mut());
+                            }
+
+                            messages
+                        }
                     }
                 }
             };
@@ -177,12 +205,7 @@ where
                 // Send a terminate message to postgres
                 Poll::Ready(None) => {
                     trace!("poll_write: at eof, terminating");
-                    let mut request = BytesMut::new();
-                    frontend::terminate(&mut request);
-
-                    Pin::new(&mut self.stream)
-                        .start_send(request.freeze())
-                        .map_err(Error::io)?;
+                    frontend::terminate(self.stream.write_buffer_mut());
 
                     trace!("poll_write: sent eof, closing");
                     trace!("poll_write: done");
@@ -205,6 +228,13 @@ where
         {
             Poll::Ready(()) => {
                 trace!("poll_flush: flushed");
+
+                // Since our codec prefers to share the buffer with the `Client`,
+                // if we don't release our share, then the `Client` would have to re-alloc
+                // the buffer when they next use it.
+                debug_assert!(self.stream.write_buffer().is_empty());
+                *self.stream.write_buffer_mut() = BytesMut::new();
+
                 Poll::Ready(Ok(()))
             }
             Poll::Pending => {

libs/proxy/tokio-postgres2/src/error/mod.rs

@@ -452,16 +452,16 @@ impl Error {
         Error(Box::new(ErrorInner { kind, cause }))
     }
 
-    pub(crate) fn closed() -> Error {
+    pub fn closed() -> Error {
         Error::new(Kind::Closed, None)
     }
 
-    pub(crate) fn unexpected_message() -> Error {
+    pub fn unexpected_message() -> Error {
         Error::new(Kind::UnexpectedMessage, None)
     }
 
     #[allow(clippy::needless_pass_by_value)]
-    pub(crate) fn db(error: ErrorResponseBody) -> Error {
+    pub fn db(error: ErrorResponseBody) -> Error {
         match DbError::parse(&mut error.fields()) {
             Ok(e) => Error::new(Kind::Db, Some(Box::new(e))),
             Err(e) => Error::new(Kind::Parse, Some(Box::new(e))),
@@ -493,7 +493,7 @@ impl Error {
         Error::new(Kind::Tls, Some(e))
     }
 
-    pub(crate) fn io(e: io::Error) -> Error {
+    pub fn io(e: io::Error) -> Error {
         Error::new(Kind::Io, Some(Box::new(e)))
     }
 

libs/proxy/tokio-postgres2/src/lib.rs

@@ -6,7 +6,6 @@ use postgres_protocol2::message::backend::ReadyForQueryBody;
 pub use crate::cancel_token::{CancelToken, RawCancelToken};
 pub use crate::client::{Client, SocketConfig};
 pub use crate::config::Config;
-pub use crate::connect_raw::RawConnection;
 pub use crate::connection::Connection;
 pub use crate::error::Error;
 pub use crate::generic_client::GenericClient;
@@ -49,8 +48,8 @@ mod cancel_token;
 mod client;
 mod codec;
 pub mod config;
-mod connect;
-mod connect_raw;
+pub mod connect;
+pub mod connect_raw;
 mod connect_socket;
 mod connect_tls;
 mod connection;

libs/safekeeper_api/src/models.rs

@@ -301,7 +301,12 @@ pub struct PullTimelineRequest {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
     pub http_hosts: Vec<String>,
-    pub ignore_tombstone: Option<bool>,
+    /// Membership configuration to switch to after pull.
+    /// It guarantees that if pull_timeline returns successfully, the timeline will
+    /// not be deleted by request with an older generation.
+    /// Storage controller always sets this field.
+    /// None is only allowed for manual pull_timeline requests.
+    pub mconf: Option<Configuration>,
 }
 
 #[derive(Debug, Serialize, Deserialize)]

libs/tracing-utils/Cargo.toml

@@ -8,7 +8,7 @@ license.workspace = true
 hyper0.workspace = true
 opentelemetry = { workspace = true, features = ["trace"] }
 opentelemetry_sdk = { workspace = true, features = ["rt-tokio"] }
-opentelemetry-otlp = { workspace = true, default-features = false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry-otlp = { workspace = true, default-features = false, features = ["http-proto", "trace", "http", "reqwest-blocking-client"] }
 opentelemetry-semantic-conventions.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tracing.workspace = true

libs/tracing-utils/src/lib.rs

@@ -104,13 +104,7 @@ fn init_tracing_internal(service_name: String, export_config: ExportConfig) -> P
     );
 
     Provider::builder()
-        .with_span_processor(
-            opentelemetry_sdk::trace::span_processor_with_async_runtime::BatchSpanProcessor::builder(
-                exporter,
-                opentelemetry_sdk::runtime::Tokio,
-            )
-            .build(),
-        )
+        .with_batch_exporter(exporter)
         .with_resource(
             opentelemetry_sdk::Resource::builder()
                 .with_service_name(service_name)

libs/tracing-utils/src/perf_span.rs

@@ -49,7 +49,7 @@ impl PerfSpan {
         }
     }
 
-    pub fn enter(&self) -> PerfSpanEntered {
+    pub fn enter(&self) -> PerfSpanEntered<'_> {
         if let Some(ref id) = self.inner.id() {
             self.dispatch.enter(id);
         }

libs/walproposer/src/api_bindings.rs

@@ -429,9 +429,11 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
     };
 
     let empty_wal_rate_limiter = crate::bindings::WalRateLimiter {
+        effective_max_wal_bytes_per_second: crate::bindings::pg_atomic_uint32 { value: 0 },
         should_limit: crate::bindings::pg_atomic_uint32 { value: 0 },
         sent_bytes: 0,
-        last_recorded_time_us: crate::bindings::pg_atomic_uint64 { value: 0 },
+        batch_start_time_us: crate::bindings::pg_atomic_uint64 { value: 0 },
+        batch_end_time_us: crate::bindings::pg_atomic_uint64 { value: 0 },
     };
 
     crate::bindings::WalproposerShmemState {

pageserver/src/bin/pageserver.rs

@@ -715,7 +715,7 @@ fn start_pageserver(
                 disk_usage_eviction_state,
                 deletion_queue.new_client(),
                 secondary_controller,
-                feature_resolver,
+                feature_resolver.clone(),
             )
             .context("Failed to initialize router state")?,
         );
@@ -841,14 +841,14 @@ fn start_pageserver(
         } else {
             None
         },
+        feature_resolver.clone(),
     );
 
-    // Spawn a Pageserver gRPC server task. It will spawn separate tasks for
-    // each stream/request.
+    // Spawn a Pageserver gRPC server task. It will spawn separate tasks for each request/stream.
+    // It uses a separate compute request Tokio runtime (COMPUTE_REQUEST_RUNTIME).
     //
-    // TODO: this uses a separate Tokio runtime for the page service. If we want
-    // other gRPC services, they will need their own port and runtime. Is this
-    // necessary?
+    // NB: this port is exposed to computes. It should only provide services that we're okay with
+    // computes accessing. Internal services should use a separate port.
     let mut page_service_grpc = None;
     if let Some(grpc_listener) = grpc_listener {
         page_service_grpc = Some(GrpcPageServiceHandler::spawn(

pageserver/src/http/routes.rs

@@ -484,6 +484,8 @@ async fn build_timeline_info_common(
         *timeline.get_applied_gc_cutoff_lsn(),
     );
 
+    let (rel_size_migration, rel_size_migrated_at) = timeline.get_rel_size_v2_status();
+
     let info = TimelineInfo {
         tenant_id: timeline.tenant_shard_id,
         timeline_id: timeline.timeline_id,
@@ -515,7 +517,8 @@ async fn build_timeline_info_common(
 
         state,
         is_archived: Some(is_archived),
-        rel_size_migration: Some(timeline.get_rel_size_v2_status()),
+        rel_size_migration: Some(rel_size_migration),
+        rel_size_migrated_at,
         is_invisible: Some(is_invisible),
 
         walreceiver_status,
@@ -930,9 +933,16 @@ async fn timeline_patch_index_part_handler(
             active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
                 .await?;
 
+        if request_data.rel_size_migration.is_none() && request_data.rel_size_migrated_at.is_some()
+        {
+            return Err(ApiError::BadRequest(anyhow!(
+                "updating rel_size_migrated_at without rel_size_migration is not allowed"
+            )));
+        }
+
         if let Some(rel_size_migration) = request_data.rel_size_migration {
             timeline
-                .update_rel_size_v2_status(rel_size_migration)
+                .update_rel_size_v2_status(rel_size_migration, request_data.rel_size_migrated_at)
                 .map_err(ApiError::InternalServerError)?;
         }
 
@@ -1995,6 +2005,10 @@ async fn put_tenant_location_config_handler(
     let state = get_state(&request);
     let conf = state.conf;
 
+    fail::fail_point!("put-location-conf-handler", |_| {
+        Err(ApiError::ResourceUnavailable("failpoint".into()))
+    });
+
     // The `Detached` state is special, it doesn't upsert a tenant, it removes
     // its local disk content and drops it from memory.
     if let LocationConfigMode::Detached = request_data.config.mode {

pageserver/src/import_datadir.rs

@@ -57,7 +57,7 @@ pub async fn import_timeline_from_postgres_datadir(
 
     // TODO this shoud be start_lsn, which is not necessarily equal to end_lsn (aka lsn)
     // Then fishing out pg_control would be unnecessary
-    let mut modification = tline.begin_modification(pgdata_lsn);
+    let mut modification = tline.begin_modification_for_import(pgdata_lsn);
     modification.init_empty()?;
 
     // Import all but pg_wal
@@ -309,7 +309,7 @@ async fn import_wal(
         waldecoder.feed_bytes(&buf);
 
         let mut nrecords = 0;
-        let mut modification = tline.begin_modification(last_lsn);
+        let mut modification = tline.begin_modification_for_import(last_lsn);
         while last_lsn <= endpoint {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 let interpreted = InterpretedWalRecord::from_bytes_filtered(
@@ -357,7 +357,7 @@ pub async fn import_basebackup_from_tar(
     ctx: &RequestContext,
 ) -> Result<()> {
     info!("importing base at {base_lsn}");
-    let mut modification = tline.begin_modification(base_lsn);
+    let mut modification = tline.begin_modification_for_import(base_lsn);
     modification.init_empty()?;
 
     let mut pg_control: Option<ControlFileData> = None;
@@ -457,7 +457,7 @@ pub async fn import_wal_from_tar(
 
         waldecoder.feed_bytes(&bytes[offset..]);
 
-        let mut modification = tline.begin_modification(last_lsn);
+        let mut modification = tline.begin_modification_for_import(last_lsn);
         while last_lsn <= end_lsn {
             if let Some((lsn, recdata)) = waldecoder.poll_decode()? {
                 let interpreted = InterpretedWalRecord::from_bytes_filtered(

pageserver/src/page_service.rs

@@ -68,6 +68,7 @@ use crate::config::PageServerConf;
 use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
 };
+use crate::feature_resolver::FeatureResolver;
 use crate::metrics::{
     self, COMPUTE_COMMANDS_COUNTERS, ComputeCommandKind, GetPageBatchBreakReason, LIVE_CONNECTIONS,
     MISROUTED_PAGESTREAM_REQUESTS, PAGESTREAM_HANDLER_RESULTS_TOTAL, SmgrOpTimer, TimelineMetrics,
@@ -139,6 +140,7 @@ pub fn spawn(
     perf_trace_dispatch: Option<Dispatch>,
     tcp_listener: tokio::net::TcpListener,
     tls_config: Option<Arc<rustls::ServerConfig>>,
+    feature_resolver: FeatureResolver,
 ) -> Listener {
     let cancel = CancellationToken::new();
     let libpq_ctx = RequestContext::todo_child(
@@ -160,6 +162,7 @@ pub fn spawn(
             conf.pg_auth_type,
             tls_config,
             conf.page_service_pipelining.clone(),
+            feature_resolver,
             libpq_ctx,
             cancel.clone(),
         )
@@ -218,6 +221,7 @@ pub async fn libpq_listener_main(
     auth_type: AuthType,
     tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
+    feature_resolver: FeatureResolver,
     listener_ctx: RequestContext,
     listener_cancel: CancellationToken,
 ) -> Connections {
@@ -261,6 +265,7 @@ pub async fn libpq_listener_main(
                     auth_type,
                     tls_config.clone(),
                     pipelining_config.clone(),
+                    feature_resolver.clone(),
                     connection_ctx,
                     connections_cancel.child_token(),
                     gate_guard,
@@ -303,6 +308,7 @@ async fn page_service_conn_main(
     auth_type: AuthType,
     tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
+    feature_resolver: FeatureResolver,
     connection_ctx: RequestContext,
     cancel: CancellationToken,
     gate_guard: GateGuard,
@@ -370,6 +376,7 @@ async fn page_service_conn_main(
         perf_span_fields,
         connection_ctx,
         cancel.clone(),
+        feature_resolver.clone(),
         gate_guard,
     );
     let pgbackend =
@@ -421,6 +428,8 @@ struct PageServerHandler {
     pipelining_config: PageServicePipeliningConfig,
     get_vectored_concurrent_io: GetVectoredConcurrentIo,
 
+    feature_resolver: FeatureResolver,
+
     gate_guard: GateGuard,
 }
 
@@ -535,6 +544,7 @@ impl timeline::handle::TenantManager<TenantManagerTypes> for TenantManagerWrappe
             match resolved {
                 ShardResolveResult::Found(tenant_shard) => break tenant_shard,
                 ShardResolveResult::NotFound => {
+                    MISROUTED_PAGESTREAM_REQUESTS.inc();
                     return Err(GetActiveTimelineError::Tenant(
                         GetActiveTenantError::NotFound(GetTenantError::NotFound(*tenant_id)),
                     ));
@@ -586,6 +596,15 @@ impl timeline::handle::TenantManager<TenantManagerTypes> for TenantManagerWrappe
     }
 }
 
+/// Whether to hold the applied GC cutoff guard when processing GetPage requests.
+/// This is determined once at the start of pagestream subprotocol handling based on
+/// feature flags, configuration, and test conditions.
+#[derive(Debug, Clone, Copy)]
+enum HoldAppliedGcCutoffGuard {
+    Yes,
+    No,
+}
+
 #[derive(thiserror::Error, Debug)]
 enum PageStreamError {
     /// We encountered an error that should prompt the client to reconnect:
@@ -729,6 +748,7 @@ enum BatchedFeMessage {
     GetPage {
         span: Span,
         shard: WeakHandle<TenantManagerTypes>,
+        applied_gc_cutoff_guard: Option<RcuReadGuard<Lsn>>,
         pages: SmallVec<[BatchedGetPageRequest; 1]>,
         batch_break_reason: GetPageBatchBreakReason,
     },
@@ -908,6 +928,7 @@ impl PageServerHandler {
         perf_span_fields: ConnectionPerfSpanFields,
         connection_ctx: RequestContext,
         cancel: CancellationToken,
+        feature_resolver: FeatureResolver,
         gate_guard: GateGuard,
     ) -> Self {
         PageServerHandler {
@@ -919,6 +940,7 @@ impl PageServerHandler {
             cancel,
             pipelining_config,
             get_vectored_concurrent_io,
+            feature_resolver,
             gate_guard,
         }
     }
@@ -958,6 +980,7 @@ impl PageServerHandler {
         ctx: &RequestContext,
         protocol_version: PagestreamProtocolVersion,
         parent_span: Span,
+        hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
     ) -> Result<Option<BatchedFeMessage>, QueryError>
     where
         IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
@@ -1195,19 +1218,27 @@ impl PageServerHandler {
                 })
                 .await?;
 
+                let applied_gc_cutoff_guard = shard.get_applied_gc_cutoff_lsn(); // hold guard
                 // We're holding the Handle
                 let effective_lsn = match Self::effective_request_lsn(
                     &shard,
                     shard.get_last_record_lsn(),
                     req.hdr.request_lsn,
                     req.hdr.not_modified_since,
-                    &shard.get_applied_gc_cutoff_lsn(),
+                    &applied_gc_cutoff_guard,
                 ) {
                     Ok(lsn) => lsn,
                     Err(e) => {
                         return respond_error!(span, e);
                     }
                 };
+                let applied_gc_cutoff_guard = match hold_gc_cutoff_guard {
+                    HoldAppliedGcCutoffGuard::Yes => Some(applied_gc_cutoff_guard),
+                    HoldAppliedGcCutoffGuard::No => {
+                        drop(applied_gc_cutoff_guard);
+                        None
+                    }
+                };
 
                 let batch_wait_ctx = if ctx.has_perf_span() {
                     Some(
@@ -1228,6 +1259,7 @@ impl PageServerHandler {
                 BatchedFeMessage::GetPage {
                     span,
                     shard: shard.downgrade(),
+                    applied_gc_cutoff_guard,
                     pages: smallvec![BatchedGetPageRequest {
                         req,
                         timer,
@@ -1328,13 +1360,28 @@ impl PageServerHandler {
                 match (eligible_batch, this_msg) {
                     (
                         BatchedFeMessage::GetPage {
-                            pages: accum_pages, ..
+                            pages: accum_pages,
+                            applied_gc_cutoff_guard: accum_applied_gc_cutoff_guard,
+                            ..
                         },
                         BatchedFeMessage::GetPage {
-                            pages: this_pages, ..
+                            pages: this_pages,
+                            applied_gc_cutoff_guard: this_applied_gc_cutoff_guard,
+                            ..
                         },
                     ) => {
                         accum_pages.extend(this_pages);
+                        // the minimum of the two guards will keep data for both alive
+                        match (&accum_applied_gc_cutoff_guard, this_applied_gc_cutoff_guard) {
+                            (None, None) => (),
+                            (None, Some(this)) => *accum_applied_gc_cutoff_guard = Some(this),
+                            (Some(_), None) => (),
+                            (Some(accum), Some(this)) => {
+                                if **accum > *this {
+                                    *accum_applied_gc_cutoff_guard = Some(this);
+                                }
+                            }
+                        };
                         Ok(())
                     }
                     #[cfg(feature = "testing")]
@@ -1649,6 +1696,7 @@ impl PageServerHandler {
             BatchedFeMessage::GetPage {
                 span,
                 shard,
+                applied_gc_cutoff_guard,
                 pages,
                 batch_break_reason,
             } => {
@@ -1668,6 +1716,7 @@ impl PageServerHandler {
                         .instrument(span.clone())
                         .await;
                         assert_eq!(res.len(), npages);
+                        drop(applied_gc_cutoff_guard);
                         res
                     },
                     span,
@@ -1749,7 +1798,7 @@ impl PageServerHandler {
     /// Coding discipline within this function: all interaction with the `pgb` connection
     /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
     /// This is so that we can shutdown page_service quickly.
-    #[instrument(skip_all)]
+    #[instrument(skip_all, fields(hold_gc_cutoff_guard))]
     async fn handle_pagerequests<IO>(
         &mut self,
         pgb: &mut PostgresBackend<IO>,
@@ -1795,6 +1844,30 @@ impl PageServerHandler {
             .take()
             .expect("implementation error: timeline_handles should not be locked");
 
