Merge pull request #5296 from neondatabase/releases/2023-09-13

Release 2023-09-13
Fix lfc_ensure_function which now disables LFC (#5294 )
2026-01-17 10:22:56 +00:00 · 2023-09-13 13:49:14 +03:00 · 2023-09-13 08:56:03 +03:00 · 2023-09-12 17:13:25 -04:00 · 2023-09-12 22:02:03 +03:00 · 2023-09-12 20:01:21 +01:00
249 changed files with 20086 additions and 5966 deletions
--- a/.dockerignore
+++ b/.dockerignore
@@ -14,10 +14,12 @@
 !pgxn/
 !proxy/
 !safekeeper/
+!s3_scrubber/
 !storage_broker/
 !trace/
 !vendor/postgres-v14/
 !vendor/postgres-v15/
+!vendor/postgres-v16/
 !workspace_hack/
 !neon_local/
 !scripts/ninstall.sh
--- a/.github/actionlint.yml
+++ b/.github/actionlint.yml
@@ -0,0 +1,8 @@
+self-hosted-runner:
+  labels:
+    - gen3
+    - large
+    - small
+    - us-east-2
+config-variables:
+  - SLACK_UPCOMING_RELEASE_CHANNEL_ID
--- a/.github/actions/allure-report-generate/action.yml
+++ b/.github/actions/allure-report-generate/action.yml
@@ -1,6 +1,13 @@
 name: 'Create Allure report'
 description: 'Generate Allure report from uploaded by actions/allure-report-store tests results'

+inputs:
+  store-test-results-into-db:
+    description: 'Whether to store test results into the database. TEST_RESULT_CONNSTR/TEST_RESULT_CONNSTR_NEW should be set'
+    type: boolean
+    required: false
+    default: false
+
 outputs:
  base-url:
    description: 'Base URL for Allure report'
@@ -139,9 +146,11 @@ runs:
        sed -i 's|<a href="." class=|<a href="https://'${BUCKET}'.s3.amazonaws.com/'${REPORT_PREFIX}'/latest/index.html?nocache='"'+Date.now()+'"'" class=|g' ${WORKDIR}/report/app.js

        # Upload a history and the final report (in this particular order to not to have duplicated history in 2 places)
-        # Use sync for the final report to delete files from previous runs
        time aws s3 mv --recursive --only-show-errors "${WORKDIR}/report/history" "s3://${BUCKET}/${REPORT_PREFIX}/latest/history"
-        time aws s3 sync --delete --only-show-errors "${WORKDIR}/report" "s3://${BUCKET}/${REPORT_PREFIX}/${GITHUB_RUN_ID}"
+
+        # Use aws s3 cp (instead of aws s3 sync) to keep files from previous runs to make old URLs work,
+        # and to keep files on the host to upload them to the database
+        time aws s3 cp --recursive --only-show-errors "${WORKDIR}/report" "s3://${BUCKET}/${REPORT_PREFIX}/${GITHUB_RUN_ID}"

        # Generate redirect
        cat <<EOF > ${WORKDIR}/index.html
@@ -170,6 +179,41 @@ runs:
          aws s3 rm "s3://${BUCKET}/${LOCK_FILE}"
        fi

+    - name: Store Allure test stat in the DB
+      if: ${{ !cancelled() && inputs.store-test-results-into-db == 'true' }}
+      shell: bash -euxo pipefail {0}
+      env:
+        COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+        REPORT_JSON_URL: ${{ steps.generate-report.outputs.report-json-url }}
+      run: |
+        export DATABASE_URL=${REGRESS_TEST_RESULT_CONNSTR}
+
+        ./scripts/pysync
+
+        poetry run python3 scripts/ingest_regress_test_result.py \
+          --revision ${COMMIT_SHA} \
+          --reference ${GITHUB_REF} \
+          --build-type unified \
+          --ingest ${WORKDIR}/report/data/suites.json
+
+    - name: Store Allure test stat in the DB (new)
+      if: ${{ !cancelled() && inputs.store-test-results-into-db == 'true' }}
+      shell: bash -euxo pipefail {0}
+      env:
+        COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+        BASE_S3_URL: ${{ steps.generate-report.outputs.base-s3-url }}
+      run: |
+        export DATABASE_URL=${REGRESS_TEST_RESULT_CONNSTR_NEW}
+
+        ./scripts/pysync
+
+        poetry run python3 scripts/ingest_regress_test_result-new-format.py \
+          --reference ${GITHUB_REF} \
+          --revision ${COMMIT_SHA} \
+          --run-id ${GITHUB_RUN_ID} \
+          --run-attempt ${GITHUB_RUN_ATTEMPT} \
+          --test-cases-dir ${WORKDIR}/report/data/test-cases
+
    - name: Cleanup
      if: always()
      shell: bash -euxo pipefail {0}
--- a/.github/actions/run-python-test-set/action.yml
+++ b/.github/actions/run-python-test-set/action.yml
@@ -70,6 +70,9 @@ runs:
        name: compatibility-snapshot-${{ inputs.build_type }}-pg${{ inputs.pg_version }}
        path: /tmp/compatibility_snapshot_pg${{ inputs.pg_version }}
        prefix: latest
+        # The lack of compatibility snapshot (for example, for the new Postgres version)
+        # shouldn't fail the whole job. Only relevant test should fail.
+        skip-if-does-not-exist: true

    - name: Checkout
      if: inputs.needs_postgres_source == 'true'
@@ -145,7 +148,11 @@ runs:

        if [ "${RERUN_FLAKY}" == "true" ]; then
          mkdir -p $TEST_OUTPUT
-          poetry run ./scripts/flaky_tests.py "${TEST_RESULT_CONNSTR}" --days 10 --output "$TEST_OUTPUT/flaky.json"
+          poetry run ./scripts/flaky_tests.py "${TEST_RESULT_CONNSTR}" \
+                                              --days 7 \
+                                              --output "$TEST_OUTPUT/flaky.json" \
+                                              --pg-version "${DEFAULT_PG_VERSION}" \
+                                              --build-type "${BUILD_TYPE}"

          EXTRA_PARAMS="--flaky-tests-json $TEST_OUTPUT/flaky.json $EXTRA_PARAMS"
        fi
--- a/.github/workflows/actionlint.yml
+++ b/.github/workflows/actionlint.yml
@@ -0,0 +1,31 @@
+name: Lint GitHub Workflows
+
+on:
+  push:
+    branches:
+      - main
+      - release
+    paths:
+      - '.github/workflows/*.ya?ml'
+  pull_request:
+    paths:
+      - '.github/workflows/*.ya?ml'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  actionlint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: reviewdog/action-actionlint@v1
+        env:
+          # SC2046 - Quote this to prevent word splitting. - https://www.shellcheck.net/wiki/SC2046
+          # SC2086 - Double quote to prevent globbing and word splitting. - https://www.shellcheck.net/wiki/SC2086
+          SHELLCHECK_OPTS: --exclude=SC2046,SC2086
+        with:
+          fail_on_error: true
+          filter_mode: nofilter
+          level: error
--- a/.github/workflows/approved-for-ci-run.yml
+++ b/.github/workflows/approved-for-ci-run.yml
@@ -2,7 +2,9 @@ name: Handle `approved-for-ci-run` label
 # This workflow helps to run CI pipeline for PRs made by external contributors (from forks).

 on:
-  pull_request:
+  pull_request_target:
+    branches:
+      - main
    types:
      # Default types that triggers a workflow ([1]):
      # - [1] https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#pull_request
@@ -17,39 +19,83 @@ on:
 env:
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PR_NUMBER: ${{ github.event.pull_request.number }}
+  BRANCH: "ci-run/pr-${{ github.event.pull_request.number }}"
+
+permissions: write-all
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}

 jobs:
  remove-label:
    # Remove `approved-for-ci-run` label if the workflow is triggered by changes in a PR.
    # The PR should be reviewed and labelled manually again.

-    runs-on: [ ubuntu-latest ]
-
    if: |
      contains(fromJSON('["opened", "synchronize", "reopened", "closed"]'), github.event.action) &&
      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')

+    runs-on: ubuntu-latest
+
    steps:
      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"

-  create-branch:
-    # Create a local branch for an `approved-for-ci-run` labelled PR to run CI pipeline in it.
-
-    runs-on: [ ubuntu-latest ]
+  create-or-update-pr-for-ci-run:
+    # Create local PR for an `approved-for-ci-run` labelled PR to run CI pipeline in it.

    if: |
      github.event.action == 'labeled' &&
      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')

+    runs-on: ubuntu-latest
+
    steps:
      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"

      - uses: actions/checkout@v3
        with:
          ref: main
+          token: ${{ secrets.CI_ACCESS_TOKEN }}

      - run: gh pr checkout "${PR_NUMBER}"

-      - run: git checkout -b "ci-run/pr-${PR_NUMBER}"
+      - run: git checkout -b "${BRANCH}"

-      - run: git push --force origin "ci-run/pr-${PR_NUMBER}"
+      - run: git push --force origin "${BRANCH}"
+
+      - name: Create a Pull Request for CI run (if required)
+        env:
+          GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+        run: |
+          cat << EOF > body.md
+            This Pull Request is created automatically to run the CI pipeline for #${PR_NUMBER}
+
+            Please do not alter or merge/close it.
+
+            Feel free to review/comment/discuss the original PR #${PR_NUMBER}.
+          EOF
+
+          ALREADY_CREATED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${HEAD} --base main --json number --jq '.[].number')"
+          if [ -z "${ALREADY_CREATED}" ]; then
+            gh pr --repo "${GITHUB_REPOSITORY}" create --title "CI run for PR #${PR_NUMBER}" \
+                                                       --body-file "body.md" \
+                                                       --head "${BRANCH}" \
+                                                       --base "main" \
+                                                       --draft
+          fi
+
+  cleanup:
+    # Close PRs and delete branchs if the original PR is closed.
+
+    if: |
+      github.event.action == 'closed' &&
+      github.event.pull_request.head.repo.full_name != github.repository
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - run: |
+          CLOSED="$(gh pr --repo ${GITHUB_REPOSITORY} list --head ${HEAD} --json 'closed' --jq '.[].closed')"
+          if [ "${CLOSED}" == "false" ]; then
+            gh pr --repo "${GITHUB_REPOSITORY}" close "${BRANCH}" --delete-branch
+          fi
--- a/.github/workflows/benchmarking.yml
+++ b/.github/workflows/benchmarking.yml
@@ -117,6 +117,7 @@ jobs:
    outputs:
      pgbench-compare-matrix: ${{ steps.pgbench-compare-matrix.outputs.matrix }}
      olap-compare-matrix: ${{ steps.olap-compare-matrix.outputs.matrix }}
+      tpch-compare-matrix: ${{ steps.tpch-compare-matrix.outputs.matrix }}

    steps:
    - name: Generate matrix for pgbench benchmark
@@ -136,11 +137,11 @@ jobs:
        }'

        if [ "$(date +%A)" = "Saturday" ]; then
-          matrix=$(echo $matrix | jq '.include += [{ "platform": "rds-postgres", "db_size": "10gb"},
+          matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres", "db_size": "10gb"},
                                                   { "platform": "rds-aurora",   "db_size": "50gb"}]')
        fi

-        echo "matrix=$(echo $matrix | jq --compact-output '.')" >> $GITHUB_OUTPUT
+        echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT

    - name: Generate matrix for OLAP benchmarks
      id: olap-compare-matrix
@@ -152,11 +153,30 @@ jobs:
        }'

        if [ "$(date +%A)" = "Saturday" ]; then
-          matrix=$(echo $matrix | jq '.include += [{ "platform": "rds-postgres" },
+          matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres" },
                                                   { "platform": "rds-aurora"   }]')
        fi

-        echo "matrix=$(echo $matrix | jq --compact-output '.')" >> $GITHUB_OUTPUT
+        echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT
+
+    - name: Generate matrix for TPC-H benchmarks
+      id: tpch-compare-matrix
+      run: |
+        matrix='{
+          "platform": [
+            "neon-captest-reuse"
+          ],
+          "scale": [
+            "10"
+          ]
+        }'
+
+        if [ "$(date +%A)" = "Saturday" ]; then
+          matrix=$(echo "$matrix" | jq '.include += [{ "platform": "rds-postgres", "scale": "10" },
+                                                   { "platform": "rds-aurora",   "scale": "10" }]')
+        fi
+
+        echo "matrix=$(echo "$matrix" | jq --compact-output '.')" >> $GITHUB_OUTPUT

  pgbench-compare:
    needs: [ generate-matrices ]
@@ -233,7 +253,11 @@ jobs:

        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT

-        psql ${CONNSTR} -c "SELECT version();"
+        QUERY="SELECT version();"
+        if [[ "${PLATFORM}" = "neon"* ]]; then
+          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+        fi
+        psql ${CONNSTR} -c "${QUERY}"

    - name: Benchmark init
      uses: ./.github/actions/run-python-test-set
@@ -358,7 +382,11 @@ jobs:

        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT

-        psql ${CONNSTR} -c "SELECT version();"
+        QUERY="SELECT version();"
+        if [[ "${PLATFORM}" = "neon"* ]]; then
+          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+        fi
+        psql ${CONNSTR} -c "${QUERY}"

    - name: ClickBench benchmark
      uses: ./.github/actions/run-python-test-set
@@ -372,6 +400,7 @@ jobs:
        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
        BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
+        TEST_OLAP_SCALE: 10

    - name: Create Allure report
      if: ${{ !cancelled() }}
@@ -398,7 +427,7 @@ jobs:

    strategy:
      fail-fast: false
-      matrix: ${{ fromJson(needs.generate-matrices.outputs.olap-compare-matrix) }}
+      matrix: ${{ fromJson(needs.generate-matrices.outputs.tpch-compare-matrix) }}

    env:
      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
@@ -407,6 +436,7 @@ jobs:
      BUILD_TYPE: remote
      SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
      PLATFORM: ${{ matrix.platform }}
+      TEST_OLAP_SCALE: ${{ matrix.scale }}

    runs-on: [ self-hosted, us-east-2, x64 ]
    container:
@@ -428,18 +458,17 @@ jobs:
        ${POSTGRES_DISTRIB_DIR}/v${DEFAULT_PG_VERSION}/bin/pgbench --version
        echo "${POSTGRES_DISTRIB_DIR}/v${DEFAULT_PG_VERSION}/bin" >> $GITHUB_PATH

-    - name: Set up Connection String
-      id: set-up-connstr
+    - name: Get Connstring Secret Name
      run: |
        case "${PLATFORM}" in
          neon-captest-reuse)
-            CONNSTR=${{ secrets.BENCHMARK_CAPTEST_TPCH_S10_CONNSTR }}
+            ENV_PLATFORM=CAPTEST_TPCH
            ;;
          rds-aurora)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_AURORA_TPCH_S10_CONNSTR }}
+            ENV_PLATFORM=RDS_AURORA_TPCH
            ;;
          rds-postgres)
-            CONNSTR=${{ secrets.BENCHMARK_RDS_POSTGRES_TPCH_S10_CONNSTR }}
+            ENV_PLATFORM=RDS_AURORA_TPCH
            ;;
          *)
            echo >&2 "Unknown PLATFORM=${PLATFORM}. Allowed only 'neon-captest-reuse', 'rds-aurora', or 'rds-postgres'"
@@ -447,9 +476,21 @@ jobs:
            ;;
        esac

+        CONNSTR_SECRET_NAME="BENCHMARK_${ENV_PLATFORM}_S${TEST_OLAP_SCALE}_CONNSTR"
+        echo "CONNSTR_SECRET_NAME=${CONNSTR_SECRET_NAME}" >> $GITHUB_ENV
+
+    - name: Set up Connection String
+      id: set-up-connstr
+      run: |
+        CONNSTR=${{ secrets[env.CONNSTR_SECRET_NAME] }}
+
        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT

-        psql ${CONNSTR} -c "SELECT version();"
+        QUERY="SELECT version();"
+        if [[ "${PLATFORM}" = "neon"* ]]; then
+          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+        fi
+        psql ${CONNSTR} -c "${QUERY}"

    - name: Run TPC-H benchmark
      uses: ./.github/actions/run-python-test-set
@@ -463,6 +504,7 @@ jobs:
        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
        BENCHMARK_CONNSTR: ${{ steps.set-up-connstr.outputs.connstr }}
+        TEST_OLAP_SCALE: ${{ matrix.scale }}

    - name: Create Allure report
      if: ${{ !cancelled() }}
@@ -534,7 +576,11 @@ jobs:

        echo "connstr=${CONNSTR}" >> $GITHUB_OUTPUT

-        psql ${CONNSTR} -c "SELECT version();"
+        QUERY="SELECT version();"
+        if [[ "${PLATFORM}" = "neon"* ]]; then
+          QUERY="${QUERY} SHOW neon.tenant_id; SHOW neon.timeline_id;"
+        fi
+        psql ${CONNSTR} -c "${QUERY}"

    - name: Run user examples
      uses: ./.github/actions/run-python-test-set
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -5,7 +5,6 @@ on:
    branches:
      - main
      - release
-      - ci-run/pr-*
  pull_request:

 defaults:
@@ -24,7 +23,30 @@ env:
  AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}

 jobs:
+  check-permissions:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Disallow PRs from forks
+      if: |
+        github.event_name == 'pull_request' &&
+        github.event.pull_request.head.repo.full_name != github.repository
+
+      run: |
+        if [ "${{ contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.pull_request.author_association) }}" = "true" ]; then
+          MESSAGE="Please create a PR from a branch of ${GITHUB_REPOSITORY} instead of a fork"
+        else
+          MESSAGE="The PR should be reviewed and labelled with 'approved-for-ci-run' to trigger a CI run"
+        fi
+
+        echo >&2 "We don't run CI for PRs from forks"
+        echo >&2 "${MESSAGE}"
+
+        exit 1
+
+
  tag:
+    needs: [ check-permissions ]
    runs-on: [ self-hosted, gen3, small ]
    container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
    outputs:
@@ -53,6 +75,7 @@ jobs:
        id: build-tag

  check-codestyle-python:
+    needs: [ check-permissions ]
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
@@ -85,6 +108,7 @@ jobs:
        run: poetry run mypy .

  check-codestyle-rust:
+    needs: [ check-permissions ]
    runs-on: [ self-hosted, gen3, large ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
@@ -151,6 +175,7 @@ jobs:
        run: cargo deny check

  build-neon:
+    needs: [ check-permissions ]
    runs-on: [ self-hosted, gen3, large ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
@@ -187,7 +212,7 @@ jobs:
          # Eventually it will be replaced by a regression test https://github.com/neondatabase/neon/pull/4603

          FAILED=false
-          for postgres in postgres-v14 postgres-v15; do
+          for postgres in postgres-v14 postgres-v15 postgres-v16; do
            expected=$(cat vendor/revisions.json | jq --raw-output '."'"${postgres}"'"')
            actual=$(git rev-parse "HEAD:vendor/${postgres}")
            if [ "${expected}" != "${actual}" ]; then
@@ -209,6 +234,10 @@ jobs:
        id: pg_v15_rev
        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v15) >> $GITHUB_OUTPUT

+      - name: Set pg 16 revision for caching
+        id: pg_v16_rev
+        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v16) >> $GITHUB_OUTPUT
+
      # Set some environment variables used by all the steps.
      #
      # CARGO_FLAGS is extra options to pass to "cargo build", "cargo test" etc.
@@ -229,10 +258,12 @@ jobs:
            cov_prefix=""
            CARGO_FLAGS="--locked --release"
          fi
-          echo "cov_prefix=${cov_prefix}" >> $GITHUB_ENV
-          echo "CARGO_FEATURES=${CARGO_FEATURES}" >> $GITHUB_ENV
-          echo "CARGO_FLAGS=${CARGO_FLAGS}" >> $GITHUB_ENV
-          echo "CARGO_HOME=${GITHUB_WORKSPACE}/.cargo" >> $GITHUB_ENV
+          {
+            echo "cov_prefix=${cov_prefix}"
+            echo "CARGO_FEATURES=${CARGO_FEATURES}"
+            echo "CARGO_FLAGS=${CARGO_FLAGS}"
+            echo "CARGO_HOME=${GITHUB_WORKSPACE}/.cargo"
+          } >> $GITHUB_ENV

      # Disabled for now
      # Don't include the ~/.cargo/registry/src directory. It contains just
@@ -267,6 +298,13 @@ jobs:
          path: pg_install/v15
          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}

+      - name: Cache postgres v16 build
+        id: cache_pg_16
+        uses: actions/cache@v3
+        with:
+          path: pg_install/v16
+          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+
      - name: Build postgres v14
        if: steps.cache_pg_14.outputs.cache-hit != 'true'
        run: mold -run make postgres-v14 -j$(nproc)
@@ -275,6 +313,10 @@ jobs:
        if: steps.cache_pg_15.outputs.cache-hit != 'true'
        run: mold -run make postgres-v15 -j$(nproc)

+      - name: Build postgres v16
+        if: steps.cache_pg_16.outputs.cache-hit != 'true'
+        run: mold -run make postgres-v16 -j$(nproc)
+
      - name: Build neon extensions
        run: mold -run make neon-pg-ext -j$(nproc)

@@ -348,17 +390,17 @@ jobs:
        uses: ./.github/actions/save-coverage-data

  regress-tests:
+    needs: [ check-permissions, build-neon ]
    runs-on: [ self-hosted, gen3, large ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      # Default shared memory is 64mb
      options: --init --shm-size=512mb
-    needs: [ build-neon ]
    strategy:
      fail-fast: false
      matrix:
        build_type: [ debug, release ]
-        pg_version: [ v14, v15 ]
+        pg_version: [ v14, v15, v16 ]
    steps:
      - name: Checkout
        uses: actions/checkout@v3
@@ -386,12 +428,12 @@ jobs:
        uses: ./.github/actions/save-coverage-data

  benchmarks:
+    needs: [ check-permissions, build-neon ]
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      # Default shared memory is 64mb
      options: --init --shm-size=512mb
-    needs: [ build-neon ]
    if: github.ref_name == 'main' || contains(github.event.pull_request.labels.*.name, 'run-benchmarks')
    strategy:
      fail-fast: false
@@ -418,12 +460,13 @@ jobs:
      # while coverage is currently collected for the debug ones

  create-test-report:
+    needs: [ check-permissions, regress-tests, coverage-report, benchmarks ]
+    if: ${{ !cancelled() && contains(fromJSON('["skipped", "success"]'), needs.check-permissions.result) }}
+
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init
-    needs: [ regress-tests, benchmarks ]
-    if: ${{ !cancelled() }}

    steps:
      - uses: actions/checkout@v3
@@ -432,6 +475,11 @@ jobs:
        if: ${{ !cancelled() }}
        id: create-allure-report
        uses: ./.github/actions/allure-report-generate
+        with:
+          store-test-results-into-db: true
+        env:
+          REGRESS_TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
+          REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}

      - uses: actions/github-script@v6
        if: ${{ !cancelled() }}
@@ -444,81 +492,40 @@ jobs:
              reportJsonUrl: "${{ steps.create-allure-report.outputs.report-json-url }}",
            }

+            const coverage = {
+              coverageUrl: "${{ needs.coverage-report.outputs.coverage-html }}",
+              summaryJsonUrl: "${{ needs.coverage-report.outputs.coverage-json }}",
+            }
+
            const script = require("./scripts/comment-test-report.js")
            await script({
              github,
              context,
              fetch,
              report,
+              coverage,
            })

-      - name: Store Allure test stat in the DB
-        if: ${{ !cancelled() && steps.create-allure-report.outputs.report-json-url }}
-        env:
-          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
-          REPORT_JSON_URL: ${{ steps.create-allure-report.outputs.report-json-url }}
-          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
-        run: |
-          ./scripts/pysync
-
-          curl --fail --output suites.json "${REPORT_JSON_URL}"
-          export BUILD_TYPE=unified
-          export DATABASE_URL="$TEST_RESULT_CONNSTR"
-
-          poetry run python3 scripts/ingest_regress_test_result.py \
-            --revision ${COMMIT_SHA} \
-            --reference ${GITHUB_REF} \
-            --build-type ${BUILD_TYPE} \
-            --ingest suites.json
-
-      - name: Store Allure test stat in the DB (new)
-        if: ${{ !cancelled() && steps.create-allure-report.outputs.report-json-url }}
-        env:
-          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
-          REPORT_JSON_URL: ${{ steps.create-allure-report.outputs.report-json-url }}
-          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
-          BASE_S3_URL: ${{ steps.create-allure-report.outputs.base-s3-url }}
-        run: |
-          aws s3 cp --only-show-errors --recursive ${BASE_S3_URL}/data/test-cases ./test-cases
-
-          ./scripts/pysync
-
-          export DATABASE_URL="$TEST_RESULT_CONNSTR"
-          poetry run python3 scripts/ingest_regress_test_result-new-format.py \
-            --reference ${GITHUB_REF} \
-            --revision ${COMMIT_SHA} \
-            --run-id ${GITHUB_RUN_ID} \
-            --run-attempt ${GITHUB_RUN_ATTEMPT} \
-            --test-cases-dir ./test-cases
-
  coverage-report:
+    needs: [ check-permissions, regress-tests ]
+
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
      options: --init
-    needs: [ regress-tests ]
    strategy:
      fail-fast: false
      matrix:
        build_type: [ debug ]
+    outputs:
+        coverage-html: ${{ steps.upload-coverage-report-new.outputs.report-url }}
+        coverage-json: ${{ steps.upload-coverage-report-new.outputs.summary-json }}
    steps:
      - name: Checkout
        uses: actions/checkout@v3
        with:
          submodules: true
-          fetch-depth: 1
-
-#      Disabled for now
-#      - name: Restore cargo deps cache
-#        id: cache_cargo
-#        uses: actions/cache@v3
-#        with:
-#          path: |
-#            ~/.cargo/registry/
-#            !~/.cargo/registry/src
-#            ~/.cargo/git/
-#            target/
-#          key: v1-${{ runner.os }}-${{ matrix.build_type }}-cargo-${{ hashFiles('rust-toolchain.toml') }}-${{ hashFiles('Cargo.lock') }}
+          fetch-depth: 0

      - name: Get Neon artifact
        uses: ./.github/actions/download
@@ -561,13 +568,45 @@ jobs:
          REPORT_URL=https://${BUCKET}.s3.amazonaws.com/code-coverage/${COMMIT_SHA}/index.html
          echo "report-url=${REPORT_URL}" >> $GITHUB_OUTPUT

+      - name: Build coverage report NEW
+        id: upload-coverage-report-new
+        env:
+          BUCKET: neon-github-public-dev
+          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+        run: |
+          BASELINE="$(git merge-base HEAD origin/main)"
+          CURRENT="${COMMIT_SHA}"
+
+          cp /tmp/coverage/report/lcov.info ./${CURRENT}.info
+
+          GENHTML_ARGS="--ignore-errors path,unmapped,empty --synthesize-missing --demangle-cpp rustfilt --output-directory lcov-html ${CURRENT}.info"
+
+          # Use differential coverage if the baseline coverage exists.
+          # It can be missing if the coverage repoer wasn't uploaded yet or tests has failed on BASELINE commit.
+          if aws s3 cp --only-show-errors s3://${BUCKET}/code-coverage/${BASELINE}/lcov.info ./${BASELINE}.info; then
+            git diff ${BASELINE} ${CURRENT} -- '*.rs' > baseline-current.diff
+
+            GENHTML_ARGS="--baseline-file ${BASELINE}.info --diff-file baseline-current.diff ${GENHTML_ARGS}"
+          fi
+
+          genhtml ${GENHTML_ARGS}
+
+          aws s3 cp --only-show-errors --recursive ./lcov-html/ s3://${BUCKET}/code-coverage/${COMMIT_SHA}/lcov
+
+          REPORT_URL=https://${BUCKET}.s3.amazonaws.com/code-coverage/${COMMIT_SHA}/lcov/index.html
+          echo "report-url=${REPORT_URL}" >> $GITHUB_OUTPUT
+
+          REPORT_URL=https://${BUCKET}.s3.amazonaws.com/code-coverage/${COMMIT_SHA}/lcov/summary.json
+          echo "summary-json=${REPORT_URL}" >> $GITHUB_OUTPUT
+
      - uses: actions/github-script@v6
        env:
          REPORT_URL: ${{ steps.upload-coverage-report.outputs.report-url }}
+          REPORT_URL_NEW: ${{ steps.upload-coverage-report-new.outputs.report-url }}
          COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
        with:
          script: |
-            const { REPORT_URL, COMMIT_SHA } = process.env
+            const { REPORT_URL, REPORT_URL_NEW, COMMIT_SHA } = process.env

            await github.rest.repos.createCommitStatus({
              owner: context.repo.owner,
@@ -578,12 +617,21 @@ jobs:
              context: 'Code coverage report',
            })

+            await github.rest.repos.createCommitStatus({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              sha: `${COMMIT_SHA}`,
+              state: 'success',
+              target_url: `${REPORT_URL_NEW}`,
+              context: 'Code coverage report NEW',
+            })
+
  trigger-e2e-tests:
+    needs: [ check-permissions, promote-images, tag ]
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
      options: --init
-    needs: [ promote-images, tag ]
    steps:
      - name: Set PR's status to pending and request a remote CI test
        run: |
@@ -624,8 +672,8 @@ jobs:
            }"

  neon-image:
+    needs: [ check-permissions, tag ]
    runs-on: [ self-hosted, gen3, large ]
-    needs: [ tag ]
    container: gcr.io/kaniko-project/executor:v1.9.2-debug
    defaults:
      run:
@@ -672,7 +720,7 @@ jobs:

  compute-tools-image:
    runs-on: [ self-hosted, gen3, large ]
-    needs: [ tag ]
+    needs: [ check-permissions, tag ]
    container: gcr.io/kaniko-project/executor:v1.9.2-debug
    defaults:
      run:
@@ -717,17 +765,17 @@ jobs:
        run: rm -rf ~/.ecr

  compute-node-image:
+    needs: [ check-permissions, tag ]
    runs-on: [ self-hosted, gen3, large ]
    container:
      image: gcr.io/kaniko-project/executor:v1.9.2-debug
      # Workaround for "Resolving download.osgeo.org (download.osgeo.org)... failed: Temporary failure in name resolution.""
      # Should be prevented by https://github.com/neondatabase/neon/issues/4281
      options: --add-host=download.osgeo.org:140.211.15.30
-    needs: [ tag ]
    strategy:
      fail-fast: false
      matrix:
-        version: [ v14, v15 ]
+        version: [ v14, v15, v16 ]
    defaults:
      run:
        shell: sh -eu {0}
@@ -771,50 +819,22 @@ jobs:
                           --destination neondatabase/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
                           --cleanup

-      # Due to a kaniko bug, we can't use cache for extensions image, thus it takes about the same amount of time as compute-node image to build (~10 min)
-      # During the transition period we need to have extensions in both places (in S3 and in compute-node image),
-      # so we won't build extension twice, but extract them from compute-node.
-      #
-      # For now we use extensions image only for new custom extensitons
-      - name: Kaniko build extensions only
-        run: |
-          # Kaniko is suposed to clean up after itself if --cleanup flag is set, but it doesn't.
-          # Despite some fixes were made in https://github.com/GoogleContainerTools/kaniko/pull/2504 (in kaniko v1.11.0),
-          # it still fails with error:
-          #   error building image: could not save file: copying file: symlink postgres /kaniko/1/usr/local/pgsql/bin/postmaster: file exists
-          #
-          # Ref https://github.com/GoogleContainerTools/kaniko/issues/1406
-          find /kaniko -maxdepth 1 -mindepth 1 -type d -regex "/kaniko/[0-9]*" -exec rm -rv {} \;
-
-          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true \
-                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache \
-                           --context . \
-                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }} \
-                           --build-arg PG_VERSION=${{ matrix.version }} \
-                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}} \
-                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com \
-                           --dockerfile Dockerfile.compute-node \
-                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
-                           --destination neondatabase/extensions-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
-                           --cleanup \
-                           --target postgres-extensions
-
      # Cleanup script fails otherwise - rm: cannot remove '/nvme/actions-runner/_work/_temp/_github_home/.ecr': Permission denied
      - name: Cleanup ECR folder
        run: rm -rf ~/.ecr

  vm-compute-node-image:
+    needs: [ check-permissions, tag, compute-node-image ]
    runs-on: [ self-hosted, gen3, large ]
-    needs: [ tag, compute-node-image ]
    strategy:
      fail-fast: false
      matrix:
-        version: [ v14, v15 ]
+        version: [ v14, v15, v16 ]
    defaults:
      run:
        shell: sh -eu {0}
    env:
-      VM_BUILDER_VERSION: v0.15.4
+      VM_BUILDER_VERSION: v0.17.5

    steps:
      - name: Checkout
@@ -835,14 +855,18 @@ jobs:

      - name: Build vm image
        run: |
-          ./vm-builder -enable-file-cache -src=369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} -dst=369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
+          ./vm-builder \
+            -enable-file-cache \
+            -cgroup-uid=postgres \
+            -src=369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
+            -dst=369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}

      - name: Pushing vm-compute-node image
        run: |
          docker push 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}

  test-images:
-    needs: [ tag, neon-image, compute-node-image, compute-tools-image ]
+    needs: [ check-permissions, tag, neon-image, compute-node-image, compute-tools-image ]
    runs-on: [ self-hosted, gen3, small ]

    steps:
@@ -885,8 +909,8 @@ jobs:
          docker compose -f ./docker-compose/docker-compose.yml down

  promote-images:
+    needs: [ check-permissions, tag, test-images, vm-compute-node-image ]
    runs-on: [ self-hosted, gen3, small ]
-    needs: [ tag, test-images, vm-compute-node-image ]
    container: golang:1.19-bullseye
    # Don't add if-condition here.
    # The job should always be run because we have dependant other jobs that shouldn't be skipped
@@ -906,6 +930,7 @@ jobs:
        run: |
          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} vm-compute-node-v14
          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} vm-compute-node-v15
+          crane pull 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} vm-compute-node-v16

      - name: Add latest tag to images
        if: |
@@ -916,10 +941,10 @@ jobs:
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:${{needs.tag.outputs.build-tag}} latest
+          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:${{needs.tag.outputs.build-tag}} latest
+          crane tag 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} latest

      - name: Push images to production ECR
        if: |
@@ -930,10 +955,10 @@ jobs:
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-tools:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v14:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v14:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-v14:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v15:latest
          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v15:latest
-          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/extensions-v15:latest
+          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/compute-node-v16:latest
+          crane copy 369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} 093970136003.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-v16:latest

      - name: Configure Docker Hub login
        run: |
@@ -945,6 +970,7 @@ jobs:
        run: |
          crane push vm-compute-node-v14 neondatabase/vm-compute-node-v14:${{needs.tag.outputs.build-tag}}
          crane push vm-compute-node-v15 neondatabase/vm-compute-node-v15:${{needs.tag.outputs.build-tag}}
+          crane push vm-compute-node-v16 neondatabase/vm-compute-node-v16:${{needs.tag.outputs.build-tag}}

      - name: Push latest tags to Docker Hub
        if: |
@@ -955,66 +981,94 @@ jobs:
          crane tag neondatabase/compute-tools:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/compute-node-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/vm-compute-node-v14:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/extensions-v14:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/compute-node-v15:${{needs.tag.outputs.build-tag}} latest
          crane tag neondatabase/vm-compute-node-v15:${{needs.tag.outputs.build-tag}} latest
-          crane tag neondatabase/extensions-v15:${{needs.tag.outputs.build-tag}} latest
+          crane tag neondatabase/compute-node-v16:${{needs.tag.outputs.build-tag}} latest
+          crane tag neondatabase/vm-compute-node-v16:${{needs.tag.outputs.build-tag}} latest

      - name: Cleanup ECR folder
        run: rm -rf ~/.ecr

-  upload-postgres-extensions-to-s3:
-    if: |
-      (github.ref_name == 'main' || github.ref_name == 'release') &&
-       github.event_name != 'workflow_dispatch'
-    runs-on: ${{ github.ref_name == 'release' && fromJSON('["self-hosted", "prod", "x64"]') || fromJSON('["self-hosted", "gen3", "small"]') }}
-    needs: [ tag, promote-images ]
-    strategy:
-      fail-fast: false
-      matrix:
-        version: [ v14, v15 ]
-
-    env:
-      EXTENSIONS_IMAGE: ${{ github.ref_name == 'release' && '093970136003' || '369495373322'}}.dkr.ecr.eu-central-1.amazonaws.com/extensions-${{ matrix.version }}:latest
-      AWS_ACCESS_KEY_ID: ${{ github.ref_name == 'release' && secrets.AWS_ACCESS_KEY_PROD || secrets.AWS_ACCESS_KEY_DEV }}
-      AWS_SECRET_ACCESS_KEY: ${{ github.ref_name == 'release' && secrets.AWS_SECRET_KEY_PROD || secrets.AWS_SECRET_KEY_DEV }}
-      S3_BUCKETS: ${{ github.ref_name == 'release' && vars.S3_EXTENSIONS_BUCKETS_PROD || vars.S3_EXTENSIONS_BUCKETS_DEV }}
-
+  trigger-custom-extensions-build-and-wait:
+    needs: [ check-permissions, tag ]
+    runs-on: ubuntu-latest
    steps:
-      - name: Pull postgres-extensions image
+      - name: Set PR's status to pending and request a remote CI test
        run: |
-          docker pull ${EXTENSIONS_IMAGE}
+          COMMIT_SHA=${{ github.event.pull_request.head.sha || github.sha }}
+          REMOTE_REPO="${{ github.repository_owner }}/build-custom-extensions"

-      - name: Create postgres-extensions container
-        id: create-container
+          curl -f -X POST \
+          https://api.github.com/repos/${{ github.repository }}/statuses/$COMMIT_SHA \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"state\": \"pending\",
+              \"context\": \"build-and-upload-extensions\",
+              \"description\": \"[$REMOTE_REPO] Remote CI job is about to start\"
+            }"
+
+          curl -f -X POST \
+          https://api.github.com/repos/$REMOTE_REPO/actions/workflows/build_and_upload_extensions.yml/dispatches \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"ref\": \"main\",
+              \"inputs\": {
+                \"ci_job_name\": \"build-and-upload-extensions\",
+                \"commit_hash\": \"$COMMIT_SHA\",
+                \"remote_repo\": \"${{ github.repository }}\",
+                \"compute_image_tag\": \"${{ needs.tag.outputs.build-tag }}\",
+                \"remote_branch_name\": \"${{ github.ref_name }}\"
+              }
+            }"
+
+      - name: Wait for extension build to finish
+        env:
+          GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
        run: |
-          EID=$(docker create ${EXTENSIONS_IMAGE} true)
-          echo "EID=${EID}" >> $GITHUB_OUTPUT
+          TIMEOUT=1800 # 30 minutes, usually it takes ~2-3 minutes, but if runners are busy, it might take longer
+          INTERVAL=15 # try each N seconds

-      - name: Extract postgres-extensions from container
-        run: |
-          rm -rf ./extensions-to-upload # Just in case
-          mkdir -p extensions-to-upload
+          last_status="" # a variable to carry the last status of the "build-and-upload-extensions" context

-          docker cp ${{ steps.create-container.outputs.EID }}:/extensions/ ./extensions-to-upload/
-          docker cp ${{ steps.create-container.outputs.EID }}:/ext_index.json ./extensions-to-upload/
+          for ((i=0; i <= TIMEOUT; i+=INTERVAL)); do
+            sleep $INTERVAL

-      - name: Upload postgres-extensions to S3
-        run: |
-          for BUCKET in $(echo ${S3_BUCKETS:-[]} | jq --raw-output '.[]'); do
-            aws s3 cp --recursive --only-show-errors ./extensions-to-upload s3://${BUCKET}/${{ needs.tag.outputs.build-tag }}/${{ matrix.version }}
+            # Get statuses for the latest commit in the PR / branch
+            gh api \
+              -H "Accept: application/vnd.github+json" \
+              -H "X-GitHub-Api-Version: 2022-11-28" \
+              "/repos/${{ github.repository }}/statuses/${{ github.event.pull_request.head.sha || github.sha }}" > statuses.json
+
+            # Get the latest status for the "build-and-upload-extensions" context
+            last_status=$(jq --raw-output '[.[] | select(.context == "build-and-upload-extensions")] | sort_by(.created_at)[-1].state' statuses.json)
+            if [ "${last_status}" = "pending" ]; then
+              # Extension build is still in progress.
+              continue
+            elif [ "${last_status}" = "success" ]; then
+              # Extension build is successful.
+              exit 0
+            else
+              # Status is neither "pending" nor "success", exit the loop and fail the job.
+              break
+            fi
          done

-      - name: Cleanup
-        if: ${{ always() && steps.create-container.outputs.EID }}
-        run: |
-          docker rm ${{ steps.create-container.outputs.EID }} || true
+          # Extension build failed, print `statuses.json` for debugging and fail the job.
+          jq '.' statuses.json
+
+          echo >&2 "Status of extension build is '${last_status}' != 'success'"
+          exit 1

  deploy:
+    needs: [ check-permissions, promote-images, tag, regress-tests, trigger-custom-extensions-build-and-wait ]
+    if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
+
    runs-on: [ self-hosted, gen3, small ]
    container: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/ansible:latest
-    needs: [ upload-postgres-extensions-to-s3, promote-images, tag, regress-tests ]
-    if: ( github.ref_name == 'main' || github.ref_name == 'release' ) && github.event_name != 'workflow_dispatch'
    steps:
      - name: Fix git ownership
        run: |
@@ -1052,20 +1106,35 @@ jobs:
          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
          retries: 5
          script: |
-            github.rest.git.createRef({
+            await github.rest.git.createRef({
              owner: context.repo.owner,
              repo: context.repo.repo,
              ref: "refs/tags/${{ needs.tag.outputs.build-tag }}",
              sha: context.sha,
            })

+      - name: Create GitHub release
+        if: github.ref_name == 'release'
+        uses: actions/github-script@v6
+        with:
+          # Retry script for 5XX server errors: https://github.com/actions/github-script#retries
+          retries: 5
+          script: |
+            await github.rest.repos.createRelease({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              tag_name: "${{ needs.tag.outputs.build-tag }}",
+              generate_release_notes: true,
+            })
+
  promote-compatibility-data:
+    needs: [ check-permissions, promote-images, tag, regress-tests ]
+    if: github.ref_name == 'release'
+
    runs-on: [ self-hosted, gen3, small ]
    container:
      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
      options: --init
-    needs: [ promote-images, tag, regress-tests ]
-    if: github.ref_name == 'release' && github.event_name != 'workflow_dispatch'
    steps:
      - name: Promote compatibility snapshot for the release
        env:
@@ -1073,7 +1142,7 @@ jobs:
          PREFIX: artifacts/latest
        run: |
          # Update compatibility snapshot for the release
-          for pg_version in v14 v15; do
+          for pg_version in v14 v15 v16; do
            for build_type in debug release; do
              OLD_FILENAME=compatibility-snapshot-${build_type}-pg${pg_version}-${GITHUB_RUN_ID}.tar.zst
              NEW_FILENAME=compatibility-snapshot-${build_type}-pg${pg_version}.tar.zst
--- a/.github/workflows/neon_extra_builds.yml
+++ b/.github/workflows/neon_extra_builds.yml
@@ -4,7 +4,6 @@ on:
  push:
    branches:
      - main
-      - ci-run/pr-*
  pull_request:

 defaults:
@@ -39,7 +38,7 @@ jobs:
          fetch-depth: 1

      - name: Install macOS postgres dependencies
-        run: brew install flex bison openssl protobuf
+        run: brew install flex bison openssl protobuf icu4c pkg-config

      - name: Set pg 14 revision for caching
        id: pg_v14_rev
@@ -49,6 +48,10 @@ jobs:
        id: pg_v15_rev
        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v15) >> $GITHUB_OUTPUT

+      - name: Set pg 16 revision for caching
+        id: pg_v16_rev
+        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v16) >> $GITHUB_OUTPUT
+
      - name: Cache postgres v14 build
        id: cache_pg_14
        uses: actions/cache@v3
@@ -63,6 +66,13 @@ jobs:
          path: pg_install/v15
          key: v1-${{ runner.os }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}

+      - name: Cache postgres v16 build
+        id: cache_pg_16
+        uses: actions/cache@v3
+        with:
+          path: pg_install/v16
+          key: v1-${{ runner.os }}-${{ env.BUILD_TYPE }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+
      - name: Set extra env for macOS
        run: |
          echo 'LDFLAGS=-L/usr/local/opt/openssl@3/lib' >> $GITHUB_ENV
@@ -86,6 +96,10 @@ jobs:
        if: steps.cache_pg_15.outputs.cache-hit != 'true'
        run: make postgres-v15 -j$(nproc)

+      - name: Build postgres v16
+        if: steps.cache_pg_16.outputs.cache-hit != 'true'
+        run: make postgres-v16 -j$(nproc)
+
      - name: Build neon extensions
        run: make neon-pg-ext -j$(nproc)

--- a/.github/workflows/release-notify.yml
+++ b/.github/workflows/release-notify.yml
@@ -0,0 +1,29 @@
+name: Notify Slack channel about upcoming release
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.number }}
+  cancel-in-progress: true
+
+on:
+  pull_request:
+    branches:
+      - release
+    types:
+      # Default types that triggers a workflow:
+      # - https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#pull_request
+      - opened
+      - synchronize
+      - reopened
+      # Additional types that we want to handle:
+      - closed
+
+jobs:
+  notify:
+    runs-on: [ ubuntu-latest ]
+
+    steps:
+      - uses: neondatabase/dev-actions/release-pr-notify@main
+        with:
+          slack-token: ${{ secrets.SLACK_BOT_TOKEN }}
+          slack-channel-id: ${{ vars.SLACK_UPCOMING_RELEASE_CHANNEL_ID || 'C05QQ9J1BRC' }} # if not set, then `#test-release-notifications`
+          github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -2,16 +2,19 @@ name: Create Release Branch

 on:
  schedule:
-    - cron: '0 10 * * 2'
+    - cron: '0 7 * * 2'
  workflow_dispatch:

 jobs:
  create_release_branch:
-    runs-on: [ubuntu-latest]
+    runs-on: [ ubuntu-latest ]
+
+    permissions:
+      contents: write # for `git push`

    steps:
    - name: Check out code
-      uses: actions/checkout@v3
+      uses: actions/checkout@v4
      with:
        ref: main

@@ -26,9 +29,16 @@ jobs:
      run: git push origin releases/${{ steps.date.outputs.date }}

    - name: Create pull request into release
-      uses: thomaseizinger/create-pull-request@e3972219c86a56550fb70708d96800d8e24ba862 # 1.3.0
-      with:
-        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        head: releases/${{ steps.date.outputs.date }}
-        base: release
-        title: Release ${{ steps.date.outputs.date }}
+      env:
+        GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+      run: |
+        cat << EOF > body.md
+          ## Release ${{ steps.date.outputs.date }}
+
+          **Please merge this PR using 'Create a merge commit'!**
+        EOF
+
+        gh pr create --title "Release ${{ steps.date.outputs.date }}" \
+                     --body-file "body.md" \
+                     --head "releases/${{ steps.date.outputs.date }}" \
+                     --base "release"
--- a/.gitmodules
+++ b/.gitmodules
@@ -6,3 +6,7 @@
 	path = vendor/postgres-v15
 	url = https://github.com/neondatabase/postgres.git
 	branch = REL_15_STABLE_neon
+[submodule "vendor/postgres-v16"]
+	path = vendor/postgres-v16
+	url = https://github.com/neondatabase/postgres.git
+	branch = REL_16_STABLE_neon
--- a/13
+++ b/13
@@ -1,11 +1,12 @@
-/compute_tools/ @neondatabase/control-plane
+/compute_tools/ @neondatabase/control-plane @neondatabase/compute
 /control_plane/ @neondatabase/compute @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/compute @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute 
-/libs/remote_storage/ @neondatabase/storage 
-/libs/safekeeper_api/ @neondatabase/safekeepers  
-/pageserver/ @neondatabase/compute @neondatabase/storage 
+/libs/postgres_ffi/ @neondatabase/compute
+/libs/remote_storage/ @neondatabase/storage
+/libs/safekeeper_api/ @neondatabase/safekeepers
+/libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
+/pageserver/ @neondatabase/compute @neondatabase/storage
 /pgxn/ @neondatabase/compute
-/proxy/ @neondatabase/control-plane 
+/proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
 /vendor/ @neondatabase/compute
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -27,3 +27,28 @@ your patch's fault. Help to fix the root cause if something else has
 broken the CI, before pushing.

 *Happy Hacking!*
+
+# How to run a CI pipeline on Pull Requests from external contributors
+_An instruction for maintainers_
+
+## TL;DR:
+- Review the PR
+- If and only if it looks **safe** (i.e. it doesn't contain any malicious code which could expose secrets or harm the CI), then:
+    - Press the "Approve and run" button in GitHub UI
+    - Add the `approved-for-ci-run` label to the PR
+
+Repeat all steps after any change to the PR.
+- When the changes are ready to get merged — merge the original PR (not the internal one)
+
+## Longer version:
+
+GitHub Actions triggered by the `pull_request` event don't share repository secrets with the forks (for security reasons).
+So, passing the CI pipeline on Pull Requests from external contributors is impossible.
+
+We're using the following approach to make it work:
+- After the review, assign the `approved-for-ci-run` label to the PR if changes look safe
+- A GitHub Action will create an internal branch and a new PR with the same changes (for example, for a PR `#1234`, it'll be a branch `ci-run/pr-1234`)
+- Because the PR is created from the internal branch, it is able to access repository secrets (that's why it's crucial to make sure that the PR doesn't contain any malicious code that could expose our secrets or intentionally harm the CI)
+- The label gets removed automatically, so to run CI again with new changes, the label should be added again (after the review)
+
+For details see [`approved-for-ci-run.yml`](.github/workflows/approved-for-ci-run.yml)
--- a/Cargo.lock
+++ b/Cargo.lock
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,4 +1,5 @@
 [workspace]
+resolver = "2"
 members = [
    "compute_tools",
    "control_plane",
@@ -7,6 +8,7 @@ members = [
    "proxy",
    "safekeeper",
    "storage_broker",
+    "s3_scrubber",
    "workspace_hack",
    "trace",
    "libs/compute_api",
@@ -23,6 +25,7 @@ members = [
    "libs/remote_storage",
    "libs/tracing-utils",
    "libs/postgres_ffi/wal_craft",
+    "libs/vm_monitor",
 ]

 [workspace.package]
@@ -36,17 +39,19 @@ async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
-aws-config = { version = "0.55", default-features = false, features=["rustls"] }
-aws-sdk-s3 = "0.27"
-aws-smithy-http = "0.55"
-aws-credential-types = "0.55"
-aws-types = "0.55"
+aws-config = { version = "0.56", default-features = false, features=["rustls"] }
+aws-sdk-s3 = "0.29"
+aws-smithy-http = "0.56"
+aws-credential-types = "0.56"
+aws-types = "0.56"
+axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
 bindgen = "0.65"
 bstr = "1.0"
 byteorder = "1.4"
 bytes = "1.0"
+cfg-if = "1.0.0"
 chrono = { version = "0.4", default-features = false, features = ["clock"] }
 clap = { version = "4.0", features = ["derive"] }
 close_fds = "0.3.2"
@@ -54,6 +59,7 @@ comfy-table = "6.1"
 const_format = "0.2"
 crc32c = "0.6"
 crossbeam-utils = "0.8.5"
+dashmap = "5.5.0"
 either = "1.8"
 enum-map = "2.4.2"
 enumset = "1.0.12"
@@ -73,6 +79,7 @@ humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
 hyper-tungstenite = "0.9"
+inotify = "0.10.2"
 itertools = "0.10"
 jsonwebtoken = "8"
 libc = "0.2"
@@ -88,7 +95,7 @@ opentelemetry = "0.19.0"
 opentelemetry-otlp = { version = "0.12.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
 opentelemetry-semantic-conventions = "0.11.0"
 parking_lot = "0.12"
-pbkdf2 = "0.12.1"
+pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pin-project-lite = "0.2"
 prometheus = {version = "0.13", default_features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.11"
@@ -100,16 +107,18 @@ reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
-rustls = "0.20"
+rustls = "0.21"
 rustls-pemfile = "1"
 rustls-split = "0.3"
 scopeguard = "1.1"
-sentry = { version = "0.30", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
+sysinfo = "0.29.2"
+sentry = { version = "0.31", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_with = "2.0"
 sha2 = "0.10.2"
 signal-hook = "0.3"
+smallvec = "1.11"
 socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
@@ -118,11 +127,11 @@ sync_wrapper = "0.1.2"
 tar = "0.4"
 test-context = "0.1"
 thiserror = "1.0"
-tls-listener = { version = "0.6", features = ["rustls", "hyper-h1"] }
+tls-listener = { version = "0.7", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }
 tokio-io-timeout = "1.2.0"
-tokio-postgres-rustls = "0.9.0"
-tokio-rustls = "0.23"
+tokio-postgres-rustls = "0.10.0"
+tokio-rustls = "0.24"
 tokio-stream = "0.1"
 tokio-tar = "0.3"
 tokio-util = { version = "0.7", features = ["io"] }
@@ -132,11 +141,11 @@ tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tracing = "0.1"
 tracing-error = "0.2.0"
 tracing-opentelemetry = "0.19.0"
-tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter"] }
+tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 url = "2.2"
 uuid = { version = "1.2", features = ["v4", "serde"] }
 walkdir = "2.3.2"
-webpki-roots = "0.23"
+webpki-roots = "0.25"
 x509-parser = "0.15"

 ## TODO replace this with tracing
@@ -168,14 +177,15 @@ storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main br
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
+vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }

 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }

 ## Build dependencies
 criterion = "0.5.1"
-rcgen = "0.10"
-rstest = "0.17"
+rcgen = "0.11"
+rstest = "0.18"
 tempfile = "3.4"
 tonic-build = "0.9"

--- a/4
+++ b/4
@@ -12,6 +12,7 @@ WORKDIR /home/nonroot

 COPY --chown=nonroot vendor/postgres-v14 vendor/postgres-v14
 COPY --chown=nonroot vendor/postgres-v15 vendor/postgres-v15
+COPY --chown=nonroot vendor/postgres-v16 vendor/postgres-v16
 COPY --chown=nonroot pgxn pgxn
 COPY --chown=nonroot Makefile Makefile
 COPY --chown=nonroot scripts/ninstall.sh scripts/ninstall.sh
@@ -39,6 +40,7 @@ ARG CACHEPOT_BUCKET=neon-github-dev

 COPY --from=pg-build /home/nonroot/pg_install/v14/include/postgresql/server pg_install/v14/include/postgresql/server
 COPY --from=pg-build /home/nonroot/pg_install/v15/include/postgresql/server pg_install/v15/include/postgresql/server
+COPY --from=pg-build /home/nonroot/pg_install/v16/include/postgresql/server pg_install/v16/include/postgresql/server
 COPY --chown=nonroot . .