+        // Evaluate the expensive feature resolver check once per pagestream subprotocol handling
+        // instead of once per GetPage request. This is shared between pipelined and serial paths.
+        let hold_gc_cutoff_guard = if cfg!(test) || cfg!(feature = "testing") {
+            HoldAppliedGcCutoffGuard::Yes
+        } else {
+            // Use the global feature resolver with the tenant ID directly, avoiding the need
+            // to get a timeline/shard which might not be available on this pageserver node.
+            let empty_properties = std::collections::HashMap::new();
+            match self.feature_resolver.evaluate_boolean(
+                "page-service-getpage-hold-applied-gc-cutoff-guard",
+                tenant_id,
+                &empty_properties,
+            ) {
+                Ok(()) => HoldAppliedGcCutoffGuard::Yes,
+                Err(_) => HoldAppliedGcCutoffGuard::No,
+            }
+        };
+        // record it in the span of handle_pagerequests so that both the request_span
+        // and the pipeline implementation spans contains the field.
+        Span::current().record(
+            "hold_gc_cutoff_guard",
+            tracing::field::debug(&hold_gc_cutoff_guard),
+        );
+
         let request_span = info_span!("request");
         let ((pgb_reader, timeline_handles), result) = match self.pipelining_config.clone() {
             PageServicePipeliningConfig::Pipelined(pipelining_config) => {
@@ -1808,6 +1881,7 @@ impl PageServerHandler {
                     pipelining_config,
                     protocol_version,
                     io_concurrency,
+                    hold_gc_cutoff_guard,
                     &ctx,
                 )
                 .await
@@ -1822,6 +1896,7 @@ impl PageServerHandler {
                     request_span,
                     protocol_version,
                     io_concurrency,
+                    hold_gc_cutoff_guard,
                     &ctx,
                 )
                 .await
@@ -1850,6 +1925,7 @@ impl PageServerHandler {
         request_span: Span,
         protocol_version: PagestreamProtocolVersion,
         io_concurrency: IoConcurrency,
+        hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
         ctx: &RequestContext,
     ) -> (
         (PostgresBackendReader<IO>, TimelineHandles),
@@ -1871,6 +1947,7 @@ impl PageServerHandler {
                 ctx,
                 protocol_version,
                 request_span.clone(),
+                hold_gc_cutoff_guard,
             )
             .await;
             let msg = match msg {
@@ -1918,6 +1995,7 @@ impl PageServerHandler {
         pipelining_config: PageServicePipeliningConfigPipelined,
         protocol_version: PagestreamProtocolVersion,
         io_concurrency: IoConcurrency,
+        hold_gc_cutoff_guard: HoldAppliedGcCutoffGuard,
         ctx: &RequestContext,
     ) -> (
         (PostgresBackendReader<IO>, TimelineHandles),
@@ -2021,6 +2099,7 @@ impl PageServerHandler {
                         &ctx,
                         protocol_version,
                         request_span.clone(),
+                        hold_gc_cutoff_guard,
                     )
                     .await;
                     let Some(read_res) = read_res.transpose() else {
@@ -2067,6 +2146,7 @@ impl PageServerHandler {
                         pages,
                         span: _,
                         shard: _,
+                        applied_gc_cutoff_guard: _,
                         batch_break_reason: _,
                     } = &mut batch
                     {
@@ -3428,8 +3508,6 @@ impl GrpcPageServiceHandler {
     /// NB: errors returned from here are intercepted in get_pages(), and may be converted to a
     /// GetPageResponse with an appropriate status code to avoid terminating the stream.
     ///
-    /// TODO: verify that the requested pages belong to this shard.
-    ///
     /// TODO: get_vectored() currently enforces a batch limit of 32. Postgres will typically send
     /// batches up to effective_io_concurrency = 100. Either we have to accept large batches, or
     /// split them up in the client or server.
@@ -3455,6 +3533,19 @@ impl GrpcPageServiceHandler {
             lsn = %req.read_lsn,
         );
 
+        for &blkno in &req.block_numbers {
+            let shard = timeline.get_shard_identity();
+            let key = rel_block_to_key(req.rel, blkno);
+            if !shard.is_key_local(&key) {
+                return Err(tonic::Status::invalid_argument(format!(
+                    "block {blkno} of relation {} requested on wrong shard {} (is on {})",
+                    req.rel,
+                    timeline.get_shard_index(),
+                    ShardIndex::new(shard.get_shard_number(&key), shard.count),
+                )));
+            }
+        }
+
         let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn(); // hold guard
         let effective_lsn = PageServerHandler::effective_request_lsn(
             &timeline,

pageserver/src/pgdatadir_mapping.rs

@@ -6,7 +6,7 @@
 //! walingest.rs handles a few things like implicit relation creation and extension.
 //! Clarify that)
 //!
-use std::collections::{HashMap, HashSet, hash_map};
+use std::collections::{BTreeSet, HashMap, HashSet, hash_map};
 use std::ops::{ControlFlow, Range};
 use std::sync::Arc;
 
@@ -227,6 +227,25 @@ impl Timeline {
             pending_nblocks: 0,
             pending_directory_entries: Vec::new(),
             pending_metadata_bytes: 0,
+            is_importing_pgdata: false,
+            lsn,
+        }
+    }
+
+    pub fn begin_modification_for_import(&self, lsn: Lsn) -> DatadirModification
+    where
+        Self: Sized,
+    {
+        DatadirModification {
+            tline: self,
+            pending_lsns: Vec::new(),
+            pending_metadata_pages: HashMap::new(),
+            pending_data_batch: None,
+            pending_deletions: Vec::new(),
+            pending_nblocks: 0,
+            pending_directory_entries: Vec::new(),
+            pending_metadata_bytes: 0,
+            is_importing_pgdata: true,
             lsn,
         }
     }
@@ -596,6 +615,50 @@ impl Timeline {
         self.get_rel_exists_in_reldir(tag, version, None, ctx).await
     }
 
+    async fn get_rel_exists_in_reldir_v1(
+        &self,
+        tag: RelTag,
+        version: Version<'_>,
+        deserialized_reldir_v1: Option<(Key, &RelDirectory)>,
+        ctx: &RequestContext,
+    ) -> Result<bool, PageReconstructError> {
+        let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
+        if let Some((cached_key, dir)) = deserialized_reldir_v1 {
+            if cached_key == key {
+                return Ok(dir.rels.contains(&(tag.relnode, tag.forknum)));
+            } else if cfg!(test) || cfg!(feature = "testing") {
+                panic!("cached reldir key mismatch: {cached_key} != {key}");
+            } else {
+                warn!("cached reldir key mismatch: {cached_key} != {key}");
+            }
+            // Fallback to reading the directory from the datadir.
+        }
+
+        let buf = version.get(self, key, ctx).await?;
+
+        let dir = RelDirectory::des(&buf)?;
+        Ok(dir.rels.contains(&(tag.relnode, tag.forknum)))
+    }
+
+    async fn get_rel_exists_in_reldir_v2(
+        &self,
+        tag: RelTag,
+        version: Version<'_>,
+        ctx: &RequestContext,
+    ) -> Result<bool, PageReconstructError> {
+        let key = rel_tag_sparse_key(tag.spcnode, tag.dbnode, tag.relnode, tag.forknum);
+        let buf = RelDirExists::decode_option(version.sparse_get(self, key, ctx).await?).map_err(
+            |_| {
+                PageReconstructError::Other(anyhow::anyhow!(
+                    "invalid reldir key: decode failed, {}",
+                    key
+                ))
+            },
+        )?;
+        let exists_v2 = buf == RelDirExists::Exists;
+        Ok(exists_v2)
+    }
+
     /// Does the relation exist? With a cached deserialized `RelDirectory`.
     ///
     /// There are some cases where the caller loops across all relations. In that specific case,
@@ -627,45 +690,134 @@ impl Timeline {
             return Ok(false);
         }
 
-        // Read path: first read the new reldir keyspace. Early return if the relation exists.
-        // Otherwise, read the old reldir keyspace.
-        // TODO: if IndexPart::rel_size_migration is `Migrated`, we only need to read from v2.
+        let (v2_status, migrated_lsn) = self.get_rel_size_v2_status();
 
-        if let RelSizeMigration::Migrated | RelSizeMigration::Migrating =
-            self.get_rel_size_v2_status()
-        {
-            // fetch directory listing (new)
-            let key = rel_tag_sparse_key(tag.spcnode, tag.dbnode, tag.relnode, tag.forknum);
-            let buf = RelDirExists::decode_option(version.sparse_get(self, key, ctx).await?)
-                .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
-            let exists_v2 = buf == RelDirExists::Exists;
-            // Fast path: if the relation exists in the new format, return true.
-            // TODO: we should have a verification mode that checks both keyspaces
-            // to ensure the relation only exists in one of them.
-            if exists_v2 {
-                return Ok(true);
+        match v2_status {
+            RelSizeMigration::Legacy => {
+                let v1_exists = self
+                    .get_rel_exists_in_reldir_v1(tag, version, deserialized_reldir_v1, ctx)
+                    .await?;
+                Ok(v1_exists)
+            }
+            RelSizeMigration::Migrating | RelSizeMigration::Migrated
+                if version.get_lsn() < migrated_lsn.unwrap_or(Lsn(0)) =>
+            {
+                // For requests below the migrated LSN, we still use the v1 read path.
+                let v1_exists = self
+                    .get_rel_exists_in_reldir_v1(tag, version, deserialized_reldir_v1, ctx)
+                    .await?;
+                Ok(v1_exists)
+            }
+            RelSizeMigration::Migrating => {
+                let v1_exists = self
+                    .get_rel_exists_in_reldir_v1(tag, version, deserialized_reldir_v1, ctx)
+                    .await?;
+                let v2_exists_res = self.get_rel_exists_in_reldir_v2(tag, version, ctx).await;
+                match v2_exists_res {
+                    Ok(v2_exists) if v1_exists == v2_exists => {}
+                    Ok(v2_exists) => {
+                        tracing::warn!(
+                            "inconsistent v1/v2 reldir keyspace for rel {}: v1_exists={}, v2_exists={}",
+                            tag,
+                            v1_exists,
+                            v2_exists
+                        );
+                    }
+                    Err(e) => {
+                        tracing::warn!("failed to get rel exists in v2: {e}");
+                    }
+                }
+                Ok(v1_exists)
+            }
+            RelSizeMigration::Migrated => {
+                let v2_exists = self.get_rel_exists_in_reldir_v2(tag, version, ctx).await?;
+                Ok(v2_exists)
             }
         }
+    }
 
-        // fetch directory listing (old)
-
-        let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
-
-        if let Some((cached_key, dir)) = deserialized_reldir_v1 {
-            if cached_key == key {
-                return Ok(dir.rels.contains(&(tag.relnode, tag.forknum)));
-            } else if cfg!(test) || cfg!(feature = "testing") {
-                panic!("cached reldir key mismatch: {cached_key} != {key}");
-            } else {
-                warn!("cached reldir key mismatch: {cached_key} != {key}");
-            }
-            // Fallback to reading the directory from the datadir.
-        }
+    async fn list_rels_v1(
+        &self,
+        spcnode: Oid,
+        dbnode: Oid,
+        version: Version<'_>,
+        ctx: &RequestContext,
+    ) -> Result<HashSet<RelTag>, PageReconstructError> {
+        let key = rel_dir_to_key(spcnode, dbnode);
         let buf = version.get(self, key, ctx).await?;
-
         let dir = RelDirectory::des(&buf)?;
-        let exists_v1 = dir.rels.contains(&(tag.relnode, tag.forknum));
-        Ok(exists_v1)
+        let rels_v1: HashSet<RelTag> =
+            HashSet::from_iter(dir.rels.iter().map(|(relnode, forknum)| RelTag {
+                spcnode,
+                dbnode,
+                relnode: *relnode,
+                forknum: *forknum,
+            }));
+        Ok(rels_v1)
+    }
+
+    async fn list_rels_v2(
+        &self,
+        spcnode: Oid,
+        dbnode: Oid,
+        version: Version<'_>,
+        ctx: &RequestContext,
+    ) -> Result<HashSet<RelTag>, PageReconstructError> {
+        let key_range = rel_tag_sparse_key_range(spcnode, dbnode);
+        let io_concurrency = IoConcurrency::spawn_from_conf(
+            self.conf.get_vectored_concurrent_io,
+            self.gate
+                .enter()
+                .map_err(|_| PageReconstructError::Cancelled)?,
+        );
+        let results = self
+            .scan(
+                KeySpace::single(key_range),
+                version.get_lsn(),
+                ctx,
+                io_concurrency,
+            )
+            .await?;
+        let mut rels = HashSet::new();
+        for (key, val) in results {
+            let val = RelDirExists::decode(&val?).map_err(|_| {
+                PageReconstructError::Other(anyhow::anyhow!(
+                    "invalid reldir key: decode failed, {}",
+                    key
+                ))
+            })?;
+            if key.field6 != 1 {
+                return Err(PageReconstructError::Other(anyhow::anyhow!(
+                    "invalid reldir key: field6 != 1, {}",
+                    key
+                )));
+            }
+            if key.field2 != spcnode {
+                return Err(PageReconstructError::Other(anyhow::anyhow!(
+                    "invalid reldir key: field2 != spcnode, {}",
+                    key
+                )));
+            }
+            if key.field3 != dbnode {
+                return Err(PageReconstructError::Other(anyhow::anyhow!(
+                    "invalid reldir key: field3 != dbnode, {}",
+                    key
+                )));
+            }
+            let tag = RelTag {
+                spcnode,
+                dbnode,
+                relnode: key.field4,
+                forknum: key.field5,
+            };
+            if val == RelDirExists::Removed {
+                debug_assert!(!rels.contains(&tag), "removed reltag in v2");
+                continue;
+            }
+            let did_not_contain = rels.insert(tag);
+            debug_assert!(did_not_contain, "duplicate reltag in v2");
+        }
+        Ok(rels)
     }
 
     /// Get a list of all existing relations in given tablespace and database.
@@ -683,60 +835,45 @@ impl Timeline {
         version: Version<'_>,
         ctx: &RequestContext,
     ) -> Result<HashSet<RelTag>, PageReconstructError> {
-        // fetch directory listing (old)
-        let key = rel_dir_to_key(spcnode, dbnode);
-        let buf = version.get(self, key, ctx).await?;
+        let (v2_status, migrated_lsn) = self.get_rel_size_v2_status();
 
-        let dir = RelDirectory::des(&buf)?;
-        let rels_v1: HashSet<RelTag> =
-            HashSet::from_iter(dir.rels.iter().map(|(relnode, forknum)| RelTag {
-                spcnode,
-                dbnode,
-                relnode: *relnode,
-                forknum: *forknum,
-            }));
-
-        if let RelSizeMigration::Legacy = self.get_rel_size_v2_status() {
-            return Ok(rels_v1);
-        }
-
-        // scan directory listing (new), merge with the old results
-        let key_range = rel_tag_sparse_key_range(spcnode, dbnode);
-        let io_concurrency = IoConcurrency::spawn_from_conf(
-            self.conf.get_vectored_concurrent_io,
-            self.gate
-                .enter()
-                .map_err(|_| PageReconstructError::Cancelled)?,
-        );
-        let results = self
-            .scan(
-                KeySpace::single(key_range),
-                version.get_lsn(),
-                ctx,
-                io_concurrency,
-            )
-            .await?;
-        let mut rels = rels_v1;
-        for (key, val) in results {
-            let val = RelDirExists::decode(&val?)
-                .map_err(|_| PageReconstructError::Other(anyhow::anyhow!("invalid reldir key")))?;
-            assert_eq!(key.field6, 1);
-            assert_eq!(key.field2, spcnode);
-            assert_eq!(key.field3, dbnode);
-            let tag = RelTag {
-                spcnode,
-                dbnode,
-                relnode: key.field4,
-                forknum: key.field5,
-            };
-            if val == RelDirExists::Removed {
-                debug_assert!(!rels.contains(&tag), "removed reltag in v2");
-                continue;
+        match v2_status {
+            RelSizeMigration::Legacy => {
+                let rels_v1 = self.list_rels_v1(spcnode, dbnode, version, ctx).await?;
+                Ok(rels_v1)
+            }
+            RelSizeMigration::Migrating | RelSizeMigration::Migrated
+                if version.get_lsn() < migrated_lsn.unwrap_or(Lsn(0)) =>
+            {
+                // For requests below the migrated LSN, we still use the v1 read path.
+                let rels_v1 = self.list_rels_v1(spcnode, dbnode, version, ctx).await?;
+                Ok(rels_v1)
+            }
+            RelSizeMigration::Migrating => {
+                let rels_v1 = self.list_rels_v1(spcnode, dbnode, version, ctx).await?;
+                let rels_v2_res = self.list_rels_v2(spcnode, dbnode, version, ctx).await;
+                match rels_v2_res {
+                    Ok(rels_v2) if rels_v1 == rels_v2 => {}
+                    Ok(rels_v2) => {
+                        tracing::warn!(
+                            "inconsistent v1/v2 reldir keyspace for db {} {}: v1_rels.len()={}, v2_rels.len()={}",
+                            spcnode,
+                            dbnode,
+                            rels_v1.len(),
+                            rels_v2.len()
+                        );
+                    }
+                    Err(e) => {
+                        tracing::warn!("failed to list rels in v2: {e}");
+                    }
+                }
+                Ok(rels_v1)
+            }
+            RelSizeMigration::Migrated => {
+                let rels_v2 = self.list_rels_v2(spcnode, dbnode, version, ctx).await?;
+                Ok(rels_v2)
             }
-            let did_not_contain = rels.insert(tag);
-            debug_assert!(did_not_contain, "duplicate reltag in v2");
         }
-        Ok(rels)
     }
 
     /// Get the whole SLRU segment
@@ -1258,10 +1395,10 @@ impl Timeline {
         let mut dbdir_cnt = 0;
         let mut rel_cnt = 0;
 
-        for (spcnode, dbnode) in dbdir.dbdirs.keys() {
+        for &(spcnode, dbnode) in dbdir.dbdirs.keys() {
             dbdir_cnt += 1;
             for rel in self
-                .list_rels(*spcnode, *dbnode, Version::at(lsn), ctx)
+                .list_rels(spcnode, dbnode, Version::at(lsn), ctx)
                 .await?
             {
                 rel_cnt += 1;
@@ -1566,6 +1703,9 @@ pub struct DatadirModification<'a> {
 
     /// An **approximation** of how many metadata bytes will be written to the EphemeralFile.
     pending_metadata_bytes: usize,
+
+    /// Whether we are importing a pgdata directory.
+    is_importing_pgdata: bool,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -1578,6 +1718,14 @@ pub enum MetricsUpdate {
     Sub(u64),
 }
 
+/// Controls the behavior of the reldir keyspace.
+pub struct RelDirMode {
+    // Whether we can read the v2 keyspace or not.
+    current_status: RelSizeMigration,
+    // Whether we should initialize the v2 keyspace or not.
+    initialize: bool,
+}
+
 impl DatadirModification<'_> {
     // When a DatadirModification is committed, we do a monolithic serialization of all its contents.  WAL records can
     // contain multiple pages, so the pageserver's record-based batch size isn't sufficient to bound this allocation: we
@@ -1933,30 +2081,49 @@ impl DatadirModification<'_> {
     }
 
     /// Returns `true` if the rel_size_v2 write path is enabled. If it is the first time that
-    /// we enable it, we also need to persist it in `index_part.json`.
-    pub fn maybe_enable_rel_size_v2(&mut self) -> anyhow::Result<bool> {
-        let status = self.tline.get_rel_size_v2_status();
+    /// we enable it, we also need to persist it in `index_part.json` (initialize is true).
+    ///
+    /// As this function is only used on the write path, we do not need to read the migrated_at
+    /// field.
+    pub fn maybe_enable_rel_size_v2(&mut self, is_create: bool) -> anyhow::Result<RelDirMode> {
+        // TODO: define the behavior of the tenant-level config flag and use feature flag to enable this feature
+
+        let (status, _) = self.tline.get_rel_size_v2_status();
         let config = self.tline.get_rel_size_v2_enabled();
         match (config, status) {
             (false, RelSizeMigration::Legacy) => {
                 // tenant config didn't enable it and we didn't write any reldir_v2 key yet
-                Ok(false)
+                Ok(RelDirMode {
+                    current_status: RelSizeMigration::Legacy,
+                    initialize: false,
+                })
             }
-            (false, RelSizeMigration::Migrating | RelSizeMigration::Migrated) => {
+            (false, status @ RelSizeMigration::Migrating | status @ RelSizeMigration::Migrated) => {
                 // index_part already persisted that the timeline has enabled rel_size_v2
-                Ok(true)
+                Ok(RelDirMode {
+                    current_status: status,
+                    initialize: false,
+                })
             }
             (true, RelSizeMigration::Legacy) => {
                 // The first time we enable it, we need to persist it in `index_part.json`
-                self.tline
-                    .update_rel_size_v2_status(RelSizeMigration::Migrating)?;
-                tracing::info!("enabled rel_size_v2");
-                Ok(true)
+                // The caller should update the reldir status once the initialization is done.
+                //
+                // Only initialize the v2 keyspace on new relation creation. No initialization
+                // during `timeline_create` (TODO: fix this, we should allow, but currently it
+                // hits consistency issues).
+                Ok(RelDirMode {
+                    current_status: RelSizeMigration::Legacy,
+                    initialize: is_create && !self.is_importing_pgdata,
+                })
             }
-            (true, RelSizeMigration::Migrating | RelSizeMigration::Migrated) => {
+            (true, status @ RelSizeMigration::Migrating | status @ RelSizeMigration::Migrated) => {
                 // index_part already persisted that the timeline has enabled rel_size_v2
                 // and we don't need to do anything
-                Ok(true)
+                Ok(RelDirMode {
+                    current_status: status,
+                    initialize: false,
+                })
             }
         }
     }
@@ -1969,8 +2136,8 @@ impl DatadirModification<'_> {
         img: Bytes,
         ctx: &RequestContext,
     ) -> Result<(), WalIngestError> {
-        let v2_enabled = self
-            .maybe_enable_rel_size_v2()
+        let v2_mode = self
+            .maybe_enable_rel_size_v2(false)
             .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
 