 # Show build caching stats to check if it was used in the end.
@@ -65,6 +67,7 @@ RUN set -e \
    && apt install -y \
        libreadline-dev \
        libseccomp-dev \
+        libicu67 \
        openssl \
        ca-certificates \
    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* \
@@ -81,6 +84,7 @@ COPY --from=build --chown=neon:neon /home/nonroot/target/release/neon_local

 COPY --from=pg-build /home/nonroot/pg_install/v14 /usr/local/v14/
 COPY --from=pg-build /home/nonroot/pg_install/v15 /usr/local/v15/
+COPY --from=pg-build /home/nonroot/pg_install/v16 /usr/local/v16/
 COPY --from=pg-build /home/nonroot/postgres_install.tar.gz /data/

 # By default, pageserver uses `.neon/` working directory in WORKDIR, so create one and fill it with the dummy config.
--- a/Dockerfile.compute-node
+++ b/Dockerfile.compute-node
@@ -74,8 +74,8 @@ RUN wget https://gitlab.com/Oslandia/SFCGAL/-/archive/v1.3.10/SFCGAL-v1.3.10.tar

 ENV PATH "/usr/local/pgsql/bin:$PATH"

-RUN wget https://download.osgeo.org/postgis/source/postgis-3.3.2.tar.gz -O postgis.tar.gz && \
-    echo "9a2a219da005a1730a39d1959a1c7cec619b1efb009b65be80ffc25bad299068 postgis.tar.gz" | sha256sum --check && \
+RUN wget https://download.osgeo.org/postgis/source/postgis-3.3.3.tar.gz -O postgis.tar.gz && \
+    echo "74eb356e3f85f14233791013360881b6748f78081cc688ff9d6f0f673a762d13 postgis.tar.gz" | sha256sum --check && \
    mkdir postgis-src && cd postgis-src && tar xvzf ../postgis.tar.gz --strip-components=1 -C . && \
    find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /before.txt &&\
    ./autogen.sh && \
@@ -124,8 +124,8 @@ COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 RUN apt update && \
    apt install -y ninja-build python3-dev libncurses5 binutils clang

-RUN wget https://github.com/plv8/plv8/archive/refs/tags/v3.1.5.tar.gz -O plv8.tar.gz && \
-    echo "1e108d5df639e4c189e1c5bdfa2432a521c126ca89e7e5a969d46899ca7bf106 plv8.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/plv8/plv8/archive/refs/tags/v3.1.8.tar.gz -O plv8.tar.gz && \
+    echo "92b10c7db39afdae97ff748c9ec54713826af222c459084ad002571b79eb3f49 plv8.tar.gz" | sha256sum --check && \
    mkdir plv8-src && cd plv8-src && tar xvzf ../plv8.tar.gz --strip-components=1 -C . && \
    export PATH="/usr/local/pgsql/bin:$PATH" && \
    make DOCKER=1 -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -172,8 +172,8 @@ RUN wget https://github.com/uber/h3/archive/refs/tags/v4.1.0.tar.gz -O h3.tar.gz
    cp -R /h3/usr / && \
    rm -rf build

-RUN wget https://github.com/zachasme/h3-pg/archive/refs/tags/v4.1.2.tar.gz -O h3-pg.tar.gz && \
-    echo "c135aa45999b2ad1326d2537c1cadef96d52660838e4ca371706c08fdea1a956 h3-pg.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/zachasme/h3-pg/archive/refs/tags/v4.1.3.tar.gz -O h3-pg.tar.gz && \
+    echo "5c17f09a820859ffe949f847bebf1be98511fb8f1bd86f94932512c00479e324 h3-pg.tar.gz" | sha256sum --check && \
    mkdir h3-pg-src && cd h3-pg-src && tar xvzf ../h3-pg.tar.gz --strip-components=1 -C . && \
    export PATH="/usr/local/pgsql/bin:$PATH" && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
@@ -211,8 +211,8 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.7.tar.gz -
 FROM build-deps AS vector-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.4.4.tar.gz -O pgvector.tar.gz && \
-    echo "1cb70a63f8928e396474796c22a20be9f7285a8a013009deb8152445b61b72e6 pgvector.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.0.tar.gz -O pgvector.tar.gz && \
+    echo "d8aa3504b215467ca528525a6de12c3f85f9891b091ce0e5864dd8a9b757f77b pgvector.tar.gz" | sha256sum --check && \
    mkdir pgvector-src && cd pgvector-src && tar xvzf ../pgvector.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -243,8 +243,8 @@ RUN wget https://github.com/michelp/pgjwt/archive/9742dab1b2f297ad3811120db7b214
 FROM build-deps AS hypopg-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/HypoPG/hypopg/archive/refs/tags/1.3.1.tar.gz -O hypopg.tar.gz && \
-    echo "e7f01ee0259dc1713f318a108f987663d60f3041948c2ada57a94b469565ca8e hypopg.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/HypoPG/hypopg/archive/refs/tags/1.4.0.tar.gz -O hypopg.tar.gz && \
+    echo "0821011743083226fc9b813c1f2ef5897a91901b57b6bea85a78e466187c6819 hypopg.tar.gz" | sha256sum --check && \
    mkdir hypopg-src && cd hypopg-src && tar xvzf ../hypopg.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -307,8 +307,8 @@ RUN wget https://github.com/theory/pgtap/archive/refs/tags/v1.2.0.tar.gz -O pgta
 FROM build-deps AS ip4r-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/RhodiumToad/ip4r/archive/refs/tags/2.4.1.tar.gz -O ip4r.tar.gz && \
-    echo "78b9f0c1ae45c22182768fe892a32d533c82281035e10914111400bf6301c726 ip4r.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/RhodiumToad/ip4r/archive/refs/tags/2.4.2.tar.gz -O ip4r.tar.gz && \
+    echo "0f7b1f159974f49a47842a8ab6751aecca1ed1142b6d5e38d81b064b2ead1b4b ip4r.tar.gz" | sha256sum --check && \
    mkdir ip4r-src && cd ip4r-src && tar xvzf ../ip4r.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -323,8 +323,8 @@ RUN wget https://github.com/RhodiumToad/ip4r/archive/refs/tags/2.4.1.tar.gz -O i
 FROM build-deps AS prefix-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/dimitri/prefix/archive/refs/tags/v1.2.9.tar.gz -O prefix.tar.gz && \
-    echo "38d30a08d0241a8bbb8e1eb8f0152b385051665a8e621c8899e7c5068f8b511e prefix.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/dimitri/prefix/archive/refs/tags/v1.2.10.tar.gz -O prefix.tar.gz && \
+    echo "4342f251432a5f6fb05b8597139d3ccde8dcf87e8ca1498e7ee931ca057a8575 prefix.tar.gz" | sha256sum --check && \
    mkdir prefix-src && cd prefix-src && tar xvzf ../prefix.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -339,8 +339,8 @@ RUN wget https://github.com/dimitri/prefix/archive/refs/tags/v1.2.9.tar.gz -O pr
 FROM build-deps AS hll-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.17.tar.gz -O hll.tar.gz && \
-    echo "9a18288e884f197196b0d29b9f178ba595b0dfc21fbf7a8699380e77fa04c1e9 hll.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.18.tar.gz -O hll.tar.gz && \
+    echo "e2f55a6f4c4ab95ee4f1b4a2b73280258c5136b161fe9d059559556079694f0e hll.tar.gz" | sha256sum --check && \
    mkdir hll-src && cd hll-src && tar xvzf ../hll.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -355,8 +355,8 @@ RUN wget https://github.com/citusdata/postgresql-hll/archive/refs/tags/v2.17.tar
 FROM build-deps AS plpgsql-check-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.3.2.tar.gz -O plpgsql_check.tar.gz && \
-    echo "9d81167c4bbeb74eebf7d60147b21961506161addc2aee537f95ad8efeae427b plpgsql_check.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.4.0.tar.gz -O plpgsql_check.tar.gz && \
+    echo "9ba58387a279b35a3bfa39ee611e5684e6cddb2ba046ddb2c5190b3bd2ca254a plpgsql_check.tar.gz" | sha256sum --check && \
    mkdir plpgsql_check-src && cd plpgsql_check-src && tar xvzf ../plpgsql_check.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
    make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config USE_PGXS=1 && \
@@ -371,12 +371,21 @@ RUN wget https://github.com/okbob/plpgsql_check/archive/refs/tags/v2.3.2.tar.gz
 FROM build-deps AS timescaledb-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

+ARG PG_VERSION
 ENV PATH "/usr/local/pgsql/bin:$PATH"

-RUN apt-get update && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        export TIMESCALEDB_VERSION=2.10.1 \
+        export TIMESCALEDB_CHECKSUM=6fca72a6ed0f6d32d2b3523951ede73dc5f9b0077b38450a029a5f411fdb8c73 \
+        ;; \
+      *) \
+        echo "TimescaleDB not supported on this PostgreSQL version. See https://github.com/timescale/timescaledb/issues/5752" && exit 0;; \
+    esac && \
+    apt-get update && \
    apt-get install -y cmake && \
-    wget https://github.com/timescale/timescaledb/archive/refs/tags/2.10.1.tar.gz -O timescaledb.tar.gz && \
-    echo "6fca72a6ed0f6d32d2b3523951ede73dc5f9b0077b38450a029a5f411fdb8c73 timescaledb.tar.gz" | sha256sum --check && \
+    wget https://github.com/timescale/timescaledb/archive/refs/tags/${TIMESCALEDB_VERSION}.tar.gz -O timescaledb.tar.gz && \
+    echo "${TIMESCALEDB_CHECKSUM} timescaledb.tar.gz" | sha256sum --check && \
    mkdir timescaledb-src && cd timescaledb-src && tar xvzf ../timescaledb.tar.gz --strip-components=1 -C . && \
    ./bootstrap -DSEND_TELEMETRY_DEFAULT:BOOL=OFF -DUSE_TELEMETRY:BOOL=OFF -DAPACHE_ONLY:BOOL=ON -DCMAKE_BUILD_TYPE=Release && \
    cd build && \
@@ -405,6 +414,10 @@ RUN case "${PG_VERSION}" in \
        export PG_HINT_PLAN_VERSION=15_1_5_0 \
        export PG_HINT_PLAN_CHECKSUM=564cbbf4820973ffece63fbf76e3c0af62c4ab23543142c7caaa682bc48918be \
        ;; \
+      "v16") \
+        export PG_HINT_PLAN_VERSION=16_1_6_0 \
+        export PG_HINT_PLAN_CHECKSUM=ce6a8040c78012000f5da7240caf6a971401412f41d33f930f09291e6c304b99 \
+        ;; \
      *) \
        echo "Export the valid PG_HINT_PLAN_VERSION variable" && exit 1 \
        ;; \
@@ -452,8 +465,8 @@ FROM build-deps AS pg-cron-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

 ENV PATH "/usr/local/pgsql/bin/:$PATH"
-RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.5.2.tar.gz -O pg_cron.tar.gz && \
-    echo "6f7f0980c03f1e2a6a747060e67bf4a303ca2a50e941e2c19daeed2b44dec744 pg_cron.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/citusdata/pg_cron/archive/refs/tags/v1.6.0.tar.gz -O pg_cron.tar.gz && \
+    echo "383a627867d730222c272bfd25cd5e151c578d73f696d32910c7db8c665cc7db pg_cron.tar.gz" | sha256sum --check && \
    mkdir pg_cron-src && cd pg_cron-src && tar xvzf ../pg_cron.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -479,8 +492,8 @@ RUN apt-get update && \
        libfreetype6-dev

 ENV PATH "/usr/local/pgsql/bin/:/usr/local/pgsql/:$PATH"
-RUN wget https://github.com/rdkit/rdkit/archive/refs/tags/Release_2023_03_1.tar.gz -O rdkit.tar.gz && \
-    echo "db346afbd0ba52c843926a2a62f8a38c7b774ffab37eaf382d789a824f21996c rdkit.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/rdkit/rdkit/archive/refs/tags/Release_2023_03_3.tar.gz -O rdkit.tar.gz && \
+    echo "bdbf9a2e6988526bfeb8c56ce3cdfe2998d60ac289078e2215374288185e8c8d rdkit.tar.gz" | sha256sum --check && \
    mkdir rdkit-src && cd rdkit-src && tar xvzf ../rdkit.tar.gz --strip-components=1 -C . && \
    cmake \
        -D RDK_BUILD_CAIRO_SUPPORT=OFF \
@@ -551,8 +564,16 @@ FROM build-deps AS pg-embedding-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

 ENV PATH "/usr/local/pgsql/bin/:$PATH"
-RUN wget https://github.com/neondatabase/pg_embedding/archive/refs/tags/0.3.5.tar.gz -O pg_embedding.tar.gz && \
-    echo "0e95b27b8b6196e2cf0a0c9ec143fe2219b82e54c5bb4ee064e76398cbe69ae9 pg_embedding.tar.gz" | sha256sum --check && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        export PG_EMBEDDING_VERSION=0.3.5 \
+        export PG_EMBEDDING_CHECKSUM=0e95b27b8b6196e2cf0a0c9ec143fe2219b82e54c5bb4ee064e76398cbe69ae9 \
+        ;; \
+      *) \
+        echo "pg_embedding not supported on this PostgreSQL version. Use pgvector instead." && exit 0;; \
+    esac && \
+    wget https://github.com/neondatabase/pg_embedding/archive/refs/tags/${PG_EMBEDDING_VERSION}.tar.gz -O pg_embedding.tar.gz && \
+    echo "${PG_EMBEDDING_CHECKSUM} pg_embedding.tar.gz" | sha256sum --check && \
    mkdir pg_embedding-src && cd pg_embedding-src && tar xvzf ../pg_embedding.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -584,6 +605,10 @@ RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgre
 # Layer "rust extensions"
 # This layer is used to build `pgx` deps
 #
+# FIXME: This needs to be updated to latest version of 'pgrx' (it was renamed from
+# 'pgx' to 'pgrx') for PostgreSQL 16. And that in turn requires bumping the pgx
+# dependency on all the rust extension that depend on it, too.
+#
 #########################################################################################
 FROM build-deps AS rust-extensions-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
@@ -598,7 +623,17 @@ USER nonroot
 WORKDIR /home/nonroot
 ARG PG_VERSION

-RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        ;; \
+      "v16") \
+        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
+	;; \
+      *) \
+        echo "unexpected PostgreSQL version ${PG_VERSION}" && exit 1 \
+        ;; \
+    esac && \
+    curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
    chmod +x rustup-init && \
    ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
    rm rustup-init && \
@@ -615,10 +650,21 @@ USER root
 #########################################################################################

 FROM rust-extensions-build AS pg-jsonschema-pg-build
+ARG PG_VERSION

 # caeab60d70b2fd3ae421ec66466a3abbb37b7ee6 made on 06/03/2023
 # there is no release tag yet, but we need it due to the superuser fix in the control file, switch to git tag after release >= 0.1.5
-RUN wget https://github.com/supabase/pg_jsonschema/archive/caeab60d70b2fd3ae421ec66466a3abbb37b7ee6.tar.gz -O pg_jsonschema.tar.gz && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        ;; \
+      "v16") \
+        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
+	;; \
+      *) \
+        echo "unexpected PostgreSQL version \"${PG_VERSION}\"" && exit 1 \
+        ;; \
+    esac && \
+    wget https://github.com/supabase/pg_jsonschema/archive/caeab60d70b2fd3ae421ec66466a3abbb37b7ee6.tar.gz -O pg_jsonschema.tar.gz && \
    echo "54129ce2e7ee7a585648dbb4cef6d73f795d94fe72f248ac01119992518469a4 pg_jsonschema.tar.gz" | sha256sum --check && \
    mkdir pg_jsonschema-src && cd pg_jsonschema-src && tar xvzf ../pg_jsonschema.tar.gz --strip-components=1 -C . && \
    sed -i 's/pgx = "0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
@@ -633,12 +679,23 @@ RUN wget https://github.com/supabase/pg_jsonschema/archive/caeab60d70b2fd3ae421e
 #########################################################################################

 FROM rust-extensions-build AS pg-graphql-pg-build
+ARG PG_VERSION

 # b4988843647450a153439be367168ed09971af85 made on 22/02/2023 (from remove-pgx-contrib-spiext branch)
 # Currently pgx version bump to >= 0.7.2  causes "call to unsafe function" compliation errors in
 # pgx-contrib-spiext. There is a branch that removes that dependency, so use it. It is on the
 # same 1.1 version we've used before.
-RUN wget https://github.com/yrashk/pg_graphql/archive/b4988843647450a153439be367168ed09971af85.tar.gz -O pg_graphql.tar.gz && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        ;; \
+      "v16") \
+        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
+	;; \
+      *) \
+        echo "unexpected PostgreSQL version" && exit 1 \
+        ;; \
+    esac && \
+    wget https://github.com/yrashk/pg_graphql/archive/b4988843647450a153439be367168ed09971af85.tar.gz -O pg_graphql.tar.gz && \
    echo "0c7b0e746441b2ec24187d0e03555faf935c2159e2839bddd14df6dafbc8c9bd pg_graphql.tar.gz" | sha256sum --check && \
    mkdir pg_graphql-src && cd pg_graphql-src && tar xvzf ../pg_graphql.tar.gz --strip-components=1 -C . && \
    sed -i 's/pgx = "~0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
@@ -656,9 +713,20 @@ RUN wget https://github.com/yrashk/pg_graphql/archive/b4988843647450a153439be367
 #########################################################################################

 FROM rust-extensions-build AS pg-tiktoken-pg-build
+ARG PG_VERSION

 # 801f84f08c6881c8aa30f405fafbf00eec386a72 made on 10/03/2023
-RUN wget https://github.com/kelvich/pg_tiktoken/archive/801f84f08c6881c8aa30f405fafbf00eec386a72.tar.gz -O pg_tiktoken.tar.gz && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        ;; \
+      "v16") \
+        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
+	;; \
+      *) \
+        echo "unexpected PostgreSQL version" && exit 1 \
+        ;; \
+    esac && \
+    wget https://github.com/kelvich/pg_tiktoken/archive/801f84f08c6881c8aa30f405fafbf00eec386a72.tar.gz -O pg_tiktoken.tar.gz && \
    echo "52f60ac800993a49aa8c609961842b611b6b1949717b69ce2ec9117117e16e4a pg_tiktoken.tar.gz" | sha256sum --check && \
    mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xvzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \
    cargo pgx install --release && \
@@ -672,8 +740,19 @@ RUN wget https://github.com/kelvich/pg_tiktoken/archive/801f84f08c6881c8aa30f405
 #########################################################################################

 FROM rust-extensions-build AS pg-pgx-ulid-build
+ARG PG_VERSION

-RUN wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.0.tar.gz -O pgx_ulid.tar.gz && \
+RUN case "${PG_VERSION}" in \
+      "v14" | "v15") \
+        ;; \
+      "v16") \
+        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
+	;; \
+      *) \
+        echo "unexpected PostgreSQL version" && exit 1 \
+        ;; \
+    esac && \
+    wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.0.tar.gz -O pgx_ulid.tar.gz && \
    echo "908b7358e6f846e87db508ae5349fb56a88ee6305519074b12f3d5b0ff09f791 pgx_ulid.tar.gz" | sha256sum --check && \
    mkdir pgx_ulid-src && cd pgx_ulid-src && tar xvzf ../pgx_ulid.tar.gz --strip-components=1 -C . && \
    sed -i 's/pgx        = "=0.7.3"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
@@ -726,6 +805,20 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) \
        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
        -C pgxn/neon_utils \
        -s install && \
+    make -j $(getconf _NPROCESSORS_ONLN) \
+        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
+        -C pgxn/neon_rmgr \
+        -s install && \
+    case "${PG_VERSION}" in \
+        "v14" | "v15") \
+        ;; \
+        "v16") \
+            echo "Skipping HNSW for PostgreSQL 16" && exit 0 \
+        ;; \
+        *) \
+            echo "unexpected PostgreSQL version" && exit 1 \
+        ;; \
+        esac && \
    make -j $(getconf _NPROCESSORS_ONLN) \
        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
        -C pgxn/hnsw \
@@ -764,29 +857,6 @@ RUN rm -r /usr/local/pgsql/include
 # if they were to be used by other libraries.
 RUN rm /usr/local/pgsql/lib/lib*.a

-#########################################################################################
-#
-# Extenstion only
-#
-#########################################################################################
-FROM python:3.9-slim-bullseye AS generate-ext-index
-ARG PG_VERSION
-ARG BUILD_TAG
-RUN apt update && apt install -y zstd
-
-# copy the control files here
-COPY --from=kq-imcx-pg-build /extensions/ /extensions/
-COPY --from=pg-anon-pg-build /extensions/ /extensions/
-COPY --from=postgis-build /extensions/ /extensions/
-COPY scripts/combine_control_files.py ./combine_control_files.py
-RUN python3 ./combine_control_files.py ${PG_VERSION} ${BUILD_TAG} --public_extensions="anon,postgis"
-
-FROM scratch AS postgres-extensions
-# After the transition this layer will include all extensitons.
-# As for now, it's only a couple for testing purposses
-COPY --from=generate-ext-index /extensions/*.tar.zst /extensions/
-COPY --from=generate-ext-index /ext_index.json /ext_index.json
-
 #########################################################################################
 #
 # Final layer
--- a/40
+++ b/40
@@ -29,6 +29,7 @@ else ifeq ($(UNAME_S),Darwin)
 	# It can be configured with OPENSSL_PREFIX variable
 	OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3)
 	PG_CONFIGURE_OPTS += --with-includes=$(OPENSSL_PREFIX)/include --with-libraries=$(OPENSSL_PREFIX)/lib
+	PG_CONFIGURE_OPTS += PKG_CONFIG_PATH=$(shell brew --prefix icu4c)/lib/pkgconfig
 	# macOS already has bison and flex in the system, but they are old and result in postgres-v14 target failure
 	# brew formulae are keg-only and not symlinked into HOMEBREW_PREFIX, force their usage
 	EXTRA_PATH_OVERRIDES += $(shell brew --prefix bison)/bin/:$(shell brew --prefix flex)/bin/:
@@ -83,6 +84,8 @@ $(POSTGRES_INSTALL_DIR)/build/%/config.status:
 # I'm not sure why it wouldn't work, but this is the only place (apart from
 # the "build-all-versions" entry points) where direct mention of PostgreSQL
 # versions is used.
+.PHONY: postgres-configure-v16
+postgres-configure-v16: $(POSTGRES_INSTALL_DIR)/build/v16/config.status
 .PHONY: postgres-configure-v15
 postgres-configure-v15: $(POSTGRES_INSTALL_DIR)/build/v15/config.status
 .PHONY: postgres-configure-v14
@@ -118,6 +121,10 @@ postgres-clean-%:
 	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/contrib/pageinspect clean
 	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/src/interfaces/libpq clean

+.PHONY: postgres-check-%
+postgres-check-%: postgres-%
+	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$* MAKELEVEL=0 check
+
 .PHONY: neon-pg-ext-%
 neon-pg-ext-%: postgres-%
 	+@echo "Compiling neon $*"
@@ -130,6 +137,11 @@ neon-pg-ext-%: postgres-%
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-walredo-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon_walredo/Makefile install
+	+@echo "Compiling neon_rmgr $*"
+	mkdir -p $(POSTGRES_INSTALL_DIR)/build/neon-rmgr-$*
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		-C $(POSTGRES_INSTALL_DIR)/build/neon-rmgr-$* \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon_rmgr/Makefile install
 	+@echo "Compiling neon_test_utils $*"
 	mkdir -p $(POSTGRES_INSTALL_DIR)/build/neon-test-utils-$*
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
@@ -140,6 +152,13 @@ neon-pg-ext-%: postgres-%
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile install
+
+# pg_embedding was temporarily released as hnsw from this repo, when we only
+# supported PostgreSQL 14 and 15
+neon-pg-ext-v14: neon-pg-ext-hnsw-v14
+neon-pg-ext-v15: neon-pg-ext-hnsw-v15
+
+neon-pg-ext-hnsw-%: postgres-headers-% postgres-%
 	+@echo "Compiling hnsw $*"
 	mkdir -p $(POSTGRES_INSTALL_DIR)/build/hnsw-$*
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
@@ -167,28 +186,39 @@ neon-pg-ext-clean-%:
 .PHONY: neon-pg-ext
 neon-pg-ext: \
 	neon-pg-ext-v14 \
-	neon-pg-ext-v15
+	neon-pg-ext-v15 \
+	neon-pg-ext-v16

 .PHONY: neon-pg-ext-clean
 neon-pg-ext-clean: \
 	neon-pg-ext-clean-v14 \
-	neon-pg-ext-clean-v15
+	neon-pg-ext-clean-v15 \
+	neon-pg-ext-clean-v16

 # shorthand to build all Postgres versions
 .PHONY: postgres
 postgres: \
 	postgres-v14 \
-	postgres-v15
+	postgres-v15 \
+	postgres-v16

 .PHONY: postgres-headers
 postgres-headers: \
 	postgres-headers-v14 \
-	postgres-headers-v15
+	postgres-headers-v15 \
+	postgres-headers-v16

 .PHONY: postgres-clean
 postgres-clean: \
 	postgres-clean-v14 \
-	postgres-clean-v15
+	postgres-clean-v15 \
+	postgres-clean-v16
+
+.PHONY: postgres-check
+postgres-check: \
+	postgres-check-v14 \
+	postgres-check-v15 \
+	postgres-check-v16

 # This doesn't remove the effects of 'configure'.
 .PHONY: clean
--- a/README.md
+++ b/README.md
@@ -29,18 +29,18 @@ See developer documentation in [SUMMARY.md](/docs/SUMMARY.md) for more informati
 ```bash
 apt install build-essential libtool libreadline-dev zlib1g-dev flex bison libseccomp-dev \
 libssl-dev clang pkg-config libpq-dev cmake postgresql-client protobuf-compiler \
-libcurl4-openssl-dev openssl python-poetry
+libcurl4-openssl-dev openssl python-poetry lsof
 ```
 * On Fedora, these packages are needed:
 ```bash
 dnf install flex bison readline-devel zlib-devel openssl-devel \
  libseccomp-devel perl clang cmake postgresql postgresql-contrib protobuf-compiler \
-  protobuf-devel libcurl-devel openssl poetry
+  protobuf-devel libcurl-devel openssl poetry lsof
 ```
 * On Arch based systems, these packages are needed:
 ```bash
 pacman -S base-devel readline zlib libseccomp openssl clang \
-postgresql-libs cmake postgresql protobuf curl
+postgresql-libs cmake postgresql protobuf curl lsof
 ```

 Building Neon requires 3.15+ version of `protoc` (protobuf-compiler). If your distribution provides an older version, you can install a newer version from [here](https://github.com/protocolbuffers/protobuf/releases).
@@ -55,7 +55,7 @@ curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
 1. Install XCode and dependencies
 ```
 xcode-select --install
-brew install protobuf openssl flex bison
+brew install protobuf openssl flex bison icu4c pkg-config

 # add openssl to PATH, required for ed25519 keys generation in neon_local
 echo 'export PATH="$(brew --prefix openssl)/bin:$PATH"' >> ~/.zshrc
--- a/clippy.toml
+++ b/clippy.toml
@@ -0,0 +1,5 @@
+disallowed-methods = [
+    "tokio::task::block_in_place",
+    # Allow this for now, to deny it later once we stop using Handle::block_on completely
+    # "tokio::runtime::Handle::block_on",
+]
--- a/compute_tools/Cargo.toml
+++ b/compute_tools/Cargo.toml
@@ -8,6 +8,7 @@ license.workspace = true
 anyhow.workspace = true
 async-compression.workspace = true
 chrono.workspace = true
+cfg-if.workspace = true
 clap.workspace = true
 flate2.workspace = true
 futures.workspace = true
@@ -23,6 +24,7 @@ tar.workspace = true
 reqwest = { workspace = true, features = ["json"] }
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tokio-postgres.workspace = true
+tokio-util.workspace = true
 tracing.workspace = true
 tracing-opentelemetry.workspace = true
 tracing-subscriber.workspace = true
@@ -34,4 +36,5 @@ utils.workspace = true
 workspace_hack.workspace = true
 toml_edit.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
+vm_monitor = { version = "0.1", path = "../libs/vm_monitor/" }
 zstd = "0.12.4"
--- a/compute_tools/README.md
+++ b/compute_tools/README.md
@@ -19,9 +19,10 @@ Also `compute_ctl` spawns two separate service threads:
 - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
  last activity requests.

-If the `vm-informant` binary is present at `/bin/vm-informant`, it will also be started. For VM
-compute nodes, `vm-informant` communicates with the VM autoscaling system. It coordinates
-downscaling and (eventually) will request immediate upscaling under resource pressure.
+If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
+`vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
+`vm-monitor` communicates with the VM autoscaling system. It coordinates
+downscaling and requests immediate upscaling under resource pressure.

 Usage example:
 ```sh
--- a/compute_tools/src/bin/compute_ctl.rs
+++ b/compute_tools/src/bin/compute_ctl.rs
@@ -20,9 +20,10 @@
 //! - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
 //!   last activity requests.
 //!
-//! If the `vm-informant` binary is present at `/bin/vm-informant`, it will also be started. For VM
-//! compute nodes, `vm-informant` communicates with the VM autoscaling system. It coordinates
-//! downscaling and (eventually) will request immediate upscaling under resource pressure.
+//! If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
+//! `vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
+//! `vm-monitor` communicates with the VM autoscaling system. It coordinates
+//! downscaling and requests immediate upscaling under resource pressure.
 //!
 //! Usage example:
 //! ```sh
@@ -35,10 +36,9 @@
 //!
 use std::collections::HashMap;
 use std::fs::File;
-use std::panic;
 use std::path::Path;
 use std::process::exit;
-use std::sync::{mpsc, Arc, Condvar, Mutex, OnceLock, RwLock};
+use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
 use std::{thread, time::Duration};

 use anyhow::{Context, Result};
@@ -147,6 +147,7 @@ fn main() -> Result<()> {
    match spec_json {
        // First, try to get cluster spec from the cli argument
        Some(json) => {
+            info!("got spec from cli argument {}", json);
            spec = Some(serde_json::from_str(json)?);
        }
        None => {
@@ -182,6 +183,7 @@ fn main() -> Result<()> {

    if let Some(spec) = spec {
        let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
+        info!("new pspec.spec: {:?}", pspec.spec);
        new_state.pspec = Some(pspec);
        spec_set = true;
    } else {
@@ -196,9 +198,7 @@ fn main() -> Result<()> {
        state: Mutex::new(new_state),
        state_changed: Condvar::new(),
        ext_remote_storage,
-        ext_remote_paths: OnceLock::new(),
        ext_download_progress: RwLock::new(HashMap::new()),
-        library_index: OnceLock::new(),
        build_tag,
    };
    let compute = Arc::new(compute_node);
@@ -271,6 +271,57 @@ fn main() -> Result<()> {
        }
    };

+    // Start the vm-monitor if directed to. The vm-monitor only runs on linux
+    // because it requires cgroups.
+    cfg_if::cfg_if! {
+        if #[cfg(target_os = "linux")] {
+            use std::env;
+            use tokio_util::sync::CancellationToken;
+            use tracing::warn;
+            let vm_monitor_addr = matches.get_one::<String>("vm-monitor-addr");
+            let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
+            let cgroup = matches.get_one::<String>("cgroup");
+            let file_cache_on_disk = matches.get_flag("file-cache-on-disk");
+
+            // Only make a runtime if we need to.
+            // Note: it seems like you can make a runtime in an inner scope and
+            // if you start a task in it it won't be dropped. However, make it
+            // in the outermost scope just to be safe.
+            let rt = match (env::var_os("AUTOSCALING"), vm_monitor_addr) {
+                (None, None) => None,
+                (None, Some(_)) => {
+                    warn!("--vm-monitor-addr option set but AUTOSCALING env var not present");
+                    None
+                }
+                (Some(_), None) => {
+                    panic!("AUTOSCALING env var present but --vm-monitor-addr option not set")
+                }
+                (Some(_), Some(_)) => Some(
+                    tokio::runtime::Builder::new_multi_thread()
+                        .worker_threads(4)
+                        .enable_all()
+                        .build()
+                        .expect("failed to create tokio runtime for monitor"),
+                ),
+            };
+
+            // This token is used internally by the monitor to clean up all threads
+            let token = CancellationToken::new();
+
+            let vm_monitor = &rt.as_ref().map(|rt| {
+                rt.spawn(vm_monitor::start(
+                    Box::leak(Box::new(vm_monitor::Args {
+                        cgroup: cgroup.cloned(),
+                        pgconnstr: file_cache_connstr.cloned(),
+                        addr: vm_monitor_addr.cloned().unwrap(),
+                        file_cache_on_disk,
+                    })),
+                    token.clone(),
+                ))
+            });
+        }
+    }
+
    // Wait for the child Postgres process forever. In this state Ctrl+C will
    // propagate to Postgres and it will be shut down as well.
    if let Some(mut pg) = pg {
@@ -284,6 +335,24 @@ fn main() -> Result<()> {
        exit_code = ecode.code()
    }

+    // Terminate the vm_monitor so it releases the file watcher on
+    // /sys/fs/cgroup/neon-postgres.
+    // Note: the vm-monitor only runs on linux because it requires cgroups.
+    cfg_if::cfg_if! {
+        if #[cfg(target_os = "linux")] {
+            if let Some(handle) = vm_monitor {
+                // Kills all threads spawned by the monitor
+                token.cancel();
+                // Kills the actual task running the monitor
+                handle.abort();
+
+                // If handle is some, rt must have been used to produce it, and
+                // hence is also some
+                rt.unwrap().shutdown_timeout(Duration::from_secs(2));
+            }
+        }
+    }
+
    // Maybe sync safekeepers again, to speed up next startup
    let compute_state = compute.state.lock().unwrap().clone();
    let pspec = compute_state.pspec.as_ref().expect("spec must be set");
@@ -393,6 +462,34 @@ fn cli() -> clap::Command {
                .long("remote-ext-config")
                .value_name("REMOTE_EXT_CONFIG"),
        )
+        // TODO(fprasx): we currently have default arguments because the cloud PR
+        // to pass them in hasn't been merged yet. We should get rid of them once
+        // the PR is merged.
+        .arg(
+            Arg::new("vm-monitor-addr")
+                .long("vm-monitor-addr")
+                .default_value("0.0.0.0:10301")
+                .value_name("VM_MONITOR_ADDR"),
+        )
+        .arg(
+            Arg::new("cgroup")
+                .long("cgroup")
+                .default_value("neon-postgres")
+                .value_name("CGROUP"),
+        )
+        .arg(
+            Arg::new("filecache-connstr")
+                .long("filecache-connstr")
+                .default_value(
+                    "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable",
+                )
+                .value_name("FILECACHE_CONNSTR"),
+        )
+        .arg(
+            Arg::new("file-cache-on-disk")
+                .long("file-cache-on-disk")
+                .action(clap::ArgAction::SetTrue),
+        )
 }

 #[test]
--- a/compute_tools/src/checker.rs
+++ b/compute_tools/src/checker.rs
@@ -1,12 +1,39 @@
-use anyhow::{anyhow, Result};
+use anyhow::{anyhow, Ok, Result};
+use postgres::Client;
 use tokio_postgres::NoTls;
-use tracing::{error, instrument};
+use tracing::{error, instrument, warn};

 use crate::compute::ComputeNode;

+/// Create a special service table for availability checks
+/// only if it does not exist already.
+pub fn create_availability_check_data(client: &mut Client) -> Result<()> {
+    let query = "
+        DO $$
+        BEGIN
+            IF NOT EXISTS(
+                SELECT 1
+                FROM pg_catalog.pg_tables
+                WHERE tablename = 'health_check'
+            )
+            THEN
+            CREATE TABLE health_check (
+                id serial primary key,
+                updated_at timestamptz default now()
+            );
+            INSERT INTO health_check VALUES (1, now())
+                ON CONFLICT (id) DO UPDATE
+                 SET updated_at = now();
+            END IF;
+        END
+        $$;";
+    client.execute(query, &[])?;
+
+    Ok(())
+}
+
 /// Update timestamp in a row in a special service table to check
 /// that we can actually write some data in this particular timeline.
-/// Create table if it's missing.
 #[instrument(skip_all)]
 pub async fn check_writability(compute: &ComputeNode) -> Result<()> {
    // Connect to the database.
@@ -24,21 +51,28 @@ pub async fn check_writability(compute: &ComputeNode) -> Result<()> {
    });

    let query = "
-    CREATE TABLE IF NOT EXISTS health_check (
-        id serial primary key,
-        updated_at timestamptz default now()
-    );
    INSERT INTO health_check VALUES (1, now())
        ON CONFLICT (id) DO UPDATE
         SET updated_at = now();";

-    let result = client.simple_query(query).await?;
-
-    if result.len() != 2 {
-        return Err(anyhow::format_err!(
-            "expected 2 query results, but got {}",
-            result.len()
-        ));
+    match client.simple_query(query).await {
+        Result::Ok(result) => {
+            if result.len() != 1 {
+                return Err(anyhow::anyhow!(
+                    "expected 1 query results, but got {}",
+                    result.len()
+                ));
+            }
+        }
+        Err(err) => {
+            if let Some(state) = err.code() {
+                if state == &tokio_postgres::error::SqlState::DISK_FULL {
+                    warn!("Tenant disk is full");
+                    return Ok(());
+                }
+            }
+            return Err(err.into());
+        }
    }

    Ok(())
--- a/compute_tools/src/compute.rs
+++ b/compute_tools/src/compute.rs
@@ -1,11 +1,12 @@
 use std::collections::HashMap;
+use std::env;
 use std::fs;
 use std::io::BufRead;
 use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
-use std::sync::{Condvar, Mutex, OnceLock, RwLock};
+use std::sync::{Condvar, Mutex, RwLock};
 use std::time::Instant;

 use anyhow::{Context, Result};
@@ -14,7 +15,6 @@ use futures::future::join_all;
 use futures::stream::FuturesUnordered;
 use futures::StreamExt;
 use postgres::{Client, NoTls};
-use regex::Regex;
 use tokio;
 use tokio_postgres;
 use tracing::{error, info, instrument, warn};
@@ -27,6 +27,7 @@ use utils::measured_stream::MeasuredReader;

 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};

+use crate::checker::create_availability_check_data;
 use crate::pg_helpers::*;
 use crate::spec::*;
 use crate::sync_sk::{check_if_synced, ping_safekeeper};
@@ -60,10 +61,6 @@ pub struct ComputeNode {
    pub state_changed: Condvar,
    ///  the S3 bucket that we search for extensions in
    pub ext_remote_storage: Option<GenericRemoteStorage>,
-    // (key: extension name, value: path to extension archive in remote storage)
-    pub ext_remote_paths: OnceLock<HashMap<String, RemotePath>>,
-    // (key: library name, value: name of extension containing this library)
-    pub library_index: OnceLock<HashMap<String, String>>,
    // key: ext_archive_name, value: started download time, download_completed?
    pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
    pub build_tag: String,
@@ -75,7 +72,6 @@ pub struct RemoteExtensionMetrics {
    num_ext_downloaded: u64,
    largest_ext_size: u64,
    total_ext_download_size: u64,
-    prep_extensions_ms: u64,
 }

 #[derive(Clone, Debug)]
@@ -181,6 +177,27 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
    }
 }

+/// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
+/// cgroup. Otherwise returns the default `Command::new(cmd)`
+///
+/// This function should be used to start postgres, as it will start it in the
+/// neon-postgres cgroup if we are a VM. This allows autoscaling to control
+/// postgres' resource usage. The cgroup will exist in VMs because vm-builder
+/// creates it during the sysinit phase of its inittab.
+fn maybe_cgexec(cmd: &str) -> Command {
+    // The cplane sets this env var for autoscaling computes.
+    // use `var_os` so we don't have to worry about the variable being valid
+    // unicode. Should never be an concern . . . but just in case
+    if env::var_os("AUTOSCALING").is_some() {
+        let mut command = Command::new("cgexec");
+        command.args(["-g", "memory:neon-postgres"]);
+        command.arg(cmd);
+        command
+    } else {
+        Command::new(cmd)
+    }
+}
+
 /// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
 /// that we give to customers
 fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
@@ -457,7 +474,7 @@ impl ComputeNode {
    pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
        let start_time = Utc::now();

-        let sync_handle = Command::new(&self.pgbin)
+        let sync_handle = maybe_cgexec(&self.pgbin)
            .args(["--sync-safekeepers"])
            .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
            .envs(if let Some(storage_auth_token) = &storage_auth_token {
@@ -592,7 +609,7 @@ impl ComputeNode {

        // Start postgres
        info!("starting postgres");
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
            .args(["-D", pgdata])
            .spawn()
            .expect("cannot start postgres process");
@@ -620,7 +637,7 @@ impl ComputeNode {
        let pgdata_path = Path::new(&self.pgdata);

        // Run postgres as a child process.
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
            .args(["-D", &self.pgdata])
            .envs(if let Some(storage_auth_token) = &storage_auth_token {
                vec![("NEON_AUTH_TOKEN", storage_auth_token)]
@@ -680,6 +697,7 @@ impl ComputeNode {
        handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
        handle_grants(spec, self.connstr.as_str())?;
        handle_extensions(spec, &mut client)?;
+        create_availability_check_data(&mut client)?;

        // 'Close' connection
        drop(client);
@@ -745,11 +763,19 @@ impl ComputeNode {
            pspec.timeline_id,
        );

+        info!(
+            "start_compute spec.remote_extensions {:?}",
+            pspec.spec.remote_extensions
+        );
+
        // This part is sync, because we need to download
        // remote shared_preload_libraries before postgres start (if any)
-        {
+        if let Some(remote_extensions) = &pspec.spec.remote_extensions {
+            // First, create control files for all availale extensions
+            extension_server::create_control_files(remote_extensions, &self.pgbin);
+
            let library_load_start_time = Utc::now();
-            let remote_ext_metrics = self.prepare_preload_libraries(&compute_state)?;
+            let remote_ext_metrics = self.prepare_preload_libraries(&pspec.spec)?;

            let library_load_time = Utc::now()
                .signed_duration_since(library_load_start_time)
@@ -761,7 +787,6 @@ impl ComputeNode {
            state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
            state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
            state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
-            state.metrics.prep_extensions_ms = remote_ext_metrics.prep_extensions_ms;
            info!(
                "Loading shared_preload_libraries took {:?}ms",
                library_load_time
@@ -918,38 +943,11 @@ LIMIT 100",
        }
    }

-    // If remote extension storage is configured,
-    // download extension control files
-    pub async fn prepare_external_extensions(&self, compute_state: &ComputeState) -> Result<()> {
-        if let Some(ref ext_remote_storage) = self.ext_remote_storage {
-            let pspec = compute_state.pspec.as_ref().expect("spec must be set");
-            let spec = &pspec.spec;
-            let custom_ext = spec.custom_extensions.clone().unwrap_or(Vec::new());
-            info!("custom extensions: {:?}", &custom_ext);
-
-            let (ext_remote_paths, library_index) = extension_server::get_available_extensions(
-                ext_remote_storage,
-                &self.pgbin,
-                &self.pgversion,
-                &custom_ext,
-                &self.build_tag,
-            )
-            .await?;
-            self.ext_remote_paths
-                .set(ext_remote_paths)
-                .expect("this is the only time we set ext_remote_paths");
-            self.library_index
-                .set(library_index)
-                .expect("this is the only time we set library_index");
-        }
-        Ok(())
-    }
-
    // download an archive, unzip and place files in correct locations
    pub async fn download_extension(
        &self,
-        ext_name: &str,
-        is_library: bool,
+        real_ext_name: String,
+        ext_path: RemotePath,
    ) -> Result<u64, DownloadError> {
        let remote_storage = self
            .ext_remote_storage
@@ -958,35 +956,6 @@ LIMIT 100",
                "Remote extensions storage is not configured",
            )))?;

-        let mut real_ext_name = ext_name;
-        if is_library {
-            // sometimes library names might have a suffix like
-            // library.so or library.so.3. We strip this off
-            // because library_index is based on the name without the file extension
-            let strip_lib_suffix = Regex::new(r"\.so.*").unwrap();
-            let lib_raw_name = strip_lib_suffix.replace(real_ext_name, "").to_string();
-
-            real_ext_name = self
-                .library_index
-                .get()
-                .expect("must have already downloaded the library_index")
-                .get(&lib_raw_name)
-                .ok_or(DownloadError::BadInput(anyhow::anyhow!(
-                    "library {} is not found",
-                    lib_raw_name
-                )))?;
-        }
-
-        let ext_path = &self
-            .ext_remote_paths
-            .get()
-            .expect("error accessing ext_remote_paths")
-            .get(real_ext_name)
-            .ok_or(DownloadError::BadInput(anyhow::anyhow!(
-                "real_ext_name {} is not found",
-                real_ext_name
-            )))?;
-
        let ext_archive_name = ext_path.object_name().expect("bad path");

        let mut first_try = false;
@@ -1039,8 +1008,8 @@ LIMIT 100",
        info!("downloading new extension {ext_archive_name}");

        let download_size = extension_server::download_extension(
-            real_ext_name,
-            ext_path,
+            &real_ext_name,
+            &ext_path,
            remote_storage,
            &self.pgbin,
        )
@@ -1058,18 +1027,19 @@ LIMIT 100",
    #[tokio::main]
    pub async fn prepare_preload_libraries(
        &self,
-        compute_state: &ComputeState,
+        spec: &ComputeSpec,
    ) -> Result<RemoteExtensionMetrics> {
        if self.ext_remote_storage.is_none() {
            return Ok(RemoteExtensionMetrics {
                num_ext_downloaded: 0,
                largest_ext_size: 0,
                total_ext_download_size: 0,
-                prep_extensions_ms: 0,
            });
        }
-        let pspec = compute_state.pspec.as_ref().expect("spec must be set");
-        let spec = &pspec.spec;
+        let remote_extensions = spec
+            .remote_extensions
+            .as_ref()
+            .ok_or(anyhow::anyhow!("Remote extensions are not configured",))?;

        info!("parse shared_preload_libraries from spec.cluster.settings");
        let mut libs_vec = Vec::new();
@@ -1081,6 +1051,7 @@ LIMIT 100",
                .collect();
        }
        info!("parse shared_preload_libraries from provided postgresql.conf");
+
        // that is used in neon_local and python tests
        if let Some(conf) = &spec.cluster.postgresql_conf {
            let conf_lines = conf.split('\n').collect::<Vec<&str>>();
@@ -1101,30 +1072,17 @@ LIMIT 100",
            libs_vec.extend(preload_libs_vec);
        }

-        info!("Download ext_index.json, find the extension paths");
-        let prep_ext_start_time = Utc::now();
-        self.prepare_external_extensions(compute_state).await?;
-        let prep_ext_time_delta = Utc::now()
-            .signed_duration_since(prep_ext_start_time)
-            .to_std()
-            .unwrap()
-            .as_millis() as u64;
-        info!("Prepare extensions took {prep_ext_time_delta}ms");
-
        // Don't try to download libraries that are not in the index.
        // Assume that they are already present locally.
-        libs_vec.retain(|lib| {
-            self.library_index
-                .get()
-                .expect("error accessing ext_remote_paths")
-                .contains_key(lib)
-        });
+        libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));

        info!("Downloading to shared preload libraries: {:?}", &libs_vec);

        let mut download_tasks = Vec::new();
        for library in &libs_vec {
-            download_tasks.push(self.download_extension(library, true));
+            let (ext_name, ext_path) =
+                remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
+            download_tasks.push(self.download_extension(ext_name, ext_path));
        }
        let results = join_all(download_tasks).await;

@@ -1132,7 +1090,6 @@ LIMIT 100",
            num_ext_downloaded: 0,
            largest_ext_size: 0,
            total_ext_download_size: 0,
-            prep_extensions_ms: prep_ext_time_delta,
        };
        for result in results {
            let download_size = match result {
--- a/compute_tools/src/config.rs
+++ b/compute_tools/src/config.rs
@@ -46,8 +46,6 @@ pub fn write_postgres_conf(
        writeln!(file, "{}", conf)?;
    }

-    write!(file, "{}", &spec.cluster.settings.as_pg_settings())?;
-
    // Add options for connecting to storage
    writeln!(file, "# Neon storage settings")?;
    if let Some(s) = &spec.pageserver_connstring {
--- a/compute_tools/src/extension_server.rs
+++ b/compute_tools/src/extension_server.rs
@@ -73,10 +73,10 @@ More specifically, here is an example ext_index.json
 */
 use anyhow::Context;
 use anyhow::{self, Result};
-use futures::future::join_all;
+use compute_api::spec::RemoteExtSpec;
+use regex::Regex;
 use remote_storage::*;
 use serde_json;
-use std::collections::HashMap;
 use std::io::Read;
 use std::num::{NonZeroU32, NonZeroUsize};
 use std::path::Path;
@@ -107,89 +107,69 @@ fn get_pg_config(argument: &str, pgbin: &str) -> String {

 pub fn get_pg_version(pgbin: &str) -> String {
    // pg_config --version returns a (platform specific) human readable string
-    // such as "PostgreSQL 15.4". We parse this to v14/v15
+    // such as "PostgreSQL 15.4". We parse this to v14/v15/v16 etc.
    let human_version = get_pg_config("--version", pgbin);
-    if human_version.contains("15") {
-        return "v15".to_string();
-    } else if human_version.contains("14") {
-        return "v14".to_string();
+    return parse_pg_version(&human_version).to_string();
+}
+
+fn parse_pg_version(human_version: &str) -> &str {
+    // Normal releases have version strings like "PostgreSQL 15.4". But there
+    // are also pre-release versions like "PostgreSQL 17devel" or "PostgreSQL
+    // 16beta2" or "PostgreSQL 17rc1". And with the --with-extra-version
+    // configure option, you can tack any string to the version number,
+    // e.g. "PostgreSQL 15.4foobar".
+    match Regex::new(r"^PostgreSQL (?<major>\d+).+")
+        .unwrap()
+        .captures(human_version)
+    {
+        Some(captures) if captures.len() == 2 => match &captures["major"] {
+            "14" => return "v14",
+            "15" => return "v15",
+            "16" => return "v16",
+            _ => {}
+        },
+        _ => {}
    }
    panic!("Unsuported postgres version {human_version}");
 }

-// download control files for enabled_extensions
-// return Hashmaps converting library names to extension names (library_index)
-// and specifying the remote path to the archive for each extension name
-pub async fn get_available_extensions(
-    remote_storage: &GenericRemoteStorage,
-    pgbin: &str,
-    pg_version: &str,
-    custom_extensions: &[String],
-    build_tag: &str,
-) -> Result<(HashMap<String, RemotePath>, HashMap<String, String>)> {
-    let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
-    let index_path = format!("{build_tag}/{pg_version}/ext_index.json");
-    let index_path = RemotePath::new(Path::new(&index_path)).context("error forming path")?;
-    info!("download ext_index.json from: {:?}", &index_path);
+#[cfg(test)]
+mod tests {
+    use super::parse_pg_version;

-    let mut download = remote_storage.download(&index_path).await?;
-    let mut ext_idx_buffer = Vec::new();
-    download
-        .download_stream
-        .read_to_end(&mut ext_idx_buffer)
-        .await?;
-    info!("ext_index downloaded");
+    #[test]
+    fn test_parse_pg_version() {
+        assert_eq!(parse_pg_version("PostgreSQL 15.4"), "v15");
+        assert_eq!(parse_pg_version("PostgreSQL 15.14"), "v15");
+        assert_eq!(
+            parse_pg_version("PostgreSQL 15.4 (Ubuntu 15.4-0ubuntu0.23.04.1)"),
+            "v15"
+        );

-    #[derive(Debug, serde::Deserialize)]
-    struct Index {
-        public_extensions: Vec<String>,
-        library_index: HashMap<String, String>,
-        extension_data: HashMap<String, ExtensionData>,
+        assert_eq!(parse_pg_version("PostgreSQL 14.15"), "v14");
+        assert_eq!(parse_pg_version("PostgreSQL 14.0"), "v14");
+        assert_eq!(
+            parse_pg_version("PostgreSQL 14.9 (Debian 14.9-1.pgdg120+1"),
+            "v14"
+        );
+
+        assert_eq!(parse_pg_version("PostgreSQL 16devel"), "v16");
+        assert_eq!(parse_pg_version("PostgreSQL 16beta1"), "v16");
+        assert_eq!(parse_pg_version("PostgreSQL 16rc2"), "v16");
+        assert_eq!(parse_pg_version("PostgreSQL 16extra"), "v16");
    }

-    #[derive(Debug, serde::Deserialize)]
-    struct ExtensionData {
-        control_data: HashMap<String, String>,
-        archive_path: String,
+    #[test]
+    #[should_panic]
+    fn test_parse_pg_unsupported_version() {
+        parse_pg_version("PostgreSQL 13.14");
    }

-    let ext_index_full = serde_json::from_slice::<Index>(&ext_idx_buffer)?;
-    let mut enabled_extensions = ext_index_full.public_extensions;
-    enabled_extensions.extend_from_slice(custom_extensions);
-    let mut library_index = ext_index_full.library_index;
-    let all_extension_data = ext_index_full.extension_data;
-    info!("library_index: {:?}", library_index);
-
-    info!("enabled_extensions: {:?}", enabled_extensions);
-    let mut ext_remote_paths = HashMap::new();
-    let mut file_create_tasks = Vec::new();
-    for extension in enabled_extensions {
-        let ext_data = &all_extension_data[&extension];
-        for (control_file, control_contents) in &ext_data.control_data {
-            let extension_name = control_file
-                .strip_suffix(".control")
-                .expect("control files must end in .control");
-            let control_path = local_sharedir.join(control_file);
-            if !control_path.exists() {
-                ext_remote_paths.insert(
-                    extension_name.to_string(),
-                    RemotePath::from_string(&ext_data.archive_path)?,
-                );
-                info!("writing file {:?}{:?}", control_path, control_contents);
-                file_create_tasks.push(tokio::fs::write(control_path, control_contents));
-            } else {
-                warn!("control file {:?} exists both locally and remotely. ignoring the remote version.", control_file);
-                // also delete this from library index
-                library_index.retain(|_, value| value != extension_name);
-            }
-        }
+    #[test]
+    #[should_panic]
+    fn test_parse_pg_incorrect_version_format() {
+        parse_pg_version("PostgreSQL 14");
    }
-    let results = join_all(file_create_tasks).await;
-    for result in results {
-        result?;
-    }
-    info!("ext_remote_paths {:?}", ext_remote_paths);
-    Ok((ext_remote_paths, library_index))
 }

 // download the archive for a given extension,
@@ -253,6 +233,34 @@ pub async fn download_extension(
    Ok(download_size)
 }

+// Create extension control files from spec
+pub fn create_control_files(remote_extensions: &RemoteExtSpec, pgbin: &str) {
+    let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
+    for (ext_name, ext_data) in remote_extensions.extension_data.iter() {
+        // Check if extension is present in public or custom.
+        // If not, then it is not allowed to be used by this compute.
+        if let Some(public_extensions) = &remote_extensions.public_extensions {
+            if !public_extensions.contains(ext_name) {
+                if let Some(custom_extensions) = &remote_extensions.custom_extensions {
+                    if !custom_extensions.contains(ext_name) {
+                        continue; // skip this extension, it is not allowed
+                    }
+                }
+            }
+        }
+
+        for (control_name, control_content) in &ext_data.control_data {
+            let control_path = local_sharedir.join(control_name);
+            if !control_path.exists() {
+                info!("writing file {:?}{:?}", control_path, control_content);
+                std::fs::write(control_path, control_content).unwrap();
+            } else {
+                warn!("control file {:?} exists both locally and remotely. ignoring the remote version.", control_path);
+            }
+        }
+    }
+}
+
 // This function initializes the necessary structs to use remote storage
 pub fn init_remote_storage(remote_ext_config: &str) -> anyhow::Result<GenericRemoteStorage> {
    #[derive(Debug, serde::Deserialize)]
--- a/compute_tools/src/http/api.rs
+++ b/compute_tools/src/http/api.rs
@@ -1,4 +1,6 @@
 use std::convert::Infallible;
+use std::net::IpAddr;
+use std::net::Ipv6Addr;
 use std::net::SocketAddr;
 use std::sync::Arc;
 use std::thread;
@@ -13,7 +15,7 @@ use hyper::{Body, Method, Request, Response, Server, StatusCode};
 use num_cpus;
 use serde_json;
 use tokio::task;
-use tracing::{error, info};
+use tracing::{error, info, warn};
 use tracing_utils::http::OtelName;

 fn status_response_from_state(state: &ComputeState) -> ComputeStatusResponse {
@@ -126,6 +128,15 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
            info!("serving {:?} POST request", route);
            info!("req.uri {:?}", req.uri());

+            // don't even try to download extensions
+            // if no remote storage is configured
+            if compute.ext_remote_storage.is_none() {
+                info!("no extensions remote storage configured");
+                let mut resp = Response::new(Body::from("no remote storage configured"));
+                *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                return resp;
+            }
+
            let mut is_library = false;
            if let Some(params) = req.uri().query() {
                info!("serving {:?} POST request with params: {}", route, params);
@@ -137,24 +148,52 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
                    return resp;
                }
            }
-
            let filename = route.split('/').last().unwrap().to_string();
            info!("serving /extension_server POST request, filename: {filename:?} is_library: {is_library}");

-            // don't even try to download extensions
-            // if no remote storage is configured
-            if compute.ext_remote_storage.is_none() {
-                info!("no extensions remote storage configured");
-                let mut resp = Response::new(Body::from("no remote storage configured"));
-                *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
-                return resp;
-            }
+            // get ext_name and path from spec
+            // don't lock compute_state for too long
+            let ext = {
+                let compute_state = compute.state.lock().unwrap();
+                let pspec = compute_state.pspec.as_ref().expect("spec must be set");
+                let spec = &pspec.spec;