         // Add it to the directory (if it doesn't exist already)
@@ -1986,17 +2153,19 @@ impl DatadirModification<'_> {
             self.put(DBDIR_KEY, Value::Image(buf.into()));
         }
         if r.is_none() {
-            // Create RelDirectory
-            // TODO: if we have fully migrated to v2, no need to create this directory
+            if v2_mode.current_status != RelSizeMigration::Legacy {
+                self.pending_directory_entries
+                    .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
+            }
+
+            // Create RelDirectory in v1 keyspace. TODO: if we have fully migrated to v2, no need to create this directory.
+            // Some code path relies on this directory to be present. We should remove it once we starts to set tenants to
+            // `RelSizeMigration::Migrated` state (currently we don't, all tenants will have `RelSizeMigration::Migrating`).
             let buf = RelDirectory::ser(&RelDirectory {
                 rels: HashSet::new(),
             })?;
             self.pending_directory_entries
                 .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
-            if v2_enabled {
-                self.pending_directory_entries
-                    .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
-            }
             self.put(
                 rel_dir_to_key(spcnode, dbnode),
                 Value::Image(Bytes::from(buf)),
@@ -2103,6 +2272,109 @@ impl DatadirModification<'_> {
         Ok(())
     }
 
+    async fn initialize_rel_size_v2_keyspace(
+        &mut self,
+        ctx: &RequestContext,
+        dbdir: &DbDirectory,
+    ) -> Result<(), WalIngestError> {
+        // Copy everything from relv1 to relv2; TODO: check if there's any key in the v2 keyspace, if so, abort.
+        tracing::info!("initializing rel_size_v2 keyspace");
+        let mut rel_cnt = 0;
+        // relmap_exists (the value of dbdirs hashmap) does not affect the migration: we need to copy things over anyways
+        for &(spcnode, dbnode) in dbdir.dbdirs.keys() {
+            let rel_dir_key = rel_dir_to_key(spcnode, dbnode);
+            let rel_dir = RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?;
+            for (relnode, forknum) in rel_dir.rels {
+                let sparse_rel_dir_key = rel_tag_sparse_key(spcnode, dbnode, relnode, forknum);
+                self.put(
+                    sparse_rel_dir_key,
+                    Value::Image(RelDirExists::Exists.encode()),
+                );
+                tracing::info!(
+                    "migrated rel_size_v2: {}",
+                    RelTag {
+                        spcnode,
+                        dbnode,
+                        relnode,
+                        forknum
+                    }
+                );
+                rel_cnt += 1;
+            }
+        }
+        tracing::info!(
+            "initialized rel_size_v2 keyspace at lsn {}: migrated {} relations",
+            self.lsn,
+            rel_cnt
+        );
+        self.tline
+            .update_rel_size_v2_status(RelSizeMigration::Migrating, Some(self.lsn))
+            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
+        Ok::<_, WalIngestError>(())
+    }
+
+    async fn put_rel_creation_v1(
+        &mut self,
+        rel: RelTag,
+        dbdir_exists: bool,
+        ctx: &RequestContext,
+    ) -> Result<(), WalIngestError> {
+        // Reldir v1 write path
+        let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
+        let mut rel_dir = if !dbdir_exists {
+            // Create the RelDirectory
+            RelDirectory::default()
+        } else {
+            // reldir already exists, fetch it
+            RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?
+        };
+
+        // Add the new relation to the rel directory entry, and write it back
+        if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
+            Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
+        }
+        if !dbdir_exists {
+            self.pending_directory_entries
+                .push((DirectoryKind::Rel, MetricsUpdate::Set(0)))
+        }
+        self.pending_directory_entries
+            .push((DirectoryKind::Rel, MetricsUpdate::Add(1)));
+        self.put(
+            rel_dir_key,
+            Value::Image(Bytes::from(RelDirectory::ser(&rel_dir)?)),
+        );
+        Ok(())
+    }
+
+    async fn put_rel_creation_v2(
+        &mut self,
+        rel: RelTag,
+        dbdir_exists: bool,
+        ctx: &RequestContext,
+    ) -> Result<(), WalIngestError> {
+        // Reldir v2 write path
+        let sparse_rel_dir_key =
+            rel_tag_sparse_key(rel.spcnode, rel.dbnode, rel.relnode, rel.forknum);
+        // check if the rel_dir_key exists in v2
+        let val = self.sparse_get(sparse_rel_dir_key, ctx).await?;
+        let val = RelDirExists::decode_option(val)
+            .map_err(|_| WalIngestErrorKind::InvalidRelDirKey(sparse_rel_dir_key))?;
+        if val == RelDirExists::Exists {
+            Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
+        }
+        self.put(
+            sparse_rel_dir_key,
+            Value::Image(RelDirExists::Exists.encode()),
+        );
+        if !dbdir_exists {
+            self.pending_directory_entries
+                .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
+        }
+        self.pending_directory_entries
+            .push((DirectoryKind::RelV2, MetricsUpdate::Add(1)));
+        Ok(())
+    }
+
     /// Create a relation fork.
     ///
     /// 'nblocks' is the initial size.
@@ -2136,66 +2408,31 @@ impl DatadirModification<'_> {
                 true
             };
 
-        let rel_dir_key = rel_dir_to_key(rel.spcnode, rel.dbnode);
-        let mut rel_dir = if !dbdir_exists {
-            // Create the RelDirectory
-            RelDirectory::default()
-        } else {
-            // reldir already exists, fetch it
-            RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?
-        };
-
-        let v2_enabled = self
-            .maybe_enable_rel_size_v2()
+        let mut v2_mode = self
+            .maybe_enable_rel_size_v2(true)
             .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
 
-        if v2_enabled {
-            if rel_dir.rels.contains(&(rel.relnode, rel.forknum)) {
-                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
+        if v2_mode.initialize {
+            if let Err(e) = self.initialize_rel_size_v2_keyspace(ctx, &dbdir).await {
+                tracing::warn!("error initializing rel_size_v2 keyspace: {}", e);
+                // TODO: circuit breaker so that it won't retry forever
+            } else {
+                v2_mode.current_status = RelSizeMigration::Migrating;
             }
-            let sparse_rel_dir_key =
-                rel_tag_sparse_key(rel.spcnode, rel.dbnode, rel.relnode, rel.forknum);
-            // check if the rel_dir_key exists in v2
-            let val = self.sparse_get(sparse_rel_dir_key, ctx).await?;
-            let val = RelDirExists::decode_option(val)
-                .map_err(|_| WalIngestErrorKind::InvalidRelDirKey(sparse_rel_dir_key))?;
-            if val == RelDirExists::Exists {
-                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
+        }
+
+        if v2_mode.current_status != RelSizeMigration::Migrated {
+            self.put_rel_creation_v1(rel, dbdir_exists, ctx).await?;
+        }
+
+        if v2_mode.current_status != RelSizeMigration::Legacy {
+            let write_v2_res = self.put_rel_creation_v2(rel, dbdir_exists, ctx).await;
+            if let Err(e) = write_v2_res {
+                if v2_mode.current_status == RelSizeMigration::Migrated {
+                    return Err(e);
+                }
+                tracing::warn!("error writing rel_size_v2 keyspace: {}", e);
             }
-            self.put(
-                sparse_rel_dir_key,
-                Value::Image(RelDirExists::Exists.encode()),
-            );
-            if !dbdir_exists {
-                self.pending_directory_entries
-                    .push((DirectoryKind::Rel, MetricsUpdate::Set(0)));
-                self.pending_directory_entries
-                    .push((DirectoryKind::RelV2, MetricsUpdate::Set(0)));
-                // We don't write `rel_dir_key -> rel_dir.rels` back to the storage in the v2 path unless it's the initial creation.
-                // TODO: if we have fully migrated to v2, no need to create this directory. Otherwise, there
-                // will be key not found errors if we don't create an empty one for rel_size_v2.
-                self.put(
-                    rel_dir_key,
-                    Value::Image(Bytes::from(RelDirectory::ser(&RelDirectory::default())?)),
-                );
-            }
-            self.pending_directory_entries
-                .push((DirectoryKind::RelV2, MetricsUpdate::Add(1)));
-        } else {
-            // Add the new relation to the rel directory entry, and write it back
-            if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
-                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
-            }
-            if !dbdir_exists {
-                self.pending_directory_entries
-                    .push((DirectoryKind::Rel, MetricsUpdate::Set(0)))
-            }
-            self.pending_directory_entries
-                .push((DirectoryKind::Rel, MetricsUpdate::Add(1)));
-            self.put(
-                rel_dir_key,
-                Value::Image(Bytes::from(RelDirectory::ser(&rel_dir)?)),
-            );
         }
 
         // Put size
@@ -2270,15 +2507,12 @@ impl DatadirModification<'_> {
         Ok(())
     }
 
-    /// Drop some relations
-    pub(crate) async fn put_rel_drops(
+    async fn put_rel_drop_v1(
         &mut self,
         drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
         ctx: &RequestContext,
-    ) -> Result<(), WalIngestError> {
-        let v2_enabled = self
-            .maybe_enable_rel_size_v2()
-            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
+    ) -> Result<BTreeSet<RelTag>, WalIngestError> {
+        let mut dropped_rels = BTreeSet::new();
         for ((spc_node, db_node), rel_tags) in drop_relations {
             let dir_key = rel_dir_to_key(spc_node, db_node);
             let buf = self.get(dir_key, ctx).await?;
@@ -2290,25 +2524,8 @@ impl DatadirModification<'_> {
                     self.pending_directory_entries
                         .push((DirectoryKind::Rel, MetricsUpdate::Sub(1)));
                     dirty = true;
+                    dropped_rels.insert(rel_tag);
                     true
-                } else if v2_enabled {
-                    // The rel is not found in the old reldir key, so we need to check the new sparse keyspace.
-                    // Note that a relation can only exist in one of the two keyspaces (guaranteed by the ingestion
-                    // logic).
-                    let key =
-                        rel_tag_sparse_key(spc_node, db_node, rel_tag.relnode, rel_tag.forknum);
-                    let val = RelDirExists::decode_option(self.sparse_get(key, ctx).await?)
-                        .map_err(|_| WalIngestErrorKind::InvalidKey(key, self.lsn))?;
-                    if val == RelDirExists::Exists {
-                        self.pending_directory_entries
-                            .push((DirectoryKind::RelV2, MetricsUpdate::Sub(1)));
-                        // put tombstone
-                        self.put(key, Value::Image(RelDirExists::Removed.encode()));
-                        // no need to set dirty to true
-                        true
-                    } else {
-                        false
-                    }
                 } else {
                     false
                 };
@@ -2331,7 +2548,67 @@ impl DatadirModification<'_> {
                 self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
             }
         }
+        Ok(dropped_rels)
+    }
 
+    async fn put_rel_drop_v2(
+        &mut self,
+        drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
+        ctx: &RequestContext,
+    ) -> Result<BTreeSet<RelTag>, WalIngestError> {
+        let mut dropped_rels = BTreeSet::new();
+        for ((spc_node, db_node), rel_tags) in drop_relations {
+            for rel_tag in rel_tags {
+                let key = rel_tag_sparse_key(spc_node, db_node, rel_tag.relnode, rel_tag.forknum);
+                let val = RelDirExists::decode_option(self.sparse_get(key, ctx).await?)
+                    .map_err(|_| WalIngestErrorKind::InvalidKey(key, self.lsn))?;
+                if val == RelDirExists::Exists {
+                    dropped_rels.insert(rel_tag);
+                    self.pending_directory_entries
+                        .push((DirectoryKind::RelV2, MetricsUpdate::Sub(1)));
+                    // put tombstone
+                    self.put(key, Value::Image(RelDirExists::Removed.encode()));
+                }
+            }
+        }
+        Ok(dropped_rels)
+    }
+
+    /// Drop some relations
+    pub(crate) async fn put_rel_drops(
+        &mut self,
+        drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
+        ctx: &RequestContext,
+    ) -> Result<(), WalIngestError> {
+        let v2_mode = self
+            .maybe_enable_rel_size_v2(false)
+            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
+        match v2_mode.current_status {
+            RelSizeMigration::Legacy => {
+                self.put_rel_drop_v1(drop_relations, ctx).await?;
+            }
+            RelSizeMigration::Migrating => {
+                let dropped_rels_v1 = self.put_rel_drop_v1(drop_relations.clone(), ctx).await?;
+                let dropped_rels_v2_res = self.put_rel_drop_v2(drop_relations, ctx).await;
+                match dropped_rels_v2_res {
+                    Ok(dropped_rels_v2) => {
+                        if dropped_rels_v1 != dropped_rels_v2 {
+                            tracing::warn!(
+                                "inconsistent v1/v2 rel drop: dropped_rels_v1.len()={}, dropped_rels_v2.len()={}",
+                                dropped_rels_v1.len(),
+                                dropped_rels_v2.len()
+                            );
+                        }
+                    }
+                    Err(e) => {
+                        tracing::warn!("error dropping rels: {}", e);
+                    }
+                }
+            }
+            RelSizeMigration::Migrated => {
+                self.put_rel_drop_v2(drop_relations, ctx).await?;
+            }
+        }
         Ok(())
     }
 

pageserver/src/tenant.rs

@@ -1205,6 +1205,7 @@ impl TenantShard {
             idempotency.clone(),
             index_part.gc_compaction.clone(),
             index_part.rel_size_migration.clone(),
+            index_part.rel_size_migrated_at,
             ctx,
         )?;
         let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
@@ -2584,6 +2585,7 @@ impl TenantShard {
             initdb_lsn,
             None,
             None,
+            None,
             ctx,
         )
         .await
@@ -2913,6 +2915,7 @@ impl TenantShard {
                     initdb_lsn,
                     None,
                     None,
+                    None,
                     ctx,
                 )
                 .await
@@ -4342,6 +4345,7 @@ impl TenantShard {
         create_idempotency: CreateTimelineIdempotency,
         gc_compaction_state: Option<GcCompactionState>,
         rel_size_v2_status: Option<RelSizeMigration>,
+        rel_size_migrated_at: Option<Lsn>,
         ctx: &RequestContext,
     ) -> anyhow::Result<(Arc<Timeline>, RequestContext)> {
         let state = match cause {
@@ -4376,6 +4380,7 @@ impl TenantShard {
             create_idempotency,
             gc_compaction_state,
             rel_size_v2_status,
+            rel_size_migrated_at,
             self.cancel.child_token(),
         );
 
@@ -5085,6 +5090,7 @@ impl TenantShard {
             src_timeline.pg_version,
         );
 
+        let (rel_size_v2_status, rel_size_migrated_at) = src_timeline.get_rel_size_v2_status();
         let (uninitialized_timeline, _timeline_ctx) = self
             .prepare_new_timeline(
                 dst_id,
@@ -5092,7 +5098,8 @@ impl TenantShard {
                 timeline_create_guard,
                 start_lsn + 1,
                 Some(Arc::clone(src_timeline)),
-                Some(src_timeline.get_rel_size_v2_status()),
+                Some(rel_size_v2_status),
+                rel_size_migrated_at,
                 ctx,
             )
             .await?;
@@ -5379,6 +5386,7 @@ impl TenantShard {
                 pgdata_lsn,
                 None,
                 None,
+                None,
                 ctx,
             )
             .await?;
@@ -5462,14 +5470,17 @@ impl TenantShard {
         start_lsn: Lsn,
         ancestor: Option<Arc<Timeline>>,
         rel_size_v2_status: Option<RelSizeMigration>,
+        rel_size_migrated_at: Option<Lsn>,
         ctx: &RequestContext,
     ) -> anyhow::Result<(UninitializedTimeline<'a>, RequestContext)> {
         let tenant_shard_id = self.tenant_shard_id;
 
         let resources = self.build_timeline_resources(new_timeline_id);
-        resources
-            .remote_client
-            .init_upload_queue_for_empty_remote(new_metadata, rel_size_v2_status.clone())?;
+        resources.remote_client.init_upload_queue_for_empty_remote(
+            new_metadata,
+            rel_size_v2_status.clone(),
+            rel_size_migrated_at,
+        )?;
 
         let (timeline_struct, timeline_ctx) = self
             .create_timeline_struct(
@@ -5482,6 +5493,7 @@ impl TenantShard {
                 create_guard.idempotency.clone(),
                 None,
                 rel_size_v2_status,
+                rel_size_migrated_at,
                 ctx,
             )
             .context("Failed to create timeline data structure")?;

pageserver/src/tenant/remote_timeline_client.rs

@@ -443,7 +443,8 @@ impl RemoteTimelineClient {
     pub fn init_upload_queue_for_empty_remote(
         &self,
         local_metadata: &TimelineMetadata,
-        rel_size_v2_status: Option<RelSizeMigration>,
+        rel_size_v2_migration: Option<RelSizeMigration>,
+        rel_size_migrated_at: Option<Lsn>,
     ) -> anyhow::Result<()> {
         // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
         // certainly no point in starting more upload tasks than this.
@@ -455,7 +456,8 @@ impl RemoteTimelineClient {
         let mut upload_queue = self.upload_queue.lock().unwrap();
         let initialized_queue =
             upload_queue.initialize_empty_remote(local_metadata, inprogress_limit)?;
-        initialized_queue.dirty.rel_size_migration = rel_size_v2_status;
+        initialized_queue.dirty.rel_size_migration = rel_size_v2_migration;
+        initialized_queue.dirty.rel_size_migrated_at = rel_size_migrated_at;
         self.update_remote_physical_size_gauge(None);
         info!("initialized upload queue as empty");
         Ok(())
@@ -994,10 +996,12 @@ impl RemoteTimelineClient {
     pub(crate) fn schedule_index_upload_for_rel_size_v2_status_update(
         self: &Arc<Self>,
         rel_size_v2_status: RelSizeMigration,
+        rel_size_migrated_at: Option<Lsn>,
     ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
         upload_queue.dirty.rel_size_migration = Some(rel_size_v2_status);
+        upload_queue.dirty.rel_size_migrated_at = rel_size_migrated_at;
         // TODO: allow this operation to bypass the validation check because we might upload the index part
         // with no layers but the flag updated. For now, we just modify the index part in memory and the next
         // upload will include the flag.

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -114,6 +114,11 @@ pub struct IndexPart {
     /// The timestamp when the timeline was marked invisible in synthetic size calculations.
     #[serde(skip_serializing_if = "Option::is_none", default)]
     pub(crate) marked_invisible_at: Option<NaiveDateTime>,
+
+    /// The LSN at which we started the rel size migration. Accesses below this LSN should be
+    /// processed with the v1 read path. Usually this LSN should be set together with `rel_size_migration`.
+    #[serde(skip_serializing_if = "Option::is_none", default)]
+    pub(crate) rel_size_migrated_at: Option<Lsn>,
 }
 
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
@@ -142,10 +147,12 @@ impl IndexPart {
     /// - 12: +l2_lsn
     /// - 13: +gc_compaction
     /// - 14: +marked_invisible_at
-    const LATEST_VERSION: usize = 14;
+    /// - 15: +rel_size_migrated_at
+    const LATEST_VERSION: usize = 15;
 
     // Versions we may see when reading from a bucket.
-    pub const KNOWN_VERSIONS: &'static [usize] = &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14];
+    pub const KNOWN_VERSIONS: &'static [usize] =
+        &[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15];
 
     pub const FILE_NAME: &'static str = "index_part.json";
 
@@ -165,6 +172,7 @@ impl IndexPart {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         }
     }
 
@@ -475,6 +483,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -524,6 +533,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -574,6 +584,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -627,6 +638,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let empty_layers_parsed = IndexPart::from_json_bytes(empty_layers_json.as_bytes()).unwrap();
@@ -675,6 +687,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -726,6 +739,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -782,6 +796,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -843,6 +858,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -905,6 +921,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -972,6 +989,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1052,6 +1070,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1133,6 +1152,7 @@ mod tests {
             l2_lsn: None,
             gc_compaction: None,
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1220,6 +1240,7 @@ mod tests {
                 last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             }),
             marked_invisible_at: None,
+            rel_size_migrated_at: None,
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
@@ -1308,6 +1329,97 @@ mod tests {
                 last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             }),
             marked_invisible_at: Some(parse_naive_datetime("2023-07-31T09:00:00.123000000")),
+            rel_size_migrated_at: None,
+        };
+
+        let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();
+        assert_eq!(part, expected);
+    }
+
+    #[test]
+    fn v15_rel_size_migrated_at_is_parsed() {
+        let example = r#"{
+            "version": 15,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata": {
+                "disk_consistent_lsn": "0/16960E8",
+                "prev_record_lsn": "0/1696070",
+                "ancestor_timeline": "e45a7f37d3ee2ff17dc14bf4f4e3f52e",
+                "ancestor_lsn": "0/0",
+                "latest_gc_cutoff_lsn": "0/1696070",
+                "initdb_lsn": "0/1696070",
+                "pg_version": 14
+            },
+            "gc_blocking": {
+                "started_at": "2024-07-19T09:00:00.123",
+                "reasons": ["DetachAncestor"]
+            },
+            "import_pgdata": {
+                "V1": {
+                    "Done": {
+                        "idempotency_key": "specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5",
+                        "started_at": "2024-11-13T09:23:42.123",
+                        "finished_at": "2024-11-13T09:42:23.123"
+                    }
+                }
+            },
+            "rel_size_migration": "legacy",
+            "l2_lsn": "0/16960E8",
+            "gc_compaction": {
+                "last_completed_lsn": "0/16960E8"
+            },
+            "marked_invisible_at": "2023-07-31T09:00:00.123",
+            "rel_size_migrated_at": "0/16960E8"
+        }"#;
+
+        let expected = IndexPart {
+            version: 15,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), LayerFileMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), LayerFileMetadata {
+                    file_size: 9007199254741001,
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::new(
+                Lsn::from_str("0/16960E8").unwrap(),
+                Some(Lsn::from_str("0/1696070").unwrap()),
+                Some(TimelineId::from_str("e45a7f37d3ee2ff17dc14bf4f4e3f52e").unwrap()),
+                Lsn::INVALID,
+                Lsn::from_str("0/1696070").unwrap(),
+                Lsn::from_str("0/1696070").unwrap(),
+                PgMajorVersion::PG14,
+            ).with_recalculated_checksum().unwrap(),
+            deleted_at: None,
+            lineage: Default::default(),
+            gc_blocking: Some(GcBlocking {
+                started_at: parse_naive_datetime("2024-07-19T09:00:00.123000000"),
+                reasons: enumset::EnumSet::from_iter([GcBlockingReason::DetachAncestor]),
+            }),
+            last_aux_file_policy: Default::default(),
+            archived_at: None,
+            import_pgdata: Some(import_pgdata::index_part_format::Root::V1(import_pgdata::index_part_format::V1::Done(import_pgdata::index_part_format::Done{
+                started_at: parse_naive_datetime("2024-11-13T09:23:42.123000000"),
+                finished_at: parse_naive_datetime("2024-11-13T09:42:23.123000000"),
+                idempotency_key: import_pgdata::index_part_format::IdempotencyKey::new("specified-by-client-218a5213-5044-4562-a28d-d024c5f057f5".to_string()),
+            }))),
+            rel_size_migration: Some(RelSizeMigration::Legacy),
+            l2_lsn: Some("0/16960E8".parse::<Lsn>().unwrap()),
+            gc_compaction: Some(GcCompactionState {
+                last_completed_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            }),
+            marked_invisible_at: Some(parse_naive_datetime("2023-07-31T09:00:00.123000000")),
+            rel_size_migrated_at: Some("0/16960E8".parse::<Lsn>().unwrap()),
         };
 
         let part = IndexPart::from_json_bytes(example.as_bytes()).unwrap();

pageserver/src/tenant/timeline.rs

@@ -70,7 +70,7 @@ use tracing::*;
 use utils::generation::Generation;
 use utils::guard_arc_swap::GuardArcSwap;
 use utils::id::TimelineId;
-use utils::logging::{MonitorSlowFutureCallback, monitor_slow_future};
+use utils::logging::{MonitorSlowFutureCallback, log_slow, monitor_slow_future};
 use utils::lsn::{AtomicLsn, Lsn, RecordLsn};
 use utils::postgres_client::PostgresClientProtocol;
 use utils::rate_limit::RateLimit;
@@ -441,7 +441,7 @@ pub struct Timeline {
     /// heatmap on demand.
     heatmap_layers_downloader: Mutex<Option<heatmap_layers_downloader::HeatmapLayersDownloader>>,
 
-    pub(crate) rel_size_v2_status: ArcSwapOption<RelSizeMigration>,
+    pub(crate) rel_size_v2_status: ArcSwap<(Option<RelSizeMigration>, Option<Lsn>)>,
 
     wait_lsn_log_slow: tokio::sync::Semaphore,
 
@@ -2894,12 +2894,9 @@ impl Timeline {
             .unwrap_or(self.conf.default_tenant_conf.rel_size_v2_enabled)
     }
 
-    pub(crate) fn get_rel_size_v2_status(&self) -> RelSizeMigration {
-        self.rel_size_v2_status
-            .load()
-            .as_ref()
-            .map(|s| s.as_ref().clone())
-            .unwrap_or(RelSizeMigration::Legacy)
+    pub(crate) fn get_rel_size_v2_status(&self) -> (RelSizeMigration, Option<Lsn>) {
+        let (status, migrated_at) = self.rel_size_v2_status.load().as_ref().clone();
+        (status.unwrap_or(RelSizeMigration::Legacy), migrated_at)
     }
 
     fn get_compaction_upper_limit(&self) -> usize {
@@ -3174,6 +3171,7 @@ impl Timeline {
         create_idempotency: crate::tenant::CreateTimelineIdempotency,
         gc_compaction_state: Option<GcCompactionState>,
         rel_size_v2_status: Option<RelSizeMigration>,
+        rel_size_migrated_at: Option<Lsn>,
         cancel: CancellationToken,
     ) -> Arc<Self> {
         let disk_consistent_lsn = metadata.disk_consistent_lsn();
@@ -3338,7 +3336,10 @@ impl Timeline {
 
                 heatmap_layers_downloader: Mutex::new(None),
 
-                rel_size_v2_status: ArcSwapOption::from_pointee(rel_size_v2_status),
+                rel_size_v2_status: ArcSwap::from_pointee((
+                    rel_size_v2_status,
+                    rel_size_migrated_at,
+                )),
 
                 wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
 
@@ -3426,11 +3427,17 @@ impl Timeline {
     pub(crate) fn update_rel_size_v2_status(
         &self,
         rel_size_v2_status: RelSizeMigration,
+        rel_size_migrated_at: Option<Lsn>,
     ) -> anyhow::Result<()> {
-        self.rel_size_v2_status
-            .store(Some(Arc::new(rel_size_v2_status.clone())));
+        self.rel_size_v2_status.store(Arc::new((
+            Some(rel_size_v2_status.clone()),
+            rel_size_migrated_at,
+        )));
         self.remote_client
-            .schedule_index_upload_for_rel_size_v2_status_update(rel_size_v2_status)
+            .schedule_index_upload_for_rel_size_v2_status_update(
+                rel_size_v2_status,
+                rel_size_migrated_at,
+            )
     }
 
     pub(crate) fn get_gc_compaction_state(&self) -> Option<GcCompactionState> {
@@ -6891,7 +6898,13 @@ impl Timeline {
 
             write_guard.store_and_unlock(new_gc_cutoff)
         };
-        waitlist.wait().await;
+        let waitlist_wait_fut = std::pin::pin!(waitlist.wait());
+        log_slow(
+            "applied_gc_cutoff waitlist wait",
+            Duration::from_secs(30),
+            waitlist_wait_fut,
+        )
+        .await;
 
         info!("GC starting");
 

pageserver/src/tenant/timeline/delete.rs

@@ -332,6 +332,7 @@ impl DeleteTimelineFlow {
                 crate::tenant::CreateTimelineIdempotency::FailWithConflict, // doesn't matter what we put here
                 None, // doesn't matter what we put here
                 None, // doesn't matter what we put here
+                None, // doesn't matter what we put here
                 ctx,
             )
             .context("create_timeline_struct")?;

pageserver/src/utilization.rs

@@ -52,7 +52,7 @@ pub(crate) fn regenerate(
     };
 
     // Express a static value for how many shards we may schedule on one node
-    const MAX_SHARDS: u32 = 5000;
+    const MAX_SHARDS: u32 = 2500;
 
     let mut doc = PageserverUtilization {
         disk_usage_bytes: used,

pgxn/neon/Makefile

@@ -33,6 +33,10 @@ SHLIB_LINK = -lcurl
 UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S), Darwin)
     SHLIB_LINK += -framework Security -framework CoreFoundation -framework SystemConfiguration
+
+    # Link against object files for the current macOS version, to avoid spurious linker warnings.
+    MACOSX_DEPLOYMENT_TARGET := $(shell xcrun --sdk macosx --show-sdk-version)
+    export MACOSX_DEPLOYMENT_TARGET
 endif
 
 EXTENSION = neon

pgxn/neon/communicator.c

@@ -79,10 +79,6 @@
 #include "access/xlogrecovery.h"
 #endif
 
-#if PG_VERSION_NUM < 160000
-typedef PGAlignedBlock PGIOAlignedBlock;
-#endif
-
 #define NEON_PANIC_CONNECTION_STATE(shard_no, elvl, message, ...) \
 	neon_shard_log(shard_no, elvl, "Broken connection state: " message, \
 				   ##__VA_ARGS__)

pgxn/neon/extension_server.c

@@ -14,7 +14,7 @@
 #include "extension_server.h"
 #include "neon_utils.h"
 
-static int	extension_server_port = 0;
+int	hadron_extension_server_port = 0;
 static int	extension_server_request_timeout = 60;
 static int	extension_server_connect_timeout = 60;
 
@@ -47,7 +47,7 @@ neon_download_extension_file_http(const char *filename, bool is_library)
 		curl_easy_setopt(handle, CURLOPT_CONNECTTIMEOUT, (long)extension_server_connect_timeout /* seconds */ );
 
 	compute_ctl_url = psprintf("http://localhost:%d/extension_server/%s%s",
-							   extension_server_port, filename, is_library ? "?is_library=true" : "");
+							   hadron_extension_server_port, filename, is_library ? "?is_library=true" : "");
 
 	elog(LOG, "Sending request to compute_ctl: %s", compute_ctl_url);
 
@@ -82,7 +82,7 @@ pg_init_extension_server()
 	DefineCustomIntVariable("neon.extension_server_port",
 							"connection string to the compute_ctl",
 							NULL,
-							&extension_server_port,
+							&hadron_extension_server_port,
 							0, 0, INT_MAX,
 							PGC_POSTMASTER,
 							0,	/* no flags required */

pgxn/neon/file_cache.c

@@ -635,6 +635,11 @@ lfc_init(void)
 							NULL);
 }
 
+/*
+ * Dump a list of pages that are currently in the LFC
+ *
+ * This is used to get a snapshot that can be used to prewarm the LFC later.
+ */
 FileCacheState*
 lfc_get_state(size_t max_entries)
 {
@@ -2267,4 +2272,3 @@ get_prewarm_info(PG_FUNCTION_ARGS)
 
 	PG_RETURN_DATUM(HeapTupleGetDatum(heap_form_tuple(tupdesc, values, nulls)));
 }
-

pgxn/neon/libpagestore.c

@@ -13,6 +13,8 @@
 #include <math.h>
 #include <sys/socket.h>
 
+#include <curl/curl.h>
+
 #include "libpq-int.h"
 
 #include "access/xlog.h"
@@ -86,6 +88,10 @@ static int pageserver_response_log_timeout = 10000;
 /* 2.5 minutes. A bit higher than highest default TCP retransmission timeout */
 static int pageserver_response_disconnect_timeout = 150000;
 
+static int	conf_refresh_reconnect_attempt_threshold = 16;
+// Hadron: timeout for refresh errors (1 minute)
+static uint64 	kRefreshErrorTimeoutUSec = 1 * USECS_PER_MINUTE;
+
 typedef struct
 {
 	char		connstring[MAX_SHARDS][MAX_PAGESERVER_CONNSTRING_SIZE];
@@ -130,7 +136,7 @@ static uint64 pagestore_local_counter = 0;
 typedef enum PSConnectionState {
 	PS_Disconnected,			/* no connection yet */
 	PS_Connecting_Startup,		/* connection starting up */
-	PS_Connecting_PageStream,	/* negotiating pagestream */ 
+	PS_Connecting_PageStream,	/* negotiating pagestream */
 	PS_Connected,				/* connected, pagestream established */
 } PSConnectionState;
 
@@ -178,6 +184,8 @@ static PageServer page_servers[MAX_SHARDS];
 static bool pageserver_flush(shardno_t shard_no);
 static void pageserver_disconnect(shardno_t shard_no);
 static void pageserver_disconnect_shard(shardno_t shard_no);
+// HADRON
+shardno_t get_num_shards(void);
 
 static bool
 PagestoreShmemIsValid(void)
@@ -286,6 +294,22 @@ AssignPageserverConnstring(const char *newval, void *extra)
 	}
 }
 
+/* BEGIN_HADRON */
+/**
+ * Return the total number of shards seen in the shard map.
+ */
+shardno_t get_num_shards(void)
+{
+	const ShardMap *shard_map;
+
+	Assert(pagestore_shared);
+	shard_map = &pagestore_shared->shard_map;
+
+	Assert(shard_map != NULL);
+	return shard_map->num_shards;
+}
+/* END_HADRON */
+
 /*
  * Get the current number of shards, and/or the connection string for a
  * particular shard from the shard map in shared memory.
@@ -383,7 +407,7 @@ get_shard_number(BufferTag *tag)
 }
 
 static inline void
-CLEANUP_AND_DISCONNECT(PageServer *shard) 
+CLEANUP_AND_DISCONNECT(PageServer *shard)
 {
 	if (shard->wes_read)
 	{
@@ -405,7 +429,7 @@ CLEANUP_AND_DISCONNECT(PageServer *shard)
  * complete the connection (e.g. due to receiving an earlier cancellation
  * during connection start).
  * Returns true if successfully connected; false if the connection failed.
- * 
+ *
  * Throws errors in unrecoverable situations, or when this backend's query
  * is canceled.
  */
@@ -1012,6 +1036,101 @@ pageserver_disconnect_shard(shardno_t shard_no)
 	shard->state = PS_Disconnected;
 }
 
+// BEGIN HADRON
+/*
+ * Nudge compute_ctl to refresh our configuration. Called when we suspect we may be
+ * connecting to the wrong pageservers due to a stale configuration.
+ *
+ * This is a best-effort operation. If we couldn't send the local loopback HTTP request
+ * to compute_ctl or if the request fails for any reason, we just log the error and move
+ * on.
+ */
+
+extern int hadron_extension_server_port;
+
+// The timestamp (usec) of the first error that occurred while trying to refresh the configuration.
+// Will be reset to 0 after a successful refresh.
+static uint64 first_recorded_refresh_error_usec = 0;
+
+// Request compute_ctl to refresh the configuration. This operation may fail, e.g., if the compute_ctl
+// is already in the configuration state. The function returns true if the caller needs to cancel the
+// current query to avoid dead/live lock.
+static bool
+hadron_request_configuration_refresh() {
+	static CURL	   *handle = NULL;
+	CURLcode	res;
+	char	   *compute_ctl_url;
+	bool cancel_query = false;
+
+	if (!lakebase_mode)
+		return false;
+
+	if (handle == NULL)
+	{
+		handle = alloc_curl_handle();
+
+		curl_easy_setopt(handle, CURLOPT_CUSTOMREQUEST, "POST");
+		curl_easy_setopt(handle, CURLOPT_TIMEOUT, 3L /* seconds */ );
+		curl_easy_setopt(handle, CURLOPT_POSTFIELDS, "");
+	}
+
+	// Set the URL
+	compute_ctl_url = psprintf("http://localhost:%d/refresh_configuration", hadron_extension_server_port);
+
+
+	elog(LOG, "Sending refresh configuration request to compute_ctl: %s", compute_ctl_url);
+
+	curl_easy_setopt(handle, CURLOPT_URL, compute_ctl_url);
+
+	res = curl_easy_perform(handle);
+	if (res != CURLE_OK )
+	{
+		elog(WARNING, "refresh_configuration request failed: %s\n", curl_easy_strerror(res));
+	}
+	else
+	{
+		long http_code = 0;
+		curl_easy_getinfo(handle, CURLINFO_RESPONSE_CODE, &http_code);
+		if ( res != CURLE_OK )
+		{
+			elog(WARNING, "compute_ctl refresh_configuration request getinfo failed: %s\n", curl_easy_strerror(res));
+		}
+		else
+		{
+			elog(LOG, "compute_ctl refresh_configuration got HTTP response: %ld\n", http_code);
+			if( http_code == 200 )
+			{
+				first_recorded_refresh_error_usec = 0;
+			}
+			else
+			{
+				if (first_recorded_refresh_error_usec == 0)
+				{
+					first_recorded_refresh_error_usec = GetCurrentTimestamp();
+				}
+				else if(GetCurrentTimestamp() - first_recorded_refresh_error_usec > kRefreshErrorTimeoutUSec)
+				{
+					{
+						first_recorded_refresh_error_usec = 0;
+						cancel_query = true;
+					}
+				}
+			}
+		}
+	}
+
+	// In regular Postgres usage, it is not necessary to manually free memory allocated by palloc (psprintf) because
+	// it will be cleaned up after the "memory context" is reset (e.g. after the query or the transaction is finished).
+	// However, the number of times this function gets called during a single query/transaction can be unbounded due to
+	// the various retry loops around calls to pageservers. Therefore, we need to manually free this memory here.
+	if (compute_ctl_url != NULL)
+	{
+		pfree(compute_ctl_url);
+	}
+	return cancel_query;
+}
+// END HADRON
+
 static bool
 pageserver_send(shardno_t shard_no, NeonRequest *request)
 {
@@ -1046,6 +1165,11 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 		while (!pageserver_connect(shard_no, shard->n_reconnect_attempts < max_reconnect_attempts ? LOG : ERROR))
 		{
 			shard->n_reconnect_attempts += 1;
+			if (shard->n_reconnect_attempts > conf_refresh_reconnect_attempt_threshold
+				&& hadron_request_configuration_refresh() )
+			{
+				neon_shard_log(shard_no, ERROR, "request failed too many times, cancelling query");
+			}
 		}
 		shard->n_reconnect_attempts = 0;
 	} else {
@@ -1153,17 +1277,26 @@ pageserver_receive(shardno_t shard_no)
 		pfree(msg);
 		pageserver_disconnect(shard_no);
 		resp = NULL;
+
+		/*
+		 * Always poke compute_ctl to request a configuration refresh if we have issues receiving data from pageservers after
+		 * successfully connecting to it. It could be an indication that we are connecting to the wrong pageservers (e.g. PS
+		 * is in secondary mode or otherwise refuses to respond our request).
+		 */
+		hadron_request_configuration_refresh();
 	}
 	else if (rc == -2)
 	{
 		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
 
 		pageserver_disconnect(shard_no);
+		hadron_request_configuration_refresh();
 		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: could not read COPY data: %s", msg);
 	}
 	else
 	{
 		pageserver_disconnect(shard_no);
+		hadron_request_configuration_refresh();
 		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc);
 	}
 
@@ -1231,21 +1364,34 @@ pageserver_try_receive(shardno_t shard_no)
 		neon_shard_log(shard_no, LOG, "pageserver_receive disconnect: psql end of copy data: %s", pchomp(PQerrorMessage(pageserver_conn)));
 		pageserver_disconnect(shard_no);
 		resp = NULL;
+		hadron_request_configuration_refresh();
 	}
 	else if (rc == -2)
 	{
 		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
 
 		pageserver_disconnect(shard_no);
+		hadron_request_configuration_refresh();
 		neon_shard_log(shard_no, LOG, "pageserver_receive disconnect: could not read COPY data: %s", msg);
 		resp = NULL;
 	}
 	else
 	{
 		pageserver_disconnect(shard_no);
+		hadron_request_configuration_refresh();
 		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc);
 	}
 