-            match compute.download_extension(&filename, is_library).await {
-                Ok(_) => Response::new(Body::from("OK")),
+                // debug only
+                info!("spec: {:?}", spec);
+
+                let remote_extensions = match spec.remote_extensions.as_ref() {
+                    Some(r) => r,
+                    None => {
+                        info!("no remote extensions spec was provided");
+                        let mut resp = Response::new(Body::from("no remote storage configured"));
+                        *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                        return resp;
+                    }
+                };
+
+                remote_extensions.get_ext(
+                    &filename,
+                    is_library,
+                    &compute.build_tag,
+                    &compute.pgversion,
+                )
+            };
+
+            match ext {
+                Ok((ext_name, ext_path)) => {
+                    match compute.download_extension(ext_name, ext_path).await {
+                        Ok(_) => Response::new(Body::from("OK")),
+                        Err(e) => {
+                            error!("extension download failed: {}", e);
+                            let mut resp = Response::new(Body::from(e.to_string()));
+                            *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                            resp
+                        }
+                    }
+                }
                Err(e) => {
-                    error!("extension download failed: {}", e);
-                    let mut resp = Response::new(Body::from(e.to_string()));
+                    warn!("extension download failed to find extension: {}", e);
+                    let mut resp = Response::new(Body::from("failed to find file"));
                    *resp.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
                    resp
                }
@@ -261,7 +300,9 @@ fn render_json_error(e: &str, status: StatusCode) -> Response<Body> {
 // Main Hyper HTTP server function that runs it and blocks waiting on it forever.
 #[tokio::main]
 async fn serve(port: u16, state: Arc<ComputeNode>) {
-    let addr = SocketAddr::from(([0, 0, 0, 0], port));
+    // this usually binds to both IPv4 and IPv6 on linux
+    // see e.g. https://github.com/rust-lang/rust/pull/34440
+    let addr = SocketAddr::new(IpAddr::from(Ipv6Addr::UNSPECIFIED), port);

    let make_service = make_service_fn(move |_conn| {
        let state = state.clone();
--- a/compute_tools/src/params.rs
+++ b/compute_tools/src/params.rs
@@ -6,4 +6,4 @@ pub const DEFAULT_LOG_LEVEL: &str = "info";
 //   https://www.postgresql.org/docs/15/auth-password.html
 //
 // So it's safe to set md5 here, as `control-plane` anyway uses SCRAM for all roles.
-pub const PG_HBA_ALL_MD5: &str = "host\tall\t\tall\t\t0.0.0.0/0\t\tmd5";
+pub const PG_HBA_ALL_MD5: &str = "host\tall\t\tall\t\tall\t\tmd5";
--- a/control_plane/Cargo.toml
+++ b/control_plane/Cargo.toml
@@ -12,6 +12,8 @@ git-version.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 postgres.workspace = true
+hex.workspace = true
+hyper.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["blocking", "json"] }
 serde.workspace = true
@@ -20,6 +22,7 @@ serde_with.workspace = true
 tar.workspace = true
 thiserror.workspace = true
 toml.workspace = true
+tokio.workspace = true
 url.workspace = true
 # Note: Do not directly depend on pageserver or safekeeper; use pageserver_api or safekeeper_api
 # instead, so that recompile times are better.
--- a/control_plane/simple.conf
+++ b/control_plane/simple.conf
@@ -1,6 +1,7 @@
 # Minimal neon environment with one safekeeper. This is equivalent to the built-in
 # defaults that you get with no --config
-[pageserver]
+[[pageservers]]
+id=1
 listen_pg_addr = '127.0.0.1:64000'
 listen_http_addr = '127.0.0.1:9898'
 pg_auth_type = 'Trust'
--- a/control_plane/src/attachment_service.rs
+++ b/control_plane/src/attachment_service.rs
@@ -0,0 +1,105 @@
+use crate::{background_process, local_env::LocalEnv};
+use anyhow::anyhow;
+use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
+use std::{path::PathBuf, process::Child};
+use utils::id::{NodeId, TenantId};
+
+pub struct AttachmentService {
+    env: LocalEnv,
+    listen: String,
+    path: PathBuf,
+}
+
+const COMMAND: &str = "attachment_service";
+
+#[serde_as]
+#[derive(Serialize, Deserialize)]
+pub struct AttachHookRequest {
+    #[serde_as(as = "DisplayFromStr")]
+    pub tenant_id: TenantId,
+    pub pageserver_id: Option<NodeId>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct AttachHookResponse {
+    pub gen: Option<u32>,
+}
+
+impl AttachmentService {
+    pub fn from_env(env: &LocalEnv) -> Self {
+        let path = env.base_data_dir.join("attachments.json");
+
+        // Makes no sense to construct this if pageservers aren't going to use it: assume
+        // pageservers have control plane API set
+        let listen_url = env.control_plane_api.clone().unwrap();
+
+        let listen = format!(
+            "{}:{}",
+            listen_url.host_str().unwrap(),
+            listen_url.port().unwrap()
+        );
+
+        Self {
+            env: env.clone(),
+            path,
+            listen,
+        }
+    }
+
+    fn pid_file(&self) -> PathBuf {
+        self.env.base_data_dir.join("attachment_service.pid")
+    }
+
+    pub fn start(&self) -> anyhow::Result<Child> {
+        let path_str = self.path.to_string_lossy();
+
+        background_process::start_process(
+            COMMAND,
+            &self.env.base_data_dir,
+            &self.env.attachment_service_bin(),
+            ["-l", &self.listen, "-p", &path_str],
+            [],
+            background_process::InitialPidFile::Create(&self.pid_file()),
+            // TODO: a real status check
+            || Ok(true),
+        )
+    }
+
+    pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
+        background_process::stop_process(immediate, COMMAND, &self.pid_file())
+    }
+
+    /// Call into the attach_hook API, for use before handing out attachments to pageservers
+    pub fn attach_hook(
+        &self,
+        tenant_id: TenantId,
+        pageserver_id: NodeId,
+    ) -> anyhow::Result<Option<u32>> {
+        use hyper::StatusCode;
+
+        let url = self
+            .env
+            .control_plane_api
+            .clone()
+            .unwrap()
+            .join("attach_hook")
+            .unwrap();
+        let client = reqwest::blocking::ClientBuilder::new()
+            .build()
+            .expect("Failed to construct http client");
+
+        let request = AttachHookRequest {
+            tenant_id,
+            pageserver_id: Some(pageserver_id),
+        };
+
+        let response = client.post(url).json(&request).send()?;
+        if response.status() != StatusCode::OK {
+            return Err(anyhow!("Unexpected status {}", response.status()));
+        }
+
+        let response = response.json::<AttachHookResponse>()?;
+        Ok(response.gen)
+    }
+}
--- a/control_plane/src/bin/attachment_service.rs
+++ b/control_plane/src/bin/attachment_service.rs
@@ -0,0 +1,273 @@
+/// The attachment service mimics the aspects of the control plane API
+/// that are required for a pageserver to operate.
+///
+/// This enables running & testing pageservers without a full-blown
+/// deployment of the Neon cloud platform.
+///
+use anyhow::anyhow;
+use clap::Parser;
+use hex::FromHex;
+use hyper::StatusCode;
+use hyper::{Body, Request, Response};
+use serde::{Deserialize, Serialize};
+use std::path::{Path, PathBuf};
+use std::{collections::HashMap, sync::Arc};
+use utils::logging::{self, LogFormat};
+
+use utils::{
+    http::{
+        endpoint::{self},
+        error::ApiError,
+        json::{json_request, json_response},
+        RequestExt, RouterBuilder,
+    },
+    id::{NodeId, TenantId},
+    tcp_listener,
+};
+
+use pageserver_api::control_api::{
+    ReAttachRequest, ReAttachResponse, ReAttachResponseTenant, ValidateRequest, ValidateResponse,
+    ValidateResponseTenant,
+};
+
+use control_plane::attachment_service::{AttachHookRequest, AttachHookResponse};
+
+#[derive(Parser)]
+#[command(author, version, about, long_about = None)]
+#[command(arg_required_else_help(true))]
+struct Cli {
+    /// Host and port to listen on, like `127.0.0.1:1234`
+    #[arg(short, long)]
+    listen: std::net::SocketAddr,
+
+    /// Path to the .json file to store state (will be created if it doesn't exist)
+    #[arg(short, long)]
+    path: PathBuf,
+}
+
+// The persistent state of each Tenant
+#[derive(Serialize, Deserialize, Clone)]
+struct TenantState {
+    // Currently attached pageserver
+    pageserver: Option<NodeId>,
+
+    // Latest generation number: next time we attach, increment this
+    // and use the incremented number when attaching
+    generation: u32,
+}
+
+fn to_hex_map<S, V>(input: &HashMap<TenantId, V>, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::Serializer,
+    V: Clone + Serialize,
+{
+    let transformed = input.iter().map(|(k, v)| (hex::encode(k), v.clone()));
+
+    transformed
+        .collect::<HashMap<String, V>>()
+        .serialize(serializer)
+}
+
+fn from_hex_map<'de, D, V>(deserializer: D) -> Result<HashMap<TenantId, V>, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+    V: Deserialize<'de>,
+{
+    let hex_map = HashMap::<String, V>::deserialize(deserializer)?;
+    hex_map
+        .into_iter()
+        .map(|(k, v)| {
+            TenantId::from_hex(k)
+                .map(|k| (k, v))
+                .map_err(serde::de::Error::custom)
+        })
+        .collect()
+}
+
+// Top level state available to all HTTP handlers
+#[derive(Serialize, Deserialize)]
+struct PersistentState {
+    #[serde(serialize_with = "to_hex_map", deserialize_with = "from_hex_map")]
+    tenants: HashMap<TenantId, TenantState>,
+
+    #[serde(skip)]
+    path: PathBuf,
+}
+
+impl PersistentState {
+    async fn save(&self) -> anyhow::Result<()> {
+        let bytes = serde_json::to_vec(self)?;
+        tokio::fs::write(&self.path, &bytes).await?;
+
+        Ok(())
+    }
+
+    async fn load(path: &Path) -> anyhow::Result<Self> {
+        let bytes = tokio::fs::read(path).await?;
+        let mut decoded = serde_json::from_slice::<Self>(&bytes)?;
+        decoded.path = path.to_owned();
+        Ok(decoded)
+    }
+
+    async fn load_or_new(path: &Path) -> Self {
+        match Self::load(path).await {
+            Ok(s) => {
+                tracing::info!("Loaded state file at {}", path.display());
+                s
+            }
+            Err(e)
+                if e.downcast_ref::<std::io::Error>()
+                    .map(|e| e.kind() == std::io::ErrorKind::NotFound)
+                    .unwrap_or(false) =>
+            {
+                tracing::info!("Will create state file at {}", path.display());
+                Self {
+                    tenants: HashMap::new(),
+                    path: path.to_owned(),
+                }
+            }
+            Err(e) => {
+                panic!("Failed to load state from '{}': {e:#} (maybe your .neon/ dir was written by an older version?)", path.display())
+            }
+        }
+    }
+}
+
+/// State available to HTTP request handlers
+#[derive(Clone)]
+struct State {
+    inner: Arc<tokio::sync::RwLock<PersistentState>>,
+}
+
+impl State {
+    fn new(persistent_state: PersistentState) -> State {
+        Self {
+            inner: Arc::new(tokio::sync::RwLock::new(persistent_state)),
+        }
+    }
+}
+
+#[inline(always)]
+fn get_state(request: &Request<Body>) -> &State {
+    request
+        .data::<Arc<State>>()
+        .expect("unknown state type")
+        .as_ref()
+}
+
+/// Pageserver calls into this on startup, to learn which tenants it should attach
+async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let reattach_req = json_request::<ReAttachRequest>(&mut req).await?;
+
+    let state = get_state(&req).inner.clone();
+    let mut locked = state.write().await;
+
+    let mut response = ReAttachResponse {
+        tenants: Vec::new(),
+    };
+    for (t, state) in &mut locked.tenants {
+        if state.pageserver == Some(reattach_req.node_id) {
+            state.generation += 1;
+            response.tenants.push(ReAttachResponseTenant {
+                id: *t,
+                generation: state.generation,
+            });
+        }
+    }
+
+    locked.save().await.map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, response)
+}
+
+/// Pageserver calls into this before doing deletions, to confirm that it still
+/// holds the latest generation for the tenants with deletions enqueued
+async fn handle_validate(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let validate_req = json_request::<ValidateRequest>(&mut req).await?;
+
+    let locked = get_state(&req).inner.read().await;
+
+    let mut response = ValidateResponse {
+        tenants: Vec::new(),
+    };
+
+    for req_tenant in validate_req.tenants {
+        if let Some(tenant_state) = locked.tenants.get(&req_tenant.id) {
+            let valid = tenant_state.generation == req_tenant.gen;
+            response.tenants.push(ValidateResponseTenant {
+                id: req_tenant.id,
+                valid,
+            });
+        }
+    }
+
+    json_response(StatusCode::OK, response)
+}
+/// Call into this before attaching a tenant to a pageserver, to acquire a generation number
+/// (in the real control plane this is unnecessary, because the same program is managing
+///  generation numbers and doing attachments).
+async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let attach_req = json_request::<AttachHookRequest>(&mut req).await?;
+
+    let state = get_state(&req).inner.clone();
+    let mut locked = state.write().await;
+
+    let tenant_state = locked
+        .tenants
+        .entry(attach_req.tenant_id)
+        .or_insert_with(|| TenantState {
+            pageserver: attach_req.pageserver_id,
+            generation: 0,
+        });
+
+    if attach_req.pageserver_id.is_some() {
+        tenant_state.generation += 1;
+    }
+    let generation = tenant_state.generation;
+
+    locked.save().await.map_err(ApiError::InternalServerError)?;
+
+    json_response(
+        StatusCode::OK,
+        AttachHookResponse {
+            gen: attach_req.pageserver_id.map(|_| generation),
+        },
+    )
+}
+
+fn make_router(persistent_state: PersistentState) -> RouterBuilder<hyper::Body, ApiError> {
+    endpoint::make_router()
+        .data(Arc::new(State::new(persistent_state)))
+        .post("/re-attach", handle_re_attach)
+        .post("/validate", handle_validate)
+        .post("/attach_hook", handle_attach_hook)
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    logging::init(
+        LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::Disabled,
+    )?;
+
+    let args = Cli::parse();
+    tracing::info!(
+        "Starting, state at {}, listening on {}",
+        args.path.to_string_lossy(),
+        args.listen
+    );
+
+    let persistent_state = PersistentState::load_or_new(&args.path).await;
+
+    let http_listener = tcp_listener::bind(args.listen)?;
+    let router = make_router(persistent_state)
+        .build()
+        .map_err(|err| anyhow!(err))?;
+    let service = utils::http::RouterService::new(router).unwrap();
+    let server = hyper::Server::from_tcp(http_listener)?.serve(service);
+
+    tracing::info!("Serving on {0}", args.listen);
+    server.await?;
+
+    Ok(())
+}
--- a/control_plane/src/bin/neon_local.rs
+++ b/control_plane/src/bin/neon_local.rs
@@ -8,6 +8,7 @@
 use anyhow::{anyhow, bail, Context, Result};
 use clap::{value_parser, Arg, ArgAction, ArgMatches, Command};
 use compute_api::spec::ComputeMode;
+use control_plane::attachment_service::AttachmentService;
 use control_plane::endpoint::ComputeControlPlane;
 use control_plane::local_env::LocalEnv;
 use control_plane::pageserver::PageServerNode;
@@ -43,14 +44,18 @@ project_git_version!(GIT_VERSION);

 const DEFAULT_PG_VERSION: &str = "15";

+const DEFAULT_PAGESERVER_CONTROL_PLANE_API: &str = "http://127.0.0.1:1234/";
+
 fn default_conf() -> String {
    format!(
        r#"
 # Default built-in configuration, defined in main.rs
+control_plane_api = '{DEFAULT_PAGESERVER_CONTROL_PLANE_API}'
+
 [broker]
 listen_addr = '{DEFAULT_BROKER_ADDR}'

-[pageserver]
+[[pageservers]]
 id = {DEFAULT_PAGESERVER_ID}
 listen_pg_addr = '{DEFAULT_PAGESERVER_PG_ADDR}'
 listen_http_addr = '{DEFAULT_PAGESERVER_HTTP_ADDR}'
@@ -61,6 +66,7 @@ http_auth_type = '{trust_auth}'
 id = {DEFAULT_SAFEKEEPER_ID}
 pg_port = {DEFAULT_SAFEKEEPER_PG_PORT}
 http_port = {DEFAULT_SAFEKEEPER_HTTP_PORT}
+
 "#,
        trust_auth = AuthType::Trust,
    )
@@ -107,6 +113,7 @@ fn main() -> Result<()> {
            "start" => handle_start_all(sub_args, &env),
            "stop" => handle_stop_all(sub_args, &env),
            "pageserver" => handle_pageserver(sub_args, &env),
+            "attachment_service" => handle_attachment_service(sub_args, &env),
            "safekeeper" => handle_safekeeper(sub_args, &env),
            "endpoint" => handle_endpoint(sub_args, &env),
            "pg" => bail!("'pg' subcommand has been renamed to 'endpoint'"),
@@ -252,7 +259,7 @@ fn get_timeline_infos(
    env: &local_env::LocalEnv,
    tenant_id: &TenantId,
 ) -> Result<HashMap<TimelineId, TimelineInfo>> {
-    Ok(PageServerNode::from_env(env)
+    Ok(get_default_pageserver(env)
        .timeline_list(tenant_id)?
        .into_iter()
        .map(|timeline_info| (timeline_info.timeline_id, timeline_info))
@@ -313,17 +320,30 @@ fn handle_init(init_match: &ArgMatches) -> anyhow::Result<LocalEnv> {
        .context("Failed to initialize neon repository")?;

    // Initialize pageserver, create initial tenant and timeline.
-    let pageserver = PageServerNode::from_env(&env);
-    pageserver
-        .initialize(&pageserver_config_overrides(init_match))
-        .unwrap_or_else(|e| {
-            eprintln!("pageserver init failed: {e:?}");
-            exit(1);
-        });
+    for ps_conf in &env.pageservers {
+        PageServerNode::from_env(&env, ps_conf)
+            .initialize(&pageserver_config_overrides(init_match))
+            .unwrap_or_else(|e| {
+                eprintln!("pageserver init failed: {e:?}");
+                exit(1);
+            });
+    }

    Ok(env)
 }

+/// The default pageserver is the one where CLI tenant/timeline operations are sent by default.
+/// For typical interactive use, one would just run with a single pageserver.  Scenarios with
+/// tenant/timeline placement across multiple pageservers are managed by python test code rather
+/// than this CLI.
+fn get_default_pageserver(env: &local_env::LocalEnv) -> PageServerNode {
+    let ps_conf = env
+        .pageservers
+        .first()
+        .expect("Config is validated to contain at least one pageserver");
+    PageServerNode::from_env(env, ps_conf)
+}
+
 fn pageserver_config_overrides(init_match: &ArgMatches) -> Vec<&str> {
    init_match
        .get_many::<String>("pageserver-config-override")
@@ -334,7 +354,7 @@ fn pageserver_config_overrides(init_match: &ArgMatches) -> Vec<&str> {
 }

 fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> anyhow::Result<()> {
-    let pageserver = PageServerNode::from_env(env);
+    let pageserver = get_default_pageserver(env);
    match tenant_match.subcommand() {
        Some(("list", _)) => {
            for t in pageserver.tenant_list()? {
@@ -342,13 +362,25 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
            }
        }
        Some(("create", create_match)) => {
-            let initial_tenant_id = parse_tenant_id(create_match)?;
            let tenant_conf: HashMap<_, _> = create_match
                .get_many::<String>("config")
                .map(|vals| vals.flat_map(|c| c.split_once(':')).collect())
                .unwrap_or_default();
-            let new_tenant_id = pageserver.tenant_create(initial_tenant_id, tenant_conf)?;
-            println!("tenant {new_tenant_id} successfully created on the pageserver");
+
+            // If tenant ID was not specified, generate one
+            let tenant_id = parse_tenant_id(create_match)?.unwrap_or_else(TenantId::generate);
+
+            let generation = if env.control_plane_api.is_some() {
+                // We must register the tenant with the attachment service, so
+                // that when the pageserver restarts, it will be re-attached.
+                let attachment_service = AttachmentService::from_env(env);
+                attachment_service.attach_hook(tenant_id, pageserver.conf.id)?
+            } else {
+                None
+            };
+
+            pageserver.tenant_create(tenant_id, generation, tenant_conf)?;
+            println!("tenant {tenant_id} successfully created on the pageserver");

            // Create an initial timeline for the new tenant
            let new_timeline_id = parse_timeline_id(create_match)?;
@@ -358,7 +390,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
                .context("Failed to parse postgres version from the argument string")?;

            let timeline_info = pageserver.timeline_create(
-                new_tenant_id,
+                tenant_id,
                new_timeline_id,
                None,
                None,
@@ -369,17 +401,17 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an

            env.register_branch_mapping(
                DEFAULT_BRANCH_NAME.to_string(),
-                new_tenant_id,
+                tenant_id,
                new_timeline_id,
            )?;

            println!(
-                "Created an initial timeline '{new_timeline_id}' at Lsn {last_record_lsn} for tenant: {new_tenant_id}",
+                "Created an initial timeline '{new_timeline_id}' at Lsn {last_record_lsn} for tenant: {tenant_id}",
            );

            if create_match.get_flag("set-default") {
-                println!("Setting tenant {new_tenant_id} as a default one");
-                env.default_tenant_id = Some(new_tenant_id);
+                println!("Setting tenant {tenant_id} as a default one");
+                env.default_tenant_id = Some(tenant_id);
            }
        }
        Some(("set-default", set_default_match)) => {
@@ -407,7 +439,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
 }

 fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -> Result<()> {
-    let pageserver = PageServerNode::from_env(env);
+    let pageserver = get_default_pageserver(env);

    match timeline_match.subcommand() {
        Some(("list", list_match)) => {
@@ -484,6 +516,7 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
                None,
                pg_version,
                ComputeMode::Primary,
+                DEFAULT_PAGESERVER_ID,
            )?;
            println!("Done");
        }
@@ -537,7 +570,6 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
        Some(ep_subcommand_data) => ep_subcommand_data,
        None => bail!("no endpoint subcommand provided"),
    };
-
    let mut cplane = ComputeControlPlane::load(env.clone())?;

    // All subcommands take an optional --tenant-id option
@@ -634,6 +666,13 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                .copied()
                .unwrap_or(false);

+            let pageserver_id =
+                if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
+                    NodeId(id_str.parse().context("while parsing pageserver id")?)
+                } else {
+                    DEFAULT_PAGESERVER_ID
+                };
+
            let mode = match (lsn, hot_standby) {
                (Some(lsn), false) => ComputeMode::Static(lsn),
                (None, true) => ComputeMode::Replica,
@@ -649,6 +688,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                http_port,
                pg_version,
                mode,
+                pageserver_id,
            )?;
        }
        "start" => {
@@ -658,6 +698,13 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                .get_one::<String>("endpoint_id")
                .ok_or_else(|| anyhow!("No endpoint ID was provided to start"))?;

+            let pageserver_id =
+                if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
+                    NodeId(id_str.parse().context("while parsing pageserver id")?)
+                } else {
+                    DEFAULT_PAGESERVER_ID
+                };
+
            let remote_ext_config = sub_args.get_one::<String>("remote-ext-config");

            // If --safekeepers argument is given, use only the listed safekeeper nodes.
@@ -677,7 +724,8 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(

            let endpoint = cplane.endpoints.get(endpoint_id.as_str());

-            let auth_token = if matches!(env.pageserver.pg_auth_type, AuthType::NeonJWT) {
+            let ps_conf = env.get_pageserver_conf(pageserver_id)?;
+            let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
                let claims = Claims::new(Some(tenant_id), Scope::Tenant);

                Some(env.generate_auth_token(&claims)?)
@@ -744,6 +792,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                    http_port,
                    pg_version,
                    mode,
+                    pageserver_id,
                )?;
                ep.start(&auth_token, safekeepers, remote_ext_config)?;
            }
@@ -768,51 +817,94 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
 }

 fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
-    let pageserver = PageServerNode::from_env(env);
+    fn get_pageserver(env: &local_env::LocalEnv, args: &ArgMatches) -> Result<PageServerNode> {
+        let node_id = if let Some(id_str) = args.get_one::<String>("pageserver-id") {
+            NodeId(id_str.parse().context("while parsing pageserver id")?)
+        } else {
+            DEFAULT_PAGESERVER_ID
+        };
+
+        Ok(PageServerNode::from_env(
+            env,
+            env.get_pageserver_conf(node_id)?,
+        ))
+    }

    match sub_match.subcommand() {
-        Some(("start", start_match)) => {
-            if let Err(e) = pageserver.start(&pageserver_config_overrides(start_match)) {
+        Some(("start", subcommand_args)) => {
+            if let Err(e) = get_pageserver(env, subcommand_args)?
+                .start(&pageserver_config_overrides(subcommand_args))
+            {
                eprintln!("pageserver start failed: {e}");
                exit(1);
            }
        }

+        Some(("stop", subcommand_args)) => {
+            let immediate = subcommand_args
+                .get_one::<String>("stop-mode")
+                .map(|s| s.as_str())
+                == Some("immediate");
+
+            if let Err(e) = get_pageserver(env, subcommand_args)?.stop(immediate) {
+                eprintln!("pageserver stop failed: {}", e);
+                exit(1);
+            }
+        }
+
+        Some(("restart", subcommand_args)) => {
+            let pageserver = get_pageserver(env, subcommand_args)?;
+            //TODO what shutdown strategy should we use here?
+            if let Err(e) = pageserver.stop(false) {
+                eprintln!("pageserver stop failed: {}", e);
+                exit(1);
+            }
+
+            if let Err(e) = pageserver.start(&pageserver_config_overrides(subcommand_args)) {
+                eprintln!("pageserver start failed: {e}");
+                exit(1);
+            }
+        }
+
+        Some(("status", subcommand_args)) => {
+            match get_pageserver(env, subcommand_args)?.check_status() {
+                Ok(_) => println!("Page server is up and running"),
+                Err(err) => {
+                    eprintln!("Page server is not available: {}", err);
+                    exit(1);
+                }
+            }
+        }
+
+        Some((sub_name, _)) => bail!("Unexpected pageserver subcommand '{}'", sub_name),
+        None => bail!("no pageserver subcommand provided"),
+    }
+    Ok(())
+}
+
+fn handle_attachment_service(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+    let svc = AttachmentService::from_env(env);
+    match sub_match.subcommand() {
+        Some(("start", _start_match)) => {
+            if let Err(e) = svc.start() {
+                eprintln!("start failed: {e}");
+                exit(1);
+            }
+        }
+
        Some(("stop", stop_match)) => {
            let immediate = stop_match
                .get_one::<String>("stop-mode")
                .map(|s| s.as_str())
                == Some("immediate");

-            if let Err(e) = pageserver.stop(immediate) {
-                eprintln!("pageserver stop failed: {}", e);
+            if let Err(e) = svc.stop(immediate) {
+                eprintln!("stop failed: {}", e);
                exit(1);
            }
        }
-
-        Some(("restart", restart_match)) => {
-            //TODO what shutdown strategy should we use here?
-            if let Err(e) = pageserver.stop(false) {
-                eprintln!("pageserver stop failed: {}", e);
-                exit(1);
-            }
-
-            if let Err(e) = pageserver.start(&pageserver_config_overrides(restart_match)) {
-                eprintln!("pageserver start failed: {e}");
-                exit(1);
-            }
-        }
-
-        Some(("status", _)) => match PageServerNode::from_env(env).check_status() {
-            Ok(_) => println!("Page server is up and running"),
-            Err(err) => {
-                eprintln!("Page server is not available: {}", err);
-                exit(1);
-            }
-        },
-
-        Some((sub_name, _)) => bail!("Unexpected pageserver subcommand '{}'", sub_name),
-        None => bail!("no pageserver subcommand provided"),
+        Some((sub_name, _)) => bail!("Unexpected attachment_service subcommand '{}'", sub_name),
+        None => bail!("no attachment_service subcommand provided"),
    }
    Ok(())
 }
@@ -825,6 +917,16 @@ fn get_safekeeper(env: &local_env::LocalEnv, id: NodeId) -> Result<SafekeeperNod
    }
 }

+// Get list of options to append to safekeeper command invocation.
+fn safekeeper_extra_opts(init_match: &ArgMatches) -> Vec<String> {
+    init_match
+        .get_many::<String>("safekeeper-extra-opt")
+        .into_iter()
+        .flatten()
+        .map(|s| s.to_owned())
+        .collect()
+}
+
 fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
    let (sub_name, sub_args) = match sub_match.subcommand() {
        Some(safekeeper_command_data) => safekeeper_command_data,
@@ -841,7 +943,9 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul

    match sub_name {
        "start" => {
-            if let Err(e) = safekeeper.start() {
+            let extra_opts = safekeeper_extra_opts(sub_args);
+
+            if let Err(e) = safekeeper.start(extra_opts) {
                eprintln!("safekeeper start failed: {}", e);
                exit(1);
            }
@@ -866,7 +970,8 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
                exit(1);
            }

-            if let Err(e) = safekeeper.start() {
+            let extra_opts = safekeeper_extra_opts(sub_args);
+            if let Err(e) = safekeeper.start(extra_opts) {
                eprintln!("safekeeper start failed: {}", e);
                exit(1);
            }
@@ -884,16 +989,28 @@ fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow

    broker::start_broker_process(env)?;

-    let pageserver = PageServerNode::from_env(env);
-    if let Err(e) = pageserver.start(&pageserver_config_overrides(sub_match)) {
-        eprintln!("pageserver {} start failed: {:#}", env.pageserver.id, e);
-        try_stop_all(env, true);
-        exit(1);
+    // Only start the attachment service if the pageserver is configured to need it
+    if env.control_plane_api.is_some() {
+        let attachment_service = AttachmentService::from_env(env);
+        if let Err(e) = attachment_service.start() {
+            eprintln!("attachment_service start failed: {:#}", e);
+            try_stop_all(env, true);
+            exit(1);
+        }
+    }
+
+    for ps_conf in &env.pageservers {
+        let pageserver = PageServerNode::from_env(env, ps_conf);
+        if let Err(e) = pageserver.start(&pageserver_config_overrides(sub_match)) {
+            eprintln!("pageserver {} start failed: {:#}", ps_conf.id, e);
+            try_stop_all(env, true);
+            exit(1);
+        }
    }

    for node in env.safekeepers.iter() {
        let safekeeper = SafekeeperNode::from_env(env, node);
-        if let Err(e) = safekeeper.start() {
+        if let Err(e) = safekeeper.start(vec![]) {
            eprintln!("safekeeper {} start failed: {:#}", safekeeper.id, e);
            try_stop_all(env, false);
            exit(1);
@@ -912,8 +1029,6 @@ fn handle_stop_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<
 }

 fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
-    let pageserver = PageServerNode::from_env(env);
-
    // Stop all endpoints
    match ComputeControlPlane::load(env.clone()) {
        Ok(cplane) => {
@@ -928,8 +1043,11 @@ fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
        }
    }

-    if let Err(e) = pageserver.stop(immediate) {
-        eprintln!("pageserver {} stop failed: {:#}", env.pageserver.id, e);
+    for ps_conf in &env.pageservers {
+        let pageserver = PageServerNode::from_env(env, ps_conf);
+        if let Err(e) = pageserver.stop(immediate) {
+            eprintln!("pageserver {} stop failed: {:#}", ps_conf.id, e);
+        }
    }

    for node in env.safekeepers.iter() {
@@ -942,6 +1060,13 @@ fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
    if let Err(e) = broker::stop_broker_process(env) {
        eprintln!("neon broker stop failed: {e:#}");
    }
+
+    if env.control_plane_api.is_some() {
+        let attachment_service = AttachmentService::from_env(env);
+        if let Err(e) = attachment_service.stop(immediate) {
+            eprintln!("attachment service stop failed: {e:#}");
+        }
+    }
 }

 fn cli() -> Command {
@@ -956,6 +1081,24 @@ fn cli() -> Command {

    let safekeeper_id_arg = Arg::new("id").help("safekeeper id").required(false);

+    // --id, when using a pageserver command
+    let pageserver_id_arg = Arg::new("pageserver-id")
+        .long("id")
+        .help("pageserver id")
+        .required(false);
+    // --pageserver-id when using a non-pageserver command
+    let endpoint_pageserver_id_arg = Arg::new("endpoint-pageserver-id")
+        .long("pageserver-id")
+        .required(false);
+
+    let safekeeper_extra_opt_arg = Arg::new("safekeeper-extra-opt")
+        .short('e')
+        .long("safekeeper-extra-opt")
+        .num_args(1)
+        .action(ArgAction::Append)
+        .help("Additional safekeeper invocation options, e.g. -e=--http-auth-public-key-path=foo")
+        .required(false);
+
    let tenant_id_arg = Arg::new("tenant-id")
        .long("tenant-id")
        .help("Tenant id. Represented as a hexadecimal string 32 symbols length")
@@ -1112,10 +1255,24 @@ fn cli() -> Command {
                .arg_required_else_help(true)
                .about("Manage pageserver")
                .subcommand(Command::new("status"))
+                .arg(pageserver_id_arg.clone())
+                .subcommand(Command::new("start").about("Start local pageserver")
+                .arg(pageserver_id_arg.clone())
+                .arg(pageserver_config_args.clone()))
+                .subcommand(Command::new("stop").about("Stop local pageserver")
+                .arg(pageserver_id_arg.clone())
+                            .arg(stop_mode_arg.clone()))
+                .subcommand(Command::new("restart").about("Restart local pageserver")
+                .arg(pageserver_id_arg.clone())
+                .arg(pageserver_config_args.clone()))
+        )
+        .subcommand(
+            Command::new("attachment_service")
+                .arg_required_else_help(true)
+                .about("Manage attachment_service")
                .subcommand(Command::new("start").about("Start local pageserver").arg(pageserver_config_args.clone()))
                .subcommand(Command::new("stop").about("Stop local pageserver")
                            .arg(stop_mode_arg.clone()))
-                .subcommand(Command::new("restart").about("Restart local pageserver").arg(pageserver_config_args.clone()))
        )
        .subcommand(
            Command::new("safekeeper")
@@ -1124,6 +1281,7 @@ fn cli() -> Command {
                .subcommand(Command::new("start")
                            .about("Start local safekeeper")
                            .arg(safekeeper_id_arg.clone())
+                            .arg(safekeeper_extra_opt_arg.clone())
                )
                .subcommand(Command::new("stop")
                            .about("Stop local safekeeper")
@@ -1134,6 +1292,7 @@ fn cli() -> Command {
                            .about("Restart local safekeeper")
                            .arg(safekeeper_id_arg)
                            .arg(stop_mode_arg.clone())
+                            .arg(safekeeper_extra_opt_arg)
                )
        )
        .subcommand(
@@ -1149,6 +1308,7 @@ fn cli() -> Command {
                    .arg(lsn_arg.clone())
                    .arg(pg_port_arg.clone())
                    .arg(http_port_arg.clone())
+                    .arg(endpoint_pageserver_id_arg.clone())
                    .arg(
                        Arg::new("config-only")
                            .help("Don't do basebackup, create endpoint directory with only config files")
@@ -1166,6 +1326,7 @@ fn cli() -> Command {
                    .arg(lsn_arg)
                    .arg(pg_port_arg)
                    .arg(http_port_arg)
+                    .arg(endpoint_pageserver_id_arg.clone())
                    .arg(pg_version_arg)
                    .arg(hot_standby_arg)
                    .arg(safekeepers_arg)
--- a/control_plane/src/endpoint.rs
+++ b/control_plane/src/endpoint.rs
@@ -70,6 +70,7 @@ pub struct EndpointConf {
    http_port: u16,
    pg_version: u32,
    skip_pg_catalog_updates: bool,
+    pageserver_id: NodeId,
 }

 //
@@ -82,19 +83,16 @@ pub struct ComputeControlPlane {
    pub endpoints: BTreeMap<String, Arc<Endpoint>>,

    env: LocalEnv,
-    pageserver: Arc<PageServerNode>,
 }

 impl ComputeControlPlane {
    // Load current endpoints from the endpoints/ subdirectories
    pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
-        let pageserver = Arc::new(PageServerNode::from_env(&env));
-
        let mut endpoints = BTreeMap::default();
        for endpoint_dir in std::fs::read_dir(env.endpoints_path())
            .with_context(|| format!("failed to list {}", env.endpoints_path().display()))?
        {
-            let ep = Endpoint::from_dir_entry(endpoint_dir?, &env, &pageserver)?;
+            let ep = Endpoint::from_dir_entry(endpoint_dir?, &env)?;
            endpoints.insert(ep.endpoint_id.clone(), Arc::new(ep));
        }

@@ -102,7 +100,6 @@ impl ComputeControlPlane {
            base_port: 55431,
            endpoints,
            env,
-            pageserver,
        })
    }

@@ -125,20 +122,29 @@ impl ComputeControlPlane {
        http_port: Option<u16>,
        pg_version: u32,
        mode: ComputeMode,
+        pageserver_id: NodeId,
    ) -> Result<Arc<Endpoint>> {
        let pg_port = pg_port.unwrap_or_else(|| self.get_port());
        let http_port = http_port.unwrap_or_else(|| self.get_port() + 1);
+        let pageserver =
+            PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
        let ep = Arc::new(Endpoint {
            endpoint_id: endpoint_id.to_owned(),
            pg_address: SocketAddr::new("127.0.0.1".parse().unwrap(), pg_port),
            http_address: SocketAddr::new("127.0.0.1".parse().unwrap(), http_port),
            env: self.env.clone(),
-            pageserver: Arc::clone(&self.pageserver),
+            pageserver,
            timeline_id,
            mode,
            tenant_id,
            pg_version,
-            skip_pg_catalog_updates: false,
+            // We don't setup roles and databases in the spec locally, so we don't need to
+            // do catalog updates. Catalog updates also include check availability
+            // data creation. Yet, we have tests that check that size and db dump
+            // before and after start are the same. So, skip catalog updates,
+            // with this we basically test a case of waking up an idle compute, where
+            // we also skip catalog updates in the cloud.
+            skip_pg_catalog_updates: true,
        });

        ep.create_endpoint_dir()?;
@@ -152,7 +158,8 @@ impl ComputeControlPlane {
                http_port,
                pg_port,
                pg_version,
-                skip_pg_catalog_updates: false,
+                skip_pg_catalog_updates: true,
+                pageserver_id,
            })?,
        )?;
        std::fs::write(
@@ -187,18 +194,14 @@ pub struct Endpoint {
    // These are not part of the endpoint as such, but the environment
    // the endpoint runs in.
    pub env: LocalEnv,
-    pageserver: Arc<PageServerNode>,
+    pageserver: PageServerNode,

    // Optimizations
    skip_pg_catalog_updates: bool,
 }

 impl Endpoint {
-    fn from_dir_entry(
-        entry: std::fs::DirEntry,
-        env: &LocalEnv,
-        pageserver: &Arc<PageServerNode>,
-    ) -> Result<Endpoint> {
+    fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
        if !entry.file_type()?.is_dir() {
            anyhow::bail!(
                "Endpoint::from_dir_entry failed: '{}' is not a directory",
@@ -214,12 +217,15 @@ impl Endpoint {
        let conf: EndpointConf =
            serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;

+        let pageserver =
+            PageServerNode::from_env(env, env.get_pageserver_conf(conf.pageserver_id)?);
+
        Ok(Endpoint {
            pg_address: SocketAddr::new("127.0.0.1".parse().unwrap(), conf.pg_port),
            http_address: SocketAddr::new("127.0.0.1".parse().unwrap(), conf.http_port),
            endpoint_id,
            env: env.clone(),
-            pageserver: Arc::clone(pageserver),
+            pageserver,
            timeline_id: conf.timeline_id,
            mode: conf.mode,
            tenant_id: conf.tenant_id,
@@ -493,7 +499,7 @@ impl Endpoint {
            pageserver_connstring: Some(pageserver_connstring),
            safekeeper_connstrings,
            storage_auth_token: auth_token.clone(),
-            custom_extensions: Some(vec![]),
+            remote_extensions: None,
        };
        let spec_path = self.endpoint_path().join("spec.json");
        std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
--- a/control_plane/src/lib.rs
+++ b/control_plane/src/lib.rs
@@ -7,6 +7,7 @@
 // local installations.
 //

+pub mod attachment_service;
 mod background_process;
 pub mod broker;
 pub mod endpoint;
--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -68,11 +68,17 @@ pub struct LocalEnv {

    pub broker: NeonBroker,

-    pub pageserver: PageServerConf,
+    /// This Vec must always contain at least one pageserver
+    pub pageservers: Vec<PageServerConf>,

    #[serde(default)]
    pub safekeepers: Vec<SafekeeperConf>,

+    // Control plane location: if None, we will not run attachment_service.  If set, this will
+    // be propagated into each pageserver's configuration.
+    #[serde(default)]
+    pub control_plane_api: Option<Url>,
+
    /// Keep human-readable aliases in memory (and persist them to config), to hide ZId hex strings from the user.
    #[serde(default)]
    // A `HashMap<String, HashMap<TenantId, TimelineId>>` would be more appropriate here,
@@ -176,32 +182,28 @@ impl LocalEnv {
    pub fn pg_distrib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
        let path = self.pg_distrib_dir.clone();

+        #[allow(clippy::manual_range_patterns)]
        match pg_version {
-            14 => Ok(path.join(format!("v{pg_version}"))),
-            15 => Ok(path.join(format!("v{pg_version}"))),
+            14 | 15 | 16 => Ok(path.join(format!("v{pg_version}"))),
            _ => bail!("Unsupported postgres version: {}", pg_version),
        }
    }

    pub fn pg_bin_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
-        match pg_version {
-            14 => Ok(self.pg_distrib_dir(pg_version)?.join("bin")),
-            15 => Ok(self.pg_distrib_dir(pg_version)?.join("bin")),
-            _ => bail!("Unsupported postgres version: {}", pg_version),
-        }
+        Ok(self.pg_distrib_dir(pg_version)?.join("bin"))
    }
    pub fn pg_lib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
-        match pg_version {
-            14 => Ok(self.pg_distrib_dir(pg_version)?.join("lib")),
-            15 => Ok(self.pg_distrib_dir(pg_version)?.join("lib")),
-            _ => bail!("Unsupported postgres version: {}", pg_version),
-        }
+        Ok(self.pg_distrib_dir(pg_version)?.join("lib"))
    }

    pub fn pageserver_bin(&self) -> PathBuf {
        self.neon_distrib_dir.join("pageserver")
    }

+    pub fn attachment_service_bin(&self) -> PathBuf {
+        self.neon_distrib_dir.join("attachment_service")
+    }
+
    pub fn safekeeper_bin(&self) -> PathBuf {
        self.neon_distrib_dir.join("safekeeper")
    }
@@ -214,15 +216,23 @@ impl LocalEnv {
        self.base_data_dir.join("endpoints")
    }

-    // TODO: move pageserver files into ./pageserver
-    pub fn pageserver_data_dir(&self) -> PathBuf {
-        self.base_data_dir.clone()
+    pub fn pageserver_data_dir(&self, pageserver_id: NodeId) -> PathBuf {
+        self.base_data_dir
+            .join(format!("pageserver_{pageserver_id}"))
    }

    pub fn safekeeper_data_dir(&self, data_dir_name: &str) -> PathBuf {
        self.base_data_dir.join("safekeepers").join(data_dir_name)
    }

+    pub fn get_pageserver_conf(&self, id: NodeId) -> anyhow::Result<&PageServerConf> {
+        if let Some(conf) = self.pageservers.iter().find(|node| node.id == id) {
+            Ok(conf)
+        } else {
+            bail!("could not find pageserver {id}")
+        }
+    }
+
    pub fn register_branch_mapping(
        &mut self,
        branch_name: String,
@@ -299,6 +309,10 @@ impl LocalEnv {
            env.neon_distrib_dir = env::current_exe()?.parent().unwrap().to_owned();
        }

+        if env.pageservers.is_empty() {
+            anyhow::bail!("Configuration must contain at least one pageserver");
+        }
+
        env.base_data_dir = base_path();

        Ok(env)
@@ -331,7 +345,7 @@ impl LocalEnv {
        // We read that in, in `create_config`, and fill any missing defaults. Then it's saved
        // to .neon/config. TODO: We lose any formatting and comments along the way, which is
        // a bit sad.
-        let mut conf_content = r#"# This file describes a locale deployment of the page server
+        let mut conf_content = r#"# This file describes a local deployment of the page server
 # and safekeeeper node. It is read by the 'neon_local' command-line
 # utility.
 "#
@@ -461,9 +475,9 @@ impl LocalEnv {
    }

    fn auth_keys_needed(&self) -> bool {
-        self.pageserver.pg_auth_type == AuthType::NeonJWT
-            || self.pageserver.http_auth_type == AuthType::NeonJWT
-            || self.safekeepers.iter().any(|sk| sk.auth_enabled)
+        self.pageservers.iter().any(|ps| {
+            ps.pg_auth_type == AuthType::NeonJWT || ps.http_auth_type == AuthType::NeonJWT
+        }) || self.safekeepers.iter().any(|sk| sk.auth_enabled)
    }
 }

--- a/control_plane/src/pageserver.rs
+++ b/control_plane/src/pageserver.rs
@@ -27,6 +27,7 @@ use utils::{
    lsn::Lsn,
 };

+use crate::local_env::PageServerConf;
 use crate::{background_process, local_env::LocalEnv};

 #[derive(Error, Debug)]
@@ -76,43 +77,40 @@ impl ResponseErrorMessageExt for Response {
 #[derive(Debug)]
 pub struct PageServerNode {
    pub pg_connection_config: PgConnectionConfig,
+    pub conf: PageServerConf,
    pub env: LocalEnv,
    pub http_client: Client,
    pub http_base_url: String,
 }

 impl PageServerNode {
-    pub fn from_env(env: &LocalEnv) -> PageServerNode {
-        let (host, port) = parse_host_port(&env.pageserver.listen_pg_addr)
-            .expect("Unable to parse listen_pg_addr");
+    pub fn from_env(env: &LocalEnv, conf: &PageServerConf) -> PageServerNode {
+        let (host, port) =
+            parse_host_port(&conf.listen_pg_addr).expect("Unable to parse listen_pg_addr");
        let port = port.unwrap_or(5432);
        Self {
            pg_connection_config: PgConnectionConfig::new_host_port(host, port),
+            conf: conf.clone(),
            env: env.clone(),
            http_client: Client::new(),
-            http_base_url: format!("http://{}/v1", env.pageserver.listen_http_addr),
+            http_base_url: format!("http://{}/v1", conf.listen_http_addr),
        }
    }

    // pageserver conf overrides defined by neon_local configuration.
    fn neon_local_overrides(&self) -> Vec<String> {
-        let id = format!("id={}", self.env.pageserver.id);
+        let id = format!("id={}", self.conf.id);
        // FIXME: the paths should be shell-escaped to handle paths with spaces, quotas etc.
        let pg_distrib_dir_param = format!(
            "pg_distrib_dir='{}'",
            self.env.pg_distrib_dir_raw().display()
        );

-        let http_auth_type_param =
-            format!("http_auth_type='{}'", self.env.pageserver.http_auth_type);
-        let listen_http_addr_param = format!(
-            "listen_http_addr='{}'",
-            self.env.pageserver.listen_http_addr
-        );
+        let http_auth_type_param = format!("http_auth_type='{}'", self.conf.http_auth_type);
+        let listen_http_addr_param = format!("listen_http_addr='{}'", self.conf.listen_http_addr);

-        let pg_auth_type_param = format!("pg_auth_type='{}'", self.env.pageserver.pg_auth_type);
-        let listen_pg_addr_param =
-            format!("listen_pg_addr='{}'", self.env.pageserver.listen_pg_addr);
+        let pg_auth_type_param = format!("pg_auth_type='{}'", self.conf.pg_auth_type);
+        let listen_pg_addr_param = format!("listen_pg_addr='{}'", self.conf.listen_pg_addr);

        let broker_endpoint_param = format!("broker_endpoint='{}'", self.env.broker.client_url());

@@ -126,10 +124,18 @@ impl PageServerNode {
            broker_endpoint_param,
        ];

-        if self.env.pageserver.http_auth_type != AuthType::Trust
-            || self.env.pageserver.pg_auth_type != AuthType::Trust
+        if let Some(control_plane_api) = &self.env.control_plane_api {
+            overrides.push(format!(
+                "control_plane_api='{}'",
+                control_plane_api.as_str()
+            ));
+        }
+
+        if self.conf.http_auth_type != AuthType::Trust || self.conf.pg_auth_type != AuthType::Trust
        {
-            overrides.push("auth_validation_public_key_path='auth_public_key.pem'".to_owned());
+            // Keys are generated in the toplevel repo dir, pageservers' workdirs
+            // are one level below that, so refer to keys with ../
+            overrides.push("auth_validation_public_key_path='../auth_public_key.pem'".to_owned());
        }
        overrides
    }
@@ -137,16 +143,12 @@ impl PageServerNode {
    /// Initializes a pageserver node by creating its config with the overrides provided.
    pub fn initialize(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
        // First, run `pageserver --init` and wait for it to write a config into FS and exit.
-        self.pageserver_init(config_overrides).with_context(|| {
-            format!(
-                "Failed to run init for pageserver node {}",
-                self.env.pageserver.id,
-            )
-        })
+        self.pageserver_init(config_overrides)
+            .with_context(|| format!("Failed to run init for pageserver node {}", self.conf.id,))
    }

    pub fn repo_path(&self) -> PathBuf {
-        self.env.pageserver_data_dir()
+        self.env.pageserver_data_dir(self.conf.id)
    }

    /// The pid file is created by the pageserver process, with its pid stored inside.
@@ -162,7 +164,7 @@ impl PageServerNode {

    fn pageserver_init(&self, config_overrides: &[&str]) -> anyhow::Result<()> {
        let datadir = self.repo_path();
-        let node_id = self.env.pageserver.id;
+        let node_id = self.conf.id;
        println!(
            "Initializing pageserver node {} at '{}' in {:?}",
            node_id,
@@ -171,6 +173,10 @@ impl PageServerNode {
        );
        io::stdout().flush()?;

+        if !datadir.exists() {
+            std::fs::create_dir(&datadir)?;
+        }
+
        let datadir_path_str = datadir.to_str().with_context(|| {
            format!("Cannot start pageserver node {node_id} in path that has no string representation: {datadir:?}")
        })?;
@@ -201,7 +207,7 @@ impl PageServerNode {
        let datadir = self.repo_path();
        print!(
            "Starting pageserver node {} at '{}' in {:?}",
-            self.env.pageserver.id,
+            self.conf.id,
            self.pg_connection_config.raw_address(),
            datadir
        );
@@ -210,7 +216,7 @@ impl PageServerNode {
        let datadir_path_str = datadir.to_str().with_context(|| {
            format!(
                "Cannot start pageserver node {} in path that has no string representation: {:?}",
-                self.env.pageserver.id, datadir,
+                self.conf.id, datadir,
            )
        })?;
        let mut args = self.pageserver_basic_args(config_overrides, datadir_path_str);
@@ -254,7 +260,7 @@ impl PageServerNode {
        // FIXME: why is this tied to pageserver's auth type? Whether or not the safekeeper
        // needs a token, and how to generate that token, seems independent to whether
        // the pageserver requires a token in incoming requests.
-        Ok(if self.env.pageserver.http_auth_type != AuthType::Trust {
+        Ok(if self.conf.http_auth_type != AuthType::Trust {
            // Generate a token to connect from the pageserver to a safekeeper
            let token = self
                .env
@@ -279,7 +285,7 @@ impl PageServerNode {

    pub fn page_server_psql_client(&self) -> anyhow::Result<postgres::Client> {
        let mut config = self.pg_connection_config.clone();
-        if self.env.pageserver.pg_auth_type == AuthType::NeonJWT {
+        if self.conf.pg_auth_type == AuthType::NeonJWT {
            let token = self
                .env
                .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
@@ -290,7 +296,7 @@ impl PageServerNode {

    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> anyhow::Result<RequestBuilder> {
        let mut builder = self.http_client.request(method, url);
-        if self.env.pageserver.http_auth_type == AuthType::NeonJWT {
+        if self.conf.http_auth_type == AuthType::NeonJWT {
            let token = self
                .env
                .generate_auth_token(&Claims::new(None, Scope::PageServerApi))?;
@@ -316,7 +322,8 @@ impl PageServerNode {

    pub fn tenant_create(
        &self,
-        new_tenant_id: Option<TenantId>,
+        new_tenant_id: TenantId,
+        generation: Option<u32>,
        settings: HashMap<&str, &str>,
    ) -> anyhow::Result<TenantId> {
        let mut settings = settings.clone();
@@ -382,11 +389,9 @@ impl PageServerNode {
                .context("Failed to parse 'gc_feedback' as bool")?,
        };

-        // If tenant ID was not specified, generate one
-        let new_tenant_id = new_tenant_id.unwrap_or(TenantId::generate());
-
        let request = models::TenantCreateRequest {
            new_tenant_id,
+            generation,
            config,
        };
        if !settings.is_empty() {
--- a/control_plane/src/safekeeper.rs
+++ b/control_plane/src/safekeeper.rs
@@ -101,7 +101,7 @@ impl SafekeeperNode {
        self.datadir_path().join("safekeeper.pid")
    }

-    pub fn start(&self) -> anyhow::Result<Child> {
+    pub fn start(&self, extra_opts: Vec<String>) -> anyhow::Result<Child> {
        print!(
            "Starting safekeeper at '{}' in '{}'",
            self.pg_connection_config.raw_address(),
@@ -161,17 +161,28 @@ impl SafekeeperNode {

        let key_path = self.env.base_data_dir.join("auth_public_key.pem");
        if self.conf.auth_enabled {
+            let key_path_string = key_path
+                .to_str()
+                .with_context(|| {
+                    format!("Key path {key_path:?} cannot be represented as a unicode string")
+                })?
+                .to_owned();
            args.extend([
-                "--auth-validation-public-key-path".to_owned(),
-                key_path
-                    .to_str()
-                    .with_context(|| {
-                        format!("Key path {key_path:?} cannot be represented as a unicode string")
-                    })?
-                    .to_owned(),
+                "--pg-auth-public-key-path".to_owned(),
+                key_path_string.clone(),
+            ]);
+            args.extend([
+                "--pg-tenant-only-auth-public-key-path".to_owned(),
+                key_path_string.clone(),
+            ]);
+            args.extend([
+                "--http-auth-public-key-path".to_owned(),
+                key_path_string.clone(),
            ]);
        }

+        args.extend(extra_opts);
+
        background_process::start_process(
            &format!("safekeeper-{id}"),
            &datadir,
--- a/deny.toml
+++ b/deny.toml
@@ -4,7 +4,12 @@
 # to your expectations and requirements.