+	/*
+	 * Always poke compute_ctl to request a configuration refresh if we have issues receiving data from pageservers after
+	 * successfully connecting to it. It could be an indication that we are connecting to the wrong pageservers (e.g. PS
+	 * is in secondary mode or otherwise refuses to respond our request).
+	 */
+	if ( rc < 0 && hadron_request_configuration_refresh() )
+	{
+		neon_shard_log(shard_no, ERROR, "refresh_configuration request failed, cancelling query");
+	}
+
 	shard->nresponses_received++;
 	return (NeonResponse *) resp;
 }
@@ -1442,6 +1588,16 @@ pg_init_libpagestore(void)
 							PGC_SU_BACKEND,
 							0,	/* no flags required */
 							NULL, NULL, NULL);
+	DefineCustomIntVariable("hadron.conf_refresh_reconnect_attempt_threshold",
+							"Threshold of the number of consecutive failed pageserver "
+							"connection attempts (per shard) before signaling "
+							"compute_ctl for a configuration refresh.",
+							NULL,
+							&conf_refresh_reconnect_attempt_threshold,
+							16, 0, INT_MAX,
+							PGC_USERSET,
+							0,
+							NULL, NULL, NULL);
 
 	DefineCustomIntVariable("neon.pageserver_response_log_timeout",
 							"pageserver response log timeout",

pgxn/neon/neon.c

@@ -1,7 +1,7 @@
 /*-------------------------------------------------------------------------
  *
  * neon.c
- *	  Main entry point into the neon exension
+ *	  Main entry point into the neon extension
  *
  *-------------------------------------------------------------------------
  */
@@ -48,6 +48,7 @@
 PG_MODULE_MAGIC;
 void		_PG_init(void);
 
+bool lakebase_mode = false;
 
 static int  running_xacts_overflow_policy;
 static bool monitor_query_exec_time = false;
@@ -507,7 +508,7 @@ _PG_init(void)
 
 	DefineCustomBoolVariable(
 							"neon.disable_logical_replication_subscribers",
-							"Disables incomming logical replication",
+							"Disable incoming logical replication",
 							NULL,
 							&disable_logical_replication_subscribers,
 							false,
@@ -566,7 +567,7 @@ _PG_init(void)
 
 	DefineCustomEnumVariable(
 							"neon.debug_compare_local",
-							"Debug mode for compaing content of pages in prefetch ring/LFC/PS and local disk",
+							"Debug mode for comparing content of pages in prefetch ring/LFC/PS and local disk",
 							NULL,
 							&debug_compare_local,
 							DEBUG_COMPARE_LOCAL_NONE,
@@ -583,6 +584,16 @@ _PG_init(void)
 							"neon_superuser",
 							PGC_POSTMASTER, 0, NULL, NULL, NULL);
 
+	DefineCustomBoolVariable(
+							"neon.lakebase_mode",
+							"Is neon running in Lakebase?",
+							NULL,
+							&lakebase_mode,
+							false,
+							PGC_POSTMASTER,
+							0,
+							NULL, NULL, NULL);
+
 	/*
 	 * Important: This must happen after other parts of the extension are
 	 * loaded, otherwise any settings to GUCs that were set before the
@@ -724,7 +735,6 @@ neon_shmem_request_hook(void)
 static void
 neon_shmem_startup_hook(void)
 {
-	/* Initialize */
 	if (prev_shmem_startup_hook)
 		prev_shmem_startup_hook();
 

pgxn/neon/neon.h

@@ -21,6 +21,7 @@ extern int	wal_acceptor_reconnect_timeout;
 extern int	wal_acceptor_connection_timeout;
 extern int	readahead_getpage_pull_timeout_ms;
 extern bool	disable_wal_prev_lsn_checks;
+extern bool	lakebase_mode;
 
 extern bool AmPrewarmWorker;
 

pgxn/neon/neon_perf_counters.h

@@ -167,11 +167,7 @@ extern neon_per_backend_counters *neon_per_backend_counters_shared;
  */
 #define NUM_NEON_PERF_COUNTER_SLOTS (MaxBackends + NUM_AUXILIARY_PROCS)
 
-#if PG_VERSION_NUM >= 170000
 #define MyNeonCounters (&neon_per_backend_counters_shared[MyProcNumber])
-#else
-#define MyNeonCounters (&neon_per_backend_counters_shared[MyProc->pgprocno])
-#endif
 
 extern void inc_getpage_wait(uint64 latency);
 extern void inc_page_cache_read_wait(uint64 latency);

pgxn/neon/neon_pgversioncompat.h

@@ -9,6 +9,10 @@
 #include "fmgr.h"
 #include "storage/buf_internals.h"
 
+#if PG_MAJORVERSION_NUM < 16
+typedef PGAlignedBlock PGIOAlignedBlock;
+#endif
+
 #if PG_MAJORVERSION_NUM < 17
 #define NRelFileInfoBackendIsTemp(rinfo) (rinfo.backend != InvalidBackendId)
 #else
@@ -158,6 +162,10 @@ InitBufferTag(BufferTag *tag, const RelFileNode *rnode,
 #define AmAutoVacuumWorkerProcess() (IsAutoVacuumWorkerProcess())
 #endif
 
+#if PG_MAJORVERSION_NUM < 17
+#define	MyProcNumber (MyProc - &ProcGlobal->allProcs[0])
+#endif
+
 #if PG_MAJORVERSION_NUM < 15
 extern void InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags);
 extern TimeLineID GetWALInsertionTimeLine(void);

pgxn/neon/pagestore_smgr.c

@@ -72,10 +72,6 @@
 #include "access/xlogrecovery.h"
 #endif
 
-#if PG_VERSION_NUM < 160000
-typedef PGAlignedBlock PGIOAlignedBlock;
-#endif
-
 #include "access/nbtree.h"
 #include "storage/bufpage.h"
 #include "access/xlog_internal.h"

pgxn/neon/relsize_cache.c

@@ -13,6 +13,7 @@
 #include "neon.h"
 #include "neon_pgversioncompat.h"
 
+#include "miscadmin.h"
 #include "pagestore_client.h"
 #include RELFILEINFO_HDR
 #include "storage/smgr.h"
@@ -23,10 +24,6 @@
 #include "utils/dynahash.h"
 #include "utils/guc.h"
 
-#if PG_VERSION_NUM >= 150000
-#include "miscadmin.h"
-#endif
-
 typedef struct
 {
 	NRelFileInfo rinfo;

pgxn/neon/walproposer.h

@@ -389,12 +389,21 @@ typedef struct PageserverFeedback
  */
 typedef struct WalRateLimiter
 {
-	/* If the value is 1, PG backends will hit backpressure. */
+	/* The effective wal write rate. Could be changed dynamically
+	based on whether PG has backpressure or not.*/
+	pg_atomic_uint32 effective_max_wal_bytes_per_second;
+	/* If the value is 1, PG backends will hit backpressure until the time has past batch_end_time_us. */
 	pg_atomic_uint32 should_limit;
 	/* The number of bytes sent in the current second. */
 	uint64		sent_bytes;
-	/* The last recorded time in microsecond. */
-	pg_atomic_uint64 last_recorded_time_us;
+	/* The timestamp when the write starts in the current batch. A batch is a time interval (e.g., )that we 
+	track and throttle writes. Most times a batch is 1s, but it could become larger if the PG overwrites the WALs
+	and we will adjust the batch accordingly to compensate (e.g., if PG writes 10MB at once and max WAL write rate
+	is 1MB/s, then the current batch will become 10s). */
+	pg_atomic_uint64 batch_start_time_us;
+	/* The timestamp (in the future) that the current batch should end and accept more writes
+	(after should_limit is set to 1). */
+	pg_atomic_uint64 batch_end_time_us;
 } WalRateLimiter;
 /* END_HADRON */
 

pgxn/neon/walproposer_pg.c

@@ -68,6 +68,14 @@ int			safekeeper_proto_version = 3;
 char	   *safekeeper_conninfo_options = "";
 /* BEGIN_HADRON */
 int         databricks_max_wal_mb_per_second = -1;
+// during throttling, we will limit the effective WAL write rate to 10KB.
+// PG can still push some WAL to SK, but at a very low rate.
+int 		databricks_throttled_max_wal_bytes_per_second = 10 * 1024;
+// The max sleep time of a batch. This is to make sure the rate limiter does not
+// overshoot too much and block PG for a very long time.
+// This is set as 5 minuetes for now. PG can send as much as 10MB of WALs to SK in one batch,
+// so this effectively caps the write rate to ~30KB/s in the worst case.
+static uint64 kRateLimitMaxBatchUSecs = 300 * USECS_PER_SEC;
 /* END_HADRON */
 
 /* Set to true in the walproposer bgw. */
@@ -86,6 +94,7 @@ static HotStandbyFeedback agg_hs_feedback;
 static void nwp_register_gucs(void);
 static void assign_neon_safekeepers(const char *newval, void *extra);
 static uint64 backpressure_lag_impl(void);
+static uint64 hadron_backpressure_lag_impl(void);
 static uint64 startup_backpressure_wrap(void);
 static bool backpressure_throttling_impl(void);
 static void walprop_register_bgworker(void);
@@ -110,6 +119,22 @@ static void rm_safekeeper_event_set(Safekeeper *to_remove, bool is_sk);
 
 static void CheckGracefulShutdown(WalProposer *wp);
 
+/* BEGIN_HADRON */
+shardno_t get_num_shards(void);
+
+static int positive_mb_to_bytes(int mb)
+{
+	if (mb <= 0)
+	{
+		return mb;
+	}
+	else
+	{
+		return mb * 1024 * 1024;
+	}
+}
+/* END_HADRON */
+
 static void
 init_walprop_config(bool syncSafekeepers)
 {
@@ -257,6 +282,16 @@ nwp_register_gucs(void)
                             PGC_SUSET,
                             GUC_UNIT_MB,
                             NULL, NULL, NULL);
+
+	DefineCustomIntVariable(
+							"databricks.throttled_max_wal_bytes_per_second",
+							"The maximum WAL bytes per second when PG is being throttled.",
+							NULL,
+							&databricks_throttled_max_wal_bytes_per_second,
+							10 * 1024, 0, INT_MAX,
+							PGC_SUSET,
+							GUC_UNIT_BYTE,
+							NULL, NULL, NULL);
     /* END_HADRON */
 }
 
@@ -395,19 +430,65 @@ assign_neon_safekeepers(const char *newval, void *extra)
 	pfree(oldval);
 }
 
-/* Check if we need to suspend inserts because of lagging replication. */
-static uint64
-backpressure_lag_impl(void)
+/* BEGIN_HADRON */
+static uint64 hadron_backpressure_lag_impl(void)
 {
 	struct WalproposerShmemState* state = NULL;
+	uint64 lag = 0;
 
-	/* BEGIN_HADRON */
 	if(max_cluster_size < 0){
 		// if max cluster size is not set, then we don't apply backpressure because we're reconfiguring PG
 		return 0;
 	}
-	/* END_HADRON */
 
+	lag = backpressure_lag_impl();
+	state = GetWalpropShmemState();
+	if ( state != NULL && databricks_max_wal_mb_per_second != -1 )
+	{
+		int old_limit = pg_atomic_read_u32(&state->wal_rate_limiter.effective_max_wal_bytes_per_second);
+		int new_limit = (lag == 0)? positive_mb_to_bytes(databricks_max_wal_mb_per_second) : databricks_throttled_max_wal_bytes_per_second;
+		if( old_limit != new_limit )
+		{
+			uint64 batch_start_time = pg_atomic_read_u64(&state->wal_rate_limiter.batch_start_time_us);
+			uint64 batch_end_time = pg_atomic_read_u64(&state->wal_rate_limiter.batch_end_time_us);
+			// the rate limit has changed, we need to reset the rate limiter's batch end time
+			pg_atomic_write_u32(&state->wal_rate_limiter.effective_max_wal_bytes_per_second, new_limit);
+			pg_atomic_write_u64(&state->wal_rate_limiter.batch_end_time_us, Min(batch_start_time + USECS_PER_SEC, batch_end_time));
+		}
+		if( new_limit == -1 )
+		{
+			return 0;
+		}
+
+		if (pg_atomic_read_u32(&state->wal_rate_limiter.should_limit) == true)
+		{
+			TimestampTz now = GetCurrentTimestamp();
+			struct WalRateLimiter *limiter = &state->wal_rate_limiter;
+			uint64 batch_end_time = pg_atomic_read_u64(&limiter->batch_end_time_us);
+			if ( now >= batch_end_time )
+			{
+				/*
+				* The backend has past the batch end time and it's time to push more WALs.
+				* If the backends are pushing WALs too fast, the wal proposer will rate limit them again.
+				*/
+				uint32 expected = true;
+				pg_atomic_compare_exchange_u32(&state->wal_rate_limiter.should_limit, &expected, false);
+				return 0;
+			}
+			return Max(lag, 1);
+		}
+		// rate limiter decides to not throttle, then return 0.
+		return 0;
+	}
+
+	return lag;
+}
+/* END_HADRON */
+
+/* Check if we need to suspend inserts because of lagging replication. */
+static uint64
+backpressure_lag_impl(void)
+{
 	if (max_replication_apply_lag > 0 || max_replication_flush_lag > 0 || max_replication_write_lag > 0)
 	{
 		XLogRecPtr	writePtr;
@@ -426,45 +507,47 @@ backpressure_lag_impl(void)
 			 LSN_FORMAT_ARGS(flushPtr),
 			 LSN_FORMAT_ARGS(applyPtr));
 
-		if ((writePtr != InvalidXLogRecPtr && max_replication_write_lag > 0 && myFlushLsn > writePtr + max_replication_write_lag * MB))
+		if (lakebase_mode)
 		{
-			return (myFlushLsn - writePtr - max_replication_write_lag * MB);
-		}
+			// in case PG does not have shard map initialized, we assume PG always has 1 shard at minimum.
+			shardno_t num_shards = Max(1, get_num_shards());
+			int tenant_max_replication_apply_lag = num_shards * max_replication_apply_lag;
+			int tenant_max_replication_flush_lag = num_shards * max_replication_flush_lag;
+			int tenant_max_replication_write_lag = num_shards * max_replication_write_lag;
 
-		if ((flushPtr != InvalidXLogRecPtr && max_replication_flush_lag > 0 && myFlushLsn > flushPtr + max_replication_flush_lag * MB))
-		{
-			return (myFlushLsn - flushPtr - max_replication_flush_lag * MB);
-		}
+			if ((writePtr != InvalidXLogRecPtr && tenant_max_replication_write_lag > 0 && myFlushLsn > writePtr + tenant_max_replication_write_lag * MB))
+			{
+				return (myFlushLsn - writePtr - tenant_max_replication_write_lag * MB);
+			}
 
-		if ((applyPtr != InvalidXLogRecPtr && max_replication_apply_lag > 0 && myFlushLsn > applyPtr + max_replication_apply_lag * MB))
+			if ((flushPtr != InvalidXLogRecPtr && tenant_max_replication_flush_lag > 0 && myFlushLsn > flushPtr + tenant_max_replication_flush_lag * MB))
+			{
+				return (myFlushLsn - flushPtr - tenant_max_replication_flush_lag * MB);
+			}
+
+			if ((applyPtr != InvalidXLogRecPtr && tenant_max_replication_apply_lag > 0 && myFlushLsn > applyPtr + tenant_max_replication_apply_lag * MB))
+			{
+				return (myFlushLsn - applyPtr - tenant_max_replication_apply_lag * MB);
+			}
+		}
+		else
 		{
-			return (myFlushLsn - applyPtr - max_replication_apply_lag * MB);
+			if ((writePtr != InvalidXLogRecPtr && max_replication_write_lag > 0 && myFlushLsn > writePtr + max_replication_write_lag * MB))
+			{
+				return (myFlushLsn - writePtr - max_replication_write_lag * MB);
+			}
+
+			if ((flushPtr != InvalidXLogRecPtr && max_replication_flush_lag > 0 && myFlushLsn > flushPtr + max_replication_flush_lag * MB))
+			{
+				return (myFlushLsn - flushPtr - max_replication_flush_lag * MB);
+			}
+
+			if ((applyPtr != InvalidXLogRecPtr && max_replication_apply_lag > 0 && myFlushLsn > applyPtr + max_replication_apply_lag * MB))
+			{
+				return (myFlushLsn - applyPtr - max_replication_apply_lag * MB);
+			}
 		}
 	}
-
-	/* BEGIN_HADRON */
-	if (databricks_max_wal_mb_per_second == -1) {
-		return 0;
-	}
-
-	state = GetWalpropShmemState();
-	if (state != NULL && !!pg_atomic_read_u32(&state->wal_rate_limiter.should_limit))
-	{
-		TimestampTz now = GetCurrentTimestamp();
-		struct WalRateLimiter *limiter = &state->wal_rate_limiter;
-		uint64 last_recorded_time = pg_atomic_read_u64(&limiter->last_recorded_time_us);
-		if (now - last_recorded_time > USECS_PER_SEC)
-		{
-			/*
-			 * The backend has past 1 second since the last recorded time and it's time to push more WALs.
-			 * If the backends are pushing WALs too fast, the wal proposer will rate limit them again.
-			 */
-			uint32 expected = true;
-			pg_atomic_compare_exchange_u32(&state->wal_rate_limiter.should_limit, &expected, false);
-		}
-		return 1;
-	}
-	/* END_HADRON */
 	return 0;
 }
 
@@ -479,9 +562,9 @@ startup_backpressure_wrap(void)
 	if (AmStartupProcess() || !IsUnderPostmaster)
 		return 0;
 
-	delay_backend_us = &backpressure_lag_impl;
+	delay_backend_us = &hadron_backpressure_lag_impl;
 
-	return backpressure_lag_impl();
+	return hadron_backpressure_lag_impl();
 }
 
 /*
@@ -511,8 +594,10 @@ WalproposerShmemInit(void)
 		pg_atomic_init_u64(&walprop_shared->backpressureThrottlingTime, 0);
 		pg_atomic_init_u64(&walprop_shared->currentClusterSize, 0);
 		/* BEGIN_HADRON */
+		pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.effective_max_wal_bytes_per_second, -1);
 		pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.should_limit, 0);
-		pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.last_recorded_time_us, 0);
+		pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.batch_start_time_us, 0);
+		pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.batch_end_time_us, 0);
 		/* END_HADRON */
 	}
 }
@@ -527,8 +612,10 @@ WalproposerShmemInit_SyncSafekeeper(void)
 	pg_atomic_init_u64(&walprop_shared->mineLastElectedTerm, 0);
 	pg_atomic_init_u64(&walprop_shared->backpressureThrottlingTime, 0);
 	/* BEGIN_HADRON */
+	pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.effective_max_wal_bytes_per_second, -1);
 	pg_atomic_init_u32(&walprop_shared->wal_rate_limiter.should_limit, 0);
-	pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.last_recorded_time_us, 0);
+	pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.batch_start_time_us, 0);
+	pg_atomic_init_u64(&walprop_shared->wal_rate_limiter.batch_end_time_us, 0);
 	/* END_HADRON */
 }
 
@@ -560,7 +647,7 @@ backpressure_throttling_impl(void)
 		return retry;
 
 	/* Calculate replicas lag */
-	lag = backpressure_lag_impl();
+	lag = hadron_backpressure_lag_impl();
 	if (lag == 0)
 		return retry;
 
@@ -646,18 +733,19 @@ walprop_pg_get_shmem_state(WalProposer *wp)
  * Record new ps_feedback in the array with shards and update min_feedback.
  */
 static PageserverFeedback
-record_pageserver_feedback(PageserverFeedback *ps_feedback)
+record_pageserver_feedback(PageserverFeedback *ps_feedback, shardno_t num_shards)
 {
 	PageserverFeedback min_feedback;
 
 	Assert(ps_feedback->present);
 	Assert(ps_feedback->shard_number < MAX_SHARDS);
+	Assert(ps_feedback->shard_number < num_shards);
 
 	SpinLockAcquire(&walprop_shared->mutex);
 
-	/* Update the number of shards */
-	if (ps_feedback->shard_number + 1 > walprop_shared->num_shards)
-		walprop_shared->num_shards = ps_feedback->shard_number + 1;
+	// Hadron: Update the num_shards from the source-of-truth (shard map) lazily when we receive
+	// a new pageserver feedback.
+	walprop_shared->num_shards = Max(walprop_shared->num_shards, num_shards);
 
 	/* Update the feedback */
 	memcpy(&walprop_shared->shard_ps_feedback[ps_feedback->shard_number], ps_feedback, sizeof(PageserverFeedback));
@@ -1475,6 +1563,7 @@ XLogBroadcastWalProposer(WalProposer *wp)
 	XLogRecPtr	endptr;
 	struct WalproposerShmemState *state = NULL;
 	TimestampTz now = 0;
+	int effective_max_wal_bytes_per_second = 0;
 
 	/* Start from the last sent position */
 	startptr = sentPtr;
@@ -1529,22 +1618,36 @@ XLogBroadcastWalProposer(WalProposer *wp)
 
 	/* BEGIN_HADRON */
 	state = GetWalpropShmemState();
-	if (databricks_max_wal_mb_per_second != -1 && state != NULL)
+	effective_max_wal_bytes_per_second = pg_atomic_read_u32(&state->wal_rate_limiter.effective_max_wal_bytes_per_second);
+	if (effective_max_wal_bytes_per_second != -1 && state != NULL)
 	{
-		uint64 max_wal_bytes = (uint64) databricks_max_wal_mb_per_second * 1024 * 1024;
 		struct WalRateLimiter *limiter = &state->wal_rate_limiter;
-		uint64 last_recorded_time = pg_atomic_read_u64(&limiter->last_recorded_time_us);
-		if (now - last_recorded_time > USECS_PER_SEC)
+		uint64 batch_end_time = pg_atomic_read_u64(&limiter->batch_end_time_us);
+		if ( now >= batch_end_time )
 		{
-			/* Reset the rate limiter */
+			// Reset the rate limiter to start a new batch
 			limiter->sent_bytes = 0;
-			pg_atomic_write_u64(&limiter->last_recorded_time_us, now);
 			pg_atomic_write_u32(&limiter->should_limit, false);
+			pg_atomic_write_u64(&limiter->batch_start_time_us, now);
+			/* tentatively assign the batch end time as 1s from now. This could result in one of the following cases:
+			1. If sent_bytes does not reach effective_max_wal_bytes_per_second in 1s,
+			then we will reset the current batch and clear sent_bytes. No throttling happens.
+			2. Otherwise, we will recompute the end time (below) based on how many bytes are actually written,
+			and throttle PG until the batch end time. */
+			pg_atomic_write_u64(&limiter->batch_end_time_us, now + USECS_PER_SEC);
 		}
 		limiter->sent_bytes += (endptr - startptr);
-		if (limiter->sent_bytes > max_wal_bytes)
+		if (limiter->sent_bytes > effective_max_wal_bytes_per_second)
 		{
+			uint64_t batch_start_time = pg_atomic_read_u64(&limiter->batch_start_time_us);
+			uint64 throttle_usecs = USECS_PER_SEC * limiter->sent_bytes / Max(effective_max_wal_bytes_per_second, 1);
+			if (throttle_usecs > kRateLimitMaxBatchUSecs){
+				elog(LOG, "throttle_usecs %lu is too large, limiting to %lu", throttle_usecs, kRateLimitMaxBatchUSecs);
+				throttle_usecs = kRateLimitMaxBatchUSecs;
+			}
+
 			pg_atomic_write_u32(&limiter->should_limit, true);
+			pg_atomic_write_u64(&limiter->batch_end_time_us, batch_start_time + throttle_usecs);
 		}
 	}
 	/* END_HADRON */
@@ -2023,19 +2126,43 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, Safekeeper *sk)
 	if (wp->config->syncSafekeepers)
 		return;
 
+
 	/* handle fresh ps_feedback */
 	if (sk->appendResponse.ps_feedback.present)
 	{
-		PageserverFeedback min_feedback = record_pageserver_feedback(&sk->appendResponse.ps_feedback);
+		shardno_t num_shards = get_num_shards();
 
-		/* Only one main shard sends non-zero currentClusterSize */
-		if (sk->appendResponse.ps_feedback.currentClusterSize > 0)
-			SetNeonCurrentClusterSize(sk->appendResponse.ps_feedback.currentClusterSize);
-
-		if (min_feedback.disk_consistent_lsn != standby_apply_lsn)
+		// During shard split, we receive ps_feedback from child shards before
+		// the split commits and our shard map GUC has been updated. We must
+		// filter out such feedback here because record_pageserver_feedback()
+		// doesn't do it.
+		//
+		// NB: what we would actually want to happen is that we only receive
+		// ps_feedback from the parent shards when the split is committed, then
+		// apply the split to our set of tracked feedback and from here on only
+		// receive ps_feedback from child shards. This filter condition doesn't
+		// do that: if we split from N parent to 2N child shards, the first N
+		// child shards' feedback messages will pass this condition, even before
+		// the split is committed. That's a bit sloppy, but OK for now.
+		if (sk->appendResponse.ps_feedback.shard_number < num_shards)
 		{
-			standby_apply_lsn = min_feedback.disk_consistent_lsn;
-			needToAdvanceSlot = true;
+			PageserverFeedback min_feedback = record_pageserver_feedback(&sk->appendResponse.ps_feedback, num_shards);
+
+			/* Only one main shard sends non-zero currentClusterSize */
+			if (sk->appendResponse.ps_feedback.currentClusterSize > 0)
+				SetNeonCurrentClusterSize(sk->appendResponse.ps_feedback.currentClusterSize);
+
+			if (min_feedback.disk_consistent_lsn != standby_apply_lsn)
+			{
+				standby_apply_lsn = min_feedback.disk_consistent_lsn;
+				needToAdvanceSlot = true;
+			}
+		}
+		else
+		{
+			// HADRON
+			elog(DEBUG2, "Ignoring pageserver feedback for unknown shard %d (current shard number %d)",
+				sk->appendResponse.ps_feedback.shard_number, num_shards);
 		}
 	}
 

proxy/Cargo.toml

@@ -33,7 +33,6 @@ env_logger.workspace = true
 framed-websockets.workspace = true
 futures.workspace = true
 hashbrown.workspace = true
-hashlink.workspace = true
 hex.workspace = true
 hmac.workspace = true
 hostname.workspace = true
@@ -54,6 +53,7 @@ json = { path = "../libs/proxy/json" }
 lasso = { workspace = true, features = ["multi-threaded"] }
 measured = { workspace = true, features = ["lasso"] }
 metrics.workspace = true
+moka.workspace = true
 once_cell.workspace = true
 opentelemetry = { workspace = true, features = ["trace"] }
 papaya = "0.2.0"
@@ -110,7 +110,7 @@ zerocopy.workspace = true
 # uncomment this to use the real subzero-core crate
 # subzero-core = { git = "https://github.com/neondatabase/subzero", rev = "396264617e78e8be428682f87469bb25429af88a", features = ["postgresql"], optional = true }
 # this is a stub for the subzero-core crate
-subzero-core = { path = "./subzero_core", features = ["postgresql"], optional = true}
+subzero-core = { path = "../libs/proxy/subzero_core", features = ["postgresql"], optional = true}
 ouroboros = { version = "0.18", optional = true }
 
 # jwt stuff

proxy/src/auth/backend/console_redirect.rs

@@ -8,11 +8,12 @@ use tracing::{info, info_span};
 
 use crate::auth::backend::ComputeUserInfo;
 use crate::cache::Cached;
+use crate::cache::node_info::CachedNodeInfo;
 use crate::compute::AuthInfo;
 use crate::config::AuthenticationConfig;
 use crate::context::RequestContext;
 use crate::control_plane::client::cplane_proxy_v1;
-use crate::control_plane::{self, CachedNodeInfo, NodeInfo};
+use crate::control_plane::{self, NodeInfo};
 use crate::error::{ReportableError, UserFacingError};
 use crate::pqproto::BeMessage;
 use crate::proxy::NeonOptions;

proxy/src/auth/backend/mod.rs

@@ -16,14 +16,14 @@ use tracing::{debug, info};
 
 use crate::auth::{self, ComputeUserInfoMaybeEndpoint, validate_password_and_exchange};
 use crate::cache::Cached;
+use crate::cache::node_info::CachedNodeInfo;
 use crate::config::AuthenticationConfig;
 use crate::context::RequestContext;
 use crate::control_plane::client::ControlPlaneClient;
 use crate::control_plane::errors::GetAuthInfoError;
 use crate::control_plane::messages::EndpointRateLimitConfig;
 use crate::control_plane::{
-    self, AccessBlockerFlags, AuthSecret, CachedNodeInfo, ControlPlaneApi, EndpointAccessControl,
-    RoleAccessControl,
+    self, AccessBlockerFlags, AuthSecret, ControlPlaneApi, EndpointAccessControl, RoleAccessControl,
 };
 use crate::intern::EndpointIdInt;
 use crate::pqproto::BeMessage;
@@ -433,11 +433,12 @@ mod tests {
     use super::auth_quirks;
     use super::jwt::JwkCache;
     use crate::auth::{ComputeUserInfoMaybeEndpoint, IpPattern};
+    use crate::cache::node_info::CachedNodeInfo;
     use crate::config::AuthenticationConfig;
     use crate::context::RequestContext;
     use crate::control_plane::messages::EndpointRateLimitConfig;
     use crate::control_plane::{
-        self, AccessBlockerFlags, CachedNodeInfo, EndpointAccessControl, RoleAccessControl,
+        self, AccessBlockerFlags, EndpointAccessControl, RoleAccessControl,
     };
     use crate::proxy::NeonOptions;
     use crate::rate_limiter::EndpointRateLimiter;

proxy/src/binary/local_proxy.rs

@@ -29,7 +29,7 @@ use crate::config::{
 };
 use crate::control_plane::locks::ApiLocks;
 use crate::http::health_server::AppMetrics;
-use crate::metrics::{Metrics, ThreadPoolMetrics};
+use crate::metrics::{Metrics, ServiceInfo, ThreadPoolMetrics};
 use crate::rate_limiter::{EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo};
 use crate::scram::threadpool::ThreadPool;
 use crate::serverless::cancel_set::CancelSet;
@@ -207,6 +207,11 @@ pub async fn run() -> anyhow::Result<()> {
         endpoint_rate_limiter,
     );
 
+    Metrics::get()
+        .service
+        .info
+        .set_label(ServiceInfo::running());
+
     match futures::future::select(pin!(maintenance_tasks.join_next()), pin!(task)).await {
         // exit immediately on maintenance task completion
         Either::Left((Some(res), _)) => match crate::error::flatten_err(res)? {},

proxy/src/binary/pg_sni_router.rs

@@ -26,7 +26,7 @@ use utils::project_git_version;
 use utils::sentry_init::init_sentry;
 
 use crate::context::RequestContext;
-use crate::metrics::{Metrics, ThreadPoolMetrics};
+use crate::metrics::{Metrics, ServiceInfo, ThreadPoolMetrics};
 use crate::pglb::TlsRequired;
 use crate::pqproto::FeStartupPacket;
 use crate::protocol2::ConnectionInfo;
@@ -135,6 +135,12 @@ pub async fn run() -> anyhow::Result<()> {
         cancellation_token.clone(),
     ))
     .map(crate::error::flatten_err);
+
+    Metrics::get()
+        .service
+        .info
+        .set_label(ServiceInfo::running());
+
     let signals_task = tokio::spawn(crate::signals::handle(cancellation_token, || {}));
 
     // the signal task cant ever succeed.

proxy/src/binary/proxy.rs

@@ -40,7 +40,7 @@ use crate::config::{
 };
 use crate::context::parquet::ParquetUploadArgs;
 use crate::http::health_server::AppMetrics;
-use crate::metrics::Metrics;
+use crate::metrics::{Metrics, ServiceInfo};
 use crate::rate_limiter::{EndpointRateLimiter, RateBucketInfo, WakeComputeRateLimiter};
 use crate::redis::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
 use crate::redis::kv_ops::RedisKVClient;
@@ -535,12 +535,7 @@ pub async fn run() -> anyhow::Result<()> {
     // add a task to flush the db_schema cache every 10 minutes
     #[cfg(feature = "rest_broker")]
     if let Some(db_schema_cache) = &config.rest_config.db_schema_cache {
-        maintenance_tasks.spawn(async move {
-            loop {
-                tokio::time::sleep(Duration::from_secs(600)).await;
-                db_schema_cache.flush();
-            }
-        });
+        maintenance_tasks.spawn(db_schema_cache.maintain());
     }
 
     if let Some(metrics_config) = &config.metric_collection {
@@ -590,6 +585,11 @@ pub async fn run() -> anyhow::Result<()> {
         }
     }
 
+    Metrics::get()
+        .service
+        .info
+        .set_label(ServiceInfo::running());
+
     let maintenance = loop {
         // get one complete task
         match futures::future::select(
@@ -711,12 +711,7 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
         info!("Using DbSchemaCache with options={db_schema_cache_config:?}");
 
         let db_schema_cache = if args.is_rest_broker {
-            Some(DbSchemaCache::new(
-                "db_schema_cache",
-                db_schema_cache_config.size,
-                db_schema_cache_config.ttl,
-                true,
-            ))
+            Some(DbSchemaCache::new(db_schema_cache_config))
         } else {
             None
         };

proxy/src/cache/common.rs

@@ -1,4 +1,16 @@
 use std::ops::{Deref, DerefMut};
+use std::time::{Duration, Instant};
+
+use moka::Expiry;
+use moka::notification::RemovalCause;
+
+use crate::control_plane::messages::ControlPlaneErrorMessage;
+use crate::metrics::{
+    CacheEviction, CacheKind, CacheOutcome, CacheOutcomeGroup, CacheRemovalCause, Metrics,
+};
+
+/// Default TTL used when caching errors from control plane.
+pub const DEFAULT_ERROR_TTL: Duration = Duration::from_secs(30);
 
 /// A generic trait which exposes types of cache's key and value,
 /// as well as the notion of cache entry invalidation.
@@ -10,20 +22,16 @@ pub(crate) trait Cache {
     /// Entry's value.
     type Value;
 
-    /// Used for entry invalidation.
-    type LookupInfo<Key>;
-
     /// Invalidate an entry using a lookup info.
     /// We don't have an empty default impl because it's error-prone.
-    fn invalidate(&self, _: &Self::LookupInfo<Self::Key>);
+    fn invalidate(&self, _: &Self::Key);
 }
 
 impl<C: Cache> Cache for &C {
     type Key = C::Key;
     type Value = C::Value;
-    type LookupInfo<Key> = C::LookupInfo<Key>;
 
-    fn invalidate(&self, info: &Self::LookupInfo<Self::Key>) {
+    fn invalidate(&self, info: &Self::Key) {
         C::invalidate(self, info);
     }
 }
@@ -31,7 +39,7 @@ impl<C: Cache> Cache for &C {
 /// Wrapper for convenient entry invalidation.
 pub(crate) struct Cached<C: Cache, V = <C as Cache>::Value> {
     /// Cache + lookup info.
-    pub(crate) token: Option<(C, C::LookupInfo<C::Key>)>,
+    pub(crate) token: Option<(C, C::Key)>,
 
     /// The value itself.
     pub(crate) value: V,
@@ -43,23 +51,6 @@ impl<C: Cache, V> Cached<C, V> {
         Self { token: None, value }
     }
 
-    pub(crate) fn take_value(self) -> (Cached<C, ()>, V) {
-        (
-            Cached {
-                token: self.token,
-                value: (),
-            },
-            self.value,
-        )
-    }
-
-    pub(crate) fn map<U>(self, f: impl FnOnce(V) -> U) -> Cached<C, U> {
-        Cached {
-            token: self.token,
-            value: f(self.value),
-        }
-    }
-
     /// Drop this entry from a cache if it's still there.
     pub(crate) fn invalidate(self) -> V {
         if let Some((cache, info)) = &self.token {
@@ -87,3 +78,91 @@ impl<C: Cache, V> DerefMut for Cached<C, V> {
         &mut self.value
     }
 }
+
+pub type ControlPlaneResult<T> = Result<T, Box<ControlPlaneErrorMessage>>;
+
+#[derive(Clone, Copy)]
+pub struct CplaneExpiry {
+    pub error: Duration,
+}
+
+impl Default for CplaneExpiry {
+    fn default() -> Self {
+        Self {
+            error: DEFAULT_ERROR_TTL,
+        }
+    }
+}
+
+impl CplaneExpiry {
+    pub fn expire_early<V>(
+        &self,
+        value: &ControlPlaneResult<V>,
+        updated: Instant,
+    ) -> Option<Duration> {
+        match value {
+            Ok(_) => None,
+            Err(err) => Some(self.expire_err_early(err, updated)),
+        }
+    }
+
+    pub fn expire_err_early(&self, err: &ControlPlaneErrorMessage, updated: Instant) -> Duration {
+        err.status
+            .as_ref()
+            .and_then(|s| s.details.retry_info.as_ref())
+            .map_or(self.error, |r| r.retry_at.into_std() - updated)
+    }
+}
+
+impl<K, V> Expiry<K, ControlPlaneResult<V>> for CplaneExpiry {
+    fn expire_after_create(
+        &self,
+        _key: &K,
+        value: &ControlPlaneResult<V>,
+        created_at: Instant,
+    ) -> Option<Duration> {
+        self.expire_early(value, created_at)
+    }
+
+    fn expire_after_update(
+        &self,
+        _key: &K,
+        value: &ControlPlaneResult<V>,
+        updated_at: Instant,
+        _duration_until_expiry: Option<Duration>,
+    ) -> Option<Duration> {
+        self.expire_early(value, updated_at)
+    }
+}
+
+pub fn eviction_listener(kind: CacheKind, cause: RemovalCause) {
+    let cause = match cause {
+        RemovalCause::Expired => CacheRemovalCause::Expired,
+        RemovalCause::Explicit => CacheRemovalCause::Explicit,
+        RemovalCause::Replaced => CacheRemovalCause::Replaced,
+        RemovalCause::Size => CacheRemovalCause::Size,
+    };
+    Metrics::get()
+        .cache
+        .evicted_total
+        .inc(CacheEviction { cache: kind, cause });
+}
+
+#[inline]
+pub fn count_cache_outcome<T>(kind: CacheKind, cache_result: Option<T>) -> Option<T> {
+    let outcome = if cache_result.is_some() {
+        CacheOutcome::Hit
+    } else {
+        CacheOutcome::Miss
+    };
+    Metrics::get().cache.request_total.inc(CacheOutcomeGroup {
+        cache: kind,
+        outcome,
+    });
+    cache_result
+}
+
+#[inline]
+pub fn count_cache_insert(kind: CacheKind) {
+    Metrics::get().cache.inserted_total.inc(kind);
+}

proxy/src/cache/mod.rs

@@ -1,6 +1,5 @@
 pub(crate) mod common;
+pub(crate) mod node_info;
 pub(crate) mod project_info;
-mod timed_lru;
 
-pub(crate) use common::{Cache, Cached};
-pub(crate) use timed_lru::TimedLru;
+pub(crate) use common::{Cached, ControlPlaneResult, CplaneExpiry};

proxy/src/cache/node_info.rs

@@ -0,0 +1,60 @@
+use crate::cache::common::{Cache, count_cache_insert, count_cache_outcome, eviction_listener};
+use crate::cache::{Cached, ControlPlaneResult, CplaneExpiry};
+use crate::config::CacheOptions;
+use crate::control_plane::NodeInfo;
+use crate::metrics::{CacheKind, Metrics};
+use crate::types::EndpointCacheKey;
+
+pub(crate) struct NodeInfoCache(moka::sync::Cache<EndpointCacheKey, ControlPlaneResult<NodeInfo>>);
+pub(crate) type CachedNodeInfo = Cached<&'static NodeInfoCache, NodeInfo>;
+
+impl Cache for NodeInfoCache {
+    type Key = EndpointCacheKey;
+    type Value = ControlPlaneResult<NodeInfo>;
+
+    fn invalidate(&self, info: &EndpointCacheKey) {
+        self.0.invalidate(info);
+    }
+}
+
+impl NodeInfoCache {
+    pub fn new(config: CacheOptions) -> Self {
+        let builder = moka::sync::Cache::builder()
+            .name("node_info")
+            .expire_after(CplaneExpiry::default());
+        let builder = config.moka(builder);
+
+        if let Some(size) = config.size {
+            Metrics::get()
+                .cache
+                .capacity
+                .set(CacheKind::NodeInfo, size as i64);
+        }
+
+        let builder = builder
+            .eviction_listener(|_k, _v, cause| eviction_listener(CacheKind::NodeInfo, cause));
+
+        Self(builder.build())
+    }
+
+    pub fn insert(&self, key: EndpointCacheKey, value: ControlPlaneResult<NodeInfo>) {
+        count_cache_insert(CacheKind::NodeInfo);
+        self.0.insert(key, value);
+    }
+
+    pub fn get(&self, key: &EndpointCacheKey) -> Option<ControlPlaneResult<NodeInfo>> {
+        count_cache_outcome(CacheKind::NodeInfo, self.0.get(key))
+    }
+
+    pub fn get_entry(
+        &'static self,
+        key: &EndpointCacheKey,
+    ) -> Option<ControlPlaneResult<CachedNodeInfo>> {
+        self.get(key).map(|res| {
+            res.map(|value| Cached {
+                token: Some((self, key.clone())),
+                value,
+            })
+        })
+    }
+}

proxy/src/cache/project_info.rs

@@ -1,84 +1,20 @@
-use std::collections::{HashMap, HashSet, hash_map};
+use std::collections::HashSet;
 use std::convert::Infallible;
-use std::time::Duration;
 