 # Root options
-targets = []
+targets = [
+    { triple = "x86_64-unknown-linux-gnu" },
+    { triple = "aarch64-unknown-linux-gnu" },
+    { triple = "aarch64-apple-darwin" },
+    { triple = "x86_64-apple-darwin" },
+]
 all-features = false
 no-default-features = false
 feature-depth = 1
@@ -18,7 +23,7 @@ vulnerability = "deny"
 unmaintained = "warn"
 yanked = "warn"
 notice = "warn"
-ignore = []
+ignore = ["RUSTSEC-2023-0052"]

 # This section is considered when running `cargo deny check licenses`
 # More documentation for the licenses section can be found here:
--- a/docker-compose/docker_compose_test.sh
+++ b/docker-compose/docker_compose_test.sh
@@ -30,7 +30,7 @@ cleanup() {
 echo "clean up containers if exists"
 cleanup

-for pg_version in 14 15; do
+for pg_version in 14 15 16; do
    echo "start containers (pg_version=$pg_version)."
    PG_VERSION=$pg_version docker compose -f $COMPOSE_FILE up --build -d

--- a/docs/rfcs/025-generation-numbers.md
+++ b/docs/rfcs/025-generation-numbers.md
@@ -0,0 +1,957 @@
+# Pageserver: split-brain safety for remote storage through generation numbers
+
+## Summary
+
+A scheme of logical "generation numbers" for tenant attachment to pageservers is proposed, along with
+changes to the remote storage format to include these generation numbers in S3 keys.
+
+Using the control plane as the issuer of these generation numbers enables strong anti-split-brain
+properties in the pageserver cluster without implementing a consensus mechanism directly
+in the pageservers.
+
+## Motivation
+
+Currently, the pageserver's remote storage format does not provide a mechanism for addressing
+split brain conditions that may happen when replacing a node or when migrating
+a tenant from one pageserver to another.
+
+From a remote storage perspective, a split brain condition occurs whenever two nodes both think
+they have the same tenant attached, and both can write to S3. This can happen in the case of a
+network partition, pathologically long delays (e.g. suspended VM), or software bugs.
+
+In the current deployment model, control plane guarantees that a tenant is attached to one
+pageserver at a time, thereby ruling out split-brain conditions resulting from dual
+attachment (however, there is always the risk of a control plane bug). This control
+plane guarantee prevents robust response to failures, as if a pageserver is unresponsive
+we may not detach from it. The mechanism in this RFC fixes this, by making it safe to
+attach to a new, different pageserver even if an unresponsive pageserver may be running.
+
+Futher, lack of safety during split-brain conditions blocks two important features where occasional
+split-brain conditions are part of the design assumptions:
+
+- seamless tenant migration ([RFC PR](https://github.com/neondatabase/neon/pull/5029))
+- automatic pageserver instance failure handling (aka "failover") (RFC TBD)
+
+### Prior art
+
+- 020-pageserver-s3-coordination.md
+- 023-the-state-of-pageserver-tenant-relocation.md
+- 026-pageserver-s3-mvcc.md
+
+This RFC has broad similarities to the proposal to implement a MVCC scheme in
+S3 object names, but this RFC avoids a general purpose transaction scheme in
+favour of more specialized "generations" that work like a transaction ID that
+always has the same lifetime as a pageserver process or tenant attachment, whichever
+is shorter.
+
+## Requirements
+
+- Accommodate storage backends with no atomic or fencing capability (i.e. work within
+  S3's limitation that there are no atomics and clients can't be fenced)
+- Don't depend on any STONITH or node fencing in the compute layer (i.e. we will not
+  assume that we can reliably kill and EC2 instance and have it die)
+- Scoped per-tenant, not per-pageserver; for _seamless tenant migration_, we need
+  per-tenant granularity, and for _failover_, we likely want to spread the workload
+  of the failed pageserver instance to a number of peers, rather than monolithically
+  moving the entire workload to another machine.
+  We do not rule out the latter case, but should not constrain ourselves to it.
+
+## Design Tenets
+
+These are not requirements, but are ideas that guide the following design:
+
+- Avoid implementing another consensus system: we already have a strongly consistent
+  database in the control plane that can do atomic operations where needed, and we also
+  have a Paxos implementation in the safekeeper.
+- Avoiding locking in to specific models of how failover will work (e.g. do not assume that
+  all the tenants on a pageserver will fail over as a unit).
+- Be strictly correct when it comes to data integrity. Occasional failures of availability
+  are tolerable, occasional data loss is not.
+
+## Non Goals
+
+The changes in this RFC intentionally isolate the design decision of how to define
+logical generations numbers and object storage format in a way that is somewhat flexible with
+respect to how actual orchestration of failover works.
+
+This RFC intentionally does not cover:
+
+- Failure detection
+- Orchestration of failover
+- Standby modes to keep data ready for fast migration
+- Intentional multi-writer operation on tenants (multi-writer scenarios are assumed to be transient split-brain situations).
+- Sharding.
+
+The interaction between this RFC and those features is discussed in [Appendix B](#appendix-b-interoperability-with-other-features)
+
+## Impacted Components
+
+pageserver, control plane, safekeeper (optional)
+
+## Implementation Part 1: Correctness
+
+### Summary
+
+- A per-tenant **generation number** is introduced to uniquely identifying tenant attachments to pageserver processes.
+
+  - This generation number increments each time the control plane modifies a tenant (`Project`)'s assigned pageserver, or when the assigned pageserver restarts.
+  - the control plane is the authority for generation numbers: only it may
+    increment a generation number.
+
+- **Object keys are suffixed** with the generation number
+- **Safety for multiply-attached tenants** is provided by the
+  generation number in the object key: the competing pageservers will not
+  try to write to the same keys.
+- **Safety in split brain for multiple nodes running with
+  the same node ID** is provided by the pageserver calling out to the control plane
+  on startup, to re-attach and thereby increment the generations of any attached tenants
+- **Safety for deletions** is achieved by deferring the DELETE from S3 to a point in time where the deleting node has validated with control plane that no attachment with a higher generation has a reference to the to-be-DELETEd key.
+- **The control plane is used to issue generation numbers** to avoid the need for
+  a built-in consensus system in the pageserver, although this could in principle
+  be changed without changing the storage format.
+
+### Generation numbers
+
+A generation number is associated with each tenant in the control plane,
+and each time the attachment status of the tenant changes, this is incremented.
+Changes in attachment status include:
+
+- Attaching the tenant to a different pageserver
+- A pageserver restarting, and "re-attaching" its tenants on startup
+
+These increments of attachment generation provide invariants we need to avoid
+split-brain issues in storage:
+
+- If two pageservers have the same tenant attached, the attachments are guaranteed to have different generation numbers, because the generation would increment
+  while attaching the second one.
+- If there are multiple pageservers running with the same node ID, all the attachments on all pageservers are guaranteed to have different generation numbers, because the generation would increment
+  when the second node started and re-attached its tenants.
+
+As long as the infrastructure does not transparently replace an underlying
+physical machine, we are totally safe. See the later [unsafe case](#unsafe-case-on-badly-behaved-infrastructure) section for details.
+
+### Object Key Changes
+
+#### Generation suffix
+
+All object keys (layer objects and index objects) will contain the attachment
+generation as a [suffix](#why-a-generation-suffix-rather-than-prefix).
+This suffix is the primary mechanism for protecting against split-brain situations, and
+enabling safe multi-attachment of tenants:
+
+- Two pageservers running with the same node ID (e.g. after a failure, where there is
+  some rogue pageserver still running) will not try to write to the same objects, because at startup they will have re-attached tenants and thereby incremented
+  generation numbers.
+- Multiple attachments (to different pageservers) of the same tenant will not try to write to the same objects, as each attachment would have a distinct generation.
+
+The generation is appended in hex format (8 byte string representing
+u32), to all our existing key names. A u32's range limit would permit
+27 restarts _per second_ over a 5 year system lifetime: orders of magnitude more than
+is realistic.
+
+The exact meaning of the generation suffix can evolve over time if necessary, for
+example if we chose to implement a failover mechanism internally to the pageservers
+rather than going via the control plane. The storage format just sees it as a number,
+with the only semantic property being that the highest numbered index is the latest.
+
+#### Index changes
+
+Since object keys now include a generation suffix, the index of these keys must also be updated. IndexPart currently stores keys and LSNs sufficient to reconstruct key names: this would be extended to store the generation as well.
+
+This will increase the size of the file, but only modestly: layers are already encoded as
+their string-ized form, so the overhead is about 10 bytes per layer. This will be less if/when
+the index storage format is migrated to a binary format from JSON.
+
+#### Visibility
+
+_This section doesn't describe code changes, but extends on the consequences of the
+object key changes given above_
+
+##### Visibility of objects to pageservers
+
+Pageservers can of course list objects in S3 at any time, but in practice their
+visible set is based on the contents of their LayerMap, which is initialized
+from the `index_part.json.???` that they load.
+
+Starting with the `index_part` from the most recent previous generation
+(see [loading index_part](#finding-the-remote-indices-for-timelines)), a pageserver
+initially has visibility of all the objects that were referenced in the loaded index.
+These objects are guaranteed to remain visible until the current generation is
+superseded, via pageservers in older generations avoiding deletions (see [deletion](#deletion)).
+
+The "most recent previous generation" is _not_ necessarily the most recent
+in terms of walltime, it is the one that is readable at the time a new generation
+starts. Consider the following sequence of a tenant being re-attached to different
+pageserver nodes:
+
+- Create + attach on PS1 in generation 1
+- PS1 Do some work, write out index_part.json-0001
+- Attach to PS2 in generation 2
+- Read index_part.json-0001
+- PS2 starts doing some work...
+- Attach to PS3 in generation 3
+- Read index_part.json-0001
+- **...PS2 finishes its work: now it writes index_part.json-0002**
+- PS3 writes out index_part.json-0003
+
+In the above sequence, the ancestry of indices is:
+
+```
+0001 -> 0002
+     |
+     -> 0003
+```
+
+This is not an issue for safety: if the 0002 references some object that is
+not in 0001, then 0003 simply does not see it, and will re-do whatever
+work was required (e.g. ingesting WAL or doing compaction). Objects referenced
+by only the 0002 index will never be read by future attachment generations, and
+will eventually be cleaned up by a scrub (see [scrubbing](#cleaning-up-orphan-objects-scrubbing)).
+
+##### Visibility of LSNs to clients
+
+Because index_part.json is now written with a generation suffix, which data
+is visible depends on which generation the reader is operating in:
+
+- If one was passively reading from S3 from outside of a pageserver, the
+  visibility of data would depend on which index_part.json-<generation> file
+  one had chosen to read from.
+- If two pageservers have the same tenant attached, they may have different
+  data visible as they're independently replaying the WAL, and maintaining
+  independent LayerMaps that are written to independent index_part.json files.
+  Data does not have to be remotely committed to be visible.
+- For a pageserver writing with a stale generation, historic LSNs
+  remain readable until another pageserver (with a higher generation suffix)
+  decides to execute GC deletions. At this point, we may think of the stale
+  attachment's generation as having logically ended: during its existence
+  the generation had a consistent view of the world.
+- For a newly attached pageserver, its highest visible LSN may appears to
+  go backwards with respect to an earlier attachment, if that earlier
+  attachment had not uploaded all data to S3 before the new attachment.
+
+### Deletion
+
+#### Generation number validation
+
+While writes are de-conflicted by writers always using their own generation number in the key,
+deletions are slightly more challenging: if a pageserver A is isolated, and the true active node is
+pageserver B, then it is dangerous for A to do any object deletions, even of objects that it wrote
+itself, because pageserver's B metadata might reference those objects.
+
+We solve this by inserting a "generation validation" step between the write of a remote index
+that un-links a particular object from the index, and the actual deletion of the object, such
+that deletions strictly obey the following ordering:
+
+1. Write out index_part.json: this guarantees that any subsequent reader of the metadata will
+   not try and read the object we unlinked.
+2. Call out to control plane to validate that the generation which we use for our attachment is still the latest.
+3. If step 2 passes, it is safe to delete the object. Why? The check-in with control plane
+   together with our visibility rules guarantees that any later generation
+   will use either the exact `index_part.json` that we uploaded in step 1, or a successor
+   of it; not an earlier one. In both cases, the `index_part.json` doesn't reference the
+   key we are deleting anymore, so, the key is invisible to any later attachment generation.
+   Hence it's safe to delete it.
+
+Note that at step 2 we are only confirming that deletions of objects _no longer referenced
+by the specific `index_part.json` written in step 1_ are safe. If we were attempting other deletions concurrently,
+these would need their own generation validation step.
+
+If step 2 fails, we may leak the object. This is safe, but has a cost: see [scrubbing](#cleaning-up-orphan-objects-scrubbing). We may avoid this entirely outside of node
+failures, if we do proper flushing of deletions on clean shutdown and clean migration.
+
+To avoid doing a huge number of control plane requests to perform generation validation,
+validation of many tenants will be done in a single request, and deletions will be queued up
+prior to validation: see [Persistent deletion queue](#persistent-deletion-queue) for more.
+
+#### `remote_consistent_lsn` updates
+
+Remote objects are not the only kind of deletion the pageserver does: it also indirectly deletes
+WAL data, by feeding back remote_consistent_lsn to safekeepers, as a signal to the safekeepers that
+they may drop data below this LSN.
+
+For the same reasons that deletion of objects must be guarded by an attachment generation number
+validation step, updates to `remote_consistent_lsn` are subject to the same rules, using
+an ordering as follows:
+
+1. upload the index_part that covers data up to LSN `L0` to S3
+2. Call out to control plane to validate that the generation which we use for our attachment is still the latest.
+3. advance the `remote_consistent_lsn` that we advertise to the safekeepers to `L0`
+
+If step 2 fails, then the `remote_consistent_lsn` advertised
+to safekeepers will not advance again until a pageserver
+with the latest generation is ready to do so.
+
+**Note:** at step 3 we are not advertising the _latest_ remote_consistent_lsn, we are
+advertising the value in the index_part that we uploaded in step 1. This provides
+a strong ordering guarantee.
+
+Internally to the pageserver, each timeline will have two remote_consistent_lsn values: the one that
+reflects its latest write to remote storage, and the one that reflects the most
+recent validation of generation number. It is only the latter value that may
+be advertised to the outside world (i.e. to the safekeeper).
+
+The control plane remains unaware of `remote_consistent_lsn`: it only has to validate
+the freshness of generation numbers, thereby granting the pageserver permission to
+share the information with the safekeeper.
+
+For convenience, in subsequent sections and RFCs we will use "deletion" to mean both deletion
+of objects in S3, and updates to the `remote_consistent_lsn`, as updates to the remote consistent
+LSN are de-facto deletions done via the safekeeper, and both kinds of deletion are subject to
+the same generation validation requirement.
+
+### Pageserver attach/startup changes
+
+#### Attachment
+
+Calls to `/v1/tenant/{tenant_id}/attach` are augmented with an additional
+`generation` field in the body.
+
+The pageserver does not persist this: a generation is only good for the lifetime
+of a process.
+
+#### Finding the remote indices for timelines
+
+Because index files are now suffixed with generation numbers, the pageserver
+cannot always GET the remote index in one request, because it can't always
+know a-priori what the latest remote index is.
+
+Typically, the most recent generation to write an index would be our own
+generation minus 1. However, this might not be the case: the previous
+node might have started and acquired a generation number, and then crashed
+before writing out a remote index.
+
+In the general case and as a fallback, the pageserver may list all the `index_part.json`
+files for a timeline, sort them by generation, and pick the highest that is `<=`
+its current generation for this attachment. The tenant should never load an index
+with an attachment generation _newer_ than its own.
+These two rules combined ensure that objects written by later generations are never visible to earlier generations.
+
+Note that if a given attachment picks an index part from an earlier generation (say n-2), but crashes & restarts before it writes its own generation's index part, next time it tries to pick an index part there may be an index part from generation n-1.
+It would pick the n-1 index part in that case, because it's sorted higher than the previous one from generation n-2.
+So, above rules guarantee no determinism in selecting the index part.
+are allowed to be attached with stale attachment generations during a multiply-attached
+phase in a migration, and in this instance if the old location's pageserver restarts,
+it should not try and load the newer generation's index.
+
+To summarize, on starting a timeline, the pageserver will:
+
+1. Issue a GET for index_part.json-<my generation - 1>
+2. If 1 failed, issue a ListObjectsv2 request for index_part.json\* and
+   pick the newest.
+
+One could optimize this further by using the control plane to record specifically
+which generation most recently wrote an index_part.json, if necessary, to increase
+the probability of finding the index_part.json in one GET. One could also improve
+the chances by having pageservers proactively write out index_part.json after they
+get a new generation ID.
+
+#### Re-attachment on startup
+
+On startup, the pageserver will call out to an new control plane `/re-attach`
+API (see [Generation API](#generation-api)). This returns a list of
+tenants that should be attached to the pageserver, and their generation numbers, which
+the control plane will increment before returning.
+
+The pageserver should still scan its local disk on startup, but should _delete_
+any local content for tenants not indicated in the `/re-attach` response: their
+absence is an implicit detach operation.
+
+**Note** if a tenant is omitted from the re-attach response, its local disk content
+will be deleted. This will change in subsequent work, when the control plane gains
+the concept of a secondary/standby location: a node with local content may revert
+to this status and retain some local content.
+
+#### Cleaning up previous generations' remote indices
+
+Deletion of old indices is not necessary for correctness, although it is necessary
+to avoid the ListObjects fallback in the previous section becoming ever more expensive.
+
+Once the new attachment has written out its index_part.json, it may asynchronously clean up historic index_part.json
+objects that were found.
+
+We may choose to implement this deletion either as an explicit step after we
+write out index_part for the first time in a pageserver's lifetime, or for
+simplicity just do it periodically as part of the background scrub (see [scrubbing](#cleaning-up-orphan-objects-scrubbing));
+
+### Control Plane Changes
+
+#### Store generations for attaching tenants
+
+- The `Project` table must store the generation number for use when
+  attaching the tenant to a new pageserver.
+- The `/v1/tenant/:tenant_id/attach` pageserver API will require the generation number,
+  which the control plane can supply by simply incrementing the `Project`'s
+  generation number each time the tenant is attached to a different server: the same database
+  transaction that changes the assigned pageserver should also change the generation number.
+
+#### Generation API
+
+This section describes an API that could be provided directly by the control plane,
+or built as a separate microservice. In earlier parts of the RFC, when we
+discuss the control plane providing generation numbers, we are referring to this API.
+
+The API endpoints used by the pageserver to acquire and validate generation
+numbers are quite simple, and only require access to some persistent and
+linerizable storage (such as a database).
+
+Building this into the control plane is proposed as a least-effort option to exploit existing infrastructure and implement generation number issuance in the same transaction that mandates it (i.e., the transaction that updates the `Project` assignment to another pageserver).
+However, this is not mandatory: this "Generation Number Issuer" could
+be built as a microservice. In practice, we will write such a miniature service
+anyway, to enable E2E pageserver/compute testing without control plane.
+
+The endpoints required by pageservers are:
+
+##### `/re-attach`
+
+- Request: `{node_id: <u32>}`
+- Response:
+  - 200 `{tenants: [{id: <TenantId>, gen: <u32>}]}`
+  - 404: unknown node_id
+  - (Future: 429: flapping detected, perhaps nodes are fighting for the same node ID,
+    or perhaps this node was in a retry loop)
+  - (On unknown tenants, omit tenant from `tenants` array)
+- Server behavior: query database for which tenants should be attached to this pageserver.
+  - for each tenant that should be attached, increment the attachment generation and
+    include the new generation in the response
+- Client behavior:
+  - for all tenants in the response, activate with the new generation number
+  - for any local disk content _not_ referenced in the response, act as if we
+    had been asked to detach it (i.e. delete local files)
+
+**Note** the `node_id` in this request will change in future if we move to ephemeral
+node IDs, to be replaced with some correlation ID that helps the control plane realize
+if a process is running with the same storage as a previous pageserver process (e.g.
+we might use EC instance ID, or we might just write some UUID to the disk the first
+time we use it)
+
+##### `/validate`
+
+- Request: `{'tenants': [{tenant: <tenant id>, attach_gen: <gen>}, ...]}'`
+- Response:
+  - 200 `{'tenants': [{tenant: <tenant id>, status: <bool>}...]}`
+  - (On unknown tenants, omit tenant from `tenants` array)
+- Purpose: enable the pageserver to discover for the given attachments whether they are still the latest.
+- Server behavior: this is a read-only operation: simply compare the generations in the request with
+  the generations known to the server, and set status to `true` if they match.
+- Client behavior: clients must not do deletions within a tenant's remote data until they have
+  received a response indicating the generation they hold for the attachment is current.
+
+#### Use of `/load` and `/ignore` APIs
+
+Because the pageserver will be changed to only attach tenants on startup
+based on the control plane's response to a `/re-attach` request, the load/ignore
+APIs no longer make sense in their current form.
+
+The `/load` API becomes functionally equivalent to attach, and will be removed:
+any location that used `/load` before should just attach instead.
+
+The `/ignore` API is equivalent to detaching, but without deleting local files.
+
+### Timeline/Branch creation & deletion
+
+All of the previous arguments for safety have described operations within
+a timeline, where we may describe a sequence that includes updates to
+index_part.json, and where reads and writes are coming from a postgres
+endpoint (writes via the safekeeper).
+
+Creating or destroying timeline is a bit different, because writes
+are coming from the control plane.
+
+We must be safe against scenarios such as:
+
+- A tenant is attached to pageserver B while pageserver A is
+  in the middle of servicing an RPC from the control plane to
+  create or delete a tenant.
+- A pageserver A has been sent a timeline creation request
+  but becomes unresponsive. The tenant is attached to a
+  different pageserver B, and the timeline creation request
+  is sent there too.
+
+#### Timeline Creation
+
+If some very slow node tries to do a timeline creation _after_
+a more recent generation node has already created the timeline
+and written some data into it, that must not cause harm. This
+is provided in timeline creations by the way all the objects
+within the timeline's remote path include a generation suffix:
+a slow node in an old generation that attempts to "create" a timeline
+that already exists will just emit an index_part.json with
+an old generation suffix.
+
+Timeline IDs are never reused, so we don't have
+to worry about the case of create/delete/create cycles. If they
+were re-used during a disaster recovery "un-delete" of a timeline,
+that special case can be handled by calling out to all available pageservers
+to check that they return 404 for the timeline, and to flush their
+deletion queues in case they had any deletions pending from the
+timeline.
+
+The above makes it safe for control plane to change the assignment of
+tenant to pageserver in control plane while a timeline creation is ongoing.
+The reason is that the creation request against the new assigned pageserver
+uses a new generation number. However, care must be taken by control plane
+to ensure that a "timeline creation successul" response from some pageserver
+is checked for the pageserver's generation for that timeline's tenant still being the latest.
+If it is not the latest, the response does not constitute a successful timeline creation.
+It is acceptable to discard such responses, the scrubber will clean up the S3 state.
+It is better to issue a timelien deletion request to the stale attachment.
+
+#### Timeline Deletion
+
+Tenant/timeline deletion operations are exempt from generation validation
+on deletes, and therefore don't have to go through the same deletion
+queue as GC/compaction layer deletions. This is because once a
+delete is issued by the control plane, it is a promise that the
+control plane will keep trying until the deletion is done, so even stale
+pageservers are permitted to go ahead and delete the objects.
+
+The implications of this for control plane are:
+
+- During timeline/tenant deletion, the control plane must wait for the deletion to
+  be truly complete (status 404) and also handle the case where the pageserver
+  becomes unavailable, either by waiting for a replacement with the same node_id,
+  or by *re-attaching the tenant elsewhere.
+
+- The control plane must persist its intent to delete
+  a timeline/tenant before issuing any RPCs, and then once it starts, it must
+  keep retrying until the tenant/timeline is gone. This is already handled
+  by using a persistent `Operation` record that is retried indefinitely.
+
+Timeline deletion may result in a special kind of object leak, where
+the latest generation attachment completes a deletion (including erasing
+all objects in the timeline path), but some slow/partitioned node is
+writing into the timeline path with a stale generation number. This would
+not be caught by any per-timeline scrubbing (see [scrubbing](#cleaning-up-orphan-objects-scrubbing)), since scrubbing happens on the
+attached pageserver, and once the timeline is deleted it isn't attached anywhere.
+This scenario should be pretty rare, and the control plane can make it even
+rarer by ensuring that if a tenant is in a multi-attached state (e.g. during
+migration), we wait for that to complete before processing the deletion. Beyond
+that, we may implement some other top-level scrub of timelines in
+an external tool, to identify any tenant/timeline paths that are not found
+in the control plane database.
+
+#### Examples
+
+- Deletion, node restarts partway through:
+  - By the time we returned 202, we have written a remote delete marker
+  - Any subsequent incarnation of the same node_id will see the remote
+    delete marker and continue to process the deletion
+  - If the original pageserver is lost permanently and no replacement
+    with the same node_id is available, then the control plane must recover
+    by re-attaching the tenant to a different node.
+- Creation, node becomes unresponsive partway through.
+  - Control plane will see HTTP request timeout, keep re-issuing
+    request to whoever is the latest attachment point for the tenant
+    until it succeeds.
+  - Stale nodes may be trying to execute timeline creation: they will
+    write out index_part.json files with
+    stale attachment generation: these will be eventually cleaned up
+    by the same mechanism as other old indices.
+
+### Unsafe case on badly behaved infrastructure
+
+This section is only relevant if running on a different environment
+than EC2 machines with ephemeral disks.
+
+If we ever run pageservers on infrastructure that might transparently restart
+a pageserver while leaving an old process running (e.g. a VM gets rescheduled
+without the old one being fenced), then there is a risk of corruption, when
+the control plane attaches the tenant, as follows:
+
+- If the control plane sends an `/attach` request to node A, then node A dies
+  and is replaced, and the control plane's retries the request without
+  incrementing that attachment ID, then it could end up with two physical nodes
+  both using the same generation number.
+- This is not an issue when using EC2 instances with ephemeral storage, as long
+  as the control plane never re-uses a node ID, but it would need re-examining
+  if running on different infrastructure.
+- To robustly protect against this class of issue, we would either:
+  - add a "node generation" to distinguish between different processes holding the
+    same node_id.
+  - or, dispense with static node_id entirely and issue an ephemeral ID to each
+    pageserver process when it starts.
+
+## Implementation Part 2: Optimizations
+
+### Persistent deletion queue
+
+Between writing our a new index_part.json that doesn't reference an object,
+and executing the deletion, an object passes through a window where it is
+only referenced in memory, and could be leaked if the pageserver is stopped
+uncleanly. That introduces conflicting incentives: on the one hand, we would
+like to delay and batch deletions to
+1. minimize the cost of the mandatory validations calls to control plane, and
+2. minimize cost for DeleteObjects requests.
+On the other hand we would also like to minimize leakage by executing
+deletions promptly.
+
+To resolve this, we may make the deletion queue persistent
+and then executing these in the background at a later time.
+
+_Note: The deletion queue's reason for existence is optimization rather than correctness,
+so there is a lot of flexibility in exactly how the it should work,
+as long as it obeys the rule to validate generations before executing deletions,
+so the following details are not essential to the overall RFC._
+
+#### Scope
+
+The deletion queue will be global per pageserver, not per-tenant. There
+are several reasons for this choice:
+
+- Use the queue as a central point to coalesce validation requests to the
+  control plane: this avoids individual `Timeline` objects ever touching
+  the control plane API, and avoids them having to know the rules about
+  validating deletions. This separation of concerns will avoid burdening
+  the already many-LoC `Timeline` type with even more responsibility.
+- Decouple the deletion queue from Tenant attachment lifetime: we may
+  "hibernate" an inactive tenant by tearing down its `Tenant`/`Timeline`
+  objects in the pageserver, without having to wait for deletions to be done.
+- Amortize the cost of I/O for the persistent queue, instead of having many
+  tiny queues.
+- Coalesce deletions into a smaller number of larger DeleteObjects calls
+
+Because of the cost of doing I/O for persistence, and the desire to coalesce
+generation validation requests across tenants, and coalesce deletions into
+larger DeleteObjects requests, there will be one deletion queue per pageserver
+rather than one per tenant. This has the added benefit that when deactivating
+a tenant, we do not have to drain their deletion queue: deletions can proceed
+for a tenant whose main `Tenant` object has been torn down.
+
+#### Flow of deletion
+
+The flow of a deletion is becomes:
+
+1. Need for deletion of an object (=> layer file) is identified.
+2. Unlink the object from all the places that reference it (=> `index_part.json`).
+3. Enqueue the deletion to a persistent queue.
+   Each entry is `tenant_id, attachment_generation, S3 key`.
+4. Validate & execute in batches:
+  4.1 For a batch of entries, call into control plane.
+  4.2 For the subset of entries that passed validation, execute a `DeleteObjects` S3 DELETE request for their S3 keys.
+
+As outlined in the Part 1 on correctness, it is critical that deletions are only
+executed once the key is not referenced anywhere in S3.
+This property is obviously upheld by the scheme above.
+
+#### We Accept Object Leakage In Acceptable Circumcstances
+
+If we crash in the flow above between (2) and (3), we lose track of unreferenced object.
+Further, enqueuing a single to the persistent queue may not be durable immediately to amortize cost of flush to disk.
+This is acceptable for now, it can be caught by [the scrubber](#cleaning-up-orphan-objects-scrubbing).
+
+There are various measures we can take to improve this in the future.
+1. Cap amount of time until enqueued entry becomes durable (timeout for flush-to-tisk)
+2. Proactively flush:
+    - On graceful shutdown, as we anticipate that some or
+      all of our attachments may be re-assigned while we are offline.
+    - On tenant detach.
+3. For each entry, keep track of whether it has passed (2).
+   Only admit entries to (4) one they have passed (2).
+   This requires re-writing / two queue entries (intent, commit) per deletion.
+
+The important take-away with any of the above is that it's not
+disastrous to leak objects in exceptional circumstances.
+
+#### Operations that may skip the queue
+
+Deletions of an entire timeline are [exempt](#Timeline-Deletion) from generation number validation. Once the
+control plane sends the deletion request, there is no requirement to retain the readability
+of any data within the timeline, and all objects within the timeline path may be deleted
+at any time from the control plane's deletion request onwards.
+
+Since deletions of smaller timelines won't have enough objects to compose a full sized
+DeleteObjects request, it is still useful to send these through the last part of the
+deletion pipeline to coalesce with other executing deletions: to enable this, the
+deletion queue should expose two input channels: one for deletions that must be
+processed in a generation-aware way, and a fast path for timeline deletions, where
+that fast path may skip validation and the persistent queue.
+
+### Cleaning up orphan objects (scrubbing)
+
+An orphan object is any object which is no longer referenced by a running node or by metadata.
+
+Examples of how orphan objects arise:
+
+- A node PUTs a layer object, then crashes before it writes the
+  index_part.json that references that layer.
+- A stale node carries on running for some time, and writes out an unbounded number of
+  objects while it believes itself to be the rightful writer for a tenant.
+- A pageserver crashes between un-linking an object from the index, and persisting
+  the object to its deletion queue.
+
+Orphan objects are functionally harmless, but have a small cost due to S3 capacity consumed. We
+may clean them up at some time in the future, but doing a ListObjectsv2 operation and cross
+referencing with the latest metadata to identify objects which are not referenced.
+
+Scrubbing will be done only by an attached pageserver (not some third party process), and deletions requested during scrub will go through the same
+validation as all other deletions: the attachment generation must be
+fresh. This avoids the possibility of a stale pageserver incorrectly
+thinking than an object written by a newer generation is stale, and deleting
+it.
+
+It is not strictly necessary that scrubbing be done by an attached
+pageserver: it could also be done externally. However, an external
+scrubber would still require the same validation procedure that
+a pageserver's deletion queue performs, before actually erasing
+objects.
+
+## Operational impact
+
+### Availability
+
+Coordination of generation numbers via the control plane introduce a dependency for certain
+operations:
+
+1. Starting new pageservers (or activating pageservers after a restart)
+2. Executing enqueued deletions
+3. Advertising updated `remote_consistent_lsn` to enable WAL trimming
+
+Item 1. would mean that some in-place restarts that previously would have resumed service even if the control plane were
+unavailable, will now not resume service to users until the control plane is available. We could
+avoid this by having a timeout on communication with the control plane, and after some timeout,
+resume service with the previous generation numbers (assuming this was persisted to disk). However,
+this is unlikely to be needed as the control plane is already an essential & highly available component. Also, having a node re-use an old generation number would complicate
+reasoning about the system, as it would break the invariant that a generation number uniquely identifies
+a tenant's attachment to a given pageserver _process_: it would merely identify the tenant's attachment
+to the pageserver _machine_ or its _on-disk-state_.
+
+Item 2. is a non-issue operationally: it's harmless to delay deletions, the only impact of objects pending deletion is
+the S3 capacity cost.
+
+Item 3. could be an issue if safekeepers are low on disk space and the control plane is unavailable for a long time. If this became an issue,
+we could adjust the safekeeper to delete segments from local disk sooner, as soon as they're uploaded to S3, rather than waiting for
+remote_consistent_lsn to advance.
+
+For a managed service, the general approach should be to make sure we are monitoring & respond fast enough
+that control plane outages are bounded in time.
+
+There is also the fact that control plane runs in a single region.
+The latency for distant regions is not a big concern for us because all request types added by this RFC are either infrequent or not in the way of the data path.
+However, we lose region isolation for the operations listed above.
+The ongoing work to split console and control will give us per-region control plane, and all operations in this RFC can be handled by these per-region control planes.
+With that in mind, we accept the trade-offs outlined in this paragraph.
+
+We will also implement an "escape hatch" config generation numbers, where in a major disaster outage,
+we may manually run pageservers with a hand-selected generation number, so that we can bring them online
+independently of a control plane.
+
+### Rollout
+
+Although there is coupling between components, we may deploy most of the new data plane components
+independently of the control plane: initially they can just use a static generation number.
+
+#### Phase 1
+
+The pageserver is deployed with some special config to:
+
+- Always act like everything is generation 1 and do not wait for a control plane issued generation on attach
+- Skip the places in deletion and remote_consistent_lsn updates where we would call into control plane
+
+#### Phase 2
+
+The control plane changes are deployed: control plane will now track and increment generation numbers.
+
+#### Phase 3
+
+The pageserver is deployed with its control-plane-dependent changes enabled: it will now require
+the control plane to service re-attach requests on startup, and handle generation
+validation requests.
+
+### On-disk backward compatibility
+
+Backward compatibility with existing data is straightforward:
+
+- When reading the index, we may assume that any layer whose metadata doesn't include
+  generations will have a path without generation suffix.
+- When locating the index file on attachment, we may use the "fallback" listing path
+  and if there is only an index without generation suffix, that is the one we load.
+
+It is not necessary to re-write existing layers: even new index files will be able
+to represent generation-less layers.
+
+### On-disk forward compatibility
+
+We will do a two phase rollout, probably over multiple releases because we will naturally
+have some of the read-side code ready before the overall functionality is ready:
+
+1. Deploy pageservers which understand the new index format and generation suffixes
+   in keys, but do not write objects with generation numbers in the keys.
+2. Deploy pageservers that write objects with generation numbers in the keys.
+
+Old pageservers will be oblivious to generation numbers. That means that they can't
+read objects with generation numbers in the name. This is why we must
+first step must deploy the ability to read, before the second step
+starts writing them.
+
+# Frequently Asked Questions
+
+## Why a generation _suffix_ rather than _prefix_?
+
+The choice is motivated by object listing, since one can list by prefix but not
+suffix.
+
+In [finding remote indices](#finding-the-remote-indices-for-timelines), we rely
+on being able to do a prefix listing for `<tenant>/<timeline>/index_part.json*`.
+That relies on the prefix listing.
+
+The converse case of using a generation prefix and listing by generation is
+not needed: one could imagine listing by generation while scrubbing (so that
+a particular generation's layers could be scrubbed), but this is not part
+of normal operations, and the [scrubber](#cleaning-up-orphan-objects-scrubbing) probably won't work that way anyway.
+
+## Wouldn't it be simpler to have a separate deletion queue per timeline?
+
+Functionally speaking, we could. That's how RemoteTimelineClient currently works,
+but this approach does not map well to a long-lived persistent queue with
+generation validation.
+
+Anything we do per-timeline generates tiny random I/O, on a pageserver with
+tens of thousands of timelines operating: to be ready for high scale, we should:
+
+- A) Amortize costs where we can (e.g. a shared deletion queue)
+- B) Expect to put tenants into a quiescent state while they're not
+  busy: i.e. we shouldn't keep a tenant alive to service its deletion queue.
+
+This was discussed in the [scope](#scope) part of the deletion queue section.
+
+# Appendix A: Examples of use in high availability/failover
+
+The generation numbers proposed in this RFC are adaptable to a variety of different
+failover scenarios and models. The sections below sketch how they would work in practice.
+
+### In-place restart of a pageserver
+
+"In-place" here means that the restart is done before any other element in the system
+has taken action in response to the node being down.
+
+- After restart, the node issues a re-attach request to the control plane, and
+  receives new generation numbers for all its attached tenants.
+- Tenants may be activated with the generation number in the re-attach response.
+- If any of its attachments were in fact stale (i.e. had be reassigned to another
+  node while this node was offline), then
+  - the re-attach response will inform the tenant about this by not including
+    the tenant of this by _not_ incrementing the generation for that attachment.
+  - This will implicitly block deletions in the tenant, but as an optimization
+    the pageserver should also proactively stop doing S3 uploads when it notices this stale-generation state.
+  - The control plane is expected to eventually detach this tenant from the
+    pageserver.
+
+If the control plane does not include a tenant in the re-attach response,
+but there is still local state for the tenant in the filesystem, the pageserver
+deletes the local state in response and does not load/active the tenant.
+See the [earlier section on pageserver startup](#pageserver-attachstartup-changes) for details.
+Control plane can use this mechanism to clean up a pageserver that has been
+down for so long that all its tenants were migrated away before it came back
+up again and asked for re-attach.
+
+### Failure of a pageserver
+
+In this context, read "failure" as the most ambiguous possible case, where
+a pageserver is unavailable to clients and control plane, but may still be executing and talking
+to S3.
+
+#### Case A: re-attachment to other nodes
+
+1. Let's say node 0 becomes unresponsive in a cluster of three nodes 0, 1, 2.
+2. Some external mechanism notices that the node is unavailable and initiates
+   movement of all tenants attached to that node to a different node according
+   to some distribution rule.
+   In this example, it would mean incrementing the generation
+   of all tenants that were attached to node 0, as each tenant's assigned pageserver changes.
+3. A tenant which is now attached to node 1 will _also_ still be attached to node
+   0, from the perspective of node 0. Node 0 will still be using its old generation,
+   node 1 will be using a newer generation.
+4. S3 writes will continue from nodes 0 and 1: there will be an index_part.json-00000001
+   \_and\* an index_part.json-00000002. Objects written under the old suffix
+   after the new attachment was created do not matter from the rest of the system's
+   perspective: the endpoints are reading from the new attachment location. Objects
+   written by node 0 are just garbage that can be cleaned up at leisure. Node 0 will
+   not do any deletions because it can't synchronize with control plane, or if it could,
+   its deletion queue processing would get errors for the validation requests.
+
+#### Case B: direct node replacement with same node_id and drive
+
+This is the scenario we would experience if running pageservers in some dynamic
+VM/container environment that would auto-replace a given node_id when it became
+unresponsive, with the node's storage supplied by some network block device
+that is attached to the replacement VM/container.
+
+1. Let's say node 0 fails, and there may be some other peers but they aren't relevant.
+2. Some external mechanism notices that the node is unavailable, and creates
+   a "new node 0" (Node 0b) which is a physically separate server. The original node 0
+   (Node 0a) may still be running, because we do not assume the environment fences nodes.
+3. On startup, node 0b re-attaches and gets higher generation numbers for
+   all tenants.
+4. S3 writes continue from nodes 0a and 0b, but the writes do not collide due to different
+   generation in the suffix, and the writes from node 0a are not visible to the rest
+   of the system because endpoints are reading only from node 0b.
+
+# Appendix B: interoperability with other features
+
+## Sharded Keyspace
+
+The design in this RFC maps neatly to a sharded keyspace design where subsets of the key space
+for a tenant are assigned to different pageservers:
+
+- the "unit of work" for attachments becomes something like a TenantShard rather than a Tenant
+- TenantShards get generation numbers just as Tenants do.
+- Write workload (ingest, compaction) for a tenant is spread out across pageservers via
+  TenantShards, but each TenantShard still has exactly one valid writer at a time.
+
+## Read replicas
+
+_This section is about a passive reader of S3 pageserver state, not a postgres
+read replica_
+
+For historical reads to LSNs below the remote persistent LSN, any node may act as a reader at any
+time: remote data is logically immutable data, and the use of deferred deletion in this RFC helps
+mitigate the fact that remote data is not _physically_ immutable (i.e. the actual data for a given
+page moves around as compaction happens).
+
+A read replica needs to be aware of generations in remote data in order to read the latest
+metadata (find the index_part.json with the latest suffix). It may either query this
+from the control plane, or find it with ListObjectsv2 request
+
+## Seamless migration
+
+To make tenant migration totally seamless, we will probably want to intentionally double-attach
+a tenant briefly, serving reads from the old node while waiting for the new node to be ready.
+
+This RFC enables that double-attachment: two nodes may be attached at the same time, with the migration destination
+having a higher generation number. The old node will be able to ingest and serve reads, but not
+do any deletes. The new node's attachment must also avoid deleting layers that the old node may
+still use. A new piece of state
+will be needed for this in the control plane's definition of an attachment.
+
+## Warm secondary locations
+
+To enable faster tenant movement after a pageserver is lost, we will probably want to spend some
+disk capacity on keeping standby locations populated with local disk data.
+
+There's no conflict between this RFC and that: implementing warm secondary locations on a per-tenant basis
+would be a separate change to the control plane to store standby location(s) for a tenant. Because
+the standbys do not write to S3, they do not need to be assigned generation numbers. When a tenant is
+re-attached to a standby location, that would increment the tenant attachment generation and this
+would work the same as any other attachment change, but with a warm cache.
+
+## Ephemeral node IDs
+
+This RFC intentionally avoids changing anything fundamental about how pageservers are identified
+and registered with the control plane, to avoid coupling the implementation of pageserver split
+brain protection with more fundamental changes in the management of the pageservers.
+
+Moving to ephemeral node IDs would provide an extra layer of
+resilience in the system, as it would prevent the control plane
+accidentally attaching to two physical nodes with the same
+generation, if somehow there were two physical nodes with
+the same node IDs (currently we rely on EC2 guarantees to
+eliminate this scenario). With ephemeral node IDs, there would be
+no possibility of that happening, no matter the behavior of
+underlying infrastructure.
+
+Nothing fundamental in the pageserver's handling of generations needs to change to handle ephemeral node IDs, since we hardly use the
+`node_id` anywhere. The `/re-attach` API would be extended
+to enable the pageserver to obtain its ephemeral ID, and provide
+some correlation identifier (e.g. EC instance ID), to help the
+control plane re-attach tenants to the same physical server that
+previously had them attached.
--- a/docs/rfcs/026-pageserver-s3-mvcc.md
+++ b/docs/rfcs/026-pageserver-s3-mvcc.md
@@ -0,0 +1,316 @@
+This is a copy from the [original Notion page](https://www.notion.so/neondatabase/Proposal-Pageserver-MVCC-S3-Storage-8a424c0c7ec5459e89d3e3f00e87657c?pvs=4), taken on 2023-08-16.
+
+This is for archival mostly.
+The RFC that we're likely to go with is https://github.com/neondatabase/neon/pull/4919.
+
+---
+
+# Proposal: Pageserver MVCC S3 Storage
+
+tl;dr: this proposal enables Control Plane to attach a tenant to a new pageserver without being 100% certain that it has been detached from the old pageserver. This enables us to automate failover if a pageserver dies (no human in the loop).
+
+# Problem Statement
+
+The current Neon architecture requires the Control Plane to guarantee that a tenant is only attached to one pageserver at a time. If a tenant is attached to multiple pageservers simultaneously, the pageservers will overwrite each other’s changes in S3 for that tenant, resulting in data loss for that tenant.
+
+The above imposes limitations on tenant relocation and future designs for high availability. For instance, Control Plane cannot relocate a tenant to another pageserver before it is 100% certain that the tenant is detached from the source pageserver. If the source pageserver is unresponsive, the tenant detach procedure cannot proceed, and Control Plane has no choice but to wait for either the source to become responsive again, or rely on a node failure detection mechanism to detect that the source pageserver is dead, and give permission to skip the detachment step. Either way, the tenant is unavailable for an extended period, and we have no means to improve it in the current architecture.
+
+Note that there is no 100% correct node failure detection mechanism, and even techniques to accelerate failure detection, such as ********************************shoot-the-other-node-in-the-head,******************************** have their limits. So, we currently rely on humans as node failure detectors: they get alerted via PagerDuty, assess the situation under high stress, and make the decision. If they make the wrong call, or the apparent dead pageserver somehow resurrects later, we’ll have data loss.
+
+Also, by relying on humans, we’re [incurring needless unscalable toil](https://sre.google/sre-book/eliminating-toil/): as Neon grows, pageserver failures will become more and more frequent because our fleet grows. Each instance will need quick response time to minimize downtime for the affected tenants, which implies higher toil, higher resulting attrition, and/or higher personnel cost.
+
+Lastly, there are foreseeable needs by operation and product such as zero-downtime relocation and automatic failover/HA. For such features, the ability to have a tenant purposefully or accidentally attached to more than one pageserver will greatly reduce risk of data loss, and improve availability.
+
+# High-Level Idea
+
+The core idea is to evolve the per-Tenant S3 state to an MVCC-like scheme, allowing multiple pageservers to operate on the same tenant S3 state without interference. To make changes to S3, pageservers acquire long-running transactions from Control Plane. After opening a transaction, Pageservers make PUTs directly against S3, but they keys include the transaction ID,  so overwrites never happen. Periodically, pageservers talk back to Control Plane to commit their transaction. This is where Control Plane enforces strict linearizability, favoring availability over work-conservation: commit is only granted if no transaction started after the one that’s requesting commit. Garbage collection is done through deadlists, and it’s simplified tremendously by above commit grant/reject policy.
+
+Minimal changes are required for safekeepers to allow WAL for a single timeline be consumed by more than one pageserver without premature truncation.
+
+**Above scheme makes it safe to attach tenants without a 100% correct node failure detection mechanism. Further, it makes it safe to interleave tenant-attachment to pageservers, unlocking new capabilities for (internal) product features:**
+
+- **Fast, Zero-Toil Failover on Network Partitions or Instance Failure**: if a pageserver is not reachable (network partition, hardware failure, overload) we want to spread its attached tenants to new pageservers to restore availability, within the range of *seconds*. We cannot afford gracious timeouts to maximize the probability that the unreachable pageserver has ceased writing to S3. This proposal enables us to attach the tenants to the replacement pageservers,  and redirect their computes, without having to wait for confirmation that the unreachable pageserver has ceased writing to S3.
+- **************************************Zero-Downtime Relocation:************************************** we want to be able to relocate tenants to different pageservers with minimized availability or a latency impact. This proposal enables us to attach the relocating Tenant to the destination Pageserver before detaching it from the source Pageserver. This can help minimize downtime because we can wait for the destination to catch up on WAL processing before redirecting Computes.
+
+# Design
+
+The core idea is to evolve the per-Tenant S3 state to a per-tenant MVCC-like scheme.
+
+To make S3 changes for a given tenant, Pageserver requests a transaction ID from control plane for that tenant. Without a transaction ID, Pageserver does not write to S3.
+
+Once Pageserver received a transaction ID it is allowed to produce new objects and overwrite objects created in this transaction. Pageserver is not allowed to delete any objects; instead, it marks the object as deleted by appending the key to the transaction’s deadlist for later deletion. Commits of transactions are serialized through Control Plane: when Pageserver wants to commit a transaction, it sends an RPC to Control Plane. Control Plane responds with a commit grant or commit reject message. Commit grant means that the transaction’s changes are now visible to subsequent transactions. Commit reject means that the transaction’s changes are not and never will be visible to another Pageserver instance, and the rejected Pageserver is to cease further activity on that tenant.
+
+## ****************************************************Commit grant/reject policy****************************************************
+
+For the purposes of Pageserver, we want **linearizability** of a tenant’s S3 state. Since our transactions are scoped per tenant, it is sufficient for linearizability to grant commit if and only if no other transaction has been started since the commit-requesting transaction started.
+
+For example, consider the case of a single tenant, attached to Pageserver A. Pageserver A has an open transaction but becomes unresponsive. Control Plane decides to relocate the tenant to another Pageserver B. It need *not* wait for A to be 100%-certainly down before B can start uploading to S3 for that tenant. Instead, B can start a new transaction right away, make progress, and get commit grants;  What about A? The transaction is RejectPending in Control Plane until A eventually becomes responsive again, tries to commit, gets a rejection, acknowledges it, and thus its transaction becomes RejectAcknowledge. If A is definitively dead, operator can also force-transition from state RejectPending to RejectAcknowledged. But critically, Control Plane doesn’t have for A’s transaction to become RejectAcknowledge before attaching the tenant to B.
+
+```mermaid
+sequenceDiagram
+
+   participant CP
+   participant A
+   participant S3
+   participant B
+
+	 CP -->> A: attach tenant
+   activate A
+	 A -->> CP: start txn
+	 CP -->> A: txn=23, last_committed_txn=22
+
+
+	 Note over CP,A: network partition
+	 CP --x A: heartbeat
+	 CP --x A: heartbeat
+
+	 Note over CP: relocate tenant to avoid downtime
+	 CP -->> B: attach tenant
+   activate B
+	 B -->> CP: start txn
+   Note over CP: mark A's txn 23 as RejectPending
+	 CP -->> B: txn=24, last-committed txn is 22
+	 B -->> S3: PUT X.layer.24<br>PUT index_part.json.24 referencing X.layer.24
+	 B -->> CP: request commit
+	 CP -->> B: granted
+   B -->> CP: start txn
+  CP -->> B: txn=25, last_committed_txn=22
+
+   A -->> S3: PUT Y.layer.23 <br> PUT index_part.json.23 referencing Y.layer.23
+  A --x CP: request commit
+	 A --x CP: request commit
+
+   Note over CP,A: partition is over
+
+   A -->> CP: request commit
+
+   Note over CP: most recently started txn is 25, not 23, reject
+
+   CP -->> A: reject
+   A -->> CP: acknowledge reject
+
+   Note over CP: mark A's txn 23 as RejectAcknowledged
+
+  deactivate A
+
+  B -->> S3: PUT 000-FFF_X-Y.layer.**************25**************<br>...
+
+  deactivate B
+
+
+```
+
+If a Pageserver gets a rejection to a commit request, it acknowledges rejection and cedes further S3 uploads for the tenant, until it receives a `/detach` request for the tenant (control plane has most likely attached the tenant to another pageserver in the meantime).
+
+In practice, Control Plane will probably extend the commit grant/reject schema above, taking into account the pageserver to which it last attached the tenant. In the above example, Control Plane could remember that the pageserver that is supposed to host the tenant is pageserver B, and reject start-txn and commit requests from pageserver A. It would also use such requests from A as a signal that A is reachable again, and retry the `/detach` .
+
+<aside>
+💡 A commit failure causes the tenant to become effectively `Broken`. Pageserver should persist this locally so it doesn’t bother ControlPlane for a new txn when Pageserver is restarted.
+
+</aside>
+
+## ********************Visibility********************
+
+We mentioned earlier that once a transaction commits, its changes are visible to subsequent transactions. But how does a given transaction know where to look for the data? There is no longer a single `index_part.json` per timeline, or a single `timelines/:timeline_id` prefix to look for; they’re all multi-versioned, suffixed by the txn number.
+The solution is: at transaction start, Pageserver receives the last-committed transaction ID from Control Plane (`last_committed_txn` in the diagram). last_commited_txn is the upper bound for what is visible for the current transaction. Control Plane keeps track of each open transaction’s last_committed_txn for purposes of garbage collection (see later paragraph).
+Equipped with last_committed_txn, Pageserver then discovers
+
+- the current index part of a timeline at `tenants/:tenant_id/timelines/:timeline_id/index_part.json.$last_committed_txn`. The `index_part.json.$last_committed_txn` has the exact same contents as the current architecture’s index_part.json, i.e. full list of layers.
+- the list of existent timelines as part of the `attach` RPC from CP;
+
+There is no other S3 state per tenant, so, that’s all the visibility required.
+An alternative to receiving the list of existent timelines from CP is to introduce a proper **********SetOfTimelines********** object in S3, and multi-version it just like above. For example, we could have a `tenants/:tenant_id/timelines.json.$txn` file that references `index_part.json.$last_committed_txn` . It can be added later if more separation between CP and PS is desired.
+
+So, the only MVCC’ed object types in this proposal are LayerFile and IndexPart (=individual timeline), but not the SetOfTimelines in a given tenant. Is this a problem? For example, the Pageserver’s garbage collection code needs to know the full set of timelines of a tenant. Otherwise it’ll make incorrect decisions. What if Pageserver A knows about timelines {R,S}, but another Pageserver B created an additional branch T, so, its set of timelines is {R,S,T}. Both pageservers will run GC code, and so, PS A may decide to delete a layer that’s still needed for branch T. Not a problem with this propsoal, because the effect of GC (i.e., layer deletion) is properly MVCC’ed.
+
+## Longevity Of Transactions & Availability
+
+Pageserver depends on Control Plane to start a new transaction. If ControlPlane is down, no new transactions can be started.
+
+Pageservers commit transactions based on a maximum amount of uncommitted changes that have accumulated in S3. A lower maximum increases dependence and load on ControlPlane which decreases availability. A higher maximum risks losing more work in the event of failover; the work will have to be re-done in a new transaction on the new node.
+
+Pageservers are persist the open txn id in local storage, so that they can resume the transaction after restart, without dependence on Control Plane.
+
+## **Operations**
+
+********PUTs:********
+
+- **layer files**
+    - current architecture: layer files are supposed to be write-once, but actually, there are edge-cases where we PUT the same layer file name twice; namely if we PUT the file to S3 but crash before uploading the index part that references it; then detach + attach, and re-run compaction, which is non-deterministic.
+    - this proposal: with transactions, we can now upload layers and index_part.json concurrently, just need to make sure layer file upload is done before we request txn commit.
+- **index part** upload: `index_part.json.$txn` may be created and subsequently overwritten multiple times in a transaction; it is an availability/work-loss trade-off how often to request a commit from CP.
+
+**************DELETEs**************: for deletion, we maintain a deadlist per transaction. It is located at `tenants/:tenant_id/deadlist/deadlist.json.$txn`. It is PUT once before the pageserver requests requests commit, and not changed after sending request to commit. An object created in the current txn need not (but can) be on the deadlist — it can be DELETEd immediately because it’s not visible to other transactions. An example use case would be an L0 layer that gets compacted within one transaction; or, if we ever start MVCC’ing the set of timelines of a tenant, a short-lived branch that is created & destroyed within one transaction.
+
+<aside>
+☝ **Deadlist Invariant:** if a an object is on a deadlist of transaction T, it is not referenced from anywhere else in the full state visible to T or any later started transaction > T.
+
+</aside>
+
+### Rationale For Deadlist.json
+
+Given that this proposal only MVCC’s layers and indexparts, one may ask why the deadlist isn’t part of indexpart. The reason is to not lose generality: the deadlist is just a list of keys; it is not necessary to understand the data format of the versioned object to process the deadlist. This is important for garbage collection / vacuuming, which we’ll come to in the next section.
+
+## Garbage Collection / Vacuuming
+
+After a transaction has reached reject-acknowledged state,  Control Plane initiates a garbage collection procedure for the aborted transaction.
+
+Control Plane is in the unique position about transaction states. Here is a sketch of the exact transaction states and what Control Plane keeps track of.
+
+```
+struct Tenant {
+  ...
+
+  txns: HashMap<TxnId, Transaction>,
+  // the most recently started txn's id; only most recently sarted can win
+  next_winner_txn: Option<TxnId>,
+}
+struct Transaction {
+  id: TxnId, // immutable
+  last_committed_txn: TxnId, // immutable; the most recent txn in state `Committed`
+                             // when self was started
+  pageserver_id: PageserverId,
+  state: enum {
+    Open,
+    Committed,
+    RejectPending,
+    RejectAcknowledged, // invariant: we know all S3 activity has ceded
+    GarbageCollected,
+  }
+}
+```
+
+Object creations & deletions by a rejected transaction have never been visible to other transactions. That is true for both RejectPending and RejectAcknowledged states. The difference is that, in RejectPending, the pageserver may still be uploading to S3, whereas in RejectAcknowledged, Control Plane can be certain that all S3 activity in the name of that transaction has ceded. So, once a transaction reaches state RejectAcknowledged state, it is safe to DELETE all objects created by that transaction, and discard the transaction’s deadlists.
+
+A transaction T in state Committed has subsequent transactions that may or may not reference the objects it created. None of the subsequent transaction can reference the objects on T’s deadlist, though, as per the Deadlist Invariant (see previous section).
+
+So, for garbage collection, we need to assess transactions in state Committed and RejectAcknowledged:
+
+- Commited: delete objects on the deadlist.
+    - We don’t need a LIST request here, the deadlist is sufficient. So, it’s really cheap.
+    - This is **not true MVCC garbage collection**; by deleting the objects on Committed transaction T ’s deadlist, we might delete data referenced by other transactions that were concurrent with T, i.e., they started while T was still open. However, the fact that T is committed means that the other transactions are RejectPending or RejectAcknowledged, so, they don’t matter. Pageservers executing these doomed RejectPending transactions must handle 404 for GETs gracefully, e.g., by trying to commit txn so they observe the rejection they’re destined to get anyways. 404’s for RejectAcknowledged is handled below.
+- RejectAcknowledged: delete all objects created in that txn, and discard deadlists.
+    - 404s / object-already-deleted type messages must be expected because of Committed garbage collection (see above)
+    - How to get this list of objects created in a txn? Open but solvable design question; Ideas:
+        - **Brute force**: within tenant prefix, search for all keys ending in `.$txn` and delete them.
+        - **WAL for PUTs**: before a txn PUTs an object, it logs to S3, or some other equivalently durable storage, that it’s going to do it. If we log to S3, this means we have to do an additional WAL PUT per “readl” PUT.
+        - ******************************LIST with reorg’ed S3 layout (preferred one right now):****************************** layout S3 key space such that `$txn` comes first, i.e., `tenants/:tenant_id/$txn/timelines/:timeline_id/*.json.$txn` . That way, when we need to GC a RejectAcknowledged txn, we just LIST the entire `tenants/:tenant_id/$txn` prefix and delete it. The cost of GC for RejectAcknowledged transactions is thus proportional to the number of objects created in that transaction.
+
+## Branches
+
+This proposal only MVCC’s layer files and and index_part.json, but leaves the tenant object not-MVCCed. We argued earlier that it’s fine to ignore this for now, because
+
+1. Control Plane can act as source-of-truth for the set of timelines, and
+2. The only operation that makes decision based on “set of timelines” is GC, which in turn only does layer deletions, and layer deletions ***are*** properly MVCC’ed.
+
+Now that we’ve introduced garbage collection, let’s elaborate a little more on (2). Recall our example from earlier: Pageserver A knows about timelines {R,S}, but another Pageserver B created an additional branch T, so, its set of timelines is {R,S,T}. Both pageservers will run GC code, and so, PS A may decide to delete a layer that’s still needed for branch T.
+
+How does the MVCC’ing of layer files protect us here? If A decides to delete that layer, it’s just on A’s transaction’s deadlist, but still present in S3 and usable by B. If A commits first, B won’t be able to commit and the layers in timeline T will be vacuumed. If B commits first, A’s deadlist is discarded and the layer continues to exist.
+
+## Safekeeper Changes
+
+We need to teach the safekeepers that there can be multiple pageservers requesting WAL for the same timeline, in order to prevent premature WAL truncation.
+
+In the current architecture, the Safekeeper service currently assumes only one Pageserver and is allowed to prune WAL older than that Pageserver’s `remote_consistent_lsn`. Safekeeper currently learns the `remote_consistent_lsn` through the walreceiver protocol.
+
+So, if we have a tenant attached to two pageservers at the same time, they will both try to stream WAL and the Safekeeper will get confused about which connection’s `remote_consistent_lsn` to use as a basis for WAL pruning.
+
+What do we need to change to make it work? We need to make sure that the Safekeepers only prune WAL up to the `remote_consistent_lsn` of the last-committed transaction.
+
+The straight-forward way to get it is to re-design WAL pruning as follows:
+
+1. Pageserver reports remote_consistent_lsn as part of transaction commit to Control Plane.
+2. Control Plane makes sure transaction state update is persisted.
+3. Control Plane (asynchronous to transaction commit) reconciles with Safekeepers to ensure WAL pruning happens.
+
+The above requires non-trivial changes, but, in the light of other planned projects such as restore-tenant-from-safekeeper-wal-backups, I think Control Plane will need to get involved in WAL pruning anyways.
+
+# How This Proposal Unlocks Future Features
+
+Let us revisit the example from the introduction where we were thinking about handling network partitions. Network partitions need to be solved first, because they’re unavoidable in distributed systems. We did that. Now let’s see how we can solve actual product problems:
+
+## **Fast, Zero-Toil Failover on Network Partitions or Instance Failure**
+
+The “Problem Statement” section outlined the current architecture’s problems with regards to network partitions or instance failure: it requires a 100% correct node-dead detector to make decisions, which doesn’t exist in reality. We rely instead on human toil: an oncall engineer has to inspect the situation and make a decision, which may be incorrect and in any case take time in the order of minutes, which means equivalent downtime for users.
+
+With this proposal, automatic failover for pageservers is trivial:
+
+If a pageserver is unresponsive from Control Plane’s / Compute’s perspective, Control Plane does the following:
+
+- attach all tenants of the unresponsive pageserver to new pageservers
+- switch over these tenants’ computes immediately;
+
+At this point, availability is restored and user pain relieved.
+
+What’s left is to somehow close the doomed transaction of the unresponsive pageserver, so that it beomes RejectAcknowledged, and GC can make progress. Since S3 is cheap, we can afford to wait a really long time here, especially if we put a soft bound on the amount of data a transaction may produce before it must commit. Procedure:
+
+1. Ensure the unresponsive pageserver is taken out of rotation for new attachments. That probably should happen as part of the routine above.
+2. Make a human operator investigate decide what to do (next morning, NO ONCALL ALERT):
+    1. Inspect the instance, investigate logs, understand root cause.
+    2. Try to re-establish connectivity between pageserver and Control Plane so that pageserver can retry commits, get rejected, ack rejection ⇒ enable GC.
+    3. Use below procedure to decomission pageserver.
+
+### Decomissioning A Pageserver (Dead or Alive-but-Unrespsonive)
+
+The solution, enabled by this proposal:
+
+1. Ensure that pageserver’s S3 credentials are revoked so that it cannot make new uploads, which wouldn’t be tracked anywhere.
+2. Let enough time pass for the S3 credential revocation to propagate. Amazon doesn’t give a guarantee here. As stated earlier, we can easily afford to wait here.
+3. Mark all Open and RejectPending transactions of that pageserver as RejectAcknowledge.
+
+Revocation of the S3 credentials is required so that, once we transition all the transactions of that pageserver to RejectAcknowledge, once garbage-collection pass is guaranteed to delete all objects that will ever exist for that pageserver. That way, we need not check *****GarbageCollected***** transactions every again.
+
+## Workflow: Zero-Downtime Relocation
+
+With zero-downtime relocation, the goal is to have the target pageserver warmed up, i.e., at the same `last_record_lsn` as the source pageserver, before switching over Computes from source to target pageserver.
+
+With this proposal, it works like so:
+
+1. Grant source pageserver its last open transaction. This one is doomed to be rejected later, unless the relocation fails.
+2. Grant target pageserver its first open transaction.
+3. Have target pageserver catch up on WAL, streaming from last-committed-txn’s remote_consistent_lsn onwards.
+4. Once target pageserver reports `last_record_lsn` close enough to source pageserver, target pageserver requests commit.
+5. Drain compute traffic from source to target pageserver. (Source can still answer requests until it tries to commit and gets reject, so, this will be quite smooth).
+
+Note that as soon as we complete step (4), the source pageserver’s transaction is doomed to be rejected later. Conversely, if the target can’t catch up fast enough, the source will make a transaction commit earlier. This will generally happen if there is a lot of write traffic coming in. The design space to make thing smooth here is large, but well explored in other areas of computing, e.g., VM live migration. We have all the important policy levers at hand, e.g.,
+
+- delaying source commits if we see target making progress
+- slowing down source consumption (need some signalling mechanism for it)
+- slowing down compute wal generation
+- …
+
+It doesn’t really matter, what’s important is that two pageservers can overlap.
+
+# Additional Trade-Offs / Remarks Brought Up During Peer Review
+
+This proposal was read by and discussed @Stas and @Dmitry Rodionov prior to publishing it with the broader team. (This does not mean they endorse this proposal!).
+
+Issues that we discussed:
+
+1. **Frequency of transactions:** If even idle tenants commit every 10min or so, that’s quite a lot of load on Control Plane. Can we minimize it by Equating Transaction Commit Period to Attachment Period? I.e. start txn on attach, commit on detach?
+    1. Would be nice, but, if a tenant is attached for 1 month, then PS dies, we lose 1 month of work.
+    2. ⇒ my solution to this problem: Adjusted this proposal to make transaction commit frequency proportional to amount of uncommitted data.
+        1. It’s ok to spend resources on active users, they pay us money to do it!
+        2. The amount of work per transaction is minimal.
+            1. In current Control Plane, it’s a small database transaction that is super unlikely to conflict with other transactions.
+            2. I have very little concerns about scalability of the commit workload on CP side because it's trivially horizontally scalable by sharding by tenant.
+        3. There's no super stringent availability requirement on control plane; if a txn can't commit because it can't reach the CP, PS can continue & retry in the background, speculating that it's CP downtime and not PS-partitioned-off scenario.
+        4. Without stringent availability requirement, there's flexibility for future changes to CP-side-implementation.
+2. ************************************************Does this proposal address mirroring / no-performance-degradation failover ?************************************************
+    1. No it doesn’t. It only provides the building block for attaching a tenant to a new pageserver without having to worry that the tenant is detached on the old pageserver.
+    2. A simple scheme to build no-performance-degradation failover on top of this proposal is to have an asynchronous read-only replica of a tenant on another pageserver in the same region.
+    3. Another more ambitious scheme to get no-performance-degradation would be [One-Pager: Layer File Spreading (Christian)](https://www.notion.so/One-Pager-Layer-File-Spreading-Christian-eb6b64182a214e11b3fceceee688d843?pvs=21); this proposal would be used in layer file spreading for risk-free automation of TenantLeader failover, which hasn’t been addressed Ithere.
+    4. In any way, failover would restart from an older S3 state, and need to re-ingest WAL before being able to server recently written pages.
+        1. Is that a show-stopper? I think not.
+        2. Is it suboptimal? Absolutely: if a pageserver instance fails, all its tenants will be distributed among the remaining pageservers (OK), and all these tenants will ask the safekeepers for WAL at the same time (BAD). So, pageserver instance failure will cause a load spike in safekeepers.
+            1. Personally I think that’s an OK trade-off to make.
+            2. There are countless options to avoid / mitigate the load spike. E.g., pro-actively streaming WAL to the standby read-only replica.
+
+3. ********************************************Does this proposal allow multiple writers for a tenant?********************************************
+    1. In abstract terms, this proposal provides a linearized history for a given S3 prefix.
+    2. In concrete terms, this proposal provides a linearized history per tenant.
+    3. There can be multiple writers at a given time, but only one of them will win to become part of the linearized history.
+4. ************************************************************************************Alternative ideas mentioned during meetings that should be turned into a written prospoal like this one:************************************************************************************
+    1. @Dmitry Rodionov : having linearized storage of index_part.json in some database that allows serializable transactions / atomic compare-and-swap PUT
+    2. @Dmitry Rodionov :
+    3. @Stas : something like this scheme, but somehow find a way to equate attachment duration with transaction duration, without losing work if pageserver dies months after attachment.
--- a/docs/rfcs/027-crash-consistent-layer-map-through-index-part.md
+++ b/docs/rfcs/027-crash-consistent-layer-map-through-index-part.md
@@ -0,0 +1,281 @@
+
+# Crash-Consistent Layer Map Updates By Leveraging `index_part.json`
+
+* Created on: Aug 23, 2023
+* Author: Christian Schwarz
+
+## Summary
+
+This RFC describes a simple scheme to make layer map updates crash consistent by leveraging the `index_part.json` in remote storage.
+Without such a mechanism, crashes can induce certain edge cases in which broadly held assumptions about system invariants don't hold.
+
+## Motivation
+
+### Background
+
+We can currently easily make complex, atomic updates to the layer map by means of an RwLock.
+If we crash or restart pageserver, we reconstruct the layer map from:
+1. local timeline directory contents
+2. remote `index_part.json` contents.
+
+The function that is responsible for this is called `Timeline::load_layer_map()`.
+The reconciliation process's behavior is the following:
+* local-only files will become part of the layer map as local-only layers and rescheduled for upload
+* For a file name that, by its name, is present locally and in the remote `index_part.json`, but where the local file has a different size (future: checksum) than the remote file, we will delete the local file and leave the remote file as a `RemoteLayer` in the layer map.
+
+### The Problem
+
+There are are cases where we need to make an atomic update to the layer map that involves **more than one layer**.
+The best example is compaction, where we need to insert the L1 layers generated from the L0 layers, and remove the L0 layers.
+As stated above, making the update to the layer map in atomic way is trivial.
+But, there is no system call API to make an atomic update to a directory that involves more than one file rename and deletion.
+Currently, we issue the system calls one by one and hope we don't crash.
+
+What happens if we crash and restart in the middle of that system call sequence?
+We will reconstruct the layer map according to the reconciliation process, taking as input whatever transitory state the timeline directory ended up in.
+
+We cannot roll back or complete the timeline directory update during which we crashed, because we keep no record of the changes we plan to make.
+
+### Problem's Implications For Compaction
+
+The implications of the above are primarily problematic for compaction.
+Specifically, the part of it that compacts L0 layers into L1 layers.
+
+Remember that compaction takes a set of L0 layers and reshuffles the delta records in them into L1 layer files.
+Once the L1 layer files are written to disk, it atomically removes the L0 layers from the layer map and adds the L1 layers to the layer map.
+It then deletes the L0 layers locally, and schedules an upload of the L1 layers and and updated index part.
+
+If we crash before deleting L0s, but after writing out L1s, the next compaction after restart will re-digest the L0s and produce new L1s.
+This means the compaction after restart will **overwrite** the previously written L1s.
+Currently we also schedule an S3 upload of the overwritten L1.
+
+If the compaction algorithm doesn't change between the two compaction runs, is deterministic, and uses the same set of L0s as input, then the second run will produce identical L1s and the overwrites will go unnoticed.
+
+*However*:
+1. the file size of the overwritten L1s may not be identical, and
+2. the bit pattern of the overwritten L1s may not be identical, and,
+3. in the future, we may want to make the compaction code non-determinstic, influenced by past access patterns, or otherwise change it, resulting in L1 overwrites with a different set of delta records than before the overwrite
+
+The items above are a problem for the [split-brain protection RFC](https://github.com/neondatabase/neon/pull/4919) because it assumes that layer files in S3 are only ever deleted, but never replaced (overPUTted).
+
+For example, if an unresponsive node A becomes active again after control plane has relocated the tenant to a new node B, the node A may overwrite some L1s.
+But node B based its world view on the version of node A's `index_part.json` from _before_ the overwrite.
+That earlier `index_part.json`` contained the file size of the pre-overwrite L1.
+If the overwritten L1 has a different file size, node B will refuse to read data from the overwritten L1.
+Effectively, the data in the L1 has become inaccessible to node B.
+If node B already uploaded an index part itself, all subsequent attachments will use node B's index part, and run into the same probem.
+
+If we ever introduce checksums instead of checking just the file size, then a mismatching bit pattern (2) will cause similar problems.
+
+In case of (1) and (2), where we know that the logical content of the layers is still the same, we can recover by manually patching the `index_part.json` of the new node to the overwritten L1's file size / checksum.
+
+But if (3) ever happens, the logical content may be different, and, we could have truly lost data.
+
+Given the above considerations, we should avoid making correctness of split-brain protection dependent on overwrites preserving _logical_ layer file contents.
+**It is a much cleaner separation of concerns to require that layer files are truly immutable in S3, i.e., PUT once and then only DELETEd, never overwritten (overPUTted).**
+
+## Design
+
+Instead of reconciling a layer map from local timeline directory contents and remote index part, this RFC proposes to view the remote index part as authoritative during timeline load.
+Local layer files will be recognized if they match what's listed in remote index part, and removed otherwise.
+
+During **timeline load**, the only thing that matters is the remote index part content.
+Essentially, timeline load becomes much like attach, except we don't need to prefix-list the remote timelines.
+The local timeline dir's `metadata` file does not matter.
+The layer files in the local timeline dir are seen as a nice-to-have cache of layer files that are in the remote index part.
+Any layer files in the local timeline dir that aren't in the remote index part are removed during startup.
+The `Timeline::load_layer_map()` no longer "merges" local timeline dir contents with the remote index part.
+Instead, it treats the remote index part as the authoritative layer map.
+If the local timeline dir contains a layer that is in the remote index part, that's nice, and we'll re-use it if file size (and in the future, check sum) match what's stated in the index part.
+If it doesn't match, we remove the file from the local timeline dir.
+
+After load, **at runtime**, nothing changes compared to what we did before this RFC.
+The procedure for single- and multi-object changes is reproduced here for reference:
+* For any new layers that the change adds:
+  * Write them to a temporary location.
+  * While holding layer map lock:
+    * Move them to the final location.
+    * Insert into layer map.
+* Make the S3 changes.
+  We won't reproduce the remote timeline client method calls here because these are subject to change.
+  Instead we reproduce the sequence of s3 changes that must result for a given single-/multi-object change:
+    * PUT layer files inserted by the change.
+    * PUT an index part that has insertions and deletions of the change.
+    * DELETE the layer files that are deleted by the change.
+
+Note that it is safe for the DELETE to be deferred arbitrarily.
+* If it never happens, we leak the object, but, that's not a correctness concern.
+* As of #4938, we don't schedule the remote timeline client operation for deletion immediately, but, only when we drop the `LayerInner`.
+* With the [split-brain protection RFC](https://github.com/neondatabase/neon/pull/4919), the deletions will be written to deletion queue for processing when it's safe to do so (see the RFC for details).
+
+## How This Solves The Problem
+
+If we crash before we've finished the S3 changes, then timeline load will reset layer map to the state that's in the S3 index part.
+The S3 change sequence above is obviously crash-consistent.
+If we crash before the index part PUT, then we leak the inserted layer files to S3.
+If we crash after the index part PUT, we leak the to-be-DELETEd layer files to S3.
+Leaking is fine, it's a pre-existing condition and not addressed in this RFC.
+
+Multi-object changes that previously created and removed files in timeline dir are now atomic because the layer map updates are atomic and crash consistent:
+* atomic layer map update at runtime, currently by using an RwLock in write mode
+* atomic `index_part.json` update in S3, as per guarantee that S3 PUT is atomic
+* local timeline dir state:
+  * irrelevant for layer map content => irrelevant for atomic updates / crash consistency
+  * if we crash after index part PUT, local layer files will be used, so, no on-demand downloads neede for them
+  * if we crash before index part PUT, local layer files will be deleted
+
+## Trade-Offs
+
+### Fundamental
+
+If we crash before finishing the index part PUT, we lose all the work that hasn't reached the S3 `index_part.json`:
+* wal ingest: we lose not-yet-uploaded L0s; load on the **safekeepers** + work for pageserver
+* compaction: we lose the entire compaction iteration work; need to re-do it again
+* gc: no change to what we have today
+
+If the work is still deemed necessary after restart, the restarted restarted pageserver will re-do this work.
+The amount of work to be re-do is capped to the lag of S3 changes to the local changes.
+Assuming upload queue allows for unlimited queue depth (that's what it does today), this means:
+* on-demand downloads that were needed to do the work: are likely still present, not lost
+* wal ingest: currently unbounded
+* L0 => L1 compaction: CPU time proportional to `O(sum(L0 size))` and upload work proportional to `O()`
+  * Compaction threshold is 10 L0s and each L0 can be up to 256M in size. Target size for L1 is 128M.
+  * In practive, most L0s are tiny due to 10minute `DEFAULT_CHECKPOINT_TIMEOUT`.
+* image layer generation: CPU time `O(sum(input data))` + upload work `O(sum(new image layer size))`
+  * I have no intuition how expensive / long-running it is in reality.
+* gc: `update_gc_info`` work (not substantial, AFAIK)
+
+To limit the amount of lost upload work, and ingest work, we can limit the upload queue depth (see suggestions in the next sub-section).
+However, to limit the amount of lost CPU work, we would need a way to make make the compaction/image-layer-generation algorithms interruptible & resumable.
+We aren't there yet, the need for it is tracked by ([#4580](https://github.com/neondatabase/neon/issues/4580)).
+However, this RFC is not constraining the design space either.
+
+### Practical
+
+#### Pageserver Restarts
+
+Pageserver crashes are very rare ; it would likely be acceptable to re-do the lost work in that case.
+However, regular pageserver restart happen frequently, e.g., during weekly deploys.
+
+In general, pageserver restart faces the problem of tenants that "take too long" to shut down.
+They are a problem because other tenants that shut down quickly are unavailble while we wait for the slow tenants to shut down.
+We currently allot 10 seconds for graceful shutdown until we SIGKILL the pageserver process (as per `pageserver.service` unit file).
+A longer budget would expose tenants that are done early to a longer downtime.
+A short budget would risk throwing away more work that'd have to be re-done after restart.
+
+In the context of this RFC, killing the process would mean losing the work that hasn't made it to S3.
+We can mitigate this problem as follows:
+0. initially, by accepting that we need to do the work again
+1. short-term, introducing measures to cap the amount of in-flight work:
+
+   - cap upload queue length, use backpressure to slow down compaction
+   - disabling compaction/image-layer-generation X minutes before `systemctl restart pageserver`
+   - introducing a read-only shutdown state for tenants that are fast to shut down;
+     that state would be equivalent to the state of a tenant in hot standby / readonly mode.
+
+2. mid term, by not restarting pageserver in place, but using [*seamless tenant migration*](https://github.com/neondatabase/neon/pull/5029) to drain a pageserver's tenants before we restart it.
+
+#### `disk_consistent_lsn` can go backwards
+
+`disk_consistent_lsn` can go backwards across restarts if we crash before we've finished the index part PUT.
+Nobody should care about it, because the only thing that matters is `remote_consistent_lsn`.
+Compute certainly doesn't care about `disk_consistent_lsn`.
+
+
+## Side-Effects Of This Design
+
+* local `metadata` is basically reduced to a cache of which timelines exist for this tenant; i.e., we can avoid a `ListObjects` requests for a tenant's timelines during tenant load.
+
+## Limitations
+
+Multi-object changes that span multiple timelines aren't covered by this RFC.
+That's fine because we currently don't need them, as evidenced by the absence
+of a Pageserver operation that holds multiple timelines' layer map lock at a time.
+
+## Impacted components
+
+Primarily pageservers.
+
+Safekeepers will experience more load when we need to re-ingest WAL because we've thrown away work.
+No changes to safekeepers are needed.
+
+## Alternatives considered
+
+### Alternative 1: WAL
+
+We could have a local WAL for timeline dir changes, as proposed here https://github.com/neondatabase/neon/issues/4418 and partially implemented here https://github.com/neondatabase/neon/pull/4422 .
+The WAL would be used to
+1. make multi-object changes atomic
+2. replace `reconcile_with_remote()` reconciliation: scheduling of layer upload would be part of WAL replay.
+
+The WAL is appealing in a local-first world, but, it's much more complex than the design described above:
+* New on-disk state to get right.
+* Forward- and backward-compatibility development costs in the future.
+
+### Alternative 2: Flow Everything Through `index_part.json`
+
+We could have gone to the other extreme and **only** update the layer map whenever we've PUT `index_part.json`.
+I.e., layer map would always be the last-persisted S3 state.
+That's axiomatically beautiful, not least because it fully separates the layer file production and consumption path (=> [layer file spreading proposal](https://www.notion.so/neondatabase/One-Pager-Layer-File-Spreading-Christian-eb6b64182a214e11b3fceceee688d843?pvs=4)).
+And it might make hot standbys / read-only pageservers less of a special case in the future.
+
+But, I have some uncertainties with regard to WAL ingestion, because it needs to be able to do some reads for the logical size feedback to safekeepers.
+
+And it's silly that we wouldn't be able to use the results of compaction or image layer generation before we're done with the upload.
+
+Lastly, a temporarily clogged-up upload queue (e.g. S3 is down) shouldn't immediately render ingestion unavailable.
+
+### Alternative 3: Sequence Numbers For Layers
+
+Instead of what's proposed in this RFC, we could use unique numbers to identify layer files:
+
+```
+# before
+tenants/$tenant/timelines/$timeline/$key_and_lsn_range
+# after
+tenants/$tenant/timelines/$timeline/$layer_file_id-$key_and_lsn_range
+```
+
+To guarantee uniqueness, the unqiue number is a sequence number, stored in `index_part.json`.
+
+This alternative does not solve atomic layer map updates.
+In our crash-during-compaction scenario above, the compaction run after the crash will not overwrite the L1s, but write/PUT new files with new sequence numbers.
+In fact, this alternative makes it worse because the data is now duplicated in the not-overwritten and overwritten L1 layer files.
+We'd need to write a deduplication pass that checks if perfectly overlapping layers have identical contents.
+
+However, this alternative is appealing because it systematically prevents overwrites at a lower level than this RFC.
+
+So, this alternative is sufficient for the needs of the split-brain safety RFC (immutable layer files locally and in S3).
+But it doesn't solve the problems with crash-during-compaction outlined earlier in this RFC, and in fact, makes it much more accute.
+The proposed design in this RFC addresses both.
+
+So, if this alternative sounds appealing, we should implement the proposal in this RFC first, then implement this alternative on top.
+That way, we avoid a phase where the crash-during-compaction problem is accute.
+
+## Related issues
+
+- https://github.com/neondatabase/neon/issues/4749
+- https://github.com/neondatabase/neon/issues/4418
+  - https://github.com/neondatabase/neon/pull/4422
+- https://github.com/neondatabase/neon/issues/5077
+- https://github.com/neondatabase/neon/issues/4088
+  - (re)resolutions:
+    - https://github.com/neondatabase/neon/pull/4696
+    - https://github.com/neondatabase/neon/pull/4094
+      - https://neondb.slack.com/archives/C033QLM5P7D/p1682519017949719
+
+Note that the test case introduced in https://github.com/neondatabase/neon/pull/4696/files#diff-13114949d1deb49ae394405d4c49558adad91150ba8a34004133653a8a5aeb76 will produce L1s with the same logical content, but, as outlined in the last paragraph of the _Problem Statement_ section above, we don't want to make that  assumption in order to fix the problem.
+
+
+## Implementation Plan
+
+1. Remove support for `remote_storage=None`, because we now rely on the existence of an index part.
+
+    - The nasty part here is to fix all the tests that fiddle with the local timeline directory.
+      Possibly they are just irrelevant with this change, but, each case will require inspection.
+
+2. Implement the design above.
+
+    - Initially, ship without the mitigations for restart and accept we will do some work twice.
+    - Measure the impact and implement one of the mitigations.
+
--- a/libs/compute_api/Cargo.toml
+++ b/libs/compute_api/Cargo.toml
@@ -10,6 +10,9 @@ chrono.workspace = true
 serde.workspace = true
 serde_with.workspace = true
 serde_json.workspace = true
+regex.workspace = true