-use async_trait::async_trait;
 use clashmap::ClashMap;
-use clashmap::mapref::one::Ref;
-use rand::Rng;
-use tokio::time::Instant;
+use moka::sync::Cache;
 use tracing::{debug, info};
 
+use crate::cache::common::{
+    ControlPlaneResult, CplaneExpiry, count_cache_insert, count_cache_outcome, eviction_listener,
+};
 use crate::config::ProjectInfoCacheOptions;
 use crate::control_plane::messages::{ControlPlaneErrorMessage, Reason};
 use crate::control_plane::{EndpointAccessControl, RoleAccessControl};
 use crate::intern::{AccountIdInt, EndpointIdInt, ProjectIdInt, RoleNameInt};
+use crate::metrics::{CacheKind, Metrics};
 use crate::types::{EndpointId, RoleName};
 
-#[async_trait]
-pub(crate) trait ProjectInfoCache {
-    fn invalidate_endpoint_access(&self, endpoint_id: EndpointIdInt);
-    fn invalidate_endpoint_access_for_project(&self, project_id: ProjectIdInt);
-    fn invalidate_endpoint_access_for_org(&self, account_id: AccountIdInt);
-    fn invalidate_role_secret_for_project(&self, project_id: ProjectIdInt, role_name: RoleNameInt);
-}
-
-struct Entry<T> {
-    expires_at: Instant,
-    value: T,
-}
-
-impl<T> Entry<T> {
-    pub(crate) fn new(value: T, ttl: Duration) -> Self {
-        Self {
-            expires_at: Instant::now() + ttl,
-            value,
-        }
-    }
-
-    pub(crate) fn get(&self) -> Option<&T> {
-        (!self.is_expired()).then_some(&self.value)
-    }
-
-    fn is_expired(&self) -> bool {
-        self.expires_at <= Instant::now()
-    }
-}
-
-struct EndpointInfo {
-    role_controls: HashMap<RoleNameInt, Entry<ControlPlaneResult<RoleAccessControl>>>,
-    controls: Option<Entry<ControlPlaneResult<EndpointAccessControl>>>,
-}
-
-type ControlPlaneResult<T> = Result<T, Box<ControlPlaneErrorMessage>>;
-
-impl EndpointInfo {
-    pub(crate) fn get_role_secret_with_ttl(
-        &self,
-        role_name: RoleNameInt,
-    ) -> Option<(ControlPlaneResult<RoleAccessControl>, Duration)> {
-        let entry = self.role_controls.get(&role_name)?;
-        let ttl = entry.expires_at - Instant::now();
-        Some((entry.get()?.clone(), ttl))
-    }
-
-    pub(crate) fn get_controls_with_ttl(
-        &self,
-    ) -> Option<(ControlPlaneResult<EndpointAccessControl>, Duration)> {
-        let entry = self.controls.as_ref()?;
-        let ttl = entry.expires_at - Instant::now();
-        Some((entry.get()?.clone(), ttl))
-    }
-
-    pub(crate) fn invalidate_endpoint(&mut self) {
-        self.controls = None;
-    }
-
-    pub(crate) fn invalidate_role_secret(&mut self, role_name: RoleNameInt) {
-        self.role_controls.remove(&role_name);
-    }
-}
-
 /// Cache for project info.
 /// This is used to cache auth data for endpoints.
 /// Invalidation is done by console notifications or by TTL (if console notifications are disabled).
@@ -86,8 +22,9 @@ impl EndpointInfo {
 /// We also store endpoint-to-project mapping in the cache, to be able to access per-endpoint data.
 /// One may ask, why the data is stored per project, when on the user request there is only data about the endpoint available?
 /// On the cplane side updates are done per project (or per branch), so it's easier to invalidate the whole project cache.
-pub struct ProjectInfoCacheImpl {
-    cache: ClashMap<EndpointIdInt, EndpointInfo>,
+pub struct ProjectInfoCache {
+    role_controls: Cache<(EndpointIdInt, RoleNameInt), ControlPlaneResult<RoleAccessControl>>,
+    ep_controls: Cache<EndpointIdInt, ControlPlaneResult<EndpointAccessControl>>,
 
     project2ep: ClashMap<ProjectIdInt, HashSet<EndpointIdInt>>,
     // FIXME(stefan): we need a way to GC the account2ep map.
@@ -96,16 +33,13 @@ pub struct ProjectInfoCacheImpl {
     config: ProjectInfoCacheOptions,
 }
 
-#[async_trait]
-impl ProjectInfoCache for ProjectInfoCacheImpl {
-    fn invalidate_endpoint_access(&self, endpoint_id: EndpointIdInt) {
+impl ProjectInfoCache {
+    pub fn invalidate_endpoint_access(&self, endpoint_id: EndpointIdInt) {
         info!("invalidating endpoint access for `{endpoint_id}`");
-        if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-            endpoint_info.invalidate_endpoint();
-        }
+        self.ep_controls.invalidate(&endpoint_id);
     }
 
-    fn invalidate_endpoint_access_for_project(&self, project_id: ProjectIdInt) {
+    pub fn invalidate_endpoint_access_for_project(&self, project_id: ProjectIdInt) {
         info!("invalidating endpoint access for project `{project_id}`");
         let endpoints = self
             .project2ep
@@ -113,13 +47,11 @@ impl ProjectInfoCache for ProjectInfoCacheImpl {
             .map(|kv| kv.value().clone())
             .unwrap_or_default();
         for endpoint_id in endpoints {
-            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_endpoint();
-            }
+            self.ep_controls.invalidate(&endpoint_id);
         }
     }
 
-    fn invalidate_endpoint_access_for_org(&self, account_id: AccountIdInt) {
+    pub fn invalidate_endpoint_access_for_org(&self, account_id: AccountIdInt) {
         info!("invalidating endpoint access for org `{account_id}`");
         let endpoints = self
             .account2ep
@@ -127,13 +59,15 @@ impl ProjectInfoCache for ProjectInfoCacheImpl {
             .map(|kv| kv.value().clone())
             .unwrap_or_default();
         for endpoint_id in endpoints {
-            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_endpoint();
-            }
+            self.ep_controls.invalidate(&endpoint_id);
         }
     }
 
-    fn invalidate_role_secret_for_project(&self, project_id: ProjectIdInt, role_name: RoleNameInt) {
+    pub fn invalidate_role_secret_for_project(
+        &self,
+        project_id: ProjectIdInt,
+        role_name: RoleNameInt,
+    ) {
         info!(
             "invalidating role secret for project_id `{}` and role_name `{}`",
             project_id, role_name,
@@ -144,47 +78,73 @@ impl ProjectInfoCache for ProjectInfoCacheImpl {
             .map(|kv| kv.value().clone())
             .unwrap_or_default();
         for endpoint_id in endpoints {
-            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_role_secret(role_name);
-            }
+            self.role_controls.invalidate(&(endpoint_id, role_name));
         }
     }
 }
 
-impl ProjectInfoCacheImpl {
+impl ProjectInfoCache {
     pub(crate) fn new(config: ProjectInfoCacheOptions) -> Self {
+        Metrics::get().cache.capacity.set(
+            CacheKind::ProjectInfoRoles,
+            (config.size * config.max_roles) as i64,
+        );
+        Metrics::get()
+            .cache
+            .capacity
+            .set(CacheKind::ProjectInfoEndpoints, config.size as i64);
+
+        // we cache errors for 30 seconds, unless retry_at is set.
+        let expiry = CplaneExpiry::default();
         Self {
-            cache: ClashMap::new(),
+            role_controls: Cache::builder()
+                .name("project_info_roles")
+                .eviction_listener(|_k, _v, cause| {
+                    eviction_listener(CacheKind::ProjectInfoRoles, cause);
+                })
+                .max_capacity(config.size * config.max_roles)
+                .time_to_live(config.ttl)
+                .expire_after(expiry)
+                .build(),
+            ep_controls: Cache::builder()
+                .name("project_info_endpoints")
+                .eviction_listener(|_k, _v, cause| {
+                    eviction_listener(CacheKind::ProjectInfoEndpoints, cause);
+                })
+                .max_capacity(config.size)
+                .time_to_live(config.ttl)
+                .expire_after(expiry)
+                .build(),
             project2ep: ClashMap::new(),
             account2ep: ClashMap::new(),
             config,
         }
     }
 
-    fn get_endpoint_cache(
-        &self,
-        endpoint_id: &EndpointId,
-    ) -> Option<Ref<'_, EndpointIdInt, EndpointInfo>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
-        self.cache.get(&endpoint_id)
-    }
-
-    pub(crate) fn get_role_secret_with_ttl(
+    pub(crate) fn get_role_secret(
         &self,
         endpoint_id: &EndpointId,
         role_name: &RoleName,
-    ) -> Option<(ControlPlaneResult<RoleAccessControl>, Duration)> {
+    ) -> Option<ControlPlaneResult<RoleAccessControl>> {
+        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
         let role_name = RoleNameInt::get(role_name)?;
-        let endpoint_info = self.get_endpoint_cache(endpoint_id)?;
-        endpoint_info.get_role_secret_with_ttl(role_name)
+
+        count_cache_outcome(
+            CacheKind::ProjectInfoRoles,
+            self.role_controls.get(&(endpoint_id, role_name)),
+        )
     }
 
-    pub(crate) fn get_endpoint_access_with_ttl(
+    pub(crate) fn get_endpoint_access(
         &self,
         endpoint_id: &EndpointId,
-    ) -> Option<(ControlPlaneResult<EndpointAccessControl>, Duration)> {
-        let endpoint_info = self.get_endpoint_cache(endpoint_id)?;
-        endpoint_info.get_controls_with_ttl()
+    ) -> Option<ControlPlaneResult<EndpointAccessControl>> {
+        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
+
+        count_cache_outcome(
+            CacheKind::ProjectInfoEndpoints,
+            self.ep_controls.get(&endpoint_id),
+        )
     }
 
     pub(crate) fn insert_endpoint_access(
@@ -203,34 +163,17 @@ impl ProjectInfoCacheImpl {
             self.insert_project2endpoint(project_id, endpoint_id);
         }
 
-        if self.cache.len() >= self.config.size {
-            // If there are too many entries, wait until the next gc cycle.
-            return;
-        }
-
         debug!(
             key = &*endpoint_id,
             "created a cache entry for endpoint access"
         );
 
-        let controls = Some(Entry::new(Ok(controls), self.config.ttl));
-        let role_controls = Entry::new(Ok(role_controls), self.config.ttl);
+        count_cache_insert(CacheKind::ProjectInfoEndpoints);
+        count_cache_insert(CacheKind::ProjectInfoRoles);
 
-        match self.cache.entry(endpoint_id) {
-            clashmap::Entry::Vacant(e) => {
-                e.insert(EndpointInfo {
-                    role_controls: HashMap::from_iter([(role_name, role_controls)]),
-                    controls,
-                });
-            }
-            clashmap::Entry::Occupied(mut e) => {
-                let ep = e.get_mut();
-                ep.controls = controls;
-                if ep.role_controls.len() < self.config.max_roles {
-                    ep.role_controls.insert(role_name, role_controls);
-                }
-            }
-        }
+        self.ep_controls.insert(endpoint_id, Ok(controls));
+        self.role_controls
+            .insert((endpoint_id, role_name), Ok(role_controls));
     }
 
     pub(crate) fn insert_endpoint_access_err(
@@ -238,55 +181,34 @@ impl ProjectInfoCacheImpl {
         endpoint_id: EndpointIdInt,
         role_name: RoleNameInt,
         msg: Box<ControlPlaneErrorMessage>,
-        ttl: Option<Duration>,
     ) {
-        if self.cache.len() >= self.config.size {
-            // If there are too many entries, wait until the next gc cycle.
-            return;
-        }
-
         debug!(
             key = &*endpoint_id,
             "created a cache entry for an endpoint access error"
         );
 
-        let ttl = ttl.unwrap_or(self.config.ttl);
-
-        let controls = if msg.get_reason() == Reason::RoleProtected {
-            // RoleProtected is the only role-specific error that control plane can give us.
-            // If a given role name does not exist, it still returns a successful response,
-            // just with an empty secret.
-            None
-        } else {
-            // We can cache all the other errors in EndpointInfo.controls,
-            // because they don't depend on what role name we pass to control plane.
-            Some(Entry::new(Err(msg.clone()), ttl))
-        };
-
-        let role_controls = Entry::new(Err(msg), ttl);
-
-        match self.cache.entry(endpoint_id) {
-            clashmap::Entry::Vacant(e) => {
-                e.insert(EndpointInfo {
-                    role_controls: HashMap::from_iter([(role_name, role_controls)]),
-                    controls,
+        // RoleProtected is the only role-specific error that control plane can give us.
+        // If a given role name does not exist, it still returns a successful response,
+        // just with an empty secret.
+        if msg.get_reason() != Reason::RoleProtected {
+            // We can cache all the other errors in ep_controls because they don't
+            // depend on what role name we pass to control plane.
+            self.ep_controls
+                .entry(endpoint_id)
+                .and_compute_with(|entry| match entry {
+                    // leave the entry alone if it's already Ok
+                    Some(entry) if entry.value().is_ok() => moka::ops::compute::Op::Nop,
+                    // replace the entry
+                    _ => {
+                        count_cache_insert(CacheKind::ProjectInfoEndpoints);
+                        moka::ops::compute::Op::Put(Err(msg.clone()))
+                    }
                 });
-            }
-            clashmap::Entry::Occupied(mut e) => {
-                let ep = e.get_mut();
-                if let Some(entry) = &ep.controls
-                    && !entry.is_expired()
-                    && entry.value.is_ok()
-                {
-                    // If we have cached non-expired, non-error controls, keep them.
-                } else {
-                    ep.controls = controls;
-                }
-                if ep.role_controls.len() < self.config.max_roles {
-                    ep.role_controls.insert(role_name, role_controls);
-                }
-            }
         }
+
+        count_cache_insert(CacheKind::ProjectInfoRoles);
+        self.role_controls
+            .insert((endpoint_id, role_name), Err(msg));
     }
 
     fn insert_project2endpoint(&self, project_id: ProjectIdInt, endpoint_id: EndpointIdInt) {
@@ -307,73 +229,35 @@ impl ProjectInfoCacheImpl {
         }
     }
 
-    pub fn maybe_invalidate_role_secret(&self, endpoint_id: &EndpointId, role_name: &RoleName) {
-        let Some(endpoint_id) = EndpointIdInt::get(endpoint_id) else {
-            return;
-        };
-        let Some(role_name) = RoleNameInt::get(role_name) else {
-            return;
-        };
-
-        let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) else {
-            return;
-        };
-
-        let entry = endpoint_info.role_controls.entry(role_name);
-        let hash_map::Entry::Occupied(role_controls) = entry else {
-            return;
-        };
-
-        if role_controls.get().is_expired() {
-            role_controls.remove();
-        }
+    pub fn maybe_invalidate_role_secret(&self, _endpoint_id: &EndpointId, _role_name: &RoleName) {
+        // TODO: Expire the value early if the key is idle.
+        // Currently not an issue as we would just use the TTL to decide, which is what already happens.
     }
 
     pub async fn gc_worker(&self) -> anyhow::Result<Infallible> {
-        let mut interval =
-            tokio::time::interval(self.config.gc_interval / (self.cache.shards().len()) as u32);
+        let mut interval = tokio::time::interval(self.config.gc_interval);
         loop {
             interval.tick().await;
-            if self.cache.len() < self.config.size {
-                // If there are not too many entries, wait until the next gc cycle.
-                continue;
-            }
-            self.gc();
+            self.ep_controls.run_pending_tasks();
+            self.role_controls.run_pending_tasks();
         }
     }
-
-    fn gc(&self) {
-        let shard = rand::rng().random_range(0..self.project2ep.shards().len());
-        debug!(shard, "project_info_cache: performing epoch reclamation");
-
-        // acquire a random shard lock
-        let mut removed = 0;
-        let shard = self.project2ep.shards()[shard].write();
-        for (_, endpoints) in shard.iter() {
-            for endpoint in endpoints {
-                self.cache.remove(endpoint);
-                removed += 1;
-            }
-        }
-        // We can drop this shard only after making sure that all endpoints are removed.
-        drop(shard);
-        info!("project_info_cache: removed {removed} endpoints");
-    }
 }
 
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
+    use std::time::Duration;
+
     use super::*;
     use crate::control_plane::messages::{Details, EndpointRateLimitConfig, ErrorInfo, Status};
     use crate::control_plane::{AccessBlockerFlags, AuthSecret};
     use crate::scram::ServerSecret;
-    use std::sync::Arc;
 
     #[tokio::test]
     async fn test_project_info_cache_settings() {
-        tokio::time::pause();
-        let cache = ProjectInfoCacheImpl::new(ProjectInfoCacheOptions {
-            size: 2,
+        let cache = ProjectInfoCache::new(ProjectInfoCacheOptions {
+            size: 1,
             max_roles: 2,
             ttl: Duration::from_secs(1),
             gc_interval: Duration::from_secs(600),
@@ -423,22 +307,17 @@ mod tests {
             },
         );
 
-        let (cached, ttl) = cache
-            .get_role_secret_with_ttl(&endpoint_id, &user1)
-            .unwrap();
+        let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
         assert_eq!(cached.unwrap().secret, secret1);
-        assert_eq!(ttl, cache.config.ttl);
 
-        let (cached, ttl) = cache
-            .get_role_secret_with_ttl(&endpoint_id, &user2)
-            .unwrap();
+        let cached = cache.get_role_secret(&endpoint_id, &user2).unwrap();
         assert_eq!(cached.unwrap().secret, secret2);
-        assert_eq!(ttl, cache.config.ttl);
 
         // Shouldn't add more than 2 roles.
         let user3: RoleName = "user3".into();
         let secret3 = Some(AuthSecret::Scram(ServerSecret::mock([3; 32])));
 
+        cache.role_controls.run_pending_tasks();
         cache.insert_endpoint_access(
             account_id,
             project_id,
@@ -455,31 +334,18 @@ mod tests {
             },
         );
 
-        assert!(
-            cache
-                .get_role_secret_with_ttl(&endpoint_id, &user3)
-                .is_none()
-        );
+        cache.role_controls.run_pending_tasks();
+        assert_eq!(cache.role_controls.entry_count(), 2);
 
-        let cached = cache
-            .get_endpoint_access_with_ttl(&endpoint_id)
-            .unwrap()
-            .0
-            .unwrap();
-        assert_eq!(cached.allowed_ips, allowed_ips);
+        tokio::time::sleep(Duration::from_secs(2)).await;
 
-        tokio::time::advance(Duration::from_secs(2)).await;
-        let cached = cache.get_role_secret_with_ttl(&endpoint_id, &user1);
-        assert!(cached.is_none());
-        let cached = cache.get_role_secret_with_ttl(&endpoint_id, &user2);
-        assert!(cached.is_none());
-        let cached = cache.get_endpoint_access_with_ttl(&endpoint_id);
-        assert!(cached.is_none());
+        cache.role_controls.run_pending_tasks();
+        assert_eq!(cache.role_controls.entry_count(), 0);
     }
 
     #[tokio::test]
     async fn test_caching_project_info_errors() {
-        let cache = ProjectInfoCacheImpl::new(ProjectInfoCacheOptions {
+        let cache = ProjectInfoCache::new(ProjectInfoCacheOptions {
             size: 10,
             max_roles: 10,
             ttl: Duration::from_secs(1),
@@ -519,34 +385,23 @@ mod tests {
             status: None,
         });
 
-        let get_role_secret = |endpoint_id, role_name| {
-            cache
-                .get_role_secret_with_ttl(endpoint_id, role_name)
-                .unwrap()
-                .0
-        };
-        let get_endpoint_access =
-            |endpoint_id| cache.get_endpoint_access_with_ttl(endpoint_id).unwrap().0;
+        let get_role_secret =
+            |endpoint_id, role_name| cache.get_role_secret(endpoint_id, role_name).unwrap();
+        let get_endpoint_access = |endpoint_id| cache.get_endpoint_access(endpoint_id).unwrap();
 
         // stores role-specific errors only for get_role_secret
-        cache.insert_endpoint_access_err(
-            (&endpoint_id).into(),
-            (&user1).into(),
-            role_msg.clone(),
-            None,
-        );
+        cache.insert_endpoint_access_err((&endpoint_id).into(), (&user1).into(), role_msg.clone());
         assert_eq!(
             get_role_secret(&endpoint_id, &user1).unwrap_err().error,
             role_msg.error
         );
-        assert!(cache.get_endpoint_access_with_ttl(&endpoint_id).is_none());
+        assert!(cache.get_endpoint_access(&endpoint_id).is_none());
 
         // stores non-role specific errors for both get_role_secret and get_endpoint_access
         cache.insert_endpoint_access_err(
             (&endpoint_id).into(),
             (&user1).into(),
             generic_msg.clone(),
-            None,
         );
         assert_eq!(
             get_role_secret(&endpoint_id, &user1).unwrap_err().error,
@@ -558,11 +413,7 @@ mod tests {
         );
 
         // error isn't returned for other roles in the same endpoint
-        assert!(
-            cache
-                .get_role_secret_with_ttl(&endpoint_id, &user2)
-                .is_none()
-        );
+        assert!(cache.get_role_secret(&endpoint_id, &user2).is_none());
 
         // success for a role does not overwrite errors for other roles
         cache.insert_endpoint_access(
@@ -590,7 +441,6 @@ mod tests {
             (&endpoint_id).into(),
             (&user2).into(),
             generic_msg.clone(),
-            None,
         );
         assert!(get_role_secret(&endpoint_id, &user2).is_err());
         assert!(get_endpoint_access(&endpoint_id).is_ok());

proxy/src/cache/timed_lru.rs

@@ -1,262 +0,0 @@
-use std::borrow::Borrow;
-use std::hash::Hash;
-use std::time::{Duration, Instant};
-
-// This seems to make more sense than `lru` or `cached`:
-//
-// * `near/nearcore` ditched `cached` in favor of `lru`
-//   (https://github.com/near/nearcore/issues?q=is%3Aissue+lru+is%3Aclosed).
-//
-// * `lru` methods use an obscure `KeyRef` type in their contraints (which is deliberately excluded from docs).
-//   This severely hinders its usage both in terms of creating wrappers and supported key types.
-//
-// On the other hand, `hashlink` has good download stats and appears to be maintained.
-use hashlink::{LruCache, linked_hash_map::RawEntryMut};
-use tracing::debug;
-
-use super::Cache;
-use super::common::Cached;
-
-/// An implementation of timed LRU cache with fixed capacity.
-/// Key properties:
-///
-/// * Whenever a new entry is inserted, the least recently accessed one is evicted.
-///   The cache also keeps track of entry's insertion time (`created_at`) and TTL (`expires_at`).
-///
-/// * If `update_ttl_on_retrieval` is `true`. When the entry is about to be retrieved, we check its expiration timestamp.
-///   If the entry has expired, we remove it from the cache; Otherwise we bump the
-///   expiration timestamp (e.g. +5mins) and change its place in LRU list to prolong
-///   its existence.
-///
-/// * There's an API for immediate invalidation (removal) of a cache entry;
-///   It's useful in case we know for sure that the entry is no longer correct.
-///   See [`Cached`] for more information.
-///
-/// * Expired entries are kept in the cache, until they are evicted by the LRU policy,
-///   or by a successful lookup (i.e. the entry hasn't expired yet).
-///   There is no background job to reap the expired records.
-///
-/// * It's possible for an entry that has not yet expired entry to be evicted
-///   before expired items. That's a bit wasteful, but probably fine in practice.
-pub(crate) struct TimedLru<K, V> {
-    /// Cache's name for tracing.
-    name: &'static str,
-
-    /// The underlying cache implementation.
-    cache: parking_lot::Mutex<LruCache<K, Entry<V>>>,
-
-    /// Default time-to-live of a single entry.
-    ttl: Duration,
-
-    update_ttl_on_retrieval: bool,
-}
-
-impl<K: Hash + Eq, V> Cache for TimedLru<K, V> {
-    type Key = K;
-    type Value = V;
-    type LookupInfo<Key> = Key;
-
-    fn invalidate(&self, info: &Self::LookupInfo<K>) {
-        self.invalidate_raw(info);
-    }
-}
-
-struct Entry<T> {
-    created_at: Instant,
-    expires_at: Instant,
-    ttl: Duration,
-    update_ttl_on_retrieval: bool,
-    value: T,
-}
-
-impl<K: Hash + Eq, V> TimedLru<K, V> {
-    /// Construct a new LRU cache with timed entries.
-    pub(crate) fn new(
-        name: &'static str,
-        capacity: usize,
-        ttl: Duration,
-        update_ttl_on_retrieval: bool,
-    ) -> Self {
-        Self {
-            name,
-            cache: LruCache::new(capacity).into(),
-            ttl,
-            update_ttl_on_retrieval,
-        }
-    }
-
-    /// Drop an entry from the cache if it's outdated.
-    #[tracing::instrument(level = "debug", fields(cache = self.name), skip_all)]
-    fn invalidate_raw(&self, key: &K) {
-        // Do costly things before taking the lock.
-        let mut cache = self.cache.lock();
-        let entry = match cache.raw_entry_mut().from_key(key) {
-            RawEntryMut::Vacant(_) => return,
-            RawEntryMut::Occupied(x) => x.remove(),
-        };
-        drop(cache); // drop lock before logging
-
-        let Entry {
-            created_at,
-            expires_at,
-            ..
-        } = entry;
-
-        debug!(
-            ?created_at,
-            ?expires_at,
-            "processed a cache entry invalidation event"
-        );
-    }
-
-    /// Try retrieving an entry by its key, then execute `extract` if it exists.
-    #[tracing::instrument(level = "debug", fields(cache = self.name), skip_all)]
-    fn get_raw<Q, R>(&self, key: &Q, extract: impl FnOnce(&K, &Entry<V>) -> R) -> Option<R>
-    where
-        K: Borrow<Q>,
-        Q: Hash + Eq + ?Sized,
-    {
-        let now = Instant::now();
-
-        // Do costly things before taking the lock.
-        let mut cache = self.cache.lock();
-        let mut raw_entry = match cache.raw_entry_mut().from_key(key) {
-            RawEntryMut::Vacant(_) => return None,
-            RawEntryMut::Occupied(x) => x,
-        };
-
-        // Immeditely drop the entry if it has expired.
-        let entry = raw_entry.get();
-        if entry.expires_at <= now {
-            raw_entry.remove();
-            return None;
-        }
-
-        let value = extract(raw_entry.key(), entry);
-        let (created_at, expires_at) = (entry.created_at, entry.expires_at);
-
-        // Update the deadline and the entry's position in the LRU list.
-        let deadline = now.checked_add(raw_entry.get().ttl).expect("time overflow");
-        if raw_entry.get().update_ttl_on_retrieval {
-            raw_entry.get_mut().expires_at = deadline;
-        }
-        raw_entry.to_back();
-
-        drop(cache); // drop lock before logging
-        debug!(
-            created_at = format_args!("{created_at:?}"),
-            old_expires_at = format_args!("{expires_at:?}"),
-            new_expires_at = format_args!("{deadline:?}"),
-            "accessed a cache entry"
-        );
-
-        Some(value)
-    }
-
-    /// Insert an entry to the cache. If an entry with the same key already
-    /// existed, return the previous value and its creation timestamp.
-    #[tracing::instrument(level = "debug", fields(cache = self.name), skip_all)]
-    fn insert_raw(&self, key: K, value: V) -> (Instant, Option<V>) {
-        self.insert_raw_ttl(key, value, self.ttl, self.update_ttl_on_retrieval)
-    }
-
-    /// Insert an entry to the cache. If an entry with the same key already
-    /// existed, return the previous value and its creation timestamp.
-    #[tracing::instrument(level = "debug", fields(cache = self.name), skip_all)]
-    fn insert_raw_ttl(
-        &self,
-        key: K,
-        value: V,
-        ttl: Duration,
-        update: bool,
-    ) -> (Instant, Option<V>) {
-        let created_at = Instant::now();
-        let expires_at = created_at.checked_add(ttl).expect("time overflow");
-
-        let entry = Entry {
-            created_at,
-            expires_at,
-            ttl,
-            update_ttl_on_retrieval: update,
-            value,
-        };
-
-        // Do costly things before taking the lock.
-        let old = self
-            .cache
-            .lock()
-            .insert(key, entry)
-            .map(|entry| entry.value);
-
-        debug!(
-            created_at = format_args!("{created_at:?}"),
-            expires_at = format_args!("{expires_at:?}"),
-            replaced = old.is_some(),
-            "created a cache entry"
-        );
-
-        (created_at, old)
-    }
-}
-
-impl<K: Hash + Eq + Clone, V: Clone> TimedLru<K, V> {
-    pub(crate) fn insert_ttl(&self, key: K, value: V, ttl: Duration) {
-        self.insert_raw_ttl(key, value, ttl, false);
-    }
-
-    #[cfg(feature = "rest_broker")]
-    pub(crate) fn insert(&self, key: K, value: V) {
-        self.insert_raw_ttl(key, value, self.ttl, self.update_ttl_on_retrieval);
-    }
-
-    pub(crate) fn insert_unit(&self, key: K, value: V) -> (Option<V>, Cached<&Self, ()>) {
-        let (_, old) = self.insert_raw(key.clone(), value);
-
-        let cached = Cached {
-            token: Some((self, key)),
-            value: (),
-        };
-
-        (old, cached)
-    }
-
-    #[cfg(feature = "rest_broker")]
-    pub(crate) fn flush(&self) {
-        let now = Instant::now();
-        let mut cache = self.cache.lock();
-
-        // Collect keys of expired entries first
-        let expired_keys: Vec<_> = cache
-            .iter()
-            .filter_map(|(key, entry)| {
-                if entry.expires_at <= now {
-                    Some(key.clone())
-                } else {
-                    None
-                }
-            })
-            .collect();
-
-        // Remove expired entries
-        for key in expired_keys {
-            cache.remove(&key);
-        }
-    }
-}
-
-impl<K: Hash + Eq, V: Clone> TimedLru<K, V> {
-    /// Retrieve a cached entry in convenient wrapper, alongside timing information.
-    pub(crate) fn get_with_created_at<Q>(
-        &self,
-        key: &Q,
-    ) -> Option<Cached<&Self, (<Self as Cache>::Value, Instant)>>
-    where
-        K: Borrow<Q> + Clone,
-        Q: Hash + Eq + ?Sized,
-    {
-        self.get_raw(key, |key, entry| Cached {
-            token: Some((self, key.clone())),
-            value: (entry.value.clone(), entry.created_at),
-        })
-    }
-}

proxy/src/cancellation.rs

@@ -429,26 +429,13 @@ impl CancellationHandler {
 /// (we'd need something like `#![feature(type_alias_impl_trait)]`).
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct CancelClosure {
-    socket_addr: SocketAddr,
-    cancel_token: RawCancelToken,
-    hostname: String, // for pg_sni router
-    user_info: ComputeUserInfo,
+    pub socket_addr: SocketAddr,
+    pub cancel_token: RawCancelToken,
+    pub hostname: String, // for pg_sni router
+    pub user_info: ComputeUserInfo,
 }
 
 impl CancelClosure {
-    pub(crate) fn new(
-        socket_addr: SocketAddr,
-        cancel_token: RawCancelToken,
-        hostname: String,
-        user_info: ComputeUserInfo,
-    ) -> Self {
-        Self {
-            socket_addr,
-            cancel_token,
-            hostname,
-            user_info,
-        }
-    }
     /// Cancels the query running on user's compute node.
     pub(crate) async fn try_cancel_query(
         &self,

proxy/src/compute/mod.rs

@@ -7,17 +7,15 @@ use std::net::{IpAddr, SocketAddr};
 use futures::{FutureExt, TryFutureExt};
 use itertools::Itertools;
 use postgres_client::config::{AuthKeys, ChannelBinding, SslMode};
+use postgres_client::connect_raw::StartupStream;
 use postgres_client::maybe_tls_stream::MaybeTlsStream;
 use postgres_client::tls::MakeTlsConnect;
-use postgres_client::{NoTls, RawCancelToken, RawConnection};
-use postgres_protocol::message::backend::NoticeResponseBody;
 use thiserror::Error;
 use tokio::net::{TcpStream, lookup_host};
 use tracing::{debug, error, info, warn};
 
-use crate::auth::backend::{ComputeCredentialKeys, ComputeUserInfo};
+use crate::auth::backend::ComputeCredentialKeys;
 use crate::auth::parse_endpoint_param;
-use crate::cancellation::CancelClosure;
 use crate::compute::tls::TlsError;
 use crate::config::ComputeConfig;
 use crate::context::RequestContext;
@@ -27,6 +25,7 @@ use crate::control_plane::messages::MetricsAuxInfo;
 use crate::error::{ReportableError, UserFacingError};
 use crate::metrics::{Metrics, NumDbConnectionsGuard};
 use crate::pqproto::StartupMessageParams;
+use crate::proxy::connect_compute::TlsNegotiation;
 use crate::proxy::neon_option;
 use crate::types::Host;
 
@@ -86,6 +85,14 @@ pub(crate) enum ConnectionError {
 
     #[error("error acquiring resource permit: {0}")]
     TooManyConnectionAttempts(#[from] ApiLockError),
+
+    #[cfg(test)]
+    #[error("retryable: {retryable}, wakeable: {wakeable}, kind: {kind:?}")]
+    TestError {
+        retryable: bool,
+        wakeable: bool,
+        kind: crate::error::ErrorKind,
+    },
 }
 
 impl UserFacingError for ConnectionError {
@@ -96,6 +103,8 @@ impl UserFacingError for ConnectionError {
                 "Failed to acquire permit to connect to the database. Too many database connection attempts are currently ongoing.".to_owned()
             }
             ConnectionError::TlsError(_) => COULD_NOT_CONNECT.to_owned(),
+            #[cfg(test)]
+            ConnectionError::TestError { .. } => self.to_string(),
         }
     }
 }
@@ -106,6 +115,8 @@ impl ReportableError for ConnectionError {
             ConnectionError::TlsError(_) => crate::error::ErrorKind::Compute,
             ConnectionError::WakeComputeError(e) => e.get_error_kind(),
             ConnectionError::TooManyConnectionAttempts(e) => e.get_error_kind(),
+            #[cfg(test)]
+            ConnectionError::TestError { kind, .. } => *kind,
         }
     }
 }
@@ -236,8 +247,7 @@ impl AuthInfo {
         &self,
         ctx: &RequestContext,
         compute: &mut ComputeConnection,
-        user_info: &ComputeUserInfo,
-    ) -> Result<PostgresSettings, PostgresError> {
+    ) -> Result<(), PostgresError> {
         // client config with stubbed connect info.
         // TODO(conrad): should we rewrite this to bypass tokio-postgres2 entirely,
         // utilising pqproto.rs.
@@ -247,39 +257,10 @@ impl AuthInfo {
         let tmp_config = self.enrich(tmp_config);
 
         let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
-        let connection = tmp_config
-            .tls_and_authenticate(&mut compute.stream, NoTls)
-            .await?;
+        tmp_config.authenticate(&mut compute.stream).await?;
         drop(pause);
 
-        let RawConnection {
-            stream: _,
-            parameters,
-            delayed_notice,
-            process_id,
-            secret_key,
-        } = connection;
-
-        tracing::Span::current().record("pid", tracing::field::display(process_id));
-
-        // NB: CancelToken is supposed to hold socket_addr, but we use connect_raw.
-        // Yet another reason to rework the connection establishing code.
-        let cancel_closure = CancelClosure::new(
-            compute.socket_addr,
-            RawCancelToken {
-                ssl_mode: compute.ssl_mode,
-                process_id,
-                secret_key,
-            },
-            compute.hostname.to_string(),
-            user_info.clone(),
-        );
-
-        Ok(PostgresSettings {
-            params: parameters,
-            cancel_closure,
-            delayed_notice,
-        })
+        Ok(())
     }
 }
 
@@ -288,6 +269,7 @@ impl ConnectInfo {
     async fn connect_raw(
         &self,
         config: &ComputeConfig,
+        tls: TlsNegotiation,
     ) -> Result<(SocketAddr, MaybeTlsStream<TcpStream, RustlsStream>), TlsError> {
         let timeout = config.timeout;
 
@@ -330,7 +312,7 @@ impl ConnectInfo {
         match connect_once(&*addrs).await {
             Ok((sockaddr, stream)) => Ok((
                 sockaddr,
-                tls::connect_tls(stream, self.ssl_mode, config, host).await?,
+                tls::connect_tls(stream, self.ssl_mode, config, host, tls).await?,
             )),
             Err(err) => {
                 warn!("couldn't connect to compute node at {host}:{port}: {err}");
@@ -343,21 +325,9 @@ impl ConnectInfo {
 pub type RustlsStream = <ComputeConfig as MakeTlsConnect<tokio::net::TcpStream>>::Stream;
 pub type MaybeRustlsStream = MaybeTlsStream<tokio::net::TcpStream, RustlsStream>;
 
-// TODO(conrad): we don't need to parse these.
-// These are just immediately forwarded back to the client.
-// We could instead stream them out instead of reading them into memory.
-pub struct PostgresSettings {
-    /// PostgreSQL connection parameters.
-    pub params: std::collections::HashMap<String, String>,
-    /// Query cancellation token.
-    pub cancel_closure: CancelClosure,
-    /// Notices received from compute after authenticating
-    pub delayed_notice: Vec<NoticeResponseBody>,
-}
-
 pub struct ComputeConnection {
     /// Socket connected to a compute node.
-    pub stream: MaybeTlsStream<tokio::net::TcpStream, RustlsStream>,
+    pub stream: StartupStream<tokio::net::TcpStream, RustlsStream>,
     /// Labels for proxy's metrics.
     pub aux: MetricsAuxInfo,
     pub hostname: Host,
@@ -373,9 +343,10 @@ impl ConnectInfo {
         ctx: &RequestContext,
         aux: &MetricsAuxInfo,
         config: &ComputeConfig,
+        tls: TlsNegotiation,
     ) -> Result<ComputeConnection, ConnectionError> {
         let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
-        let (socket_addr, stream) = self.connect_raw(config).await?;
+        let (socket_addr, stream) = self.connect_raw(config, tls).await?;
         drop(pause);
 
         tracing::Span::current().record("compute_id", tracing::field::display(&aux.compute_id));
@@ -390,6 +361,7 @@ impl ConnectInfo {
             ctx.get_testodrome_id().unwrap_or_default(),
         );
 
+        let stream = StartupStream::new(stream);
         let connection = ComputeConnection {
             stream,
             socket_addr,

proxy/src/compute/tls.rs

@@ -7,6 +7,7 @@ use thiserror::Error;
 use tokio::io::{AsyncRead, AsyncWrite};
 
 use crate::pqproto::request_tls;
+use crate::proxy::connect_compute::TlsNegotiation;
 use crate::proxy::retry::CouldRetry;
 
 #[derive(Debug, Error)]
@@ -35,6 +36,7 @@ pub async fn connect_tls<S, T>(
     mode: SslMode,
     tls: &T,
     host: &str,
+    negotiation: TlsNegotiation,
 ) -> Result<MaybeTlsStream<S, T::Stream>, TlsError>
 where
     S: AsyncRead + AsyncWrite + Unpin + Send,
@@ -49,12 +51,15 @@ where
         SslMode::Prefer | SslMode::Require => {}
     }
 
-    if !request_tls(&mut stream).await? {
-        if SslMode::Require == mode {
-            return Err(TlsError::Required);
-        }
-
-        return Ok(MaybeTlsStream::Raw(stream));
+    match negotiation {
+        // No TLS request needed
+        TlsNegotiation::Direct => {}
+        // TLS request successful
+        TlsNegotiation::Postgres if request_tls(&mut stream).await? => {}
+        // TLS request failed but is required
+        TlsNegotiation::Postgres if SslMode::Require == mode => return Err(TlsError::Required),
+        // TLS request failed but is not required
+        TlsNegotiation::Postgres => return Ok(MaybeTlsStream::Raw(stream)),
     }
 
     Ok(MaybeTlsStream::Tls(