 utils = { path = "../utils" }
+remote_storage = { version = "0.1", path = "../remote_storage/" }
+
 workspace_hack.workspace = true
--- a/libs/compute_api/src/responses.rs
+++ b/libs/compute_api/src/responses.rs
@@ -107,7 +107,6 @@ pub struct ComputeMetrics {
    pub num_ext_downloaded: u64,
    pub largest_ext_size: u64, // these are measured in bytes
    pub total_ext_download_size: u64,
-    pub prep_extensions_ms: u64,
 }

 /// Response of the `/computes/{compute_id}/spec` control-plane API.
--- a/libs/compute_api/src/spec.rs
+++ b/libs/compute_api/src/spec.rs
@@ -3,11 +3,16 @@
 //! The spec.json file is used to pass information to 'compute_ctl'. It contains
 //! all the information needed to start up the right version of PostgreSQL,
 //! and connect it to the storage nodes.
+use std::collections::HashMap;
+
 use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;

+use regex::Regex;
+use remote_storage::RemotePath;
+
 /// String type alias representing Postgres identifier and
 /// intended to be used for DB / role names.
 pub type PgIdent = String;
@@ -61,8 +66,78 @@ pub struct ComputeSpec {
    /// the pageserver and safekeepers.
    pub storage_auth_token: Option<String>,

-    // list of prefixes to search for custom extensions in remote extension storage
+    // information about available remote extensions
+    pub remote_extensions: Option<RemoteExtSpec>,
+}
+
+#[derive(Clone, Debug, Default, Deserialize, Serialize)]
+pub struct RemoteExtSpec {
+    pub public_extensions: Option<Vec<String>>,
    pub custom_extensions: Option<Vec<String>>,
+    pub library_index: HashMap<String, String>,
+    pub extension_data: HashMap<String, ExtensionData>,
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+pub struct ExtensionData {
+    pub control_data: HashMap<String, String>,
+    pub archive_path: String,
+}
+
+impl RemoteExtSpec {
+    pub fn get_ext(
+        &self,
+        ext_name: &str,
+        is_library: bool,
+        build_tag: &str,
+        pg_major_version: &str,
+    ) -> anyhow::Result<(String, RemotePath)> {
+        let mut real_ext_name = ext_name;
+        if is_library {
+            // sometimes library names might have a suffix like
+            // library.so or library.so.3. We strip this off
+            // because library_index is based on the name without the file extension
+            let strip_lib_suffix = Regex::new(r"\.so.*").unwrap();
+            let lib_raw_name = strip_lib_suffix.replace(real_ext_name, "").to_string();
+
+            real_ext_name = self
+                .library_index
+                .get(&lib_raw_name)
+                .ok_or(anyhow::anyhow!("library {} is not found", lib_raw_name))?;
+        }
+
+        // Check if extension is present in public or custom.
+        // If not, then it is not allowed to be used by this compute.
+        if let Some(public_extensions) = &self.public_extensions {
+            if !public_extensions.contains(&real_ext_name.to_string()) {
+                if let Some(custom_extensions) = &self.custom_extensions {
+                    if !custom_extensions.contains(&real_ext_name.to_string()) {
+                        return Err(anyhow::anyhow!("extension {} is not found", real_ext_name));
+                    }
+                }
+            }
+        }
+
+        match self.extension_data.get(real_ext_name) {
+            Some(_ext_data) => {
+                // Construct the path to the extension archive
+                // BUILD_TAG/PG_MAJOR_VERSION/extensions/EXTENSION_NAME.tar.zst
+                //
+                // Keep it in sync with path generation in
+                // https://github.com/neondatabase/build-custom-extensions/tree/main
+                let archive_path_str =
+                    format!("{build_tag}/{pg_major_version}/extensions/{real_ext_name}.tar.zst");
+                Ok((
+                    real_ext_name.to_string(),
+                    RemotePath::from_string(&archive_path_str)?,
+                ))
+            }
+            None => Err(anyhow::anyhow!(
+                "real_ext_name {} is not found",
+                real_ext_name
+            )),
+        }
+    }
 }

 #[serde_as]
--- a/libs/compute_api/tests/cluster_spec.json
+++ b/libs/compute_api/tests/cluster_spec.json
@@ -205,5 +205,43 @@
            "name": "zenith new",
            "new_name": "zenith \"new\""
        }
-    ]
+    ],
+    "remote_extensions": {
+        "library_index": {
+          "anon": "anon",
+          "postgis-3": "postgis",
+          "libpgrouting-3.4": "postgis",
+          "postgis_raster-3": "postgis",
+          "postgis_sfcgal-3": "postgis",
+          "postgis_topology-3": "postgis",
+          "address_standardizer-3": "postgis"
+        },
+        "extension_data": {
+          "anon": {
+            "archive_path": "5834329303/v15/extensions/anon.tar.zst",
+            "control_data": {
+              "anon.control": "# PostgreSQL Anonymizer (anon) extension\ncomment = ''Data anonymization tools''\ndefault_version = ''1.1.0''\ndirectory=''extension/anon''\nrelocatable = false\nrequires = ''pgcrypto''\nsuperuser = false\nmodule_pathname = ''$libdir/anon''\ntrusted = true\n"
+            }
+          },
+          "postgis": {
+            "archive_path": "5834329303/v15/extensions/postgis.tar.zst",
+            "control_data": {
+              "postgis.control": "# postgis extension\ncomment = ''PostGIS geometry and geography spatial types and functions''\ndefault_version = ''3.3.2''\nmodule_pathname = ''$libdir/postgis-3''\nrelocatable = false\ntrusted = true\n",
+              "pgrouting.control": "# pgRouting Extension\ncomment = ''pgRouting Extension''\ndefault_version = ''3.4.2''\nmodule_pathname = ''$libdir/libpgrouting-3.4''\nrelocatable = true\nrequires = ''plpgsql''\nrequires = ''postgis''\ntrusted = true\n",
+              "postgis_raster.control": "# postgis_raster extension\ncomment = ''PostGIS raster types and functions''\ndefault_version = ''3.3.2''\nmodule_pathname = ''$libdir/postgis_raster-3''\nrelocatable = false\nrequires = postgis\ntrusted = true\n",
+              "postgis_sfcgal.control": "# postgis topology extension\ncomment = ''PostGIS SFCGAL functions''\ndefault_version = ''3.3.2''\nrelocatable = true\nrequires = postgis\ntrusted = true\n",
+              "postgis_topology.control": "# postgis topology extension\ncomment = ''PostGIS topology spatial types and functions''\ndefault_version = ''3.3.2''\nrelocatable = false\nschema = topology\nrequires = postgis\ntrusted = true\n",
+              "address_standardizer.control": "# address_standardizer extension\ncomment = ''Used to parse an address into constituent elements. Generally used to support geocoding address normalization step.''\ndefault_version = ''3.3.2''\nrelocatable = true\ntrusted = true\n",
+              "postgis_tiger_geocoder.control": "# postgis tiger geocoder extension\ncomment = ''PostGIS tiger geocoder and reverse geocoder''\ndefault_version = ''3.3.2''\nrelocatable = false\nschema = tiger\nrequires = ''postgis,fuzzystrmatch''\nsuperuser= false\ntrusted = true\n",
+              "address_standardizer_data_us.control": "# address standardizer us dataset\ncomment = ''Address Standardizer US dataset example''\ndefault_version = ''3.3.2''\nrelocatable = true\ntrusted = true\n"
+            }
+          }
+        },
+        "custom_extensions": [
+          "anon"
+        ],
+        "public_extensions": [
+          "postgis"
+        ]
+      }
 }
--- a/libs/pageserver_api/src/control_api.rs
+++ b/libs/pageserver_api/src/control_api.rs
@@ -0,0 +1,52 @@
+//! Types in this file are for pageserver's upward-facing API calls to the control plane,
+//! required for acquiring and validating tenant generation numbers.
+//!
+//! See docs/rfcs/025-generation-numbers.md
+
+use serde::{Deserialize, Serialize};
+use serde_with::{serde_as, DisplayFromStr};
+use utils::id::{NodeId, TenantId};
+
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachRequest {
+    pub node_id: NodeId,
+}
+
+#[serde_as]
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachResponseTenant {
+    #[serde_as(as = "DisplayFromStr")]
+    pub id: TenantId,
+    pub generation: u32,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachResponse {
+    pub tenants: Vec<ReAttachResponseTenant>,
+}
+
+#[serde_as]
+#[derive(Serialize, Deserialize)]
+pub struct ValidateRequestTenant {
+    #[serde_as(as = "DisplayFromStr")]
+    pub id: TenantId,
+    pub gen: u32,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateRequest {
+    pub tenants: Vec<ValidateRequestTenant>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateResponse {
+    pub tenants: Vec<ValidateResponseTenant>,
+}
+
+#[serde_as]
+#[derive(Serialize, Deserialize)]
+pub struct ValidateResponseTenant {
+    #[serde_as(as = "DisplayFromStr")]
+    pub id: TenantId,
+    pub valid: bool,
+}
--- a/libs/pageserver_api/src/lib.rs
+++ b/libs/pageserver_api/src/lib.rs
@@ -1,6 +1,7 @@
 use const_format::formatcp;

 /// Public API types
+pub mod control_api;
 pub mod models;
 pub mod reltag;

--- a/libs/pageserver_api/src/models.rs
+++ b/libs/pageserver_api/src/models.rs
@@ -194,10 +194,22 @@ pub struct TimelineCreateRequest {
 pub struct TenantCreateRequest {
    #[serde_as(as = "DisplayFromStr")]
    pub new_tenant_id: TenantId,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub generation: Option<u32>,
    #[serde(flatten)]
    pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
 }

+#[serde_as]
+#[derive(Deserialize, Debug)]
+#[serde(deny_unknown_fields)]
+pub struct TenantLoadRequest {
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub generation: Option<u32>,
+}
+
 impl std::ops::Deref for TenantCreateRequest {
    type Target = TenantConfig;

@@ -241,15 +253,6 @@ pub struct StatusResponse {
    pub id: NodeId,
 }

-impl TenantCreateRequest {
-    pub fn new(new_tenant_id: TenantId) -> TenantCreateRequest {
-        TenantCreateRequest {
-            new_tenant_id,
-            config: TenantConfig::default(),
-        }
-    }
-}
-
 #[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
@@ -293,9 +296,11 @@ impl TenantConfigRequest {
    }
 }

-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Deserialize)]
 pub struct TenantAttachRequest {
    pub config: TenantAttachConfig,
+    #[serde(default)]
+    pub generation: Option<u32>,
 }

 /// Newtype to enforce deny_unknown_fields on TenantConfig for
@@ -376,6 +381,8 @@ pub struct TimelineInfo {
    pub pg_version: u32,

    pub state: TimelineState,
+
+    pub walreceiver_status: String,
 }

 #[derive(Debug, Clone, Serialize)]
--- a/libs/postgres_ffi/README.md
+++ b/libs/postgres_ffi/README.md
@@ -10,9 +10,11 @@ should be auto-generated too, but that's a TODO.
 The PostgreSQL on-disk file format is not portable across different
 CPU architectures and operating systems. It is also subject to change
 in each major PostgreSQL version. Currently, this module supports
-PostgreSQL v14 and v15: bindings and code that depends on them are version-specific.
-This code is organized in modules: `postgres_ffi::v14` and `postgres_ffi::v15`
-Version independend code is explicitly exported into shared `postgres_ffi`.
+PostgreSQL v14, v15 and v16: bindings and code that depends on them are
+version-specific.
+This code is organized in modules `postgres_ffi::v14`, `postgres_ffi::v15` and
+`postgres_ffi::v16`. Version independent code is explicitly exported into
+shared `postgres_ffi`.


 TODO: Currently, there is also some code that deals with WAL records
--- a/libs/postgres_ffi/build.rs
+++ b/libs/postgres_ffi/build.rs
@@ -56,7 +56,7 @@ fn main() -> anyhow::Result<()> {
        PathBuf::from("pg_install")
    };

-    for pg_version in &["v14", "v15"] {
+    for pg_version in &["v14", "v15", "v16"] {
        let mut pg_install_dir_versioned = pg_install_dir.join(pg_version);
        if pg_install_dir_versioned.is_relative() {
            let cwd = env::current_dir().context("Failed to get current_dir")?;
@@ -125,6 +125,7 @@ fn main() -> anyhow::Result<()> {
            .allowlist_var("PG_CONTROLFILEDATA_OFFSETOF_CRC")
            .allowlist_type("PageHeaderData")
            .allowlist_type("DBState")
+            .allowlist_type("RelMapFile")
            // Because structs are used for serialization, tell bindgen to emit
            // explicit padding fields.
            .explicit_padding(true)
--- a/libs/postgres_ffi/src/lib.rs
+++ b/libs/postgres_ffi/src/lib.rs
@@ -51,11 +51,59 @@ macro_rules! for_all_postgres_versions {
    ($macro:tt) => {
        $macro!(v14);
        $macro!(v15);
+        $macro!(v16);
    };
 }

 for_all_postgres_versions! { postgres_ffi }

+/// dispatch_pgversion
+///
+/// Run a code block in a context where the postgres_ffi bindings for a
+/// specific (supported) PostgreSQL version are `use`-ed in scope under the pgv
+/// identifier.
+/// If the provided pg_version is not supported, we panic!(), unless the
+/// optional third argument was provided (in which case that code will provide
+/// the default handling instead).
+///
+/// Use like
+///
+/// dispatch_pgversion!(my_pgversion, { pgv::constants::XLOG_DBASE_CREATE })
+/// dispatch_pgversion!(my_pgversion, pgv::constants::XLOG_DBASE_CREATE)
+///
+/// Other uses are for macro-internal purposes only and strictly unsupported.
+///
+#[macro_export]
+macro_rules! dispatch_pgversion {
+    ($version:expr, $code:expr) => {
+        dispatch_pgversion!($version, $code, panic!("Unknown PostgreSQL version {}", $version))
+    };
+    ($version:expr, $code:expr, $invalid_pgver_handling:expr) => {
+        dispatch_pgversion!(
+            $version => $code,
+            default = $invalid_pgver_handling,
+            pgversions = [
+                14 : v14,
+                15 : v15,
+                16 : v16,
+            ]
+        )
+    };
+    ($pgversion:expr => $code:expr,
+     default = $default:expr,
+     pgversions = [$($sv:literal : $vsv:ident),+ $(,)?]) => {
+        match ($pgversion) {
+            $($sv => {
+                use $crate::$vsv as pgv;
+                $code
+            },)+
+            _ => {
+                $default
+            }
+        }
+    };
+}
+
 pub mod pg_constants;
 pub mod relfile_utils;

@@ -90,13 +138,7 @@ pub use v14::xlog_utils::XLogFileName;
 pub use v14::bindings::DBState_DB_SHUTDOWNED;

 pub fn bkpimage_is_compressed(bimg_info: u8, version: u32) -> anyhow::Result<bool> {
-    match version {
-        14 => Ok(bimg_info & v14::bindings::BKPIMAGE_IS_COMPRESSED != 0),
-        15 => Ok(bimg_info & v15::bindings::BKPIMAGE_COMPRESS_PGLZ != 0
-            || bimg_info & v15::bindings::BKPIMAGE_COMPRESS_LZ4 != 0
-            || bimg_info & v15::bindings::BKPIMAGE_COMPRESS_ZSTD != 0),
-        _ => anyhow::bail!("Unknown version {}", version),
-    }
+    dispatch_pgversion!(version, Ok(pgv::bindings::bkpimg_is_compressed(bimg_info)))
 }

 pub fn generate_wal_segment(
@@ -107,11 +149,11 @@ pub fn generate_wal_segment(
 ) -> Result<Bytes, SerializeError> {
    assert_eq!(segno, lsn.segment_number(WAL_SEGMENT_SIZE));

-    match pg_version {
-        14 => v14::xlog_utils::generate_wal_segment(segno, system_id, lsn),
-        15 => v15::xlog_utils::generate_wal_segment(segno, system_id, lsn),
-        _ => Err(SerializeError::BadInput),
-    }
+    dispatch_pgversion!(
+        pg_version,
+        pgv::xlog_utils::generate_wal_segment(segno, system_id, lsn),
+        Err(SerializeError::BadInput)
+    )
 }

 pub fn generate_pg_control(
@@ -120,11 +162,11 @@ pub fn generate_pg_control(
    lsn: Lsn,
    pg_version: u32,
 ) -> anyhow::Result<(Bytes, u64)> {
-    match pg_version {
-        14 => v14::xlog_utils::generate_pg_control(pg_control_bytes, checkpoint_bytes, lsn),
-        15 => v15::xlog_utils::generate_pg_control(pg_control_bytes, checkpoint_bytes, lsn),
-        _ => anyhow::bail!("Unknown version {}", pg_version),
-    }
+    dispatch_pgversion!(
+        pg_version,
+        pgv::xlog_utils::generate_pg_control(pg_control_bytes, checkpoint_bytes, lsn),
+        anyhow::bail!("Unknown version {}", pg_version)
+    )
 }

 // PG timeline is always 1, changing it doesn't have any useful meaning in Neon.
@@ -196,8 +238,6 @@ pub fn fsm_logical_to_physical(addr: BlockNumber) -> BlockNumber {
 }

 pub mod waldecoder {
-
-    use crate::{v14, v15};
    use bytes::{Buf, Bytes, BytesMut};
    use std::num::NonZeroU32;
    use thiserror::Error;
@@ -248,22 +288,17 @@ pub mod waldecoder {
        }

        pub fn poll_decode(&mut self) -> Result<Option<(Lsn, Bytes)>, WalDecodeError> {
-            match self.pg_version {
-                // This is a trick to support both versions simultaneously.
-                // See WalStreamDecoderHandler comments.
-                14 => {
-                    use self::v14::waldecoder_handler::WalStreamDecoderHandler;
+            dispatch_pgversion!(
+                self.pg_version,
+                {
+                    use pgv::waldecoder_handler::WalStreamDecoderHandler;
                    self.poll_decode_internal()
-                }
-                15 => {
-                    use self::v15::waldecoder_handler::WalStreamDecoderHandler;
-                    self.poll_decode_internal()
-                }
-                _ => Err(WalDecodeError {
+                },
+                Err(WalDecodeError {
                    msg: format!("Unknown version {}", self.pg_version),
                    lsn: self.lsn,
-                }),
-            }
+                })
+            )
        }
    }
 }
--- a/libs/postgres_ffi/src/pg_constants.rs
+++ b/libs/postgres_ffi/src/pg_constants.rs
@@ -163,6 +163,20 @@ pub const RM_HEAP2_ID: u8 = 9;
 pub const RM_HEAP_ID: u8 = 10;
 pub const RM_LOGICALMSG_ID: u8 = 21;

+// from neon_rmgr.h
+pub const RM_NEON_ID: u8 = 134;
+
+pub const XLOG_NEON_HEAP_INIT_PAGE: u8 = 0x80;
+
+pub const XLOG_NEON_HEAP_INSERT: u8 = 0x00;
+pub const XLOG_NEON_HEAP_DELETE: u8 = 0x10;
+pub const XLOG_NEON_HEAP_UPDATE: u8 = 0x20;
+pub const XLOG_NEON_HEAP_HOT_UPDATE: u8 = 0x30;
+pub const XLOG_NEON_HEAP_LOCK: u8 = 0x40;
+pub const XLOG_NEON_HEAP_MULTI_INSERT: u8 = 0x50;
+
+pub const XLOG_NEON_HEAP_VISIBLE: u8 = 0x40;
+
 // from xlogreader.h
 pub const XLR_INFO_MASK: u8 = 0x0F;
 pub const XLR_RMGR_INFO_MASK: u8 = 0xF0;
--- a/libs/postgres_ffi/src/pg_constants_v14.rs
+++ b/libs/postgres_ffi/src/pg_constants_v14.rs
@@ -3,3 +3,8 @@ pub const XLOG_DBASE_DROP: u8 = 0x10;

 pub const BKPIMAGE_IS_COMPRESSED: u8 = 0x02; /* page image is compressed */
 pub const BKPIMAGE_APPLY: u8 = 0x04; /* page image should be restored during replay */
+pub const SIZEOF_RELMAPFILE: usize = 512; /* sizeof(RelMapFile) in relmapper.c */
+
+pub fn bkpimg_is_compressed(bimg_info: u8) -> bool {
+    (bimg_info & BKPIMAGE_IS_COMPRESSED) != 0
+}
--- a/libs/postgres_ffi/src/pg_constants_v15.rs
+++ b/libs/postgres_ffi/src/pg_constants_v15.rs
@@ -1,10 +1,18 @@
 pub const XACT_XINFO_HAS_DROPPED_STATS: u32 = 1u32 << 8;

 pub const XLOG_DBASE_CREATE_FILE_COPY: u8 = 0x00;
-pub const XLOG_DBASE_CREATE_WAL_LOG: u8 = 0x00;
+pub const XLOG_DBASE_CREATE_WAL_LOG: u8 = 0x10;
 pub const XLOG_DBASE_DROP: u8 = 0x20;

 pub const BKPIMAGE_APPLY: u8 = 0x02; /* page image should be restored during replay */
 pub const BKPIMAGE_COMPRESS_PGLZ: u8 = 0x04; /* page image is compressed */
 pub const BKPIMAGE_COMPRESS_LZ4: u8 = 0x08; /* page image is compressed */
 pub const BKPIMAGE_COMPRESS_ZSTD: u8 = 0x10; /* page image is compressed */
+
+pub const SIZEOF_RELMAPFILE: usize = 512; /* sizeof(RelMapFile) in relmapper.c */
+
+pub fn bkpimg_is_compressed(bimg_info: u8) -> bool {
+    const ANY_COMPRESS_FLAG: u8 = BKPIMAGE_COMPRESS_PGLZ | BKPIMAGE_COMPRESS_LZ4 | BKPIMAGE_COMPRESS_ZSTD;
+
+    (bimg_info & ANY_COMPRESS_FLAG) != 0
+}
--- a/libs/postgres_ffi/src/pg_constants_v16.rs
+++ b/libs/postgres_ffi/src/pg_constants_v16.rs
@@ -0,0 +1,18 @@
+pub const XACT_XINFO_HAS_DROPPED_STATS: u32 = 1u32 << 8;
+
+pub const XLOG_DBASE_CREATE_FILE_COPY: u8 = 0x00;
+pub const XLOG_DBASE_CREATE_WAL_LOG: u8 = 0x10;
+pub const XLOG_DBASE_DROP: u8 = 0x20;
+
+pub const BKPIMAGE_APPLY: u8 = 0x02; /* page image should be restored during replay */
+pub const BKPIMAGE_COMPRESS_PGLZ: u8 = 0x04; /* page image is compressed */
+pub const BKPIMAGE_COMPRESS_LZ4: u8 = 0x08; /* page image is compressed */
+pub const BKPIMAGE_COMPRESS_ZSTD: u8 = 0x10; /* page image is compressed */
+
+pub const SIZEOF_RELMAPFILE: usize = 524; /* sizeof(RelMapFile) in relmapper.c */
+
+pub fn bkpimg_is_compressed(bimg_info: u8) -> bool {
+    const ANY_COMPRESS_FLAG: u8 = BKPIMAGE_COMPRESS_PGLZ | BKPIMAGE_COMPRESS_LZ4 | BKPIMAGE_COMPRESS_ZSTD;
+
+    (bimg_info & ANY_COMPRESS_FLAG) != 0
+}
--- a/libs/postgres_ffi/wal_craft/src/lib.rs
+++ b/libs/postgres_ffi/wal_craft/src/lib.rs
@@ -49,9 +49,9 @@ impl Conf {
    pub fn pg_distrib_dir(&self) -> anyhow::Result<PathBuf> {
        let path = self.pg_distrib_dir.clone();

+        #[allow(clippy::manual_range_patterns)]
        match self.pg_version {
-            14 => Ok(path.join(format!("v{}", self.pg_version))),
-            15 => Ok(path.join(format!("v{}", self.pg_version))),
+            14 | 15 | 16 => Ok(path.join(format!("v{}", self.pg_version))),
            _ => bail!("Unsupported postgres version: {}", self.pg_version),
        }
    }
@@ -250,11 +250,18 @@ fn craft_internal<C: postgres::GenericClient>(
    let (mut intermediate_lsns, last_lsn) = f(client, initial_lsn)?;
    let last_lsn = match last_lsn {
        None => client.pg_current_wal_insert_lsn()?,
-        Some(last_lsn) => match last_lsn.cmp(&client.pg_current_wal_insert_lsn()?) {
-            Ordering::Less => bail!("Some records were inserted after the crafted WAL"),
-            Ordering::Equal => last_lsn,
-            Ordering::Greater => bail!("Reported LSN is greater than insert_lsn"),
-        },
+        Some(last_lsn) => {
+            let insert_lsn = client.pg_current_wal_insert_lsn()?;
+            match last_lsn.cmp(&insert_lsn) {
+                Ordering::Less => bail!(
+                    "Some records were inserted after the crafted WAL: {} vs {}",
+                    last_lsn,
+                    insert_lsn
+                ),
+                Ordering::Equal => last_lsn,
+                Ordering::Greater => bail!("Reported LSN is greater than insert_lsn"),
+            }
+        }
    };
    if !intermediate_lsns.starts_with(&[initial_lsn]) {
        intermediate_lsns.insert(0, initial_lsn);
@@ -363,8 +370,9 @@ impl Crafter for LastWalRecordXlogSwitchEndsOnPageBoundary {
        );
        ensure!(
            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ == XLOG_SIZE_OF_XLOG_SHORT_PHD,
-            "XLOG_SWITCH message ended not on page boundary: {}",
-            after_xlog_switch
+            "XLOG_SWITCH message ended not on page boundary: {}, offset = {}",
+            after_xlog_switch,
+            u64::from(after_xlog_switch) as usize % XLOG_BLCKSZ
        );
        Ok((vec![before_xlog_switch, after_xlog_switch], next_segment))
    }
--- a/libs/pq_proto/src/lib.rs
+++ b/libs/pq_proto/src/lib.rs
@@ -959,7 +959,7 @@ mod tests {
        let make_params = |options| StartupMessageParams::new([("options", options)]);

        let params = StartupMessageParams::new([]);
-        assert!(matches!(params.options_escaped(), None));
+        assert!(params.options_escaped().is_none());

        let params = make_params("");
        assert!(split_options(&params).is_empty());
--- a/libs/remote_storage/src/local_fs.rs
+++ b/libs/remote_storage/src/local_fs.rs
@@ -148,21 +148,55 @@ impl RemoteStorage for LocalFs {
            Some(folder) => folder.with_base(&self.storage_root),
            None => self.storage_root.clone(),
        };
-        let mut files = vec![];
-        let mut directory_queue = vec![full_path.clone()];

+        // If we were given a directory, we may use it as our starting point.
+        // Otherwise, we must go up to the parent directory.  This is because
+        // S3 object list prefixes can be arbitrary strings, but when reading
+        // the local filesystem we need a directory to start calling read_dir on.
+        let mut initial_dir = full_path.clone();
+        match fs::metadata(full_path.clone()).await {
+            Ok(meta) => {
+                if !meta.is_dir() {
+                    // It's not a directory: strip back to the parent
+                    initial_dir.pop();
+                }
+            }
+            Err(e) if e.kind() == ErrorKind::NotFound => {
+                // It's not a file that exists: strip the prefix back to the parent directory
+                initial_dir.pop();
+            }
+            Err(e) => {
+                // Unexpected I/O error
+                anyhow::bail!(e)
+            }
+        }
+
+        // Note that PathBuf starts_with only considers full path segments, but
+        // object prefixes are arbitrary strings, so we need the strings for doing
+        // starts_with later.
+        let prefix = full_path.to_string_lossy();
+
+        let mut files = vec![];
+        let mut directory_queue = vec![initial_dir.clone()];
        while let Some(cur_folder) = directory_queue.pop() {
            let mut entries = fs::read_dir(cur_folder.clone()).await?;
            while let Some(entry) = entries.next_entry().await? {
                let file_name: PathBuf = entry.file_name().into();
                let full_file_name = cur_folder.clone().join(&file_name);
-                let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
-                files.push(file_remote_path.clone());
-                if full_file_name.is_dir() {
-                    directory_queue.push(full_file_name);
+                if full_file_name
+                    .to_str()
+                    .map(|s| s.starts_with(prefix.as_ref()))
+                    .unwrap_or(false)
+                {
+                    let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
+                    files.push(file_remote_path.clone());
+                    if full_file_name.is_dir() {
+                        directory_queue.push(full_file_name);
+                    }
                }
            }
        }
+
        Ok(files)
    }

--- a/libs/remote_storage/src/s3_bucket.rs
+++ b/libs/remote_storage/src/s3_bucket.rs
@@ -573,7 +573,7 @@ mod tests {

    #[test]
    fn relative_path() {
-        let all_paths = vec!["", "some/path", "some/path/"];
+        let all_paths = ["", "some/path", "some/path/"];
        let all_paths: Vec<RemotePath> = all_paths
            .iter()
            .map(|x| RemotePath::new(Path::new(x)).expect("bad path"))
--- a/libs/remote_storage/src/simulate_failures.rs
+++ b/libs/remote_storage/src/simulate_failures.rs
@@ -71,6 +71,13 @@ impl UnreliableWrapper {
            }
        }
    }
+
+    async fn delete_inner(&self, path: &RemotePath, attempt: bool) -> anyhow::Result<()> {
+        if attempt {
+            self.attempt(RemoteOp::Delete(path.clone()))?;
+        }
+        self.inner.delete(path).await
+    }
 }

 #[async_trait::async_trait]
@@ -122,15 +129,15 @@ impl RemoteStorage for UnreliableWrapper {
    }

    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
-        self.attempt(RemoteOp::Delete(path.clone()))?;
-        self.inner.delete(path).await
+        self.delete_inner(path, true).await
    }

    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
        self.attempt(RemoteOp::DeleteObjects(paths.to_vec()))?;
        let mut error_counter = 0;
        for path in paths {
-            if (self.delete(path).await).is_err() {
+            // Dont record attempt because it was already recorded above
+            if (self.delete_inner(path, false).await).is_err() {
                error_counter += 1;
            }
        }
--- a/libs/safekeeper_api/src/models.rs
+++ b/libs/safekeeper_api/src/models.rs
@@ -31,6 +31,8 @@ fn lsn_invalid() -> Lsn {
 #[serde_as]
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct SkTimelineInfo {
+    /// Term.
+    pub term: Option<u64>,
    /// Term of the last entry.
    pub last_log_term: Option<u64>,
    /// LSN of the last record.
@@ -58,4 +60,6 @@ pub struct SkTimelineInfo {
    /// A connection string to use for WAL receiving.
    #[serde(default)]
    pub safekeeper_connstr: Option<String>,
+    #[serde(default)]
+    pub http_connstr: Option<String>,
 }
--- a/libs/utils/Cargo.toml
+++ b/libs/utils/Cargo.toml
@@ -26,6 +26,7 @@ serde_json.workspace = true
 signal-hook.workspace = true
 thiserror.workspace = true
 tokio.workspace = true
+tokio-util.workspace = true
 tracing.workspace = true
 tracing-error.workspace = true
 tracing-subscriber = { workspace = true, features = ["json", "registry"] }
@@ -37,6 +38,7 @@ url.workspace = true
 uuid.workspace = true

 pq_proto.workspace = true
+postgres_connection.workspace = true
 metrics.workspace = true
 workspace_hack.workspace = true

--- a/libs/utils/scripts/restore_from_wal.sh
+++ b/libs/utils/scripts/restore_from_wal.sh
@@ -9,11 +9,12 @@ PORT=$4
 SYSID=$(od -A n -j 24 -N 8 -t d8 "$WAL_PATH"/000000010000000000000002* | cut -c 3-)
 rm -fr "$DATA_DIR"
 env -i LD_LIBRARY_PATH="$PG_BIN"/../lib "$PG_BIN"/initdb -E utf8 -U cloud_admin -D "$DATA_DIR" --sysid="$SYSID"
-echo port="$PORT" >> "$DATA_DIR"/postgresql.conf
+echo "port=$PORT" >> "$DATA_DIR"/postgresql.conf
+echo "shared_preload_libraries='\$libdir/neon_rmgr.so'" >> "$DATA_DIR"/postgresql.conf
 REDO_POS=0x$("$PG_BIN"/pg_controldata -D "$DATA_DIR" | grep -F "REDO location"| cut -c 42-)
 declare -i WAL_SIZE=$REDO_POS+114
-"$PG_BIN"/pg_ctl -D "$DATA_DIR" -l logfile start
-"$PG_BIN"/pg_ctl -D "$DATA_DIR" -l logfile stop -m immediate
+"$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" start
+"$PG_BIN"/pg_ctl -D "$DATA_DIR" -l "$DATA_DIR/logfile.log" stop -m immediate
 cp "$DATA_DIR"/pg_wal/000000010000000000000001 .
 cp "$WAL_PATH"/* "$DATA_DIR"/pg_wal/
 for partial in "$DATA_DIR"/pg_wal/*.partial ; do mv "$partial" "${partial%.partial}" ; done
--- a/libs/utils/src/backoff.rs
+++ b/libs/utils/src/backoff.rs
@@ -1,18 +1,31 @@
 use std::fmt::{Debug, Display};

 use futures::Future;
+use tokio_util::sync::CancellationToken;

 pub const DEFAULT_BASE_BACKOFF_SECONDS: f64 = 0.1;
 pub const DEFAULT_MAX_BACKOFF_SECONDS: f64 = 3.0;

-pub async fn exponential_backoff(n: u32, base_increment: f64, max_seconds: f64) {
+pub async fn exponential_backoff(
+    n: u32,
+    base_increment: f64,
+    max_seconds: f64,
+    cancel: &CancellationToken,
+) {
    let backoff_duration_seconds =
        exponential_backoff_duration_seconds(n, base_increment, max_seconds);
    if backoff_duration_seconds > 0.0 {
        tracing::info!(
            "Backoff: waiting {backoff_duration_seconds} seconds before processing with the task",
        );
-        tokio::time::sleep(std::time::Duration::from_secs_f64(backoff_duration_seconds)).await;
+
+        drop(
+            tokio::time::timeout(
+                std::time::Duration::from_secs_f64(backoff_duration_seconds),
+                cancel.cancelled(),
+            )
+            .await,
+        )
    }
 }

@@ -24,28 +37,57 @@ pub fn exponential_backoff_duration_seconds(n: u32, base_increment: f64, max_sec
    }
 }

+/// Configure cancellation for a retried operation: when to cancel (the token), and
+/// what kind of error to return on cancellation
+pub struct Cancel<E, CF>
+where
+    E: Display + Debug + 'static,
+    CF: Fn() -> E,
+{
+    token: CancellationToken,
+    on_cancel: CF,
+}
+
+impl<E, CF> Cancel<E, CF>
+where
+    E: Display + Debug + 'static,
+    CF: Fn() -> E,
+{
+    pub fn new(token: CancellationToken, on_cancel: CF) -> Self {
+        Self { token, on_cancel }
+    }
+}
+
 /// retries passed operation until one of the following conditions are met:
 /// Encountered error is considered as permanent (non-retryable)
 /// Retries have been exhausted.
 /// `is_permanent` closure should be used to provide distinction between permanent/non-permanent errors
 /// When attempts cross `warn_threshold` function starts to emit log warnings.
 /// `description` argument is added to log messages. Its value should identify the `op` is doing
-pub async fn retry<T, O, F, E>(
+/// `cancel` argument is required: any time we are looping on retry, we should be using a CancellationToken
+/// to drop out promptly on shutdown.
+pub async fn retry<T, O, F, E, CF>(
    mut op: O,
    is_permanent: impl Fn(&E) -> bool,
    warn_threshold: u32,
    max_retries: u32,
    description: &str,
+    cancel: Cancel<E, CF>,
 ) -> Result<T, E>
 where
    // Not std::error::Error because anyhow::Error doesnt implement it.
    // For context see https://github.com/dtolnay/anyhow/issues/63
-    E: Display + Debug,
+    E: Display + Debug + 'static,
    O: FnMut() -> F,
    F: Future<Output = Result<T, E>>,
+    CF: Fn() -> E,
 {
    let mut attempts = 0;
    loop {
+        if cancel.token.is_cancelled() {
+            return Err((cancel.on_cancel)());
+        }
+
        let result = op().await;
        match result {
            Ok(_) => {
@@ -80,6 +122,7 @@ where
            attempts,
            DEFAULT_BASE_BACKOFF_SECONDS,
            DEFAULT_MAX_BACKOFF_SECONDS,
+            &cancel.token,
        )
        .await;
        attempts += 1;
@@ -132,6 +175,7 @@ mod tests {
            1,
            1,
            "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
        )
        .await;

@@ -157,6 +201,7 @@ mod tests {
            2,
            2,
            "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
        )
        .await
        .unwrap();
@@ -179,6 +224,7 @@ mod tests {
            2,
            2,
            "work",
+            Cancel::new(CancellationToken::new(), || -> io::Error { unreachable!() }),
        )
        .await
        .unwrap_err();
--- a/libs/utils/src/crashsafe.rs
+++ b/libs/utils/src/crashsafe.rs
@@ -111,6 +111,10 @@ pub fn fsync(path: &Path) -> io::Result<()> {
        .map_err(|e| io::Error::new(e.kind(), format!("Failed to fsync file {path:?}: {e}")))
 }

+pub async fn fsync_async(path: impl AsRef<std::path::Path>) -> Result<(), std::io::Error> {
+    tokio::fs::File::open(path).await?.sync_all().await
+}
+
 #[cfg(test)]
 mod tests {
    use tempfile::tempdir;
--- a/libs/utils/src/generation.rs
+++ b/libs/utils/src/generation.rs
@@ -0,0 +1,138 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+/// Tenant generations are used to provide split-brain safety and allow
+/// multiple pageservers to attach the same tenant concurrently.
+///
+/// See docs/rfcs/025-generation-numbers.md for detail on how generation
+/// numbers are used.
+#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord)]
+pub enum Generation {
+    // Generations with this magic value will not add a suffix to S3 keys, and will not
+    // be included in persisted index_part.json.  This value is only to be used
+    // during migration from pre-generation metadata to generation-aware metadata,
+    // and should eventually go away.
+    //
+    // A special Generation is used rather than always wrapping Generation in an Option,
+    // so that code handling generations doesn't have to be aware of the legacy
+    // case everywhere it touches a generation.
+    None,
+    // Generations with this magic value may never be used to construct S3 keys:
+    // we will panic if someone tries to.  This is for Tenants in the "Broken" state,
+    // so that we can satisfy their constructor with a Generation without risking
+    // a code bug using it in an S3 write (broken tenants should never write)
+    Broken,
+    Valid(u32),
+}
+
+/// The Generation type represents a number associated with a Tenant, which
+/// increments every time the tenant is attached to a new pageserver, or
+/// an attached pageserver restarts.
+///
+/// It is included as a suffix in S3 keys, as a protection against split-brain
+/// scenarios where pageservers might otherwise issue conflicting writes to
+/// remote storage
+impl Generation {
+    /// Create a new Generation that represents a legacy key format with
+    /// no generation suffix
+    pub fn none() -> Self {
+        Self::None
+    }
+
+    // Create a new generation that will panic if you try to use get_suffix
+    pub fn broken() -> Self {
+        Self::Broken
+    }
+
+    pub fn new(v: u32) -> Self {
+        Self::Valid(v)
+    }
+
+    pub fn is_none(&self) -> bool {
+        matches!(self, Self::None)
+    }
+
+    #[track_caller]
+    pub fn get_suffix(&self) -> String {
+        match self {
+            Self::Valid(v) => {
+                format!("-{:08x}", v)
+            }
+            Self::None => "".into(),
+            Self::Broken => {
+                panic!("Tried to use a broken generation");
+            }
+        }
+    }
+
+    /// `suffix` is the part after "-" in a key
+    ///
+    /// Returns None if parsing was unsuccessful
+    pub fn parse_suffix(suffix: &str) -> Option<Generation> {
+        u32::from_str_radix(suffix, 16).map(Generation::new).ok()
+    }
+
+    #[track_caller]
+    pub fn previous(&self) -> Generation {
+        match self {
+            Self::Valid(n) => {
+                if *n == 0 {
+                    // Since a tenant may be upgraded from a pre-generations state, interpret the "previous" generation
+                    // to 0 as being "no generation".
+                    Self::None
+                } else {
+                    Self::Valid(n - 1)
+                }
+            }
+            Self::None => Self::None,
+            Self::Broken => panic!("Attempted to use a broken generation"),
+        }
+    }
+}
+
+impl Serialize for Generation {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        if let Self::Valid(v) = self {
+            v.serialize(serializer)
+        } else {
+            // We should never be asked to serialize a None or Broken.  Structures
+            // that include an optional generation should convert None to an
+            // Option<Generation>::None
+            Err(serde::ser::Error::custom(
+                "Tried to serialize invalid generation ({self})",
+            ))
+        }
+    }
+}
+
+impl<'de> Deserialize<'de> for Generation {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        Ok(Self::Valid(u32::deserialize(deserializer)?))
+    }
+}
+
+// We intentionally do not implement Display for Generation, to reduce the
+// risk of a bug where the generation is used in a format!() string directly
+// instead of using get_suffix().
+impl Debug for Generation {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Valid(v) => {
+                write!(f, "{:08x}", v)
+            }
+            Self::None => {
+                write!(f, "<none>")
+            }
+            Self::Broken => {
+                write!(f, "<broken>")
+            }
+        }
+    }
+}
--- a/libs/utils/src/lib.rs
+++ b/libs/utils/src/lib.rs
@@ -27,6 +27,9 @@ pub mod id;
 // http endpoint utils
 pub mod http;

+// definition of the Generation type for pageserver attachment APIs
+pub mod generation;
+
 // common log initialisation routine
 pub mod logging;

@@ -58,6 +61,8 @@ pub mod serde_regex;

 pub mod pageserver_feedback;

+pub mod postgres_client;
+
 pub mod tracing_span_assert;

 pub mod rate_limit;
@@ -68,44 +73,6 @@ pub mod completion;
 /// Reporting utilities
 pub mod error;

-mod failpoint_macro_helpers {
-
-    /// use with fail::cfg("$name", "return(2000)")
-    ///
-    /// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the
-    /// specified time (in milliseconds). The main difference is that we use async
-    /// tokio sleep function. Another difference is that we print lines to the log,
-    /// which can be useful in tests to check that the failpoint was hit.
-    #[macro_export]
-    macro_rules! failpoint_sleep_millis_async {
-        ($name:literal) => {{
-            // If the failpoint is used with a "return" action, set should_sleep to the
-            // returned value (as string). Otherwise it's set to None.
-            let should_sleep = (|| {
-                ::fail::fail_point!($name, |x| x);
-                ::std::option::Option::None
-            })();
-
-            // Sleep if the action was a returned value
-            if let ::std::option::Option::Some(duration_str) = should_sleep {
-                $crate::failpoint_sleep_helper($name, duration_str).await
-            }
-        }};
-    }
-
-    // Helper function used by the macro. (A function has nicer scoping so we
-    // don't need to decorate everything with "::")
-    pub async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
-        let millis = duration_str.parse::<u64>().unwrap();
-        let d = std::time::Duration::from_millis(millis);
-
-        tracing::info!("failpoint {:?}: sleeping for {:?}", name, d);
-        tokio::time::sleep(d).await;
-        tracing::info!("failpoint {:?}: sleep done", name);
-    }
-}
-pub use failpoint_macro_helpers::failpoint_sleep_helper;
-
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:
--- a/libs/utils/src/postgres_client.rs
+++ b/libs/utils/src/postgres_client.rs
@@ -0,0 +1,37 @@
+//! Postgres client connection code common to other crates (safekeeper and
+//! pageserver) which depends on tenant/timeline ids and thus not fitting into
+//! postgres_connection crate.
+
+use anyhow::Context;
+use postgres_connection::{parse_host_port, PgConnectionConfig};
+
+use crate::id::TenantTimelineId;
+
+/// Create client config for fetching WAL from safekeeper on particular timeline.
+/// listen_pg_addr_str is in form host:\[port\].
+pub fn wal_stream_connection_config(
+    TenantTimelineId {
+        tenant_id,
+        timeline_id,
+    }: TenantTimelineId,
+    listen_pg_addr_str: &str,
+    auth_token: Option<&str>,
+    availability_zone: Option<&str>,
+) -> anyhow::Result<PgConnectionConfig> {
+    let (host, port) =
+        parse_host_port(listen_pg_addr_str).context("Unable to parse listen_pg_addr_str")?;
+    let port = port.unwrap_or(5432);
+    let mut connstr = PgConnectionConfig::new_host_port(host, port)
+        .extend_options([
+            "-c".to_owned(),
+            format!("timeline_id={}", timeline_id),
+            format!("tenant_id={}", tenant_id),
+        ])
+        .set_password(auth_token.map(|s| s.to_owned()));
+
+    if let Some(availability_zone) = availability_zone {
+        connstr = connstr.extend_options([format!("availability_zone={}", availability_zone)]);
+    }
+
+    Ok(connstr)
+}
--- a/libs/vm_monitor/Cargo.toml
+++ b/libs/vm_monitor/Cargo.toml
@@ -0,0 +1,31 @@
+[package]
+name = "vm_monitor"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[[bin]]
+name = "vm-monitor"
+path = "./src/bin/monitor.rs"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow.workspace = true
+axum.workspace = true
+clap.workspace = true
+futures.workspace = true
+inotify.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+sysinfo.workspace = true
+tokio.workspace = true
+tokio-postgres.workspace = true
+tokio-stream.workspace = true
+tokio-util.workspace = true
+tracing.workspace = true
+tracing-subscriber.workspace = true
+workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+
+[target.'cfg(target_os = "linux")'.dependencies]
+cgroups-rs = "0.3.3"
--- a/libs/vm_monitor/README.md
+++ b/libs/vm_monitor/README.md
@@ -0,0 +1,34 @@
+# `vm-monitor`
+
+The `vm-monitor` (or just monitor) is a core component of the autoscaling system,
+along with the `autoscale-scheduler` and the `autoscaler-agent`s. The monitor has
+two primary roles: 1) notifying agents when immediate upscaling is necessary due
+to memory conditions and 2) managing Postgres' file cache and a cgroup to carry
+out upscaling and downscaling decisions.
+
+## More on scaling
+
+We scale CPU and memory using NeonVM, our in-house QEMU tool for use with Kubernetes.
+To control thresholds for receiving memory usage notifications, we start Postgres
+in the `neon-postgres` cgroup and set its `memory.{max,high}`.
+
+* See also: [`neondatabase/autoscaling`](https://github.com/neondatabase/autoscaling/)
+* See also: [`neondatabase/vm-monitor`](https://github.com/neondatabase/vm-monitor/),
+where initial development of the monitor happened. The repository is no longer
+maintained but the commit history may be useful for debugging.
+
+## Structure
+
+The `vm-monitor` is loosely comprised of a few systems. These are:
+* the server: this is just a simple `axum` server that accepts requests and
+upgrades them to websocket connections. The server only allows one connection at
+a time. This means that upon receiving a new connection, the server will terminate
+and old one if it exists.
+* the filecache: a struct that allows communication with the Postgres file cache.
+On startup, we connect to the filecache and hold on to the connection for the
+entire monitor lifetime.
+* the cgroup watcher: the `CgroupWatcher` manages the `neon-postgres` cgroup by
+listening for `memory.high` events and setting its `memory.{high,max}` values.
+* the runner: the runner marries the filecache and cgroup watcher together,
+communicating with the agent throught the `Dispatcher`, and then calling filecache
+and cgroup watcher functions as needed to upscale and downscale
--- a/libs/vm_monitor/src/bin/monitor.rs
+++ b/libs/vm_monitor/src/bin/monitor.rs
@@ -0,0 +1,33 @@
+// We expose a standalone binary _and_ start the monitor in `compute_ctl` so that
+// we can test the monitor as part of the entire autoscaling system in
+// neondatabase/autoscaling.
+//
+// The monitor was previously started by vm-builder, and for testing purposes,
+// we can mimic that setup with this binary.
+
+#[cfg(target_os = "linux")]
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    use clap::Parser;
+    use tokio_util::sync::CancellationToken;
+    use tracing_subscriber::EnvFilter;
+    use vm_monitor::Args;
+
+    let subscriber = tracing_subscriber::fmt::Subscriber::builder()
+        .json()
+        .with_file(true)
+        .with_line_number(true)
+        .with_span_list(true)
+        .with_env_filter(EnvFilter::from_default_env())
+        .finish();
+    tracing::subscriber::set_global_default(subscriber)?;
+
+    let args: &'static Args = Box::leak(Box::new(Args::parse()));
+    let token = CancellationToken::new();
+    vm_monitor::start(args, token).await
+}
+
+#[cfg(not(target_os = "linux"))]
+fn main() {
+    panic!("the monitor requires cgroups, which are only available on linux")
+}
--- a/libs/vm_monitor/src/cgroup.rs
+++ b/libs/vm_monitor/src/cgroup.rs
@@ -0,0 +1,693 @@
+use std::{
+    fmt::{Debug, Display},
+    fs,
+    pin::pin,
+    sync::atomic::{AtomicU64, Ordering},
+};
+
+use anyhow::{anyhow, bail, Context};
+use cgroups_rs::{
+    freezer::FreezerController,
+    hierarchies::{self, is_cgroup2_unified_mode, UNIFIED_MOUNTPOINT},
+    memory::MemController,
+    MaxValue,
+    Subsystem::{Freezer, Mem},
+};
+use inotify::{EventStream, Inotify, WatchMask};
+use tokio::sync::mpsc::{self, error::TryRecvError};
+use tokio::time::{Duration, Instant};
+use tokio_stream::{Stream, StreamExt};
+use tracing::{info, warn};
+
+use crate::protocol::Resources;
+use crate::MiB;
+
+/// Monotonically increasing counter of the number of memory.high events
+/// the cgroup has experienced.
+///
+/// We use this to determine if a modification to the `memory.events` file actually
+/// changed the `high` field. If not, we don't care about the change. When we
+/// read the file, we check the `high` field in the file against `MEMORY_EVENT_COUNT`
+/// to see if it changed since last time.
+pub static MEMORY_EVENT_COUNT: AtomicU64 = AtomicU64::new(0);
+
+/// Monotonically increasing counter that gives each cgroup event a unique id.
+///
+/// This allows us to answer questions like "did this upscale arrive before this
+/// memory.high?". This static is also used by the `Sequenced` type to "tag" values
+/// with a sequence number. As such, prefer to used the `Sequenced` type rather
+/// than this static directly.
+static EVENT_SEQUENCE_NUMBER: AtomicU64 = AtomicU64::new(0);
+
+/// A memory event type reported in memory.events.
+#[derive(Debug, Eq, PartialEq, Copy, Clone)]
+pub enum MemoryEvent {
+    Low,
+    High,
+    Max,
+    Oom,
+    OomKill,
+    OomGroupKill,
+}
+
+impl MemoryEvent {
+    fn as_str(&self) -> &str {
+        match self {
+            MemoryEvent::Low => "low",
+            MemoryEvent::High => "high",
+            MemoryEvent::Max => "max",
+            MemoryEvent::Oom => "oom",
+            MemoryEvent::OomKill => "oom_kill",
+            MemoryEvent::OomGroupKill => "oom_group_kill",
+        }
+    }
+}
+
+impl Display for MemoryEvent {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.as_str())
+    }
+}
+
+/// Configuration for a `CgroupWatcher`
+#[derive(Debug, Clone)]
+pub struct Config {
+    // The target difference between the total memory reserved for the cgroup
+    // and the value of the cgroup's memory.high.
+    //
+    // In other words, memory.high + oom_buffer_bytes will equal the total memory that the cgroup may
+    // use (equal to system memory, minus whatever's taken out for the file cache).
+    oom_buffer_bytes: u64,
+
+    // The amount of memory, in bytes, below a proposed new value for
+    // memory.high that the cgroup's memory usage must be for us to downscale
+    //
+    // In other words, we can downscale only when:
+    //
+    //   memory.current + memory_high_buffer_bytes < (proposed) memory.high
+    //
+    // TODO: there's some minor issues with this approach -- in particular, that we might have
+    // memory in use by the kernel's page cache that we're actually ok with getting rid of.
+    pub(crate) memory_high_buffer_bytes: u64,
+
+    // The maximum duration, in milliseconds, that we're allowed to pause
+    // the cgroup for while waiting for the autoscaler-agent to upscale us
+    max_upscale_wait: Duration,
+
+    // The required minimum time, in milliseconds, that we must wait before re-freezing
+    // the cgroup while waiting for the autoscaler-agent to upscale us.
+    do_not_freeze_more_often_than: Duration,
+
+    // The amount of memory, in bytes, that we should periodically increase memory.high
+    // by while waiting for the autoscaler-agent to upscale us.
+    //
+    // This exists to avoid the excessive throttling that happens when a cgroup is above its
+    // memory.high for too long. See more here:
+    // https://github.com/neondatabase/autoscaling/issues/44#issuecomment-1522487217
+    memory_high_increase_by_bytes: u64,
+
+    // The period, in milliseconds, at which we should repeatedly increase the value
+    // of the cgroup's memory.high while we're waiting on upscaling and memory.high
+    // is still being hit.
+    //
+    // Technically speaking, this actually serves as a rate limit to moderate responding to
+    // memory.high events, but these are roughly equivalent if the process is still allocating
+    // memory.
+    memory_high_increase_every: Duration,
+}
+
+impl Config {
+    /// Calculate the new value for the cgroups memory.high based on system memory
+    pub fn calculate_memory_high_value(&self, total_system_mem: u64) -> u64 {
+        total_system_mem.saturating_sub(self.oom_buffer_bytes)
+    }
+}
+
+impl Default for Config {
+    fn default() -> Self {
+        Self {
+            oom_buffer_bytes: 100 * MiB,
+            memory_high_buffer_bytes: 100 * MiB,
+            // while waiting for upscale, don't freeze for more than 20ms every 1s
+            max_upscale_wait: Duration::from_millis(20),
+            do_not_freeze_more_often_than: Duration::from_millis(1000),
+            // while waiting for upscale, increase memory.high by 10MiB every 25ms
+            memory_high_increase_by_bytes: 10 * MiB,
+            memory_high_increase_every: Duration::from_millis(25),
+        }
+    }
+}
+
+/// Used to represent data that is associated with a certain point in time, such
+/// as an upscale request or memory.high event.
+///
+/// Internally, creating a `Sequenced` uses a static atomic counter to obtain
+/// a unique sequence number. Sequence numbers are monotonically increasing,
+/// allowing us to answer questions like "did this upscale happen after this
+/// memory.high event?" by comparing the sequence numbers of the two events.
+#[derive(Debug, Clone)]
+pub struct Sequenced<T> {
+    seqnum: u64,
+    data: T,
+}
+
+impl<T> Sequenced<T> {
+    pub fn new(data: T) -> Self {
+        Self {
+            seqnum: EVENT_SEQUENCE_NUMBER.fetch_add(1, Ordering::AcqRel),
+            data,
+        }
+    }
+}
+
+/// Responds to `MonitorEvents` to manage the cgroup: preventing it from being
+/// OOM killed or throttling.
+///
+/// The `CgroupWatcher` primarily achieves this by reading from a stream of
+/// `MonitorEvent`s. See `main_signals_loop` for details on how to keep the
+/// cgroup happy.
+#[derive(Debug)]
+pub struct CgroupWatcher {
+    pub config: Config,
+
+    /// The sequence number of the last upscale.
+    ///
+    /// If we receive a memory.high event that has a _lower_ sequence number than
+    /// `last_upscale_seqnum`, then we know it occured before the upscale, and we
+    /// can safely ignore it.
+    ///
+    /// Note: Like the `events` field, this doesn't _need_ interior mutability but we
+    /// use it anyways so that methods take `&self`, not `&mut self`.
+    last_upscale_seqnum: AtomicU64,
+
+    /// A channel on which we send messages to request upscale from the dispatcher.
+    upscale_requester: mpsc::Sender<()>,
+
+    /// The actual cgroup we are watching and managing.
+    cgroup: cgroups_rs::Cgroup,
+}
+
+/// Read memory.events for the desired event type.
+///
+/// `path` specifies the path to the desired `memory.events` file.
+/// For more info, see the `memory.events` section of the [kernel docs]
+/// <https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>
+fn get_event_count(path: &str, event: MemoryEvent) -> anyhow::Result<u64> {
+    let contents = fs::read_to_string(path)
+        .with_context(|| format!("failed to read memory.events from {path}"))?;
+
+    // Then contents of the file look like:
+    // low 42
+    // high 101
+    // ...
+    contents
+        .lines()
+        .filter_map(|s| s.split_once(' '))
+        .find(|(e, _)| *e == event.as_str())
+        .ok_or_else(|| anyhow!("failed to find entry for memory.{event} events in {path}"))
+        .and_then(|(_, count)| {
+            count
+                .parse::<u64>()
+                .with_context(|| format!("failed to parse memory.{event} as u64"))
+        })
+}
+
+/// Create an event stream that produces events whenever the file at the provided
+/// path is modified.
+fn create_file_watcher(path: &str) -> anyhow::Result<EventStream<[u8; 1024]>> {
+    info!("creating file watcher for {path}");
+    let inotify = Inotify::init().context("failed to initialize file watcher")?;
+    inotify
+        .watches()
+        .add(path, WatchMask::MODIFY)
+        .with_context(|| format!("failed to start watching {path}"))?;
+    inotify
+        // The inotify docs use [0u8; 1024] so we'll just copy them. We only need
+        // to store one event at a time - if the event gets written over, that's
+        // ok. We still see that there is an event. For more information, see:
+        // https://man7.org/linux/man-pages/man7/inotify.7.html
+        .into_event_stream([0u8; 1024])
+        .context("failed to start inotify event stream")
+}
+
+impl CgroupWatcher {
+    /// Create a new `CgroupWatcher`.
+    #[tracing::instrument(skip_all, fields(%name))]
+    pub fn new(
+        name: String,
+        // A channel on which to send upscale requests
+        upscale_requester: mpsc::Sender<()>,
+    ) -> anyhow::Result<(Self, impl Stream<Item = Sequenced<u64>>)> {
+        // TODO: clarify exactly why we need v2
+        // Make sure cgroups v2 (aka unified) are supported
+        if !is_cgroup2_unified_mode() {
+            anyhow::bail!("cgroups v2 not supported");
+        }
+        let cgroup = cgroups_rs::Cgroup::load(hierarchies::auto(), &name);
+
+        // Start monitoring the cgroup for memory events. In general, for
+        // cgroups v2 (aka unified), metrics are reported in files like
+        // > `/sys/fs/cgroup/{name}/{metric}`
+        // We are looking for `memory.high` events, which are stored in the
+        // file `memory.events`. For more info, see the `memory.events` section
+        // of https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
+        let path = format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name);
+        let memory_events = create_file_watcher(&path)
+            .with_context(|| format!("failed to create event watcher for {path}"))?
+            // This would be nice with with .inspect_err followed by .ok
+            .filter_map(move |_| match get_event_count(&path, MemoryEvent::High) {
+                Ok(high) => Some(high),
+                Err(error) => {
+                    // TODO: Might want to just panic here
+                    warn!(?error, "failed to read high events count from {}", &path);
+                    None
+                }
+            })
+            // Only report the event if the memory.high count increased
+            .filter_map(|high| {
+                if MEMORY_EVENT_COUNT.fetch_max(high, Ordering::AcqRel) < high {
+                    Some(high)
+                } else {
+                    None
+                }
+            })
+            .map(Sequenced::new);
+
+        let initial_count = get_event_count(
+            &format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name),
+            MemoryEvent::High,
+        )?;
+
+        info!(initial_count, "initial memory.high event count");
+
+        // Hard update `MEMORY_EVENT_COUNT` since there could have been processes
+        // running in the cgroup before that caused it to be non-zero.
+        MEMORY_EVENT_COUNT.fetch_max(initial_count, Ordering::AcqRel);
+
+        Ok((
+            Self {
+                cgroup,
+                upscale_requester,
+                last_upscale_seqnum: AtomicU64::new(0),
+                config: Default::default(),
+            },
+            memory_events,
+        ))
+    }
+
+    /// The entrypoint for the `CgroupWatcher`.
+    #[tracing::instrument(skip_all)]
+    pub async fn watch<E>(
+        &self,
+        // These are ~dependency injected~ (fancy, I know) because this function
+        // should never return.
+        // -> therefore: when we tokio::spawn it, we don't await the JoinHandle.
+        // -> therefore: if we want to stick it in an Arc so many threads can access
+        //    it, methods can never take mutable access.
+        //     - note: we use the Arc strategy so that a) we can call this function
+        //             right here and b) the runner can call the set/get_memory methods
+        // -> since calling recv() on a tokio::sync::mpsc::Receiver takes &mut self,
+        //    we just pass them in here instead of holding them in fields, as that
+        //    would require this method to take &mut self.
+        mut upscales: mpsc::Receiver<Sequenced<Resources>>,
+        events: E,
+    ) -> anyhow::Result<()>
+    where
+        E: Stream<Item = Sequenced<u64>>,
+    {
+        // There are several actions might do when receiving a `memory.high`,
+        // such as freezing the cgroup, or increasing its `memory.high`. We don't
+        // want to do these things too often (because postgres needs to run, and
+        // we only have so much memory). These timers serve as rate limits for this.
+        let mut wait_to_freeze = pin!(tokio::time::sleep(Duration::ZERO));
+        let mut wait_to_increase_memory_high = pin!(tokio::time::sleep(Duration::ZERO));
+        let mut events = pin!(events);
+
+        // Are we waiting to be upscaled? Could be true if we request upscale due
+        // to a memory.high event and it does not arrive in time.
+        let mut waiting_on_upscale = false;
+
+        loop {
+            tokio::select! {
+                upscale = upscales.recv() => {
+                    let Sequenced { seqnum, data } = upscale
+                        .context("failed to listen on upscale notification channel")?;
+                    self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+                    info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
+                }
+                event = events.next() => {
+                    let Some(Sequenced { seqnum, .. }) = event else {
+                        bail!("failed to listen for memory.high events")
+                    };
+                    // The memory.high came before our last upscale, so we consider
+                    // it resolved
+                    if self.last_upscale_seqnum.fetch_max(seqnum, Ordering::AcqRel) > seqnum {
+                        info!(
+                            "received memory.high event, but it came before our last upscale -> ignoring it"
+                        );
+                        continue;
+                    }
+
+                    // The memory.high came after our latest upscale. We don't
+                    // want to do anything yet, so peek the next event in hopes
+                    // that it's an upscale.
+                    if let Some(upscale_num) = self
+                        .upscaled(&mut upscales)
+                        .context("failed to check if we were upscaled")?
+                    {
+                        if upscale_num > seqnum {
+                            info!(
+                                "received memory.high event, but it came before our last upscale -> ignoring it"
+                            );
+                            continue;
+                        }
+                    }
+
+                    // If it's been long enough since we last froze, freeze the
+                    // cgroup and request upscale
+                    if wait_to_freeze.is_elapsed() {
+                        info!("received memory.high event -> requesting upscale");
+                        waiting_on_upscale = self
+                            .handle_memory_high_event(&mut upscales)
+                            .await
+                            .context("failed to handle upscale")?;
+                        wait_to_freeze
+                            .as_mut()
+                            .reset(Instant::now() + self.config.do_not_freeze_more_often_than);
+                        continue;
+                    }
+
+                    // Ok, we can't freeze, just request upscale
+                    if !waiting_on_upscale {
+                        info!("received memory.high event, but too soon to refreeze -> requesting upscale");
+
+                        // Make check to make sure we haven't been upscaled in the
+                        // meantine (can happen if the agent independently decides
+                        // to upscale us again)
+                        if self
+                            .upscaled(&mut upscales)
+                            .context("failed to check if we were upscaled")?
+                            .is_some()
+                        {
+                            info!("no need to request upscaling because we got upscaled");
+                            continue;
+                        }
+                        self.upscale_requester
+                            .send(())
+                            .await
+                            .context("failed to request upscale")?;
+                        continue;
+                    }
+
+                    // Shoot, we can't freeze or and we're still waiting on upscale,
+                    // increase memory.high to reduce throttling
+                    if wait_to_increase_memory_high.is_elapsed() {
+                        info!(
+                            "received memory.high event, \
+                            but too soon to refreeze and already requested upscale \
+                            -> increasing memory.high"
+                        );
+
+                        // Make check to make sure we haven't been upscaled in the
+                        // meantine (can happen if the agent independently decides
+                        // to upscale us again)
+                        if self
+                            .upscaled(&mut upscales)
+                            .context("failed to check if we were upscaled")?
+                            .is_some()
+                        {
+                            info!("no need to increase memory.high because got upscaled");
+                            continue;
+                        }
+
+                        // Request upscale anyways (the agent will handle deduplicating
+                        // requests)
+                        self.upscale_requester
+                            .send(())
+                            .await
+                            .context("failed to request upscale")?;
+
+                        let memory_high =
+                            self.get_high_bytes().context("failed to get memory.high")?;
+                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
+                        info!(
+                            current_high_bytes = memory_high,
+                            new_high_bytes = new_high,
+                            "updating memory.high"
+                        );
+                        self.set_high_bytes(new_high)
+                            .context("failed to set memory.high")?;
+                        wait_to_increase_memory_high
+                            .as_mut()
+                            .reset(Instant::now() + self.config.memory_high_increase_every)
+                    }
+
+                    // we can't do anything
+                }
+            };
+        }
+    }
+
+    /// Handle a `memory.high`, returning whether we are still waiting on upscale
+    /// by the time the function returns.
+    ///
+    /// The general plan for handling a `memory.high` event is as follows:
+    /// 1. Freeze the cgroup
+    /// 2. Start a timer for `self.config.max_upscale_wait`
+    /// 3. Request upscale
+    /// 4. After the timer elapses or we receive upscale, thaw the cgroup.
+    /// 5. Return whether or not we are still waiting for upscale. If we are,
+    ///    we'll increase the cgroups memory.high to avoid getting oom killed
+    #[tracing::instrument(skip_all)]
+    async fn handle_memory_high_event(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<bool> {
+        // Immediately freeze the cgroup before doing anything else.
+        info!("received memory.high event -> freezing cgroup");
+        self.freeze().context("failed to freeze cgroup")?;
+
+        // We'll use this for logging durations
+        let start_time = Instant::now();
+
+        // Await the upscale until we have to unfreeze
+        let timed =
+            tokio::time::timeout(self.config.max_upscale_wait, self.await_upscale(upscales));
+
+        // Request the upscale
+        info!(
+            wait = ?self.config.max_upscale_wait,
+            "sending request for immediate upscaling",
+        );
+        self.upscale_requester
+            .send(())
+            .await
+            .context("failed to request upscale")?;
+
+        let waiting_on_upscale = match timed.await {
+            Ok(Ok(())) => {
+                info!(elapsed = ?start_time.elapsed(), "received upscale in time");
+                false
+            }
+            // **important**: unfreeze the cgroup before ?-reporting the error
+            Ok(Err(e)) => {
+                info!("error waiting for upscale -> thawing cgroup");
+                self.thaw()
+                    .context("failed to thaw cgroup after errored waiting for upscale")?;
+                Err(e.context("failed to await upscale"))?
+            }
+            Err(_) => {
+                info!(elapsed = ?self.config.max_upscale_wait, "timed out waiting for upscale");
+                true
+            }
+        };
+
+        info!("thawing cgroup");
+        self.thaw().context("failed to thaw cgroup")?;
+
+        Ok(waiting_on_upscale)
+    }
+
+    /// Checks whether we were just upscaled, returning the upscale's sequence
+    /// number if so.
+    #[tracing::instrument(skip_all)]
+    fn upscaled(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<Option<u64>> {
+        let Sequenced { seqnum, data } = match upscales.try_recv() {
+            Ok(upscale) => upscale,
+            Err(TryRecvError::Empty) => return Ok(None),
+            Err(TryRecvError::Disconnected) => {
+                bail!("upscale notification channel was disconnected")
+            }
+        };
+
+        // Make sure to update the last upscale sequence number
+        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+        info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
+        Ok(Some(seqnum))
+    }
+
+    /// Await an upscale event, discarding any `memory.high` events received in
+    /// the process.
+    ///
+    /// This is used in `handle_memory_high_event`, where we need to listen
+    /// for upscales in particular so we know if we can thaw the cgroup early.
+    #[tracing::instrument(skip_all)]
+    async fn await_upscale(
+        &self,
+        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+    ) -> anyhow::Result<()> {
+        let Sequenced { seqnum, .. } = upscales
+            .recv()
+            .await
+            .context("error listening for upscales")?;
+
+        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
+        Ok(())
+    }
+
+    /// Get the cgroup's name.
+    pub fn path(&self) -> &str {
+        self.cgroup.path()
+    }
+}
+
+/// Represents a set of limits we apply to a cgroup to control memory usage.
+///
+/// Setting these values also affects the thresholds for receiving usage alerts.
+#[derive(Debug)]
+pub struct MemoryLimits {
+    high: u64,
+    max: u64,
+}
+
+impl MemoryLimits {
+    pub fn new(high: u64, max: u64) -> Self {
+        Self { max, high }
+    }
+}
+
+// Methods for manipulating the actual cgroup
+impl CgroupWatcher {
+    /// Get a handle on the freezer subsystem.
+    fn freezer(&self) -> anyhow::Result<&FreezerController> {
+        if let Some(Freezer(freezer)) = self
+            .cgroup
+            .subsystems()
+            .iter()
+            .find(|sub| matches!(sub, Freezer(_)))
+        {
+            Ok(freezer)
+        } else {
+            anyhow::bail!("could not find freezer subsystem")
+        }
+    }
+
+    /// Attempt to freeze the cgroup.
+    pub fn freeze(&self) -> anyhow::Result<()> {
+        self.freezer()
+            .context("failed to get freezer subsystem")?
+            .freeze()
+            .context("failed to freeze")
+    }
+
+    /// Attempt to thaw the cgroup.
+    pub fn thaw(&self) -> anyhow::Result<()> {
+        self.freezer()
+            .context("failed to get freezer subsystem")?
+            .thaw()
+            .context("failed to thaw")
+    }
+
+    /// Get a handle on the memory subsystem.
+    ///
+    /// Note: this method does not require `self.memory_update_lock` because
+    /// getting a handle to the subsystem does not access any of the files we
+    /// care about, such as memory.high and memory.events
+    fn memory(&self) -> anyhow::Result<&MemController> {
+        if let Some(Mem(memory)) = self
+            .cgroup
+            .subsystems()
+            .iter()
+            .find(|sub| matches!(sub, Mem(_)))
+        {
+            Ok(memory)
+        } else {
+            anyhow::bail!("could not find memory subsystem")
+        }
+    }
+
+    /// Get cgroup current memory usage.
+    pub fn current_memory_usage(&self) -> anyhow::Result<u64> {
+        Ok(self
+            .memory()
+            .context("failed to get memory subsystem")?
+            .memory_stat()
+            .usage_in_bytes)
+    }
+
+    /// Set cgroup memory.high threshold.
+    pub fn set_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
+        self.memory()
+            .context("failed to get memory subsystem")?
+            .set_mem(cgroups_rs::memory::SetMemory {
+                low: None,
+                high: Some(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64)),
+                min: None,
+                max: None,
+            })
+            .context("failed to set memory.high")
+    }
+
+    /// Set cgroup memory.high and memory.max.
+    pub fn set_limits(&self, limits: &MemoryLimits) -> anyhow::Result<()> {
+        info!(
+            limits.high,
+            limits.max,
+            path = self.path(),
+            "writing new memory limits",
+        );
+        self.memory()
+            .context("failed to get memory subsystem while setting memory limits")?
+            .set_mem(cgroups_rs::memory::SetMemory {
+                min: None,
+                low: None,
+                high: Some(MaxValue::Value(
+                    u64::min(limits.high, i64::MAX as u64) as i64
+                )),
+                max: Some(MaxValue::Value(u64::min(limits.max, i64::MAX as u64) as i64)),
+            })
+            .context("failed to set memory limits")
+    }
+
+    /// Given some amount of available memory, set the desired cgroup memory limits
+    pub fn set_memory_limits(&mut self, available_memory: u64) -> anyhow::Result<()> {
+        let new_high = self.config.calculate_memory_high_value(available_memory);
+        let limits = MemoryLimits::new(new_high, available_memory);
+        info!(
+            path = self.path(),
+            memory = ?limits,
+            "setting cgroup memory",
+        );
+        self.set_limits(&limits)
+            .context("failed to set cgroup memory limits")?;
+        Ok(())
+    }
+
+    /// Get memory.high threshold.
+    pub fn get_high_bytes(&self) -> anyhow::Result<u64> {
+        let high = self
+            .memory()
+            .context("failed to get memory subsystem while getting memory statistics")?
+            .get_mem()
+            .map(|mem| mem.high)
+            .context("failed to get memory statistics from subsystem")?;
+        match high {
+            Some(MaxValue::Max) => Ok(i64::MAX as u64),
+            Some(MaxValue::Value(high)) => Ok(high as u64),
+            None => anyhow::bail!("failed to read memory.high from memory subsystem"),
+        }
+    }
+}
--- a/libs/vm_monitor/src/dispatcher.rs
+++ b/libs/vm_monitor/src/dispatcher.rs
@@ -0,0 +1,153 @@
+//! Managing the websocket connection and other signals in the monitor.
+//!
+//! Contains types that manage the interaction (not data interchange, see `protocol`)
+//! between agent and monitor, allowing us to to process and send messages in a
+//! straightforward way. The dispatcher also manages that signals that come from
+//! the cgroup (requesting upscale), and the signals that go to the cgroup
+//! (notifying it of upscale).
+
+use anyhow::{bail, Context};
+use axum::extract::ws::{Message, WebSocket};
+use futures::{
+    stream::{SplitSink, SplitStream},
+    SinkExt, StreamExt,
+};
+use tokio::sync::mpsc;
+use tracing::info;
+
+use crate::cgroup::Sequenced;
+use crate::protocol::{
+    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, Resources, PROTOCOL_MAX_VERSION,
+    PROTOCOL_MIN_VERSION,
+};
+
+/// The central handler for all communications in the monitor.
+///
+/// The dispatcher has two purposes:
+/// 1. Manage the connection to the agent, sending and receiving messages.
+/// 2. Communicate with the cgroup manager, notifying it when upscale is received,
+///    and sending a message to the agent when the cgroup manager requests
+///    upscale.
+#[derive(Debug)]
+pub struct Dispatcher {
+    /// We read agent messages of of `source`
+    pub(crate) source: SplitStream<WebSocket>,
+
+    /// We send messages to the agent through `sink`
+    sink: SplitSink<WebSocket, Message>,
+
+    /// Used to notify the cgroup when we are upscaled.
+    pub(crate) notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
+
+    /// When the cgroup requests upscale it will send on this channel. In response
+    /// we send an `UpscaleRequst` to the agent.
+    pub(crate) request_upscale_events: mpsc::Receiver<()>,
+
+    /// The protocol version we have agreed to use with the agent. This is negotiated
+    /// during the creation of the dispatcher, and should be the highest shared protocol
+    /// version.
+    ///
+    // NOTE: currently unused, but will almost certainly be used in the futures
+    // as the protocol changes
+    #[allow(unused)]
+    pub(crate) proto_version: ProtocolVersion,
+}
+
+impl Dispatcher {
+    /// Creates a new dispatcher using the passed-in connection.
+    ///
+    /// Performs a negotiation with the agent to determine the highest protocol
+    /// version that both support. This consists of two steps:
+    /// 1. Wait for the agent to sent the range of protocols it supports.
+    /// 2. Send a protocol version that works for us as well, or an error if there
+    ///    is no compatible version.
+    pub async fn new(
+        stream: WebSocket,
+        notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
+        request_upscale_events: mpsc::Receiver<()>,
+    ) -> anyhow::Result<Self> {
+        let (mut sink, mut source) = stream.split();
+
+        // Figure out the highest protocol version we both support
+        info!("waiting for agent to send protocol version range");
+        let Some(message) = source.next().await else {
+            bail!("websocket connection closed while performing protocol handshake")
+        };
+
+        let message = message.context("failed to read protocol version range off connection")?;
+
+        let Message::Text(message_text) = message else {
+            // All messages should be in text form, since we don't do any
+            // pinging/ponging. See nhooyr/websocket's implementation and the
+            // agent for more info
+            bail!("received non-text message during proocol handshake: {message:?}")
+        };
+
+        let monitor_range = ProtocolRange {
+            min: PROTOCOL_MIN_VERSION,
+            max: PROTOCOL_MAX_VERSION,
+        };
+
+        let agent_range: ProtocolRange = serde_json::from_str(&message_text)
+            .context("failed to deserialize protocol version range")?;
+
+        info!(range = ?agent_range, "received protocol version range");
+
+        let highest_shared_version = match monitor_range.highest_shared_version(&agent_range) {
+            Ok(version) => {
+                sink.send(Message::Text(
+                    serde_json::to_string(&ProtocolResponse::Version(version)).unwrap(),
+                ))
+                .await
+                .context("failed to notify agent of negotiated protocol version")?;
+                version
+            }
+            Err(e) => {
+                sink.send(Message::Text(
+                    serde_json::to_string(&ProtocolResponse::Error(format!(
+                        "Received protocol version range {} which does not overlap with {}",
+                        agent_range, monitor_range
+                    )))
+                    .unwrap(),
+                ))
+                .await
+                .context("failed to notify agent of no overlap between protocol version ranges")?;
+                Err(e).context("error determining suitable protocol version range")?
+            }
+        };
+
+        Ok(Self {
+            sink,
+            source,
+            notify_upscale_events,
+            request_upscale_events,
+            proto_version: highest_shared_version,
+        })
+    }
+
+    /// Notify the cgroup manager that we have received upscale and wait for
+    /// the acknowledgement.
+    #[tracing::instrument(skip_all, fields(?resources))]
+    pub async fn notify_upscale(&self, resources: Sequenced<Resources>) -> anyhow::Result<()> {
+        self.notify_upscale_events
+            .send(resources)
+            .await
+            .context("failed to send resources and oneshot sender across channel")
+    }
+
+    /// Send a message to the agent.
+    ///
+    /// Although this function is small, it has one major benefit: it is the only
+    /// way to send data accross the connection, and you can only pass in a proper
+    /// `MonitorMessage`. Without safeguards like this, it's easy to accidentally
+    /// serialize the wrong thing and send it, since `self.sink.send` will take
+    /// any string.
+    pub async fn send(&mut self, message: OutboundMsg) -> anyhow::Result<()> {
+        info!(?message, "sending message");
+        let json = serde_json::to_string(&message).context("failed to serialize message")?;
+        self.sink
+            .send(Message::Text(json))
+            .await
+            .context("stream error sending message")
+    }
+}
--- a/libs/vm_monitor/src/filecache.rs
+++ b/libs/vm_monitor/src/filecache.rs
@@ -0,0 +1,316 @@
+//! Logic for configuring and scaling the Postgres file cache.
+
+use std::num::NonZeroU64;
+
+use crate::MiB;
+use anyhow::{anyhow, Context};
+use tokio_postgres::{types::ToSql, Client, NoTls, Row};
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info};
+
+/// Manages Postgres' file cache by keeping a connection open.
+#[derive(Debug)]
+pub struct FileCacheState {
+    client: Client,
+    conn_str: String,
+    pub(crate) config: FileCacheConfig,
+
+    /// A token for cancelling spawned threads during shutdown.
+    token: CancellationToken,
+}
+
+#[derive(Debug)]
+pub struct FileCacheConfig {
+    /// Whether the file cache is *actually* stored in memory (e.g. by writing to
+    /// a tmpfs or shmem file). If true, the size of the file cache will be counted against the
+    /// memory available for the cgroup.
+    pub(crate) in_memory: bool,
+
+    /// The size of the file cache, in terms of the size of the resource it consumes
+    /// (currently: only memory)
+    ///
+    /// For example, setting `resource_multipler = 0.75` gives the cache a target size of 75% of total
+    /// resources.
+    ///
+    /// This value must be strictly between 0 and 1.
+    resource_multiplier: f64,
+
+    /// The required minimum amount of memory, in bytes, that must remain available
+    /// after subtracting the file cache.
+    ///
+    /// This value must be non-zero.
+    min_remaining_after_cache: NonZeroU64,
+
+    /// Controls the rate of increase in the file cache's size as it grows from zero
+    /// (when total resources equals min_remaining_after_cache) to the desired size based on
+    /// `resource_multiplier`.
+    ///
+    /// A `spread_factor` of zero means that all additional resources will go to the cache until it
+    /// reaches the desired size. Setting `spread_factor` to N roughly means "for every 1 byte added to
+    /// the cache's size, N bytes are reserved for the rest of the system, until the cache gets to
+    /// its desired size".
+    ///
+    /// This value must be >= 0, and must retain an increase that is more than what would be given by
+    /// `resource_multiplier`. For example, setting `resource_multiplier` = 0.75 but `spread_factor` = 1
+    /// would be invalid, because `spread_factor` would induce only 50% usage - never reaching the 75%
+    /// as desired by `resource_multiplier`.
+    ///
+    /// `spread_factor` is too large if `(spread_factor + 1) * resource_multiplier >= 1`.
+    spread_factor: f64,
+}
+
+impl FileCacheConfig {
+    pub fn default_in_memory() -> Self {
+        Self {
+            in_memory: true,
+            // 75 %
+            resource_multiplier: 0.75,
+            // 640 MiB; (512 + 128)
+            min_remaining_after_cache: NonZeroU64::new(640 * MiB).unwrap(),
+            // ensure any increase in file cache size is split 90-10 with 10% to other memory
+            spread_factor: 0.1,
+        }
+    }
+
+    pub fn default_on_disk() -> Self {
+        Self {
+            in_memory: false,
+            resource_multiplier: 0.75,
+            // 256 MiB - lower than when in memory because overcommitting is safe; if we don't have
+            // memory, the kernel will just evict from its page cache, rather than e.g. killing
+            // everything.
+            min_remaining_after_cache: NonZeroU64::new(256 * MiB).unwrap(),
+            spread_factor: 0.1,
+        }
+    }
+
+    /// Make sure fields of the config are consistent.
+    pub fn validate(&self) -> anyhow::Result<()> {
+        // Single field validity
+        anyhow::ensure!(
+            0.0 < self.resource_multiplier && self.resource_multiplier < 1.0,
+            "resource_multiplier must be between 0.0 and 1.0 exclusive, got {}",
+            self.resource_multiplier
+        );
+        anyhow::ensure!(
+            self.spread_factor >= 0.0,
+            "spread_factor must be >= 0, got {}",
+            self.spread_factor
+        );
+
+        // Check that `resource_multiplier` and `spread_factor` are valid w.r.t. each other.
+        //
+        // As shown in `calculate_cache_size`, we have two lines resulting from `resource_multiplier` and
+        // `spread_factor`, respectively. They are:
+        //
+        //                 `total`           `min_remaining_after_cache`
+        //   size = ————————————————————— - —————————————————————————————
+        //           `spread_factor` + 1         `spread_factor` + 1
+        //
+        // and
+        //
+        //   size = `resource_multiplier` × total
+        //
+        // .. where `total` is the total resources. These are isomorphic to the typical 'y = mx + b'
+        // form, with y = "size" and x = "total".
+        //
+        // These lines intersect at:
+        //
+        //               `min_remaining_after_cache`
+        //   ———————————————————————————————————————————————————
+        //    1 - `resource_multiplier` × (`spread_factor` + 1)
+        //
+        // We want to ensure that this value (a) exists, and (b) is >= `min_remaining_after_cache`. This is
+        // guaranteed when '`resource_multiplier` × (`spread_factor` + 1)' is less than 1.
+        // (We also need it to be >= 0, but that's already guaranteed.)
+
+        let intersect_factor = self.resource_multiplier * (self.spread_factor + 1.0);
+        anyhow::ensure!(
+            intersect_factor < 1.0,
+            "incompatible resource_multipler and spread_factor"
+        );
+        Ok(())
+    }
+
+    /// Calculate the desired size of the cache, given the total memory
+    pub fn calculate_cache_size(&self, total: u64) -> u64 {
+        // *Note*: all units are in bytes, until the very last line.
+        let available = total.saturating_sub(self.min_remaining_after_cache.get());
+        if available == 0 {
+            return 0;
+        }
+
+        // Conversions to ensure we don't overflow from floating-point ops
+        let size_from_spread =
+            i64::max(0, (available as f64 / (1.0 + self.spread_factor)) as i64) as u64;
+
+        let size_from_normal = (total as f64 * self.resource_multiplier) as u64;
+
+        let byte_size = u64::min(size_from_spread, size_from_normal);
+
+        // The file cache operates in units of mebibytes, so the sizes we produce should
+        // be rounded to a mebibyte. We round down to be conservative.
+        byte_size / MiB * MiB
+    }
+}
+
+impl FileCacheState {
+    /// Connect to the file cache.
+    #[tracing::instrument(skip_all, fields(%conn_str, ?config))]
+    pub async fn new(
+        conn_str: &str,
+        config: FileCacheConfig,
+        token: CancellationToken,
+    ) -> anyhow::Result<Self> {
+        config.validate().context("file cache config is invalid")?;
+
+        info!(conn_str, "connecting to Postgres file cache");
+        let client = FileCacheState::connect(conn_str, token.clone())
+            .await
+            .context("failed to connect to postgres file cache")?;
+
+        let conn_str = conn_str.to_string();
+        Ok(Self {
+            client,
+            config,
+            conn_str,
+            token,
+        })
+    }
+
+    /// Connect to Postgres.
+    ///
+    /// Aborts the spawned thread if the kill signal is received. This is not
+    /// a method as it is called in [`FileCacheState::new`].
+    #[tracing::instrument(skip_all, fields(%conn_str))]
+    async fn connect(conn_str: &str, token: CancellationToken) -> anyhow::Result<Client> {
+        let (client, conn) = tokio_postgres::connect(conn_str, NoTls)
+            .await
+            .context("failed to connect to pg client")?;
+
+        // The connection object performs the actual communication with the database,
+        // so spawn it off to run on its own. See tokio-postgres docs.
+        crate::spawn_with_cancel(
+            token,
+            |res| {
+                if let Err(error) = res {
+                    error!(%error, "postgres error")
+                }
+            },
+            conn,
+        );
+
+        Ok(client)
+    }
+
+    /// Execute a query with a retry if necessary.
+    ///
+    /// If the initial query fails, we restart the database connection and attempt
+    /// if again.
+    #[tracing::instrument(skip_all, fields(%statement))]
+    pub async fn query_with_retry(
+        &mut self,
+        statement: &str,
+        params: &[&(dyn ToSql + Sync)],
+    ) -> anyhow::Result<Vec<Row>> {
+        match self
+            .client
+            .query(statement, params)
+            .await
+            .context("failed to execute query")
+        {
+            Ok(rows) => Ok(rows),
+            Err(e) => {
+                error!(error = ?e, "postgres error: {e} -> retrying");
+
+                let client = FileCacheState::connect(&self.conn_str, self.token.clone())
+                    .await
+                    .context("failed to connect to postgres file cache")?;
+                info!("successfully reconnected to postgres client");
+
+                // Replace the old client and attempt the query with the new one
+                self.client = client;
+                self.client
+                    .query(statement, params)
+                    .await
+                    .context("failed to execute query a second time")
+            }
+        }
+    }
+
+    /// Get the current size of the file cache.
+    #[tracing::instrument(skip_all)]
+    pub async fn get_file_cache_size(&mut self) -> anyhow::Result<u64> {
+        self.query_with_retry(
+            // The file cache GUC variable is in MiB, but the conversion with
+            // pg_size_bytes means that the end result we get is in bytes.
+            "SELECT pg_size_bytes(current_setting('neon.file_cache_size_limit'));",
+            &[],
+        )
+        .await
+        .context("failed to query pg for file cache size")?
+        .first()
+        .ok_or_else(|| anyhow!("file cache size query returned no rows"))?
+        // pg_size_bytes returns a bigint which is the same as an i64.
+        .try_get::<_, i64>(0)
+        // Since the size of the table is not negative, the cast is sound.
+        .map(|bytes| bytes as u64)
+        .context("failed to extract file cache size from query result")
+    }
+
+    /// Attempt to set the file cache size, returning the size it was actually
+    /// set to.
+    #[tracing::instrument(skip_all, fields(%num_bytes))]
+    pub async fn set_file_cache_size(&mut self, num_bytes: u64) -> anyhow::Result<u64> {
+        let max_bytes = self
+            // The file cache GUC variable is in MiB, but the conversion with pg_size_bytes
+            // means that the end result we get is in bytes.
+            .query_with_retry(
+                "SELECT pg_size_bytes(current_setting('neon.max_file_cache_size'));",
+                &[],
+            )
+            .await
+            .context("failed to query pg for max file cache size")?
+            .first()
+            .ok_or_else(|| anyhow!("max file cache size query returned no rows"))?
+            .try_get::<_, i64>(0)
+            .map(|bytes| bytes as u64)
+            .context("failed to extract max file cache size from query result")?;
+
+        let max_mb = max_bytes / MiB;
+        let num_mb = u64::min(num_bytes, max_bytes) / MiB;
+
+        let capped = if num_bytes > max_bytes {
+            " (capped by maximum size)"
+        } else {
+            ""
+        };
+
+        info!(
+            size = num_mb,
+            max = max_mb,
+            "updating file cache size {capped}",
+        );
+
+        // note: even though the normal ways to get the cache size produce values with trailing "MB"
+        // (hence why we call pg_size_bytes in `get_file_cache_size`'s query), the format
+        // it expects to set the value is "integer number of MB" without trailing units.
+        // For some reason, this *really* wasn't working with normal arguments, so that's
+        // why we're constructing the query here.
+        self.client
+            .query(
+                &format!("ALTER SYSTEM SET neon.file_cache_size_limit = {};", num_mb),
+                &[],
+            )
+            .await
+            .context("failed to change file cache size limit")?;
+
+        // must use pg_reload_conf to have the settings change take effect
+        self.client
+            .execute("SELECT pg_reload_conf();", &[])
+            .await
+            .context("failed to reload config")?;
+
+        Ok(num_mb * MiB)
+    }
+}
--- a/libs/vm_monitor/src/lib.rs
+++ b/libs/vm_monitor/src/lib.rs
@@ -0,0 +1,218 @@
+#![cfg(target_os = "linux")]
+
+use anyhow::Context;
+use axum::{
+    extract::{ws::WebSocket, State, WebSocketUpgrade},
+    response::Response,
+};
+use axum::{routing::get, Router, Server};
+use clap::Parser;
+use futures::Future;
+use std::{fmt::Debug, time::Duration};
+use sysinfo::{RefreshKind, System, SystemExt};
+use tokio::{sync::broadcast, task::JoinHandle};
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info};
+
+use runner::Runner;
+
+// Code that interfaces with agent
+pub mod dispatcher;
+pub mod protocol;
+
+pub mod cgroup;
+pub mod filecache;
+pub mod runner;
+
+/// The vm-monitor is an autoscaling component started by compute_ctl.
+///
+/// It carries out autoscaling decisions (upscaling/downscaling) and responds to
+/// memory pressure by making requests to the autoscaler-agent.
+#[derive(Debug, Parser)]
+pub struct Args {
+    /// The name of the cgroup we should monitor for memory.high events. This
+    /// is the cgroup that postgres should be running in.
+    #[arg(short, long)]
+    pub cgroup: Option<String>,
+
+    /// The connection string for the Postgres file cache we should manage.
+    #[arg(short, long)]
+    pub pgconnstr: Option<String>,
+
+    /// Flag to signal that the Postgres file cache is on disk (i.e. not in memory aside from the
+    /// kernel's page cache), and therefore should not count against available memory.
+    //
+    // NB: Ideally this flag would directly refer to whether the file cache is in memory (rather
+    // than a roundabout way, via whether it's on disk), but in order to be backwards compatible
+    // during the switch away from an in-memory file cache, we had to default to the previous
+    // behavior.
+    #[arg(long)]
+    pub file_cache_on_disk: bool,
+
+    /// The address we should listen on for connection requests. For the
+    /// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
+    #[arg(short, long)]
+    pub addr: String,
+}
+
+impl Args {
+    pub fn addr(&self) -> &str {
+        &self.addr
+    }
+}
+
+/// The number of bytes in one mebibyte.
+#[allow(non_upper_case_globals)]
+const MiB: u64 = 1 << 20;
+
+/// Convert a quantity in bytes to a quantity in mebibytes, generally for display
+/// purposes. (Most calculations in this crate use bytes directly)
+pub fn bytes_to_mebibytes(bytes: u64) -> f32 {
+    (bytes as f32) / (MiB as f32)
+}
+
+pub fn get_total_system_memory() -> u64 {
+    System::new_with_specifics(RefreshKind::new().with_memory()).total_memory()
+}
+
+/// Global app state for the Axum server
+#[derive(Debug, Clone)]
+pub struct ServerState {
+    /// Used to close old connections.
+    ///
+    /// When a new connection is made, we send a message signalling to the old
+    /// connection to close.
+    pub sender: broadcast::Sender<()>,
+
+    /// Used to cancel all spawned threads in the monitor.
+    pub token: CancellationToken,
+
+    // The CLI args
+    pub args: &'static Args,
+}
+
+/// Spawn a thread that may get cancelled by the provided [`CancellationToken`].
+///
+/// This is mainly meant to be called with futures that will be pending for a very
+/// long time, or are not mean to return. If it is not desirable for the future to
+/// ever resolve, such as in the case of [`cgroup::CgroupWatcher::watch`], the error can
+/// be logged with `f`.
+pub fn spawn_with_cancel<T, F>(
+    token: CancellationToken,
+    f: F,
+    future: T,
+) -> JoinHandle<Option<T::Output>>
+where
+    T: Future + Send + 'static,
+    T::Output: Send + 'static,
+    F: FnOnce(&T::Output) + Send + 'static,
+{
+    tokio::spawn(async move {
+        tokio::select! {
+            _ = token.cancelled() => {
+                info!("received global kill signal");
+                None
+            }
+            res = future => {
+                f(&res);
+                Some(res)
+            }
+        }
+    })
+}
+
+/// The entrypoint to the binary.
+///
+/// Set up tracing, parse arguments, and start an http server.
+pub async fn start(args: &'static Args, token: CancellationToken) -> anyhow::Result<()> {
+    // This channel is used to close old connections. When a new connection is
+    // made, we send a message signalling to the old connection to close.
+    let (sender, _) = tokio::sync::broadcast::channel::<()>(1);
+
+    let app = Router::new()
+        // This route gets upgraded to a websocket connection. We only support
+        // one connection at a time, which we enforce by killing old connections
+        // when we receive a new one.
+        .route("/monitor", get(ws_handler))
+        .with_state(ServerState {
+            sender,
+            token,
+            args,
+        });
+
+    let addr = args.addr();
+    let bound = Server::try_bind(&addr.parse().expect("parsing address should not fail"))
+        .with_context(|| format!("failed to bind to {addr}"))?;
+
+    info!(addr, "server bound");
+
+    bound
+        .serve(app.into_make_service())
+        .await
+        .context("server exited")?;
+
+    Ok(())
+}
+
+/// Handles incoming websocket connections.
+///
+/// If we are already to connected to an agent, we kill that old connection
+/// and accept the new one.
+#[tracing::instrument(name = "/monitor", skip_all, fields(?args))]
+pub async fn ws_handler(
+    ws: WebSocketUpgrade,
+    State(ServerState {
+        sender,
+        token,
+        args,
+    }): State<ServerState>,
+) -> Response {
+    // Kill the old monitor
+    info!("closing old connection if there is one");
+    let _ = sender.send(());
+
+    // Start the new one. Wow, the cycle of death and rebirth
+    let closer = sender.subscribe();
+    ws.on_upgrade(|ws| start_monitor(ws, args, closer, token))
+}
+
+/// Starts the monitor. If startup fails or the monitor exits, an error will
+/// be logged and our internal state will be reset to allow for new connections.
+#[tracing::instrument(skip_all)]
+async fn start_monitor(
+    ws: WebSocket,
+    args: &Args,
+    kill: broadcast::Receiver<()>,
+    token: CancellationToken,
+) {
+    info!(
+        ?args,
+        "accepted new websocket connection -> starting monitor"
+    );
+    let timeout = Duration::from_secs(4);
+    let monitor = tokio::time::timeout(
+        timeout,
+        Runner::new(Default::default(), args, ws, kill, token),
+    )
+    .await;
+    let mut monitor = match monitor {
+        Ok(Ok(monitor)) => monitor,
+        Ok(Err(error)) => {
+            error!(?error, "failed to create monitor");
+            return;
+        }
+        Err(_) => {
+            error!(
+                ?timeout,
+                "creating monitor timed out (probably waiting to receive protocol range)"
+            );
+            return;
+        }
+    };
+    info!("connected to agent");
+
+    match monitor.run().await {
+        Ok(()) => info!("monitor was killed due to new connection"),
+        Err(e) => error!(error = ?e, "monitor terminated unexpectedly"),
+    }
+}
--- a/libs/vm_monitor/src/protocol.rs
+++ b/libs/vm_monitor/src/protocol.rs
@@ -0,0 +1,241 @@
+//! Types representing protocols and actual agent-monitor messages.
+//!
+//! The pervasive use of serde modifiers throughout this module is to ease
+//! serialization on the go side. Because go does not have enums (which model
+//! messages well), it is harder to model messages, and we accomodate that with
+//! serde.
+//!
+//! *Note*: the agent sends and receives messages in different ways.
+//!
+//! The agent serializes messages in the form and then sends them. The use
+//! of `#[serde(tag = "type", content = "content")]` allows us to use `Type`
+//! to determine how to deserialize `Content`.
+//! ```ignore
+//! struct {
+//!     Content any
+//!     Type    string
+//!     Id      uint64
+//! }
+//! ```
+//! and receives messages in the form:
+//! ```ignore
+//! struct {
+//!     {fields embedded}
+//!     Type string
+//!     Id   uint64
+//! }
+//! ```
+//! After reading the type field, the agent will decode the entire message
+//! again, this time into the correct type using the embedded fields.
+//! Because the agent cannot just extract the json contained in a certain field
+//! (it initially deserializes to `map[string]interface{}`), we keep the fields
+//! at the top level, so the entire piece of json can be deserialized into a struct,
+//! such as a `DownscaleResult`, with the `Type` and `Id` fields ignored.
+
+use core::fmt;
+use std::cmp;
+
+use serde::{de::Error, Deserialize, Serialize};
+
+/// A Message we send to the agent.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct OutboundMsg {
+    #[serde(flatten)]
+    pub(crate) inner: OutboundMsgKind,
+    pub(crate) id: usize,
+}
+
+impl OutboundMsg {
+    pub fn new(inner: OutboundMsgKind, id: usize) -> Self {
+        Self { inner, id }
+    }
+}
+
+/// The different underlying message types we can send to the agent.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+#[serde(tag = "type")]
+pub enum OutboundMsgKind {
+    /// Indicates that the agent sent an invalid message, i.e, we couldn't
+    /// properly deserialize it.
+    InvalidMessage { error: String },
+    /// Indicates that we experienced an internal error while processing a message.
+    /// For example, if a cgroup operation fails while trying to handle an upscale,
+    /// we return `InternalError`.
+    InternalError { error: String },
+    /// Returned to the agent once we have finished handling an upscale. If the
+    /// handling was unsuccessful, an `InternalError` will get returned instead.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    UpscaleConfirmation {},
+    /// Indicates to the monitor that we are urgently requesting resources.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    UpscaleRequest {},
+    /// Returned to the agent once we have finished attempting to downscale. If
+    /// an error occured trying to do so, an `InternalError` will get returned instead.
+    /// However, if we are simply unsuccessful (for example, do to needing the resources),
+    /// that gets included in the `DownscaleResult`.
+    DownscaleResult {
+        // FIXME for the future (once the informant is deprecated)
+        // As of the time of writing, the agent/informant version of this struct is
+        // called api.DownscaleResult. This struct has uppercase fields which are
+        // serialized as such. Thus, we serialize using uppercase names so we don't
+        // have to make a breaking change to the agent<->informant protocol. Once
+        // the informant has been superseded by the monitor, we can add the correct
+        // struct tags to api.DownscaleResult without causing a breaking change,
+        // since we don't need to support the agent<->informant protocol anymore.
+        #[serde(rename = "Ok")]
+        ok: bool,
+        #[serde(rename = "Status")]
+        status: String,
+    },
+    /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
+    /// agent.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    HealthCheck {},
+}
+
+/// A message received form the agent.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct InboundMsg {
+    #[serde(flatten)]
+    pub(crate) inner: InboundMsgKind,
+    pub(crate) id: usize,
+}
+
+/// The different underlying message types we can receive from the agent.
+#[derive(Serialize, Deserialize, Debug, Clone)]
+#[serde(tag = "type", content = "content")]
+pub enum InboundMsgKind {
+    /// Indicates that the we sent an invalid message, i.e, we couldn't
+    /// properly deserialize it.
+    InvalidMessage { error: String },
+    /// Indicates that the informan experienced an internal error while processing
+    /// a message. For example, if it failed to request upsacle from the agent, it
+    /// would return an `InternalError`.
+    InternalError { error: String },
+    /// Indicates to us that we have been granted more resources. We should respond
+    /// with an `UpscaleConfirmation` when done handling the resources (increasins
+    /// file cache size, cgorup memory limits).
+    UpscaleNotification { granted: Resources },
+    /// A request to reduce resource usage. We should response with a `DownscaleResult`,
+    /// when done.
+    DownscaleRequest { target: Resources },
+    /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
+    /// agent.
+    /// *Note*: this is a struct variant because of the way go serializes struct{}
+    HealthCheck {},
+}
+
+/// Represents the resources granted to a VM.
+#[derive(Serialize, Deserialize, Debug, Clone, Copy)]
+// Renamed because the agent has multiple resources types:
+// `Resources` (milliCPU/memory slots)
+// `Allocation` (vCPU/bytes) <- what we correspond to
+#[serde(rename(serialize = "Allocation", deserialize = "Allocation"))]
+pub struct Resources {
+    /// Number of vCPUs
+    pub(crate) cpu: f64,
+    /// Bytes of memory
+    pub(crate) mem: u64,
+}
+
+impl Resources {
+    pub fn new(cpu: f64, mem: u64) -> Self {
+        Self { cpu, mem }
+    }
+}
+
+pub const PROTOCOL_MIN_VERSION: ProtocolVersion = ProtocolVersion::V1_0;
+pub const PROTOCOL_MAX_VERSION: ProtocolVersion = ProtocolVersion::V1_0;
+
+#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Ord, Eq, Serialize, Deserialize)]
+pub struct ProtocolVersion(u8);
+
+impl ProtocolVersion {
+    /// Represents v1.0 of the agent<-> monitor protocol - the initial version
+    ///
+    /// Currently the latest version.
+    const V1_0: ProtocolVersion = ProtocolVersion(1);
+}
+
+impl fmt::Display for ProtocolVersion {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match *self {
+            ProtocolVersion(0) => f.write_str("<invalid: zero>"),
+            ProtocolVersion::V1_0 => f.write_str("v1.0"),
+            other => write!(f, "<unknown: {other}>"),
+        }
+    }
+}
+
+/// A set of protocol bounds that determines what we are speaking.
+///
+/// These bounds are inclusive.
+#[derive(Debug)]
+pub struct ProtocolRange {
+    pub min: ProtocolVersion,
+    pub max: ProtocolVersion,
+}
+
+// Use a custom deserialize impl to ensure that `self.min <= self.max`
+impl<'de> Deserialize<'de> for ProtocolRange {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        #[derive(Deserialize)]
+        struct InnerProtocolRange {
+            min: ProtocolVersion,
+            max: ProtocolVersion,
+        }
+        let InnerProtocolRange { min, max } = InnerProtocolRange::deserialize(deserializer)?;
+        if min > max {
+            Err(D::Error::custom(format!(
+                "min version = {min} is greater than max version = {max}",
+            )))
+        } else {
+            Ok(ProtocolRange { min, max })
+        }
+    }
+}
+
+impl fmt::Display for ProtocolRange {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        if self.min == self.max {
+            f.write_fmt(format_args!("{}", self.max))
+        } else {
+            f.write_fmt(format_args!("{} to {}", self.min, self.max))
+        }
+    }
+}
+
+impl ProtocolRange {
+    /// Find the highest shared version between two `ProtocolRange`'s
+    pub fn highest_shared_version(&self, other: &Self) -> anyhow::Result<ProtocolVersion> {
+        // We first have to make sure the ranges are overlapping. Once we know
+        // this, we can merge the ranges by taking the max of the mins and the
+        // mins of the maxes.
+        if self.min > other.max {
+            anyhow::bail!(
+                "Non-overlapping bounds: other.max = {} was less than self.min = {}",
+                other.max,
+                self.min,
+            )
+        } else if self.max < other.min {
+            anyhow::bail!(
+                "Non-overlappinng bounds: self.max = {} was less than other.min = {}",
+                self.max,
+                other.min
+            )
+        } else {
+            Ok(cmp::min(self.max, other.max))
+        }
+    }
+}
+
+/// We send this to the monitor after negotiating which protocol to use
+#[derive(Serialize, Debug)]
+#[serde(rename_all = "camelCase")]
+pub enum ProtocolResponse {
+    Error(String),
+    Version(ProtocolVersion),
+}
--- a/libs/vm_monitor/src/runner.rs
+++ b/libs/vm_monitor/src/runner.rs
@@ -0,0 +1,478 @@
+//! Exposes the `Runner`, which handles messages received from agent and
+//! sends upscale requests.
+//!
+//! This is the "Monitor" part of the monitor binary and is the main entrypoint for
+//! all functionality.
+
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+use std::{fmt::Debug, mem};
+
+use anyhow::{bail, Context};
+use axum::extract::ws::{Message, WebSocket};
+use futures::StreamExt;
+use tokio::sync::broadcast;
+use tokio::sync::mpsc;
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info, warn};
+
+use crate::cgroup::{CgroupWatcher, MemoryLimits, Sequenced};
+use crate::dispatcher::Dispatcher;
+use crate::filecache::{FileCacheConfig, FileCacheState};
+use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
+use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args, MiB};
+
+/// Central struct that interacts with agent, dispatcher, and cgroup to handle
+/// signals from the agent.
+#[derive(Debug)]
+pub struct Runner {
+    config: Config,
+    filecache: Option<FileCacheState>,
+    cgroup: Option<Arc<CgroupWatcher>>,
+    dispatcher: Dispatcher,
+
+    /// We "mint" new message ids by incrementing this counter and taking the value.
+    ///
+    /// **Note**: This counter is always odd, so that we avoid collisions between the IDs generated
+    /// by us vs the autoscaler-agent.
+    counter: usize,
+
+    last_upscale_request_at: Option<Instant>,
+
+    /// A signal to kill the main thread produced by `self.run()`. This is triggered
+    /// when the server receives a new connection. When the thread receives the
+    /// signal off this channel, it will gracefully shutdown.
+    kill: broadcast::Receiver<()>,
+}
+
+/// Configuration for a `Runner`
+#[derive(Debug)]
+pub struct Config {
+    /// `sys_buffer_bytes` gives the estimated amount of memory, in bytes, that the kernel uses before
+    /// handing out the rest to userspace. This value is the estimated difference between the
+    /// *actual* physical memory and the amount reported by `grep MemTotal /proc/meminfo`.
+    ///
+    /// For more information, refer to `man 5 proc`, which defines MemTotal as "Total usable RAM
+    /// (i.e., physical RAM minus a few reserved bits and the kernel binary code)".
+    ///
+    /// We only use `sys_buffer_bytes` when calculating the system memory from the *external* memory
+    /// size, rather than the self-reported memory size, according to the kernel.
+    ///
+    /// TODO: this field is only necessary while we still have to trust the autoscaler-agent's
+    /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
+    /// should be removed once we have a better solution there.
+    sys_buffer_bytes: u64,
+}
+
+impl Default for Config {
+    fn default() -> Self {
+        Self {
+            sys_buffer_bytes: 100 * MiB,
+        }
+    }
+}
+
+impl Runner {
+    /// Create a new monitor.
+    #[tracing::instrument(skip_all, fields(?config, ?args))]
+    pub async fn new(
+        config: Config,
+        args: &Args,
+        ws: WebSocket,
+        kill: broadcast::Receiver<()>,
+        token: CancellationToken,
+    ) -> anyhow::Result<Runner> {
+        anyhow::ensure!(
+            config.sys_buffer_bytes != 0,
+            "invalid monitor Config: sys_buffer_bytes cannot be 0"
+        );
+
+        // *NOTE*: the dispatcher and cgroup manager talk through these channels
+        // so make sure they each get the correct half, nothing is droppped, etc.
+        let (notified_send, notified_recv) = mpsc::channel(1);
+        let (requesting_send, requesting_recv) = mpsc::channel(1);
+
+        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+            .await
+            .context("error creating new dispatcher")?;
+
+        let mut state = Runner {
+            config,
+            filecache: None,
+            cgroup: None,
+            dispatcher,
+            counter: 1, // NB: must be odd, see the comment about the field for more.
+            last_upscale_request_at: None,
+            kill,
+        };
+
+        let mut file_cache_reserved_bytes = 0;
+        let mem = get_total_system_memory();
+
+        // We need to process file cache initialization before cgroup initialization, so that the memory
+        // allocated to the file cache is appropriately taken into account when we decide the cgroup's
+        // memory limits.
+        if let Some(connstr) = &args.pgconnstr {
+            info!("initializing file cache");
+            let config = match args.file_cache_on_disk {
+                true => FileCacheConfig::default_on_disk(),
+                false => FileCacheConfig::default_in_memory(),
+            };
+
+            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
+                .await
+                .context("failed to create file cache")?;
+
+            let size = file_cache
+                .get_file_cache_size()
+                .await
+                .context("error getting file cache size")?;
+
+            let new_size = file_cache.config.calculate_cache_size(mem);
+            info!(
+                initial = bytes_to_mebibytes(size),
+                new = bytes_to_mebibytes(new_size),
+                "setting initial file cache size",
+            );
+
+            // note: even if size == new_size, we want to explicitly set it, just
+            // to make sure that we have the permissions to do so
+            let actual_size = file_cache
+                .set_file_cache_size(new_size)
+                .await
+                .context("failed to set file cache size, possibly due to inadequate permissions")?;
+            if actual_size != new_size {
+                info!("file cache size actually got set to {actual_size}")
+            }
+            // Mark the resources given to the file cache as reserved, but only if it's in memory.
+            if !args.file_cache_on_disk {
+                file_cache_reserved_bytes = actual_size;
+            }
+
+            state.filecache = Some(file_cache);
+        }
+
+        if let Some(name) = &args.cgroup {
+            let (mut cgroup, cgroup_event_stream) =
+                CgroupWatcher::new(name.clone(), requesting_send)
+                    .context("failed to create cgroup manager")?;
+
+            let available = mem - file_cache_reserved_bytes;
+
+            cgroup
+                .set_memory_limits(available)
+                .context("failed to set cgroup memory limits")?;
+
+            let cgroup = Arc::new(cgroup);
+
+            // Some might call this . . . cgroup v2
+            let cgroup_clone = Arc::clone(&cgroup);
+
+            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
+                cgroup_clone.watch(notified_recv, cgroup_event_stream).await
+            });
+
+            state.cgroup = Some(cgroup);
+        } else {
+            // *NOTE*: We need to forget the sender so that its drop impl does not get ran.
+            // This allows us to poll it in `Monitor::run` regardless of whether we
+            // are managing a cgroup or not. If we don't forget it, all receives will
+            // immediately return an error because the sender is droped and it will
+            // claim all select! statements, effectively turning `Monitor::run` into
+            // `loop { fail to receive }`.
+            mem::forget(requesting_send);
+        }
+
+        Ok(state)
+    }
+
+    /// Attempt to downscale filecache + cgroup
+    #[tracing::instrument(skip_all, fields(?target))]
+    pub async fn try_downscale(&mut self, target: Resources) -> anyhow::Result<(bool, String)> {
+        // Nothing to adjust
+        if self.cgroup.is_none() && self.filecache.is_none() {
+            info!("no action needed for downscale (no cgroup or file cache enabled)");
+            return Ok((
+                true,
+                "monitor is not managing cgroup or file cache".to_string(),
+            ));
+        }
+
+        let requested_mem = target.mem;
+        let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
+        let expected_file_cache_mem_usage = self
+            .filecache
+            .as_ref()
+            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
+            .unwrap_or(0);
+        let mut new_cgroup_mem_high = 0;
+        if let Some(cgroup) = &self.cgroup {
+            new_cgroup_mem_high = cgroup
+                .config
+                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+
+            let current = cgroup
+                .current_memory_usage()
+                .context("failed to fetch cgroup memory")?;
+
+            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+                let status = format!(
+                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
+                    "calculated memory.high too low",
+                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    bytes_to_mebibytes(current),
+                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                );
+
+                info!(status, "discontinuing downscale");
+
+                return Ok((false, status));
+            }
+        }
+
+        // The downscaling has been approved. Downscale the file cache, then the cgroup.
+        let mut status = vec![];
+        let mut file_cache_mem_usage = 0;
+        if let Some(file_cache) = &mut self.filecache {
+            let actual_usage = file_cache
+                .set_file_cache_size(expected_file_cache_mem_usage)
+                .await
+                .context("failed to set file cache size")?;
+            if file_cache.config.in_memory {
+                file_cache_mem_usage = actual_usage;
+            }
+            let message = format!(
+                "set file cache size to {} MiB (in memory = {})",
+                bytes_to_mebibytes(actual_usage),
+                file_cache.config.in_memory,
+            );
+            info!("downscale: {message}");
+            status.push(message);
+        }
+
+        if let Some(cgroup) = &self.cgroup {
+            let available_memory = usable_system_memory - file_cache_mem_usage;
+
+            if file_cache_mem_usage != expected_file_cache_mem_usage {
+                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+            }
+
+            let limits = MemoryLimits::new(
+                // new_cgroup_mem_high is initialized to 0 but it is guarancontextd to not be here
+                // since it is properly initialized in the previous cgroup if let block
+                new_cgroup_mem_high,
+                available_memory,
+            );
+            cgroup
+                .set_limits(&limits)
+                .context("failed to set cgroup memory limits")?;
+
+            let message = format!(
+                "set cgroup memory.high to {} MiB, of new max {} MiB",
+                bytes_to_mebibytes(new_cgroup_mem_high),
+                bytes_to_mebibytes(available_memory)
+            );
+            info!("downscale: {message}");
+            status.push(message);
+        }
+
+        // TODO: make this status thing less jank
+        let status = status.join("; ");
+        Ok((true, status))
+    }
+
+    /// Handle new resources
+    #[tracing::instrument(skip_all, fields(?resources))]
+    pub async fn handle_upscale(&mut self, resources: Resources) -> anyhow::Result<()> {
+        if self.filecache.is_none() && self.cgroup.is_none() {
+            info!("no action needed for upscale (no cgroup or file cache enabled)");
+            return Ok(());
+        }
+
+        let new_mem = resources.mem;
+        let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
+
+        // Get the file cache's expected contribution to the memory usage
+        let mut file_cache_mem_usage = 0;
+        if let Some(file_cache) = &mut self.filecache {
+            let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
+            info!(
+                target = bytes_to_mebibytes(expected_usage),
+                total = bytes_to_mebibytes(new_mem),
+                "updating file cache size",
+            );
+
+            let actual_usage = file_cache
+                .set_file_cache_size(expected_usage)
+                .await
+                .context("failed to set file cache size")?;
+            if file_cache.config.in_memory {
+                file_cache_mem_usage = actual_usage;
+            }
+
+            if actual_usage != expected_usage {
+                warn!(
+                    "file cache was set to a different size that we wanted: target = {} Mib, actual= {} Mib",
+                    bytes_to_mebibytes(expected_usage),
+                    bytes_to_mebibytes(actual_usage)
+                )
+            }
+        }
+
+        if let Some(cgroup) = &self.cgroup {
+            let available_memory = usable_system_memory - file_cache_mem_usage;
+            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+            info!(
+                target = bytes_to_mebibytes(new_cgroup_mem_high),
+                total = bytes_to_mebibytes(new_mem),
+                name = cgroup.path(),
+                "updating cgroup memory.high",
+            );
+            let limits = MemoryLimits::new(new_cgroup_mem_high, available_memory);
+            cgroup
+                .set_limits(&limits)
+                .context("failed to set file cache size")?;
+        }
+
+        Ok(())
+    }
+
+    /// Take in a message and perform some action, such as downscaling or upscaling,
+    /// and return a message to be send back.
+    #[tracing::instrument(skip_all, fields(%id, message = ?inner))]
+    pub async fn process_message(
+        &mut self,
+        InboundMsg { inner, id }: InboundMsg,
+    ) -> anyhow::Result<Option<OutboundMsg>> {
+        match inner {
+            InboundMsgKind::UpscaleNotification { granted } => {
+                self.handle_upscale(granted)
+                    .await
+                    .context("failed to handle upscale")?;
+                self.dispatcher
+                    .notify_upscale(Sequenced::new(granted))
+                    .await
+                    .context("failed to notify notify cgroup of upscale")?;
+                Ok(Some(OutboundMsg::new(
+                    OutboundMsgKind::UpscaleConfirmation {},
+                    id,
+                )))
+            }
+            InboundMsgKind::DownscaleRequest { target } => self
+                .try_downscale(target)
+                .await
+                .context("failed to downscale")
+                .map(|(ok, status)| {
+                    Some(OutboundMsg::new(
+                        OutboundMsgKind::DownscaleResult { ok, status },
+                        id,
+                    ))
+                }),
+            InboundMsgKind::InvalidMessage { error } => {
+                warn!(
+                    %error, id, "received notification of an invalid message we sent"
+                );
+                Ok(None)
+            }
+            InboundMsgKind::InternalError { error } => {
+                warn!(error, id, "agent experienced an internal error");
+                Ok(None)
+            }
+            InboundMsgKind::HealthCheck {} => {
+                Ok(Some(OutboundMsg::new(OutboundMsgKind::HealthCheck {}, id)))
+            }
+        }
+    }
+
+    // TODO: don't propagate errors, probably just warn!?
+    #[tracing::instrument(skip_all)]
+    pub async fn run(&mut self) -> anyhow::Result<()> {
+        info!("starting dispatcher");
+        loop {
+            tokio::select! {
+                signal = self.kill.recv() => {
+                    match signal {
+                        Ok(()) => return Ok(()),
+                        Err(e) => bail!("failed to receive kill signal: {e}")
+                    }
+                }
+                // we need to propagate an upscale request
+                request = self.dispatcher.request_upscale_events.recv() => {
+                    if request.is_none() {
+                        bail!("failed to listen for upscale event from cgroup")
+                    }
+
+                    // If it's been less than 1 second since the last time we requested upscaling,
+                    // ignore the event, to avoid spamming the agent (otherwise, this can happen
+                    // ~1k times per second).
+                    if let Some(t) = self.last_upscale_request_at {
+                        let elapsed = t.elapsed();
+                        if elapsed < Duration::from_secs(1) {
+                            info!(elapsed_millis = elapsed.as_millis(), "cgroup asked for upscale but too soon to forward the request, ignoring");
+                            continue;
+                        }
+                    }
+
+                    self.last_upscale_request_at = Some(Instant::now());
+
+                    info!("cgroup asking for upscale; forwarding request");
+                    self.counter += 2; // Increment, preserving parity (i.e. keep the
+                                       // counter odd). See the field comment for more.
+                    self.dispatcher
+                        .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
+                        .await
+                        .context("failed to send message")?;
+                }
+                // there is a message from the agent
+                msg = self.dispatcher.source.next() => {
+                    if let Some(msg) = msg {
+                        // Don't use 'message' as a key as the string also uses
+                        // that for its key
+                        info!(?msg, "received message");
+                        match msg {
+                            Ok(msg) => {
+                                let message: InboundMsg = match msg {
+                                    Message::Text(text) => {
+                                        serde_json::from_str(&text).context("failed to deserialize text message")?
+                                    }
+                                    other => {
+                                        warn!(
+                                            // Don't use 'message' as a key as the
+                                            // string also uses that for its key
+                                            msg = ?other,
+                                            "agent should only send text messages but received different type"
+                                        );
+                                        continue
+                                    },
+                                };
+
+                                let out = match self.process_message(message.clone()).await {
+                                    Ok(Some(out)) => out,
+                                    Ok(None) => continue,
+                                    Err(e) => {
+                                        let error = e.to_string();
+                                        warn!(?error, "error handling message");
+                                        OutboundMsg::new(
+                                            OutboundMsgKind::InternalError {
+                                                error
+                                            },
+                                            message.id
+                                        )
+                                    }
+                                };
+
+                                self.dispatcher
+                                    .send(out)
+                                    .await
+                                    .context("failed to send message")?;
+                            }
+                            Err(e) => warn!("{e}"),
+                        }
+                    } else {
+                        anyhow::bail!("dispatcher connection closed")
+                    }
+                }
+            }
+        }
+    }
+}
--- a/pageserver/Cargo.toml
+++ b/pageserver/Cargo.toml
@@ -51,6 +51,7 @@ serde.workspace = true
 serde_json = { workspace = true, features = ["raw_value"] }
 serde_with.workspace = true
 signal-hook.workspace = true
+smallvec = { workspace = true, features = ["write"] }
 svg_fmt.workspace = true
 sync_wrapper.workspace = true
 tokio-tar.workspace = true
--- a/pageserver/benches/bench_layer_map.rs
+++ b/pageserver/benches/bench_layer_map.rs
@@ -215,7 +215,6 @@ fn bench_sequential(c: &mut Criterion) {
            TimelineId::generate(),
            zero.add(10 * i32)..zero.add(10 * i32 + 1),
            Lsn(i),
-            false,
            0,
        );
        updates.insert_historic(layer);
--- a/pageserver/ctl/src/layer_map_analyzer.rs
+++ b/pageserver/ctl/src/layer_map_analyzer.rs
@@ -3,6 +3,7 @@
 //! Currently it only analyzes holes, which are regions within the layer range that the layer contains no updates for. In the future it might do more analysis (maybe key quantiles?) but it should never return sensitive data.

 use anyhow::Result;
+use pageserver::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use std::cmp::Ordering;
 use std::collections::BinaryHeap;
 use std::ops::Range;
@@ -10,7 +11,7 @@ use std::{fs, path::Path, str};

 use pageserver::page_cache::PAGE_SZ;
 use pageserver::repository::{Key, KEY_SIZE};
-use pageserver::tenant::block_io::{BlockReader, FileBlockReader};
+use pageserver::tenant::block_io::FileBlockReader;
 use pageserver::tenant::disk_btree::{DiskBtreeReader, VisitDirection};
 use pageserver::tenant::storage_layer::delta_layer::{Summary, DELTA_KEY_SIZE};
 use pageserver::tenant::storage_layer::range_overlaps;
@@ -96,8 +97,8 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {

 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
 async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
-    let file = FileBlockReader::new(VirtualFile::open(path)?);
-    let summary_blk = file.read_blk(0)?;
+    let file = FileBlockReader::new(VirtualFile::open(path).await?);
+    let summary_blk = file.read_blk(0).await?;
    let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
    let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
        actual_summary.index_start_blk,
@@ -142,12 +143,12 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
    let mut total_delta_layers = 0usize;
    let mut total_image_layers = 0usize;
    let mut total_excess_layers = 0usize;
-    for tenant in fs::read_dir(storage_path.join("tenants"))? {
+    for tenant in fs::read_dir(storage_path.join(TENANTS_SEGMENT_NAME))? {
        let tenant = tenant?;
        if !tenant.file_type()?.is_dir() {
            continue;
        }
-        for timeline in fs::read_dir(tenant.path().join("timelines"))? {
+        for timeline in fs::read_dir(tenant.path().join(TIMELINES_SEGMENT_NAME))? {
            let timeline = timeline?;
            if !timeline.file_type()?.is_dir() {
                continue;
--- a/pageserver/ctl/src/layers.rs
+++ b/pageserver/ctl/src/layers.rs
@@ -5,6 +5,7 @@ use clap::Subcommand;
 use pageserver::tenant::block_io::BlockCursor;
 use pageserver::tenant::disk_btree::DiskBtreeReader;
 use pageserver::tenant::storage_layer::delta_layer::{BlobRef, Summary};
+use pageserver::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use pageserver::{page_cache, virtual_file};
 use pageserver::{
    repository::{Key, KEY_SIZE},
@@ -44,13 +45,11 @@ pub(crate) enum LayerCmd {
 }

 async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
-    use pageserver::tenant::block_io::BlockReader;
-
    let path = path.as_ref();
    virtual_file::init(10);
    page_cache::init(100);
-    let file = FileBlockReader::new(VirtualFile::open(path)?);
-    let summary_blk = file.read_blk(0)?;
+    let file = FileBlockReader::new(VirtualFile::open(path).await?);
+    let summary_blk = file.read_blk(0).await?;
    let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
    let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
        actual_summary.index_start_blk,
@@ -70,7 +69,7 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
            },
        )
        .await?;
-    let cursor = BlockCursor::new(&file);
+    let cursor = BlockCursor::new_fileblockreader(&file);
    for (k, v) in all {
        let value = cursor.read_blob(v.pos()).await?;
        println!("key:{} value_len:{}", k, value.len());
@@ -82,13 +81,13 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
 pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
    match cmd {
        LayerCmd::List { path } => {
-            for tenant in fs::read_dir(path.join("tenants"))? {
+            for tenant in fs::read_dir(path.join(TENANTS_SEGMENT_NAME))? {
                let tenant = tenant?;
                if !tenant.file_type()?.is_dir() {
                    continue;
                }
                println!("tenant {}", tenant.file_name().to_string_lossy());
-                for timeline in fs::read_dir(tenant.path().join("timelines"))? {
+                for timeline in fs::read_dir(tenant.path().join(TIMELINES_SEGMENT_NAME))? {
                    let timeline = timeline?;
                    if !timeline.file_type()?.is_dir() {
                        continue;
@@ -103,9 +102,9 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
            timeline,
        } => {
            let timeline_path = path
-                .join("tenants")
+                .join(TENANTS_SEGMENT_NAME)
                .join(tenant)
-                .join("timelines")
+                .join(TIMELINES_SEGMENT_NAME)
                .join(timeline);
            let mut idx = 0;
            for layer in fs::read_dir(timeline_path)? {
--- a/pageserver/src/basebackup.rs
+++ b/pageserver/src/basebackup.rs
@@ -25,6 +25,7 @@ use crate::context::RequestContext;
 use crate::tenant::Timeline;
 use pageserver_api::reltag::{RelTag, SlruKind};

+use postgres_ffi::dispatch_pgversion;
 use postgres_ffi::pg_constants::{DEFAULTTABLESPACE_OID, GLOBALTABLESPACE_OID};
 use postgres_ffi::pg_constants::{PGDATA_SPECIAL_FILES, PGDATA_SUBDIRS, PG_HBA};
 use postgres_ffi::relfile_utils::{INIT_FORKNUM, MAIN_FORKNUM};
@@ -323,14 +324,25 @@ where
                .timeline
                .get_relmap_file(spcnode, dbnode, self.lsn, self.ctx)
                .await?;
-            ensure!(img.len() == 512);
+
+            ensure!(
+                img.len()
+                    == dispatch_pgversion!(
+                        self.timeline.pg_version,
+                        pgv::bindings::SIZEOF_RELMAPFILE
+                    )
+            );
+
            Some(img)
        } else {
            None
        };

        if spcnode == GLOBALTABLESPACE_OID {
-            let pg_version_str = self.timeline.pg_version.to_string();
+            let pg_version_str = match self.timeline.pg_version {
+                14 | 15 => self.timeline.pg_version.to_string(),
+                ver => format!("{ver}\x0A"),
+            };
            let header = new_tar_header("PG_VERSION", pg_version_str.len() as u64)?;
            self.ar.append(&header, pg_version_str.as_bytes()).await?;

@@ -374,7 +386,10 @@ where
            if let Some(img) = relmap_img {
                let dst_path = format!("base/{}/PG_VERSION", dbnode);

-                let pg_version_str = self.timeline.pg_version.to_string();
+                let pg_version_str = match self.timeline.pg_version {
+                    14 | 15 => self.timeline.pg_version.to_string(),
+                    ver => format!("{ver}\x0A"),
+                };
                let header = new_tar_header(&dst_path, pg_version_str.len() as u64)?;
                self.ar.append(&header, pg_version_str.as_bytes()).await?;

--- a/pageserver/src/bin/pageserver.rs
+++ b/pageserver/src/bin/pageserver.rs
@@ -6,11 +6,12 @@ use std::{env, ops::ControlFlow, path::Path, str::FromStr};

 use anyhow::{anyhow, Context};
 use clap::{Arg, ArgAction, Command};
-use fail::FailScenario;
+
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::WALRECEIVER_RUNTIME;
+use pageserver::tenant::TenantSharedResources;
 use remote_storage::GenericRemoteStorage;
 use tokio::time::Instant;
 use tracing::*;
@@ -121,7 +122,7 @@ fn main() -> anyhow::Result<()> {
    }

    // Initialize up failpoints support
-    let scenario = FailScenario::setup();
+    let scenario = pageserver::failpoint_support::init();

    // Basic initialization of things that don't change after startup
    virtual_file::init(conf.max_file_descriptors);
@@ -382,9 +383,12 @@ fn start_pageserver(

    BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
        conf,
-        broker_client.clone(),
-        remote_storage.clone(),
+        TenantSharedResources {
+            broker_client: broker_client.clone(),
+            remote_storage: remote_storage.clone(),
+        },
        order,
+        shutdown_pageserver.clone(),
    ))?;

    BACKGROUND_RUNTIME.spawn({
@@ -473,16 +477,19 @@ fn start_pageserver(
    {
        let _rt_guard = MGMT_REQUEST_RUNTIME.enter();

-        let router = http::make_router(
-            conf,
-            launch_ts,
-            http_auth,
-            broker_client.clone(),
-            remote_storage,
-            disk_usage_eviction_state,
-        )?
-        .build()
-        .map_err(|err| anyhow!(err))?;
+        let router_state = Arc::new(
+            http::routes::State::new(
+                conf,
+                http_auth.clone(),
+                remote_storage,
+                broker_client.clone(),
+                disk_usage_eviction_state,
+            )
+            .context("Failed to initialize router state")?,
+        );
+        let router = http::make_router(router_state, launch_ts, http_auth.clone())?
+            .build()
+            .map_err(|err| anyhow!(err))?;
        let service = utils::http::RouterService::new(router).unwrap();
        let server = hyper::Server::from_tcp(http_listener)?
            .serve(service)
--- a/pageserver/src/config.rs
+++ b/pageserver/src/config.rs
@@ -32,7 +32,8 @@ use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{
-    TENANT_ATTACHING_MARKER_FILENAME, TENANT_DELETED_MARKER_FILE_NAME, TIMELINES_SEGMENT_NAME,
+    TENANTS_SEGMENT_NAME, TENANT_ATTACHING_MARKER_FILENAME, TENANT_DELETED_MARKER_FILE_NAME,
+    TIMELINES_SEGMENT_NAME,
 };
 use crate::{
    IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TIMELINE_DELETE_MARK_SUFFIX,
@@ -72,7 +73,7 @@ pub mod defaults {
    /// Default built-in configuration file.
    ///
    pub const DEFAULT_CONFIG_FILE: &str = formatcp!(
-        r###"
+        r#"
 # Initial configuration file created by 'pageserver --init'
 #listen_pg_addr = '{DEFAULT_PG_LISTEN_ADDR}'
 #listen_http_addr = '{DEFAULT_HTTP_LISTEN_ADDR}'
@@ -117,7 +118,7 @@ pub mod defaults {

 [remote_storage]

-"###
+"#
    );
 }

@@ -204,6 +205,8 @@ pub struct PageServerConf {
    /// has it's initial logical size calculated. Not running background tasks for some seconds is
    /// not terrible.
    pub background_task_maximum_delay: Duration,
+
+    pub control_plane_api: Option<Url>,
 }

 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -278,6 +281,8 @@ struct PageServerConfigBuilder {
    ondemand_download_behavior_treat_error_as_warn: BuilderValue<bool>,

    background_task_maximum_delay: BuilderValue<Duration>,
+
+    control_plane_api: BuilderValue<Option<Url>>,
 }

 impl Default for PageServerConfigBuilder {
@@ -340,6 +345,8 @@ impl Default for PageServerConfigBuilder {
                DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY,
            )
            .unwrap()),
+
+            control_plane_api: Set(None),
        }
    }
 }
@@ -468,6 +475,10 @@ impl PageServerConfigBuilder {
        self.background_task_maximum_delay = BuilderValue::Set(delay);
    }

+    pub fn control_plane_api(&mut self, api: Url) {
+        self.control_plane_api = BuilderValue::Set(Some(api))
+    }
+
    pub fn build(self) -> anyhow::Result<PageServerConf> {
        let concurrent_tenant_size_logical_size_queries = self
            .concurrent_tenant_size_logical_size_queries
@@ -553,6 +564,9 @@ impl PageServerConfigBuilder {
            background_task_maximum_delay: self
                .background_task_maximum_delay
                .ok_or(anyhow!("missing background_task_maximum_delay"))?,
+            control_plane_api: self
+                .control_plane_api
+                .ok_or(anyhow!("missing control_plane_api"))?,
        })
    }
 }
@@ -563,7 +577,7 @@ impl PageServerConf {
    //

    pub fn tenants_path(&self) -> PathBuf {
-        self.workdir.join("tenants")
+        self.workdir.join(TENANTS_SEGMENT_NAME)
    }

    pub fn tenant_path(&self, tenant_id: &TenantId) -> PathBuf {
@@ -643,23 +657,6 @@ impl PageServerConf {
            .join(METADATA_FILE_NAME)
    }

-    /// Files on the remote storage are stored with paths, relative to the workdir.
-    /// That path includes in itself both tenant and timeline ids, allowing to have a unique remote storage path.
-    ///
-    /// Errors if the path provided does not start from pageserver's workdir.
-    pub fn remote_path(&self, local_path: &Path) -> anyhow::Result<RemotePath> {
-        local_path
-            .strip_prefix(&self.workdir)
-            .context("Failed to strip workdir prefix")
-            .and_then(RemotePath::new)
-            .with_context(|| {
-                format!(
-                    "Failed to resolve remote part of path {:?} for base {:?}",
-                    local_path, self.workdir
-                )
-            })
-    }
-
    /// Turns storage remote path of a file into its local path.
    pub fn local_path(&self, remote_path: &RemotePath) -> PathBuf {
        remote_path.with_base(&self.workdir)
@@ -671,26 +668,18 @@ impl PageServerConf {
    pub fn pg_distrib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
        let path = self.pg_distrib_dir.clone();

+        #[allow(clippy::manual_range_patterns)]
        match pg_version {
-            14 => Ok(path.join(format!("v{pg_version}"))),
-            15 => Ok(path.join(format!("v{pg_version}"))),
+            14 | 15 | 16 => Ok(path.join(format!("v{pg_version}"))),
            _ => bail!("Unsupported postgres version: {}", pg_version),
        }
    }

    pub fn pg_bin_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
-        match pg_version {
-            14 => Ok(self.pg_distrib_dir(pg_version)?.join("bin")),
-            15 => Ok(self.pg_distrib_dir(pg_version)?.join("bin")),
-            _ => bail!("Unsupported postgres version: {}", pg_version),
-        }
+        Ok(self.pg_distrib_dir(pg_version)?.join("bin"))
    }
    pub fn pg_lib_dir(&self, pg_version: u32) -> anyhow::Result<PathBuf> {
-        match pg_version {
-            14 => Ok(self.pg_distrib_dir(pg_version)?.join("lib")),
-            15 => Ok(self.pg_distrib_dir(pg_version)?.join("lib")),
-            _ => bail!("Unsupported postgres version: {}", pg_version),
-        }
+        Ok(self.pg_distrib_dir(pg_version)?.join("lib"))
    }

    /// Parse a configuration file (pageserver.toml) into a PageServerConf struct,
@@ -758,6 +747,7 @@ impl PageServerConf {
                },
                "ondemand_download_behavior_treat_error_as_warn" => builder.ondemand_download_behavior_treat_error_as_warn(parse_toml_bool(key, item)?),
                "background_task_maximum_delay" => builder.background_task_maximum_delay(parse_toml_duration(key, item)?),
+                "control_plane_api" => builder.control_plane_api(parse_toml_string(key, item)?.parse().context("failed to parse control plane URL")?),
                _ => bail!("unrecognized pageserver option '{key}'"),
            }
        }
@@ -926,6 +916,7 @@ impl PageServerConf {
            test_remote_failures: 0,
            ondemand_download_behavior_treat_error_as_warn: false,
            background_task_maximum_delay: Duration::ZERO,
+            control_plane_api: None,
        }
    }
 }
@@ -1149,6 +1140,7 @@ background_task_maximum_delay = '334 s'
                background_task_maximum_delay: humantime::parse_duration(
                    defaults::DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY
                )?,
+                control_plane_api: None
            },
            "Correct defaults should be used when no config values are provided"
        );
@@ -1204,6 +1196,7 @@ background_task_maximum_delay = '334 s'
                test_remote_failures: 0,
                ondemand_download_behavior_treat_error_as_warn: false,
                background_task_maximum_delay: Duration::from_secs(334),
+                control_plane_api: None
            },
            "Should be able to parse all basic config values correctly"
        );
--- a/pageserver/src/control_plane_client.rs
+++ b/pageserver/src/control_plane_client.rs
@@ -0,0 +1,119 @@
+use std::collections::HashMap;
+
+use hyper::StatusCode;
+use pageserver_api::control_api::{ReAttachRequest, ReAttachResponse};
+use tokio_util::sync::CancellationToken;
+use url::Url;
+use utils::{
+    backoff,
+    generation::Generation,
+    id::{NodeId, TenantId},
+};
+
+use crate::config::PageServerConf;
+
+// Backoffs when control plane requests do not succeed: compromise between reducing load
+// on control plane, and retrying frequently when we are blocked on a control plane
+// response to make progress.
+const BACKOFF_INCREMENT: f64 = 0.1;
+const BACKOFF_MAX: f64 = 10.0;
+
+/// The Pageserver's client for using the control plane API: this is a small subset
+/// of the overall control plane API, for dealing with generations (see docs/rfcs/025-generation-numbers.md)
+pub(crate) struct ControlPlaneClient {
+    http_client: reqwest::Client,
+    base_url: Url,
+    node_id: NodeId,
+    cancel: CancellationToken,
+}
+
+impl ControlPlaneClient {
+    /// A None return value indicates that the input `conf` object does not have control
+    /// plane API enabled.
+    pub(crate) fn new(conf: &'static PageServerConf, cancel: &CancellationToken) -> Option<Self> {
+        let mut url = match conf.control_plane_api.as_ref() {
+            Some(u) => u.clone(),
+            None => return None,
+        };
+
+        if let Ok(mut segs) = url.path_segments_mut() {
+            // This ensures that `url` ends with a slash if it doesn't already.
+            // That way, we can subsequently use join() to safely attach extra path elements.
+            segs.pop_if_empty().push("");
+        }
+
+        let client = reqwest::ClientBuilder::new()
+            .build()
+            .expect("Failed to construct http client");
+
+        Some(Self {
+            http_client: client,
+            base_url: url,
+            node_id: conf.id,
+            cancel: cancel.clone(),
+        })
+    }
+
+    async fn try_re_attach(
+        &self,
+        url: Url,
+        request: &ReAttachRequest,
+    ) -> anyhow::Result<ReAttachResponse> {
+        match self.http_client.post(url).json(request).send().await {
+            Err(e) => Err(anyhow::Error::from(e)),
+            Ok(r) => {
+                if r.status() == StatusCode::OK {
+                    r.json::<ReAttachResponse>()
+                        .await
+                        .map_err(anyhow::Error::from)
+                } else {
+                    Err(anyhow::anyhow!("Unexpected status {}", r.status()))
+                }
+            }
+        }
+    }
+
+    /// Block until we get a successful response
+    pub(crate) async fn re_attach(&self) -> anyhow::Result<HashMap<TenantId, Generation>> {
+        let re_attach_path = self
+            .base_url
+            .join("re-attach")
+            .expect("Failed to build re-attach path");
+        let request = ReAttachRequest {
+            node_id: self.node_id,
+        };
+
+        let mut attempt = 0;
+        loop {
+            let result = self.try_re_attach(re_attach_path.clone(), &request).await;
+            match result {
+                Ok(res) => {
+                    tracing::info!(
+                        "Received re-attach response with {} tenants",
+                        res.tenants.len()
+                    );
+
+                    return Ok(res
+                        .tenants
+                        .into_iter()
+                        .map(|t| (t.id, Generation::new(t.generation)))
+                        .collect::<HashMap<_, _>>());
+                }
+                Err(e) => {
+                    tracing::error!("Error re-attaching tenants, retrying: {e:#}");
+                    backoff::exponential_backoff(
+                        attempt,
+                        BACKOFF_INCREMENT,
+                        BACKOFF_MAX,
+                        &self.cancel,
+                    )
+                    .await;
+                    if self.cancel.is_cancelled() {
+                        return Err(anyhow::anyhow!("Shutting down"));
+                    }
+                    attempt += 1;
+                }
+            }
+        }
+    }
+}
--- a/pageserver/src/failpoint_support.rs
+++ b/pageserver/src/failpoint_support.rs
@@ -0,0 +1,86 @@
+/// use with fail::cfg("$name", "return(2000)")
+///
+/// The effect is similar to a "sleep(2000)" action, i.e. we sleep for the
+/// specified time (in milliseconds). The main difference is that we use async
+/// tokio sleep function. Another difference is that we print lines to the log,
+/// which can be useful in tests to check that the failpoint was hit.
+#[macro_export]
+macro_rules! __failpoint_sleep_millis_async {
+    ($name:literal) => {{
+        // If the failpoint is used with a "return" action, set should_sleep to the
+        // returned value (as string). Otherwise it's set to None.
+        let should_sleep = (|| {
+            ::fail::fail_point!($name, |x| x);
+            ::std::option::Option::None
+        })();
+
+        // Sleep if the action was a returned value
+        if let ::std::option::Option::Some(duration_str) = should_sleep {
+            $crate::failpoint_support::failpoint_sleep_helper($name, duration_str).await
+        }
+    }};
+}
+pub use __failpoint_sleep_millis_async as sleep_millis_async;
+
+// Helper function used by the macro. (A function has nicer scoping so we
+// don't need to decorate everything with "::")
+#[doc(hidden)]
+pub(crate) async fn failpoint_sleep_helper(name: &'static str, duration_str: String) {
+    let millis = duration_str.parse::<u64>().unwrap();
+    let d = std::time::Duration::from_millis(millis);
+
+    tracing::info!("failpoint {:?}: sleeping for {:?}", name, d);
+    tokio::time::sleep(d).await;
+    tracing::info!("failpoint {:?}: sleep done", name);
+}
+
+pub fn init() -> fail::FailScenario<'static> {
+    // The failpoints lib provides support for parsing the `FAILPOINTS` env var.
+    // We want non-default behavior for `exit`, though, so, we handle it separately.
+    //
+    // Format for FAILPOINTS is "name=actions" separated by ";".
+    let actions = std::env::var("FAILPOINTS");
+    if actions.is_ok() {
+        std::env::remove_var("FAILPOINTS");
+    } else {
+        // let the library handle non-utf8, or nothing for not present
+    }
+
+    let scenario = fail::FailScenario::setup();
+
+    if let Ok(val) = actions {
+        val.split(';')
+            .enumerate()
+            .map(|(i, s)| s.split_once('=').ok_or((i, s)))
+            .for_each(|res| {
+                let (name, actions) = match res {
+                    Ok(t) => t,
+                    Err((i, s)) => {
+                        panic!(
+                            "startup failpoints: missing action on the {}th failpoint; try `{s}=return`",
+                            i + 1,
+                        );
+                    }
+                };
+                if let Err(e) = apply_failpoint(name, actions) {
+                    panic!("startup failpoints: failed to apply failpoint {name}={actions}: {e}");
+                }
+            });
+    }
+
+    scenario
+}
+
+pub(crate) fn apply_failpoint(name: &str, actions: &str) -> Result<(), String> {
+    if actions == "exit" {
+        fail::cfg_callback(name, exit_failpoint)
+    } else {
+        fail::cfg(name, actions)
+    }
+}
+
+#[inline(never)]
+fn exit_failpoint() {
+    tracing::info!("Exit requested by failpoint");
+    std::process::exit(1);
+}
--- a/pageserver/src/http/openapi_spec.yml
+++ b/pageserver/src/http/openapi_spec.yml
@@ -383,7 +383,6 @@ paths:
        schema:
          type: string
          format: hex
-
    post:
      description: |
        Schedules attach operation to happen in the background for the given tenant.
@@ -1020,6 +1019,9 @@ components:
      properties:
        config:
          $ref: '#/components/schemas/TenantConfig'
+        generation:
+          type: integer
+          description: Attachment generation number.
    TenantConfigRequest:
      allOf:
        - $ref: '#/components/schemas/TenantConfig'
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -8,9 +8,10 @@ use anyhow::{anyhow, Context, Result};
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
-use pageserver_api::models::{DownloadRemoteLayersTaskSpawnRequest, TenantAttachRequest};
+use pageserver_api::models::{
+    DownloadRemoteLayersTaskSpawnRequest, TenantAttachRequest, TenantLoadRequest,
+};
 use remote_storage::GenericRemoteStorage;
-use storage_broker::BrokerClientChannel;
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -32,11 +33,13 @@ use crate::tenant::mgr::{
 };
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
-use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
+use crate::tenant::timeline::Timeline;
+use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
 use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
 use utils::{
    auth::JwtAuth,
+    generation::Generation,
    http::{
        endpoint::{self, attach_openapi_ui, auth_middleware, check_permission_with},
        error::{ApiError, HttpErrorBody},
@@ -51,7 +54,7 @@ use utils::{
 // Imports only used for testing APIs
 use super::models::ConfigureFailpointsRequest;

-struct State {
+pub struct State {
    conf: &'static PageServerConf,
    auth: Option<Arc<JwtAuth>>,
    allowlist_routes: Vec<Uri>,
@@ -61,7 +64,7 @@ struct State {
 }

 impl State {
-    fn new(
+    pub fn new(
        conf: &'static PageServerConf,
        auth: Option<Arc<JwtAuth>>,
        remote_storage: Option<GenericRemoteStorage>,
@@ -282,6 +285,8 @@ async fn build_timeline_info_common(
    let state = timeline.current_state();
    let remote_consistent_lsn = timeline.get_remote_consistent_lsn().unwrap_or(Lsn(0));

+    let walreceiver_status = timeline.walreceiver_status();
+
    let info = TimelineInfo {
        tenant_id: timeline.tenant_id,
        timeline_id: timeline.timeline_id,
@@ -302,6 +307,8 @@ async fn build_timeline_info_common(
        pg_version: timeline.pg_version,

        state,
+
+        walreceiver_status,
    };
    Ok(info)
 }
@@ -472,7 +479,7 @@ async fn tenant_attach_handler(
    check_permission(&request, Some(tenant_id))?;

    let maybe_body: Option<TenantAttachRequest> = json_request_or_empty_body(&mut request).await?;
-    let tenant_conf = match maybe_body {
+    let tenant_conf = match &maybe_body {
        Some(request) => TenantConfOpt::try_from(&*request.config).map_err(ApiError::BadRequest)?,
        None => TenantConfOpt::default(),
    };
@@ -483,10 +490,13 @@ async fn tenant_attach_handler(

    let state = get_state(&request);

+    let generation = get_request_generation(state, maybe_body.as_ref().and_then(|r| r.generation))?;
+
    if let Some(remote_storage) = &state.remote_storage {
        mgr::attach_tenant(
            state.conf,
            tenant_id,
+            generation,
            tenant_conf,
            state.broker_client.clone(),
            remote_storage.clone(),
@@ -517,7 +527,6 @@ async fn timeline_delete_handler(
        .instrument(info_span!("timeline_delete", %tenant_id, %timeline_id))
        .await?;

-    // FIXME: needs to be an error for console to retry it. Ideally Accepted should be used and retried until 404.
    json_response(StatusCode::ACCEPTED, ())
 }

@@ -539,7 +548,7 @@ async fn tenant_detach_handler(
 }

 async fn tenant_load_handler(
-    request: Request<Body>,
+    mut request: Request<Body>,
    _cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
@@ -547,10 +556,18 @@ async fn tenant_load_handler(

    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);

+    let maybe_body: Option<TenantLoadRequest> = json_request_or_empty_body(&mut request).await?;
+
    let state = get_state(&request);
+
+    // The /load request is only usable when control_plane_api is not set.  Once it is set, callers
+    // should always use /attach instead.
+    let generation = get_request_generation(state, maybe_body.as_ref().and_then(|r| r.generation))?;
+
    mgr::load_tenant(
        state.conf,
        tenant_id,
+        generation,
        state.broker_client.clone(),
        state.remote_storage.clone(),
        &ctx,
@@ -852,6 +869,21 @@ pub fn html_response(status: StatusCode, data: String) -> Result<Response<Body>,
    Ok(response)
 }

+/// Helper for requests that may take a generation, which is mandatory
+/// when control_plane_api is set, but otherwise defaults to Generation::none()
+fn get_request_generation(state: &State, req_gen: Option<u32>) -> Result<Generation, ApiError> {
+    if state.conf.control_plane_api.is_some() {
+        req_gen
+            .map(Generation::new)
+            .ok_or(ApiError::BadRequest(anyhow!(
+                "generation attribute missing"
+            )))
+    } else {
+        // Legacy mode: all tenants operate with no generation
+        Ok(Generation::none())
+    }
+}
+
 async fn tenant_create_handler(
    mut request: Request<Body>,
    _cancel: CancellationToken,
@@ -868,14 +900,17 @@ async fn tenant_create_handler(
    let tenant_conf =
        TenantConfOpt::try_from(&request_data.config).map_err(ApiError::BadRequest)?;

-    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
-
    let state = get_state(&request);

+    let generation = get_request_generation(state, request_data.generation)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+
    let new_tenant = mgr::create_tenant(
        state.conf,
        tenant_conf,
        target_tenant_id,
+        generation,
        state.broker_client.clone(),
        state.remote_storage.clone(),
        &ctx,
@@ -980,14 +1015,7 @@ async fn failpoints_handler(

        // We recognize one extra "action" that's not natively recognized
        // by the failpoints crate: exit, to immediately kill the process
-        let cfg_result = if fp.actions == "exit" {
-            fail::cfg_callback(fp.name, || {
-                info!("Exit requested by failpoint");
-                std::process::exit(1);
-            })
-        } else {
-            fail::cfg(fp.name, &fp.actions)
-        };
+        let cfg_result = crate::failpoint_support::apply_failpoint(&fp.name, &fp.actions);

        if let Err(err_msg) = cfg_result {
            return Err(ApiError::BadRequest(anyhow!(
@@ -1329,12 +1357,9 @@ where
 }

 pub fn make_router(
-    conf: &'static PageServerConf,
+    state: Arc<State>,
    launch_ts: &'static LaunchTimestamp,
    auth: Option<Arc<JwtAuth>>,
-    broker_client: BrokerClientChannel,
-    remote_storage: Option<GenericRemoteStorage>,
-    disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 ) -> anyhow::Result<RouterBuilder<hyper::Body, ApiError>> {
    let spec = include_bytes!("openapi_spec.yml");
    let mut router = attach_openapi_ui(endpoint::make_router(), spec, "/swagger.yml", "/v1/doc");
@@ -1358,16 +1383,7 @@ pub fn make_router(
    );

    Ok(router
-        .data(Arc::new(
-            State::new(
-                conf,
-                auth,
-                remote_storage,
-                broker_client,
-                disk_usage_eviction_state,
-            )
-            .context("Failed to initialize router state")?,
-        ))
+        .data(state)
        .get("/v1/status", |r| api_handler(r, status_handler))
        .put("/v1/failpoints", |r| {
            testing_api_handler("manage failpoints", r, failpoints_handler)
--- a/pageserver/src/lib.rs
+++ b/pageserver/src/lib.rs
@@ -3,6 +3,7 @@ pub mod basebackup;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
+mod control_plane_client;
 pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;
@@ -21,6 +22,8 @@ pub mod walingest;
 pub mod walrecord;
 pub mod walredo;

+pub mod failpoint_support;
+
 use std::path::Path;

 use crate::task_mgr::TaskKind;
--- a/pageserver/src/metrics.rs
+++ b/pageserver/src/metrics.rs
@@ -6,7 +6,7 @@ use metrics::{
    HistogramVec, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
-use strum::VariantNames;
+use strum::{EnumCount, IntoEnumIterator, VariantNames};
 use strum_macros::{EnumVariantNames, IntoStaticStr};
 use utils::id::{TenantId, TimelineId};

@@ -537,7 +537,7 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
    30.000,   // 30000 ms
 ];

-/// Tracks time taken by fs operations near VirtualFile.
+/// VirtualFile fs operation variants.
 ///
 /// Operations:
 /// - open ([`std::fs::OpenOptions::open`])
@@ -548,15 +548,66 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
 /// - seek (modify internal position or file length query)
 /// - fsync ([`std::fs::File::sync_all`])
 /// - metadata ([`std::fs::File::metadata`])
-pub(crate) static STORAGE_IO_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
-        "pageserver_io_operations_seconds",
-        "Time spent in IO operations",
-        &["operation"],
-        STORAGE_IO_TIME_BUCKETS.into()
-    )
-    .expect("failed to define a metric")
-});
+#[derive(
+    Debug, Clone, Copy, strum_macros::EnumCount, strum_macros::EnumIter, strum_macros::FromRepr,
+)]
+pub(crate) enum StorageIoOperation {
+    Open,
+    Close,
+    CloseByReplace,
+    Read,
+    Write,
+    Seek,
+    Fsync,
+    Metadata,
+}
+
+impl StorageIoOperation {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            StorageIoOperation::Open => "open",
+            StorageIoOperation::Close => "close",
+            StorageIoOperation::CloseByReplace => "close-by-replace",
+            StorageIoOperation::Read => "read",
+            StorageIoOperation::Write => "write",
+            StorageIoOperation::Seek => "seek",
+            StorageIoOperation::Fsync => "fsync",
+            StorageIoOperation::Metadata => "metadata",
+        }
+    }
+}
+
+/// Tracks time taken by fs operations near VirtualFile.
+#[derive(Debug)]
+pub(crate) struct StorageIoTime {
+    metrics: [Histogram; StorageIoOperation::COUNT],
+}
+
+impl StorageIoTime {
+    fn new() -> Self {
+        let storage_io_histogram_vec = register_histogram_vec!(
+            "pageserver_io_operations_seconds",
+            "Time spent in IO operations",
+            &["operation"],
+            STORAGE_IO_TIME_BUCKETS.into()
+        )
+        .expect("failed to define a metric");
+        let metrics = std::array::from_fn(|i| {
+            let op = StorageIoOperation::from_repr(i).unwrap();
+            let metric = storage_io_histogram_vec
+                .get_metric_with_label_values(&[op.as_str()])
+                .unwrap();
+            metric
+        });
+        Self { metrics }
+    }
+
+    pub(crate) fn get(&self, op: StorageIoOperation) -> &Histogram {
+        &self.metrics[op as usize]
+    }
+}
+
+pub(crate) static STORAGE_IO_TIME_METRIC: Lazy<StorageIoTime> = Lazy::new(StorageIoTime::new);

 const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];

@@ -570,23 +621,160 @@ pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

-const SMGR_QUERY_TIME_OPERATIONS: &[&str] = &[
-    "get_rel_exists",
-    "get_rel_size",
-    "get_page_at_lsn",
-    "get_db_size",
-];
+#[derive(Debug)]
+struct GlobalAndPerTimelineHistogram {
+    global: Histogram,
+    per_tenant_timeline: Histogram,
+}

-pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+impl GlobalAndPerTimelineHistogram {
+    fn observe(&self, value: f64) {
+        self.global.observe(value);
+        self.per_tenant_timeline.observe(value);
+    }
+}
+
+struct GlobalAndPerTimelineHistogramTimer<'a> {
+    h: &'a GlobalAndPerTimelineHistogram,
+    start: std::time::Instant,
+}
+
+impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
+    fn drop(&mut self) {
+        let elapsed = self.start.elapsed();
+        self.h.observe(elapsed.as_secs_f64());
+    }
+}
+
+#[derive(
+    Debug,
+    Clone,
+    Copy,
+    IntoStaticStr,
+    strum_macros::EnumCount,
+    strum_macros::EnumIter,
+    strum_macros::FromRepr,
+)]
+#[strum(serialize_all = "snake_case")]
+pub enum SmgrQueryType {
+    GetRelExists,
+    GetRelSize,
+    GetPageAtLsn,
+    GetDbSize,
+}
+
+#[derive(Debug)]
+pub struct SmgrQueryTimePerTimeline {
+    metrics: [GlobalAndPerTimelineHistogram; SmgrQueryType::COUNT],
+}
+
+static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
    register_histogram_vec!(
        "pageserver_smgr_query_seconds",
-        "Time spent on smgr query handling",
+        "Time spent on smgr query handling, aggegated by query type and tenant/timeline.",
        &["smgr_query_type", "tenant_id", "timeline_id"],
        CRITICAL_OP_BUCKETS.into(),
    )
    .expect("failed to define a metric")
 });

+static SMGR_QUERY_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
+    register_histogram_vec!(
+        "pageserver_smgr_query_seconds_global",
+        "Time spent on smgr query handling, aggregated by query type.",
+        &["smgr_query_type"],
+        CRITICAL_OP_BUCKETS.into(),
+    )
+    .expect("failed to define a metric")
+});
+
+impl SmgrQueryTimePerTimeline {
+    pub(crate) fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
+        let tenant_id = tenant_id.to_string();
+        let timeline_id = timeline_id.to_string();
+        let metrics = std::array::from_fn(|i| {
+            let op = SmgrQueryType::from_repr(i).unwrap();
+            let global = SMGR_QUERY_TIME_GLOBAL
+                .get_metric_with_label_values(&[op.into()])
+                .unwrap();
+            let per_tenant_timeline = SMGR_QUERY_TIME_PER_TENANT_TIMELINE
+                .get_metric_with_label_values(&[op.into(), &tenant_id, &timeline_id])
+                .unwrap();
+            GlobalAndPerTimelineHistogram {
+                global,
+                per_tenant_timeline,
+            }
+        });
+        Self { metrics }
+    }
+    pub(crate) fn start_timer(&self, op: SmgrQueryType) -> impl Drop + '_ {
+        let metric = &self.metrics[op as usize];
+        GlobalAndPerTimelineHistogramTimer {
+            h: metric,
+            start: std::time::Instant::now(),
+        }
+    }
+}
+
+#[cfg(test)]
+mod smgr_query_time_tests {
+    use strum::IntoEnumIterator;
+    use utils::id::{TenantId, TimelineId};
+
+    // Regression test, we used hard-coded string constants before using an enum.
+    #[test]
+    fn op_label_name() {
+        use super::SmgrQueryType::*;
+        let expect: [(super::SmgrQueryType, &'static str); 4] = [
+            (GetRelExists, "get_rel_exists"),
+            (GetRelSize, "get_rel_size"),
+            (GetPageAtLsn, "get_page_at_lsn"),
+            (GetDbSize, "get_db_size"),
+        ];
+        for (op, expect) in expect {
+            let actual: &'static str = op.into();
+            assert_eq!(actual, expect);
+        }
+    }
+
+    #[test]
+    fn basic() {
+        let ops: Vec<_> = super::SmgrQueryType::iter().collect();
+
+        for op in &ops {
+            let tenant_id = TenantId::generate();
+            let timeline_id = TimelineId::generate();
+            let metrics = super::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);
+
+            let get_counts = || {
+                let global: u64 = ops
+                    .iter()
+                    .map(|op| metrics.metrics[*op as usize].global.get_sample_count())
+                    .sum();
+                let per_tenant_timeline: u64 = ops
+                    .iter()
+                    .map(|op| {
+                        metrics.metrics[*op as usize]
+                            .per_tenant_timeline
+                            .get_sample_count()
+                    })
+                    .sum();
+                (global, per_tenant_timeline)
+            };
+
+            let (pre_global, pre_per_tenant_timeline) = get_counts();
+            assert_eq!(pre_per_tenant_timeline, 0);
+
+            let timer = metrics.start_timer(*op);
+            drop(timer);
+
+            let (post_global, post_per_tenant_timeline) = get_counts();
+            assert_eq!(post_per_tenant_timeline, 1);
+            assert!(post_global > pre_global);
+        }
+    }
+}
+
 // keep in sync with control plane Go code so that we can validate
 // compute's basebackup_ms metric with our perspective in the context of SLI/SLO.
 static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
@@ -1028,6 +1216,12 @@ impl TimelineMetrics {
            ),
        }
    }
+
+    pub fn record_new_file_metrics(&self, sz: u64) {
+        self.resident_physical_size_gauge.add(sz);
+        self.num_persistent_files_created.inc_by(1);
+        self.persistent_bytes_written.inc_by(sz);
+    }
 }

 impl Drop for TimelineMetrics {
@@ -1045,6 +1239,12 @@ impl Drop for TimelineMetrics {
            .write()
            .unwrap()
            .remove(tenant_id, timeline_id);
+
+        // The following metrics are born outside of the TimelineMetrics lifecycle but still
+        // removed at the end of it. The idea is to have the metrics outlive the
+        // entity during which they're observed, e.g., the smgr metrics shall
+        // outlive an individual smgr connection, but not the timeline.
+
        for op in StorageTimeOperation::VARIANTS {
            let _ =
                STORAGE_TIME_SUM_PER_TIMELINE.remove_label_values(&[op, tenant_id, timeline_id]);
@@ -1056,8 +1256,12 @@ impl Drop for TimelineMetrics {
            let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, timeline_id]);
        }

-        for op in SMGR_QUERY_TIME_OPERATIONS {
-            let _ = SMGR_QUERY_TIME.remove_label_values(&[op, tenant_id, timeline_id]);
+        for op in SmgrQueryType::iter() {
+            let _ = SMGR_QUERY_TIME_PER_TENANT_TIMELINE.remove_label_values(&[
+                op.into(),
+                tenant_id,
+                timeline_id,
+            ]);
        }
    }
 }
--- a/pageserver/src/page_cache.rs
+++ b/pageserver/src/page_cache.rs
@@ -10,6 +10,42 @@
 //! PostgreSQL buffer size, and a Slot struct for each buffer to contain
 //! information about what's stored in the buffer.
 //!
+//! # Types Of Pages
+//!
+//! [`PageCache`] only supports immutable pages.
+//! Hence there is no need to worry about coherency.
+//!
+//! Two types of pages are supported:
+//!
+//! * **Materialized pages**, filled & used by page reconstruction
+//! * **Immutable File pages**, filled & used by [`crate::tenant::block_io`] and [`crate::tenant::ephemeral_file`].
+//!
+//! Note that [`crate::tenant::ephemeral_file::EphemeralFile`] is generally mutable, but, it's append-only.
+//! It uses the page cache only for the blocks that are already fully written and immutable.
+//!
+//! # Filling The Page Cache
+//!
+//! Page cache maps from a cache key to a buffer slot.
+//! The cache key uniquely identifies the piece of data that is being cached.
+//!
+//! The cache key for **materialized pages** is  [`TenantId`], [`TimelineId`], [`Key`], and [`Lsn`].
+//! Use [`PageCache::memorize_materialized_page`] and [`PageCache::lookup_materialized_page`] for fill & access.
+//!
+//! The cache key for **immutable file** pages is [`FileId`] and a block number.
+//! Users of page cache that wish to page-cache an arbitrary (immutable!) on-disk file do the following:
+//! * Have a mechanism to deterministically associate the on-disk file with a [`FileId`].
+//! * Get a [`FileId`] using [`next_file_id`].
+//! * Use the mechanism to associate the on-disk file with the returned [`FileId`].
+//! * Use [`PageCache::read_immutable_buf`] to get a [`ReadBufResult`].
+//! * If the page was already cached, it'll be the [`ReadBufResult::Found`] variant that contains
+//!   a read guard for the page. Just use it.
+//! * If the page was not cached, it'll be the [`ReadBufResult::NotFound`] variant that contains
+//!   a write guard for the page. Fill the page with the contents of the on-disk file.
+//!   Then call [`PageWriteGuard::mark_valid`] to mark the page as valid.
+//!   Then try again to [`PageCache::read_immutable_buf`].
+//!   Unless there's high cache pressure, the page should now be cached.
+//!   (TODO: allow downgrading the write guard to a read guard to ensure forward progress.)
+//!
 //! # Locking
 //!
 //! There are two levels of locking involved: There's one lock for the "mapping"
@@ -39,21 +75,16 @@
 use std::{
    collections::{hash_map::Entry, HashMap},
    convert::TryInto,
-    sync::{
-        atomic::{AtomicU8, AtomicUsize, Ordering},
-        RwLock, RwLockReadGuard, RwLockWriteGuard, TryLockError,
-    },
+    sync::atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
 };

 use anyhow::Context;
 use once_cell::sync::OnceCell;
-use tracing::error;
 use utils::{
    id::{TenantId, TimelineId},
    lsn::Lsn,
 };

-use crate::tenant::writeback_ephemeral_file;
 use crate::{metrics::PageCacheSizeMetrics, repository::Key};

 static PAGE_CACHE: OnceCell<PageCache> = OnceCell::new();
@@ -87,6 +118,17 @@ pub fn get() -> &'static PageCache {
 pub const PAGE_SZ: usize = postgres_ffi::BLCKSZ as usize;
 const MAX_USAGE_COUNT: u8 = 5;

+/// See module-level comment.
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
+pub struct FileId(u64);
+
+static NEXT_ID: AtomicU64 = AtomicU64::new(1);
+
+/// See module-level comment.
+pub fn next_file_id() -> FileId {
+    FileId(NEXT_ID.fetch_add(1, Ordering::Relaxed))
+}
+
 ///
 /// CacheKey uniquely identifies a "thing" to cache in the page cache.
 ///
@@ -97,12 +139,8 @@ enum CacheKey {
        hash_key: MaterializedPageHashKey,
        lsn: Lsn,
    },
-    EphemeralPage {
-        file_id: u64,
-        blkno: u32,
-    },
    ImmutableFilePage {
-        file_id: u64,
+        file_id: FileId,
        blkno: u32,
    },
 }
@@ -121,14 +159,13 @@ struct Version {
 }

 struct Slot {
-    inner: RwLock<SlotInner>,
+    inner: tokio::sync::RwLock<SlotInner>,
    usage_count: AtomicU8,
 }

 struct SlotInner {
    key: Option<CacheKey>,
    buf: &'static mut [u8; PAGE_SZ],
-    dirty: bool,
 }

 impl Slot {
@@ -163,6 +200,11 @@ impl Slot {
            Err(usage_count) => usage_count,
        }
    }
+
+    /// Sets the usage count to a specific value.
+    fn set_usage_count(&self, count: u8) {
+        self.usage_count.store(count, Ordering::Relaxed);
+    }
 }

 pub struct PageCache {
@@ -175,11 +217,9 @@ pub struct PageCache {
    ///
    /// If you add support for caching different kinds of objects, each object kind
    /// can have a separate mapping map, next to this field.
-    materialized_page_map: RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,
+    materialized_page_map: std::sync::RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,

-    ephemeral_page_map: RwLock<HashMap<(u64, u32), usize>>,
-
-    immutable_page_map: RwLock<HashMap<(u64, u32), usize>>,
+    immutable_page_map: std::sync::RwLock<HashMap<(FileId, u32), usize>>,

    /// The actual buffers with their metadata.
    slots: Box<[Slot]>,
@@ -195,7 +235,7 @@ pub struct PageCache {
 /// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
 /// until the guard is dropped.
 ///
-pub struct PageReadGuard<'i>(RwLockReadGuard<'i, SlotInner>);
+pub struct PageReadGuard<'i>(tokio::sync::RwLockReadGuard<'i, SlotInner>);

 impl std::ops::Deref for PageReadGuard<'_> {
    type Target = [u8; PAGE_SZ];
@@ -222,9 +262,10 @@ impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
 /// to initialize.
 ///
 pub struct PageWriteGuard<'i> {
-    inner: RwLockWriteGuard<'i, SlotInner>,
+    inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,

    // Are the page contents currently valid?
+    // Used to mark pages as invalid that are assigned but not yet filled with data.
    valid: bool,
 }

@@ -258,14 +299,6 @@ impl PageWriteGuard<'_> {
        );
        self.valid = true;
    }
-    pub fn mark_dirty(&mut self) {
-        // only ephemeral pages can be dirty ATM.
-        assert!(matches!(
-            self.inner.key,
-            Some(CacheKey::EphemeralPage { .. })
-        ));
-        self.inner.dirty = true;
-    }
 }

 impl Drop for PageWriteGuard<'_> {
@@ -280,7 +313,6 @@ impl Drop for PageWriteGuard<'_> {
            let self_key = self.inner.key.as_ref().unwrap();
            PAGE_CACHE.get().unwrap().remove_mapping(self_key);
            self.inner.key = None;
-            self.inner.dirty = false;
        }
    }
 }
@@ -308,7 +340,7 @@ impl PageCache {
    /// The 'lsn' is an upper bound, this will return the latest version of
    /// the given block, but not newer than 'lsn'. Returns the actual LSN of the
    /// returned page.
-    pub fn lookup_materialized_page(
+    pub async fn lookup_materialized_page(
        &self,
        tenant_id: TenantId,
        timeline_id: TimelineId,
@@ -328,7 +360,7 @@ impl PageCache {
            lsn,
        };

-        if let Some(guard) = self.try_lock_for_read(&mut cache_key) {
+        if let Some(guard) = self.try_lock_for_read(&mut cache_key).await {
            if let CacheKey::MaterializedPage {
                hash_key: _,
                lsn: available_lsn,
@@ -355,7 +387,7 @@ impl PageCache {
    ///
    /// Store an image of the given page in the cache.
    ///
-    pub fn memorize_materialized_page(
+    pub async fn memorize_materialized_page(
        &self,
        tenant_id: TenantId,
        timeline_id: TimelineId,
@@ -372,7 +404,7 @@ impl PageCache {
            lsn,
        };

-        match self.lock_for_write(&cache_key)? {
+        match self.lock_for_write(&cache_key).await? {
            WriteBufResult::Found(write_guard) => {
                // We already had it in cache. Another thread must've put it there
                // concurrently. Check that it had the same contents that we
@@ -388,68 +420,16 @@ impl PageCache {
        Ok(())
    }

-    // Section 1.2: Public interface functions for working with Ephemeral pages.
+    // Section 1.2: Public interface functions for working with immutable file pages.

-    pub fn read_ephemeral_buf(&self, file_id: u64, blkno: u32) -> anyhow::Result<ReadBufResult> {
-        let mut cache_key = CacheKey::EphemeralPage { file_id, blkno };
-
-        self.lock_for_read(&mut cache_key)
-    }
-
-    pub fn write_ephemeral_buf(&self, file_id: u64, blkno: u32) -> anyhow::Result<WriteBufResult> {
-        let cache_key = CacheKey::EphemeralPage { file_id, blkno };
-
-        self.lock_for_write(&cache_key)
-    }
-
-    /// Immediately drop all buffers belonging to given file, without writeback
-    pub fn drop_buffers_for_ephemeral(&self, drop_file_id: u64) {
-        for slot_idx in 0..self.slots.len() {
-            let slot = &self.slots[slot_idx];
-
-            let mut inner = slot.inner.write().unwrap();
-            if let Some(key) = &inner.key {
-                match key {
-                    CacheKey::EphemeralPage { file_id, blkno: _ } if *file_id == drop_file_id => {
-                        // remove mapping for old buffer
-                        self.remove_mapping(key);
-                        inner.key = None;
-                        inner.dirty = false;
-                    }
-                    _ => {}
-                }
-            }
-        }
-    }
-
-    // Section 1.3: Public interface functions for working with immutable file pages.
-
-    pub fn read_immutable_buf(&self, file_id: u64, blkno: u32) -> anyhow::Result<ReadBufResult> {
+    pub async fn read_immutable_buf(
+        &self,
+        file_id: FileId,
+        blkno: u32,
+    ) -> anyhow::Result<ReadBufResult> {
        let mut cache_key = CacheKey::ImmutableFilePage { file_id, blkno };

-        self.lock_for_read(&mut cache_key)
-    }
-
-    /// Immediately drop all buffers belonging to given file, without writeback
-    pub fn drop_buffers_for_immutable(&self, drop_file_id: u64) {
-        for slot_idx in 0..self.slots.len() {
-            let slot = &self.slots[slot_idx];
-
-            let mut inner = slot.inner.write().unwrap();
-            if let Some(key) = &inner.key {
-                match key {
-                    CacheKey::ImmutableFilePage { file_id, blkno: _ }
-                        if *file_id == drop_file_id =>
-                    {
-                        // remove mapping for old buffer
-                        self.remove_mapping(key);
-                        inner.key = None;
-                        inner.dirty = false;
-                    }
-                    _ => {}
-                }
-            }
-        }
+        self.lock_for_read(&mut cache_key).await
    }

    //
@@ -469,14 +449,14 @@ impl PageCache {
    ///
    /// If no page is found, returns None and *cache_key is left unmodified.
    ///
-    fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
+    async fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
        let cache_key_orig = cache_key.clone();
        if let Some(slot_idx) = self.search_mapping(cache_key) {
            // The page was found in the mapping. Lock the slot, and re-check
            // that it's still what we expected (because we released the mapping
            // lock already, another thread could have evicted the page)
            let slot = &self.slots[slot_idx];
-            let inner = slot.inner.read().unwrap();
+            let inner = slot.inner.read().await;
            if inner.key.as_ref() == Some(cache_key) {
                slot.inc_usage_count();
                return Some(PageReadGuard(inner));
@@ -517,15 +497,11 @@ impl PageCache {
    /// }
    /// ```
    ///
-    fn lock_for_read(&self, cache_key: &mut CacheKey) -> anyhow::Result<ReadBufResult> {
+    async fn lock_for_read(&self, cache_key: &mut CacheKey) -> anyhow::Result<ReadBufResult> {
        let (read_access, hit) = match cache_key {
            CacheKey::MaterializedPage { .. } => {
                unreachable!("Materialized pages use lookup_materialized_page")
            }
-            CacheKey::EphemeralPage { .. } => (
-                &crate::metrics::PAGE_CACHE.read_accesses_ephemeral,
-                &crate::metrics::PAGE_CACHE.read_hits_ephemeral,
-            ),
            CacheKey::ImmutableFilePage { .. } => (
                &crate::metrics::PAGE_CACHE.read_accesses_immutable,
                &crate::metrics::PAGE_CACHE.read_hits_immutable,
@@ -536,7 +512,7 @@ impl PageCache {
        let mut is_first_iteration = true;
        loop {
            // First check if the key already exists in the cache.
-            if let Some(read_guard) = self.try_lock_for_read(cache_key) {
+            if let Some(read_guard) = self.try_lock_for_read(cache_key).await {
                if is_first_iteration {
                    hit.inc();
                }
@@ -566,8 +542,7 @@ impl PageCache {
            // Make the slot ready
            let slot = &self.slots[slot_idx];
            inner.key = Some(cache_key.clone());
-            inner.dirty = false;
-            slot.usage_count.store(1, Ordering::Relaxed);
+            slot.set_usage_count(1);

            return Ok(ReadBufResult::NotFound(PageWriteGuard {
                inner,
@@ -580,13 +555,13 @@ impl PageCache {
    /// found, returns None.
    ///
    /// When locking a page for writing, the search criteria is always "exact".
-    fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
+    async fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
        if let Some(slot_idx) = self.search_mapping_for_write(cache_key) {
            // The page was found in the mapping. Lock the slot, and re-check
            // that it's still what we expected (because we don't released the mapping
            // lock already, another thread could have evicted the page)
            let slot = &self.slots[slot_idx];
-            let inner = slot.inner.write().unwrap();
+            let inner = slot.inner.write().await;
            if inner.key.as_ref() == Some(cache_key) {
                slot.inc_usage_count();
                return Some(PageWriteGuard { inner, valid: true });
@@ -599,10 +574,10 @@ impl PageCache {
    ///
    /// Similar to lock_for_read(), but the returned buffer is write-locked and
    /// may be modified by the caller even if it's already found in the cache.
-    fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
+    async fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
        loop {
            // First check if the key already exists in the cache.
-            if let Some(write_guard) = self.try_lock_for_write(cache_key) {
+            if let Some(write_guard) = self.try_lock_for_write(cache_key).await {
                return Ok(WriteBufResult::Found(write_guard));
            }

@@ -628,8 +603,7 @@ impl PageCache {
            // Make the slot ready
            let slot = &self.slots[slot_idx];
            inner.key = Some(cache_key.clone());
-            inner.dirty = false;
-            slot.usage_count.store(1, Ordering::Relaxed);
+            slot.set_usage_count(1);

            return Ok(WriteBufResult::NotFound(PageWriteGuard {
                inner,
@@ -667,10 +641,6 @@ impl PageCache {
                *lsn = version.lsn;
                Some(version.slot_idx)
            }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let map = self.ephemeral_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
-            }
            CacheKey::ImmutableFilePage { file_id, blkno } => {
                let map = self.immutable_page_map.read().unwrap();
                Some(*map.get(&(*file_id, *blkno))?)
@@ -694,10 +664,6 @@ impl PageCache {
                    None
                }
            }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let map = self.ephemeral_page_map.read().unwrap();
-                Some(*map.get(&(*file_id, *blkno))?)
-            }
            CacheKey::ImmutableFilePage { file_id, blkno } => {
                let map = self.immutable_page_map.read().unwrap();
                Some(*map.get(&(*file_id, *blkno))?)
@@ -731,12 +697,6 @@ impl PageCache {
                    panic!("could not find old key in mapping")
                }
            }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let mut map = self.ephemeral_page_map.write().unwrap();
-                map.remove(&(*file_id, *blkno))
-                    .expect("could not find old key in mapping");
-                self.size_metrics.current_bytes_ephemeral.sub_page_sz(1);
-            }
            CacheKey::ImmutableFilePage { file_id, blkno } => {
                let mut map = self.immutable_page_map.write().unwrap();
                map.remove(&(*file_id, *blkno))
@@ -776,17 +736,7 @@ impl PageCache {
                    }
                }
            }
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                let mut map = self.ephemeral_page_map.write().unwrap();
-                match map.entry((*file_id, *blkno)) {
-                    Entry::Occupied(entry) => Some(*entry.get()),
-                    Entry::Vacant(entry) => {
-                        entry.insert(slot_idx);
-                        self.size_metrics.current_bytes_ephemeral.add_page_sz(1);
-                        None
-                    }
-                }
-            }
+
            CacheKey::ImmutableFilePage { file_id, blkno } => {
                let mut map = self.immutable_page_map.write().unwrap();
                match map.entry((*file_id, *blkno)) {
@@ -808,7 +758,7 @@ impl PageCache {
    /// Find a slot to evict.
    ///
    /// On return, the slot is empty and write-locked.
-    fn find_victim(&self) -> anyhow::Result<(usize, RwLockWriteGuard<SlotInner>)> {
+    fn find_victim(&self) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
        let iter_limit = self.slots.len() * 10;
        let mut iters = 0;
        loop {
@@ -820,10 +770,7 @@ impl PageCache {
            if slot.dec_usage_count() == 0 {
                let mut inner = match slot.inner.try_write() {
                    Ok(inner) => inner,
-                    Err(TryLockError::Poisoned(err)) => {
-                        anyhow::bail!("buffer lock was poisoned: {err:?}")
-                    }
-                    Err(TryLockError::WouldBlock) => {
+                    Err(_err) => {
                        // If we have looped through the whole buffer pool 10 times
                        // and still haven't found a victim buffer, something's wrong.
                        // Maybe all the buffers were in locked. That could happen in
@@ -837,25 +784,8 @@ impl PageCache {
                    }
                };
                if let Some(old_key) = &inner.key {
-                    if inner.dirty {
-                        if let Err(err) = Self::writeback(old_key, inner.buf) {
-                            // Writing the page to disk failed.
-                            //
-                            // FIXME: What to do here, when? We could propagate the error to the
-                            // caller, but victim buffer is generally unrelated to the original
-                            // call. It can even belong to a different tenant. Currently, we
-                            // report the error to the log and continue the clock sweep to find
-                            // a different victim. But if the problem persists, the page cache
-                            // could fill up with dirty pages that we cannot evict, and we will
-                            // loop retrying the writebacks indefinitely.
-                            error!("writeback of buffer {:?} failed: {}", old_key, err);
-                            continue;
-                        }
-                    }
-
                    // remove mapping for old buffer
                    self.remove_mapping(old_key);
-                    inner.dirty = false;
                    inner.key = None;
                }
                return Ok((slot_idx, inner));
@@ -863,39 +793,19 @@ impl PageCache {
        }
    }

-    fn writeback(cache_key: &CacheKey, buf: &[u8]) -> Result<(), std::io::Error> {
-        match cache_key {
-            CacheKey::MaterializedPage {
-                hash_key: _,
-                lsn: _,
-            } => Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
-                "unexpected dirty materialized page",
-            )),
-            CacheKey::EphemeralPage { file_id, blkno } => {
-                writeback_ephemeral_file(*file_id, *blkno, buf)
-            }
-            CacheKey::ImmutableFilePage {
-                file_id: _,
-                blkno: _,
-            } => Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
-                "unexpected dirty immutable page",
-            )),
-        }
-    }
-
    /// Initialize a new page cache
    ///
    /// This should be called only once at page server startup.
    fn new(num_pages: usize) -> Self {
        assert!(num_pages > 0, "page cache size must be > 0");

+        // We could use Vec::leak here, but that potentially also leaks
+        // uninitialized reserved capacity. With into_boxed_slice and Box::leak
+        // this is avoided.
        let page_buffer = Box::leak(vec![0u8; num_pages * PAGE_SZ].into_boxed_slice());

        let size_metrics = &crate::metrics::PAGE_CACHE_SIZE;
        size_metrics.max_bytes.set_page_sz(num_pages);
-        size_metrics.current_bytes_ephemeral.set_page_sz(0);
        size_metrics.current_bytes_immutable.set_page_sz(0);
        size_metrics.current_bytes_materialized_page.set_page_sz(0);

@@ -905,11 +815,7 @@ impl PageCache {
                let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();

                Slot {
-                    inner: RwLock::new(SlotInner {
-                        key: None,
-                        buf,
-                        dirty: false,
-                    }),
+                    inner: tokio::sync::RwLock::new(SlotInner { key: None, buf }),
                    usage_count: AtomicU8::new(0),
                }
            })
@@ -917,7 +823,6 @@ impl PageCache {

        Self {
            materialized_page_map: Default::default(),
-            ephemeral_page_map: Default::default(),
            immutable_page_map: Default::default(),
            slots,
            next_evict_slot: AtomicUsize::new(0),
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -50,7 +50,8 @@ use crate::basebackup;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir::import_wal_from_tar;
-use crate::metrics::{LIVE_CONNECTIONS_COUNT, SMGR_QUERY_TIME};
+use crate::metrics;
+use crate::metrics::LIVE_CONNECTIONS_COUNT;
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant;
@@ -306,39 +307,6 @@ async fn page_service_conn_main(
    }
 }

-struct PageRequestMetrics {
-    get_rel_exists: metrics::Histogram,
-    get_rel_size: metrics::Histogram,
-    get_page_at_lsn: metrics::Histogram,
-    get_db_size: metrics::Histogram,
-}
-
-impl PageRequestMetrics {
-    fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
-        let tenant_id = tenant_id.to_string();
-        let timeline_id = timeline_id.to_string();
-
-        let get_rel_exists =
-            SMGR_QUERY_TIME.with_label_values(&["get_rel_exists", &tenant_id, &timeline_id]);
-
-        let get_rel_size =
-            SMGR_QUERY_TIME.with_label_values(&["get_rel_size", &tenant_id, &timeline_id]);
-
-        let get_page_at_lsn =
-            SMGR_QUERY_TIME.with_label_values(&["get_page_at_lsn", &tenant_id, &timeline_id]);
-
-        let get_db_size =
-            SMGR_QUERY_TIME.with_label_values(&["get_db_size", &tenant_id, &timeline_id]);
-
-        Self {
-            get_rel_exists,
-            get_rel_size,
-            get_page_at_lsn,
-            get_db_size,
-        }
-    }
-}
-
 struct PageServerHandler {
    _conf: &'static PageServerConf,
    broker_client: storage_broker::BrokerClientChannel,
@@ -406,7 +374,7 @@ impl PageServerHandler {
        pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
        pgb.flush().await?;

-        let metrics = PageRequestMetrics::new(&tenant_id, &timeline_id);
+        let metrics = metrics::SmgrQueryTimePerTimeline::new(&tenant_id, &timeline_id);

        loop {
            let msg = tokio::select! {
@@ -446,21 +414,21 @@ impl PageServerHandler {

            let response = match neon_fe_msg {
                PagestreamFeMessage::Exists(req) => {
-                    let _timer = metrics.get_rel_exists.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetRelExists);
                    self.handle_get_rel_exists_request(&timeline, &req, &ctx)
                        .await
                }
                PagestreamFeMessage::Nblocks(req) => {
-                    let _timer = metrics.get_rel_size.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetRelSize);
                    self.handle_get_nblocks_request(&timeline, &req, &ctx).await
                }
                PagestreamFeMessage::GetPage(req) => {
-                    let _timer = metrics.get_page_at_lsn.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetPageAtLsn);
                    self.handle_get_page_at_lsn_request(&timeline, &req, &ctx)
                        .await
                }
                PagestreamFeMessage::DbSize(req) => {
-                    let _timer = metrics.get_db_size.start_timer();
+                    let _timer = metrics.start_timer(metrics::SmgrQueryType::GetDbSize);
                    self.handle_db_size_request(&timeline, &req, &ctx).await
                }
            };
@@ -501,7 +469,9 @@ impl PageServerHandler {
        // Create empty timeline
        info!("creating new timeline");
        let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
-        let timeline = tenant.create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)?;
+        let timeline = tenant
+            .create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)
+            .await?;

        // TODO mark timeline as not ready until it reaches end_lsn.
        // We might have some wal to import as well, and we should prevent compute
@@ -984,8 +954,8 @@ where
                false
            };

-            metrics::metric_vec_duration::observe_async_block_duration_by_result(
-                &*crate::metrics::BASEBACKUP_QUERY_TIME,
+            ::metrics::metric_vec_duration::observe_async_block_duration_by_result(
+                &*metrics::BASEBACKUP_QUERY_TIME,
                async move {
                    self.handle_basebackup_request(
                        pgb,
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
--- a/pageserver/src/tenant/blob_io.rs
+++ b/pageserver/src/tenant/blob_io.rs
@@ -12,14 +12,12 @@
 //! len >= 128: 1XXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
 use crate::page_cache::PAGE_SZ;
-use crate::tenant::block_io::{BlockCursor, BlockReader};
+use crate::tenant::block_io::BlockCursor;
+use crate::virtual_file::VirtualFile;
 use std::cmp::min;
 use std::io::{Error, ErrorKind};

-impl<R> BlockCursor<R>
-where
-    R: BlockReader,
-{
+impl<'a> BlockCursor<'a> {
    /// Read a blob into a new buffer.
    pub async fn read_blob(&self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
        let mut buf = Vec::new();
@@ -36,7 +34,7 @@ where
        let mut blknum = (offset / PAGE_SZ as u64) as u32;
        let mut off = (offset % PAGE_SZ as u64) as usize;

-        let mut buf = self.read_blk(blknum)?;
+        let mut buf = self.read_blk(blknum).await?;

        // peek at the first byte, to determine if it's a 1- or 4-byte length
        let first_len_byte = buf[off];
@@ -52,7 +50,7 @@ where
                // it is split across two pages
                len_buf[..thislen].copy_from_slice(&buf[off..PAGE_SZ]);
                blknum += 1;
-                buf = self.read_blk(blknum)?;
+                buf = self.read_blk(blknum).await?;
                len_buf[thislen..].copy_from_slice(&buf[0..4 - thislen]);
                off = 4 - thislen;
            } else {
@@ -73,7 +71,7 @@ where
            if page_remain == 0 {
                // continue on next page
                blknum += 1;
-                buf = self.read_blk(blknum)?;
+                buf = self.read_blk(blknum).await?;
                off = 0;
                page_remain = PAGE_SZ;
            }
@@ -86,35 +84,24 @@ where
    }
 }

+/// A wrapper of `VirtualFile` that allows users to write blobs.
 ///
-/// Abstract trait for a data sink that you can write blobs to.
-///
-pub trait BlobWriter {
-    /// Write a blob of data. Returns the offset that it was written to,
-    /// which can be used to retrieve the data later.
-    fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error>;
-}
-
-///
-/// An implementation of BlobWriter to write blobs to anything that
-/// implements std::io::Write.
-///
-pub struct WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
-    inner: W,
+/// If a `BlobWriter` is dropped, the internal buffer will be
+/// discarded. You need to call [`flush_buffer`](Self::flush_buffer)
+/// manually before dropping.
+pub struct BlobWriter<const BUFFERED: bool> {
+    inner: VirtualFile,
    offset: u64,
+    /// A buffer to save on write calls, only used if BUFFERED=true
+    buf: Vec<u8>,
 }

-impl<W> WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
-    pub fn new(inner: W, start_offset: u64) -> Self {
-        WriteBlobWriter {
+impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
+    pub fn new(inner: VirtualFile, start_offset: u64) -> Self {
+        Self {
            inner,
            offset: start_offset,
+            buf: Vec::with_capacity(Self::CAPACITY),
        }
    }

@@ -122,28 +109,79 @@ where
        self.offset
    }

-    /// Access the underlying Write object.
-    ///
-    /// NOTE: WriteBlobWriter keeps track of the current write offset. If
-    /// you write something directly to the inner Write object, it makes the
-    /// internally tracked 'offset' to go out of sync. So don't do that.
-    pub fn into_inner(self) -> W {
-        self.inner
-    }
-}
+    const CAPACITY: usize = if BUFFERED { PAGE_SZ } else { 0 };

-impl<W> BlobWriter for WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
-    fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error> {
+    #[inline(always)]
+    /// Writes the given buffer directly to the underlying `VirtualFile`.
+    /// You need to make sure that the internal buffer is empty, otherwise
+    /// data will be written in wrong order.
+    async fn write_all_unbuffered(&mut self, src_buf: &[u8]) -> Result<(), Error> {
+        self.inner.write_all(src_buf).await?;
+        self.offset += src_buf.len() as u64;
+        Ok(())
+    }
+
+    #[inline(always)]
+    /// Flushes the internal buffer to the underlying `VirtualFile`.
+    pub async fn flush_buffer(&mut self) -> Result<(), Error> {
+        self.inner.write_all(&self.buf).await?;
+        self.buf.clear();
+        Ok(())
+    }
+
+    #[inline(always)]
+    /// Writes as much of `src_buf` into the internal buffer as it fits
+    fn write_into_buffer(&mut self, src_buf: &[u8]) -> usize {
+        let remaining = Self::CAPACITY - self.buf.len();
+        let to_copy = src_buf.len().min(remaining);
+        self.buf.extend_from_slice(&src_buf[..to_copy]);
+        self.offset += to_copy as u64;
+        to_copy
+    }
+
+    /// Internal, possibly buffered, write function
+    async fn write_all(&mut self, mut src_buf: &[u8]) -> Result<(), Error> {
+        if !BUFFERED {
+            assert!(self.buf.is_empty());
+            self.write_all_unbuffered(src_buf).await?;
+            return Ok(());
+        }
+        let remaining = Self::CAPACITY - self.buf.len();
+        // First try to copy as much as we can into the buffer
+        if remaining > 0 {
+            let copied = self.write_into_buffer(src_buf);
+            src_buf = &src_buf[copied..];
+        }
+        // Then, if the buffer is full, flush it out
+        if self.buf.len() == Self::CAPACITY {
+            self.flush_buffer().await?;
+        }
+        // Finally, write the tail of src_buf:
+        // If it wholly fits into the buffer without
+        // completely filling it, then put it there.
+        // If not, write it out directly.
+        if !src_buf.is_empty() {
+            assert_eq!(self.buf.len(), 0);
+            if src_buf.len() < Self::CAPACITY {
+                let copied = self.write_into_buffer(src_buf);
+                // We just verified above that src_buf fits into our internal buffer.
+                assert_eq!(copied, src_buf.len());
+            } else {
+                self.write_all_unbuffered(src_buf).await?;
+            }
+        }
+        Ok(())
+    }
+
+    /// Write a blob of data. Returns the offset that it was written to,
+    /// which can be used to retrieve the data later.
+    pub async fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error> {
        let offset = self.offset;

        if srcbuf.len() < 128 {
            // Short blob. Write a 1-byte length header
            let len_buf = srcbuf.len() as u8;
-            self.inner.write_all(&[len_buf])?;
-            self.offset += 1;
+            self.write_all(&[len_buf]).await?;
        } else {
            // Write a 4-byte length header
            if srcbuf.len() > 0x7fff_ffff {
@@ -154,11 +192,153 @@ where
            }
            let mut len_buf = ((srcbuf.len()) as u32).to_be_bytes();
            len_buf[0] |= 0x80;
-            self.inner.write_all(&len_buf)?;
-            self.offset += 4;
+            self.write_all(&len_buf).await?;
        }
-        self.inner.write_all(srcbuf)?;
-        self.offset += srcbuf.len() as u64;
+        self.write_all(srcbuf).await?;
        Ok(offset)
    }
 }
+
+impl BlobWriter<true> {
+    /// Access the underlying `VirtualFile`.
+    ///
+    /// This function flushes the internal buffer before giving access
+    /// to the underlying `VirtualFile`.
+    pub async fn into_inner(mut self) -> Result<VirtualFile, Error> {
+        self.flush_buffer().await?;
+        Ok(self.inner)
+    }
+
+    /// Access the underlying `VirtualFile`.
+    ///
+    /// Unlike [`into_inner`](Self::into_inner), this doesn't flush
+    /// the internal buffer before giving access.
+    pub fn into_inner_no_flush(self) -> VirtualFile {
+        self.inner
+    }
+}
+
+impl BlobWriter<false> {
+    /// Access the underlying `VirtualFile`.
+    pub fn into_inner(self) -> VirtualFile {
+        self.inner
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::tenant::block_io::BlockReaderRef;
+    use rand::{Rng, SeedableRng};
+
+    async fn round_trip_test<const BUFFERED: bool>(blobs: &[Vec<u8>]) -> Result<(), Error> {
+        let temp_dir = tempfile::tempdir()?;
+        let path = temp_dir.path().join("file");
+
+        // Write part (in block to drop the file)
+        let mut offsets = Vec::new();
+        {
+            let file = VirtualFile::create(&path).await?;
+            let mut wtr = BlobWriter::<BUFFERED>::new(file, 0);
+            for blob in blobs.iter() {
+                let offs = wtr.write_blob(blob).await?;
+                offsets.push(offs);
+            }
+            // Write out one page worth of zeros so that we can
+            // read again with read_blk
+            let offs = wtr.write_blob(&vec![0; PAGE_SZ]).await?;
+            println!("Writing final blob at offs={offs}");
+            wtr.flush_buffer().await?;
+        }
+
+        let file = VirtualFile::open(&path).await?;
+        let rdr = BlockReaderRef::VirtualFile(&file);
+        let rdr = BlockCursor::new(rdr);
+        for (idx, (blob, offset)) in blobs.iter().zip(offsets.iter()).enumerate() {
+            let blob_read = rdr.read_blob(*offset).await?;
+            assert_eq!(
+                blob, &blob_read,
+                "mismatch for idx={idx} at offset={offset}"
+            );
+        }
+        Ok(())
+    }
+
+    fn random_array(len: usize) -> Vec<u8> {
+        let mut rng = rand::thread_rng();
+        (0..len).map(|_| rng.gen()).collect::<_>()
+    }
+
+    #[tokio::test]
+    async fn test_one() -> Result<(), Error> {
+        let blobs = &[vec![12, 21, 22]];
+        round_trip_test::<false>(blobs).await?;
+        round_trip_test::<true>(blobs).await?;
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_hello_simple() -> Result<(), Error> {
+        let blobs = &[
+            vec![0, 1, 2, 3],
+            b"Hello, World!".to_vec(),
+            Vec::new(),
+            b"foobar".to_vec(),
+        ];
+        round_trip_test::<false>(blobs).await?;
+        round_trip_test::<true>(blobs).await?;
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_really_big_array() -> Result<(), Error> {
+        let blobs = &[
+            b"test".to_vec(),
+            random_array(10 * PAGE_SZ),
+            b"foobar".to_vec(),
+        ];
+        round_trip_test::<false>(blobs).await?;
+        round_trip_test::<true>(blobs).await?;
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_arrays_inc() -> Result<(), Error> {
+        let blobs = (0..PAGE_SZ / 8)
+            .map(|v| random_array(v * 16))
+            .collect::<Vec<_>>();
+        round_trip_test::<false>(&blobs).await?;
+        round_trip_test::<true>(&blobs).await?;
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_arrays_random_size() -> Result<(), Error> {
+        let mut rng = rand::rngs::StdRng::seed_from_u64(42);
+        let blobs = (0..1024)
+            .map(|_| {
+                let mut sz: u16 = rng.gen();
+                // Make 50% of the arrays small
+                if rng.gen() {
+                    sz |= 63;
+                }
+                random_array(sz.into())
+            })
+            .collect::<Vec<_>>();
+        round_trip_test::<false>(&blobs).await?;
+        round_trip_test::<true>(&blobs).await?;
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_arrays_page_boundary() -> Result<(), Error> {
+        let blobs = &[
+            random_array(PAGE_SZ - 4),
+            random_array(PAGE_SZ - 4),
+            random_array(PAGE_SZ - 4),
+        ];
+        round_trip_test::<false>(blobs).await?;
+        round_trip_test::<true>(blobs).await?;
+        Ok(())
+    }
+}
--- a/pageserver/src/tenant/block_io.rs
+++ b/pageserver/src/tenant/block_io.rs
@@ -2,11 +2,12 @@
 //! Low-level Block-oriented I/O functions
 //!

+use super::ephemeral_file::EphemeralFile;
+use super::storage_layer::delta_layer::{Adapter, DeltaLayerInner};
 use crate::page_cache::{self, PageReadGuard, ReadBufResult, PAGE_SZ};
+use crate::virtual_file::VirtualFile;
 use bytes::Bytes;
 use std::ops::{Deref, DerefMut};
-use std::os::unix::fs::FileExt;
-use std::sync::atomic::AtomicU64;

 /// This is implemented by anything that can read 8 kB (PAGE_SZ)
 /// blocks, using the page cache
@@ -14,68 +15,83 @@ use std::sync::atomic::AtomicU64;
 /// There are currently two implementations: EphemeralFile, and FileBlockReader
 /// below.
 pub trait BlockReader {
-    ///
-    /// Read a block. Returns a "lease" object that can be used to
-    /// access to the contents of the page. (For the page cache, the
-    /// lease object represents a lock on the buffer.)
-    ///
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error>;
-
    ///
    /// Create a new "cursor" for reading from this reader.
    ///
    /// A cursor caches the last accessed page, allowing for faster
    /// access if the same block is accessed repeatedly.
-    fn block_cursor(&self) -> BlockCursor<&Self>
-    where
-        Self: Sized,
-    {
-        BlockCursor::new(self)
-    }
+    fn block_cursor(&self) -> BlockCursor<'_>;
 }

 impl<B> BlockReader for &B
 where
    B: BlockReader,
 {
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        (*self).read_blk(blknum)
+    fn block_cursor(&self) -> BlockCursor<'_> {
+        (*self).block_cursor()
    }
 }

-/// A block accessible for reading
-///
-/// During builds with `#[cfg(test)]`, this is a proper enum
-/// with two variants to support testing code. During normal
-/// builds, it just has one variant and is thus a cheap newtype
-/// wrapper of [`PageReadGuard`]
-pub enum BlockLease {
+/// Reference to an in-memory copy of an immutable on-disk block.
+pub enum BlockLease<'a> {
    PageReadGuard(PageReadGuard<'static>),
+    EphemeralFileMutableTail(&'a [u8; PAGE_SZ]),
    #[cfg(test)]
-    Rc(std::rc::Rc<[u8; PAGE_SZ]>),
+    Arc(std::sync::Arc<[u8; PAGE_SZ]>),
 }

-impl From<PageReadGuard<'static>> for BlockLease {
-    fn from(value: PageReadGuard<'static>) -> Self {
+impl From<PageReadGuard<'static>> for BlockLease<'static> {
+    fn from(value: PageReadGuard<'static>) -> BlockLease<'static> {
        BlockLease::PageReadGuard(value)
    }
 }

 #[cfg(test)]
-impl From<std::rc::Rc<[u8; PAGE_SZ]>> for BlockLease {
-    fn from(value: std::rc::Rc<[u8; PAGE_SZ]>) -> Self {
-        BlockLease::Rc(value)
+impl<'a> From<std::sync::Arc<[u8; PAGE_SZ]>> for BlockLease<'a> {
+    fn from(value: std::sync::Arc<[u8; PAGE_SZ]>) -> Self {
+        BlockLease::Arc(value)
    }
 }

-impl Deref for BlockLease {
+impl<'a> Deref for BlockLease<'a> {
    type Target = [u8; PAGE_SZ];

    fn deref(&self) -> &Self::Target {
        match self {
            BlockLease::PageReadGuard(v) => v.deref(),
+            BlockLease::EphemeralFileMutableTail(v) => v,
            #[cfg(test)]
-            BlockLease::Rc(v) => v.deref(),
+            BlockLease::Arc(v) => v.deref(),
+        }
+    }
+}
+
+/// Provides the ability to read blocks from different sources,
+/// similar to using traits for this purpose.
+///
+/// Unlike traits, we also support the read function to be async though.
+pub(crate) enum BlockReaderRef<'a> {
+    FileBlockReader(&'a FileBlockReader),
+    EphemeralFile(&'a EphemeralFile),
+    Adapter(Adapter<&'a DeltaLayerInner>),
+    #[cfg(test)]
+    TestDisk(&'a super::disk_btree::tests::TestDisk),
+    #[cfg(test)]
+    VirtualFile(&'a VirtualFile),
+}
+
+impl<'a> BlockReaderRef<'a> {
+    #[inline(always)]
+    async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+        use BlockReaderRef::*;
+        match self {
+            FileBlockReader(r) => r.read_blk(blknum).await,
+            EphemeralFile(r) => r.read_blk(blknum).await,
+            Adapter(r) => r.read_blk(blknum).await,
+            #[cfg(test)]
+            TestDisk(r) => r.read_blk(blknum),
+            #[cfg(test)]
+            VirtualFile(r) => r.read_blk(blknum).await,
        }
    }
 }
@@ -89,7 +105,7 @@ impl Deref for BlockLease {
 ///
 /// ```no_run
 /// # use pageserver::tenant::block_io::{BlockReader, FileBlockReader};
-/// # let reader: FileBlockReader<std::fs::File> = unimplemented!("stub");
+/// # let reader: FileBlockReader = unimplemented!("stub");
 /// let cursor = reader.block_cursor();
 /// let buf = cursor.read_blk(1);
 /// // do stuff with 'buf'
@@ -97,65 +113,68 @@ impl Deref for BlockLease {
 /// // do stuff with 'buf'
 /// ```
 ///
-pub struct BlockCursor<R>
-where
-    R: BlockReader,
-{
-    reader: R,
+pub struct BlockCursor<'a> {
+    reader: BlockReaderRef<'a>,
 }

-impl<R> BlockCursor<R>
-where
-    R: BlockReader,
-{
-    pub fn new(reader: R) -> Self {
+impl<'a> BlockCursor<'a> {
+    pub(crate) fn new(reader: BlockReaderRef<'a>) -> Self {
        BlockCursor { reader }
    }
+    // Needed by cli
+    pub fn new_fileblockreader(reader: &'a FileBlockReader) -> Self {
+        BlockCursor {
+            reader: BlockReaderRef::FileBlockReader(reader),
+        }
+    }

-    pub fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        self.reader.read_blk(blknum)
+    /// Read a block.
+    ///
+    /// Returns a "lease" object that can be used to
+    /// access to the contents of the page. (For the page cache, the
+    /// lease object represents a lock on the buffer.)
+    #[inline(always)]
+    pub async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+        self.reader.read_blk(blknum).await
    }
 }
-static NEXT_ID: AtomicU64 = AtomicU64::new(1);

 /// An adapter for reading a (virtual) file using the page cache.
 ///
 /// The file is assumed to be immutable. This doesn't provide any functions
 /// for modifying the file, nor for invalidating the cache if it is modified.
-pub struct FileBlockReader<F> {
-    pub file: F,
+pub struct FileBlockReader {
+    pub file: VirtualFile,

    /// Unique ID of this file, used as key in the page cache.
-    file_id: u64,
+    file_id: page_cache::FileId,
 }

-impl<F> FileBlockReader<F>
-where
-    F: FileExt,
-{
-    pub fn new(file: F) -> Self {
-        let file_id = NEXT_ID.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+impl FileBlockReader {
+    pub fn new(file: VirtualFile) -> Self {
+        let file_id = page_cache::next_file_id();

        FileBlockReader { file_id, file }
    }

    /// Read a page from the underlying file into given buffer.
-    fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), std::io::Error> {
+    async fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), std::io::Error> {
        assert!(buf.len() == PAGE_SZ);
-        self.file.read_exact_at(buf, blkno as u64 * PAGE_SZ as u64)
+        self.file
+            .read_exact_at(buf, blkno as u64 * PAGE_SZ as u64)
+            .await
    }
-}
-
-impl<F> BlockReader for FileBlockReader<F>
-where
-    F: FileExt,
-{
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        // Look up the right page
+    /// Read a block.
+    ///
+    /// Returns a "lease" object that can be used to
+    /// access to the contents of the page. (For the page cache, the
+    /// lease object represents a lock on the buffer.)
+    pub async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
        let cache = page_cache::get();
        loop {
            match cache
                .read_immutable_buf(self.file_id, blknum)
+                .await
                .map_err(|e| {
                    std::io::Error::new(
                        std::io::ErrorKind::Other,
@@ -165,7 +184,7 @@ where
                ReadBufResult::Found(guard) => break Ok(guard.into()),
                ReadBufResult::NotFound(mut write_guard) => {
                    // Read the page from disk into the buffer
-                    self.fill_buffer(write_guard.deref_mut(), blknum)?;
+                    self.fill_buffer(write_guard.deref_mut(), blknum).await?;
                    write_guard.mark_valid();

                    // Swap for read lock
@@ -176,6 +195,12 @@ where
    }
 }

+impl BlockReader for FileBlockReader {
+    fn block_cursor(&self) -> BlockCursor<'_> {
+        BlockCursor::new(BlockReaderRef::FileBlockReader(self))
+    }
+}
+
 ///
 /// Trait for block-oriented output
 ///
--- a/pageserver/src/tenant/delete.rs
+++ b/pageserver/src/tenant/delete.rs
@@ -7,6 +7,7 @@ use anyhow::Context;
 use pageserver_api::models::TenantState;
 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 use tokio::sync::OwnedMutexGuard;
+use tokio_util::sync::CancellationToken;
 use tracing::{error, info, instrument, warn, Instrument, Span};

 use utils::{
@@ -82,6 +83,8 @@ async fn create_remote_delete_mark(
        FAILED_UPLOAD_WARN_THRESHOLD,
        FAILED_REMOTE_OP_RETRIES,
        "mark_upload",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
    )
    .await
    .context("mark_upload")?;
@@ -171,6 +174,8 @@ async fn remove_tenant_remote_delete_mark(
            FAILED_UPLOAD_WARN_THRESHOLD,
            FAILED_REMOTE_OP_RETRIES,
            "remove_tenant_remote_delete_mark",
+            // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+            backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
        )
        .await
        .context("remove_tenant_remote_delete_mark")?;
@@ -212,6 +217,19 @@ async fn cleanup_remaining_fs_traces(
        ))?
    });

+    // Make sure previous deletions are ordered before mark removal.
+    // Otherwise there is no guarantee that they reach the disk before mark deletion.
+    // So its possible for mark to reach disk first and for other deletions
+    // to be reordered later and thus missed if a crash occurs.
+    // Note that we dont need to sync after mark file is removed
+    // because we can tolerate the case when mark file reappears on startup.
+    let tenant_path = &conf.tenant_path(tenant_id);
+    if tenant_path.exists() {
+        crashsafe::fsync_async(&conf.tenant_path(tenant_id))
+            .await
+            .context("fsync_pre_mark_remove")?;
+    }
+
    rm(conf.tenant_deleted_mark_file_path(tenant_id), false).await?;

    fail::fail_point!("tenant-delete-before-remove-tenant-dir", |_| {
@@ -225,6 +243,32 @@ async fn cleanup_remaining_fs_traces(
    Ok(())
 }

+pub(crate) async fn remote_delete_mark_exists(
+    conf: &PageServerConf,
+    tenant_id: &TenantId,
+    remote_storage: &GenericRemoteStorage,
+) -> anyhow::Result<bool> {
+    // If remote storage is there we rely on it
+    let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_id).context("path")?;
+
+    let result = backoff::retry(
+        || async { remote_storage.download(&remote_mark_path).await },
+        |e| matches!(e, DownloadError::NotFound),
+        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
+        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
+        "fetch_tenant_deletion_mark",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
+    )
+    .await;
+
+    match result {
+        Ok(_) => Ok(true),
+        Err(DownloadError::NotFound) => Ok(false),
+        Err(e) => Err(anyhow::anyhow!(e)).context("remote_delete_mark_exists")?,
+    }
+}
+
 /// Orchestrates tenant shut down of all tasks, removes its in-memory structures,
 /// and deletes its data from both disk and s3.
 /// The sequence of steps:
@@ -238,8 +282,9 @@ async fn cleanup_remaining_fs_traces(
 /// It is resumable from any step in case a crash/restart occurs.
 /// There are three entrypoints to the process:
 /// 1. [`DeleteTenantFlow::run`] this is the main one called by a management api handler.
-/// 2. [`DeleteTenantFlow::resume`] is called during restarts when local or remote deletion marks are still there.
-/// Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
+/// 2. [`DeleteTenantFlow::resume_from_load`] is called during restarts when local or remote deletion marks are still there.
+/// 3. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
+///  Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
 #[derive(Default)]
 pub enum DeleteTenantFlow {
    #[default]
@@ -359,26 +404,14 @@ impl DeleteTenantFlow {
            None => return Ok(None),
        };

-        // If remote storage is there we rely on it
-        let remote_mark_path = remote_tenant_delete_mark_path(conf, &tenant_id)?;
-
-        let result = backoff::retry(
-            || async { remote_storage.download(&remote_mark_path).await },
-            |e| matches!(e, DownloadError::NotFound),
-            SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
-            SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
-            "fetch_tenant_deletion_mark",
-        )
-        .await;
-
-        match result {
-            Ok(_) => Ok(acquire(tenant)),
-            Err(DownloadError::NotFound) => Ok(None),
-            Err(e) => Err(anyhow::anyhow!(e)).context("should_resume_deletion")?,
+        if remote_delete_mark_exists(conf, &tenant_id, remote_storage).await? {
+            Ok(acquire(tenant))
+        } else {
+            Ok(None)
        }
    }

-    pub(crate) async fn resume(
+    pub(crate) async fn resume_from_load(
        guard: DeletionGuard,
        tenant: &Arc<Tenant>,
        init_order: Option<&InitializationOrder>,
@@ -388,7 +421,7 @@ impl DeleteTenantFlow {
        let (_, progress) = completion::channel();

        tenant
-            .set_stopping(progress, true)
+            .set_stopping(progress, true, false)
            .await
            .expect("cant be stopping or broken");

@@ -416,6 +449,31 @@ impl DeleteTenantFlow {
        .await
    }

+    pub(crate) async fn resume_from_attach(
+        guard: DeletionGuard,
+        tenant: &Arc<Tenant>,
+        tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        ctx: &RequestContext,
+    ) -> Result<(), DeleteTenantError> {
+        let (_, progress) = completion::channel();
+
+        tenant
+            .set_stopping(progress, false, true)
+            .await
+            .expect("cant be stopping or broken");
+
+        tenant.attach(ctx).await.context("attach")?;
+
+        Self::background(
+            guard,
+            tenant.conf,
+            tenant.remote_storage.clone(),
+            tenants,
+            tenant,
+        )
+        .await
+    }
+
    async fn prepare(
        tenants: &tokio::sync::RwLock<TenantsMap>,
        tenant_id: TenantId,
--- a/Show More
+++ b/Show More